I(z X,¬p ; z X ). Definition 3 (p-Specific Information). Given a phenomenon p ∈ P and an agent X with internal state z X ∈ Z X , the p-specific information z X,p is the object where I(z X,p ; p) > 0, I(z X,p ; p ′ ) = 0 ∀p ′ ∈ P, p ′ = p and the mutual information I(z X,p ; z X ) is maximised, i.e. there exists no z ′ X,p such that I(z ′ X,p ; p) > 0, I(z ′ X,p , p ′ ) = 0 ∀p ′ ∈ P, p ′ = p and I(z ′ X,p ; z X ) > I(z X,p ; z X ). Definition 4 (p-Background Information). Given a phenomenon p ∈ P and an agent X with internal state z X ∈ Z X and p-specific information z X,p , the p-background information2 z X, * ¬p is the object where I(z X,p ; z X, * ¬p ) = 0 and I(z X, * ¬p ; z X ) is maximised, i.e. there exists no z ′ X, * ¬p such that I(z X,p ; z ′ X, * ¬p ) = 0 and I(z ′ X, * ¬p ; z X ) > I(z X, * ¬p ; z X )."},{"figure_ref":[],"heading":"Information Compression","publication_ref":[],"table_ref":[],"text":"With these partitions we can define how compressed the p-relevant information is:\nDefinition 5 (p-Compression Factor). Suppose a phenomenon p ∈ P and an agent X. The p-compression factor c : X × P → (0, 1] is given as the ratio of the Kolmogorov complexity of the p-relevant information object to the size of the agent's encoding of that information:\nc(X, p) = K(z X, * p ) | z X, * p X | (5)"},{"figure_ref":[],"heading":"Information Utilisation","publication_ref":["b9"],"table_ref":[],"text":"Next, we will attempt to formalise the performance stance on understanding, i.e. we will try to define X's information utilisation of p. To do this, we will need to construct a set of 'fair tests of p' for X. We will start by noting: (1) A fair test for X should require X's background information;\n(2) a test of p should require information about p. We will use the formalisation of rational decision-making, AIXI, to 'benchmark' how information is utilised in a given environment. Unlike a typical test-taker, AIXI enters into an environment with no prior knowledge, and thus we must present any priors to AIXI as a part of its percept sequence. Therefore, to decide whether or not a given task meets the criteria outlined above we will construct a 'meta-task' for AIXI where relevant observations are prepended to the task.\nDefinition 6 ((X, p)-tests). Given a phenomenon p ∈ P and an agent X with internal state z X ∈ Z X , we start with the set of bounded-test environments E, we define the set of (X, p)-tests, E X,p , as follows:\nE X,p = µ ∈ E : V AIXI (p,b)•µ = V * µ > 0, V AIXI (p)•µ = V AIXI (b)•µ = V AIXI µ = 0 (6)\nWhere b is a shorthand for the p-background information b = z X, * ¬p , and x • µ denotes the construction of a new environment µ ′ such that:\n∀x i ∈ x, ∀a |x|, µ ′ (e k | a I(z X,¬p ; z X ). Definition 3 (p-Specific Information). Given a phenomenon p ∈ P and an agent X with internal state z X ∈ Z X , the p-specific information z X,p is the object where I(z X,p ; p) > 0, I(z X,p ; p ′ ) = 0 ∀p ′ ∈ P, p ′ = p and the mutual information I(z X,p ; z X ) is maximised, i.e. there exists no z ′ X,p such that I(z ′ X,p ; p) > 0, I(z ′ X,p , p ′ ) = 0 ∀p ′ ∈ P, p ′ = p and I(z ′ X,p ; z X ) > I(z X,p ; z X ). Definition 4 (p-Background Information). Given a phenomenon p ∈ P and an agent X with internal state z X ∈ Z X and p-specific information z X,p , the p-background information2 z X, * ¬p is the object where I(z X,p ; z X, * ¬p ) = 0 and I(z X, * ¬p ; z X ) is maximised, i.e. there exists no z ′ X, * ¬p such that I(z X,p ; z ′ X, * ¬p ) = 0 and I(z ′ X, * ¬p ; z X ) > I(z X, * ¬p ; z X ).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Information Compression\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"With these partitions we can define how compressed the p-relevant information is:\\nDefinition 5 (p-Compression Factor). Suppose a phenomenon p ∈ P and an agent X. The p-compression factor c : X × P → (0, 1] is given as the ratio of the Kolmogorov complexity of the p-relevant information object to the size of the agent's encoding of that information:\\nc(X, p) = K(z X, * p ) | z X, * p X | (5)\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Information Utilisation\",\n \"publication_ref\": [\n \"b9\"\n ],\n \"table_ref\": [],\n \"text\": \"Next, we will attempt to formalise the performance stance on understanding, i.e. we will try to define X's information utilisation of p. To do this, we will need to construct a set of 'fair tests of p' for X. We will start by noting: (1) A fair test for X should require X's background information;\\n(2) a test of p should require information about p. We will use the formalisation of rational decision-making, AIXI, to 'benchmark' how information is utilised in a given environment. Unlike a typical test-taker, AIXI enters into an environment with no prior knowledge, and thus we must present any priors to AIXI as a part of its percept sequence. Therefore, to decide whether or not a given task meets the criteria outlined above we will construct a 'meta-task' for AIXI where relevant observations are prepended to the task.\\nDefinition 6 ((X, p)-tests). Given a phenomenon p ∈ P and an agent X with internal state z X ∈ Z X , we start with the set of bounded-test environments E, we define the set of (X, p)-tests, E X,p , as follows:\\nE X,p = µ ∈ E : V AIXI (p,b)•µ = V * µ > 0, V AIXI (p)•µ = V AIXI (b)•µ = V AIXI µ = 0 (6)\\nWhere b is a shorthand for the p-background information b = z X, * ¬p , and x • µ denotes the construction of a new environment µ ′ such that:\\n∀x i ∈ x, ∀a |x|, µ ′ (e k | a 0, V AIXI (p)•µ = V AIXI (b)•µ = V AIXI µ = 0 (6)"},{"formula_coordinates":[6,192.96,388.5,287.68,26.04],"formula_id":"formula_5","formula_text":"∀x i ∈ x, ∀a |x|, µ ′ (e k | a 0, V AIXI (p)•µ = V AIXI (b)•µ = V AIXI µ = 0 (6)\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 192.96,\n 388.5,\n 287.68,\n 26.04\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"∀x i ∈ x, ∀a |x|, µ ′ (e k | a 0 where 0 is a zero vector, z is a variable that conditions the policy (e.g. observations and/or latent memory), and d τ is a distance metric over T i . Definition 2 (Positive Signalling). Let m = (m 0 , . . . , m T ) be a sequence of messages sent by an agent over the course of a trajectory of length T , and similarly for observations o = (o 0 , . . . , o T ), and actions a = (a 0 , . . . , a T ). An agent exhibits positive signalling if m is statistically dependent on either a or o."},{"figure_ref":[],"heading":"One-Way Communication Problem Formulation","publication_ref":["b8","b5","b6","b2","b17","b12"],"table_ref":[],"text":"We start to formulate the problem of cooperative language acquisition with the simplest case involving two agents: a speaker A and a listener B. For a particular interaction i the speaker emits a message m i ∈ A c * A that is received by the listener, and then the listener takes actions that lead it on a trajectory τ i ∈ T B in a Dec-POMDP sampled from a given domain. We are only considering the case of one-way communication with this set-up, but we will discuss two-way communication, i.e. dialogues, in Section 6. To make this more precise, we assume that A and B are agents sampled from a pool of players operating in an ad hoc team, where the domain D is a set of referential games, i.e. a class of Dec-POMDPs based on Lewis signalling games (Lewis, 1969;Lazaridou et al., 2017;Lee et al., 2018). We will assume that the listener is exhibiting positive listening to the messages sent by the speaker, and the speaker is positive signalling an 'intended' target trajectory2 , τ ⊙ i ∈ T B . We can denote this by saying that the observer observes:\nm i ∼ π A (m | τ ⊙ i ) and τ i ∼ π B (τ | m i )\n, where π A and π B denote the policies followed by each agent respectively.3 Before moving on, let us define a running example game to illustrate the setting. Suppose that the speaker has access to a shopping list and a map of the supermarket, and must write a note for the listener to observe, who then must retrieve the items as quickly as possible.\nThe cooperative language acquisition task is to construct an agent X, who we will call the observer, which is able to take on the roles of either the speaker or the listener and successfully communicate with others. So, if X is taking on the role of the speaker, given some τ that they intend for B to follow, they should emit a message that maximises the probability that B does so. If X is acting as the listener, and receives some m ∼ π A (m | τ ⊙ ) they should estimate τ ⊙ and follow this trajectory. With this, given a dataset of interactions between speakers and listeners, m i , τ i ∼ D AB , we can define the following sub-problems: Problem 1 (The Forward Problem (Signalling)). Find a function β(m | τ, θ β ) parameterised by θ β , which we call the Broca function, such that m maximises the probability that the listening agent B follows the trajectory τ upon receiving m, i.e.:\nθ * β = arg max θ β τ ∈TB E m∼β(m|τ,θ β ) π B (τ | m) (1)\nProblem 2 (The Backward Problem (Listening)). Find a function ν(τ | m, θ ν ) parameterised by θ ν , which we call the Wernicke function. Given the message m is from the speaking agent A and is intended to invoke the trajectory τ ⊙ , the function ν should maximise the probability of τ ⊙ .\nθ * ν = arg max θν τ ⊙ ∈TB E m∼πA(m|τ ⊙ ) ν(τ ⊙ | m, θ ν ) (2)\n4 Finding Broca and Wernicke\nFirstly, we can directly model the forward problem with the data available. We estimate the parameters θ β by mapping from observed trajectories to messages received by B:\nθ * β = arg min θ β mi,τi∈DAB d m (m i , mi )\nwhere mi = arg max\nm β(m | τ i , θ β )(3)\nGiven a distance function d m over messages. Put differently, we are aiming to find θ β such that the Broca function can produce the message that caused a given trajectory in the data.\nTo place this into our running example, we have data regarding the notes that were sent to the shopper (m i ), and paths through the shop that the shopper took (τ i ), and we are learning the relationship between notes and paths.\nHowever, the backward problem is much trickier given that we are never able to directly observe the intended trajectory τ ⊙ i for the message m i sent by the speaker. If we assume that the speaker is optimal, i.e. it always sends the perfect message to invoke the intended actions in B, then τ ⊙ i = τ i and thus we can optimise the reverse mapping as in the forward problem (i.e. messages to trajectories). But what can we do if we wish to relax this? Instead of modelling the speaker as perfectly optimal, we can assume 'soft-optimality', otherwise known as Boltzmann-rationality4 . We will do this in two parts: first, we will assume that given the target trajectory τ ⊙ , the speaker is more likely to send messages that are 'closer to optimal', for which we need some notion of semantic distance between messages. Secondly, we will assume that the speaker is more likely to pick target trajectories for the listener that yield a high return in the Dec-POMDP. Put formally: 4)\nP (τ ⊙ ) ∝ exp(V (τ ⊙ )) (\nP (m|τ ⊙ ) ∝ exp(-S B (m * B (τ ⊙ ), m))\n(5) Where, V is the expected return of a given trajectory, m * B (τ ⊙ ) is the optimal message to send to B to maximise the chance that B takes the trajectory τ ⊙ , and S B is a measure of the semantic distance between two messages for B. These latter two are defined as follows:\nm * B (τ ) = arg max m π B (τ |m) (6) S B (m 1 , m 2 ) = d τ π B (τ | m 1 ), π B (τ | m 2 ) (7)\nIn other words, the semantic distance is a function of the difference in actions that B takes (characterised by the distance function over trajectories d τ ) as a result of different messages. Thus, it is mathematically similar to Lowe et al.'s (2019) definition of positive listening, and philosophically close to the various approaches to 'meaning' that couple information and action (Haig, 2017;Wittgenstein, 1953;Peirce, 1878). Additionally, note that if we substitute these definitions into P (m|τ ⊙ ):\nP (m|τ ⊙ ) ∝ exp -d τ π B (τ | arg max m π B (τ ⊙ |m)), π B (τ | m) ∝ exp -d τ τ ⊙ , π B (τ | m) (8)\nThus the expression involving the semantic distance captures the intuition that the more optimal messages are the ones that, in expectation, lead to trajectories that are closer to the target. With our assumptions in place, we now insert this into the backward problem. For a given interaction m i , τ i ∼ D AB we can express most probable target trajectory τ ⊙ i with the maximum a posteriori estimate:\nτ ⊙ i = arg max τ ⊙ P (τ ⊙ | m i ) = arg max τ ⊙ P (τ ⊙ )P (m i | τ ⊙ ) = arg max τ ⊙ V (τ ⊙ ) -αd τ (τ ⊙ , π B (τ | m i )) = arg max τ ⊙ V (τ ⊙ ) -αd τ (τ ⊙ , τ i ) (9)\nWhere α is a hyperparameter controlling our prior on the relative optimality of the speakers ability to effectively communicate versus pick the optimal trajectory (similar to Jeon et al. ( 2020)). Therefore, to estimate the parameters of the Wernicke function θ ν :\nθ * ν = arg min θν mi,τi∈DAB (αd m (τ ⊙ i , τ i ) -V (τ ⊙ i ))\nwhere\nτ ⊙ i = arg max τ ν(τ | m i , θ ν )(10)\nAgain, let us contextualise this within the example of the shoppers. If the speaker's note contained roughly the right set of instructions, but is perhaps slightly confusing in a way that throws off the shopper (perhaps impossible directions), the Wernicke function will not emulate the shoppers confusion. Instead, the estimate of the intended trajectory can take into account the ambiguity or inconsistency and try and figure out what would be a successful path through the supermarket. Comparatively, when we assume optimality of the speaker, we are forced to conclude that the shoppers confusion was intended."},{"figure_ref":[],"heading":"Related Work","publication_ref":["b14","b10","b18","b0","b6","b5","b1"],"table_ref":[],"text":"A close area of related work is inverse reinforcement learning (Russell, 1998;Ng & Russell, 2000). Namely, the modelling of the speaker is similar to IRL, where instead of there being a hidden reward function influencing the agents' actions, there is a target trajectory for the listener. Further, the Boltzmann-rational model used is very similar to approaches in IRL (Zietbart et al., 2008;Finn et al., 2016). In the field of emergent communication, the works of Lee et al. (2018) and Lazaridou et al. (2017) both present frameworks for grounding learning agents in human natural language. They do this by using text annotated images rather than data from direct human communication in a cooperative setting. Finally, outside of AI, in the economics literature there has been work modelling how an uniformed listener may extract information from informed debaters (Glazer & Rubinstein, 2001). But because of markedly different assumptions, it does not tackle the problem of the current paper."},{"figure_ref":[],"heading":"Discussion and Conclusion","publication_ref":["b4","b7","b15"],"table_ref":[],"text":"In this paper we have presented the first steps towards constructing an agent that, given data regarding the interactions of language-users, can find the meaning behind messages received, as well as optimally convey a recommended trajectory to a listener. Yet, there are still several directions of further work to be explored.\nFirstly, how do we extend the system to dialogues, i.e. two-way communication? Potentially the system naturally captures dialogues as both agents can play the roles of speakers and listeners simultaneously (or interchangeably). For instance, suppose that in the supermarket example the speaker and the shopper held a phone call, and the shopper asks a question for clarification on directions. The shopper does not have an exact intended trajectory for the speaker's response, because if they knew this they would not need to ask the question. However, this does not necessarily pose a problem for the framework presented in this paper.\nAlthough we have referred to the target trajectory as \"intentional\" it is not necessary for a speaker to know the full details of the trajectory. This applies so long as they help the listener to find the closest trajectory that maximises reward, which they may do so by adding their own private information.\nSecondly, but no less critically, is the problem of empirically testing this framework by constructing an agent. There are several suitable test environments, for example, gridworld games that are similar to the supermarket example (Kajić et al., 2020;Leibo et al., 2017), or more communication focused problems such as the game of Taboo. In this game, one person has to get another to say a hidden word, but they are forbidden from revealing certain pieces of helpful information5 . Finally, this work could be extended from the passive case of observing data, to a situation where the learner is engaged with the language-users, perhaps for example with an active learning approach (Settles, 2009)."},{"figure_ref":[],"heading":"Acknowledgements","publication_ref":[],"table_ref":[],"text":"Work done by DC is thanks to the UKRI Centre for Doctoral Training in Safe and Trusted AI (EPSRC Project EP/S023356/1).\nWe would like to thank Francis Rhys Ward, Nandi Schoots, Richard Willis, Mattia Villani and Charles Higgins for their help."}],"string":"[\n {\n \"figure_ref\": [],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b11\",\n \"b16\"\n ],\n \"table_ref\": [],\n \"text\": \"Typically, in the field of emergent communication a group of agents learn to interact with one another through communication channels in order to facilitate coordination in a shared environment, i.e. a Dec-POMDP (Oliehoek & Amato, 2016). The agents learn highly effective communication strategies, but they tend to be brittle in the sense that they are unable to coordinate with agents that they have not encountered before. This construction does not naturally lend itself to systems that require a machine to communicate with a human, or enter within a community of humans using language to coordinate. In this paper, we frame this as a problem of cooperative language acquisition, where the goal is to adopt the language of a community of agents so as to coordinate with them.\\nMore precisely, we place the problem in the context of ad hoc team play (Stone et al., 2010). In ad hoc team play, we are given a set of competent1 agents and a domain of coordination tasks, and the problem is to design new agents that are capable of achieving success when playing with randomly sampled teammates. In our problem, we assume that there exists a community of language-users that define the pool of players who, by means of their shared language, are all successful ad hoc team players. Therefore, the cooperative language acquisition problem is defined as the case of designing a new agent to join this pool of players by observing a sample of interactions from the community.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Background\",\n \"publication_ref\": [\n \"b9\"\n ],\n \"table_ref\": [],\n \"text\": \"A decentralised partially-observable Markov decision process (Dec-POMDP) is described by a 7-tuple (S, {A i }, T, R, {Ω i }, O, γ), where S is a set of states, {A i } is a set of action sets, T is a transition function, R is a reward function, {Ω i } is a set of observation sets, O is an observation function, and γ is a discount factor. We will talk of trajectories for a given agent i, which are sequences of state-action-reward tuples τ ∈ T i = (S × A i × R) * . Each agent i follows a policy π i that maps an observation sequence to a distribution over actions. We denote the distribution over future trajectories that a policy induces as π(τ |.). The return of a trajectory is computed as the discounted sum of rewards:\\nV (τ ) = |τ | k=0 γ k r k\\nFollowing Lowe et al. (2019), we suppose that each agent's action sets can be expressed as\\nA i = A c i ∪ A e i ,\\nwhere A c i is a set of communicative actions and A e i is a set of environment actions. Communicative actions are sent to a target agent by a dedicated cheap-talk channel (there is no cost to communication), meaning they appear in the receiver's observation at the next time step. We also use Lowe et al.'s (2019) definitions of positive listening and positive signalling: Definition 1 (Positive Listening). An agent i with the policy π i exhibits positive listening if there exists a message generated by a signaller j, m ∈ A c j , such that d τ (π i (τ |z, 0), π i (τ |z, m)) > 0 where 0 is a zero vector, z is a variable that conditions the policy (e.g. observations and/or latent memory), and d τ is a distance metric over T i . Definition 2 (Positive Signalling). Let m = (m 0 , . . . , m T ) be a sequence of messages sent by an agent over the course of a trajectory of length T , and similarly for observations o = (o 0 , . . . , o T ), and actions a = (a 0 , . . . , a T ). An agent exhibits positive signalling if m is statistically dependent on either a or o.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"One-Way Communication Problem Formulation\",\n \"publication_ref\": [\n \"b8\",\n \"b5\",\n \"b6\",\n \"b2\",\n \"b17\",\n \"b12\"\n ],\n \"table_ref\": [],\n \"text\": \"We start to formulate the problem of cooperative language acquisition with the simplest case involving two agents: a speaker A and a listener B. For a particular interaction i the speaker emits a message m i ∈ A c * A that is received by the listener, and then the listener takes actions that lead it on a trajectory τ i ∈ T B in a Dec-POMDP sampled from a given domain. We are only considering the case of one-way communication with this set-up, but we will discuss two-way communication, i.e. dialogues, in Section 6. To make this more precise, we assume that A and B are agents sampled from a pool of players operating in an ad hoc team, where the domain D is a set of referential games, i.e. a class of Dec-POMDPs based on Lewis signalling games (Lewis, 1969;Lazaridou et al., 2017;Lee et al., 2018). We will assume that the listener is exhibiting positive listening to the messages sent by the speaker, and the speaker is positive signalling an 'intended' target trajectory2 , τ ⊙ i ∈ T B . We can denote this by saying that the observer observes:\\nm i ∼ π A (m | τ ⊙ i ) and τ i ∼ π B (τ | m i )\\n, where π A and π B denote the policies followed by each agent respectively.3 Before moving on, let us define a running example game to illustrate the setting. Suppose that the speaker has access to a shopping list and a map of the supermarket, and must write a note for the listener to observe, who then must retrieve the items as quickly as possible.\\nThe cooperative language acquisition task is to construct an agent X, who we will call the observer, which is able to take on the roles of either the speaker or the listener and successfully communicate with others. So, if X is taking on the role of the speaker, given some τ that they intend for B to follow, they should emit a message that maximises the probability that B does so. If X is acting as the listener, and receives some m ∼ π A (m | τ ⊙ ) they should estimate τ ⊙ and follow this trajectory. With this, given a dataset of interactions between speakers and listeners, m i , τ i ∼ D AB , we can define the following sub-problems: Problem 1 (The Forward Problem (Signalling)). Find a function β(m | τ, θ β ) parameterised by θ β , which we call the Broca function, such that m maximises the probability that the listening agent B follows the trajectory τ upon receiving m, i.e.:\\nθ * β = arg max θ β τ ∈TB E m∼β(m|τ,θ β ) π B (τ | m) (1)\\nProblem 2 (The Backward Problem (Listening)). Find a function ν(τ | m, θ ν ) parameterised by θ ν , which we call the Wernicke function. Given the message m is from the speaking agent A and is intended to invoke the trajectory τ ⊙ , the function ν should maximise the probability of τ ⊙ .\\nθ * ν = arg max θν τ ⊙ ∈TB E m∼πA(m|τ ⊙ ) ν(τ ⊙ | m, θ ν ) (2)\\n4 Finding Broca and Wernicke\\nFirstly, we can directly model the forward problem with the data available. We estimate the parameters θ β by mapping from observed trajectories to messages received by B:\\nθ * β = arg min θ β mi,τi∈DAB d m (m i , mi )\\nwhere mi = arg max\\nm β(m | τ i , θ β )(3)\\nGiven a distance function d m over messages. Put differently, we are aiming to find θ β such that the Broca function can produce the message that caused a given trajectory in the data.\\nTo place this into our running example, we have data regarding the notes that were sent to the shopper (m i ), and paths through the shop that the shopper took (τ i ), and we are learning the relationship between notes and paths.\\nHowever, the backward problem is much trickier given that we are never able to directly observe the intended trajectory τ ⊙ i for the message m i sent by the speaker. If we assume that the speaker is optimal, i.e. it always sends the perfect message to invoke the intended actions in B, then τ ⊙ i = τ i and thus we can optimise the reverse mapping as in the forward problem (i.e. messages to trajectories). But what can we do if we wish to relax this? Instead of modelling the speaker as perfectly optimal, we can assume 'soft-optimality', otherwise known as Boltzmann-rationality4 . We will do this in two parts: first, we will assume that given the target trajectory τ ⊙ , the speaker is more likely to send messages that are 'closer to optimal', for which we need some notion of semantic distance between messages. Secondly, we will assume that the speaker is more likely to pick target trajectories for the listener that yield a high return in the Dec-POMDP. Put formally: 4)\\nP (τ ⊙ ) ∝ exp(V (τ ⊙ )) (\\nP (m|τ ⊙ ) ∝ exp(-S B (m * B (τ ⊙ ), m))\\n(5) Where, V is the expected return of a given trajectory, m * B (τ ⊙ ) is the optimal message to send to B to maximise the chance that B takes the trajectory τ ⊙ , and S B is a measure of the semantic distance between two messages for B. These latter two are defined as follows:\\nm * B (τ ) = arg max m π B (τ |m) (6) S B (m 1 , m 2 ) = d τ π B (τ | m 1 ), π B (τ | m 2 ) (7)\\nIn other words, the semantic distance is a function of the difference in actions that B takes (characterised by the distance function over trajectories d τ ) as a result of different messages. Thus, it is mathematically similar to Lowe et al.'s (2019) definition of positive listening, and philosophically close to the various approaches to 'meaning' that couple information and action (Haig, 2017;Wittgenstein, 1953;Peirce, 1878). Additionally, note that if we substitute these definitions into P (m|τ ⊙ ):\\nP (m|τ ⊙ ) ∝ exp -d τ π B (τ | arg max m π B (τ ⊙ |m)), π B (τ | m) ∝ exp -d τ τ ⊙ , π B (τ | m) (8)\\nThus the expression involving the semantic distance captures the intuition that the more optimal messages are the ones that, in expectation, lead to trajectories that are closer to the target. With our assumptions in place, we now insert this into the backward problem. For a given interaction m i , τ i ∼ D AB we can express most probable target trajectory τ ⊙ i with the maximum a posteriori estimate:\\nτ ⊙ i = arg max τ ⊙ P (τ ⊙ | m i ) = arg max τ ⊙ P (τ ⊙ )P (m i | τ ⊙ ) = arg max τ ⊙ V (τ ⊙ ) -αd τ (τ ⊙ , π B (τ | m i )) = arg max τ ⊙ V (τ ⊙ ) -αd τ (τ ⊙ , τ i ) (9)\\nWhere α is a hyperparameter controlling our prior on the relative optimality of the speakers ability to effectively communicate versus pick the optimal trajectory (similar to Jeon et al. ( 2020)). Therefore, to estimate the parameters of the Wernicke function θ ν :\\nθ * ν = arg min θν mi,τi∈DAB (αd m (τ ⊙ i , τ i ) -V (τ ⊙ i ))\\nwhere\\nτ ⊙ i = arg max τ ν(τ | m i , θ ν )(10)\\nAgain, let us contextualise this within the example of the shoppers. If the speaker's note contained roughly the right set of instructions, but is perhaps slightly confusing in a way that throws off the shopper (perhaps impossible directions), the Wernicke function will not emulate the shoppers confusion. Instead, the estimate of the intended trajectory can take into account the ambiguity or inconsistency and try and figure out what would be a successful path through the supermarket. Comparatively, when we assume optimality of the speaker, we are forced to conclude that the shoppers confusion was intended.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Work\",\n \"publication_ref\": [\n \"b14\",\n \"b10\",\n \"b18\",\n \"b0\",\n \"b6\",\n \"b5\",\n \"b1\"\n ],\n \"table_ref\": [],\n \"text\": \"A close area of related work is inverse reinforcement learning (Russell, 1998;Ng & Russell, 2000). Namely, the modelling of the speaker is similar to IRL, where instead of there being a hidden reward function influencing the agents' actions, there is a target trajectory for the listener. Further, the Boltzmann-rational model used is very similar to approaches in IRL (Zietbart et al., 2008;Finn et al., 2016). In the field of emergent communication, the works of Lee et al. (2018) and Lazaridou et al. (2017) both present frameworks for grounding learning agents in human natural language. They do this by using text annotated images rather than data from direct human communication in a cooperative setting. Finally, outside of AI, in the economics literature there has been work modelling how an uniformed listener may extract information from informed debaters (Glazer & Rubinstein, 2001). But because of markedly different assumptions, it does not tackle the problem of the current paper.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Discussion and Conclusion\",\n \"publication_ref\": [\n \"b4\",\n \"b7\",\n \"b15\"\n ],\n \"table_ref\": [],\n \"text\": \"In this paper we have presented the first steps towards constructing an agent that, given data regarding the interactions of language-users, can find the meaning behind messages received, as well as optimally convey a recommended trajectory to a listener. Yet, there are still several directions of further work to be explored.\\nFirstly, how do we extend the system to dialogues, i.e. two-way communication? Potentially the system naturally captures dialogues as both agents can play the roles of speakers and listeners simultaneously (or interchangeably). For instance, suppose that in the supermarket example the speaker and the shopper held a phone call, and the shopper asks a question for clarification on directions. The shopper does not have an exact intended trajectory for the speaker's response, because if they knew this they would not need to ask the question. However, this does not necessarily pose a problem for the framework presented in this paper.\\nAlthough we have referred to the target trajectory as \\\"intentional\\\" it is not necessary for a speaker to know the full details of the trajectory. This applies so long as they help the listener to find the closest trajectory that maximises reward, which they may do so by adding their own private information.\\nSecondly, but no less critically, is the problem of empirically testing this framework by constructing an agent. There are several suitable test environments, for example, gridworld games that are similar to the supermarket example (Kajić et al., 2020;Leibo et al., 2017), or more communication focused problems such as the game of Taboo. In this game, one person has to get another to say a hidden word, but they are forbidden from revealing certain pieces of helpful information5 . Finally, this work could be extended from the passive case of observing data, to a situation where the learner is engaged with the language-users, perhaps for example with an active learning approach (Settles, 2009).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Acknowledgements\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Work done by DC is thanks to the UKRI Centre for Doctoral Training in Safe and Trusted AI (EPSRC Project EP/S023356/1).\\nWe would like to thank Francis Rhys Ward, Nandi Schoots, Richard Willis, Mattia Villani and Charles Higgins for their help.\"\n }\n]"},"abstract":{"kind":"string","value":"In this paper, we propose and consider the problem of cooperative language acquisition as a particular form of the ad hoc team play problem. We then present a probabilistic model for inferring a speaker's intentions and a listener's semantics from observing communications between a team of language-users. This model builds on the assumptions that speakers are engaged in positive signalling and listeners are exhibiting positive listening, which is to say the messages convey hidden information from the listener, that then causes them to change their behaviour. Further, it accounts for potential sub-optimality in the speaker's ability to convey the right information (according to the given task). Finally, we discuss further work for testing and developing this framework."},"title":{"kind":"string","value":"Joining the Conversation: Towards Language Acquisition for Ad Hoc Team Play"},"figures":{"kind":"list like","value":[],"string":"[]"},"authors":{"kind":"string","value":"Dylan R Cope; Peter Mcburney"},"references":{"kind":"list like","value":[{"authors":"Chelsea Finn; Sergey Levine; Pieter Abbeel","journal":"","ref_id":"b0","title":"Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization","year":"2016"},{"authors":"Jacob Glazer; Ariel Rubinstein","journal":"Games and Economic Behavior","ref_id":"b1","title":"Debates and Decisions: On a Rationale of Argumentation Rules","year":"2001"},{"authors":"David Haig","journal":"","ref_id":"b2","title":"Making Sense: Information Interpreted as Meaning","year":"2017"},{"authors":"Smitha Hong Jun Jeon; Anca Milli; Dragan","journal":"","ref_id":"b3","title":"Reward-rational (implicit) choice: A unifying formalism for reward learning","year":"2020"},{"authors":"Ivana Kajić; Eser Aygün; Aygün Aygün; Doina Precup","journal":"","ref_id":"b4","title":"Learning to cooperate: Emergent communication in multi-agent navigation","year":"2020"},{"authors":"Angeliki Lazaridou; Alexander Peysakhovich; Marco Baroni","journal":"","ref_id":"b5","title":"Multi-Agent Cooperation and the Emergence of (Natural) Language","year":"2017"},{"authors":"Jason Lee; Kyunghyun Cho; Jason Weston; Douwe Kiela","journal":"","ref_id":"b6","title":"Emergent Translation in Multi-Agent Communication","year":"2018"},{"authors":"Vinicius Joel Z Leibo; Marc Zambaldi; Janusz Lanctot; Thore Marecki; Graepel","journal":"","ref_id":"b7","title":"Multiagent Reinforcement Learning in Sequential Social Dilemmas","year":"2017"},{"authors":"David K Lewis","journal":"Wiley-Blackwell","ref_id":"b8","title":"Convention: A Philosophical Study","year":"1969"},{"authors":"Ryan Lowe; Jakob Foerster; Y-Lan Boureau; Joelle Pineau; Yann Dauphin","journal":"","ref_id":"b9","title":"On the Pitfalls of Measuring Emergent Communication","year":"2019"},{"authors":"Andrew Y Ng; Stuart Russell","journal":"","ref_id":"b10","title":"Algorithms for Inverse Reinforcement Learning","year":"2000"},{"authors":"Frans A Oliehoek; Christopher Amato","journal":"Springer International Publishing","ref_id":"b11","title":"A Concise Introduction to Decentralized POMDPs","year":"2016"},{"authors":"Peirce Charles","journal":"Popular Science Monthly","ref_id":"b12","title":"How to Make Our Ideas Clear","year":"1878"},{"authors":"Dagmar Michael Rovatsos; Gábor Gromann; Bella","journal":"","ref_id":"b13","title":"The Taboo Challenge Competition","year":"2017"},{"authors":"Stuart Russell","journal":"ACM Press","ref_id":"b14","title":"Learning agents for uncertain environments (extended abstract)","year":"1998"},{"authors":"Burr Settles","journal":"","ref_id":"b15","title":"Active Learning Literature Survey","year":"2009"},{"authors":"Peter Stone; A Gal; Sarit Kaminka; Jeffrey S Kraus; Rosenschein","journal":"","ref_id":"b16","title":"Ad Hoc Autonomous Agent Teams: Collaboration without Pre-Coordination","year":"2010"},{"authors":"Ludwig Wittgenstein","journal":"Macmillan Publishers","ref_id":"b17","title":"Philosophical Investigations","year":"1953"},{"authors":"Brian D Zietbart; Andrew Maas; Andrew Bagnell; Anind K Dey","journal":"","ref_id":"b18","title":"Maximum Entropy Inverse Reinforcement Learning","year":"2008"}],"string":"[\n {\n \"authors\": \"Chelsea Finn; Sergey Levine; Pieter Abbeel\",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Jacob Glazer; Ariel Rubinstein\",\n \"journal\": \"Games and Economic Behavior\",\n \"ref_id\": \"b1\",\n \"title\": \"Debates and Decisions: On a Rationale of Argumentation Rules\",\n \"year\": \"2001\"\n },\n {\n \"authors\": \"David Haig\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"Making Sense: Information Interpreted as Meaning\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Smitha Hong Jun Jeon; Anca Milli; Dragan\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"Reward-rational (implicit) choice: A unifying formalism for reward learning\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Ivana Kajić; Eser Aygün; Aygün Aygün; Doina Precup\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"Learning to cooperate: Emergent communication in multi-agent navigation\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Angeliki Lazaridou; Alexander Peysakhovich; Marco Baroni\",\n \"journal\": \"\",\n \"ref_id\": \"b5\",\n \"title\": \"Multi-Agent Cooperation and the Emergence of (Natural) Language\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Jason Lee; Kyunghyun Cho; Jason Weston; Douwe Kiela\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"Emergent Translation in Multi-Agent Communication\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Vinicius Joel Z Leibo; Marc Zambaldi; Janusz Lanctot; Thore Marecki; Graepel\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"Multiagent Reinforcement Learning in Sequential Social Dilemmas\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"David K Lewis\",\n \"journal\": \"Wiley-Blackwell\",\n \"ref_id\": \"b8\",\n \"title\": \"Convention: A Philosophical Study\",\n \"year\": \"1969\"\n },\n {\n \"authors\": \"Ryan Lowe; Jakob Foerster; Y-Lan Boureau; Joelle Pineau; Yann Dauphin\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"On the Pitfalls of Measuring Emergent Communication\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Andrew Y Ng; Stuart Russell\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"Algorithms for Inverse Reinforcement Learning\",\n \"year\": \"2000\"\n },\n {\n \"authors\": \"Frans A Oliehoek; Christopher Amato\",\n \"journal\": \"Springer International Publishing\",\n \"ref_id\": \"b11\",\n \"title\": \"A Concise Introduction to Decentralized POMDPs\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Peirce Charles\",\n \"journal\": \"Popular Science Monthly\",\n \"ref_id\": \"b12\",\n \"title\": \"How to Make Our Ideas Clear\",\n \"year\": \"1878\"\n },\n {\n \"authors\": \"Dagmar Michael Rovatsos; Gábor Gromann; Bella\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"The Taboo Challenge Competition\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Stuart Russell\",\n \"journal\": \"ACM Press\",\n \"ref_id\": \"b14\",\n \"title\": \"Learning agents for uncertain environments (extended abstract)\",\n \"year\": \"1998\"\n },\n {\n \"authors\": \"Burr Settles\",\n \"journal\": \"\",\n \"ref_id\": \"b15\",\n \"title\": \"Active Learning Literature Survey\",\n \"year\": \"2009\"\n },\n {\n \"authors\": \"Peter Stone; A Gal; Sarit Kaminka; Jeffrey S Kraus; Rosenschein\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"Ad Hoc Autonomous Agent Teams: Collaboration without Pre-Coordination\",\n \"year\": \"2010\"\n },\n {\n \"authors\": \"Ludwig Wittgenstein\",\n \"journal\": \"Macmillan Publishers\",\n \"ref_id\": \"b17\",\n \"title\": \"Philosophical Investigations\",\n \"year\": \"1953\"\n },\n {\n \"authors\": \"Brian D Zietbart; Andrew Maas; Andrew Bagnell; Anind K Dey\",\n \"journal\": \"\",\n \"ref_id\": \"b18\",\n \"title\": \"Maximum Entropy Inverse Reinforcement Learning\",\n \"year\": \"2008\"\n }\n]"},"formulas":{"kind":"list like","value":[{"formula_coordinates":[1,382.08,691.37,81.03,14.81],"formula_id":"formula_0","formula_text":"V (τ ) = |τ | k=0 γ k r k"},{"formula_coordinates":[2,108,94.15,63.73,12.63],"formula_id":"formula_1","formula_text":"A i = A c i ∪ A e i ,"},{"formula_coordinates":[2,338.4,404.21,162.31,13.73],"formula_id":"formula_2","formula_text":"m i ∼ π A (m | τ ⊙ i ) and τ i ∼ π B (τ | m i )"},{"formula_coordinates":[2,210.72,589.37,293.32,23.05],"formula_id":"formula_3","formula_text":"θ * β = arg max θ β τ ∈TB E m∼β(m|τ,θ β ) π B (τ | m) (1)"},{"formula_coordinates":[2,200.64,665.33,303.4,23.53],"formula_id":"formula_4","formula_text":"θ * ν = arg max θν τ ⊙ ∈TB E m∼πA(m|τ ⊙ ) ν(τ ⊙ | m, θ ν ) (2)"},{"formula_coordinates":[3,246.36,133.01,150.15,22.73],"formula_id":"formula_5","formula_text":"θ * β = arg min θ β mi,τi∈DAB d m (m i , mi )"},{"formula_coordinates":[3,283.32,149.61,220.72,27.97],"formula_id":"formula_6","formula_text":"m β(m | τ i , θ β )(3)"},{"formula_coordinates":[3,239.76,368.33,255.77,11.8],"formula_id":"formula_7","formula_text":"P (τ ⊙ ) ∝ exp(V (τ ⊙ )) ("},{"formula_coordinates":[3,228.24,384.05,155.44,13.25],"formula_id":"formula_8","formula_text":"P (m|τ ⊙ ) ∝ exp(-S B (m * B (τ ⊙ ), m))"},{"formula_coordinates":[3,213.6,434.45,290.44,33.85],"formula_id":"formula_9","formula_text":"m * B (τ ) = arg max m π B (τ |m) (6) S B (m 1 , m 2 ) = d τ π B (τ | m 1 ), π B (τ | m 2 ) (7)"},{"formula_coordinates":[3,167.04,541.73,337,36.65],"formula_id":"formula_10","formula_text":"P (m|τ ⊙ ) ∝ exp -d τ π B (τ | arg max m π B (τ ⊙ |m)), π B (τ | m) ∝ exp -d τ τ ⊙ , π B (τ | m) (8)"},{"formula_coordinates":[3,188.04,640.97,316,64.49],"formula_id":"formula_11","formula_text":"τ ⊙ i = arg max τ ⊙ P (τ ⊙ | m i ) = arg max τ ⊙ P (τ ⊙ )P (m i | τ ⊙ ) = arg max τ ⊙ V (τ ⊙ ) -αd τ (τ ⊙ , π B (τ | m i )) = arg max τ ⊙ V (τ ⊙ ) -αd τ (τ ⊙ , τ i ) (9)"},{"formula_coordinates":[4,222.6,121.13,198.16,22.85],"formula_id":"formula_12","formula_text":"θ * ν = arg min θν mi,τi∈DAB (αd m (τ ⊙ i , τ i ) -V (τ ⊙ i ))"},{"formula_coordinates":[4,220.56,138.21,283.48,28.57],"formula_id":"formula_13","formula_text":"τ ⊙ i = arg max τ ν(τ | m i , θ ν )(10)"}],"string":"[\n {\n \"formula_coordinates\": [\n 1,\n 382.08,\n 691.37,\n 81.03,\n 14.81\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"V (τ ) = |τ | k=0 γ k r k\"\n },\n {\n \"formula_coordinates\": [\n 2,\n 108,\n 94.15,\n 63.73,\n 12.63\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"A i = A c i ∪ A e i ,\"\n },\n {\n \"formula_coordinates\": [\n 2,\n 338.4,\n 404.21,\n 162.31,\n 13.73\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"m i ∼ π A (m | τ ⊙ i ) and τ i ∼ π B (τ | m i )\"\n },\n {\n \"formula_coordinates\": [\n 2,\n 210.72,\n 589.37,\n 293.32,\n 23.05\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"θ * β = arg max θ β τ ∈TB E m∼β(m|τ,θ β ) π B (τ | m) (1)\"\n },\n {\n \"formula_coordinates\": [\n 2,\n 200.64,\n 665.33,\n 303.4,\n 23.53\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"θ * ν = arg max θν τ ⊙ ∈TB E m∼πA(m|τ ⊙ ) ν(τ ⊙ | m, θ ν ) (2)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 246.36,\n 133.01,\n 150.15,\n 22.73\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"θ * β = arg min θ β mi,τi∈DAB d m (m i , mi )\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 283.32,\n 149.61,\n 220.72,\n 27.97\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"m β(m | τ i , θ β )(3)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 239.76,\n 368.33,\n 255.77,\n 11.8\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"P (τ ⊙ ) ∝ exp(V (τ ⊙ )) (\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 228.24,\n 384.05,\n 155.44,\n 13.25\n ],\n \"formula_id\": \"formula_8\",\n \"formula_text\": \"P (m|τ ⊙ ) ∝ exp(-S B (m * B (τ ⊙ ), m))\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 213.6,\n 434.45,\n 290.44,\n 33.85\n ],\n \"formula_id\": \"formula_9\",\n \"formula_text\": \"m * B (τ ) = arg max m π B (τ |m) (6) S B (m 1 , m 2 ) = d τ π B (τ | m 1 ), π B (τ | m 2 ) (7)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 167.04,\n 541.73,\n 337,\n 36.65\n ],\n \"formula_id\": \"formula_10\",\n \"formula_text\": \"P (m|τ ⊙ ) ∝ exp -d τ π B (τ | arg max m π B (τ ⊙ |m)), π B (τ | m) ∝ exp -d τ τ ⊙ , π B (τ | m) (8)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 188.04,\n 640.97,\n 316,\n 64.49\n ],\n \"formula_id\": \"formula_11\",\n \"formula_text\": \"τ ⊙ i = arg max τ ⊙ P (τ ⊙ | m i ) = arg max τ ⊙ P (τ ⊙ )P (m i | τ ⊙ ) = arg max τ ⊙ V (τ ⊙ ) -αd τ (τ ⊙ , π B (τ | m i )) = arg max τ ⊙ V (τ ⊙ ) -αd τ (τ ⊙ , τ i ) (9)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 222.6,\n 121.13,\n 198.16,\n 22.85\n ],\n \"formula_id\": \"formula_12\",\n \"formula_text\": \"θ * ν = arg min θν mi,τi∈DAB (αd m (τ ⊙ i , τ i ) -V (τ ⊙ i ))\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 220.56,\n 138.21,\n 283.48,\n 28.57\n ],\n \"formula_id\": \"formula_13\",\n \"formula_text\": \"τ ⊙ i = arg max τ ν(τ | m i , θ ν )(10)\"\n }\n]"}}},{"rowIdx":381,"cells":{"doi":{"kind":"string","value":""},"pub_date":{"kind":"string","value":"2023-05-20"},"sections":{"kind":"list like","value":[{"figure_ref":[],"heading":"I. INTRODUCTION","publication_ref":["b0","b2","b3","b4","b5","b7","b8","b9","b10","b11","b12","b14","b15","b18","b19","b22","b15","b20","b23","b24","b25","b26","b28"],"table_ref":[],"text":"High Dynamic Range Imaging (HDRI) [1]- [3], encompassing comprehensive scene content with optimal exposure, has gained considerable attention in recent years. As a pivotal technique for computer vision, HDRI not only delivers visually appealing observations in harmony with the human visual system, but also integrates essential features for a range of downstream vision applications, including object detection [4], [5], visual enhancement [6]- [8] and semantic segmentation [9], [10]. Unfortunately, due to constraints in photography Fig. 1. Visual comparison with representative MEF methods on general and misaligned fusion scenarios. The left figure plots the ranking with these competitors under general and misaligned scenarios. Other figures shows the obvious comparison with patch-based scheme MEFSPD [11] and learningbased method MEFGAN [12].\nequipment (e.g., smartphones and single-lens reflex cameras) that capture images with limited dynamic ranges, these images experience varying degrees of luminance degradation, leading to corrupted over/under-exposed regions. As a result, Low Dynamic Range (LDR) images are plagued by color distortion and loss of detail, hindering the accurate portrayal of complete natural scenes. Consequently, the generation of well-exposed HDR images remains both a challenging and significant research topic.\nRecently, a growing number of researchers have endeavored to create cutting-edge HDRI hardware devices capable of producing an extensive range of illumination, thereby addressing the limitations inherent to traditional digital cameras [13]- [15]. Nevertheless, due to elevated production costs and suboptimal efficiency, these intricately designed devices face challenges in achieving widespread adoption in real-world applications. As an alternative, Multi-Exposure Fusion (MEF) offers an efficacious solution for generating HDR images by assimilating characteristic texture information from a collection of LDR images captured under diverse exposures. This strategy adeptly bypasses hardware-specific limitations while maintaining a lower computational cost. Within the existing literature, conventional frameworks [16]- [19] and learning-based ones [20]- [23] constitute the predominant categories of MEF techniques. Despite these advancements, MEF continues to grapple with certain hurdles that impede its overall efficacy.\nIt is imperative to highlight that current learning-based methods neglect the essential adaptive preservation in multi-exposure image fusion. Specifically, a variety of approaches utilize direct fusion rules for feature aggregation, such as summation [16] and multiplication [21]. Regrettably, these rudimentary fusion rules fall short in effectively aggregating information from LDR pairs with markedly distinct characteristics, thereby failing to preserve critical information (e.g., pixel intensity and texture details) appropriately. Furthermore, given the considerable variation in multi-exposure image distributions, manually designed architectures face difficulties in flexibly adapting to disparate data distributions. In addition, due to the unavoidable movements and shaking of imaging devices, minor pixel misalignments in LDR pairs are commonplace. Existing methods seldom tackle this issue, leading to fused images characterized by blurred details and compromised structure. In essence, there is an urgent demand for an all-encompassing, robust, and efficient learning approach that not only delivers promising visual realism enhancement but also ensures high efficiency and stability across a wide array of scenes.\nTo be more specific, the second challenge stems from the computational efficiency of existing methods. Present MEF methods, encompassing both conventional and learningbased approaches, rely heavily on handcrafted architectures and operations. In terms of conventional frameworks, various transformations such as wavelet transform [24], multi-scale representation [25], and Laplacian pyramid [26] are proposed to enable feature fusion through handcrafted mechanisms. Unfortunately, these manual designs for feature extraction and fusion rules demand substantial fine-tuning and a significant amount of experiential knowledge. Most of these methods utilize numerical optimization, which in turn impacts the inference efficiency and robustness in real-world applications. Furthermore, in recent years, the powerful feature extraction capabilities of CNN-based learning have led to the increasing dominance of end-to-end models in MEF development, considerably improving performance concerning statistical metrics and visual effects. For architectural construction, a variety of learnable mechanisms [27]- [29] have been introduced to forge connections between LDR pairs and HDR outputs. We contend that current neural architectures for MEF largely borrow effective practices from other vision tasks without paying adequate attention to MEF-specific characteristics. As a consequence, these simplistic cascaded architectures, marked by increased width and depth, possess an excessive number of parameters, making them prone to feature redundancy."},{"figure_ref":[],"heading":"A. Contributions","publication_ref":[],"table_ref":[],"text":"To partially overcome these limitations, we present a versatile architecture search-based multi-exposure fusion approach. Specifically, we first develop a fusion-centric hyper-architecture, adhering to two primary principles: selfalignment and detail repletion. We initially propose scene relighting to map source images onto the same illumination, enhancing over/under-exposed details and providing improved support for subsequent feature aggregation. We introduce deformable alignment to achieve accurate feature registration, minimizing artifacts and blurring. Subsequently, we propose the detail repletion module to refine the coarse fusion results, leading to richer texture details. Next, we establish a flexible search space encompassing more effective operations, enabling the selection of an optimal network architecture. Considering hardware latency constraints, we employ a differentiable architecture search to automatically discover a compact and efficient model for image fusion. Consequently, our method effectively produces vivid colors, abundant details, and ghostingfree results, as intuitively illustrated in Fig. 1. Our core contributions can be summarized as follows:\n• Tackling pixel misalignment and detail enhancement as critical components of multi-exposure image fusion, we introduce a self-alignment strategy that combines robust registration and detail repletion to effectively preserve texture. This approach adeptly mitigates artifacts while maintaining the intricate structure inherent to the source images. "},{"figure_ref":[],"heading":"II. RELATED WORKS","publication_ref":[],"table_ref":[],"text":"In this part, we first briefly overview the related literature on multi-exposed image fusion, which contain two representative categories of methods, i.e., traditional numerical schemes and learning-based networks. Then we introduce the development of related architecture search methods."},{"figure_ref":[],"heading":"A. Traditional Numerical Schemes","publication_ref":["b29","b30","b31","b32","b15","b33","b34","b36","b34","b35","b37","b38","b39","b40"],"table_ref":[],"text":"In the past decades, various handcrafted numerical strategies are proposed to achieve multi-exposure fusion. These schemes can be roughly divided into transform-based, gradientbased, weighting-based, and patch-based methods. In detail, transform-based schemes first extract features and introduce fusion strategies based on the informativeness measurement. Diverse multi-scale transformations are developed to comprehensively utilize the principled features from each scale, such as wavelet transform [30], contourlet transform [31], pyramid transform [32] and dense invariant transform [33]. For instance, guided filters [16] is proposed to decompose images into base and details parts on the spatial domain. By introducing the weighted average technique, this method can fuse the consistent feature comprehensively for various fusion scenarios, including multi-exposure, multi-focus, and multimodal image fusion. Another representative technique is the Gaussian pyramids transform [34], which fuse source images to enhance the under/over-exposed regions progressively.\nPatch-based methods [35]- [37] are robust for the fusion scenarios but suffer from artifacts and blurred boundaries. Ma et al. [35], [36] introduced a patch-based method to measure the structural information and use the decomposition strategy to extract the richest features to form the fused images. Kou et al. [38], [39] present the gradient-domain smoothing to realize edge preservation instead of Gaussian smoothing and avoid the inference of halo artifacts. Furthermore, tone-mapping-based methods are developed to achieve HDR construction with various LDR images. Sparse representation methods [40], [41] are widely utilized for multi-exposure image fusion. These schemes utilize the overcomplete dictionary to capture the features of source images and fuse the features utilized by the corresponding sparse coefficients. These methods benefit from the extraction of sharp structures and abundant textures. Traditional numerical schemes rely on the handcrafted feature extraction and inflexible fusion strategies. In this way, traditional schemes cannot adequately perform for challenging fusion scenarios (e.g., extreme exposure variation). Moreover, these schemes are also limited by the huge computation resource and the fusion performance can be reduced drastically when facing large exposure intervals."},{"figure_ref":[],"heading":"B. Learning-based Schemes","publication_ref":["b41","b42","b43","b44","b27","b26","b20","b45","b46","b11","b21","b28","b45","b11"],"table_ref":[],"text":"With the flourishing progress of the deep learning paradigm, learning-based methods realize the promising improvement in the quantity and quality of multi-exposure fusion tasks, compared with traditional methods. Supervised by the MEF-SSIM metric, DeepFuse pioneered the first learning framework to aggregate the luminance components, and utilized a weighted fusion strategy to fuse color and brightness components. Ma et al. proposed MEF-Net [42] to predict the weighted map by feeding down-sampled images. Zhang et al. presented the network IFCNN [43] to adopt the element-wise feature fusion based on the features extracted from two independent branches. The above networks either trained by unsatisfied metric (e.g., MEF-SSIM [44]) or based on the local pixel-wise feature fusion, are easy to result in the color distortion and global structural inconsistency. On the other hand, There are various unified learning-based schemes to uniformly address diverse image fusion tasks. Zhang et al. proposes a densenet with the squeeze and decomposition principle, called SDNet [45] to realize the versatile image fusion framework. A multi-decoder-based framework is introduced with a shared encoder to realize the unified fusion [28]. U2Fusion [27] utilizes the feature similarity based on the gradient to measure the differences between source and fusion images. These versatile fusion methods pursuit discoverer the similarity of tasks, inevitably lacking task-oriented consideration, thus leading to color distortion and structural detail degradation.\nLately, attention mechanisms are widely used for multiexposure fusion. Liu et al. propose a hierarchical attention module [21], [46] to investigate the sufficient information on both under/over-exposed images. Yan et al. [47] introduce the dual spatial attention module to remove the ghosts and misalignments of adjoint frames. Xu et al. [12] introduce the non-local self-attention block to capture the long-range dependency of all regions with a generative adversarial network. Similarly, Li et al. design various attention modules [22] (e.g., coordinate and self-attention) to extract the texture details from source images. Han et al. [29] decouple the MEF task into two deep perceptual enhancements including content detail extraction and color correction. We emphasize that the mainstream learning-based schemes put amount of effort to design attention modules (e.g., hierarchical attention [46], nonlocal attention [12]) to realize the visual-pleasant realistic color correction. However, these schemes mostly utilize the heuristic attention architectures, which cannot achieve the adaptive feature extraction among diverse fusion scenes and suffer from the slow inference time with abundant parameters. Thus, the texture information from source images cannot be sufficiently investigated to generate reasonable results."},{"figure_ref":[],"heading":"C. Neural Architecture Search","publication_ref":["b47","b48","b49","b50","b51","b52","b5","b53","b54","b55"],"table_ref":[],"text":"The appearance of Neural Architecture Search (NAS) takes the network design to enter into a new phase. Current NAS techniques can be categorized as three mainstream methodologies, including reinforcement learning [48], [49], evolutionary algorithms [50], [51], and differentiable search methods. The classical reinforcement learning-based and evolutionary algorithms are limited by massive computations with unaffordable resources. Lately, imposing continuous relaxation into the architecture representation, differentiable search strategy [52], [53] has attracted plenty of attention and widely leveraged for various image enhancement tasks. For instance, Liu et al. [6] introduced the Retinex theory to composite cooperative architecture search for low-light enhancement. Zhang et al. [54] construct the hierarchical super-net with primitive search space and strategy to discover the suitable architecture for image restoration. Li et al. [55] proposes the all-in-one architectures of extracting the shared image features for diverse image restoration using a series of physics principles-guided operations. As for image fusion, cell-level and operation-level search spaces [56] are proposed to address infrared-visible image fusion. We argue that current search strategies have obtained unprecedented attention, but mostly ignore the taskspecific module design, efficient search space, and the tradeoff between inference time. Thus, in this manuscript, we elaborately design principled super-net and search space with introducing hardware-sensitive constraint, aiming to provide a compact architecture for MEF tasks."},{"figure_ref":["fig_0","fig_0"],"heading":"III. PROPOSED METHOD","publication_ref":["b56","b57","b58","b59"],"table_ref":[],"text":"We first introduce a robust multi-exposure fusion framework to address misalignment between source images and visual aesthetics of fused images. Then we introduce a hardwarelatency constrained architecture search with corresponding loss functions to discover the nimble network for fast inference. Concrete components are schematically illustrated in Fig. 2.\nA. Robust Multi-Exposure Fusion Framework 1) Self-alignment Module: As aforementioned, in realworld, the misalignment of image pairs caused by device shaking and movement is almost inevitable. On the other hand, due to the extreme exposure intervals of pairs, it is untoward to straightforwardly utilize alignment techniques (e.g., optical flow [57], registration module [58], [59]), which may produce texture artifacts under inexact alignment. Thus, we conquer this obstacle into two steps: scene relighting for illumiation correction and deformable aligning for feature registration.\nScene Relighting Sub-Module: In essence, based on the Retinex theory, we propose a recurrent adaptative attention mechanism for scene relighting, aiming to push the images into similar illumination domain. We introduce two Scene Relighting Sub-Modules (SRSM) for each image to restrain the degree of illumination of source images into the similar domain for following alignment and detail enhancement. Furthermore, instead of targeting to restore the normal-light scene from single sources, SRSM aims to leverage the illumination map to preserve the comprehensive structures.\nDenoted the intermediate results as I S U and I S O and SRSM as S, the illumination correction can be formulated as\nI S i = I i ⊗ S(I i ), i ∈ {U; O},(1)\nwhere ⊗ denotes the element-wise multiplication. O and U represent under/over exposed images. Noting that, we exploit a recurrent gradual scheme to cascade SRSM, aiming to realize the progressive illumination correction. The stage-wise attention maps can benefit the procedure of complementary feature learning fully and elaborately. As shown at left part of Fig. 2, rather than utilizing heuristic handcrafted methodology, we leverage the differentiable architecture search to construct this module for fast sceneadaption. In detail, we first utilize one 3 × 3 convolution to transfer image into feature domain. Then we set two candidate operations to extract scene features. Max pooling and average pooling are hierarchically embed to realize the amplification of salient feature for illumination estimation completely. Then we leverage one undetermined convolution layer to boost the information richness of features and utilize one 3 × 3 convolution with sigmoid function to generate three-channel illumination map with range [0,1].\nDeformable Aligning Sub-Module: Few multi-exposure fusion methods consider the misalignment of source images, which are based on pre-registered pairs. However, in the realworld scenes, misalignment of over/under exposed images would damage the visual quality with serious ghosts artifacts, due to the movement of image devices or targets. Moreover, introducing learning-based optical flow methods would lead to the huge computation of pixel motion. The lack of real optical flow as ground truth for pre-training limits their performance. Thus, we introduce the Pyramid, Cascading and Deformable Convolution (PCD) mechanism [60] to establish Deformable Aligning Sub-Module (DASM) based on the supervision of visual quality metrics. We only consider DASM under the misalignment scenario.\nSpecifically, DASM first employs diverse strided convolution to generate pyramid features F U and F O based on the intermediate results from SRSM, we utilize deformable convolutions to conduct the feature-level alignment by coarseto-fine manner. Denoted the DASM as A, we can obtain the comprehensive feature as\nF A = A(F U , F O ) + F O ,(2)\nwhere F A represents the fused features based on the summation of aligned source features. Similarly, instead of introducing the original PCD network, we employ architecture search scheme to rebuild the structure (i.e., replacing different kernels of deformable convolutions) to accommodate itself into multiexposure fusion task.\n2) Detail Repletion Module: Then we introduce Detail Repletion Module (DRM) to enhance the textural details of complementary feature. In order to preserve the spatial structures, we utilize successive structure under the same resolution to promote the information richness. Specifically, inspired by effective residual learning mechanisms (e.g., residual dense blocks, dilated dense block and residual on residual), we introduce a residual operator-based search space to discover a suitable dense structure. We also introduce the attention mechanism to finally address the global color distortion. Thus, denoted the network as R and output as y, we can formulate the optimization procedure as\ny = F A R(F A ).(3)\nIn a word, DRM not only targets to strengthen feature representation of details from the fused features, but also protect the integral normal illumination. Specifically, we employ four candidate operators to composite this module. Lastly, we utilize one 3 × 3 convolution layer with sigmoid function to estimate the illumination map. In the following, we will discuss the concrete strategy to search compact MEF framework."},{"figure_ref":["fig_0"],"heading":"B. Automatic Architecture Construction","publication_ref":["b5","b51","b53","b51"],"table_ref":[],"text":"In this part, we introduce the detailed search space and strategy for the light-weight effective architectures.\n1) Principle-driven Search Space: Different from recent NAS-based schemes [6], [52], [54], which introduces the single operators (e.g., one-layer convolution and primitive pooling operations) to composite the search space, without the deep investigation of principles for module-related characteristics, we construct the principle-driven search space. As shown in the bottom part of Fig. 2, normal convolutions (denoted as \"C\") and dilated convolutions (denoted as \"DC\") with different kernel size k × k, k ∈ {1, 3, 5, 7} are utilized for the SRSM, which are consisted by three layers of convolutions for feature representation and dimension changing. In order to persevere the sufficient features to recover the complementary information, we add the skip connection to establish the residual learning, which are denoted as \"RConv\" and \"RDConv\" respectively. Similarly, DASM also can be searched using three kinds of deformable convolutions, denoted as \"3-DeC\", \"5-DeC\" and \"7-DeC\" respectively.\n2) Compact Architecture Search: In this paper, following with the continuous relaxation [52], we introduce the architecture weight α to connect the operators from search space O for the super-net construction. The continuous relaxation from layer i to layer j is formulated as:\nF j = O i→j (F i ); O i→j = O∈O α i→j O(F i ),(4)\nwhere the relaxation operator is denoted as O and O∈O α i→j = 1. In order to obtain the desired architecture with high performance and fast inference time, we also establish the continuous relaxation with operation latency. In this way, we can obtain the inference time of this super-net:\nR(α; LAT) = M O∈O α i→j LAT(O),(5)\nwhere M denotes the number of search blocks. Thus, we introduce the summation of operation latency R(α; LAT) as the constraint for architecture search objective, which can be expressed as:\nmin α val (α; ω * ) + ηR(α; LAT),(6)\nwhere val and ω * are the validation loss and optimal parameters based on the training data. Introducing the differentiable search strategy, we conducted the search of whole super-net."},{"figure_ref":[],"heading":"C. Loss Functions","publication_ref":["b60","b61"],"table_ref":[],"text":"Focusing on the texture details preservation, color information promotion and global scene consistency, we leverage three categories of loss functions to train the proposed network, including pixel-intensity loss Int , gradient loss Gra and global-adversarial loss Dis by supervised learning with ground truth y gt . On the whole, the total loss measurement Total is denoted as:\nTotal = Int + β 1 Gra + β 2 Dis .(7)\nwhere {β 1 , β 2 } are a series of trade-off parameters.\nIn order to realize the same intensity distribution with ground truth (denoted as y gt ), we impose the 1 distance to measure the discrepancy, which can be formulated as:\nInt = 1 HW y -y gt 1 ,(8)\nwhere H, W denote the height and width of image. Due to the interference by noises and corrupted exposures, source images lack partial details. We utilize the Sobel operator to preserve the fine-grained texture details.\nGra = 1 HW ∇y -∇y gt 2 ,(9)\nBecause of the information deficiency of local region, it is untoward to obtain the global consistency of color distribution. Thus, we introduce the discriminator D from PatchGAN [61] to judge the generated results with a global activation map. By this constraint, color distribution of whole scene can be guaranteed. We introduce the gradient-penalty wasserstein training strategy [62] to conduct the generative adversarial learning. Dis is formulated as:\nE x∼Pfake D(y)-E x∼Preal D(y gt )+ηE x∼Pfake [( ∇ y D(y) 2 -1) 2 ].(10)\nIV. EXPERIMENTAL RESULTS In this section, we first introduce the detailed configurations of the architecture search and training procedure. Then we conduct the subjective and objective comparisons on general multi-exposure image fusion and misaligned multi-exposure image fusion with eleven methods, which demonstrates the remarkable performances and robust generalization ability of the proposed method."},{"figure_ref":[],"heading":"A. Search and Training Configurations","publication_ref":["b62","b34"],"table_ref":[],"text":"1) Datasets: We used the widely-used SICE dataset [63] to train and evaluate the performance of our network. This dataset contains diverse sequences of scenes with varying exposure ratios. Each sequence has a well-exposed ground truth. For our general multi-exposure image fusion task, we randomly selected 258 pixel-level registered pairs for training and 100 pairs for testing with a significant exposure difference from each sequence. For the misaligned multi-exposure image fusion task, we selected 100 pairs with noticeable unregistered pixels to create a dataset for misaligned scenarios. Additionally, we introduced a dataset [35] without ground truth to verify the generalization ability of the network. 2) Architecture Search: Specifically, different from the original architecture weight sharing, each search bock is with the unique architecture weights to integrate the operations from diverse sub-search spaces. We utilize 100 pairs from the training dataset to divide the search-training and validation datasets homogeneously. Firstly, the whole super-net is pretrained with 10 epochs to obtain well initialized ω. Then we conduct the differentiable architecture search with 300 epochs. SGD optimizer and cosine delay schedule with initial learning rate 3e -4 are introduced to optimize the neural parameters. Adam optimizer is introduced to update the architecture with learning rate 1e -4 . The η at Eq. ( 6) is empirically set to 0.5 to balance the performance and inference time (denoted as \"Ours\"). The faster version is also provided based on the constraint η = 1 and denoted as \"Ours * \". Moreover, 80 unregistered pairs are utilized to search for the specific network for the misaligned scenarios. Int is defined as the training and validation loss for architecture search."},{"figure_ref":[],"heading":"3) Network Training:","publication_ref":["b63","b64"],"table_ref":[],"text":"The β 1 and β 2 of Eq.( 7) are set to 0.75 and 0.05 respectively. Data augmentation, such as the random crop, horizontal and vertical flipping, rotating are utilized for the training procedure. With patch size of 128 × 128, we train the network 2000 epochs. This network is trained with Adam optimizer and introduce the cosine annealing strategy to delay the learning rate from 1e -4 to 1e -10 progressively.\nWe introduce two categories of metrics to measure the visual quality of generated results, including the referencebased measurements (PSNR and SSIM) and visual perception metrics (LPIPS [64] and FSIM [65]). PSNR measures the differences of pixel intensity between outputs and ground truths. SSIM can provide the similarly measurement from the luminance, contrast and structure aspects. LPIPS utilizes the deep features to measure the perceptual similarity incline with human visual perception. FSIM is also used for the evaluation, which represents the salient low-level features based on phase congruency and gradient magnitude."},{"figure_ref":["fig_1","fig_2","fig_3"],"heading":"B. General Multi-Exposure Image Fusion","publication_ref":["b10","b36","b32","b16","b19","b42","b44","b26","b11","b45","b28","b34"],"table_ref":[],"text":"To demonstrate the effectiveness and remarkable advantages of the proposed methods, we comprehensively compare our methods with eleven competitors including the traditional methods, i.e., MESPD [11], FMMEF [37], MEFDSIFT [33], DEMEF [17] and deep learning methods, i.e., DeepFuse [20], IFCNN [43], SDNet [45], U2Fusion [27], MEFGAN [12], HALDER [46] and DPEMEF [29]. Then we evaluate the proposed scheme with these methods from three aspects, i.e., the objective, subjective comparison, and computation complexity.\n1) Subjective Comparison: We select two representative pairs to demonstrate the superiority of our proposed method, which are shown at Fig. 3 and Fig. 4. In the below part of each image, we also illustrate related signal maps of the line marked by \"blue\" compared with ground truth, to show the obvious differences. Since these image pairs have a challenging distinct exposure gap, it is untoward to design the multi-exposure fusion scheme to preserve the suitable brightness, structural details, and abundant color distribution. From these noticeable comparisons, we can conclude our methods contain three significant advantages.\nFirstly, our scheme realizes the consistency of global brightness. Patch-based schemes (e.g., MESPD, DEMEF, and FM- MEF) cannot achieve brightness consistency, which introduces significant artifacts, caused by the extreme difference in exposure time. From the signal maps of DEMEF and FMMEF at the first scene, the fused results are over-exposure with higher pixel intensities compared with ground truth. Meanwhile, the sky is degraded by extreme-low exposure time, shown in the local under-exposed regions. In contrast, our method can recover the promising brightness with normal illumination. Secondly, current learning-based schemes are easily trended to color distortion. For instance, SDNet, U2Fusion, and HALDER methods cannot realize the vivid color details, including the bushes in the first scene and the flowers in the second scene. This illustration is also reflected in the corresponding signal maps. These methods cannot achieve large signal changes and are with a moderate reflection of pixel intensity. Our method and MEFGAN can preserve the promising color distribution with remarkable improvement. More importantly, the maintenance of texture details is a critical goal for this fusion task. From the two blew-up regions, it is obvious to observe that, most fusion schemes can not maintain sufficient details, e.g., the cliff painting and tower.\nEspecially in the second scene, corrupted by a large difference in exposure time and severe illumination, it is difficult to recover the concrete details by simple feature aggregation and direct fusion principles. Our results are visual-friendly, which is an incline with the human vision system.\nOn the other hand, we also provide another comparison based on the dataset [35] without ground truth in Fig. 5. We select two pairs with extreme exposure variance to illustrate the effectiveness of our scheme with nine state-ofthe-art multi-exposed image fusion. As shown in the first sequence, the information (e.g., cloud layer) under low exposure cannot be recovered clearly. Thus, these details are hard to be highlighted and recovered from under-exposed images (e.g., SDNet and U2Fusion marked by the green boxes). We can clearly observe that attention-based methods (MEFGAN, HALDER, and Ours) can achieve abundant detail preservations. Especially, our network can effectively promote visual perception to render sufficient details. Furthermore, our method can accomplish vivid color enhancement. Most of the results appear in the local over-exposure region. In contrast, our method achieves abundant texture details (e.g., bushes) and consistent color distribution (e.g., computer screen). Either existing patch-based optimization methods or learning-based schemes cannot address the global consistency of illumination. "},{"figure_ref":[],"heading":"2) Objective Comparison:","publication_ref":[],"table_ref":[],"text":"To demonstrate the superiority of the proposed scheme, we utilize four different metrics, including PSNR, SSIM, LPIPS, and FSIM to measure the visual quality of diverse methods. The whole numerical results are reported in Table . I. We introduce two versions to conduct the comparison, which are named \"Ours\" and \"Ours * \" respectively. The difference between both versions is utilizing diverse latency constraints, where \"Ours * \" focuses more on the inference time. Obviously, our scheme achieves consistent performance improvement in terms of these metrics. Compared with representative learning-based schemes MEFGAN and HALDER, our scheme promotes 0.8 dB and 1.0 dB drastically. On the other hand, it can be clearly seen that we obtain the second best numerical results. However, the patch-based fusion scheme obtains the suboptimal numerical performance, which indicates our scheme can effectively preserve sufficient textural details and structural information. We also utilize the LPIPS to measure the distortion at feature levels. Our scheme can reduce almost 24.6% of LPIPS compared with HALDER, which demonstrates better visual quality in line with the human perception system. Obviously, our scheme also achieves the best results.\n3) Computation Efficiency Analyses: We also conduct a comparison under computation efficiency, which is a critical point for real-world deployment. The concert numerical results among these competitors under the metrics of parameters and runtime on the SICE dataset are reported in Table . II. Obviously, most learning-based methods achieve fast inference time than traditional numerical-based schemes (e.g., DEMEF and MEFDSIFT). Furthermore, though SDNet and IFCNN have fewer parameters, and are limited by direct fusion rules and inefficient architectures, their performances are interfered. More importantly, both our methods realize the faster inference time. Compared with the latest learning-based method DPEMEF, the fastest version (ours * ) can effectively promote 68.11% runtime and reduce 98.68% parameters, which demonstrates high efficiency with visual-pleasant fused results.\n4) Fusion with Arbitrary Exposure Ratios: In order to verify the generalization ability to address the inputs with arbitrary exposure ratios, we select two representative sequences, which is shown in Fig. 6. Since we utilize the pairs with the Fig. 6. Visual results about the source image sequences with different exposure ratios. ( 1) and ( 2) are under-exposed images, (3), ( 4) and ( 5) are over-exposed images. (a)-(c) are fused by inputs of ( 1) and ( 3)-( 5). (d)-(f) are fused by inputs of ( 2) and ( 3)-( 5)."},{"figure_ref":[],"heading":"Under Exposure","publication_ref":[],"table_ref":[],"text":"Under Exposure\nOver Exposure Over exposure ratios for training and testing, the proposed scheme is robust enough to handle different exposure ratios. These sequences contain two under-exposed images and three over-exposed images. We can clearly observe that the scheme can obtain the consistent visual-pleasant fused results with natural color correction, generated by the image pairs with diverse exposure ranges. Several characteristics can be found from the fused result of this sequence. Firstly, the proposed scheme is with large capacity for wide exposure difference to preserve the textural details and maintain suitable illumination distribution. Secondly, though without the training procedure of small exposure difference, the fused images generated by (a) and (b) contain sufficient scene details, e.g., grasses and sunset glow. Finally, our result (i.e., (c)) fused by the large difference of exposure is close to the ground truth. Several fused images (e.g., (b) and (d)) have more vivid scene representation compared with ground truth."},{"figure_ref":[],"heading":"C. Misaligned Multi-Exposure Image Fusion","publication_ref":[],"table_ref":[],"text":"Misaligned multi-exposure image fusion is a challenging scenario due to the camera movement and device shaking, whereas current methods for MEF are easy to generate blurs without the consideration of pixel registration. We also illustrate the performance by numerical and visual comparisons."},{"figure_ref":["fig_4"],"heading":"1) Visual Comparison:","publication_ref":[],"table_ref":[],"text":"We select four misaligned pairs from the testing dataset to demonstrate the effectiveness of proposed framework in Fig. 7. The first two rows illustrate the scene with large pixel movements. Most methods cannot preserve sufficient details. The results generated by MEF-GAN and FMMEF methods have obvious artifacts and cannot preserve the normal color distribution. More importantly, our scheme successfully realizes the uniform promotion of pixel alignment and visual enhancement, which can effectively address diverse levels of pixel alignment. 2) Numerical Comparison: In Table . III, we report the numerical results compared with various representative multiexposure image fusion methods. We utilize five typical metrics to measure the visual quality of fused images. Since existing methods often assume the image pairs are well-registered, these methods cannot obtain promising quantitative results. We can observe that the proposed method achieves the consistent best numerical results in terms of these four metrics. These results demonstrate the superiority of our methods on the misaligned MEF. "},{"figure_ref":[],"heading":"V. ABLATION STUDY","publication_ref":[],"table_ref":[],"text":"In this section, we conduct sufficient experiments with numerical and visual evaluations to verify the effectiveness of proposed modules, loss functions and architecture search."},{"figure_ref":["fig_5","fig_5"],"heading":"A. Effectiveness of Scene Relighting.","publication_ref":[],"table_ref":[],"text":"In this part, we first validate the effectiveness of proposed SRSM and validate the suitable cascaded numbers for MEF task. The role of scene relighting is to gradually preserve the scene information and constrain the level of illumination for following feature aggregation. The ablation experiment about the cascaded number of SRSM is conducted, where the quantitative and qualitative comparisons are shown at Table . IV and Fig. 8 respectively. Firstly, we illustrate the necessary of proposed SRSM. The version without SRSM only concatenate the inputs to feed into the DRM, without the procedure of illumination adjustment. We can clearly observe that, directly processing compromises the image quality, which reduce the numerical performance drastically. As shown in Fig. 8, we can obtain the output image is over-exposed, cannot render sufficient detail and preserve the normal light distribution. Compared by pie charts, which depicts the proportion of RGB channels. The version w/o SRSM cannot restore the normal color distribution, which leads to the distortion of color and details. Then we evaluate the cascaded number of SRSM. By introducing the cascaded SRSMs, we propose the recurrent attention mechanisms to extract the sufficient features. Cascading two modules can achieve the best numerical performance. Increasing the number of SRSM obtain the moderate improvement. "},{"figure_ref":["fig_4"],"heading":"B. Effectiveness of Deformable Alignment","publication_ref":["b56"],"table_ref":[],"text":"We further evaluate the advantages of self-alignment mechanism. Three variants of comparisons are conducted, including \"w/o DASM\", optical flow-based and changing the position before relighting. Self-alignment module targets to align the unregistered pixels of image pairs. In the architecture construction, we put DASM after the SRSM. Optical flow-based schemes is to firstly utilize the RAFT [57] to align image pairs. Numerical results and visual comparisons are depicted at Table . V and Fig. 7 respectively. From the numerical results, we can observe the effectiveness of proposed mechanisms. Our scheme improves 2.65dB of PSNR, 23.55% of VIF and 7.16% of FSIM. From the visual comparison, the fused result of \"w/o DASM\" cannot preserve enough texture details, such as the grasses. On the other hand, we can see that, the optical flow-based schemes cannot improve the quantitative results due to the inaccurate motion estimation caused by different illumination. The fused images contains more obvious artifacts. Meanwhile, we also discuss the suitable position of relighting. Obviously, the variant \"Before SRSM\" obtains the sub-optimal numerical and visual results compared with the final scheme, but has remarkable performance promotion compared with previous two variants. Furthermore, based on the similar intensity distribution by SRSM, the final scheme can realize the consistent improvement. "},{"figure_ref":["fig_6","fig_6","fig_6"],"heading":"C. Training Losses Analyses","publication_ref":[],"table_ref":[],"text":"also perform the detailed analyses to evaluate the effectiveness of diverse training strategies (i.e., combinations of loss functions). In this part, we gradually introduce the loss functions to composited three schemes, including \"w/ Int \", \"w/ Int + Gra \" and our scheme. Related visual results are plotted at Fig. 10. From the objective comparison, Gra can effectively preserve the edge information, which reflects on the structural measurement SSIM and feature-level metric LPIPS. As shown in Fig. 10, visual results scheme \" w/ Int + Gra \" provide flourishing textural details, e.g., the details of grasses and floors. Meanwhile, introducing Gra can effectively remove the artifact such as the shape of tree on the first row. Our scheme is combined with three categories of losses, i.e., Int for pixel intensity, Gra for structural detail and Dis for color distribution. Thus, our scheme can further improve the visual quality, obtaining with the highest numerical results. For instance, our final scheme obtain the vivid color distribution, without any color distortion (e.g., the color of wall), shown at the second row in Fig. 10."},{"figure_ref":[],"heading":"Inputs","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Int","publication_ref":[],"table_ref":[],"text":"Int+ Gra Total "},{"figure_ref":[],"heading":"D. Search Space Analyses","publication_ref":[],"table_ref":[],"text":"We also verify the basic properties of search space, where the concrete performances are reported at Table. VII. From the fusion performance, we can observe that, dilated convolutions (3-DC and 7-DC) have higher quantitative results (e.g., PSNR, LPIPS and FSIM) compared with normal convolutions. On the other hand, we can directly observe that 3 × 3 convolution has the fast inference speed but has the sub-optimal statistical results. Under the constrained of hardware latency, our searched scheme actually achieve the balance between visual quality and inference speed."},{"figure_ref":[],"heading":"E. Hardware-sensitive Analyses","publication_ref":[],"table_ref":[],"text":"We also analyse the influence of trade-off parameter η, which controls the influences of hardware-sensitive latency constraint. The numerical results are reported in Table . VIII. \"C32\" and \"C64\" denote the version with 32 and 64 channels. When η = 0.5, the balance between fusion quality and inference requirement can be guaranteed simultaneously. The concrete architectures under diverse η are plotted in Fig. 11. The previous three layers illustrate the architecture of SAM. The last four layers shows the structure of DRM with residual connection. In detail, without the constraint of latency, NAS scheme chooses operator with large receptive field to better capture the large features. Moreover, we also can conclude that 3 × 3 convolution can effectively extract features with high efficiency, which is widely leveraged for SRSM under the latency constraint. As for DRM, {3-C, 1-C} with skip connection is a low-weight combination for the detail compensation, as shown at the subfigure (c) and (d). "},{"figure_ref":[],"heading":"VI. CONCLUDING REMARKS","publication_ref":[],"table_ref":[],"text":"In this paper, we proposed a robust multi-exposure image fusion framework to address various scenarios, including the aligned and misaligned image pairs. We divided the fusion procedure into two parts: self-alignment for feature-wise alignment and detail repletion to enhance texture details visually. By utilizing a hardware-friendly architecture search strategy and incorporating a task-oriented search space, we discovered a highly efficient and compact architecture for MEF. Furthermore, we conducted comprehensive subjective and objective comparisons to demonstrate the outstanding performance of our method compared to various state-of-the-arts."},{"figure_ref":[],"heading":"","publication_ref":[],"table_ref":[],"text":"The source code will be available at https://github.com/LiuZhu-CV"},{"figure_ref":[],"heading":"","publication_ref":[],"table_ref":[],"text":"This work is partially supported by the National Key R&D Program of China (2020YFB1313503), the National Natural Science Foundation of China (U22B2052), the Fundamental Research Funds for the Central Universities and the Major Key Project of PCL (PCL2021A12)."}],"string":"[\n {\n \"figure_ref\": [],\n \"heading\": \"I. INTRODUCTION\",\n \"publication_ref\": [\n \"b0\",\n \"b2\",\n \"b3\",\n \"b4\",\n \"b5\",\n \"b7\",\n \"b8\",\n \"b9\",\n \"b10\",\n \"b11\",\n \"b12\",\n \"b14\",\n \"b15\",\n \"b18\",\n \"b19\",\n \"b22\",\n \"b15\",\n \"b20\",\n \"b23\",\n \"b24\",\n \"b25\",\n \"b26\",\n \"b28\"\n ],\n \"table_ref\": [],\n \"text\": \"High Dynamic Range Imaging (HDRI) [1]- [3], encompassing comprehensive scene content with optimal exposure, has gained considerable attention in recent years. As a pivotal technique for computer vision, HDRI not only delivers visually appealing observations in harmony with the human visual system, but also integrates essential features for a range of downstream vision applications, including object detection [4], [5], visual enhancement [6]- [8] and semantic segmentation [9], [10]. Unfortunately, due to constraints in photography Fig. 1. Visual comparison with representative MEF methods on general and misaligned fusion scenarios. The left figure plots the ranking with these competitors under general and misaligned scenarios. Other figures shows the obvious comparison with patch-based scheme MEFSPD [11] and learningbased method MEFGAN [12].\\nequipment (e.g., smartphones and single-lens reflex cameras) that capture images with limited dynamic ranges, these images experience varying degrees of luminance degradation, leading to corrupted over/under-exposed regions. As a result, Low Dynamic Range (LDR) images are plagued by color distortion and loss of detail, hindering the accurate portrayal of complete natural scenes. Consequently, the generation of well-exposed HDR images remains both a challenging and significant research topic.\\nRecently, a growing number of researchers have endeavored to create cutting-edge HDRI hardware devices capable of producing an extensive range of illumination, thereby addressing the limitations inherent to traditional digital cameras [13]- [15]. Nevertheless, due to elevated production costs and suboptimal efficiency, these intricately designed devices face challenges in achieving widespread adoption in real-world applications. As an alternative, Multi-Exposure Fusion (MEF) offers an efficacious solution for generating HDR images by assimilating characteristic texture information from a collection of LDR images captured under diverse exposures. This strategy adeptly bypasses hardware-specific limitations while maintaining a lower computational cost. Within the existing literature, conventional frameworks [16]- [19] and learning-based ones [20]- [23] constitute the predominant categories of MEF techniques. Despite these advancements, MEF continues to grapple with certain hurdles that impede its overall efficacy.\\nIt is imperative to highlight that current learning-based methods neglect the essential adaptive preservation in multi-exposure image fusion. Specifically, a variety of approaches utilize direct fusion rules for feature aggregation, such as summation [16] and multiplication [21]. Regrettably, these rudimentary fusion rules fall short in effectively aggregating information from LDR pairs with markedly distinct characteristics, thereby failing to preserve critical information (e.g., pixel intensity and texture details) appropriately. Furthermore, given the considerable variation in multi-exposure image distributions, manually designed architectures face difficulties in flexibly adapting to disparate data distributions. In addition, due to the unavoidable movements and shaking of imaging devices, minor pixel misalignments in LDR pairs are commonplace. Existing methods seldom tackle this issue, leading to fused images characterized by blurred details and compromised structure. In essence, there is an urgent demand for an all-encompassing, robust, and efficient learning approach that not only delivers promising visual realism enhancement but also ensures high efficiency and stability across a wide array of scenes.\\nTo be more specific, the second challenge stems from the computational efficiency of existing methods. Present MEF methods, encompassing both conventional and learningbased approaches, rely heavily on handcrafted architectures and operations. In terms of conventional frameworks, various transformations such as wavelet transform [24], multi-scale representation [25], and Laplacian pyramid [26] are proposed to enable feature fusion through handcrafted mechanisms. Unfortunately, these manual designs for feature extraction and fusion rules demand substantial fine-tuning and a significant amount of experiential knowledge. Most of these methods utilize numerical optimization, which in turn impacts the inference efficiency and robustness in real-world applications. Furthermore, in recent years, the powerful feature extraction capabilities of CNN-based learning have led to the increasing dominance of end-to-end models in MEF development, considerably improving performance concerning statistical metrics and visual effects. For architectural construction, a variety of learnable mechanisms [27]- [29] have been introduced to forge connections between LDR pairs and HDR outputs. We contend that current neural architectures for MEF largely borrow effective practices from other vision tasks without paying adequate attention to MEF-specific characteristics. As a consequence, these simplistic cascaded architectures, marked by increased width and depth, possess an excessive number of parameters, making them prone to feature redundancy.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A. Contributions\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"To partially overcome these limitations, we present a versatile architecture search-based multi-exposure fusion approach. Specifically, we first develop a fusion-centric hyper-architecture, adhering to two primary principles: selfalignment and detail repletion. We initially propose scene relighting to map source images onto the same illumination, enhancing over/under-exposed details and providing improved support for subsequent feature aggregation. We introduce deformable alignment to achieve accurate feature registration, minimizing artifacts and blurring. Subsequently, we propose the detail repletion module to refine the coarse fusion results, leading to richer texture details. Next, we establish a flexible search space encompassing more effective operations, enabling the selection of an optimal network architecture. Considering hardware latency constraints, we employ a differentiable architecture search to automatically discover a compact and efficient model for image fusion. Consequently, our method effectively produces vivid colors, abundant details, and ghostingfree results, as intuitively illustrated in Fig. 1. Our core contributions can be summarized as follows:\\n• Tackling pixel misalignment and detail enhancement as critical components of multi-exposure image fusion, we introduce a self-alignment strategy that combines robust registration and detail repletion to effectively preserve texture. This approach adeptly mitigates artifacts while maintaining the intricate structure inherent to the source images. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"II. RELATED WORKS\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this part, we first briefly overview the related literature on multi-exposed image fusion, which contain two representative categories of methods, i.e., traditional numerical schemes and learning-based networks. Then we introduce the development of related architecture search methods.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A. Traditional Numerical Schemes\",\n \"publication_ref\": [\n \"b29\",\n \"b30\",\n \"b31\",\n \"b32\",\n \"b15\",\n \"b33\",\n \"b34\",\n \"b36\",\n \"b34\",\n \"b35\",\n \"b37\",\n \"b38\",\n \"b39\",\n \"b40\"\n ],\n \"table_ref\": [],\n \"text\": \"In the past decades, various handcrafted numerical strategies are proposed to achieve multi-exposure fusion. These schemes can be roughly divided into transform-based, gradientbased, weighting-based, and patch-based methods. In detail, transform-based schemes first extract features and introduce fusion strategies based on the informativeness measurement. Diverse multi-scale transformations are developed to comprehensively utilize the principled features from each scale, such as wavelet transform [30], contourlet transform [31], pyramid transform [32] and dense invariant transform [33]. For instance, guided filters [16] is proposed to decompose images into base and details parts on the spatial domain. By introducing the weighted average technique, this method can fuse the consistent feature comprehensively for various fusion scenarios, including multi-exposure, multi-focus, and multimodal image fusion. Another representative technique is the Gaussian pyramids transform [34], which fuse source images to enhance the under/over-exposed regions progressively.\\nPatch-based methods [35]- [37] are robust for the fusion scenarios but suffer from artifacts and blurred boundaries. Ma et al. [35], [36] introduced a patch-based method to measure the structural information and use the decomposition strategy to extract the richest features to form the fused images. Kou et al. [38], [39] present the gradient-domain smoothing to realize edge preservation instead of Gaussian smoothing and avoid the inference of halo artifacts. Furthermore, tone-mapping-based methods are developed to achieve HDR construction with various LDR images. Sparse representation methods [40], [41] are widely utilized for multi-exposure image fusion. These schemes utilize the overcomplete dictionary to capture the features of source images and fuse the features utilized by the corresponding sparse coefficients. These methods benefit from the extraction of sharp structures and abundant textures. Traditional numerical schemes rely on the handcrafted feature extraction and inflexible fusion strategies. In this way, traditional schemes cannot adequately perform for challenging fusion scenarios (e.g., extreme exposure variation). Moreover, these schemes are also limited by the huge computation resource and the fusion performance can be reduced drastically when facing large exposure intervals.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B. Learning-based Schemes\",\n \"publication_ref\": [\n \"b41\",\n \"b42\",\n \"b43\",\n \"b44\",\n \"b27\",\n \"b26\",\n \"b20\",\n \"b45\",\n \"b46\",\n \"b11\",\n \"b21\",\n \"b28\",\n \"b45\",\n \"b11\"\n ],\n \"table_ref\": [],\n \"text\": \"With the flourishing progress of the deep learning paradigm, learning-based methods realize the promising improvement in the quantity and quality of multi-exposure fusion tasks, compared with traditional methods. Supervised by the MEF-SSIM metric, DeepFuse pioneered the first learning framework to aggregate the luminance components, and utilized a weighted fusion strategy to fuse color and brightness components. Ma et al. proposed MEF-Net [42] to predict the weighted map by feeding down-sampled images. Zhang et al. presented the network IFCNN [43] to adopt the element-wise feature fusion based on the features extracted from two independent branches. The above networks either trained by unsatisfied metric (e.g., MEF-SSIM [44]) or based on the local pixel-wise feature fusion, are easy to result in the color distortion and global structural inconsistency. On the other hand, There are various unified learning-based schemes to uniformly address diverse image fusion tasks. Zhang et al. proposes a densenet with the squeeze and decomposition principle, called SDNet [45] to realize the versatile image fusion framework. A multi-decoder-based framework is introduced with a shared encoder to realize the unified fusion [28]. U2Fusion [27] utilizes the feature similarity based on the gradient to measure the differences between source and fusion images. These versatile fusion methods pursuit discoverer the similarity of tasks, inevitably lacking task-oriented consideration, thus leading to color distortion and structural detail degradation.\\nLately, attention mechanisms are widely used for multiexposure fusion. Liu et al. propose a hierarchical attention module [21], [46] to investigate the sufficient information on both under/over-exposed images. Yan et al. [47] introduce the dual spatial attention module to remove the ghosts and misalignments of adjoint frames. Xu et al. [12] introduce the non-local self-attention block to capture the long-range dependency of all regions with a generative adversarial network. Similarly, Li et al. design various attention modules [22] (e.g., coordinate and self-attention) to extract the texture details from source images. Han et al. [29] decouple the MEF task into two deep perceptual enhancements including content detail extraction and color correction. We emphasize that the mainstream learning-based schemes put amount of effort to design attention modules (e.g., hierarchical attention [46], nonlocal attention [12]) to realize the visual-pleasant realistic color correction. However, these schemes mostly utilize the heuristic attention architectures, which cannot achieve the adaptive feature extraction among diverse fusion scenes and suffer from the slow inference time with abundant parameters. Thus, the texture information from source images cannot be sufficiently investigated to generate reasonable results.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C. Neural Architecture Search\",\n \"publication_ref\": [\n \"b47\",\n \"b48\",\n \"b49\",\n \"b50\",\n \"b51\",\n \"b52\",\n \"b5\",\n \"b53\",\n \"b54\",\n \"b55\"\n ],\n \"table_ref\": [],\n \"text\": \"The appearance of Neural Architecture Search (NAS) takes the network design to enter into a new phase. Current NAS techniques can be categorized as three mainstream methodologies, including reinforcement learning [48], [49], evolutionary algorithms [50], [51], and differentiable search methods. The classical reinforcement learning-based and evolutionary algorithms are limited by massive computations with unaffordable resources. Lately, imposing continuous relaxation into the architecture representation, differentiable search strategy [52], [53] has attracted plenty of attention and widely leveraged for various image enhancement tasks. For instance, Liu et al. [6] introduced the Retinex theory to composite cooperative architecture search for low-light enhancement. Zhang et al. [54] construct the hierarchical super-net with primitive search space and strategy to discover the suitable architecture for image restoration. Li et al. [55] proposes the all-in-one architectures of extracting the shared image features for diverse image restoration using a series of physics principles-guided operations. As for image fusion, cell-level and operation-level search spaces [56] are proposed to address infrared-visible image fusion. We argue that current search strategies have obtained unprecedented attention, but mostly ignore the taskspecific module design, efficient search space, and the tradeoff between inference time. Thus, in this manuscript, we elaborately design principled super-net and search space with introducing hardware-sensitive constraint, aiming to provide a compact architecture for MEF tasks.\"\n },\n {\n \"figure_ref\": [\n \"fig_0\",\n \"fig_0\"\n ],\n \"heading\": \"III. PROPOSED METHOD\",\n \"publication_ref\": [\n \"b56\",\n \"b57\",\n \"b58\",\n \"b59\"\n ],\n \"table_ref\": [],\n \"text\": \"We first introduce a robust multi-exposure fusion framework to address misalignment between source images and visual aesthetics of fused images. Then we introduce a hardwarelatency constrained architecture search with corresponding loss functions to discover the nimble network for fast inference. Concrete components are schematically illustrated in Fig. 2.\\nA. Robust Multi-Exposure Fusion Framework 1) Self-alignment Module: As aforementioned, in realworld, the misalignment of image pairs caused by device shaking and movement is almost inevitable. On the other hand, due to the extreme exposure intervals of pairs, it is untoward to straightforwardly utilize alignment techniques (e.g., optical flow [57], registration module [58], [59]), which may produce texture artifacts under inexact alignment. Thus, we conquer this obstacle into two steps: scene relighting for illumiation correction and deformable aligning for feature registration.\\nScene Relighting Sub-Module: In essence, based on the Retinex theory, we propose a recurrent adaptative attention mechanism for scene relighting, aiming to push the images into similar illumination domain. We introduce two Scene Relighting Sub-Modules (SRSM) for each image to restrain the degree of illumination of source images into the similar domain for following alignment and detail enhancement. Furthermore, instead of targeting to restore the normal-light scene from single sources, SRSM aims to leverage the illumination map to preserve the comprehensive structures.\\nDenoted the intermediate results as I S U and I S O and SRSM as S, the illumination correction can be formulated as\\nI S i = I i ⊗ S(I i ), i ∈ {U; O},(1)\\nwhere ⊗ denotes the element-wise multiplication. O and U represent under/over exposed images. Noting that, we exploit a recurrent gradual scheme to cascade SRSM, aiming to realize the progressive illumination correction. The stage-wise attention maps can benefit the procedure of complementary feature learning fully and elaborately. As shown at left part of Fig. 2, rather than utilizing heuristic handcrafted methodology, we leverage the differentiable architecture search to construct this module for fast sceneadaption. In detail, we first utilize one 3 × 3 convolution to transfer image into feature domain. Then we set two candidate operations to extract scene features. Max pooling and average pooling are hierarchically embed to realize the amplification of salient feature for illumination estimation completely. Then we leverage one undetermined convolution layer to boost the information richness of features and utilize one 3 × 3 convolution with sigmoid function to generate three-channel illumination map with range [0,1].\\nDeformable Aligning Sub-Module: Few multi-exposure fusion methods consider the misalignment of source images, which are based on pre-registered pairs. However, in the realworld scenes, misalignment of over/under exposed images would damage the visual quality with serious ghosts artifacts, due to the movement of image devices or targets. Moreover, introducing learning-based optical flow methods would lead to the huge computation of pixel motion. The lack of real optical flow as ground truth for pre-training limits their performance. Thus, we introduce the Pyramid, Cascading and Deformable Convolution (PCD) mechanism [60] to establish Deformable Aligning Sub-Module (DASM) based on the supervision of visual quality metrics. We only consider DASM under the misalignment scenario.\\nSpecifically, DASM first employs diverse strided convolution to generate pyramid features F U and F O based on the intermediate results from SRSM, we utilize deformable convolutions to conduct the feature-level alignment by coarseto-fine manner. Denoted the DASM as A, we can obtain the comprehensive feature as\\nF A = A(F U , F O ) + F O ,(2)\\nwhere F A represents the fused features based on the summation of aligned source features. Similarly, instead of introducing the original PCD network, we employ architecture search scheme to rebuild the structure (i.e., replacing different kernels of deformable convolutions) to accommodate itself into multiexposure fusion task.\\n2) Detail Repletion Module: Then we introduce Detail Repletion Module (DRM) to enhance the textural details of complementary feature. In order to preserve the spatial structures, we utilize successive structure under the same resolution to promote the information richness. Specifically, inspired by effective residual learning mechanisms (e.g., residual dense blocks, dilated dense block and residual on residual), we introduce a residual operator-based search space to discover a suitable dense structure. We also introduce the attention mechanism to finally address the global color distortion. Thus, denoted the network as R and output as y, we can formulate the optimization procedure as\\ny = F A R(F A ).(3)\\nIn a word, DRM not only targets to strengthen feature representation of details from the fused features, but also protect the integral normal illumination. Specifically, we employ four candidate operators to composite this module. Lastly, we utilize one 3 × 3 convolution layer with sigmoid function to estimate the illumination map. In the following, we will discuss the concrete strategy to search compact MEF framework.\"\n },\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"B. Automatic Architecture Construction\",\n \"publication_ref\": [\n \"b5\",\n \"b51\",\n \"b53\",\n \"b51\"\n ],\n \"table_ref\": [],\n \"text\": \"In this part, we introduce the detailed search space and strategy for the light-weight effective architectures.\\n1) Principle-driven Search Space: Different from recent NAS-based schemes [6], [52], [54], which introduces the single operators (e.g., one-layer convolution and primitive pooling operations) to composite the search space, without the deep investigation of principles for module-related characteristics, we construct the principle-driven search space. As shown in the bottom part of Fig. 2, normal convolutions (denoted as \\\"C\\\") and dilated convolutions (denoted as \\\"DC\\\") with different kernel size k × k, k ∈ {1, 3, 5, 7} are utilized for the SRSM, which are consisted by three layers of convolutions for feature representation and dimension changing. In order to persevere the sufficient features to recover the complementary information, we add the skip connection to establish the residual learning, which are denoted as \\\"RConv\\\" and \\\"RDConv\\\" respectively. Similarly, DASM also can be searched using three kinds of deformable convolutions, denoted as \\\"3-DeC\\\", \\\"5-DeC\\\" and \\\"7-DeC\\\" respectively.\\n2) Compact Architecture Search: In this paper, following with the continuous relaxation [52], we introduce the architecture weight α to connect the operators from search space O for the super-net construction. The continuous relaxation from layer i to layer j is formulated as:\\nF j = O i→j (F i ); O i→j = O∈O α i→j O(F i ),(4)\\nwhere the relaxation operator is denoted as O and O∈O α i→j = 1. In order to obtain the desired architecture with high performance and fast inference time, we also establish the continuous relaxation with operation latency. In this way, we can obtain the inference time of this super-net:\\nR(α; LAT) = M O∈O α i→j LAT(O),(5)\\nwhere M denotes the number of search blocks. Thus, we introduce the summation of operation latency R(α; LAT) as the constraint for architecture search objective, which can be expressed as:\\nmin α val (α; ω * ) + ηR(α; LAT),(6)\\nwhere val and ω * are the validation loss and optimal parameters based on the training data. Introducing the differentiable search strategy, we conducted the search of whole super-net.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C. Loss Functions\",\n \"publication_ref\": [\n \"b60\",\n \"b61\"\n ],\n \"table_ref\": [],\n \"text\": \"Focusing on the texture details preservation, color information promotion and global scene consistency, we leverage three categories of loss functions to train the proposed network, including pixel-intensity loss Int , gradient loss Gra and global-adversarial loss Dis by supervised learning with ground truth y gt . On the whole, the total loss measurement Total is denoted as:\\nTotal = Int + β 1 Gra + β 2 Dis .(7)\\nwhere {β 1 , β 2 } are a series of trade-off parameters.\\nIn order to realize the same intensity distribution with ground truth (denoted as y gt ), we impose the 1 distance to measure the discrepancy, which can be formulated as:\\nInt = 1 HW y -y gt 1 ,(8)\\nwhere H, W denote the height and width of image. Due to the interference by noises and corrupted exposures, source images lack partial details. We utilize the Sobel operator to preserve the fine-grained texture details.\\nGra = 1 HW ∇y -∇y gt 2 ,(9)\\nBecause of the information deficiency of local region, it is untoward to obtain the global consistency of color distribution. Thus, we introduce the discriminator D from PatchGAN [61] to judge the generated results with a global activation map. By this constraint, color distribution of whole scene can be guaranteed. We introduce the gradient-penalty wasserstein training strategy [62] to conduct the generative adversarial learning. Dis is formulated as:\\nE x∼Pfake D(y)-E x∼Preal D(y gt )+ηE x∼Pfake [( ∇ y D(y) 2 -1) 2 ].(10)\\nIV. EXPERIMENTAL RESULTS In this section, we first introduce the detailed configurations of the architecture search and training procedure. Then we conduct the subjective and objective comparisons on general multi-exposure image fusion and misaligned multi-exposure image fusion with eleven methods, which demonstrates the remarkable performances and robust generalization ability of the proposed method.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A. Search and Training Configurations\",\n \"publication_ref\": [\n \"b62\",\n \"b34\"\n ],\n \"table_ref\": [],\n \"text\": \"1) Datasets: We used the widely-used SICE dataset [63] to train and evaluate the performance of our network. This dataset contains diverse sequences of scenes with varying exposure ratios. Each sequence has a well-exposed ground truth. For our general multi-exposure image fusion task, we randomly selected 258 pixel-level registered pairs for training and 100 pairs for testing with a significant exposure difference from each sequence. For the misaligned multi-exposure image fusion task, we selected 100 pairs with noticeable unregistered pixels to create a dataset for misaligned scenarios. Additionally, we introduced a dataset [35] without ground truth to verify the generalization ability of the network. 2) Architecture Search: Specifically, different from the original architecture weight sharing, each search bock is with the unique architecture weights to integrate the operations from diverse sub-search spaces. We utilize 100 pairs from the training dataset to divide the search-training and validation datasets homogeneously. Firstly, the whole super-net is pretrained with 10 epochs to obtain well initialized ω. Then we conduct the differentiable architecture search with 300 epochs. SGD optimizer and cosine delay schedule with initial learning rate 3e -4 are introduced to optimize the neural parameters. Adam optimizer is introduced to update the architecture with learning rate 1e -4 . The η at Eq. ( 6) is empirically set to 0.5 to balance the performance and inference time (denoted as \\\"Ours\\\"). The faster version is also provided based on the constraint η = 1 and denoted as \\\"Ours * \\\". Moreover, 80 unregistered pairs are utilized to search for the specific network for the misaligned scenarios. Int is defined as the training and validation loss for architecture search.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"3) Network Training:\",\n \"publication_ref\": [\n \"b63\",\n \"b64\"\n ],\n \"table_ref\": [],\n \"text\": \"The β 1 and β 2 of Eq.( 7) are set to 0.75 and 0.05 respectively. Data augmentation, such as the random crop, horizontal and vertical flipping, rotating are utilized for the training procedure. With patch size of 128 × 128, we train the network 2000 epochs. This network is trained with Adam optimizer and introduce the cosine annealing strategy to delay the learning rate from 1e -4 to 1e -10 progressively.\\nWe introduce two categories of metrics to measure the visual quality of generated results, including the referencebased measurements (PSNR and SSIM) and visual perception metrics (LPIPS [64] and FSIM [65]). PSNR measures the differences of pixel intensity between outputs and ground truths. SSIM can provide the similarly measurement from the luminance, contrast and structure aspects. LPIPS utilizes the deep features to measure the perceptual similarity incline with human visual perception. FSIM is also used for the evaluation, which represents the salient low-level features based on phase congruency and gradient magnitude.\"\n },\n {\n \"figure_ref\": [\n \"fig_1\",\n \"fig_2\",\n \"fig_3\"\n ],\n \"heading\": \"B. General Multi-Exposure Image Fusion\",\n \"publication_ref\": [\n \"b10\",\n \"b36\",\n \"b32\",\n \"b16\",\n \"b19\",\n \"b42\",\n \"b44\",\n \"b26\",\n \"b11\",\n \"b45\",\n \"b28\",\n \"b34\"\n ],\n \"table_ref\": [],\n \"text\": \"To demonstrate the effectiveness and remarkable advantages of the proposed methods, we comprehensively compare our methods with eleven competitors including the traditional methods, i.e., MESPD [11], FMMEF [37], MEFDSIFT [33], DEMEF [17] and deep learning methods, i.e., DeepFuse [20], IFCNN [43], SDNet [45], U2Fusion [27], MEFGAN [12], HALDER [46] and DPEMEF [29]. Then we evaluate the proposed scheme with these methods from three aspects, i.e., the objective, subjective comparison, and computation complexity.\\n1) Subjective Comparison: We select two representative pairs to demonstrate the superiority of our proposed method, which are shown at Fig. 3 and Fig. 4. In the below part of each image, we also illustrate related signal maps of the line marked by \\\"blue\\\" compared with ground truth, to show the obvious differences. Since these image pairs have a challenging distinct exposure gap, it is untoward to design the multi-exposure fusion scheme to preserve the suitable brightness, structural details, and abundant color distribution. From these noticeable comparisons, we can conclude our methods contain three significant advantages.\\nFirstly, our scheme realizes the consistency of global brightness. Patch-based schemes (e.g., MESPD, DEMEF, and FM- MEF) cannot achieve brightness consistency, which introduces significant artifacts, caused by the extreme difference in exposure time. From the signal maps of DEMEF and FMMEF at the first scene, the fused results are over-exposure with higher pixel intensities compared with ground truth. Meanwhile, the sky is degraded by extreme-low exposure time, shown in the local under-exposed regions. In contrast, our method can recover the promising brightness with normal illumination. Secondly, current learning-based schemes are easily trended to color distortion. For instance, SDNet, U2Fusion, and HALDER methods cannot realize the vivid color details, including the bushes in the first scene and the flowers in the second scene. This illustration is also reflected in the corresponding signal maps. These methods cannot achieve large signal changes and are with a moderate reflection of pixel intensity. Our method and MEFGAN can preserve the promising color distribution with remarkable improvement. More importantly, the maintenance of texture details is a critical goal for this fusion task. From the two blew-up regions, it is obvious to observe that, most fusion schemes can not maintain sufficient details, e.g., the cliff painting and tower.\\nEspecially in the second scene, corrupted by a large difference in exposure time and severe illumination, it is difficult to recover the concrete details by simple feature aggregation and direct fusion principles. Our results are visual-friendly, which is an incline with the human vision system.\\nOn the other hand, we also provide another comparison based on the dataset [35] without ground truth in Fig. 5. We select two pairs with extreme exposure variance to illustrate the effectiveness of our scheme with nine state-ofthe-art multi-exposed image fusion. As shown in the first sequence, the information (e.g., cloud layer) under low exposure cannot be recovered clearly. Thus, these details are hard to be highlighted and recovered from under-exposed images (e.g., SDNet and U2Fusion marked by the green boxes). We can clearly observe that attention-based methods (MEFGAN, HALDER, and Ours) can achieve abundant detail preservations. Especially, our network can effectively promote visual perception to render sufficient details. Furthermore, our method can accomplish vivid color enhancement. Most of the results appear in the local over-exposure region. In contrast, our method achieves abundant texture details (e.g., bushes) and consistent color distribution (e.g., computer screen). Either existing patch-based optimization methods or learning-based schemes cannot address the global consistency of illumination. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"2) Objective Comparison:\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"To demonstrate the superiority of the proposed scheme, we utilize four different metrics, including PSNR, SSIM, LPIPS, and FSIM to measure the visual quality of diverse methods. The whole numerical results are reported in Table . I. We introduce two versions to conduct the comparison, which are named \\\"Ours\\\" and \\\"Ours * \\\" respectively. The difference between both versions is utilizing diverse latency constraints, where \\\"Ours * \\\" focuses more on the inference time. Obviously, our scheme achieves consistent performance improvement in terms of these metrics. Compared with representative learning-based schemes MEFGAN and HALDER, our scheme promotes 0.8 dB and 1.0 dB drastically. On the other hand, it can be clearly seen that we obtain the second best numerical results. However, the patch-based fusion scheme obtains the suboptimal numerical performance, which indicates our scheme can effectively preserve sufficient textural details and structural information. We also utilize the LPIPS to measure the distortion at feature levels. Our scheme can reduce almost 24.6% of LPIPS compared with HALDER, which demonstrates better visual quality in line with the human perception system. Obviously, our scheme also achieves the best results.\\n3) Computation Efficiency Analyses: We also conduct a comparison under computation efficiency, which is a critical point for real-world deployment. The concert numerical results among these competitors under the metrics of parameters and runtime on the SICE dataset are reported in Table . II. Obviously, most learning-based methods achieve fast inference time than traditional numerical-based schemes (e.g., DEMEF and MEFDSIFT). Furthermore, though SDNet and IFCNN have fewer parameters, and are limited by direct fusion rules and inefficient architectures, their performances are interfered. More importantly, both our methods realize the faster inference time. Compared with the latest learning-based method DPEMEF, the fastest version (ours * ) can effectively promote 68.11% runtime and reduce 98.68% parameters, which demonstrates high efficiency with visual-pleasant fused results.\\n4) Fusion with Arbitrary Exposure Ratios: In order to verify the generalization ability to address the inputs with arbitrary exposure ratios, we select two representative sequences, which is shown in Fig. 6. Since we utilize the pairs with the Fig. 6. Visual results about the source image sequences with different exposure ratios. ( 1) and ( 2) are under-exposed images, (3), ( 4) and ( 5) are over-exposed images. (a)-(c) are fused by inputs of ( 1) and ( 3)-( 5). (d)-(f) are fused by inputs of ( 2) and ( 3)-( 5).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Under Exposure\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Under Exposure\\nOver Exposure Over exposure ratios for training and testing, the proposed scheme is robust enough to handle different exposure ratios. These sequences contain two under-exposed images and three over-exposed images. We can clearly observe that the scheme can obtain the consistent visual-pleasant fused results with natural color correction, generated by the image pairs with diverse exposure ranges. Several characteristics can be found from the fused result of this sequence. Firstly, the proposed scheme is with large capacity for wide exposure difference to preserve the textural details and maintain suitable illumination distribution. Secondly, though without the training procedure of small exposure difference, the fused images generated by (a) and (b) contain sufficient scene details, e.g., grasses and sunset glow. Finally, our result (i.e., (c)) fused by the large difference of exposure is close to the ground truth. Several fused images (e.g., (b) and (d)) have more vivid scene representation compared with ground truth.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C. Misaligned Multi-Exposure Image Fusion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Misaligned multi-exposure image fusion is a challenging scenario due to the camera movement and device shaking, whereas current methods for MEF are easy to generate blurs without the consideration of pixel registration. We also illustrate the performance by numerical and visual comparisons.\"\n },\n {\n \"figure_ref\": [\n \"fig_4\"\n ],\n \"heading\": \"1) Visual Comparison:\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We select four misaligned pairs from the testing dataset to demonstrate the effectiveness of proposed framework in Fig. 7. The first two rows illustrate the scene with large pixel movements. Most methods cannot preserve sufficient details. The results generated by MEF-GAN and FMMEF methods have obvious artifacts and cannot preserve the normal color distribution. More importantly, our scheme successfully realizes the uniform promotion of pixel alignment and visual enhancement, which can effectively address diverse levels of pixel alignment. 2) Numerical Comparison: In Table . III, we report the numerical results compared with various representative multiexposure image fusion methods. We utilize five typical metrics to measure the visual quality of fused images. Since existing methods often assume the image pairs are well-registered, these methods cannot obtain promising quantitative results. We can observe that the proposed method achieves the consistent best numerical results in terms of these four metrics. These results demonstrate the superiority of our methods on the misaligned MEF. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"V. ABLATION STUDY\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this section, we conduct sufficient experiments with numerical and visual evaluations to verify the effectiveness of proposed modules, loss functions and architecture search.\"\n },\n {\n \"figure_ref\": [\n \"fig_5\",\n \"fig_5\"\n ],\n \"heading\": \"A. Effectiveness of Scene Relighting.\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this part, we first validate the effectiveness of proposed SRSM and validate the suitable cascaded numbers for MEF task. The role of scene relighting is to gradually preserve the scene information and constrain the level of illumination for following feature aggregation. The ablation experiment about the cascaded number of SRSM is conducted, where the quantitative and qualitative comparisons are shown at Table . IV and Fig. 8 respectively. Firstly, we illustrate the necessary of proposed SRSM. The version without SRSM only concatenate the inputs to feed into the DRM, without the procedure of illumination adjustment. We can clearly observe that, directly processing compromises the image quality, which reduce the numerical performance drastically. As shown in Fig. 8, we can obtain the output image is over-exposed, cannot render sufficient detail and preserve the normal light distribution. Compared by pie charts, which depicts the proportion of RGB channels. The version w/o SRSM cannot restore the normal color distribution, which leads to the distortion of color and details. Then we evaluate the cascaded number of SRSM. By introducing the cascaded SRSMs, we propose the recurrent attention mechanisms to extract the sufficient features. Cascading two modules can achieve the best numerical performance. Increasing the number of SRSM obtain the moderate improvement. \"\n },\n {\n \"figure_ref\": [\n \"fig_4\"\n ],\n \"heading\": \"B. Effectiveness of Deformable Alignment\",\n \"publication_ref\": [\n \"b56\"\n ],\n \"table_ref\": [],\n \"text\": \"We further evaluate the advantages of self-alignment mechanism. Three variants of comparisons are conducted, including \\\"w/o DASM\\\", optical flow-based and changing the position before relighting. Self-alignment module targets to align the unregistered pixels of image pairs. In the architecture construction, we put DASM after the SRSM. Optical flow-based schemes is to firstly utilize the RAFT [57] to align image pairs. Numerical results and visual comparisons are depicted at Table . V and Fig. 7 respectively. From the numerical results, we can observe the effectiveness of proposed mechanisms. Our scheme improves 2.65dB of PSNR, 23.55% of VIF and 7.16% of FSIM. From the visual comparison, the fused result of \\\"w/o DASM\\\" cannot preserve enough texture details, such as the grasses. On the other hand, we can see that, the optical flow-based schemes cannot improve the quantitative results due to the inaccurate motion estimation caused by different illumination. The fused images contains more obvious artifacts. Meanwhile, we also discuss the suitable position of relighting. Obviously, the variant \\\"Before SRSM\\\" obtains the sub-optimal numerical and visual results compared with the final scheme, but has remarkable performance promotion compared with previous two variants. Furthermore, based on the similar intensity distribution by SRSM, the final scheme can realize the consistent improvement. \"\n },\n {\n \"figure_ref\": [\n \"fig_6\",\n \"fig_6\",\n \"fig_6\"\n ],\n \"heading\": \"C. Training Losses Analyses\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"also perform the detailed analyses to evaluate the effectiveness of diverse training strategies (i.e., combinations of loss functions). In this part, we gradually introduce the loss functions to composited three schemes, including \\\"w/ Int \\\", \\\"w/ Int + Gra \\\" and our scheme. Related visual results are plotted at Fig. 10. From the objective comparison, Gra can effectively preserve the edge information, which reflects on the structural measurement SSIM and feature-level metric LPIPS. As shown in Fig. 10, visual results scheme \\\" w/ Int + Gra \\\" provide flourishing textural details, e.g., the details of grasses and floors. Meanwhile, introducing Gra can effectively remove the artifact such as the shape of tree on the first row. Our scheme is combined with three categories of losses, i.e., Int for pixel intensity, Gra for structural detail and Dis for color distribution. Thus, our scheme can further improve the visual quality, obtaining with the highest numerical results. For instance, our final scheme obtain the vivid color distribution, without any color distortion (e.g., the color of wall), shown at the second row in Fig. 10.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Inputs\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Int\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Int+ Gra Total \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"D. Search Space Analyses\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We also verify the basic properties of search space, where the concrete performances are reported at Table. VII. From the fusion performance, we can observe that, dilated convolutions (3-DC and 7-DC) have higher quantitative results (e.g., PSNR, LPIPS and FSIM) compared with normal convolutions. On the other hand, we can directly observe that 3 × 3 convolution has the fast inference speed but has the sub-optimal statistical results. Under the constrained of hardware latency, our searched scheme actually achieve the balance between visual quality and inference speed.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"E. Hardware-sensitive Analyses\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We also analyse the influence of trade-off parameter η, which controls the influences of hardware-sensitive latency constraint. The numerical results are reported in Table . VIII. \\\"C32\\\" and \\\"C64\\\" denote the version with 32 and 64 channels. When η = 0.5, the balance between fusion quality and inference requirement can be guaranteed simultaneously. The concrete architectures under diverse η are plotted in Fig. 11. The previous three layers illustrate the architecture of SAM. The last four layers shows the structure of DRM with residual connection. In detail, without the constraint of latency, NAS scheme chooses operator with large receptive field to better capture the large features. Moreover, we also can conclude that 3 × 3 convolution can effectively extract features with high efficiency, which is widely leveraged for SRSM under the latency constraint. As for DRM, {3-C, 1-C} with skip connection is a low-weight combination for the detail compensation, as shown at the subfigure (c) and (d). \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"VI. CONCLUDING REMARKS\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this paper, we proposed a robust multi-exposure image fusion framework to address various scenarios, including the aligned and misaligned image pairs. We divided the fusion procedure into two parts: self-alignment for feature-wise alignment and detail repletion to enhance texture details visually. By utilizing a hardware-friendly architecture search strategy and incorporating a task-oriented search space, we discovered a highly efficient and compact architecture for MEF. Furthermore, we conducted comprehensive subjective and objective comparisons to demonstrate the outstanding performance of our method compared to various state-of-the-arts.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The source code will be available at https://github.com/LiuZhu-CV\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This work is partially supported by the National Key R&D Program of China (2020YFB1313503), the National Natural Science Foundation of China (U22B2052), the Fundamental Research Funds for the Central Universities and the Major Key Project of PCL (PCL2021A12).\"\n }\n]"},"abstract":{"kind":"string","value":"In recent years, deep learning-based methods have achieved remarkable progress in multi-exposure image fusion. However, existing methods rely on aligned image pairs, inevitably generating artifacts when faced with device shaking in real-world scenarios. Moreover, these learning-based methods are built on handcrafted architectures and operations by increasing network depth or width, neglecting different exposure characteristics. As a result, these direct cascaded architectures with redundant parameters fail to achieve highly effective inference time and lead to massive computation. To alleviate these issues, in this paper, we propose a search-based paradigm, involving selfalignment and detail repletion modules for robust multi-exposure image fusion. By utilizing scene relighting and deformable convolutions, the self-alignment module can accurately align images despite camera movement. Furthermore, by imposing a hardware-sensitive constraint, we introduce neural architecture search to discover compact and efficient networks, investigating effective feature representation for fusion. We realize the stateof-the-art performance in comparison to various competitive schemes, yielding a 4.02% and 29.34% improvement in PSNR for general and misaligned scenarios, respectively, while reducing inference time by 68.1%."},"title":{"kind":"string","value":"Embracing Compact and Robust Architectures for Multi-Exposure Image Fusion"},"figures":{"kind":"list like","value":[{"figure_caption":"Fig. 2 .2Fig. 2. Schematic diagram of the proposed architecture. The super-architecture for multi-exposure fusion consists of Self-Alignment Module (SAM) and Detail Repletion Module (DRM). Search spaces of SAM and for DRM are also illustrated respectively.","figure_data":"","figure_id":"fig_0","figure_label":"2","figure_type":"figure"},{"figure_caption":"Fig. 3 .3Fig. 3. Qualitative comparison with nine state-of-the-art methods. The signal maps provide the differences of pixel intensity with the ground truth.","figure_data":"","figure_id":"fig_1","figure_label":"3","figure_type":"figure"},{"figure_caption":"Fig. 4 .4Fig. 4. Qualitative comparison with nine state-of-the-art methods. The signal maps provide the differences of pixel intensity with ground truth.","figure_data":"","figure_id":"fig_2","figure_label":"4","figure_type":"figure"},{"figure_caption":"Fig. 5 .5Fig. 5. Qualitative comparison with nine state-of-the-art methods on the dataset [35] without ground truth.","figure_data":"","figure_id":"fig_3","figure_label":"5","figure_type":"figure"},{"figure_caption":"Fig. 7 .7Fig. 7. Qualitative results compared with the state-of-the-arts on the misaligned multi-exposure fusion.","figure_data":"","figure_id":"fig_4","figure_label":"7","figure_type":"figure"},{"figure_caption":"Fig. 8 .8Fig. 8. Visual comparison of effectiveness for SRSM.","figure_data":"","figure_id":"fig_5","figure_label":"8","figure_type":"figure"},{"figure_caption":"Fig. 10 .10Fig. 10. Visual results and error maps obtained by the different loss functions.","figure_data":"","figure_id":"fig_6","figure_label":"10","figure_type":"figure"},{"figure_caption":"Trade= 5 Fig. 11 .511Fig. 11. Heatmaps of the searched architectures based on different trade-off parameter η. The selected operators are marked by red boxes.","figure_data":"","figure_id":"fig_7","figure_label":"511","figure_type":"figure"},{"figure_caption":"COMPARISON OF PROPOSED METHODS WITH A SERIES OF MULTI-EXPOSURE FUSION SCHEMES.","figure_data":"5314.5015.1013.2417.5619.14 17.4217.6719.7119.9119.2320.71 ↑4.02% 20.54SSIM ↑ 0.7180.7660.8080.7450.7420.795 0.7530.7180.7570.7630.8440.8250.822LPIPS ↓ 0.2270.2310.2090.2070.2640.190 0.2480.2240.2730.1750.1430.132 ↓7.69% 0.138FSIM ↑ 0.8800.8770.8830.8710.8990.887 0.8290.8530.9060.9210.8860.924 ↑0.32% 0.924","figure_id":"tab_1","figure_label":"I","figure_type":"table"},{"figure_caption":"EFFICIENCY COMPARISON INCLUDING PARAMETERS AND AVERAGED RUNTIME ON THE SICE DATASET.","figure_data":"MetricsMESPD FMMEF MEFDSIFT DEMEF DeepFuse IFCNN SDNetU2Fusion MEFGAN HALDER DPEMEF Ours Ours *PlatformMatlab MatlabMatlabMatlab Tensorflow Pytorch Tensorflow Tensorflow Tensorflow Pytorch Pytorch Pytorch PytorchDeviceCPUCPUCPUCPUGPUGPUGPUGPUGPUGPUGPUGPUGPUParameters (M) ↓----0.0890.0830.0670.6593.1574.71151.931.879 0.684Runtime (s) ↓0.1830.2350.5570.2170.2030.0220.8850.3050.8620.1410.0680.047 0.021TABLE IIINUMERICAL RESULTS COMPARED WITH REPRESENTATIVE METHODS FOR MISALIGNED MULTI-EXPOSURE IMAGE FUSION.Metrics MESPD FMMEF MEFDSIFT DEMEF DeepFuse IFCNN SDNet U2Fusion MEFGAN HALDER DPEMEFOursPSNR ↑13.1215.1113.3613.7312.1817.0416.3416.71↑0.4230.4540.4400.4360.3880.4240.4340.4640.4640.4950.4350.681 ↑37.6%LPIPS ↓ 0.2800.2650.2860.2770.3840.2930.2660.2850.3590.2700.2660.187 ↓29.4%FSIM ↑0.7960.8000.7920.7930.7920.7760.7850.8010.8010.8210.8070.913 ↑11.2%","figure_id":"tab_2","figure_label":"II","figure_type":"table"},{"figure_caption":"COMPARISONS OF EFFECTIVENESS WITH SRSM.","figure_data":"Numberw/o SRSM Cascade-1 Cascade-2 Cascaded-3PSNR ↑9.5720.1520.7120.56SSIM ↑0.6110.8190.8250.825LPIPS ↓0.4070.1500.1320.136FSIM ↑0.7870.9320.9240.929","figure_id":"tab_3","figure_label":"IV","figure_type":"table"},{"figure_caption":"COMPARISONS OF EFFECTIVENESS WITH DASM.","figure_data":"Metricw/o DASMFlow-BasedBefore RelightingOursPSNR ↑19.3919.1921.7322.04SSIM ↑0.5610.5400.6750.681LPIPS ↓0.3250.3310.2370.187FSIM ↑0.8520.8380.9070.913","figure_id":"tab_4","figure_label":"V","figure_type":"table"},{"figure_caption":"THE EFFECTIVENESS OF LOSS FUNCTIONS.","figure_data":"Metricw/ Intw/ Int+ Graw/ Int+ DisTotalPSNR ↑20.6220.6320.6620.71SSIM ↑0.7990.8230.8210.825LPIPS ↓0.1450.1340.1360.132FSIM ↑0.9320.9130.9270.924TABLE VIIQUANTITATIVE COMPARISION OF SEARCH SPACE.3-C20.660.8240.1320.9260.0423-DC20.680.8250.1310.9310.0445-C20.580.8220.1340.9300.1245-DC20.490.8180.1350.9300.0887-C20.520.8200.1350.9280.1957-DC20.800.8230.1400.9330.153Ours20.710.8250.1320.9240.047","figure_id":"tab_5","figure_label":"VI","figure_type":"table"},{"figure_caption":"-OFF η FOR HARDWARE-SENSITIVE ANALYSIS.","figure_data":"","figure_id":"tab_6","figure_label":"VIII","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Fig. 2 .2Fig. 2. Schematic diagram of the proposed architecture. The super-architecture for multi-exposure fusion consists of Self-Alignment Module (SAM) and Detail Repletion Module (DRM). Search spaces of SAM and for DRM are also illustrated respectively.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 3 .3Fig. 3. Qualitative comparison with nine state-of-the-art methods. The signal maps provide the differences of pixel intensity with the ground truth.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 4 .4Fig. 4. Qualitative comparison with nine state-of-the-art methods. The signal maps provide the differences of pixel intensity with ground truth.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 5 .5Fig. 5. Qualitative comparison with nine state-of-the-art methods on the dataset [35] without ground truth.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"5\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 7 .7Fig. 7. Qualitative results compared with the state-of-the-arts on the misaligned multi-exposure fusion.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"7\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 8 .8Fig. 8. Visual comparison of effectiveness for SRSM.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_5\",\n \"figure_label\": \"8\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 10 .10Fig. 10. Visual results and error maps obtained by the different loss functions.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_6\",\n \"figure_label\": \"10\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Trade= 5 Fig. 11 .511Fig. 11. Heatmaps of the searched architectures based on different trade-off parameter η. The selected operators are marked by red boxes.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_7\",\n \"figure_label\": \"511\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"COMPARISON OF PROPOSED METHODS WITH A SERIES OF MULTI-EXPOSURE FUSION SCHEMES.\",\n \"figure_data\": \"5314.5015.1013.2417.5619.14 17.4217.6719.7119.9119.2320.71 ↑4.02% 20.54SSIM ↑ 0.7180.7660.8080.7450.7420.795 0.7530.7180.7570.7630.8440.8250.822LPIPS ↓ 0.2270.2310.2090.2070.2640.190 0.2480.2240.2730.1750.1430.132 ↓7.69% 0.138FSIM ↑ 0.8800.8770.8830.8710.8990.887 0.8290.8530.9060.9210.8860.924 ↑0.32% 0.924\",\n \"figure_id\": \"tab_1\",\n \"figure_label\": \"I\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"EFFICIENCY COMPARISON INCLUDING PARAMETERS AND AVERAGED RUNTIME ON THE SICE DATASET.\",\n \"figure_data\": \"MetricsMESPD FMMEF MEFDSIFT DEMEF DeepFuse IFCNN SDNetU2Fusion MEFGAN HALDER DPEMEF Ours Ours *PlatformMatlab MatlabMatlabMatlab Tensorflow Pytorch Tensorflow Tensorflow Tensorflow Pytorch Pytorch Pytorch PytorchDeviceCPUCPUCPUCPUGPUGPUGPUGPUGPUGPUGPUGPUGPUParameters (M) ↓----0.0890.0830.0670.6593.1574.71151.931.879 0.684Runtime (s) ↓0.1830.2350.5570.2170.2030.0220.8850.3050.8620.1410.0680.047 0.021TABLE IIINUMERICAL RESULTS COMPARED WITH REPRESENTATIVE METHODS FOR MISALIGNED MULTI-EXPOSURE IMAGE FUSION.Metrics MESPD FMMEF MEFDSIFT DEMEF DeepFuse IFCNN SDNet U2Fusion MEFGAN HALDER DPEMEFOursPSNR ↑13.1215.1113.3613.7312.1817.0416.3416.71↑0.4230.4540.4400.4360.3880.4240.4340.4640.4640.4950.4350.681 ↑37.6%LPIPS ↓ 0.2800.2650.2860.2770.3840.2930.2660.2850.3590.2700.2660.187 ↓29.4%FSIM ↑0.7960.8000.7920.7930.7920.7760.7850.8010.8010.8210.8070.913 ↑11.2%\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"II\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"COMPARISONS OF EFFECTIVENESS WITH SRSM.\",\n \"figure_data\": \"Numberw/o SRSM Cascade-1 Cascade-2 Cascaded-3PSNR ↑9.5720.1520.7120.56SSIM ↑0.6110.8190.8250.825LPIPS ↓0.4070.1500.1320.136FSIM ↑0.7870.9320.9240.929\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"IV\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"COMPARISONS OF EFFECTIVENESS WITH DASM.\",\n \"figure_data\": \"Metricw/o DASMFlow-BasedBefore RelightingOursPSNR ↑19.3919.1921.7322.04SSIM ↑0.5610.5400.6750.681LPIPS ↓0.3250.3310.2370.187FSIM ↑0.8520.8380.9070.913\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"V\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"THE EFFECTIVENESS OF LOSS FUNCTIONS.\",\n \"figure_data\": \"Metricw/ Intw/ Int+ Graw/ Int+ DisTotalPSNR ↑20.6220.6320.6620.71SSIM ↑0.7990.8230.8210.825LPIPS ↓0.1450.1340.1360.132FSIM ↑0.9320.9130.9270.924TABLE VIIQUANTITATIVE COMPARISION OF SEARCH SPACE.3-C20.660.8240.1320.9260.0423-DC20.680.8250.1310.9310.0445-C20.580.8220.1340.9300.1245-DC20.490.8180.1350.9300.0887-C20.520.8200.1350.9280.1957-DC20.800.8230.1400.9330.153Ours20.710.8250.1320.9240.047\",\n \"figure_id\": \"tab_5\",\n \"figure_label\": \"VI\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"-OFF η FOR HARDWARE-SENSITIVE ANALYSIS.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_6\",\n \"figure_label\": \"VIII\",\n \"figure_type\": \"table\"\n }\n]"},"authors":{"kind":"string","value":"Zhu Liu; Jinyuan Liu; Guanyao Wu; Risheng Liu"},"references":{"kind":"list like","value":[{"authors":"X Zhang","journal":"Information Fusion","ref_id":"b0","title":"Benchmarking and comparing multi-exposure image fusion algorithms","year":"2021"},{"authors":"L Wang; K.-J Yoon","journal":"IEEE transactions on pattern analysis and machine intelligence","ref_id":"b1","title":"Deep learning for hdr imaging: State-ofthe-art and future trends","year":"2021"},{"authors":"J Luo; W Ren; X Gao; X Cao","journal":"IEEE Transactions on Image Processing","ref_id":"b2","title":"Multi-exposure image fusion via deformable self-attention","year":"2023"},{"authors":"T Ma; L Ma; X Fan; Z Luo; R Liu","journal":"","ref_id":"b3","title":"Pia: Parallel architecture with illumination allocator for joint enhancement and detection in lowlight","year":"2022"},{"authors":"J Liu; X Fan; Z Huang; G Wu; R Liu; W Zhong; Z Luo","journal":"","ref_id":"b4","title":"Target-aware dual adversarial learning and a multi-scenario multimodality benchmark to fuse infrared and visible for object detection","year":"2022"},{"authors":"R Liu; L Ma; J Zhang; X Fan; Z Luo","journal":"","ref_id":"b5","title":"Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement","year":"2021"},{"authors":"R Liu; L Ma; T Ma; X Fan; Z Luo","journal":"IEEE Transactions on Pattern Analysis and Machine Intelligence","ref_id":"b6","title":"Learning with nested scene modeling and cooperative architecture search for low-light vision","year":"2022"},{"authors":"H Zhang; J Ma","journal":"Information Fusion","ref_id":"b7","title":"Iid-mef: A multi-exposure fusion network based on intrinsic image decomposition","year":"2023"},{"authors":"L Ma; T Ma; R Liu; X Fan; Z Luo","journal":"","ref_id":"b8","title":"Toward fast, flexible, and robust low-light image enhancement","year":"2022"},{"authors":"X Xue; J He; L Ma; Y Wang; X Fan; R Liu","journal":"","ref_id":"b9","title":"Best of both worlds: See and understand clearly in the dark","year":"2022"},{"authors":"H Li; T N Chan; X Qi; W Xie","journal":"IEEE Transactions on Circuits and Systems for Video Technology","ref_id":"b10","title":"Detail-preserving multiexposure fusion with edge-preserving structural patch decomposition","year":"2021"},{"authors":"H Xu; J Ma; X.-P Zhang","journal":"IEEE Transactions on Image Processing","ref_id":"b11","title":"Mef-gan: Multi-exposure image fusion via generative adversarial networks","year":"2020"},{"authors":"P E Debevec; J Malik","journal":"","ref_id":"b12","title":"Recovering high dynamic range radiance maps from photographs","year":"2008"},{"authors":"J Munkberg; P Clarberg; J Hasselgren; T Akenine-Möller","journal":"ACM Transactions on Graphics (TOG)","ref_id":"b13","title":"High dynamic range texture compression for graphics hardware","year":"2006"},{"authors":"S K Nayar; T Mitsunaga","journal":"IEEE","ref_id":"b14","title":"High dynamic range imaging: Spatially varying pixel exposures","year":"2000"},{"authors":"S Li; X Kang; J Hu","journal":"IEEE Transactions on Image processing","ref_id":"b15","title":"Image fusion with guided filtering","year":"2013"},{"authors":"Q Wang; W Chen; X Wu; Z Li","journal":"IEEE Transactions on Circuits and Systems for Video Technology","ref_id":"b16","title":"Detail-enhanced multi-scale exposure fusion in yuv color space","year":"2020"},{"authors":"N Hayat; M Imran","journal":"Journal of Visual Communication and Image Representation","ref_id":"b17","title":"Ghost-free multi exposure image fusion technique using dense sift descriptor and guided filter","year":"2019"},{"authors":"S Hu; W Zhang","journal":"IEEE","ref_id":"b18","title":"Exploiting patch-based correlation for ghost removal in exposure fusion","year":"2017"},{"authors":"K Ram Prabhakar; V Sai; R Srikar; Venkatesh; Babu","journal":"","ref_id":"b19","title":"Deepfuse: A deep unsupervised approach for exposure fusion with extreme exposure image pairs","year":"2017"},{"authors":"J Liu; J Shang; R Liu; X Fan","journal":"IEEE Transactions on Circuits and Systems for Video Technology","ref_id":"b20","title":"Attention-guided global-local adversarial learning for detail-preserving multi-exposure image fusion","year":"2022"},{"authors":"J Li; J Liu; S Zhou; Q Zhang; N K Kasabov","journal":"IEEE Transactions on Circuits and Systems for Video Technology","ref_id":"b21","title":"Learning a coordinated network for detail-refinement multi-exposure image fusion","year":"2022"},{"authors":"J Liu; G Wu; J Luan; Z Jiang; R Liu; X Fan","journal":"Information Fusion","ref_id":"b22","title":"Holoco: Holistic and local contrastive learning network for multi-exposure image fusion","year":"2023"},{"authors":"W Zhang; X Liu; W Wang; Y Zeng","journal":"International Journal of Advanced Robotic Systems","ref_id":"b23","title":"Multi-exposure image fusion based on wavelet transform","year":"2018"},{"authors":"Y Yang; D Zhang; W Wan; S Huang","journal":"IEEE Transactions on Instrumentation and Measurement","ref_id":"b24","title":"Multi-scale exposure fusion based on multi-visual feature measurement and detail enhancement representation","year":"2022"},{"authors":"S Li; B Yang; J Hu","journal":"Information Fusion","ref_id":"b25","title":"Performance comparison of different multiresolution transforms for image fusion","year":"2011"},{"authors":"H Xu; J Ma; J Jiang; X Guo; H Ling","journal":"IEEE Transactions on Pattern Analysis and Machine Intelligence","ref_id":"b26","title":"U2fusion: A unified unsupervised image fusion network","year":"2022"},{"authors":"Z Li; J Liu; R Liu; X Fan; Z Luo; W Gao","journal":"IEEE","ref_id":"b27","title":"Multiple taskoriented encoders for unified image fusion","year":"2021"},{"authors":"D Han; L Li; X Guo; J Ma","journal":"Information Fusion","ref_id":"b28","title":"Multi-exposure image fusion via deep perceptual enhancement","year":"2022"},{"authors":"J J Lewis; R J O'callaghan; S G Nikolov; D R Bull; N Canagarajah","journal":"Information fusion","ref_id":"b29","title":"Pixel-and region-based image fusion with complex wavelets","year":"2007"},{"authors":"M Qiguang; W Baoshu","journal":"IEEE","ref_id":"b30","title":"A novel image fusion method using contourlet transform","year":"2006"},{"authors":"J Shen; Y Zhao; S Yan; X Li","journal":"IEEE Trans. Cybern","ref_id":"b31","title":"Exposure fusion using boosting laplacian pyramid","year":"2014"},{"authors":"Y Liu; Z Wang","journal":"Journal of Visual Communication and Image Representation","ref_id":"b32","title":"Dense sift for ghost-free multi-exposure fusion","year":"2015"},{"authors":"S Li; X Kang; L Fang; J Hu; H Yin","journal":"information Fusion","ref_id":"b33","title":"Pixel-level image fusion: A survey of the state of the art","year":"2017"},{"authors":"K Ma; H Li; H Yong; Z Wang; D Meng; L Zhang","journal":"IEEE Transactions on Image Processing","ref_id":"b34","title":"Robust multiexposure image fusion: a structural patch decomposition approach","year":"2017"},{"authors":"K Ma; Z Wang","journal":"IEEE","ref_id":"b35","title":"Multi-exposure image fusion: A patch-wise approach","year":"2015"},{"authors":"H Li; K Ma; H Yong; L Zhang","journal":"IEEE Transactions on Image Processing","ref_id":"b36","title":"Fast multi-scale structural patch decomposition for multi-exposure image fusion","year":"2020"},{"authors":"F Kou; Z Li; C Wen; W Chen","journal":"IEEE","ref_id":"b37","title":"Multi-scale exposure fusion via gradient domain guided image filtering","year":"2017"},{"authors":"","journal":"Journal of Visual Communication and Image Representation","ref_id":"b38","title":"Edge-preserving smoothing pyramid based multi-scale exposure fusion","year":"2018"},{"authors":"Q Yan; J Sun; H Li; Y Zhu; Y Zhang","journal":"Neurocomputing","ref_id":"b39","title":"High dynamic range imaging by sparse representation","year":"2017"},{"authors":"J Wang; H Liu; N He","journal":"Neurocomputing","ref_id":"b40","title":"Exposure fusion based on sparse representation using approximate k-svd","year":"2014"},{"authors":"K Ma; Z Duanmu; H Zhu; Y Fang; Z Wang","journal":"IEEE Transactions on Image Processing","ref_id":"b41","title":"Deep guided learning for fast multi-exposure image fusion","year":"2019"},{"authors":"Y Zhang; Y Liu; P Sun; H Yan; X Zhao; L Zhang","journal":"Information Fusion","ref_id":"b42","title":"Ifcnn: A general image fusion framework based on convolutional neural network","year":"2020"},{"authors":"K Ma; K Zeng; Z Wang","journal":"IEEE Transactions on Image Processing","ref_id":"b43","title":"Perceptual quality assessment for multi-exposure image fusion","year":"2015"},{"authors":"H Zhang; J Ma","journal":"International Journal of Computer Vision","ref_id":"b44","title":"Sdnet: A versatile squeeze-and-decomposition network for real-time image fusion","year":"2021"},{"authors":"J Liu; J Shang; R Liu; X Fan","journal":"IEEE","ref_id":"b45","title":"Halder: Hierarchical attentionguided learning with detail-refinement for multi-exposure image fusion","year":"2021"},{"authors":"Q Yan; D Gong; Q Shi; A V D Hengel; C Shen; I Reid; Y Zhang","journal":"","ref_id":"b46","title":"Attention-guided network for ghost-free high dynamic range imaging","year":"2019"},{"authors":"M Guo; Z Zhong; W Wu; D Lin; J Yan","journal":"","ref_id":"b47","title":"Irlas: Inverse reinforcement learning for architecture search","year":"2019"},{"authors":"Y Gao; H Yang; P Zhang; C Zhou; Y Hu","journal":"","ref_id":"b48","title":"Graph neural architecture search","year":"2020"},{"authors":"Y Chen; G Meng; Q Zhang; S Xiang; C Huang; L Mu; X Wang","journal":"","ref_id":"b49","title":"Renas: Reinforced evolutionary neural architecture search","year":"2019"},{"authors":"X Zhang; Z Huang; N Wang; S Xiang; C Pan","journal":"IEEE Transactions on Pattern Analysis and Machine Intelligence","ref_id":"b50","title":"You only search once: Single shot neural architecture search via direct sparse optimization","year":"2020"},{"authors":"H Liu; K Simonyan; Y Yang","journal":"","ref_id":"b51","title":"Darts: Differentiable architecture search","year":"2018"},{"authors":"R Liu; J Gao; J Zhang; D Meng; Z Lin","journal":"IEEE Transactions on Pattern Analysis and Machine Intelligence","ref_id":"b52","title":"Investigating bilevel optimization for learning and vision from a unified perspective: A survey and beyond","year":"2021"},{"authors":"H Zhang; Y Li; H Chen; C Gong; Z Bai; C Shen","journal":"International Journal of Computer Vision","ref_id":"b53","title":"Memoryefficient hierarchical neural architecture search for image restoration","year":"2022"},{"authors":"R Li; R T Tan; L.-F Cheong","journal":"","ref_id":"b54","title":"All in one bad weather removal using architectural search","year":"2020"},{"authors":"R Liu; Z Liu; J Liu; X Fan","journal":"","ref_id":"b55","title":"Searching a hierarchically aggregated fusion architecture for fast multi-modality image fusion","year":"2021"},{"authors":"Z Teed; J Deng","journal":"Springer","ref_id":"b56","title":"Raft: Recurrent all-pairs field transforms for optical flow","year":"2020"},{"authors":"D Wang; J Liu; X Fan; R Liu","journal":"","ref_id":"b57","title":"Unsupervised misaligned infrared and visible image fusion via cross-modality image generation and registration","year":"2022"},{"authors":"Z Huang; J Liu; X Fan; R Liu; W Zhong; Z Luo","journal":"","ref_id":"b58","title":"Reconet: Recurrent correction network for fast and efficient multi-modality image fusion","year":"2022"},{"authors":"X Wang; K C Chan; K Yu; C Dong; C Change Loy","journal":"","ref_id":"b59","title":"Edvr: Video restoration with enhanced deformable convolutional networks","year":"2019"},{"authors":"A Liu; X Liu; J Fan; Y Ma; A Zhang; H Xie; D Tao","journal":"","ref_id":"b60","title":"Perceptualsensitive gan for generating adversarial patches","year":"2019"},{"authors":"I Gulrajani; F Ahmed; M Arjovsky; V Dumoulin; A C Courville","journal":"Advances in neural information processing systems","ref_id":"b61","title":"Improved training of wasserstein gans","year":"2017"},{"authors":"J Cai; S Gu; L Zhang","journal":"IEEE Transactions on Image Processing","ref_id":"b62","title":"Learning a deep single image contrast enhancer from multi-exposure images","year":"2018"},{"authors":"R Zhang; P Isola; A A Efros; E Shechtman; O Wang","journal":"","ref_id":"b63","title":"The unreasonable effectiveness of deep features as a perceptual metric","year":"2018"},{"authors":"L Zhang; L Zhang; X Mou; D Zhang","journal":"IEEE transactions on Image Processing","ref_id":"b64","title":"Fsim: A feature similarity index for image quality assessment","year":"2011"}],"string":"[\n {\n \"authors\": \"X Zhang\",\n \"journal\": \"Information Fusion\",\n \"ref_id\": \"b0\",\n \"title\": \"Benchmarking and comparing multi-exposure image fusion algorithms\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"L Wang; K.-J Yoon\",\n \"journal\": \"IEEE transactions on pattern analysis and machine intelligence\",\n \"ref_id\": \"b1\",\n \"title\": \"Deep learning for hdr imaging: State-ofthe-art and future trends\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"J Luo; W Ren; X Gao; X Cao\",\n \"journal\": \"IEEE Transactions on Image Processing\",\n \"ref_id\": \"b2\",\n \"title\": \"Multi-exposure image fusion via deformable self-attention\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"T Ma; L Ma; X Fan; Z Luo; R Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"Pia: Parallel architecture with illumination allocator for joint enhancement and detection in lowlight\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"J Liu; X Fan; Z Huang; G Wu; R Liu; W Zhong; Z Luo\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"Target-aware dual adversarial learning and a multi-scenario multimodality benchmark to fuse infrared and visible for object detection\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"R Liu; L Ma; J Zhang; X Fan; Z Luo\",\n \"journal\": \"\",\n \"ref_id\": \"b5\",\n \"title\": \"Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"R Liu; L Ma; T Ma; X Fan; Z Luo\",\n \"journal\": \"IEEE Transactions on Pattern Analysis and Machine Intelligence\",\n \"ref_id\": \"b6\",\n \"title\": \"Learning with nested scene modeling and cooperative architecture search for low-light vision\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"H Zhang; J Ma\",\n \"journal\": \"Information Fusion\",\n \"ref_id\": \"b7\",\n \"title\": \"Iid-mef: A multi-exposure fusion network based on intrinsic image decomposition\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"L Ma; T Ma; R Liu; X Fan; Z Luo\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"Toward fast, flexible, and robust low-light image enhancement\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"X Xue; J He; L Ma; Y Wang; X Fan; R Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"Best of both worlds: See and understand clearly in the dark\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"H Li; T N Chan; X Qi; W Xie\",\n \"journal\": \"IEEE Transactions on Circuits and Systems for Video Technology\",\n \"ref_id\": \"b10\",\n \"title\": \"Detail-preserving multiexposure fusion with edge-preserving structural patch decomposition\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"H Xu; J Ma; X.-P Zhang\",\n \"journal\": \"IEEE Transactions on Image Processing\",\n \"ref_id\": \"b11\",\n \"title\": \"Mef-gan: Multi-exposure image fusion via generative adversarial networks\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"P E Debevec; J Malik\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"Recovering high dynamic range radiance maps from photographs\",\n \"year\": \"2008\"\n },\n {\n \"authors\": \"J Munkberg; P Clarberg; J Hasselgren; T Akenine-Möller\",\n \"journal\": \"ACM Transactions on Graphics (TOG)\",\n \"ref_id\": \"b13\",\n \"title\": \"High dynamic range texture compression for graphics hardware\",\n \"year\": \"2006\"\n },\n {\n \"authors\": \"S K Nayar; T Mitsunaga\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b14\",\n \"title\": \"High dynamic range imaging: Spatially varying pixel exposures\",\n \"year\": \"2000\"\n },\n {\n \"authors\": \"S Li; X Kang; J Hu\",\n \"journal\": \"IEEE Transactions on Image processing\",\n \"ref_id\": \"b15\",\n \"title\": \"Image fusion with guided filtering\",\n \"year\": \"2013\"\n },\n {\n \"authors\": \"Q Wang; W Chen; X Wu; Z Li\",\n \"journal\": \"IEEE Transactions on Circuits and Systems for Video Technology\",\n \"ref_id\": \"b16\",\n \"title\": \"Detail-enhanced multi-scale exposure fusion in yuv color space\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"N Hayat; M Imran\",\n \"journal\": \"Journal of Visual Communication and Image Representation\",\n \"ref_id\": \"b17\",\n \"title\": \"Ghost-free multi exposure image fusion technique using dense sift descriptor and guided filter\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"S Hu; W Zhang\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b18\",\n \"title\": \"Exploiting patch-based correlation for ghost removal in exposure fusion\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"K Ram Prabhakar; V Sai; R Srikar; Venkatesh; Babu\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"Deepfuse: A deep unsupervised approach for exposure fusion with extreme exposure image pairs\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"J Liu; J Shang; R Liu; X Fan\",\n \"journal\": \"IEEE Transactions on Circuits and Systems for Video Technology\",\n \"ref_id\": \"b20\",\n \"title\": \"Attention-guided global-local adversarial learning for detail-preserving multi-exposure image fusion\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"J Li; J Liu; S Zhou; Q Zhang; N K Kasabov\",\n \"journal\": \"IEEE Transactions on Circuits and Systems for Video Technology\",\n \"ref_id\": \"b21\",\n \"title\": \"Learning a coordinated network for detail-refinement multi-exposure image fusion\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"J Liu; G Wu; J Luan; Z Jiang; R Liu; X Fan\",\n \"journal\": \"Information Fusion\",\n \"ref_id\": \"b22\",\n \"title\": \"Holoco: Holistic and local contrastive learning network for multi-exposure image fusion\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"W Zhang; X Liu; W Wang; Y Zeng\",\n \"journal\": \"International Journal of Advanced Robotic Systems\",\n \"ref_id\": \"b23\",\n \"title\": \"Multi-exposure image fusion based on wavelet transform\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Y Yang; D Zhang; W Wan; S Huang\",\n \"journal\": \"IEEE Transactions on Instrumentation and Measurement\",\n \"ref_id\": \"b24\",\n \"title\": \"Multi-scale exposure fusion based on multi-visual feature measurement and detail enhancement representation\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"S Li; B Yang; J Hu\",\n \"journal\": \"Information Fusion\",\n \"ref_id\": \"b25\",\n \"title\": \"Performance comparison of different multiresolution transforms for image fusion\",\n \"year\": \"2011\"\n },\n {\n \"authors\": \"H Xu; J Ma; J Jiang; X Guo; H Ling\",\n \"journal\": \"IEEE Transactions on Pattern Analysis and Machine Intelligence\",\n \"ref_id\": \"b26\",\n \"title\": \"U2fusion: A unified unsupervised image fusion network\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Z Li; J Liu; R Liu; X Fan; Z Luo; W Gao\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b27\",\n \"title\": \"Multiple taskoriented encoders for unified image fusion\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"D Han; L Li; X Guo; J Ma\",\n \"journal\": \"Information Fusion\",\n \"ref_id\": \"b28\",\n \"title\": \"Multi-exposure image fusion via deep perceptual enhancement\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"J J Lewis; R J O'callaghan; S G Nikolov; D R Bull; N Canagarajah\",\n \"journal\": \"Information fusion\",\n \"ref_id\": \"b29\",\n \"title\": \"Pixel-and region-based image fusion with complex wavelets\",\n \"year\": \"2007\"\n },\n {\n \"authors\": \"M Qiguang; W Baoshu\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b30\",\n \"title\": \"A novel image fusion method using contourlet transform\",\n \"year\": \"2006\"\n },\n {\n \"authors\": \"J Shen; Y Zhao; S Yan; X Li\",\n \"journal\": \"IEEE Trans. Cybern\",\n \"ref_id\": \"b31\",\n \"title\": \"Exposure fusion using boosting laplacian pyramid\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"Y Liu; Z Wang\",\n \"journal\": \"Journal of Visual Communication and Image Representation\",\n \"ref_id\": \"b32\",\n \"title\": \"Dense sift for ghost-free multi-exposure fusion\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"S Li; X Kang; L Fang; J Hu; H Yin\",\n \"journal\": \"information Fusion\",\n \"ref_id\": \"b33\",\n \"title\": \"Pixel-level image fusion: A survey of the state of the art\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"K Ma; H Li; H Yong; Z Wang; D Meng; L Zhang\",\n \"journal\": \"IEEE Transactions on Image Processing\",\n \"ref_id\": \"b34\",\n \"title\": \"Robust multiexposure image fusion: a structural patch decomposition approach\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"K Ma; Z Wang\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b35\",\n \"title\": \"Multi-exposure image fusion: A patch-wise approach\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"H Li; K Ma; H Yong; L Zhang\",\n \"journal\": \"IEEE Transactions on Image Processing\",\n \"ref_id\": \"b36\",\n \"title\": \"Fast multi-scale structural patch decomposition for multi-exposure image fusion\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"F Kou; Z Li; C Wen; W Chen\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b37\",\n \"title\": \"Multi-scale exposure fusion via gradient domain guided image filtering\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"Journal of Visual Communication and Image Representation\",\n \"ref_id\": \"b38\",\n \"title\": \"Edge-preserving smoothing pyramid based multi-scale exposure fusion\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Q Yan; J Sun; H Li; Y Zhu; Y Zhang\",\n \"journal\": \"Neurocomputing\",\n \"ref_id\": \"b39\",\n \"title\": \"High dynamic range imaging by sparse representation\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"J Wang; H Liu; N He\",\n \"journal\": \"Neurocomputing\",\n \"ref_id\": \"b40\",\n \"title\": \"Exposure fusion based on sparse representation using approximate k-svd\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"K Ma; Z Duanmu; H Zhu; Y Fang; Z Wang\",\n \"journal\": \"IEEE Transactions on Image Processing\",\n \"ref_id\": \"b41\",\n \"title\": \"Deep guided learning for fast multi-exposure image fusion\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Y Zhang; Y Liu; P Sun; H Yan; X Zhao; L Zhang\",\n \"journal\": \"Information Fusion\",\n \"ref_id\": \"b42\",\n \"title\": \"Ifcnn: A general image fusion framework based on convolutional neural network\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"K Ma; K Zeng; Z Wang\",\n \"journal\": \"IEEE Transactions on Image Processing\",\n \"ref_id\": \"b43\",\n \"title\": \"Perceptual quality assessment for multi-exposure image fusion\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"H Zhang; J Ma\",\n \"journal\": \"International Journal of Computer Vision\",\n \"ref_id\": \"b44\",\n \"title\": \"Sdnet: A versatile squeeze-and-decomposition network for real-time image fusion\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"J Liu; J Shang; R Liu; X Fan\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b45\",\n \"title\": \"Halder: Hierarchical attentionguided learning with detail-refinement for multi-exposure image fusion\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Q Yan; D Gong; Q Shi; A V D Hengel; C Shen; I Reid; Y Zhang\",\n \"journal\": \"\",\n \"ref_id\": \"b46\",\n \"title\": \"Attention-guided network for ghost-free high dynamic range imaging\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"M Guo; Z Zhong; W Wu; D Lin; J Yan\",\n \"journal\": \"\",\n \"ref_id\": \"b47\",\n \"title\": \"Irlas: Inverse reinforcement learning for architecture search\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Y Gao; H Yang; P Zhang; C Zhou; Y Hu\",\n \"journal\": \"\",\n \"ref_id\": \"b48\",\n \"title\": \"Graph neural architecture search\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Y Chen; G Meng; Q Zhang; S Xiang; C Huang; L Mu; X Wang\",\n \"journal\": \"\",\n \"ref_id\": \"b49\",\n \"title\": \"Renas: Reinforced evolutionary neural architecture search\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"X Zhang; Z Huang; N Wang; S Xiang; C Pan\",\n \"journal\": \"IEEE Transactions on Pattern Analysis and Machine Intelligence\",\n \"ref_id\": \"b50\",\n \"title\": \"You only search once: Single shot neural architecture search via direct sparse optimization\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"H Liu; K Simonyan; Y Yang\",\n \"journal\": \"\",\n \"ref_id\": \"b51\",\n \"title\": \"Darts: Differentiable architecture search\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"R Liu; J Gao; J Zhang; D Meng; Z Lin\",\n \"journal\": \"IEEE Transactions on Pattern Analysis and Machine Intelligence\",\n \"ref_id\": \"b52\",\n \"title\": \"Investigating bilevel optimization for learning and vision from a unified perspective: A survey and beyond\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"H Zhang; Y Li; H Chen; C Gong; Z Bai; C Shen\",\n \"journal\": \"International Journal of Computer Vision\",\n \"ref_id\": \"b53\",\n \"title\": \"Memoryefficient hierarchical neural architecture search for image restoration\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"R Li; R T Tan; L.-F Cheong\",\n \"journal\": \"\",\n \"ref_id\": \"b54\",\n \"title\": \"All in one bad weather removal using architectural search\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"R Liu; Z Liu; J Liu; X Fan\",\n \"journal\": \"\",\n \"ref_id\": \"b55\",\n \"title\": \"Searching a hierarchically aggregated fusion architecture for fast multi-modality image fusion\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Z Teed; J Deng\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b56\",\n \"title\": \"Raft: Recurrent all-pairs field transforms for optical flow\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"D Wang; J Liu; X Fan; R Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b57\",\n \"title\": \"Unsupervised misaligned infrared and visible image fusion via cross-modality image generation and registration\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Z Huang; J Liu; X Fan; R Liu; W Zhong; Z Luo\",\n \"journal\": \"\",\n \"ref_id\": \"b58\",\n \"title\": \"Reconet: Recurrent correction network for fast and efficient multi-modality image fusion\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"X Wang; K C Chan; K Yu; C Dong; C Change Loy\",\n \"journal\": \"\",\n \"ref_id\": \"b59\",\n \"title\": \"Edvr: Video restoration with enhanced deformable convolutional networks\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"A Liu; X Liu; J Fan; Y Ma; A Zhang; H Xie; D Tao\",\n \"journal\": \"\",\n \"ref_id\": \"b60\",\n \"title\": \"Perceptualsensitive gan for generating adversarial patches\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"I Gulrajani; F Ahmed; M Arjovsky; V Dumoulin; A C Courville\",\n \"journal\": \"Advances in neural information processing systems\",\n \"ref_id\": \"b61\",\n \"title\": \"Improved training of wasserstein gans\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"J Cai; S Gu; L Zhang\",\n \"journal\": \"IEEE Transactions on Image Processing\",\n \"ref_id\": \"b62\",\n \"title\": \"Learning a deep single image contrast enhancer from multi-exposure images\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"R Zhang; P Isola; A A Efros; E Shechtman; O Wang\",\n \"journal\": \"\",\n \"ref_id\": \"b63\",\n \"title\": \"The unreasonable effectiveness of deep features as a perceptual metric\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"L Zhang; L Zhang; X Mou; D Zhang\",\n \"journal\": \"IEEE transactions on Image Processing\",\n \"ref_id\": \"b64\",\n \"title\": \"Fsim: A feature similarity index for image quality assessment\",\n \"year\": \"2011\"\n }\n]"},"formulas":{"kind":"list like","value":[{"formula_coordinates":[4,118.74,706.98,181.29,12.19],"formula_id":"formula_0","formula_text":"I S i = I i ⊗ S(I i ), i ∈ {U; O},(1)"},{"formula_coordinates":[4,390.5,496.27,172.53,9.88],"formula_id":"formula_1","formula_text":"F A = A(F U , F O ) + F O ,(2)"},{"formula_coordinates":[4,400.78,739.05,162.25,9.88],"formula_id":"formula_2","formula_text":"y = F A R(F A ).(3)"},{"formula_coordinates":[5,80.97,489.53,219.05,20.09],"formula_id":"formula_3","formula_text":"F j = O i→j (F i ); O i→j = O∈O α i→j O(F i ),(4)"},{"formula_coordinates":[5,97.98,597.14,202.04,20.09],"formula_id":"formula_4","formula_text":"R(α; LAT) = M O∈O α i→j LAT(O),(5)"},{"formula_coordinates":[5,109.93,687.19,190.1,16.21],"formula_id":"formula_5","formula_text":"min α val (α; ω * ) + ηR(α; LAT),(6)"},{"formula_coordinates":[5,376.12,162.19,186.92,9.85],"formula_id":"formula_6","formula_text":"Total = Int + β 1 Gra + β 2 Dis .(7)"},{"formula_coordinates":[5,391.23,230.32,171.81,22.31],"formula_id":"formula_7","formula_text":"Int = 1 HW y -y gt 1 ,(8)"},{"formula_coordinates":[5,382.93,306.87,180.11,22.31],"formula_id":"formula_8","formula_text":"Gra = 1 HW ∇y -∇y gt 2 ,(9)"},{"formula_coordinates":[5,311.98,432.54,256.09,22.98],"formula_id":"formula_9","formula_text":"E x∼Pfake D(y)-E x∼Preal D(y gt )+ηE x∼Pfake [( ∇ y D(y) 2 -1) 2 ].(10)"}],"string":"[\n {\n \"formula_coordinates\": [\n 4,\n 118.74,\n 706.98,\n 181.29,\n 12.19\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"I S i = I i ⊗ S(I i ), i ∈ {U; O},(1)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 390.5,\n 496.27,\n 172.53,\n 9.88\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"F A = A(F U , F O ) + F O ,(2)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 400.78,\n 739.05,\n 162.25,\n 9.88\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"y = F A R(F A ).(3)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 80.97,\n 489.53,\n 219.05,\n 20.09\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"F j = O i→j (F i ); O i→j = O∈O α i→j O(F i ),(4)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 97.98,\n 597.14,\n 202.04,\n 20.09\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"R(α; LAT) = M O∈O α i→j LAT(O),(5)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 109.93,\n 687.19,\n 190.1,\n 16.21\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"min α val (α; ω * ) + ηR(α; LAT),(6)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 376.12,\n 162.19,\n 186.92,\n 9.85\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"Total = Int + β 1 Gra + β 2 Dis .(7)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 391.23,\n 230.32,\n 171.81,\n 22.31\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"Int = 1 HW y -y gt 1 ,(8)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 382.93,\n 306.87,\n 180.11,\n 22.31\n ],\n \"formula_id\": \"formula_8\",\n \"formula_text\": \"Gra = 1 HW ∇y -∇y gt 2 ,(9)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 311.98,\n 432.54,\n 256.09,\n 22.98\n ],\n \"formula_id\": \"formula_9\",\n \"formula_text\": \"E x∼Pfake D(y)-E x∼Preal D(y gt )+ηE x∼Pfake [( ∇ y D(y) 2 -1) 2 ].(10)\"\n }\n]"}}},{"rowIdx":382,"cells":{"doi":{"kind":"string","value":"10.48550/arxiv.2011.00362"},"pub_date":{"kind":"string","value":"2023-05-20"},"sections":{"kind":"list like","value":[{"figure_ref":[],"heading":"INTRODUCTION","publication_ref":[],"table_ref":[],"text":"The key challenge that we seek to address is that of identifying learned features in a model's intermediate representations that are more or less likely to be robust to distributional shift. Our approach starts from the intuition that for high-entropy features in a model's training distribution, it will have learned a better understanding for when the feature is relevant. More precisely, it will be better at distinguishing the presence or absence of the feature across different situations. Consider a hypothetical data set containing photographs from two safari trips, where each trip contains the same people on the same safari, but driving around in different trucks. Suppose that it is useful for the given task to identify which of the two trips a given image corresponds to; we might expect the model to be particularly good at distinguishing between the trucks. On the other hand, if a rare tree appears in exactly one photograph, the model may have learned to recognise the specific pattern of pixels in that photograph corresponding to the tree, but it might not have the capability to recognise the tree in new situations.\nAs models have increased in performance within the bounds of the i.i.d. assumption, recent years have seen growing interest in the OOD behaviour of machine learning systems. While many approaches have studied OOD detection or the effects of external changes to a model's training regime on OOD behaviour (e.g. domain randomization or auxiliary loss functions), to the best of our knowledge our proposal of the entropy of an intermediate representation as a guide to its effects OOD is a novel approach. In this paper we demonstrate that the removal of low-entropy representations via the masking of learned discrete bits can notably improve OOD performance."},{"figure_ref":["fig_0","fig_3"],"heading":"TASK AND MODEL DESCRIPTION","publication_ref":["b12"],"table_ref":[],"text":"To learn representations of a domain we train an encoder network to produce a representation r of a given input x * . This representation is given to a distinguisher network that is tasked with identifying x * from a set of k images composed of x * and k -1 distractor inputs arranged randomly. We use the CIFAR-10 dataset (Krizhevsky, 2009) as the training distribution. The labels from the dataset are discarded and an unsupervised k-contrast task is constructed by pairing each image with k -1 distractor images, shuffling, and giving the distinguisher k inputs to choose from. The same preprocessing is later used when out-of-distribution datasets are introduced. See Figure 1for an example of a contrastive task and Figure 4 in the Supplementary Material for the full architecture. It is important to note that we use a 'soft-discretization' technique (Foerster et al., 2016) on the intermediate representation r such that it can be learned with gradient-descent, but each dimension can be mapped to a binary digit at test time with no loss in performance. While the use of a communication channel to discretize representations poses optimization challenges, it also provides a large benefit when it comes to computing the entropy values of each bit in the representation. The computation is reduced from approximating an integral to the simple formula for the entropy of a binary variable, as outlined in Section 3.1. This allows us to run a greater number of experiments with higher precision than if we had used continuous representations. This unsupervised contrastive learning task was chosen as it can be easily transferred to different data distributions. A task such as image classification limits the available datasets as it requires the out-of-distribution testing data to have the same (or at least overlapping) image labels."},{"figure_ref":[],"heading":"ENTROPY-BASED MASKING","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":["fig_5"],"heading":"ENTROPY OF REPRESENTATION BITS","publication_ref":[],"table_ref":[],"text":"Each representation r produced by an encoder network consists of a number of bits |r|, referred to as the representation length. By considering each bit at index i as a random variable B i we can compute the binary entropy of the bit on a given dataset D:\nH(B i | D) = -p log 2 p -(1 -p) log 2 (1 -p), where p = P (B i = 1 | D).\n(1)\nEntropy close to 1 means that the bit is 0 or 1 with roughly equal probability of p = 0.5. Very low entropy means that the bit is either almost always 0 or almost always 1. We notice that for smaller representation lengths and/or few distractors the distribution tends to skew towards higher entropy bits. In separate experiments where we further varied representation lengths, we find that for smaller |r| equal to 8, 16 or 32, all bits have entropy higher than 0.8, which makes studying bits based on entropy variation uninteresting for these representation lengths. For a visualization of these entropy values see Figure 5 in the Supplementary Material. Representation lengths of 64, 128, 256 and 512 all lead to a wide range of entropy values. A theoretical analysis of the optimal bit-entropy can be found in Section B of the Supplementary Material."},{"figure_ref":[],"heading":"BIT MASKING STRATEGIES","publication_ref":[],"table_ref":[],"text":"In this paper we are interested in the effects of strategically 'removing' parts of the model's intermediate representation, i.e. obscuring bits in r. It is important to note that we are only applying masking at test time. The masking is not used to train any of the models. The mask is defined by a set masking variables m i ∈ {0, 1} for each bit r i in the representation. The masked bit ri is computed:\nri = m i r i + (1 -m i ) 1 2 . (2\n)\nIn other words, when the masking variable m i = 0 then ri = 0.5, and otherwise ri = r i . In this paper we use three masking strategies; Random Masking, Top-Entropy Masking, and Bottom-Entropy Masking. In order to construct a mask with any of these strategies, we define a masking proportion p mask that represents the percentage of bits in r that should be masked.\nTo construct any mask M = {m 1 , . . . , m |r| } we will need to choose l mask = p mask • |r| bits to remove. For a random mask we draw l mask masking variables from M at random with uniform probability and without replacement, and set them to 0, we set the remaining |r| -l mask variables to 1. To construct a top-entropy mask we compute the entropy for each bit h i = H(B i | D) and sort these values in descending order. We then take the bits associated with the first l mask entropy values (i.e. highest entropy) and set their corresponding masking variables to zero. Likewise, for the bottom-entropy mask we take the last l mask bits and remove those instead."},{"figure_ref":[],"heading":"EXPERIMENTAL RESULTS","publication_ref":["b16"],"table_ref":[],"text":"Figure 2: Accuracy of CIFAR-10 pre-trained models on OOD datasets (on the y-axis) against accuracy on CIFAR-10 (on the x-axis). The dashed line (which coincides with the green and blue lines) is the y = x line.\nWe trained 54 encoder-distinguisher pairs 1 on CIFAR-10 and removed models that did not converge, resulting in 51 trained models. Models were trained with varying combinations of representation lengths and number of distractors:\n(|r|, k) ∈ {64, 128, 256, 512} × {3, 5, 10, 20}.\nSee Table 1 for the test accuracy statistics for the models on the k-contrast CIFAR-10 training distributions. See Section C in the Supplementary Material for a full description of the training methodology.\nTo evaluate the effects of distributional shifts we test our 51 trained models on the CIFAR-100 Krizhevsky ( 2009 In Figure 2 we demonstrate the shift in performance that results from applying the models to the new datasets. Following Taori et al. (Taori et al., 2020), plotting the relationship between InD and OOD performance in this manner allows us to study distributional shift while controlling for the variations in initial accuracy. The y = x line is plotted with a black dashed line, however, it is obscured by the regression lines for CIFAR-100 and Stanford Online Products. This tells us that there is no distributional shift for these datasets, i.e. no loss in performance. For this reason, we drop these datasets from all further out-of-distribution analysis. For the other datasets, we see in order of increased degradation: Plant Village, Colorectal Histology, and MNIST."},{"figure_ref":[],"heading":"ANALYSIS OF MASKING EFFECTS IN-DISTRIBUTION","publication_ref":[],"table_ref":[],"text":"Before moving onto the out-of-distribution case, we first examine the effects of applying the different masking strategies to the models that we trained on CIFAR-10, with the CIFAR-10 test data. For each of the 51 successfully trained models we evaluated the accuracy without any masking, and with each 1 A sweep of 3 runs for each pair of (|r|, k) plus 6 initial separate runs. of the different masking strategies for masking proportions between 0.15 and 0.5 at 0.05 intervals. We found that for any masking proportion, removing the top-entropy bits is more damaging to accuracy than masking out bottom-entropy bits. In light of general insights from information theory, this result is not too surprising. The highest entropy bits necessarily convey the most information, and so it follows that their removal should lead to the largest drop in performance.\nIn general, we did not expect any of the masking strategies to provide a benefit when applied within the training distribution. Yet, we saw that with a small masking proportion (around p mask < 0.3) we see an increase in accuracy for low-entropy and random masks. Our initial hypothesis was that the masking may be 'undoing' overfitting to the training set. But for each of the trained models we have verified that there is no overfitting (see Section D.1 in the Supplementary Material for a visualization)."},{"figure_ref":["fig_2"],"heading":"ANALYSIS OF MASKING EFFECTS OUT-OF-DISTRIBUTION (OOD)","publication_ref":["b6","b2","b2"],"table_ref":["tab_1"],"text":"In order to understand the effects of masking on accuracy in the OOD setting we measure the mean change in accuracy of a masking strategy under various circumstances. We also report the standard deviations associated with these estimates. As in the case of in-distribution masking we evaluated the masking strategies for a sweep of masking proportions between 0.15 and 0.5 at 0.05 intervals. We cut-off the maximum masking proportion p mask ≤ 0.25 for all further analysis as beyond that threshold masking has an almost universally negative effect. The overall mean accuracy changes can be seen in Table 2. We see that masking the bottom-entropy or random bits produces the highest increase, albeit with a large variance.\nThis variance can be understood and disentangled by separating the low-k models from the high-k models. What we see is that the benefits of bottom-entropy masking are more prevalent for low-k models. This is visualized in Figure 3 where we illustrate the effective robustness of each of the masking strategies on the three OOD datasets. In the Supplementary Material Section D.2 we include plots for all values of k and p mask that we tested. Effective robustness is a concept introduced by Taori et al. ( 2020) as a way to understand the efficacy of a method for increasing robustness to distributional shift. By plotting the baseline regression line for unaltered models with differing in-distribution accuracy values on the diagram we can observe whether a proposed robustness method moves towards the y = x line (i.e. no degradation). Crucially, with these plots, we are able to account for each model's performance on the training distribution. Hence, despite the large variance in the performance of models trained across various k and |r| values2 , we are able to discern the effects of the masking interventions.\nIn our case, we see that -as is consistent with previous results -for each dataset the top-entropy masking moves below the dashed green line showing the baseline unmasked models. On the other hand, the random masking and bottom-entropy masking lines move closer to y = x (as compared to the no masking lines). For Plant Village we see that almost all of the in-distribution accuracy is recovered. For MNIST we find the most substantial jump, and the largest benefit of bottom-entropy over random masking. 2021). However, we apply entropy in an entirely different context, namely, we calculate the entropy of latent variables to estimate how robust they will be to distributional shift. Relative entropy (KL-divergence) is a popular measure and is notably used in the Bits-Back method Hinton & van Camp (1993), Flamich et al. (2020) to calculate the optimal compression rate in latent variables. Images that are traditionally compressed by a variational auto-encoder have now been compressed with code-length close to this theoretical optimum Flamich et al. (2020).\nContrastive representation learning takes many forms; in computer vision alone there are many approaches for applying deep learning to multiple inputs and producing representations to distinguish between them; see Jaiswal et al. (2020) for a review. To our knowledge, there are no existing suitable state-of-the-art (SOTA) methods for OOD robustness in contrastive learning to benchmark our proposals against."},{"figure_ref":["fig_2","fig_3","fig_3","fig_0"],"heading":"CONCLUSION","publication_ref":["b19"],"table_ref":[],"text":"In this paper we have investigated the out-of-distribution effects of using different post-hoc strategies to remove bits from discrete intermediate representations in an unsupervised contrastive learning task.\nWe have studied how the difficulty of the task (more distractors) impacts the entropy distribution of the learned representations and shown that removing low-entropy bits can improve the performance of models out-of-distribution (Section 4.2), notably almost entirely restoring in-distribution performance for one of our datasets (see Figure 3). However, the results also present mysteries that prompt further experiments and analysis. At the time of writing, we do not have a clear understanding of why the removal of bits within the training distribution should increase performance, as we would expect the encoder to learn an optimal protocol.\nNext, there is a need for a deeper understanding of the conditions in which our results hold. Within our experimentation, we found that the effect (of harm from low-entropy features OOD) was less pronounced for models trained on the more difficult tasks (higher numbers of distractors). From our data, it is unclear if this relationship represents something fundamental or if it is a side-effect of these models generally performing to a lower standard. One of the most important avenues of further work is in testing if other systems built on top of the learned representations in this paper inherit the same OOD robustness under low-entropy masking.\nA NETWORK ARCHITECTURES The encoder network is composed of a convolutional network (CNN) that takes a 32 × 32 × 3 dimensional tensor as input (CNN A in Figure 4), followed by: a 3 × 3 convolutional layer with 64 filters and ReLU activation; two 3 × 3 convolutional layers with 64 filters, ReLU activation, and a stride-length of 2; a flatten layer; and finally a dense layer without any activation that projects into R |r| , where |r| is a hyperparameter controlling the 'representation length' of r. Next, between the encoder and the distinguisher, there is a discretize/regularize unit (Foerster et al., 2016). Following the literature in which this component was developed, we will refer to this as a communication channel (see in green in Figure 4). The channel is a differentiable unit that, during training, 'soft discretizes' activations passed through it by applying Gaussian white noise (GWN) and a sigmoid function. Then at test time we 'hard discretize' the activations by passing through a sigmoid function and emitting 0 if the result is less than 0.5 and 1 otherwise. This enables the end-to-end learning of a discrete representation via backpropagation from the output of the distinguisher. We configure the channel with a fixed GWN standard deviation of 0.5 during training.\nThe distinguisher network is composed of another convolutional network (CNN B in Figure 1) with exactly the same input and layers as the CNN in the encoder (initialized separately and no parameter sharing), except projecting to a fixed embedding size of 128. This CNN is shared for each of the 'possible answer' images, producing embeddings that are each concatenated with the representation r from the encoder (i.e. the output of the communication channel) and fed into a transformer network (Vaswani et al., 2017) as tokens. The transformer is composed of two self-attention encoder layers with 3 heads of dimension 64, and a dropout rate of 0.1. After the transformer layers each token is projected onto a single dimension without activation. This is then taken as the log-probability (logit) that the corresponding possible answer is correct. The networks are trained together with a sparse categorical crossentropy loss on these logits and the index of the correct answer. The use of a transformer and a shared encoder for the input images means that a model trained, for example, on a 3-contrast dataset (k = 3) can be tested on a 5-contrast dataset without any modification."},{"figure_ref":[],"heading":"B THEORETICAL RESULT","publication_ref":[],"table_ref":[],"text":"Consider the following abstracted and idealized version of the contrastive learning game. An encoder receives an input, and communicates features in that input via bits. A distinguisher has to identify the original input from a set of k (distractor) inputs, based on the communicated features. The encoder and distinguisher win if the distinguisher correctly identifies the original input. The encoder and distinguisher need to decide on a communication protocol before playing the game. Each bit corresponds to one feature. The encoder sends a 1 if a given feature is present and a 0 otherwise.\nThe question we're answering in this section is: what is the optimal feature occurrence (or bit entropy) for a feature when the encoder can choose b bits, and the distinguisher has to choose between k inputs.\nBelow we calculate that the optimal strategy is to use l independent features that are each present in exactly half of the images. The chance of the receiver picking out the right image depends on k.\nFor this calculation we will assume the encoder can choose b = 2 bits, i.e. can communicate two features x and y. Let f x and f y be the frequency of respectively feature x and feature y in the dataset.\nTo answer the question we will calculate the values of f x and f y that maximize the chance of winning.\nLet c x be the random variable that represents: the correct input has feature x, and c y the variable that represents: the correct input has feature y. We assume that these variables are independent. Let v be the random variable that represents the number of inputs in the set of k inputs that the distinguisher gets to see, that have both feature x and feature y. Note that P (c x , c y ) = f x • f y . Below we calculate that\nP (win|c x , c y ) = Σ k v=1 1 v • (f x f y ) v-1 • (1 -f x f y ) k-v • k -1 v -1 .\nTo do so we introduce one more helper variable ṽ which represents the number of inputs in the set of k inputs that the distinguisher gets to see, that have both feature x and feature y, but excluding the correct input.\nWe now calculate\nP (win|c x , c y ) = Σ k v=1 P (win|c x , c y , V = v) • P ( Ṽ = v -1|c x , c y ) = Σ k v=1 P (win|c x , c y , V = v) • P ( Ṽ = v -1)\nNote that P (win|c x , c y , V = v) = 1 v and\nP ( Ṽ = v -1) = (f x f y ) v-1 • (1 -f x f y ) k-1-(v-1) • k -1 v -1 Hence P (win|c x , c y ) = Σ k v=1 1 v • (f x f y ) v-1 • (1 -f x f y ) k-v • k -1 v -1 .\nApplying the Bionomial theorem gives us the following equality\nP (win|c x , c y )P (c x , c y ) = f x f y • Σ k v=1 1 v • (f x f y ) v-1 • (1 -f x f y ) k-v • k -1 v -1 = Σ k v=1 1 v • (f x f y ) v • (1 -f x f y ) k-v • k -1 v -1 = 1 k Σ k v=1 (f x f y ) v • (1 -f x f y ) k-v • k v = 1 k (f x f y + 1 -f x f y ) k -(1 -f x f y ) k = 1 k 1 -(1 -f x f y ) k = 1 k - 1 k (1 -f x f y ) k\nWe can write similar equations for c ¬x and c ¬y and combining them results in\nP (win) = 4 k - 1 k (1 -f x • f y ) k +(1 -f x • (1 -f y )) k +(1 -(1 -f x ) • f y ) k +(1 -(1 -f x ) • (1 -f y )) k .\nMore generally, for arbitrary number of bits b and feature frequencies f 1 , . . . , f b we find\nP (win) = 2 b k - 1 k (1 -f 1 • • • f b ) k + (1 -(1 -f 1 )f 2 • • • f b ) k + . . . + (1 -(1 -f 1 ) • • • (1 -f b )) k\nThe derivative of P (win) with respect to f 1 is\n∂P (win) ∂f 1 = f 2 • • • f b (1 -f 1 • • • f b ) k-1 -f 2 • • • f b (1 -(1 -f 1 )f 2 • • • f b ) k-1 + . . . -(1 -f 2 ) • • • (1 -f b )(1 -(1 -f 1 ) • • • (1 -f b )) k-1\nWhen f 1 = 0.5 the components with a factor of f 1 compensate for the ones with a factor of (1 -f 1 ), and so the derivative is 0 for f 1 = 0.5. Deriving with respect to other feature values gives analogues results. That is, one optimal feature occurrence value for maximizing P (win) is 0.5."},{"figure_ref":["fig_5"],"heading":"C TRAINING METHODOLOGY","publication_ref":["b10","b0"],"table_ref":[],"text":"In order to prevent overfitting and the representation of 'trivial features' (e.g. specific pixel values) in the representations, during training we use a stack of image augmentation layers independently applied prior to each image encoder. This involves a random rotation of up to 0.1 radians, a random contrast shift of up to 10%, a random translation of up to 10% along both axes, and a random zoom of up to 10% (all with a nearest-neighbour filling of blank pixels).\nThe models were optimized using Adam (Kingma & Ba, 2015) with a learning rate of 0.001. The batch size used for training was dependent on the number of distractors, and each epoch iterated through the entire training dataset. See Table 1 for the full breakdown of test accuracy values for trained models, i.e. the mean and standard deviations for the proportion of occasions where the distinguisher was correctly able to identifier x * by using r.\nAll of the code was implemented with Tensorflow 2 (Abadi et al., 2015) and datasets were pulled from Tensorflow Datasets3 (TFDS) (TF Devs, 2022). CIFAR-10 was split into the default TFDS training and test sets (50,000 training images and 10,000 test images). Training and analysis were performed with an NVIDIA RTX 3090 GPU.\nWe trained 54 independent encoder-distinguisher pairs4 for 10 epochs on CIFAR-10 and removed models that did not converge (as defined by not reaching an 80% drop in loss), resulting in 51 trained models (taken as the best performing checkpoint). Models were trained with varying combinations of representation lengths and number of distractors: (|r|, k) ∈ {64, 128, 256, 512} × {3, 5, 10, 20}. We also trained models with representation lengths 8, 16 and 32, visualizations of which can be found in Figure 5, which we discarded because their bit entropies were too homogeneous to meaningfully study the effect of masking out low versus high entropy bits. "},{"figure_ref":[],"heading":"D EXPERIMENTS","publication_ref":[],"table_ref":[],"text":"The code for the experiments can be found at the following repository: [URL removed to preserve anonymity]"},{"figure_ref":["fig_6","fig_7","fig_8","fig_9"],"heading":"D.1 OVERFITTING ANALYSIS","publication_ref":[],"table_ref":[],"text":"In Figure 6 we see that the test and training accuracies are very similar (with the test accuracy even being slightly higher) and so no overfitting has happened. Figure 7 shows the OOD accuracies for each dataset (using the data of all the values of k and all the analysed representation lengths). Figure 8 shows the accuracies for each dataset and each value of k.\nFigure 9 shows the accuracies for each dataset and each representation length |r|. "},{"figure_ref":[],"heading":"ACKNOWLEDGEMENTS","publication_ref":[],"table_ref":[],"text":"Work done by both authors is thanks to the UKRI Centre for Doctoral Training in Safe and Trusted AI (EPSRC Project EP/S023356/1)."},{"figure_ref":[],"heading":"D.3 OOD MEAN ACCURACY CHANGE FROM MASKS","publication_ref":[],"table_ref":[],"text":"The tables in this section are the same as Table 2 in Section ??, except separated by different values of k. Figure 10 is a visualisation of the data along with the 'distance out-of-distribution' for each k value. "}],"string":"[\n {\n \"figure_ref\": [],\n \"heading\": \"INTRODUCTION\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The key challenge that we seek to address is that of identifying learned features in a model's intermediate representations that are more or less likely to be robust to distributional shift. Our approach starts from the intuition that for high-entropy features in a model's training distribution, it will have learned a better understanding for when the feature is relevant. More precisely, it will be better at distinguishing the presence or absence of the feature across different situations. Consider a hypothetical data set containing photographs from two safari trips, where each trip contains the same people on the same safari, but driving around in different trucks. Suppose that it is useful for the given task to identify which of the two trips a given image corresponds to; we might expect the model to be particularly good at distinguishing between the trucks. On the other hand, if a rare tree appears in exactly one photograph, the model may have learned to recognise the specific pattern of pixels in that photograph corresponding to the tree, but it might not have the capability to recognise the tree in new situations.\\nAs models have increased in performance within the bounds of the i.i.d. assumption, recent years have seen growing interest in the OOD behaviour of machine learning systems. While many approaches have studied OOD detection or the effects of external changes to a model's training regime on OOD behaviour (e.g. domain randomization or auxiliary loss functions), to the best of our knowledge our proposal of the entropy of an intermediate representation as a guide to its effects OOD is a novel approach. In this paper we demonstrate that the removal of low-entropy representations via the masking of learned discrete bits can notably improve OOD performance.\"\n },\n {\n \"figure_ref\": [\n \"fig_0\",\n \"fig_3\"\n ],\n \"heading\": \"TASK AND MODEL DESCRIPTION\",\n \"publication_ref\": [\n \"b12\"\n ],\n \"table_ref\": [],\n \"text\": \"To learn representations of a domain we train an encoder network to produce a representation r of a given input x * . This representation is given to a distinguisher network that is tasked with identifying x * from a set of k images composed of x * and k -1 distractor inputs arranged randomly. We use the CIFAR-10 dataset (Krizhevsky, 2009) as the training distribution. The labels from the dataset are discarded and an unsupervised k-contrast task is constructed by pairing each image with k -1 distractor images, shuffling, and giving the distinguisher k inputs to choose from. The same preprocessing is later used when out-of-distribution datasets are introduced. See Figure 1for an example of a contrastive task and Figure 4 in the Supplementary Material for the full architecture. It is important to note that we use a 'soft-discretization' technique (Foerster et al., 2016) on the intermediate representation r such that it can be learned with gradient-descent, but each dimension can be mapped to a binary digit at test time with no loss in performance. While the use of a communication channel to discretize representations poses optimization challenges, it also provides a large benefit when it comes to computing the entropy values of each bit in the representation. The computation is reduced from approximating an integral to the simple formula for the entropy of a binary variable, as outlined in Section 3.1. This allows us to run a greater number of experiments with higher precision than if we had used continuous representations. This unsupervised contrastive learning task was chosen as it can be easily transferred to different data distributions. A task such as image classification limits the available datasets as it requires the out-of-distribution testing data to have the same (or at least overlapping) image labels.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"ENTROPY-BASED MASKING\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [\n \"fig_5\"\n ],\n \"heading\": \"ENTROPY OF REPRESENTATION BITS\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Each representation r produced by an encoder network consists of a number of bits |r|, referred to as the representation length. By considering each bit at index i as a random variable B i we can compute the binary entropy of the bit on a given dataset D:\\nH(B i | D) = -p log 2 p -(1 -p) log 2 (1 -p), where p = P (B i = 1 | D).\\n(1)\\nEntropy close to 1 means that the bit is 0 or 1 with roughly equal probability of p = 0.5. Very low entropy means that the bit is either almost always 0 or almost always 1. We notice that for smaller representation lengths and/or few distractors the distribution tends to skew towards higher entropy bits. In separate experiments where we further varied representation lengths, we find that for smaller |r| equal to 8, 16 or 32, all bits have entropy higher than 0.8, which makes studying bits based on entropy variation uninteresting for these representation lengths. For a visualization of these entropy values see Figure 5 in the Supplementary Material. Representation lengths of 64, 128, 256 and 512 all lead to a wide range of entropy values. A theoretical analysis of the optimal bit-entropy can be found in Section B of the Supplementary Material.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"BIT MASKING STRATEGIES\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this paper we are interested in the effects of strategically 'removing' parts of the model's intermediate representation, i.e. obscuring bits in r. It is important to note that we are only applying masking at test time. The masking is not used to train any of the models. The mask is defined by a set masking variables m i ∈ {0, 1} for each bit r i in the representation. The masked bit ri is computed:\\nri = m i r i + (1 -m i ) 1 2 . (2\\n)\\nIn other words, when the masking variable m i = 0 then ri = 0.5, and otherwise ri = r i . In this paper we use three masking strategies; Random Masking, Top-Entropy Masking, and Bottom-Entropy Masking. In order to construct a mask with any of these strategies, we define a masking proportion p mask that represents the percentage of bits in r that should be masked.\\nTo construct any mask M = {m 1 , . . . , m |r| } we will need to choose l mask = p mask • |r| bits to remove. For a random mask we draw l mask masking variables from M at random with uniform probability and without replacement, and set them to 0, we set the remaining |r| -l mask variables to 1. To construct a top-entropy mask we compute the entropy for each bit h i = H(B i | D) and sort these values in descending order. We then take the bits associated with the first l mask entropy values (i.e. highest entropy) and set their corresponding masking variables to zero. Likewise, for the bottom-entropy mask we take the last l mask bits and remove those instead.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"EXPERIMENTAL RESULTS\",\n \"publication_ref\": [\n \"b16\"\n ],\n \"table_ref\": [],\n \"text\": \"Figure 2: Accuracy of CIFAR-10 pre-trained models on OOD datasets (on the y-axis) against accuracy on CIFAR-10 (on the x-axis). The dashed line (which coincides with the green and blue lines) is the y = x line.\\nWe trained 54 encoder-distinguisher pairs 1 on CIFAR-10 and removed models that did not converge, resulting in 51 trained models. Models were trained with varying combinations of representation lengths and number of distractors:\\n(|r|, k) ∈ {64, 128, 256, 512} × {3, 5, 10, 20}.\\nSee Table 1 for the test accuracy statistics for the models on the k-contrast CIFAR-10 training distributions. See Section C in the Supplementary Material for a full description of the training methodology.\\nTo evaluate the effects of distributional shifts we test our 51 trained models on the CIFAR-100 Krizhevsky ( 2009 In Figure 2 we demonstrate the shift in performance that results from applying the models to the new datasets. Following Taori et al. (Taori et al., 2020), plotting the relationship between InD and OOD performance in this manner allows us to study distributional shift while controlling for the variations in initial accuracy. The y = x line is plotted with a black dashed line, however, it is obscured by the regression lines for CIFAR-100 and Stanford Online Products. This tells us that there is no distributional shift for these datasets, i.e. no loss in performance. For this reason, we drop these datasets from all further out-of-distribution analysis. For the other datasets, we see in order of increased degradation: Plant Village, Colorectal Histology, and MNIST.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"ANALYSIS OF MASKING EFFECTS IN-DISTRIBUTION\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Before moving onto the out-of-distribution case, we first examine the effects of applying the different masking strategies to the models that we trained on CIFAR-10, with the CIFAR-10 test data. For each of the 51 successfully trained models we evaluated the accuracy without any masking, and with each 1 A sweep of 3 runs for each pair of (|r|, k) plus 6 initial separate runs. of the different masking strategies for masking proportions between 0.15 and 0.5 at 0.05 intervals. We found that for any masking proportion, removing the top-entropy bits is more damaging to accuracy than masking out bottom-entropy bits. In light of general insights from information theory, this result is not too surprising. The highest entropy bits necessarily convey the most information, and so it follows that their removal should lead to the largest drop in performance.\\nIn general, we did not expect any of the masking strategies to provide a benefit when applied within the training distribution. Yet, we saw that with a small masking proportion (around p mask < 0.3) we see an increase in accuracy for low-entropy and random masks. Our initial hypothesis was that the masking may be 'undoing' overfitting to the training set. But for each of the trained models we have verified that there is no overfitting (see Section D.1 in the Supplementary Material for a visualization).\"\n },\n {\n \"figure_ref\": [\n \"fig_2\"\n ],\n \"heading\": \"ANALYSIS OF MASKING EFFECTS OUT-OF-DISTRIBUTION (OOD)\",\n \"publication_ref\": [\n \"b6\",\n \"b2\",\n \"b2\"\n ],\n \"table_ref\": [\n \"tab_1\"\n ],\n \"text\": \"In order to understand the effects of masking on accuracy in the OOD setting we measure the mean change in accuracy of a masking strategy under various circumstances. We also report the standard deviations associated with these estimates. As in the case of in-distribution masking we evaluated the masking strategies for a sweep of masking proportions between 0.15 and 0.5 at 0.05 intervals. We cut-off the maximum masking proportion p mask ≤ 0.25 for all further analysis as beyond that threshold masking has an almost universally negative effect. The overall mean accuracy changes can be seen in Table 2. We see that masking the bottom-entropy or random bits produces the highest increase, albeit with a large variance.\\nThis variance can be understood and disentangled by separating the low-k models from the high-k models. What we see is that the benefits of bottom-entropy masking are more prevalent for low-k models. This is visualized in Figure 3 where we illustrate the effective robustness of each of the masking strategies on the three OOD datasets. In the Supplementary Material Section D.2 we include plots for all values of k and p mask that we tested. Effective robustness is a concept introduced by Taori et al. ( 2020) as a way to understand the efficacy of a method for increasing robustness to distributional shift. By plotting the baseline regression line for unaltered models with differing in-distribution accuracy values on the diagram we can observe whether a proposed robustness method moves towards the y = x line (i.e. no degradation). Crucially, with these plots, we are able to account for each model's performance on the training distribution. Hence, despite the large variance in the performance of models trained across various k and |r| values2 , we are able to discern the effects of the masking interventions.\\nIn our case, we see that -as is consistent with previous results -for each dataset the top-entropy masking moves below the dashed green line showing the baseline unmasked models. On the other hand, the random masking and bottom-entropy masking lines move closer to y = x (as compared to the no masking lines). For Plant Village we see that almost all of the in-distribution accuracy is recovered. For MNIST we find the most substantial jump, and the largest benefit of bottom-entropy over random masking. 2021). However, we apply entropy in an entirely different context, namely, we calculate the entropy of latent variables to estimate how robust they will be to distributional shift. Relative entropy (KL-divergence) is a popular measure and is notably used in the Bits-Back method Hinton & van Camp (1993), Flamich et al. (2020) to calculate the optimal compression rate in latent variables. Images that are traditionally compressed by a variational auto-encoder have now been compressed with code-length close to this theoretical optimum Flamich et al. (2020).\\nContrastive representation learning takes many forms; in computer vision alone there are many approaches for applying deep learning to multiple inputs and producing representations to distinguish between them; see Jaiswal et al. (2020) for a review. To our knowledge, there are no existing suitable state-of-the-art (SOTA) methods for OOD robustness in contrastive learning to benchmark our proposals against.\"\n },\n {\n \"figure_ref\": [\n \"fig_2\",\n \"fig_3\",\n \"fig_3\",\n \"fig_0\"\n ],\n \"heading\": \"CONCLUSION\",\n \"publication_ref\": [\n \"b19\"\n ],\n \"table_ref\": [],\n \"text\": \"In this paper we have investigated the out-of-distribution effects of using different post-hoc strategies to remove bits from discrete intermediate representations in an unsupervised contrastive learning task.\\nWe have studied how the difficulty of the task (more distractors) impacts the entropy distribution of the learned representations and shown that removing low-entropy bits can improve the performance of models out-of-distribution (Section 4.2), notably almost entirely restoring in-distribution performance for one of our datasets (see Figure 3). However, the results also present mysteries that prompt further experiments and analysis. At the time of writing, we do not have a clear understanding of why the removal of bits within the training distribution should increase performance, as we would expect the encoder to learn an optimal protocol.\\nNext, there is a need for a deeper understanding of the conditions in which our results hold. Within our experimentation, we found that the effect (of harm from low-entropy features OOD) was less pronounced for models trained on the more difficult tasks (higher numbers of distractors). From our data, it is unclear if this relationship represents something fundamental or if it is a side-effect of these models generally performing to a lower standard. One of the most important avenues of further work is in testing if other systems built on top of the learned representations in this paper inherit the same OOD robustness under low-entropy masking.\\nA NETWORK ARCHITECTURES The encoder network is composed of a convolutional network (CNN) that takes a 32 × 32 × 3 dimensional tensor as input (CNN A in Figure 4), followed by: a 3 × 3 convolutional layer with 64 filters and ReLU activation; two 3 × 3 convolutional layers with 64 filters, ReLU activation, and a stride-length of 2; a flatten layer; and finally a dense layer without any activation that projects into R |r| , where |r| is a hyperparameter controlling the 'representation length' of r. Next, between the encoder and the distinguisher, there is a discretize/regularize unit (Foerster et al., 2016). Following the literature in which this component was developed, we will refer to this as a communication channel (see in green in Figure 4). The channel is a differentiable unit that, during training, 'soft discretizes' activations passed through it by applying Gaussian white noise (GWN) and a sigmoid function. Then at test time we 'hard discretize' the activations by passing through a sigmoid function and emitting 0 if the result is less than 0.5 and 1 otherwise. This enables the end-to-end learning of a discrete representation via backpropagation from the output of the distinguisher. We configure the channel with a fixed GWN standard deviation of 0.5 during training.\\nThe distinguisher network is composed of another convolutional network (CNN B in Figure 1) with exactly the same input and layers as the CNN in the encoder (initialized separately and no parameter sharing), except projecting to a fixed embedding size of 128. This CNN is shared for each of the 'possible answer' images, producing embeddings that are each concatenated with the representation r from the encoder (i.e. the output of the communication channel) and fed into a transformer network (Vaswani et al., 2017) as tokens. The transformer is composed of two self-attention encoder layers with 3 heads of dimension 64, and a dropout rate of 0.1. After the transformer layers each token is projected onto a single dimension without activation. This is then taken as the log-probability (logit) that the corresponding possible answer is correct. The networks are trained together with a sparse categorical crossentropy loss on these logits and the index of the correct answer. The use of a transformer and a shared encoder for the input images means that a model trained, for example, on a 3-contrast dataset (k = 3) can be tested on a 5-contrast dataset without any modification.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B THEORETICAL RESULT\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Consider the following abstracted and idealized version of the contrastive learning game. An encoder receives an input, and communicates features in that input via bits. A distinguisher has to identify the original input from a set of k (distractor) inputs, based on the communicated features. The encoder and distinguisher win if the distinguisher correctly identifies the original input. The encoder and distinguisher need to decide on a communication protocol before playing the game. Each bit corresponds to one feature. The encoder sends a 1 if a given feature is present and a 0 otherwise.\\nThe question we're answering in this section is: what is the optimal feature occurrence (or bit entropy) for a feature when the encoder can choose b bits, and the distinguisher has to choose between k inputs.\\nBelow we calculate that the optimal strategy is to use l independent features that are each present in exactly half of the images. The chance of the receiver picking out the right image depends on k.\\nFor this calculation we will assume the encoder can choose b = 2 bits, i.e. can communicate two features x and y. Let f x and f y be the frequency of respectively feature x and feature y in the dataset.\\nTo answer the question we will calculate the values of f x and f y that maximize the chance of winning.\\nLet c x be the random variable that represents: the correct input has feature x, and c y the variable that represents: the correct input has feature y. We assume that these variables are independent. Let v be the random variable that represents the number of inputs in the set of k inputs that the distinguisher gets to see, that have both feature x and feature y. Note that P (c x , c y ) = f x • f y . Below we calculate that\\nP (win|c x , c y ) = Σ k v=1 1 v • (f x f y ) v-1 • (1 -f x f y ) k-v • k -1 v -1 .\\nTo do so we introduce one more helper variable ṽ which represents the number of inputs in the set of k inputs that the distinguisher gets to see, that have both feature x and feature y, but excluding the correct input.\\nWe now calculate\\nP (win|c x , c y ) = Σ k v=1 P (win|c x , c y , V = v) • P ( Ṽ = v -1|c x , c y ) = Σ k v=1 P (win|c x , c y , V = v) • P ( Ṽ = v -1)\\nNote that P (win|c x , c y , V = v) = 1 v and\\nP ( Ṽ = v -1) = (f x f y ) v-1 • (1 -f x f y ) k-1-(v-1) • k -1 v -1 Hence P (win|c x , c y ) = Σ k v=1 1 v • (f x f y ) v-1 • (1 -f x f y ) k-v • k -1 v -1 .\\nApplying the Bionomial theorem gives us the following equality\\nP (win|c x , c y )P (c x , c y ) = f x f y • Σ k v=1 1 v • (f x f y ) v-1 • (1 -f x f y ) k-v • k -1 v -1 = Σ k v=1 1 v • (f x f y ) v • (1 -f x f y ) k-v • k -1 v -1 = 1 k Σ k v=1 (f x f y ) v • (1 -f x f y ) k-v • k v = 1 k (f x f y + 1 -f x f y ) k -(1 -f x f y ) k = 1 k 1 -(1 -f x f y ) k = 1 k - 1 k (1 -f x f y ) k\\nWe can write similar equations for c ¬x and c ¬y and combining them results in\\nP (win) = 4 k - 1 k (1 -f x • f y ) k +(1 -f x • (1 -f y )) k +(1 -(1 -f x ) • f y ) k +(1 -(1 -f x ) • (1 -f y )) k .\\nMore generally, for arbitrary number of bits b and feature frequencies f 1 , . . . , f b we find\\nP (win) = 2 b k - 1 k (1 -f 1 • • • f b ) k + (1 -(1 -f 1 )f 2 • • • f b ) k + . . . + (1 -(1 -f 1 ) • • • (1 -f b )) k\\nThe derivative of P (win) with respect to f 1 is\\n∂P (win) ∂f 1 = f 2 • • • f b (1 -f 1 • • • f b ) k-1 -f 2 • • • f b (1 -(1 -f 1 )f 2 • • • f b ) k-1 + . . . -(1 -f 2 ) • • • (1 -f b )(1 -(1 -f 1 ) • • • (1 -f b )) k-1\\nWhen f 1 = 0.5 the components with a factor of f 1 compensate for the ones with a factor of (1 -f 1 ), and so the derivative is 0 for f 1 = 0.5. Deriving with respect to other feature values gives analogues results. That is, one optimal feature occurrence value for maximizing P (win) is 0.5.\"\n },\n {\n \"figure_ref\": [\n \"fig_5\"\n ],\n \"heading\": \"C TRAINING METHODOLOGY\",\n \"publication_ref\": [\n \"b10\",\n \"b0\"\n ],\n \"table_ref\": [],\n \"text\": \"In order to prevent overfitting and the representation of 'trivial features' (e.g. specific pixel values) in the representations, during training we use a stack of image augmentation layers independently applied prior to each image encoder. This involves a random rotation of up to 0.1 radians, a random contrast shift of up to 10%, a random translation of up to 10% along both axes, and a random zoom of up to 10% (all with a nearest-neighbour filling of blank pixels).\\nThe models were optimized using Adam (Kingma & Ba, 2015) with a learning rate of 0.001. The batch size used for training was dependent on the number of distractors, and each epoch iterated through the entire training dataset. See Table 1 for the full breakdown of test accuracy values for trained models, i.e. the mean and standard deviations for the proportion of occasions where the distinguisher was correctly able to identifier x * by using r.\\nAll of the code was implemented with Tensorflow 2 (Abadi et al., 2015) and datasets were pulled from Tensorflow Datasets3 (TFDS) (TF Devs, 2022). CIFAR-10 was split into the default TFDS training and test sets (50,000 training images and 10,000 test images). Training and analysis were performed with an NVIDIA RTX 3090 GPU.\\nWe trained 54 independent encoder-distinguisher pairs4 for 10 epochs on CIFAR-10 and removed models that did not converge (as defined by not reaching an 80% drop in loss), resulting in 51 trained models (taken as the best performing checkpoint). Models were trained with varying combinations of representation lengths and number of distractors: (|r|, k) ∈ {64, 128, 256, 512} × {3, 5, 10, 20}. We also trained models with representation lengths 8, 16 and 32, visualizations of which can be found in Figure 5, which we discarded because their bit entropies were too homogeneous to meaningfully study the effect of masking out low versus high entropy bits. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"D EXPERIMENTS\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The code for the experiments can be found at the following repository: [URL removed to preserve anonymity]\"\n },\n {\n \"figure_ref\": [\n \"fig_6\",\n \"fig_7\",\n \"fig_8\",\n \"fig_9\"\n ],\n \"heading\": \"D.1 OVERFITTING ANALYSIS\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In Figure 6 we see that the test and training accuracies are very similar (with the test accuracy even being slightly higher) and so no overfitting has happened. Figure 7 shows the OOD accuracies for each dataset (using the data of all the values of k and all the analysed representation lengths). Figure 8 shows the accuracies for each dataset and each value of k.\\nFigure 9 shows the accuracies for each dataset and each representation length |r|. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"ACKNOWLEDGEMENTS\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Work done by both authors is thanks to the UKRI Centre for Doctoral Training in Safe and Trusted AI (EPSRC Project EP/S023356/1).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"D.3 OOD MEAN ACCURACY CHANGE FROM MASKS\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The tables in this section are the same as Table 2 in Section ??, except separated by different values of k. Figure 10 is a visualisation of the data along with the 'distance out-of-distribution' for each k value. \"\n }\n]"},"abstract":{"kind":"string","value":"We study the relationship between the entropy of intermediate representations and a model's robustness to distributional shift. We train models consisting of two feed-forward networks end-to-end separated by a discrete n-bit channel on an unsupervised contrastive learning task. Different masking strategies are applied after training that remove a proportion of low-entropy bits, high-entropy bits, or randomly selected bits, and the effects on performance are compared to the baseline accuracy with no mask. We hypothesize that the entropy of a bit serves as a guide to its usefulness out-of-distribution (OOD). Through experiment on three OOD datasets we demonstrate that the removal of low-entropy bits can notably benefit OOD performance. Conversely, we find that top-entropy masking disproportionately harms performance both in-distribution (InD) and OOD."},"title":{"kind":"string","value":"Generalization LOW-ENTROPY LATENT VARIABLES HURT OUT-OF-DISTRIBUTION PERFORMANCE"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 1 :1Figure 1: An example of a contrastive task (k = 3). For a given dataset, the distinguisher is shown k images, among which k-1 distractor images, and has to predict the correct image.","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"), Stanford Online Products Song et al. (2016), Colorectal Histology Kather et al. (2016), Plant Village Hughes & Salathe (2015), and MNIST LeCun et al. (1999) datasets.","figure_data":"","figure_id":"fig_1","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 3 :3Figure 3: Effective robustness plots for low-k models. y = x shown as black dashed line.","figure_data":"","figure_id":"fig_2","figure_label":"3","figure_type":"figure"},{"figure_caption":"Figure 4 :4Figure 4: Architecture diagram (k = 5). The encoder is shown as the purple and green components, and the distinguisher is the orange and red.","figure_data":"","figure_id":"fig_3","figure_label":"4","figure_type":"figure"},{"figure_caption":"P(win) = P (win|c x , c y ) • P (c x , c y ) +P (win|c x , c ¬y ) • P (c x , c ¬y ) +P (win|c ¬x , c y ) • P (c ¬x , c y ) +P (win|c ¬x , c ¬y ) • P (c ¬x , c ¬y )","figure_data":"","figure_id":"fig_4","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 5 :5Figure5: The x-axis represents entropy percentile of bits in the representation. The y-axis shows the entropy values of bits (measured on CIFAR-10). In other words, we take the list of bits and sort them by entropy, and then plot the sorted line as percentiles in order to compare the distributions of different lengths. The translucent regions show the error bars from various training runs. We can see that for lower |r| values, the entropy distributions do not tend to go below 0.8.","figure_data":"","figure_id":"fig_5","figure_label":"5","figure_type":"figure"},{"figure_caption":"Figure 6 :6Figure 6: For different values of k the blue line shows the training accuracy and the orange line shows the test accuracy.","figure_data":"","figure_id":"fig_6","figure_label":"6","figure_type":"figure"},{"figure_caption":"Figure 7 :7Figure 7: The y-axis represents the accuracy and the x-axis the masking proportion. Different masking strategies are represented by different colors.","figure_data":"","figure_id":"fig_7","figure_label":"7","figure_type":"figure"},{"figure_caption":"Figure 8 :8Figure 8: The y-axis shows the accuracy and the x-axis shows different masking proportions. Masking strategies are indicated by color.","figure_data":"","figure_id":"fig_8","figure_label":"8","figure_type":"figure"},{"figure_caption":"Figure 9 :9Figure 9: The y-axis shows the accuracy and the x-axis shows different masking proportions. Masking strategies are indicated by color.","figure_data":"","figure_id":"fig_9","figure_label":"9","figure_type":"figure"},{"figure_caption":"","figure_data":"","figure_id":"","figure_label":"","figure_type":"figure"},{"figure_caption":"Mean accuracy shift (in percentage points) after masking with each strategy. After running paired t-tests we find that all of these accuracy shifts are statistically significant (with p = 0.05). work adds to the toolkit of methods to aid in understanding and improving robustness to distributional shift, which for example includes forms of data augmentationHendrycks et al. (2021) and abstaining from making a prediction in the face of uncertaintyThulasidasan et al. (2021). For a general overview of problems and methods in OOD robustness seeShen et al. (2015).Below we reference some notable entropy-based methods that have a different purpose than improving OOD robustness.Chatterjee & Mishchenko (2019) use low entropy (or \"rare\") signals to analyze the extent to which a model is overfitted to the training distribution. Entropy-based methods have also been used widely in the adjacent problem of OOD detection. For example, predictive entropy measures the uncertainty of the prediction of a sample given a training distribution and is used to calculate the extent to which a sample is OODKirsch et al. (","figure_data":"CIFAR-10Colorectal HistologyMNISTPlant VillageMasked Bottom Entropy1.6 ± 8.0-2.0 ± 14.39.4 ± 15.63.0 ± 23.7Masked Top Entropy-4.3 ± 21.4-7.8 ± 19.0-16.6 ± 5.3 -18.5 ± 21.8Random Mask2.5 ± 12.33.4 ± 10.94.2 ± 13.72.1 ± 19.65 RELATED WORK","figure_id":"tab_1","figure_label":"2","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 1 :1Figure 1: An example of a contrastive task (k = 3). For a given dataset, the distinguisher is shown k images, among which k-1 distractor images, and has to predict the correct image.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"), Stanford Online Products Song et al. (2016), Colorectal Histology Kather et al. (2016), Plant Village Hughes & Salathe (2015), and MNIST LeCun et al. (1999) datasets.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 :3Figure 3: Effective robustness plots for low-k models. y = x shown as black dashed line.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 4 :4Figure 4: Architecture diagram (k = 5). The encoder is shown as the purple and green components, and the distinguisher is the orange and red.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"P(win) = P (win|c x , c y ) • P (c x , c y ) +P (win|c x , c ¬y ) • P (c x , c ¬y ) +P (win|c ¬x , c y ) • P (c ¬x , c y ) +P (win|c ¬x , c ¬y ) • P (c ¬x , c ¬y )\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 5 :5Figure5: The x-axis represents entropy percentile of bits in the representation. The y-axis shows the entropy values of bits (measured on CIFAR-10). In other words, we take the list of bits and sort them by entropy, and then plot the sorted line as percentiles in order to compare the distributions of different lengths. The translucent regions show the error bars from various training runs. We can see that for lower |r| values, the entropy distributions do not tend to go below 0.8.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_5\",\n \"figure_label\": \"5\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 6 :6Figure 6: For different values of k the blue line shows the training accuracy and the orange line shows the test accuracy.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_6\",\n \"figure_label\": \"6\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 7 :7Figure 7: The y-axis represents the accuracy and the x-axis the masking proportion. Different masking strategies are represented by different colors.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_7\",\n \"figure_label\": \"7\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 8 :8Figure 8: The y-axis shows the accuracy and the x-axis shows different masking proportions. Masking strategies are indicated by color.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_8\",\n \"figure_label\": \"8\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 9 :9Figure 9: The y-axis shows the accuracy and the x-axis shows different masking proportions. Masking strategies are indicated by color.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_9\",\n \"figure_label\": \"9\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"\",\n \"figure_data\": \"\",\n \"figure_id\": \"\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Mean accuracy shift (in percentage points) after masking with each strategy. After running paired t-tests we find that all of these accuracy shifts are statistically significant (with p = 0.05). work adds to the toolkit of methods to aid in understanding and improving robustness to distributional shift, which for example includes forms of data augmentationHendrycks et al. (2021) and abstaining from making a prediction in the face of uncertaintyThulasidasan et al. (2021). For a general overview of problems and methods in OOD robustness seeShen et al. (2015).Below we reference some notable entropy-based methods that have a different purpose than improving OOD robustness.Chatterjee & Mishchenko (2019) use low entropy (or \\\"rare\\\") signals to analyze the extent to which a model is overfitted to the training distribution. Entropy-based methods have also been used widely in the adjacent problem of OOD detection. For example, predictive entropy measures the uncertainty of the prediction of a sample given a training distribution and is used to calculate the extent to which a sample is OODKirsch et al. (\",\n \"figure_data\": \"CIFAR-10Colorectal HistologyMNISTPlant VillageMasked Bottom Entropy1.6 ± 8.0-2.0 ± 14.39.4 ± 15.63.0 ± 23.7Masked Top Entropy-4.3 ± 21.4-7.8 ± 19.0-16.6 ± 5.3 -18.5 ± 21.8Random Mask2.5 ± 12.33.4 ± 10.94.2 ± 13.72.1 ± 19.65 RELATED WORK\",\n \"figure_id\": \"tab_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n }\n]"},"authors":{"kind":"string","value":"Nandi Schoots; Dylan Cope"},"references":{"kind":"list like","value":[{"authors":"Martín Abadi; Ashish Agarwal; Paul Barham; Eugene Brevdo; Zhifeng Chen; Craig Citro; Greg S Corrado; Andy Davis; Jeffrey Dean; Matthieu Devin; Sanjay Ghemawat; Ian Goodfellow; Andrew Harp; Geoffrey Irving; Michael Isard; Jia Yangqing; Rafal Jozefowicz; Lukasz Kaiser; Manjunath Kudlur; Josh Levenberg; Dandelion Mané; Rajat Monga; Sherry Moore; Derek Murray; Chris Olah; Mike Schuster; Jonathon Shlens; Benoit Steiner; Ilya Sutskever; Kunal Talwar; Paul Tucker; Vincent Vanhoucke; Vijay Vasudevan; Fernanda Viégas; Oriol Vinyals; Pete Warden; Martin Wattenberg; Martin Wicke; Yuan Yu; Xiaoqiang Zheng","journal":"","ref_id":"b0","title":"TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems","year":"2015"},{"authors":"Satrajit Chatterjee; Alan Mishchenko","journal":"","ref_id":"b1","title":"Coherent gradients: An approach to understanding generalization in gradient descent-based optimization","year":"2019"},{"authors":"Gergely Flamich; Marton Havasi; José Miguel Hernández-Lobato","journal":"CoRR","ref_id":"b2","title":"Compressing images by encoding their latent representations with relative entropy coding","year":"2020"},{"authors":"Jakob Foerster; Alexandros Ioannis; Nando Assael; Shimon De Freitas; Whiteson","journal":"","ref_id":"b3","title":"Learning to Communicate with Deep Multi-Agent Reinforcement Learning","year":""},{"authors":" Curran Associates; Inc","journal":"","ref_id":"b4","title":"","year":"2016"},{"authors":"Dan Hendrycks; Steven Basart; Norman Mu; Saurav Kadavath; Frank Wang; Evan Dorundo; Rahul Desai; Tyler Zhu; Samyak Parajuli; Mike Guo; Dawn Song; Jacob Steinhardt; Justin Gilmer","journal":"IEEE","ref_id":"b5","title":"The many faces of robustness: A critical analysis of out-of-distribution generalization","year":"2021"},{"authors":"Geoffrey E Hinton; Drew Van Camp","journal":"Association for Computing Machinery","ref_id":"b6","title":"Keeping the neural networks simple by minimizing the description length of the weights","year":"1993"},{"authors":"P David; Marcel Hughes; Salathe","journal":"","ref_id":"b7","title":"An open access repository of images on plant health to enable the development of mobile disease diagnostics through machine learning and crowdsourcing","year":"2015"},{"authors":"Ashish Jaiswal; Ramesh Ashwin; Mohammad Zaki Babu; Debapriya Zadeh; Fillia Banerjee; Makedon","journal":"Technologies","ref_id":"b8","title":"A Survey on Contrastive Self-supervised Learning","year":""},{"authors":"Jakob Nikolas Kather; Cleo-Aron Weis; Francesco Bianconi; Susanne M Melchers; Lothar R Schad; Timo Gaiser; Alexander Marx; Frank Gerrit Zollner","journal":"Scientific Reports","ref_id":"b9","title":"Multi-class texture analysis in colorectal cancer histology","year":"2016"},{"authors":"D P Kingma; L J Ba","journal":"","ref_id":"b10","title":"Adam: A Method for Stochastic Optimization","year":"2015"},{"authors":"Andreas Kirsch; Jishnu Mukhoti; Joost Amersfoort; H S Philip; Yarin Torr; Gal","journal":"","ref_id":"b11","title":"On pitfalls in ood detection: Entropy considered harmful","year":"2021"},{"authors":"Alex Krizhevsky","journal":"","ref_id":"b12","title":"Learning multiple layers of features from tiny images","year":"2009"},{"authors":"Yann Lecun; Corinna Cortes; Chris Burges","journal":"","ref_id":"b13","title":"MNIST handwritten digit database","year":"1999"},{"authors":"Zheyan Shen; Jiashuo Liu; Yue He; Xingxuan Zhang; Renzhe Xu; Han Yu; Peng Cui","journal":"Journal of Latex Class Files","ref_id":"b14","title":"Towards Out-Of-Distribution Generalization: A Survey","year":"2015"},{"authors":"Hyun Oh Song; Yu Xiang; Stefanie Jegelka; Silvio Savarese","journal":"","ref_id":"b15","title":"Deep Metric Learning via Lifted Structured Feature Embedding","year":"2016"},{"authors":"Rohan Taori; Achal Dave; Vaishaal Shankar; Nicholas Carlini; Benjamin Recht; Ludwig Schmidt","journal":"","ref_id":"b16","title":"Measuring Robustness to Natural Distribution Shifts in Image Classification","year":"2020"},{"authors":" Tf Devs","journal":"","ref_id":"b17","title":"TensorFlow Datasets: A collection of ready-to-use datasets","year":"2022"},{"authors":"Sushil Sunil Thulasidasan; Sayera Thapa; Gopinath Dhaubhadel; Tanmoy Chennupati; Jeff A Bhattacharya; Bilmes","journal":"IEEE","ref_id":"b18","title":"An effective baseline for robustness to distributional shift","year":"2021"},{"authors":"Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Łukasz Kaiser; Illia Polosukhin","journal":"","ref_id":"b19","title":"Attention Is All You Need","year":"2017"}],"string":"[\n {\n \"authors\": \"Martín Abadi; Ashish Agarwal; Paul Barham; Eugene Brevdo; Zhifeng Chen; Craig Citro; Greg S Corrado; Andy Davis; Jeffrey Dean; Matthieu Devin; Sanjay Ghemawat; Ian Goodfellow; Andrew Harp; Geoffrey Irving; Michael Isard; Jia Yangqing; Rafal Jozefowicz; Lukasz Kaiser; Manjunath Kudlur; Josh Levenberg; Dandelion Mané; Rajat Monga; Sherry Moore; Derek Murray; Chris Olah; Mike Schuster; Jonathon Shlens; Benoit Steiner; Ilya Sutskever; Kunal Talwar; Paul Tucker; Vincent Vanhoucke; Vijay Vasudevan; Fernanda Viégas; Oriol Vinyals; Pete Warden; Martin Wattenberg; Martin Wicke; Yuan Yu; Xiaoqiang Zheng\",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Satrajit Chatterjee; Alan Mishchenko\",\n \"journal\": \"\",\n \"ref_id\": \"b1\",\n \"title\": \"Coherent gradients: An approach to understanding generalization in gradient descent-based optimization\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Gergely Flamich; Marton Havasi; José Miguel Hernández-Lobato\",\n \"journal\": \"CoRR\",\n \"ref_id\": \"b2\",\n \"title\": \"Compressing images by encoding their latent representations with relative entropy coding\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Jakob Foerster; Alexandros Ioannis; Nando Assael; Shimon De Freitas; Whiteson\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"Learning to Communicate with Deep Multi-Agent Reinforcement Learning\",\n \"year\": \"\"\n },\n {\n \"authors\": \" Curran Associates; Inc\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Dan Hendrycks; Steven Basart; Norman Mu; Saurav Kadavath; Frank Wang; Evan Dorundo; Rahul Desai; Tyler Zhu; Samyak Parajuli; Mike Guo; Dawn Song; Jacob Steinhardt; Justin Gilmer\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b5\",\n \"title\": \"The many faces of robustness: A critical analysis of out-of-distribution generalization\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Geoffrey E Hinton; Drew Van Camp\",\n \"journal\": \"Association for Computing Machinery\",\n \"ref_id\": \"b6\",\n \"title\": \"Keeping the neural networks simple by minimizing the description length of the weights\",\n \"year\": \"1993\"\n },\n {\n \"authors\": \"P David; Marcel Hughes; Salathe\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"An open access repository of images on plant health to enable the development of mobile disease diagnostics through machine learning and crowdsourcing\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Ashish Jaiswal; Ramesh Ashwin; Mohammad Zaki Babu; Debapriya Zadeh; Fillia Banerjee; Makedon\",\n \"journal\": \"Technologies\",\n \"ref_id\": \"b8\",\n \"title\": \"A Survey on Contrastive Self-supervised Learning\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Jakob Nikolas Kather; Cleo-Aron Weis; Francesco Bianconi; Susanne M Melchers; Lothar R Schad; Timo Gaiser; Alexander Marx; Frank Gerrit Zollner\",\n \"journal\": \"Scientific Reports\",\n \"ref_id\": \"b9\",\n \"title\": \"Multi-class texture analysis in colorectal cancer histology\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"D P Kingma; L J Ba\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"Adam: A Method for Stochastic Optimization\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Andreas Kirsch; Jishnu Mukhoti; Joost Amersfoort; H S Philip; Yarin Torr; Gal\",\n \"journal\": \"\",\n \"ref_id\": \"b11\",\n \"title\": \"On pitfalls in ood detection: Entropy considered harmful\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Alex Krizhevsky\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"Learning multiple layers of features from tiny images\",\n \"year\": \"2009\"\n },\n {\n \"authors\": \"Yann Lecun; Corinna Cortes; Chris Burges\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"MNIST handwritten digit database\",\n \"year\": \"1999\"\n },\n {\n \"authors\": \"Zheyan Shen; Jiashuo Liu; Yue He; Xingxuan Zhang; Renzhe Xu; Han Yu; Peng Cui\",\n \"journal\": \"Journal of Latex Class Files\",\n \"ref_id\": \"b14\",\n \"title\": \"Towards Out-Of-Distribution Generalization: A Survey\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Hyun Oh Song; Yu Xiang; Stefanie Jegelka; Silvio Savarese\",\n \"journal\": \"\",\n \"ref_id\": \"b15\",\n \"title\": \"Deep Metric Learning via Lifted Structured Feature Embedding\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Rohan Taori; Achal Dave; Vaishaal Shankar; Nicholas Carlini; Benjamin Recht; Ludwig Schmidt\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"Measuring Robustness to Natural Distribution Shifts in Image Classification\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \" Tf Devs\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"TensorFlow Datasets: A collection of ready-to-use datasets\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Sushil Sunil Thulasidasan; Sayera Thapa; Gopinath Dhaubhadel; Tanmoy Chennupati; Jeff A Bhattacharya; Bilmes\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b18\",\n \"title\": \"An effective baseline for robustness to distributional shift\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Łukasz Kaiser; Illia Polosukhin\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"Attention Is All You Need\",\n \"year\": \"2017\"\n }\n]"},"formulas":{"kind":"list like","value":[{"formula_coordinates":[2,151.59,454.79,308.81,10.59],"formula_id":"formula_0","formula_text":"H(B i | D) = -p log 2 p -(1 -p) log 2 (1 -p), where p = P (B i = 1 | D)."},{"formula_coordinates":[2,256.32,661.71,243.81,22.31],"formula_id":"formula_1","formula_text":"ri = m i r i + (1 -m i ) 1 2 . (2"},{"formula_coordinates":[2,500.13,668.76,3.87,8.64],"formula_id":"formula_2","formula_text":")"},{"formula_coordinates":[3,106.83,258.74,190.95,8.74],"formula_id":"formula_3","formula_text":"(|r|, k) ∈ {64, 128, 256, 512} × {3, 5, 10, 20}."},{"formula_coordinates":[9,174.64,263.52,262.72,22.31],"formula_id":"formula_4","formula_text":"P (win|c x , c y ) = Σ k v=1 1 v • (f x f y ) v-1 • (1 -f x f y ) k-v • k -1 v -1 ."},{"formula_coordinates":[9,170.79,346.74,270.41,29.58],"formula_id":"formula_5","formula_text":"P (win|c x , c y ) = Σ k v=1 P (win|c x , c y , V = v) • P ( Ṽ = v -1|c x , c y ) = Σ k v=1 P (win|c x , c y , V = v) • P ( Ṽ = v -1)"},{"formula_coordinates":[9,108,403.21,329.36,64.12],"formula_id":"formula_6","formula_text":"P ( Ṽ = v -1) = (f x f y ) v-1 • (1 -f x f y ) k-1-(v-1) • k -1 v -1 Hence P (win|c x , c y ) = Σ k v=1 1 v • (f x f y ) v-1 • (1 -f x f y ) k-v • k -1 v -1 ."},{"formula_coordinates":[9,143.45,490.96,317.76,152.67],"formula_id":"formula_7","formula_text":"P (win|c x , c y )P (c x , c y ) = f x f y • Σ k v=1 1 v • (f x f y ) v-1 • (1 -f x f y ) k-v • k -1 v -1 = Σ k v=1 1 v • (f x f y ) v • (1 -f x f y ) k-v • k -1 v -1 = 1 k Σ k v=1 (f x f y ) v • (1 -f x f y ) k-v • k v = 1 k (f x f y + 1 -f x f y ) k -(1 -f x f y ) k = 1 k 1 -(1 -f x f y ) k = 1 k - 1 k (1 -f x f y ) k"},{"formula_coordinates":[9,211,665.58,189.99,68.62],"formula_id":"formula_8","formula_text":"P (win) = 4 k - 1 k (1 -f x • f y ) k +(1 -f x • (1 -f y )) k +(1 -(1 -f x ) • f y ) k +(1 -(1 -f x ) • (1 -f y )) k ."},{"formula_coordinates":[10,108,119.7,397.48,23.89],"formula_id":"formula_9","formula_text":"P (win) = 2 b k - 1 k (1 -f 1 • • • f b ) k + (1 -(1 -f 1 )f 2 • • • f b ) k + . . . + (1 -(1 -f 1 ) • • • (1 -f b )) k"},{"formula_coordinates":[10,174.62,209.53,263.46,68.28],"formula_id":"formula_10","formula_text":"∂P (win) ∂f 1 = f 2 • • • f b (1 -f 1 • • • f b ) k-1 -f 2 • • • f b (1 -(1 -f 1 )f 2 • • • f b ) k-1 + . . . -(1 -f 2 ) • • • (1 -f b )(1 -(1 -f 1 ) • • • (1 -f b )) k-1"}],"string":"[\n {\n \"formula_coordinates\": [\n 2,\n 151.59,\n 454.79,\n 308.81,\n 10.59\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"H(B i | D) = -p log 2 p -(1 -p) log 2 (1 -p), where p = P (B i = 1 | D).\"\n },\n {\n \"formula_coordinates\": [\n 2,\n 256.32,\n 661.71,\n 243.81,\n 22.31\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"ri = m i r i + (1 -m i ) 1 2 . (2\"\n },\n {\n \"formula_coordinates\": [\n 2,\n 500.13,\n 668.76,\n 3.87,\n 8.64\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \")\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 106.83,\n 258.74,\n 190.95,\n 8.74\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"(|r|, k) ∈ {64, 128, 256, 512} × {3, 5, 10, 20}.\"\n },\n {\n \"formula_coordinates\": [\n 9,\n 174.64,\n 263.52,\n 262.72,\n 22.31\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"P (win|c x , c y ) = Σ k v=1 1 v • (f x f y ) v-1 • (1 -f x f y ) k-v • k -1 v -1 .\"\n },\n {\n \"formula_coordinates\": [\n 9,\n 170.79,\n 346.74,\n 270.41,\n 29.58\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"P (win|c x , c y ) = Σ k v=1 P (win|c x , c y , V = v) • P ( Ṽ = v -1|c x , c y ) = Σ k v=1 P (win|c x , c y , V = v) • P ( Ṽ = v -1)\"\n },\n {\n \"formula_coordinates\": [\n 9,\n 108,\n 403.21,\n 329.36,\n 64.12\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"P ( Ṽ = v -1) = (f x f y ) v-1 • (1 -f x f y ) k-1-(v-1) • k -1 v -1 Hence P (win|c x , c y ) = Σ k v=1 1 v • (f x f y ) v-1 • (1 -f x f y ) k-v • k -1 v -1 .\"\n },\n {\n \"formula_coordinates\": [\n 9,\n 143.45,\n 490.96,\n 317.76,\n 152.67\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"P (win|c x , c y )P (c x , c y ) = f x f y • Σ k v=1 1 v • (f x f y ) v-1 • (1 -f x f y ) k-v • k -1 v -1 = Σ k v=1 1 v • (f x f y ) v • (1 -f x f y ) k-v • k -1 v -1 = 1 k Σ k v=1 (f x f y ) v • (1 -f x f y ) k-v • k v = 1 k (f x f y + 1 -f x f y ) k -(1 -f x f y ) k = 1 k 1 -(1 -f x f y ) k = 1 k - 1 k (1 -f x f y ) k\"\n },\n {\n \"formula_coordinates\": [\n 9,\n 211,\n 665.58,\n 189.99,\n 68.62\n ],\n \"formula_id\": \"formula_8\",\n \"formula_text\": \"P (win) = 4 k - 1 k (1 -f x • f y ) k +(1 -f x • (1 -f y )) k +(1 -(1 -f x ) • f y ) k +(1 -(1 -f x ) • (1 -f y )) k .\"\n },\n {\n \"formula_coordinates\": [\n 10,\n 108,\n 119.7,\n 397.48,\n 23.89\n ],\n \"formula_id\": \"formula_9\",\n \"formula_text\": \"P (win) = 2 b k - 1 k (1 -f 1 • • • f b ) k + (1 -(1 -f 1 )f 2 • • • f b ) k + . . . + (1 -(1 -f 1 ) • • • (1 -f b )) k\"\n },\n {\n \"formula_coordinates\": [\n 10,\n 174.62,\n 209.53,\n 263.46,\n 68.28\n ],\n \"formula_id\": \"formula_10\",\n \"formula_text\": \"∂P (win) ∂f 1 = f 2 • • • f b (1 -f 1 • • • f b ) k-1 -f 2 • • • f b (1 -(1 -f 1 )f 2 • • • f b ) k-1 + . . . -(1 -f 2 ) • • • (1 -f b )(1 -(1 -f 1 ) • • • (1 -f b )) k-1\"\n }\n]"}}},{"rowIdx":383,"cells":{"doi":{"kind":"string","value":""},"pub_date":{"kind":"string","value":"2023-08-15"},"sections":{"kind":"list like","value":[{"figure_ref":[],"heading":"Introduction","publication_ref":["b47","b54","b60","b3","b10","b11"],"table_ref":[],"text":"Estimating a map along with camera poses from a collection of images is a long-standing and challenging problem in computer vision [9,41,49] with relevant applications in domains such as autonomous driving and augmented reality. Recently, deep visual detectors and descriptors have shown increased resilience to extreme viewpoint and appearance changes [4, 48,55,61,64]. However, a common limitation among all these deep descriptors and detectors is the lack of a probabilistic formulation for detection noise. Consequently, downstream pose estimation relies on the assumption of constant spatial covariances, as shown in Figure 1 (a), leading to suboptimal results.\nWe propose to model the spatial covariance of detected keypoints in deep detectors. We find that recent detectors share a common design, where a deep convolutional backbone is used to predict a score map that assigns a \"probabil-ity\" to a pixel being a point of interest. We exploit this common design space, to propose two post-hoc methods that estimate a covariance matrix for each keypoint detected in any pretrained feature detector. Our simplest method uses the score at a given pixel to initialize an isotropic covariance matrix. We illustrate the learned score maps (lower triangle) overlaid on exemplar images, along with our proposed isotropic covariances in Figure 1 (b) for two state-ofthe-art deep feature detectors, Superpoint [11] and D2Net [12]. We also suggest a theoretically-motivated method to estimate the full spatial covariance matrix of detected keypoints using the local structure tensor. The structure tensor models the local saliency of the detections in the score map. Figure 1 (c) shows the deduced full covariances. These covariances capture the larger uncertainty along edges on the learned score map. To the best of our knowledge, we are the first to model spatial uncertainties of deep detectors.\nWe show in a series of experiments that our proposed methods for modeling the spatial covariances of detected features are directly related with the errors in matching. Accounting for them allows us to improve performance in tasks such as solving the perspective-n-point (PnP) problem and nonlinear optimizations."},{"figure_ref":[],"heading":"Related Work","publication_ref":["b50","b5","b0","b31","b17","b22","b56","b13","b25","b6","b7","b27","b38","b58","b62","b13","b14","b39","b29","b4","b10","b11","b42","b57","b59","b61","b10","b11","b42","b57","b4","b23","b53","b55","b49"],"table_ref":[],"text":"We propose to model the spatial covariances of learned local features. Modeling the spatial uncertainty of features is not a new idea, and has been studied for hand-crafted detectors. In this section, we first review hand-crafted methods and how uncertainty estimation has been proposed for them. We then describe recent progress in learned detectors.\nHandcrafted local features. Harris et al. [22] pioneeres rotationally-invariant corner detection by using a heuristic over the eigenvalues of the local structure tensor, while Shi et al. [51] proposes its smallest eigenvalue for detection. Mikolajczyk and Schmid [37] robustifies it against scale and affine transformations. On the other hand, SIFT (DoG) [34] popularizes detection (and description) of blobs. SURF [6] reduces its execution time by using integral images and (A-)KAZE [1,2] proposes non-linear diffusion to improve the invariance to changes in scale. Lastly, FAST [45] and its extensions [32,46] stand out for achieving the lowest execution times, thanks to only requiring intensity comparisons between the neighboring pixels of an image patch.\nUncertainty quantification in handcrafted detectors. Inclusion of spatial uncertainty of local features for estimation of geometry has been extensively studied [18,23,25,57]. However, its quantification on classical local features is still recognized as an open problem in the literature [14,26,27] and which has not been addressed yet with learned detections. Several works have shown the benefits of a precise quantification of uncertainty. For this purpose, they adapt the quantifications to specific classical detectors [8,17,28,39,59,63], assume planar surfaces [14,42] and require an accurate offline calibration [15,16]. Only recently [40] proposes a learning-based approach of spatial covariances, trained per detector, in order to weigh normalized epipolar errors [30,33]. In contrast to previous works, we propose a general formulation for their quantification, directly applicable on state-of-the-art learned detectors, seamlessly fitting them by leveraging their common characteristics.\nLearned local features and lack of uncertainty. The dominant approach [5,11,12,35,43,58,60,62] consists on training CNNs to regress a score map over which detections are extracted via non-maximum-suppresion (NMS). Superpoint [11] is an efficient detector-descriptor of corners, robust to noise and image transformations thanks to a synthetic pre-training followed by homographic adaptation on real images. D2Net [12] shows the applicability, to the problem of local feature detection and description, of classification networks [52]. R2D2 [43] proposes a reliability measure, used to discard unmatchable points. Similarly, DISK [58] bases its learning on matches, and KeyNet [5], inspired by classical systems, proposes to learn from spatial image gradients. This work experiments with Superpoint, D2Net, R2D2, and KeyNet, but our proposal is applicable to the rest of the systems. Finally, recent works [24,54,56] exploit attention mechanisms to match pixels without explicitly using detectors. Although their inclusion in geometric estimation pipelines can be engineered [50], they suffer from lack of repeatability. Because of this, in this work, we focus on the quantification of spatial uncertainty for learned detectors."},{"figure_ref":["fig_0"],"heading":"Method","publication_ref":[],"table_ref":[],"text":"Our framework takes as input a RGB image I of spatial dimensions H × W , and outputs local features together with their spatial uncertainties. It is composed of two main com-ponents as shown in Fig. 2. (1) A pretrained feature detector and (2) our novel and detector-agnostic uncertainty module. Our methods work with any pretrained state-ofthe-art detectors, taking their learned score maps as input, and outputting reliable covariance estimates for each detected keypoint. We show that the estimated covariances are well-calibrated and improve downstream tasks such as PnP and motion-only bundle adjustment."},{"figure_ref":["fig_1"],"heading":"Overview","publication_ref":["b4","b10","b11","b42","b57","b59","b61","b22","b56"],"table_ref":[],"text":"Pre-trained local-feature detector. Our framework is directly applicable to the vast majority of learned detectors [5,11,12,35,43,58,60,62]. These detectors share a common architectural design, using convolutional backbones to predict a score map S ∈ R kH×kW with k ∈ R + , followed by non-maximum-suppression (NMS) to extract a sparse set of features. We leverage this standard design to make a detector-agnostic, post-hoc uncertainty module that takes the estimated score maps as input and estimates the spatial uncertainty based on the peakiness in a local region around the detected point. Our approach can be applied to these detectors without any training or fine-tuning.\nDetector-Agnostic Feature Uncertainties. Instead of considering the 2D position of detections as deterministic locations, we propose to model their spatial uncertainties. More formally, we consider that the spatial location x i ∈ R 2 of a local feature i, detected in I, stems from perturbing its true location x i,true ∈ R 2 , with random noise ξ i ∈ R 2\nx i = x i,true + ξ i ,(1)\nwhose probability distribution we want to estimate. We follow the dominant model in computer vision [23,25,57], which uses second order statistics to describe the spatial uncertainty of each location\nE{ξ i } = 0 , Σ i := Cov(ξ i ) = E{ξ i ξ ⊤ i } , ∀i . (2)\nThereby our goal is to quantify each covariance matrix Σ i . To this end, we propose two methods: (1) A point-wise estimator of isotropic covariances (Sec. 3.2) that uses the inverse of the scores at the detected point, and (2) a full covariance estimator based on the local structure tensor (Sec. 3.3) which models the local saliency to estimate the uncertainty in all directions. We show that both approaches lead to reliable uncertainty estimates, benefiting downstream tasks. Intuitively, a peaky score map at the location of a detected feature will yield low spatial uncertainty, whereas a flatter score map will yield larger spatial uncertainty. Figure 3 qualitatively shows how our deduced 2D uncertainties relate to 3D uncertainties. In the Supplemental we explore the agnostic behavior of our covariances. \nϕ([∇ x S, ∇ y S]) = [(∇ x S) 2 , (∇ y S) 2 , ∇ x S∇ y S]"},{"figure_ref":[],"heading":"Point-wise Estimation","publication_ref":[],"table_ref":[],"text":"As the simplest estimator of the spatial uncertainty, we propose to use the regressed score of each local feature to create an isotropic covariance matrix\nΣ i := 1 S(x i ) I 2×2 = 1/S(x i ) 0 0 1/S(x i ) .(3)\nFig 1 shows that this estimator yields isotropic predictions of uncertainty (equal in all directions), so it only quantifies the relative scale regardless of the learned local structure."},{"figure_ref":["fig_2"],"heading":"Structure Tensor","publication_ref":["b4","b10","b11","b42","b46","b43","b19","b35"],"table_ref":[],"text":"Quantification of local saliency motivates the use of the local structure tensor,\nC i ∈ R 2×2 [22]. Defining [∇ x S i , ∇ y S i ] := ∂S/∂x| xi as the spatial gradient of S eval- uated at x i , C i in its local neighborhood W i (window of size u × v) is given by C i := j∈Wi w j ∂S ∂x ⊤ xj ∂S ∂x xj = j∈Wi w j (∇ x S j ) 2 ∇ x S j ∇ y S j ∇ y S j ∇ x S j (∇ y S j ) 2 ,(4)\nPoint-wise (Sec. 3.2)\nStructure Tensor (Sec. 3.3) Zoom Key.Net [5] Superpoint [11] D2Net [12] R2D2 [43] Figure 5. Qualitative comparison. Our estimated covariance matrices for each detected local feature are illustrated with uncertainty ellipses for four state-of-the-art pretrained detectors. Our pointwise covariance estimator yields isotropic uncertainty predictions (equal in all directions). Our proposed local structure tensor estimates the directionality of the spatial uncertainty, resulting in anisotropic predictions. For instance, Superpoint (which is trained to detect corners) has high uncertainty patterns corresponding to lines. For visualization purposes, a uniform sample of local feature detections is shown.\nwith w j ∈ R + being the weight of pixel j, preferably defined by a Gaussian centered at x i [47]. As such, C i is a positive semidefinite (PSD) matrix, resulting from averaging the directionality of gradients and hence not canceling opposite ones (see Fig. 4).\nThe reason why C i captures the local saliency lies in the auto-correlation function, c : R u×v → R, which averages local changes in S given small displacements δx [22]:\nc i = j∈Wi w j (S(x j ) -S(x j + δx)) 2 ,(5)\nwith S(x j ) indicating the score at x j . Linearly approximating S(x j + δx) ≈ S(x j ) + ∂S/∂x| xj δx, yields\nc i ≈ j∈Wi w j (S(x j ) -S(x j ) - ∂S ∂x xj δx) 2 ,(6)\n= j∈Wi w j δx ⊤ ∂S ∂x ⊤ xj ∂S ∂x xj δx = δx ⊤ C i δx . (7)\nThus, extreme saliency directions are obtained by solving:\nmax δx min δx δx ⊤ C i δx, s.t. ∥δx∥ = 1 ,(8)\nwhere the constraint ∥δx∥ = 1 ensures non-degenerate directions, which can be obtained with Lagrange multipliers [44] i.e. by defining the Lagrangian\nL(δx, λ) := δx ⊤ C i δx -λ(δx ⊤ δx -1) ,(9)\ndifferentiating w.r.t. δx and setting it to 0:\n2δx ⊤ C i -2λδx ⊤ = 0 ⇒ C i δx = λδx ,(10)\nwe conclude that directions of extreme saliency correspond to the eigenvectors of C i . Since inverting a matrix, does the same to its eigenvalues without affecting its eigenvectors 1 ,\nΣ i := C -1\ni results in a proper covariance matrix (PSD) assigning greater uncertainty in the direction of less saliency and vice-versa. [20,29] and motivated by the principle of maximum entropy [36] of the local features' scores, distributed independently by"},{"figure_ref":[],"heading":"Statistical interpretation. Under a Gaussian model of aleatoric uncertainty, common in deep learning","publication_ref":["b27","b56"],"table_ref":[],"text":"N (S(x i + t i ), σ 2 ) ,(11)\ninspired by [28], we can set up a parametric optimization of t ∈ R 2 which maximizes the likelihood, L(t i | S) with S := {S(x j ) | x j ∈ W i }, of the observations:\nS(x i ) = S(x i + t i ) + ε i , with ε i ∼ N (0, σ 2 ) ,(12)\n1 Let A ∈ R m×m and det A ̸ = 0, then λ -1 v = A -1 v, with λ and v being the eigenvalues and eigenvectors of A.\nwhere S(x i ) is the observed score, perturbed by the random noise ε i , independently affecting the rest of observations. Thus, the optimization is formulated as follows\nti = arg max ti j∈Wi 1 σ √ 2π exp - (S(x j ) -S(x j + t i )) 2 2σ 2 . (13\n)\nIts solution, or maximum-likelihood-estimation (MLE), is known a priori: ti = 0, which is unbiased given our statistical model (Eq. 12), and coherent with Equation 2.\nWith these conditions, the inverse of the Fisher information matrix, I( t), evaluated at the MLE, imposes a lower bound in the covariance matrix of the estimator ti = 0, known as Cramer-Rao Lower Bound (CRLB) [21]:\nVar( ti ) ≥ I( ti ) -1 .(14)\nI( t) is defined as E{s ⊤ i s i }, being the variance of the loglikelihood derivative (known as score)\ns i := ∂ log L(t i | S)/∂t i , since E{s i } = 0 [21].\nIn our case, it is given by\ns i = xj ∈Wi d j := S(x j ) -S(x j + t i ) σ 2 ∂S(x j + t i ) ∂t i .(15)\nDue to the linearity of expectation and the independence of observations,\nE{d ⊤ j d k } = 0, ∀j ̸ = k. Thereby I( ti ) = j∈Wi E{d ⊤ j d j } ti .(16)\nSince derivatives of Eq. 15 are applied on our deterministic model, they can go out of the expectation, and evaluating them on ti = 0 lead to\nI( ti ) = xj ∈Wi 1 σ 4 ∂S(x j ) ∂x ⊤ xj ∂S(x j ) ∂x xj E{(S(x j ) -S(x j + t i )) 2 } ti .(17)\nLastly, E{(S(x j ) -S(x j + t i )) 2 } ti = E{ε 2 i } = σ 2 since E{ε i } = 0, implying that our Fisher information matrix is\nI( ti ) = 1 σ 2 xj ∈Wi ∂S(x j ) ∂x ⊤ xj ∂S(x j ) ∂x xj ,(18)\nat the MLE. It matches the local structure tensor (Eq. 4) up to a scale factor Var(ε i ) -1 = σ -2 , unknown a priori.\nRecalling the CRLB (Eq. 14), although achievable only asymptotically [21,57], it motivates Σ i := C -1 i as an upto-scale covariance matrix of each location x i ."},{"figure_ref":[],"heading":"Experiments","publication_ref":[],"table_ref":[],"text":"Implementation details. In our experiments, we compute the structure tensor (and relative uncertainty) independently of the learned detector: for each local feature i, spatial differentiation at S(x j ), ∀j ∈ W i , is done with Sobel filters. Integration in W i is done with a 7 × 7 window, MMA overall 0.42 0.71"},{"figure_ref":[],"heading":"D2Net","publication_ref":[],"table_ref":[],"text":"Full Isotropic "},{"figure_ref":[],"heading":"SuperPoint","publication_ref":["b2","b4","b10","b11","b42","b18","b52"],"table_ref":[],"text":"Full Isotropic -uncertainty + Figure 6. Uncertainty and matching accuracy on HPatches [3]. We show MMA for each detector. We organize the results into 10 uncertainty ranges, ranging from lowest to highest (x-axis), each containing an equal number of matches. Our covariances accurately model less accurate matches, with the full-based approach (structure tensor) showing higher sensitivity in D2Net, Key.Net and R2D2.\nthe result of using an isotropic Gaussian filter with σ = 1 of cutoff frequency 3σ. Throughout all the experiments, we evaluate and extract our covariances using the detectors of state-of-the-art learned systems: Key.Net [5], Superpoint [11], D2Net [12] and R2D2 [43]. The score map used with Superpoint is the one prior to the channel-wise softmax to avoid the alteration of learned patterns crossing the boundaries of each grid. For the rest of the systems, we directly use their regressed score map. Diverse qualitative examples are shown in Figure 5.\nFigure 7. Qualitative results on KITTI [19]. Our covariances correctly model detections with unclear spatial locations, such as faraway corners or those located at edges.\nFigure 8. Qualitative results on TUM-RGBD [53]. We assign more spatial uncertainty at a priori less distinctive regions, such as papers and edges of the keyboard."},{"figure_ref":[],"heading":"Matching accuracy.","publication_ref":["b2","b11","b11","b52","b18"],"table_ref":[],"text":"We first test the relation between our estimated covariances with the accuracy of local-feature matching. Intuitively, local features detected with higher uncertainty should relate to less accurate matches, and vice versa. For this purpose, we consider the widely adopted HPatches dataset [3]. HPatches contains 116 sequences of 6 images each. 57 sequences exhibit significant illumination changes while the other 59 sequences undergo viewpoint changes.\nEvaluation protocol. We base our evaluation on the one proposed by [12]. First, extraction of local features and, in our case, covariance matrices of their locations is performed for all images. For every sequence, pairwise matching is done between a reference image r and each remaining image i by using Mutual Nearest Neighbor search (MNN). We then compute the reprojection errors and their covariances with the homographies, H i,r , provided by the dataset:\ne i,r := cart(x i -H i,r xr ) ,(19)\nΣ ei,r := JΣ xr J ⊤ + Σ xi ,(20)\nwhere cart(•) maps from homogeneous to Cartesian coordinates, and J := ∂e i,r /∂x r , i.e. we linearly propagate each covariance matrix Σ xr of the reference locations.\nTo quantify the uncertainty of the match with a scalar, we use the biggest eigenvalue of the corresponding Σ ei,r . Based on them, all matches gathered in the dataset are distributed in 10 ranges from lowest to highest uncertainty estimates, such that each range has the same number of matches. To quantify the accuracy in matching at each range, we choose the mean matching accuracy error (MMA), which represents the average percentage of matches with a corresponding value of ∥e i,r ∥ below a threshold. We use the same thresholds as in [12]. Finally, we compute the mean of all the MMA values at each range. This process is repeated for all the evaluated detectors and with the proposed full and isotropic covariances.\nResults. Figure 6 shows the averaged mean matching accuracy, MMA, at each uncertainty range. Ranges are ordered from lowest (1) to highest (10) estimated uncer- [53] and KITTI [19]. We report the cumulative errors for camera pose rotation and translation. Practically all estimations converge to acceptable thresholds when using our 2D full and 2D isotropic covariances. This is also apparent when using 3D covariances derived from our 2D ones. Without our covariances, a significant percentage of poses fail to localize. tainty. As can be seen, it exists a direct relationship between matching accuracy and both full and isotropic covariance estimates. With full covariances and for all evaluated ranges, lower uncertainty estimates imply higher accuracy in matching. However, this is not always the case when using isotropic covariances. As can be seen with R2D2, there is a certain increase in MMA for more uncertain matches. Additionally, there is a higher sensitivity of MMA to the uncertainty estimates stemming from full covariances on D2Net, Key.Net, and R2D2. This motivates the need for taking into account the learned local structure when quantifying the spatial uncertainty of the local feature, rather than basing it only on the regressed scalar estimate of the regressed score map."},{"figure_ref":[],"heading":"Geometry estimation","publication_ref":["b58","b18","b52","b12","b22","b30","b58","b18","b52"],"table_ref":[],"text":"To test the influence of our uncertainties in 3D-geometry estimation, we follow the evaluation proposed in [59]. It covers common stages in geometric estimation pipelines such as solving the perspective-n-point problem and motion-only bundle adjustment. The data used consists in the three sequences 00-02 of KITTI [19] and the first three 'freiburg 1' monocular RGB sequences of TUM-RGBD [53].\nEvaluation protocol. KITTI is used with a temporal win-dow of two left frames, while three are used in TUM RGB-D (each with a pose distance > 2.5 cm). Features and our 2D covariance matrices are extracted with the evaluated detectors. Pairwise matching is done across frames with MNN. In TUM, since more than two images are used, we form feature tracks (set of 2D local features corresponding to the same 3D point) with the track separation algorithm of [13]. Matched features are triangulated with ground-truth camera poses and DLT algorithm [23], and refined with 2D-covariance-weighted Levenberg-Marquardt (LM), producing also covariance matrices for 3D point coordinates. The next frame is used for evaluation. After matching it to the reference images we obtain 2D-3D matches which are then processed with P3P LO-RANSAC [10] to filter potential outliers. Given the potential inliers, and when using no uncertainty, we choose EPnP [31] as the nonminimal PnP solver. Otherwise, when leveraging our proposed 2D covariances, and optionally, the 3D covariances from LM, we use our implementation (validation in Supp.) of EPnPU [59]. Finally, the estimated camera pose is refined with a covariance-weighted motion-only bundle adjustment. In the Supplemental, we detail how the inclusion of our uncertainty estimates is done in the previous tasks.\nTo quantify the accuracy in pose estimation at each 1. Motion estimation from 2D-3D correspondences on KITTI [19] and TUM-RGBD [53]. Sequences are specified at the leftmost column ('all' is their aggregation). erot and et are the mean absolute rotation (in 0.1×deg.) and translation (in cm.) errors. We compare estimations without using uncertainty (first two columns) and with 2D and 3D uncertainties via the proposed full and isotropic covariances. Errors are consistently reduced when using uncertainty estimates. The best result for each sequence-detector is in bold.\nsequence, we use the absolute rotation error in degrees: e rot = arccos(0.5 trace(R ⊤ true R -1)), and the absolute translation error e t = ∥t true -t∥ in cm., where R true , t true is the ground-truth pose and R, t is the estimated one."},{"figure_ref":["fig_3"],"heading":"Results","publication_ref":["b58"],"table_ref":[],"text":". Following [59], in Table 1, we report the mean errors obtained across sequences of both datasets. As can be seen, taking into account the proposed uncertainties is a key aspect to converge pose estimations across sequences. Figures 7 and8 show qualitative examples of why uncertainties help geometric estimations by assigning more uncertainty to distant or less reliable keypoints. This behavior can be understood better by having a look at the cumulative error curves. In Figure 9, it is shown that practically all pose estimations obtained with methods leveraging the proposed covariances, fall under acceptable error thresholds, whereas the ones from the baseline do not."},{"figure_ref":[],"heading":"Limitations","publication_ref":["b30","b58","b56","b52","b18"],"table_ref":[],"text":"The proposed covariances modeling the spatial uncertainty of the learned local-features are up-to-scale. This is not an issue for common 3D geometric estimation algorithms, such as solving linear systems [31,59] or nonlinear leastsquares optimizations [57], as their solutions depend only on the relative weight imposed by the covariance matrices. However, this limitation hinders reasoning about the covariances in pixel units. For instance, extracting the absolute scale factor would facilitate the use of robust cost functions, pointing towards a potential direction for future work.\nAdditionally, while we achieved improvements in 3D geometric estimation tasks on the standard datasets TUM-RGBD [53] and KITTI [19], their exposure to effects like illumination changes or different types of camera motions might be limited. These effects may pose challenges to our approach, as we believe it is subject to the equivariance of the noise of the learned score maps to such changes."},{"figure_ref":[],"heading":"Conclusions","publication_ref":["b58","b30","b30","b30"],"table_ref":[],"text":"In this paper, we formulate, for the first time in the literature, detector-agnostic models for the spatial covariances of deep local features. Specifically, we proposed two methods based on their learned score maps: one using local-feature scores directly, and another theoretically-motivated method using local structure tensors. Our experiments on KITTI and TUM show that our covariances are well calibrated, significantly benefiting 3D-geometry estimation tasks.\nPnP. Consider a set of n 3D points {p w,i ∈ R3 | i ∈ 1 . . n} expressed in an absolute reference system {w}, along with their corresponding 2D points {x i ∈ R 2 | i ∈ 1 . . n} in an image captured by a camera with calibration matrix K ∈ R 3×3 . The Perspective-n-Point problem (PnP) involves finding the rotation R cw ∈ SO(3) and translation t cw ∈ R 3 that transform the 3D points to the camera's reference system {c}: p c,i := R cw p w,i + t cw .\nEPnP(U). To leverage our proposed covariance matrices, we adopt the recent EPnPU [59] as our PnP solver. EPnPU extends EPnP [31] to take into account the uncertainty of the observations. To understand how uncertainties are included, consider the algebraic residual r i (i ∈ 1 . . n):\nr i := x(1,2) i - x(3) i x i ,(21)\nwhere xi represents, in homogeneous coordinates, the estimated 2D location corresponding to p w,i :\nxi := K(R c,w p w,i + t c,w ) ,(22)\nwhere x(1,2) i (resp.\nx(3) i ) represents its two first entries (resp. third entry). Given this, we seek to estimate the covariance matrix of each residual, Σ ri := Cov(r i ), that stems from the covariance matrices of the observations Σ xi and Σ pw,i .\nEach residual can be linearized by instead considering a set of control points as the unknowns 2 of the problem, whose concatenation we denote here as c ∈ R 12 , and a matrix block M i ∈ R 2×12 that depends only on the input data [31]. Thus, the solution to c is found in the null space of the matrix formed by the concatenation of each residual [31,Eq. 7]:\nMc = 0 , with M := M ⊤ 1 • • • M ⊤ n ⊤ . (23\n)\nThanks to this, we can weigh the influence of the residuals according to their projections onto the directions of extreme uncertainty:\nM ⊤ Σ -1 r Mc = 0 ,(24)\nwith Σ r := diag(Σ r1 , . . . , Σ rn ) .\n(25)\n2 To ease uncertainty propagation from p w,i to xi , we instead parameterize xi with the pose (Rcw, tcw), as in Eq. (22). Because of this, we need an initial rough pose estimate. This pose estimate is obtained at no additional cost during the outlier filtering step with P3P LO-RANSAC.\nΣ ri derivation. For the noise ν ∈ R 3 affecting the coordinates of each 3D point, we assume a similar model to the one proposed for the 2D location x i of each local feature (Eqs. ( 1) and (2)):\np w,i = E{p w,i } + ν ,(26)\nE{ν} = 0 , Σ pw,i := E{νν ⊤ } .(27)\nwhere, as we do in our case, the covariance matrix Σ pw,i ∈ R 3×3 can be estimated after the convergence of the nonlinear optimization of p w,i (Appendix A.2). Plugging Eq. ( 26) in Eq. ( 22) leads to\nxi = KRp w,i + Kt = E{x i } + ζ ,(28)\nwhere E{ζ} = 0, because ζ := KRν, and E{ν} = 0 (by definition in Eq. ( 26)). According to Eq. ( 28), we can propagate the uncertainty of p w,i to xi by:\nΣ xi := Λ w w ⊤ γ := KRΣ pw,i R ⊤ K ⊤ .(29)\nEach residual can then be expressed as a random variable as follows:\nr i = x(1,2) i - x(3) i x i ,(30)\n= (E{x\n(1,2) i } + ζ (1,2) ) -(E{x(3)\ni } + ζ (3) )(E{x i } + ξ) ,(31)\n= E{x (1,2) i - x(3) i x i } + ζ (1,2) -E{x (3) i }ξ -ζ (3) E{x i } -ζ (3) ξ ,(32)\nwhere we have used the linearity of the expectation and assumed independence between ξ and ζ. Following Eq. ( 30) and the fact that\nE{r i } = E{x (1,2) i - x(3\n) i x i } since E{ξ} = 0 2 , E{ζ} = 0 3 , the derivation of Σ ri follows as in Tab. 2."},{"figure_ref":[],"heading":"A.2. Nonlinear optimizations","publication_ref":["b22","b56","b22","b30","b58"],"table_ref":[],"text":"Uncertainty-based geometric refinement. Commonly, initial estimates of camera poses and 3D points (obtained, in our case, using EPnP(U) and multi-view DLT triangulation [23], respectively) are not geometrically optimal 3 . Therefore, they are typically optimized using iterative algorithms such as Gauss-Newton or Levenberg-Marquardt [57] by minimizing reprojection errors. This approach is considered in the literature as the gold standard [23]. known beforehand. Comparisons are done with OpenCV's implementation of EPnP [31]. Since EPnPU is an extension of EPnP's algorithm to include uncertainties, results should improve accordingly when leveraging them. Additionally, we compare our implementation when using just identity covariance matrices, to verify that it matches the behavior of EPnP.\nAs shown in Figure 10, our EPnPU implementation gives results very similar to the ones reported in [59]. In turn it improves the results of OpenCV's EPnP, whose behavior also matches our EPnPU implementation when just using identity covariances matrices."},{"figure_ref":["fig_4","fig_5"],"heading":"C. Interpretability","publication_ref":["b10","b4","b11","b42"],"table_ref":[],"text":"Our proposed methods for quantifying the uncertainty of the locations are based on the learned score maps, independently of the detector that has learned it. Depending on its training, systems learn to focus on different input image pat- terns. For instance, SuperPoint [11] is trained to detect corners, while Key.Net [5], D2Net [12] and R2D2 [43] do not directly impose such constraint. Additionally, all of them use different learning objectives.\nTo explore what kind of locations get assigned low uncertainty estimates, we set up a toy experiment inspired by DeepDream [38]. As depicted in Fig. 11 we update a 20 × 20 synthetic input patch via gradient descent, such that we minimize the biggest eigenvalue of the covariance matrix (computed by using the regressed score map) at the center pixel. We downweight the gradients located at the extremes of the receptive field and, as in [38], we smooth the gradients with a Gaussian filter.\nResults after convergence are shown in Fig. 12. We obtain distinctive blob/corner-like regions by minimizing both uncertainty estimates. This highlights the detectoragnostic behavior of our two methods. Interestingly, excepting Key.Net, the generated patterns are slightly different depending on the method. We attribute this to the the fact that the full approach takes into account the surrounding learned patterns, which in this case increases the saliency of the generated input image pattern."},{"figure_ref":[],"heading":"","publication_ref":[],"table_ref":[],"text":"Acknowledgements. The authors thank Alejandro Fontán Villacampa for his thoughtful comments and help with the experiments. This work was supported by the Ministerio de Universidades Scholarship FPU21/04468."},{"figure_ref":[],"heading":"Supplementary Material","publication_ref":["b56","b22","b58","b58","b30","b58","b58"],"table_ref":[],"text":"A. Accounting for uncertainty A.1. Perspective-n-Point problem\n= E{ζ (1,2) ( ( ( ( ( ( ( ( ( E{E{x (3) i }ζ (1,2) ξ ⊤ } E{ζ (1,2) }=E{ξ}=0 -E{ζ (3) ζ (1,2) E{x i } ⊤ } w E{xi} ⊤ -( ( ( ( ( ( ( ( E{ζ (3) ζ (1,2) ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( E{(\nTable 2. Derivation of Σr i (Appendix A.1).\nConsider a 3D point j expressed in {w}: p w,j ∈ R 3 , that is observed from a camera i whose pose is: R cw,i ∈ SO(3), t cw,i ∈ R 3 , then we define the reprojection error as\nwhere x ij ∈ R 2 is the 2D location in the image of a local feature corresponding to p w,j and π(•) is the projection function. Thereby, the variables of interest (poses and/or 3D points), which we represent for convenience in vectorized form as y, are refined with the following optimization\nwith e(y) := . . . e ij (y) ⊤ . . . ⊤ , (40)\nwhere W ij := Σ -1 ij represents the inverse of the covariance matrix of e ij (we use the identity matrix W ij := I 2 in the no-uncertainty baseline), thus weighing each e ij according to its projection onto the directions of extreme uncertainty.\nIn our experiments, we approximate the solution of Eq. ( 38) with ten iterations of Levenberg-Marquardt (LM) [57] i.e. we iteratively update y ← y ⊕ ∆y within the manifold of the unknowns y [7], by composing them with increments ∆y computed by solving the following linear system of equations\nwhere J := ∂e ∂∆y 0 is the Jacobian matrix of the residuals. As such, LM imposes a penalization of the magnitude of ∆y that is controlled with the damping factor λ ∈ R. We initialize λ as 10 -3 times the average of diag(J ⊤ WJ), as recommended in [23]. We also follow the recommended protocol of [23, App. A6.2] for updating λ at each iteration. [59] implementation with known synthetic noise. Our implementation, labeled as 'EPnPU imp' obtains pose errors akin to the ones of [59], improving as it should over EPnP [31], labeled as 'EPnP OpenCV', whose behavior is matched when using our implementation with identity covariance matrices, labeled as 'EPnP imp'.\nCovariances for 3D-points refinement. Since we use ground-truth camera poses for triangulation in our experiments, we directly use our proposed 2D covariances Σ ij := Σ xij as the covariance matrix of each error e ij . After convergence, we estimate the covariance matrix Σ pw,j of the 3D point as the inverse of the Hessian, following [59].\nCovariances for motion-only bundle-adjustment. The covariance matrix of each error e j 4 , when considering 3D noise, is estimated at each iteration following [59]:\ni.e. linearly propagating Σ pw,j and assuming independence between the distributions of x j and p w,j . On the other hand, if only 2D noise is considered, we consider Σ j := Σ xj ."},{"figure_ref":[],"heading":"B. Validation of EPnPU implementation","publication_ref":["b4","b10","b11","b42","b58","b58"],"table_ref":[],"text":"Since all learned detectors used in our paper [5,11,12,43] are implemented in Python, but EPnPU [59] is originally written in MATLAB 5 , we reimplemented this last one in Python to ease the workflow. As validation, we followed the same synthetic experiments done in [59], where the noise affecting the simulated observations (2D and 3D points) is 4 We drop here the subscript i to avoid clutter since just one camera is considered in motion-only BA. "}],"string":"[\n {\n \"figure_ref\": [],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b47\",\n \"b54\",\n \"b60\",\n \"b3\",\n \"b10\",\n \"b11\"\n ],\n \"table_ref\": [],\n \"text\": \"Estimating a map along with camera poses from a collection of images is a long-standing and challenging problem in computer vision [9,41,49] with relevant applications in domains such as autonomous driving and augmented reality. Recently, deep visual detectors and descriptors have shown increased resilience to extreme viewpoint and appearance changes [4, 48,55,61,64]. However, a common limitation among all these deep descriptors and detectors is the lack of a probabilistic formulation for detection noise. Consequently, downstream pose estimation relies on the assumption of constant spatial covariances, as shown in Figure 1 (a), leading to suboptimal results.\\nWe propose to model the spatial covariance of detected keypoints in deep detectors. We find that recent detectors share a common design, where a deep convolutional backbone is used to predict a score map that assigns a \\\"probabil-ity\\\" to a pixel being a point of interest. We exploit this common design space, to propose two post-hoc methods that estimate a covariance matrix for each keypoint detected in any pretrained feature detector. Our simplest method uses the score at a given pixel to initialize an isotropic covariance matrix. We illustrate the learned score maps (lower triangle) overlaid on exemplar images, along with our proposed isotropic covariances in Figure 1 (b) for two state-ofthe-art deep feature detectors, Superpoint [11] and D2Net [12]. We also suggest a theoretically-motivated method to estimate the full spatial covariance matrix of detected keypoints using the local structure tensor. The structure tensor models the local saliency of the detections in the score map. Figure 1 (c) shows the deduced full covariances. These covariances capture the larger uncertainty along edges on the learned score map. To the best of our knowledge, we are the first to model spatial uncertainties of deep detectors.\\nWe show in a series of experiments that our proposed methods for modeling the spatial covariances of detected features are directly related with the errors in matching. Accounting for them allows us to improve performance in tasks such as solving the perspective-n-point (PnP) problem and nonlinear optimizations.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Work\",\n \"publication_ref\": [\n \"b50\",\n \"b5\",\n \"b0\",\n \"b31\",\n \"b17\",\n \"b22\",\n \"b56\",\n \"b13\",\n \"b25\",\n \"b6\",\n \"b7\",\n \"b27\",\n \"b38\",\n \"b58\",\n \"b62\",\n \"b13\",\n \"b14\",\n \"b39\",\n \"b29\",\n \"b4\",\n \"b10\",\n \"b11\",\n \"b42\",\n \"b57\",\n \"b59\",\n \"b61\",\n \"b10\",\n \"b11\",\n \"b42\",\n \"b57\",\n \"b4\",\n \"b23\",\n \"b53\",\n \"b55\",\n \"b49\"\n ],\n \"table_ref\": [],\n \"text\": \"We propose to model the spatial covariances of learned local features. Modeling the spatial uncertainty of features is not a new idea, and has been studied for hand-crafted detectors. In this section, we first review hand-crafted methods and how uncertainty estimation has been proposed for them. We then describe recent progress in learned detectors.\\nHandcrafted local features. Harris et al. [22] pioneeres rotationally-invariant corner detection by using a heuristic over the eigenvalues of the local structure tensor, while Shi et al. [51] proposes its smallest eigenvalue for detection. Mikolajczyk and Schmid [37] robustifies it against scale and affine transformations. On the other hand, SIFT (DoG) [34] popularizes detection (and description) of blobs. SURF [6] reduces its execution time by using integral images and (A-)KAZE [1,2] proposes non-linear diffusion to improve the invariance to changes in scale. Lastly, FAST [45] and its extensions [32,46] stand out for achieving the lowest execution times, thanks to only requiring intensity comparisons between the neighboring pixels of an image patch.\\nUncertainty quantification in handcrafted detectors. Inclusion of spatial uncertainty of local features for estimation of geometry has been extensively studied [18,23,25,57]. However, its quantification on classical local features is still recognized as an open problem in the literature [14,26,27] and which has not been addressed yet with learned detections. Several works have shown the benefits of a precise quantification of uncertainty. For this purpose, they adapt the quantifications to specific classical detectors [8,17,28,39,59,63], assume planar surfaces [14,42] and require an accurate offline calibration [15,16]. Only recently [40] proposes a learning-based approach of spatial covariances, trained per detector, in order to weigh normalized epipolar errors [30,33]. In contrast to previous works, we propose a general formulation for their quantification, directly applicable on state-of-the-art learned detectors, seamlessly fitting them by leveraging their common characteristics.\\nLearned local features and lack of uncertainty. The dominant approach [5,11,12,35,43,58,60,62] consists on training CNNs to regress a score map over which detections are extracted via non-maximum-suppresion (NMS). Superpoint [11] is an efficient detector-descriptor of corners, robust to noise and image transformations thanks to a synthetic pre-training followed by homographic adaptation on real images. D2Net [12] shows the applicability, to the problem of local feature detection and description, of classification networks [52]. R2D2 [43] proposes a reliability measure, used to discard unmatchable points. Similarly, DISK [58] bases its learning on matches, and KeyNet [5], inspired by classical systems, proposes to learn from spatial image gradients. This work experiments with Superpoint, D2Net, R2D2, and KeyNet, but our proposal is applicable to the rest of the systems. Finally, recent works [24,54,56] exploit attention mechanisms to match pixels without explicitly using detectors. Although their inclusion in geometric estimation pipelines can be engineered [50], they suffer from lack of repeatability. Because of this, in this work, we focus on the quantification of spatial uncertainty for learned detectors.\"\n },\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"Method\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Our framework takes as input a RGB image I of spatial dimensions H × W , and outputs local features together with their spatial uncertainties. It is composed of two main com-ponents as shown in Fig. 2. (1) A pretrained feature detector and (2) our novel and detector-agnostic uncertainty module. Our methods work with any pretrained state-ofthe-art detectors, taking their learned score maps as input, and outputting reliable covariance estimates for each detected keypoint. We show that the estimated covariances are well-calibrated and improve downstream tasks such as PnP and motion-only bundle adjustment.\"\n },\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"Overview\",\n \"publication_ref\": [\n \"b4\",\n \"b10\",\n \"b11\",\n \"b42\",\n \"b57\",\n \"b59\",\n \"b61\",\n \"b22\",\n \"b56\"\n ],\n \"table_ref\": [],\n \"text\": \"Pre-trained local-feature detector. Our framework is directly applicable to the vast majority of learned detectors [5,11,12,35,43,58,60,62]. These detectors share a common architectural design, using convolutional backbones to predict a score map S ∈ R kH×kW with k ∈ R + , followed by non-maximum-suppression (NMS) to extract a sparse set of features. We leverage this standard design to make a detector-agnostic, post-hoc uncertainty module that takes the estimated score maps as input and estimates the spatial uncertainty based on the peakiness in a local region around the detected point. Our approach can be applied to these detectors without any training or fine-tuning.\\nDetector-Agnostic Feature Uncertainties. Instead of considering the 2D position of detections as deterministic locations, we propose to model their spatial uncertainties. More formally, we consider that the spatial location x i ∈ R 2 of a local feature i, detected in I, stems from perturbing its true location x i,true ∈ R 2 , with random noise ξ i ∈ R 2\\nx i = x i,true + ξ i ,(1)\\nwhose probability distribution we want to estimate. We follow the dominant model in computer vision [23,25,57], which uses second order statistics to describe the spatial uncertainty of each location\\nE{ξ i } = 0 , Σ i := Cov(ξ i ) = E{ξ i ξ ⊤ i } , ∀i . (2)\\nThereby our goal is to quantify each covariance matrix Σ i . To this end, we propose two methods: (1) A point-wise estimator of isotropic covariances (Sec. 3.2) that uses the inverse of the scores at the detected point, and (2) a full covariance estimator based on the local structure tensor (Sec. 3.3) which models the local saliency to estimate the uncertainty in all directions. We show that both approaches lead to reliable uncertainty estimates, benefiting downstream tasks. Intuitively, a peaky score map at the location of a detected feature will yield low spatial uncertainty, whereas a flatter score map will yield larger spatial uncertainty. Figure 3 qualitatively shows how our deduced 2D uncertainties relate to 3D uncertainties. In the Supplemental we explore the agnostic behavior of our covariances. \\nϕ([∇ x S, ∇ y S]) = [(∇ x S) 2 , (∇ y S) 2 , ∇ x S∇ y S]\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Point-wise Estimation\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"As the simplest estimator of the spatial uncertainty, we propose to use the regressed score of each local feature to create an isotropic covariance matrix\\nΣ i := 1 S(x i ) I 2×2 = 1/S(x i ) 0 0 1/S(x i ) .(3)\\nFig 1 shows that this estimator yields isotropic predictions of uncertainty (equal in all directions), so it only quantifies the relative scale regardless of the learned local structure.\"\n },\n {\n \"figure_ref\": [\n \"fig_2\"\n ],\n \"heading\": \"Structure Tensor\",\n \"publication_ref\": [\n \"b4\",\n \"b10\",\n \"b11\",\n \"b42\",\n \"b46\",\n \"b43\",\n \"b19\",\n \"b35\"\n ],\n \"table_ref\": [],\n \"text\": \"Quantification of local saliency motivates the use of the local structure tensor,\\nC i ∈ R 2×2 [22]. Defining [∇ x S i , ∇ y S i ] := ∂S/∂x| xi as the spatial gradient of S eval- uated at x i , C i in its local neighborhood W i (window of size u × v) is given by C i := j∈Wi w j ∂S ∂x ⊤ xj ∂S ∂x xj = j∈Wi w j (∇ x S j ) 2 ∇ x S j ∇ y S j ∇ y S j ∇ x S j (∇ y S j ) 2 ,(4)\\nPoint-wise (Sec. 3.2)\\nStructure Tensor (Sec. 3.3) Zoom Key.Net [5] Superpoint [11] D2Net [12] R2D2 [43] Figure 5. Qualitative comparison. Our estimated covariance matrices for each detected local feature are illustrated with uncertainty ellipses for four state-of-the-art pretrained detectors. Our pointwise covariance estimator yields isotropic uncertainty predictions (equal in all directions). Our proposed local structure tensor estimates the directionality of the spatial uncertainty, resulting in anisotropic predictions. For instance, Superpoint (which is trained to detect corners) has high uncertainty patterns corresponding to lines. For visualization purposes, a uniform sample of local feature detections is shown.\\nwith w j ∈ R + being the weight of pixel j, preferably defined by a Gaussian centered at x i [47]. As such, C i is a positive semidefinite (PSD) matrix, resulting from averaging the directionality of gradients and hence not canceling opposite ones (see Fig. 4).\\nThe reason why C i captures the local saliency lies in the auto-correlation function, c : R u×v → R, which averages local changes in S given small displacements δx [22]:\\nc i = j∈Wi w j (S(x j ) -S(x j + δx)) 2 ,(5)\\nwith S(x j ) indicating the score at x j . Linearly approximating S(x j + δx) ≈ S(x j ) + ∂S/∂x| xj δx, yields\\nc i ≈ j∈Wi w j (S(x j ) -S(x j ) - ∂S ∂x xj δx) 2 ,(6)\\n= j∈Wi w j δx ⊤ ∂S ∂x ⊤ xj ∂S ∂x xj δx = δx ⊤ C i δx . (7)\\nThus, extreme saliency directions are obtained by solving:\\nmax δx min δx δx ⊤ C i δx, s.t. ∥δx∥ = 1 ,(8)\\nwhere the constraint ∥δx∥ = 1 ensures non-degenerate directions, which can be obtained with Lagrange multipliers [44] i.e. by defining the Lagrangian\\nL(δx, λ) := δx ⊤ C i δx -λ(δx ⊤ δx -1) ,(9)\\ndifferentiating w.r.t. δx and setting it to 0:\\n2δx ⊤ C i -2λδx ⊤ = 0 ⇒ C i δx = λδx ,(10)\\nwe conclude that directions of extreme saliency correspond to the eigenvectors of C i . Since inverting a matrix, does the same to its eigenvalues without affecting its eigenvectors 1 ,\\nΣ i := C -1\\ni results in a proper covariance matrix (PSD) assigning greater uncertainty in the direction of less saliency and vice-versa. [20,29] and motivated by the principle of maximum entropy [36] of the local features' scores, distributed independently by\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Statistical interpretation. Under a Gaussian model of aleatoric uncertainty, common in deep learning\",\n \"publication_ref\": [\n \"b27\",\n \"b56\"\n ],\n \"table_ref\": [],\n \"text\": \"N (S(x i + t i ), σ 2 ) ,(11)\\ninspired by [28], we can set up a parametric optimization of t ∈ R 2 which maximizes the likelihood, L(t i | S) with S := {S(x j ) | x j ∈ W i }, of the observations:\\nS(x i ) = S(x i + t i ) + ε i , with ε i ∼ N (0, σ 2 ) ,(12)\\n1 Let A ∈ R m×m and det A ̸ = 0, then λ -1 v = A -1 v, with λ and v being the eigenvalues and eigenvectors of A.\\nwhere S(x i ) is the observed score, perturbed by the random noise ε i , independently affecting the rest of observations. Thus, the optimization is formulated as follows\\nti = arg max ti j∈Wi 1 σ √ 2π exp - (S(x j ) -S(x j + t i )) 2 2σ 2 . (13\\n)\\nIts solution, or maximum-likelihood-estimation (MLE), is known a priori: ti = 0, which is unbiased given our statistical model (Eq. 12), and coherent with Equation 2.\\nWith these conditions, the inverse of the Fisher information matrix, I( t), evaluated at the MLE, imposes a lower bound in the covariance matrix of the estimator ti = 0, known as Cramer-Rao Lower Bound (CRLB) [21]:\\nVar( ti ) ≥ I( ti ) -1 .(14)\\nI( t) is defined as E{s ⊤ i s i }, being the variance of the loglikelihood derivative (known as score)\\ns i := ∂ log L(t i | S)/∂t i , since E{s i } = 0 [21].\\nIn our case, it is given by\\ns i = xj ∈Wi d j := S(x j ) -S(x j + t i ) σ 2 ∂S(x j + t i ) ∂t i .(15)\\nDue to the linearity of expectation and the independence of observations,\\nE{d ⊤ j d k } = 0, ∀j ̸ = k. Thereby I( ti ) = j∈Wi E{d ⊤ j d j } ti .(16)\\nSince derivatives of Eq. 15 are applied on our deterministic model, they can go out of the expectation, and evaluating them on ti = 0 lead to\\nI( ti ) = xj ∈Wi 1 σ 4 ∂S(x j ) ∂x ⊤ xj ∂S(x j ) ∂x xj E{(S(x j ) -S(x j + t i )) 2 } ti .(17)\\nLastly, E{(S(x j ) -S(x j + t i )) 2 } ti = E{ε 2 i } = σ 2 since E{ε i } = 0, implying that our Fisher information matrix is\\nI( ti ) = 1 σ 2 xj ∈Wi ∂S(x j ) ∂x ⊤ xj ∂S(x j ) ∂x xj ,(18)\\nat the MLE. It matches the local structure tensor (Eq. 4) up to a scale factor Var(ε i ) -1 = σ -2 , unknown a priori.\\nRecalling the CRLB (Eq. 14), although achievable only asymptotically [21,57], it motivates Σ i := C -1 i as an upto-scale covariance matrix of each location x i .\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiments\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Implementation details. In our experiments, we compute the structure tensor (and relative uncertainty) independently of the learned detector: for each local feature i, spatial differentiation at S(x j ), ∀j ∈ W i , is done with Sobel filters. Integration in W i is done with a 7 × 7 window, MMA overall 0.42 0.71\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"D2Net\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Full Isotropic \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"SuperPoint\",\n \"publication_ref\": [\n \"b2\",\n \"b4\",\n \"b10\",\n \"b11\",\n \"b42\",\n \"b18\",\n \"b52\"\n ],\n \"table_ref\": [],\n \"text\": \"Full Isotropic -uncertainty + Figure 6. Uncertainty and matching accuracy on HPatches [3]. We show MMA for each detector. We organize the results into 10 uncertainty ranges, ranging from lowest to highest (x-axis), each containing an equal number of matches. Our covariances accurately model less accurate matches, with the full-based approach (structure tensor) showing higher sensitivity in D2Net, Key.Net and R2D2.\\nthe result of using an isotropic Gaussian filter with σ = 1 of cutoff frequency 3σ. Throughout all the experiments, we evaluate and extract our covariances using the detectors of state-of-the-art learned systems: Key.Net [5], Superpoint [11], D2Net [12] and R2D2 [43]. The score map used with Superpoint is the one prior to the channel-wise softmax to avoid the alteration of learned patterns crossing the boundaries of each grid. For the rest of the systems, we directly use their regressed score map. Diverse qualitative examples are shown in Figure 5.\\nFigure 7. Qualitative results on KITTI [19]. Our covariances correctly model detections with unclear spatial locations, such as faraway corners or those located at edges.\\nFigure 8. Qualitative results on TUM-RGBD [53]. We assign more spatial uncertainty at a priori less distinctive regions, such as papers and edges of the keyboard.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Matching accuracy.\",\n \"publication_ref\": [\n \"b2\",\n \"b11\",\n \"b11\",\n \"b52\",\n \"b18\"\n ],\n \"table_ref\": [],\n \"text\": \"We first test the relation between our estimated covariances with the accuracy of local-feature matching. Intuitively, local features detected with higher uncertainty should relate to less accurate matches, and vice versa. For this purpose, we consider the widely adopted HPatches dataset [3]. HPatches contains 116 sequences of 6 images each. 57 sequences exhibit significant illumination changes while the other 59 sequences undergo viewpoint changes.\\nEvaluation protocol. We base our evaluation on the one proposed by [12]. First, extraction of local features and, in our case, covariance matrices of their locations is performed for all images. For every sequence, pairwise matching is done between a reference image r and each remaining image i by using Mutual Nearest Neighbor search (MNN). We then compute the reprojection errors and their covariances with the homographies, H i,r , provided by the dataset:\\ne i,r := cart(x i -H i,r xr ) ,(19)\\nΣ ei,r := JΣ xr J ⊤ + Σ xi ,(20)\\nwhere cart(•) maps from homogeneous to Cartesian coordinates, and J := ∂e i,r /∂x r , i.e. we linearly propagate each covariance matrix Σ xr of the reference locations.\\nTo quantify the uncertainty of the match with a scalar, we use the biggest eigenvalue of the corresponding Σ ei,r . Based on them, all matches gathered in the dataset are distributed in 10 ranges from lowest to highest uncertainty estimates, such that each range has the same number of matches. To quantify the accuracy in matching at each range, we choose the mean matching accuracy error (MMA), which represents the average percentage of matches with a corresponding value of ∥e i,r ∥ below a threshold. We use the same thresholds as in [12]. Finally, we compute the mean of all the MMA values at each range. This process is repeated for all the evaluated detectors and with the proposed full and isotropic covariances.\\nResults. Figure 6 shows the averaged mean matching accuracy, MMA, at each uncertainty range. Ranges are ordered from lowest (1) to highest (10) estimated uncer- [53] and KITTI [19]. We report the cumulative errors for camera pose rotation and translation. Practically all estimations converge to acceptable thresholds when using our 2D full and 2D isotropic covariances. This is also apparent when using 3D covariances derived from our 2D ones. Without our covariances, a significant percentage of poses fail to localize. tainty. As can be seen, it exists a direct relationship between matching accuracy and both full and isotropic covariance estimates. With full covariances and for all evaluated ranges, lower uncertainty estimates imply higher accuracy in matching. However, this is not always the case when using isotropic covariances. As can be seen with R2D2, there is a certain increase in MMA for more uncertain matches. Additionally, there is a higher sensitivity of MMA to the uncertainty estimates stemming from full covariances on D2Net, Key.Net, and R2D2. This motivates the need for taking into account the learned local structure when quantifying the spatial uncertainty of the local feature, rather than basing it only on the regressed scalar estimate of the regressed score map.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Geometry estimation\",\n \"publication_ref\": [\n \"b58\",\n \"b18\",\n \"b52\",\n \"b12\",\n \"b22\",\n \"b30\",\n \"b58\",\n \"b18\",\n \"b52\"\n ],\n \"table_ref\": [],\n \"text\": \"To test the influence of our uncertainties in 3D-geometry estimation, we follow the evaluation proposed in [59]. It covers common stages in geometric estimation pipelines such as solving the perspective-n-point problem and motion-only bundle adjustment. The data used consists in the three sequences 00-02 of KITTI [19] and the first three 'freiburg 1' monocular RGB sequences of TUM-RGBD [53].\\nEvaluation protocol. KITTI is used with a temporal win-dow of two left frames, while three are used in TUM RGB-D (each with a pose distance > 2.5 cm). Features and our 2D covariance matrices are extracted with the evaluated detectors. Pairwise matching is done across frames with MNN. In TUM, since more than two images are used, we form feature tracks (set of 2D local features corresponding to the same 3D point) with the track separation algorithm of [13]. Matched features are triangulated with ground-truth camera poses and DLT algorithm [23], and refined with 2D-covariance-weighted Levenberg-Marquardt (LM), producing also covariance matrices for 3D point coordinates. The next frame is used for evaluation. After matching it to the reference images we obtain 2D-3D matches which are then processed with P3P LO-RANSAC [10] to filter potential outliers. Given the potential inliers, and when using no uncertainty, we choose EPnP [31] as the nonminimal PnP solver. Otherwise, when leveraging our proposed 2D covariances, and optionally, the 3D covariances from LM, we use our implementation (validation in Supp.) of EPnPU [59]. Finally, the estimated camera pose is refined with a covariance-weighted motion-only bundle adjustment. In the Supplemental, we detail how the inclusion of our uncertainty estimates is done in the previous tasks.\\nTo quantify the accuracy in pose estimation at each 1. Motion estimation from 2D-3D correspondences on KITTI [19] and TUM-RGBD [53]. Sequences are specified at the leftmost column ('all' is their aggregation). erot and et are the mean absolute rotation (in 0.1×deg.) and translation (in cm.) errors. We compare estimations without using uncertainty (first two columns) and with 2D and 3D uncertainties via the proposed full and isotropic covariances. Errors are consistently reduced when using uncertainty estimates. The best result for each sequence-detector is in bold.\\nsequence, we use the absolute rotation error in degrees: e rot = arccos(0.5 trace(R ⊤ true R -1)), and the absolute translation error e t = ∥t true -t∥ in cm., where R true , t true is the ground-truth pose and R, t is the estimated one.\"\n },\n {\n \"figure_ref\": [\n \"fig_3\"\n ],\n \"heading\": \"Results\",\n \"publication_ref\": [\n \"b58\"\n ],\n \"table_ref\": [],\n \"text\": \". Following [59], in Table 1, we report the mean errors obtained across sequences of both datasets. As can be seen, taking into account the proposed uncertainties is a key aspect to converge pose estimations across sequences. Figures 7 and8 show qualitative examples of why uncertainties help geometric estimations by assigning more uncertainty to distant or less reliable keypoints. This behavior can be understood better by having a look at the cumulative error curves. In Figure 9, it is shown that practically all pose estimations obtained with methods leveraging the proposed covariances, fall under acceptable error thresholds, whereas the ones from the baseline do not.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Limitations\",\n \"publication_ref\": [\n \"b30\",\n \"b58\",\n \"b56\",\n \"b52\",\n \"b18\"\n ],\n \"table_ref\": [],\n \"text\": \"The proposed covariances modeling the spatial uncertainty of the learned local-features are up-to-scale. This is not an issue for common 3D geometric estimation algorithms, such as solving linear systems [31,59] or nonlinear leastsquares optimizations [57], as their solutions depend only on the relative weight imposed by the covariance matrices. However, this limitation hinders reasoning about the covariances in pixel units. For instance, extracting the absolute scale factor would facilitate the use of robust cost functions, pointing towards a potential direction for future work.\\nAdditionally, while we achieved improvements in 3D geometric estimation tasks on the standard datasets TUM-RGBD [53] and KITTI [19], their exposure to effects like illumination changes or different types of camera motions might be limited. These effects may pose challenges to our approach, as we believe it is subject to the equivariance of the noise of the learned score maps to such changes.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusions\",\n \"publication_ref\": [\n \"b58\",\n \"b30\",\n \"b30\",\n \"b30\"\n ],\n \"table_ref\": [],\n \"text\": \"In this paper, we formulate, for the first time in the literature, detector-agnostic models for the spatial covariances of deep local features. Specifically, we proposed two methods based on their learned score maps: one using local-feature scores directly, and another theoretically-motivated method using local structure tensors. Our experiments on KITTI and TUM show that our covariances are well calibrated, significantly benefiting 3D-geometry estimation tasks.\\nPnP. Consider a set of n 3D points {p w,i ∈ R3 | i ∈ 1 . . n} expressed in an absolute reference system {w}, along with their corresponding 2D points {x i ∈ R 2 | i ∈ 1 . . n} in an image captured by a camera with calibration matrix K ∈ R 3×3 . The Perspective-n-Point problem (PnP) involves finding the rotation R cw ∈ SO(3) and translation t cw ∈ R 3 that transform the 3D points to the camera's reference system {c}: p c,i := R cw p w,i + t cw .\\nEPnP(U). To leverage our proposed covariance matrices, we adopt the recent EPnPU [59] as our PnP solver. EPnPU extends EPnP [31] to take into account the uncertainty of the observations. To understand how uncertainties are included, consider the algebraic residual r i (i ∈ 1 . . n):\\nr i := x(1,2) i - x(3) i x i ,(21)\\nwhere xi represents, in homogeneous coordinates, the estimated 2D location corresponding to p w,i :\\nxi := K(R c,w p w,i + t c,w ) ,(22)\\nwhere x(1,2) i (resp.\\nx(3) i ) represents its two first entries (resp. third entry). Given this, we seek to estimate the covariance matrix of each residual, Σ ri := Cov(r i ), that stems from the covariance matrices of the observations Σ xi and Σ pw,i .\\nEach residual can be linearized by instead considering a set of control points as the unknowns 2 of the problem, whose concatenation we denote here as c ∈ R 12 , and a matrix block M i ∈ R 2×12 that depends only on the input data [31]. Thus, the solution to c is found in the null space of the matrix formed by the concatenation of each residual [31,Eq. 7]:\\nMc = 0 , with M := M ⊤ 1 • • • M ⊤ n ⊤ . (23\\n)\\nThanks to this, we can weigh the influence of the residuals according to their projections onto the directions of extreme uncertainty:\\nM ⊤ Σ -1 r Mc = 0 ,(24)\\nwith Σ r := diag(Σ r1 , . . . , Σ rn ) .\\n(25)\\n2 To ease uncertainty propagation from p w,i to xi , we instead parameterize xi with the pose (Rcw, tcw), as in Eq. (22). Because of this, we need an initial rough pose estimate. This pose estimate is obtained at no additional cost during the outlier filtering step with P3P LO-RANSAC.\\nΣ ri derivation. For the noise ν ∈ R 3 affecting the coordinates of each 3D point, we assume a similar model to the one proposed for the 2D location x i of each local feature (Eqs. ( 1) and (2)):\\np w,i = E{p w,i } + ν ,(26)\\nE{ν} = 0 , Σ pw,i := E{νν ⊤ } .(27)\\nwhere, as we do in our case, the covariance matrix Σ pw,i ∈ R 3×3 can be estimated after the convergence of the nonlinear optimization of p w,i (Appendix A.2). Plugging Eq. ( 26) in Eq. ( 22) leads to\\nxi = KRp w,i + Kt = E{x i } + ζ ,(28)\\nwhere E{ζ} = 0, because ζ := KRν, and E{ν} = 0 (by definition in Eq. ( 26)). According to Eq. ( 28), we can propagate the uncertainty of p w,i to xi by:\\nΣ xi := Λ w w ⊤ γ := KRΣ pw,i R ⊤ K ⊤ .(29)\\nEach residual can then be expressed as a random variable as follows:\\nr i = x(1,2) i - x(3) i x i ,(30)\\n= (E{x\\n(1,2) i } + ζ (1,2) ) -(E{x(3)\\ni } + ζ (3) )(E{x i } + ξ) ,(31)\\n= E{x (1,2) i - x(3) i x i } + ζ (1,2) -E{x (3) i }ξ -ζ (3) E{x i } -ζ (3) ξ ,(32)\\nwhere we have used the linearity of the expectation and assumed independence between ξ and ζ. Following Eq. ( 30) and the fact that\\nE{r i } = E{x (1,2) i - x(3\\n) i x i } since E{ξ} = 0 2 , E{ζ} = 0 3 , the derivation of Σ ri follows as in Tab. 2.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.2. Nonlinear optimizations\",\n \"publication_ref\": [\n \"b22\",\n \"b56\",\n \"b22\",\n \"b30\",\n \"b58\"\n ],\n \"table_ref\": [],\n \"text\": \"Uncertainty-based geometric refinement. Commonly, initial estimates of camera poses and 3D points (obtained, in our case, using EPnP(U) and multi-view DLT triangulation [23], respectively) are not geometrically optimal 3 . Therefore, they are typically optimized using iterative algorithms such as Gauss-Newton or Levenberg-Marquardt [57] by minimizing reprojection errors. This approach is considered in the literature as the gold standard [23]. known beforehand. Comparisons are done with OpenCV's implementation of EPnP [31]. Since EPnPU is an extension of EPnP's algorithm to include uncertainties, results should improve accordingly when leveraging them. Additionally, we compare our implementation when using just identity covariance matrices, to verify that it matches the behavior of EPnP.\\nAs shown in Figure 10, our EPnPU implementation gives results very similar to the ones reported in [59]. In turn it improves the results of OpenCV's EPnP, whose behavior also matches our EPnPU implementation when just using identity covariances matrices.\"\n },\n {\n \"figure_ref\": [\n \"fig_4\",\n \"fig_5\"\n ],\n \"heading\": \"C. Interpretability\",\n \"publication_ref\": [\n \"b10\",\n \"b4\",\n \"b11\",\n \"b42\"\n ],\n \"table_ref\": [],\n \"text\": \"Our proposed methods for quantifying the uncertainty of the locations are based on the learned score maps, independently of the detector that has learned it. Depending on its training, systems learn to focus on different input image pat- terns. For instance, SuperPoint [11] is trained to detect corners, while Key.Net [5], D2Net [12] and R2D2 [43] do not directly impose such constraint. Additionally, all of them use different learning objectives.\\nTo explore what kind of locations get assigned low uncertainty estimates, we set up a toy experiment inspired by DeepDream [38]. As depicted in Fig. 11 we update a 20 × 20 synthetic input patch via gradient descent, such that we minimize the biggest eigenvalue of the covariance matrix (computed by using the regressed score map) at the center pixel. We downweight the gradients located at the extremes of the receptive field and, as in [38], we smooth the gradients with a Gaussian filter.\\nResults after convergence are shown in Fig. 12. We obtain distinctive blob/corner-like regions by minimizing both uncertainty estimates. This highlights the detectoragnostic behavior of our two methods. Interestingly, excepting Key.Net, the generated patterns are slightly different depending on the method. We attribute this to the the fact that the full approach takes into account the surrounding learned patterns, which in this case increases the saliency of the generated input image pattern.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Acknowledgements. The authors thank Alejandro Fontán Villacampa for his thoughtful comments and help with the experiments. This work was supported by the Ministerio de Universidades Scholarship FPU21/04468.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Supplementary Material\",\n \"publication_ref\": [\n \"b56\",\n \"b22\",\n \"b58\",\n \"b58\",\n \"b30\",\n \"b58\",\n \"b58\"\n ],\n \"table_ref\": [],\n \"text\": \"A. Accounting for uncertainty A.1. Perspective-n-Point problem\\n= E{ζ (1,2) ( ( ( ( ( ( ( ( ( E{E{x (3) i }ζ (1,2) ξ ⊤ } E{ζ (1,2) }=E{ξ}=0 -E{ζ (3) ζ (1,2) E{x i } ⊤ } w E{xi} ⊤ -( ( ( ( ( ( ( ( E{ζ (3) ζ (1,2) ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( E{(\\nTable 2. Derivation of Σr i (Appendix A.1).\\nConsider a 3D point j expressed in {w}: p w,j ∈ R 3 , that is observed from a camera i whose pose is: R cw,i ∈ SO(3), t cw,i ∈ R 3 , then we define the reprojection error as\\nwhere x ij ∈ R 2 is the 2D location in the image of a local feature corresponding to p w,j and π(•) is the projection function. Thereby, the variables of interest (poses and/or 3D points), which we represent for convenience in vectorized form as y, are refined with the following optimization\\nwith e(y) := . . . e ij (y) ⊤ . . . ⊤ , (40)\\nwhere W ij := Σ -1 ij represents the inverse of the covariance matrix of e ij (we use the identity matrix W ij := I 2 in the no-uncertainty baseline), thus weighing each e ij according to its projection onto the directions of extreme uncertainty.\\nIn our experiments, we approximate the solution of Eq. ( 38) with ten iterations of Levenberg-Marquardt (LM) [57] i.e. we iteratively update y ← y ⊕ ∆y within the manifold of the unknowns y [7], by composing them with increments ∆y computed by solving the following linear system of equations\\nwhere J := ∂e ∂∆y 0 is the Jacobian matrix of the residuals. As such, LM imposes a penalization of the magnitude of ∆y that is controlled with the damping factor λ ∈ R. We initialize λ as 10 -3 times the average of diag(J ⊤ WJ), as recommended in [23]. We also follow the recommended protocol of [23, App. A6.2] for updating λ at each iteration. [59] implementation with known synthetic noise. Our implementation, labeled as 'EPnPU imp' obtains pose errors akin to the ones of [59], improving as it should over EPnP [31], labeled as 'EPnP OpenCV', whose behavior is matched when using our implementation with identity covariance matrices, labeled as 'EPnP imp'.\\nCovariances for 3D-points refinement. Since we use ground-truth camera poses for triangulation in our experiments, we directly use our proposed 2D covariances Σ ij := Σ xij as the covariance matrix of each error e ij . After convergence, we estimate the covariance matrix Σ pw,j of the 3D point as the inverse of the Hessian, following [59].\\nCovariances for motion-only bundle-adjustment. The covariance matrix of each error e j 4 , when considering 3D noise, is estimated at each iteration following [59]:\\ni.e. linearly propagating Σ pw,j and assuming independence between the distributions of x j and p w,j . On the other hand, if only 2D noise is considered, we consider Σ j := Σ xj .\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B. Validation of EPnPU implementation\",\n \"publication_ref\": [\n \"b4\",\n \"b10\",\n \"b11\",\n \"b42\",\n \"b58\",\n \"b58\"\n ],\n \"table_ref\": [],\n \"text\": \"Since all learned detectors used in our paper [5,11,12,43] are implemented in Python, but EPnPU [59] is originally written in MATLAB 5 , we reimplemented this last one in Python to ease the workflow. As validation, we followed the same synthetic experiments done in [59], where the noise affecting the simulated observations (2D and 3D points) is 4 We drop here the subscript i to avoid clutter since just one camera is considered in motion-only BA. \"\n }\n]"},"abstract":{"kind":"string","value":"Superpoint D2Net (a) Constant Covariance (b) Isotropic Covariance (c) Full Covariance Figure 1. Detector-agnostic, post-hoc uncertainty for learned local-feature detectors. State-of-the-art deep feature detectors, such as Superpoint [11] and D2Net [12], do not estimate the spatial uncertainty of their detections. This corresponds to assuming constant covariances (see (a)), which lead to suboptimal performance. We propose two methods to model local features' uncertainty: (b) a pointwise isotropic covariance and (c) a structure-based full covariance estimate. We demonstrate that modeling uncertainties leads to improved performance in downstream tasks such as solving the perspective-n-point problem and nonlinear optimizations."},"title":{"kind":"string","value":"DAC: Detector-Agnostic Spatial Covariances for Deep Local Features"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 2 .2Figure 2. Method overview. By exploiting the dominant approach in learned local-feature detectors (represented by D) of regressing, for each input image I, a map of scores S := D(I), over which detections are done, we propose to quantify the uncertainty of each detected location xi by a mapping U(S, xi) agnostic to D.","figure_data":"","figure_id":"fig_0","figure_label":"2","figure_type":"figure"},{"figure_caption":"Figure 3 .3Figure 3. Influence of 2D uncertainties in 3D geometry. The learned score maps for two images and matched local features with highest and lowest estimated uncertainty -represented w. ellipses-for two points detected in two views. More uncertain 2D points present significantly higher 3D reprojection errors.","figure_data":"","figure_id":"fig_1","figure_label":"3","figure_type":"figure"},{"figure_caption":"Figure 4 .4Figure 4. Structure tensor as kernel. The direction of local gradients [∇xS, ∇yS] := ∂S/∂x can be averaged without loss of information with the local structure tensor. It acts as a kernel ϕ : R 2 → R 3 , mapping vectors with 180 • of difference (opposite directions) to the same point in R 3 .","figure_data":"","figure_id":"fig_2","figure_label":"4","figure_type":"figure"},{"figure_caption":"Figure 9 .9Figure9. Evaluation in TUM-RGBD[53] and KITTI[19]. We report the cumulative errors for camera pose rotation and translation. Practically all estimations converge to acceptable thresholds when using our 2D full and 2D isotropic covariances. This is also apparent when using 3D covariances derived from our 2D ones. Without our covariances, a significant percentage of poses fail to localize.","figure_data":"","figure_id":"fig_3","figure_label":"9","figure_type":"figure"},{"figure_caption":"Figure 11 .11Figure 11. Interpretability. We update a 20 × 20 input patch I by minimizing our uncertainty estimates (full and isotropic) at the center of the patch. Gradients are downweighted and smoothed, as in [38], to favor convergence.","figure_data":"","figure_id":"fig_4","figure_label":"11","figure_type":"figure"},{"figure_caption":"Figure 12 .12Figure 12. Interpretability results. Minimizing both uncertainty measures lead to distinctive (blob-like and corner-like) regions generated in the input patch. Differences obtained depending on learned detector highlights the detector-agnostic behavior of the proposals.","figure_data":"","figure_id":"fig_5","figure_label":"12","figure_type":"figure"}],"string":"[\n {\n \"figure_caption\": \"Figure 2 .2Figure 2. Method overview. By exploiting the dominant approach in learned local-feature detectors (represented by D) of regressing, for each input image I, a map of scores S := D(I), over which detections are done, we propose to quantify the uncertainty of each detected location xi by a mapping U(S, xi) agnostic to D.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 .3Figure 3. Influence of 2D uncertainties in 3D geometry. The learned score maps for two images and matched local features with highest and lowest estimated uncertainty -represented w. ellipses-for two points detected in two views. More uncertain 2D points present significantly higher 3D reprojection errors.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 4 .4Figure 4. Structure tensor as kernel. The direction of local gradients [∇xS, ∇yS] := ∂S/∂x can be averaged without loss of information with the local structure tensor. It acts as a kernel ϕ : R 2 → R 3 , mapping vectors with 180 • of difference (opposite directions) to the same point in R 3 .\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 9 .9Figure9. Evaluation in TUM-RGBD[53] and KITTI[19]. We report the cumulative errors for camera pose rotation and translation. Practically all estimations converge to acceptable thresholds when using our 2D full and 2D isotropic covariances. This is also apparent when using 3D covariances derived from our 2D ones. Without our covariances, a significant percentage of poses fail to localize.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"9\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 11 .11Figure 11. Interpretability. We update a 20 × 20 input patch I by minimizing our uncertainty estimates (full and isotropic) at the center of the patch. Gradients are downweighted and smoothed, as in [38], to favor convergence.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"11\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 12 .12Figure 12. Interpretability results. Minimizing both uncertainty measures lead to distinctive (blob-like and corner-like) regions generated in the input patch. Differences obtained depending on learned detector highlights the detector-agnostic behavior of the proposals.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_5\",\n \"figure_label\": \"12\",\n \"figure_type\": \"figure\"\n }\n]"},"authors":{"kind":"string","value":"Javier Tirado-Garín; Frederik Warburg; Javier Civera"},"references":{"kind":"list like","value":[{"authors":"Pablo Fernández Alcantarilla; Adrien Bartoli; Andrew J Davison","journal":"Springer","ref_id":"b0","title":"Kaze features","year":"2012"},{"authors":"Pablo Fernández Alcantarilla; Jesús Nuevo; Adrien Bartoli","journal":"","ref_id":"b1","title":"Fast explicit diffusion for accelerated features in nonlinear scale spaces","year":"2013"},{"authors":"Vassileios Balntas; Karel Lenc; Andrea Vedaldi; Krystian Mikolajczyk","journal":"","ref_id":"b2","title":"Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors","year":"2017"},{"authors":" O León Barbed; Javier Franc ¸ois Chadebecq; Morlana; M M José; Ana C Montiel; Murillo","journal":"Springer","ref_id":"b3","title":"Superpoint features in endoscopy","year":"2022-09-18"},{"authors":"Axel Barroso; -Laguna ; Krystian Mikolajczyk","journal":"IEEE Transactions on Pattern Analysis and Machine Intelligence","ref_id":"b4","title":"Key. net: Keypoint detection by handcrafted and learned cnn filters revisited","year":"2022"},{"authors":"Herbert Bay; Andreas Ess; Tinne Tuytelaars; Luc Van Gool","journal":"Computer vision and image understanding","ref_id":"b5","title":"Speeded-up robust features (surf)","year":"2008"},{"authors":"Luis José; Blanco-Claraco","journal":"","ref_id":"b6","title":"A tutorial on SE(3) transformation parameterizations and on-manifold optimization","year":"2022"},{"authors":"J Michael; Wojciech Brooks; Darren Chojnacki; Anton Gawley; Van Den; Hengel","journal":"IEEE","ref_id":"b7","title":"What value covariance information in estimating vision parameters","year":"2001"},{"authors":"Carlos Campos; Richard Elvira; Juan J Gómez Rodríguez; José Mm Montiel; Juan D Tardós","journal":"IEEE Transactions on Robotics","ref_id":"b8","title":"Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam","year":"2021"},{"authors":"Ondřej Chum; Jiří Matas; Josef Kittler","journal":"Springer","ref_id":"b9","title":"Locally optimized ransac","year":"2003"},{"authors":"Tomasz Daniel Detone; Andrew Malisiewicz; Rabinovich","journal":"","ref_id":"b10","title":"Superpoint: Self-supervised interest point detection and description","year":"2018"},{"authors":"Mihai Dusmanu; Ignacio Rocco; Tomas Pajdla; Marc Pollefeys; Josef Sivic; Akihiko Torii; Torsten Sattler","journal":"","ref_id":"b11","title":"D2-net: A trainable cnn for joint description and detection of local features","year":"2019"},{"authors":"Mihai Dusmanu; Johannes L Schönberger; Marc Pollefeys","journal":"Springer","ref_id":"b12","title":"Multi-view optimization of local feature geometry","year":"2020"},{"authors":"Luis Ferraz; Xavier Binefa; Francesc Moreno-Noguer","journal":"BMVA Press","ref_id":"b13","title":"Leveraging feature uncertainty in the pnp problem","year":"2014"},{"authors":"Alejandro Fontan; Laura Oliva; Javier Civera; Rudolph Triebel","journal":"IEEE Robotics and Automation Letters","ref_id":"b14","title":"Model for multi-view residual covariances based on perspective deformation","year":"2022"},{"authors":"Alejandro Fontan; Riccardo Giubilato; Laura Oliva; Javier Civera; Rudolph Triebel","journal":"IEEE Robotics and Automation Letters","ref_id":"b15","title":"Sid-slam: Semi-direct information-driven rgb-d slam","year":"2023"},{"authors":"Wolfgang Förstner; Eberhard Gülch","journal":"","ref_id":"b16","title":"A fast operator for detection and precise location of distinct points, corners and centres of circular features","year":"1987"},{"authors":"Wolfgang Förstner; Bernhard P Wrobel","journal":"Springer","ref_id":"b17","title":"Photogrammetric computer vision","year":"2016"},{"authors":"Andreas Geiger; Philip Lenz; Christoph Stiller; Raquel Urtasun","journal":"The International Journal of Robotics Research","ref_id":"b18","title":"Vision meets robotics: The kitti dataset","year":"2013"},{"authors":"Martin Fredrik K Gustafsson; Thomas B Danelljan; Schon","journal":"","ref_id":"b19","title":"Evaluating scalable bayesian deep learning methods for robust computer vision","year":"2020"},{"authors":"Wolfgang Härdle; Léopold Simar","journal":"Springer","ref_id":"b20","title":"Applied multivariate statistical analysis","year":"2007"},{"authors":"Chris Harris; Mike Stephens","journal":"Citeseer","ref_id":"b21","title":"A combined corner and edge detector","year":"1988"},{"authors":"R I Hartley; A Zisserman","journal":"Cambridge University Press","ref_id":"b22","title":"Multiple View Geometry in Computer Vision","year":"2004"},{"authors":"Wei Jiang; Eduard Trulls; Jan Hosang; Andrea Tagliasacchi; Kwang Moo; Yi ","journal":"","ref_id":"b23","title":"Cotr: Correspondence transformer for matching across images","year":"2021"},{"authors":"Kenichi Kanatani","journal":"Elsevier Science Inc","ref_id":"b24","title":"Statistical Optimization for Geometric Computation: Theory and Practice","year":"1996"},{"authors":"Kenichi Kanatani","journal":"Systems and Computers in Japan","ref_id":"b25","title":"For geometric inference from images, what kind of statistical model is necessary?","year":"2004"},{"authors":"Kenichi Kanatani","journal":"International Journal of Computer Vision","ref_id":"b26","title":"Statistical optimization for geometric fitting: Theoretical accuracy bound and high order error analysis","year":"2008"},{"authors":"Yasushi Kanazawa; Ken-Ichi Kanatani","journal":"","ref_id":"b27","title":"Do we really have to consider covariance matrices for image features?","year":"2001"},{"authors":"Alex Kendall; Yarin Gal","journal":"Advances in neural information processing systems","ref_id":"b28","title":"What uncertainties do we need in bayesian deep learning for computer vision","year":"2017"},{"authors":"Seong Hun; Lee ; Javier Civera","journal":"","ref_id":"b29","title":"Geometric interpretations of the normalized epipolar error","year":"2020"},{"authors":"Vincent Lepetit; Francesc Moreno-Noguer; Pascal Fua","journal":"International journal of computer vision","ref_id":"b30","title":"Epnp: An accurate o (n) solution to the pnp problem","year":"2009"},{"authors":"Stefan Leutenegger; Margarita Chli; Roland Y Siegwart","journal":"Ieee","ref_id":"b31","title":"Brisk: Binary robust invariant scalable keypoints","year":"2011"},{"authors":"H Christopher; Longuet- Higgins","journal":"Nature","ref_id":"b32","title":"A computer algorithm for reconstructing a scene from two projections","year":"1981"},{"authors":" David G Lowe","journal":"International journal of computer vision","ref_id":"b33","title":"Distinctive image features from scaleinvariant keypoints","year":"2004"},{"authors":"Zixin Luo; Lei Zhou; Xuyang Bai; Hongkai Chen; Jiahui Zhang; Yao Yao; Shiwei Li; Tian Fang; Long Quan","journal":"","ref_id":"b34","title":"Aslfeat: Learning local features of accurate shape and localization","year":"2020"},{"authors":"Jochen Meidow; Christian Beder; Wolfgang Förstner","journal":"ISPRS Journal of Photogrammetry and Remote Sensing","ref_id":"b35","title":"Reasoning with uncertain points, straight lines, and straight line segments in 2d","year":"2009"},{"authors":"Krystian Mikolajczyk; Cordelia Schmid","journal":"International journal of computer vision","ref_id":"b36","title":"Scale & affine invariant interest point detectors","year":"2004"},{"authors":"Alexander Mordvintsev; Christopher Olah; Mike Tyka","journal":"","ref_id":"b37","title":"Inceptionism: Going deeper into neural networks","year":"2015"},{"authors":"Dominik Muhle; Lukas Koestler; Nikolaus Demmel; Florian Bernard; Daniel Cremers","journal":"","ref_id":"b38","title":"The probabilistic normal epipolar constraint for frame-to-frame rotation optimization under uncertain feature positions","year":"2022"},{"authors":"Dominik Muhle; Lukas Koestler; Krishna Murthy Jatavallabhula; Daniel Cremers","journal":"","ref_id":"b39","title":"Learning correspondence uncertainty via differentiable nonlinear least squares","year":"2023"},{"authors":"Raul Mur; -Artal ; Juan D Tardós","journal":"IEEE transactions on robotics","ref_id":"b40","title":"Orb-slam2: An opensource slam system for monocular, stereo, and rgb-d cameras","year":"2017"},{"authors":"Songyou Peng; Peter Sturm","journal":"","ref_id":"b41","title":"Calibration wizard: A guidance system for camera calibration based on modelling geometric and corner uncertainty","year":"2019"},{"authors":"Jerome Revaud; Cesar De Souza; Martin Humenberger; Philippe Weinzaepfel","journal":"Advances in neural information processing systems","ref_id":"b42","title":"R2d2: Reliable and repeatable detector and descriptor","year":"2019"},{"authors":"Rockafellar Tyrrell","journal":"SIAM review","ref_id":"b43","title":"Lagrange multipliers and optimality","year":"1993"},{"authors":"Edward Rosten; Reid Porter; Tom Drummond","journal":"IEEE transactions on pattern analysis and machine intelligence","ref_id":"b44","title":"Faster and better: A machine learning approach to corner detection","year":"2008"},{"authors":"Ethan Rublee; Vincent Rabaud; Kurt Konolige; Gary Bradski","journal":"Ieee","ref_id":"b45","title":"Orb: An efficient alternative to sift or surf","year":"2011"},{"authors":"Javier Sánchez; Nelson Monzón; Agustín Salgado De; La Nuez","journal":"Image Processing On Line","ref_id":"b46","title":"An analysis and implementation of the harris corner detector","year":"2018"},{"authors":"Paul-Edouard Sarlin; Cesar Cadena; Roland Siegwart; Marcin Dymczyk","journal":"","ref_id":"b47","title":"From coarse to fine: Robust hierarchical localization at large scale","year":"2019"},{"authors":"L Johannes; Jan-Michael Schonberger; Frahm","journal":"","ref_id":"b48","title":"Structurefrom-motion revisited","year":"2016"},{"authors":"Zehong Shen; Jiaming Sun; Yuang Wang; Xingyi He; Hujun Bao; Xiaowei Zhou","journal":"IEEE Transactions on Pattern Analysis and Machine Intelligence","ref_id":"b49","title":"Semi-dense feature matching with transformers and its applications in multiple-view geometry","year":"2022"},{"authors":"Jianbo Shi","journal":"IEEE","ref_id":"b50","title":"Good features to track","year":"1994"},{"authors":"K Simonyan; Zisserman","journal":"Computational and Biological Learning Society","ref_id":"b51","title":"Very deep convolutional networks for large-scale image recognition","year":"2015"},{"authors":"Jürgen Sturm; Nikolas Engelhard; Felix Endres; Wolfram Burgard; Daniel Cremers","journal":"IEEE","ref_id":"b52","title":"A benchmark for the evaluation of rgb-d slam systems","year":"2012"},{"authors":"Jiaming Sun; Zehong Shen; Yuang Wang; Hujun Bao; Xiaowei Zhou","journal":"","ref_id":"b53","title":"Loftr: Detector-free local feature matching with transformers","year":"2021"},{"authors":"Jiaming Sun; Zihao Wang; Siyu Zhang; Xingyi He; Hongcheng Zhao; Guofeng Zhang; Xiaowei Zhou","journal":"","ref_id":"b54","title":"Onepose: One-shot object pose estimation without cad models","year":"2022"},{"authors":"Shitao Tang; Jiahui Zhang; Siyu Zhu; Ping Tan","journal":"ICLR","ref_id":"b55","title":"Quadtree attention for vision transformers","year":"2022"},{"authors":"Bill Triggs; Richard I Philip F Mclauchlan; Andrew W Hartley; Fitzgibbon","journal":"Springer","ref_id":"b56","title":"Bundle adjustment-a modern synthesis","year":"1999"},{"authors":"Michał Tyszkiewicz; Pascal Fua; Eduard Trulls","journal":"Advances in Neural Information Processing Systems","ref_id":"b57","title":"Disk: Learning local features with policy gradient","year":"2020"},{"authors":"Alexander Vakhitov; Luis Ferraz; Antonio Agudo; Francesc Moreno-Noguer","journal":"","ref_id":"b58","title":"Uncertainty-aware camera pose estimation from points and lines","year":"2021"},{"authors":"Yannick Verdie; Kwang Yi; Pascal Fua; Vincent Lepetit","journal":"","ref_id":"b59","title":"Tilde: A temporally invariant learned detector","year":"2015"},{"authors":"Kuan Xu; Yuefan Hao; Chen Wang; Lihua Xie","journal":"","ref_id":"b60","title":"Airvo: An illumination-robust point-line visual odometry","year":"2022"},{"authors":"Kwang Moo; Yi ; Eduard Trulls; Vincent Lepetit; Pascal Fua","journal":"Springer","ref_id":"b61","title":"Lift: Learned invariant feature transform","year":"2016"},{"authors":"Bernhard Zeisl; Pierre Fite Georgel; Florian Schweiger; G Eckehard; Nassir Steinbach; G Navab; Munich","journal":"","ref_id":"b62","title":"Estimation of location uncertainty for scale invariant features points","year":"2009"},{"authors":"Qunjie Zhou; Sérgio Agostinho; Aljoša Ošep; Laura Leal-Taixé","journal":"Springer","ref_id":"b63","title":"Is geometry enough for matching in visual localization","year":"2022"}],"string":"[\n {\n \"authors\": \"Pablo Fernández Alcantarilla; Adrien Bartoli; Andrew J Davison\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b0\",\n \"title\": \"Kaze features\",\n \"year\": \"2012\"\n },\n {\n \"authors\": \"Pablo Fernández Alcantarilla; Jesús Nuevo; Adrien Bartoli\",\n \"journal\": \"\",\n \"ref_id\": \"b1\",\n \"title\": \"Fast explicit diffusion for accelerated features in nonlinear scale spaces\",\n \"year\": \"2013\"\n },\n {\n \"authors\": \"Vassileios Balntas; Karel Lenc; Andrea Vedaldi; Krystian Mikolajczyk\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \" O León Barbed; Javier Franc ¸ois Chadebecq; Morlana; M M José; Ana C Montiel; Murillo\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b3\",\n \"title\": \"Superpoint features in endoscopy\",\n \"year\": \"2022-09-18\"\n },\n {\n \"authors\": \"Axel Barroso; -Laguna ; Krystian Mikolajczyk\",\n \"journal\": \"IEEE Transactions on Pattern Analysis and Machine Intelligence\",\n \"ref_id\": \"b4\",\n \"title\": \"Key. net: Keypoint detection by handcrafted and learned cnn filters revisited\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Herbert Bay; Andreas Ess; Tinne Tuytelaars; Luc Van Gool\",\n \"journal\": \"Computer vision and image understanding\",\n \"ref_id\": \"b5\",\n \"title\": \"Speeded-up robust features (surf)\",\n \"year\": \"2008\"\n },\n {\n \"authors\": \"Luis José; Blanco-Claraco\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"A tutorial on SE(3) transformation parameterizations and on-manifold optimization\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"J Michael; Wojciech Brooks; Darren Chojnacki; Anton Gawley; Van Den; Hengel\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b7\",\n \"title\": \"What value covariance information in estimating vision parameters\",\n \"year\": \"2001\"\n },\n {\n \"authors\": \"Carlos Campos; Richard Elvira; Juan J Gómez Rodríguez; José Mm Montiel; Juan D Tardós\",\n \"journal\": \"IEEE Transactions on Robotics\",\n \"ref_id\": \"b8\",\n \"title\": \"Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Ondřej Chum; Jiří Matas; Josef Kittler\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b9\",\n \"title\": \"Locally optimized ransac\",\n \"year\": \"2003\"\n },\n {\n \"authors\": \"Tomasz Daniel Detone; Andrew Malisiewicz; Rabinovich\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"Superpoint: Self-supervised interest point detection and description\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Mihai Dusmanu; Ignacio Rocco; Tomas Pajdla; Marc Pollefeys; Josef Sivic; Akihiko Torii; Torsten Sattler\",\n \"journal\": \"\",\n \"ref_id\": \"b11\",\n \"title\": \"D2-net: A trainable cnn for joint description and detection of local features\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Mihai Dusmanu; Johannes L Schönberger; Marc Pollefeys\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b12\",\n \"title\": \"Multi-view optimization of local feature geometry\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Luis Ferraz; Xavier Binefa; Francesc Moreno-Noguer\",\n \"journal\": \"BMVA Press\",\n \"ref_id\": \"b13\",\n \"title\": \"Leveraging feature uncertainty in the pnp problem\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"Alejandro Fontan; Laura Oliva; Javier Civera; Rudolph Triebel\",\n \"journal\": \"IEEE Robotics and Automation Letters\",\n \"ref_id\": \"b14\",\n \"title\": \"Model for multi-view residual covariances based on perspective deformation\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Alejandro Fontan; Riccardo Giubilato; Laura Oliva; Javier Civera; Rudolph Triebel\",\n \"journal\": \"IEEE Robotics and Automation Letters\",\n \"ref_id\": \"b15\",\n \"title\": \"Sid-slam: Semi-direct information-driven rgb-d slam\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Wolfgang Förstner; Eberhard Gülch\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"A fast operator for detection and precise location of distinct points, corners and centres of circular features\",\n \"year\": \"1987\"\n },\n {\n \"authors\": \"Wolfgang Förstner; Bernhard P Wrobel\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b17\",\n \"title\": \"Photogrammetric computer vision\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Andreas Geiger; Philip Lenz; Christoph Stiller; Raquel Urtasun\",\n \"journal\": \"The International Journal of Robotics Research\",\n \"ref_id\": \"b18\",\n \"title\": \"Vision meets robotics: The kitti dataset\",\n \"year\": \"2013\"\n },\n {\n \"authors\": \"Martin Fredrik K Gustafsson; Thomas B Danelljan; Schon\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"Evaluating scalable bayesian deep learning methods for robust computer vision\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Wolfgang Härdle; Léopold Simar\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b20\",\n \"title\": \"Applied multivariate statistical analysis\",\n \"year\": \"2007\"\n },\n {\n \"authors\": \"Chris Harris; Mike Stephens\",\n \"journal\": \"Citeseer\",\n \"ref_id\": \"b21\",\n \"title\": \"A combined corner and edge detector\",\n \"year\": \"1988\"\n },\n {\n \"authors\": \"R I Hartley; A Zisserman\",\n \"journal\": \"Cambridge University Press\",\n \"ref_id\": \"b22\",\n \"title\": \"Multiple View Geometry in Computer Vision\",\n \"year\": \"2004\"\n },\n {\n \"authors\": \"Wei Jiang; Eduard Trulls; Jan Hosang; Andrea Tagliasacchi; Kwang Moo; Yi \",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"Cotr: Correspondence transformer for matching across images\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Kenichi Kanatani\",\n \"journal\": \"Elsevier Science Inc\",\n \"ref_id\": \"b24\",\n \"title\": \"Statistical Optimization for Geometric Computation: Theory and Practice\",\n \"year\": \"1996\"\n },\n {\n \"authors\": \"Kenichi Kanatani\",\n \"journal\": \"Systems and Computers in Japan\",\n \"ref_id\": \"b25\",\n \"title\": \"For geometric inference from images, what kind of statistical model is necessary?\",\n \"year\": \"2004\"\n },\n {\n \"authors\": \"Kenichi Kanatani\",\n \"journal\": \"International Journal of Computer Vision\",\n \"ref_id\": \"b26\",\n \"title\": \"Statistical optimization for geometric fitting: Theoretical accuracy bound and high order error analysis\",\n \"year\": \"2008\"\n },\n {\n \"authors\": \"Yasushi Kanazawa; Ken-Ichi Kanatani\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"Do we really have to consider covariance matrices for image features?\",\n \"year\": \"2001\"\n },\n {\n \"authors\": \"Alex Kendall; Yarin Gal\",\n \"journal\": \"Advances in neural information processing systems\",\n \"ref_id\": \"b28\",\n \"title\": \"What uncertainties do we need in bayesian deep learning for computer vision\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Seong Hun; Lee ; Javier Civera\",\n \"journal\": \"\",\n \"ref_id\": \"b29\",\n \"title\": \"Geometric interpretations of the normalized epipolar error\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Vincent Lepetit; Francesc Moreno-Noguer; Pascal Fua\",\n \"journal\": \"International journal of computer vision\",\n \"ref_id\": \"b30\",\n \"title\": \"Epnp: An accurate o (n) solution to the pnp problem\",\n \"year\": \"2009\"\n },\n {\n \"authors\": \"Stefan Leutenegger; Margarita Chli; Roland Y Siegwart\",\n \"journal\": \"Ieee\",\n \"ref_id\": \"b31\",\n \"title\": \"Brisk: Binary robust invariant scalable keypoints\",\n \"year\": \"2011\"\n },\n {\n \"authors\": \"H Christopher; Longuet- Higgins\",\n \"journal\": \"Nature\",\n \"ref_id\": \"b32\",\n \"title\": \"A computer algorithm for reconstructing a scene from two projections\",\n \"year\": \"1981\"\n },\n {\n \"authors\": \" David G Lowe\",\n \"journal\": \"International journal of computer vision\",\n \"ref_id\": \"b33\",\n \"title\": \"Distinctive image features from scaleinvariant keypoints\",\n \"year\": \"2004\"\n },\n {\n \"authors\": \"Zixin Luo; Lei Zhou; Xuyang Bai; Hongkai Chen; Jiahui Zhang; Yao Yao; Shiwei Li; Tian Fang; Long Quan\",\n \"journal\": \"\",\n \"ref_id\": \"b34\",\n \"title\": \"Aslfeat: Learning local features of accurate shape and localization\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Jochen Meidow; Christian Beder; Wolfgang Förstner\",\n \"journal\": \"ISPRS Journal of Photogrammetry and Remote Sensing\",\n \"ref_id\": \"b35\",\n \"title\": \"Reasoning with uncertain points, straight lines, and straight line segments in 2d\",\n \"year\": \"2009\"\n },\n {\n \"authors\": \"Krystian Mikolajczyk; Cordelia Schmid\",\n \"journal\": \"International journal of computer vision\",\n \"ref_id\": \"b36\",\n \"title\": \"Scale & affine invariant interest point detectors\",\n \"year\": \"2004\"\n },\n {\n \"authors\": \"Alexander Mordvintsev; Christopher Olah; Mike Tyka\",\n \"journal\": \"\",\n \"ref_id\": \"b37\",\n \"title\": \"Inceptionism: Going deeper into neural networks\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Dominik Muhle; Lukas Koestler; Nikolaus Demmel; Florian Bernard; Daniel Cremers\",\n \"journal\": \"\",\n \"ref_id\": \"b38\",\n \"title\": \"The probabilistic normal epipolar constraint for frame-to-frame rotation optimization under uncertain feature positions\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Dominik Muhle; Lukas Koestler; Krishna Murthy Jatavallabhula; Daniel Cremers\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"Learning correspondence uncertainty via differentiable nonlinear least squares\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Raul Mur; -Artal ; Juan D Tardós\",\n \"journal\": \"IEEE transactions on robotics\",\n \"ref_id\": \"b40\",\n \"title\": \"Orb-slam2: An opensource slam system for monocular, stereo, and rgb-d cameras\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Songyou Peng; Peter Sturm\",\n \"journal\": \"\",\n \"ref_id\": \"b41\",\n \"title\": \"Calibration wizard: A guidance system for camera calibration based on modelling geometric and corner uncertainty\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Jerome Revaud; Cesar De Souza; Martin Humenberger; Philippe Weinzaepfel\",\n \"journal\": \"Advances in neural information processing systems\",\n \"ref_id\": \"b42\",\n \"title\": \"R2d2: Reliable and repeatable detector and descriptor\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Rockafellar Tyrrell\",\n \"journal\": \"SIAM review\",\n \"ref_id\": \"b43\",\n \"title\": \"Lagrange multipliers and optimality\",\n \"year\": \"1993\"\n },\n {\n \"authors\": \"Edward Rosten; Reid Porter; Tom Drummond\",\n \"journal\": \"IEEE transactions on pattern analysis and machine intelligence\",\n \"ref_id\": \"b44\",\n \"title\": \"Faster and better: A machine learning approach to corner detection\",\n \"year\": \"2008\"\n },\n {\n \"authors\": \"Ethan Rublee; Vincent Rabaud; Kurt Konolige; Gary Bradski\",\n \"journal\": \"Ieee\",\n \"ref_id\": \"b45\",\n \"title\": \"Orb: An efficient alternative to sift or surf\",\n \"year\": \"2011\"\n },\n {\n \"authors\": \"Javier Sánchez; Nelson Monzón; Agustín Salgado De; La Nuez\",\n \"journal\": \"Image Processing On Line\",\n \"ref_id\": \"b46\",\n \"title\": \"An analysis and implementation of the harris corner detector\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Paul-Edouard Sarlin; Cesar Cadena; Roland Siegwart; Marcin Dymczyk\",\n \"journal\": \"\",\n \"ref_id\": \"b47\",\n \"title\": \"From coarse to fine: Robust hierarchical localization at large scale\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"L Johannes; Jan-Michael Schonberger; Frahm\",\n \"journal\": \"\",\n \"ref_id\": \"b48\",\n \"title\": \"Structurefrom-motion revisited\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Zehong Shen; Jiaming Sun; Yuang Wang; Xingyi He; Hujun Bao; Xiaowei Zhou\",\n \"journal\": \"IEEE Transactions on Pattern Analysis and Machine Intelligence\",\n \"ref_id\": \"b49\",\n \"title\": \"Semi-dense feature matching with transformers and its applications in multiple-view geometry\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jianbo Shi\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b50\",\n \"title\": \"Good features to track\",\n \"year\": \"1994\"\n },\n {\n \"authors\": \"K Simonyan; Zisserman\",\n \"journal\": \"Computational and Biological Learning Society\",\n \"ref_id\": \"b51\",\n \"title\": \"Very deep convolutional networks for large-scale image recognition\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Jürgen Sturm; Nikolas Engelhard; Felix Endres; Wolfram Burgard; Daniel Cremers\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b52\",\n \"title\": \"A benchmark for the evaluation of rgb-d slam systems\",\n \"year\": \"2012\"\n },\n {\n \"authors\": \"Jiaming Sun; Zehong Shen; Yuang Wang; Hujun Bao; Xiaowei Zhou\",\n \"journal\": \"\",\n \"ref_id\": \"b53\",\n \"title\": \"Loftr: Detector-free local feature matching with transformers\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Jiaming Sun; Zihao Wang; Siyu Zhang; Xingyi He; Hongcheng Zhao; Guofeng Zhang; Xiaowei Zhou\",\n \"journal\": \"\",\n \"ref_id\": \"b54\",\n \"title\": \"Onepose: One-shot object pose estimation without cad models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Shitao Tang; Jiahui Zhang; Siyu Zhu; Ping Tan\",\n \"journal\": \"ICLR\",\n \"ref_id\": \"b55\",\n \"title\": \"Quadtree attention for vision transformers\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Bill Triggs; Richard I Philip F Mclauchlan; Andrew W Hartley; Fitzgibbon\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b56\",\n \"title\": \"Bundle adjustment-a modern synthesis\",\n \"year\": \"1999\"\n },\n {\n \"authors\": \"Michał Tyszkiewicz; Pascal Fua; Eduard Trulls\",\n \"journal\": \"Advances in Neural Information Processing Systems\",\n \"ref_id\": \"b57\",\n \"title\": \"Disk: Learning local features with policy gradient\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Alexander Vakhitov; Luis Ferraz; Antonio Agudo; Francesc Moreno-Noguer\",\n \"journal\": \"\",\n \"ref_id\": \"b58\",\n \"title\": \"Uncertainty-aware camera pose estimation from points and lines\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Yannick Verdie; Kwang Yi; Pascal Fua; Vincent Lepetit\",\n \"journal\": \"\",\n \"ref_id\": \"b59\",\n \"title\": \"Tilde: A temporally invariant learned detector\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Kuan Xu; Yuefan Hao; Chen Wang; Lihua Xie\",\n \"journal\": \"\",\n \"ref_id\": \"b60\",\n \"title\": \"Airvo: An illumination-robust point-line visual odometry\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Kwang Moo; Yi ; Eduard Trulls; Vincent Lepetit; Pascal Fua\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b61\",\n \"title\": \"Lift: Learned invariant feature transform\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Bernhard Zeisl; Pierre Fite Georgel; Florian Schweiger; G Eckehard; Nassir Steinbach; G Navab; Munich\",\n \"journal\": \"\",\n \"ref_id\": \"b62\",\n \"title\": \"Estimation of location uncertainty for scale invariant features points\",\n \"year\": \"2009\"\n },\n {\n \"authors\": \"Qunjie Zhou; Sérgio Agostinho; Aljoša Ošep; Laura Leal-Taixé\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b63\",\n \"title\": \"Is geometry enough for matching in visual localization\",\n \"year\": \"2022\"\n }\n]"},"formulas":{"kind":"list like","value":[{"formula_coordinates":[3,130.91,436.47,155.46,10.62],"formula_id":"formula_0","formula_text":"x i = x i,true + ξ i ,(1)"},{"formula_coordinates":[3,61.02,519.88,225.34,13.03],"formula_id":"formula_1","formula_text":"E{ξ i } = 0 , Σ i := Cov(ξ i ) = E{ξ i ξ ⊤ i } , ∀i . (2)"},{"formula_coordinates":[3,333.66,275.96,186.92,11.23],"formula_id":"formula_2","formula_text":"ϕ([∇ x S, ∇ y S]) = [(∇ x S) 2 , (∇ y S) 2 , ∇ x S∇ y S]"},{"formula_coordinates":[3,336.03,531.58,209.08,23.23],"formula_id":"formula_3","formula_text":"Σ i := 1 S(x i ) I 2×2 = 1/S(x i ) 0 0 1/S(x i ) .(3)"},{"formula_coordinates":[3,308.86,636.1,236.25,79.14],"formula_id":"formula_4","formula_text":"C i ∈ R 2×2 [22]. Defining [∇ x S i , ∇ y S i ] := ∂S/∂x| xi as the spatial gradient of S eval- uated at x i , C i in its local neighborhood W i (window of size u × v) is given by C i := j∈Wi w j ∂S ∂x ⊤ xj ∂S ∂x xj = j∈Wi w j (∇ x S j ) 2 ∇ x S j ∇ y S j ∇ y S j ∇ x S j (∇ y S j ) 2 ,(4)"},{"formula_coordinates":[5,90.52,178.65,195.85,22.13],"formula_id":"formula_5","formula_text":"c i = j∈Wi w j (S(x j ) -S(x j + δx)) 2 ,(5)"},{"formula_coordinates":[5,61.33,244.1,225.04,26.8],"formula_id":"formula_6","formula_text":"c i ≈ j∈Wi w j (S(x j ) -S(x j ) - ∂S ∂x xj δx) 2 ,(6)"},{"formula_coordinates":[5,71.72,275.97,214.64,30],"formula_id":"formula_7","formula_text":"= j∈Wi w j δx ⊤ ∂S ∂x ⊤ xj ∂S ∂x xj δx = δx ⊤ C i δx . (7)"},{"formula_coordinates":[5,91.72,336.01,194.65,30.53],"formula_id":"formula_8","formula_text":"max δx min δx δx ⊤ C i δx, s.t. ∥δx∥ = 1 ,(8)"},{"formula_coordinates":[5,84.39,418.17,201.97,11.72],"formula_id":"formula_9","formula_text":"L(δx, λ) := δx ⊤ C i δx -λ(δx ⊤ δx -1) ,(9)"},{"formula_coordinates":[5,84.81,458.61,201.55,11.72],"formula_id":"formula_10","formula_text":"2δx ⊤ C i -2λδx ⊤ = 0 ⇒ C i δx = λδx ,(10)"},{"formula_coordinates":[5,50.11,514.55,45.74,11.87],"formula_id":"formula_11","formula_text":"Σ i := C -1"},{"formula_coordinates":[5,126.48,610.64,159.88,11.72],"formula_id":"formula_12","formula_text":"N (S(x i + t i ), σ 2 ) ,(11)"},{"formula_coordinates":[5,55.09,674.99,231.27,11.72],"formula_id":"formula_13","formula_text":"S(x i ) = S(x i + t i ) + ε i , with ε i ∼ N (0, σ 2 ) ,(12)"},{"formula_coordinates":[5,314.51,115.84,226.45,23.62],"formula_id":"formula_14","formula_text":"ti = arg max ti j∈Wi 1 σ √ 2π exp - (S(x j ) -S(x j + t i )) 2 2σ 2 . (13"},{"formula_coordinates":[5,540.96,121.94,4.15,8.64],"formula_id":"formula_15","formula_text":")"},{"formula_coordinates":[5,385.48,236.99,159.63,11.03],"formula_id":"formula_16","formula_text":"Var( ti ) ≥ I( ti ) -1 .(14)"},{"formula_coordinates":[5,308.86,269.66,236.25,21.95],"formula_id":"formula_17","formula_text":"s i := ∂ log L(t i | S)/∂t i , since E{s i } = 0 [21]."},{"formula_coordinates":[5,314.74,300.08,230.37,24.15],"formula_id":"formula_18","formula_text":"s i = xj ∈Wi d j := S(x j ) -S(x j + t i ) σ 2 ∂S(x j + t i ) ∂t i .(15)"},{"formula_coordinates":[5,363.96,344.08,181.16,44.08],"formula_id":"formula_19","formula_text":"E{d ⊤ j d k } = 0, ∀j ̸ = k. Thereby I( ti ) = j∈Wi E{d ⊤ j d j } ti .(16)"},{"formula_coordinates":[5,315.97,437.51,229.15,48.12],"formula_id":"formula_20","formula_text":"I( ti ) = xj ∈Wi 1 σ 4 ∂S(x j ) ∂x ⊤ xj ∂S(x j ) ∂x xj E{(S(x j ) -S(x j + t i )) 2 } ti .(17)"},{"formula_coordinates":[5,329.72,527.25,215.39,30.01],"formula_id":"formula_21","formula_text":"I( ti ) = 1 σ 2 xj ∈Wi ∂S(x j ) ∂x ⊤ xj ∂S(x j ) ∂x xj ,(18)"},{"formula_coordinates":[6,375,438.19,170.11,10],"formula_id":"formula_22","formula_text":"e i,r := cart(x i -H i,r xr ) ,(19)"},{"formula_coordinates":[6,367.99,452.99,177.12,11.72],"formula_id":"formula_23","formula_text":"Σ ei,r := JΣ xr J ⊤ + Σ xi ,(20)"},{"formula_coordinates":[12,120.73,338.9,165.64,11.28],"formula_id":"formula_24","formula_text":"r i := x(1,2) i - x(3) i x i ,(21)"},{"formula_coordinates":[12,111.41,394.35,174.95,10],"formula_id":"formula_25","formula_text":"xi := K(R c,w p w,i + t c,w ) ,(22)"},{"formula_coordinates":[12,61.86,570.07,220.35,14.58],"formula_id":"formula_26","formula_text":"Mc = 0 , with M := M ⊤ 1 • • • M ⊤ n ⊤ . (23"},{"formula_coordinates":[12,282.21,574.4,4.15,8.64],"formula_id":"formula_27","formula_text":")"},{"formula_coordinates":[12,125.3,639.15,161.06,12.95],"formula_id":"formula_28","formula_text":"M ⊤ Σ -1 r Mc = 0 ,(24)"},{"formula_coordinates":[12,356.08,184.78,189.03,9.68],"formula_id":"formula_29","formula_text":"p w,i = E{p w,i } + ν ,(26)"},{"formula_coordinates":[12,357.74,199.27,187.37,11.72],"formula_id":"formula_30","formula_text":"E{ν} = 0 , Σ pw,i := E{νν ⊤ } .(27)"},{"formula_coordinates":[12,354.42,279.75,190.69,9.79],"formula_id":"formula_31","formula_text":"xi = KRp w,i + Kt = E{x i } + ζ ,(28)"},{"formula_coordinates":[12,342.1,346.72,203.01,20.81],"formula_id":"formula_32","formula_text":"Σ xi := Λ w w ⊤ γ := KRΣ pw,i R ⊤ K ⊤ .(29)"},{"formula_coordinates":[12,326.15,412.72,218.96,11.07],"formula_id":"formula_33","formula_text":"r i = x(1,2) i - x(3) i x i ,(30)"},{"formula_coordinates":[12,356.33,427.7,73.47,29.85],"formula_id":"formula_34","formula_text":"(1,2) i } + ζ (1,2) ) -(E{x(3)"},{"formula_coordinates":[12,387.84,439.23,157.28,20.52],"formula_id":"formula_35","formula_text":"i } + ζ (3) )(E{x i } + ξ) ,(31)"},{"formula_coordinates":[12,336.96,463.66,208.16,30.04],"formula_id":"formula_36","formula_text":"= E{x (1,2) i - x(3) i x i } + ζ (1,2) -E{x (3) i }ξ -ζ (3) E{x i } -ζ (3) ξ ,(32)"},{"formula_coordinates":[12,373.39,527.19,93.84,14.07],"formula_id":"formula_37","formula_text":"E{r i } = E{x (1,2) i - x(3"}],"string":"[\n {\n \"formula_coordinates\": [\n 3,\n 130.91,\n 436.47,\n 155.46,\n 10.62\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"x i = x i,true + ξ i ,(1)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 61.02,\n 519.88,\n 225.34,\n 13.03\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"E{ξ i } = 0 , Σ i := Cov(ξ i ) = E{ξ i ξ ⊤ i } , ∀i . (2)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 333.66,\n 275.96,\n 186.92,\n 11.23\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"ϕ([∇ x S, ∇ y S]) = [(∇ x S) 2 , (∇ y S) 2 , ∇ x S∇ y S]\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 336.03,\n 531.58,\n 209.08,\n 23.23\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"Σ i := 1 S(x i ) I 2×2 = 1/S(x i ) 0 0 1/S(x i ) .(3)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 308.86,\n 636.1,\n 236.25,\n 79.14\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"C i ∈ R 2×2 [22]. Defining [∇ x S i , ∇ y S i ] := ∂S/∂x| xi as the spatial gradient of S eval- uated at x i , C i in its local neighborhood W i (window of size u × v) is given by C i := j∈Wi w j ∂S ∂x ⊤ xj ∂S ∂x xj = j∈Wi w j (∇ x S j ) 2 ∇ x S j ∇ y S j ∇ y S j ∇ x S j (∇ y S j ) 2 ,(4)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 90.52,\n 178.65,\n 195.85,\n 22.13\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"c i = j∈Wi w j (S(x j ) -S(x j + δx)) 2 ,(5)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 61.33,\n 244.1,\n 225.04,\n 26.8\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"c i ≈ j∈Wi w j (S(x j ) -S(x j ) - ∂S ∂x xj δx) 2 ,(6)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 71.72,\n 275.97,\n 214.64,\n 30\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"= j∈Wi w j δx ⊤ ∂S ∂x ⊤ xj ∂S ∂x xj δx = δx ⊤ C i δx . (7)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 91.72,\n 336.01,\n 194.65,\n 30.53\n ],\n \"formula_id\": \"formula_8\",\n \"formula_text\": \"max δx min δx δx ⊤ C i δx, s.t. ∥δx∥ = 1 ,(8)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 84.39,\n 418.17,\n 201.97,\n 11.72\n ],\n \"formula_id\": \"formula_9\",\n \"formula_text\": \"L(δx, λ) := δx ⊤ C i δx -λ(δx ⊤ δx -1) ,(9)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 84.81,\n 458.61,\n 201.55,\n 11.72\n ],\n \"formula_id\": \"formula_10\",\n \"formula_text\": \"2δx ⊤ C i -2λδx ⊤ = 0 ⇒ C i δx = λδx ,(10)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 50.11,\n 514.55,\n 45.74,\n 11.87\n ],\n \"formula_id\": \"formula_11\",\n \"formula_text\": \"Σ i := C -1\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 126.48,\n 610.64,\n 159.88,\n 11.72\n ],\n \"formula_id\": \"formula_12\",\n \"formula_text\": \"N (S(x i + t i ), σ 2 ) ,(11)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 55.09,\n 674.99,\n 231.27,\n 11.72\n ],\n \"formula_id\": \"formula_13\",\n \"formula_text\": \"S(x i ) = S(x i + t i ) + ε i , with ε i ∼ N (0, σ 2 ) ,(12)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 314.51,\n 115.84,\n 226.45,\n 23.62\n ],\n \"formula_id\": \"formula_14\",\n \"formula_text\": \"ti = arg max ti j∈Wi 1 σ √ 2π exp - (S(x j ) -S(x j + t i )) 2 2σ 2 . (13\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 540.96,\n 121.94,\n 4.15,\n 8.64\n ],\n \"formula_id\": \"formula_15\",\n \"formula_text\": \")\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 385.48,\n 236.99,\n 159.63,\n 11.03\n ],\n \"formula_id\": \"formula_16\",\n \"formula_text\": \"Var( ti ) ≥ I( ti ) -1 .(14)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 308.86,\n 269.66,\n 236.25,\n 21.95\n ],\n \"formula_id\": \"formula_17\",\n \"formula_text\": \"s i := ∂ log L(t i | S)/∂t i , since E{s i } = 0 [21].\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 314.74,\n 300.08,\n 230.37,\n 24.15\n ],\n \"formula_id\": \"formula_18\",\n \"formula_text\": \"s i = xj ∈Wi d j := S(x j ) -S(x j + t i ) σ 2 ∂S(x j + t i ) ∂t i .(15)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 363.96,\n 344.08,\n 181.16,\n 44.08\n ],\n \"formula_id\": \"formula_19\",\n \"formula_text\": \"E{d ⊤ j d k } = 0, ∀j ̸ = k. Thereby I( ti ) = j∈Wi E{d ⊤ j d j } ti .(16)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 315.97,\n 437.51,\n 229.15,\n 48.12\n ],\n \"formula_id\": \"formula_20\",\n \"formula_text\": \"I( ti ) = xj ∈Wi 1 σ 4 ∂S(x j ) ∂x ⊤ xj ∂S(x j ) ∂x xj E{(S(x j ) -S(x j + t i )) 2 } ti .(17)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 329.72,\n 527.25,\n 215.39,\n 30.01\n ],\n \"formula_id\": \"formula_21\",\n \"formula_text\": \"I( ti ) = 1 σ 2 xj ∈Wi ∂S(x j ) ∂x ⊤ xj ∂S(x j ) ∂x xj ,(18)\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 375,\n 438.19,\n 170.11,\n 10\n ],\n \"formula_id\": \"formula_22\",\n \"formula_text\": \"e i,r := cart(x i -H i,r xr ) ,(19)\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 367.99,\n 452.99,\n 177.12,\n 11.72\n ],\n \"formula_id\": \"formula_23\",\n \"formula_text\": \"Σ ei,r := JΣ xr J ⊤ + Σ xi ,(20)\"\n },\n {\n \"formula_coordinates\": [\n 12,\n 120.73,\n 338.9,\n 165.64,\n 11.28\n ],\n \"formula_id\": \"formula_24\",\n \"formula_text\": \"r i := x(1,2) i - x(3) i x i ,(21)\"\n },\n {\n \"formula_coordinates\": [\n 12,\n 111.41,\n 394.35,\n 174.95,\n 10\n ],\n \"formula_id\": \"formula_25\",\n \"formula_text\": \"xi := K(R c,w p w,i + t c,w ) ,(22)\"\n },\n {\n \"formula_coordinates\": [\n 12,\n 61.86,\n 570.07,\n 220.35,\n 14.58\n ],\n \"formula_id\": \"formula_26\",\n \"formula_text\": \"Mc = 0 , with M := M ⊤ 1 • • • M ⊤ n ⊤ . (23\"\n },\n {\n \"formula_coordinates\": [\n 12,\n 282.21,\n 574.4,\n 4.15,\n 8.64\n ],\n \"formula_id\": \"formula_27\",\n \"formula_text\": \")\"\n },\n {\n \"formula_coordinates\": [\n 12,\n 125.3,\n 639.15,\n 161.06,\n 12.95\n ],\n \"formula_id\": \"formula_28\",\n \"formula_text\": \"M ⊤ Σ -1 r Mc = 0 ,(24)\"\n },\n {\n \"formula_coordinates\": [\n 12,\n 356.08,\n 184.78,\n 189.03,\n 9.68\n ],\n \"formula_id\": \"formula_29\",\n \"formula_text\": \"p w,i = E{p w,i } + ν ,(26)\"\n },\n {\n \"formula_coordinates\": [\n 12,\n 357.74,\n 199.27,\n 187.37,\n 11.72\n ],\n \"formula_id\": \"formula_30\",\n \"formula_text\": \"E{ν} = 0 , Σ pw,i := E{νν ⊤ } .(27)\"\n },\n {\n \"formula_coordinates\": [\n 12,\n 354.42,\n 279.75,\n 190.69,\n 9.79\n ],\n \"formula_id\": \"formula_31\",\n \"formula_text\": \"xi = KRp w,i + Kt = E{x i } + ζ ,(28)\"\n },\n {\n \"formula_coordinates\": [\n 12,\n 342.1,\n 346.72,\n 203.01,\n 20.81\n ],\n \"formula_id\": \"formula_32\",\n \"formula_text\": \"Σ xi := Λ w w ⊤ γ := KRΣ pw,i R ⊤ K ⊤ .(29)\"\n },\n {\n \"formula_coordinates\": [\n 12,\n 326.15,\n 412.72,\n 218.96,\n 11.07\n ],\n \"formula_id\": \"formula_33\",\n \"formula_text\": \"r i = x(1,2) i - x(3) i x i ,(30)\"\n },\n {\n \"formula_coordinates\": [\n 12,\n 356.33,\n 427.7,\n 73.47,\n 29.85\n ],\n \"formula_id\": \"formula_34\",\n \"formula_text\": \"(1,2) i } + ζ (1,2) ) -(E{x(3)\"\n },\n {\n \"formula_coordinates\": [\n 12,\n 387.84,\n 439.23,\n 157.28,\n 20.52\n ],\n \"formula_id\": \"formula_35\",\n \"formula_text\": \"i } + ζ (3) )(E{x i } + ξ) ,(31)\"\n },\n {\n \"formula_coordinates\": [\n 12,\n 336.96,\n 463.66,\n 208.16,\n 30.04\n ],\n \"formula_id\": \"formula_36\",\n \"formula_text\": \"= E{x (1,2) i - x(3) i x i } + ζ (1,2) -E{x (3) i }ξ -ζ (3) E{x i } -ζ (3) ξ ,(32)\"\n },\n {\n \"formula_coordinates\": [\n 12,\n 373.39,\n 527.19,\n 93.84,\n 14.07\n ],\n \"formula_id\": \"formula_37\",\n \"formula_text\": \"E{r i } = E{x (1,2) i - x(3\"\n }\n]"}}},{"rowIdx":384,"cells":{"doi":{"kind":"string","value":""},"pub_date":{"kind":"string","value":"2023-05-20"},"sections":{"kind":"list like","value":[{"figure_ref":["fig_0","fig_0"],"heading":"Introduction","publication_ref":["b0","b1","b2","b3","b4","b5","b6","b7","b8","b9","b13","b14","b15","b5","b6","b7","b8","b16","b17","b18","b19","b20","b21","b22","b23","b24","b1","b11","b25","b15","b26","b27","b28","b29","b12","b4","b29","b30","b28","b31","b32","b27","b33","b34","b35","b13","b14","b15","b36","b37","b38","b39","b40","b41","b7","b5","b8","b42","b28","b43","b43","b12","b3","b44","b26","b30","b45","b5","b46","b47","b20","b19","b18","b20","b24","b48","b49","b50","b22","b23"],"table_ref":[],"text":"Human-Object Interaction Detection (HOI), a core element in human-centric vision perception tasks such as human activity recognition [1], motion tracking [2], and anomaly behavior detection [3], has attracted considerable attention over the past decades. HOI primarily involves localizing correlated human-object 2D positions within an image and identifying their interactions. Despite numerous models proposed in recent years [4][5][6][7][8][9], practical implementation remains challenging due to the inherent ambiguity of verbs and subtle distinctions among interaction categories.\nWe investigate the problems of the current methods from both data and model aspects. From the data perspective, current datasets characterize an HOI instance as a 〈human, object, interaction〉 triplet, and complete with comprehensive annotations, i.e., human and object bounding boxes and respective interaction types. We carefully analyze the widely-used HICO-DET dataset [10], which encompasses 38, 118 training images with 600 distinct types of real-world HOI triplets, and have identified its several limitations that hinder the effective learning of interaction representations: (1) Class imbalance. As demonstrated in Fig. 1-(a), we observe it has a significant long-tail distribution in the HOI triplet types. This imbalanced label distribution is not conducive to learning, particularly when identifying interaction categories with only subtle variances. (2) Small data size. Upon further investigation, we discover that among its 600 interaction categories, 51 categories have only a single image available, and 435 categories have less than 100 images. Such limited data severely impacts the learning of these rare categories. (3) Limited diversity. Compared with existing datasets, real-world scenarios typically exhibit more diversity due to the complex variation in human appearance attributes, environmental backgrounds, and shooting perspectives. This diversity causes a considerable decline in the performance of existing methods when the scene content changes.\nFrom the model standpoint, the majority of existing HOI detectors, regardless of whether they employ two-stage [11-13, 4, 5] or one-stage [14][15][16][6][7][8][9] strategies, are built upon pre-trained object detectors to enhance the initial localization of humans and objects. However, it remains challenging to carry out effective interaction prediction solely based on representations of humans and objects, sometimes with their spatial relationship. Specifically, verb concepts and their nuanced contextual information, including human posture, orientation, attention area and overall circumstances, can critically influence prediction. Extracting this semantic information from data is typically inefficient and cumbersome, significantly limiting existing methods' performance.\nRecently, text-to-image diffusion models [17][18][19][20][21] trained on massive internet-scale data have achieved significant performance in conditioned image generation, which provides high-quality, versatile, and semantically controllable image generation. Such an advantage offers us an effective means to generate rich and diverse realistic HOI images. In practice, these models utilize cross-attention mechanisms between text embeddings and visual representations, signifying a substantial correlation between their feature spaces and semantic concepts in language. Inspired by this, we use DAAM [22] to generate pixel-level linguistic concept heatmaps based on the image features of this diffusion model. As depicted in Fig. 1-(b), beyond the noun concepts highlighted in prior studies [23][24][25], we find that the internal representation space of a frozen text-to-image diffusion model is highly relevant to verb concepts and their associated contexts. However, the challenge of extracting these verb-associated representations from the diffusion model for downstream HOI task still exist.\nTo address the aforementioned issues, we introduce DiffHOI, a novel HOI detection scheme based on text-to-image diffusion models (e.g., Stable Diffusion). For the first time, DiffHOI tries to leverage the generative and representative capabilities to benefit the HOI task, i.e., the extracted powerful verb-associated contextual representations of SD. DiffHOI consists of three components: the pretrained human-object detector, interaction decoder, and object-interaction classifiers. We introduce an adapter-style tuning approach to align global and local semantic associated representations from the SD and CLIP model in the interaction decoder. These representations serve a critical role in comprehending the nuanced disparities in interactions and reducing interaction prediction ambiguity.\nTo fill in the shortcomings of existing long-tail HOI datasets, we present SynHOI, a class-balance, large-scale, and high-diversity synthetic HOI dataset with over 140K fully annotated HOI images, which can effectively facilitate learning interaction representations. To make the flow of dataset production scalable, we present an automatic pipeline, including the HOIPrompt design, automatic labeling and filtering, and quality verification, designed to continually scale up the generation of diverse and high-precision HOI-annotated data. Therefore, our contributions can be summarized as follows: (1) We introduce a novel scheme, DiffHOI, which leverages both the generative and representation capacities of pre-trained text-to-image diffusion models to enhance the performance of HOI detection tasks. (2) We present an automatic and scalable pipeline to generate realistic annotated HOI images characterized by varied attributes and scene contexts and propose a class-balance, large-scale, and high-diversity synthetic HOI dataset, namely SynHOI. (3) Extensive experimental results demonstrate the proposed adapter-style tuning, together with the proposed SynHOI dataset, significantly improves the performance of HOI tasks under the regular and zero-shot settings and achieves the new state-of-the-art, i.e., 41.50 mAP on HICO-DET.\n2 Related Work HOI Detection. HOI detection task primarily encompasses three sub-problems, i.e., object detection, human-object pairing, and interaction recognition. Previous HOI detectors can generally be divided into one-stage and two-stage paradigms. The two-stage strategy employs an off-the-shelf detector to determine the locations and classes of objects, followed by specially-designed modules for humanobject association and interaction recognition. Most methods are dedicated to exploring additional feature streams to improve interaction classification, such as the appearance stream [12,26,16,27], spatial stream [28][29][30], pose and body-parts stream [13,5,30], semantic stream [31,29,32], and graph network [33,28,[34][35][36]. Instead, the one-stage strategy detects HOI triplets in a single forward pass by assigning human and object proposals to predefined anchors and then estimating potential interactions [14][15][16]37]. Recently, the DETR-based HOI detectors [38][39][40][41][42] have gained prominence in this paradigm, and they formulate the HOI detection task as a set prediction problem, avoiding complex post-processing. In particular, many methods [8,6,9,43] demonstrate promising performance improvements by disentangling human-object detection and interaction classification as two decoders in a cascade manner. Our work builds on the top of the transformer-based HOI detection strategy and focuses on enhancing the learning of a dedicated interaction decoder.\nZero-shot HOI Detection. Zero-shot HOI detection has emerged as a field aiming to identify unseen HOI triplet categories not present in the training data. Previous research [29,44,44,13,4,45,27,31] has addressed this task in a compositional manner, by disentangling the reasoning process on actions and objects during training. This approach enables the recognition of unseen HOI triplets during inference. With the advancements in Vision-Language Models, such as CLIP [46], recent research [6,47,48] has shifted focus toward transferring knowledge from CLIP to recognize unseen HOI concepts. This shift has resulted in a notable performance improvement in zero-shot settings. In this work, we aim to further explore the potential benefits of the Text-to-Image Diffusion model [21] in enhancing zero-shot HOI detection.\nDiffusion Models for Image Generation and Vision Perception. Text-to-image diffusion models, such as DALLE-2 [20], Imagen [19], and Stable Diffusion [21], have shown considerable potentials in generating photorealistic images from free-form text prompts. This capability is attributed to the strong semantic correspondence between visual and language elements learned from a vast corpus of image-caption pairs. Recent research has proposed utilizing diffusion models to augment real datasets, assisting the training of downstream tasks [25,[49][50][51]. Our work concentrates on generating synthetic HOI data to enhance HOI detection performance. Moreover, text-to-image diffusion models pretrained on large-scale image-text pairs offer a high degree of control through customizable prompts. This aspect suits them for downstream tasks [23,24], where the additional feature representations from pretraining can bolster performance. This work first explores using a frozen text-image diffusion model for HOI detection."},{"figure_ref":[],"heading":"SynHOI-A Synthetic HOI Dataset","publication_ref":[],"table_ref":[],"text":"This section introduces how the proposed high-quality synthetic HOI dataset SynHOI is built via an automatic and scalable pipeline. \"urban\" → \"outdoor\" \"partial view\" → \"front view\" Generated annotations Generated annotations \"white woman\" → \"asian man\" \"outdoor\" → \"secluded\" \"oblique view\" → \"back view\" "},{"figure_ref":["fig_5","fig_5","fig_5"],"heading":"Construction Process","publication_ref":["b9","b51","b52","b53"],"table_ref":[],"text":"The HOIPrompt design. To address limitations in current datasets , we propose pre-defined HOIPrompts, as illustrated in Fig. 2 (a). Complete HOI triplets are formed by sampling verb and noun combinations from the HICO-DET dataset [10], creating \"a {race} {age & gender} verb-ing an object.\" To describe a person's appearance, we use the format \"a {race} {age & gender},\" randomly selecting elements from the HOIPrompts. We generate images with combinations likely to occur together by analyzing co-occurring HOI triplets. For interaction environments, we select an adjective from a range of options to describe the atmosphere (denoted as \"{environment}\"). Photographic information is represented by four components (\"{quality},\" \"{lighting},\" \"{view},\" and \"{camera}\"), aligning the synthetic images with real HOI data and providing camera angle diversity. We further enhance diversity and quality through negative prompts and random model configurations. Overall, we generate 259, 806 images using HOIPrompts at this stage.\nAutomatic Labeling and Filtering. We design a three-step process to automatically annotate and filter the synthetic images. Firstly, we utilize a state-of-the-art detection model [52] to detect objects within each image. Secondly, we discard any images in which the confidence score of the detected object specified in the corresponding HOI triplet(s) prompt is below a threshold of 0.5. Thirdly, we associate humans with objects in the images and assign the appropriate HOI category from the prompts to the human-object combination(s). In practice, we assign the HOI category to the person closest to the center of the object's bounding box. If not all humans in the image have an HOI label, we select the object closest to the corresponding human's bounding box. Upon completing the automated labeling and filtering process, we obtain a new synthetic dataset, namely SynHOI, including 146, 772 annotated images with complete HOI labels associated with the detected human-object interactions.\nVisualization and Manual Verification. We develop a visualization tool to facilitate the manual inspection and filtering of any incorrect HOI annotations. We incorporate manual efforts to sample and inspect the annotated results to ensure their quality. During this inspection, we observe that the SOTA detector trained on COCO [53] performs well in detecting humans and objects, indicating that the data distribution in SynHOI closely resembles that of natural images in COCO. However, due to the inherent ambiguity of verbs, a small number of synthetic interactions in SynHOI may be incompletely accurate, which also exists in the HICO-DET dataset. To address this issue, we construct a subset comprising 5% data of SynHOI (over 8K), namely SynHOI-Sub, during the sampling and inspection process. This subset has undergone meticulous manual examination, resulting in annotations that are verified to be completely accurate. High-diversity. SynHOI exhibits a high level of diversity, offering a wide range of visually distinct images. Fig. 2-(b) demonstrates the impact of random variations in people's descriptions, environments, and photographic information within the HOIPrompts on the diversity of synthetic images.\nHigh-quality. SynHOI showcases high-quality HOI annotations. First, we employ CLIPScore [54] to measure the similarity between the synthetic images and the corresponding HOI triplet prompts. The SynHOI dataset achieves a high CLIPScore of 0.805, indicating a faithful reflection of the HOI triplet information in the synthetic images. Second, Fig. 2-(b) provides evidence of the high quality of detection annotations in SynHOI, attributed to the effectiveness of the SOTA detector and the alignment of SynHOI with real-world data distributions. As mentioned earlier, we also release a carefully verified subset of SynHOI, called SynHOI-Sub."},{"figure_ref":[],"heading":"Methodology","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":["fig_4"],"heading":"Framework Overview","publication_ref":["b54","b20","b45"],"table_ref":[],"text":"The proposed DiffHOI framework is presented in Figure 3, consisting of three components. Primarily, we employ a transformer-based object detector D obj [55] to extract local feature representations of humans and objects. Using these extracted features, initial representations of interactions are computed and fed into the human-object interaction decoder, referred to as D int , which subsequently updates these interaction representations. Additionally, we employ the frozen Stable Diffusion (SD) Model F sd [21] to generate semantically associated image feature maps, ensuring that distinct regions of these feature maps respond to various types of semantic information (see Sec.4.2). Furthermore, we introduce the CLIP model [46] and utilize its frozen image encoder F img clip to extract comprehensive semantic representations that align with text descriptors. Two trainable scene-aware adaptors are introduced to adapt these semantic representations to suit the SD Model and interaction classifier (see Sec. 4.3). In practice, both the object classifier and interaction classifier are generated using the CLIP text encoder F text clip , which is applied during both the training and inference phases. Pair-wise Human-Object Localization. Given an input image x, we first utilize the image encoder to extract feature maps V , which is then adopted to localize humans and objects by using humanobject decoder D obj . We further input pair-wise human-object queries Q {h,o} ∈ R 2×N ×C into the D obj and update them to Q {h,o} ∈ R 2×N ×C , where N is the number of paired queries and C indicates the channel dimension. We use Q h,o to predict the human bounding box B h and object bounding box B o , whose dimension are N × 4. To predict the class label of each object, we further employ the text encoder F text clip of CLIP to extract the object-aware embedding T o clip ∈ R K1×C . We further leverage T o clip to conduct the dot-product with Q o ∈ R N ×C to predict the final object category distributions P o ∈ R N ×K1 , where K 1 denotes the total number of object classes.\nHuman-Object Interaction Recognition. We design a novel interaction decoder D int to perform better interaction understanding, where we exploit three types of visual representations as the input, i.e., the V from object detector, the V sd from Stable Diffusion Model F sd , and the v clip from the image encoder F img clip of CLIP. Specifically, we first perform average pooling on the human-object queries Q h,o , resulting in interaction queries Qi ∈ R N ×C . Then, we feed Q i into the interaction decoder D int to apply self-attention and subsequent cross-attention with the sum of V and V sd . This process updates the interaction queries from Q i to Q i ∈ R N ×C . Moreover, we leverage v clip ∈ R C to enhance each of the interaction query in Q i , performing element-wise addition. To predict the interaction label of each human-object pair, we utilize the text encoder F text clip of CLIP to extract the interaction-aware embedding T i clip ∈ R K2×C . Finally, we employ dot-product between T i clip and Q i to predict the human-object interaction category distributions P i ∈ R N ×K2 , where K 2 is the number of interaction categories."},{"figure_ref":["fig_0","fig_4","fig_4"],"heading":"Local Semantic Association via Stable Diffusion","publication_ref":["b20","b5","b46","b47"],"table_ref":[],"text":"The output features V from the object detector only retain human and object-oriented semantic information. In such a case, the interactional relationship and its corresponding context remain limited, resulting in sub-optimal interaction predictions. Recenlty, Stable Diffusion [21] (SD) has been primarily designed to produce high-quality images utilizing textual descriptors. For the HOI task, we find that it be used to encourage diverse semantic information embedded in the textual representations to be associated with specific local regions of the images, as shown in Fig. 1. Inspired by the above observations, the feature maps derived from the SD model are expected to exhibit superior semantic correspondence within local regions, establishing an association between the semantic information of the text and the corresponding regions within the image. Accordingly, we utilize the UNet-based diffusion model to extract the feature maps, which are not only associated with the noun concepts, e.g., the human and objects, but also include the verb concepts and the corresponding contextual details. Different from using a series of denoising steps to generate an image, we directly feed the input image x into the UNet (denoted as F sd ) and perform a single forward pass via the network. The output multi-scale image features are as V sd = F sd (x, A x ), where A x denotes the text-related representation that corresponds to x. Typically, A x can be obtained by utilizing a text encoder, such as the CLIP text encoder, to encode the description of the image. However, as a discriminative task, the description of the HOI image is not available in the inference. To address this, we replace it with the global feature representation of x extracted from the CLIP image encoder F img clip , which bears significant relevance to the textual representation. Please refer to Sec. 4.3 for more details. As illustrated in the blocks in the red background of Fig. 3, we proceed to combine the features V and V sd to augment the representation ability of the verb concepts within V and use it to generate more informative Q i for the final interaction prediction. Two Scene-aware Adaptors. Given that v clip represents a global feature, it inherently encompasses the contextual information of the overall scene. As shown in Fig. 3, we introduce two scene-aware adaptors, denoted as α and β, to project v clip into feature spaces more consistent with the SD model and interaction predictor. Regarding the scene-aware adaptor α, since the CLIP model is trained to align global visual and textual representations within a latent space, the v clip can be employed as a substitute for the textual representation. Hence, we can train an MLP to directly project v clip to a textual space suitable for the SD model. As for the scene-aware adaptor β, we project v clip through it and incorporate the adapted version into each interaction query of Q i . This adjustment allows for the tuning of these interaction queries to align more effectively with the CLIP-based classifiers T i clip . CLIP-based Classifiers. Inspired by [6,47,48], we use the CLIP text encoder F text clip to generate both the object classifier and interaction classifier. Specifically, we initiate the process by transforming each object category or HOI category into a sentence using the hand-crafted template, i.e. \"A photo of a [Object]\" or \"A photo of a person [Verb-ing] a [Object]\". Then these sentences can be encoded to obtain the object category embeddings T o clip or interaction category embeddings T i clip . Formally, the final object category distributions P o and HOI category distributions P i can be calculated as,\nP o = softmax(Q o * T o clip )(1)\nP i = softmax(Q i * T i clip )(2)\nwhere\nQ o ∈ R N ×C and Q o ∈ R N ×C\ndenotes object queries and interaction queries, respectively, and softmax indicates the row-wise softmax operation."},{"figure_ref":[],"heading":"Loss Function","publication_ref":["b5","b7"],"table_ref":[],"text":"Following the query-based methods [6,8], we employ the Hungarian algorithm to match predictions to each ground-truth. The overall loss is computed between the matched predictions and their corresponding ground-truths, which includes the box regression loss L b , the intersection-over-union loss L g , the object class loss L o c , and the interaction category loss L i c ,\nL = λ b L b + λ g L g + λ o c L o c + λ i c L i c ,(3)\nwhere L b and L g contain both human and object localization. λ b , λ g , λ o c and λ i c are used to adjust the weights of each loss component."},{"figure_ref":[],"heading":"Experiments","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Experimental Settings","publication_ref":["b10","b55","b7","b5","b5","b54","b56","b51","b57","b5","b65","b22","b20"],"table_ref":[],"text":"Due to the page limit, we leave the detailed experimental settings in the Appendix.\nDatasets and Evaluation Metrics. We evaluate our models on two widely used datasets: HICO-DET [11] and V-COCO [56]. The mean Average Precision (mAP) is used as the evaluation metric, following standard protocols [8,6].\nZero-Shot Setting. We conduct zero-shot experiments on HICO-Det, following the settings in [6]: Rare First UC (RF-UC), Non-rare First UC (NF-UC), Unseen Object (UO) and Unseen Verb (UV).\nImplementation Details. We implement two variant architectures of DiffHOI: DiffHOI-S, and DiffHOI-L, where 'S' and 'L' refer to small and large, respectively. For DiffHOI-S, we use ResNet-50 as the backbone and a six-layer vanilla Transformer encoder [55] as the feature extractor. Both the human-object decoder and interaction decoder are three-layer vanilla Transformer decoders. For DiffHOI-L, we employ Swin-L [57] as the backbone. In this variant, we replace all the transformer layers with deformable Transformer layers [52,58]. We fine-tune the CLIP-based interaction and object classifiers for regular settings with a small learning rate of 10 -5 while freezing them for zero-shot settings. The diffusion model is frozen during training for all the settings. Two Key Representations. As described in Sec. 4.1, we introduce V sd derived from stable diffusion F sd and v clip obtained from the image encoder F img clip of CLIP, into the interaction decoder. Tab. 6 demonstrates their effectiveness. Initially, we establish a baseline model following the design of GEN-VLKT [6] while excluding the knowledge distillation component. Subsequently, the incorporation of V sd leads to significant improvements, particularly in terms of rare mAP. This suggests that the internal features in stable diffusion play a crucial role in enhancing the representation of rare categories in the HICO-DET dataset. Additionally, the integration of v clip further aids in improving the non-rare AP, as it aligns the output queries with the interaction classifier. Diffusion Time Steps. We investigate the effectiveness of different diffusion steps in extracting interaction-aware features V sd , similarly to [66,23]. Diffusion models control the noise distortion added to the input image by varying the value of t. Stable diffusion [21] uses a total of 1000 time steps. We set t values to 0, 100, 500 for ablation studies. As demonstrated in Tab. 7, the best performance is achieved when t = 0. It is worth noting that using V sd extracted from input images with higher noise levels would decrease performance and potentially impact the learning of interactions."},{"figure_ref":[],"heading":"Conclusion","publication_ref":[],"table_ref":[],"text":"We investigate the issues of the current HOI detection methods and present a novel scheme, namely DiffHOI, which leverages both the generative and representation capacities of a pre-trained text-toimage diffusion model to benefit the HOI detection task. Particularly, we release a class-balance, large-scale, and high-diversity synthetic HOI dataset called SynHOI to address the long-tail issue in previous datasets and develop an automatic and scalable pipeline to scale up the generation of diverse and high-precision HOI-annotated data. Extensive experimental results demonstrate that our method significantly outperforms the prior state-of-the-art in regular and zero-shot detection tasks. We hope this work could inspire further research in related fields."}],"string":"[\n {\n \"figure_ref\": [\n \"fig_0\",\n \"fig_0\"\n ],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b0\",\n \"b1\",\n \"b2\",\n \"b3\",\n \"b4\",\n \"b5\",\n \"b6\",\n \"b7\",\n \"b8\",\n \"b9\",\n \"b13\",\n \"b14\",\n \"b15\",\n \"b5\",\n \"b6\",\n \"b7\",\n \"b8\",\n \"b16\",\n \"b17\",\n \"b18\",\n \"b19\",\n \"b20\",\n \"b21\",\n \"b22\",\n \"b23\",\n \"b24\",\n \"b1\",\n \"b11\",\n \"b25\",\n \"b15\",\n \"b26\",\n \"b27\",\n \"b28\",\n \"b29\",\n \"b12\",\n \"b4\",\n \"b29\",\n \"b30\",\n \"b28\",\n \"b31\",\n \"b32\",\n \"b27\",\n \"b33\",\n \"b34\",\n \"b35\",\n \"b13\",\n \"b14\",\n \"b15\",\n \"b36\",\n \"b37\",\n \"b38\",\n \"b39\",\n \"b40\",\n \"b41\",\n \"b7\",\n \"b5\",\n \"b8\",\n \"b42\",\n \"b28\",\n \"b43\",\n \"b43\",\n \"b12\",\n \"b3\",\n \"b44\",\n \"b26\",\n \"b30\",\n \"b45\",\n \"b5\",\n \"b46\",\n \"b47\",\n \"b20\",\n \"b19\",\n \"b18\",\n \"b20\",\n \"b24\",\n \"b48\",\n \"b49\",\n \"b50\",\n \"b22\",\n \"b23\"\n ],\n \"table_ref\": [],\n \"text\": \"Human-Object Interaction Detection (HOI), a core element in human-centric vision perception tasks such as human activity recognition [1], motion tracking [2], and anomaly behavior detection [3], has attracted considerable attention over the past decades. HOI primarily involves localizing correlated human-object 2D positions within an image and identifying their interactions. Despite numerous models proposed in recent years [4][5][6][7][8][9], practical implementation remains challenging due to the inherent ambiguity of verbs and subtle distinctions among interaction categories.\\nWe investigate the problems of the current methods from both data and model aspects. From the data perspective, current datasets characterize an HOI instance as a 〈human, object, interaction〉 triplet, and complete with comprehensive annotations, i.e., human and object bounding boxes and respective interaction types. We carefully analyze the widely-used HICO-DET dataset [10], which encompasses 38, 118 training images with 600 distinct types of real-world HOI triplets, and have identified its several limitations that hinder the effective learning of interaction representations: (1) Class imbalance. As demonstrated in Fig. 1-(a), we observe it has a significant long-tail distribution in the HOI triplet types. This imbalanced label distribution is not conducive to learning, particularly when identifying interaction categories with only subtle variances. (2) Small data size. Upon further investigation, we discover that among its 600 interaction categories, 51 categories have only a single image available, and 435 categories have less than 100 images. Such limited data severely impacts the learning of these rare categories. (3) Limited diversity. Compared with existing datasets, real-world scenarios typically exhibit more diversity due to the complex variation in human appearance attributes, environmental backgrounds, and shooting perspectives. This diversity causes a considerable decline in the performance of existing methods when the scene content changes.\\nFrom the model standpoint, the majority of existing HOI detectors, regardless of whether they employ two-stage [11-13, 4, 5] or one-stage [14][15][16][6][7][8][9] strategies, are built upon pre-trained object detectors to enhance the initial localization of humans and objects. However, it remains challenging to carry out effective interaction prediction solely based on representations of humans and objects, sometimes with their spatial relationship. Specifically, verb concepts and their nuanced contextual information, including human posture, orientation, attention area and overall circumstances, can critically influence prediction. Extracting this semantic information from data is typically inefficient and cumbersome, significantly limiting existing methods' performance.\\nRecently, text-to-image diffusion models [17][18][19][20][21] trained on massive internet-scale data have achieved significant performance in conditioned image generation, which provides high-quality, versatile, and semantically controllable image generation. Such an advantage offers us an effective means to generate rich and diverse realistic HOI images. In practice, these models utilize cross-attention mechanisms between text embeddings and visual representations, signifying a substantial correlation between their feature spaces and semantic concepts in language. Inspired by this, we use DAAM [22] to generate pixel-level linguistic concept heatmaps based on the image features of this diffusion model. As depicted in Fig. 1-(b), beyond the noun concepts highlighted in prior studies [23][24][25], we find that the internal representation space of a frozen text-to-image diffusion model is highly relevant to verb concepts and their associated contexts. However, the challenge of extracting these verb-associated representations from the diffusion model for downstream HOI task still exist.\\nTo address the aforementioned issues, we introduce DiffHOI, a novel HOI detection scheme based on text-to-image diffusion models (e.g., Stable Diffusion). For the first time, DiffHOI tries to leverage the generative and representative capabilities to benefit the HOI task, i.e., the extracted powerful verb-associated contextual representations of SD. DiffHOI consists of three components: the pretrained human-object detector, interaction decoder, and object-interaction classifiers. We introduce an adapter-style tuning approach to align global and local semantic associated representations from the SD and CLIP model in the interaction decoder. These representations serve a critical role in comprehending the nuanced disparities in interactions and reducing interaction prediction ambiguity.\\nTo fill in the shortcomings of existing long-tail HOI datasets, we present SynHOI, a class-balance, large-scale, and high-diversity synthetic HOI dataset with over 140K fully annotated HOI images, which can effectively facilitate learning interaction representations. To make the flow of dataset production scalable, we present an automatic pipeline, including the HOIPrompt design, automatic labeling and filtering, and quality verification, designed to continually scale up the generation of diverse and high-precision HOI-annotated data. Therefore, our contributions can be summarized as follows: (1) We introduce a novel scheme, DiffHOI, which leverages both the generative and representation capacities of pre-trained text-to-image diffusion models to enhance the performance of HOI detection tasks. (2) We present an automatic and scalable pipeline to generate realistic annotated HOI images characterized by varied attributes and scene contexts and propose a class-balance, large-scale, and high-diversity synthetic HOI dataset, namely SynHOI. (3) Extensive experimental results demonstrate the proposed adapter-style tuning, together with the proposed SynHOI dataset, significantly improves the performance of HOI tasks under the regular and zero-shot settings and achieves the new state-of-the-art, i.e., 41.50 mAP on HICO-DET.\\n2 Related Work HOI Detection. HOI detection task primarily encompasses three sub-problems, i.e., object detection, human-object pairing, and interaction recognition. Previous HOI detectors can generally be divided into one-stage and two-stage paradigms. The two-stage strategy employs an off-the-shelf detector to determine the locations and classes of objects, followed by specially-designed modules for humanobject association and interaction recognition. Most methods are dedicated to exploring additional feature streams to improve interaction classification, such as the appearance stream [12,26,16,27], spatial stream [28][29][30], pose and body-parts stream [13,5,30], semantic stream [31,29,32], and graph network [33,28,[34][35][36]. Instead, the one-stage strategy detects HOI triplets in a single forward pass by assigning human and object proposals to predefined anchors and then estimating potential interactions [14][15][16]37]. Recently, the DETR-based HOI detectors [38][39][40][41][42] have gained prominence in this paradigm, and they formulate the HOI detection task as a set prediction problem, avoiding complex post-processing. In particular, many methods [8,6,9,43] demonstrate promising performance improvements by disentangling human-object detection and interaction classification as two decoders in a cascade manner. Our work builds on the top of the transformer-based HOI detection strategy and focuses on enhancing the learning of a dedicated interaction decoder.\\nZero-shot HOI Detection. Zero-shot HOI detection has emerged as a field aiming to identify unseen HOI triplet categories not present in the training data. Previous research [29,44,44,13,4,45,27,31] has addressed this task in a compositional manner, by disentangling the reasoning process on actions and objects during training. This approach enables the recognition of unseen HOI triplets during inference. With the advancements in Vision-Language Models, such as CLIP [46], recent research [6,47,48] has shifted focus toward transferring knowledge from CLIP to recognize unseen HOI concepts. This shift has resulted in a notable performance improvement in zero-shot settings. In this work, we aim to further explore the potential benefits of the Text-to-Image Diffusion model [21] in enhancing zero-shot HOI detection.\\nDiffusion Models for Image Generation and Vision Perception. Text-to-image diffusion models, such as DALLE-2 [20], Imagen [19], and Stable Diffusion [21], have shown considerable potentials in generating photorealistic images from free-form text prompts. This capability is attributed to the strong semantic correspondence between visual and language elements learned from a vast corpus of image-caption pairs. Recent research has proposed utilizing diffusion models to augment real datasets, assisting the training of downstream tasks [25,[49][50][51]. Our work concentrates on generating synthetic HOI data to enhance HOI detection performance. Moreover, text-to-image diffusion models pretrained on large-scale image-text pairs offer a high degree of control through customizable prompts. This aspect suits them for downstream tasks [23,24], where the additional feature representations from pretraining can bolster performance. This work first explores using a frozen text-image diffusion model for HOI detection.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"SynHOI-A Synthetic HOI Dataset\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This section introduces how the proposed high-quality synthetic HOI dataset SynHOI is built via an automatic and scalable pipeline. \\\"urban\\\" → \\\"outdoor\\\" \\\"partial view\\\" → \\\"front view\\\" Generated annotations Generated annotations \\\"white woman\\\" → \\\"asian man\\\" \\\"outdoor\\\" → \\\"secluded\\\" \\\"oblique view\\\" → \\\"back view\\\" \"\n },\n {\n \"figure_ref\": [\n \"fig_5\",\n \"fig_5\",\n \"fig_5\"\n ],\n \"heading\": \"Construction Process\",\n \"publication_ref\": [\n \"b9\",\n \"b51\",\n \"b52\",\n \"b53\"\n ],\n \"table_ref\": [],\n \"text\": \"The HOIPrompt design. To address limitations in current datasets , we propose pre-defined HOIPrompts, as illustrated in Fig. 2 (a). Complete HOI triplets are formed by sampling verb and noun combinations from the HICO-DET dataset [10], creating \\\"a {race} {age & gender} verb-ing an object.\\\" To describe a person's appearance, we use the format \\\"a {race} {age & gender},\\\" randomly selecting elements from the HOIPrompts. We generate images with combinations likely to occur together by analyzing co-occurring HOI triplets. For interaction environments, we select an adjective from a range of options to describe the atmosphere (denoted as \\\"{environment}\\\"). Photographic information is represented by four components (\\\"{quality},\\\" \\\"{lighting},\\\" \\\"{view},\\\" and \\\"{camera}\\\"), aligning the synthetic images with real HOI data and providing camera angle diversity. We further enhance diversity and quality through negative prompts and random model configurations. Overall, we generate 259, 806 images using HOIPrompts at this stage.\\nAutomatic Labeling and Filtering. We design a three-step process to automatically annotate and filter the synthetic images. Firstly, we utilize a state-of-the-art detection model [52] to detect objects within each image. Secondly, we discard any images in which the confidence score of the detected object specified in the corresponding HOI triplet(s) prompt is below a threshold of 0.5. Thirdly, we associate humans with objects in the images and assign the appropriate HOI category from the prompts to the human-object combination(s). In practice, we assign the HOI category to the person closest to the center of the object's bounding box. If not all humans in the image have an HOI label, we select the object closest to the corresponding human's bounding box. Upon completing the automated labeling and filtering process, we obtain a new synthetic dataset, namely SynHOI, including 146, 772 annotated images with complete HOI labels associated with the detected human-object interactions.\\nVisualization and Manual Verification. We develop a visualization tool to facilitate the manual inspection and filtering of any incorrect HOI annotations. We incorporate manual efforts to sample and inspect the annotated results to ensure their quality. During this inspection, we observe that the SOTA detector trained on COCO [53] performs well in detecting humans and objects, indicating that the data distribution in SynHOI closely resembles that of natural images in COCO. However, due to the inherent ambiguity of verbs, a small number of synthetic interactions in SynHOI may be incompletely accurate, which also exists in the HICO-DET dataset. To address this issue, we construct a subset comprising 5% data of SynHOI (over 8K), namely SynHOI-Sub, during the sampling and inspection process. This subset has undergone meticulous manual examination, resulting in annotations that are verified to be completely accurate. High-diversity. SynHOI exhibits a high level of diversity, offering a wide range of visually distinct images. Fig. 2-(b) demonstrates the impact of random variations in people's descriptions, environments, and photographic information within the HOIPrompts on the diversity of synthetic images.\\nHigh-quality. SynHOI showcases high-quality HOI annotations. First, we employ CLIPScore [54] to measure the similarity between the synthetic images and the corresponding HOI triplet prompts. The SynHOI dataset achieves a high CLIPScore of 0.805, indicating a faithful reflection of the HOI triplet information in the synthetic images. Second, Fig. 2-(b) provides evidence of the high quality of detection annotations in SynHOI, attributed to the effectiveness of the SOTA detector and the alignment of SynHOI with real-world data distributions. As mentioned earlier, we also release a carefully verified subset of SynHOI, called SynHOI-Sub.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Methodology\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [\n \"fig_4\"\n ],\n \"heading\": \"Framework Overview\",\n \"publication_ref\": [\n \"b54\",\n \"b20\",\n \"b45\"\n ],\n \"table_ref\": [],\n \"text\": \"The proposed DiffHOI framework is presented in Figure 3, consisting of three components. Primarily, we employ a transformer-based object detector D obj [55] to extract local feature representations of humans and objects. Using these extracted features, initial representations of interactions are computed and fed into the human-object interaction decoder, referred to as D int , which subsequently updates these interaction representations. Additionally, we employ the frozen Stable Diffusion (SD) Model F sd [21] to generate semantically associated image feature maps, ensuring that distinct regions of these feature maps respond to various types of semantic information (see Sec.4.2). Furthermore, we introduce the CLIP model [46] and utilize its frozen image encoder F img clip to extract comprehensive semantic representations that align with text descriptors. Two trainable scene-aware adaptors are introduced to adapt these semantic representations to suit the SD Model and interaction classifier (see Sec. 4.3). In practice, both the object classifier and interaction classifier are generated using the CLIP text encoder F text clip , which is applied during both the training and inference phases. Pair-wise Human-Object Localization. Given an input image x, we first utilize the image encoder to extract feature maps V , which is then adopted to localize humans and objects by using humanobject decoder D obj . We further input pair-wise human-object queries Q {h,o} ∈ R 2×N ×C into the D obj and update them to Q {h,o} ∈ R 2×N ×C , where N is the number of paired queries and C indicates the channel dimension. We use Q h,o to predict the human bounding box B h and object bounding box B o , whose dimension are N × 4. To predict the class label of each object, we further employ the text encoder F text clip of CLIP to extract the object-aware embedding T o clip ∈ R K1×C . We further leverage T o clip to conduct the dot-product with Q o ∈ R N ×C to predict the final object category distributions P o ∈ R N ×K1 , where K 1 denotes the total number of object classes.\\nHuman-Object Interaction Recognition. We design a novel interaction decoder D int to perform better interaction understanding, where we exploit three types of visual representations as the input, i.e., the V from object detector, the V sd from Stable Diffusion Model F sd , and the v clip from the image encoder F img clip of CLIP. Specifically, we first perform average pooling on the human-object queries Q h,o , resulting in interaction queries Qi ∈ R N ×C . Then, we feed Q i into the interaction decoder D int to apply self-attention and subsequent cross-attention with the sum of V and V sd . This process updates the interaction queries from Q i to Q i ∈ R N ×C . Moreover, we leverage v clip ∈ R C to enhance each of the interaction query in Q i , performing element-wise addition. To predict the interaction label of each human-object pair, we utilize the text encoder F text clip of CLIP to extract the interaction-aware embedding T i clip ∈ R K2×C . Finally, we employ dot-product between T i clip and Q i to predict the human-object interaction category distributions P i ∈ R N ×K2 , where K 2 is the number of interaction categories.\"\n },\n {\n \"figure_ref\": [\n \"fig_0\",\n \"fig_4\",\n \"fig_4\"\n ],\n \"heading\": \"Local Semantic Association via Stable Diffusion\",\n \"publication_ref\": [\n \"b20\",\n \"b5\",\n \"b46\",\n \"b47\"\n ],\n \"table_ref\": [],\n \"text\": \"The output features V from the object detector only retain human and object-oriented semantic information. In such a case, the interactional relationship and its corresponding context remain limited, resulting in sub-optimal interaction predictions. Recenlty, Stable Diffusion [21] (SD) has been primarily designed to produce high-quality images utilizing textual descriptors. For the HOI task, we find that it be used to encourage diverse semantic information embedded in the textual representations to be associated with specific local regions of the images, as shown in Fig. 1. Inspired by the above observations, the feature maps derived from the SD model are expected to exhibit superior semantic correspondence within local regions, establishing an association between the semantic information of the text and the corresponding regions within the image. Accordingly, we utilize the UNet-based diffusion model to extract the feature maps, which are not only associated with the noun concepts, e.g., the human and objects, but also include the verb concepts and the corresponding contextual details. Different from using a series of denoising steps to generate an image, we directly feed the input image x into the UNet (denoted as F sd ) and perform a single forward pass via the network. The output multi-scale image features are as V sd = F sd (x, A x ), where A x denotes the text-related representation that corresponds to x. Typically, A x can be obtained by utilizing a text encoder, such as the CLIP text encoder, to encode the description of the image. However, as a discriminative task, the description of the HOI image is not available in the inference. To address this, we replace it with the global feature representation of x extracted from the CLIP image encoder F img clip , which bears significant relevance to the textual representation. Please refer to Sec. 4.3 for more details. As illustrated in the blocks in the red background of Fig. 3, we proceed to combine the features V and V sd to augment the representation ability of the verb concepts within V and use it to generate more informative Q i for the final interaction prediction. Two Scene-aware Adaptors. Given that v clip represents a global feature, it inherently encompasses the contextual information of the overall scene. As shown in Fig. 3, we introduce two scene-aware adaptors, denoted as α and β, to project v clip into feature spaces more consistent with the SD model and interaction predictor. Regarding the scene-aware adaptor α, since the CLIP model is trained to align global visual and textual representations within a latent space, the v clip can be employed as a substitute for the textual representation. Hence, we can train an MLP to directly project v clip to a textual space suitable for the SD model. As for the scene-aware adaptor β, we project v clip through it and incorporate the adapted version into each interaction query of Q i . This adjustment allows for the tuning of these interaction queries to align more effectively with the CLIP-based classifiers T i clip . CLIP-based Classifiers. Inspired by [6,47,48], we use the CLIP text encoder F text clip to generate both the object classifier and interaction classifier. Specifically, we initiate the process by transforming each object category or HOI category into a sentence using the hand-crafted template, i.e. \\\"A photo of a [Object]\\\" or \\\"A photo of a person [Verb-ing] a [Object]\\\". Then these sentences can be encoded to obtain the object category embeddings T o clip or interaction category embeddings T i clip . Formally, the final object category distributions P o and HOI category distributions P i can be calculated as,\\nP o = softmax(Q o * T o clip )(1)\\nP i = softmax(Q i * T i clip )(2)\\nwhere\\nQ o ∈ R N ×C and Q o ∈ R N ×C\\ndenotes object queries and interaction queries, respectively, and softmax indicates the row-wise softmax operation.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Loss Function\",\n \"publication_ref\": [\n \"b5\",\n \"b7\"\n ],\n \"table_ref\": [],\n \"text\": \"Following the query-based methods [6,8], we employ the Hungarian algorithm to match predictions to each ground-truth. The overall loss is computed between the matched predictions and their corresponding ground-truths, which includes the box regression loss L b , the intersection-over-union loss L g , the object class loss L o c , and the interaction category loss L i c ,\\nL = λ b L b + λ g L g + λ o c L o c + λ i c L i c ,(3)\\nwhere L b and L g contain both human and object localization. λ b , λ g , λ o c and λ i c are used to adjust the weights of each loss component.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiments\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experimental Settings\",\n \"publication_ref\": [\n \"b10\",\n \"b55\",\n \"b7\",\n \"b5\",\n \"b5\",\n \"b54\",\n \"b56\",\n \"b51\",\n \"b57\",\n \"b5\",\n \"b65\",\n \"b22\",\n \"b20\"\n ],\n \"table_ref\": [],\n \"text\": \"Due to the page limit, we leave the detailed experimental settings in the Appendix.\\nDatasets and Evaluation Metrics. We evaluate our models on two widely used datasets: HICO-DET [11] and V-COCO [56]. The mean Average Precision (mAP) is used as the evaluation metric, following standard protocols [8,6].\\nZero-Shot Setting. We conduct zero-shot experiments on HICO-Det, following the settings in [6]: Rare First UC (RF-UC), Non-rare First UC (NF-UC), Unseen Object (UO) and Unseen Verb (UV).\\nImplementation Details. We implement two variant architectures of DiffHOI: DiffHOI-S, and DiffHOI-L, where 'S' and 'L' refer to small and large, respectively. For DiffHOI-S, we use ResNet-50 as the backbone and a six-layer vanilla Transformer encoder [55] as the feature extractor. Both the human-object decoder and interaction decoder are three-layer vanilla Transformer decoders. For DiffHOI-L, we employ Swin-L [57] as the backbone. In this variant, we replace all the transformer layers with deformable Transformer layers [52,58]. We fine-tune the CLIP-based interaction and object classifiers for regular settings with a small learning rate of 10 -5 while freezing them for zero-shot settings. The diffusion model is frozen during training for all the settings. Two Key Representations. As described in Sec. 4.1, we introduce V sd derived from stable diffusion F sd and v clip obtained from the image encoder F img clip of CLIP, into the interaction decoder. Tab. 6 demonstrates their effectiveness. Initially, we establish a baseline model following the design of GEN-VLKT [6] while excluding the knowledge distillation component. Subsequently, the incorporation of V sd leads to significant improvements, particularly in terms of rare mAP. This suggests that the internal features in stable diffusion play a crucial role in enhancing the representation of rare categories in the HICO-DET dataset. Additionally, the integration of v clip further aids in improving the non-rare AP, as it aligns the output queries with the interaction classifier. Diffusion Time Steps. We investigate the effectiveness of different diffusion steps in extracting interaction-aware features V sd , similarly to [66,23]. Diffusion models control the noise distortion added to the input image by varying the value of t. Stable diffusion [21] uses a total of 1000 time steps. We set t values to 0, 100, 500 for ablation studies. As demonstrated in Tab. 7, the best performance is achieved when t = 0. It is worth noting that using V sd extracted from input images with higher noise levels would decrease performance and potentially impact the learning of interactions.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We investigate the issues of the current HOI detection methods and present a novel scheme, namely DiffHOI, which leverages both the generative and representation capacities of a pre-trained text-toimage diffusion model to benefit the HOI detection task. Particularly, we release a class-balance, large-scale, and high-diversity synthetic HOI dataset called SynHOI to address the long-tail issue in previous datasets and develop an automatic and scalable pipeline to scale up the generation of diverse and high-precision HOI-annotated data. Extensive experimental results demonstrate that our method significantly outperforms the prior state-of-the-art in regular and zero-shot detection tasks. We hope this work could inspire further research in related fields.\"\n }\n]"},"abstract":{"kind":"string","value":"This paper investigates the problem of the current HOI detection methods and introduces DiffHOI, a novel HOI detection scheme grounded on a pre-trained text-image diffusion model, which enhances the detector's performance via improved data diversity and HOI representation. We demonstrate that the internal representation space of a frozen text-to-image diffusion model is highly relevant to verb concepts and their corresponding context. Accordingly, we propose an adapterstyle tuning method to extract the various semantic associated representation from a frozen diffusion model and CLIP model to enhance the human and object representations from the pre-trained detector, further reducing the ambiguity in interaction prediction. Moreover, to fill in the gaps of HOI datasets, we propose SynHOI, a class-balance, large-scale, and high-diversity synthetic dataset containing over 140K HOI images with fully triplet annotations. It is built using an automatic and scalable pipeline designed to scale up the generation of diverse and high-precision HOI-annotated data. SynHOI could effectively relieve the long-tail issue in existing datasets and facilitate learning interaction representations. Extensive experiments demonstrate that DiffHOI significantly outperforms the state-of-the-art in regular detection (i.e., 41.50 mAP) and zero-shot detection. Furthermore, SynHOI can improve the performance of model-agnostic and backbone-agnostic HOI detection, particularly exhibiting outstanding an 11.55% mAP improvement in rare classes."},"title":{"kind":"string","value":"Boosting Human-Object Interaction Detection with Text-to-Image Diffusion Model"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 1 :1Figure 1: We show a) the long-tail distribution issue in HICO-DET and b) the high correlation between HOI text (i.e., nouns and verbs) and internal image features within the frozen stable diffusion.","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"(a) Examples of HOIPrompts \"black young woman\" → \"latino teen\"","figure_data":"","figure_id":"fig_1","figure_label":"","figure_type":"figure"},{"figure_caption":"Examples of generated images and annotations","figure_data":"","figure_id":"fig_2","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 2 :2Figure 2: Illustration of a) HOIPrompts and b) how HOIPrompts guide the text-to-image generation process to enhance diversity. 3.1 Construction Process","figure_data":"","figure_id":"fig_3","figure_label":"2","figure_type":"figure"},{"figure_caption":"Figure 3 :3Figure 3: Overview of DiffHOI, comprising a pretrained human-object decoder, a novel interaction decoder, and CLIP-based object and interaction classifiers.","figure_data":"","figure_id":"fig_4","figure_label":"3","figure_type":"figure"},{"figure_caption":"3. 22Data Characteristics Large-scale. SynHOI consists of 146, 772 images, 157, 358 person bounding boxes, 165, 423 object bounding boxes, and 282, 140 HOI triplet instances. It provides approximately four times the amount of training data compared to HICO-DET. Class-balance. SynHOI can effectively address the long-tail issue in previous datasets, where 343 HOI categories have fewer than 50 images in HICO-DET. Combining SynHOI with HICO-DET reduces the number of HOI categories with fewer than 50 images to only three (refer to Fig. 1-(a)).","figure_data":"","figure_id":"fig_5","figure_label":"2","figure_type":"figure"},{"figure_caption":"4. 33CLIP Representation and Scene-aware Adaptor CLIP Preliminaries. The CLIP model is trained to align visual and textual representations. It comprises an image encoder F img clip and a text encoder F text clip , each followed by a linear layer, projecting image and text representations into a shared latent space. Given an image x, F img clip extracts a global visual representation v clip ∈ R C , while F text clip extracts a global text representation T clip ∈ R K×C for all K predefined categories. The prediction distribution over K categories is then calculated as S = softmax(T clip * v clip ), where * denotes the matrix-vector multiplication.","figure_data":"","figure_id":"fig_6","figure_label":"3","figure_type":"figure"},{"figure_caption":"Performance comparison on HICO-DET in terms of mAP. † means pre-training using our SynHOI dataset. The underlined highlights the compared results.5.2 Comparison to State-of-the-Art for Regular HOI DetectionTab. 1 and Tab. 2 present the performance comparison between DiffHOI and other state-of-the-art methods, which are grouped into one-stage and two-stage methods. For HICO-DET, DiffHOI-L outperforms all existing one-stage and two-stage methods by a significant margin in all evaluation settings. Notably, it achieves a new state-of-the-art performance of 40.84 mAP in the Default Full setting. Using the same backbone Swin-L, DiffHOI-L demonstrates a performance improvement of 3.45 mAP compared to the current one-stage state-of-the-art method FGAHOI[42]. Furthermore, by incorporating our SynHOI dataset, the performance is further boosted to 41.50 mAP. For V-COCO, DiffHOI-L also surpasses the previous state-of-the-art methods with role AP scores of 65.7 on Scenario 1 and 68.2 on Scenario 2.","figure_data":"DefaultKnown ObjectMethodBackboneFullRare Non-RareFullRare Non-RareTwo-stage methodsSTIP [59]ResNet-5032.2228.1533.4335.2931.4336.45DEFR [60]ViT-B/1632.3533.4532.02---UPT [61]ResNet-10132.6228.6233.8136.0831.4137.47ViPLO [36]ViT-B/3234.9533.8335.2838.1536.7738.56ViPLO [36]ViT-B/1637.2235.4537.7540.6138.8241.15One-stage methodsQPIC [39]ResNet-10129.9023.9231.6932.3826.0634.27MSTR [62]ResNet-5031.1725.3132.9234.0228.8335.572SSRT [63]ResNet-10131.3424.3133.32---CDN [8]ResNet-10132.0727.1933.5334.7929.4836.38DOQ [64]ResNet-5033.2829.1934.50---IF [65]ResNet-5033.5130.3034.4636.2833.1637.21GEN-VLKT [6] ResNet-5033.7529.2535.1036.7832.7537.99QAHOI [41]Swin-T28.4722.4430.2730.9924.8332.84QAHOI [41]Swin-L35.7829.8037.5637.5931.6639.36FGAHOI [42]Swin-T29.9422.2432.2432.4824.1634.97FGAHOI [42]Swin-L37.1830.7139.1138.9331.9341.02DiffHOI-SResNet-5034.41↑ 1.96% 31.0735.4037.31↑ 1.44% 34.5638.14DiffHOI-LSwin-L40.63↑ 9.28% 38.1041.3843.14↑ 10.81% 40.2444.01DiffHOI-L †Swin-L41.50↑ 11.61% 39.9641.9643.62↑ 12.05% 41.4144.28MethodAP (Scenario 1) AP (Scenario 2)Two-stage methodsSCG [35]54.260.9UPT [61]61.367.1ViPLO [36]60.966.6One-stage methodsQPIC [39]58.360.7MSTR [62]62.065.2CDN [8]63.965.9IF [65]63.065.2GEN-VLKT [6]62.464.5ParMap [7]63.065.1DiffHOI-S61.163.5DiffHOI-L65.768.2","figure_id":"tab_0","figure_label":"1","figure_type":"table"},{"figure_caption":"Performance comparison on V-COCO.","figure_data":"MethodTypeUnseen SeenFullGEN-VLKT[6] RF-UC21.3632.91 30.56DiffHOI-SRF-UC24.1332.93 31.08DiffHOI-LRF-UC28.7638.01 36.16GEN-VLKT[6] NF-UC25.0523.38 23.71DiffHOI-SNF-UC26.5725.55 25.75DiffHOI-LNF-UC29.4531.68 31.24GEN-VLKT[6]UO10.5128.92 25.63DiffHOI-SUO9.4229.79 26.22DiffHOI-LUO5.7535.08 30.11GEN-VLKT[6]UV20.9630.23 28.74DiffHOI-SUV23.1030.91 29.72DiffHOI-LUV24.2036.81 35.04","figure_id":"tab_1","figure_label":"2","figure_type":"table"},{"figure_caption":"Zero-shot performance comparison on HICO-DET.5.3 Comparison to State-of-the-Art for Zero-Shot HOI DetectionTo comprehensively assess the zero-shot capability of DiffHOI, we conduct experiments on various zero-shot settings, as shown in Tab. 3. It demonstrates that DiffHOI-L outperforms the state-of-theart method-GEN-VLKT[6] in all zero-shot settings. Notably, DiffHOI-S with the same backbone achieves remarkable improvements, with a +2.04 mAP gain under the NF-UC setting for all categories and a +2.14 mAP improvement for rare categories under the UO setting.5.4 Investigating the Usefulness of the SynHOI DatasetAs outlined in Sec. 3, we develop the SynHOI dataset, consisting of over 140K synthetic images that are generated, filtered, and annotated automatically. Additionally, we present a subset of SynHOI,","figure_data":"MethodBackboneFullRareNon-RareGEN-VLKT ‡ ResNet5033.0429.1034.21GEN-VLKT [4]ResNet50 34.43↑ 4.21% 32.46↑ 11.55% 35.02↑ 2.37%DiffHOI-LSwin-L40.6338.1041.38DiffHOI-LSwin-L41.50↑ 2.14% 39.96↑ 4.88%6 41.96↑ 1.40%","figure_id":"tab_2","figure_label":"3","figure_type":"table"},{"figure_caption":"Effectiveness of pre-training using Syn-HOI and fine-tuning on HICO-DET. ‡ indicates that our experiments reproduce the results.","figure_data":"MethodBackboneFullRareNon-RareGEN-VLKT ‡ [4] ResNet5033.0429.1034.21GEN-VLKT [4]ResNet50 33.73↑ 2.09% 30.08↑ 3.37% 34.82↑ 1.78%DiffHOI-LSwin-L40.6338.1041.38DiffHOI-LSwin-L41.42↑ 1.94% 39.94↑ 4.83% 41.87↑ 1.18%","figure_id":"tab_3","figure_label":"4","figure_type":"table"},{"figure_caption":"Effectiveness of joint-training using SynHOI-sub and HICO-DET, evaluated on HICO-DET. named as SynHOI-Sub, which includes over 8K data that has undergone manual inspection to ensure the quality of annotations. Here, we explore two strategies to leverage SynHOI and SynHOI-Sub to benefit the HOI detection task.Pre-training. An intuitive strategy is to use SynHOI to pre-train an HOI detector and fine-tune the model on the target dataset. As shown in Tab. 4, SynHOI could improve the performance of model-agnostic and backbone-agnostic HOI detection. Notably, there is a significant improvement in the rare AP when evaluating 138 HOI categories with fewer than 10 training instances in HICO-DET. Specifically, the rare AP increases significantly by 3.36 mAP for GEN-VLKT and 1.86 mAP for DiffHOI. It demonstrates that SynHOI can effectively relieve the long-tail issue and boost the interaction modeling in HICO-DET.Joint-training with HICO-DET. SynHOI-Sub contains more diverse and class-balance images with accurate annotations than HICO-DET. We employ it in conjunction with HICO-DET to perform joint training of the models. As shown in Tab. 5, the inclusion of SynHOI-Sub leads to an improvement of 0.69 mAP for GEN-VLKT and 0.58 mAP for DiffHOI, which can also contribute to improving the rare mAP and enhancing the learning of interactions within the rare categories.","figure_data":"5.5 Ablation StudyV sd v clipFullRare Non-rare31.99 29.6332.7032.92 31.2933.4134.41 31.0735.40","figure_id":"tab_4","figure_label":"5","figure_type":"table"},{"figure_caption":"Ablation results of two key representations.","figure_data":"Time stepFullRare Non-Rare034.41 31.0735.4010034.03 30.5835.0250033.59 29.8034.71","figure_id":"tab_5","figure_label":"6","figure_type":"table"},{"figure_caption":"Ablation results of different diffusion time steps.","figure_data":"","figure_id":"tab_6","figure_label":"7","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 1 :1Figure 1: We show a) the long-tail distribution issue in HICO-DET and b) the high correlation between HOI text (i.e., nouns and verbs) and internal image features within the frozen stable diffusion.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"(a) Examples of HOIPrompts \\\"black young woman\\\" → \\\"latino teen\\\"\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Examples of generated images and annotations\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 2 :2Figure 2: Illustration of a) HOIPrompts and b) how HOIPrompts guide the text-to-image generation process to enhance diversity. 3.1 Construction Process\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 :3Figure 3: Overview of DiffHOI, comprising a pretrained human-object decoder, a novel interaction decoder, and CLIP-based object and interaction classifiers.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"3. 22Data Characteristics Large-scale. SynHOI consists of 146, 772 images, 157, 358 person bounding boxes, 165, 423 object bounding boxes, and 282, 140 HOI triplet instances. It provides approximately four times the amount of training data compared to HICO-DET. Class-balance. SynHOI can effectively address the long-tail issue in previous datasets, where 343 HOI categories have fewer than 50 images in HICO-DET. Combining SynHOI with HICO-DET reduces the number of HOI categories with fewer than 50 images to only three (refer to Fig. 1-(a)).\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_5\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"4. 33CLIP Representation and Scene-aware Adaptor CLIP Preliminaries. The CLIP model is trained to align visual and textual representations. It comprises an image encoder F img clip and a text encoder F text clip , each followed by a linear layer, projecting image and text representations into a shared latent space. Given an image x, F img clip extracts a global visual representation v clip ∈ R C , while F text clip extracts a global text representation T clip ∈ R K×C for all K predefined categories. The prediction distribution over K categories is then calculated as S = softmax(T clip * v clip ), where * denotes the matrix-vector multiplication.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_6\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Performance comparison on HICO-DET in terms of mAP. † means pre-training using our SynHOI dataset. The underlined highlights the compared results.5.2 Comparison to State-of-the-Art for Regular HOI DetectionTab. 1 and Tab. 2 present the performance comparison between DiffHOI and other state-of-the-art methods, which are grouped into one-stage and two-stage methods. For HICO-DET, DiffHOI-L outperforms all existing one-stage and two-stage methods by a significant margin in all evaluation settings. Notably, it achieves a new state-of-the-art performance of 40.84 mAP in the Default Full setting. Using the same backbone Swin-L, DiffHOI-L demonstrates a performance improvement of 3.45 mAP compared to the current one-stage state-of-the-art method FGAHOI[42]. Furthermore, by incorporating our SynHOI dataset, the performance is further boosted to 41.50 mAP. For V-COCO, DiffHOI-L also surpasses the previous state-of-the-art methods with role AP scores of 65.7 on Scenario 1 and 68.2 on Scenario 2.\",\n \"figure_data\": \"DefaultKnown ObjectMethodBackboneFullRare Non-RareFullRare Non-RareTwo-stage methodsSTIP [59]ResNet-5032.2228.1533.4335.2931.4336.45DEFR [60]ViT-B/1632.3533.4532.02---UPT [61]ResNet-10132.6228.6233.8136.0831.4137.47ViPLO [36]ViT-B/3234.9533.8335.2838.1536.7738.56ViPLO [36]ViT-B/1637.2235.4537.7540.6138.8241.15One-stage methodsQPIC [39]ResNet-10129.9023.9231.6932.3826.0634.27MSTR [62]ResNet-5031.1725.3132.9234.0228.8335.572SSRT [63]ResNet-10131.3424.3133.32---CDN [8]ResNet-10132.0727.1933.5334.7929.4836.38DOQ [64]ResNet-5033.2829.1934.50---IF [65]ResNet-5033.5130.3034.4636.2833.1637.21GEN-VLKT [6] ResNet-5033.7529.2535.1036.7832.7537.99QAHOI [41]Swin-T28.4722.4430.2730.9924.8332.84QAHOI [41]Swin-L35.7829.8037.5637.5931.6639.36FGAHOI [42]Swin-T29.9422.2432.2432.4824.1634.97FGAHOI [42]Swin-L37.1830.7139.1138.9331.9341.02DiffHOI-SResNet-5034.41↑ 1.96% 31.0735.4037.31↑ 1.44% 34.5638.14DiffHOI-LSwin-L40.63↑ 9.28% 38.1041.3843.14↑ 10.81% 40.2444.01DiffHOI-L †Swin-L41.50↑ 11.61% 39.9641.9643.62↑ 12.05% 41.4144.28MethodAP (Scenario 1) AP (Scenario 2)Two-stage methodsSCG [35]54.260.9UPT [61]61.367.1ViPLO [36]60.966.6One-stage methodsQPIC [39]58.360.7MSTR [62]62.065.2CDN [8]63.965.9IF [65]63.065.2GEN-VLKT [6]62.464.5ParMap [7]63.065.1DiffHOI-S61.163.5DiffHOI-L65.768.2\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Performance comparison on V-COCO.\",\n \"figure_data\": \"MethodTypeUnseen SeenFullGEN-VLKT[6] RF-UC21.3632.91 30.56DiffHOI-SRF-UC24.1332.93 31.08DiffHOI-LRF-UC28.7638.01 36.16GEN-VLKT[6] NF-UC25.0523.38 23.71DiffHOI-SNF-UC26.5725.55 25.75DiffHOI-LNF-UC29.4531.68 31.24GEN-VLKT[6]UO10.5128.92 25.63DiffHOI-SUO9.4229.79 26.22DiffHOI-LUO5.7535.08 30.11GEN-VLKT[6]UV20.9630.23 28.74DiffHOI-SUV23.1030.91 29.72DiffHOI-LUV24.2036.81 35.04\",\n \"figure_id\": \"tab_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Zero-shot performance comparison on HICO-DET.5.3 Comparison to State-of-the-Art for Zero-Shot HOI DetectionTo comprehensively assess the zero-shot capability of DiffHOI, we conduct experiments on various zero-shot settings, as shown in Tab. 3. It demonstrates that DiffHOI-L outperforms the state-of-theart method-GEN-VLKT[6] in all zero-shot settings. Notably, DiffHOI-S with the same backbone achieves remarkable improvements, with a +2.04 mAP gain under the NF-UC setting for all categories and a +2.14 mAP improvement for rare categories under the UO setting.5.4 Investigating the Usefulness of the SynHOI DatasetAs outlined in Sec. 3, we develop the SynHOI dataset, consisting of over 140K synthetic images that are generated, filtered, and annotated automatically. Additionally, we present a subset of SynHOI,\",\n \"figure_data\": \"MethodBackboneFullRareNon-RareGEN-VLKT ‡ ResNet5033.0429.1034.21GEN-VLKT [4]ResNet50 34.43↑ 4.21% 32.46↑ 11.55% 35.02↑ 2.37%DiffHOI-LSwin-L40.6338.1041.38DiffHOI-LSwin-L41.50↑ 2.14% 39.96↑ 4.88%6 41.96↑ 1.40%\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Effectiveness of pre-training using Syn-HOI and fine-tuning on HICO-DET. ‡ indicates that our experiments reproduce the results.\",\n \"figure_data\": \"MethodBackboneFullRareNon-RareGEN-VLKT ‡ [4] ResNet5033.0429.1034.21GEN-VLKT [4]ResNet50 33.73↑ 2.09% 30.08↑ 3.37% 34.82↑ 1.78%DiffHOI-LSwin-L40.6338.1041.38DiffHOI-LSwin-L41.42↑ 1.94% 39.94↑ 4.83% 41.87↑ 1.18%\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Effectiveness of joint-training using SynHOI-sub and HICO-DET, evaluated on HICO-DET. named as SynHOI-Sub, which includes over 8K data that has undergone manual inspection to ensure the quality of annotations. Here, we explore two strategies to leverage SynHOI and SynHOI-Sub to benefit the HOI detection task.Pre-training. An intuitive strategy is to use SynHOI to pre-train an HOI detector and fine-tune the model on the target dataset. As shown in Tab. 4, SynHOI could improve the performance of model-agnostic and backbone-agnostic HOI detection. Notably, there is a significant improvement in the rare AP when evaluating 138 HOI categories with fewer than 10 training instances in HICO-DET. Specifically, the rare AP increases significantly by 3.36 mAP for GEN-VLKT and 1.86 mAP for DiffHOI. It demonstrates that SynHOI can effectively relieve the long-tail issue and boost the interaction modeling in HICO-DET.Joint-training with HICO-DET. SynHOI-Sub contains more diverse and class-balance images with accurate annotations than HICO-DET. We employ it in conjunction with HICO-DET to perform joint training of the models. As shown in Tab. 5, the inclusion of SynHOI-Sub leads to an improvement of 0.69 mAP for GEN-VLKT and 0.58 mAP for DiffHOI, which can also contribute to improving the rare mAP and enhancing the learning of interactions within the rare categories.\",\n \"figure_data\": \"5.5 Ablation StudyV sd v clipFullRare Non-rare31.99 29.6332.7032.92 31.2933.4134.41 31.0735.40\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"5\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Ablation results of two key representations.\",\n \"figure_data\": \"Time stepFullRare Non-Rare034.41 31.0735.4010034.03 30.5835.0250033.59 29.8034.71\",\n \"figure_id\": \"tab_5\",\n \"figure_label\": \"6\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Ablation results of different diffusion time steps.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_6\",\n \"figure_label\": \"7\",\n \"figure_type\": \"table\"\n }\n]"},"authors":{"kind":"string","value":"Jie Yang; Bingliang Li; Fengyu Yang; Ailing Zeng; ‡ Lei Zhang; Ruimao Zhang"},"references":{"kind":"list like","value":[{"authors":"Fabian Caba Heilbron; Victor Escorcia; Bernard Ghanem; Juan Niebles","journal":"IEEE","ref_id":"b0","title":"Activitynet: A large-scale video benchmark for human activity understanding","year":"2015"},{"authors":"Xinyu Yi; Yuxiao Zhou; Marc Habermann; Soshi Shimada; Vladislav Golyanik; Christian Theobalt; Feng Xu","journal":"","ref_id":"b1","title":"Physical inertial poser (pip): Physics-aware real-time human motion tracking from sparse inertial sensors","year":"2022"},{"authors":"Wen Liu; Weixin Luo; Dongze Lian; Shenghua Gao","journal":"","ref_id":"b2","title":"Future frame prediction for anomaly detection-a new baseline","year":"2018"},{"authors":"Zhi Hou; Xiaojiang Peng; Yu Qiao; Dacheng Tao","journal":"Springer","ref_id":"b3","title":"Visual compositional learning for humanobject interaction detection","year":"2020"},{"authors":"Bo Wan; Desen Zhou; Yongfei Liu; Rongjie Li; Xuming He","journal":"","ref_id":"b4","title":"Pose-aware multi-level feature network for human object interaction detection","year":"2019"},{"authors":"Yue Liao; Aixi Zhang; Miao Lu; Yongliang Wang; Xiaobo Li; Si Liu","journal":"","ref_id":"b5","title":"Gen-vlkt: Simplify association and enhance interaction understanding for hoi detection","year":"2022"},{"authors":"Xiaoqian Wu; Yong-Lu Li; Xinpeng Liu; Junyi Zhang; Yuzhe Wu; Cewu Lu","journal":"Springer","ref_id":"b6","title":"Mining cross-person cues for body-part interactiveness learning in hoi detection","year":"2022"},{"authors":"Aixi Zhang; Yue Liao; Si Liu; Miao Lu; Yongliang Wang; Chen Gao; Xiaobo Li","journal":"Advances in Neural Information Processing Systems","ref_id":"b7","title":"Mining the benefits of two-stage and one-stage hoi detection","year":"2021"},{"authors":"Desen Zhou; Zhichao Liu; Jian Wang; Leshan Wang; Tao Hu; Errui Ding; Jingdong Wang","journal":"","ref_id":"b8","title":"Human-object interaction detection via disentangled transformer","year":"2022"},{"authors":"Yu-Wei Chao; Yunfan Liu; Xieyang Liu; Huayi Zeng; Jia Deng","journal":"","ref_id":"b9","title":"Learning to detect humanobject interactions","year":"2018"},{"authors":"Yu-Wei Chao; Yunfan Liu; Xieyang Liu; Huayi Zeng; Jia Deng","journal":"IEEE","ref_id":"b10","title":"Learning to detect humanobject interactions","year":"2018"},{"authors":"Chen Gao; Yuliang Zou; Jia-Bin Huang","journal":"","ref_id":"b11","title":"ican: Instance-centric attention network for human-object interaction detection","year":"2018"},{"authors":"Tanmay Gupta; Alexander Schwing; Derek Hoiem","journal":"","ref_id":"b12","title":"No-frills human-object interaction detection: Factorization, layout encodings, and training techniques","year":"2019"},{"authors":"Yue Liao; Si Liu; Fei Wang; Yanjie Chen; Chen Qian; Jiashi Feng","journal":"","ref_id":"b13","title":"Ppdm: Parallel point detection and matching for real-time human-object interaction detection","year":"2020"},{"authors":"Tiancai Wang; Tong Yang; Martin Danelljan; Fahad Shahbaz Khan; Xiangyu Zhang; Jian Sun","journal":"","ref_id":"b14","title":"Learning human-object interaction detection using interaction points","year":"2020"},{"authors":"Bumsoo Kim; Taeho Choi; Jaewoo Kang; Hyunwoo J Kim","journal":"Springer","ref_id":"b15","title":"Union-level detector towards real-time human-object interaction detection","year":"2020"},{"authors":"Yogesh Balaji; Seungjun Nah; Xun Huang; Arash Vahdat; Jiaming Song; Karsten Kreis; Miika Aittala; Timo Aila; Samuli Laine; Bryan Catanzaro","journal":"","ref_id":"b16","title":"ediffi: Text-to-image diffusion models with an ensemble of expert denoisers","year":"2022"},{"authors":"Yufan Zhou; Ruiyi Zhang; Changyou Chen; Chunyuan Li; Chris Tensmeyer; Tong Yu; Jiuxiang Gu; Jinhui Xu; Tong Sun","journal":"","ref_id":"b17","title":"Towards language-free training for text-to-image generation","year":"2022"},{"authors":"Chitwan Saharia; William Chan; Saurabh Saxena; Lala Li; Jay Whang; Emily L Denton; Kamyar Ghasemipour; Raphael Gontijo Lopes; Burcu Karagol Ayan; Tim Salimans","journal":"Advances in Neural Information Processing Systems","ref_id":"b18","title":"Photorealistic text-to-image diffusion models with deep language understanding","year":"2022"},{"authors":"Aditya Ramesh; Prafulla Dhariwal; Alex Nichol; Casey Chu; Mark Chen","journal":"","ref_id":"b19","title":"Hierarchical text-conditional image generation with clip latents","year":"2022"},{"authors":"Robin Rombach; Andreas Blattmann; Dominik Lorenz; Patrick Esser; Björn Ommer","journal":"","ref_id":"b20","title":"Highresolution image synthesis with latent diffusion models","year":"2022"},{"authors":"Raphael Tang; Akshat Pandey; Zhiying Jiang; Gefei Yang; Karun Kumar; Jimmy Lin; Ferhan Ture","journal":"","ref_id":"b21","title":"What the daam: Interpreting stable diffusion using cross attention","year":"2022"},{"authors":"Jiarui Xu; Sifei Liu; Arash Vahdat; Wonmin Byeon; Xiaolong Wang; Shalini De Mello","journal":"","ref_id":"b22","title":"Open-vocabulary panoptic segmentation with text-to-image diffusion models","year":"2023"},{"authors":"Ziyi Li; Qinye Zhou; Xiaoyun Zhang; Ya Zhang; Yanfeng Wang; Weidi Xie","journal":"","ref_id":"b23","title":"Guiding text-to-image diffusion model towards grounded generation","year":"2023"},{"authors":"Jordan Shipard; Arnold Wiliem; Kien Nguyen Thanh; Wei Xiang; Clinton Fookes","journal":"","ref_id":"b24","title":"Diversity is definitely needed: Improving model-agnostic zero-shot classification via stable diffusion","year":""},{"authors":"Yong-Lu Li; Siyuan Zhou; Xijie Huang; Liang Xu; Ze Ma; Hao-Shu Fang; Yanfeng Wang; Cewu Lu","journal":"","ref_id":"b25","title":"Transferable interactiveness knowledge for human-object interaction detection","year":"2019"},{"authors":"Zhi Hou; Baosheng Yu; Yu Qiao; Xiaojiang Peng; Dacheng Tao","journal":"","ref_id":"b26","title":"Detecting human-object interaction via fabricated compositional learning","year":"2021"},{"authors":"Bingjie Xu; Yongkang Wong; Junnan Li; Qi Zhao; Mohan S Kankanhalli","journal":"","ref_id":"b27","title":"Learning to detect human-object interactions with knowledge","year":"2019"},{"authors":"Ankan Bansal; Sai Saketh Rambhatla; Abhinav Shrivastava; Rama Chellappa","journal":"","ref_id":"b28","title":"Detecting human-object interactions via functional generalization","year":"2020"},{"authors":"Yong-Lu Li; Xinpeng Liu; Han Lu; Shiyi Wang; Junqi Liu; Jiefeng Li; Cewu Lu","journal":"","ref_id":"b29","title":"Detailed 2d-3d joint representation for human-object interaction","year":"2020"},{"authors":"Ye Liu; Junsong Yuan; Chang Wen; Chen ","journal":"","ref_id":"b30","title":"Consnet: Learning graph for zero-shot human-object interaction detection","year":"2020"},{"authors":"Chen Gao; Jiarui Xu; Yuliang Zou; Jia-Bin Huang","journal":"Springer","ref_id":"b31","title":"Drg: Dual relation graph for humanobject interaction detection","year":"2020"},{"authors":"Siyuan Qi; Wenguan Wang; Baoxiong Jia; Jianbing Shen; Song-Chun Zhu","journal":"","ref_id":"b32","title":"Learning human-object interactions by graph parsing neural networks","year":"2018"},{"authors":"Hai Wang; Wei-Shi Zheng; Ling Yingbiao","journal":"Springer","ref_id":"b33","title":"Contextual heterogeneous graph network for human-object interaction detection","year":"2020"},{"authors":"Frederic Z Zhang; Dylan Campbell; Stephen Gould","journal":"","ref_id":"b34","title":"Spatially conditioned graphs for detecting human-object interactions","year":"2021"},{"authors":"Jeeseung Park; Jin-Woo Park; Jong-Seok Lee","journal":"","ref_id":"b35","title":"Viplo: Vision transformer based pose-conditioned self-loop graph for human-object interaction detection","year":"2023"},{"authors":"Yichen Hao-Shu Fang; Dian Xie; Cewu Shao; Lu","journal":"","ref_id":"b36","title":"Dirv: Dense interaction region voting for end-to-end human-object interaction detection","year":"2021"},{"authors":"Masato Tamura; Hiroki Ohashi; Tomoaki Yoshinaga","journal":"","ref_id":"b37","title":"Qpic: Query-based pairwise humanobject interaction detection with image-wide contextual information","year":"2021-06"},{"authors":"Masato Tamura; Hiroki Ohashi; Tomoaki Yoshinaga","journal":"","ref_id":"b38","title":"Qpic: Query-based pairwise humanobject interaction detection with image-wide contextual information","year":"2021"},{"authors":"Bumsoo Kim; Junhyun Lee; Jaewoo Kang; Eun-Sol Kim; Hyunwoo J Kim","journal":"","ref_id":"b39","title":"Hotr: End-toend human-object interaction detection with transformers","year":"2021"},{"authors":"Junwen Chen; Keiji Yanai","journal":"","ref_id":"b40","title":"Qahoi: query-based anchors for human-object interaction detection","year":"2021"},{"authors":"Shuailei Ma; Yuefeng Wang; Shanze Wang; Ying Wei","journal":"","ref_id":"b41","title":"Fgahoi: Fine-grained anchors for human-object interaction detection","year":"2023"},{"authors":"Mingfei Chen; Yue Liao; Si Liu; Zhiyuan Chen; Fei Wang; Chen Qian","journal":"","ref_id":"b42","title":"Reformulating hoi detection as adaptive set prediction","year":"2021"},{"authors":"Julia Peyre; Ivan Laptev; Cordelia Schmid; Josef Sivic","journal":"","ref_id":"b43","title":"Detecting unseen visual relations using analogies","year":"2019"},{"authors":"Zhi Hou; Baosheng Yu; Yu Qiao; Xiaojiang Peng; Dacheng Tao","journal":"","ref_id":"b44","title":"Affordance transfer learning for human-object interaction detection","year":"2021"},{"authors":"Alec Radford; Jong Wook Kim; Chris Hallacy; Aditya Ramesh; Gabriel Goh; Sandhini Agarwal; Girish Sastry; Amanda Askell; Pamela Mishkin; Jack Clark","journal":"PMLR","ref_id":"b45","title":"Learning transferable visual models from natural language supervision","year":"2021"},{"authors":"Mingrui Wu; Jiaxin Gu; Yunhang Shen; Mingbao Lin; Chao Chen; Sun ; Rongrong Ji","journal":"","ref_id":"b46","title":"End-to-end zero-shot hoi detection via vision and language knowledge distillation","year":"2022"},{"authors":"Shan Ning; Longtian Qiu; Yongfei Liu; Xuming He","journal":"","ref_id":"b47","title":"Hoiclip: Efficient knowledge transfer for hoi detection with vision-language models","year":"2023"},{"authors":"Brandon Trabucco; Kyle Doherty; Max Gurinas; Ruslan Salakhutdinov","journal":"","ref_id":"b48","title":"Effective data augmentation with diffusion models","year":"2023"},{"authors":"Hritik Bansal; Aditya Grover","journal":"","ref_id":"b49","title":"Leaving reality to imagination: Robust classification via generated datasets","year":"2023"},{"authors":"Zebin You; Yong Zhong; Fan Bao; Jiacheng Sun; Chongxuan Li; Jun Zhu","journal":"","ref_id":"b50","title":"Diffusion models and semi-supervised learners benefit mutually with few labels","year":"2023"},{"authors":"Hao Zhang; Feng Li; Shilong Liu; Lei Zhang; Hang Su; Jun Zhu; Lionel M Ni; Heung-Yeung Shum","journal":"","ref_id":"b51","title":"Dino: Detr with improved denoising anchor boxes for end-to-end object detection","year":"2022"},{"authors":"Tsung-Yi Lin; Michael Maire; Serge Belongie; James Hays; Pietro Perona; Deva Ramanan; Piotr Dollár; C Lawrence; Zitnick ","journal":"Springer","ref_id":"b52","title":"Microsoft coco: Common objects in context","year":"2014"},{"authors":"Jack Hessel; Ari Holtzman; Maxwell Forbes; Ronan Le Bras; Yejin Choi","journal":"","ref_id":"b53","title":"Clipscore: A reference-free evaluation metric for image captioning","year":"2021"},{"authors":"Nicolas Carion; Francisco Massa; Gabriel Synnaeve; Nicolas Usunier; Alexander Kirillov; Sergey Zagoruyko","journal":"Springer","ref_id":"b54","title":"End-to-end object detection with transformers","year":"2020"},{"authors":"Saurabh Gupta; Jitendra Malik","journal":"","ref_id":"b55","title":"Visual semantic role labeling","year":"2015"},{"authors":"Ze Liu; Yutong Lin; Yue Cao; Han Hu; Yixuan Wei; Zheng Zhang; Stephen Lin; Baining Guo","journal":"","ref_id":"b56","title":"Swin transformer: Hierarchical vision transformer using shifted windows","year":"2021"},{"authors":"Xizhou Zhu; Weijie Su; Lewei Lu; Bin Li; Xiaogang Wang; Jifeng Dai","journal":"","ref_id":"b57","title":"Deformable detr: Deformable transformers for end-to-end object detection","year":"2020"},{"authors":"Yong Zhang; Yingwei Pan; Ting Yao; Rui Huang; Tao Mei; Chang-Wen Chen","journal":"","ref_id":"b58","title":"Exploring structure-aware transformer over interaction proposals for human-object interaction detection","year":"2022"},{"authors":"Ying Jin; Yinpeng Chen; Lijuan Wang; Jianfeng Wang; Pei Yu; Lin Liang; Jenq-Neng Hwang; Zicheng Liu","journal":"","ref_id":"b59","title":"The overlooked classifier in human-object interaction recognition","year":"2022"},{"authors":"Frederic Z Zhang; Dylan Campbell; Stephen Gould","journal":"","ref_id":"b60","title":"Efficient two-stage detection of humanobject interactions with a novel unary-pairwise transformer","year":"2022"},{"authors":"Bumsoo Kim; Jonghwan Mun; Minchul Kyoung-Woon On; Junhyun Shin; Eun-Sol Lee; Kim","journal":"","ref_id":"b61","title":"Mstr: Multi-scale transformer for end-to-end human-object interaction detection","year":"2022"},{"authors":"Hao Asm Iftekhar; Kaustav Chen; Xinyu Kundu; Joseph Li; Davide Tighe; Modolo","journal":"","ref_id":"b62","title":"What to look at and where: Semantic and spatial refined transformer for detecting human-object interactions","year":"2022"},{"authors":"Xian Qu; Changxing Ding; Xingao Li; Xubin Zhong; Dacheng Tao","journal":"","ref_id":"b63","title":"Distillation using oracle queries for transformer-based human-object interaction detection","year":"2022"},{"authors":"Xinpeng Liu; Yong-Lu Li; Xiaoqian Wu; Yu-Wing Tai; Cewu Lu; Chi-Keung Tang","journal":"","ref_id":"b64","title":"Interactiveness field in human-object interactions","year":"2022"},{"authors":"Dmitry Baranchuk; Ivan Rubachev; Andrey Voynov; Valentin Khrulkov; Artem Babenko","journal":"","ref_id":"b65","title":"Label-efficient semantic segmentation with diffusion models","year":"2021"}],"string":"[\n {\n \"authors\": \"Fabian Caba Heilbron; Victor Escorcia; Bernard Ghanem; Juan Niebles\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b0\",\n \"title\": \"Activitynet: A large-scale video benchmark for human activity understanding\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Xinyu Yi; Yuxiao Zhou; Marc Habermann; Soshi Shimada; Vladislav Golyanik; Christian Theobalt; Feng Xu\",\n \"journal\": \"\",\n \"ref_id\": \"b1\",\n \"title\": \"Physical inertial poser (pip): Physics-aware real-time human motion tracking from sparse inertial sensors\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Wen Liu; Weixin Luo; Dongze Lian; Shenghua Gao\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"Future frame prediction for anomaly detection-a new baseline\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Zhi Hou; Xiaojiang Peng; Yu Qiao; Dacheng Tao\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b3\",\n \"title\": \"Visual compositional learning for humanobject interaction detection\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Bo Wan; Desen Zhou; Yongfei Liu; Rongjie Li; Xuming He\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"Pose-aware multi-level feature network for human object interaction detection\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Yue Liao; Aixi Zhang; Miao Lu; Yongliang Wang; Xiaobo Li; Si Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b5\",\n \"title\": \"Gen-vlkt: Simplify association and enhance interaction understanding for hoi detection\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Xiaoqian Wu; Yong-Lu Li; Xinpeng Liu; Junyi Zhang; Yuzhe Wu; Cewu Lu\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b6\",\n \"title\": \"Mining cross-person cues for body-part interactiveness learning in hoi detection\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Aixi Zhang; Yue Liao; Si Liu; Miao Lu; Yongliang Wang; Chen Gao; Xiaobo Li\",\n \"journal\": \"Advances in Neural Information Processing Systems\",\n \"ref_id\": \"b7\",\n \"title\": \"Mining the benefits of two-stage and one-stage hoi detection\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Desen Zhou; Zhichao Liu; Jian Wang; Leshan Wang; Tao Hu; Errui Ding; Jingdong Wang\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"Human-object interaction detection via disentangled transformer\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Yu-Wei Chao; Yunfan Liu; Xieyang Liu; Huayi Zeng; Jia Deng\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"Learning to detect humanobject interactions\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Yu-Wei Chao; Yunfan Liu; Xieyang Liu; Huayi Zeng; Jia Deng\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b10\",\n \"title\": \"Learning to detect humanobject interactions\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Chen Gao; Yuliang Zou; Jia-Bin Huang\",\n \"journal\": \"\",\n \"ref_id\": \"b11\",\n \"title\": \"ican: Instance-centric attention network for human-object interaction detection\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Tanmay Gupta; Alexander Schwing; Derek Hoiem\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"No-frills human-object interaction detection: Factorization, layout encodings, and training techniques\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Yue Liao; Si Liu; Fei Wang; Yanjie Chen; Chen Qian; Jiashi Feng\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"Ppdm: Parallel point detection and matching for real-time human-object interaction detection\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Tiancai Wang; Tong Yang; Martin Danelljan; Fahad Shahbaz Khan; Xiangyu Zhang; Jian Sun\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"Learning human-object interaction detection using interaction points\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Bumsoo Kim; Taeho Choi; Jaewoo Kang; Hyunwoo J Kim\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b15\",\n \"title\": \"Union-level detector towards real-time human-object interaction detection\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Yogesh Balaji; Seungjun Nah; Xun Huang; Arash Vahdat; Jiaming Song; Karsten Kreis; Miika Aittala; Timo Aila; Samuli Laine; Bryan Catanzaro\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"ediffi: Text-to-image diffusion models with an ensemble of expert denoisers\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Yufan Zhou; Ruiyi Zhang; Changyou Chen; Chunyuan Li; Chris Tensmeyer; Tong Yu; Jiuxiang Gu; Jinhui Xu; Tong Sun\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"Towards language-free training for text-to-image generation\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Chitwan Saharia; William Chan; Saurabh Saxena; Lala Li; Jay Whang; Emily L Denton; Kamyar Ghasemipour; Raphael Gontijo Lopes; Burcu Karagol Ayan; Tim Salimans\",\n \"journal\": \"Advances in Neural Information Processing Systems\",\n \"ref_id\": \"b18\",\n \"title\": \"Photorealistic text-to-image diffusion models with deep language understanding\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Aditya Ramesh; Prafulla Dhariwal; Alex Nichol; Casey Chu; Mark Chen\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"Hierarchical text-conditional image generation with clip latents\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Robin Rombach; Andreas Blattmann; Dominik Lorenz; Patrick Esser; Björn Ommer\",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"Highresolution image synthesis with latent diffusion models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Raphael Tang; Akshat Pandey; Zhiying Jiang; Gefei Yang; Karun Kumar; Jimmy Lin; Ferhan Ture\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"What the daam: Interpreting stable diffusion using cross attention\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jiarui Xu; Sifei Liu; Arash Vahdat; Wonmin Byeon; Xiaolong Wang; Shalini De Mello\",\n \"journal\": \"\",\n \"ref_id\": \"b22\",\n \"title\": \"Open-vocabulary panoptic segmentation with text-to-image diffusion models\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Ziyi Li; Qinye Zhou; Xiaoyun Zhang; Ya Zhang; Yanfeng Wang; Weidi Xie\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"Guiding text-to-image diffusion model towards grounded generation\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Jordan Shipard; Arnold Wiliem; Kien Nguyen Thanh; Wei Xiang; Clinton Fookes\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Diversity is definitely needed: Improving model-agnostic zero-shot classification via stable diffusion\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Yong-Lu Li; Siyuan Zhou; Xijie Huang; Liang Xu; Ze Ma; Hao-Shu Fang; Yanfeng Wang; Cewu Lu\",\n \"journal\": \"\",\n \"ref_id\": \"b25\",\n \"title\": \"Transferable interactiveness knowledge for human-object interaction detection\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Zhi Hou; Baosheng Yu; Yu Qiao; Xiaojiang Peng; Dacheng Tao\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"Detecting human-object interaction via fabricated compositional learning\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Bingjie Xu; Yongkang Wong; Junnan Li; Qi Zhao; Mohan S Kankanhalli\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"Learning to detect human-object interactions with knowledge\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Ankan Bansal; Sai Saketh Rambhatla; Abhinav Shrivastava; Rama Chellappa\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"Detecting human-object interactions via functional generalization\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Yong-Lu Li; Xinpeng Liu; Han Lu; Shiyi Wang; Junqi Liu; Jiefeng Li; Cewu Lu\",\n \"journal\": \"\",\n \"ref_id\": \"b29\",\n \"title\": \"Detailed 2d-3d joint representation for human-object interaction\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Ye Liu; Junsong Yuan; Chang Wen; Chen \",\n \"journal\": \"\",\n \"ref_id\": \"b30\",\n \"title\": \"Consnet: Learning graph for zero-shot human-object interaction detection\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Chen Gao; Jiarui Xu; Yuliang Zou; Jia-Bin Huang\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b31\",\n \"title\": \"Drg: Dual relation graph for humanobject interaction detection\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Siyuan Qi; Wenguan Wang; Baoxiong Jia; Jianbing Shen; Song-Chun Zhu\",\n \"journal\": \"\",\n \"ref_id\": \"b32\",\n \"title\": \"Learning human-object interactions by graph parsing neural networks\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Hai Wang; Wei-Shi Zheng; Ling Yingbiao\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b33\",\n \"title\": \"Contextual heterogeneous graph network for human-object interaction detection\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Frederic Z Zhang; Dylan Campbell; Stephen Gould\",\n \"journal\": \"\",\n \"ref_id\": \"b34\",\n \"title\": \"Spatially conditioned graphs for detecting human-object interactions\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Jeeseung Park; Jin-Woo Park; Jong-Seok Lee\",\n \"journal\": \"\",\n \"ref_id\": \"b35\",\n \"title\": \"Viplo: Vision transformer based pose-conditioned self-loop graph for human-object interaction detection\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Yichen Hao-Shu Fang; Dian Xie; Cewu Shao; Lu\",\n \"journal\": \"\",\n \"ref_id\": \"b36\",\n \"title\": \"Dirv: Dense interaction region voting for end-to-end human-object interaction detection\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Masato Tamura; Hiroki Ohashi; Tomoaki Yoshinaga\",\n \"journal\": \"\",\n \"ref_id\": \"b37\",\n \"title\": \"Qpic: Query-based pairwise humanobject interaction detection with image-wide contextual information\",\n \"year\": \"2021-06\"\n },\n {\n \"authors\": \"Masato Tamura; Hiroki Ohashi; Tomoaki Yoshinaga\",\n \"journal\": \"\",\n \"ref_id\": \"b38\",\n \"title\": \"Qpic: Query-based pairwise humanobject interaction detection with image-wide contextual information\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Bumsoo Kim; Junhyun Lee; Jaewoo Kang; Eun-Sol Kim; Hyunwoo J Kim\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"Hotr: End-toend human-object interaction detection with transformers\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Junwen Chen; Keiji Yanai\",\n \"journal\": \"\",\n \"ref_id\": \"b40\",\n \"title\": \"Qahoi: query-based anchors for human-object interaction detection\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Shuailei Ma; Yuefeng Wang; Shanze Wang; Ying Wei\",\n \"journal\": \"\",\n \"ref_id\": \"b41\",\n \"title\": \"Fgahoi: Fine-grained anchors for human-object interaction detection\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Mingfei Chen; Yue Liao; Si Liu; Zhiyuan Chen; Fei Wang; Chen Qian\",\n \"journal\": \"\",\n \"ref_id\": \"b42\",\n \"title\": \"Reformulating hoi detection as adaptive set prediction\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Julia Peyre; Ivan Laptev; Cordelia Schmid; Josef Sivic\",\n \"journal\": \"\",\n \"ref_id\": \"b43\",\n \"title\": \"Detecting unseen visual relations using analogies\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Zhi Hou; Baosheng Yu; Yu Qiao; Xiaojiang Peng; Dacheng Tao\",\n \"journal\": \"\",\n \"ref_id\": \"b44\",\n \"title\": \"Affordance transfer learning for human-object interaction detection\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Alec Radford; Jong Wook Kim; Chris Hallacy; Aditya Ramesh; Gabriel Goh; Sandhini Agarwal; Girish Sastry; Amanda Askell; Pamela Mishkin; Jack Clark\",\n \"journal\": \"PMLR\",\n \"ref_id\": \"b45\",\n \"title\": \"Learning transferable visual models from natural language supervision\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Mingrui Wu; Jiaxin Gu; Yunhang Shen; Mingbao Lin; Chao Chen; Sun ; Rongrong Ji\",\n \"journal\": \"\",\n \"ref_id\": \"b46\",\n \"title\": \"End-to-end zero-shot hoi detection via vision and language knowledge distillation\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Shan Ning; Longtian Qiu; Yongfei Liu; Xuming He\",\n \"journal\": \"\",\n \"ref_id\": \"b47\",\n \"title\": \"Hoiclip: Efficient knowledge transfer for hoi detection with vision-language models\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Brandon Trabucco; Kyle Doherty; Max Gurinas; Ruslan Salakhutdinov\",\n \"journal\": \"\",\n \"ref_id\": \"b48\",\n \"title\": \"Effective data augmentation with diffusion models\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Hritik Bansal; Aditya Grover\",\n \"journal\": \"\",\n \"ref_id\": \"b49\",\n \"title\": \"Leaving reality to imagination: Robust classification via generated datasets\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Zebin You; Yong Zhong; Fan Bao; Jiacheng Sun; Chongxuan Li; Jun Zhu\",\n \"journal\": \"\",\n \"ref_id\": \"b50\",\n \"title\": \"Diffusion models and semi-supervised learners benefit mutually with few labels\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Hao Zhang; Feng Li; Shilong Liu; Lei Zhang; Hang Su; Jun Zhu; Lionel M Ni; Heung-Yeung Shum\",\n \"journal\": \"\",\n \"ref_id\": \"b51\",\n \"title\": \"Dino: Detr with improved denoising anchor boxes for end-to-end object detection\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Tsung-Yi Lin; Michael Maire; Serge Belongie; James Hays; Pietro Perona; Deva Ramanan; Piotr Dollár; C Lawrence; Zitnick \",\n \"journal\": \"Springer\",\n \"ref_id\": \"b52\",\n \"title\": \"Microsoft coco: Common objects in context\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"Jack Hessel; Ari Holtzman; Maxwell Forbes; Ronan Le Bras; Yejin Choi\",\n \"journal\": \"\",\n \"ref_id\": \"b53\",\n \"title\": \"Clipscore: A reference-free evaluation metric for image captioning\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Nicolas Carion; Francisco Massa; Gabriel Synnaeve; Nicolas Usunier; Alexander Kirillov; Sergey Zagoruyko\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b54\",\n \"title\": \"End-to-end object detection with transformers\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Saurabh Gupta; Jitendra Malik\",\n \"journal\": \"\",\n \"ref_id\": \"b55\",\n \"title\": \"Visual semantic role labeling\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Ze Liu; Yutong Lin; Yue Cao; Han Hu; Yixuan Wei; Zheng Zhang; Stephen Lin; Baining Guo\",\n \"journal\": \"\",\n \"ref_id\": \"b56\",\n \"title\": \"Swin transformer: Hierarchical vision transformer using shifted windows\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Xizhou Zhu; Weijie Su; Lewei Lu; Bin Li; Xiaogang Wang; Jifeng Dai\",\n \"journal\": \"\",\n \"ref_id\": \"b57\",\n \"title\": \"Deformable detr: Deformable transformers for end-to-end object detection\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Yong Zhang; Yingwei Pan; Ting Yao; Rui Huang; Tao Mei; Chang-Wen Chen\",\n \"journal\": \"\",\n \"ref_id\": \"b58\",\n \"title\": \"Exploring structure-aware transformer over interaction proposals for human-object interaction detection\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Ying Jin; Yinpeng Chen; Lijuan Wang; Jianfeng Wang; Pei Yu; Lin Liang; Jenq-Neng Hwang; Zicheng Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b59\",\n \"title\": \"The overlooked classifier in human-object interaction recognition\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Frederic Z Zhang; Dylan Campbell; Stephen Gould\",\n \"journal\": \"\",\n \"ref_id\": \"b60\",\n \"title\": \"Efficient two-stage detection of humanobject interactions with a novel unary-pairwise transformer\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Bumsoo Kim; Jonghwan Mun; Minchul Kyoung-Woon On; Junhyun Shin; Eun-Sol Lee; Kim\",\n \"journal\": \"\",\n \"ref_id\": \"b61\",\n \"title\": \"Mstr: Multi-scale transformer for end-to-end human-object interaction detection\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Hao Asm Iftekhar; Kaustav Chen; Xinyu Kundu; Joseph Li; Davide Tighe; Modolo\",\n \"journal\": \"\",\n \"ref_id\": \"b62\",\n \"title\": \"What to look at and where: Semantic and spatial refined transformer for detecting human-object interactions\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Xian Qu; Changxing Ding; Xingao Li; Xubin Zhong; Dacheng Tao\",\n \"journal\": \"\",\n \"ref_id\": \"b63\",\n \"title\": \"Distillation using oracle queries for transformer-based human-object interaction detection\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Xinpeng Liu; Yong-Lu Li; Xiaoqian Wu; Yu-Wing Tai; Cewu Lu; Chi-Keung Tang\",\n \"journal\": \"\",\n \"ref_id\": \"b64\",\n \"title\": \"Interactiveness field in human-object interactions\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Dmitry Baranchuk; Ivan Rubachev; Andrey Voynov; Valentin Khrulkov; Artem Babenko\",\n \"journal\": \"\",\n \"ref_id\": \"b65\",\n \"title\": \"Label-efficient semantic segmentation with diffusion models\",\n \"year\": \"2021\"\n }\n]"},"formulas":{"kind":"list like","value":[{"formula_coordinates":[7,247.2,303.93,256.81,12.95],"formula_id":"formula_0","formula_text":"P o = softmax(Q o * T o clip )(1)"},{"formula_coordinates":[7,248.31,321.26,255.69,12.95],"formula_id":"formula_1","formula_text":"P i = softmax(Q i * T i clip )(2)"},{"formula_coordinates":[7,135.1,338.39,124.87,12.19],"formula_id":"formula_2","formula_text":"Q o ∈ R N ×C and Q o ∈ R N ×C"},{"formula_coordinates":[7,233.78,452.58,270.22,12.69],"formula_id":"formula_3","formula_text":"L = λ b L b + λ g L g + λ o c L o c + λ i c L i c ,(3)"}],"string":"[\n {\n \"formula_coordinates\": [\n 7,\n 247.2,\n 303.93,\n 256.81,\n 12.95\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"P o = softmax(Q o * T o clip )(1)\"\n },\n {\n \"formula_coordinates\": [\n 7,\n 248.31,\n 321.26,\n 255.69,\n 12.95\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"P i = softmax(Q i * T i clip )(2)\"\n },\n {\n \"formula_coordinates\": [\n 7,\n 135.1,\n 338.39,\n 124.87,\n 12.19\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"Q o ∈ R N ×C and Q o ∈ R N ×C\"\n },\n {\n \"formula_coordinates\": [\n 7,\n 233.78,\n 452.58,\n 270.22,\n 12.69\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"L = λ b L b + λ g L g + λ o c L o c + λ i c L i c ,(3)\"\n }\n]"}}},{"rowIdx":385,"cells":{"doi":{"kind":"string","value":""},"pub_date":{"kind":"string","value":"2023-05-20"},"sections":{"kind":"list like","value":[{"figure_ref":[],"heading":"Introduction","publication_ref":["b32"],"table_ref":[],"text":"The standard training strategy of a modern Neural Image Captioning system includes a policy gradient method, called Self-Critical Sequence Training [27] (shortened as SCST) which is designed to maximize the evaluation score given to the outputs. In this work, we discuss the problems caused by the lack of transparency from the research community over the inclusion or omission of the End-of-Sequence token during the optimization. An easy-to-overlook implementation detail that can significantly increase the performance of any model despite yielding worse descriptions.\nThe lack of awareness of the impact of the End-of-Sequence () omission and the lack of explicit information on the SCST implementation during the reporting of results pose an obstacle to scientific progress as they make it challenging to compare established works and evaluate new ones. Our paper attempts to spread awareness about the issue and proposes a solution to increase transparency in future works. This paper is structured as follows: in Section 2, we discuss the problem of the End-of-Sequence omission and why it is a problem for the research community; in Section 3, we provide a qualitative and quantitative analysis of the issue and we sample some of the recent works in Image Captioning to demonstrate its pervasiveness and provide some practical examples of its impact; In Section 4, we propose a possible solution with the help of a Python library called SacreEOS; in Section 5, we mention some of the literature approaches, and, finally, we draw our conclusions in Section 6."},{"figure_ref":[],"heading":"Problem Description","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"CIDEr Optimization","publication_ref":["b35","b29","b21","b8","b7","b35","b32"],"table_ref":[],"text":"CIDEr [30] is an n-gram-based metric that evaluates the caption semantic content according to its similarities to the ground truths. Compared to the other metrics [24,16,3,2], it exploits the entire corpus of reference descriptions in the attempt of backing the evaluation with the consensus of the majority of people. In particular, each n-gram w k in sequence Z is weighted according to the tf-idf term g n k (Z) defined as:\nh n k (Z) w l ∈Ω h n l (Z) • log( |I| Ii∈I min(1, q h n k (V i q )) ) (1\n)\nwhere Ω is the set possible n-grams in the corpus, I is the set of corpus images and h n k (Z), h n k (V i j ) represent the number of occurrences of n-gram w k in the sequence Z and in the j-th ground truth of image I i ∈ I. The CIDEr and its alternative (CIDEr-D), compute the similarity between the candidate and reference description as the number of matching n-grams, weighted according to Equation 1. We refer to [30] for additional details of the formula since they are unnecessary for the sake of the discussion.\nThe standard training practice of the Image Captioning model consists of a pre-training phase using the Cross-Entropy loss followed by a CIDer-D optimization by means of a policy gradient method called Self-Critical Sequence Training [27]. The latter minimizes the negative expected reward:\nL R (θ) = -E y 1:T ∼p θ [r(y 1:T )] (2\n)\nwhere r is the CIDEr function, and its gradient is approximated as follows:\n∇ θ L R (θ) ≈ -(r(y s 1:T ) -r(y b 1:T ))∇ θ log p θ (y s 1:T )(3)\nwhere y s 1:T are the sampled captions and y b 1:T are the base predictions."},{"figure_ref":[],"heading":"The End-of-Sequence token in SCST","publication_ref":[],"table_ref":[],"text":"Two properties are desirable in an image description: completeness and correctness. While the first goal is pursued by the reward maximization, the SCST algorithm provides no explicit control over the latter, which is instead implicitly encouraged by the sequentiality of the decoding process. A token predicted at a specific time step also determines the most likely n-grams in the following ones. Since all n-grams are extracted from linguistically correct references, the final description will be correct, at least locally. Unfortunately, the CIDEr score does not consider a sentence's global correctness, and this aspect can be easily exploited by the SCST if not carefully implemented. In particular, the algorithm is allowed to produce incomplete descriptions using trivial sentence fragments that almost certainly match some parts of any set of references. This is the reason why the standard SCST implementation includes the special End-of-Sequence token, abbreviated as , in the definition of the n-grams space. With this precaution, the reward function encourages a correct sentence termination leveraging the fact that the tf-idf of the token out-weights those of function words."},{"figure_ref":["fig_0","fig_0"],"heading":"The problem of the omission","publication_ref":[],"table_ref":[],"text":"The inclusion or exclusion of the token in the SCST algorithm represents a small and easy-to-overlook detail that significantly impacts a captioning system's performance. In case the token is omitted, the descriptions generated by the network are often terminated by trivial sentence fragments such as \"and a\", \"in the\", \"on top of\" and \"in front of\" (more examples in Figure 1). However, despite the presence of artifacts, they achieve superior performances on popular benchmarks compared to the correct ones (Figure 1). In particular, the number of additional points yielded by the artifacts can even be greater than the range of values in which different models developed around the same period typically compete. Therefore, the Image Captioning research field is currently suffering from a lack of transparency and, in some cases lack of awareness over the importance of the token in the SCST. The problem can be described from multiple perspectives:\n-If details over the token in the SCST implementation are unavailable, omitted, or simply overlooked, it becomes difficult to compare models in the literature fairly. -Researchers that are aware of the issue are given the difficult choice between less competitive results and poorly formulated outputs. -Finally, researchers that are not aware of the issue (especially the newcomers in the field of Image Captioning) are indirectly encouraged to adopt the implementations that generate compromised sentences because of their superior performances.\n3 Omission Impact Analysis"},{"figure_ref":[],"heading":"Experimental Setup","publication_ref":["b34","b22","b16","b6"],"table_ref":[],"text":"For the qualitative and quantitative analysis of artifacts we implement1 the Transformer [29] with 3 layers, d model =512 and d f f =2048, trained on the COCO 2014 [17] data set using the Karpathy split [11]. The Faster-RCNN backbone provided by [1] is adopted. The learning procedure consists of a first training step on Cross Entropy loss for 8 epochs followed by the CIDEr-D optimization for 20 epochs. The following configurations are adopted:\n1. batch size of 48, a learning rate of 2e-4 annealed by 0.8 every 2 epochs and warm-up of 10000 in case of Cross Entropy Loss; 2. batch size of 48, a learning rate of 1e-4 annealed by 0.8 every 2 epochs during the SCST.\nOptimization details are provided only for the sake of reproducibility since the artifacts discussed in this work arise regardless of the architecture and optimization details. For the ensemble results, 4 model instances are generated with the aforementioned method differing only in the initialization seed. In the experiments, for each seed, the SCST in the Standard and No configurations optimize the same pre-trained model."},{"figure_ref":["fig_1","fig_1","fig_1"],"heading":"Artifacts Analysis","publication_ref":[],"table_ref":[],"text":"The token can be omitted in two aspects of SCST:\n1. during the reward computation; 2. during the initialization of tf-idfs;\nwhich leads to 4 implementation instances in case sampled descriptions are tokenized consistently with respect to the ground-truths. Prediction w/ ) and (tf-idf Init. w/o , Prediction w/o ) configuration referred as \"Standard\" and \"No\" respectively throughout the rest of this work.\nIn the No configuration, results are affected by 8 classes of artifacts depending on how sequences are terminated, with the last token belonging to A={\"in\", \"a\", \"of\", \"the\", \"with\", \"on\", \"and\" \"*\"}, where \"*\" represents all the possible remaining cases. While all elements in the set A are just simplifications of longer trivial fragments such as \"and a\", \"in a\", \"with a\" and \"in front of\", the case of \"on\" may seem acceptable but the token is often part of uncommon formulations such as \"a beach with a surfboard on\" and \"a street with a bus on\". Nevertheless, \"on\" represents only a small fraction of all instances, which mostly end with the \"a\" token instead (see Figure 2.c).\nFigure 2.a showcases the number of artifacts converging to 50% of the whole testing set as the number of epochs increases. Thus, both correct and compromised sentences are produced by the omission, which means the network learns to inject the fragments following a non-trivial and unpredictable criteria for each sequence.\nFigure 2.b and Table 3.2 showcase that a single model trained with SCST in the No configuration consistently outperforms the standard one across all seeds, often by a large margin, with a maximum gain of +2.8 and +4.3 CIDEr-D in the offline test and validation set respectively. Whereas, by removing the artifacts from the latter predictions we observed the opposite trend with a maximum performance decrease of -2.3 and -2.0. Therefore, the increase in score is mostly due to the artifacts and the omission poses an obstacle to the generation of semantically meaningful content. Similar behaviour is observed for ensemble performances (referred as )."},{"figure_ref":[],"heading":"Literature classification","publication_ref":["b13","b28"],"table_ref":["tab_2"],"text":"We sample recent works in the research literature and classify each of them according to the way SCST is implemented. In Section 3.2 we observed that only half of the evaluated sentences are compromised, which means that if a paper provides only a few correct captioning examples, it is not enough to determine whether the token was omitted or not. Because of that, the classification is made through code inspection. The classes and the respective criteria are defined as follows:\n-Standard: token is included in both SCST initialization and reward computation or complete results on either test or validation set are provided; -No: token is omitted in both initialization and reward computation; -Unknown: the code was not found or it was not available at the time this work was completed.\nTable 3.3 showcases that only 12 of 25 works are confirmed to follow the Standard implementation, 8 fall in the No category and 5 are unknown. The State-of-the-art architectures in 2019 [8] and 2020 [23] achieved 129.6 and 131.4 CIDEr-D scores respectively, which showcases the gradual improvement process of the research activity and provides an example of the magnitude of improvements over the years. Unfortunately, such a difference in performance can be lower than the additional score yielded by artifacts (see Section 3.2). For instance, if AoANet adopted the No configuration, its score would have been comparable to the State-of-the-art performances of the following year (X-Transformer) (see Table 4).\nThe amount of No implementations in the last years confirms the phenomena described in Section 2.3.\nTable 3. SCST classification of recent Image Captioning works and their respective performances on the MS-COCO 2014 task. The offline case reports the CIDEr-D score of a single model in contrast to the online evaluation server results where an ensemble is adopted instead with some exceptions denoted with \" \". "},{"figure_ref":[],"heading":"SacreEOS","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"SacreEOS signature","publication_ref":[],"table_ref":[],"text":"The lack of transparency and awareness over the token in SCST originates from an easy-to-overlook implementation detail. Therefore, the natural solution is to disseminate awareness of the issue. To achieve this goal we introduce SacreEOS, a Python library whose main functionality consists of the generation of signatures that uniquely identify the key aspects of the SCST implementation. In particular, how the token is handled. The sharing of the SacreEOS signature accomplishes three objectives:\n1. it increases transparency and eases the comparison of models; 2. it informs the reader about the presence or absence of artifacts (those related to the omission) in the results; 3. last but not least, it spreads awareness of the problem.\nWe believe this is especially useful in cases of works that do not release the code to the public.\nEstablished researchers and existing implementations can manually generate the signature using the SacreEOS command line interface. The tool simply asks a few questions regarding the technical aspects of SCST, therefore it does not require any code integration. For new projects instead, SacreEOS consists of an SCST implementation helper, in this case, the signature is provided automatically. "},{"figure_ref":[],"heading":"Implementation helper and limitations of the approach","publication_ref":[],"table_ref":[],"text":"In addition to the functionality of signature generation, the SacreEOS library optionally provides helpful classes to ease the implementation of SCST in future projects. In particular, it covers the following aspects:\n-SCST class selection. Given the number of established works implemented in both Standard and No configurations, it is out of the scope of this paper to decide which one is the \"correct\" one (the library provides no default option in this regard). However, the tool helps the user to make informed decisions. Classes are currently defined by the reward metric, the reward base and whether the token is included or omitted in both initialization and reward computation. -SCST initialization. The library initializes the tf-idfs for the reward computation and performs input checks according to the selected class. -SCST reward computation. The library currently supports the following reward functions CIDEr, CIDEr-D, CIDEr-R and BLEU. Results are consistent with the official repositories 2 . Each function is implemented in both Python and C, users can optionally enable the latter version to increase efficiency. -Signature generation. In this case the SacreEOS signature is automatically determined by the class selection and does not require user intervention.\nThe library includes an intricate collection of assertions and input checks on all implementation levels, taylored to each specific class. Nevertheless, the SacreEOS does not prevent misreporting. In case the signature is manually generated, it relies on the user to provide the correct data."},{"figure_ref":[],"heading":"Related works","publication_ref":["b32","b15","b24","b33","b30","b29"],"table_ref":[],"text":"The work of [27] mentioned the role of the End-of-Sequence token. However, it only provided a few qualitative examples and did not report numerical details. Several works in the past focused on improving the evaluation of Image Captioning systems but they mostly proposed alternatives to the CIDEr metric, such as TIGEr [10], SPIDEr [19], and CIDEr-R [28]. None of them addressed the issue discussed in this work.\nThe main inspiration of SacreEOS is SacreBLEU [25], in the field of Machine Translation, where ambiguities can arise from different tokenization and detokenization choices that ultimately affect the BLEU score [24]."},{"figure_ref":[],"heading":"Conclusion","publication_ref":[],"table_ref":[],"text":"Our work discussed the role of in the Self-Critical Sequence Training and how the lack of transparency and awareness over its function pose an obstacle 2 CIDEr, CIDEr-D, BLEU: github.com/vrama91/cider CIDEr-R: github.com/gabrielsantosrv/coco-caption to the scientific progress in the Image Captioning field. We described the source of the problem from a qualitative and quantitative perspective. We classified recent works in the scientific literature according to the SCST configuration to showcase the pervasiveness and the importance of the matter. Finally, we proposed a possible solution that consists of sharing a unique signature with the help of a Python library called SacreEOS, to enable fair model comparisons and spread awareness regarding the issue."}],"string":"[\n {\n \"figure_ref\": [],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b32\"\n ],\n \"table_ref\": [],\n \"text\": \"The standard training strategy of a modern Neural Image Captioning system includes a policy gradient method, called Self-Critical Sequence Training [27] (shortened as SCST) which is designed to maximize the evaluation score given to the outputs. In this work, we discuss the problems caused by the lack of transparency from the research community over the inclusion or omission of the End-of-Sequence token during the optimization. An easy-to-overlook implementation detail that can significantly increase the performance of any model despite yielding worse descriptions.\\nThe lack of awareness of the impact of the End-of-Sequence () omission and the lack of explicit information on the SCST implementation during the reporting of results pose an obstacle to scientific progress as they make it challenging to compare established works and evaluate new ones. Our paper attempts to spread awareness about the issue and proposes a solution to increase transparency in future works. This paper is structured as follows: in Section 2, we discuss the problem of the End-of-Sequence omission and why it is a problem for the research community; in Section 3, we provide a qualitative and quantitative analysis of the issue and we sample some of the recent works in Image Captioning to demonstrate its pervasiveness and provide some practical examples of its impact; In Section 4, we propose a possible solution with the help of a Python library called SacreEOS; in Section 5, we mention some of the literature approaches, and, finally, we draw our conclusions in Section 6.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Problem Description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"CIDEr Optimization\",\n \"publication_ref\": [\n \"b35\",\n \"b29\",\n \"b21\",\n \"b8\",\n \"b7\",\n \"b35\",\n \"b32\"\n ],\n \"table_ref\": [],\n \"text\": \"CIDEr [30] is an n-gram-based metric that evaluates the caption semantic content according to its similarities to the ground truths. Compared to the other metrics [24,16,3,2], it exploits the entire corpus of reference descriptions in the attempt of backing the evaluation with the consensus of the majority of people. In particular, each n-gram w k in sequence Z is weighted according to the tf-idf term g n k (Z) defined as:\\nh n k (Z) w l ∈Ω h n l (Z) • log( |I| Ii∈I min(1, q h n k (V i q )) ) (1\\n)\\nwhere Ω is the set possible n-grams in the corpus, I is the set of corpus images and h n k (Z), h n k (V i j ) represent the number of occurrences of n-gram w k in the sequence Z and in the j-th ground truth of image I i ∈ I. The CIDEr and its alternative (CIDEr-D), compute the similarity between the candidate and reference description as the number of matching n-grams, weighted according to Equation 1. We refer to [30] for additional details of the formula since they are unnecessary for the sake of the discussion.\\nThe standard training practice of the Image Captioning model consists of a pre-training phase using the Cross-Entropy loss followed by a CIDer-D optimization by means of a policy gradient method called Self-Critical Sequence Training [27]. The latter minimizes the negative expected reward:\\nL R (θ) = -E y 1:T ∼p θ [r(y 1:T )] (2\\n)\\nwhere r is the CIDEr function, and its gradient is approximated as follows:\\n∇ θ L R (θ) ≈ -(r(y s 1:T ) -r(y b 1:T ))∇ θ log p θ (y s 1:T )(3)\\nwhere y s 1:T are the sampled captions and y b 1:T are the base predictions.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"The End-of-Sequence token in SCST\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Two properties are desirable in an image description: completeness and correctness. While the first goal is pursued by the reward maximization, the SCST algorithm provides no explicit control over the latter, which is instead implicitly encouraged by the sequentiality of the decoding process. A token predicted at a specific time step also determines the most likely n-grams in the following ones. Since all n-grams are extracted from linguistically correct references, the final description will be correct, at least locally. Unfortunately, the CIDEr score does not consider a sentence's global correctness, and this aspect can be easily exploited by the SCST if not carefully implemented. In particular, the algorithm is allowed to produce incomplete descriptions using trivial sentence fragments that almost certainly match some parts of any set of references. This is the reason why the standard SCST implementation includes the special End-of-Sequence token, abbreviated as , in the definition of the n-grams space. With this precaution, the reward function encourages a correct sentence termination leveraging the fact that the tf-idf of the token out-weights those of function words.\"\n },\n {\n \"figure_ref\": [\n \"fig_0\",\n \"fig_0\"\n ],\n \"heading\": \"The problem of the omission\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The inclusion or exclusion of the token in the SCST algorithm represents a small and easy-to-overlook detail that significantly impacts a captioning system's performance. In case the token is omitted, the descriptions generated by the network are often terminated by trivial sentence fragments such as \\\"and a\\\", \\\"in the\\\", \\\"on top of\\\" and \\\"in front of\\\" (more examples in Figure 1). However, despite the presence of artifacts, they achieve superior performances on popular benchmarks compared to the correct ones (Figure 1). In particular, the number of additional points yielded by the artifacts can even be greater than the range of values in which different models developed around the same period typically compete. Therefore, the Image Captioning research field is currently suffering from a lack of transparency and, in some cases lack of awareness over the importance of the token in the SCST. The problem can be described from multiple perspectives:\\n-If details over the token in the SCST implementation are unavailable, omitted, or simply overlooked, it becomes difficult to compare models in the literature fairly. -Researchers that are aware of the issue are given the difficult choice between less competitive results and poorly formulated outputs. -Finally, researchers that are not aware of the issue (especially the newcomers in the field of Image Captioning) are indirectly encouraged to adopt the implementations that generate compromised sentences because of their superior performances.\\n3 Omission Impact Analysis\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experimental Setup\",\n \"publication_ref\": [\n \"b34\",\n \"b22\",\n \"b16\",\n \"b6\"\n ],\n \"table_ref\": [],\n \"text\": \"For the qualitative and quantitative analysis of artifacts we implement1 the Transformer [29] with 3 layers, d model =512 and d f f =2048, trained on the COCO 2014 [17] data set using the Karpathy split [11]. The Faster-RCNN backbone provided by [1] is adopted. The learning procedure consists of a first training step on Cross Entropy loss for 8 epochs followed by the CIDEr-D optimization for 20 epochs. The following configurations are adopted:\\n1. batch size of 48, a learning rate of 2e-4 annealed by 0.8 every 2 epochs and warm-up of 10000 in case of Cross Entropy Loss; 2. batch size of 48, a learning rate of 1e-4 annealed by 0.8 every 2 epochs during the SCST.\\nOptimization details are provided only for the sake of reproducibility since the artifacts discussed in this work arise regardless of the architecture and optimization details. For the ensemble results, 4 model instances are generated with the aforementioned method differing only in the initialization seed. In the experiments, for each seed, the SCST in the Standard and No configurations optimize the same pre-trained model.\"\n },\n {\n \"figure_ref\": [\n \"fig_1\",\n \"fig_1\",\n \"fig_1\"\n ],\n \"heading\": \"Artifacts Analysis\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The token can be omitted in two aspects of SCST:\\n1. during the reward computation; 2. during the initialization of tf-idfs;\\nwhich leads to 4 implementation instances in case sampled descriptions are tokenized consistently with respect to the ground-truths. Prediction w/ ) and (tf-idf Init. w/o , Prediction w/o ) configuration referred as \\\"Standard\\\" and \\\"No\\\" respectively throughout the rest of this work.\\nIn the No configuration, results are affected by 8 classes of artifacts depending on how sequences are terminated, with the last token belonging to A={\\\"in\\\", \\\"a\\\", \\\"of\\\", \\\"the\\\", \\\"with\\\", \\\"on\\\", \\\"and\\\" \\\"*\\\"}, where \\\"*\\\" represents all the possible remaining cases. While all elements in the set A are just simplifications of longer trivial fragments such as \\\"and a\\\", \\\"in a\\\", \\\"with a\\\" and \\\"in front of\\\", the case of \\\"on\\\" may seem acceptable but the token is often part of uncommon formulations such as \\\"a beach with a surfboard on\\\" and \\\"a street with a bus on\\\". Nevertheless, \\\"on\\\" represents only a small fraction of all instances, which mostly end with the \\\"a\\\" token instead (see Figure 2.c).\\nFigure 2.a showcases the number of artifacts converging to 50% of the whole testing set as the number of epochs increases. Thus, both correct and compromised sentences are produced by the omission, which means the network learns to inject the fragments following a non-trivial and unpredictable criteria for each sequence.\\nFigure 2.b and Table 3.2 showcase that a single model trained with SCST in the No configuration consistently outperforms the standard one across all seeds, often by a large margin, with a maximum gain of +2.8 and +4.3 CIDEr-D in the offline test and validation set respectively. Whereas, by removing the artifacts from the latter predictions we observed the opposite trend with a maximum performance decrease of -2.3 and -2.0. Therefore, the increase in score is mostly due to the artifacts and the omission poses an obstacle to the generation of semantically meaningful content. Similar behaviour is observed for ensemble performances (referred as ).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Literature classification\",\n \"publication_ref\": [\n \"b13\",\n \"b28\"\n ],\n \"table_ref\": [\n \"tab_2\"\n ],\n \"text\": \"We sample recent works in the research literature and classify each of them according to the way SCST is implemented. In Section 3.2 we observed that only half of the evaluated sentences are compromised, which means that if a paper provides only a few correct captioning examples, it is not enough to determine whether the token was omitted or not. Because of that, the classification is made through code inspection. The classes and the respective criteria are defined as follows:\\n-Standard: token is included in both SCST initialization and reward computation or complete results on either test or validation set are provided; -No: token is omitted in both initialization and reward computation; -Unknown: the code was not found or it was not available at the time this work was completed.\\nTable 3.3 showcases that only 12 of 25 works are confirmed to follow the Standard implementation, 8 fall in the No category and 5 are unknown. The State-of-the-art architectures in 2019 [8] and 2020 [23] achieved 129.6 and 131.4 CIDEr-D scores respectively, which showcases the gradual improvement process of the research activity and provides an example of the magnitude of improvements over the years. Unfortunately, such a difference in performance can be lower than the additional score yielded by artifacts (see Section 3.2). For instance, if AoANet adopted the No configuration, its score would have been comparable to the State-of-the-art performances of the following year (X-Transformer) (see Table 4).\\nThe amount of No implementations in the last years confirms the phenomena described in Section 2.3.\\nTable 3. SCST classification of recent Image Captioning works and their respective performances on the MS-COCO 2014 task. The offline case reports the CIDEr-D score of a single model in contrast to the online evaluation server results where an ensemble is adopted instead with some exceptions denoted with \\\" \\\". \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"SacreEOS\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"SacreEOS signature\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The lack of transparency and awareness over the token in SCST originates from an easy-to-overlook implementation detail. Therefore, the natural solution is to disseminate awareness of the issue. To achieve this goal we introduce SacreEOS, a Python library whose main functionality consists of the generation of signatures that uniquely identify the key aspects of the SCST implementation. In particular, how the token is handled. The sharing of the SacreEOS signature accomplishes three objectives:\\n1. it increases transparency and eases the comparison of models; 2. it informs the reader about the presence or absence of artifacts (those related to the omission) in the results; 3. last but not least, it spreads awareness of the problem.\\nWe believe this is especially useful in cases of works that do not release the code to the public.\\nEstablished researchers and existing implementations can manually generate the signature using the SacreEOS command line interface. The tool simply asks a few questions regarding the technical aspects of SCST, therefore it does not require any code integration. For new projects instead, SacreEOS consists of an SCST implementation helper, in this case, the signature is provided automatically. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Implementation helper and limitations of the approach\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In addition to the functionality of signature generation, the SacreEOS library optionally provides helpful classes to ease the implementation of SCST in future projects. In particular, it covers the following aspects:\\n-SCST class selection. Given the number of established works implemented in both Standard and No configurations, it is out of the scope of this paper to decide which one is the \\\"correct\\\" one (the library provides no default option in this regard). However, the tool helps the user to make informed decisions. Classes are currently defined by the reward metric, the reward base and whether the token is included or omitted in both initialization and reward computation. -SCST initialization. The library initializes the tf-idfs for the reward computation and performs input checks according to the selected class. -SCST reward computation. The library currently supports the following reward functions CIDEr, CIDEr-D, CIDEr-R and BLEU. Results are consistent with the official repositories 2 . Each function is implemented in both Python and C, users can optionally enable the latter version to increase efficiency. -Signature generation. In this case the SacreEOS signature is automatically determined by the class selection and does not require user intervention.\\nThe library includes an intricate collection of assertions and input checks on all implementation levels, taylored to each specific class. Nevertheless, the SacreEOS does not prevent misreporting. In case the signature is manually generated, it relies on the user to provide the correct data.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related works\",\n \"publication_ref\": [\n \"b32\",\n \"b15\",\n \"b24\",\n \"b33\",\n \"b30\",\n \"b29\"\n ],\n \"table_ref\": [],\n \"text\": \"The work of [27] mentioned the role of the End-of-Sequence token. However, it only provided a few qualitative examples and did not report numerical details. Several works in the past focused on improving the evaluation of Image Captioning systems but they mostly proposed alternatives to the CIDEr metric, such as TIGEr [10], SPIDEr [19], and CIDEr-R [28]. None of them addressed the issue discussed in this work.\\nThe main inspiration of SacreEOS is SacreBLEU [25], in the field of Machine Translation, where ambiguities can arise from different tokenization and detokenization choices that ultimately affect the BLEU score [24].\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Our work discussed the role of in the Self-Critical Sequence Training and how the lack of transparency and awareness over its function pose an obstacle 2 CIDEr, CIDEr-D, BLEU: github.com/vrama91/cider CIDEr-R: github.com/gabrielsantosrv/coco-caption to the scientific progress in the Image Captioning field. We described the source of the problem from a qualitative and quantitative perspective. We classified recent works in the scientific literature according to the SCST configuration to showcase the pervasiveness and the importance of the matter. Finally, we proposed a possible solution that consists of sharing a unique signature with the help of a Python library called SacreEOS, to enable fair model comparisons and spread awareness regarding the issue.\"\n }\n]"},"abstract":{"kind":"string","value":"The Image Captioning research field is currently compromised by the lack of transparency and awareness over the End-of-Sequence token () in the Self-Critical Sequence Training. If the token is omitted, a model can boost its performance up to +4.1 CIDEr-D using trivial sentence fragments. While this phenomenon poses an obstacle to a fair evaluation and comparison of established works, people involved in new projects are given the arduous choice between lower scores and unsatisfactory descriptions due to the competitive nature of the research. This work proposes to solve the problem by spreading awareness of the issue itself. In particular, we invite future works to share a simple and informative signature with the help of a library called SacreEOS."},"title":{"kind":"string","value":"A request for clarity over the End of Sequence token in the Self-Critical Sequence Training"},"figures":{"kind":"list like","value":[{"figure_caption":"Fig. 1 .1Fig. 1. Captions generated by the same model (the Transformer [29]) trained with different implementations of SCST on the MS-COCO [17] data set. (Left) The model is optimized by the standard SCST and achieves 125.8 CIDEr-D on the validation set. (Right) The model is optimized by an implementation of SCST in which the token is omitted and achieves 130.1 CIDEr-D on the validation set.","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"Fig. 2 .2Fig. 2. a) The number of artifacts in the No configuration on 5000 test set predictions. b) Average CIDEr-D score of 4 training instances (different seeds) in the Standard and No configuration, \"Cleaned\" denotes the No performance in case artifacts are removed before the evaluation. c) Artifacts distribution. Sequences terminated by \"a\" account for 89.8% of all cases (top). Histogram of sequences terminated by \"a\" (bottom).","figure_data":"","figure_id":"fig_1","figure_label":"2","figure_type":"figure"},{"figure_caption":"Format and signature examples are the following: Format: _+++ Examples: STANDARD_w/oInit+Cider-D[n5,s6.0]+average[nspi5]+1.0.0 NOMODE_wInit+Cider-D[n4,s6.0]+greedy[nspi5]+1.0.0 NOMODE_w/oInit+BLEU[n4]+average[nspi5]+1.0.0","figure_data":"","figure_id":"fig_2","figure_label":"","figure_type":"figure"},{"figure_caption":"Table1reports the impact of each configuration over the final descriptions. Two cases are the focus of this work since most popular implementations fall into the (tf-idf Init. w/ , Impact of the token in SCST over the final CIDEr-D score and outputs. \"tf-idf Init.\" refers to the ground truth sentences involved in the calculation of document frequencies, and \"Predictions\" refers to the sampled predictions and respective references.","figure_data":"tf-idf Init.tf-idf Init.w/ w/o Reward baseline score lower scorew/ no artifacts with artifactsRewardlower score higher scorew/o with artifacts with artifactsKarpathy test splitKarpathy validation splitStandardNo (ε) / δCleaned / δ StandardNo (ε) / δCleaned / δSeed 1 128.4 131.2 (48.3%) / +2.8 127.8 / -0.6 125.8 130.1 (47.5%) / +4.3 126.4 / +0.6Seed 2 129.0 130.9 (49.3%) / +1.9 127.4 / -1.6 127.0 129.9 (48.1%) / +2.9 126.2 / -0.8Seed 3 129.0 131.0 (50.3%) / +2.0 127.5 / -1.5 127.2 129.3 (47.6%) / +2.1 125.7 / -1.5Seed 4 129.1 130.7 (50.4%) / +1.6 126.8 / -2.3 128.0 130.0 (50.6%) / +2.0 126.0 / -2.0Avg128.9 130.9 (49.6%) / +2.0 127.3 / -1.1 126.9 129.8 (48.6%) / +2.8 126.0 / -0.9133.0 134.9 (50.2%) / +1.9 131.2 / -1.8 131.8 133.8 (49.5%) / +2.0 129.8 / -2.0","figure_id":"tab_0","figure_label":"1","figure_type":"table"},{"figure_caption":"Performance comparison the CIDEr-D optimization in Standard and No training. \"Cleaned\" refers to the No results but artifacts are removed prior to the evaluation.refers to the ensemble of the four models and ε represents the percentage of artifacts.","figure_data":"","figure_id":"tab_1","figure_label":"2","figure_type":"table"},{"figure_caption":"CIDEr-D performance increase observed in open source projects when the SCST configuration is changed from Standard into No mode. Training details can be found in the respective works or repositories.","figure_data":"a Prefix https://github.com/","figure_id":"tab_2","figure_label":"4","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Fig. 1 .1Fig. 1. Captions generated by the same model (the Transformer [29]) trained with different implementations of SCST on the MS-COCO [17] data set. (Left) The model is optimized by the standard SCST and achieves 125.8 CIDEr-D on the validation set. (Right) The model is optimized by an implementation of SCST in which the token is omitted and achieves 130.1 CIDEr-D on the validation set.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 2 .2Fig. 2. a) The number of artifacts in the No configuration on 5000 test set predictions. b) Average CIDEr-D score of 4 training instances (different seeds) in the Standard and No configuration, \\\"Cleaned\\\" denotes the No performance in case artifacts are removed before the evaluation. c) Artifacts distribution. Sequences terminated by \\\"a\\\" account for 89.8% of all cases (top). Histogram of sequences terminated by \\\"a\\\" (bottom).\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Format and signature examples are the following: Format: _+++ Examples: STANDARD_w/oInit+Cider-D[n5,s6.0]+average[nspi5]+1.0.0 NOMODE_wInit+Cider-D[n4,s6.0]+greedy[nspi5]+1.0.0 NOMODE_w/oInit+BLEU[n4]+average[nspi5]+1.0.0\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Table1reports the impact of each configuration over the final descriptions. Two cases are the focus of this work since most popular implementations fall into the (tf-idf Init. w/ , Impact of the token in SCST over the final CIDEr-D score and outputs. \\\"tf-idf Init.\\\" refers to the ground truth sentences involved in the calculation of document frequencies, and \\\"Predictions\\\" refers to the sampled predictions and respective references.\",\n \"figure_data\": \"tf-idf Init.tf-idf Init.w/ w/o Reward baseline score lower scorew/ no artifacts with artifactsRewardlower score higher scorew/o with artifacts with artifactsKarpathy test splitKarpathy validation splitStandardNo (ε) / δCleaned / δ StandardNo (ε) / δCleaned / δSeed 1 128.4 131.2 (48.3%) / +2.8 127.8 / -0.6 125.8 130.1 (47.5%) / +4.3 126.4 / +0.6Seed 2 129.0 130.9 (49.3%) / +1.9 127.4 / -1.6 127.0 129.9 (48.1%) / +2.9 126.2 / -0.8Seed 3 129.0 131.0 (50.3%) / +2.0 127.5 / -1.5 127.2 129.3 (47.6%) / +2.1 125.7 / -1.5Seed 4 129.1 130.7 (50.4%) / +1.6 126.8 / -2.3 128.0 130.0 (50.6%) / +2.0 126.0 / -2.0Avg128.9 130.9 (49.6%) / +2.0 127.3 / -1.1 126.9 129.8 (48.6%) / +2.8 126.0 / -0.9133.0 134.9 (50.2%) / +1.9 131.2 / -1.8 131.8 133.8 (49.5%) / +2.0 129.8 / -2.0\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Performance comparison the CIDEr-D optimization in Standard and No training. \\\"Cleaned\\\" refers to the No results but artifacts are removed prior to the evaluation.refers to the ensemble of the four models and ε represents the percentage of artifacts.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"CIDEr-D performance increase observed in open source projects when the SCST configuration is changed from Standard into No mode. Training details can be found in the respective works or repositories.\",\n \"figure_data\": \"a Prefix https://github.com/\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n }\n]"},"authors":{"kind":"string","value":"Jia Cheng Hu; Roberto Cavicchioli; Alessandro Capotondi"},"references":{"kind":"list like","value":[{"authors":"","journal":"","ref_id":"b0","title":"Year Work Offline Online SCST Code inspection a (commit) 2018 GCN-LSTM","year":""},{"authors":"","journal":"","ref_id":"b1","title":"Standard ruotianluo/self-critical","year":""},{"authors":"","journal":"Standard LuoweiZhou/VLP","ref_id":"b2","title":"d85)","year":"2021"},{"authors":"","journal":"Standard jchenghu/ExpansionNet_v","ref_id":"b3","title":"d130) 2022 BLIP [15] 136.7","year":""},{"authors":"","journal":"","ref_id":"b4","title":"Standard OFA-Sys/OFA (1809b55","year":"2022"},{"authors":"","journal":"","ref_id":"b5","title":"GT-RIPL/Xmodal-Ctx (d927eec) Bibliography","year":""},{"authors":"Peter Anderson","journal":"","ref_id":"b6","title":"Bottom-up and top-down attention for image captioning and visual question answering","year":"2018"},{"authors":"Peter Anderson","journal":"Springer","ref_id":"b7","title":"Spice: Semantic propositional image caption evaluation","year":"2016"},{"authors":"Satanjeev Banerjee; Alon Lavie","journal":"","ref_id":"b8","title":"METEOR: An automatic metric for MT evaluation with improved correlation with human judgments","year":"2005"},{"authors":"Manuele Barraco","journal":"","ref_id":"b9","title":"CaMEL: Mean Teacher Learning for Image Captioning","year":"2022"},{"authors":"Marcella Cornia","journal":"","ref_id":"b10","title":"Meshed-memory transformer for image captioning","year":"2020"},{"authors":"Simao Herdade","journal":"Advances in Neural Information Processing Systems","ref_id":"b11","title":"Image captioning: Transforming objects into words","year":"2019"},{"authors":"Jia Cheng Hu; Roberto Cavicchioli; Alessandro Capotondi","journal":"","ref_id":"b12","title":"Expan-sionNet v2: Block Static Expansion in fast end to end training for Image Captioning","year":"2022"},{"authors":"Lun Huang","journal":"","ref_id":"b13","title":"Attention on attention for image captioning","year":"2019"},{"authors":"Jiayi Ji","journal":"","ref_id":"b14","title":"Improving image captioning by leveraging intra-and interlayer global representation in transformer network","year":"2021"},{"authors":"Ming Jiang","journal":"","ref_id":"b15","title":"Tiger: Text-to-image grounding for image caption evaluation","year":"2019"},{"authors":"Andrej Karpathy; Li Fei-Fei","journal":"","ref_id":"b16","title":"Deep visual-semantic alignments for generating image descriptions","year":"2015"},{"authors":"Lei Ke","journal":"","ref_id":"b17","title":"Reflective decoding network for image captioning","year":"2019"},{"authors":"Chia-Wen Kuo; Zsolt Kira","journal":"","ref_id":"b18","title":"Beyond a Pre-Trained Object Detector: Cross-Modal Textual and Visual Context for Image Captioning","year":"2022"},{"authors":"Jingyu Li","journal":"","ref_id":"b19","title":"ER-SAN: Enhanced-Adaptive Relation Self-Attention Network for Image Captioning","year":"2022-07"},{"authors":"Junnan Li","journal":"","ref_id":"b20","title":"Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation","year":"2022"},{"authors":"Chin-Yew Lin","journal":"","ref_id":"b21","title":"Rouge: A package for automatic evaluation of summaries","year":"2004"},{"authors":"Tsung-Yi Lin","journal":"Springer","ref_id":"b22","title":"Microsoft coco: Common objects in context","year":"2014"},{"authors":"Bing Liu","journal":"","ref_id":"b23","title":"Show, Deconfound and Tell: Image Captioning With Causal Inference","year":"2022"},{"authors":"Siqi Liu","journal":"","ref_id":"b24","title":"Improved image captioning via policy gradient optimization of spider","year":"2017"},{"authors":"Ruotian Luo","journal":"","ref_id":"b25","title":"A better variant of self-critical sequence training","year":"2020"},{"authors":"Yunpeng Luo","journal":"","ref_id":"b26","title":"Dual-level collaborative transformer for image captioning","year":"2021"},{"authors":"Masanori Van-Quang Nguyen; Takayuki Suganuma; Okatani","journal":"","ref_id":"b27","title":"GRIT: Faster and Better Image captioning Transformer Using Dual Visual Features","year":"2022"},{"authors":"Yingwei Pan","journal":"","ref_id":"b28","title":"X-Linear Attention Networks for Image Captioning","year":"2020"},{"authors":"Kishore Papineni","journal":"","ref_id":"b29","title":"Bleu: a method for automatic evaluation of machine translation","year":"2002"},{"authors":"Matt Post","journal":"","ref_id":"b30","title":"A call for clarity in reporting BLEU scores","year":"2018"},{"authors":"Yu Qin","journal":"","ref_id":"b31","title":"Look back and predict forward in image captioning","year":"2019"},{"authors":"Rennie Steven","journal":"","ref_id":"b32","title":"Self-critical sequence training for image captioning","year":"2017"},{"authors":"Gabriel Oliveira Dos Santos; Esther Luna Colombini; Sandra Avila","journal":"","ref_id":"b33","title":"Cider-r: Robust consensus-based image description evaluation","year":"2021"},{"authors":"Ashish Vaswani","journal":"","ref_id":"b34","title":"Attention is all you need","year":"2017"},{"authors":"Ramakrishna Vedantam; Lawrence Zitnick; Devi Parikh","journal":"","ref_id":"b35","title":"Cider: Consensus-based image description evaluation","year":"2015"},{"authors":"Peng Wang","journal":"PMLR","ref_id":"b36","title":"Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework","year":"2022"},{"authors":"Weixuan Wang; Zhihong Chen; Haifeng Hu","journal":"","ref_id":"b37","title":"Hierarchical attention network for image captioning","year":"2019"},{"authors":"Yiyu Wang; Jungang Xu; Yingfei Sun","journal":"","ref_id":"b38","title":"End-to-End Transformer Based Model for Image Captioning","year":"2022"},{"authors":"Yang Xu","journal":"","ref_id":"b39","title":"Auto-encoding scene graphs for image captioning","year":"2019"},{"authors":"Ting Yao","journal":"","ref_id":"b40","title":"Exploring visual relationship for image captioning","year":"2018"},{"authors":"Pengpeng Zeng","journal":"","ref_id":"b41","title":"S2 Transformer for Image Captioning","year":""},{"authors":"","journal":"","ref_id":"b42","title":"Main Track. International Joint Conferences on Artificial Intelligence Organization","year":"2022-07"},{"authors":"Xuying Zhang","journal":"","ref_id":"b43","title":"RSTNet: Captioning with adaptive attention on visual and non-visual words","year":"2021"},{"authors":"Luowei Zhou","journal":"","ref_id":"b44","title":"Unified vision-language pre-training for image captioning and vqa","year":"2020"}],"string":"[\n {\n \"authors\": \"\",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"Year Work Offline Online SCST Code inspection a (commit) 2018 GCN-LSTM\",\n \"year\": \"\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"\",\n \"ref_id\": \"b1\",\n \"title\": \"Standard ruotianluo/self-critical\",\n \"year\": \"\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"Standard LuoweiZhou/VLP\",\n \"ref_id\": \"b2\",\n \"title\": \"d85)\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"Standard jchenghu/ExpansionNet_v\",\n \"ref_id\": \"b3\",\n \"title\": \"d130) 2022 BLIP [15] 136.7\",\n \"year\": \"\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"Standard OFA-Sys/OFA (1809b55\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"\",\n \"ref_id\": \"b5\",\n \"title\": \"GT-RIPL/Xmodal-Ctx (d927eec) Bibliography\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Peter Anderson\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"Bottom-up and top-down attention for image captioning and visual question answering\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Peter Anderson\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b7\",\n \"title\": \"Spice: Semantic propositional image caption evaluation\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Satanjeev Banerjee; Alon Lavie\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"METEOR: An automatic metric for MT evaluation with improved correlation with human judgments\",\n \"year\": \"2005\"\n },\n {\n \"authors\": \"Manuele Barraco\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"CaMEL: Mean Teacher Learning for Image Captioning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Marcella Cornia\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"Meshed-memory transformer for image captioning\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Simao Herdade\",\n \"journal\": \"Advances in Neural Information Processing Systems\",\n \"ref_id\": \"b11\",\n \"title\": \"Image captioning: Transforming objects into words\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Jia Cheng Hu; Roberto Cavicchioli; Alessandro Capotondi\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"Expan-sionNet v2: Block Static Expansion in fast end to end training for Image Captioning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Lun Huang\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"Attention on attention for image captioning\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Jiayi Ji\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"Improving image captioning by leveraging intra-and interlayer global representation in transformer network\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Ming Jiang\",\n \"journal\": \"\",\n \"ref_id\": \"b15\",\n \"title\": \"Tiger: Text-to-image grounding for image caption evaluation\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Andrej Karpathy; Li Fei-Fei\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"Deep visual-semantic alignments for generating image descriptions\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Lei Ke\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"Reflective decoding network for image captioning\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Chia-Wen Kuo; Zsolt Kira\",\n \"journal\": \"\",\n \"ref_id\": \"b18\",\n \"title\": \"Beyond a Pre-Trained Object Detector: Cross-Modal Textual and Visual Context for Image Captioning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jingyu Li\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"ER-SAN: Enhanced-Adaptive Relation Self-Attention Network for Image Captioning\",\n \"year\": \"2022-07\"\n },\n {\n \"authors\": \"Junnan Li\",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Chin-Yew Lin\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"Rouge: A package for automatic evaluation of summaries\",\n \"year\": \"2004\"\n },\n {\n \"authors\": \"Tsung-Yi Lin\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b22\",\n \"title\": \"Microsoft coco: Common objects in context\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"Bing Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"Show, Deconfound and Tell: Image Captioning With Causal Inference\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Siqi Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Improved image captioning via policy gradient optimization of spider\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Ruotian Luo\",\n \"journal\": \"\",\n \"ref_id\": \"b25\",\n \"title\": \"A better variant of self-critical sequence training\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Yunpeng Luo\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"Dual-level collaborative transformer for image captioning\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Masanori Van-Quang Nguyen; Takayuki Suganuma; Okatani\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"GRIT: Faster and Better Image captioning Transformer Using Dual Visual Features\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Yingwei Pan\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"X-Linear Attention Networks for Image Captioning\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Kishore Papineni\",\n \"journal\": \"\",\n \"ref_id\": \"b29\",\n \"title\": \"Bleu: a method for automatic evaluation of machine translation\",\n \"year\": \"2002\"\n },\n {\n \"authors\": \"Matt Post\",\n \"journal\": \"\",\n \"ref_id\": \"b30\",\n \"title\": \"A call for clarity in reporting BLEU scores\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Yu Qin\",\n \"journal\": \"\",\n \"ref_id\": \"b31\",\n \"title\": \"Look back and predict forward in image captioning\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Rennie Steven\",\n \"journal\": \"\",\n \"ref_id\": \"b32\",\n \"title\": \"Self-critical sequence training for image captioning\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Gabriel Oliveira Dos Santos; Esther Luna Colombini; Sandra Avila\",\n \"journal\": \"\",\n \"ref_id\": \"b33\",\n \"title\": \"Cider-r: Robust consensus-based image description evaluation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Ashish Vaswani\",\n \"journal\": \"\",\n \"ref_id\": \"b34\",\n \"title\": \"Attention is all you need\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Ramakrishna Vedantam; Lawrence Zitnick; Devi Parikh\",\n \"journal\": \"\",\n \"ref_id\": \"b35\",\n \"title\": \"Cider: Consensus-based image description evaluation\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Peng Wang\",\n \"journal\": \"PMLR\",\n \"ref_id\": \"b36\",\n \"title\": \"Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Weixuan Wang; Zhihong Chen; Haifeng Hu\",\n \"journal\": \"\",\n \"ref_id\": \"b37\",\n \"title\": \"Hierarchical attention network for image captioning\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Yiyu Wang; Jungang Xu; Yingfei Sun\",\n \"journal\": \"\",\n \"ref_id\": \"b38\",\n \"title\": \"End-to-End Transformer Based Model for Image Captioning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Yang Xu\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"Auto-encoding scene graphs for image captioning\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Ting Yao\",\n \"journal\": \"\",\n \"ref_id\": \"b40\",\n \"title\": \"Exploring visual relationship for image captioning\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Pengpeng Zeng\",\n \"journal\": \"\",\n \"ref_id\": \"b41\",\n \"title\": \"S2 Transformer for Image Captioning\",\n \"year\": \"\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"\",\n \"ref_id\": \"b42\",\n \"title\": \"Main Track. International Joint Conferences on Artificial Intelligence Organization\",\n \"year\": \"2022-07\"\n },\n {\n \"authors\": \"Xuying Zhang\",\n \"journal\": \"\",\n \"ref_id\": \"b43\",\n \"title\": \"RSTNet: Captioning with adaptive attention on visual and non-visual words\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Luowei Zhou\",\n \"journal\": \"\",\n \"ref_id\": \"b44\",\n \"title\": \"Unified vision-language pre-training for image captioning and vqa\",\n \"year\": \"2020\"\n }\n]"},"formulas":{"kind":"list like","value":[{"formula_coordinates":[2,221.17,240.39,255.18,32.89],"formula_id":"formula_0","formula_text":"h n k (Z) w l ∈Ω h n l (Z) • log( |I| Ii∈I min(1, q h n k (V i q )) ) (1"},{"formula_coordinates":[2,476.35,248.65,4.24,8.8],"formula_id":"formula_1","formula_text":")"},{"formula_coordinates":[2,247.76,424.29,228.59,10.38],"formula_id":"formula_2","formula_text":"L R (θ) = -E y 1:T ∼p θ [r(y 1:T )] (2"},{"formula_coordinates":[2,476.35,424.29,4.24,8.8],"formula_id":"formula_3","formula_text":")"},{"formula_coordinates":[2,205.93,464.65,274.66,12.69],"formula_id":"formula_4","formula_text":"∇ θ L R (θ) ≈ -(r(y s 1:T ) -r(y b 1:T ))∇ θ log p θ (y s 1:T )(3)"}],"string":"[\n {\n \"formula_coordinates\": [\n 2,\n 221.17,\n 240.39,\n 255.18,\n 32.89\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"h n k (Z) w l ∈Ω h n l (Z) • log( |I| Ii∈I min(1, q h n k (V i q )) ) (1\"\n },\n {\n \"formula_coordinates\": [\n 2,\n 476.35,\n 248.65,\n 4.24,\n 8.8\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \")\"\n },\n {\n \"formula_coordinates\": [\n 2,\n 247.76,\n 424.29,\n 228.59,\n 10.38\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"L R (θ) = -E y 1:T ∼p θ [r(y 1:T )] (2\"\n },\n {\n \"formula_coordinates\": [\n 2,\n 476.35,\n 424.29,\n 4.24,\n 8.8\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \")\"\n },\n {\n \"formula_coordinates\": [\n 2,\n 205.93,\n 464.65,\n 274.66,\n 12.69\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"∇ θ L R (θ) ≈ -(r(y s 1:T ) -r(y b 1:T ))∇ θ log p θ (y s 1:T )(3)\"\n }\n]"}}},{"rowIdx":386,"cells":{"doi":{"kind":"string","value":""},"pub_date":{"kind":"string","value":"2023-05-25"},"sections":{"kind":"list like","value":[{"figure_ref":["fig_1","fig_2"],"heading":"Introduction","publication_ref":["b32","b2","b42","b1","b5","b19","b6","b29","b30","b14","b13","b31","b13","b41","b20","b12","b41","b9","b16","b35"],"table_ref":["tab_0"],"text":"Current neural machine translation (NMT) has achieved great triumph (Sutskever et al., 2014;Bahdanau et al., 2015;Zhu et al., 2020), however in the cost of creating large-scale parallel sentences, which obstructs the development of NMT for the minor languages. Unsupervised NMT (UMT) has thus been proposed to relieve the reliance of parallel corpora (Artetxe et al., 2018;Chen et al., 2018). The core idea of UMT is to align the representation spaces between two languages with alternative pivot signals rather than parallel sentences, such as bilingual lexicons (Lample et al., 2018), multilingual language models (LM) (Conneau and Lample, 2019) and back-translation technique (Sennrich et al., 2016). Recent trends have considered the incorporation of visual information, i.e., multimodal machine translation (MMT) (Specia et al., 2016;Huang et al., 2016). Intuitively, visual modality can serve as language-agnostic signals, pivoting different languages by grounding the same textual semantics into the common visual space. Therefore, solving UMT with visual contents as pivot becomes a promising solution, a.k.a., unsupervised MMT (UMMT) (Huang et al., 2020;Su et al., 2019).\nUMMT systems are trained with only the textimage pairs (), which can be easier to collect than the parallel source-target sentence pairs () (Huang et al., 2020). Although exempting the parallel sentences for training, UMMT still requires such text-image pairs as inputs for testing. Yet such assumption might be unrealistic, because in most of the real-world scenarios such as online translation systems, paired images are not available during inference. Especially for some scarce languages, the pairs have difficult access. In other words, practical UMMT systems should not only avoid the parallel sentences during training, but also the text-image pairs during inference. As summarized in Table 1, although some existing MMT researches exempt the testing-time visual inputs (Zhang et al., 2020;Li et al., 2022), they all unfortunately are supervised methods, relying on large-scale parallel sentences for training. As emphasized above, the visual information is vital to UMMT. However, for both the existing supervised and unsupervised MMT studies, they may suffer from ineffective and insufficient modeling of visual pivot features. For example, most of MMT models perform vision-language (VL) grounding over the whole image and text (Huang et al., 2019;Zhang et al., 2020), where such coarse-grained representation learning can cause mismatching and sacrifice the subtle VL semantics. Fang and Feng (2022) recently introduce a fine-grained VL alignment learning via phrase-level grounding, while without a holistic understanding of the visual scene, such local-level method may lead to incomplete or missing alignments.\nIn this work, we present a novel UMMT method that solves all aforementioned challenges. First of all, to better represent the visual (also the textual) inputs, we consider incorporating the visual scene graph (VSG) (Johnson et al., 2015) and language scene graph (LSG) (Wang et al., 2018). The scene graphs (SG) advance in intrinsically depicting the semantic structures of texts or images with rich details (cf. Fig. 1), which offers a holistic viewpoint for more effective pivoting learning. Then, we build the UMMT framework as illustrated in Fig. 2. The input src text and paired image are first transformed into LSG and VSG, which are further fused into a mixed SG, and then translated into the tgt-side LSG. And the tgt sentence will be finally produced conditioned on the tgt LSG. Several SGbased pivoting learning strategies are proposed for unsupervised training of UMMT system. In addition, to support pure-text (image-free) input during inference, we devise a novel visual scene hallucination module, which dynamically generates a hallucinated VSG from the LSG compensatively. Our system is evaluated on the standard MMT Multi30K and NMT WMT data. Extensive experimental results verify that the proposed method outperforms strong baselines on unsupervised multimodal translation by above 5 BLEU score on average. We further reveal the efficacy of the visual scene hallucination mechanism in relieving the reliance on image inputs during inference. Our SG-pivoting based UMMT helps yield translations with higher completeness, relevance and fluency, and especially obtains improvements on the longer sentences.\nOverall, we make the following contributions: ▶ 1) We are the first to study the inferencetime image-free unsupervised multimodal machine translation, solved with a novel visual scene hallucination mechanism. ▶ 2) We leverage the SGs to better represent the visual and language inputs. Moreover, we design SG-based graph pivoting learning strategies for UMMT training. ▶ 3) Our model achieves huge boosts over strong baselines on benchmark data. Code is available at https: //github.com/scofield7419/UMMT-VSH."},{"figure_ref":[],"heading":"Related Work","publication_ref":["b32","b2","b23","b33","b7","b30","b14","b29","b21","b19","b6","b25","b5","b31","b13","b18","b35","b16","b15","b39","b41","b9"],"table_ref":[],"text":"Neural machine translation has achieved notable development in the era of deep learning (Sutskever et al., 2014;Bahdanau et al., 2015;Luong et al., 2015). The constructions of powerful neural models and training paradigms as well as the collection of large-scale parallel corpora are the driving forces to NMT's success (Vaswani et al., 2017;Devlin et al., 2019). The key of NMT is to learn a good mapping between two (or more) languages. In recent years, visual information has been intro-duced for stronger NMT (i.e., multimodal machine translation), by enhancing the alignments of language latent spaces with visual grounding (Specia et al., 2016;Huang et al., 2016). Intuitively, people speaking different languages can actually refer to the same physical visual contents and conceptions.\nUnsupervised machine translation aims to learn cross-lingual mapping without the use of largescale parallel corpora. The setting is practically meaningful to those minor languages with hard data accessibility. The basic idea is to leverage alternative pivoting contents to compensate the parallel signals based on the back-translation method (Sennrich et al., 2016), such as third-languages (Li et al., 2020), bilingual lexicons (Lample et al., 2018) or multilingual LM (Conneau and Lample, 2019). The visual information can also serve as pivot signals for UMT, i.e., unsupervised multimodal machine translation. Comparing to the standard MMT that trains with triples, UMMT takes as input only the . So far, few studies have explored the UMMT setting, most of which try to enhance the back-translation with multimodal alignment mechanism (Nakayama and Nishida, 2017;Chen et al., 2018;Su et al., 2019;Huang et al., 2020).\nScene graph describes a scene of an image or text into a structure layout, by connecting discrete objects with attributes and with other objects via pairwise relations (Krishna et al., 2017;Wang et al., 2018). As the SGs carry rich contextual and semantic information, they are widely integrated into downstream tasks for enhancements, e.g., image retrieval (Johnson et al., 2015), image generation (Johnson et al., 2018) and image captioning (Yang et al., 2019). This work inherits wisdom, incorporating both the visual scene graph and language scene graph as pivots for UMMT.\nAll the UMMT researches assume that the pairs are required during inference, yet we notice that this can be actually unrealistic. We thus propose a visual hallucination mechanism, achieving the inference-time image-free goal. There are relevant studies on supervised MMT that manage to avoid image inputs (with text only) during inference. The visual retrieval-base methods (Zhang et al., 2020;Fang and Feng, 2022) sentence. Differently, we consider generating the visual scene graph with richer and holistic visual structural information.\n3 Scene Graph-based Translation System"},{"figure_ref":[],"heading":"Problem Definition","publication_ref":[],"table_ref":[],"text":"In UMMT, no parallel translation pairs are available. This work considers an inference-time imagefree UMMT. During training, the data availability is ∈ and the corresponding srcside LSG x and VSG, where X are the src-side sentences, and Z are the paired images. During inference, the model generates tgt-side sentences y ∈ Y based on the inputs of only x ∈ X and the corresponding LSG x , while the visual scene VSG ′ is hallucinated from LSG x . In both training and inference, y will be generated from the intermediate tgt-side language scene graph LSG y , which is produced from LSG x and VSG (or VSG ′ )."},{"figure_ref":["fig_2"],"heading":"Framework","publication_ref":[],"table_ref":[],"text":"As shown in Fig. 2, the system first represents the src-side LSG x and VSG features with two GCN graph encoders, respectively. Then the SG fus-ing&mapping module integrates and transforms two SG representations into a unified one as tgtside LSG, i.e., LSG y . Another GSN model further encodes the LSG y , where the representations are used to generate the tgt sentence (i.e., translation)."},{"figure_ref":["fig_3"],"heading":"Scene Graph Generating and Encoding","publication_ref":["b24"],"table_ref":[],"text":"We first employ two off-the-shelf SG parsers to obtain the LSG and VSG, separately (detailed in the experiment part). For simplicity, here we unify the notations of LSG and VSG as SG. We denote a SG as G=(V, E), where V are the nodes (including object o, attribute a and relation r types), and E are the edges e i,j between any pair of nodes v i ∈ V .\nWe then encode both the VSG and LSG with two spatial Graph Convolution Networks (GCN) (Marcheggiani and Titov, 2017) respectively, which is formulated as:\nr 1 , • • • , r n = GCN(G) ,(1)\nwhere r i is the representation of node v i . We here denote r L i as LSG's node representation, and r V i as VSG's node representation.\nVisual Scene Hallucinating During inference, the visual scene hallucination (VSH) module is activated to perform two-step inference to generate the hallucinated VSG ′ , as illustrated in Fig. 3.\nStep1: sketching skeleton aims to build the skeleton VSG. We copy all the nodes from the raw LSG x to the target VSG, and transform the textual entity nodes into the visual object nodes.\nStep2: completing vision aims to enrich and augment the skeleton VSG into a more realistic one. It is indispensable to add new nodes and edges in the skeleton VSG, since in real scenarios, visual scenes are much more concrete and vivid than textual scenes. Specifically, we develop a node augmentor and a relation augmentor, where the former decides whether to attach a new node to an existing one, and the later decides whether to create an edge between two disjoint nodes. To ensure the fidelity of the hallucinated VSG ′ , during training, the node augmentor and relation augmentor will be updated (i.e., with the learning target L VSH ) with the input LSG and VSG supervisions. Appendix §A.1 details the VSH module.\nSG Fusing&Mapping Now we fuse the heterogeneous LSG x and VSG into one unified scene graph with a mixed view. The key idea is to merge the information from two SGs serving similar roles.\nIn particular, we first measure the representation similarity of each pair of nodes from two GCNs. For those pairs with high alignment scores, we merge them as one by averaging their representations, and for those not, we take the union structures from two SGs. This results in a pseudo tgt-side LSG y . We then use another GCN model for further representation propagation. Finally, we employ a graph-to-text generator to transform the LSG y representations to the tgt sentence y. Appendix §A.2 presents all the technical details in this part."},{"figure_ref":["fig_4"],"heading":"Learning with Scene Graph Pivoting","publication_ref":[],"table_ref":[],"text":"In this part, based on the SG pivot we introduce several learning strategies to accomplish the unsupervised training of machine translation. We mainly consider 1) cross-SG visual-language learning, and 2) SG-pivoted back-translation training. Fig. 4 illustrates these learning strategies."},{"figure_ref":[],"heading":"Cross-SG Visual-language Learning","publication_ref":["b22","b38","b11","b15"],"table_ref":[],"text":"The visual-language SG cross-learning aims to enhance the structural correspondence between the LSG and VSG. Via cross-learning we also teach the SG encoders to automatically learn to highlight those shared visual-language information while deactivating those trivial substructures, i.e., denoising.\nCross-modal SG Aligning The idea is to encourage the text and visual nodes that serve a similar role in VSG and LSG to be closer. To align the fine-grained structures between SGs, we adopt the contrastive learning (CL) technique (Logeswaran and Lee, 2018;Yan et al., 2021;Fei et al., 2022;Huang et al., 2022). In particular, CL learns effec-tive representation by pulling semantically close content pairs together, while pushing apart those different ones. Technically, we measure the similarities between pairs of nodes from two VSG and LSG:\nsi,j = (r L i ) T • r V j ||r L i || ||r V j || .(2)\nA threshold value α is pre-defined to decide the alignment confidence, i.e., pairs with s i,j > α are considered similar. Then we put on the CL loss:\nL CMA = - i∈LSG x , j * ∈VSG log exp(si,j * /τ ) Z ,(3)\nZ = i∈LSG x , k∈VSG, k̸ =j * exp(s i,k /τ ) ,(4)\nwhere τ >0 is an annealing factor. j * means a positive pair with i, i.e., s i,j * >α.\nCross-modal Cross-reconstruction We further strengthen the correspondence between VSG and LSG via cross-modal cross-reconstruction. Specifically, we try to reconstruct the input sentence from the VSG, and the image representations from the LSG. In this way we force both two SGs to focus on the VL-shared parts. To realize VSG→x we employ the aforementioned graph-to-text generator.\nFor LSG→z, we use the graph-to-image generator (Johnson et al., 2018). The learning loss can be marked as L REC ."},{"figure_ref":[],"heading":"SG-pivoted Back-translation Training","publication_ref":["b29","b13"],"table_ref":[],"text":"Back-translation is a key method to realize unsupervised machine translation (Sennrich et al., 2016).\nIn this work, we further aid the back-translation with structural SG pivoting.\nVisual-concomitant Back-translation We perform the back-translation with the SG pivoting. We denote the X →Y translation direction as y=F xz→y (x, z), and Y→Z as x=F yz→x (y, z).\nAs we only have src-side sentences, the backtranslation is uni-directional, i.e., x→ȳ→x.\nL VCB = E[-log p yz→x (x|F xz→y (x, z), z)] . (5)\nCaptioning-pivoted Back-translation Image captioning is partially similar to MMT besides the non-text part of the input. Inspired by Huang et al. (2020), based on the SG pivoting, we incorporate two captioning procedures, Z→X and Z→Y, to generate pseudo parallel sentences for back-translation and better align the language latent spaces. We denote Z→X as x=C z→x (z), Z→Y as ȳ=C z→y (z). The back-translation loss will be:\nL CPB = E[-log p(x|F xz→y (x, z), z)] + E[-log p(ȳ|F yz→x (ȳ, z), z)] .(6)\n⋆ Remarks In the initial stage, each of the above learning objectives will be executed separately, in a certain order, so as to maintain a stable and effective UMMT system. We first perform L CMA and L REC , because the cross-SG visual-language learning is responsible for aligning the VL SGs, based on which the high-level translation can happen. Then we perform back-translation training L VCB and L CPB , together with VSH updating L VSH .\nOnce the system tends to converge, we put them all together for further fine-tuning:\nL = L CMA + L REC + L VCB + L CPB + L VSH . (7)\n5 Experiments"},{"figure_ref":[],"heading":"Setups","publication_ref":["b8","b3","b4","b20","b26","b27","b40","b0","b28","b17","b33","b5","b31","b13","b9","b41","b9","b20"],"table_ref":[],"text":"The experiments are carried out on Multi30K data (Elliott et al., 2016), a benchmark for MMT, where each image comes with three parallel descriptions in English/German/French. Following Huang et al.\n(2020), we mainly consider the English-French (En↔Fr) and English-German (En↔De). For each translation direction, we only use the src sentence & img for training, and only the src sentence for testing. We also test on the WMT16 En→Ro and WMT14 En→De, En→Fr. WMT (Bojar et al., 2014(Bojar et al., , 2016) ) is widely-used text-only translation corpora, where following Li et al. (2022), we use CLIP (Radford et al., 2021) to retrieve images from Multi30K for sentences.\nFollowing prior research, we employ the Faster-RCNN (Ren et al., 2015) as an object detector, and MOTIFS (Zellers et al., 2018) as a relation classifier and an attribute classifier, where these three together form a VSG generator. For LSG generation, we convert the sentences into dependency trees with a parser (Anderson et al., 2018), which is then transformed into the scene graph based on certain rules (Schuster et al., 2015). For text preprocessing, we use Moses (Koehn et al., 2007) for tokenization and apply the byte pair encoding (BPE) technique. We use Transformer (Vaswani et al., 2017) as the underlying text-encoder to offer representations for GCN, and use the FasterRCNN to encode visual feature representations. All GCN encoders and other feature embeddings have the same dimension of 1,024, and all GCN encoders are with two layers.\nWe mainly compare with the existing UMMT models: Game-MMT (Chen et al., 2018), UMMT (Su et al., 2019) and PVP (Huang et al., 2020).\nTo achieve a fair comparison on the inference-time image-free setup, we also re-implement the UMMT and PVP by integrating the phrase-level retrievalbased visual hallucination method (Fang and Feng, 2022). All models use the same fair configurations, and we do not use pre-trained LM. On WMT we also test the supervised MMT setup, where we use these baselines: UVR (Zhang et al., 2020), RMMT (Wu et al., 2021b), PUVR (Fang and Feng, 2022) and VALHALLA (Li et al., 2022). We report the BLEU and METEOR scores for model evaluation.\nOur results are computed with a model averaging over 5 latest checkpoints with significance test. Our experiments are based on the NVIDIA A100 Tensor Core GPUs."},{"figure_ref":[],"heading":"Main Results","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Results on Multi30K","publication_ref":["b29","b13"],"table_ref":["tab_3","tab_4"],"text":"In Table 2 we show the overall results on Multi30K data. First, we inspect the performance where gold-paired images are given as inputs for testing. We see that our method (Ours # ), by integrating the LSG and VSG information, shows clear superiority over baselines on all translation jobs, while ablating the SGs, the performance drops rapidly. This shows the importance of leveraging scene graphs for more effective Ablation Study In Table 3 we quantify the contribution of each objective of scene graph pivoting learning via ablation study. Each learning strategy exhibits considerable impacts on the overall performance, where the captioning-pivoted (man in shirt and pants playing football.) mann in t-shirt und shorts tritt fußball vom tee.\n(man in t-shirt and shorts kicks football off the tee.) (two bicycles stand behind two people sitting on the grass near a body of water.) back-translation influences the results the biggest, with an average 4.3 BLEU score. Overall, two SG-pivoted back-translation training targets show much higher influences than the two cross-SG visual-language learning objectives. When removing both two back-translation targets, we witness the most dramatic decrease, i.e., average -5.7 BLEU. This validates the long-standing finding that the back-translation mechanism is key to unsupervised translation (Sennrich et al., 2016;Huang et al., 2020). Table 6: Vision-language aligning evaluation. For our models, we transform the hallucinated VSG into an image via a graph-to-image generator. We use CLIP to measure the VL relevance score."},{"figure_ref":[],"heading":"Results on WMT","publication_ref":[],"table_ref":[],"text":"unsupervised MMT. We can find that our unsupervised method only loses within 1 BLEU score to supervised models, e.g., UVR and PUVR."},{"figure_ref":["fig_6","fig_6","fig_7"],"heading":"Further Analyses and Discussions","publication_ref":["b24","b20"],"table_ref":["tab_7","tab_7","tab_7"],"text":"In this part we try to dive deep into the model, presenting in-depth analyses to reveal what and how our proposed method really works and improves.\n• Integration of the vision and language SGs helps gain a holistic understanding of input.\nBoth VSG and LSG advance in comprehensively depicting the intrinsic structure of the content semantics, which ensures a holistic understanding of the input texts and images. By encoding the vision and language SGs, it is expected to completely capture the key components from src inputs, and thus achieve better translations. However, without such structural features, some information may be lost during the translation. In Table 5 via human evalua- tion we can see that our system obtains significantly higher scores in terms of the completeness, comparing to those baselines without considering SGs. Also in Fig. 5, we can find that the baseline system PVP * (PR), with only the local-level phrase-level visual retrieval, has frequently missed the key entities during the translation, e.g., the object 'tee' in case#2.\n• SG-based multimodal feature modeling helps achieve more accurate alignment between vision and language. Another merit to integrating the SGs is that the fine-grained graph modeling of visual and language scenes obviously aids more precise multimodal feature alignment. In this way, the translated texts have higher fidelity to the original texts. Inaccurate multimodal alignment without considering the SG modeling will otherwise lead to worse ambiguity. Observing the ambiguity in Table 5, we see that our model exhibits the lowest ambiguity. In Fig. 5 for the case#3, PVP * (PR) confuses the verb 'saw' as 'see' as it fails to accurately refer 'saw' to a certain lumbering tool, while ours gives a correct prediction. Besides, accurate multimodal alignment greatly enhances the utility of visual information. In Table 6 we compare the relevance of vision-language counterparts by different models, where our model gives the highest performance on both the overall text-image matching and the regional phrase-object matching. In addition, two proposed cross-SG learning targets display big impacts on the VL-aligning ability.\n• The longer and more complex the sentences, the higher the translation quality benefiting from the SGs features. In this work, we investigate the SG structures to model the input texts.\nGraph modeling of the texts has proven effective for resolving the long-range dependency issue (Marcheggiani and Titov, 2017;Li et al., 2022). In Fig. 6 we group the translation performance based on the lengths of source sentences. We see that our SG-based model gives very considerable gains over the two non-SG baselines, where the longer the sentences the higher the improvements.\n• Incorporating SGs into MMT advances in more fluent translation. Also, modeling the semantic scene graph of the input features contributes a lot to the language fluency of the translation texts.\nLooking at the Fluency item in Table 5, we find that our system gives the best fluency with the lowest grammar errors.\n• SG-based visual scene hallucination mechanism helps gain rich and correct visual features. Different from the baseline retrieval-based methods that directly obtain the whole images (or local regions), our proposed VSH mechanism instead compensatively generates the VSGs from the given LSGs. In this way, the hallucinated visual features enjoy two-fold advantages. On the one hand, the pseudo VSG has high correspondence with the textual one, both of which will enhance the shared feature learning between the two modalities. On the other hand, the hallucinated VSG will produce some vision-specific scene components and structures, providing additional clues to facilitate back to the textual features for overall better semantic understanding. Fig. 7 illustrates the node increasing rate during the vision scene graph hallucination. We see that the numbers of all three types of nodes increase, to different extents, where object nodes grow rapidest. Also, during the two transition steps of the VSH mechanism we get two VSGs, skeleton VSG and hallucinated VSG. From Fig. 8 we see that after two full hallucination steps, we can obtain high-fidelity vision features, demonstrating the necessity of the second completing-vision step."},{"figure_ref":[],"heading":"Conclusion","publication_ref":[],"table_ref":[],"text":"We investigate an inference-time image-free setup in unsupervised multimodal machine translation.\nIn specific, we integrate the visual and language scene graph to learn the fine-grained visionlanguage representations. "},{"figure_ref":[],"heading":"A Appendix","publication_ref":[],"table_ref":[],"text":"In §3.2 we give a brief induction to the overall model framework. Here we extend the details of each module of the scene graph-based multimodal translation backbone. In Fig. 9 we outline our framework."},{"figure_ref":["fig_3","fig_9"],"heading":"A.1 Visual Scene Hallucination Learning Module","publication_ref":["b34"],"table_ref":[],"text":"First of all, we note that VSH only will be activated to produce VSG hallucination at inference time. During the training phase, we construct the VSG vocabularies of different VSG nodes. We denote the object vocabulary as D o , which caches the object nodes from parsed VSG of training images; denote the attribute vocabulary as D a , which caches the attribute nodes; and denote the relation vocabulary as D r , which caches the relation nodes.\nThose vocabularies will be used to provide basic ingredients for VSG hallucination. At inference time, VSH is activated to perform two-step inference to generate the hallucinated VSG ′ . The process is illustrated in Fig. 3.\nStep1: Sketching Skeleton This step builds the skeleton VSG from the raw LSG. Specifically, we only need to transform the textual entity nodes into the visual object nodes, while keeping unchanged the whole graph topology. As for the attribute nodes and the relation nodes, we directly copy them into the VSG, as they are all text-based labels that are applicable in VSG. Then we transform the textual entity nodes into the visual object nodes. For each textual entity node in LSG, we employ the edges, which is illustrated in Fig. 11.\n▶ For the node augmentor, we first traverse all the object nodes in the skeleton VSG. For each object node v i , we then perform k-order routing over its neighbor nodes. We denote its neighbor nodes as\nV na i = {• • • , v k , • • • }.\nThen we use the attention to learn the neighbor influence to v i , and obtain the k-order feature representation h i of v i :\nα n k = exp r i • r k v * k ∈V na i exp r i • r * k h na i = r i + k α n k • r k .\nwhere r i and r k is the node representations of v i and v k , which are obtained from GCN encoder.\nThen we use a classifier to make prediction over the total vocabularies of D o and D a , to determine which node v′ i (either an object or an attribute node) should be attached to v i , if any: (FFN([h na i ; r i )) . ▶ For the relation augmentor, we first traverse all the node-pairs (object or attribute nodes, excluding the relation nodes) in the VSG, i.e., v i &v j . Then, for each node in the pair we use a triaffine attention (Wang et al., 2019;Wu et al., 2021a) to directly determine which new relation type ê′ i,j should be built between them, if exists:\nv′ i ← Softmax D na (FFN(h na i )) , where D na = D o ∪ D a ∪\nh pa i-j = Sigmoid( r i 1 T (r j ) T W r i-j 1 ) , ê′ i,j ← Softmax D pa (FFN(h pa i-j )) ,\nwhere D pa = D r ∪ {ϵ}, where the dummy token ϵ indicates no new edge should be created between two nodes. The new edge ê′ i,j has a relation label. r i-j is the representation of the path from v i to v j , which is obtained by the pooling function over all the nodes in the path:\nh pa i-j = Pool(r i , • • • , r j ) . Note that the triaffine scorer is effective in modeling the high-order ternary relations, which will provide a precise determination on whether to add a new edge.\nDuring training, the node augmentor and the relation augmentor are trained and updated based on the gold LSG and VSG, to learn the correct mapping between LSG and VSG. Such supervised learning is also important for ensuring that the final generated hallucinated visual scenes are basically coincident with the caption text, instead of random or groundless vision scenes."},{"figure_ref":[],"heading":"A.2 SG Fusing&Mapping Module","publication_ref":[],"table_ref":[],"text":"Here we extend the contents in § 3.2. As shown in Fig. 9, first of all, the SG fusing module merges the LSG x and VSG into a mixed cross-modal scene graph, such that the merged scene graph are highly compact with less redundant. Before the merging, we first measure the similarity of each pair of node representations via cosine distance:\ns f i,j = (r L i ) T • r V j ||r L i || ||r V j ||\n. This is a similar process as in Eq. ( 2). For those pairs with high alignment scores, i.e., s i,j > α (we use the same pre-defined threshold as in crossmodal alignment learning), we consider them as serving a similar role. Since we will perform the cross-modal SG aligning learning L CMA , the accuracy of the alignment between LSG x and VSG can be guaranteed. Then, we average the representations of the image-text node pair from their GCNs. And for the rest of nodes in LSG x and VSG, we take the union structures of them. The resulting mixed SG fully inherits the semantic-rich scene nodes from both the textual SG and the visual SG, which will benefit the following text generation. Now we treat the mixed SG as a pseudo tgt-side LSG y . We use another GCN to model LSG y for further feature propagation: r y 1 , • • • , r y m = GCN(V SG y ) . The initial node representations of GCN are from the GCNs of VSG and LSG x , i.e., r L and r V as in Eq. (1). Based on the node representation r y i of VSG y , we finally employ a graph-to-text model2 to generate the final tgt-side sentence. Specifically, all the node representation r i will be first summarized into one unified graph-level feature via pooling: r y = Pool(r y 1 , • • • , r y m ) . Then, an autoregressive sequential decoder (Se-qDec) will take r y to generate tgt-side token over the tgt-side vocabulary at each setp, sequentially: e i = SeqDec(e ≤i , r y ) , ŷi ← Softmax(e i ) ."},{"figure_ref":[],"heading":"Acknowledgments","publication_ref":[],"table_ref":[],"text":"This research is supported by the National Natural Science Foundation of China (No. 62176180), and also the Sea-NExT Joint Lab."},{"figure_ref":[],"heading":"Limitations","publication_ref":[],"table_ref":[],"text":"Our paper has the following potential limitations. First of all, we take advantage of the external scene graph structures to achieve the inference-time visual hallucination and secure significant improvements of the target task, while it could be a doubleedged sword. This makes our method subject to the quality of the external structure parsers. When the parsed structures of visual scene graphs and language scene graphs are with much noise, it will deteriorate our methods. Fortunately, the existing scene graph parsers have already achieved satisfactory performance for the majority language (e.g., English), which can meet our demands. Second, the effectiveness of our approach depends on the availability of good-quality images, which however shares the pitfalls associated with the standard unsupervised multimodal translation setup."},{"figure_ref":[],"heading":"Node augmentor","publication_ref":[],"table_ref":[],"text":"Skeleton VSG CLIP tool 1 to search for the best matching visual node (proposal) in D o as the counterpart visual object, resulting in the skeleton VSG. After this step, we obtain the sketch structure of the target VSG.\nStep2: Completing Vision This step completes the skeleton VSG into a more realistic one, i.e., the final hallucinated VSG ′ . With the skeleton VSG at hand, we aim to further enrich skeleton VSG. Because intuitively, in actual world the visual scenes are always much more concrete and vivid than textual scenes. For example, given a caption text 'boys are playing baseball on playground', the LSG only mentions 'boys', 'baseball' and 'playground' objects. But imaginarily, there must be a 'baseball bat' in the scene of vision, and also both the pairs of 'boys'-'playground' and 'baseball'-'playground' has 'on' relation. Thus it is indispensable to add new nodes and more edges, i.e., scene graph augmentation. To reach the goal, we propose a node augmentor and a relation augmentor, as shown in Fig. 10. First of all, we downgrade all the relation nodes as the edge itself, i.e., an edge with a relation label. By this, we obtain a VSG that only contains object and attribute nodes, and labeled 1 https://github.com/openai/CLIP"}],"string":"[\n {\n \"figure_ref\": [\n \"fig_1\",\n \"fig_2\"\n ],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b32\",\n \"b2\",\n \"b42\",\n \"b1\",\n \"b5\",\n \"b19\",\n \"b6\",\n \"b29\",\n \"b30\",\n \"b14\",\n \"b13\",\n \"b31\",\n \"b13\",\n \"b41\",\n \"b20\",\n \"b12\",\n \"b41\",\n \"b9\",\n \"b16\",\n \"b35\"\n ],\n \"table_ref\": [\n \"tab_0\"\n ],\n \"text\": \"Current neural machine translation (NMT) has achieved great triumph (Sutskever et al., 2014;Bahdanau et al., 2015;Zhu et al., 2020), however in the cost of creating large-scale parallel sentences, which obstructs the development of NMT for the minor languages. Unsupervised NMT (UMT) has thus been proposed to relieve the reliance of parallel corpora (Artetxe et al., 2018;Chen et al., 2018). The core idea of UMT is to align the representation spaces between two languages with alternative pivot signals rather than parallel sentences, such as bilingual lexicons (Lample et al., 2018), multilingual language models (LM) (Conneau and Lample, 2019) and back-translation technique (Sennrich et al., 2016). Recent trends have considered the incorporation of visual information, i.e., multimodal machine translation (MMT) (Specia et al., 2016;Huang et al., 2016). Intuitively, visual modality can serve as language-agnostic signals, pivoting different languages by grounding the same textual semantics into the common visual space. Therefore, solving UMT with visual contents as pivot becomes a promising solution, a.k.a., unsupervised MMT (UMMT) (Huang et al., 2020;Su et al., 2019).\\nUMMT systems are trained with only the textimage pairs (), which can be easier to collect than the parallel source-target sentence pairs () (Huang et al., 2020). Although exempting the parallel sentences for training, UMMT still requires such text-image pairs as inputs for testing. Yet such assumption might be unrealistic, because in most of the real-world scenarios such as online translation systems, paired images are not available during inference. Especially for some scarce languages, the pairs have difficult access. In other words, practical UMMT systems should not only avoid the parallel sentences during training, but also the text-image pairs during inference. As summarized in Table 1, although some existing MMT researches exempt the testing-time visual inputs (Zhang et al., 2020;Li et al., 2022), they all unfortunately are supervised methods, relying on large-scale parallel sentences for training. As emphasized above, the visual information is vital to UMMT. However, for both the existing supervised and unsupervised MMT studies, they may suffer from ineffective and insufficient modeling of visual pivot features. For example, most of MMT models perform vision-language (VL) grounding over the whole image and text (Huang et al., 2019;Zhang et al., 2020), where such coarse-grained representation learning can cause mismatching and sacrifice the subtle VL semantics. Fang and Feng (2022) recently introduce a fine-grained VL alignment learning via phrase-level grounding, while without a holistic understanding of the visual scene, such local-level method may lead to incomplete or missing alignments.\\nIn this work, we present a novel UMMT method that solves all aforementioned challenges. First of all, to better represent the visual (also the textual) inputs, we consider incorporating the visual scene graph (VSG) (Johnson et al., 2015) and language scene graph (LSG) (Wang et al., 2018). The scene graphs (SG) advance in intrinsically depicting the semantic structures of texts or images with rich details (cf. Fig. 1), which offers a holistic viewpoint for more effective pivoting learning. Then, we build the UMMT framework as illustrated in Fig. 2. The input src text and paired image are first transformed into LSG and VSG, which are further fused into a mixed SG, and then translated into the tgt-side LSG. And the tgt sentence will be finally produced conditioned on the tgt LSG. Several SGbased pivoting learning strategies are proposed for unsupervised training of UMMT system. In addition, to support pure-text (image-free) input during inference, we devise a novel visual scene hallucination module, which dynamically generates a hallucinated VSG from the LSG compensatively. Our system is evaluated on the standard MMT Multi30K and NMT WMT data. Extensive experimental results verify that the proposed method outperforms strong baselines on unsupervised multimodal translation by above 5 BLEU score on average. We further reveal the efficacy of the visual scene hallucination mechanism in relieving the reliance on image inputs during inference. Our SG-pivoting based UMMT helps yield translations with higher completeness, relevance and fluency, and especially obtains improvements on the longer sentences.\\nOverall, we make the following contributions: ▶ 1) We are the first to study the inferencetime image-free unsupervised multimodal machine translation, solved with a novel visual scene hallucination mechanism. ▶ 2) We leverage the SGs to better represent the visual and language inputs. Moreover, we design SG-based graph pivoting learning strategies for UMMT training. ▶ 3) Our model achieves huge boosts over strong baselines on benchmark data. Code is available at https: //github.com/scofield7419/UMMT-VSH.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Work\",\n \"publication_ref\": [\n \"b32\",\n \"b2\",\n \"b23\",\n \"b33\",\n \"b7\",\n \"b30\",\n \"b14\",\n \"b29\",\n \"b21\",\n \"b19\",\n \"b6\",\n \"b25\",\n \"b5\",\n \"b31\",\n \"b13\",\n \"b18\",\n \"b35\",\n \"b16\",\n \"b15\",\n \"b39\",\n \"b41\",\n \"b9\"\n ],\n \"table_ref\": [],\n \"text\": \"Neural machine translation has achieved notable development in the era of deep learning (Sutskever et al., 2014;Bahdanau et al., 2015;Luong et al., 2015). The constructions of powerful neural models and training paradigms as well as the collection of large-scale parallel corpora are the driving forces to NMT's success (Vaswani et al., 2017;Devlin et al., 2019). The key of NMT is to learn a good mapping between two (or more) languages. In recent years, visual information has been intro-duced for stronger NMT (i.e., multimodal machine translation), by enhancing the alignments of language latent spaces with visual grounding (Specia et al., 2016;Huang et al., 2016). Intuitively, people speaking different languages can actually refer to the same physical visual contents and conceptions.\\nUnsupervised machine translation aims to learn cross-lingual mapping without the use of largescale parallel corpora. The setting is practically meaningful to those minor languages with hard data accessibility. The basic idea is to leverage alternative pivoting contents to compensate the parallel signals based on the back-translation method (Sennrich et al., 2016), such as third-languages (Li et al., 2020), bilingual lexicons (Lample et al., 2018) or multilingual LM (Conneau and Lample, 2019). The visual information can also serve as pivot signals for UMT, i.e., unsupervised multimodal machine translation. Comparing to the standard MMT that trains with triples, UMMT takes as input only the . So far, few studies have explored the UMMT setting, most of which try to enhance the back-translation with multimodal alignment mechanism (Nakayama and Nishida, 2017;Chen et al., 2018;Su et al., 2019;Huang et al., 2020).\\nScene graph describes a scene of an image or text into a structure layout, by connecting discrete objects with attributes and with other objects via pairwise relations (Krishna et al., 2017;Wang et al., 2018). As the SGs carry rich contextual and semantic information, they are widely integrated into downstream tasks for enhancements, e.g., image retrieval (Johnson et al., 2015), image generation (Johnson et al., 2018) and image captioning (Yang et al., 2019). This work inherits wisdom, incorporating both the visual scene graph and language scene graph as pivots for UMMT.\\nAll the UMMT researches assume that the pairs are required during inference, yet we notice that this can be actually unrealistic. We thus propose a visual hallucination mechanism, achieving the inference-time image-free goal. There are relevant studies on supervised MMT that manage to avoid image inputs (with text only) during inference. The visual retrieval-base methods (Zhang et al., 2020;Fang and Feng, 2022) sentence. Differently, we consider generating the visual scene graph with richer and holistic visual structural information.\\n3 Scene Graph-based Translation System\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Problem Definition\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In UMMT, no parallel translation pairs are available. This work considers an inference-time imagefree UMMT. During training, the data availability is ∈ and the corresponding srcside LSG x and VSG, where X are the src-side sentences, and Z are the paired images. During inference, the model generates tgt-side sentences y ∈ Y based on the inputs of only x ∈ X and the corresponding LSG x , while the visual scene VSG ′ is hallucinated from LSG x . In both training and inference, y will be generated from the intermediate tgt-side language scene graph LSG y , which is produced from LSG x and VSG (or VSG ′ ).\"\n },\n {\n \"figure_ref\": [\n \"fig_2\"\n ],\n \"heading\": \"Framework\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"As shown in Fig. 2, the system first represents the src-side LSG x and VSG features with two GCN graph encoders, respectively. Then the SG fus-ing&mapping module integrates and transforms two SG representations into a unified one as tgtside LSG, i.e., LSG y . Another GSN model further encodes the LSG y , where the representations are used to generate the tgt sentence (i.e., translation).\"\n },\n {\n \"figure_ref\": [\n \"fig_3\"\n ],\n \"heading\": \"Scene Graph Generating and Encoding\",\n \"publication_ref\": [\n \"b24\"\n ],\n \"table_ref\": [],\n \"text\": \"We first employ two off-the-shelf SG parsers to obtain the LSG and VSG, separately (detailed in the experiment part). For simplicity, here we unify the notations of LSG and VSG as SG. We denote a SG as G=(V, E), where V are the nodes (including object o, attribute a and relation r types), and E are the edges e i,j between any pair of nodes v i ∈ V .\\nWe then encode both the VSG and LSG with two spatial Graph Convolution Networks (GCN) (Marcheggiani and Titov, 2017) respectively, which is formulated as:\\nr 1 , • • • , r n = GCN(G) ,(1)\\nwhere r i is the representation of node v i . We here denote r L i as LSG's node representation, and r V i as VSG's node representation.\\nVisual Scene Hallucinating During inference, the visual scene hallucination (VSH) module is activated to perform two-step inference to generate the hallucinated VSG ′ , as illustrated in Fig. 3.\\nStep1: sketching skeleton aims to build the skeleton VSG. We copy all the nodes from the raw LSG x to the target VSG, and transform the textual entity nodes into the visual object nodes.\\nStep2: completing vision aims to enrich and augment the skeleton VSG into a more realistic one. It is indispensable to add new nodes and edges in the skeleton VSG, since in real scenarios, visual scenes are much more concrete and vivid than textual scenes. Specifically, we develop a node augmentor and a relation augmentor, where the former decides whether to attach a new node to an existing one, and the later decides whether to create an edge between two disjoint nodes. To ensure the fidelity of the hallucinated VSG ′ , during training, the node augmentor and relation augmentor will be updated (i.e., with the learning target L VSH ) with the input LSG and VSG supervisions. Appendix §A.1 details the VSH module.\\nSG Fusing&Mapping Now we fuse the heterogeneous LSG x and VSG into one unified scene graph with a mixed view. The key idea is to merge the information from two SGs serving similar roles.\\nIn particular, we first measure the representation similarity of each pair of nodes from two GCNs. For those pairs with high alignment scores, we merge them as one by averaging their representations, and for those not, we take the union structures from two SGs. This results in a pseudo tgt-side LSG y . We then use another GCN model for further representation propagation. Finally, we employ a graph-to-text generator to transform the LSG y representations to the tgt sentence y. Appendix §A.2 presents all the technical details in this part.\"\n },\n {\n \"figure_ref\": [\n \"fig_4\"\n ],\n \"heading\": \"Learning with Scene Graph Pivoting\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this part, based on the SG pivot we introduce several learning strategies to accomplish the unsupervised training of machine translation. We mainly consider 1) cross-SG visual-language learning, and 2) SG-pivoted back-translation training. Fig. 4 illustrates these learning strategies.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Cross-SG Visual-language Learning\",\n \"publication_ref\": [\n \"b22\",\n \"b38\",\n \"b11\",\n \"b15\"\n ],\n \"table_ref\": [],\n \"text\": \"The visual-language SG cross-learning aims to enhance the structural correspondence between the LSG and VSG. Via cross-learning we also teach the SG encoders to automatically learn to highlight those shared visual-language information while deactivating those trivial substructures, i.e., denoising.\\nCross-modal SG Aligning The idea is to encourage the text and visual nodes that serve a similar role in VSG and LSG to be closer. To align the fine-grained structures between SGs, we adopt the contrastive learning (CL) technique (Logeswaran and Lee, 2018;Yan et al., 2021;Fei et al., 2022;Huang et al., 2022). In particular, CL learns effec-tive representation by pulling semantically close content pairs together, while pushing apart those different ones. Technically, we measure the similarities between pairs of nodes from two VSG and LSG:\\nsi,j = (r L i ) T • r V j ||r L i || ||r V j || .(2)\\nA threshold value α is pre-defined to decide the alignment confidence, i.e., pairs with s i,j > α are considered similar. Then we put on the CL loss:\\nL CMA = - i∈LSG x , j * ∈VSG log exp(si,j * /τ ) Z ,(3)\\nZ = i∈LSG x , k∈VSG, k̸ =j * exp(s i,k /τ ) ,(4)\\nwhere τ >0 is an annealing factor. j * means a positive pair with i, i.e., s i,j * >α.\\nCross-modal Cross-reconstruction We further strengthen the correspondence between VSG and LSG via cross-modal cross-reconstruction. Specifically, we try to reconstruct the input sentence from the VSG, and the image representations from the LSG. In this way we force both two SGs to focus on the VL-shared parts. To realize VSG→x we employ the aforementioned graph-to-text generator.\\nFor LSG→z, we use the graph-to-image generator (Johnson et al., 2018). The learning loss can be marked as L REC .\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"SG-pivoted Back-translation Training\",\n \"publication_ref\": [\n \"b29\",\n \"b13\"\n ],\n \"table_ref\": [],\n \"text\": \"Back-translation is a key method to realize unsupervised machine translation (Sennrich et al., 2016).\\nIn this work, we further aid the back-translation with structural SG pivoting.\\nVisual-concomitant Back-translation We perform the back-translation with the SG pivoting. We denote the X →Y translation direction as y=F xz→y (x, z), and Y→Z as x=F yz→x (y, z).\\nAs we only have src-side sentences, the backtranslation is uni-directional, i.e., x→ȳ→x.\\nL VCB = E[-log p yz→x (x|F xz→y (x, z), z)] . (5)\\nCaptioning-pivoted Back-translation Image captioning is partially similar to MMT besides the non-text part of the input. Inspired by Huang et al. (2020), based on the SG pivoting, we incorporate two captioning procedures, Z→X and Z→Y, to generate pseudo parallel sentences for back-translation and better align the language latent spaces. We denote Z→X as x=C z→x (z), Z→Y as ȳ=C z→y (z). The back-translation loss will be:\\nL CPB = E[-log p(x|F xz→y (x, z), z)] + E[-log p(ȳ|F yz→x (ȳ, z), z)] .(6)\\n⋆ Remarks In the initial stage, each of the above learning objectives will be executed separately, in a certain order, so as to maintain a stable and effective UMMT system. We first perform L CMA and L REC , because the cross-SG visual-language learning is responsible for aligning the VL SGs, based on which the high-level translation can happen. Then we perform back-translation training L VCB and L CPB , together with VSH updating L VSH .\\nOnce the system tends to converge, we put them all together for further fine-tuning:\\nL = L CMA + L REC + L VCB + L CPB + L VSH . (7)\\n5 Experiments\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Setups\",\n \"publication_ref\": [\n \"b8\",\n \"b3\",\n \"b4\",\n \"b20\",\n \"b26\",\n \"b27\",\n \"b40\",\n \"b0\",\n \"b28\",\n \"b17\",\n \"b33\",\n \"b5\",\n \"b31\",\n \"b13\",\n \"b9\",\n \"b41\",\n \"b9\",\n \"b20\"\n ],\n \"table_ref\": [],\n \"text\": \"The experiments are carried out on Multi30K data (Elliott et al., 2016), a benchmark for MMT, where each image comes with three parallel descriptions in English/German/French. Following Huang et al.\\n(2020), we mainly consider the English-French (En↔Fr) and English-German (En↔De). For each translation direction, we only use the src sentence & img for training, and only the src sentence for testing. We also test on the WMT16 En→Ro and WMT14 En→De, En→Fr. WMT (Bojar et al., 2014(Bojar et al., , 2016) ) is widely-used text-only translation corpora, where following Li et al. (2022), we use CLIP (Radford et al., 2021) to retrieve images from Multi30K for sentences.\\nFollowing prior research, we employ the Faster-RCNN (Ren et al., 2015) as an object detector, and MOTIFS (Zellers et al., 2018) as a relation classifier and an attribute classifier, where these three together form a VSG generator. For LSG generation, we convert the sentences into dependency trees with a parser (Anderson et al., 2018), which is then transformed into the scene graph based on certain rules (Schuster et al., 2015). For text preprocessing, we use Moses (Koehn et al., 2007) for tokenization and apply the byte pair encoding (BPE) technique. We use Transformer (Vaswani et al., 2017) as the underlying text-encoder to offer representations for GCN, and use the FasterRCNN to encode visual feature representations. All GCN encoders and other feature embeddings have the same dimension of 1,024, and all GCN encoders are with two layers.\\nWe mainly compare with the existing UMMT models: Game-MMT (Chen et al., 2018), UMMT (Su et al., 2019) and PVP (Huang et al., 2020).\\nTo achieve a fair comparison on the inference-time image-free setup, we also re-implement the UMMT and PVP by integrating the phrase-level retrievalbased visual hallucination method (Fang and Feng, 2022). All models use the same fair configurations, and we do not use pre-trained LM. On WMT we also test the supervised MMT setup, where we use these baselines: UVR (Zhang et al., 2020), RMMT (Wu et al., 2021b), PUVR (Fang and Feng, 2022) and VALHALLA (Li et al., 2022). We report the BLEU and METEOR scores for model evaluation.\\nOur results are computed with a model averaging over 5 latest checkpoints with significance test. Our experiments are based on the NVIDIA A100 Tensor Core GPUs.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Main Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Results on Multi30K\",\n \"publication_ref\": [\n \"b29\",\n \"b13\"\n ],\n \"table_ref\": [\n \"tab_3\",\n \"tab_4\"\n ],\n \"text\": \"In Table 2 we show the overall results on Multi30K data. First, we inspect the performance where gold-paired images are given as inputs for testing. We see that our method (Ours # ), by integrating the LSG and VSG information, shows clear superiority over baselines on all translation jobs, while ablating the SGs, the performance drops rapidly. This shows the importance of leveraging scene graphs for more effective Ablation Study In Table 3 we quantify the contribution of each objective of scene graph pivoting learning via ablation study. Each learning strategy exhibits considerable impacts on the overall performance, where the captioning-pivoted (man in shirt and pants playing football.) mann in t-shirt und shorts tritt fußball vom tee.\\n(man in t-shirt and shorts kicks football off the tee.) (two bicycles stand behind two people sitting on the grass near a body of water.) back-translation influences the results the biggest, with an average 4.3 BLEU score. Overall, two SG-pivoted back-translation training targets show much higher influences than the two cross-SG visual-language learning objectives. When removing both two back-translation targets, we witness the most dramatic decrease, i.e., average -5.7 BLEU. This validates the long-standing finding that the back-translation mechanism is key to unsupervised translation (Sennrich et al., 2016;Huang et al., 2020). Table 6: Vision-language aligning evaluation. For our models, we transform the hallucinated VSG into an image via a graph-to-image generator. We use CLIP to measure the VL relevance score.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Results on WMT\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"unsupervised MMT. We can find that our unsupervised method only loses within 1 BLEU score to supervised models, e.g., UVR and PUVR.\"\n },\n {\n \"figure_ref\": [\n \"fig_6\",\n \"fig_6\",\n \"fig_7\"\n ],\n \"heading\": \"Further Analyses and Discussions\",\n \"publication_ref\": [\n \"b24\",\n \"b20\"\n ],\n \"table_ref\": [\n \"tab_7\",\n \"tab_7\",\n \"tab_7\"\n ],\n \"text\": \"In this part we try to dive deep into the model, presenting in-depth analyses to reveal what and how our proposed method really works and improves.\\n• Integration of the vision and language SGs helps gain a holistic understanding of input.\\nBoth VSG and LSG advance in comprehensively depicting the intrinsic structure of the content semantics, which ensures a holistic understanding of the input texts and images. By encoding the vision and language SGs, it is expected to completely capture the key components from src inputs, and thus achieve better translations. However, without such structural features, some information may be lost during the translation. In Table 5 via human evalua- tion we can see that our system obtains significantly higher scores in terms of the completeness, comparing to those baselines without considering SGs. Also in Fig. 5, we can find that the baseline system PVP * (PR), with only the local-level phrase-level visual retrieval, has frequently missed the key entities during the translation, e.g., the object 'tee' in case#2.\\n• SG-based multimodal feature modeling helps achieve more accurate alignment between vision and language. Another merit to integrating the SGs is that the fine-grained graph modeling of visual and language scenes obviously aids more precise multimodal feature alignment. In this way, the translated texts have higher fidelity to the original texts. Inaccurate multimodal alignment without considering the SG modeling will otherwise lead to worse ambiguity. Observing the ambiguity in Table 5, we see that our model exhibits the lowest ambiguity. In Fig. 5 for the case#3, PVP * (PR) confuses the verb 'saw' as 'see' as it fails to accurately refer 'saw' to a certain lumbering tool, while ours gives a correct prediction. Besides, accurate multimodal alignment greatly enhances the utility of visual information. In Table 6 we compare the relevance of vision-language counterparts by different models, where our model gives the highest performance on both the overall text-image matching and the regional phrase-object matching. In addition, two proposed cross-SG learning targets display big impacts on the VL-aligning ability.\\n• The longer and more complex the sentences, the higher the translation quality benefiting from the SGs features. In this work, we investigate the SG structures to model the input texts.\\nGraph modeling of the texts has proven effective for resolving the long-range dependency issue (Marcheggiani and Titov, 2017;Li et al., 2022). In Fig. 6 we group the translation performance based on the lengths of source sentences. We see that our SG-based model gives very considerable gains over the two non-SG baselines, where the longer the sentences the higher the improvements.\\n• Incorporating SGs into MMT advances in more fluent translation. Also, modeling the semantic scene graph of the input features contributes a lot to the language fluency of the translation texts.\\nLooking at the Fluency item in Table 5, we find that our system gives the best fluency with the lowest grammar errors.\\n• SG-based visual scene hallucination mechanism helps gain rich and correct visual features. Different from the baseline retrieval-based methods that directly obtain the whole images (or local regions), our proposed VSH mechanism instead compensatively generates the VSGs from the given LSGs. In this way, the hallucinated visual features enjoy two-fold advantages. On the one hand, the pseudo VSG has high correspondence with the textual one, both of which will enhance the shared feature learning between the two modalities. On the other hand, the hallucinated VSG will produce some vision-specific scene components and structures, providing additional clues to facilitate back to the textual features for overall better semantic understanding. Fig. 7 illustrates the node increasing rate during the vision scene graph hallucination. We see that the numbers of all three types of nodes increase, to different extents, where object nodes grow rapidest. Also, during the two transition steps of the VSH mechanism we get two VSGs, skeleton VSG and hallucinated VSG. From Fig. 8 we see that after two full hallucination steps, we can obtain high-fidelity vision features, demonstrating the necessity of the second completing-vision step.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We investigate an inference-time image-free setup in unsupervised multimodal machine translation.\\nIn specific, we integrate the visual and language scene graph to learn the fine-grained visionlanguage representations. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A Appendix\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In §3.2 we give a brief induction to the overall model framework. Here we extend the details of each module of the scene graph-based multimodal translation backbone. In Fig. 9 we outline our framework.\"\n },\n {\n \"figure_ref\": [\n \"fig_3\",\n \"fig_9\"\n ],\n \"heading\": \"A.1 Visual Scene Hallucination Learning Module\",\n \"publication_ref\": [\n \"b34\"\n ],\n \"table_ref\": [],\n \"text\": \"First of all, we note that VSH only will be activated to produce VSG hallucination at inference time. During the training phase, we construct the VSG vocabularies of different VSG nodes. We denote the object vocabulary as D o , which caches the object nodes from parsed VSG of training images; denote the attribute vocabulary as D a , which caches the attribute nodes; and denote the relation vocabulary as D r , which caches the relation nodes.\\nThose vocabularies will be used to provide basic ingredients for VSG hallucination. At inference time, VSH is activated to perform two-step inference to generate the hallucinated VSG ′ . The process is illustrated in Fig. 3.\\nStep1: Sketching Skeleton This step builds the skeleton VSG from the raw LSG. Specifically, we only need to transform the textual entity nodes into the visual object nodes, while keeping unchanged the whole graph topology. As for the attribute nodes and the relation nodes, we directly copy them into the VSG, as they are all text-based labels that are applicable in VSG. Then we transform the textual entity nodes into the visual object nodes. For each textual entity node in LSG, we employ the edges, which is illustrated in Fig. 11.\\n▶ For the node augmentor, we first traverse all the object nodes in the skeleton VSG. For each object node v i , we then perform k-order routing over its neighbor nodes. We denote its neighbor nodes as\\nV na i = {• • • , v k , • • • }.\\nThen we use the attention to learn the neighbor influence to v i , and obtain the k-order feature representation h i of v i :\\nα n k = exp r i • r k v * k ∈V na i exp r i • r * k h na i = r i + k α n k • r k .\\nwhere r i and r k is the node representations of v i and v k , which are obtained from GCN encoder.\\nThen we use a classifier to make prediction over the total vocabularies of D o and D a , to determine which node v′ i (either an object or an attribute node) should be attached to v i , if any: (FFN([h na i ; r i )) . ▶ For the relation augmentor, we first traverse all the node-pairs (object or attribute nodes, excluding the relation nodes) in the VSG, i.e., v i &v j . Then, for each node in the pair we use a triaffine attention (Wang et al., 2019;Wu et al., 2021a) to directly determine which new relation type ê′ i,j should be built between them, if exists:\\nv′ i ← Softmax D na (FFN(h na i )) , where D na = D o ∪ D a ∪\\nh pa i-j = Sigmoid( r i 1 T (r j ) T W r i-j 1 ) , ê′ i,j ← Softmax D pa (FFN(h pa i-j )) ,\\nwhere D pa = D r ∪ {ϵ}, where the dummy token ϵ indicates no new edge should be created between two nodes. The new edge ê′ i,j has a relation label. r i-j is the representation of the path from v i to v j , which is obtained by the pooling function over all the nodes in the path:\\nh pa i-j = Pool(r i , • • • , r j ) . Note that the triaffine scorer is effective in modeling the high-order ternary relations, which will provide a precise determination on whether to add a new edge.\\nDuring training, the node augmentor and the relation augmentor are trained and updated based on the gold LSG and VSG, to learn the correct mapping between LSG and VSG. Such supervised learning is also important for ensuring that the final generated hallucinated visual scenes are basically coincident with the caption text, instead of random or groundless vision scenes.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.2 SG Fusing&Mapping Module\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Here we extend the contents in § 3.2. As shown in Fig. 9, first of all, the SG fusing module merges the LSG x and VSG into a mixed cross-modal scene graph, such that the merged scene graph are highly compact with less redundant. Before the merging, we first measure the similarity of each pair of node representations via cosine distance:\\ns f i,j = (r L i ) T • r V j ||r L i || ||r V j ||\\n. This is a similar process as in Eq. ( 2). For those pairs with high alignment scores, i.e., s i,j > α (we use the same pre-defined threshold as in crossmodal alignment learning), we consider them as serving a similar role. Since we will perform the cross-modal SG aligning learning L CMA , the accuracy of the alignment between LSG x and VSG can be guaranteed. Then, we average the representations of the image-text node pair from their GCNs. And for the rest of nodes in LSG x and VSG, we take the union structures of them. The resulting mixed SG fully inherits the semantic-rich scene nodes from both the textual SG and the visual SG, which will benefit the following text generation. Now we treat the mixed SG as a pseudo tgt-side LSG y . We use another GCN to model LSG y for further feature propagation: r y 1 , • • • , r y m = GCN(V SG y ) . The initial node representations of GCN are from the GCNs of VSG and LSG x , i.e., r L and r V as in Eq. (1). Based on the node representation r y i of VSG y , we finally employ a graph-to-text model2 to generate the final tgt-side sentence. Specifically, all the node representation r i will be first summarized into one unified graph-level feature via pooling: r y = Pool(r y 1 , • • • , r y m ) . Then, an autoregressive sequential decoder (Se-qDec) will take r y to generate tgt-side token over the tgt-side vocabulary at each setp, sequentially: e i = SeqDec(e ≤i , r y ) , ŷi ← Softmax(e i ) .\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Acknowledgments\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This research is supported by the National Natural Science Foundation of China (No. 62176180), and also the Sea-NExT Joint Lab.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Limitations\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Our paper has the following potential limitations. First of all, we take advantage of the external scene graph structures to achieve the inference-time visual hallucination and secure significant improvements of the target task, while it could be a doubleedged sword. This makes our method subject to the quality of the external structure parsers. When the parsed structures of visual scene graphs and language scene graphs are with much noise, it will deteriorate our methods. Fortunately, the existing scene graph parsers have already achieved satisfactory performance for the majority language (e.g., English), which can meet our demands. Second, the effectiveness of our approach depends on the availability of good-quality images, which however shares the pitfalls associated with the standard unsupervised multimodal translation setup.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Node augmentor\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Skeleton VSG CLIP tool 1 to search for the best matching visual node (proposal) in D o as the counterpart visual object, resulting in the skeleton VSG. After this step, we obtain the sketch structure of the target VSG.\\nStep2: Completing Vision This step completes the skeleton VSG into a more realistic one, i.e., the final hallucinated VSG ′ . With the skeleton VSG at hand, we aim to further enrich skeleton VSG. Because intuitively, in actual world the visual scenes are always much more concrete and vivid than textual scenes. For example, given a caption text 'boys are playing baseball on playground', the LSG only mentions 'boys', 'baseball' and 'playground' objects. But imaginarily, there must be a 'baseball bat' in the scene of vision, and also both the pairs of 'boys'-'playground' and 'baseball'-'playground' has 'on' relation. Thus it is indispensable to add new nodes and more edges, i.e., scene graph augmentation. To reach the goal, we propose a node augmentor and a relation augmentor, as shown in Fig. 10. First of all, we downgrade all the relation nodes as the edge itself, i.e., an edge with a relation label. By this, we obtain a VSG that only contains object and attribute nodes, and labeled 1 https://github.com/openai/CLIP\"\n }\n]"},"abstract":{"kind":"string","value":"In this work, we investigate a more realistic unsupervised multimodal machine translation (UMMT) setup, inference-time image-free UMMT, where the model is trained with sourcetext image pairs, and tested with only sourcetext inputs. First, we represent the input images and texts with the visual and language scene graphs (SG), where such fine-grained visionlanguage features ensure a holistic understanding of the semantics. To enable pure-text input during inference, we devise a visual scene hallucination mechanism that dynamically generates pseudo visual SG from the given textual SG. Several SG-pivoting based learning objectives are introduced for unsupervised translation training. On the benchmark Multi30K data, our SG-based method outperforms the best-performing baseline by significant BLEU scores on the task and setup, helping yield translations with better completeness, relevance and fluency without relying on paired images. Further in-depth analyses reveal how our model advances in the task setting."},"title":{"kind":"string","value":"Scene Graph as Pivoting: Inference-time Image-free Unsupervised Multimodal Machine Translation with Visual Scene Hallucination"},"figures":{"kind":"list like","value":[{"figure_caption":"✓• Unsupervised MMT Chen et al. (2018) Su et al. (2019)Huang et al. (2020) ","figure_data":"","figure_id":"fig_0","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 1 :1Figure 1: Representing the texts and images via language scene graphs (LSG) and visual scene graphs (VSG). In a scene graph, object, attribute, relation nodes are shown in green, orange, purple respectively.","figure_data":"","figure_id":"fig_1","figure_label":"1","figure_type":"figure"},{"figure_caption":"Figure 2 :2Figure 2: The high-level overview of our SG-based UMMT model. During training, src-side sentences with paired images are used as inputs, together with the corresponding LSG and VSG. Testing phase only takes src-side sentences, where the visual hallucination module is activated to generate VSG from text sources.","figure_data":"","figure_id":"fig_2","figure_label":"2","figure_type":"figure"},{"figure_caption":"Figure 3 :3Figure 3: The illustration of the visual scene hallucination (VSH) module, including two steps of inference.","figure_data":"","figure_id":"fig_3","figure_label":"3","figure_type":"figure"},{"figure_caption":"Figure 4 :4Figure 4: Illustrations of the learning strategies for unsupervised multimodal machine translation.","figure_data":"","figure_id":"fig_4","figure_label":"4","figure_type":"figure"},{"figure_caption":"two bicycles stand behind two people sitting on the grass near a body of water. hinter zwei Personen, die auf dem gras in der nähe eines gewässers sitzen. stand behind two man with the herbaceous plants near the ocean.)zwei fahrräder stehen hinter zwei mann mit den eingetopften graspflanzen in der nähe des meeres.man in t-shirt and shorts kicking football off tee. und hose, der fußball spielt.","figure_data":"","figure_id":"fig_5","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 5 :5Figure 5: Qualitative results of inference-time image-free UMMT (En→De).","figure_data":"","figure_id":"fig_6","figure_label":"5","figure_type":"figure"},{"figure_caption":"Figure 6 :6Figure 6: BLEU scores under different sentence lengths.","figure_data":"","figure_id":"fig_7","figure_label":"6","figure_type":"figure"},{"figure_caption":"Figure 7 :Figure 8 :78Figure 7: Growing rate of nodes in hallucinated VSG.","figure_data":"","figure_id":"fig_8","figure_label":"78","figure_type":"figure"},{"figure_caption":"Figure 11 :11Figure 11: Degeneration of the relation node to the labeled edge.","figure_data":"","figure_id":"fig_9","figure_label":"11","figure_type":"figure"},{"figure_caption":"|V SG ← LSG) , L VSH = L N A + L P A .","figure_data":"","figure_id":"fig_10","figure_label":"","figure_type":"figure"},{"figure_caption":"Practical unsupervised MMT requires the avoidance of not only parallel sentences during training, but also the paired image during inference (testing).","figure_data":"","figure_id":"tab_0","figure_label":"1","figure_type":"table"},{"figure_caption":"Results of UMMT on Multi30K data. 'Ours # ': using paired images for testing instead of visual hallucination. 'UMMT * /PVP * ': re-implemented baselines with phrase-level retrieval-based visual hallucination. In the brackets are the improvements of our model over the best-performing baseline(s).","figure_data":"En→Fr En←Fr En→De En←De Avg.Ours50.645.532.033.6 40.4-L CMA49.244.330.932.6 39.3(-1.1)-L REC48.743.930.332.1 38.8(-1.6)-L VCB47.042.228.730.1 37.0(-3.4)-L CPB45.941.627.629.2 36.1(-4.3)-L CMA &L REC 47.242.529.230.9 37.5(-2.9)-L CPB &L VCB 44.640.026.327.7 34.7(-5.7)","figure_id":"tab_3","figure_label":"2","figure_type":"table"},{"figure_caption":"Ablating different learning strategies.","figure_data":"","figure_id":"tab_4","figure_label":"3","figure_type":"table"},{"figure_caption":"Human evaluations are rated on a Likert 10scale, where the results are averaged on En→De and De→En. PVP","figure_data":"","figure_id":"tab_7","figure_label":"5","figure_type":"table"},{"figure_caption":"Table 4 further compares the translation results on WMT corpora under supervised/unsupervised MMT. It is unsurprising to see that MMT models trained with supervision from parallel sentences are overall better than the unsupervised ones. However, our UMMT system effectively narrows the gap between supervised and","figure_data":"67.4±6.8-PVP * (PR)-88.9±5.4Ours86.8±4.791.4±3.8-L CMA76.5±5.580.3±4.3-L REC70.1±5.277.5±4.0-L CMA &L REC68.6±6.172.8±4.8","figure_id":"tab_8","figure_label":"","figure_type":"table"},{"figure_caption":"{ϵ}, including an additional dummy token ϵ indicating no new node to be attached to v i . And if the predicted node is an object node, an additional relation classifier will determine what is the relation label ê′ between v′","figure_data":"iand v i :ê′ ← Softmax D r","figure_id":"tab_10","figure_label":"","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"✓• Unsupervised MMT Chen et al. (2018) Su et al. (2019)Huang et al. (2020) \",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 1 :1Figure 1: Representing the texts and images via language scene graphs (LSG) and visual scene graphs (VSG). In a scene graph, object, attribute, relation nodes are shown in green, orange, purple respectively.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 2 :2Figure 2: The high-level overview of our SG-based UMMT model. During training, src-side sentences with paired images are used as inputs, together with the corresponding LSG and VSG. Testing phase only takes src-side sentences, where the visual hallucination module is activated to generate VSG from text sources.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 :3Figure 3: The illustration of the visual scene hallucination (VSH) module, including two steps of inference.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 4 :4Figure 4: Illustrations of the learning strategies for unsupervised multimodal machine translation.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"two bicycles stand behind two people sitting on the grass near a body of water. hinter zwei Personen, die auf dem gras in der nähe eines gewässers sitzen. stand behind two man with the herbaceous plants near the ocean.)zwei fahrräder stehen hinter zwei mann mit den eingetopften graspflanzen in der nähe des meeres.man in t-shirt and shorts kicking football off tee. und hose, der fußball spielt.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_5\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 5 :5Figure 5: Qualitative results of inference-time image-free UMMT (En→De).\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_6\",\n \"figure_label\": \"5\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 6 :6Figure 6: BLEU scores under different sentence lengths.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_7\",\n \"figure_label\": \"6\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 7 :Figure 8 :78Figure 7: Growing rate of nodes in hallucinated VSG.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_8\",\n \"figure_label\": \"78\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 11 :11Figure 11: Degeneration of the relation node to the labeled edge.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_9\",\n \"figure_label\": \"11\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"|V SG ← LSG) , L VSH = L N A + L P A .\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_10\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Practical unsupervised MMT requires the avoidance of not only parallel sentences during training, but also the paired image during inference (testing).\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Results of UMMT on Multi30K data. 'Ours # ': using paired images for testing instead of visual hallucination. 'UMMT * /PVP * ': re-implemented baselines with phrase-level retrieval-based visual hallucination. In the brackets are the improvements of our model over the best-performing baseline(s).\",\n \"figure_data\": \"En→Fr En←Fr En→De En←De Avg.Ours50.645.532.033.6 40.4-L CMA49.244.330.932.6 39.3(-1.1)-L REC48.743.930.332.1 38.8(-1.6)-L VCB47.042.228.730.1 37.0(-3.4)-L CPB45.941.627.629.2 36.1(-4.3)-L CMA &L REC 47.242.529.230.9 37.5(-2.9)-L CPB &L VCB 44.640.026.327.7 34.7(-5.7)\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Ablating different learning strategies.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Human evaluations are rated on a Likert 10scale, where the results are averaged on En→De and De→En. PVP\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_7\",\n \"figure_label\": \"5\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Table 4 further compares the translation results on WMT corpora under supervised/unsupervised MMT. It is unsurprising to see that MMT models trained with supervision from parallel sentences are overall better than the unsupervised ones. However, our UMMT system effectively narrows the gap between supervised and\",\n \"figure_data\": \"67.4±6.8-PVP * (PR)-88.9±5.4Ours86.8±4.791.4±3.8-L CMA76.5±5.580.3±4.3-L REC70.1±5.277.5±4.0-L CMA &L REC68.6±6.172.8±4.8\",\n \"figure_id\": \"tab_8\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"{ϵ}, including an additional dummy token ϵ indicating no new node to be attached to v i . And if the predicted node is an object node, an additional relation classifier will determine what is the relation label ê′ between v′\",\n \"figure_data\": \"iand v i :ê′ ← Softmax D r\",\n \"figure_id\": \"tab_10\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n }\n]"},"authors":{"kind":"string","value":"Hao Fei; Qian Liu; Meishan Zhang; Min Zhang; Tat-Seng Chua"},"references":{"kind":"list like","value":[{"authors":"Peter Anderson; Xiaodong He; Chris Buehler; Damien Teney; Mark Johnson; Stephen Gould; Lei Zhang","journal":"","ref_id":"b0","title":"Bottom-up and top-down attention for image captioning and visual question answering","year":"2018"},{"authors":"Mikel Artetxe; Gorka Labaka; Eneko Agirre; Kyunghyun Cho","journal":"","ref_id":"b1","title":"Unsupervised neural machine translation","year":"2018"},{"authors":"Dzmitry Bahdanau; Kyunghyun Cho; Yoshua Bengio","journal":"","ref_id":"b2","title":"Neural machine translation by jointly learning to align and translate","year":"2015"},{"authors":"Ondřej Bojar; Christian Buck; Christian Federmann; Barry Haddow; Philipp Koehn; Johannes Leveling; Christof Monz; Pavel Pecina; Matt Post; Herve Saint-Amand; Radu Soricut; Lucia Specia; Aleš Tamchyna","journal":"","ref_id":"b3","title":"Findings of the 2014 workshop on statistical machine translation","year":"2014"},{"authors":"Ondřej Bojar; Rajen Chatterjee; Christian Federmann; Yvette Graham; Barry Haddow; Matthias Huck; Antonio Jimeno Yepes; Philipp Koehn; Varvara Logacheva; Christof Monz; Matteo Negri; Aurélie Névéol; Mariana Neves; Martin Popel; Matt Post; Raphael Rubino; Carolina Scarton; Lucia Specia; Marco Turchi; Karin Verspoor; Marcos Zampieri","journal":"","ref_id":"b4","title":"Findings of the 2016 conference on machine translation","year":"2016"},{"authors":"Yun Chen; Yang Liu; O K Victor; Li","journal":"","ref_id":"b5","title":"Zeroresource neural machine translation with multi-agent communication game","year":"2018"},{"authors":"Alexis Conneau; Guillaume Lample","journal":"","ref_id":"b6","title":"Crosslingual language model pretraining","year":"2019"},{"authors":"Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova","journal":"","ref_id":"b7","title":"BERT: Pre-training of deep bidirectional transformers for language understanding","year":"2019"},{"authors":"Desmond Elliott; Stella Frank; Khalil Sima'an; Lucia Specia","journal":"","ref_id":"b8","title":"Multi30K: Multilingual English-German image descriptions","year":"2016"},{"authors":"Qingkai Fang; Yang Feng","journal":"","ref_id":"b9","title":"Neural machine translation with phrase-level universal visual representations","year":"2022"},{"authors":"Shengqiong Hao Fei; Yafeng Wu; Meishan Ren; Zhang","journal":"","ref_id":"b10","title":"Matching structure for dual learning","year":"2022"},{"authors":"Chengyu Huang; Zheng Zhang; Hao Fei; Lizi Liao","journal":"","ref_id":"b11","title":"Conversation disentanglement with bi-level contrastive learning","year":"2022"},{"authors":"Po-Yao Huang; Xiaojun Chang; Alexander Hauptmann","journal":"","ref_id":"b12","title":"Multi-head attention with diversity for learning grounded multilingual multimodal representations","year":"2019"},{"authors":"Po-Yao Huang; Junjie Hu; Xiaojun Chang; Alexander Hauptmann","journal":"","ref_id":"b13","title":"Unsupervised multimodal neural machine translation with pseudo visual pivoting","year":"2020"},{"authors":"Po-Yao Huang; Frederick Liu; Sz-Rung Shiang; Jean Oh; Chris Dyer","journal":"","ref_id":"b14","title":"Attention-based multimodal neural machine translation","year":"2016"},{"authors":"Justin Johnson; Agrim Gupta; Li Fei-Fei","journal":"","ref_id":"b15","title":"Image generation from scene graphs","year":"2018"},{"authors":"Justin Johnson; Ranjay Krishna; Michael Stark; Li-Jia Li; David A Shamma; Michael S Bernstein; Li Fei-Fei","journal":"","ref_id":"b16","title":"Image retrieval using scene graphs","year":"2015"},{"authors":"Philipp Koehn; Hieu Hoang; Alexandra Birch; Chris Callison-Burch; Marcello Federico; Nicola Bertoldi; Brooke Cowan; Wade Shen; Christine Moran; Richard Zens; Chris Dyer; Ondřej Bojar; Alexandra Constantin; Evan Herbst","journal":"","ref_id":"b17","title":"Moses: Open source toolkit for statistical machine translation","year":"2007"},{"authors":"Ranjay Krishna; Yuke Zhu; Oliver Groth; Justin Johnson; Kenji Hata; Joshua Kravitz; Stephanie Chen; Yannis Kalantidis; Li-Jia Li; David A Shamma; Michael S Bernstein; Li Fei-Fei","journal":"International Journal of Computer Vision","ref_id":"b18","title":"Visual genome: Connecting language and vision using crowdsourced dense image annotations","year":"2017"},{"authors":"Guillaume Lample; Alexis Conneau; Ludovic Denoyer; Marc'aurelio Ranzato","journal":"","ref_id":"b19","title":"Unsupervised machine translation using monolingual corpora only","year":"2018"},{"authors":"Yi Li; Rameswar Panda; Yoon Kim; Chun-Fu Richard Chen; Rogério Feris; David D Cox; Nuno Vasconcelos","journal":"","ref_id":"b20","title":"VALHALLA: visual hallucination for machine translation","year":"2022"},{"authors":"Zuchao Li; Hai Zhao; Rui Wang; Masao Utiyama; Eiichiro Sumita","journal":"","ref_id":"b21","title":"Reference language based unsupervised neural machine translation","year":"2020"},{"authors":"Lajanugen Logeswaran; Honglak Lee","journal":"","ref_id":"b22","title":"An efficient framework for learning sentence representations","year":"2018"},{"authors":"Thang Luong; Hieu Pham; Christopher D Manning","journal":"","ref_id":"b23","title":"Effective approaches to attention-based neural machine translation","year":"2015"},{"authors":"Diego Marcheggiani; Ivan Titov","journal":"","ref_id":"b24","title":"Encoding sentences with graph convolutional networks for semantic role labeling","year":"2017"},{"authors":"Hideki Nakayama; Noriki Nishida","journal":"Machine Translation","ref_id":"b25","title":"Zeroresource machine translation by multimodal encoderdecoder network with multimedia pivot","year":"2017"},{"authors":"Alec Radford; Jong Wook Kim; Chris Hallacy; Aditya Ramesh; Gabriel Goh; Sandhini Agarwal; Girish Sastry; Amanda Askell; Pamela Mishkin; Jack Clark; Gretchen Krueger; Ilya Sutskever","journal":"","ref_id":"b26","title":"Learning transferable visual models from natural language supervision","year":"2021"},{"authors":"Kaiming Shaoqing Ren; Ross B He; Jian Girshick; Sun","journal":"","ref_id":"b27","title":"Faster R-CNN: towards real-time object detection with region proposal networks","year":"2015"},{"authors":"Sebastian Schuster; Ranjay Krishna; Angel Chang; Li Fei-Fei; Christopher D Manning","journal":"","ref_id":"b28","title":"Generating semantically precise scene graphs from textual descriptions for improved image retrieval","year":"2015"},{"authors":"Rico Sennrich; Barry Haddow; Alexandra Birch","journal":"","ref_id":"b29","title":"Improving neural machine translation models with monolingual data","year":"2016"},{"authors":"Lucia Specia; Stella Frank; Khalil Sima'an; Desmond Elliott","journal":"","ref_id":"b30","title":"A shared task on multimodal machine translation and crosslingual image description","year":"2016"},{"authors":"Yuanhang Su; Kai Fan; C.-C. Jay Nguyen Bach; Fei Kuo; Huang","journal":"","ref_id":"b31","title":"Unsupervised multi-modal neural machine translation","year":"2019"},{"authors":"Ilya Sutskever; Oriol Vinyals; Quoc V Le","journal":"","ref_id":"b32","title":"Sequence to sequence learning with neural networks","year":"2014"},{"authors":"Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Lukasz Kaiser; Illia Polosukhin","journal":"","ref_id":"b33","title":"Attention is all you need","year":"2017"},{"authors":"Xinyu Wang; Jingxian Huang; Kewei Tu","journal":"","ref_id":"b34","title":"Second-order semantic dependency parsing with endto-end neural networks","year":"2019"},{"authors":"Yu-Siang Wang; Chenxi Liu; Xiaohui Zeng; Alan Yuille","journal":"","ref_id":"b35","title":"Scene graph parsing as dependency parsing","year":"2018"},{"authors":"Shengqiong Wu; Hao Fei; Yafeng Ren; Donghong Ji; Jingye Li; ; ","journal":"","ref_id":"b36","title":"Learn from syntax: Improving pair-wise aspect and opinion terms extraction with rich syntactic knowledge","year":"2021"},{"authors":"Zhiyong Wu; Lingpeng Kong; Wei Bi; Xiang Li; Ben Kao","journal":"","ref_id":"b37","title":"Good for misconceived reasons: An empirical revisiting on the need for visual context in multimodal machine translation","year":"2021"},{"authors":"Yuanmeng Yan; Rumei Li; Sirui Wang; Fuzheng Zhang; Wei Wu; Weiran Xu","journal":"","ref_id":"b38","title":"ConSERT: A contrastive framework for self-supervised sentence representation transfer","year":"2021"},{"authors":"Xu Yang; Kaihua Tang; Hanwang Zhang; Jianfei Cai","journal":"","ref_id":"b39","title":"Auto-encoding scene graphs for image captioning","year":"2019"},{"authors":"Rowan Zellers; Mark Yatskar; Sam Thomson; Yejin Choi","journal":"","ref_id":"b40","title":"Neural motifs: Scene graph parsing with global context","year":"2018"},{"authors":"Zhuosheng Zhang; Kehai Chen; Rui Wang; Masao Utiyama; Eiichiro Sumita; Zuchao Li; Hai Zhao","journal":"","ref_id":"b41","title":"Neural machine translation with universal visual representation","year":"2020"},{"authors":"Jinhua Zhu; Yingce Xia; Lijun Wu; Di He; Tao Qin; Wengang Zhou; Houqiang Li; Tie-Yan Liu","journal":"","ref_id":"b42","title":"Incorporating BERT into neural machine translation","year":"2020"}],"string":"[\n {\n \"authors\": \"Peter Anderson; Xiaodong He; Chris Buehler; Damien Teney; Mark Johnson; Stephen Gould; Lei Zhang\",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"Bottom-up and top-down attention for image captioning and visual question answering\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Mikel Artetxe; Gorka Labaka; Eneko Agirre; Kyunghyun Cho\",\n \"journal\": \"\",\n \"ref_id\": \"b1\",\n \"title\": \"Unsupervised neural machine translation\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Dzmitry Bahdanau; Kyunghyun Cho; Yoshua Bengio\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"Neural machine translation by jointly learning to align and translate\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Ondřej Bojar; Christian Buck; Christian Federmann; Barry Haddow; Philipp Koehn; Johannes Leveling; Christof Monz; Pavel Pecina; Matt Post; Herve Saint-Amand; Radu Soricut; Lucia Specia; Aleš Tamchyna\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"Findings of the 2014 workshop on statistical machine translation\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"Ondřej Bojar; Rajen Chatterjee; Christian Federmann; Yvette Graham; Barry Haddow; Matthias Huck; Antonio Jimeno Yepes; Philipp Koehn; Varvara Logacheva; Christof Monz; Matteo Negri; Aurélie Névéol; Mariana Neves; Martin Popel; Matt Post; Raphael Rubino; Carolina Scarton; Lucia Specia; Marco Turchi; Karin Verspoor; Marcos Zampieri\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"Findings of the 2016 conference on machine translation\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Yun Chen; Yang Liu; O K Victor; Li\",\n \"journal\": \"\",\n \"ref_id\": \"b5\",\n \"title\": \"Zeroresource neural machine translation with multi-agent communication game\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Alexis Conneau; Guillaume Lample\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"Crosslingual language model pretraining\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"BERT: Pre-training of deep bidirectional transformers for language understanding\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Desmond Elliott; Stella Frank; Khalil Sima'an; Lucia Specia\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"Multi30K: Multilingual English-German image descriptions\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Qingkai Fang; Yang Feng\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"Neural machine translation with phrase-level universal visual representations\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Shengqiong Hao Fei; Yafeng Wu; Meishan Ren; Zhang\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"Matching structure for dual learning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Chengyu Huang; Zheng Zhang; Hao Fei; Lizi Liao\",\n \"journal\": \"\",\n \"ref_id\": \"b11\",\n \"title\": \"Conversation disentanglement with bi-level contrastive learning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Po-Yao Huang; Xiaojun Chang; Alexander Hauptmann\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"Multi-head attention with diversity for learning grounded multilingual multimodal representations\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Po-Yao Huang; Junjie Hu; Xiaojun Chang; Alexander Hauptmann\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"Unsupervised multimodal neural machine translation with pseudo visual pivoting\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Po-Yao Huang; Frederick Liu; Sz-Rung Shiang; Jean Oh; Chris Dyer\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"Attention-based multimodal neural machine translation\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Justin Johnson; Agrim Gupta; Li Fei-Fei\",\n \"journal\": \"\",\n \"ref_id\": \"b15\",\n \"title\": \"Image generation from scene graphs\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Justin Johnson; Ranjay Krishna; Michael Stark; Li-Jia Li; David A Shamma; Michael S Bernstein; Li Fei-Fei\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"Image retrieval using scene graphs\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Philipp Koehn; Hieu Hoang; Alexandra Birch; Chris Callison-Burch; Marcello Federico; Nicola Bertoldi; Brooke Cowan; Wade Shen; Christine Moran; Richard Zens; Chris Dyer; Ondřej Bojar; Alexandra Constantin; Evan Herbst\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"Moses: Open source toolkit for statistical machine translation\",\n \"year\": \"2007\"\n },\n {\n \"authors\": \"Ranjay Krishna; Yuke Zhu; Oliver Groth; Justin Johnson; Kenji Hata; Joshua Kravitz; Stephanie Chen; Yannis Kalantidis; Li-Jia Li; David A Shamma; Michael S Bernstein; Li Fei-Fei\",\n \"journal\": \"International Journal of Computer Vision\",\n \"ref_id\": \"b18\",\n \"title\": \"Visual genome: Connecting language and vision using crowdsourced dense image annotations\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Guillaume Lample; Alexis Conneau; Ludovic Denoyer; Marc'aurelio Ranzato\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"Unsupervised machine translation using monolingual corpora only\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Yi Li; Rameswar Panda; Yoon Kim; Chun-Fu Richard Chen; Rogério Feris; David D Cox; Nuno Vasconcelos\",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"VALHALLA: visual hallucination for machine translation\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Zuchao Li; Hai Zhao; Rui Wang; Masao Utiyama; Eiichiro Sumita\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"Reference language based unsupervised neural machine translation\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Lajanugen Logeswaran; Honglak Lee\",\n \"journal\": \"\",\n \"ref_id\": \"b22\",\n \"title\": \"An efficient framework for learning sentence representations\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Thang Luong; Hieu Pham; Christopher D Manning\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"Effective approaches to attention-based neural machine translation\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Diego Marcheggiani; Ivan Titov\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Encoding sentences with graph convolutional networks for semantic role labeling\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Hideki Nakayama; Noriki Nishida\",\n \"journal\": \"Machine Translation\",\n \"ref_id\": \"b25\",\n \"title\": \"Zeroresource machine translation by multimodal encoderdecoder network with multimedia pivot\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Alec Radford; Jong Wook Kim; Chris Hallacy; Aditya Ramesh; Gabriel Goh; Sandhini Agarwal; Girish Sastry; Amanda Askell; Pamela Mishkin; Jack Clark; Gretchen Krueger; Ilya Sutskever\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"Learning transferable visual models from natural language supervision\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Kaiming Shaoqing Ren; Ross B He; Jian Girshick; Sun\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"Faster R-CNN: towards real-time object detection with region proposal networks\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Sebastian Schuster; Ranjay Krishna; Angel Chang; Li Fei-Fei; Christopher D Manning\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"Generating semantically precise scene graphs from textual descriptions for improved image retrieval\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Rico Sennrich; Barry Haddow; Alexandra Birch\",\n \"journal\": \"\",\n \"ref_id\": \"b29\",\n \"title\": \"Improving neural machine translation models with monolingual data\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Lucia Specia; Stella Frank; Khalil Sima'an; Desmond Elliott\",\n \"journal\": \"\",\n \"ref_id\": \"b30\",\n \"title\": \"A shared task on multimodal machine translation and crosslingual image description\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Yuanhang Su; Kai Fan; C.-C. Jay Nguyen Bach; Fei Kuo; Huang\",\n \"journal\": \"\",\n \"ref_id\": \"b31\",\n \"title\": \"Unsupervised multi-modal neural machine translation\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Ilya Sutskever; Oriol Vinyals; Quoc V Le\",\n \"journal\": \"\",\n \"ref_id\": \"b32\",\n \"title\": \"Sequence to sequence learning with neural networks\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Lukasz Kaiser; Illia Polosukhin\",\n \"journal\": \"\",\n \"ref_id\": \"b33\",\n \"title\": \"Attention is all you need\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Xinyu Wang; Jingxian Huang; Kewei Tu\",\n \"journal\": \"\",\n \"ref_id\": \"b34\",\n \"title\": \"Second-order semantic dependency parsing with endto-end neural networks\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Yu-Siang Wang; Chenxi Liu; Xiaohui Zeng; Alan Yuille\",\n \"journal\": \"\",\n \"ref_id\": \"b35\",\n \"title\": \"Scene graph parsing as dependency parsing\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Shengqiong Wu; Hao Fei; Yafeng Ren; Donghong Ji; Jingye Li; ; \",\n \"journal\": \"\",\n \"ref_id\": \"b36\",\n \"title\": \"Learn from syntax: Improving pair-wise aspect and opinion terms extraction with rich syntactic knowledge\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Zhiyong Wu; Lingpeng Kong; Wei Bi; Xiang Li; Ben Kao\",\n \"journal\": \"\",\n \"ref_id\": \"b37\",\n \"title\": \"Good for misconceived reasons: An empirical revisiting on the need for visual context in multimodal machine translation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Yuanmeng Yan; Rumei Li; Sirui Wang; Fuzheng Zhang; Wei Wu; Weiran Xu\",\n \"journal\": \"\",\n \"ref_id\": \"b38\",\n \"title\": \"ConSERT: A contrastive framework for self-supervised sentence representation transfer\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Xu Yang; Kaihua Tang; Hanwang Zhang; Jianfei Cai\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"Auto-encoding scene graphs for image captioning\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Rowan Zellers; Mark Yatskar; Sam Thomson; Yejin Choi\",\n \"journal\": \"\",\n \"ref_id\": \"b40\",\n \"title\": \"Neural motifs: Scene graph parsing with global context\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Zhuosheng Zhang; Kehai Chen; Rui Wang; Masao Utiyama; Eiichiro Sumita; Zuchao Li; Hai Zhao\",\n \"journal\": \"\",\n \"ref_id\": \"b41\",\n \"title\": \"Neural machine translation with universal visual representation\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Jinhua Zhu; Yingce Xia; Lijun Wu; Di He; Tao Qin; Wengang Zhou; Houqiang Li; Tie-Yan Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b42\",\n \"title\": \"Incorporating BERT into neural machine translation\",\n \"year\": \"2020\"\n }\n]"},"formulas":{"kind":"list like","value":[{"formula_coordinates":[4,126.28,335.6,163.59,11.11],"formula_id":"formula_0","formula_text":"r 1 , • • • , r n = GCN(G) ,(1)"},{"formula_coordinates":[5,141.85,137.39,147.88,24.09],"formula_id":"formula_1","formula_text":"si,j = (r L i ) T • r V j ||r L i || ||r V j || .(2)"},{"formula_coordinates":[5,93.7,209.69,196.03,26.75],"formula_id":"formula_2","formula_text":"L CMA = - i∈LSG x , j * ∈VSG log exp(si,j * /τ ) Z ,(3)"},{"formula_coordinates":[5,106.76,242.15,182.98,20.94],"formula_id":"formula_3","formula_text":"Z = i∈LSG x , k∈VSG, k̸ =j * exp(s i,k /τ ) ,(4)"},{"formula_coordinates":[5,77.21,603.63,212.65,13.09],"formula_id":"formula_4","formula_text":"L VCB = E[-log p yz→x (x|F xz→y (x, z), z)] . (5)"},{"formula_coordinates":[5,95.95,744.15,193.92,29.21],"formula_id":"formula_5","formula_text":"L CPB = E[-log p(x|F xz→y (x, z), z)] + E[-log p(ȳ|F yz→x (ȳ, z), z)] .(6)"},{"formula_coordinates":[5,313.82,225.41,211.32,10.59],"formula_id":"formula_6","formula_text":"L = L CMA + L REC + L VCB + L CPB + L VSH . (7)"},{"formula_coordinates":[12,347.44,663.13,99.3,14],"formula_id":"formula_7","formula_text":"V na i = {• • • , v k , • • • }."},{"formula_coordinates":[12,351.92,702.34,123.17,58.75],"formula_id":"formula_8","formula_text":"α n k = exp r i • r k v * k ∈V na i exp r i • r * k h na i = r i + k α n k • r k ."},{"formula_coordinates":[13,70.47,141.62,170.55,31.94],"formula_id":"formula_9","formula_text":"v′ i ← Softmax D na (FFN(h na i )) , where D na = D o ∪ D a ∪"},{"formula_coordinates":[13,76.82,359.33,206.36,54.45],"formula_id":"formula_10","formula_text":"h pa i-j = Sigmoid( r i 1 T (r j ) T W r i-j 1 ) , ê′ i,j ← Softmax D pa (FFN(h pa i-j )) ,"},{"formula_coordinates":[13,371.09,185.95,82.33,31.07],"formula_id":"formula_11","formula_text":"s f i,j = (r L i ) T • r V j ||r L i || ||r V j ||"}],"string":"[\n {\n \"formula_coordinates\": [\n 4,\n 126.28,\n 335.6,\n 163.59,\n 11.11\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"r 1 , • • • , r n = GCN(G) ,(1)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 141.85,\n 137.39,\n 147.88,\n 24.09\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"si,j = (r L i ) T • r V j ||r L i || ||r V j || .(2)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 93.7,\n 209.69,\n 196.03,\n 26.75\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"L CMA = - i∈LSG x , j * ∈VSG log exp(si,j * /τ ) Z ,(3)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 106.76,\n 242.15,\n 182.98,\n 20.94\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"Z = i∈LSG x , k∈VSG, k̸ =j * exp(s i,k /τ ) ,(4)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 77.21,\n 603.63,\n 212.65,\n 13.09\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"L VCB = E[-log p yz→x (x|F xz→y (x, z), z)] . (5)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 95.95,\n 744.15,\n 193.92,\n 29.21\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"L CPB = E[-log p(x|F xz→y (x, z), z)] + E[-log p(ȳ|F yz→x (ȳ, z), z)] .(6)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 313.82,\n 225.41,\n 211.32,\n 10.59\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"L = L CMA + L REC + L VCB + L CPB + L VSH . (7)\"\n },\n {\n \"formula_coordinates\": [\n 12,\n 347.44,\n 663.13,\n 99.3,\n 14\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"V na i = {• • • , v k , • • • }.\"\n },\n {\n \"formula_coordinates\": [\n 12,\n 351.92,\n 702.34,\n 123.17,\n 58.75\n ],\n \"formula_id\": \"formula_8\",\n \"formula_text\": \"α n k = exp r i • r k v * k ∈V na i exp r i • r * k h na i = r i + k α n k • r k .\"\n },\n {\n \"formula_coordinates\": [\n 13,\n 70.47,\n 141.62,\n 170.55,\n 31.94\n ],\n \"formula_id\": \"formula_9\",\n \"formula_text\": \"v′ i ← Softmax D na (FFN(h na i )) , where D na = D o ∪ D a ∪\"\n },\n {\n \"formula_coordinates\": [\n 13,\n 76.82,\n 359.33,\n 206.36,\n 54.45\n ],\n \"formula_id\": \"formula_10\",\n \"formula_text\": \"h pa i-j = Sigmoid( r i 1 T (r j ) T W r i-j 1 ) , ê′ i,j ← Softmax D pa (FFN(h pa i-j )) ,\"\n },\n {\n \"formula_coordinates\": [\n 13,\n 371.09,\n 185.95,\n 82.33,\n 31.07\n ],\n \"formula_id\": \"formula_11\",\n \"formula_text\": \"s f i,j = (r L i ) T • r V j ||r L i || ||r V j ||\"\n }\n]"}}},{"rowIdx":387,"cells":{"doi":{"kind":"string","value":""},"pub_date":{"kind":"string","value":"2023-06-04"},"sections":{"kind":"list like","value":[{"figure_ref":[],"heading":"Introduction","publication_ref":["b5","b37","b23","b3","b7","b26","b17","b22","b34","b32","b15"],"table_ref":[],"text":"Relation extraction (RE) aims at extracting from the plain texts the meaningful entity mentions paired with semantic relations. One widelyacknowledged key bottleneck of RE is called the long-range dependence (LRD) issue, i.e., the decay of dependence clues of two mention entities with increasing distance in between (Culotta and Sorensen, 2004;Zhang et al., 2018;Fei et al., 2021). Fortunately, prior work extensively reveals that the syntactic dependency trees help resolve LRD issue effectively, by taking advantage of the close relevance between the dependency structure and the relational RE pair (Miwa and Bansal, 2016;Can et al., 2019). In cross-lingual RE, likewise, the universal dependency trees (de Marneffe et al., 2021) are leveraged as effective language-persistent features Copying current node from SRC or TGT trees into forest, if the node has no alignment in opposite tree at this layer.\n在In马歇尔Marshall的家乡德国城hometown Germantown他是位大胆的法官，,他 于1801年in the year of 1801被by亚当斯提名为to be首席大法官。 Synthesized code-mixed text Figure 2: A real example to construct a code-mixed UD forest. The raw sentence is selected from ACE05 data. We exemplify the transfer from English (EN) to Chinese (ZH).\nbe exacerbated in UD-based model transfer, cf. Fig. 1(a). Given that UD has a universal annotation standard, inevitably, there is still a syntax discrepancy between the two languages due to their intrinsic linguistic nature. We show (cf. §3 for more discussion) that between the parallel sentences in English and Arabic, around 30% words are misaligned and over 35% UD word-pairs have no correspondence. Such structural discrepancies consequently undermine the model transfer efficacy.\nOne alternative solution is using annotation projection (Padó and Lapata, 2009;Kim et al., 2010;McDonald et al., 2013;Xiao and Guo, 2015). The main idea is directly synthesizing the pseudo TGTside training data, so that the TGT-side linguistic features (i.e., UD trees) are well preserved. However, it could be a double side of the sword in the annotation projection paradigm. It manages to learn the language-specific features, while at the cost of losing some high-efficient structural knowledge from SRC-side UD, thus leading to the SRC-biased UD feature transfer. As illustrated in Fig. 1(b), the dependence paths in the SRC UD tree that effectively solves the LRD issues for the task are sacrificed when transforming the SRC tree into the TGT tree. This motivates us to pursue an unbiased and holistic UD-based XRE transfer by considering both the SRC and TGT UD syntax features. To reach the goal, in this work, we propose combining the view of model transfer and annotation projection paradigm, and constructing a type of codemixed UD forests. Technically, we first project the SRC training instances and TGT predicting instances into the opposite languages, respectively.\nThen, we parse the parallel UD trees of both sides respectively via existing UD parsers. Next, merge each pair of SRC and TGT UD trees together into the code-mixed UD forest, in which the wellaligned word pairs are merged to the TGT ones in the forest, and the unaligned words will all be kept in the forest. With these code-mixed syntactic features, the gap between training and predicting phases can be closed, as depicted in Fig. 1(c).\nWe encode the UD forest with the graph attention model (GAT; Velickovic et al., 2018) for feature encoding. We perform experiments on the representative XRE benchmark, ACE05 (hristopher Walker et al., 2006), where the transfer results from English to Chinese and Arabic show that the proposed code-mixed forests bring significant improvement over the current best-performing UD-based system, obtaining the new SoTA results. Further analyses verify that 1) the code-mixed UD forests help maintain the debiased cross-lingual transfer of RE task, and 2) the larger the difference between SRC and TGT languages, the bigger the boosts offered by code-mixed forests. To our knowledge, we are the first taking the complementary advantages of annotation projection and model transfer paradigm for unbiased XRE transfer. We verify that the gap between training and predicting of UD-based XRE can be bridged by synthesizing a type of code-mixed UD forests. The resource can be found at https://github.com/ scofield7419/XLSIE/."},{"figure_ref":[],"heading":"Related Work","publication_ref":["b4","b35","b33","b10","b21","b30","b38","b18","b25","b2","b24","b6","b20","b39","b19","b16","b1","b28","b28","b27","b36","b36"],"table_ref":[],"text":"Different from the sequential type of information extraction (IE), e.g., named entity recognition (NER) (Cucerzan and Yarowsky, 1999), RE not only detects the mentions but also recognizes the semantic relations between mentions. RE has long received extensive research attention within the last decades (Zelenko et al., 2002). Within the community, research has revealed that the syntactic dependency trees share close correlations with RE or broad-covering information extraction tasks in structure (Fei et al., 2021;Wu et al., 2021;Fei et al., 2022), and thus the former is frequently leveraged as supporting features for enhancing RE. In XRE, the key relational features between words need to be transferred between languages, which motivates the incorporation of UD tree features that have consistent annotations and principles across various languages. Thus, UD-based systems extensively achieve the current SoTA XRE (Lu et al., 2020;Taghizadeh and Faili, 2021;Zhang et al., 2021). This work inherits the prior wisdom, and leverages the UD features.\nModel transfer (Kozhevnikov and Titov, 2013;Ni and Florian, 2019;Fei et al., 2020b) and annotation projection (Björkelund et al., 2009;Mulcaire et al., 2018;Daza and Frank, 2019;Fei et al., 2020a;Lou et al., 2022) are two mainstream avenues in structural cross-lingual transfer track. The former trains a model on SRC annotations and them make predictions with TGT instances, i.e., transferring the shared language-invariant features. The latter directly synthesizes the pseudo training instances in TGT language based on some parallel sentences, in which the TGT-specific features are retained to the largest extent. As we indicated earlier, in both two paradigms the UD tree features can be unfortunately biased during the transfer, thus leading to the underutilization of UD resource. This work considers a holistic viewpoint, integrating both the two cross-lingual transfer schemes and combining both the SRC and TGT syntax trees by code mixing.\nSeveral prior studies have shown that combining the raw SRC and pseudo TGT (from projection) data for training helps better transfer. It is shown that although the two data are semantically identical, SRC data still can offer some complementary language-biased features (Fei et al., 2020a,b;Zhen et al., 2021). Yet we emphasize that different from regular cross-lingual text classification or sequential prediction, XRE relies particularly on the syntactic structure features, e.g., UD, and thus needs a more fine-grained approach for SRC-TGT data ensembling, instead of simply instance stack-ing. Thus, we propose merging the SRC and TGT syntax trees into the code-mixed forests.\nCode mixing has been explored in several different NLP applications (Labutov and Lipson, 2014;Joshi et al., 2016;Banerjee et al., 2018;Samanta et al., 2019), where the core idea is creating data piece containing words from different languages simultaneously. For example, Samanta et al. (2019) introduce a novel data augmentation method for enhancing the recognition of code-switched sentiment analysis, where they replace the constituent phrases with code-mixed alternatives. Qin et al. (2020) propose generating code-switching data to augment the existing multilingual language models for better zero-shot cross-lingual tasks. While we notice that most of the works focus on the development of code-mixed sequential texts, this work considers the one for structural syntax trees. Our work is partially similar to Zhang et al. (2019) on the code-mixed UD tree construction. But ours differentiate theirs in that Zhang et al. (2019) target better UD parsing itself, while we aim to improve downstream tasks.\n3 Observations on UD Bias"},{"figure_ref":[],"heading":"Bias Source Analysis","publication_ref":[],"table_ref":[],"text":"As mentioned, even though UD trees define consistent annotations across languages, it still falls short on wiping all syntactic bias. This is inevitably caused by the underlying linguistic disparity deeply embedded in the language itself. Observing the linguistic discrepancies between different languages, we can summarize them into following three levels:\n1) Word-level Changes.\n• Word number. The words referring to same semantics in different languages vary, e.g., in English one single-token word may be translated in Chinese with more than one token. • Part of speech. In different languages a parallel lexicon may come with different part of speech. • Word order. Also it is a common case that the word order varies among parallel sentences in different languages.\n2) Phrase-level Change.\n• Modification type. A modifier of a phrasal constituent can be changed when translating into another languages. For example, in English, 'in the distance' is often an adverbial modifier, while its counterpart in Chinese '遥 远的' plays a role of an attribute modifier. • Change of pronouns. English grammar has strict structure, while in some other languages the grammar structures may not strict. For example, in English, it is often case to use relative pronouns (e.g., which, that, who) to refer to the prior mentions, while in other languages, such as Chinese, the personal pronouns (e.g., which, that, who) will be used to refer the prior mentions. • Constituency order change. Some constituent phrases will be reorganized and reordered from one language to another language, due to the differences in grammar rules.\n3) Sentence-level Change.\n• Transformation between active and passive sentences. In English it could be frequent to use the passive forms of sentences, while being translated into other languages the forms will be transformed into active types, where the words and phrases in the whole sentences can be reversed. • Transformation between clause and main sentence. In English the attributive clauses and noun clauses are often used as subordinate components, while they can be translated into two parallel clauses in other languages. • Change of reading order of sentences. The majority of the languages in this world have the reading order of from-left-to-right, such as English, French, etc. But some languages, e.g., under Afro-Asiatic family, Arabic, Hebrew, Persian, Sindhi and Urdu languages read from right to left."},{"figure_ref":[],"heading":"UD Bias Statistics","publication_ref":[],"table_ref":[],"text":"In Fig. 3 we present the statistics of such bias between the parallel UD trees in different languages, such as the misaligned words, mismatched UD (w ↷ i w j ) pair and UD path of (e ↷ s • • • ↷ e o ) relational pair. Fig. 3(a) reveals that languages under different families show distinct divergences. And the more different of languages, the greater the divergences (e.g., English to Arabic). Fig. 3(b) indicates that complex sentences (e.g., compound sentences) bring larger bias; and in the real world, complex sentences are much more ubiquitous than simple ones. Also, the mismatch goes worse when the UD core predicates are nouns instead of verbs. "},{"figure_ref":[],"heading":"Code-mixed UD Forest Construction","publication_ref":[],"table_ref":[],"text":"To eliminate such discrepancies for unbiased UDfeature transfer, we build the code-mixed UD forests, via the following six steps."},{"figure_ref":[],"heading":"▶","publication_ref":[],"table_ref":[],"text":"Step 1: translating a sentence x Src in SRC language to the one x Tgt in TGT language.1 This step is to generate a pseudo parallel sentence pair in both TGT and SRC languages. We accomplish this by using the state-of-the-art Google Translation API. 2 We denote the parallel sentences as or ."},{"figure_ref":[],"heading":"▶","publication_ref":[],"table_ref":[],"text":"Step 2: obtaining the word alignment scores. Meanwhile, we employ the Awesome-align toolkit3 to obtain the word alignment confidence M ={m i↔j } between word pair w i ∈ x Src and w j ∈ x Tgt in parallel sentences."},{"figure_ref":[],"heading":"▶","publication_ref":[],"table_ref":[],"text":"Step 3: parsing UD trees for parallel sentences. Then, we use the UD parsers in SRC and SRC languages respectively to parse the UD syntax trees for two parallel sentences, respectively. We adopt the UDPipe 4 as our UD parsers, which are trained separately on different UD annotated data 5 . We denote the SRC UD tree as T Src , and the pseudo TGT UD tree as T Tgt . Note that the UD trees in all languages share the same dependency labels, Algorithm 1 Process of constructing code-mixed UD forests Input: T SRC , T TGT , M , threshold θ, empty forest F = Φ. Output: Code-mixed UD forest F.\n1: def Construct (T SRC , T TGT , M , F)\n▷ breadth-first top-down traverse.\n2: is_root = True ▷ a flag for traversing the predicate only once."},{"figure_ref":[],"heading":"3:","publication_ref":[],"table_ref":[],"text":"F.w cur = ROOT ▷ creating ROOT node for F. if is_root then 7: return aligned_pairs, nonaligned_nodes i.e., with the same (as much as possible) annotation standards. In Appendix §A we list the dependency labels which are the commonly occurred types.\nw merged ="},{"figure_ref":[],"heading":"▶","publication_ref":[],"table_ref":[],"text":"Step 4: projecting and merging the labels of training data. For the training set, we also need to project the annotations (relational subjectobject pairs) of sentences in SRC languages to TGT pseudo sentences. Note that this step is not needed for the testing set. The projection is based on the open source6 , during which the word alignment scores at step-2 are used. We can denote the SRC annotation as y, and the pseudo TGT label as y.\nWe then merge the annotation from both SRC and TGT viewpoints, into the code-mixed one Y , for later training use. Specifically, for the node that is kept in the final code-mixed forest, we will keep its labels; and for those nodes that are filtered, the annotations are replaced by their correspondences.\n▶ Step 5: merging the SRC and TGT UD trees into a code-mixed forest. Finally, based on the SRC UD tree and the TGT UD tree, we construct the code-mixed UD forest. We mainly perform breadth-first top-down traversal over each pair of nodes T Src and T Tgt , layer by layer. The traversal starts from their ROOT node. We first create a ROOT node as the initiation of the codemixed forest. We design two types of actions for the forest merging process:\n• Merging current pair of nodes w i ∈ T Src from SRC tree and w j ∈ T Tgt from TGT tree into the forest F, if the current two nodes are confidently aligned at same dependency layer. We check the word alignment confidence m i↔j between the two nodes, and if the confidence is above a pre-defined threshold θ, i.e., m i↔j > θ, we treat them as confidently aligned.\n• Copying current node from SRC tree T Src or TGT tree T Tgt into the forest F, once the node has no significant alignment in the opposite tree at this layer. In Algorithm 1 we formulate in detail the process of code-mixed forest construction. Also, we note that when moving the nodes from two separate UD trees into the forest, the attached dependency labels are also copied. When two nodes are merged, we only choose the label of the TGT-side node. Finally, the resulting forest F looks like code-mixing, and is structurally compact."},{"figure_ref":[],"heading":"▶","publication_ref":[],"table_ref":[],"text":"Step 6: assembling code-mixed texts. Also we need to synthesize a code-mixed text X based on the raw SRC text x Src and the pseudo TGT text x Tgt . The code-mixed text X will also be used as inputs together with the forest, into the forest encoder. We directly replace the SRC words with the TGT words that have been determined significantly aligned at Step-5."},{"figure_ref":[],"heading":"XRE with Code-mixed UD Forest","publication_ref":["b32","b8"],"table_ref":[],"text":"Along with the UD forest F Src , we also assemble the code-mixed sequential text X Src from the SRC and translated pseudo-TGT sentences (i.e., x Src and x Tgt ), and the same for the TGT sentences X Tgt . An XRE system, being trained with SRC-side annotated data (, Y Src ), needs to determine the label Y Tgt of relational pair e r ↷ s e o given a TGT sentence and UD forest ().\nThe XRE system takes as input X={w i } n and F. We use the multilingual language model (MLM) for representing the input code-mixed sentence X:\nH = {h1, • • • , hn} = MLM(X) , (1\n)\nwhere X is the code-mixed sentential text. We then formulate the code-mixed forest F as a graph, G=, where E={e i,j } n×n is the edge between word pair (i.e., initiated with e i,j =0/1, meaning dis-/connecting), V ={w i } n are the words. We main the node embeddings r i for each node v i . We adopt the GAT model (Velickovic et al., 2018) for the backbone forest encoding:\nρi,j = Softmax(GeLU(U T [W1ri; W2rj])) ,(2)\nui = σ( j ρi,jW3r 1 j ) ,(3)\nwhere W 3/4/5 and U are all trainable parameters. σ is the sigmoid function. GeLU is a Gaussian error linear activation function. Note that the firstlayer representations of r i is initialized with h i .\nH and U are then concatenated as the resulting feature representation:\nĤ = H ⊕ U .(4)\nXRE aims to determine the semantic relation labels between two given mention entities. For example, given a sentence 'John Smith works at Google', RE should identify that there is a relationship of \"works at\" between the entities \"John Smith\" and \"Google\". Our XRE model needs to predict the relation label y. We adopt the biaffine decoder (Dozat and Manning, 2017) to make prediction:\ny = Softmax(h T s • W1 • ho + W2 • Pool( Ĥ)) . (5)\nHere both h s and h o are given.\n6 Experiments"},{"figure_ref":[],"heading":"Setups","publication_ref":["b15","b0","b0"],"table_ref":["tab_1"],"text":"We consider the ACE05 (hristopher Walker et al., 2006) dataset, which includes English (EN), Chinese (ZH) and Arabic (AR). We give the data statistics in Table 1 The multilingual BERT is used. We use two-layer GAT for forest encoding, with a 768-d hidden size. We mainly consider the transfer from EN to one other language. Following most cross-lingual works (Fei et al., 2020b;Ahmad et al., 2021), we train the XRE model with fixed 300 iterations without early-stopping. We make comparisons between three setups: 1) using only raw SRC training data with the model transfer, 2) using only the pseudo TGT (via annotation projection) for training, and 3) using both the above SRC and TGT data. Each setting uses both the texts and UD tree (or forest) features. The baseline uses the same GAT model for syntax encoding, marked as Syn-Baseline. For setup 1)&2) we also test the transfer with only text inputs, removing the syntax features, marked as TxtBaseline. Besides, for setup 1) we cite current SoTA performances as references. We use F1 to measure the RE performance, following Ahmad et al. (2021). All experiments are undergone five times and the average value is reported."},{"figure_ref":[],"heading":"Data Inspection","publication_ref":[],"table_ref":["tab_4"],"text":"We also show in Table 3 the differences in average sequential and syntactic (shortest dependency path) distances between the subjects and objects of the relational triplets. As seen, the syntactic distances between subject-object pairs are clearly shortened in the view of syntactic dependency trees, which indicates the imperative to incorporate the tree structure features. However, the syntactic distances between different languages vary, i.e., more complex languages have longer syntactic distances. Such discrepancy reflects the necessity of employing our proposed UD debiasing methods to bridge the gap."},{"figure_ref":[],"heading":"Main Results","publication_ref":[],"table_ref":["tab_3"],"text":"From Table 2, we can see that UD features offer exceptional boosts (M1 vs. M2, M4 vs. M5). And annotation projection methods outperform model transfer ones (i.e., M1&M2&M3 vs. M4&M5) by offering direct TGT-side features. Interestingly, in both two transfer paradigms, the improvements from UD become weak on the language pairs with bigger divergences. For example, the improvement on EN→DE outweighs the ones on EN→ZH. Furthermore, using our proposed code-mixed syntax forests is significantly better than using standalone SRC or TGT (or the simple combination) UD features (M7 vs. M2&M5&M6) on all transfers with big margins. For example, our system outperforms SoTA UD-based systems with averaged +4.8%(=67.2-62.4) F1. This evidently verifies the necessity to create the code-mixed forests, i.e., bringing unbiased UD features for transfer. Also, we find that the more the difference between the two languages, the bigger the improvements from forests. The ablation of code-mixed texts also shows the contribution of the sequential textual features, which indirectly demonstrates the larger efficacy of the structural code-mixed UD forests."},{"figure_ref":["fig_3"],"heading":"Probing Unbiasedness of Code-mixed UD Forest","publication_ref":[],"table_ref":[],"text":"Fig. 4 plots the change of the syntax distances of RE pairs during the transfer with different syntax trees. We see that the use of SRC UD trees shows clear bias (with larger inclination angles) during the transfer, while the use of TGT UD trees and codemixed forests comes with less change of syntax distances. Also, we can see from the figure that the inference paths between objects and subjects of RE tasks are clearly shortened with the forests (in orange color), compared to the uses of SRC/TGT UD trees."},{"figure_ref":[],"heading":"Change during Code-mixed UD Forest Merge","publication_ref":[],"table_ref":["tab_6"],"text":"Here we make statistics of how many words are merged and kept during the UD tree merging, respectively. The statistics are shown in Table 4. We can see that the distance between EN-ZH is shorter than that between EN-AR. For example, the length of code-mixed EN-ZH UD forests (sentences) is 31.63, while for EN-AR the length is 40.44. Also, EN-ZH UD forests have a higher to 21.4% merging rate, while EN-AR UD forests have 16.6% merging rate. This demonstrates that the more divergences of languages, the lower the merging rate of the code-mixed forest."},{"figure_ref":["fig_4"],"heading":"Impacts of θ on Controlling the Quality of Merged Forest","publication_ref":[],"table_ref":[],"text":"In §4 of step-5, we describe that we use a threshold θ to control the aligning during the UD tree merging. Intuitively, the large the threshold θ, the lower the alignment rate. When θ → 0, most of the SRC and TGT nodes in two parallel UD trees can find their counterparts but the alignments are most likely to be wrong, thus hurting the quality of the resulting code-mixed UD forests. When θ → 1, none of the SRC and TGT nodes in two parallel UD trees can be aligned, and both two UD trees are copied and co-existed in the resulting code-mixed UD forests. In such case, the integration of such forests is equivalent to the annotation projection methods where we directly use both the raw SRC UD feature and the translated pseudo TGT UD tree feature. In Fig. 5 we now study the influences of using different code-mixed forest features generated with different merging rates (θ). We see that with a threshold of θ=0.5, the performances are consistently the best."},{"figure_ref":[],"heading":"Performances on Different Types of Sentence","publication_ref":[],"table_ref":["tab_7"],"text":"In Table 5 we show the results under different types of sentences. We directly select 500 short sentences (with length < 12) as simple sentences; and select 500 lengthy sentences (with length > 35) as complex sentences. As can be seen, with the code-mixed forest features, the system shows very notable improvements in complex sentences. For example, on the EN→ZH we obtain 15.9(=57.2-41.3)% F1 improvement, and on the EN→AR the boost increases strikingly to 25.2(=67.3-42.1)% F1. However, such enhancements are not very significant in handling simple sentences. This indicates that the code-mixed UD forest features can espe- cially enhance the effectiveness on the hard case, i.e., the transfer between those pairs with greater divergences will receive stronger enhancements from our methods."},{"figure_ref":[],"heading":"Conclusion and Future Work","publication_ref":["b14","b13"],"table_ref":[],"text":"Universal dependencies (UD) have been served as effective language-consistent syntactic features for cross-lingual relation extraction (XRE). In this work, we reveal the intrinsic language discrepancies with respect to the UD structural annotations, which limit the utility of the UD features. We enhance the efficacy of UD features for an unbiased UD-based transfer, by constructing code-mixed UD forests from both the source and target UD trees. Experimental results demonstrate that the UD forests effectively debias the syntactic disparity in the UD-based XRE transfer, especially for those language pairs with larger gaps. Leveraging the syntactic dependency features is a long-standing practice for strengthening the performance of RE tasks. In this work, we propose a novel type of syntactic feature, code-mixed UD forests, for cross-lingual relation extraction. We note that this feature can be applied broadly to other cross-lingual structured information extraction tasks that share the same task definition besides RE, such as event detection (ED) (Halpin and Moore, 2006) and semantic role labeling (SRL) (Gildea and Jurafsky, 2000). Besides, how to fur-ther increase the utility of the UD forests with a better modeling method is a promising research direction, i.e., filtering the noisy structures in the UD forests.\n2.5-4.0 9 as well as GNU GPL 3.0 10 . Our use of UD treebanks comply with all these license terms is at non-commercial purpose. The software tools (i.e., UDPipe parsers) are provided under GNU GPL V2. Our use of UDPipe tools complies with the term."},{"figure_ref":[],"heading":"Acknowledgments","publication_ref":[],"table_ref":[],"text":"This research is supported by the National Natural Science Foundation of China (No. 62176180), and also the Sea-NExT Joint Lab."},{"figure_ref":[],"heading":"Limitations","publication_ref":[],"table_ref":[],"text":"Although showing great prominence, our proposed method has the following limitations. First of all, our method relies on the availability of annotated UD trees of both the source and target languages, as we need to use the annotations to parse the syntax trees for our own sentences. Fortunately, UD project covers over 100 languages, where most of the languages, even the minor ones, will have the UD resources. At the same time, our method will be influenced by the quality of UD parsers. Secondly, our method also uses the external translation systems to produce the pseudo parallel sentences, where our method may largely subject to the quality of the translators. Again luckily, current neural machine translation systems have been well developed and established, i.e., Google Translation. Only when handling very scare languages where the current translation systems fail to give satisfactory performances, our method will fail."},{"figure_ref":[],"heading":"Ethics Statement","publication_ref":[],"table_ref":[],"text":"In this work, we construct a type of code-mixed UD forest based on the existing UD resources. We note that all the data construction has been accomplished automatically, and we have not created any new annotations with additional human labor. Specifically, we use the UD v2.10 resource, which is a collection of linguistic data and tools that are open-sourced. Each of treebanks of UD has its own license terms, including the CC BY-SA 4.0 8 and CC BY-NC-SA"},{"figure_ref":[],"heading":"A The universal dependency labels","publication_ref":[],"table_ref":[],"text":"In Table 6, we list the dependency labels which are the commonly occurred types. Please refer to Stanford dependency 11 "}],"string":"[\n {\n \"figure_ref\": [],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b5\",\n \"b37\",\n \"b23\",\n \"b3\",\n \"b7\",\n \"b26\",\n \"b17\",\n \"b22\",\n \"b34\",\n \"b32\",\n \"b15\"\n ],\n \"table_ref\": [],\n \"text\": \"Relation extraction (RE) aims at extracting from the plain texts the meaningful entity mentions paired with semantic relations. One widelyacknowledged key bottleneck of RE is called the long-range dependence (LRD) issue, i.e., the decay of dependence clues of two mention entities with increasing distance in between (Culotta and Sorensen, 2004;Zhang et al., 2018;Fei et al., 2021). Fortunately, prior work extensively reveals that the syntactic dependency trees help resolve LRD issue effectively, by taking advantage of the close relevance between the dependency structure and the relational RE pair (Miwa and Bansal, 2016;Can et al., 2019). In cross-lingual RE, likewise, the universal dependency trees (de Marneffe et al., 2021) are leveraged as effective language-persistent features Copying current node from SRC or TGT trees into forest, if the node has no alignment in opposite tree at this layer.\\n在In马歇尔Marshall的家乡德国城hometown Germantown他是位大胆的法官，,他 于1801年in the year of 1801被by亚当斯提名为to be首席大法官。 Synthesized code-mixed text Figure 2: A real example to construct a code-mixed UD forest. The raw sentence is selected from ACE05 data. We exemplify the transfer from English (EN) to Chinese (ZH).\\nbe exacerbated in UD-based model transfer, cf. Fig. 1(a). Given that UD has a universal annotation standard, inevitably, there is still a syntax discrepancy between the two languages due to their intrinsic linguistic nature. We show (cf. §3 for more discussion) that between the parallel sentences in English and Arabic, around 30% words are misaligned and over 35% UD word-pairs have no correspondence. Such structural discrepancies consequently undermine the model transfer efficacy.\\nOne alternative solution is using annotation projection (Padó and Lapata, 2009;Kim et al., 2010;McDonald et al., 2013;Xiao and Guo, 2015). The main idea is directly synthesizing the pseudo TGTside training data, so that the TGT-side linguistic features (i.e., UD trees) are well preserved. However, it could be a double side of the sword in the annotation projection paradigm. It manages to learn the language-specific features, while at the cost of losing some high-efficient structural knowledge from SRC-side UD, thus leading to the SRC-biased UD feature transfer. As illustrated in Fig. 1(b), the dependence paths in the SRC UD tree that effectively solves the LRD issues for the task are sacrificed when transforming the SRC tree into the TGT tree. This motivates us to pursue an unbiased and holistic UD-based XRE transfer by considering both the SRC and TGT UD syntax features. To reach the goal, in this work, we propose combining the view of model transfer and annotation projection paradigm, and constructing a type of codemixed UD forests. Technically, we first project the SRC training instances and TGT predicting instances into the opposite languages, respectively.\\nThen, we parse the parallel UD trees of both sides respectively via existing UD parsers. Next, merge each pair of SRC and TGT UD trees together into the code-mixed UD forest, in which the wellaligned word pairs are merged to the TGT ones in the forest, and the unaligned words will all be kept in the forest. With these code-mixed syntactic features, the gap between training and predicting phases can be closed, as depicted in Fig. 1(c).\\nWe encode the UD forest with the graph attention model (GAT; Velickovic et al., 2018) for feature encoding. We perform experiments on the representative XRE benchmark, ACE05 (hristopher Walker et al., 2006), where the transfer results from English to Chinese and Arabic show that the proposed code-mixed forests bring significant improvement over the current best-performing UD-based system, obtaining the new SoTA results. Further analyses verify that 1) the code-mixed UD forests help maintain the debiased cross-lingual transfer of RE task, and 2) the larger the difference between SRC and TGT languages, the bigger the boosts offered by code-mixed forests. To our knowledge, we are the first taking the complementary advantages of annotation projection and model transfer paradigm for unbiased XRE transfer. We verify that the gap between training and predicting of UD-based XRE can be bridged by synthesizing a type of code-mixed UD forests. The resource can be found at https://github.com/ scofield7419/XLSIE/.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Work\",\n \"publication_ref\": [\n \"b4\",\n \"b35\",\n \"b33\",\n \"b10\",\n \"b21\",\n \"b30\",\n \"b38\",\n \"b18\",\n \"b25\",\n \"b2\",\n \"b24\",\n \"b6\",\n \"b20\",\n \"b39\",\n \"b19\",\n \"b16\",\n \"b1\",\n \"b28\",\n \"b28\",\n \"b27\",\n \"b36\",\n \"b36\"\n ],\n \"table_ref\": [],\n \"text\": \"Different from the sequential type of information extraction (IE), e.g., named entity recognition (NER) (Cucerzan and Yarowsky, 1999), RE not only detects the mentions but also recognizes the semantic relations between mentions. RE has long received extensive research attention within the last decades (Zelenko et al., 2002). Within the community, research has revealed that the syntactic dependency trees share close correlations with RE or broad-covering information extraction tasks in structure (Fei et al., 2021;Wu et al., 2021;Fei et al., 2022), and thus the former is frequently leveraged as supporting features for enhancing RE. In XRE, the key relational features between words need to be transferred between languages, which motivates the incorporation of UD tree features that have consistent annotations and principles across various languages. Thus, UD-based systems extensively achieve the current SoTA XRE (Lu et al., 2020;Taghizadeh and Faili, 2021;Zhang et al., 2021). This work inherits the prior wisdom, and leverages the UD features.\\nModel transfer (Kozhevnikov and Titov, 2013;Ni and Florian, 2019;Fei et al., 2020b) and annotation projection (Björkelund et al., 2009;Mulcaire et al., 2018;Daza and Frank, 2019;Fei et al., 2020a;Lou et al., 2022) are two mainstream avenues in structural cross-lingual transfer track. The former trains a model on SRC annotations and them make predictions with TGT instances, i.e., transferring the shared language-invariant features. The latter directly synthesizes the pseudo training instances in TGT language based on some parallel sentences, in which the TGT-specific features are retained to the largest extent. As we indicated earlier, in both two paradigms the UD tree features can be unfortunately biased during the transfer, thus leading to the underutilization of UD resource. This work considers a holistic viewpoint, integrating both the two cross-lingual transfer schemes and combining both the SRC and TGT syntax trees by code mixing.\\nSeveral prior studies have shown that combining the raw SRC and pseudo TGT (from projection) data for training helps better transfer. It is shown that although the two data are semantically identical, SRC data still can offer some complementary language-biased features (Fei et al., 2020a,b;Zhen et al., 2021). Yet we emphasize that different from regular cross-lingual text classification or sequential prediction, XRE relies particularly on the syntactic structure features, e.g., UD, and thus needs a more fine-grained approach for SRC-TGT data ensembling, instead of simply instance stack-ing. Thus, we propose merging the SRC and TGT syntax trees into the code-mixed forests.\\nCode mixing has been explored in several different NLP applications (Labutov and Lipson, 2014;Joshi et al., 2016;Banerjee et al., 2018;Samanta et al., 2019), where the core idea is creating data piece containing words from different languages simultaneously. For example, Samanta et al. (2019) introduce a novel data augmentation method for enhancing the recognition of code-switched sentiment analysis, where they replace the constituent phrases with code-mixed alternatives. Qin et al. (2020) propose generating code-switching data to augment the existing multilingual language models for better zero-shot cross-lingual tasks. While we notice that most of the works focus on the development of code-mixed sequential texts, this work considers the one for structural syntax trees. Our work is partially similar to Zhang et al. (2019) on the code-mixed UD tree construction. But ours differentiate theirs in that Zhang et al. (2019) target better UD parsing itself, while we aim to improve downstream tasks.\\n3 Observations on UD Bias\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Bias Source Analysis\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"As mentioned, even though UD trees define consistent annotations across languages, it still falls short on wiping all syntactic bias. This is inevitably caused by the underlying linguistic disparity deeply embedded in the language itself. Observing the linguistic discrepancies between different languages, we can summarize them into following three levels:\\n1) Word-level Changes.\\n• Word number. The words referring to same semantics in different languages vary, e.g., in English one single-token word may be translated in Chinese with more than one token. • Part of speech. In different languages a parallel lexicon may come with different part of speech. • Word order. Also it is a common case that the word order varies among parallel sentences in different languages.\\n2) Phrase-level Change.\\n• Modification type. A modifier of a phrasal constituent can be changed when translating into another languages. For example, in English, 'in the distance' is often an adverbial modifier, while its counterpart in Chinese '遥 远的' plays a role of an attribute modifier. • Change of pronouns. English grammar has strict structure, while in some other languages the grammar structures may not strict. For example, in English, it is often case to use relative pronouns (e.g., which, that, who) to refer to the prior mentions, while in other languages, such as Chinese, the personal pronouns (e.g., which, that, who) will be used to refer the prior mentions. • Constituency order change. Some constituent phrases will be reorganized and reordered from one language to another language, due to the differences in grammar rules.\\n3) Sentence-level Change.\\n• Transformation between active and passive sentences. In English it could be frequent to use the passive forms of sentences, while being translated into other languages the forms will be transformed into active types, where the words and phrases in the whole sentences can be reversed. • Transformation between clause and main sentence. In English the attributive clauses and noun clauses are often used as subordinate components, while they can be translated into two parallel clauses in other languages. • Change of reading order of sentences. The majority of the languages in this world have the reading order of from-left-to-right, such as English, French, etc. But some languages, e.g., under Afro-Asiatic family, Arabic, Hebrew, Persian, Sindhi and Urdu languages read from right to left.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"UD Bias Statistics\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In Fig. 3 we present the statistics of such bias between the parallel UD trees in different languages, such as the misaligned words, mismatched UD (w ↷ i w j ) pair and UD path of (e ↷ s • • • ↷ e o ) relational pair. Fig. 3(a) reveals that languages under different families show distinct divergences. And the more different of languages, the greater the divergences (e.g., English to Arabic). Fig. 3(b) indicates that complex sentences (e.g., compound sentences) bring larger bias; and in the real world, complex sentences are much more ubiquitous than simple ones. Also, the mismatch goes worse when the UD core predicates are nouns instead of verbs. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Code-mixed UD Forest Construction\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"To eliminate such discrepancies for unbiased UDfeature transfer, we build the code-mixed UD forests, via the following six steps.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"▶\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Step 1: translating a sentence x Src in SRC language to the one x Tgt in TGT language.1 This step is to generate a pseudo parallel sentence pair in both TGT and SRC languages. We accomplish this by using the state-of-the-art Google Translation API. 2 We denote the parallel sentences as or .\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"▶\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Step 2: obtaining the word alignment scores. Meanwhile, we employ the Awesome-align toolkit3 to obtain the word alignment confidence M ={m i↔j } between word pair w i ∈ x Src and w j ∈ x Tgt in parallel sentences.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"▶\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Step 3: parsing UD trees for parallel sentences. Then, we use the UD parsers in SRC and SRC languages respectively to parse the UD syntax trees for two parallel sentences, respectively. We adopt the UDPipe 4 as our UD parsers, which are trained separately on different UD annotated data 5 . We denote the SRC UD tree as T Src , and the pseudo TGT UD tree as T Tgt . Note that the UD trees in all languages share the same dependency labels, Algorithm 1 Process of constructing code-mixed UD forests Input: T SRC , T TGT , M , threshold θ, empty forest F = Φ. Output: Code-mixed UD forest F.\\n1: def Construct (T SRC , T TGT , M , F)\\n▷ breadth-first top-down traverse.\\n2: is_root = True ▷ a flag for traversing the predicate only once.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"3:\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"F.w cur = ROOT ▷ creating ROOT node for F. if is_root then 7: return aligned_pairs, nonaligned_nodes i.e., with the same (as much as possible) annotation standards. In Appendix §A we list the dependency labels which are the commonly occurred types.\\nw merged =\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"▶\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Step 4: projecting and merging the labels of training data. For the training set, we also need to project the annotations (relational subjectobject pairs) of sentences in SRC languages to TGT pseudo sentences. Note that this step is not needed for the testing set. The projection is based on the open source6 , during which the word alignment scores at step-2 are used. We can denote the SRC annotation as y, and the pseudo TGT label as y.\\nWe then merge the annotation from both SRC and TGT viewpoints, into the code-mixed one Y , for later training use. Specifically, for the node that is kept in the final code-mixed forest, we will keep its labels; and for those nodes that are filtered, the annotations are replaced by their correspondences.\\n▶ Step 5: merging the SRC and TGT UD trees into a code-mixed forest. Finally, based on the SRC UD tree and the TGT UD tree, we construct the code-mixed UD forest. We mainly perform breadth-first top-down traversal over each pair of nodes T Src and T Tgt , layer by layer. The traversal starts from their ROOT node. We first create a ROOT node as the initiation of the codemixed forest. We design two types of actions for the forest merging process:\\n• Merging current pair of nodes w i ∈ T Src from SRC tree and w j ∈ T Tgt from TGT tree into the forest F, if the current two nodes are confidently aligned at same dependency layer. We check the word alignment confidence m i↔j between the two nodes, and if the confidence is above a pre-defined threshold θ, i.e., m i↔j > θ, we treat them as confidently aligned.\\n• Copying current node from SRC tree T Src or TGT tree T Tgt into the forest F, once the node has no significant alignment in the opposite tree at this layer. In Algorithm 1 we formulate in detail the process of code-mixed forest construction. Also, we note that when moving the nodes from two separate UD trees into the forest, the attached dependency labels are also copied. When two nodes are merged, we only choose the label of the TGT-side node. Finally, the resulting forest F looks like code-mixing, and is structurally compact.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"▶\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Step 6: assembling code-mixed texts. Also we need to synthesize a code-mixed text X based on the raw SRC text x Src and the pseudo TGT text x Tgt . The code-mixed text X will also be used as inputs together with the forest, into the forest encoder. We directly replace the SRC words with the TGT words that have been determined significantly aligned at Step-5.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"XRE with Code-mixed UD Forest\",\n \"publication_ref\": [\n \"b32\",\n \"b8\"\n ],\n \"table_ref\": [],\n \"text\": \"Along with the UD forest F Src , we also assemble the code-mixed sequential text X Src from the SRC and translated pseudo-TGT sentences (i.e., x Src and x Tgt ), and the same for the TGT sentences X Tgt . An XRE system, being trained with SRC-side annotated data (, Y Src ), needs to determine the label Y Tgt of relational pair e r ↷ s e o given a TGT sentence and UD forest ().\\nThe XRE system takes as input X={w i } n and F. We use the multilingual language model (MLM) for representing the input code-mixed sentence X:\\nH = {h1, • • • , hn} = MLM(X) , (1\\n)\\nwhere X is the code-mixed sentential text. We then formulate the code-mixed forest F as a graph, G=, where E={e i,j } n×n is the edge between word pair (i.e., initiated with e i,j =0/1, meaning dis-/connecting), V ={w i } n are the words. We main the node embeddings r i for each node v i . We adopt the GAT model (Velickovic et al., 2018) for the backbone forest encoding:\\nρi,j = Softmax(GeLU(U T [W1ri; W2rj])) ,(2)\\nui = σ( j ρi,jW3r 1 j ) ,(3)\\nwhere W 3/4/5 and U are all trainable parameters. σ is the sigmoid function. GeLU is a Gaussian error linear activation function. Note that the firstlayer representations of r i is initialized with h i .\\nH and U are then concatenated as the resulting feature representation:\\nĤ = H ⊕ U .(4)\\nXRE aims to determine the semantic relation labels between two given mention entities. For example, given a sentence 'John Smith works at Google', RE should identify that there is a relationship of \\\"works at\\\" between the entities \\\"John Smith\\\" and \\\"Google\\\". Our XRE model needs to predict the relation label y. We adopt the biaffine decoder (Dozat and Manning, 2017) to make prediction:\\ny = Softmax(h T s • W1 • ho + W2 • Pool( Ĥ)) . (5)\\nHere both h s and h o are given.\\n6 Experiments\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Setups\",\n \"publication_ref\": [\n \"b15\",\n \"b0\",\n \"b0\"\n ],\n \"table_ref\": [\n \"tab_1\"\n ],\n \"text\": \"We consider the ACE05 (hristopher Walker et al., 2006) dataset, which includes English (EN), Chinese (ZH) and Arabic (AR). We give the data statistics in Table 1 The multilingual BERT is used. We use two-layer GAT for forest encoding, with a 768-d hidden size. We mainly consider the transfer from EN to one other language. Following most cross-lingual works (Fei et al., 2020b;Ahmad et al., 2021), we train the XRE model with fixed 300 iterations without early-stopping. We make comparisons between three setups: 1) using only raw SRC training data with the model transfer, 2) using only the pseudo TGT (via annotation projection) for training, and 3) using both the above SRC and TGT data. Each setting uses both the texts and UD tree (or forest) features. The baseline uses the same GAT model for syntax encoding, marked as Syn-Baseline. For setup 1)&2) we also test the transfer with only text inputs, removing the syntax features, marked as TxtBaseline. Besides, for setup 1) we cite current SoTA performances as references. We use F1 to measure the RE performance, following Ahmad et al. (2021). All experiments are undergone five times and the average value is reported.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Data Inspection\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_4\"\n ],\n \"text\": \"We also show in Table 3 the differences in average sequential and syntactic (shortest dependency path) distances between the subjects and objects of the relational triplets. As seen, the syntactic distances between subject-object pairs are clearly shortened in the view of syntactic dependency trees, which indicates the imperative to incorporate the tree structure features. However, the syntactic distances between different languages vary, i.e., more complex languages have longer syntactic distances. Such discrepancy reflects the necessity of employing our proposed UD debiasing methods to bridge the gap.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Main Results\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_3\"\n ],\n \"text\": \"From Table 2, we can see that UD features offer exceptional boosts (M1 vs. M2, M4 vs. M5). And annotation projection methods outperform model transfer ones (i.e., M1&M2&M3 vs. M4&M5) by offering direct TGT-side features. Interestingly, in both two transfer paradigms, the improvements from UD become weak on the language pairs with bigger divergences. For example, the improvement on EN→DE outweighs the ones on EN→ZH. Furthermore, using our proposed code-mixed syntax forests is significantly better than using standalone SRC or TGT (or the simple combination) UD features (M7 vs. M2&M5&M6) on all transfers with big margins. For example, our system outperforms SoTA UD-based systems with averaged +4.8%(=67.2-62.4) F1. This evidently verifies the necessity to create the code-mixed forests, i.e., bringing unbiased UD features for transfer. Also, we find that the more the difference between the two languages, the bigger the improvements from forests. The ablation of code-mixed texts also shows the contribution of the sequential textual features, which indirectly demonstrates the larger efficacy of the structural code-mixed UD forests.\"\n },\n {\n \"figure_ref\": [\n \"fig_3\"\n ],\n \"heading\": \"Probing Unbiasedness of Code-mixed UD Forest\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Fig. 4 plots the change of the syntax distances of RE pairs during the transfer with different syntax trees. We see that the use of SRC UD trees shows clear bias (with larger inclination angles) during the transfer, while the use of TGT UD trees and codemixed forests comes with less change of syntax distances. Also, we can see from the figure that the inference paths between objects and subjects of RE tasks are clearly shortened with the forests (in orange color), compared to the uses of SRC/TGT UD trees.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Change during Code-mixed UD Forest Merge\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_6\"\n ],\n \"text\": \"Here we make statistics of how many words are merged and kept during the UD tree merging, respectively. The statistics are shown in Table 4. We can see that the distance between EN-ZH is shorter than that between EN-AR. For example, the length of code-mixed EN-ZH UD forests (sentences) is 31.63, while for EN-AR the length is 40.44. Also, EN-ZH UD forests have a higher to 21.4% merging rate, while EN-AR UD forests have 16.6% merging rate. This demonstrates that the more divergences of languages, the lower the merging rate of the code-mixed forest.\"\n },\n {\n \"figure_ref\": [\n \"fig_4\"\n ],\n \"heading\": \"Impacts of θ on Controlling the Quality of Merged Forest\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In §4 of step-5, we describe that we use a threshold θ to control the aligning during the UD tree merging. Intuitively, the large the threshold θ, the lower the alignment rate. When θ → 0, most of the SRC and TGT nodes in two parallel UD trees can find their counterparts but the alignments are most likely to be wrong, thus hurting the quality of the resulting code-mixed UD forests. When θ → 1, none of the SRC and TGT nodes in two parallel UD trees can be aligned, and both two UD trees are copied and co-existed in the resulting code-mixed UD forests. In such case, the integration of such forests is equivalent to the annotation projection methods where we directly use both the raw SRC UD feature and the translated pseudo TGT UD tree feature. In Fig. 5 we now study the influences of using different code-mixed forest features generated with different merging rates (θ). We see that with a threshold of θ=0.5, the performances are consistently the best.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Performances on Different Types of Sentence\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_7\"\n ],\n \"text\": \"In Table 5 we show the results under different types of sentences. We directly select 500 short sentences (with length < 12) as simple sentences; and select 500 lengthy sentences (with length > 35) as complex sentences. As can be seen, with the code-mixed forest features, the system shows very notable improvements in complex sentences. For example, on the EN→ZH we obtain 15.9(=57.2-41.3)% F1 improvement, and on the EN→AR the boost increases strikingly to 25.2(=67.3-42.1)% F1. However, such enhancements are not very significant in handling simple sentences. This indicates that the code-mixed UD forest features can espe- cially enhance the effectiveness on the hard case, i.e., the transfer between those pairs with greater divergences will receive stronger enhancements from our methods.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion and Future Work\",\n \"publication_ref\": [\n \"b14\",\n \"b13\"\n ],\n \"table_ref\": [],\n \"text\": \"Universal dependencies (UD) have been served as effective language-consistent syntactic features for cross-lingual relation extraction (XRE). In this work, we reveal the intrinsic language discrepancies with respect to the UD structural annotations, which limit the utility of the UD features. We enhance the efficacy of UD features for an unbiased UD-based transfer, by constructing code-mixed UD forests from both the source and target UD trees. Experimental results demonstrate that the UD forests effectively debias the syntactic disparity in the UD-based XRE transfer, especially for those language pairs with larger gaps. Leveraging the syntactic dependency features is a long-standing practice for strengthening the performance of RE tasks. In this work, we propose a novel type of syntactic feature, code-mixed UD forests, for cross-lingual relation extraction. We note that this feature can be applied broadly to other cross-lingual structured information extraction tasks that share the same task definition besides RE, such as event detection (ED) (Halpin and Moore, 2006) and semantic role labeling (SRL) (Gildea and Jurafsky, 2000). Besides, how to fur-ther increase the utility of the UD forests with a better modeling method is a promising research direction, i.e., filtering the noisy structures in the UD forests.\\n2.5-4.0 9 as well as GNU GPL 3.0 10 . Our use of UD treebanks comply with all these license terms is at non-commercial purpose. The software tools (i.e., UDPipe parsers) are provided under GNU GPL V2. Our use of UDPipe tools complies with the term.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Acknowledgments\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This research is supported by the National Natural Science Foundation of China (No. 62176180), and also the Sea-NExT Joint Lab.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Limitations\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Although showing great prominence, our proposed method has the following limitations. First of all, our method relies on the availability of annotated UD trees of both the source and target languages, as we need to use the annotations to parse the syntax trees for our own sentences. Fortunately, UD project covers over 100 languages, where most of the languages, even the minor ones, will have the UD resources. At the same time, our method will be influenced by the quality of UD parsers. Secondly, our method also uses the external translation systems to produce the pseudo parallel sentences, where our method may largely subject to the quality of the translators. Again luckily, current neural machine translation systems have been well developed and established, i.e., Google Translation. Only when handling very scare languages where the current translation systems fail to give satisfactory performances, our method will fail.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Ethics Statement\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this work, we construct a type of code-mixed UD forest based on the existing UD resources. We note that all the data construction has been accomplished automatically, and we have not created any new annotations with additional human labor. Specifically, we use the UD v2.10 resource, which is a collection of linguistic data and tools that are open-sourced. Each of treebanks of UD has its own license terms, including the CC BY-SA 4.0 8 and CC BY-NC-SA\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A The universal dependency labels\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In Table 6, we list the dependency labels which are the commonly occurred types. Please refer to Stanford dependency 11 \"\n }\n]"},"abstract":{"kind":"string","value":"Latest efforts on cross-lingual relation extraction (XRE) aggressively leverage the languageconsistent structural features from the universal dependency (UD) resource, while they may largely suffer from biased transfer (e.g., either target-biased or source-biased) due to the inevitable linguistic disparity between languages. In this work, we investigate an unbiased UDbased XRE transfer by constructing a type of code-mixed UD forest. We first translate the sentence of the source language to the parallel target-side language, for both of which we parse the UD tree respectively. Then, we merge the source-/target-side UD structures as a unified code-mixed UD forest. With such forest features, the gaps of UD-based XRE between the training and predicting phases can be effectively closed. We conduct experiments on the ACE XRE benchmark datasets, where the results demonstrate that the proposed code-mixed UD forests help unbiased UD-based XRE transfer, with which we achieve significant XRE performance gains."},"title":{"kind":"string","value":"Constructing Code-mixed Universal Dependency Forest for Unbiased Cross-lingual Relation Extraction"},"figures":{"kind":"list like","value":[{"figure_caption":"Merging current pair of nodes from SRC and TGT trees as one into forest, if the two nodes are aligned at same dependency layer.","figure_data":"","figure_id":"fig_0","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 3: Statistics of mismatching items of UD trees.","figure_data":"","figure_id":"fig_1","figure_label":"","figure_type":"figure"},{"figure_caption":"SRC ̸ = Φ) or (T TGT ̸ = Φ) or (opt_nodes̸ = Φ) do 6:","figure_data":"","figure_id":"fig_2","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 4 :4Figure 4: Change of syntax distance (shortest path) of relational pair in different UD trees.","figure_data":"","figure_id":"fig_3","figure_label":"4","figure_type":"figure"},{"figure_caption":"Figure 5 :5Figure 5: Transfer performances by using code-mixed forests generated with different merging rates (θ).","figure_data":"","figure_id":"fig_4","figure_label":"5","figure_type":"figure"},{"figure_caption":"Merge(T SRC .ROOT, T TGT .ROOT) ▷ merging from ROOT in T SRC and T TGT . merged .next SRC = T SRC .ROOT.GetChildNodes() 9:w merged .next TGT = T TGT .ROOT.GetChildNodes()","figure_data":"10:F.w cur .SetChild(w merged , 'root')11:opt_nodes.enqueue(w merged )12:is_root = False13:else14:F.w cur = opt_nodes.dequeue()15:17:w merged = Merge(w SRC i , w TGT j )18:21:opt_nodes.enqueue(w merged )22:end for23:for w i ∈ nonaligned_nodes do24:25:end for26:end if27:end while28:return F32:aligned_pairs = []33:for m i↔j ∈ M do34:if m i↔j > θ then35:aligned_pairs.Append(nodes_a[i], nodes_b[j], nodes_b[i].arc )36:nodes_a.Remove(w i )37:nodes_a.Remove(w j )38:end if39:end for40:nonaligned_nodes = nodes_a.union(nodes_b)▷ words with no salient alignments.","figure_id":"tab_0","figure_label":"","figure_type":"table"},{"figure_caption":"7 Data statistics. The numbers are documents.","figure_data":"LanguageTrainDevTestEN4796060ZH5076363AR3234040","figure_id":"tab_1","figure_label":"1","figure_type":"table"},{"figure_caption":"Main results of cross-lingual RE transfer tasks from English language to other languages, by different models and features. M6 uses two separate instances (texts and UD trees) for training, including the raw SRC one and the pseudo TGT one. M7 uses the SRC-TGT merged one as ours, i.e., code-mixed texts and forests.","figure_data":"ENZHAR•Sequential Distance4.83.925.8•Syntactic Distance2.22.65.1","figure_id":"tab_3","figure_label":"2","figure_type":"table"},{"figure_caption":"Sequential and syntactic (shortest dependency path) distances (words) between the subjects and objects of the relational triplets.","figure_data":"","figure_id":"tab_4","figure_label":"3","figure_type":"table"},{"figure_caption":"The statistics of the words before and after constructing code-mixed data.","figure_data":"Words per SentenceBefore MergingAfter MergingSRC (EN)TGTSumCode-mixedMerged (Rate)EN-ZH15.3224.9140.2331.638.6 (21.4%)EN-AR15.3233.1248.4440.448.0 (16.6%)EN→ZHEN→AR• Simple SentenceSynBaseline(+T SRC )66.178.2SynBaseline(+T T GT )68.780.6SynBaseline(+F)71.382.4• Complex SentenceSynBaseline(+T SRC )39.537.4SynBaseline(+T T GT )41.342.1SynBaseline(+F)57.267.3","figure_id":"tab_6","figure_label":"4","figure_type":"table"},{"figure_caption":"Comparisons under different types of sentences.","figure_data":"","figure_id":"tab_7","figure_label":"5","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Merging current pair of nodes from SRC and TGT trees as one into forest, if the two nodes are aligned at same dependency layer.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3: Statistics of mismatching items of UD trees.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"SRC ̸ = Φ) or (T TGT ̸ = Φ) or (opt_nodes̸ = Φ) do 6:\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 4 :4Figure 4: Change of syntax distance (shortest path) of relational pair in different UD trees.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 5 :5Figure 5: Transfer performances by using code-mixed forests generated with different merging rates (θ).\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"5\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Merge(T SRC .ROOT, T TGT .ROOT) ▷ merging from ROOT in T SRC and T TGT . merged .next SRC = T SRC .ROOT.GetChildNodes() 9:w merged .next TGT = T TGT .ROOT.GetChildNodes()\",\n \"figure_data\": \"10:F.w cur .SetChild(w merged , 'root')11:opt_nodes.enqueue(w merged )12:is_root = False13:else14:F.w cur = opt_nodes.dequeue()15:17:w merged = Merge(w SRC i , w TGT j )18:21:opt_nodes.enqueue(w merged )22:end for23:for w i ∈ nonaligned_nodes do24:25:end for26:end if27:end while28:return F32:aligned_pairs = []33:for m i↔j ∈ M do34:if m i↔j > θ then35:aligned_pairs.Append(nodes_a[i], nodes_b[j], nodes_b[i].arc )36:nodes_a.Remove(w i )37:nodes_a.Remove(w j )38:end if39:end for40:nonaligned_nodes = nodes_a.union(nodes_b)▷ words with no salient alignments.\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"7 Data statistics. The numbers are documents.\",\n \"figure_data\": \"LanguageTrainDevTestEN4796060ZH5076363AR3234040\",\n \"figure_id\": \"tab_1\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Main results of cross-lingual RE transfer tasks from English language to other languages, by different models and features. M6 uses two separate instances (texts and UD trees) for training, including the raw SRC one and the pseudo TGT one. M7 uses the SRC-TGT merged one as ours, i.e., code-mixed texts and forests.\",\n \"figure_data\": \"ENZHAR•Sequential Distance4.83.925.8•Syntactic Distance2.22.65.1\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Sequential and syntactic (shortest dependency path) distances (words) between the subjects and objects of the relational triplets.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"The statistics of the words before and after constructing code-mixed data.\",\n \"figure_data\": \"Words per SentenceBefore MergingAfter MergingSRC (EN)TGTSumCode-mixedMerged (Rate)EN-ZH15.3224.9140.2331.638.6 (21.4%)EN-AR15.3233.1248.4440.448.0 (16.6%)EN→ZHEN→AR• Simple SentenceSynBaseline(+T SRC )66.178.2SynBaseline(+T T GT )68.780.6SynBaseline(+F)71.382.4• Complex SentenceSynBaseline(+T SRC )39.537.4SynBaseline(+T T GT )41.342.1SynBaseline(+F)57.267.3\",\n \"figure_id\": \"tab_6\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Comparisons under different types of sentences.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_7\",\n \"figure_label\": \"5\",\n \"figure_type\": \"table\"\n }\n]"},"authors":{"kind":"string","value":"Hao Fei; Meishan Zhang; Min Zhang; Tat-Seng Chua"},"references":{"kind":"list like","value":[{"authors":"Ahmad Wasi Uddin; Nanyun Peng; Kai-Wei Chang","journal":"","ref_id":"b0","title":"GATE: graph attention transformer encoder for cross-lingual relation and event extraction","year":"2021"},{"authors":"Suman Banerjee; Nikita Moghe; Siddhartha Arora; Mitesh M Khapra","journal":"","ref_id":"b1","title":"A dataset for building code-mixed goal oriented conversation systems","year":"2018"},{"authors":"Anders Björkelund; Love Hafdell; Pierre Nugues","journal":"","ref_id":"b2","title":"Multilingual semantic role labeling","year":"2009"},{"authors":"Duy-Cat Can; Hoang-Quynh Le; Quang-Thuy Ha; Nigel Collier","journal":"","ref_id":"b3","title":"A richer-but-smarter shortest dependency path with attentive augmentation for relation extraction","year":"2019"},{"authors":"Silviu Cucerzan; David Yarowsky","journal":"","ref_id":"b4","title":"Language independent named entity recognition combining morphological and contextual evidence","year":"1999"},{"authors":"Aron Culotta; Jeffrey Sorensen","journal":"","ref_id":"b5","title":"Dependency tree kernels for relation extraction","year":"2004"},{"authors":"Angel Daza; Anette Frank","journal":"","ref_id":"b6","title":"Translate and label! an encoder-decoder approach for cross-lingual semantic role labeling","year":"2019"},{"authors":"Marie-Catherine De Marneffe; Christopher D Manning; Joakim Nivre; Daniel Zeman","journal":"Comput. Linguistics","ref_id":"b7","title":"Universal dependencies","year":"2021"},{"authors":"Timothy Dozat; Christopher D Manning","journal":"","ref_id":"b8","title":"Deep biaffine attention for neural dependency parsing","year":"2017"},{"authors":"Fei Hao Fei; Bobo Li; Donghong Li; Ji","journal":"","ref_id":"b9","title":"Encoder-decoder based unified semantic role labeling with label-aware syntax","year":"2021"},{"authors":"Shengqiong Hao Fei; Jingye Wu; Bobo Li; Fei Li; Libo Li; Meishan Qin; Min Zhang; Tat-Seng Zhang; Chua","journal":"","ref_id":"b10","title":"Lasuie: Unifying information extraction with latent adaptive structure-aware generative language model","year":"2022"},{"authors":"Meishan Hao Fei; Donghong Zhang; Ji","journal":"","ref_id":"b11","title":"a. Cross-lingual semantic role labeling with highquality translated training corpus","year":"2020"},{"authors":"Meishan Hao Fei; Fei Zhang; Donghong Li; Ji","journal":"IEEE ACM Trans. Audio Speech Lang. Process","ref_id":"b12","title":"Cross-lingual semantic role labeling with model transfer","year":"2020"},{"authors":"Daniel Gildea; Daniel Jurafsky","journal":"","ref_id":"b13","title":"Automatic labeling of semantic roles","year":"2000"},{"authors":"Harry Halpin; Johanna D Moore","journal":"","ref_id":"b14","title":"Event extraction in a plot advice agent","year":"2006"},{"authors":"Stephanie Walker; Julie Strassel; Kazuaki Medero; Maeda","journal":"","ref_id":"b15","title":"Ace 2005 multilingual training corpus","year":"2006"},{"authors":"Aditya Joshi; Ameya Prabhu; Manish Shrivastava; Vasudeva Varma","journal":"","ref_id":"b16","title":"Towards sub-word level compositions for sentiment analysis of Hindi-English code mixed text","year":"2016"},{"authors":"Seokhwan Kim; Minwoo Jeong; Jonghoon Lee; Gary Geunbae; Lee ","journal":"","ref_id":"b17","title":"A cross-lingual annotation projection approach for relation detection","year":"2010"},{"authors":"Mikhail Kozhevnikov; Ivan Titov","journal":"","ref_id":"b18","title":"Crosslingual transfer of semantic role labeling models","year":"2013"},{"authors":"Igor Labutov; Hod Lipson","journal":"","ref_id":"b19","title":"Generating codeswitched text for lexical learning","year":"2014"},{"authors":"Chenwei Lou; Jun Gao; Changlong Yu; Wei Wang; Huan Zhao; Weiwei Tu; Ruifeng Xu","journal":"","ref_id":"b20","title":"Translation-based implicit annotation projection for zero-shot cross-lingual event argument extraction","year":"2022"},{"authors":"Di Lu; Ananya Subburathinam; Heng Ji; Jonathan May; Shih-Fu Chang; Avi Sil; Clare Voss","journal":"","ref_id":"b21","title":"Crosslingual structure transfer for zero-resource event extraction","year":"2020"},{"authors":"Ryan Mcdonald; Joakim Nivre; Yvonne Quirmbach-Brundage; Yoav Goldberg; Dipanjan Das; Kuzman Ganchev; Keith Hall; Slav Petrov; Hao Zhang; Oscar Täckström; Claudia Bedini; Núria Bertomeu Castelló; Jungmee Lee","journal":"","ref_id":"b22","title":"Universal dependency annotation for multilingual parsing","year":"2013"},{"authors":"Makoto Miwa; Mohit Bansal","journal":"","ref_id":"b23","title":"End-to-end relation extraction using LSTMs on sequences and tree structures","year":"2016"},{"authors":"Phoebe Mulcaire; Swabha Swayamdipta; Noah A Smith","journal":"","ref_id":"b24","title":"Polyglot semantic role labeling","year":"2018"},{"authors":"Jian Ni; Radu Florian","journal":"","ref_id":"b25","title":"Neural cross-lingual relation extraction based on bilingual word embedding mapping","year":"2019"},{"authors":"Sebastian Padó; Mirella Lapata","journal":"J. Artif. Intell. Res","ref_id":"b26","title":"Cross-lingual annotation projection for semantic roles","year":"2009"},{"authors":"Libo Qin; Minheng Ni; Yue Zhang; Wanxiang Che","journal":"","ref_id":"b27","title":"Cosda-ml: Multi-lingual code-switching data augmentation for zero-shot cross-lingual NLP","year":"2020"},{"authors":"Bidisha Samanta; Niloy Ganguly; Soumen Chakrabarti","journal":"","ref_id":"b28","title":"Improved sentiment detection via label transfer from monolingual to synthetic codeswitched text","year":"2019"},{"authors":"Ananya Subburathinam; Di Lu; Heng Ji; Jonathan May; Shih-Fu Chang; Avirup Sil; Clare Voss","journal":"","ref_id":"b29","title":"Cross-lingual structure transfer for relation and event extraction","year":"2019"},{"authors":"Nasrin Taghizadeh; Heshaam Faili","journal":"ACM Trans. Asian Low Resour. Lang. Inf. Process","ref_id":"b30","title":"Crosslingual adaptation using universal dependencies","year":"2021"},{"authors":"Nasrin Taghizadeh; Heshaam Faili","journal":"Comput. Speech Lang","ref_id":"b31","title":"Crosslingual transfer learning for relation extraction using universal dependencies","year":"2022"},{"authors":"Petar Velickovic; Guillem Cucurull; Arantxa Casanova; Adriana Romero; Pietro Liò; Yoshua Bengio","journal":"","ref_id":"b32","title":"Graph attention networks","year":"2018"},{"authors":"Shengqiong Wu; Hao Fei; Yafeng Ren; Donghong Ji; Jingye Li","journal":"","ref_id":"b33","title":"Learn from syntax: Improving pair-wise aspect and opinion terms extraction with rich syntactic knowledge","year":"2021"},{"authors":"Min Xiao; Yuhong Guo","journal":"","ref_id":"b34","title":"Annotation projection-based representation learning for crosslingual dependency parsing","year":"2015"},{"authors":"Dmitry Zelenko; Chinatsu Aone; Anthony Richardella","journal":"","ref_id":"b35","title":"Kernel methods for relation extraction","year":"2002"},{"authors":"Meishan Zhang; Yue Zhang; Guohong Fu","journal":"","ref_id":"b36","title":"Cross-lingual dependency parsing using code-mixed TreeBank","year":"2019"},{"authors":"Yuhao Zhang; Peng Qi; Christopher D Manning","journal":"","ref_id":"b37","title":"Graph convolution over pruned dependency trees improves relation extraction","year":"2018"},{"authors":"Zhisong Zhang; Emma Strubell; Eduard Hovy","journal":"","ref_id":"b38","title":"On the benefit of syntactic supervision for crosslingual transfer in semantic role labeling","year":"2021"},{"authors":"Ranran Zhen; Rui Wang; Guohong Fu; Chengguo Lv; Meishan Zhang","journal":"","ref_id":"b39","title":"Chinese opinion role labeling with corpus translation: A pivot study","year":"2021"}],"string":"[\n {\n \"authors\": \"Ahmad Wasi Uddin; Nanyun Peng; Kai-Wei Chang\",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"GATE: graph attention transformer encoder for cross-lingual relation and event extraction\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Suman Banerjee; Nikita Moghe; Siddhartha Arora; Mitesh M Khapra\",\n \"journal\": \"\",\n \"ref_id\": \"b1\",\n \"title\": \"A dataset for building code-mixed goal oriented conversation systems\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Anders Björkelund; Love Hafdell; Pierre Nugues\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"Multilingual semantic role labeling\",\n \"year\": \"2009\"\n },\n {\n \"authors\": \"Duy-Cat Can; Hoang-Quynh Le; Quang-Thuy Ha; Nigel Collier\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"A richer-but-smarter shortest dependency path with attentive augmentation for relation extraction\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Silviu Cucerzan; David Yarowsky\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"Language independent named entity recognition combining morphological and contextual evidence\",\n \"year\": \"1999\"\n },\n {\n \"authors\": \"Aron Culotta; Jeffrey Sorensen\",\n \"journal\": \"\",\n \"ref_id\": \"b5\",\n \"title\": \"Dependency tree kernels for relation extraction\",\n \"year\": \"2004\"\n },\n {\n \"authors\": \"Angel Daza; Anette Frank\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"Translate and label! an encoder-decoder approach for cross-lingual semantic role labeling\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Marie-Catherine De Marneffe; Christopher D Manning; Joakim Nivre; Daniel Zeman\",\n \"journal\": \"Comput. Linguistics\",\n \"ref_id\": \"b7\",\n \"title\": \"Universal dependencies\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Timothy Dozat; Christopher D Manning\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"Deep biaffine attention for neural dependency parsing\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Fei Hao Fei; Bobo Li; Donghong Li; Ji\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"Encoder-decoder based unified semantic role labeling with label-aware syntax\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Shengqiong Hao Fei; Jingye Wu; Bobo Li; Fei Li; Libo Li; Meishan Qin; Min Zhang; Tat-Seng Zhang; Chua\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"Lasuie: Unifying information extraction with latent adaptive structure-aware generative language model\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Meishan Hao Fei; Donghong Zhang; Ji\",\n \"journal\": \"\",\n \"ref_id\": \"b11\",\n \"title\": \"a. Cross-lingual semantic role labeling with highquality translated training corpus\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Meishan Hao Fei; Fei Zhang; Donghong Li; Ji\",\n \"journal\": \"IEEE ACM Trans. Audio Speech Lang. Process\",\n \"ref_id\": \"b12\",\n \"title\": \"Cross-lingual semantic role labeling with model transfer\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Daniel Gildea; Daniel Jurafsky\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"Automatic labeling of semantic roles\",\n \"year\": \"2000\"\n },\n {\n \"authors\": \"Harry Halpin; Johanna D Moore\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"Event extraction in a plot advice agent\",\n \"year\": \"2006\"\n },\n {\n \"authors\": \"Stephanie Walker; Julie Strassel; Kazuaki Medero; Maeda\",\n \"journal\": \"\",\n \"ref_id\": \"b15\",\n \"title\": \"Ace 2005 multilingual training corpus\",\n \"year\": \"2006\"\n },\n {\n \"authors\": \"Aditya Joshi; Ameya Prabhu; Manish Shrivastava; Vasudeva Varma\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"Towards sub-word level compositions for sentiment analysis of Hindi-English code mixed text\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Seokhwan Kim; Minwoo Jeong; Jonghoon Lee; Gary Geunbae; Lee \",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"A cross-lingual annotation projection approach for relation detection\",\n \"year\": \"2010\"\n },\n {\n \"authors\": \"Mikhail Kozhevnikov; Ivan Titov\",\n \"journal\": \"\",\n \"ref_id\": \"b18\",\n \"title\": \"Crosslingual transfer of semantic role labeling models\",\n \"year\": \"2013\"\n },\n {\n \"authors\": \"Igor Labutov; Hod Lipson\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"Generating codeswitched text for lexical learning\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"Chenwei Lou; Jun Gao; Changlong Yu; Wei Wang; Huan Zhao; Weiwei Tu; Ruifeng Xu\",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"Translation-based implicit annotation projection for zero-shot cross-lingual event argument extraction\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Di Lu; Ananya Subburathinam; Heng Ji; Jonathan May; Shih-Fu Chang; Avi Sil; Clare Voss\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"Crosslingual structure transfer for zero-resource event extraction\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Ryan Mcdonald; Joakim Nivre; Yvonne Quirmbach-Brundage; Yoav Goldberg; Dipanjan Das; Kuzman Ganchev; Keith Hall; Slav Petrov; Hao Zhang; Oscar Täckström; Claudia Bedini; Núria Bertomeu Castelló; Jungmee Lee\",\n \"journal\": \"\",\n \"ref_id\": \"b22\",\n \"title\": \"Universal dependency annotation for multilingual parsing\",\n \"year\": \"2013\"\n },\n {\n \"authors\": \"Makoto Miwa; Mohit Bansal\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"End-to-end relation extraction using LSTMs on sequences and tree structures\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Phoebe Mulcaire; Swabha Swayamdipta; Noah A Smith\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Polyglot semantic role labeling\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Jian Ni; Radu Florian\",\n \"journal\": \"\",\n \"ref_id\": \"b25\",\n \"title\": \"Neural cross-lingual relation extraction based on bilingual word embedding mapping\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Sebastian Padó; Mirella Lapata\",\n \"journal\": \"J. Artif. Intell. Res\",\n \"ref_id\": \"b26\",\n \"title\": \"Cross-lingual annotation projection for semantic roles\",\n \"year\": \"2009\"\n },\n {\n \"authors\": \"Libo Qin; Minheng Ni; Yue Zhang; Wanxiang Che\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"Cosda-ml: Multi-lingual code-switching data augmentation for zero-shot cross-lingual NLP\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Bidisha Samanta; Niloy Ganguly; Soumen Chakrabarti\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"Improved sentiment detection via label transfer from monolingual to synthetic codeswitched text\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Ananya Subburathinam; Di Lu; Heng Ji; Jonathan May; Shih-Fu Chang; Avirup Sil; Clare Voss\",\n \"journal\": \"\",\n \"ref_id\": \"b29\",\n \"title\": \"Cross-lingual structure transfer for relation and event extraction\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Nasrin Taghizadeh; Heshaam Faili\",\n \"journal\": \"ACM Trans. Asian Low Resour. Lang. Inf. Process\",\n \"ref_id\": \"b30\",\n \"title\": \"Crosslingual adaptation using universal dependencies\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Nasrin Taghizadeh; Heshaam Faili\",\n \"journal\": \"Comput. Speech Lang\",\n \"ref_id\": \"b31\",\n \"title\": \"Crosslingual transfer learning for relation extraction using universal dependencies\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Petar Velickovic; Guillem Cucurull; Arantxa Casanova; Adriana Romero; Pietro Liò; Yoshua Bengio\",\n \"journal\": \"\",\n \"ref_id\": \"b32\",\n \"title\": \"Graph attention networks\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Shengqiong Wu; Hao Fei; Yafeng Ren; Donghong Ji; Jingye Li\",\n \"journal\": \"\",\n \"ref_id\": \"b33\",\n \"title\": \"Learn from syntax: Improving pair-wise aspect and opinion terms extraction with rich syntactic knowledge\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Min Xiao; Yuhong Guo\",\n \"journal\": \"\",\n \"ref_id\": \"b34\",\n \"title\": \"Annotation projection-based representation learning for crosslingual dependency parsing\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Dmitry Zelenko; Chinatsu Aone; Anthony Richardella\",\n \"journal\": \"\",\n \"ref_id\": \"b35\",\n \"title\": \"Kernel methods for relation extraction\",\n \"year\": \"2002\"\n },\n {\n \"authors\": \"Meishan Zhang; Yue Zhang; Guohong Fu\",\n \"journal\": \"\",\n \"ref_id\": \"b36\",\n \"title\": \"Cross-lingual dependency parsing using code-mixed TreeBank\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Yuhao Zhang; Peng Qi; Christopher D Manning\",\n \"journal\": \"\",\n \"ref_id\": \"b37\",\n \"title\": \"Graph convolution over pruned dependency trees improves relation extraction\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Zhisong Zhang; Emma Strubell; Eduard Hovy\",\n \"journal\": \"\",\n \"ref_id\": \"b38\",\n \"title\": \"On the benefit of syntactic supervision for crosslingual transfer in semantic role labeling\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Ranran Zhen; Rui Wang; Guohong Fu; Chengguo Lv; Meishan Zhang\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"Chinese opinion role labeling with corpus translation: A pivot study\",\n \"year\": \"2021\"\n }\n]"},"formulas":{"kind":"list like","value":[{"formula_coordinates":[5,76.98,124.77,168.47,11.33],"formula_id":"formula_0","formula_text":"1: def Construct (T SRC , T TGT , M , F)"},{"formula_coordinates":[5,122.14,207.59,45.26,10.77],"formula_id":"formula_1","formula_text":"w merged ="},{"formula_coordinates":[6,351.22,248.22,170.31,8.09],"formula_id":"formula_2","formula_text":"H = {h1, • • • , hn} = MLM(X) , (1"},{"formula_coordinates":[6,521.52,248.54,3.48,7.77],"formula_id":"formula_3","formula_text":")"},{"formula_coordinates":[6,332.35,374.29,192.66,10.33],"formula_id":"formula_4","formula_text":"ρi,j = Softmax(GeLU(U T [W1ri; W2rj])) ,(2)"},{"formula_coordinates":[6,369.96,389.93,155.05,19.94],"formula_id":"formula_5","formula_text":"ui = σ( j ρi,jW3r 1 j ) ,(3)"},{"formula_coordinates":[6,390.33,497.1,134.68,10.25],"formula_id":"formula_6","formula_text":"Ĥ = H ⊕ U .(4)"},{"formula_coordinates":[6,328.21,633.92,196.8,11.13],"formula_id":"formula_7","formula_text":"y = Softmax(h T s • W1 • ho + W2 • Pool( Ĥ)) . (5)"}],"string":"[\n {\n \"formula_coordinates\": [\n 5,\n 76.98,\n 124.77,\n 168.47,\n 11.33\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"1: def Construct (T SRC , T TGT , M , F)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 122.14,\n 207.59,\n 45.26,\n 10.77\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"w merged =\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 351.22,\n 248.22,\n 170.31,\n 8.09\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"H = {h1, • • • , hn} = MLM(X) , (1\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 521.52,\n 248.54,\n 3.48,\n 7.77\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \")\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 332.35,\n 374.29,\n 192.66,\n 10.33\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"ρi,j = Softmax(GeLU(U T [W1ri; W2rj])) ,(2)\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 369.96,\n 389.93,\n 155.05,\n 19.94\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"ui = σ( j ρi,jW3r 1 j ) ,(3)\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 390.33,\n 497.1,\n 134.68,\n 10.25\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"Ĥ = H ⊕ U .(4)\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 328.21,\n 633.92,\n 196.8,\n 11.13\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"y = Softmax(h T s • W1 • ho + W2 • Pool( Ĥ)) . (5)\"\n }\n]"}}},{"rowIdx":388,"cells":{"doi":{"kind":"string","value":"10.48550/ARXIV.2002.05709"},"pub_date":{"kind":"string","value":"2023-05-20"},"sections":{"kind":"list like","value":[{"figure_ref":["fig_0"],"heading":"Introduction","publication_ref":["b12","b6","b12","b18","b17","b16","b18","b17","b16","b0","b23","b18","b9","b37","b20","b39","b11","b13","b8","b2","b16","b0","b17"],"table_ref":[],"text":"Continual learning aims to continually train models with new tasks without forgetting previously learned tasks (Ke and Liu, 2022;De Lange et al., 2022). It has become a promising direction for NLP models to incrementally learn new tasks/domains/classes as humans do (Ke and Liu, 2022). A typical scenario aims to enable NLP models to solve various tasks in an incremental manner, namely the task-incremental continual learning scenario, which is our study setting in this paper. A salient challenge for continual learning is that continually learned models usually suffer from cartographic forgetting (CF), i.e. the performance on previously learned tasks decreases after training on the new one (Lopez-Paz and Ranzato, 2017).\nVarious training strategies have been proposed to mitigate CF (Li and Hoiem, 2017;Kirkpatrick et al., 2017;Lopez-Paz and Ranzato, 2017). Under the fixed model structure, regularization-based methods design regularization terms to control the shift of representations learned from previous tasks (Li and Hoiem, 2017;Kirkpatrick et al., 2017;Aljundi et al., 2018). Rehearsal-based methods save the data samples from previous tasks into a memory buffer and re-train the model to recover knowledge during training on the current task (Riemer et al., 2019;Lopez-Paz and Ranzato, 2017;de Masson D'Autume et al., 2019). However, most continual learning methods are designed to recover the learned knowledge or mitigate the representation of forgetting.\nSeldom considers adapting the classification criterion for the newly learned representations. For example, in supervised contrastive learning, the contrastive objective is designed to pull the data representations with the same labels together and push representations with different labels away (Chen et al., 2020a;Gao et al., 2021;Zhang et al., 2021;Neelakantan et al., 2022;Zhao et al., 2022). Representations of the training samples can be saved as a classification criterion, after which an instancebased method such as a k Nearest Neighbor (kNN) module can be leveraged for inference (Kassner and Schütze, 2020;Khandelwal et al., 2020). After learning the new task, we can feed the saved sample into models for new classification criteria and mitigate the problem of CF. For example in Figure 1, although the representations have decayed for learning the new task, the saved samples adapt to serve as the classification criterion in kNN modules.\nInspired by the above motivation, we investigate the use of supervised contrastive learning for taskincremental continual learning (SCCL). After supervised contrastive learning on each task, we use a K-means module to select several samples and save them into a memory buffer while maintaining the learned representation distribution. In addition, to mitigate the representation drift when training the model for new tasks, we use an instance-wise relation distillation (IRD) term (Fang et al., 2020;Cha et al., 2021) and a memory replay module (de Masson D'Autume et al., 2019) to maintain the learned knowledge. During inference, the saved samples are fed into the trained model to obtain updated representations and a kNN module is used for classification.\nExperimental results show that our proposed model can achieve state-of-the-art performance compared with standard cross-entropy-based (CE) baselines. We additionally extend different continual learning strategies (Kirkpatrick et al., 2017;Aljundi et al., 2018;Li and Hoiem, 2017) to the supervised contrastive continual learning framework, which gives stronger results than corresponding CE-based methods, showing the advantage of contrastive learning with a kNN classifier in continual learning scenarios. We further analyze the effectiveness of each module in our paper through ablation studies. To our knowledge, we are the first to propose a supervised contrastive learning framework for task-incremental continual learning, without any augmented parameters. The code will be released when accepted."},{"figure_ref":[],"heading":"Related Work","publication_ref":["b33","b26","b17","b16","b24","b16","b0","b17","b17","b24","b1","b31","b9","b14","b19","b2"],"table_ref":[],"text":"Continual Learning Various continual learning methods have been proposed to mitigate the problem of CF. The methods can be broadly divided into architecture-based methods (Yoon et al., 2018;Serra et al., 2018), regularization-based methods Li and Hoiem (2017); Kirkpatrick et al. (2017), and rehearsal-based methods (Rolnick et al., 2019). Under the fixed model structure, regularizationbased methods (Kirkpatrick et al., 2017;Aljundi et al., 2018;Li and Hoiem, 2017) optimize network parameters on the current task while constraining the representation drift. For example, Li and Hoiem (2017) propose learning without forgetting (LwF) to tackle this problem, which regularizes the model output of current data close to those trained for the previous model. Another category of fixed-structure strategies (rehearsal-based) stores a limited subset of samples from previous tasks to mitigate CF such as ER (Rolnick et al., 2019), RM (Bang et al., 2021), andiCaRL (Rebuffi et al., 2017).\nContrastive Learning Contrastive learning is initially introduced in self-supervised settings and proved to subsume or significantly outperform traditional contrastive losses such as triplet loss (Chen et al., 2020b;Wu et al., 2018;Gao et al., 2021;?). For example, Khosla et al. (2020) first propose the idea of self-supervised contrastive learning and prove that the method is more robust to natural corruptions, stable to hyper-parameter settings, and has strong transfer performance. Luo et al. (2022) uses supervised contrastive learning combined with a kNN inference module for cross-domain sentiment analysis, showing a stronger generalization ability compared with standard CE-based methods. Cha et al. (2021) propose a contrastive continual learning method, Co 2 L, for class-incremental continual learning. The method uses an asymmetric supervised contrastive loss to enlarge the distance between representations of previous and new tasks. However, there are significant differences between our model and Co 2 L. First, the asymmetric contrastive loss of Co 2 L is unsuitable for taskincremental continual learning, because a representation can be predicted as different labels according to task objectives. Second, Co 2 L uses a decoupled classification layer for inference, i.e. it learns rep- resentations first and then learns a linear classifier separately, causing low extensibility and high complexity of the model. In contrast, we use a kNN module as the classification criteria to enhance the extensibility of the model. Third, Co 2 L only considers the representation drift from the view of regularization. But we also mitigate the problem of representation drift by feeding memory data into the current model to obtain updated classification criteria."},{"figure_ref":["fig_1"],"heading":"Method","publication_ref":[],"table_ref":[],"text":"The overall SCCL framework is illustrated in Figure 2, consisting of four parts. First, we introduce the contrastive learning objective of SCCL in Section 3.1. Second, the selection of learned representations is shown in Section 3.2. Third, an instancewise distillation module and a memory replay module are introduced to preserve learned knowledge in Section 3.3 and 3.4, respectively. Fourth, the kNN inference procedure is shown in 3.5, respectively. The training algorithm is shown in Algorithm 1.\nFormally, a model learns several tasks denoted as {T i }, i = 1, 2, ..., n (i is the number of tasks). Each task T i contains a limited set of labels C i . During the training of the task T i , only the corresponding data D i = {(x i j , y i j )} are available, where x i j is the input text and y i j ∈ C i is the corresponding label. In the scenario of task-incremental continual training, the task id can be observed when carrying out inference, and for generality, we consider the label set\nC j ∩ C k = ∅, if i = j."},{"figure_ref":[],"heading":"Supervised Contrastive Continual","publication_ref":["b14"],"table_ref":[],"text":"Training (SCCL)\nDuring the learning on the task T i , we first feed the input x i j into a pre-trained language model to obtain hidden states. The hidden states of a special token\n[CLS] (the beginning token of the pre-trained language model) are regarded as the representation of the input sequence:\nh i j = N orm(LM i (x i j )[CLS]),(1)\nwhere N orm(•) refers to normalization, LM i is the language model encoder trained for the task T i , and LM 0 is the initial pre-trained language model. We denote the data samples in a mini-batch as A (we omit the corner mark i during task T i for simplicity). For each data sample j, we denote N (j) ≡ A/{j}, and the positive neighbor set of it as P (j) = {u|y u = y j and u ∈ N (j)}. To push the representations with different labels away, and pull them with the same labels together, we use supervised contrastive learning objective following Khosla et al. (2020):\nL cl = j∈A -1 |P (j)| p∈P (j) log exp(h i j • h i p /κ) a∈N (j) exp(h i j • h i a /κ) (2)\nwhere κ is the hyper-parameter of temperature."},{"figure_ref":[],"heading":"Sample Selection","publication_ref":[],"table_ref":[],"text":"After training on each task T i , we select m samples from training data of D i to keep the representation distribution with respect to the labels (Algorithm 1 (18-22)). In particular, we adopt a K-means module to aggregate the data D i (c) of each label (c ∈ C i ) to clusters. Then we randomly select samples according to the data density to keep representation distribution, which can be formulated as:\nM c = Sample(Kmeans(D i (c)), c, m |C i | ). (3)\nThe selected samples for task T i are the union of selected data for each label c that\nM i = ∪ c∈C i M c .\nM i is saved in the memory buffer and serves as the classification criteria for task T i in the continual learning process."},{"figure_ref":[],"heading":"Instance-wise Relation Distillation (IRD)","publication_ref":["b8","b2"],"table_ref":[],"text":"To preserve the knowledge learned for previous tasks, inspired by Fang et al. (2020) and Cha et al. (2021), we use an instance-wise relation distillation term to control representation drift (Algorithm 1 (7-9)). During the learning on task T i , i > 1, the normalized instance-wise similarity in the minibatch A is calculated as:\ns i j,p = exp(h i j • h i p /τ ) a∈N (j) exp(h i j • h i a /τ ) ,(4)\nwhere N (j) ≡ A/{j}, the representations are encoded by the model LM i and τ is the hyperparameter temperature. Then the IRD regularization term follows:\nL IRD = 1 |A| 2 j p s i-1 j,p log s i j,p .(5)\nThe IRD regularization term aims to estimate the discrepancy of current representations to those learned in the previous model, and mitigate the representation drift through optimization. In this way, the knowledge of previous models is preserved and the CF problem can be mitigated.\nThe overall training objective can be denoted as follows:\nL = L cl + L IRD .(6)"},{"figure_ref":[],"heading":"Memory Replay (MR)","publication_ref":[],"table_ref":[],"text":"To make full use of the memory buffer saved during training, we use a memory replay module (de Masson D' Autume et al., 2019) to further recover the knowledge learned in the previous tasks (Algorithm 1 (14-16)). In the training on the task T i , i > 1, we revisit the samples in the memory buffer and train the model with the same loss in Eq (2) after training every f step on the current task."},{"figure_ref":[],"heading":"Inference","publication_ref":["b13"],"table_ref":[],"text":"After learning the task T i , we can obtain the model LM i . During the inference for previous tasks T u , u <= i, we feed each test data x u j into LM i and obtain the corresponding representation h u j . Then we retrieve the k buffered data from M u whose cosine similarity with h u j is the largest. Note that the representations of buffered data are obtained using the current model, which can adapt to the representation drift for parameter update. We denote the k nearest neighbors as (h u k , y u k ) ∈ K u j . The retrieved set is converted to a probability distribution over the labels by applying a softmax with \nM c = Sample(Kmeans(D i (c)), c, m |C i | ); 21: M i = M i ∪ M c ; 22:\nend for 23: M = M + M i ; 24: end for temperature T to the similarity. Using the temperature T > 1 can flatten the distribution, and prevent over-fitting to the most similar searches (Khandelwal et al., 2020). The probability distribution on the labels is expressed as follows:\np k (y j ) ∝ (h u k ,y u k )∈K u j 1 y j =y u k • exp( h u j • h u k T ),(7)\nand the label with the largest probability is taken as the prediction result.\n4 Experimental Setting"},{"figure_ref":[],"heading":"Tasks","publication_ref":["b28","b10","b29","b30","b21","b38","b38"],"table_ref":["tab_1"],"text":"We adopt classification tasks from the benchmark GLUE (Wang et al., 2018) and those from MBPA++ (Huang et al., 2021;de Masson D'Autume et al., 2019). We select dissimilar tasks to form the task sequences, i.e. there are no overlap labels between each task. The tasks contain 1) CoLA (Warstadt et al., 2019), requiring the model to determine whether a sentence is linguistically acceptable; 2) MNLI (Williams et al., 2017) mation; 3) QNLI1 , requiring deciding whether the answer answers the question; 4) QQP, (parsed from SQuAD (Rajpurkar et al., 2016)), testing whether a pair of Quora questions are synonymous; 5) Yelp (Zhang et al., 2015), requiring detecting the sentiment of a sentence; 6) AG (Zhang et al., 2015), requiring to classify the topics of the news.\nThe sequences can be divided into 2 types with respect to the task lengths: 1) a sequence of 4 classification tasks containing AG, Yelp, QNLI, and MRPC; 2) a sequence of 6 classification tasks containing AG, MRPC, MNLI, CoLA, Yelp, and QNLI. Without losing generality the orders are randomly selected and the task orders for experiments are shown in Table 1."},{"figure_ref":[],"heading":"Evaluation Metrics","publication_ref":["b18"],"table_ref":[],"text":"We adopt the metrics of average accuracy (ACC) and backward transfer (BWT) to evaluate the performance of the continual learning model (Lopez-Paz and Ranzato, 2017). The model trained after the task T i is evaluated on the test set of earlier tasks T j (j <= i), and the test accuracy is denoted as R i j . The metrics are shown as follows:\nACC = 1 n n i=1 R n i (8) BW T = 1 n -1 n-1 i=1 R n i -R i i ,(9)\nwhere the former evaluates the overall performance of the final trained model, and the latter calculates the knowledge forgetting during the continual training procedure."},{"figure_ref":[],"heading":"Baselines","publication_ref":["b10","b2","b32","b23"],"table_ref":[],"text":"We not only compare our model with several CEbased continual learning methods but extend training strategies of them to our contrastive learning framework (i.e. training with contrastive learning and inferring with kNN) to verify the effectiveness of contrastive learning in mitigating CF. We also compare our model with the competitive models IRDB (Huang et al., 2021) and Co 2 L (Cha et al., 2021). The shared hyper-parameters are kept the same as SCCL in baselines. The model details are as follows:\n• Fine-tune (CE, CL) (Yogatama et al., 2019) modifies the parameters of the pre-trained language model to adapt to a new task without any augmented strategies and additional loss. • Experience Replay (ER) (Riemer et al., 2019) stores a small subset of samples from previous tasks and replays those to prevent models from forgetting past knowledge. (Yu et al., 2020b) trains on all the tasks simultaneously, i.e. the data of different tasks are mixed up for training. It does not suffer from catastrophic forgetting and represents an upper bound on model performance."},{"figure_ref":[],"heading":"Implementation Details","publication_ref":["b15"],"table_ref":[],"text":"We adopt the officially released roberta-base from HuggingFace 2 as our backbone network. We train our model on 1 GPU (A100 80G) using the Adam optimizer (Kingma and Ba, 2014). For all the models, the batch size is 96, the learning rate is 3e-5, "},{"figure_ref":[],"heading":"Results","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Overall Results","publication_ref":[],"table_ref":["tab_4"],"text":"The overall results of our experiments are shown in Table 2. First, our model SCCL achieves ACCs of 79.20%, 80.05%, 80.24%, 78.36%, 79.00%, and 78.55% in Order 1-6, respectively, which are 2.37%, 0.9%, 0.66%, 3.28%, 2.49%, and 2.40% higher than the second-best performance of the continual learning baselines. It shows that the performance of the continually learned model is wellmaintained in SCCL, but the problem of CF still exists. SCCL achieves state-of-the-art ACCs compared with the baseline models, indicating the effectiveness of our proposed framework. We also observe that the performance variance is small in the SCCL model for different orders, which implies that our models are not sensitive to the order of task sequences.\nSecond, the results of BWT range from -3.75% to 0.57% in SCCL for Orders 1-6, which demonstrates knowledge forgetting during the continual learning procedure. The results of SCCL are rela-tively higher than CE-based models, indicating that SCCL suffers from a milder impact of CF. Note that the BWT of SCCL is 0.57% in Order 4, which indicates that SCCL can even backward transfer the knowledge from the current tasks to previous tasks. But compared with CL-LwF, CL-MAS, and CL-EWC, the values of ACCs in SCCL are higher, but BWTs are adverse. It implies that using the regularization-based strategies, the fine-tuning performance is destructed for explicit control of representations. In this way, BWTs become low since the fine-tuning performance on downstream tasks is relatively weak.\nThird, the extended CL-based models achieve stronger performance than corresponding standard CE-based models. For example, the model CL-LwF achieves ACCs of 76.53%, 79.15%, 79.58%, 68.24%, 72.39%, and 71.48%, which are 4.44%, 7.02%, 6.04%, 0.01%, 9.24% and 3.56% higher than those of CE-LwF. The results of CL, CL-MAS, and CL-EWC are in a similar pattern. The results reflect that contrastive learning with a kNN classifier for continual learning has a stronger ability to overcome CF. But we observe that Co 2 L achieves relatively low performance compared with our model, which proves that Co 2 L is not effective for task-incremental learning. It can be explained that Co 2 L keeps the knowledge of classes and separate the tasks with clear boundary, by using asymmetric supervised contrastive loss, which makes it difficult to distinguish a representation for different task purposes.\nFinally, we observe a significant variance in the results of different task orders for regularizationbased methods. For example, ACCs of CL-EWC range from 75.89% to 66.74%. But in CE-ER or SCCL the variance is less drastic, such as CE-ER ranging from 76.90% to 75.08% and SCCL ranging from 80.24% to 78.55%. The phenomenon may result from that knowledge forgetting of previous tasks increases step by step for information los, but no samples help recover such information in regularization-based methods."},{"figure_ref":[],"heading":"Ablation Study","publication_ref":[],"table_ref":["tab_4"],"text":"We show the ablation study of memory replay and IRD in the last two rows in Table 2. The ACCs of the models w/o memory replay range from 71.22%, to 80.27% for Order 1-6, which are 2.01%, 1.42%, -0.03%, 2.71%, 7.78%, and 3.91% lower than SCCL, respectively. It shows the effectiveness of memory replay, without which ACC also becomes less robust to task orders. Then ACCs of the models w/o IRD are 1.63%, 0.32%, 0.76%, 4.49%, 2.11%, and 4.23% lower than SCCL for Order 1-6, respectively. We observe that the models w/o IRD are more robust to task orders, which implies that rehearsal-based methods are less sensitive to task sequences. Comparing the model w/o IRD with CE-ER, the model performance are also higher than those of CE-ER, which uses almost the same training strategy. The phenomenon demonstrates the effectiveness of contrastive learning in overcoming CF.\nWe also compare our model with Co 2 L in abla-tion studies (Figure 4). "},{"figure_ref":[],"heading":"Detailed Results","publication_ref":[],"table_ref":[],"text":"As an example, we show the detailed results of Order 3 in several models (Figure 3). and MRPC. The accuracy of QNLI decreases by 17.38%, that of Yelp decreases by 11.59%, and that of MRPC decreases by 14.84%. As for our model SCCL, we observe that the test performance is 88.70%, 87.24%, 86.24% and 85.68% after training on tasks QNLI, Yelp, MRPC, and AG, respectively. It shows that the performance of SCCL decreases as the training precedes, but within a small range (3.02%). The results on Yelp and MRPC are in a similar pattern. It demonstrates that our model has a strong ability to overcome CF."},{"figure_ref":["fig_4"],"heading":"Visualization","publication_ref":[],"table_ref":[],"text":"We use t-SNE to visualize the representations of QNLI in Order 3 of the training models, CE, CE-ER, CL, and SCCL (Figure 5). As we observe in CE the representations of the test data are clearly separated into two clusters after training on the task QNLI. When finishing the continual learning, the representations become nearly uniformly distributed on the feature space and the model only achieves an accuracy of 50.21%. It demonstrates that catastrophic forgetting is significant due to representation drift. In the model CL, the representations drift severely as well, but the distribution is less uniform compared with CE. Typically, we can clearly at the upper right of the distribution, there are more memory samples with label 1, and the test samples with label 1 also gather in the position, indicating correct classification based on kNN. The test performance achieves 61.74%, but is still 26.94% lower than the initial model. The phenomenon shows representations during continual learning drift less significantly and the saved sam-ples (the classification criterion) also drift, which maintains some correct inferences. But CF is still a salient problem in contrastive learning.\nBut in CE-ER, the boundary of the representations becomes indistinct, and the accuracy of QNLI decreases from 87.79% to 70.11% after continual learning. It indicates that the representations are less effective compared with the initially trained, i.e. CF is significant in CE-ER. But the representations in SCCL are still clearly divided into two parts according to the labels. The representations of the memory samples are among the according clusters, implying the performance on the task QNLI is well-maintained. Correspondingly, the accuracy at the end of learning is 85.68% based on SCCL, only 3.02% lower than the initial performance. It shows that in SCCL the representation drift slightly and the classification criterion is well-maintained, resulting in a satisfactory performance."},{"figure_ref":[],"heading":"Conclusion","publication_ref":["b9"],"table_ref":[],"text":"In this paper, we proposed a supervised contrastive learning model for task-incremental continual learning (SCCL) to boost the extensibility of continual learning. The model used contrastive learning to learn representations and a kNN module was adopted for inference, together with an instance-wise distillation and a memory replay module to maintain previously learned knowledge.\nWith extensive experiments, our model achieved state-of-the-art performance compared with standard CE-based methods. Ablation studies and visualizations also proved the effectiveness of our model in solving the problem of CF.\nOur model SCCL is specific for task-incremental continual learning scenarios, but not suitable for class-incremental scenarios. In class-incremental scenarios, the representations of current classes should be designed to be far away from previous ones. For simplicity, we do not consider data augmentation in our model, so the batch size should be large enough to contain positive pairs for each label. But data augmentation (such as two different dropout representations (Gao et al., 2021)) is a plug-and-play module for our model if there are plenty of labels in each task."},{"figure_ref":[],"heading":"A Data Statistics","publication_ref":[],"table_ref":["tab_6"],"text":"We show the data statistics in Table 3. "},{"figure_ref":[],"heading":"B kNN Sensitivity","publication_ref":[],"table_ref":[],"text":"We show the sensitivity of SCCL to the number of k in the kNN module (Figure 6). We find that the performance of our model fluctuates from 80.24% to 80.28% (a significantly small range) in our method, indicating the representations in our model cluster well in the feature space and are robust to the hyperparameter k. But the performance in CL fluctuates more severely, ranging from 59.02% to 60.12%. The best performance is achieved when k = 10 and decreases with the increase of k, which means the representations drift significantly and the clusters become less reliable. The experiment demonstrates the effectiveness of IRD regularization term and the memory replay module in maintaining the representation distribution, and without them the representations drift significantly, suffering from the CF problem. "}],"string":"[\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b12\",\n \"b6\",\n \"b12\",\n \"b18\",\n \"b17\",\n \"b16\",\n \"b18\",\n \"b17\",\n \"b16\",\n \"b0\",\n \"b23\",\n \"b18\",\n \"b9\",\n \"b37\",\n \"b20\",\n \"b39\",\n \"b11\",\n \"b13\",\n \"b8\",\n \"b2\",\n \"b16\",\n \"b0\",\n \"b17\"\n ],\n \"table_ref\": [],\n \"text\": \"Continual learning aims to continually train models with new tasks without forgetting previously learned tasks (Ke and Liu, 2022;De Lange et al., 2022). It has become a promising direction for NLP models to incrementally learn new tasks/domains/classes as humans do (Ke and Liu, 2022). A typical scenario aims to enable NLP models to solve various tasks in an incremental manner, namely the task-incremental continual learning scenario, which is our study setting in this paper. A salient challenge for continual learning is that continually learned models usually suffer from cartographic forgetting (CF), i.e. the performance on previously learned tasks decreases after training on the new one (Lopez-Paz and Ranzato, 2017).\\nVarious training strategies have been proposed to mitigate CF (Li and Hoiem, 2017;Kirkpatrick et al., 2017;Lopez-Paz and Ranzato, 2017). Under the fixed model structure, regularization-based methods design regularization terms to control the shift of representations learned from previous tasks (Li and Hoiem, 2017;Kirkpatrick et al., 2017;Aljundi et al., 2018). Rehearsal-based methods save the data samples from previous tasks into a memory buffer and re-train the model to recover knowledge during training on the current task (Riemer et al., 2019;Lopez-Paz and Ranzato, 2017;de Masson D'Autume et al., 2019). However, most continual learning methods are designed to recover the learned knowledge or mitigate the representation of forgetting.\\nSeldom considers adapting the classification criterion for the newly learned representations. For example, in supervised contrastive learning, the contrastive objective is designed to pull the data representations with the same labels together and push representations with different labels away (Chen et al., 2020a;Gao et al., 2021;Zhang et al., 2021;Neelakantan et al., 2022;Zhao et al., 2022). Representations of the training samples can be saved as a classification criterion, after which an instancebased method such as a k Nearest Neighbor (kNN) module can be leveraged for inference (Kassner and Schütze, 2020;Khandelwal et al., 2020). After learning the new task, we can feed the saved sample into models for new classification criteria and mitigate the problem of CF. For example in Figure 1, although the representations have decayed for learning the new task, the saved samples adapt to serve as the classification criterion in kNN modules.\\nInspired by the above motivation, we investigate the use of supervised contrastive learning for taskincremental continual learning (SCCL). After supervised contrastive learning on each task, we use a K-means module to select several samples and save them into a memory buffer while maintaining the learned representation distribution. In addition, to mitigate the representation drift when training the model for new tasks, we use an instance-wise relation distillation (IRD) term (Fang et al., 2020;Cha et al., 2021) and a memory replay module (de Masson D'Autume et al., 2019) to maintain the learned knowledge. During inference, the saved samples are fed into the trained model to obtain updated representations and a kNN module is used for classification.\\nExperimental results show that our proposed model can achieve state-of-the-art performance compared with standard cross-entropy-based (CE) baselines. We additionally extend different continual learning strategies (Kirkpatrick et al., 2017;Aljundi et al., 2018;Li and Hoiem, 2017) to the supervised contrastive continual learning framework, which gives stronger results than corresponding CE-based methods, showing the advantage of contrastive learning with a kNN classifier in continual learning scenarios. We further analyze the effectiveness of each module in our paper through ablation studies. To our knowledge, we are the first to propose a supervised contrastive learning framework for task-incremental continual learning, without any augmented parameters. The code will be released when accepted.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Work\",\n \"publication_ref\": [\n \"b33\",\n \"b26\",\n \"b17\",\n \"b16\",\n \"b24\",\n \"b16\",\n \"b0\",\n \"b17\",\n \"b17\",\n \"b24\",\n \"b1\",\n \"b31\",\n \"b9\",\n \"b14\",\n \"b19\",\n \"b2\"\n ],\n \"table_ref\": [],\n \"text\": \"Continual Learning Various continual learning methods have been proposed to mitigate the problem of CF. The methods can be broadly divided into architecture-based methods (Yoon et al., 2018;Serra et al., 2018), regularization-based methods Li and Hoiem (2017); Kirkpatrick et al. (2017), and rehearsal-based methods (Rolnick et al., 2019). Under the fixed model structure, regularizationbased methods (Kirkpatrick et al., 2017;Aljundi et al., 2018;Li and Hoiem, 2017) optimize network parameters on the current task while constraining the representation drift. For example, Li and Hoiem (2017) propose learning without forgetting (LwF) to tackle this problem, which regularizes the model output of current data close to those trained for the previous model. Another category of fixed-structure strategies (rehearsal-based) stores a limited subset of samples from previous tasks to mitigate CF such as ER (Rolnick et al., 2019), RM (Bang et al., 2021), andiCaRL (Rebuffi et al., 2017).\\nContrastive Learning Contrastive learning is initially introduced in self-supervised settings and proved to subsume or significantly outperform traditional contrastive losses such as triplet loss (Chen et al., 2020b;Wu et al., 2018;Gao et al., 2021;?). For example, Khosla et al. (2020) first propose the idea of self-supervised contrastive learning and prove that the method is more robust to natural corruptions, stable to hyper-parameter settings, and has strong transfer performance. Luo et al. (2022) uses supervised contrastive learning combined with a kNN inference module for cross-domain sentiment analysis, showing a stronger generalization ability compared with standard CE-based methods. Cha et al. (2021) propose a contrastive continual learning method, Co 2 L, for class-incremental continual learning. The method uses an asymmetric supervised contrastive loss to enlarge the distance between representations of previous and new tasks. However, there are significant differences between our model and Co 2 L. First, the asymmetric contrastive loss of Co 2 L is unsuitable for taskincremental continual learning, because a representation can be predicted as different labels according to task objectives. Second, Co 2 L uses a decoupled classification layer for inference, i.e. it learns rep- resentations first and then learns a linear classifier separately, causing low extensibility and high complexity of the model. In contrast, we use a kNN module as the classification criteria to enhance the extensibility of the model. Third, Co 2 L only considers the representation drift from the view of regularization. But we also mitigate the problem of representation drift by feeding memory data into the current model to obtain updated classification criteria.\"\n },\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"Method\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The overall SCCL framework is illustrated in Figure 2, consisting of four parts. First, we introduce the contrastive learning objective of SCCL in Section 3.1. Second, the selection of learned representations is shown in Section 3.2. Third, an instancewise distillation module and a memory replay module are introduced to preserve learned knowledge in Section 3.3 and 3.4, respectively. Fourth, the kNN inference procedure is shown in 3.5, respectively. The training algorithm is shown in Algorithm 1.\\nFormally, a model learns several tasks denoted as {T i }, i = 1, 2, ..., n (i is the number of tasks). Each task T i contains a limited set of labels C i . During the training of the task T i , only the corresponding data D i = {(x i j , y i j )} are available, where x i j is the input text and y i j ∈ C i is the corresponding label. In the scenario of task-incremental continual training, the task id can be observed when carrying out inference, and for generality, we consider the label set\\nC j ∩ C k = ∅, if i = j.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Supervised Contrastive Continual\",\n \"publication_ref\": [\n \"b14\"\n ],\n \"table_ref\": [],\n \"text\": \"Training (SCCL)\\nDuring the learning on the task T i , we first feed the input x i j into a pre-trained language model to obtain hidden states. The hidden states of a special token\\n[CLS] (the beginning token of the pre-trained language model) are regarded as the representation of the input sequence:\\nh i j = N orm(LM i (x i j )[CLS]),(1)\\nwhere N orm(•) refers to normalization, LM i is the language model encoder trained for the task T i , and LM 0 is the initial pre-trained language model. We denote the data samples in a mini-batch as A (we omit the corner mark i during task T i for simplicity). For each data sample j, we denote N (j) ≡ A/{j}, and the positive neighbor set of it as P (j) = {u|y u = y j and u ∈ N (j)}. To push the representations with different labels away, and pull them with the same labels together, we use supervised contrastive learning objective following Khosla et al. (2020):\\nL cl = j∈A -1 |P (j)| p∈P (j) log exp(h i j • h i p /κ) a∈N (j) exp(h i j • h i a /κ) (2)\\nwhere κ is the hyper-parameter of temperature.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Sample Selection\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"After training on each task T i , we select m samples from training data of D i to keep the representation distribution with respect to the labels (Algorithm 1 (18-22)). In particular, we adopt a K-means module to aggregate the data D i (c) of each label (c ∈ C i ) to clusters. Then we randomly select samples according to the data density to keep representation distribution, which can be formulated as:\\nM c = Sample(Kmeans(D i (c)), c, m |C i | ). (3)\\nThe selected samples for task T i are the union of selected data for each label c that\\nM i = ∪ c∈C i M c .\\nM i is saved in the memory buffer and serves as the classification criteria for task T i in the continual learning process.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Instance-wise Relation Distillation (IRD)\",\n \"publication_ref\": [\n \"b8\",\n \"b2\"\n ],\n \"table_ref\": [],\n \"text\": \"To preserve the knowledge learned for previous tasks, inspired by Fang et al. (2020) and Cha et al. (2021), we use an instance-wise relation distillation term to control representation drift (Algorithm 1 (7-9)). During the learning on task T i , i > 1, the normalized instance-wise similarity in the minibatch A is calculated as:\\ns i j,p = exp(h i j • h i p /τ ) a∈N (j) exp(h i j • h i a /τ ) ,(4)\\nwhere N (j) ≡ A/{j}, the representations are encoded by the model LM i and τ is the hyperparameter temperature. Then the IRD regularization term follows:\\nL IRD = 1 |A| 2 j p s i-1 j,p log s i j,p .(5)\\nThe IRD regularization term aims to estimate the discrepancy of current representations to those learned in the previous model, and mitigate the representation drift through optimization. In this way, the knowledge of previous models is preserved and the CF problem can be mitigated.\\nThe overall training objective can be denoted as follows:\\nL = L cl + L IRD .(6)\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Memory Replay (MR)\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"To make full use of the memory buffer saved during training, we use a memory replay module (de Masson D' Autume et al., 2019) to further recover the knowledge learned in the previous tasks (Algorithm 1 (14-16)). In the training on the task T i , i > 1, we revisit the samples in the memory buffer and train the model with the same loss in Eq (2) after training every f step on the current task.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Inference\",\n \"publication_ref\": [\n \"b13\"\n ],\n \"table_ref\": [],\n \"text\": \"After learning the task T i , we can obtain the model LM i . During the inference for previous tasks T u , u <= i, we feed each test data x u j into LM i and obtain the corresponding representation h u j . Then we retrieve the k buffered data from M u whose cosine similarity with h u j is the largest. Note that the representations of buffered data are obtained using the current model, which can adapt to the representation drift for parameter update. We denote the k nearest neighbors as (h u k , y u k ) ∈ K u j . The retrieved set is converted to a probability distribution over the labels by applying a softmax with \\nM c = Sample(Kmeans(D i (c)), c, m |C i | ); 21: M i = M i ∪ M c ; 22:\\nend for 23: M = M + M i ; 24: end for temperature T to the similarity. Using the temperature T > 1 can flatten the distribution, and prevent over-fitting to the most similar searches (Khandelwal et al., 2020). The probability distribution on the labels is expressed as follows:\\np k (y j ) ∝ (h u k ,y u k )∈K u j 1 y j =y u k • exp( h u j • h u k T ),(7)\\nand the label with the largest probability is taken as the prediction result.\\n4 Experimental Setting\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Tasks\",\n \"publication_ref\": [\n \"b28\",\n \"b10\",\n \"b29\",\n \"b30\",\n \"b21\",\n \"b38\",\n \"b38\"\n ],\n \"table_ref\": [\n \"tab_1\"\n ],\n \"text\": \"We adopt classification tasks from the benchmark GLUE (Wang et al., 2018) and those from MBPA++ (Huang et al., 2021;de Masson D'Autume et al., 2019). We select dissimilar tasks to form the task sequences, i.e. there are no overlap labels between each task. The tasks contain 1) CoLA (Warstadt et al., 2019), requiring the model to determine whether a sentence is linguistically acceptable; 2) MNLI (Williams et al., 2017) mation; 3) QNLI1 , requiring deciding whether the answer answers the question; 4) QQP, (parsed from SQuAD (Rajpurkar et al., 2016)), testing whether a pair of Quora questions are synonymous; 5) Yelp (Zhang et al., 2015), requiring detecting the sentiment of a sentence; 6) AG (Zhang et al., 2015), requiring to classify the topics of the news.\\nThe sequences can be divided into 2 types with respect to the task lengths: 1) a sequence of 4 classification tasks containing AG, Yelp, QNLI, and MRPC; 2) a sequence of 6 classification tasks containing AG, MRPC, MNLI, CoLA, Yelp, and QNLI. Without losing generality the orders are randomly selected and the task orders for experiments are shown in Table 1.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Evaluation Metrics\",\n \"publication_ref\": [\n \"b18\"\n ],\n \"table_ref\": [],\n \"text\": \"We adopt the metrics of average accuracy (ACC) and backward transfer (BWT) to evaluate the performance of the continual learning model (Lopez-Paz and Ranzato, 2017). The model trained after the task T i is evaluated on the test set of earlier tasks T j (j <= i), and the test accuracy is denoted as R i j . The metrics are shown as follows:\\nACC = 1 n n i=1 R n i (8) BW T = 1 n -1 n-1 i=1 R n i -R i i ,(9)\\nwhere the former evaluates the overall performance of the final trained model, and the latter calculates the knowledge forgetting during the continual training procedure.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Baselines\",\n \"publication_ref\": [\n \"b10\",\n \"b2\",\n \"b32\",\n \"b23\"\n ],\n \"table_ref\": [],\n \"text\": \"We not only compare our model with several CEbased continual learning methods but extend training strategies of them to our contrastive learning framework (i.e. training with contrastive learning and inferring with kNN) to verify the effectiveness of contrastive learning in mitigating CF. We also compare our model with the competitive models IRDB (Huang et al., 2021) and Co 2 L (Cha et al., 2021). The shared hyper-parameters are kept the same as SCCL in baselines. The model details are as follows:\\n• Fine-tune (CE, CL) (Yogatama et al., 2019) modifies the parameters of the pre-trained language model to adapt to a new task without any augmented strategies and additional loss. • Experience Replay (ER) (Riemer et al., 2019) stores a small subset of samples from previous tasks and replays those to prevent models from forgetting past knowledge. (Yu et al., 2020b) trains on all the tasks simultaneously, i.e. the data of different tasks are mixed up for training. It does not suffer from catastrophic forgetting and represents an upper bound on model performance.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Implementation Details\",\n \"publication_ref\": [\n \"b15\"\n ],\n \"table_ref\": [],\n \"text\": \"We adopt the officially released roberta-base from HuggingFace 2 as our backbone network. We train our model on 1 GPU (A100 80G) using the Adam optimizer (Kingma and Ba, 2014). For all the models, the batch size is 96, the learning rate is 3e-5, \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Overall Results\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_4\"\n ],\n \"text\": \"The overall results of our experiments are shown in Table 2. First, our model SCCL achieves ACCs of 79.20%, 80.05%, 80.24%, 78.36%, 79.00%, and 78.55% in Order 1-6, respectively, which are 2.37%, 0.9%, 0.66%, 3.28%, 2.49%, and 2.40% higher than the second-best performance of the continual learning baselines. It shows that the performance of the continually learned model is wellmaintained in SCCL, but the problem of CF still exists. SCCL achieves state-of-the-art ACCs compared with the baseline models, indicating the effectiveness of our proposed framework. We also observe that the performance variance is small in the SCCL model for different orders, which implies that our models are not sensitive to the order of task sequences.\\nSecond, the results of BWT range from -3.75% to 0.57% in SCCL for Orders 1-6, which demonstrates knowledge forgetting during the continual learning procedure. The results of SCCL are rela-tively higher than CE-based models, indicating that SCCL suffers from a milder impact of CF. Note that the BWT of SCCL is 0.57% in Order 4, which indicates that SCCL can even backward transfer the knowledge from the current tasks to previous tasks. But compared with CL-LwF, CL-MAS, and CL-EWC, the values of ACCs in SCCL are higher, but BWTs are adverse. It implies that using the regularization-based strategies, the fine-tuning performance is destructed for explicit control of representations. In this way, BWTs become low since the fine-tuning performance on downstream tasks is relatively weak.\\nThird, the extended CL-based models achieve stronger performance than corresponding standard CE-based models. For example, the model CL-LwF achieves ACCs of 76.53%, 79.15%, 79.58%, 68.24%, 72.39%, and 71.48%, which are 4.44%, 7.02%, 6.04%, 0.01%, 9.24% and 3.56% higher than those of CE-LwF. The results of CL, CL-MAS, and CL-EWC are in a similar pattern. The results reflect that contrastive learning with a kNN classifier for continual learning has a stronger ability to overcome CF. But we observe that Co 2 L achieves relatively low performance compared with our model, which proves that Co 2 L is not effective for task-incremental learning. It can be explained that Co 2 L keeps the knowledge of classes and separate the tasks with clear boundary, by using asymmetric supervised contrastive loss, which makes it difficult to distinguish a representation for different task purposes.\\nFinally, we observe a significant variance in the results of different task orders for regularizationbased methods. For example, ACCs of CL-EWC range from 75.89% to 66.74%. But in CE-ER or SCCL the variance is less drastic, such as CE-ER ranging from 76.90% to 75.08% and SCCL ranging from 80.24% to 78.55%. The phenomenon may result from that knowledge forgetting of previous tasks increases step by step for information los, but no samples help recover such information in regularization-based methods.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Ablation Study\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_4\"\n ],\n \"text\": \"We show the ablation study of memory replay and IRD in the last two rows in Table 2. The ACCs of the models w/o memory replay range from 71.22%, to 80.27% for Order 1-6, which are 2.01%, 1.42%, -0.03%, 2.71%, 7.78%, and 3.91% lower than SCCL, respectively. It shows the effectiveness of memory replay, without which ACC also becomes less robust to task orders. Then ACCs of the models w/o IRD are 1.63%, 0.32%, 0.76%, 4.49%, 2.11%, and 4.23% lower than SCCL for Order 1-6, respectively. We observe that the models w/o IRD are more robust to task orders, which implies that rehearsal-based methods are less sensitive to task sequences. Comparing the model w/o IRD with CE-ER, the model performance are also higher than those of CE-ER, which uses almost the same training strategy. The phenomenon demonstrates the effectiveness of contrastive learning in overcoming CF.\\nWe also compare our model with Co 2 L in abla-tion studies (Figure 4). \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Detailed Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"As an example, we show the detailed results of Order 3 in several models (Figure 3). and MRPC. The accuracy of QNLI decreases by 17.38%, that of Yelp decreases by 11.59%, and that of MRPC decreases by 14.84%. As for our model SCCL, we observe that the test performance is 88.70%, 87.24%, 86.24% and 85.68% after training on tasks QNLI, Yelp, MRPC, and AG, respectively. It shows that the performance of SCCL decreases as the training precedes, but within a small range (3.02%). The results on Yelp and MRPC are in a similar pattern. It demonstrates that our model has a strong ability to overcome CF.\"\n },\n {\n \"figure_ref\": [\n \"fig_4\"\n ],\n \"heading\": \"Visualization\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We use t-SNE to visualize the representations of QNLI in Order 3 of the training models, CE, CE-ER, CL, and SCCL (Figure 5). As we observe in CE the representations of the test data are clearly separated into two clusters after training on the task QNLI. When finishing the continual learning, the representations become nearly uniformly distributed on the feature space and the model only achieves an accuracy of 50.21%. It demonstrates that catastrophic forgetting is significant due to representation drift. In the model CL, the representations drift severely as well, but the distribution is less uniform compared with CE. Typically, we can clearly at the upper right of the distribution, there are more memory samples with label 1, and the test samples with label 1 also gather in the position, indicating correct classification based on kNN. The test performance achieves 61.74%, but is still 26.94% lower than the initial model. The phenomenon shows representations during continual learning drift less significantly and the saved sam-ples (the classification criterion) also drift, which maintains some correct inferences. But CF is still a salient problem in contrastive learning.\\nBut in CE-ER, the boundary of the representations becomes indistinct, and the accuracy of QNLI decreases from 87.79% to 70.11% after continual learning. It indicates that the representations are less effective compared with the initially trained, i.e. CF is significant in CE-ER. But the representations in SCCL are still clearly divided into two parts according to the labels. The representations of the memory samples are among the according clusters, implying the performance on the task QNLI is well-maintained. Correspondingly, the accuracy at the end of learning is 85.68% based on SCCL, only 3.02% lower than the initial performance. It shows that in SCCL the representation drift slightly and the classification criterion is well-maintained, resulting in a satisfactory performance.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [\n \"b9\"\n ],\n \"table_ref\": [],\n \"text\": \"In this paper, we proposed a supervised contrastive learning model for task-incremental continual learning (SCCL) to boost the extensibility of continual learning. The model used contrastive learning to learn representations and a kNN module was adopted for inference, together with an instance-wise distillation and a memory replay module to maintain previously learned knowledge.\\nWith extensive experiments, our model achieved state-of-the-art performance compared with standard CE-based methods. Ablation studies and visualizations also proved the effectiveness of our model in solving the problem of CF.\\nOur model SCCL is specific for task-incremental continual learning scenarios, but not suitable for class-incremental scenarios. In class-incremental scenarios, the representations of current classes should be designed to be far away from previous ones. For simplicity, we do not consider data augmentation in our model, so the batch size should be large enough to contain positive pairs for each label. But data augmentation (such as two different dropout representations (Gao et al., 2021)) is a plug-and-play module for our model if there are plenty of labels in each task.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A Data Statistics\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_6\"\n ],\n \"text\": \"We show the data statistics in Table 3. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B kNN Sensitivity\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We show the sensitivity of SCCL to the number of k in the kNN module (Figure 6). We find that the performance of our model fluctuates from 80.24% to 80.28% (a significantly small range) in our method, indicating the representations in our model cluster well in the feature space and are robust to the hyperparameter k. But the performance in CL fluctuates more severely, ranging from 59.02% to 60.12%. The best performance is achieved when k = 10 and decreases with the increase of k, which means the representations drift significantly and the clusters become less reliable. The experiment demonstrates the effectiveness of IRD regularization term and the memory replay module in maintaining the representation distribution, and without them the representations drift significantly, suffering from the CF problem. \"\n }\n]"},"abstract":{"kind":"string","value":"Task-incremental continual learning refers to continually training a model in a sequence of tasks while overcoming the problem of catastrophic forgetting (CF). The issue arrives for the reason that the learned representations are forgotten for learning new tasks, and the decision boundary is destructed. Previous studies mostly consider how to recover the representations of learned tasks. It is seldom considered to adapt the decision boundary for new representations and in this paper we propose a Supervised Contrastive learning framework with adaptive classification criterion for Continual Learning (SCCL), In our method, a contrastive loss is used to directly learn representations for different tasks and a limited number of data samples are saved as the classification criterion. During inference, the saved data samples are fed into the current model to obtain updated representations, and a k Nearest Neighbour module is used for classification. In this way, the extensible model can solve the learned tasks with adaptive criteria of saved samples. To mitigate CF, we further use an instance-wise relation distillation regularization term and a memory replay module to maintain the information of previous tasks. Experiments show that SCCL achieves stateof-the-art performance and has a stronger ability to overcome CF compared with the classification baselines."},"title":{"kind":"string","value":"Mitigating Catastrophic Forgetting in Task-Incremental Continual Learning with Adaptive Classification Criterion"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 1 :1Figure 1: Illustration of representations after contrastive continual learning on a task before and after learning a new task.","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"Figure 2 :2Figure 2: The model framework of SCCL contains four main modules: (1) the supervised contrastive learning for each task; (2) the explicit control of catastrophic forgetting with IRD knowledge distillation and memory replay; (3) the selection of learned representations; (4) a kNN inference module.","figure_data":"","figure_id":"fig_1","figure_label":"2","figure_type":"figure"},{"figure_caption":"Figure 3 :Figure 4 :34Figure 3: Detailed results during continual learning procedure for different strategies in Order 3.","figure_data":"","figure_id":"fig_2","figure_label":"34","figure_type":"figure"},{"figure_caption":"First, we replace the kNN module of SCCL with a decoupled linear classifier like(Cha et al., 2021) (SCCL-CLS), where ACCs are slightly smaller than SCCL. It indicates that the kNN module in SCCL can achieve satisfactory performance without additional training on the final representations of contrastive learning. Then we replace the decoupled linear classifier of Co 2 L with our kNN module (Co 2 L-kNN), and we observe an increase in performance. It implies that the representations learned by Co 2 L are not separated clearly in the feature space, thus a trained linear layer is less effective for classification. But K-means selection of the samples and kNN inference module can estimate the representation distribution more precisely, resulting in better performance. Note that the results of SCCL are also stronger than Co 2 L-kNN, which indicates the effectiveness of our model on task-incremental continual learning.","figure_data":"","figure_id":"fig_3","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 55Figure 5: t-SNE visualization of the representations of QNLI samples learned based on the different continual learning methods in Order 5. 'E' refers to the representations at the end of continual learning.","figure_data":"","figure_id":"fig_4","figure_label":"5","figure_type":"figure"},{"figure_caption":"Algorithm 1 SCCL Training Input: A set of training task {T i } n , the corresponding data set {D i } n , sets of disjoint classes {C i } n . Training steps S and memory replay frequency f . Memory buffer size m. Initial pre-trained language model LM 0 . Output: Trained language model encoder LM n and memory buffer M.","figure_data":"1: Load pre-trained language model LM 0 ;2: M = []3: for i = 1, ..., n do4:for t = 1, ..., S do5:Draw mini-batch A from D i ;6:Calculate L cl of A with LM i (Eq (1-2));7:if i > 1 then8:Calculate LIRD of A (Eq (5));9:L = L cl + LIRD;10:else11:L = L cl ;12:end if13:Update model parameters with L;14:if i % f == 0 then15:Update model parameters with memory relay;16:end if17:end for18:for c ∈ C i do19:Obtain k-means clusters of data with label c;20:","figure_id":"tab_0","figure_label":"","figure_type":"table"},{"figure_caption":"containing 433k sentence pairs annotated with textual entailment infor-Orders 1 AG → Yelp → QNLI→ MRPC 2 MRPC → QNLI → Yelp →AG 3 QNLI →Yelp →MRPC→AG 4 AG→MRPC →CoLA→MNLI →Yelp→ QNLI 5 QNLI →Yelp →MNLI→CoLA →MRPC→ AG 6 MNLI → AG →QNLI→ MRPC →Yelp→ CoLA Different task orders for our experiments.","figure_data":"","figure_id":"tab_1","figure_label":"1","figure_type":"table"},{"figure_caption":"Continual Learning results on 6 different tasks. 'CE' refers to the standard cross-entropy-based methods, and 'CL' refers to extended contrastive-learning-based methods with continual learning strategies. '-' for not acquirable. All the results are averaged on 5 different random seeds.","figure_data":"61.87 -22.81","figure_id":"tab_4","figure_label":"2","figure_type":"table"},{"figure_caption":"Statistics for different classification tasks.","figure_data":"TaskType#Train #Test #LabelsAGNews16000 76004QNLIQ & A800052662YelpSentiment20000 76005CoLA Linguistics652710422MNLIInference12000 98153MRPC Paraphrase407417252","figure_id":"tab_6","figure_label":"3","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 1 :1Figure 1: Illustration of representations after contrastive continual learning on a task before and after learning a new task.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 2 :2Figure 2: The model framework of SCCL contains four main modules: (1) the supervised contrastive learning for each task; (2) the explicit control of catastrophic forgetting with IRD knowledge distillation and memory replay; (3) the selection of learned representations; (4) a kNN inference module.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 :Figure 4 :34Figure 3: Detailed results during continual learning procedure for different strategies in Order 3.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"34\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"First, we replace the kNN module of SCCL with a decoupled linear classifier like(Cha et al., 2021) (SCCL-CLS), where ACCs are slightly smaller than SCCL. It indicates that the kNN module in SCCL can achieve satisfactory performance without additional training on the final representations of contrastive learning. Then we replace the decoupled linear classifier of Co 2 L with our kNN module (Co 2 L-kNN), and we observe an increase in performance. It implies that the representations learned by Co 2 L are not separated clearly in the feature space, thus a trained linear layer is less effective for classification. But K-means selection of the samples and kNN inference module can estimate the representation distribution more precisely, resulting in better performance. Note that the results of SCCL are also stronger than Co 2 L-kNN, which indicates the effectiveness of our model on task-incremental continual learning.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 55Figure 5: t-SNE visualization of the representations of QNLI samples learned based on the different continual learning methods in Order 5. 'E' refers to the representations at the end of continual learning.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"5\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Algorithm 1 SCCL Training Input: A set of training task {T i } n , the corresponding data set {D i } n , sets of disjoint classes {C i } n . Training steps S and memory replay frequency f . Memory buffer size m. Initial pre-trained language model LM 0 . Output: Trained language model encoder LM n and memory buffer M.\",\n \"figure_data\": \"1: Load pre-trained language model LM 0 ;2: M = []3: for i = 1, ..., n do4:for t = 1, ..., S do5:Draw mini-batch A from D i ;6:Calculate L cl of A with LM i (Eq (1-2));7:if i > 1 then8:Calculate LIRD of A (Eq (5));9:L = L cl + LIRD;10:else11:L = L cl ;12:end if13:Update model parameters with L;14:if i % f == 0 then15:Update model parameters with memory relay;16:end if17:end for18:for c ∈ C i do19:Obtain k-means clusters of data with label c;20:\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"containing 433k sentence pairs annotated with textual entailment infor-Orders 1 AG → Yelp → QNLI→ MRPC 2 MRPC → QNLI → Yelp →AG 3 QNLI →Yelp →MRPC→AG 4 AG→MRPC →CoLA→MNLI →Yelp→ QNLI 5 QNLI →Yelp →MNLI→CoLA →MRPC→ AG 6 MNLI → AG →QNLI→ MRPC →Yelp→ CoLA Different task orders for our experiments.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_1\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Continual Learning results on 6 different tasks. 'CE' refers to the standard cross-entropy-based methods, and 'CL' refers to extended contrastive-learning-based methods with continual learning strategies. '-' for not acquirable. All the results are averaged on 5 different random seeds.\",\n \"figure_data\": \"61.87 -22.81\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Statistics for different classification tasks.\",\n \"figure_data\": \"TaskType#Train #Test #LabelsAGNews16000 76004QNLIQ & A800052662YelpSentiment20000 76005CoLA Linguistics652710422MNLIInference12000 98153MRPC Paraphrase407417252\",\n \"figure_id\": \"tab_6\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n }\n]"},"authors":{"kind":"string","value":"Yun Luo; Xiaotian Lin; Zhen Yang; Fandong Meng; Jie Zhou; Yue Zhang"},"references":{"kind":"list like","value":[{"authors":"Rahaf Aljundi; Francesca Babiloni; Mohamed Elhoseiny; Marcus Rohrbach; Tinne Tuytelaars","journal":"","ref_id":"b0","title":"Memory aware synapses: Learning what (not) to forget","year":"2018"},{"authors":"Jihwan Bang; Heesu Kim; Youngjoon Yoo; Jung-Woo Ha; Jonghyun Choi","journal":"","ref_id":"b1","title":"Rainbow memory: Continual learning with a memory of diverse samples","year":"2021"},{"authors":"Hyuntak Cha; Jaeho Lee; Jinwoo Shin","journal":"","ref_id":"b2","title":"Co2l: Contrastive continual learning","year":"2021"},{"authors":"Ting Chen; Simon Kornblith; Mohammad Norouzi; Geoffrey Hinton","journal":"","ref_id":"b3","title":"a. A simple framework for contrastive learning of visual representations","year":"2020"},{"authors":"Ting Chen; Simon Kornblith; Mohammad Norouzi; Geoffrey Hinton","journal":"","ref_id":"b4","title":"A simple framework for contrastive learning of visual representations","year":"2020"},{"authors":" Pmlr","journal":"","ref_id":"b5","title":"","year":""},{"authors":"Matthias De Lange; Rahaf Aljundi; Marc Masana; Sarah Parisot; Xu Jia; Aleš Leonardis; Gregory Slabaugh; Tinne Tuytelaars","journal":"IEEE Transactions on Pattern Analysis and Machine Intelligence","ref_id":"b6","title":"A continual learning survey: Defying forgetting in classification tasks","year":"2022"},{"authors":"Cyprien De Masson D'autume; Sebastian Ruder; Lingpeng Kong; Dani Yogatama","journal":"Advances in Neural Information Processing Systems","ref_id":"b7","title":"Episodic memory in lifelong language learning","year":"2019"},{"authors":"Zhiyuan Fang; Jianfeng Wang; Lijuan Wang; Lei Zhang; Yezhou Yang; Zicheng Liu","journal":"","ref_id":"b8","title":"Seed: Self-supervised distillation for visual representation","year":"2020"},{"authors":"Tianyu Gao; Xingcheng Yao; Danqi Chen","journal":"Association for Computational Linguistics","ref_id":"b9","title":"SimCSE: Simple contrastive learning of sentence embeddings","year":"2021"},{"authors":"Yufan Huang; Yanzhe Zhang; Jiaao Chen; Xuezhi Wang; Diyi Yang","journal":"","ref_id":"b10","title":"Continual learning for text classification with information disentanglement based regularization","year":"2021"},{"authors":"Nora Kassner; Hinrich Schütze","journal":"Association for Computational Linguistics","ref_id":"b11","title":"BERT-kNN: Adding a kNN search component to pretrained language models for better QA","year":"2020"},{"authors":"Zixuan Ke; Bing Liu","journal":"","ref_id":"b12","title":"Continual learning of natural language processing tasks: A survey","year":"2022"},{"authors":"Urvashi Khandelwal; Angela Fan; Dan Jurafsky; Luke Zettlemoyer; Mike Lewis","journal":"","ref_id":"b13","title":"Nearest neighbor machine translation","year":"2020"},{"authors":"Prannay Khosla; Piotr Teterwak; Chen Wang; Aaron Sarna; Yonglong Tian; Phillip Isola; Aaron Maschinot; Ce Liu; Dilip Krishnan","journal":"Advances in Neural Information Processing Systems","ref_id":"b14","title":"Supervised contrastive learning","year":"2020"},{"authors":"P Diederik; Jimmy Kingma; Ba","journal":"","ref_id":"b15","title":"Adam: A method for stochastic optimization","year":"2014"},{"authors":"James Kirkpatrick; Razvan Pascanu; Neil Rabinowitz; Joel Veness; Guillaume Desjardins; Andrei A Rusu; Kieran Milan; John Quan; Tiago Ramalho; Agnieszka Grabska-Barwinska","journal":"Proceedings of the national academy of sciences","ref_id":"b16","title":"Overcoming catastrophic forgetting in neural networks","year":"2017"},{"authors":"Zhizhong Li; Derek Hoiem","journal":"IEEE transactions on pattern analysis and machine intelligence","ref_id":"b17","title":"Learning without forgetting","year":"2017"},{"authors":"David Lopez; - Paz; Marc'aurelio Ranzato","journal":"Advances in neural information processing systems","ref_id":"b18","title":"Gradient episodic memory for continual learning","year":"2017"},{"authors":"Yun Luo; Fang Guo; Zihan Liu; Yue Zhang","journal":"","ref_id":"b19","title":"Mere contrastive learning for cross-domain sentiment analysis","year":"2022"},{"authors":"Arvind Neelakantan; Tao Xu; Raul Puri; Alec Radford; Jesse Michael Han; Jerry Tworek; Qiming Yuan; Nikolas Tezak; Jong Wook Kim; Chris Hallacy","journal":"","ref_id":"b20","title":"Text and code embeddings by contrastive pretraining","year":"2022"},{"authors":"Pranav Rajpurkar; Jian Zhang; Konstantin Lopyrev; Percy Liang","journal":"Association for Computational Linguistics","ref_id":"b21","title":"SQuAD: 100,000+ questions for machine comprehension of text","year":"2016"},{"authors":" Sylvestre-Alvise; Alexander Rebuffi; Georg Kolesnikov; Christoph H Sperl; Lampert","journal":"","ref_id":"b22","title":"icarl: Incremental classifier and representation learning","year":"2017"},{"authors":"Matthew Riemer; Ignacio Cases; Robert Ajemian; Miao Liu; Irina Rish; Yuhai Tu; Gerald Tesauro","journal":"ICLR","ref_id":"b23","title":"Learning to learn without forgetting by maximizing transfer and minimizing interference","year":"2019"},{"authors":"David Rolnick; Arun Ahuja; Jonathan Schwarz; Timothy Lillicrap; Gregory Wayne","journal":"","ref_id":"b24","title":"Experience replay for continual learning","year":"2019"},{"authors":"Curran Associates; Inc ","journal":"","ref_id":"b25","title":"","year":""},{"authors":"Joan Serra; Didac Suris; Marius Miron; Alexandros Karatzoglou","journal":"","ref_id":"b26","title":"Overcoming catastrophic forgetting with hard attention to the task","year":"2018"},{"authors":" Pmlr","journal":"","ref_id":"b27","title":"","year":""},{"authors":"Alex Wang; Amanpreet Singh; Julian Michael; Felix Hill; Omer Levy; Samuel Bowman","journal":"","ref_id":"b28","title":"Glue: A multi-task benchmark and analysis platform for natural language understanding","year":"2018"},{"authors":"Alex Warstadt; Amanpreet Singh; Samuel R Bowman","journal":"Transactions of the Association for Computational Linguistics","ref_id":"b29","title":"Neural network acceptability judgments","year":"2019"},{"authors":"Adina Williams; Nikita Nangia; Samuel R Bowman","journal":"","ref_id":"b30","title":"A broad-coverage challenge corpus for sentence understanding through inference","year":"2017"},{"authors":"Zhirong Wu; Yuanjun Xiong; Stella X Yu; Dahua Lin","journal":"","ref_id":"b31","title":"Unsupervised feature learning via nonparametric instance discrimination","year":"2018"},{"authors":"Dani Yogatama; Cyprien De Masson D'autume; Jerome Connor; Tomas Kocisky; Mike Chrzanowski; Lingpeng Kong; Angeliki Lazaridou; Wang Ling; Lei Yu; Chris Dyer","journal":"","ref_id":"b32","title":"Learning and evaluating general linguistic intelligence","year":"2019"},{"authors":"Jaehong Yoon; Eunho Yang; Jeongtae Lee; Sung Ju Hwang","journal":"ICLR","ref_id":"b33","title":"Lifelong learning with dynamically expandable networks","year":"2018"},{"authors":"Lu Yu; Bartłomiej Twardowski; Xialei Liu; Luis Herranz; Kai Wang; Yongmei Cheng; Shangling Jui; Joost Van De Weijer","journal":"","ref_id":"b34","title":"Semantic drift compensation for class-incremental learning","year":"2020"},{"authors":"Tianhe Yu; Saurabh Kumar; Abhishek Gupta; Sergey Levine; Karol Hausman; Chelsea Finn","journal":"","ref_id":"b35","title":"Gradient surgery for multi-task learning","year":"2020"},{"authors":"Curran Associates; Inc ","journal":"","ref_id":"b36","title":"","year":""},{"authors":"Dejiao Zhang; Shang-Wen Li; Wei Xiao; Henghui Zhu; Ramesh Nallapati; Andrew O Arnold; Bing Xiang","journal":"Association for Computational Linguistics","ref_id":"b37","title":"Pairwise supervised contrastive learning of sentence representations","year":"2021"},{"authors":"Xiang Zhang; Junbo Zhao; Yann Lecun","journal":"Advances in neural information processing systems","ref_id":"b38","title":"Character-level convolutional networks for text classification","year":"2015"},{"authors":"Kang Zhao; Hua Xu; Jiangong Yang; Kai Gao","journal":"Association for Computational Linguistics","ref_id":"b39","title":"Consistent representation learning for continual relation extraction","year":"2022"}],"string":"[\n {\n \"authors\": \"Rahaf Aljundi; Francesca Babiloni; Mohamed Elhoseiny; Marcus Rohrbach; Tinne Tuytelaars\",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"Memory aware synapses: Learning what (not) to forget\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Jihwan Bang; Heesu Kim; Youngjoon Yoo; Jung-Woo Ha; Jonghyun Choi\",\n \"journal\": \"\",\n \"ref_id\": \"b1\",\n \"title\": \"Rainbow memory: Continual learning with a memory of diverse samples\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Hyuntak Cha; Jaeho Lee; Jinwoo Shin\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"Co2l: Contrastive continual learning\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Ting Chen; Simon Kornblith; Mohammad Norouzi; Geoffrey Hinton\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"a. A simple framework for contrastive learning of visual representations\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Ting Chen; Simon Kornblith; Mohammad Norouzi; Geoffrey Hinton\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"A simple framework for contrastive learning of visual representations\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \" Pmlr\",\n \"journal\": \"\",\n \"ref_id\": \"b5\",\n \"title\": \"\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Matthias De Lange; Rahaf Aljundi; Marc Masana; Sarah Parisot; Xu Jia; Aleš Leonardis; Gregory Slabaugh; Tinne Tuytelaars\",\n \"journal\": \"IEEE Transactions on Pattern Analysis and Machine Intelligence\",\n \"ref_id\": \"b6\",\n \"title\": \"A continual learning survey: Defying forgetting in classification tasks\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Cyprien De Masson D'autume; Sebastian Ruder; Lingpeng Kong; Dani Yogatama\",\n \"journal\": \"Advances in Neural Information Processing Systems\",\n \"ref_id\": \"b7\",\n \"title\": \"Episodic memory in lifelong language learning\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Zhiyuan Fang; Jianfeng Wang; Lijuan Wang; Lei Zhang; Yezhou Yang; Zicheng Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"Seed: Self-supervised distillation for visual representation\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Tianyu Gao; Xingcheng Yao; Danqi Chen\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b9\",\n \"title\": \"SimCSE: Simple contrastive learning of sentence embeddings\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Yufan Huang; Yanzhe Zhang; Jiaao Chen; Xuezhi Wang; Diyi Yang\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"Continual learning for text classification with information disentanglement based regularization\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Nora Kassner; Hinrich Schütze\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b11\",\n \"title\": \"BERT-kNN: Adding a kNN search component to pretrained language models for better QA\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Zixuan Ke; Bing Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"Continual learning of natural language processing tasks: A survey\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Urvashi Khandelwal; Angela Fan; Dan Jurafsky; Luke Zettlemoyer; Mike Lewis\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"Nearest neighbor machine translation\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Prannay Khosla; Piotr Teterwak; Chen Wang; Aaron Sarna; Yonglong Tian; Phillip Isola; Aaron Maschinot; Ce Liu; Dilip Krishnan\",\n \"journal\": \"Advances in Neural Information Processing Systems\",\n \"ref_id\": \"b14\",\n \"title\": \"Supervised contrastive learning\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"P Diederik; Jimmy Kingma; Ba\",\n \"journal\": \"\",\n \"ref_id\": \"b15\",\n \"title\": \"Adam: A method for stochastic optimization\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"James Kirkpatrick; Razvan Pascanu; Neil Rabinowitz; Joel Veness; Guillaume Desjardins; Andrei A Rusu; Kieran Milan; John Quan; Tiago Ramalho; Agnieszka Grabska-Barwinska\",\n \"journal\": \"Proceedings of the national academy of sciences\",\n \"ref_id\": \"b16\",\n \"title\": \"Overcoming catastrophic forgetting in neural networks\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Zhizhong Li; Derek Hoiem\",\n \"journal\": \"IEEE transactions on pattern analysis and machine intelligence\",\n \"ref_id\": \"b17\",\n \"title\": \"Learning without forgetting\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"David Lopez; - Paz; Marc'aurelio Ranzato\",\n \"journal\": \"Advances in neural information processing systems\",\n \"ref_id\": \"b18\",\n \"title\": \"Gradient episodic memory for continual learning\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Yun Luo; Fang Guo; Zihan Liu; Yue Zhang\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"Mere contrastive learning for cross-domain sentiment analysis\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Arvind Neelakantan; Tao Xu; Raul Puri; Alec Radford; Jesse Michael Han; Jerry Tworek; Qiming Yuan; Nikolas Tezak; Jong Wook Kim; Chris Hallacy\",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"Text and code embeddings by contrastive pretraining\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Pranav Rajpurkar; Jian Zhang; Konstantin Lopyrev; Percy Liang\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b21\",\n \"title\": \"SQuAD: 100,000+ questions for machine comprehension of text\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \" Sylvestre-Alvise; Alexander Rebuffi; Georg Kolesnikov; Christoph H Sperl; Lampert\",\n \"journal\": \"\",\n \"ref_id\": \"b22\",\n \"title\": \"icarl: Incremental classifier and representation learning\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Matthew Riemer; Ignacio Cases; Robert Ajemian; Miao Liu; Irina Rish; Yuhai Tu; Gerald Tesauro\",\n \"journal\": \"ICLR\",\n \"ref_id\": \"b23\",\n \"title\": \"Learning to learn without forgetting by maximizing transfer and minimizing interference\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"David Rolnick; Arun Ahuja; Jonathan Schwarz; Timothy Lillicrap; Gregory Wayne\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Experience replay for continual learning\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Curran Associates; Inc \",\n \"journal\": \"\",\n \"ref_id\": \"b25\",\n \"title\": \"\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Joan Serra; Didac Suris; Marius Miron; Alexandros Karatzoglou\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"Overcoming catastrophic forgetting with hard attention to the task\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \" Pmlr\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Alex Wang; Amanpreet Singh; Julian Michael; Felix Hill; Omer Levy; Samuel Bowman\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"Glue: A multi-task benchmark and analysis platform for natural language understanding\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Alex Warstadt; Amanpreet Singh; Samuel R Bowman\",\n \"journal\": \"Transactions of the Association for Computational Linguistics\",\n \"ref_id\": \"b29\",\n \"title\": \"Neural network acceptability judgments\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Adina Williams; Nikita Nangia; Samuel R Bowman\",\n \"journal\": \"\",\n \"ref_id\": \"b30\",\n \"title\": \"A broad-coverage challenge corpus for sentence understanding through inference\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Zhirong Wu; Yuanjun Xiong; Stella X Yu; Dahua Lin\",\n \"journal\": \"\",\n \"ref_id\": \"b31\",\n \"title\": \"Unsupervised feature learning via nonparametric instance discrimination\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Dani Yogatama; Cyprien De Masson D'autume; Jerome Connor; Tomas Kocisky; Mike Chrzanowski; Lingpeng Kong; Angeliki Lazaridou; Wang Ling; Lei Yu; Chris Dyer\",\n \"journal\": \"\",\n \"ref_id\": \"b32\",\n \"title\": \"Learning and evaluating general linguistic intelligence\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Jaehong Yoon; Eunho Yang; Jeongtae Lee; Sung Ju Hwang\",\n \"journal\": \"ICLR\",\n \"ref_id\": \"b33\",\n \"title\": \"Lifelong learning with dynamically expandable networks\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Lu Yu; Bartłomiej Twardowski; Xialei Liu; Luis Herranz; Kai Wang; Yongmei Cheng; Shangling Jui; Joost Van De Weijer\",\n \"journal\": \"\",\n \"ref_id\": \"b34\",\n \"title\": \"Semantic drift compensation for class-incremental learning\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Tianhe Yu; Saurabh Kumar; Abhishek Gupta; Sergey Levine; Karol Hausman; Chelsea Finn\",\n \"journal\": \"\",\n \"ref_id\": \"b35\",\n \"title\": \"Gradient surgery for multi-task learning\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Curran Associates; Inc \",\n \"journal\": \"\",\n \"ref_id\": \"b36\",\n \"title\": \"\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Dejiao Zhang; Shang-Wen Li; Wei Xiao; Henghui Zhu; Ramesh Nallapati; Andrew O Arnold; Bing Xiang\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b37\",\n \"title\": \"Pairwise supervised contrastive learning of sentence representations\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Xiang Zhang; Junbo Zhao; Yann Lecun\",\n \"journal\": \"Advances in neural information processing systems\",\n \"ref_id\": \"b38\",\n \"title\": \"Character-level convolutional networks for text classification\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Kang Zhao; Hua Xu; Jiangong Yang; Kai Gao\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b39\",\n \"title\": \"Consistent representation learning for continual relation extraction\",\n \"year\": \"2022\"\n }\n]"},"formulas":{"kind":"list like","value":[{"formula_coordinates":[3,149.64,680.8,100.08,11.76],"formula_id":"formula_0","formula_text":"C j ∩ C k = ∅, if i = j."},{"formula_coordinates":[3,345.45,300.89,178.96,14.19],"formula_id":"formula_1","formula_text":"h i j = N orm(LM i (x i j )[CLS]),(1)"},{"formula_coordinates":[3,314.2,487.38,210.21,37.81],"formula_id":"formula_2","formula_text":"L cl = j∈A -1 |P (j)| p∈P (j) log exp(h i j • h i p /κ) a∈N (j) exp(h i j • h i a /κ) (2)"},{"formula_coordinates":[3,313.22,676,211.19,24.43],"formula_id":"formula_3","formula_text":"M c = Sample(Kmeans(D i (c)), c, m |C i | ). (3)"},{"formula_coordinates":[3,448.51,720.97,77.81,13.42],"formula_id":"formula_4","formula_text":"M i = ∪ c∈C i M c ."},{"formula_coordinates":[4,107.49,190.07,181.64,30.32],"formula_id":"formula_5","formula_text":"s i j,p = exp(h i j • h i p /τ ) a∈N (j) exp(h i j • h i a /τ ) ,(4)"},{"formula_coordinates":[4,103.07,288.07,186.07,28.55],"formula_id":"formula_6","formula_text":"L IRD = 1 |A| 2 j p s i-1 j,p log s i j,p .(5)"},{"formula_coordinates":[4,140.66,435.49,148.47,10.77],"formula_id":"formula_7","formula_text":"L = L cl + L IRD .(6)"},{"formula_coordinates":[4,306.14,339.7,198.4,33.44],"formula_id":"formula_8","formula_text":"M c = Sample(Kmeans(D i (c)), c, m |C i | ); 21: M i = M i ∪ M c ; 22:"},{"formula_coordinates":[4,312.31,500.91,212.1,35.02],"formula_id":"formula_9","formula_text":"p k (y j ) ∝ (h u k ,y u k )∈K u j 1 y j =y u k • exp( h u j • h u k T ),(7)"},{"formula_coordinates":[5,112.83,504.97,176.3,75.25],"formula_id":"formula_10","formula_text":"ACC = 1 n n i=1 R n i (8) BW T = 1 n -1 n-1 i=1 R n i -R i i ,(9)"}],"string":"[\n {\n \"formula_coordinates\": [\n 3,\n 149.64,\n 680.8,\n 100.08,\n 11.76\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"C j ∩ C k = ∅, if i = j.\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 345.45,\n 300.89,\n 178.96,\n 14.19\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"h i j = N orm(LM i (x i j )[CLS]),(1)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 314.2,\n 487.38,\n 210.21,\n 37.81\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"L cl = j∈A -1 |P (j)| p∈P (j) log exp(h i j • h i p /κ) a∈N (j) exp(h i j • h i a /κ) (2)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 313.22,\n 676,\n 211.19,\n 24.43\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"M c = Sample(Kmeans(D i (c)), c, m |C i | ). (3)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 448.51,\n 720.97,\n 77.81,\n 13.42\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"M i = ∪ c∈C i M c .\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 107.49,\n 190.07,\n 181.64,\n 30.32\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"s i j,p = exp(h i j • h i p /τ ) a∈N (j) exp(h i j • h i a /τ ) ,(4)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 103.07,\n 288.07,\n 186.07,\n 28.55\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"L IRD = 1 |A| 2 j p s i-1 j,p log s i j,p .(5)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 140.66,\n 435.49,\n 148.47,\n 10.77\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"L = L cl + L IRD .(6)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 306.14,\n 339.7,\n 198.4,\n 33.44\n ],\n \"formula_id\": \"formula_8\",\n \"formula_text\": \"M c = Sample(Kmeans(D i (c)), c, m |C i | ); 21: M i = M i ∪ M c ; 22:\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 312.31,\n 500.91,\n 212.1,\n 35.02\n ],\n \"formula_id\": \"formula_9\",\n \"formula_text\": \"p k (y j ) ∝ (h u k ,y u k )∈K u j 1 y j =y u k • exp( h u j • h u k T ),(7)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 112.83,\n 504.97,\n 176.3,\n 75.25\n ],\n \"formula_id\": \"formula_10\",\n \"formula_text\": \"ACC = 1 n n i=1 R n i (8) BW T = 1 n -1 n-1 i=1 R n i -R i i ,(9)\"\n }\n]"}}},{"rowIdx":389,"cells":{"doi":{"kind":"string","value":"10.1093/acprof:oso/9780199547548.003.0003"},"pub_date":{"kind":"string","value":"2023-05-20"},"sections":{"kind":"list like","value":[{"figure_ref":[],"heading":"Introduction","publication_ref":["b12","b7","b8","b12","b39","b6","b5","b16","b14","b27","b29","b15","b16","b4","b42"],"table_ref":[],"text":"Maltese is a Semitic language that has been shaped by an extensive history of contact with non-Semitic languages. A large influx of Sicilian, Italian, and English words over the course of hundreds of years has influenced the Maltese lexicon and grammar, making it a prime case study for those interested in the effects of language contact on morphological systems. Semitic languages are notable for their use of root-and-pattern (a.k.a. templatic) morphology in which inflectional or derivational forms of a lexeme may be related via the non-concatenative interleaving of consonants and vowels. In Maltese, some lexemes of non-Semitic origin have integrated into the native morphology to take both concatenative as well as non-concatenative patterns of Semitic origin. Non-Semitic morphological markers have also entered the grammar and may be found on lexemes of both non-Semitic and Semitic origin.\nThis study applies methods from computational modeling and information theory to investigate factors shaping the organization of the modern Maltese lexicon. Contextualized within frameworks of analogical classification and usage-based accounts of contact-induced language change, we quantify the extent to which the phonology and etymology of Maltese lexemes are predictive of nominal plural inflection in the language. The results indicate that system-level phonology, hypothesized to capture analogical pressures, and etymology, hypothesized to capture conservative pressures that resist analogical change, are predictive of Maltese plural inflection in non-redundant ways, with phonology being more predictive than etymology overall.\nBecause Maltese is a Semitic language, we are also interested in the extent to which these factors are predictive of the type of morphology (either concatenative or non-concatenative) relating singular-plural pairs in the language. Our results show that both phonology and etymology are twice as predictive of a lexeme's plural allomorph(s) as compared to its concatenative type. This suggests that the analogical processes hypothesized to inform speakers' morphological intuitions are most sensitive to phonological similarities across surface forms, regardless of typological differences distinguishing concatenative and non-concatenative relationships. This study provides quantitative evidence for the role of analogical classification based on phonological similarity at the word level as a structuring principle of Maltese nominal plural morphology.\n2 Morphology in Contact: Maltese as a \"Hybrid\" Language?\nMaltese is a descendant of the Siculo Arabic variety spoken by settlers of the Maltese islands beginning in the year 1048 (Fabri, 2010;Brincat, 2011).\nWhile the language is Semitic with respect to its genetic classification, isolation and centuries of foreign colonization led to the development of Maltese as a distinct language shaped by Sicilian, Italian, and English influence. Written records from as early as 1240 acknowledge Maltese as its own language (Brincat, 2017), but it was not until 1934 that Maltese was declared an official language of Malta, along with English and Italian (Fabri, 2010). Italian was revoked as an official language in 1936, but its influence on the Maltese lexicon and grammar remains.\nMuch of the existing literature on Maltese describes the language as having a \"split lexicon\" or a \"hybrid morphology\" (e.g., Spagnol, 2011;Borg and Gatt, 2017). These characterizations reflect an etymological divide in the lexical stock. Semitic nouns in the language mostly form the plural with Semitic affixes or root-and-pattern templates, while non-Semitic nouns show a less strong tendency to form the plural with non-Semitic affixes. At the same time, hundreds of non-Semitic nouns inflect using Semitic patterns and are found in nearly all plural classes (Borg and Azzopardi-Alexander, 1997). Integration in the opposite direction is also found for a smaller number Semitic nouns which inflect using non-Semitic affixes. Maltese thus represents a partial, but not total, example of what has variously been called a \"stratal effect\" (Gardani, 2021) or \"code compartmentalization\" (Friedman, 2013) or \"compartmentalized morphology\" (Matras, 2015), in which native and borrowed morphological exponents in a language are restricted to applying to lexemes of the same etymological origin.\nIt is common in contact linguistics to describe outcomes of language contact as compositions of distinct linguistic systems, even in cases of extensive borrowing or codeswitching (e.g., Myers-Scotton, 1997;Gardani, 2020). Such descriptions are sometimes intended as theoretical analyses. For example, Gardani (2021) treats the stratal effect not simply as an empirically observable pattern, but as a synchronic constraint within the grammar that is psychologically real for speakers: \"... a restriction on the application domain of non-native morphological formatives in a recipient language...\" (Gardani, 2021, 132) that enforces the boundaries of etymologically-defined morphological subsystems. However, we find the a priori assumption that stratal effects reflect distinct and psychologically real morphological subsystems to be problematic inasmuch as it conflates the property to be explained -that language contact can result (to greater or lesser degree) in compartmentalized morphology -with the mechanisms that produce and reinforce that compartmentalization. Stated differently, reification of the stratal effect as a mechanism of the grammar obscures important questions: Given that speakers do not generally know the etymological origins of words, how do they classify words into morphological patterns? What is the relationship between the processes that they use to do this and the stratal effect (or lack thereof) as an empirically observable outcome of language contact?\nIn this study we examine the (partial) stratal effect found in Maltese noun morphology, examining its relationship to factors known to be important outside of contact situations to how speakers classify words into morphological patterns. In particular we analyze the relative strength of a word's phonology and etymology as predictors of its nominal plural morphology and look at the relevance of these factors for the organization of the Maltese lexicon. It is important to note that we are not interested in etymology directly and we do not assume that speakers have or use direct knowledge of the etymology of words. We instead use etymology as a way to estimate the influence of conservative forces on morphological classification. We assume that the predictive power of etymology applies to words which have retained their etymological plurals, in some cases resisting pressures to conform to other parts of the language system. The conservative forces which resist these pressures include token frequency (Krause-Lerche, 2022).\nAdditionally, as a related question, we ask whether there is evidence in Maltese for distinct morphological subsystems (\"hybrid morphology\") in theoretical terms. This question is interesting in part because characterizations of Maltese as having hybrid morphology have also suggested, sometimes explicitly, that the non-concatenative morphology native to Semitic languages should be analyzed as distinct from concatenative morphology, both Semitic and non-Semitic. Moreover, research on morphological integration in Semitic languages has tended to focus specifically on the extent to which foreign words make use of native root-and-template morphology, as compared to affixation (e.g., Bensoukas, 2018;Ziani, 2020). However, since the vast majority of suffixal allomorphs in Maltese are of Semitic origin, division of the lexicon along etymological lines does not correspond to a split according to concatenative vs. non-concatenative morphology, as is sometimes implied. We test whether morphological type is a distinct factor in the stratal effect. Specifically, we ask whether there is support for analyzing root-and-pattern (templatic) plural morphology and affixal plural morphology as distinct subsystems.\nWe compare the results of two models: the first uses a lexeme's phonology and etymology to predict its concatenative type, either affixal or templatic. The second uses the same information to predict its inflectional allomorph, i.e., the specific affix or template found on the lexeme's plural form. Comparisons across factors within each model provide insight into the extent to which phonology and etymology are informative about plural morphology, and thus are likely to have played a role in the development of the language over centuries of contact with speakers of non-Semitic languages. Comparisons across the two models offer insight into the extent to which templatic and affixal morphological patterns operate as distinct subsystems in Maltese."},{"figure_ref":[],"heading":"Analogy and Language Change","publication_ref":["b2","b22","b13","b23","b19","b40","b28","b26"],"table_ref":[],"text":"We take an analogical approach, using the term analogy to refer broadly to any similarity-based, paradigmatic influence of one word on the morphological behavior of another. The importance of analogy as a mechanism of language change is well established in the field of historical linguistics (Anttila, 1977;Hock, 1991;Fertig, 2013;Joseph, 2013), but it is most often discussed with respect to its role in language-internal change, independent of the effects of language contact. In contact linguistics, the idea that (phonologically-based) analogy plays a role in whether and how borrowed words are morphologically integrated into a recipient language has a long history, going back to at least Haugen (1950) and Weinreich (1953). However, most analyses of lexical and morphological borrowing focus on the potential and observed outcomes of contact (see Matras and Adamou, 2020, for an overview), often with little to no discussion of the exact ways in which analogy is hypothesized to play a role.\nTo examine the role of analogy, we take a cue from Matras (2009), who proposes a usage-based model of language contact in which a multilingual individual draws on a unified repertoire of linguistic resources. In this section we elaborate on how such a perspective can help in understanding the role of analogy, specifically analogical classification, in contact-induced morphological change and the development of the Maltese lexicon."},{"figure_ref":[],"heading":"The Paradigm Cell Filling Problem","publication_ref":["b0","b0","b38","b18","b35","b36"],"table_ref":[],"text":"Analysis of the analogical mechanism hypothesized to drive morphological integration in contact may be understood as an extension of the Paradigm Cell Filling Problem (PCFP), a line of research in theoretical morphology that seeks to identify the information available to speakers that allows them to infer and produce grammatically inflected surface forms (Ackerman et al., 2009). Most quantitative analyses of the PCFP to date take an analogical approach: speakers are hypothesized to rely on emergent similarities and paradigmatic relations among previously-acquired words in the lexicon to inform their intuitions when inflecting or processing rare or novel word forms (see, e.g., Ackerman et al., 2009;Sims and Parker, 2016;Guzmán Naranjo, 2020;Parker et al., 2022).\nMatras's ( 2009) usage-based model of language contact is directly compatible with analogical approaches to the PCFP. Since multilingual speakers are assumed to have access to a unified linguistic repertoire corresponding to all of their languages, this full repertoire may be drawn upon to make morphological generalizations. Combinations of generalizations from different languages during speech production may result in linguistic innovations or morphologically adapted \"nonce borrowings\" (Poplack et al., 1988). Over time, some of these may be conventionalized and perpetuated throughout the larger speech community, leading to contact-induced language change.\nWe may therefore specify the PCFP with respect to language contact as follows: what guides speakers' grammatical intuitions when adapting and integrating lexemes in multilingual contexts, and how may conventionalized integration of borrowed linguistic material affect the intuitions of a monolingual speaker when producing inflected word forms?"},{"figure_ref":[],"heading":"Computational Modeling of the PCFP","publication_ref":["b21","b41","b34","b1"],"table_ref":[],"text":"A number of recent studies in computational linguistics have applied machine learning methods to analyze the kinds and amounts of information that may be available to speakers when solving the PCFP (in monolingual contexts). For example, Guzmán Naranjo (2020) uses a Long Shortterm Memory Network (LSTM, Hochreiter and Schmidhuber, 1996) to quantify the respective informativity of stem phonology, lexical semantics, and affixal exponents as predictors of nominal inflection class organization in Russian. His results indicate that while each factor contributes predictive information, more information about inflection class is contributed by stem phonology than by any individual affix. Furthermore, the contributions of the three predictors are additive, indicating a level of nonredundancy in their informativity. Williams et al. (2020) also employ the representational power of an LSTM to quantify the extent to which phonology and lexical semantics are predictive of a noun's declension class in German and Czech. As opposed to model accuracy, they measure the amount of Mutual Information, in bits, shared by phonology, semantics, and declension class systems in each language. They find that, while phonology is more predictive than semantics overall in both languages, the relative informativity of phonology and semantics varies greatly across the two languages and across individual declension classes within each language.\nDawdy-Hesterberg and Pierrehumbert (2014) take an analogical approach to modeling plural formation in Modern Standard Arabic. The authors use a Generalized Context Model (GCM, Nosofsky, 1990) to quantify the extent to which phonological factors, specifically similarities in consonant-vowel (CV) template (a.k.a. \"broken plural\" allomorph), segmental properties (in terms of natural classes), and lexical gang size (Alegre and Gordon, 1999), predict the form of a plural noun in Arabic. Their results indicate that all three factors are predictive to varying degrees, suggesting phonological representations that are both fine-grained, i.e., at the segmental level, and coarse-grained, i.e., with respect to gang size and CV template, may serve as a basis for analogical processing and morphological organization in Arabic.\nFinally, Nieder et al. (2021a,d) use both computational and psycholinguistic methods to investigate the role of analogical classification in the nominal plural system of Maltese. The authors find that plural forms in Maltese may be predicted with a reasonable degree of accuracy based on their phonological similarity to attested plural forms, modulated by the frequency distribution of plural allomorphs in the language. However, the authors do not specifically measure etymology as a predictor, leaving open the question of how non-Semitic words were integrated into the morphological system. In other words, it is unclear from their results whether phonology is predictive independently of etymology, or only as an indicator of etymological origin."},{"figure_ref":[],"heading":"Methods","publication_ref":["b41"],"table_ref":[],"text":"The current study adapts the methods proposed by Williams et al. (2020) to quantify the relative contributions of phonology and etymology as predictors of inflectional organization in Maltese. We use a character-level LSTM classifier trained to make inferences about a word's plural class by abstracting over the phonology of each word form as a whole. We then quantify the influence of phonology on Maltese nominal plural inflection using Mutual Information, an information theoretic measure of interpredictability among two or more systems. We compare our results to the predictive strength of the word's etymological origin using the same measures, quantifying the balance of analogical and conservative factors hypothesized to shape the integration of foreign lexemes into the grammar."},{"figure_ref":[],"heading":"Data","publication_ref":["b17","b3","b41"],"table_ref":[],"text":"This study merges data from two collections compiled by Nieder et al. (2021b,c) into a single dataset consisting of 3,174 singular-plural noun pairs. Each pair is tagged for etymological origin, either Semitic or non-Semitic. The original data was manually compiled from the MLRS Korpus Malti v. 2.0 and 3.0 (Gatt and Čéplö, 2013) and supplemented with Schembri's (2012) collection of Maltese CV templates. Etymological information was sourced from a digitized version of Aquilina's (2006) Maltese-English dictionary. Plural nouns in the data are classified as taking one of 12 different suffixes (\"sound plurals\") or 11 different non-concatenative CV templates (\"broken plurals\"), forming a nominal plural inflection system composed of 23 different inflection classes (Nieder et al., 2021b). Maltese is the only standardized Semitic language written in a Latin script, using an orthography that \"represents the phonology of the language admirably\" according to Hoberman (2007, 258). For this reason, we analyze nouns using their original orthography, as in Williams et al. (2020). Over 135 nouns in the dataset take more than one plural form. Of these, 78 nouns may take both broken and sound plurals. In this study, we account for these nouns by representing each pair separately at the allomorph level, whereas in the binary prediction model of the lexeme's concatenative type (concatenative vs. non-concatenative) we include a noun only once per type. For example, the word LIBSA 'dress' may take the sound plural libsiet and the broken plurals lbies and lbiesi. The lexeme LIBSA is therefore included in the model three times in the allomorph prediction setting, but only twice in the type prediction setting."},{"figure_ref":[],"heading":"Non-Semitic","publication_ref":["b41"],"table_ref":["tab_0"],"text":"Following Williams et al. (2020), we remove all classes with fewer than 20 lexemes, leaving a total of 13 plural allomorph classes in our model. Table 1 shows the full distribution of allomorphs according to etymology and concatenative type. Note that lexemes that take more than one allomorph are counted more than once."},{"figure_ref":[],"heading":"Formal Notation","publication_ref":["b41"],"table_ref":[],"text":"Following Williams et al. (2020), we can define a lexeme as a tuple (w i , e i , c i ) where for the i th lexeme, w i = the lexeme's phonological form, e i = the lexeme's etymological origin, and c i = the lexeme's inflection class. We assume the lexemes follow a probability distribution p(w, e, c), approximated by the corpus. We can define the space of K inflection classes as C = {1, ..., K}, so that c i ∈ C and define C as the random variable associated with C. For a set of lexemes derived from N etymological origins, we can define an etymological space as E = {1, ..., N } so that e i ∈ E and define E as the random variable associated with E. Each noun may be associated with one of two genders g i from the space of genders G specific to Maltese. Finally, we define the space of word forms as the Kleene closure over a language's alphabet Σ, so that w i ∈ Σ*, with W as the random variable associated with Σ*."},{"figure_ref":["fig_0"],"heading":"Mutual Information (MI)","publication_ref":["b10","b41"],"table_ref":[],"text":"Mutual Information (MI) is an information theoretic measure that quantifies the degree of interpredictability among two or more systems. For example, the MI shared by the nominal plural inflection class system C and phonological system W in Maltese may be calculated as follows:\nMI(C; W ) = H(C) -H(C|W )(1)\nThis may be generalized to consider the amount of redundant information shared by inflection class, phonology, and etymology E as follows:\nMI(C; E; W ) = MI(C; W ) -MI(C; W |E) (2)\nBecause a language's grammatical gender system is known to interact with its inflectional morphology in non-deterministic ways (Corbett and Fraser, 2000), we follow Williams et al. (2020) and condition all relevant measures on gender:\nMI(C; W |G) = H(C|G) -H(C|W, G) (3)\nThe intuitive reasoning behind Equations 1 -3 may be seen in Figure 1, in which each colored circle represents H|G, the total entropy, conditioned on gender, of the three interacting systems under analysis.\nFinally, since our corpus is only a sample of the language, we note that all calculations are estimates. However, while estimates over the finite inflection class and etymology systems can be empirically calculated using the corpus, the infinite number of possible word forms in the Σ* means calculations involving W must be further approximated. Methods for estimating the entropy of both kinds of systems are described in detail in the following sections."},{"figure_ref":[],"heading":"Techniques for Estimating Entropy","publication_ref":[],"table_ref":[],"text":"We use plug-in estimation to obtain entropy values for C and E, calculating the distribution p(c) for c ∈ C (or alternatively, p(e) for e ∈ E) and using this to estimate H(C) in Equation 1 above."},{"figure_ref":[],"heading":"Approximating Conditional Entropy","publication_ref":["b9"],"table_ref":[],"text":"H(C|E) may be similarly calculated using plugin estimation. However, given the infinite number of possible word forms in Σ*, an estimate for H(C|W ) cannot be calculated directly from the corpus. We therefore approximate this value using cross-entropy, which has been mathematically proven to be an upper bound on conditional entropy (Brown et al., 1992). We use the cross-entropy loss obtained from a computational model that has been trained to predict the plural class c i associated with a singular noun w i to approximate the cross-entropy of the system:\nH(C|W ) ≤ - 1 M M i=1 log q(c i |w i ) (4)\nWe note that as the amount of data in the corpus increases, i.e., as M → ∞, the above value approaches the true cross-entropy value."},{"figure_ref":[],"heading":"Normalized Mutual Information (NMI)","publication_ref":[],"table_ref":[],"text":"To compare results across models and across languages, we normalize MI values by dividing by the total entropy of the inflection class system. For example, the NMI shared by a Maltese noun's phonology and plural inflection may be calculated as:\nNMI(C; W ) = MI(C; W ) H(C) (5)"},{"figure_ref":[],"heading":"Model Details","publication_ref":["b41","b41"],"table_ref":[],"text":"We adapt the LSTM classifier implemented in Williams et al. (2020) to estimate the probability that a plural class c is associated with a given input noun w of gender g, i.e., q(c|w, g) in Equation 4. embedded and input into the model's initial hidden state. The model is trained using Adam (Kingma and Ba, 2015) with model hyperparameters, including the number of training epochs and the number and sizes of hidden layers, optimized using the Bayesian optimization technique implemented in Williams et al. (2020). The model then learns a probability distribution that serves to approximate q(c|w, g).\nFollowing training, we test the model on a heldout dataset and use the model's cross-entropy loss to serve as an approximate upper bound on the conditional entropy H(C|W, G). We use 10-fold cross validation to make full use of the dataset for our approximations. To estimate q(c|w, e, g), we concatenate a binary character representing the word's etymology onto the end of the noun to serve as model input and follow the same procedure."},{"figure_ref":[],"heading":"Results","publication_ref":[],"table_ref":["tab_1"],"text":"NMI and H(C|G) values for C defined as concatenative type and plural allomorph, respectively, are presented in Table 2. The largest NMI value we obtain, NMI(E; W |G), indicates that more than half of the information needed to predict a word's etymology is shared with its phonology. In other words, it is often not difficult to guess the origin of a Maltese word based on how it sounds. Note that this value is consistent across models, as it does not depend on C."},{"figure_ref":[],"heading":"Concatenative Type","publication_ref":[],"table_ref":["tab_1"],"text":"Results for the model predicting a noun's concatenative type are in Table 2. Note first that the entropy H(C|G) of the plural inflection class system defined at the level of concatenative type is calculated to be 0.81, indicating that, given its gender, predicting whether a random Maltese noun takes concate-native or non-concatenative morphology is more predictable than chance, although not by much. We find phonology, indicated by NMI(C; W |G), to be more predictive than etymology, indicated by NMI(C; E|G). Crucially, each of these bipartite NMI values exceeds the tripartite mutual information NMI(C; E; W |G) shared across all three systems. This indicates that while a non-trivial amount of predictive information is shared across all three systems, phonology and etymology are each predictive of concatenative type in partially non-redundant ways. This suggests that both analogical and conservative forces are likely to have played a role in the development of the Maltese nominal plural system."},{"figure_ref":[],"heading":"Plural Allomorph","publication_ref":[],"table_ref":["tab_1"],"text":"In an analogical model of inflection in which singular inflected forms and their plural counterparts share a direct relationship in the lexicon, the predictive principles structuring the morphological system are expected to be most evident when defining an inflection class system at the level of the allomorph.\nWe first note that the entropy H(C|G) calculated over the plural class distribution defined at the allomorph level is nearly three times as high as the entropy of C when defined as a noun's concatenative type. This is reflective of the higher degree of unpredictability associated with a non-uniform distribution of nouns over a greater number of inflection classes. When comparing across the allomorph and concatenative type models it is thus important to normalize for the fact that predicting allomorphs is more difficult than predicting concatenative type. However, even calculations normalized in this way show that the interpredictability among phonology, etymology, and plural inflection, indicated by the NMI values in Table 2, are all twice as high at the allomorph level as they are for concatenative type. In other words, a noun's singular form reduces the relative uncertainty about its plural allomorph twice as much as it reduces the uncertainty about whether that allomorph is concatenative. This suggests the analogical and conservative pressures hypothesized to shape morphological organization are more sensitive to correspondences at the word level than to typological similarities with respect to concatenativity.\nAdditionally, the general tendency found at the level of concatenative type still follows when classes are defined at the level of individual allomorphs: phonology shares more information with inflection class than does etymology, with each factor contributing some amount of non-redundant information. This illustrates one key advantage of the methods employed in this study, namely the ability to disentangle the independent contributions of either predictor from the degree to which both exert redundant organizational pressure towards the same end.\nFor example, given the fact that phonology and etymology are themselves mutually informative, we cannot uniquely interpret either bipartite measure of MI, that is, NMI(C; W |G) or NMI(C; E|G), as indicative of the forces hypothesized to shape the integration of linguistic material in contact. Rather, evidence for analogical structuring of the Maltese plural system at the allomorph level is specifically indicated by the positive difference between NMI(C; W |G) and NMI(C; E; W |G). Conservative pressures, such as those associated with high token-frequency items (Krause-Lerche, 2022), are similarly indicated by the extent to which NMI(C; E|G) exceeds NMI(C; E; W |G)."},{"figure_ref":["fig_1","fig_1","fig_2"],"heading":"Variation Across Allomorph Classes","publication_ref":["b41"],"table_ref":["tab_2"],"text":"Closer examination of the model's predictions reveals an effect of type frequency, with larger inflection classes predicted more often than smaller classes. Table 3 reports the accuracy of all models in which singular noun phonology W is a predictor. Since all models achieve an overall accuracy above a majority baseline, the NMI values we obtain may be reliably interpreted as empirical minimums. However, as can be seen in Figure 2, the model's incorrect predictions do not clearly distinguish between sound and broken classes; nouns with a sound plural allomorph may be misclassified as taking a broken plural template, and nouns taking a broken plural may be incorrectly predicted to take a sound plural.\nIf speakers are sensitive to differences between concatenative and non-concatenative allomorphs grouped into high-level macro classes (morphological subsystems), we might expect some degree of observable within-class coherence with respect to either or both of the phonology and etymology of words exhibiting a particular morphological behavior. Specifically, we would expect a pattern of predictions in which the LSTM is able to first Instead, as seen in Figure 2, we do not find such evidence. Rather, we find evidence for coherence at the allomorph level, specifically, for phonological patterns as a predictor of inflectional organization and driver of inflectional behavior at the allomorph level.\nFinally, as in Williams et al. (2020), we also conduct an analysis of the partial Pointwise Mutual Information (PMI) shared between phonology W and class C with respect to the surprisal H(C = c|G) for each class, defined at the allomorph level. Figure 3 shows this distribution, with allomorph classes presented in order of increasing type frequency (and thus decreasing surprisal). We note that Maltese noun classes are each only par- "},{"figure_ref":[],"heading":"Discussion","publication_ref":[],"table_ref":[],"text":"In this paper we used an LSTM to help estimate the kinds and amounts of information that may be available to speakers when \"solving\" the PCFP. Overall, our results provide quantitative evidence for the role of both word phonology and etymology (as a stand-in for conservative factors) in shaping the Maltese lexicon.\nSpecifically, we found that the extent to which a Maltese singular noun's phonology predicts its plural morphology exceeds that of etymology in nonredundant ways. This suggests that analogical pressures from phonological correspondences across the lexicon shape nominal plural inflection in Maltese, independently of the etymological source language for some word or morphological pattern.\nOur results also show an independent contribution of etymology as a predictor. We hypothesize that this captures conservative pressures theorized to resist analogical change, including token frequency (Krause-Lerche, 2022). It may also reflect associative correlations from the use of lexemes of a common etymology in similar contexts, strengthening their coherence as a subsystem in the multilingual repertoire and encouraging the maintenance of a noun's original morphology. Further work is needed to investigate these possibilities.\nIn language contact situations such as that of Maltese, it is likely that an influx of foreign lexemes and increased productivity of foreign affixes affect both the size and character (e.g., phonology) of nominal plural classes relative to each other over time. This in turn is likely to affect subsequent classification and integration of words into the inflectional morphology of the language.\nIn general, our results do not support characterizations of Maltese in which concatenative and non-concatenative morphologies co-exist as discrete systems within the lexicon. While a singular noun's phonology and etymology are each somewhat predictive of its concatenative type, they are twice as predictive of the actual plural allomorph(s) with which the lexeme is associated. This suggests that systematic relationships at the word level organize the morphology of Maltese, in turn shaping the language as new words are integrated and inflected."},{"figure_ref":[],"heading":"Conclusion","publication_ref":[],"table_ref":[],"text":"This study extends previous work in information theory, computational modeling, and theoretical morphology to provide quantitative evidence for the role of phonology as an analogical force in the morphological organization of Maltese. We ground this in a usage-based account of multilingualism and contact-induced change in which speakers are hypothesized to make use of analogical reasoning, among other language-general cognitive functions, when integrating novel words and patterns within a unified linguistic repertoire. The same processes that guide synchronic language use are proposed to be responsible for the diachronic effects of contactinduced language change. Specifically, it is hypothesized that speakers draw on similarities across multiple dimensions -including but not limited to phonological patterns, semantic and indexical meaning, pragmatic function, and contexts of useto collaboratively construct and adapt grammatical systems of linguistic communication over time.\nIn the case of Maltese, our findings indicate that while a lexeme's phonology and etymology are themselves highly interpredictable, each contributes non-redundant information to reduce uncertainty when predicting the lexeme's plural inflection. While the etymology of a noun is somewhat predictive of its plural inflection, the word's phonology plays a much greater role. This synchronic analysis has diachronic implications. Our results suggest that analogical pressures from phonological similarities across the lexicon may have guided speakers' inflectional behavior when code mixing over the course of the development of the language to result in the conventionalized forms observed in modern Maltese. However, further diachronic study is needed to confirm this interpretation.\nContrary to a hypothesis in which concatenative and non-concatenative systems operate as separate subsystems within a \"split\" or \"hybrid\" morphology, our results indicate correspondences at the level of individual wordforms and affixes are driving speakers' morphological behavior. Specifically, the phonology and etymology of a lexeme are twice as predictive of its plural allomorph than its concatenative type. Further investigation into Maltese nouns attested to take plural forms of both concatenative types may provide additional insight into the ways in which concatenative type affects speakers' behavior, if at all. Future work should also consider additional factors known to shape inflection class systems, for example by integrating semantic word vectors into the model. Finally, additional comparisons implementing these methods across corpora in a variety of languages will continue to shed light on the factors shaping morphological systems cross-linguistically. "},{"figure_ref":[],"heading":"A Nominal Plural Allomorphs in Maltese","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Acknowledgments","publication_ref":[],"table_ref":[],"text":"We thank Jessica Nieder and Adam Ussishkin for generously sharing an abundance of digital resources and helpful feedback, Sarah Caruana for her insight into the Maltese language, and Christian Clark and Andrew Duffy for their contributions to an initial version of this project.\nThis material is based on work supported by the National Science Foundation under grant BCS-2217554 (Neural discovery of abstract inflectional structure, PI Micha Elsner, Co-PI Andrea Sims)."}],"string":"[\n {\n \"figure_ref\": [],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b12\",\n \"b7\",\n \"b8\",\n \"b12\",\n \"b39\",\n \"b6\",\n \"b5\",\n \"b16\",\n \"b14\",\n \"b27\",\n \"b29\",\n \"b15\",\n \"b16\",\n \"b4\",\n \"b42\"\n ],\n \"table_ref\": [],\n \"text\": \"Maltese is a Semitic language that has been shaped by an extensive history of contact with non-Semitic languages. A large influx of Sicilian, Italian, and English words over the course of hundreds of years has influenced the Maltese lexicon and grammar, making it a prime case study for those interested in the effects of language contact on morphological systems. Semitic languages are notable for their use of root-and-pattern (a.k.a. templatic) morphology in which inflectional or derivational forms of a lexeme may be related via the non-concatenative interleaving of consonants and vowels. In Maltese, some lexemes of non-Semitic origin have integrated into the native morphology to take both concatenative as well as non-concatenative patterns of Semitic origin. Non-Semitic morphological markers have also entered the grammar and may be found on lexemes of both non-Semitic and Semitic origin.\\nThis study applies methods from computational modeling and information theory to investigate factors shaping the organization of the modern Maltese lexicon. Contextualized within frameworks of analogical classification and usage-based accounts of contact-induced language change, we quantify the extent to which the phonology and etymology of Maltese lexemes are predictive of nominal plural inflection in the language. The results indicate that system-level phonology, hypothesized to capture analogical pressures, and etymology, hypothesized to capture conservative pressures that resist analogical change, are predictive of Maltese plural inflection in non-redundant ways, with phonology being more predictive than etymology overall.\\nBecause Maltese is a Semitic language, we are also interested in the extent to which these factors are predictive of the type of morphology (either concatenative or non-concatenative) relating singular-plural pairs in the language. Our results show that both phonology and etymology are twice as predictive of a lexeme's plural allomorph(s) as compared to its concatenative type. This suggests that the analogical processes hypothesized to inform speakers' morphological intuitions are most sensitive to phonological similarities across surface forms, regardless of typological differences distinguishing concatenative and non-concatenative relationships. This study provides quantitative evidence for the role of analogical classification based on phonological similarity at the word level as a structuring principle of Maltese nominal plural morphology.\\n2 Morphology in Contact: Maltese as a \\\"Hybrid\\\" Language?\\nMaltese is a descendant of the Siculo Arabic variety spoken by settlers of the Maltese islands beginning in the year 1048 (Fabri, 2010;Brincat, 2011).\\nWhile the language is Semitic with respect to its genetic classification, isolation and centuries of foreign colonization led to the development of Maltese as a distinct language shaped by Sicilian, Italian, and English influence. Written records from as early as 1240 acknowledge Maltese as its own language (Brincat, 2017), but it was not until 1934 that Maltese was declared an official language of Malta, along with English and Italian (Fabri, 2010). Italian was revoked as an official language in 1936, but its influence on the Maltese lexicon and grammar remains.\\nMuch of the existing literature on Maltese describes the language as having a \\\"split lexicon\\\" or a \\\"hybrid morphology\\\" (e.g., Spagnol, 2011;Borg and Gatt, 2017). These characterizations reflect an etymological divide in the lexical stock. Semitic nouns in the language mostly form the plural with Semitic affixes or root-and-pattern templates, while non-Semitic nouns show a less strong tendency to form the plural with non-Semitic affixes. At the same time, hundreds of non-Semitic nouns inflect using Semitic patterns and are found in nearly all plural classes (Borg and Azzopardi-Alexander, 1997). Integration in the opposite direction is also found for a smaller number Semitic nouns which inflect using non-Semitic affixes. Maltese thus represents a partial, but not total, example of what has variously been called a \\\"stratal effect\\\" (Gardani, 2021) or \\\"code compartmentalization\\\" (Friedman, 2013) or \\\"compartmentalized morphology\\\" (Matras, 2015), in which native and borrowed morphological exponents in a language are restricted to applying to lexemes of the same etymological origin.\\nIt is common in contact linguistics to describe outcomes of language contact as compositions of distinct linguistic systems, even in cases of extensive borrowing or codeswitching (e.g., Myers-Scotton, 1997;Gardani, 2020). Such descriptions are sometimes intended as theoretical analyses. For example, Gardani (2021) treats the stratal effect not simply as an empirically observable pattern, but as a synchronic constraint within the grammar that is psychologically real for speakers: \\\"... a restriction on the application domain of non-native morphological formatives in a recipient language...\\\" (Gardani, 2021, 132) that enforces the boundaries of etymologically-defined morphological subsystems. However, we find the a priori assumption that stratal effects reflect distinct and psychologically real morphological subsystems to be problematic inasmuch as it conflates the property to be explained -that language contact can result (to greater or lesser degree) in compartmentalized morphology -with the mechanisms that produce and reinforce that compartmentalization. Stated differently, reification of the stratal effect as a mechanism of the grammar obscures important questions: Given that speakers do not generally know the etymological origins of words, how do they classify words into morphological patterns? What is the relationship between the processes that they use to do this and the stratal effect (or lack thereof) as an empirically observable outcome of language contact?\\nIn this study we examine the (partial) stratal effect found in Maltese noun morphology, examining its relationship to factors known to be important outside of contact situations to how speakers classify words into morphological patterns. In particular we analyze the relative strength of a word's phonology and etymology as predictors of its nominal plural morphology and look at the relevance of these factors for the organization of the Maltese lexicon. It is important to note that we are not interested in etymology directly and we do not assume that speakers have or use direct knowledge of the etymology of words. We instead use etymology as a way to estimate the influence of conservative forces on morphological classification. We assume that the predictive power of etymology applies to words which have retained their etymological plurals, in some cases resisting pressures to conform to other parts of the language system. The conservative forces which resist these pressures include token frequency (Krause-Lerche, 2022).\\nAdditionally, as a related question, we ask whether there is evidence in Maltese for distinct morphological subsystems (\\\"hybrid morphology\\\") in theoretical terms. This question is interesting in part because characterizations of Maltese as having hybrid morphology have also suggested, sometimes explicitly, that the non-concatenative morphology native to Semitic languages should be analyzed as distinct from concatenative morphology, both Semitic and non-Semitic. Moreover, research on morphological integration in Semitic languages has tended to focus specifically on the extent to which foreign words make use of native root-and-template morphology, as compared to affixation (e.g., Bensoukas, 2018;Ziani, 2020). However, since the vast majority of suffixal allomorphs in Maltese are of Semitic origin, division of the lexicon along etymological lines does not correspond to a split according to concatenative vs. non-concatenative morphology, as is sometimes implied. We test whether morphological type is a distinct factor in the stratal effect. Specifically, we ask whether there is support for analyzing root-and-pattern (templatic) plural morphology and affixal plural morphology as distinct subsystems.\\nWe compare the results of two models: the first uses a lexeme's phonology and etymology to predict its concatenative type, either affixal or templatic. The second uses the same information to predict its inflectional allomorph, i.e., the specific affix or template found on the lexeme's plural form. Comparisons across factors within each model provide insight into the extent to which phonology and etymology are informative about plural morphology, and thus are likely to have played a role in the development of the language over centuries of contact with speakers of non-Semitic languages. Comparisons across the two models offer insight into the extent to which templatic and affixal morphological patterns operate as distinct subsystems in Maltese.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Analogy and Language Change\",\n \"publication_ref\": [\n \"b2\",\n \"b22\",\n \"b13\",\n \"b23\",\n \"b19\",\n \"b40\",\n \"b28\",\n \"b26\"\n ],\n \"table_ref\": [],\n \"text\": \"We take an analogical approach, using the term analogy to refer broadly to any similarity-based, paradigmatic influence of one word on the morphological behavior of another. The importance of analogy as a mechanism of language change is well established in the field of historical linguistics (Anttila, 1977;Hock, 1991;Fertig, 2013;Joseph, 2013), but it is most often discussed with respect to its role in language-internal change, independent of the effects of language contact. In contact linguistics, the idea that (phonologically-based) analogy plays a role in whether and how borrowed words are morphologically integrated into a recipient language has a long history, going back to at least Haugen (1950) and Weinreich (1953). However, most analyses of lexical and morphological borrowing focus on the potential and observed outcomes of contact (see Matras and Adamou, 2020, for an overview), often with little to no discussion of the exact ways in which analogy is hypothesized to play a role.\\nTo examine the role of analogy, we take a cue from Matras (2009), who proposes a usage-based model of language contact in which a multilingual individual draws on a unified repertoire of linguistic resources. In this section we elaborate on how such a perspective can help in understanding the role of analogy, specifically analogical classification, in contact-induced morphological change and the development of the Maltese lexicon.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"The Paradigm Cell Filling Problem\",\n \"publication_ref\": [\n \"b0\",\n \"b0\",\n \"b38\",\n \"b18\",\n \"b35\",\n \"b36\"\n ],\n \"table_ref\": [],\n \"text\": \"Analysis of the analogical mechanism hypothesized to drive morphological integration in contact may be understood as an extension of the Paradigm Cell Filling Problem (PCFP), a line of research in theoretical morphology that seeks to identify the information available to speakers that allows them to infer and produce grammatically inflected surface forms (Ackerman et al., 2009). Most quantitative analyses of the PCFP to date take an analogical approach: speakers are hypothesized to rely on emergent similarities and paradigmatic relations among previously-acquired words in the lexicon to inform their intuitions when inflecting or processing rare or novel word forms (see, e.g., Ackerman et al., 2009;Sims and Parker, 2016;Guzmán Naranjo, 2020;Parker et al., 2022).\\nMatras's ( 2009) usage-based model of language contact is directly compatible with analogical approaches to the PCFP. Since multilingual speakers are assumed to have access to a unified linguistic repertoire corresponding to all of their languages, this full repertoire may be drawn upon to make morphological generalizations. Combinations of generalizations from different languages during speech production may result in linguistic innovations or morphologically adapted \\\"nonce borrowings\\\" (Poplack et al., 1988). Over time, some of these may be conventionalized and perpetuated throughout the larger speech community, leading to contact-induced language change.\\nWe may therefore specify the PCFP with respect to language contact as follows: what guides speakers' grammatical intuitions when adapting and integrating lexemes in multilingual contexts, and how may conventionalized integration of borrowed linguistic material affect the intuitions of a monolingual speaker when producing inflected word forms?\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Computational Modeling of the PCFP\",\n \"publication_ref\": [\n \"b21\",\n \"b41\",\n \"b34\",\n \"b1\"\n ],\n \"table_ref\": [],\n \"text\": \"A number of recent studies in computational linguistics have applied machine learning methods to analyze the kinds and amounts of information that may be available to speakers when solving the PCFP (in monolingual contexts). For example, Guzmán Naranjo (2020) uses a Long Shortterm Memory Network (LSTM, Hochreiter and Schmidhuber, 1996) to quantify the respective informativity of stem phonology, lexical semantics, and affixal exponents as predictors of nominal inflection class organization in Russian. His results indicate that while each factor contributes predictive information, more information about inflection class is contributed by stem phonology than by any individual affix. Furthermore, the contributions of the three predictors are additive, indicating a level of nonredundancy in their informativity. Williams et al. (2020) also employ the representational power of an LSTM to quantify the extent to which phonology and lexical semantics are predictive of a noun's declension class in German and Czech. As opposed to model accuracy, they measure the amount of Mutual Information, in bits, shared by phonology, semantics, and declension class systems in each language. They find that, while phonology is more predictive than semantics overall in both languages, the relative informativity of phonology and semantics varies greatly across the two languages and across individual declension classes within each language.\\nDawdy-Hesterberg and Pierrehumbert (2014) take an analogical approach to modeling plural formation in Modern Standard Arabic. The authors use a Generalized Context Model (GCM, Nosofsky, 1990) to quantify the extent to which phonological factors, specifically similarities in consonant-vowel (CV) template (a.k.a. \\\"broken plural\\\" allomorph), segmental properties (in terms of natural classes), and lexical gang size (Alegre and Gordon, 1999), predict the form of a plural noun in Arabic. Their results indicate that all three factors are predictive to varying degrees, suggesting phonological representations that are both fine-grained, i.e., at the segmental level, and coarse-grained, i.e., with respect to gang size and CV template, may serve as a basis for analogical processing and morphological organization in Arabic.\\nFinally, Nieder et al. (2021a,d) use both computational and psycholinguistic methods to investigate the role of analogical classification in the nominal plural system of Maltese. The authors find that plural forms in Maltese may be predicted with a reasonable degree of accuracy based on their phonological similarity to attested plural forms, modulated by the frequency distribution of plural allomorphs in the language. However, the authors do not specifically measure etymology as a predictor, leaving open the question of how non-Semitic words were integrated into the morphological system. In other words, it is unclear from their results whether phonology is predictive independently of etymology, or only as an indicator of etymological origin.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Methods\",\n \"publication_ref\": [\n \"b41\"\n ],\n \"table_ref\": [],\n \"text\": \"The current study adapts the methods proposed by Williams et al. (2020) to quantify the relative contributions of phonology and etymology as predictors of inflectional organization in Maltese. We use a character-level LSTM classifier trained to make inferences about a word's plural class by abstracting over the phonology of each word form as a whole. We then quantify the influence of phonology on Maltese nominal plural inflection using Mutual Information, an information theoretic measure of interpredictability among two or more systems. We compare our results to the predictive strength of the word's etymological origin using the same measures, quantifying the balance of analogical and conservative factors hypothesized to shape the integration of foreign lexemes into the grammar.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Data\",\n \"publication_ref\": [\n \"b17\",\n \"b3\",\n \"b41\"\n ],\n \"table_ref\": [],\n \"text\": \"This study merges data from two collections compiled by Nieder et al. (2021b,c) into a single dataset consisting of 3,174 singular-plural noun pairs. Each pair is tagged for etymological origin, either Semitic or non-Semitic. The original data was manually compiled from the MLRS Korpus Malti v. 2.0 and 3.0 (Gatt and Čéplö, 2013) and supplemented with Schembri's (2012) collection of Maltese CV templates. Etymological information was sourced from a digitized version of Aquilina's (2006) Maltese-English dictionary. Plural nouns in the data are classified as taking one of 12 different suffixes (\\\"sound plurals\\\") or 11 different non-concatenative CV templates (\\\"broken plurals\\\"), forming a nominal plural inflection system composed of 23 different inflection classes (Nieder et al., 2021b). Maltese is the only standardized Semitic language written in a Latin script, using an orthography that \\\"represents the phonology of the language admirably\\\" according to Hoberman (2007, 258). For this reason, we analyze nouns using their original orthography, as in Williams et al. (2020). Over 135 nouns in the dataset take more than one plural form. Of these, 78 nouns may take both broken and sound plurals. In this study, we account for these nouns by representing each pair separately at the allomorph level, whereas in the binary prediction model of the lexeme's concatenative type (concatenative vs. non-concatenative) we include a noun only once per type. For example, the word LIBSA 'dress' may take the sound plural libsiet and the broken plurals lbies and lbiesi. The lexeme LIBSA is therefore included in the model three times in the allomorph prediction setting, but only twice in the type prediction setting.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Non-Semitic\",\n \"publication_ref\": [\n \"b41\"\n ],\n \"table_ref\": [\n \"tab_0\"\n ],\n \"text\": \"Following Williams et al. (2020), we remove all classes with fewer than 20 lexemes, leaving a total of 13 plural allomorph classes in our model. Table 1 shows the full distribution of allomorphs according to etymology and concatenative type. Note that lexemes that take more than one allomorph are counted more than once.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Formal Notation\",\n \"publication_ref\": [\n \"b41\"\n ],\n \"table_ref\": [],\n \"text\": \"Following Williams et al. (2020), we can define a lexeme as a tuple (w i , e i , c i ) where for the i th lexeme, w i = the lexeme's phonological form, e i = the lexeme's etymological origin, and c i = the lexeme's inflection class. We assume the lexemes follow a probability distribution p(w, e, c), approximated by the corpus. We can define the space of K inflection classes as C = {1, ..., K}, so that c i ∈ C and define C as the random variable associated with C. For a set of lexemes derived from N etymological origins, we can define an etymological space as E = {1, ..., N } so that e i ∈ E and define E as the random variable associated with E. Each noun may be associated with one of two genders g i from the space of genders G specific to Maltese. Finally, we define the space of word forms as the Kleene closure over a language's alphabet Σ, so that w i ∈ Σ*, with W as the random variable associated with Σ*.\"\n },\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"Mutual Information (MI)\",\n \"publication_ref\": [\n \"b10\",\n \"b41\"\n ],\n \"table_ref\": [],\n \"text\": \"Mutual Information (MI) is an information theoretic measure that quantifies the degree of interpredictability among two or more systems. For example, the MI shared by the nominal plural inflection class system C and phonological system W in Maltese may be calculated as follows:\\nMI(C; W ) = H(C) -H(C|W )(1)\\nThis may be generalized to consider the amount of redundant information shared by inflection class, phonology, and etymology E as follows:\\nMI(C; E; W ) = MI(C; W ) -MI(C; W |E) (2)\\nBecause a language's grammatical gender system is known to interact with its inflectional morphology in non-deterministic ways (Corbett and Fraser, 2000), we follow Williams et al. (2020) and condition all relevant measures on gender:\\nMI(C; W |G) = H(C|G) -H(C|W, G) (3)\\nThe intuitive reasoning behind Equations 1 -3 may be seen in Figure 1, in which each colored circle represents H|G, the total entropy, conditioned on gender, of the three interacting systems under analysis.\\nFinally, since our corpus is only a sample of the language, we note that all calculations are estimates. However, while estimates over the finite inflection class and etymology systems can be empirically calculated using the corpus, the infinite number of possible word forms in the Σ* means calculations involving W must be further approximated. Methods for estimating the entropy of both kinds of systems are described in detail in the following sections.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Techniques for Estimating Entropy\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We use plug-in estimation to obtain entropy values for C and E, calculating the distribution p(c) for c ∈ C (or alternatively, p(e) for e ∈ E) and using this to estimate H(C) in Equation 1 above.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Approximating Conditional Entropy\",\n \"publication_ref\": [\n \"b9\"\n ],\n \"table_ref\": [],\n \"text\": \"H(C|E) may be similarly calculated using plugin estimation. However, given the infinite number of possible word forms in Σ*, an estimate for H(C|W ) cannot be calculated directly from the corpus. We therefore approximate this value using cross-entropy, which has been mathematically proven to be an upper bound on conditional entropy (Brown et al., 1992). We use the cross-entropy loss obtained from a computational model that has been trained to predict the plural class c i associated with a singular noun w i to approximate the cross-entropy of the system:\\nH(C|W ) ≤ - 1 M M i=1 log q(c i |w i ) (4)\\nWe note that as the amount of data in the corpus increases, i.e., as M → ∞, the above value approaches the true cross-entropy value.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Normalized Mutual Information (NMI)\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"To compare results across models and across languages, we normalize MI values by dividing by the total entropy of the inflection class system. For example, the NMI shared by a Maltese noun's phonology and plural inflection may be calculated as:\\nNMI(C; W ) = MI(C; W ) H(C) (5)\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Model Details\",\n \"publication_ref\": [\n \"b41\",\n \"b41\"\n ],\n \"table_ref\": [],\n \"text\": \"We adapt the LSTM classifier implemented in Williams et al. (2020) to estimate the probability that a plural class c is associated with a given input noun w of gender g, i.e., q(c|w, g) in Equation 4. embedded and input into the model's initial hidden state. The model is trained using Adam (Kingma and Ba, 2015) with model hyperparameters, including the number of training epochs and the number and sizes of hidden layers, optimized using the Bayesian optimization technique implemented in Williams et al. (2020). The model then learns a probability distribution that serves to approximate q(c|w, g).\\nFollowing training, we test the model on a heldout dataset and use the model's cross-entropy loss to serve as an approximate upper bound on the conditional entropy H(C|W, G). We use 10-fold cross validation to make full use of the dataset for our approximations. To estimate q(c|w, e, g), we concatenate a binary character representing the word's etymology onto the end of the noun to serve as model input and follow the same procedure.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Results\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_1\"\n ],\n \"text\": \"NMI and H(C|G) values for C defined as concatenative type and plural allomorph, respectively, are presented in Table 2. The largest NMI value we obtain, NMI(E; W |G), indicates that more than half of the information needed to predict a word's etymology is shared with its phonology. In other words, it is often not difficult to guess the origin of a Maltese word based on how it sounds. Note that this value is consistent across models, as it does not depend on C.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Concatenative Type\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_1\"\n ],\n \"text\": \"Results for the model predicting a noun's concatenative type are in Table 2. Note first that the entropy H(C|G) of the plural inflection class system defined at the level of concatenative type is calculated to be 0.81, indicating that, given its gender, predicting whether a random Maltese noun takes concate-native or non-concatenative morphology is more predictable than chance, although not by much. We find phonology, indicated by NMI(C; W |G), to be more predictive than etymology, indicated by NMI(C; E|G). Crucially, each of these bipartite NMI values exceeds the tripartite mutual information NMI(C; E; W |G) shared across all three systems. This indicates that while a non-trivial amount of predictive information is shared across all three systems, phonology and etymology are each predictive of concatenative type in partially non-redundant ways. This suggests that both analogical and conservative forces are likely to have played a role in the development of the Maltese nominal plural system.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Plural Allomorph\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_1\"\n ],\n \"text\": \"In an analogical model of inflection in which singular inflected forms and their plural counterparts share a direct relationship in the lexicon, the predictive principles structuring the morphological system are expected to be most evident when defining an inflection class system at the level of the allomorph.\\nWe first note that the entropy H(C|G) calculated over the plural class distribution defined at the allomorph level is nearly three times as high as the entropy of C when defined as a noun's concatenative type. This is reflective of the higher degree of unpredictability associated with a non-uniform distribution of nouns over a greater number of inflection classes. When comparing across the allomorph and concatenative type models it is thus important to normalize for the fact that predicting allomorphs is more difficult than predicting concatenative type. However, even calculations normalized in this way show that the interpredictability among phonology, etymology, and plural inflection, indicated by the NMI values in Table 2, are all twice as high at the allomorph level as they are for concatenative type. In other words, a noun's singular form reduces the relative uncertainty about its plural allomorph twice as much as it reduces the uncertainty about whether that allomorph is concatenative. This suggests the analogical and conservative pressures hypothesized to shape morphological organization are more sensitive to correspondences at the word level than to typological similarities with respect to concatenativity.\\nAdditionally, the general tendency found at the level of concatenative type still follows when classes are defined at the level of individual allomorphs: phonology shares more information with inflection class than does etymology, with each factor contributing some amount of non-redundant information. This illustrates one key advantage of the methods employed in this study, namely the ability to disentangle the independent contributions of either predictor from the degree to which both exert redundant organizational pressure towards the same end.\\nFor example, given the fact that phonology and etymology are themselves mutually informative, we cannot uniquely interpret either bipartite measure of MI, that is, NMI(C; W |G) or NMI(C; E|G), as indicative of the forces hypothesized to shape the integration of linguistic material in contact. Rather, evidence for analogical structuring of the Maltese plural system at the allomorph level is specifically indicated by the positive difference between NMI(C; W |G) and NMI(C; E; W |G). Conservative pressures, such as those associated with high token-frequency items (Krause-Lerche, 2022), are similarly indicated by the extent to which NMI(C; E|G) exceeds NMI(C; E; W |G).\"\n },\n {\n \"figure_ref\": [\n \"fig_1\",\n \"fig_1\",\n \"fig_2\"\n ],\n \"heading\": \"Variation Across Allomorph Classes\",\n \"publication_ref\": [\n \"b41\"\n ],\n \"table_ref\": [\n \"tab_2\"\n ],\n \"text\": \"Closer examination of the model's predictions reveals an effect of type frequency, with larger inflection classes predicted more often than smaller classes. Table 3 reports the accuracy of all models in which singular noun phonology W is a predictor. Since all models achieve an overall accuracy above a majority baseline, the NMI values we obtain may be reliably interpreted as empirical minimums. However, as can be seen in Figure 2, the model's incorrect predictions do not clearly distinguish between sound and broken classes; nouns with a sound plural allomorph may be misclassified as taking a broken plural template, and nouns taking a broken plural may be incorrectly predicted to take a sound plural.\\nIf speakers are sensitive to differences between concatenative and non-concatenative allomorphs grouped into high-level macro classes (morphological subsystems), we might expect some degree of observable within-class coherence with respect to either or both of the phonology and etymology of words exhibiting a particular morphological behavior. Specifically, we would expect a pattern of predictions in which the LSTM is able to first Instead, as seen in Figure 2, we do not find such evidence. Rather, we find evidence for coherence at the allomorph level, specifically, for phonological patterns as a predictor of inflectional organization and driver of inflectional behavior at the allomorph level.\\nFinally, as in Williams et al. (2020), we also conduct an analysis of the partial Pointwise Mutual Information (PMI) shared between phonology W and class C with respect to the surprisal H(C = c|G) for each class, defined at the allomorph level. Figure 3 shows this distribution, with allomorph classes presented in order of increasing type frequency (and thus decreasing surprisal). We note that Maltese noun classes are each only par- \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Discussion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this paper we used an LSTM to help estimate the kinds and amounts of information that may be available to speakers when \\\"solving\\\" the PCFP. Overall, our results provide quantitative evidence for the role of both word phonology and etymology (as a stand-in for conservative factors) in shaping the Maltese lexicon.\\nSpecifically, we found that the extent to which a Maltese singular noun's phonology predicts its plural morphology exceeds that of etymology in nonredundant ways. This suggests that analogical pressures from phonological correspondences across the lexicon shape nominal plural inflection in Maltese, independently of the etymological source language for some word or morphological pattern.\\nOur results also show an independent contribution of etymology as a predictor. We hypothesize that this captures conservative pressures theorized to resist analogical change, including token frequency (Krause-Lerche, 2022). It may also reflect associative correlations from the use of lexemes of a common etymology in similar contexts, strengthening their coherence as a subsystem in the multilingual repertoire and encouraging the maintenance of a noun's original morphology. Further work is needed to investigate these possibilities.\\nIn language contact situations such as that of Maltese, it is likely that an influx of foreign lexemes and increased productivity of foreign affixes affect both the size and character (e.g., phonology) of nominal plural classes relative to each other over time. This in turn is likely to affect subsequent classification and integration of words into the inflectional morphology of the language.\\nIn general, our results do not support characterizations of Maltese in which concatenative and non-concatenative morphologies co-exist as discrete systems within the lexicon. While a singular noun's phonology and etymology are each somewhat predictive of its concatenative type, they are twice as predictive of the actual plural allomorph(s) with which the lexeme is associated. This suggests that systematic relationships at the word level organize the morphology of Maltese, in turn shaping the language as new words are integrated and inflected.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This study extends previous work in information theory, computational modeling, and theoretical morphology to provide quantitative evidence for the role of phonology as an analogical force in the morphological organization of Maltese. We ground this in a usage-based account of multilingualism and contact-induced change in which speakers are hypothesized to make use of analogical reasoning, among other language-general cognitive functions, when integrating novel words and patterns within a unified linguistic repertoire. The same processes that guide synchronic language use are proposed to be responsible for the diachronic effects of contactinduced language change. Specifically, it is hypothesized that speakers draw on similarities across multiple dimensions -including but not limited to phonological patterns, semantic and indexical meaning, pragmatic function, and contexts of useto collaboratively construct and adapt grammatical systems of linguistic communication over time.\\nIn the case of Maltese, our findings indicate that while a lexeme's phonology and etymology are themselves highly interpredictable, each contributes non-redundant information to reduce uncertainty when predicting the lexeme's plural inflection. While the etymology of a noun is somewhat predictive of its plural inflection, the word's phonology plays a much greater role. This synchronic analysis has diachronic implications. Our results suggest that analogical pressures from phonological similarities across the lexicon may have guided speakers' inflectional behavior when code mixing over the course of the development of the language to result in the conventionalized forms observed in modern Maltese. However, further diachronic study is needed to confirm this interpretation.\\nContrary to a hypothesis in which concatenative and non-concatenative systems operate as separate subsystems within a \\\"split\\\" or \\\"hybrid\\\" morphology, our results indicate correspondences at the level of individual wordforms and affixes are driving speakers' morphological behavior. Specifically, the phonology and etymology of a lexeme are twice as predictive of its plural allomorph than its concatenative type. Further investigation into Maltese nouns attested to take plural forms of both concatenative types may provide additional insight into the ways in which concatenative type affects speakers' behavior, if at all. Future work should also consider additional factors known to shape inflection class systems, for example by integrating semantic word vectors into the model. Finally, additional comparisons implementing these methods across corpora in a variety of languages will continue to shed light on the factors shaping morphological systems cross-linguistically. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A Nominal Plural Allomorphs in Maltese\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Acknowledgments\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We thank Jessica Nieder and Adam Ussishkin for generously sharing an abundance of digital resources and helpful feedback, Sarah Caruana for her insight into the Maltese language, and Christian Clark and Andrew Duffy for their contributions to an initial version of this project.\\nThis material is based on work supported by the National Science Foundation under grant BCS-2217554 (Neural discovery of abstract inflectional structure, PI Micha Elsner, Co-PI Andrea Sims).\"\n }\n]"},"abstract":{"kind":"string","value":"Maltese is often described as having a hybrid morphological system resulting from extensive contact between Semitic and Romance language varieties. Such a designation reflects an etymological divide as much as it does a larger tradition in the literature to consider concatenative and non-concatenative morphological patterns as distinct in the language architecture. Using a combination of computational modeling and information theoretic methods, we quantify the extent to which the phonology and etymology of a Maltese singular noun may predict the morphological process (affixal vs. templatic) as well as the specific plural allomorph (affix or template) relating a singular noun to its associated plural form(s) in the lexicon. The results indicate phonological pressures shape the organization of the Maltese lexicon with predictive power that extends beyond that of a word's etymology, in line with analogical theories of language change in contact."},"title":{"kind":"string","value":"Analogy in Contact: Modeling Maltese Plural Inflection"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 1 :1Figure 1: Tripartite Mutual Information","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"Figure 2 :2Figure 2: Confusion matrix: predicting plural allomorph from singular phonology and gender","figure_data":"","figure_id":"fig_1","figure_label":"2","figure_type":"figure"},{"figure_caption":"Figure 3 :3Figure 3: Partial Pointwise Mutual Information (PMI) shared by word form and class for each allomorph class","figure_data":"","figure_id":"fig_2","figure_label":"3","figure_type":"figure"},{"figure_caption":"Distribution of Maltese nominal plural allomorphs by lexeme etymology and concatenative type","figure_data":"SemiticTotalLexemeLexeme(%)Non-Semitic Affix1,2742142%Semitic Affix41668435%Semitic Template24053723%Total (%)62%38%100%","figure_id":"tab_0","figure_label":"1","figure_type":"table"},{"figure_caption":"Normalized Mutual Information measures for plural class C defined with respect to TYPE vs.","figure_data":"TYPE ALLO.","figure_id":"tab_1","figure_label":"2","figure_type":"table"},{"figure_caption":"Model accuracy for all models predicting Etymology E or Plural Class C (Type vs. Allomorph) using the Phonology W of singular nouns in Maltese identify a lexeme's concatenative type before predicting, possibly incorrectly, an allomorph of that specific type.","figure_data":"TargetModelAccuracyETYM. (E)MI(E; W |G) Baseline0.90 0.62MI(C; W |G)0.80TYPE (C)MI(C; E; W |G)0.81Baseline0.77MI(C; W |G)0.65ALLOMORPHMI(C; E; W |G)0.68(C)Baseline0.40","figure_id":"tab_2","figure_label":"3","figure_type":"table"},{"figure_caption":"Sound plural allomorphs in Maltese, fromNieder et al. (2021b) ","figure_data":"Sound PluralSingular PluralGlossAllomorphkartakarti'paper'-iommommijiet 'mother'-ijietrixarixiet'feather'-ietgiddieb giddieba 'liar'-ameèlusmeèlusin 'freed'-inkuxinkuxins'cushions' -striqtriqat'street'-atsidsidien'owner'-ienbaèribaèrin'sailor'-nèatièatjin'guilty'-jinspallaspallejn'shoulder' -ejnsieqsaqajn'foot'-ajnqiegèqiegèan'bottom'-anBroken PluralSingular PluralGlossAllomorphfardalfradal'apron'CCVVCVCbirrabirer'beer'(C)CVCVCkbirkbar'big'CCVVCftiraftajjar'type of bread' CCVjjVCbitèabtieèi'yard'CCVVCVsiderisdra'chest'VCCCVmaridmorda'sick person'CVCCVgèodda gèodod'tool'(gè)VCVCelfeluf'thousand'VCVCgèarefgèorrief 'wise man'CVCCVVC(V)gèamagèomja'blind person' (gè)VCCV","figure_id":"tab_3","figure_label":"4","figure_type":"table"},{"figure_caption":"Broken plural allomorphs in Maltese, fromNieder et al. (2021b) ","figure_data":"","figure_id":"tab_4","figure_label":"5","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 1 :1Figure 1: Tripartite Mutual Information\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 2 :2Figure 2: Confusion matrix: predicting plural allomorph from singular phonology and gender\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 :3Figure 3: Partial Pointwise Mutual Information (PMI) shared by word form and class for each allomorph class\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Distribution of Maltese nominal plural allomorphs by lexeme etymology and concatenative type\",\n \"figure_data\": \"SemiticTotalLexemeLexeme(%)Non-Semitic Affix1,2742142%Semitic Affix41668435%Semitic Template24053723%Total (%)62%38%100%\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Normalized Mutual Information measures for plural class C defined with respect to TYPE vs.\",\n \"figure_data\": \"TYPE ALLO.\",\n \"figure_id\": \"tab_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Model accuracy for all models predicting Etymology E or Plural Class C (Type vs. Allomorph) using the Phonology W of singular nouns in Maltese identify a lexeme's concatenative type before predicting, possibly incorrectly, an allomorph of that specific type.\",\n \"figure_data\": \"TargetModelAccuracyETYM. (E)MI(E; W |G) Baseline0.90 0.62MI(C; W |G)0.80TYPE (C)MI(C; E; W |G)0.81Baseline0.77MI(C; W |G)0.65ALLOMORPHMI(C; E; W |G)0.68(C)Baseline0.40\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Sound plural allomorphs in Maltese, fromNieder et al. (2021b) \",\n \"figure_data\": \"Sound PluralSingular PluralGlossAllomorphkartakarti'paper'-iommommijiet 'mother'-ijietrixarixiet'feather'-ietgiddieb giddieba 'liar'-ameèlusmeèlusin 'freed'-inkuxinkuxins'cushions' -striqtriqat'street'-atsidsidien'owner'-ienbaèribaèrin'sailor'-nèatièatjin'guilty'-jinspallaspallejn'shoulder' -ejnsieqsaqajn'foot'-ajnqiegèqiegèan'bottom'-anBroken PluralSingular PluralGlossAllomorphfardalfradal'apron'CCVVCVCbirrabirer'beer'(C)CVCVCkbirkbar'big'CCVVCftiraftajjar'type of bread' CCVjjVCbitèabtieèi'yard'CCVVCVsiderisdra'chest'VCCCVmaridmorda'sick person'CVCCVgèodda gèodod'tool'(gè)VCVCelfeluf'thousand'VCVCgèarefgèorrief 'wise man'CVCCVVC(V)gèamagèomja'blind person' (gè)VCCV\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Broken plural allomorphs in Maltese, fromNieder et al. (2021b) \",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"5\",\n \"figure_type\": \"table\"\n }\n]"},"authors":{"kind":"string","value":"Sara Court; Andrea D Sims; Micha Elsner"},"references":{"kind":"list like","value":[{"authors":"James P Farrell Ackerman; Robert Blevins; Malouf","journal":"Oxford University Press","ref_id":"b0","title":"Parts and wholes: Patterns of relatedness in complex morphological systems and why they matter","year":"2009"},{"authors":"Maria Alegre; Peter Gordon","journal":"Brain and Language","ref_id":"b1","title":"Rule-based versus associative processes in derivational morphology","year":"1999"},{"authors":"Raimo Anttila","journal":"Mouton","ref_id":"b2","title":"Analogy","year":"1977"},{"authors":"Joseph Aquilina","journal":"Midsea Books","ref_id":"b3","title":"Concise Maltese-English, English-Maltese Dictionary","year":"2006"},{"authors":"Karim Bensoukas","journal":"International Journal of Arabic Linguistics","ref_id":"b4","title":"Concurrent cognate and contact-induced plural traits in Afro-Asiatic: Amazigh id-and Arabic -at plurals","year":"2018"},{"authors":"Albert J Borg; Marie Azzopardi-Alexander","journal":"Routledge","ref_id":"b5","title":"Maltese","year":"1997"},{"authors":"Claudia Borg; Albert Gatt","journal":"Association for Computational Linguistics","ref_id":"b6","title":"Morphological analysis for the Maltese language: The challenges of a hybrid system","year":"2017"},{"authors":"Joseph M Brincat","journal":"Midsea Books","ref_id":"b7","title":"Maltese and other languages: A linguistic history of Malta","year":"2011"},{"authors":"Joseph M Brincat","journal":"Lexicographica","ref_id":"b8","title":"Maltese: Blending Semitic, Romance and Germanic lexemes","year":"2017"},{"authors":"F Peter; Stephen A Brown; Vincent J Della Pietra; Jennifer C Della Pietra; Robert L Lai; Mercer","journal":"Computational Linguistics","ref_id":"b9","title":"An estimate of an upper bound for the entropy of English","year":"1992"},{"authors":"G Greville; Norman M Corbett; Fraser","journal":"Cambridge University Press","ref_id":"b10","title":"Gender assignment: A typology and a model","year":"2000"},{"authors":"Lisa Garnand; Dawdy-Hesterberg ; Janet Breckenridge; Pierrehumbert ","journal":"Language, Cognition and Neuroscience","ref_id":"b11","title":"Learnability and generalisation of Arabic broken plural nouns","year":"2014"},{"authors":"Ray Fabri","journal":"Revue belge de Philologie et d'Histoire","ref_id":"b12","title":"Maltese","year":"2010"},{"authors":"L David; Fertig","journal":"Edinburgh University Press","ref_id":"b13","title":"Analogy and morphological change","year":"2013"},{"authors":"A Victor; Friedman","journal":"Romani Studies","ref_id":"b14","title":"Compartmentalized grammar: The variable (non)-integration of Turkish verbal conjugation in Romani dialects","year":"2013"},{"authors":"Francesco Gardani","journal":"Morphology","ref_id":"b15","title":"Borrowing matter and pattern in morphology: An overview","year":"2020"},{"authors":"Francesco Gardani","journal":"Word Structure","ref_id":"b16","title":"On how morphology spreads","year":"2021"},{"authors":"Albert Gatt; Slavomír Čéplö","journal":"","ref_id":"b17","title":"Digital corpora and other electronic resources for Maltese","year":"2013"},{"authors":"Guzmán Matías; Naranjo","journal":"Morphology","ref_id":"b18","title":"Analogy, complexity and predictability in the Russian nominal inflection system","year":"2020"},{"authors":"Einar Haugen","journal":"Language","ref_id":"b19","title":"The analysis of linguistic borrowing","year":"1950"},{"authors":"Robert D Hoberman","journal":"Morphologies of Asia and Africa","ref_id":"b20","title":"Maltese morphology","year":"2007"},{"authors":"Sepp Hochreiter; Jürgen Schmidhuber","journal":"Advances in Neural Information Processing Systems","ref_id":"b21","title":"LSTM can solve hard long time lag problems","year":"1996"},{"authors":"Hans Henrich; Hock ","journal":"De Gruyter Mouton","ref_id":"b22","title":"Principles of historical linguistics","year":"1991"},{"authors":"Brian D Joseph","journal":"Working papers in linguistics","ref_id":"b23","title":"On phonically based analogy","year":"2013"},{"authors":"P Diederik; Jimmy Kingma; Ba","journal":"","ref_id":"b24","title":"Adam: A method for stochastic optimization","year":"2015"},{"authors":"Anne Krause-Lerche","journal":"Corpus Linguistics and Linguistic Theory","ref_id":"b25","title":"Conservation in ongoing analogical change: The measurement and effect(s) of token frequency","year":"2022"},{"authors":"Yaron Matras","journal":"Cambridge University Press","ref_id":"b26","title":"Language contact","year":"2009"},{"authors":"Yaron Matras","journal":"De Gruyter Mouton","ref_id":"b27","title":"Why is the borrowing of inflectional morphology dispreferred?","year":"2015"},{"authors":"Yaron Matras; Evangelia Adamou","journal":"Routledge","ref_id":"b28","title":"Borrowing","year":"2020"},{"authors":"Carol Myers-Scotton","journal":"Clarendon","ref_id":"b29","title":"Duelling languages: Grammatical structure in codeswitching","year":"1997"},{"authors":"Jessica Nieder; Fabian Tomaschek; Enum Cohrs; Ruben Van De Vijver","journal":"Language, Cognition and Neuroscience","ref_id":"b30","title":"Modelling Maltese noun plural classes without morphemes","year":"2021"},{"authors":"Jessica Nieder; Fabian Tomaschek; Enum Cohrs; Ruben Van De Vijver","journal":"","ref_id":"b31","title":"Modeling Maltese noun plural classes without morphemes: Supplemental materials","year":"2021"},{"authors":"Jessica Nieder; Ruben Van De; Yu-Ying Vijver; R H Chuang; Baayen","journal":"","ref_id":"b32","title":"A discriminative lexicon approach to word comprehension, production and processing: Maltese plurals: Supplemental materials","year":"2021"},{"authors":"Jessica Nieder; Ruben Van De; Holger Vijver; Mitterer","journal":"Morphology","ref_id":"b33","title":"d. Knowledge of Maltese singular-plural mappings","year":"2021"},{"authors":"Robert M Nosofsky","journal":"Journal of Mathematical Psychology","ref_id":"b34","title":"Relations between exemplar-similarity and likelihood models of classification","year":"1990"},{"authors":"Jeff Parker; Robert Reynolds; Andrea D Sims","journal":"Cambridge University Press","ref_id":"b35","title":"Network structure and inflection class predictability: Modeling the emergence of Marginal Detraction","year":"2022"},{"authors":"Shana Poplack; David Sankoff; Christophe Miller","journal":"Linguistics","ref_id":"b36","title":"The social correlates and linguistic processes of lexical borrowing and assimilation","year":"1988"},{"authors":"Tamara Schembri","journal":"Brockmeyer","ref_id":"b37","title":"The broken plural in Maltese: A description","year":"2012"},{"authors":"Andrea D Sims; Jeff Parker","journal":"Word Structure","ref_id":"b38","title":"How inflection class systems work: On the informativity of implicative structure","year":"2016"},{"authors":"Michael Spagnol","journal":"","ref_id":"b39","title":"A tale of two morphologies: Verb structure and argument alternations in Maltese","year":"2011"},{"authors":"Uriel Weinreich","journal":"De Gruyter Mouton","ref_id":"b40","title":"Languages in contact: Findings and problems","year":"1953"},{"authors":"Adina Williams; Tiago Pimentel; Hagen Blix; D Arya; Eleanor Mccarthy; Ryan Chodroff; Cotterell","journal":"Association for Computational Linguistics","ref_id":"b41","title":"Predicting declension class from form and meaning","year":"2020"},{"authors":"Zoubida Ziani","journal":"International Journal of Arabic Linguistics","ref_id":"b42","title":"The morphology of borrowings and its relevance to lexical organization in Moroccan Arabic","year":"2020"}],"string":"[\n {\n \"authors\": \"James P Farrell Ackerman; Robert Blevins; Malouf\",\n \"journal\": \"Oxford University Press\",\n \"ref_id\": \"b0\",\n \"title\": \"Parts and wholes: Patterns of relatedness in complex morphological systems and why they matter\",\n \"year\": \"2009\"\n },\n {\n \"authors\": \"Maria Alegre; Peter Gordon\",\n \"journal\": \"Brain and Language\",\n \"ref_id\": \"b1\",\n \"title\": \"Rule-based versus associative processes in derivational morphology\",\n \"year\": \"1999\"\n },\n {\n \"authors\": \"Raimo Anttila\",\n \"journal\": \"Mouton\",\n \"ref_id\": \"b2\",\n \"title\": \"Analogy\",\n \"year\": \"1977\"\n },\n {\n \"authors\": \"Joseph Aquilina\",\n \"journal\": \"Midsea Books\",\n \"ref_id\": \"b3\",\n \"title\": \"Concise Maltese-English, English-Maltese Dictionary\",\n \"year\": \"2006\"\n },\n {\n \"authors\": \"Karim Bensoukas\",\n \"journal\": \"International Journal of Arabic Linguistics\",\n \"ref_id\": \"b4\",\n \"title\": \"Concurrent cognate and contact-induced plural traits in Afro-Asiatic: Amazigh id-and Arabic -at plurals\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Albert J Borg; Marie Azzopardi-Alexander\",\n \"journal\": \"Routledge\",\n \"ref_id\": \"b5\",\n \"title\": \"Maltese\",\n \"year\": \"1997\"\n },\n {\n \"authors\": \"Claudia Borg; Albert Gatt\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b6\",\n \"title\": \"Morphological analysis for the Maltese language: The challenges of a hybrid system\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Joseph M Brincat\",\n \"journal\": \"Midsea Books\",\n \"ref_id\": \"b7\",\n \"title\": \"Maltese and other languages: A linguistic history of Malta\",\n \"year\": \"2011\"\n },\n {\n \"authors\": \"Joseph M Brincat\",\n \"journal\": \"Lexicographica\",\n \"ref_id\": \"b8\",\n \"title\": \"Maltese: Blending Semitic, Romance and Germanic lexemes\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"F Peter; Stephen A Brown; Vincent J Della Pietra; Jennifer C Della Pietra; Robert L Lai; Mercer\",\n \"journal\": \"Computational Linguistics\",\n \"ref_id\": \"b9\",\n \"title\": \"An estimate of an upper bound for the entropy of English\",\n \"year\": \"1992\"\n },\n {\n \"authors\": \"G Greville; Norman M Corbett; Fraser\",\n \"journal\": \"Cambridge University Press\",\n \"ref_id\": \"b10\",\n \"title\": \"Gender assignment: A typology and a model\",\n \"year\": \"2000\"\n },\n {\n \"authors\": \"Lisa Garnand; Dawdy-Hesterberg ; Janet Breckenridge; Pierrehumbert \",\n \"journal\": \"Language, Cognition and Neuroscience\",\n \"ref_id\": \"b11\",\n \"title\": \"Learnability and generalisation of Arabic broken plural nouns\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"Ray Fabri\",\n \"journal\": \"Revue belge de Philologie et d'Histoire\",\n \"ref_id\": \"b12\",\n \"title\": \"Maltese\",\n \"year\": \"2010\"\n },\n {\n \"authors\": \"L David; Fertig\",\n \"journal\": \"Edinburgh University Press\",\n \"ref_id\": \"b13\",\n \"title\": \"Analogy and morphological change\",\n \"year\": \"2013\"\n },\n {\n \"authors\": \"A Victor; Friedman\",\n \"journal\": \"Romani Studies\",\n \"ref_id\": \"b14\",\n \"title\": \"Compartmentalized grammar: The variable (non)-integration of Turkish verbal conjugation in Romani dialects\",\n \"year\": \"2013\"\n },\n {\n \"authors\": \"Francesco Gardani\",\n \"journal\": \"Morphology\",\n \"ref_id\": \"b15\",\n \"title\": \"Borrowing matter and pattern in morphology: An overview\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Francesco Gardani\",\n \"journal\": \"Word Structure\",\n \"ref_id\": \"b16\",\n \"title\": \"On how morphology spreads\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Albert Gatt; Slavomír Čéplö\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"Digital corpora and other electronic resources for Maltese\",\n \"year\": \"2013\"\n },\n {\n \"authors\": \"Guzmán Matías; Naranjo\",\n \"journal\": \"Morphology\",\n \"ref_id\": \"b18\",\n \"title\": \"Analogy, complexity and predictability in the Russian nominal inflection system\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Einar Haugen\",\n \"journal\": \"Language\",\n \"ref_id\": \"b19\",\n \"title\": \"The analysis of linguistic borrowing\",\n \"year\": \"1950\"\n },\n {\n \"authors\": \"Robert D Hoberman\",\n \"journal\": \"Morphologies of Asia and Africa\",\n \"ref_id\": \"b20\",\n \"title\": \"Maltese morphology\",\n \"year\": \"2007\"\n },\n {\n \"authors\": \"Sepp Hochreiter; Jürgen Schmidhuber\",\n \"journal\": \"Advances in Neural Information Processing Systems\",\n \"ref_id\": \"b21\",\n \"title\": \"LSTM can solve hard long time lag problems\",\n \"year\": \"1996\"\n },\n {\n \"authors\": \"Hans Henrich; Hock \",\n \"journal\": \"De Gruyter Mouton\",\n \"ref_id\": \"b22\",\n \"title\": \"Principles of historical linguistics\",\n \"year\": \"1991\"\n },\n {\n \"authors\": \"Brian D Joseph\",\n \"journal\": \"Working papers in linguistics\",\n \"ref_id\": \"b23\",\n \"title\": \"On phonically based analogy\",\n \"year\": \"2013\"\n },\n {\n \"authors\": \"P Diederik; Jimmy Kingma; Ba\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Adam: A method for stochastic optimization\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Anne Krause-Lerche\",\n \"journal\": \"Corpus Linguistics and Linguistic Theory\",\n \"ref_id\": \"b25\",\n \"title\": \"Conservation in ongoing analogical change: The measurement and effect(s) of token frequency\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Yaron Matras\",\n \"journal\": \"Cambridge University Press\",\n \"ref_id\": \"b26\",\n \"title\": \"Language contact\",\n \"year\": \"2009\"\n },\n {\n \"authors\": \"Yaron Matras\",\n \"journal\": \"De Gruyter Mouton\",\n \"ref_id\": \"b27\",\n \"title\": \"Why is the borrowing of inflectional morphology dispreferred?\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Yaron Matras; Evangelia Adamou\",\n \"journal\": \"Routledge\",\n \"ref_id\": \"b28\",\n \"title\": \"Borrowing\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Carol Myers-Scotton\",\n \"journal\": \"Clarendon\",\n \"ref_id\": \"b29\",\n \"title\": \"Duelling languages: Grammatical structure in codeswitching\",\n \"year\": \"1997\"\n },\n {\n \"authors\": \"Jessica Nieder; Fabian Tomaschek; Enum Cohrs; Ruben Van De Vijver\",\n \"journal\": \"Language, Cognition and Neuroscience\",\n \"ref_id\": \"b30\",\n \"title\": \"Modelling Maltese noun plural classes without morphemes\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Jessica Nieder; Fabian Tomaschek; Enum Cohrs; Ruben Van De Vijver\",\n \"journal\": \"\",\n \"ref_id\": \"b31\",\n \"title\": \"Modeling Maltese noun plural classes without morphemes: Supplemental materials\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Jessica Nieder; Ruben Van De; Yu-Ying Vijver; R H Chuang; Baayen\",\n \"journal\": \"\",\n \"ref_id\": \"b32\",\n \"title\": \"A discriminative lexicon approach to word comprehension, production and processing: Maltese plurals: Supplemental materials\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Jessica Nieder; Ruben Van De; Holger Vijver; Mitterer\",\n \"journal\": \"Morphology\",\n \"ref_id\": \"b33\",\n \"title\": \"d. Knowledge of Maltese singular-plural mappings\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Robert M Nosofsky\",\n \"journal\": \"Journal of Mathematical Psychology\",\n \"ref_id\": \"b34\",\n \"title\": \"Relations between exemplar-similarity and likelihood models of classification\",\n \"year\": \"1990\"\n },\n {\n \"authors\": \"Jeff Parker; Robert Reynolds; Andrea D Sims\",\n \"journal\": \"Cambridge University Press\",\n \"ref_id\": \"b35\",\n \"title\": \"Network structure and inflection class predictability: Modeling the emergence of Marginal Detraction\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Shana Poplack; David Sankoff; Christophe Miller\",\n \"journal\": \"Linguistics\",\n \"ref_id\": \"b36\",\n \"title\": \"The social correlates and linguistic processes of lexical borrowing and assimilation\",\n \"year\": \"1988\"\n },\n {\n \"authors\": \"Tamara Schembri\",\n \"journal\": \"Brockmeyer\",\n \"ref_id\": \"b37\",\n \"title\": \"The broken plural in Maltese: A description\",\n \"year\": \"2012\"\n },\n {\n \"authors\": \"Andrea D Sims; Jeff Parker\",\n \"journal\": \"Word Structure\",\n \"ref_id\": \"b38\",\n \"title\": \"How inflection class systems work: On the informativity of implicative structure\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Michael Spagnol\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"A tale of two morphologies: Verb structure and argument alternations in Maltese\",\n \"year\": \"2011\"\n },\n {\n \"authors\": \"Uriel Weinreich\",\n \"journal\": \"De Gruyter Mouton\",\n \"ref_id\": \"b40\",\n \"title\": \"Languages in contact: Findings and problems\",\n \"year\": \"1953\"\n },\n {\n \"authors\": \"Adina Williams; Tiago Pimentel; Hagen Blix; D Arya; Eleanor Mccarthy; Ryan Chodroff; Cotterell\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b41\",\n \"title\": \"Predicting declension class from form and meaning\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Zoubida Ziani\",\n \"journal\": \"International Journal of Arabic Linguistics\",\n \"ref_id\": \"b42\",\n \"title\": \"The morphology of borrowings and its relevance to lexical organization in Moroccan Arabic\",\n \"year\": \"2020\"\n }\n]"},"formulas":{"kind":"list like","value":[{"formula_coordinates":[5,344.83,235.88,179.58,9.81],"formula_id":"formula_0","formula_text":"MI(C; W ) = H(C) -H(C|W )(1)"},{"formula_coordinates":[5,311.6,304.77,212.81,9.81],"formula_id":"formula_1","formula_text":"MI(C; E; W ) = MI(C; W ) -MI(C; W |E) (2)"},{"formula_coordinates":[5,321.06,404.33,203.35,9.81],"formula_id":"formula_2","formula_text":"MI(C; W |G) = H(C|G) -H(C|W, G) (3)"},{"formula_coordinates":[6,105.67,440.98,183.46,33.71],"formula_id":"formula_3","formula_text":"H(C|W ) ≤ - 1 M M i=1 log q(c i |w i ) (4)"},{"formula_coordinates":[6,119.77,626.97,169.36,24.43],"formula_id":"formula_4","formula_text":"NMI(C; W ) = MI(C; W ) H(C) (5)"}],"string":"[\n {\n \"formula_coordinates\": [\n 5,\n 344.83,\n 235.88,\n 179.58,\n 9.81\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"MI(C; W ) = H(C) -H(C|W )(1)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 311.6,\n 304.77,\n 212.81,\n 9.81\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"MI(C; E; W ) = MI(C; W ) -MI(C; W |E) (2)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 321.06,\n 404.33,\n 203.35,\n 9.81\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"MI(C; W |G) = H(C|G) -H(C|W, G) (3)\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 105.67,\n 440.98,\n 183.46,\n 33.71\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"H(C|W ) ≤ - 1 M M i=1 log q(c i |w i ) (4)\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 119.77,\n 626.97,\n 169.36,\n 24.43\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"NMI(C; W ) = MI(C; W ) H(C) (5)\"\n }\n]"}}},{"rowIdx":390,"cells":{"doi":{"kind":"string","value":"10.18653/v1/2020.acl-main.632"},"pub_date":{"kind":"string","value":"2023-11-08"},"sections":{"kind":"list like","value":[{"figure_ref":[],"heading":"Introduction","publication_ref":["b21","b3","b26","b32","b2","b8","b20","b27","b12","b25","b24","b23","b14","b11","b19","b19","b11","b11","b9","b13","b30","b17","b20","b6","b7","b22","b15","b27","b12","b12"],"table_ref":[],"text":"Reliable analysis of arguments in natural language holds the promise to support applications such as automated grading (Ludwig et al., 2021), and tackling misinformation and targeted speech (Alhindi, Tariq, 2023). Computational argument analysis has been relatively popular through tasks like argument extraction (Chakrabarty et al., 2019), evidence mining (Rinott et al., 2015), relation assignment (Trautmann et al., 2020), writing support (Stab and Gurevych, 2014b) and claim generation (Bilu and Slonim, 2016). A particularly challenging task is argument quality assessment (Fromm et al., 2022), which addresses the cogency, effectiveness, and reasonableness of an argument (Wachsmuth et al., 2017b) pertaining to a topic. Assessing the quality of the argument involves analyzing the objective evidence, relevant assumptions, and structural soundness, making the overall task difficult.\nResearch on argument quality assessment has focused on extracting textual patterns using various learning frameworks and content features (Lauscher et al., 2022). It has been widely recognized that contextualizing arguments with implicit knowledge, such as extracting claim revisions (Skitalinskaya et al., 2021) and generating explicit conclusions (Gurcke et al., 2021) can be informative for reasoning models. However, we note that: 1) these methods fail to generalize to novel arguments where this information is not available, and 2) no prior work has considered jointly a comprehensive set of such contextualization strategies.\nTo bridge these gaps, we propose a novel framework called SPARK (Scoring the Pragmatics of Arguments via Relevant Knowledge), which incorporates augmentation strategies based on a large language model (LLM), GPT 3.5 (OpenAI, 2022), and elements from argumentation literature (Nickerson, 2020;Mulyati et al., 2023;Harvey, 2009), specifically, feedback, assumptions, arguments with similar quality, and counter-arguments. SPARK processes the original argument and topic and its augmentations separately using a dualencoder Transformer architecture with a multi-head cross-attention layer. We demonstrate the effectiveness of SPARK 's augmentations and architecture using both in-domain and out-of-domain datasets. Our entire code is made available at (anonymized).\nTask formulation. Inspired by prior work (Gretz et al., 2020;Lauscher et al., 2020), we formalize argumentation quality assessment as a regression task of predicting the quality of a natural language argument. Given a topic and an argument, we consider three quality indicators (Lauscher et al., 2020): 1) cogency, which evaluates the relevance and sufficiency of the argument's premise in relation to the conclusion, 2) effectiveness, which measures the argument's persuasive power based on factors like arrangement, clarity, and appropriateness, and 3) reasonableness, which determines the argument's ability to resolve the debate's issue (Wachsmuth et al., 2017a). The overall quality of an argument can be estimated by averaging these three metrics (Gretz et al., 2020). Connection to prior studies. The introduction of benchmarks for argument quality (Stab and Gurevych, 2014a;Gretz et al., 2020) has inspired various methods based on logistic regression (Ghosh et al., 2016), fully connected and recurrent neural networks (Habernal and Gurevych, 2016), and fine-tuned Transformers (Toledo et al., 2019). Hulpus et al. (2019) explained that contextualizing an argument with implicit knowledge is essential to understanding its quality. Lauscher et al. (2022) categorized knowledge used by current argument assessment research so far under linguistic, task-specific, and argument-specific sections. To mimic human reasoning over arguments, prior work has incorporated users' prior beliefs as predictors of argument persuasiveness (Durmus and Cardie, 2018), trained classifiers for different audience groups (El Baff et al., 2020), utilized user history to predict persuasion (Al Khatib et al., 2020), augmented arguments with supporting or refuting documents (Marro et al., 2022), and augmented arguments with visual features (Hasan et al., 2021). Most similar to SPARK, Skitalinskaya et al. (2021) leverage comparison between revisions of the same claims, while Gurcke et al. (2021) generate conclusions to assess argument sufficiency. However, revisions are rarely provided for novel arguments, whereas generated conclusions are argument-specific and may not generalize well (Gurcke et al., 2021). Addressing prior work limitations, SPARK implements four wellmotivated augmentation strategies to enhance novel arguments and utilizes an attention-based dual encoder model for effective reasoning."},{"figure_ref":["fig_0","fig_0","fig_0","fig_0","fig_1"],"heading":"SPARK","publication_ref":["b23","b24","b27","b16","b14","b10","b33"],"table_ref":[],"text":"Augmentation strategies. We devise four augmentation techniques to contextualize arguments. We generate the augmentations by prompting GPT-3.5 (OpenAI, 2022) (see appendix for details). Feedback. Constructive feedback in the form of comments and suggestions helps comprehension and domain knowledge acquisition (Mulyati et al., 2023). We hypothesize that assessing argument strengths and weaknesses helps argument ranking, as in Figure 1, where feedback identifies emotional appeal, insufficient evidence, and generalization. We prompt the LLM to generate writing feedback for a topic-argument pair in a zero-shot setting. Assumptions. Unstated assumptions frequently introduce bias in arguments (Nickerson, 2020). Making assumptions explicit can reveal these hidden biases, which may aid in assessing argument persuasiveness and relevance. One such assumption in Figure 1 is that people do not take driving seriously. We employ an LLM to extract the argument's underlying assumptions in a zero-shot setting. Similar-quality instance. Inspired by prior work on claim revisions (Skitalinskaya et al., 2021), we hypothesize that retrieving arguments with similar quality at training time leads to generalizable model learning. For this purpose, we derive a synthetic argument with similar reasonableness, cogency, and effectiveness to the original one (Figure 1). We generate this synthetic argument in a few-shot setting, where the LLM has access to example arguments alongside their quality scores covering the full 1-5 range. Since this augmentation uses ground-truth information that is not available during inference, we randomly replace synthetic arguments with None at training time with a probability of P = 0.5, thus familiarizing the model with the absence of similar arguments during testing where it only sees None. This technique is similar to distillation (Hinton et al., 2015), where the encoder's (student) goal is to utilize the LLM's (teacher) ranking-based argument generations (soft labels) to learn the argument quality ranking task. Counter-arguments. Counter-arguments provide objections, alternatives, and doubts of a skeptical reader (Harvey, 2009). We expect that contrasting the strengths and weaknesses of two opposite arguments will aid quality assessment. The example counter-argument in Figure 1 makes a firm claim that alternative activities (e.g., eating) lead to distractions, and solely blaming phones is unfair. We ask the LLM to provide a counter-argument for a topic and an argument in a zero-shot setting. Dual-encoder architecture. To consider the argument together with the augmentations, we employ a dual BERT encoder (Figure 2) as an improvement to the architecture by Gillick et al. (2018). The first encoder embeds the topic and argument, whereas the second embeds the augmentations. The second encoder can store individual augmentations or their concatenation, arbitrarily fixed to Similar quality argument [SEP] Feedback [SEP] Assumptions [SEP] Counter-argument. Notably, the dual encoder can effectively store all of the augmentation data without truncating information in practically all cases (see subsection A.4 for the augmentation lengths). We use a multi-head crossattention layer (Vaswani et al., 2017) to enable the model to weigh each augmentation according to the argument-topic pair. We pass the attention outputs to a mean pooler, whose output is fed into three separate regressor heads, one per quality metric."},{"figure_ref":[],"heading":"Experiments","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Baselines","publication_ref":["b5","b37","b18","b4","b31"],"table_ref":[],"text":"Scoring models. We compare our dual BERT with a standard BERT model to study the effect of disjoint embeddings. We use bert-base-uncased (Devlin et al., 2019) as an encoder. We compare dual BERT to XLNet (Yang et al., 2019), as they can both handle more than 512 tokens. Finally, we utilize GPT-3.5 in a zero-shot setting to gauge the model's ability to accomplish the task directly, without a dual encoder. We provide GPT-3.5 with the definitions of each metric and prompt it to individually rate each argument by a float between 1 and 5 with respect to the topic. Alternative augmentation strategies. We evaluate the impact of using all augmentations together or one at a time, against ablated baselines without augmentations. We include two alternative augmentation methods: Wikipedia paragraphs extracted using dense passage retrieval (DPR) (Karpukhin et al., 2020), and augmentations generated using smaller models, namely, Flan-T5-XL (Chung et al., 2022) and Llama-2 (7B) (Touvron et al., 2023)."},{"figure_ref":[],"heading":"Datasets and Evaluation","publication_ref":["b19","b11","b19"],"table_ref":[],"text":"We use GAQCorpus (Lauscher et al., 2020) as our training dataset for its diversity of domains (reviews, QA, and debates) and quality metrics (cogency, effectiveness, and reasonableness). We also use IBM-30K (Gretz et al., 2020) to test SPARK 's generalization on out-of-domain data. For IBM-30K, we perform weighted averaging (WA) of the three metric scores, which is supported by the high correlation between IBM-30K's WA and the GAQ-Corpus metrics (Lauscher et al., 2020). We report the Pearson (ρ) and Spearman (σ) correlation coefficients between the predictions and ground truth."},{"figure_ref":[],"heading":"Results","publication_ref":[],"table_ref":["tab_0"],"text":"We investigate whether augmentations help language models assess the quality of arguments more effectively (Q1); how augmentation strategies compare to each other (Q2); and whether human quality judgments align with their model utility (Q3).\nEffect of augmentation on in-domain performance (Q1). Table 1 shows that the overall best-performing combination uses Dual BERT with all four augmentations combined, which improves the Spearman correlation over the baseline BERT by 0.08-0.17 across the three metrics. The improvement is the largest for effectiveness, where the Spearman correlation increases by 61%. While both single BERT and XLNet benefit from SPARK 's augmentations as well, their performance is consistently lower than using the dual encoder. GPT-3 alone often performs better than the other baselines but lags significantly behind SPARK, showing the importance of the dual encoder. Among the augmentations, the benefit of our four augmentations declines when using a smaller generative model (Flan T5 and Llama-2), augmenting via DPR, or using the dual BERT with a masked second encoder. Thus, merely adding text to the second encoder does not by itself bring higher performance. The gap between SPARK and the baselines increases in the zero-shot setting on the IBM-Rank-30K dataset. Here, augmentation with DPR, Flan-T5, and Llama-2 is consistently inferior, as is the SPARK augmentation of the singleencoder methods. In summary, SPARK effectively combines dual encoding and data augmentation for strong task accuracy and generalization."},{"figure_ref":[],"heading":"Comparison of augmentation variants (Q2).","publication_ref":[],"table_ref":["tab_1"],"text":"On the in-domain task, the dual encoder performs the best when it has access to the information from the four SPARK augmentations simultaneously. Among them, feedback is most effective for predicting argument cogency and reasonableness as it exposes flaws that directly relate to these metrics. Meanwhile, contextualizing through similar-quality arguments is optimal for predicting effectiveness, which we attribute to illustrating the connection between quality scores and the argument structure, format, and wording. On out-ofdomain data, the best performance is obtained by the feedback-augmented dual BERT, which even outperforms using all augmentations. This is illustrated in Table 3, where the first two arguments receive positive feedback, with space for improvement by further elaboration or addressing of alternatives, directing SPARK to increase its score.\nThe third argument receives more critical feedback, causing SPARK to decrease its score. While GPT-3.5 is generally able to highlight the salient points of an argument and provide valid criticism, we also note an occasional bias of the model towards maintaining a neutral or positive argument stance (as in the case of the libertarianism argument).\nHuman judgment of augmentations (Q3). To validate the alignment of the augmentations with human utility, we performed a human study where we asked participants to score the validity, infor-mativeness, and relevance of each augmentation strategy. The participants were asked to score 50 randomly sampled in-domain data points on these three metrics using a Likert scale of 1 (lowest) to 5 (highest). The results in Table 2 show that the augmentations are perceived by people as highly valid, informative, and relevant. Assumptions and counter-arguments were found to be consistently more valid, informative, and relevant than the other augmentations. Curiously, the participants judged the feedback informativeness to be lower, explaining that it often summarizes the argument instead of giving writing suggestions. This finding provides a cue that highly effective augmentations for models may not be perceived as informative by people. SPARK alleviates this issue by effectively combining the complementary augmentations and delegating the weighting of their utility to the model."},{"figure_ref":[],"heading":"Conclusions","publication_ref":[],"table_ref":[],"text":"This paper enhances argument quality estimation models by providing contextualized feedback, inferred assumptions, and similar quality arguments or counter-arguments. We employ a dual-encoder Transformer to compare the argument and additional evidence effectively. Experimental results demonstrate that the best performance is achieved with the combination of all augmentations, indicating their complementary insights. SPARK outperforms single BERT, and XLNet in handling longer sequences, surpassing baselines and alternative augmentations in both in-and out-of-domain scenarios. Feedback augmentation is the single most effective augmentation strategy for models, despite being scored the least informative by human participants.\nThe limitations in this paper primarily stem from the generative capabilities and hallucination tendencies of the LLM used for augmentation. For example, despite constraining the output format for the assumption augmentation task, the LLM still generates \"No assumptions\" as output after listing a set of valid assumptions. As future work, further augmentation studies must be performed to analyze and improve the prompts to minimize the misunderstanding and biases of the LLM. Moreover, evaluating on a larger set of domains and other argumentative tasks such as logical fallacy detection is an important next step in investigating the generalizability of our method. While in theory, SPARK can be directly applied to such tasks, it remains to be seen to which extent the current architecture and augmentation strategies will generalize to other tasks.\nFinally, our analysis in this paper focuses on the comparison of methods and strategies, yet, we do not dive deep into the specific differences in performance across the three quality metrics which is an important future direction.\nTable 3 shows augmentation examples for one topic-argument pair. The topic questions if cell phones distract people and the argument agrees with it in context to distracted driving due to cell phones.\nFeedback: The feedback discusses how the argument takes a clear position but appears overly emotional while answering, which may influence the persuasiveness of the argument. Additionally, the feedback regarding lack of evidence other than personal experience and the lack of discussion on phones causing a distraction beyond driving aim to help improve understanding while ensuring that the Feedback -The argument takes a clear position that cell phones cause distraction while driving.\n-The writer expresses strong emotions regarding the topic, which may influence the persuasiveness of the argument.\n-The argument does not provide any evidence or supporting details beyond the writer's personal experience and emotions.\n-The argument does not address whether cell phones cause distractions in other contexts beyond driving. Assumptions -Cell phones cause a distraction when driving.\n-Bluetooth or speaker use of cell phones can also be a distraction while driving.\n-People don't take driving as seriously as they should. Similar quality argument Yes, cell phones can be a major distraction to people. People are constantly checking their phones for notifications, messages, and other tasks that can take away from what they should be focusing on. This can lead to people not paying attention in class, not being productive at work, and not being present in conversations. Cell phones are an amazing invention, but they can also be a major distraction."},{"figure_ref":[],"heading":"Counter-Argument","publication_ref":[],"table_ref":[],"text":"While it is true that distracted driving can be a major problem, it is unfair to blame cell phones as the sole cause of distraction. Many other factors can also contribute to distracted driving, such as eating, applying makeup, or even talking with passengers. Furthermore, cell phones can also be used in a responsible manner while driving, such as with the use of hands-free devices or by pulling over to take a call. Therefore, instead of demonizing cell phones, we should focus on educating drivers on safe driving practices and the dangers of distracted driving in general.\nTable 3: Augmentation examples for the topic Do cell phones cause a distraction to people? with argument: If you mean do they cause a distraction when driving the answer is a resounding YES!!! That includes bluetooth or speaker. Nothing makes me more angry when people don't take driving as seriously as they should. It is a major killer and if you don't follow the rules and focus you are putting other people in danger.\nEffectiveness Score Generated similar quality argument to: Do you equate legality with morality? 1.0 Yes, I think that if the law says something is wrong, then it is wrong. People should follow the law and not do anything illegal."},{"figure_ref":[],"heading":"2.0","publication_ref":[],"table_ref":[],"text":"No, I do not think that legality necessarily equates to morality. Laws can often be outdated, biased, or simply not reflective of what is right and wrong. For example, slavery used to be legal in many countries, yet it was obviously immoral. So legality and morality do not always align."},{"figure_ref":[],"heading":"3.0","publication_ref":[],"table_ref":[],"text":"No, I do not equate legality with morality. Just because something is legal does not make it moral, and just because something is illegal does not make it immoral. Laws are made by people and can be changed, while morality is often determined by an individual's personal values and beliefs. Therefore, legality should not be used as a measure of morality. 4.0\nNo, I do not equate legality with morality. I believe that there are laws that should be followed in order to maintain the peace and order of a society, but there are also moral standards that must be adhered to for the same purpose. For example, it may be legal in some places to drive over the speed limit, but it is still morally wrong to do so. 5.0 I do not necessarily equate legality with morality. Legality is based on laws, which can be changed to reflect the morality of a society. Morality, in contrast, is based on principles and values that don't necessarily have to be enforced by law. For example, while it may be legal to drive over the speed limit, it is not necessarily moral to do so. model pays careful attention to the topic in context."},{"figure_ref":[],"heading":"Assumptions:","publication_ref":[],"table_ref":["tab_2","tab_2"],"text":"The LLM lists several important assumptions in the proposed argument with respect to the topic. The first is the base assumption of the author's perspective, which is that cell phones cause distraction during driving. The LLM extracts the sentence \" people don't take driving as seriously as they should\" and labels it as an assumption because this is a faulty generalization to apply a general rule to all people, which is, in this case, people not taking driving seriously.\nSimilar quality argument: The generated similar quality argument tries to replicate the structural pattern of the given argument. The similar quality instance contains open-ended, long-winded sentences such as \"Cell phones are an amazing invention, but they can also be a major distraction\" which reduce the score of the argument, similar to the original argument. We also see that the LLM understands the ranking progression in a few shot setting and similar to the original argument, the similar quality argument also focuses solely on driving based cell phone distractions.\nCounter argument: We notice that the LLM recognizes that the original argument only discusses distracted driving and so it only produces a counter argument of the stance that distracted driving is not the only cause for distraction. The response discusses the safe use of cell phones such as hands free, etc and advocates educating drivers on the effects of cell phone based distracted driving.\nA.3 Impact of effectiveness score on GPT 3.5 outputs for similar quality arguments Table 4 discusses the impact of effectiveness score on the generated similar quality argument. As can be seen in Table 4, the generated argument with an effectiveness score of 1.0 oversimplifies the relationship between legality and morality and treats related laws as fixed. Comparatively, the argument with an effectiveness score of 2.0 provides an example of slavery which enhances the effectiveness of the argument. Despite the addition of this example, the second argument lacks elaboration on why the example is immoral and fails to provide relevant evidence. The argument generated, given an effectiveness score of 3.0, recognizes that the law is not the sole arbiter of morality and that laws are subject to change. It does not only highlight the potential flaws in legal systems but also addresses the distinction between personal values and the law. However, this argument oversimplifies morality by implying that personal values and beliefs solely determine morality and lacks supporting evidence for the statement: just because something is legal does not make it moral, and just because something is illegal does not make it immoral.. The argument generated with an effectiveness score of 4.0 considers the coexistence of legal and moral standards. The addition of a specific example in this argument adds concreteness and strengthens its persua-siveness. However, the argument can be further strengthened by acknowledging a broader range of situations where legality and morality may diverge. Finally, the argument ranked with the highest effectiveness score emphasizes the independence of morality from legal enforcement, which makes it even more persuasive. The contrasting comparison adds clarity to the flow of the argument and hence makes it better than all previously generated arguments."},{"figure_ref":["fig_2"],"heading":"A.4 Distribution analysis of augmentation lengths across splits","publication_ref":[],"table_ref":[],"text":"In this subsection, we conduct a distribution analysis on the augmentation input sizes to justify the use of the dual BERT architecture. Based on the findings presented in The validation split has a minimum tokenized length of 144 tokens, a maximum length of 703 tokens, and an average length of 318.55 tokens. The testing split, on the other hand, has a minimum length of 136 tokens, a maximum length of 613 tokens, and an average length of 324.60 tokens. The percentage of data points in the validation and testing splits as seen in Figure 3 that exceed BERT's Hence, we can conclude that the second BERT encoder tasked with embedding the augmentations is able to capture all the information in the augmentations without truncating the augmentations."},{"figure_ref":[],"heading":"A.5 Impact of feedback","publication_ref":[],"table_ref":["tab_6"],"text":"Table 6 shows the impact of providing feedback on the quality scores given by SPARK for three arguments.\nA.6 Questions for the human study Our human study poses the following three targeted questions to the participants:\n1. How valid is the information provided by the augmentation with respect to the background of the argument?\n2. How informative is the augmentation for the task of argument quality analysis?\n3. How relevant is the augmentation to help with the task of assessing the quality of the argument?"},{"figure_ref":[],"heading":"A Appendix","publication_ref":[],"table_ref":[],"text":"A.1 GPT 3.5 prompt templates"},{"figure_ref":[],"heading":"Feedback","publication_ref":[],"table_ref":[],"text":"The feedback on writing is sampled by considering both the topic and the argument related to the topic. To ensure brevity, we output the feedback in bullet point format. We follow the format below for sampling the feedback from the LLM:\nGive concise writing feedback for the following argument in context with the topic, preferably in bullet points: Topic: topic Argument: argument."},{"figure_ref":[],"heading":"Assumptions","publication_ref":[],"table_ref":[],"text":"Similar to feedback, assumptions are sampled in bullet point format to ensure brevity. Additionally, to constrain the hallucinations of the LLM, we restrict it to output \"No assumptions\" for the cases where it does not find assumptions or biases. We use the below prompt to sample this assumptions list:\nSummarize the assumptions, if any, in the following argument in a bullet format otherwise return \"No assumptions\" Topic: topic Argument: argument."},{"figure_ref":[],"heading":"Similar quality argument","publication_ref":[],"table_ref":[],"text":"To sample a similar quality argument, we use the following template:\nCogency Score: cogency score Effectiveness Score: effectiveness score Reasonableness Score: reasonableness score Topic: topic\nWe use ten samples in the few shot setting with two each from every integer ranking from 1-5 on the ranking scale for each metric. Finally, we prompt the LLM to generate the argument with respect to the cogency, effectiveness and robustness scores."},{"figure_ref":[],"heading":"Counter-argument","publication_ref":[],"table_ref":[],"text":"The counter-argument is generated using the given argument and topic, and the following template:\nGive a counter-argument for the following argument with respect to the Topic: topic Argument: argument"},{"figure_ref":[],"heading":"A.2 Augmentation examples","publication_ref":[],"table_ref":[],"text":""}],"string":"[\n {\n \"figure_ref\": [],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b21\",\n \"b3\",\n \"b26\",\n \"b32\",\n \"b2\",\n \"b8\",\n \"b20\",\n \"b27\",\n \"b12\",\n \"b25\",\n \"b24\",\n \"b23\",\n \"b14\",\n \"b11\",\n \"b19\",\n \"b19\",\n \"b11\",\n \"b11\",\n \"b9\",\n \"b13\",\n \"b30\",\n \"b17\",\n \"b20\",\n \"b6\",\n \"b7\",\n \"b22\",\n \"b15\",\n \"b27\",\n \"b12\",\n \"b12\"\n ],\n \"table_ref\": [],\n \"text\": \"Reliable analysis of arguments in natural language holds the promise to support applications such as automated grading (Ludwig et al., 2021), and tackling misinformation and targeted speech (Alhindi, Tariq, 2023). Computational argument analysis has been relatively popular through tasks like argument extraction (Chakrabarty et al., 2019), evidence mining (Rinott et al., 2015), relation assignment (Trautmann et al., 2020), writing support (Stab and Gurevych, 2014b) and claim generation (Bilu and Slonim, 2016). A particularly challenging task is argument quality assessment (Fromm et al., 2022), which addresses the cogency, effectiveness, and reasonableness of an argument (Wachsmuth et al., 2017b) pertaining to a topic. Assessing the quality of the argument involves analyzing the objective evidence, relevant assumptions, and structural soundness, making the overall task difficult.\\nResearch on argument quality assessment has focused on extracting textual patterns using various learning frameworks and content features (Lauscher et al., 2022). It has been widely recognized that contextualizing arguments with implicit knowledge, such as extracting claim revisions (Skitalinskaya et al., 2021) and generating explicit conclusions (Gurcke et al., 2021) can be informative for reasoning models. However, we note that: 1) these methods fail to generalize to novel arguments where this information is not available, and 2) no prior work has considered jointly a comprehensive set of such contextualization strategies.\\nTo bridge these gaps, we propose a novel framework called SPARK (Scoring the Pragmatics of Arguments via Relevant Knowledge), which incorporates augmentation strategies based on a large language model (LLM), GPT 3.5 (OpenAI, 2022), and elements from argumentation literature (Nickerson, 2020;Mulyati et al., 2023;Harvey, 2009), specifically, feedback, assumptions, arguments with similar quality, and counter-arguments. SPARK processes the original argument and topic and its augmentations separately using a dualencoder Transformer architecture with a multi-head cross-attention layer. We demonstrate the effectiveness of SPARK 's augmentations and architecture using both in-domain and out-of-domain datasets. Our entire code is made available at (anonymized).\\nTask formulation. Inspired by prior work (Gretz et al., 2020;Lauscher et al., 2020), we formalize argumentation quality assessment as a regression task of predicting the quality of a natural language argument. Given a topic and an argument, we consider three quality indicators (Lauscher et al., 2020): 1) cogency, which evaluates the relevance and sufficiency of the argument's premise in relation to the conclusion, 2) effectiveness, which measures the argument's persuasive power based on factors like arrangement, clarity, and appropriateness, and 3) reasonableness, which determines the argument's ability to resolve the debate's issue (Wachsmuth et al., 2017a). The overall quality of an argument can be estimated by averaging these three metrics (Gretz et al., 2020). Connection to prior studies. The introduction of benchmarks for argument quality (Stab and Gurevych, 2014a;Gretz et al., 2020) has inspired various methods based on logistic regression (Ghosh et al., 2016), fully connected and recurrent neural networks (Habernal and Gurevych, 2016), and fine-tuned Transformers (Toledo et al., 2019). Hulpus et al. (2019) explained that contextualizing an argument with implicit knowledge is essential to understanding its quality. Lauscher et al. (2022) categorized knowledge used by current argument assessment research so far under linguistic, task-specific, and argument-specific sections. To mimic human reasoning over arguments, prior work has incorporated users' prior beliefs as predictors of argument persuasiveness (Durmus and Cardie, 2018), trained classifiers for different audience groups (El Baff et al., 2020), utilized user history to predict persuasion (Al Khatib et al., 2020), augmented arguments with supporting or refuting documents (Marro et al., 2022), and augmented arguments with visual features (Hasan et al., 2021). Most similar to SPARK, Skitalinskaya et al. (2021) leverage comparison between revisions of the same claims, while Gurcke et al. (2021) generate conclusions to assess argument sufficiency. However, revisions are rarely provided for novel arguments, whereas generated conclusions are argument-specific and may not generalize well (Gurcke et al., 2021). Addressing prior work limitations, SPARK implements four wellmotivated augmentation strategies to enhance novel arguments and utilizes an attention-based dual encoder model for effective reasoning.\"\n },\n {\n \"figure_ref\": [\n \"fig_0\",\n \"fig_0\",\n \"fig_0\",\n \"fig_0\",\n \"fig_1\"\n ],\n \"heading\": \"SPARK\",\n \"publication_ref\": [\n \"b23\",\n \"b24\",\n \"b27\",\n \"b16\",\n \"b14\",\n \"b10\",\n \"b33\"\n ],\n \"table_ref\": [],\n \"text\": \"Augmentation strategies. We devise four augmentation techniques to contextualize arguments. We generate the augmentations by prompting GPT-3.5 (OpenAI, 2022) (see appendix for details). Feedback. Constructive feedback in the form of comments and suggestions helps comprehension and domain knowledge acquisition (Mulyati et al., 2023). We hypothesize that assessing argument strengths and weaknesses helps argument ranking, as in Figure 1, where feedback identifies emotional appeal, insufficient evidence, and generalization. We prompt the LLM to generate writing feedback for a topic-argument pair in a zero-shot setting. Assumptions. Unstated assumptions frequently introduce bias in arguments (Nickerson, 2020). Making assumptions explicit can reveal these hidden biases, which may aid in assessing argument persuasiveness and relevance. One such assumption in Figure 1 is that people do not take driving seriously. We employ an LLM to extract the argument's underlying assumptions in a zero-shot setting. Similar-quality instance. Inspired by prior work on claim revisions (Skitalinskaya et al., 2021), we hypothesize that retrieving arguments with similar quality at training time leads to generalizable model learning. For this purpose, we derive a synthetic argument with similar reasonableness, cogency, and effectiveness to the original one (Figure 1). We generate this synthetic argument in a few-shot setting, where the LLM has access to example arguments alongside their quality scores covering the full 1-5 range. Since this augmentation uses ground-truth information that is not available during inference, we randomly replace synthetic arguments with None at training time with a probability of P = 0.5, thus familiarizing the model with the absence of similar arguments during testing where it only sees None. This technique is similar to distillation (Hinton et al., 2015), where the encoder's (student) goal is to utilize the LLM's (teacher) ranking-based argument generations (soft labels) to learn the argument quality ranking task. Counter-arguments. Counter-arguments provide objections, alternatives, and doubts of a skeptical reader (Harvey, 2009). We expect that contrasting the strengths and weaknesses of two opposite arguments will aid quality assessment. The example counter-argument in Figure 1 makes a firm claim that alternative activities (e.g., eating) lead to distractions, and solely blaming phones is unfair. We ask the LLM to provide a counter-argument for a topic and an argument in a zero-shot setting. Dual-encoder architecture. To consider the argument together with the augmentations, we employ a dual BERT encoder (Figure 2) as an improvement to the architecture by Gillick et al. (2018). The first encoder embeds the topic and argument, whereas the second embeds the augmentations. The second encoder can store individual augmentations or their concatenation, arbitrarily fixed to Similar quality argument [SEP] Feedback [SEP] Assumptions [SEP] Counter-argument. Notably, the dual encoder can effectively store all of the augmentation data without truncating information in practically all cases (see subsection A.4 for the augmentation lengths). We use a multi-head crossattention layer (Vaswani et al., 2017) to enable the model to weigh each augmentation according to the argument-topic pair. We pass the attention outputs to a mean pooler, whose output is fed into three separate regressor heads, one per quality metric.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiments\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Baselines\",\n \"publication_ref\": [\n \"b5\",\n \"b37\",\n \"b18\",\n \"b4\",\n \"b31\"\n ],\n \"table_ref\": [],\n \"text\": \"Scoring models. We compare our dual BERT with a standard BERT model to study the effect of disjoint embeddings. We use bert-base-uncased (Devlin et al., 2019) as an encoder. We compare dual BERT to XLNet (Yang et al., 2019), as they can both handle more than 512 tokens. Finally, we utilize GPT-3.5 in a zero-shot setting to gauge the model's ability to accomplish the task directly, without a dual encoder. We provide GPT-3.5 with the definitions of each metric and prompt it to individually rate each argument by a float between 1 and 5 with respect to the topic. Alternative augmentation strategies. We evaluate the impact of using all augmentations together or one at a time, against ablated baselines without augmentations. We include two alternative augmentation methods: Wikipedia paragraphs extracted using dense passage retrieval (DPR) (Karpukhin et al., 2020), and augmentations generated using smaller models, namely, Flan-T5-XL (Chung et al., 2022) and Llama-2 (7B) (Touvron et al., 2023).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Datasets and Evaluation\",\n \"publication_ref\": [\n \"b19\",\n \"b11\",\n \"b19\"\n ],\n \"table_ref\": [],\n \"text\": \"We use GAQCorpus (Lauscher et al., 2020) as our training dataset for its diversity of domains (reviews, QA, and debates) and quality metrics (cogency, effectiveness, and reasonableness). We also use IBM-30K (Gretz et al., 2020) to test SPARK 's generalization on out-of-domain data. For IBM-30K, we perform weighted averaging (WA) of the three metric scores, which is supported by the high correlation between IBM-30K's WA and the GAQ-Corpus metrics (Lauscher et al., 2020). We report the Pearson (ρ) and Spearman (σ) correlation coefficients between the predictions and ground truth.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Results\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_0\"\n ],\n \"text\": \"We investigate whether augmentations help language models assess the quality of arguments more effectively (Q1); how augmentation strategies compare to each other (Q2); and whether human quality judgments align with their model utility (Q3).\\nEffect of augmentation on in-domain performance (Q1). Table 1 shows that the overall best-performing combination uses Dual BERT with all four augmentations combined, which improves the Spearman correlation over the baseline BERT by 0.08-0.17 across the three metrics. The improvement is the largest for effectiveness, where the Spearman correlation increases by 61%. While both single BERT and XLNet benefit from SPARK 's augmentations as well, their performance is consistently lower than using the dual encoder. GPT-3 alone often performs better than the other baselines but lags significantly behind SPARK, showing the importance of the dual encoder. Among the augmentations, the benefit of our four augmentations declines when using a smaller generative model (Flan T5 and Llama-2), augmenting via DPR, or using the dual BERT with a masked second encoder. Thus, merely adding text to the second encoder does not by itself bring higher performance. The gap between SPARK and the baselines increases in the zero-shot setting on the IBM-Rank-30K dataset. Here, augmentation with DPR, Flan-T5, and Llama-2 is consistently inferior, as is the SPARK augmentation of the singleencoder methods. In summary, SPARK effectively combines dual encoding and data augmentation for strong task accuracy and generalization.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Comparison of augmentation variants (Q2).\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_1\"\n ],\n \"text\": \"On the in-domain task, the dual encoder performs the best when it has access to the information from the four SPARK augmentations simultaneously. Among them, feedback is most effective for predicting argument cogency and reasonableness as it exposes flaws that directly relate to these metrics. Meanwhile, contextualizing through similar-quality arguments is optimal for predicting effectiveness, which we attribute to illustrating the connection between quality scores and the argument structure, format, and wording. On out-ofdomain data, the best performance is obtained by the feedback-augmented dual BERT, which even outperforms using all augmentations. This is illustrated in Table 3, where the first two arguments receive positive feedback, with space for improvement by further elaboration or addressing of alternatives, directing SPARK to increase its score.\\nThe third argument receives more critical feedback, causing SPARK to decrease its score. While GPT-3.5 is generally able to highlight the salient points of an argument and provide valid criticism, we also note an occasional bias of the model towards maintaining a neutral or positive argument stance (as in the case of the libertarianism argument).\\nHuman judgment of augmentations (Q3). To validate the alignment of the augmentations with human utility, we performed a human study where we asked participants to score the validity, infor-mativeness, and relevance of each augmentation strategy. The participants were asked to score 50 randomly sampled in-domain data points on these three metrics using a Likert scale of 1 (lowest) to 5 (highest). The results in Table 2 show that the augmentations are perceived by people as highly valid, informative, and relevant. Assumptions and counter-arguments were found to be consistently more valid, informative, and relevant than the other augmentations. Curiously, the participants judged the feedback informativeness to be lower, explaining that it often summarizes the argument instead of giving writing suggestions. This finding provides a cue that highly effective augmentations for models may not be perceived as informative by people. SPARK alleviates this issue by effectively combining the complementary augmentations and delegating the weighting of their utility to the model.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusions\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This paper enhances argument quality estimation models by providing contextualized feedback, inferred assumptions, and similar quality arguments or counter-arguments. We employ a dual-encoder Transformer to compare the argument and additional evidence effectively. Experimental results demonstrate that the best performance is achieved with the combination of all augmentations, indicating their complementary insights. SPARK outperforms single BERT, and XLNet in handling longer sequences, surpassing baselines and alternative augmentations in both in-and out-of-domain scenarios. Feedback augmentation is the single most effective augmentation strategy for models, despite being scored the least informative by human participants.\\nThe limitations in this paper primarily stem from the generative capabilities and hallucination tendencies of the LLM used for augmentation. For example, despite constraining the output format for the assumption augmentation task, the LLM still generates \\\"No assumptions\\\" as output after listing a set of valid assumptions. As future work, further augmentation studies must be performed to analyze and improve the prompts to minimize the misunderstanding and biases of the LLM. Moreover, evaluating on a larger set of domains and other argumentative tasks such as logical fallacy detection is an important next step in investigating the generalizability of our method. While in theory, SPARK can be directly applied to such tasks, it remains to be seen to which extent the current architecture and augmentation strategies will generalize to other tasks.\\nFinally, our analysis in this paper focuses on the comparison of methods and strategies, yet, we do not dive deep into the specific differences in performance across the three quality metrics which is an important future direction.\\nTable 3 shows augmentation examples for one topic-argument pair. The topic questions if cell phones distract people and the argument agrees with it in context to distracted driving due to cell phones.\\nFeedback: The feedback discusses how the argument takes a clear position but appears overly emotional while answering, which may influence the persuasiveness of the argument. Additionally, the feedback regarding lack of evidence other than personal experience and the lack of discussion on phones causing a distraction beyond driving aim to help improve understanding while ensuring that the Feedback -The argument takes a clear position that cell phones cause distraction while driving.\\n-The writer expresses strong emotions regarding the topic, which may influence the persuasiveness of the argument.\\n-The argument does not provide any evidence or supporting details beyond the writer's personal experience and emotions.\\n-The argument does not address whether cell phones cause distractions in other contexts beyond driving. Assumptions -Cell phones cause a distraction when driving.\\n-Bluetooth or speaker use of cell phones can also be a distraction while driving.\\n-People don't take driving as seriously as they should. Similar quality argument Yes, cell phones can be a major distraction to people. People are constantly checking their phones for notifications, messages, and other tasks that can take away from what they should be focusing on. This can lead to people not paying attention in class, not being productive at work, and not being present in conversations. Cell phones are an amazing invention, but they can also be a major distraction.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Counter-Argument\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"While it is true that distracted driving can be a major problem, it is unfair to blame cell phones as the sole cause of distraction. Many other factors can also contribute to distracted driving, such as eating, applying makeup, or even talking with passengers. Furthermore, cell phones can also be used in a responsible manner while driving, such as with the use of hands-free devices or by pulling over to take a call. Therefore, instead of demonizing cell phones, we should focus on educating drivers on safe driving practices and the dangers of distracted driving in general.\\nTable 3: Augmentation examples for the topic Do cell phones cause a distraction to people? with argument: If you mean do they cause a distraction when driving the answer is a resounding YES!!! That includes bluetooth or speaker. Nothing makes me more angry when people don't take driving as seriously as they should. It is a major killer and if you don't follow the rules and focus you are putting other people in danger.\\nEffectiveness Score Generated similar quality argument to: Do you equate legality with morality? 1.0 Yes, I think that if the law says something is wrong, then it is wrong. People should follow the law and not do anything illegal.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"2.0\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"No, I do not think that legality necessarily equates to morality. Laws can often be outdated, biased, or simply not reflective of what is right and wrong. For example, slavery used to be legal in many countries, yet it was obviously immoral. So legality and morality do not always align.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"3.0\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"No, I do not equate legality with morality. Just because something is legal does not make it moral, and just because something is illegal does not make it immoral. Laws are made by people and can be changed, while morality is often determined by an individual's personal values and beliefs. Therefore, legality should not be used as a measure of morality. 4.0\\nNo, I do not equate legality with morality. I believe that there are laws that should be followed in order to maintain the peace and order of a society, but there are also moral standards that must be adhered to for the same purpose. For example, it may be legal in some places to drive over the speed limit, but it is still morally wrong to do so. 5.0 I do not necessarily equate legality with morality. Legality is based on laws, which can be changed to reflect the morality of a society. Morality, in contrast, is based on principles and values that don't necessarily have to be enforced by law. For example, while it may be legal to drive over the speed limit, it is not necessarily moral to do so. model pays careful attention to the topic in context.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Assumptions:\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_2\",\n \"tab_2\"\n ],\n \"text\": \"The LLM lists several important assumptions in the proposed argument with respect to the topic. The first is the base assumption of the author's perspective, which is that cell phones cause distraction during driving. The LLM extracts the sentence \\\" people don't take driving as seriously as they should\\\" and labels it as an assumption because this is a faulty generalization to apply a general rule to all people, which is, in this case, people not taking driving seriously.\\nSimilar quality argument: The generated similar quality argument tries to replicate the structural pattern of the given argument. The similar quality instance contains open-ended, long-winded sentences such as \\\"Cell phones are an amazing invention, but they can also be a major distraction\\\" which reduce the score of the argument, similar to the original argument. We also see that the LLM understands the ranking progression in a few shot setting and similar to the original argument, the similar quality argument also focuses solely on driving based cell phone distractions.\\nCounter argument: We notice that the LLM recognizes that the original argument only discusses distracted driving and so it only produces a counter argument of the stance that distracted driving is not the only cause for distraction. The response discusses the safe use of cell phones such as hands free, etc and advocates educating drivers on the effects of cell phone based distracted driving.\\nA.3 Impact of effectiveness score on GPT 3.5 outputs for similar quality arguments Table 4 discusses the impact of effectiveness score on the generated similar quality argument. As can be seen in Table 4, the generated argument with an effectiveness score of 1.0 oversimplifies the relationship between legality and morality and treats related laws as fixed. Comparatively, the argument with an effectiveness score of 2.0 provides an example of slavery which enhances the effectiveness of the argument. Despite the addition of this example, the second argument lacks elaboration on why the example is immoral and fails to provide relevant evidence. The argument generated, given an effectiveness score of 3.0, recognizes that the law is not the sole arbiter of morality and that laws are subject to change. It does not only highlight the potential flaws in legal systems but also addresses the distinction between personal values and the law. However, this argument oversimplifies morality by implying that personal values and beliefs solely determine morality and lacks supporting evidence for the statement: just because something is legal does not make it moral, and just because something is illegal does not make it immoral.. The argument generated with an effectiveness score of 4.0 considers the coexistence of legal and moral standards. The addition of a specific example in this argument adds concreteness and strengthens its persua-siveness. However, the argument can be further strengthened by acknowledging a broader range of situations where legality and morality may diverge. Finally, the argument ranked with the highest effectiveness score emphasizes the independence of morality from legal enforcement, which makes it even more persuasive. The contrasting comparison adds clarity to the flow of the argument and hence makes it better than all previously generated arguments.\"\n },\n {\n \"figure_ref\": [\n \"fig_2\"\n ],\n \"heading\": \"A.4 Distribution analysis of augmentation lengths across splits\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this subsection, we conduct a distribution analysis on the augmentation input sizes to justify the use of the dual BERT architecture. Based on the findings presented in The validation split has a minimum tokenized length of 144 tokens, a maximum length of 703 tokens, and an average length of 318.55 tokens. The testing split, on the other hand, has a minimum length of 136 tokens, a maximum length of 613 tokens, and an average length of 324.60 tokens. The percentage of data points in the validation and testing splits as seen in Figure 3 that exceed BERT's Hence, we can conclude that the second BERT encoder tasked with embedding the augmentations is able to capture all the information in the augmentations without truncating the augmentations.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.5 Impact of feedback\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_6\"\n ],\n \"text\": \"Table 6 shows the impact of providing feedback on the quality scores given by SPARK for three arguments.\\nA.6 Questions for the human study Our human study poses the following three targeted questions to the participants:\\n1. How valid is the information provided by the augmentation with respect to the background of the argument?\\n2. How informative is the augmentation for the task of argument quality analysis?\\n3. How relevant is the augmentation to help with the task of assessing the quality of the argument?\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A Appendix\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"A.1 GPT 3.5 prompt templates\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Feedback\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The feedback on writing is sampled by considering both the topic and the argument related to the topic. To ensure brevity, we output the feedback in bullet point format. We follow the format below for sampling the feedback from the LLM:\\nGive concise writing feedback for the following argument in context with the topic, preferably in bullet points: Topic: topic Argument: argument.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Assumptions\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Similar to feedback, assumptions are sampled in bullet point format to ensure brevity. Additionally, to constrain the hallucinations of the LLM, we restrict it to output \\\"No assumptions\\\" for the cases where it does not find assumptions or biases. We use the below prompt to sample this assumptions list:\\nSummarize the assumptions, if any, in the following argument in a bullet format otherwise return \\\"No assumptions\\\" Topic: topic Argument: argument.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Similar quality argument\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"To sample a similar quality argument, we use the following template:\\nCogency Score: cogency score Effectiveness Score: effectiveness score Reasonableness Score: reasonableness score Topic: topic\\nWe use ten samples in the few shot setting with two each from every integer ranking from 1-5 on the ranking scale for each metric. Finally, we prompt the LLM to generate the argument with respect to the cogency, effectiveness and robustness scores.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Counter-argument\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The counter-argument is generated using the given argument and topic, and the following template:\\nGive a counter-argument for the following argument with respect to the Topic: topic Argument: argument\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.2 Augmentation examples\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n }\n]"},"abstract":{"kind":"string","value":"Automatic assessment of the quality of arguments has been recognized as a challenging task with significant implications for misinformation and targeted speech. While real-world arguments are tightly anchored in context, existing computational methods analyze their quality in isolation, which affects their accuracy and generalizability. We propose SPARK : a novel method for scoring argument quality based on contextualization via relevant knowledge. We devise four augmentations that leverage large language models to provide feedback, infer hidden assumptions, supply a similar-quality argument, or give a counter-argument. SPARK uses a dual-encoder Transformer architecture to enable the original argument and its augmentation to be considered jointly. Our experiments in both in-domain and zero-shot setups show that SPARK consistently outperforms existing techniques across multiple metrics."},"title":{"kind":"string","value":"Contextualizing Argument Quality Assessment with Relevant Knowledge"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 1 :1Figure 1: Overview of SPARK.","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"Figure 2 :2Figure 2: Dual BERT encoder architecture.","figure_data":"","figure_id":"fig_1","figure_label":"2","figure_type":"figure"},{"figure_caption":"Figure 3 :3Figure 3: Distributions for augmentation lengths for the training, validation and testing splits respectively.","figure_data":"","figure_id":"fig_2","figure_label":"3","figure_type":"figure"},{"figure_caption":"Performance of Dual-BERT model with augmentations applied compared to the baseline models. The performance of the model achieving the best scores per metric is boldfaced and the second best score is underlined.","figure_data":"GAQ Corpus (in-domain)IBM-30K (ZS)","figure_id":"tab_0","figure_label":"1","figure_type":"table"},{"figure_caption":"Validity, informativeness, and relevance scores of the augmentations for argument scoring.","figure_data":"","figure_id":"tab_1","figure_label":"2","figure_type":"table"},{"figure_caption":"Similar quality examples.","figure_data":"","figure_id":"tab_2","figure_label":"4","figure_type":"table"},{"figure_caption":"Distribution properties across splits for the concatenated augmentations.","figure_data":"","figure_id":"tab_4","figure_label":"5","figure_type":"table"},{"figure_caption":"","figure_data":", the training split ex-","figure_id":"tab_5","figure_label":"5","figure_type":"table"},{"figure_caption":"Feedback examples for the out-of-domain dataset.","figure_data":"TopicArgumentScoreScoreGroundFeedback augmentationbe-aftertruthforeaug-scoreaug-men-men-tationtationThe use of publicA centralized system of0.6280.6750.828 -Clear and concise argument presenteddefenders shouldcriminal defense would-Supports the idea of a centralized system ofbe mandatorymean that all people wouldcriminal defensehave access to the same-Highlights the importance of equal accessstandard of legal counsel,to legal counselmeaning that wealth and-Addresses the issue of wealth and powerpower can't be used to avoidinfluencing justicejustice.-Could benefit from further elaboration orevidence to strengthen the argumentWe should ban al-algorithmic trading has been0.5850.6310.948 -Clear and concise argument presentedgorithmic tradingresponsible for several mini-Provides specific examples to support argu-market collapses, since com-mentputer systems lack the hu--Could benefit from further elaboration onman sensitivity to look out-the potential consequences of mini marketside the stream of meaning-collapsesless numbers to a wider con--Could also benefit from addressing potentialtext.counterarguments or alternative solutions tothe issueWe should adoptlibertarianism is a justifica-0.5400.5190.666The argument is not specific enough aboutlibertarianismtion for greed and exploita-what adopting libertarianism entailstion-It assumes that libertarianism automaticallyleads to working together for a greater goodand favoring the less well off, which is notnecessarily true-The argument could benefit from provid-ing concrete examples of how libertarianismwould benefit society as a whole-It is unclear how giving freedom of choicewould lead to greater societal benefits, andthis point could be expanded upon.token limit are only 1.1% (13 examples) and 1.31%(15 examples), respectively.","figure_id":"tab_6","figure_label":"6","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 1 :1Figure 1: Overview of SPARK.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 2 :2Figure 2: Dual BERT encoder architecture.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 :3Figure 3: Distributions for augmentation lengths for the training, validation and testing splits respectively.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Performance of Dual-BERT model with augmentations applied compared to the baseline models. The performance of the model achieving the best scores per metric is boldfaced and the second best score is underlined.\",\n \"figure_data\": \"GAQ Corpus (in-domain)IBM-30K (ZS)\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Validity, informativeness, and relevance scores of the augmentations for argument scoring.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Similar quality examples.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Distribution properties across splits for the concatenated augmentations.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"5\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"\",\n \"figure_data\": \", the training split ex-\",\n \"figure_id\": \"tab_5\",\n \"figure_label\": \"5\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Feedback examples for the out-of-domain dataset.\",\n \"figure_data\": \"TopicArgumentScoreScoreGroundFeedback augmentationbe-aftertruthforeaug-scoreaug-men-men-tationtationThe use of publicA centralized system of0.6280.6750.828 -Clear and concise argument presenteddefenders shouldcriminal defense would-Supports the idea of a centralized system ofbe mandatorymean that all people wouldcriminal defensehave access to the same-Highlights the importance of equal accessstandard of legal counsel,to legal counselmeaning that wealth and-Addresses the issue of wealth and powerpower can't be used to avoidinfluencing justicejustice.-Could benefit from further elaboration orevidence to strengthen the argumentWe should ban al-algorithmic trading has been0.5850.6310.948 -Clear and concise argument presentedgorithmic tradingresponsible for several mini-Provides specific examples to support argu-market collapses, since com-mentputer systems lack the hu--Could benefit from further elaboration onman sensitivity to look out-the potential consequences of mini marketside the stream of meaning-collapsesless numbers to a wider con--Could also benefit from addressing potentialtext.counterarguments or alternative solutions tothe issueWe should adoptlibertarianism is a justifica-0.5400.5190.666The argument is not specific enough aboutlibertarianismtion for greed and exploita-what adopting libertarianism entailstion-It assumes that libertarianism automaticallyleads to working together for a greater goodand favoring the less well off, which is notnecessarily true-The argument could benefit from provid-ing concrete examples of how libertarianismwould benefit society as a whole-It is unclear how giving freedom of choicewould lead to greater societal benefits, andthis point could be expanded upon.token limit are only 1.1% (13 examples) and 1.31%(15 examples), respectively.\",\n \"figure_id\": \"tab_6\",\n \"figure_label\": \"6\",\n \"figure_type\": \"table\"\n }\n]"},"authors":{"kind":"string","value":"Darshan Deshpande; Zhivar Sourati; Filip Ilievski; Fred Morstatter"},"references":{"kind":"list like","value":[{"authors":"Al Khalid; Michael Khatib; Shahbaz Völske; Nikolay Syed; Benno Kolyada; Stein","journal":"Association for Computational Linguistics","ref_id":"b0","title":"Exploiting personal characteristics of debaters for predicting persuasiveness","year":"2020"},{"authors":" Alhindi","journal":"","ref_id":"b1","title":"Computational models of argument structure and argument quality for understanding misinformation","year":"2023"},{"authors":"Yonatan Bilu; Noam Slonim","journal":"Association for Computational Linguistics","ref_id":"b2","title":"Claim synthesis via predicate recycling","year":"2016"},{"authors":"Tuhin Chakrabarty; Christopher Hidey; Smaranda Muresan; Kathy Mckeown; Alyssa Hwang","journal":"","ref_id":"b3","title":"AMPERSAND: Argument mining for PERSuAsive oNline discussions","year":"2019"},{"authors":"Chung Hyung Won; Le Hou; Shayne Longpre; Barret Zoph; Yi Tay; William Fedus; Eric Li; Xuezhi Wang; Mostafa Dehghani; Siddhartha Brahma","journal":"","ref_id":"b4","title":"Scaling instruction-finetuned language models","year":"2022"},{"authors":"Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova","journal":"Association for Computational Linguistics","ref_id":"b5","title":"BERT: Pre-training of deep bidirectional transformers for language understanding","year":"2019"},{"authors":"Esin Durmus; Claire Cardie","journal":"Association for Computational Linguistics","ref_id":"b6","title":"Exploring the role of prior beliefs for argument persuasion","year":"2018"},{"authors":"Roxanne El Baff; Henning Wachsmuth; Khalid ; Al Khatib; Benno Stein","journal":"Association for Computational Linguistics","ref_id":"b7","title":"Analyzing the Persuasive Effect of Style in News Editorial Argumentation","year":"2020"},{"authors":"Michael Fromm; Max Berrendorf; Johanna Reiml; Isabelle Mayerhofer; Siddharth Bhargava; Evgeniy Faerman; Thomas Seidl","journal":"","ref_id":"b8","title":"Towards a holistic view on argument quality prediction","year":"2022"},{"authors":"Debanjan Ghosh; Aquila Khanam; Yubo Han; Smaranda Muresan","journal":"Association for Computational Linguistics","ref_id":"b9","title":"Coarse-grained argumentation features for scoring persuasive essays","year":"2016"},{"authors":"Daniel Gillick; Alessandro Presta; Gaurav Singh Tomar","journal":"","ref_id":"b10","title":"End-to-end retrieval in continuous space","year":"2018"},{"authors":"Shai Gretz; Roni Friedman; Edo Cohen-Karlik; Assaf Toledo; Dan Lahav; Ranit Aharonov; Noam Slonim","journal":"","ref_id":"b11","title":"A large-scale dataset for argument quality ranking: Construction and analysis","year":"2020"},{"authors":"Timon Gurcke; Milad Alshomary; Henning Wachsmuth","journal":"","ref_id":"b12","title":"Assessing the sufficiency of arguments through conclusion generation","year":"2021"},{"authors":"Ivan Habernal; Iryna Gurevych","journal":"Association for Computational Linguistics","ref_id":"b13","title":"What makes a convincing argument? empirical analysis and detecting attributes of convincingness in web argumentation","year":"2016"},{"authors":"Gordon Harvey","journal":"","ref_id":"b14","title":"A brief guide to the elements of the academic essay","year":"2009"},{"authors":"Md Kamrul Hasan; James Spann; Masum Hasan; Md Saiful Islam; Kurtis Haut; Rada Mihalcea; Ehsan Hoque","journal":"Association for Computational Linguistics","ref_id":"b15","title":"Hitting your MARQ: Multimodal ARgument quality assessment in long debate video","year":"2021"},{"authors":"Geoffrey Hinton; Oriol Vinyals; Jeff Dean","journal":"","ref_id":"b16","title":"Distilling the knowledge in a neural network","year":"2015"},{"authors":"Ioana Hulpus; Jonathan Kobbe; Christian Meilicke; Heiner Stuckenschmidt; Maria Becker; Juri Opitz; Vivi Nastase; Anette Frank","journal":"","ref_id":"b17","title":"Towards explaining natural language arguments with background knowledge","year":"2019"},{"authors":"Vladimir Karpukhin; Barlas Oguz; Sewon Min; Patrick Lewis; Ledell Wu; Sergey Edunov; Danqi Chen; Wen-Tau Yih","journal":"Association for Computational Linguistics","ref_id":"b18","title":"Dense passage retrieval for opendomain question answering","year":"2020"},{"authors":"Anne Lauscher; Lily Ng; Courtney Napoles; Joel Tetreault","journal":"International Committee on Computational Linguistics","ref_id":"b19","title":"Rhetoric, logic, and dialectic: Advancing theory-based argument quality assessment in natural language processing","year":"2020"},{"authors":"Anne Lauscher; Henning Wachsmuth; Iryna Gurevych; Goran Glavaš","journal":"Transactions of the Association for Computational Linguistics","ref_id":"b20","title":"Scientia Potentia Est-On the Role of Knowledge in Computational Argumentation","year":"2022"},{"authors":"Sabrina Ludwig; Christian Mayer; Christopher Hansen; Kerstin Eilers; Steffen Brandt","journal":"Psych","ref_id":"b21","title":"Automated essay scoring using transformer models","year":"2021"},{"authors":"Santiago Marro; Elena Cabrio; Serena Villata","journal":"Association for Computational Linguistics","ref_id":"b22","title":"Graph embeddings for argumentation quality assessment","year":"2022"},{"authors":"Yeti Mulyati; Daris Hadianto","journal":"International Journal of Instruction","ref_id":"b23","title":"Enhancing argumentative writing via online peer feedbackbased essay: A quasi-experiment study","year":"2023"},{"authors":"Raymond S Nickerson","journal":"Cambridge University Press","ref_id":"b24","title":"Biases, Misconceptions, and the Like","year":"2020"},{"authors":" Openai","journal":"","ref_id":"b25","title":"Chatgpt","year":"2022-04-30"},{"authors":"Ruty Rinott; Lena Dankin; Carlos Alzate Perez; Mitesh M Khapra; Ehud Aharoni; Noam Slonim","journal":"Association for Computational Linguistics","ref_id":"b26","title":"Show me your evidence -an automatic method for context dependent evidence detection","year":"2015"},{"authors":"Gabriella Skitalinskaya; Jonas Klaff; Henning Wachsmuth","journal":"","ref_id":"b27","title":"Learning from revisions: Quality assessment of claims in argumentation at scale","year":"2021"},{"authors":"Christian Stab; Iryna Gurevych","journal":"Dublin City University and Association for Computational Linguistics","ref_id":"b28","title":"a. Annotating argument components and relations in persuasive essays","year":"2014"},{"authors":"Christian Stab; Iryna Gurevych","journal":"Association for Computational Linguistics","ref_id":"b29","title":"Identifying argumentative discourse structures in persuasive essays","year":"2014"},{"authors":"Assaf Toledo; Shai Gretz; Edo Cohen-Karlik; Roni Friedman; Elad Venezian; Dan Lahav; Michal Jacovi; Ranit Aharonov; Noam Slonim","journal":"Association for Computational Linguistics","ref_id":"b30","title":"Automatic argument quality assessment -new datasets and methods","year":"2019"},{"authors":"Hugo Touvron; Louis Martin; Kevin Stone; Peter Albert; Amjad Almahairi; Yasmine Babaei; Nikolay Bashlykov; Soumya Batra; Prajjwal Bhargava; Shruti Bhosale; Dan Bikel; Lukas Blecher; Cristian Canton Ferrer; Moya Chen; Guillem Cucurull; David Esiobu; Jude Fernandes; Jeremy Fu; Wenyin Fu; Brian Fuller; Cynthia Gao; Vedanuj Goswami; Naman Goyal; Anthony Hartshorn; Saghar Hosseini; Rui Hou; Hakan Inan; Marcin Kardas; Viktor Kerkez; Madian Khabsa; Isabel Kloumann; Artem Korenev; Punit Singh Koura; Marie-Anne Lachaux; Thibaut Lavril; Jenya Lee; Diana Liskovich; Yinghai Lu; Yuning Mao; Xavier Martinet; Todor Mihaylov; Pushkar Mishra; Igor Molybog; Yixin Nie; Andrew Poulton; Jeremy Reizenstein; Rashi Rungta; Kalyan Saladi; Alan Schelten; Ruan Silva; Eric Michael Smith; Ranjan Subramanian; Ellen Xiaoqing; Binh Tan; Ross Tang; Adina Taylor; Jian Williams; Puxin Xiang Kuan; Zheng Xu; Iliyan Yan; Yuchen Zarov; Angela Zhang; Melanie Fan; Sharan Kambadur; Aurelien Narang; Robert Rodriguez; Sergey Stojnic; Thomas Edunov; Scialom","journal":"","ref_id":"b31","title":"Llama 2: Open foundation and finetuned chat models","year":"2023"},{"authors":"Dietrich Trautmann; Michael Fromm; Thomas Volker Tresp; Hinrich Seidl; Schütze","journal":"Datenbank-Spektrum","ref_id":"b32","title":"Relational and fine-grained argument mining","year":"2020"},{"authors":"Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Ł Ukasz Kaiser; Illia Polosukhin","journal":"Advances in Neural Information Processing Systems","ref_id":"b33","title":"Attention is all you need","year":"2017"},{"authors":"Curran Associates; Inc ","journal":"","ref_id":"b34","title":"","year":""},{"authors":"Henning Wachsmuth; Nona Naderi; Ivan Habernal; Yufang Hou; Graeme Hirst; Iryna Gurevych; Benno Stein","journal":"Association for Computational Linguistics","ref_id":"b35","title":"Argumentation quality assessment: Theory vs. practice","year":"2017"},{"authors":"Henning Wachsmuth; Nona Naderi; Yufang Hou; Yonatan Bilu; Tim Alberdingk Vinodkumar Prabhakaran; Graeme Thijm; Benno Hirst; Stein","journal":"Association for Computational Linguistics","ref_id":"b36","title":"Computational argumentation quality assessment in natural language","year":"2017"},{"authors":"Zhilin Yang; Zihang Dai; Yiming Yang; Jaime Carbonell; Russ R Salakhutdinov; Quoc V Le","journal":"","ref_id":"b37","title":"Xlnet: Generalized autoregressive pretraining for language understanding","year":"2019"}],"string":"[\n {\n \"authors\": \"Al Khalid; Michael Khatib; Shahbaz Völske; Nikolay Syed; Benno Kolyada; Stein\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b0\",\n \"title\": \"Exploiting personal characteristics of debaters for predicting persuasiveness\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \" Alhindi\",\n \"journal\": \"\",\n \"ref_id\": \"b1\",\n \"title\": \"Computational models of argument structure and argument quality for understanding misinformation\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Yonatan Bilu; Noam Slonim\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b2\",\n \"title\": \"Claim synthesis via predicate recycling\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Tuhin Chakrabarty; Christopher Hidey; Smaranda Muresan; Kathy Mckeown; Alyssa Hwang\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"AMPERSAND: Argument mining for PERSuAsive oNline discussions\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Chung Hyung Won; Le Hou; Shayne Longpre; Barret Zoph; Yi Tay; William Fedus; Eric Li; Xuezhi Wang; Mostafa Dehghani; Siddhartha Brahma\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"Scaling instruction-finetuned language models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b5\",\n \"title\": \"BERT: Pre-training of deep bidirectional transformers for language understanding\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Esin Durmus; Claire Cardie\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b6\",\n \"title\": \"Exploring the role of prior beliefs for argument persuasion\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Roxanne El Baff; Henning Wachsmuth; Khalid ; Al Khatib; Benno Stein\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b7\",\n \"title\": \"Analyzing the Persuasive Effect of Style in News Editorial Argumentation\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Michael Fromm; Max Berrendorf; Johanna Reiml; Isabelle Mayerhofer; Siddharth Bhargava; Evgeniy Faerman; Thomas Seidl\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"Towards a holistic view on argument quality prediction\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Debanjan Ghosh; Aquila Khanam; Yubo Han; Smaranda Muresan\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b9\",\n \"title\": \"Coarse-grained argumentation features for scoring persuasive essays\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Daniel Gillick; Alessandro Presta; Gaurav Singh Tomar\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"End-to-end retrieval in continuous space\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Shai Gretz; Roni Friedman; Edo Cohen-Karlik; Assaf Toledo; Dan Lahav; Ranit Aharonov; Noam Slonim\",\n \"journal\": \"\",\n \"ref_id\": \"b11\",\n \"title\": \"A large-scale dataset for argument quality ranking: Construction and analysis\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Timon Gurcke; Milad Alshomary; Henning Wachsmuth\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"Assessing the sufficiency of arguments through conclusion generation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Ivan Habernal; Iryna Gurevych\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b13\",\n \"title\": \"What makes a convincing argument? empirical analysis and detecting attributes of convincingness in web argumentation\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Gordon Harvey\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"A brief guide to the elements of the academic essay\",\n \"year\": \"2009\"\n },\n {\n \"authors\": \"Md Kamrul Hasan; James Spann; Masum Hasan; Md Saiful Islam; Kurtis Haut; Rada Mihalcea; Ehsan Hoque\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b15\",\n \"title\": \"Hitting your MARQ: Multimodal ARgument quality assessment in long debate video\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Geoffrey Hinton; Oriol Vinyals; Jeff Dean\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"Distilling the knowledge in a neural network\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Ioana Hulpus; Jonathan Kobbe; Christian Meilicke; Heiner Stuckenschmidt; Maria Becker; Juri Opitz; Vivi Nastase; Anette Frank\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"Towards explaining natural language arguments with background knowledge\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Vladimir Karpukhin; Barlas Oguz; Sewon Min; Patrick Lewis; Ledell Wu; Sergey Edunov; Danqi Chen; Wen-Tau Yih\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b18\",\n \"title\": \"Dense passage retrieval for opendomain question answering\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Anne Lauscher; Lily Ng; Courtney Napoles; Joel Tetreault\",\n \"journal\": \"International Committee on Computational Linguistics\",\n \"ref_id\": \"b19\",\n \"title\": \"Rhetoric, logic, and dialectic: Advancing theory-based argument quality assessment in natural language processing\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Anne Lauscher; Henning Wachsmuth; Iryna Gurevych; Goran Glavaš\",\n \"journal\": \"Transactions of the Association for Computational Linguistics\",\n \"ref_id\": \"b20\",\n \"title\": \"Scientia Potentia Est-On the Role of Knowledge in Computational Argumentation\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Sabrina Ludwig; Christian Mayer; Christopher Hansen; Kerstin Eilers; Steffen Brandt\",\n \"journal\": \"Psych\",\n \"ref_id\": \"b21\",\n \"title\": \"Automated essay scoring using transformer models\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Santiago Marro; Elena Cabrio; Serena Villata\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b22\",\n \"title\": \"Graph embeddings for argumentation quality assessment\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Yeti Mulyati; Daris Hadianto\",\n \"journal\": \"International Journal of Instruction\",\n \"ref_id\": \"b23\",\n \"title\": \"Enhancing argumentative writing via online peer feedbackbased essay: A quasi-experiment study\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Raymond S Nickerson\",\n \"journal\": \"Cambridge University Press\",\n \"ref_id\": \"b24\",\n \"title\": \"Biases, Misconceptions, and the Like\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \" Openai\",\n \"journal\": \"\",\n \"ref_id\": \"b25\",\n \"title\": \"Chatgpt\",\n \"year\": \"2022-04-30\"\n },\n {\n \"authors\": \"Ruty Rinott; Lena Dankin; Carlos Alzate Perez; Mitesh M Khapra; Ehud Aharoni; Noam Slonim\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b26\",\n \"title\": \"Show me your evidence -an automatic method for context dependent evidence detection\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Gabriella Skitalinskaya; Jonas Klaff; Henning Wachsmuth\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"Learning from revisions: Quality assessment of claims in argumentation at scale\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Christian Stab; Iryna Gurevych\",\n \"journal\": \"Dublin City University and Association for Computational Linguistics\",\n \"ref_id\": \"b28\",\n \"title\": \"a. Annotating argument components and relations in persuasive essays\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"Christian Stab; Iryna Gurevych\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b29\",\n \"title\": \"Identifying argumentative discourse structures in persuasive essays\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"Assaf Toledo; Shai Gretz; Edo Cohen-Karlik; Roni Friedman; Elad Venezian; Dan Lahav; Michal Jacovi; Ranit Aharonov; Noam Slonim\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b30\",\n \"title\": \"Automatic argument quality assessment -new datasets and methods\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Hugo Touvron; Louis Martin; Kevin Stone; Peter Albert; Amjad Almahairi; Yasmine Babaei; Nikolay Bashlykov; Soumya Batra; Prajjwal Bhargava; Shruti Bhosale; Dan Bikel; Lukas Blecher; Cristian Canton Ferrer; Moya Chen; Guillem Cucurull; David Esiobu; Jude Fernandes; Jeremy Fu; Wenyin Fu; Brian Fuller; Cynthia Gao; Vedanuj Goswami; Naman Goyal; Anthony Hartshorn; Saghar Hosseini; Rui Hou; Hakan Inan; Marcin Kardas; Viktor Kerkez; Madian Khabsa; Isabel Kloumann; Artem Korenev; Punit Singh Koura; Marie-Anne Lachaux; Thibaut Lavril; Jenya Lee; Diana Liskovich; Yinghai Lu; Yuning Mao; Xavier Martinet; Todor Mihaylov; Pushkar Mishra; Igor Molybog; Yixin Nie; Andrew Poulton; Jeremy Reizenstein; Rashi Rungta; Kalyan Saladi; Alan Schelten; Ruan Silva; Eric Michael Smith; Ranjan Subramanian; Ellen Xiaoqing; Binh Tan; Ross Tang; Adina Taylor; Jian Williams; Puxin Xiang Kuan; Zheng Xu; Iliyan Yan; Yuchen Zarov; Angela Zhang; Melanie Fan; Sharan Kambadur; Aurelien Narang; Robert Rodriguez; Sergey Stojnic; Thomas Edunov; Scialom\",\n \"journal\": \"\",\n \"ref_id\": \"b31\",\n \"title\": \"Llama 2: Open foundation and finetuned chat models\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Dietrich Trautmann; Michael Fromm; Thomas Volker Tresp; Hinrich Seidl; Schütze\",\n \"journal\": \"Datenbank-Spektrum\",\n \"ref_id\": \"b32\",\n \"title\": \"Relational and fine-grained argument mining\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Ł Ukasz Kaiser; Illia Polosukhin\",\n \"journal\": \"Advances in Neural Information Processing Systems\",\n \"ref_id\": \"b33\",\n \"title\": \"Attention is all you need\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Curran Associates; Inc \",\n \"journal\": \"\",\n \"ref_id\": \"b34\",\n \"title\": \"\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Henning Wachsmuth; Nona Naderi; Ivan Habernal; Yufang Hou; Graeme Hirst; Iryna Gurevych; Benno Stein\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b35\",\n \"title\": \"Argumentation quality assessment: Theory vs. practice\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Henning Wachsmuth; Nona Naderi; Yufang Hou; Yonatan Bilu; Tim Alberdingk Vinodkumar Prabhakaran; Graeme Thijm; Benno Hirst; Stein\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b36\",\n \"title\": \"Computational argumentation quality assessment in natural language\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Zhilin Yang; Zihang Dai; Yiming Yang; Jaime Carbonell; Russ R Salakhutdinov; Quoc V Le\",\n \"journal\": \"\",\n \"ref_id\": \"b37\",\n \"title\": \"Xlnet: Generalized autoregressive pretraining for language understanding\",\n \"year\": \"2019\"\n }\n]"},"formulas":{"kind":"list like","value":[],"string":"[]"}}},{"rowIdx":391,"cells":{"doi":{"kind":"string","value":"10.18653/v1/2021.emnlp-main.532"},"pub_date":{"kind":"string","value":""},"sections":{"kind":"list like","value":[{"figure_ref":[],"heading":"Introduction","publication_ref":["b18","b15","b38","b47","b52"],"table_ref":[],"text":"When recurrent neural networks such as LSTM (Hochreiter and Schmidhuber, 1997) are the mainstream language model (LM) architecture, pointer networks, or so-called copy mechanisms (Gu et al., 2016), have been shown to improve the state-of-the-art LMs for next word prediction (Merity et al., 2017) and summarizations (See et al., 2017) by a large margin. However, after transformer (Vaswani et al., 2017) becomes the dominating LM architectures, the pointer networks are rarely used in the state-of-the-art pretrained LMs.One major reason is that the attention mechanism in every transformer layer can learn to copy the words from the context, so it "},{"figure_ref":["fig_7","fig_7","fig_0"],"heading":"Input context Ct","publication_ref":["b5","b44","b45","b60","b5","b5","b15","b60","b5"],"table_ref":[],"text":"Figure 1: Illustration of the softmax bottleneck and pointer network using an example from Chang and McCallum (2022). GPT-2 cannot output both king or woman as the possible next word due to the parallelogram structure in the output word embedding space, while the pointer network could solve this by directly copying words from the context. The standard softmax estimate the probabilities of outputting king and woman by the dot products between the hidden state h ct,V and their global word embeddings. By contrast, The pointer networks compute the dot products between the projected current hidden state h ct,S and projected hidden states h e,. for king and woman to estimate their probabilities.\nseems to be redundant to add a copying mechanism on top of the transformer.\nIn this paper, we demonstrate that the architectures like pointer networks can still substantially improve the state-of-the-art transformer LM architectures such as GPT-2 (Radford et al., 2019) and T5 (Raffel et al., 2020) mainly due to breaking the bottleneck of their final softmax layer (Yang et al., 2018;Chang and McCallum, 2022).\nIn Figure 1, we illustrate a simple example from Chang and McCallum (2022) word, most LMs would try to output a hidden state h ct,V that is close to all the next word possibilities. For example, when the next word should be either king or woman with similar probabilities, the ideal hidden state is supposed to be the average of the global output word embeddings of king and woman. However, there might be other interfering words (queen and man in this case) between the ideal next word candidates, which force the LM to output the wrong distribution.\nTo solve this problem, we can let the LMs predict the probability of copying the words in the context separately by paying attention to the previous hidden states (Gu et al., 2016) and we call this kind of architecture pointer networks in this paper. That is, we can compute the dot products with the hidden states of king h e,k and the hidden states of woman h e,w rather than with their global output word embeddings in order to estimate the probabilities of copying these two words in the context. Our experiments show that the pointer networks consistently improve the performance of GPT-2 in next word prediction and the quality of summarization from T5 and BART.\nContrary to the mainstream explanation in previous pointer network literature, we discover that most of the improvements in our experiments do not come from the attention mechanism. To study these improvements, we propose a very simple pointer network variant that does not use any previous hidden states and we show that the proposed method can achieve similar improvements.\nAs shown in Figure 2, we simply project the last hidden state into two embeddings. One embedding h ct,S is to compute the dot product with the context words, and h ct,V is for the dot product of the other words. Then, the GPT-2 can output the hidden state for context words h ct,S as the average embedding of the king and woman without interfered by the words of man and queen that are handled by h ct,V . We call this method context partition. In addition to words in the context, we can also use another embedding for the top-k likely next words. This can be viewed as a very simple and efficient alternative to a reranker, so we call it reranker partition.\nIn our experiments, we show that the context partition performs similarly to pointer networks while combining a pointer network, context partition, and reranker partition would significantly outperform each individual method. Compared to the state-ofthe-art solutions for alleviating the softmax bottleneck such as mixture of softmax (Yang et al., 2018;Chang and McCallum, 2022), our proposed method is more efficient while achieving lower perplexity on GPT-2. Furthermore, we find that adding a very expensive word-by-word reranker only improves our method slightly, which suggested the difficulty of further improving the final softmax layer over the proposed alternatives.\nIn the text completion task using GPT-2, we find that the proposed softmax alternatives reduce hallucination by copying more proper nouns from the context even though we did not provide any partof-speech information during training. In summarization, our methods and pointer networks output a more specific summary, increase the factuality, and consistently improve 9 metrics, especially in the smaller language models. Finally, we show that the softmax bottleneck problem is not completely solved in GPT-3.5 in the limitation section."},{"figure_ref":[],"heading":"Main Contributions","publication_ref":[],"table_ref":[],"text":"• We propose a series of efficient softmax alternatives that unify the ideas of pointer network, reranker, multiple embeddings, and vocabulary partitioning. 1• We evaluate the proposed softmax alternatives in text completion tasks and summarization tasks using various metrics to identify where our methods improve the most.\n• Our experiments indicate pointer networks and our proposed alternatives can still improve the modern transformer-based LMs. By breaking the softmax bottleneck, our methods learn to sometimes copy the context words to reduce generation hallucination and sometimes exclude the context words to reduce the repetition. Besides, we find that the softmax bottleneck problem won't be completely solved by the huge size of GPT-3.5."},{"figure_ref":[],"heading":"Background","publication_ref":[],"table_ref":[],"text":"Before introducing our method, we would first briefly review the problem we are solving and its state-of-the-art solutions."},{"figure_ref":[],"heading":"Softmax Bottleneck Problem","publication_ref":["b5"],"table_ref":[],"text":"Most LMs use a softmax layer to compute the final probability of predicting the word x:\nP M (x|c t ) = exp(Logit(x, c t ))\nx exp(Logit(x , c t ))\n,\nwhere c t is the context words. Typically, the logit Logit(x, c t ) = (h M ct ) T w x , h M ct is the M th-layer hidden state given the input context c t and w x is the output word embeddings for x.\nOne problem is that the output word embeddings w x are global and independent to the context. After pretraining, the similar words would have similar output word embeddings. However, the similarity structure in the word embedding space might prevent LMs from outputting the desired distribution. The parallelogram structure among the embeddings of king, queen, woman, and man is a simple example. Chang and McCallum (2022) generalize this observation and show that some words in a small subspace would create some multi-mode distributions that a LM cannot output using a single hidden state h ct in the softmax layer."},{"figure_ref":[],"heading":"Mixture of Softmax Method","publication_ref":["b60"],"table_ref":[],"text":"To overcome the bottleneck, one natural solution is to have multiple hidden states and each hidden state corresponds to a group of possible words (Yang et al., 2018). For example, we can have one hidden state for king and another hidden state for woman.\nOne major concern of this mixture of softmax (MoS) approach is the computational overhead. MoS needs to compute the final softmax multiple times and merge their resulting distributions. That is, we need to compute the dot products between every hidden state and all the words in the vocabulary, which is expensive especially when the vocabulary size is large. "},{"figure_ref":[],"heading":"Multiple Input State Enhancement","publication_ref":["b5","b5"],"table_ref":[],"text":"In MoS, the multiple hidden states come from the linear projections of the last hidden state. Chang and McCallum (2022) point out that the total degree of freedom among the multiple hidden states is limited by the dimensionality of the hidden state.\nTo allow LMs to move multiple hidden states more freely, Chang and McCallum (2022) propose to concatenate the projection of a block of hidden state with the last hidden state h M ct so as to increase its dimensionality:\nq ct = h M ct ⊕ GELU L h (⊕ i,m h M -m c t-i ) ,(2)\nwhere GELU is the non-linear transformation used in GPT-2 and L h is a linear transformation that allows us to consider more hidden states without significantly increasing the model size.\n⊕ i,m h M -m c t-i\nis the concatenation of a block of hidden states. We set the block size to be 3x3 in our GPT-2 experiments and 1x3 in our summarization experiments (i.e., considering the last 3 hidden states in the last layer as shown in Figure 3)."},{"figure_ref":[],"heading":"Methods","publication_ref":[],"table_ref":[],"text":"To break the softmax bottleneck more efficiently compared to MoS, our overall strategy is simple.\nIf we can identify a small partition of words that are very likely to become the next word, we can just compute the dot products between a hidden state and the embeddings of these likely words instead of all the words as in MoS. For example, if we can identify king and woman are much more likely to appear than queen and man, we can only compute the dot product between a hidden state and the embeddings of king and woman without being interfered by other words. Specifically, when we compute the next word probability in Equation 1, the logit of the word x given the context c \nt Logit(x, c t ) = f T ct,S e x if x ∈ S f T ct,V w x O/W ,(3)\nL f LD L f C\nFigure 3: Architectures of our method for T5/BART that computes Logit CEP R in Equation 6. In GPT-2, we use same architecture except that we take the 3x3 input hidden state block rather than the 1x3 block and there are no encoder-related components, which are marked by dotted lines.\nwhere f ct,S = L f S (q ct ) and f ct,V = L f V (q ct ) are the linear projections of the hidden state concatenation q ct in Equation 2. As shown in Table 1, different softmax alternatives have different ways of constructing this set S and use different word embeddings e x .\nTo simplify our explanation, we will focus on the decoder-only LM (i.e., GPT-2) first and extend our method to encoder-decoder LM (i.e., T5 and BART)."},{"figure_ref":[],"heading":"GPT-2","publication_ref":[],"table_ref":[],"text":"We will explain each softmax alternative individually and their connections to previous work such as pointer networks or rerankers."},{"figure_ref":[],"heading":"Pointer Network (P) as Local Word Embedding","publication_ref":["b38","b38"],"table_ref":[],"text":"Similar to Pointer Sentinel (PS) (Merity et al., 2017), we treat the words in the context differently (S = {x|x ∈ c t }) and let their word embeddings e x come from the previous hidden states:\ne x = f x,ct,LD = t i=1 1 c i t =x L f LD (q c i t ) t i=1 1 c i t =x ,(4)\nwhere c i t is the ith input words in the context c t , L f LD is a linear layer, and\n1 c i t =x = 1 if c i t = x.\nAs a result, we can use the GPT-2 model to not only predict the hidden state f ct,S = f ct,P D = L f P D (q ct ) and f ct,V but also predict the word embedding of context words e x . Unlike the global word embedding w x , the local word embedding e x is context-dependent, so the LM can break the softmax bottleneck by adjusting the similarity of words based on the context. For example, GPT-2 could increase the similarity between e king and e woman to output the high probability for both words easily.\nWe call this version of pointer network local decoder (LD) embedding, which has some minor differences compared to PS (Merity et al., 2017) and other variants. For example, we merge their logits while PS merges their probabilities. PS does not do normalization when computing e x . In our experiments, we would show that these pointer network variants all have very similar improvements in modern LMs."},{"figure_ref":[],"heading":"Context Partition (C)","publication_ref":[],"table_ref":[],"text":"To understand the source of the improvements from pointer networks, we simplify their architectures by setting the word embedding e x = w x and the partition S is still the set of context words. Although much simpler, the LM with this context partition method can still break the softmax bottleneck by properly coordinating the hidden state specifically for the context words f ct,S = f ct,C = L f C (q ct ) and the hidden state for other words f ct,V . Compared to the pointer network, one advantage of context partition is that the LM can still leverage the learned global word similarity when estimating the probabilities of context words."},{"figure_ref":[],"heading":"Reranker Partition (R)","publication_ref":[],"table_ref":[],"text":"In some cases, the possible next words might not be mentioned in the context. For example, in the context My favorite actor is Ryan [MASK], the next word could be Reynolds, Gosling, or the last names of other Ryan. Hence, using only the context partition does not completely solve the multimodal distribution problem.\nInspired by the idea of the reranker, we set S to be the top k words with the highest logits f T ct,V w x . In practice, finding an ideal k could be difficult. When k is small, the reranker partition might not include the very likely next word. When k is large, the reranker partition might not be able to separate the output candidates and the interfering words. To alleviate the problem, we can have multiple reranker partitions and use different hidden state embeddings (e.g., f ct,R1 and f ct,R2 ) for different partitions."},{"figure_ref":[],"heading":"Hybrid Approach (CPR)","publication_ref":[],"table_ref":[],"text":"Local embeddings in the pointer networks and global embeddings in the context partition are complementary. Using local embeddings is representational powerful while using global embedding can leverage the global similarity of words. Hence, we can combine the two methods by summing their dot products.\nFor the methods that use different S, we can simply determine an order of computing the dot products and let the later dot products overwrite the existing values. In our experiments, we always use the order illustrated in Figure 3. That is, we compute the logits (Logit CP R (x, c t )) by\n f T ct,C w x + f T ct,P D f x,ct,LD if x ∈ c t f T ct,R1 w x if x ∈ W (k 1 ) -c t f T ct,R2 w x if x ∈ W (k 2 ) -W (k 1 ) -c t f T ct,V w x O/W ,(5)\nwhere W (k 2 ) is the top k 2 words with the highest f T ct,V w x and W (k 1 ) is the top k 1 words with the highest max(f T ct,V w x , f T ct,R2 w x )."},{"figure_ref":[],"heading":"T5 and BART","publication_ref":[],"table_ref":[],"text":"In the encoder-decoder architectures, our local decoder embedding, context partition, and reranker partitions are still applicable. Besides, we can leverage the words in the encoder input to further improve the performance."},{"figure_ref":[],"heading":"Encoder Partition (E) and Local Encoder Embedding (P)","publication_ref":[],"table_ref":[],"text":"Similar to the context partition, the encoder partition handles the words in the encoder input I differently by setting S = {x|x ∈ I} and using the global word embedding e x = w x .\nAs in Equation 4, we can also let the hidden states in the last layer pass through another linear layer L f LE () to predict the embeddings of the words in the encoder input. The method is called local encoder (LE) embedding."},{"figure_ref":[],"heading":"Hybrid Approach (CEPR)","publication_ref":[],"table_ref":[],"text":"Similar to GPT-2, we combine local encoder embedding and encoder partition for computing the probabilities of the words that are in the encoder context but not in the decoder context. As shown in Figure 3, we compute Logit CEP R (x, c t ) by\n f T ct,C w x + f T ct,P D f x,ct,LD if x ∈ c t f T ct,E w x + f T ct,P E f x,I,LE if x ∈ I -c t f T ct,R1 w x if x ∈ W (k 1 ) -c t -I f T ct,V w x O/W ,(6)\nwhich is the same as Equation 5 except that we add the encoder partition and local encoder embedding, and we remove the second reranker partition."},{"figure_ref":[],"heading":"Experiments","publication_ref":["b47"],"table_ref":[],"text":"The pointer network was a popular technique in language modeling (Merity et al., 2017) and summarization (See et al., 2017). Thus, we also focus on these two fundamental applications."},{"figure_ref":[],"heading":"GPT-2","publication_ref":["b5","b44"],"table_ref":[],"text":"We follow the setup in Chang and McCallum (2022) to continue training GPT-2 on Wikipedia 2021 and OpenWebText (Radford et al., 2019)."},{"figure_ref":[],"heading":"Perplexity Comparison","publication_ref":[],"table_ref":["tab_3","tab_5","tab_6","tab_13"],"text":"In Table 2, we first compare their predictions on the next word distribution using the testing data perplexity, which is a standard metric in the LM architecture studies. In the To know how well our method breaks the softmax bottleneck, we implement a word-by-word reranker model on GPT-2, which appends the most likely 100 words to the context when predicting each next word (see Appendix C.3 for more details). In Table 3, we show that our efficient softmax alternative Softmax + CPR:20,100 + Mi achieves significantly lower perplexity. Furthermore, the word-by-word reranker is at least 10 times slower during training. Combining word-by-word reranker with our method only improves the perplexity very 2 Notice that the pointer networks from the previous work were originally designed for RNN. To add them on top of the transformer based LMs and make it more comparable to our methods, we simplify their architectures a little. Please see Appendix C.2 for more details. Table 4: ROUGE-1 F1 (%) of different methods on GPT-2. We compare the scores between the generated text and the reference (i.e., continuation), and between the generation and context. More methods and metrics are reported in Table 8.\nslightly, which suggests the challenges of further improving LM by breaking softmax bottleneck."},{"figure_ref":[],"heading":"Generated Text Comparison","publication_ref":[],"table_ref":["tab_13"],"text":"Next, we would like to understand how the distribution improvement affects the text generation. We sample some contexts in the test set of Wikipedia 2021 and compare the generated text quality of the different models given the contexts. The quality is measured by the ROUGE-1 F1 scores between the generated text and the actual continuation. To know how much the different models copy from the context, we also report the ROUGE-1 scores between the generation and the contexts. 8, we compare methods using more metrics to further support the conclusion."},{"figure_ref":[],"heading":"The results in","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Qualitative Analysis","publication_ref":[],"table_ref":["tab_8"],"text":"In Table 5, we visualize some distributions to explain our improvements. The softmax layer of GPT-2 is unable to properly learn to copy or exclude the word from the input context. For example, Softmax + Mi and MoS + Mi might output \"There are plates, keys, scissors, toys, and balloons in front of me, and I pick up the phone\", which causes a hallucination problem, while Softmax + CPR:20,100 + Mi and Pointer Sentinel (PS) + Mi can output the mentioned options with similar probability by copying the words in the context. In addition, GPT-2, MoS, and PS + Mi are very likely to output \"I like tennis, baseball, golf, basketball, and tennis\". This repetition problem happens because the next word should be some words similar to the listed sports names except for the sports that have been mentioned and the softmax layer has difficulties in outputting a donut-shape next word distribution in embedding space. In contrast, Softmax + CPR:20,100 + Mi can learn to exclude the listed sports by putting very negative logits on the context words, which yield the desired donut-shape distribution."},{"figure_ref":[],"heading":"T5 and BART in Summarization","publication_ref":["b34","b26","b43","b24","b4"],"table_ref":["tab_15","tab_16","tab_9"],"text":"We In the main paper, we evaluate the quality of summaries using four metrics. ROUGE-1 F1 (Lin, 2004) measures the unigram overlapping between the generated summary and the ground truth summary; CIDEr (Vedantam et al., 2015) adds a tfidf weighting on the n-gram overlapping score to emphasize correct prediction of rare phrases; factCC (Kryscinski et al., 2020) evaluates the factuality of the summary; MAUVE (Pillutla et al., 2021) compares the word distribution of summary and ground truth in a quantized embedding space. To further support our conclusions, we also compare the quality measured by several other metrics and their model sizes in Table 9 andTable 10.\nThe results are reported in Table 6. Similar to the GPT-2 experiments, the results are generally better as we combine more partitions and local embedding approaches. This demonstrates that we can directly fine-tune the LMs with our softmax alternatives without expensive pretraining.\nUnlike the GPT-2 experiments, multiple input hidden state enhancement (Mi) is not very effective, so we mainly compare the methods without Mi (i.e., q ct = h M ct , unlike Equation 2). We hypothesize one possible reason is that we haven't pretrained the T5 and BART with our softmax alternatives.\nOur improvements are larger in smaller models. This is probably because in a smaller word embedding space, there are more likely to be interfering words between the desired next word possibilities. Compared to our methods, the pointer networks perform well in BART-base but usually perform worse in other LMs. We need further investigations in the future to explore the reasons.\nCompared to ROUGE-1 score, the improvement percentage of CIDEr is overall higher. One major problem of the summarization LMs is that the generated summary contains too many commonly used phrases (King et al., 2022) and our considerably higher CIDEr scores indicate the alleviation of the problem. Our improvement on the factCC is also significant (Cao and Wang, 2021). Finally, our MAUVE improvement percentage on Book-Sum Paragraph dataset could reach around 30% in T5-Small. We hypothesize this is because we often mention the global entity names in the news (e.g., Obama) while the meaning of names in stories (e.g., John) is often defined by the context."},{"figure_ref":[],"heading":"Related Work","publication_ref":["b5","b5"],"table_ref":[],"text":"Repetition and hallucination are two common problems in language generation tasks. One common solution for repetition is to avoid outputting the words in the context, which is often called unlike- Our analyses demonstrate that parts of the hallucination and repetition problem come from the softmax bottleneck. The findings provide an explanation for the effectiveness of prior studies such as the above reranker approaches and pointer networks (Li et (Chang and McCallum, 2022) to explain the improvement of a pointer network. Their empirical results also support our conclusion that softmax bottleneck is a major reason that causes the factuality problem of LMs.\nOur work is motivated and inspired by Chang and McCallum (2022). In their work, they also propose to use different hidden states for different vocabulary partitions, but their partitioning is global and needs to be combined with the mixture of softmax (MoS) approach, which adds a significant overhead compared to the standard softmax layer. Our dynamic partitioning methods not only perform better but greatly reduce the overhead by removing the reliance on MoS."},{"figure_ref":[],"heading":"Conclusion","publication_ref":["b5"],"table_ref":[],"text":"Since the transformer becomes the mainstream encoder and decoder for LMs, the output softmax layer seems to be the only reasonable option for computing the word probability distribution. Although being simple and efficient, the softmax layer is inherently limited while the existing solutions are relatively slow (Chang and McCallum, 2022). This work proposes a series of softmax alternatives that can improve the text generation models without increasing the computational costs significantly. Our experiments suggest that the main improvement of the pointer network on top of a transformer comes from breaking the softmax bottleneck. Our results also indicate that the alternatives could alleviate some problems of hallucination, repetition, and too generic generation. Furthermore, all of the proposed alternatives can be applied to the LMs that have already been pretrained using softmax without requiring retraining from scratch. For the practitioner, we recommend using all the partitioning methods together to get the best performance, or using only the simple context partition to keep the architecture simple while getting the majority of the gain."},{"figure_ref":[],"heading":"Acknowledgement","publication_ref":[],"table_ref":[],"text":"We thank Nadar Akoury and the anonymous reviewers for their constructive feedback. This work was supported in part by the Center for Data Science and the Center for Intelligent Information Retrieval, in part by the Chan Zuckerberg Initiative under the project Scientific Knowledge Base Construction, in part by the IBM Research AI through the AI Horizons Network, in part using high performance computing equipment obtained under a grant from the Collaborative R&D Fund managed by the Massachusetts Technology Collaborative, and in part by the National Science Foundation (NSF) grant numbers IIS-1922090 and IIS-1763618. Any opinions, findings, conclusions, or recommendations expressed in this material are those of the authors and do not necessarily reflect those of the sponsor."},{"figure_ref":["fig_2","fig_2","fig_2","fig_2"],"heading":"Limitations","publication_ref":["b33","b23"],"table_ref":["tab_8","tab_8"],"text":"In our experiments, we find that the improvement of our methods tend to be larger in relatively smaller language models. Due to our limited access of computational resources, we are not able to try our methods on larger LMs. To know if a larger LM still suffers from the softmax bottleneck problem, we input the examples we used in Table 5 to GPT-3.5 and report their results in Figure 4.\nWe find that although GPT-3.5 greatly reduces the chance of hallucination compared to GPT-2, the next word distribution is still not ideal. For example, in Figure 4a, although the incorrect answer queen receives only a small probability, GPT-3.5 puts around 67% probability on woman. Similarly, even though GPT-3.5 is unlikely to hallucinate the sentence: There are plates, keys, scissors, toys, and balloons in front of me, and I pick up the phone as GPT-2, Figure 4b and Figure 4d show that the output distribution is still heavily biased toward one of the options and the most likely next word could change if the order of the options in the context changes. These results suggest that increasing model size indeed alleviates the softmax bottleneck problem but the problem is not completely solved even if a huge hidden state size (12k) and model size (175B) are used (Brown et al., 2020). We expect that adding our methods to the large LMs could rectify the biased distributions as shown in our experiments on smaller LMs (Table 5). Therefore, although improving smaller LMs has already had wide applications in practice, trying our methods on a larger LM is a promising next step, which we haven't been able to do.\nThe current implementation of our methods also has some room for improvements. Our codes currently contain some unnecessary computation to circumvent the restrictions of PyTorch library, so we should be able to further accelerate it by writing CUDA code. Furthermore, our codes haven't supported the pretraining of BART or T5. We expect that completing the future work could make our method faster and better.\nSince the focus of this paper is improving the architecture of general transformer decoder, our evaluation of each application is not as comprehensive as the studies for a particular application. For example, although we test our methods using many metrics and the metrics show a consistent trend, there are many other factuality metrics we haven't tried (Li et al., 2022). We also haven't conducted human evaluation to further verify our conclusion because conducting human evaluation properly is challenging (Karpinska et al., 2021) and time-consuming. In addition, if we include more words in a context partition, the performance might be better at the cost of extra computational overhead. We leave the analyses of the tradeoff as future work."},{"figure_ref":[],"heading":"Ethics Statement","publication_ref":["b29"],"table_ref":[],"text":"In our experiments, we find that our methods usually copy more words from the context or encoder input. The tendency might have some potential issues. For example, our improvements might be reduced on the languages with more morphology. Furthermore, in some summarization applications, increasing the factuality by increasing the extractiveness might not be ideal (Ladhak et al., 2022;Goyal et al., 2022a).\nAs described in Section 2.1, one major limita-tion of the popular softmax layer is its global word embeddings. The problem would become more serious when there are more tokens whose meanings are locally defined (e.g., names in the BookSum dataset). Our methods would be more useful in those circumstances and might alleviate some biases described in Finally, our results show that when there are some uncertainties in the next word (e.g., could be king or woman), existing LMs could have some difficulties of copying the words from the context and our methods alleviate the problem. Thus, our methods should also be able to improve the lexically controllable language generation models that put the desired keywords into the context such as Goldfarb-Tarrant et al. ( 2019 "},{"figure_ref":[],"heading":"A Appendix Overview","publication_ref":[],"table_ref":[],"text":"In the appendix, we first analyze our methods using more metrics in Appendix B and describe what we learn from the results. Next, we provide some details of our methods and baselines in Appendix C. Finally, we specify some experiment setups and hyperparameters in Appendix D."},{"figure_ref":["fig_4"],"heading":"B More Results and Analysis","publication_ref":[],"table_ref":[],"text":"In this section, we will report more results and provide more detailed analyses accordingly to investigate the advantages of different methods. 2020) demonstrate that the loss decreases linearly as the log of the model size increases. Therefore, a new architecture needs to perform better than the old architecture with a similar model size to verify that the improvement does not come from memorizing more information through the extra parameters. From the loss versus log(model size) curve in Figure 5, we can see that our proposed methods are significantly better than MoS and slightly better than a pointer network baseline as the model becomes larger."},{"figure_ref":[],"heading":"B.1 GPT-2 Experiments","publication_ref":["b48","b8","b5","b39","b5","b5"],"table_ref":["tab_13","tab_12"],"text":"We use the following metrics to measure the text generated by GPT-2.\n• ROUGE-1 (R1): The prediction F1 for unigram in the actual continuation.\n• ROUGE-1 Context (R1C): The prediction F1 for unigram in the context. • ROUGE-1 Proper (R1P): The same as ROUGE-1 except that only the proper nouns are considered. We measure this metric because the correctness of the entity name prediction is critical to the factuality of the generation.\n• ROUGE-1 Proper Context (R1PC): The same as ROUGE-1 Context (R1C) except that only the proper nouns are considered.\n• ROUGE-2 (R2): The prediction F1 for bigram in the actual continuation.\n• Proper Noun Ratio (P Ratio): The average number of proper nouns in the generation divided by the average number of proper nouns in the actual continuation. The LMs usually generate fewer proper nouns compared to the actual continuation (See et al., 2019), so the values are usually lower than 1. The P Ratio closer to 1 is better.\n• CIDEr (Vedantam et al., 2015): A metric for measuring the quality and specificity of the generation.\n• NIST (Doddington, 2002): Similar to CIDEr. CIDEr uses tf-idf to weigh the n-gram while NIST measures the information gain.\nThe results are reported in Table 8. In terms of R1, R2, CIDEr, and NIST, our proposed methods such as Softmax + C + Mi and Softmax + CPR:20,100 + Mi are significantly better than the pointer network baselines PS + Mi and PG + Mi. Comparing with Softmax + CPR:20,100 + Mi, PS + Mi has a significantly higher P Ratio and R1PC but similar R1P. This indicates that PS + Mi copies more proper nouns from the context while there is a similar number of proper nouns that are in actual continuation, so Softmax + CPR:20,100 + Mi actually has a higher accuracy on the proper noun prediction.\nIn text corpus such as Wikipedia, we do not know the ground truth next word distribution and which context leads to multiple probable next words, so we cannot quantitatively analyze the improvement on the ambiguous contexts. To alleviate the concern, we test our methods on the synthetic dataset constructed by Chang and McCallum (2022). The dataset is built using templates and Google analogy dataset (Mikolov et al., 2013), so we know the ground truth next word distribution. The dataset consists of the ambiguous contexts such as I went to Paris and Germany before, and I love one of the places more, which is, where the next word is either the diagonal words of the parallelogram such as Paris and Germany or the edge words such as Paris and France. For the details of the experimental setup, please refer to Chang and McCallum (2022).\nIn Table 7, we can see that Softmax + CPR:20,100 + Mi achieves the lowest perplexity in all subsets and outperforms the Softmax + Mi baseline by a large margin, especially in the diagonal subset where the ground truth word embedding distribution has multiple modes. Notice that the performance of MoS + Mi is worse than what reported in Chang and McCallum (2022) probably because we shared the input and output word embeddings."},{"figure_ref":["fig_4","fig_6"],"heading":"B.2 Summarization","publication_ref":["b47"],"table_ref":["tab_15","tab_16","tab_9"],"text":"Compared to Figure 5, Figure 6 shows that our methods improve the loss of T5 in CNN/DM more than GPT-2 in Wikipedia.\nIn Table 9 and Table 10, we compare the different summarization models by their model size, evaluation losses, inference time, and other metrics which we use in subsection B.1. The pointer network baselines and our methods significantly improve most metrics over the softmax baseline, which is used ubiquitously in nearly all LMs. Although our method generally improves less on the T5-Base model, the percentages of additional parameters and inference time overhead are much smaller. Although our methods tend to improve less in larger language model, we still improve BART Large very significantly in NIST, CIDEr, and MAUVE, and Mi seems to become more effective in BART Large.\nThe testing set of SAMSUM dataset only has 819 samples, so some metrics such as R1 and R2 are not as stable as other three datasets. PG (See et al., 2017) for T5-Small and T5-Base perform much worse in SAMSUM dataset. We hypothesize that it is because the dialog input in SAMSUM dataset is very different from the pretraining data of T5, which makes training PG unstable.\nIn most datasets and models, the R Ratio from our method is significantly closer to 1 than the softmax baseline, which means the average number of proper nouns in our summaries is much closer to the average number of proper nouns in the humanwritten summary. For example, in BookSum Paragraph, we improve its R Ratio by 26%, which partially explains our large MAUVE improvement in Table 6. Notice that our methods do not always output more proper nouns. For example, for BART Base in CNN/DM dataset, our methods reduce the R Ratio of the softmax baseline, which is larger than 1. This shows that our methods could learn when we should copy the proper nouns according to the training data."},{"figure_ref":[],"heading":"C Method Details","publication_ref":[],"table_ref":[],"text":"We describe some details of our methods and baselines in this section. "},{"figure_ref":[],"heading":"C.1 Proposed Methods","publication_ref":[],"table_ref":[],"text":"To allow us to start from existing LMs that are pretrained using softmax, we keep the modified softmax layer initially working almost the same as the original softmax layer. We initialize the linear transformation weights of L f P D (), L f LD (), L f P E (), and L f LE () as 10 -10 • I. The other linear weights L f . () are initialized as the identity matrix I.\nIn the local decoder embedding method Softmax + P + Mi, the initialization would give the 0 logit to all context words. To solve the issue, we revise Equation 3 a little and compute Logit P (x, c t ) by\nf T ct,V w x + f T ct,P D f x,ct,LD if x ∈ c t f T ct,V w x O/W .(7)\nThat is, we initially rely on the original softmax layer to compute all the logits and let the term f T ct,P D f x,ct,LD gradually influences the logits of the context words.\nIn MoS + CPR:20,100 + Mi, our proposed method only revises the logit in one of the softmax."},{"figure_ref":[],"heading":"C.2 Pointer Network Baselines","publication_ref":["b15","b47","b38"],"table_ref":["tab_8"],"text":"The pointer networks are originally designed for RNN, so we are unable to use exactly the same formula proposed in the papers. Nevertheless, we try our best to adapt the pointer networks for the transformer encoder while keeping the gist of the formulas. In all methods, to let the results more comparable to our methods, we use f ct,P E and L f LE to determine the probability of copying the words from the context, and use f T ct,V w x to determine the probability of generating all the words in the vocabulary.\nIn CopyNet (Gu et al., 2016), we compute the probability of outputting the word x as Notice that CopyNet needs to sum up the exponential of dot products, which often causes overflow problems in GPT-2. We can set b to be a large negative value initially to solve the problem, but its perplexity is much worse than the other two pointer network variants. Thus, we choose to skip the CopyNet in the GPT-2 experiments.\nIn Pointer Generator (See et al., 2017), we compute the probability of x using\nP rob(x|I, c t ) = p gen exp f T ct,V w x Z V +(1 -p gen ) |I| j=1 1 I j =x P E (j|I, c t ),(9)\nwhere P E (j|I, c t ) = exp v T tanh(f c t ,P E +L f LE (h M I j )+b) Z E , p gen = σ(q T h M ct + b ptr ), the normalization term Z V =\nx∈V exp f T ct,V w x and Z E = |I| j=1 exp v T tanh(f ct,P E + L f LE (h M I j ) + b) . We skip the coverage mechanism in the pointer generator paper to make it more comparable to other methods. In T5 experiments, its training loss is sometimes very large, so we set b ptr as 3 initially to keep the p gen close to 1 (i.e., turn pointer part off initially). In other experiments, we set b ptr = 0.\nIn Pointer Sentinel (Merity et al., 2017), the probability of x is computed by\nP rob(x|I, ct) = g exp f T c t ,V wx ZV + |I| j=1 1 I j =x exp f T c t ,P E tanh(L f LE (h M I j )) + b Zp ,(10)\ng = exp(q T h M c t ) Zp , and Z p exp f T ct,P E tanh(L f LE (h M I j )) + b . In our experiments, we find that the pointer network variants usually have similar performance (except that PG sometimes performs much worse in summarization due to some training stability issues). This suggests that the differences in the pointer network variants often do not influence the performance significantly, which justifies our simplification of the formulas in the original paper and supports our conclusion that the improvement comes from breaking the softmax bottleneck.\nNotice that in the above pointer network variants, the pointer part can only increase the probability of the context words from the generator part. As a result, it cannot alleviate the repetition problem in the last example of Table 5."},{"figure_ref":[],"heading":"C.3 Word-by-word Reranker Baseline","publication_ref":[],"table_ref":[],"text":"We illustrate our word-by-word reranker (wbwR) in Figure 7. The method has two stages. In the first stage, we compute the logits using the projected hidden state f ct,V and retrieve the top k words. At the second stage, we append the top k words to the input context along with the hidden state f ct,R for reranking the context words. 3 We use the same positional embeddings for all candidates to encourage the model to change the ranking of the words. Next, we use the hidden states corresponding to the candidates to compute their local word embeddings as f x,ct,LD . Finally, we re-estimate the probabilities of top k words by\nf T ct,V w x + f T ct,R f x,ct,LD if x ∈ W (k) f T ct,V w x O/W .(11)\nTo improve the quality of our top k candidates, the final loss is the addition of the wbwR loss at the second stage and the loss of the original softmax layer that only uses the logits from f T ct,V w x at the first stage. When we combine the wbwR with Softmax + CPR:20,100 + Mi, we simply use Softmax + CPR:20,100 + Mi at the first stage and use the wbwR to overwrite the logits of Softmax + CPR:20,100 + Mi at the second stage.\nUsing this method, we can update the embeddings of the words that are not in the context and allow the candidates to interact with the input context to determine their probabilities as the classic two-stage reranker while keeping the model size roughly unchanged. Nevertheless, the method can only change the probability of the top k words and its computational overhead and memory requirement prevents us from using a very large k."},{"figure_ref":[],"heading":"Global Word Embeddings","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"GPT-2 encoder ……","publication_ref":[],"table_ref":[],"text":"After debating whether to bow to the king or the woman first, the jester decided on the Unlike the standard GPT-2, we cannot get the probability of all positions in one forward pass because the input contexts are different when computing the probability at each position and the input of the second stage reranker depends on the results of the previous forward pass at the first stage. To speed up, we reuse the computed hidden states and batchify the forward passes.\nIn our implementation, we first get the top k candidates corresponding to all tokens in the stage1 (just original GPT2) as the input of stage2 reranker. To avoid recalculating the hidden states of the context at stage2, we store the hidden states using the past-key-value in Hugging Face and only compute the hidden states corresponding to the top k candidate tokens at stage2.\nWe divide the computation of the whole input sequence into several blocks as shown in Figure 8. In each block, we input a batch containing the last few tokens and top k candidates into the GPT-2 while reusing the hidden states of their common contexts from stage1. In this way, we can increase parallelism by increasing the block size if the GPU memory allows it.\nEven though we spent substantial effort on optimizing the wbwR, the method is still too slow to be practically useful. Even if we use four RTX 8000 (a faster GPU with a larger memory), our wbwR implementation is still around 10 times slower than our proposed Softmax + CPR:20,100 + Mi that uses only one RTX 2080."},{"figure_ref":[],"heading":"D Experiment Details","publication_ref":[],"table_ref":[],"text":"For the reproducibility, we provide some experimental configuration in this section.\nPlease see our codes (https: //github.com/iesl/Softmax-CPR) for more details."},{"figure_ref":[],"heading":"D.1 GPT-2 Experimental Details","publication_ref":["b5"],"table_ref":[],"text":"We mostly follow the experimental setup Chang and McCallum (2022) except that we share the input and output word embeddings as in the standard GPT-2 models. As in Chang and McCallum (2022), we use the last 2% of the corpus as the test set and the 2% before that as the validation set. 4 In 4 We do not shuffle the corpus before splitting the datasets. We found that our improvement could be even larger if we In the text completion experiment, we generate 360k continuations with a length of 50 given the prompts in Wikipedia. We first sample 40k sequences in the test data of Wikipedia 2021. Next, we use the first 20, 70, and 120 words in the sequence as our context and let the different models generate the next 50 words as continuations. The references are the actual next 50 words. All the methods use Top-K sampling and K=5."},{"figure_ref":[],"heading":"D.2 Summarization Experimental Details","publication_ref":["b27"],"table_ref":[],"text":"BookSum dataset (Kryściński et al., 2021) includes three summarization tasks: Summarizing a book, a chapter, and a paragraph. We test our methods using the paragraph summarization task due to the input length restriction of BART and T5. The dataset is constructed by automatically aligning the paragraphs in a chapter with the sentences in a chapter summary, which introduces noise to the dataset. Similarly, XSUM uses the first sentence in news instead of manually-written summary as the ground truth reference. The relatively noisier datasets such as XSUM and BookSum Paragraph, and smaller dataset like SAMSUM could test the stability of the methods. The sizes of the summarization datasets could be found in Table 11.\nshuffle the corpus to let the training data distribution closer to the testing data distribution.\nWe conduct the summarization experiments based on a summarization example code from Hugging Face 5 . Most of our hyperparameters use the default value in the code. In our preliminary study, our improvement is not sensitive to the hyperparameter choice (e.g., the improvement gap is similar across different numbers of epochs). Thus, we do not tune the hyperparameters for each method or for each dataset unless we cannot reach a low training loss at the end.\nIn CNN/DM, XSUM, and SAMSUM datasets, We train models for 3 epochs. In BookSum datasets, We trained models for 5 epochs. 6 The learning rate is set to be 5e -05 except for BART Large model in BookSum, where we use 1e -05 to stablize the training of all methods.\nAll the experiments use batch size 8 and AdamW with betas=(0.9,0.999), epsilon=1e -06, weight-decay=1.2e -6. During the generation, we used Top-K sampling (K=10) as our decoding method. The maximum summary length is set as 128 and maximum input length is 1024. We use warmup for the first 1000 steps in all the experiments, which allows us to change the architecture of T5 and BART more significantly (e.g., using Mi) without having a training stability issue.\nThe k in the reranker partition and the block size of multiple input hidden states (Mi) is coarsely tuned based on validation performance of CNN/DM. Unlike considering the top 100 words in the open-end text completion using GPT-2, we find that reranking the top 20 words is sufficient for our summarization models, probably because next words are easier to predict in the summarization task.\nFor our evaluation metrics, we use the default setting for ROUGE 7 and set use_stemmer=True. When reporting the ROUGE scores, we follow the conventions to show their percentages. We use the default setting for MAUVE 8 , CIDER 9 , NIST 10 . For MAUVE, we insert a new line symbol after every sentence as in the original Hugging Face "},{"figure_ref":[],"heading":"D.3 Computational Environment and Software","publication_ref":["b59","b19"],"table_ref":[],"text":"We implement our methods by revising the Hugging Face library (Wolf et al., 2020). From Hugging Face, we load the pretrained LMs including GPT-2 Small 12 , GPT-2 Medium 13 , T5-Small 14 , T5-Base 15 , BART Base 16 , and BART Large 17 . We use SpaCy (Honnibal et al., 2020) to detect the proper nouns.\nFor GPT-2 Medium, T5-Base, and BART Large, we use NVIDIA GeForce RTX 8000 to train the model and for other smaller models, we use NVIDIA GeForce RTX 2080. Most of experiments could be done within one week. In all the inference time experiments, we use NVIDIA GeForce GTX TITAN X, batch size 4 for GPT-2, and batch size 8 for BART and T5."}],"string":"[\n {\n \"figure_ref\": [],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b18\",\n \"b15\",\n \"b38\",\n \"b47\",\n \"b52\"\n ],\n \"table_ref\": [],\n \"text\": \"When recurrent neural networks such as LSTM (Hochreiter and Schmidhuber, 1997) are the mainstream language model (LM) architecture, pointer networks, or so-called copy mechanisms (Gu et al., 2016), have been shown to improve the state-of-the-art LMs for next word prediction (Merity et al., 2017) and summarizations (See et al., 2017) by a large margin. However, after transformer (Vaswani et al., 2017) becomes the dominating LM architectures, the pointer networks are rarely used in the state-of-the-art pretrained LMs.One major reason is that the attention mechanism in every transformer layer can learn to copy the words from the context, so it \"\n },\n {\n \"figure_ref\": [\n \"fig_7\",\n \"fig_7\",\n \"fig_0\"\n ],\n \"heading\": \"Input context Ct\",\n \"publication_ref\": [\n \"b5\",\n \"b44\",\n \"b45\",\n \"b60\",\n \"b5\",\n \"b5\",\n \"b15\",\n \"b60\",\n \"b5\"\n ],\n \"table_ref\": [],\n \"text\": \"Figure 1: Illustration of the softmax bottleneck and pointer network using an example from Chang and McCallum (2022). GPT-2 cannot output both king or woman as the possible next word due to the parallelogram structure in the output word embedding space, while the pointer network could solve this by directly copying words from the context. The standard softmax estimate the probabilities of outputting king and woman by the dot products between the hidden state h ct,V and their global word embeddings. By contrast, The pointer networks compute the dot products between the projected current hidden state h ct,S and projected hidden states h e,. for king and woman to estimate their probabilities.\\nseems to be redundant to add a copying mechanism on top of the transformer.\\nIn this paper, we demonstrate that the architectures like pointer networks can still substantially improve the state-of-the-art transformer LM architectures such as GPT-2 (Radford et al., 2019) and T5 (Raffel et al., 2020) mainly due to breaking the bottleneck of their final softmax layer (Yang et al., 2018;Chang and McCallum, 2022).\\nIn Figure 1, we illustrate a simple example from Chang and McCallum (2022) word, most LMs would try to output a hidden state h ct,V that is close to all the next word possibilities. For example, when the next word should be either king or woman with similar probabilities, the ideal hidden state is supposed to be the average of the global output word embeddings of king and woman. However, there might be other interfering words (queen and man in this case) between the ideal next word candidates, which force the LM to output the wrong distribution.\\nTo solve this problem, we can let the LMs predict the probability of copying the words in the context separately by paying attention to the previous hidden states (Gu et al., 2016) and we call this kind of architecture pointer networks in this paper. That is, we can compute the dot products with the hidden states of king h e,k and the hidden states of woman h e,w rather than with their global output word embeddings in order to estimate the probabilities of copying these two words in the context. Our experiments show that the pointer networks consistently improve the performance of GPT-2 in next word prediction and the quality of summarization from T5 and BART.\\nContrary to the mainstream explanation in previous pointer network literature, we discover that most of the improvements in our experiments do not come from the attention mechanism. To study these improvements, we propose a very simple pointer network variant that does not use any previous hidden states and we show that the proposed method can achieve similar improvements.\\nAs shown in Figure 2, we simply project the last hidden state into two embeddings. One embedding h ct,S is to compute the dot product with the context words, and h ct,V is for the dot product of the other words. Then, the GPT-2 can output the hidden state for context words h ct,S as the average embedding of the king and woman without interfered by the words of man and queen that are handled by h ct,V . We call this method context partition. In addition to words in the context, we can also use another embedding for the top-k likely next words. This can be viewed as a very simple and efficient alternative to a reranker, so we call it reranker partition.\\nIn our experiments, we show that the context partition performs similarly to pointer networks while combining a pointer network, context partition, and reranker partition would significantly outperform each individual method. Compared to the state-ofthe-art solutions for alleviating the softmax bottleneck such as mixture of softmax (Yang et al., 2018;Chang and McCallum, 2022), our proposed method is more efficient while achieving lower perplexity on GPT-2. Furthermore, we find that adding a very expensive word-by-word reranker only improves our method slightly, which suggested the difficulty of further improving the final softmax layer over the proposed alternatives.\\nIn the text completion task using GPT-2, we find that the proposed softmax alternatives reduce hallucination by copying more proper nouns from the context even though we did not provide any partof-speech information during training. In summarization, our methods and pointer networks output a more specific summary, increase the factuality, and consistently improve 9 metrics, especially in the smaller language models. Finally, we show that the softmax bottleneck problem is not completely solved in GPT-3.5 in the limitation section.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Main Contributions\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"• We propose a series of efficient softmax alternatives that unify the ideas of pointer network, reranker, multiple embeddings, and vocabulary partitioning. 1• We evaluate the proposed softmax alternatives in text completion tasks and summarization tasks using various metrics to identify where our methods improve the most.\\n• Our experiments indicate pointer networks and our proposed alternatives can still improve the modern transformer-based LMs. By breaking the softmax bottleneck, our methods learn to sometimes copy the context words to reduce generation hallucination and sometimes exclude the context words to reduce the repetition. Besides, we find that the softmax bottleneck problem won't be completely solved by the huge size of GPT-3.5.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Background\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Before introducing our method, we would first briefly review the problem we are solving and its state-of-the-art solutions.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Softmax Bottleneck Problem\",\n \"publication_ref\": [\n \"b5\"\n ],\n \"table_ref\": [],\n \"text\": \"Most LMs use a softmax layer to compute the final probability of predicting the word x:\\nP M (x|c t ) = exp(Logit(x, c t ))\\nx exp(Logit(x , c t ))\\n,\\nwhere c t is the context words. Typically, the logit Logit(x, c t ) = (h M ct ) T w x , h M ct is the M th-layer hidden state given the input context c t and w x is the output word embeddings for x.\\nOne problem is that the output word embeddings w x are global and independent to the context. After pretraining, the similar words would have similar output word embeddings. However, the similarity structure in the word embedding space might prevent LMs from outputting the desired distribution. The parallelogram structure among the embeddings of king, queen, woman, and man is a simple example. Chang and McCallum (2022) generalize this observation and show that some words in a small subspace would create some multi-mode distributions that a LM cannot output using a single hidden state h ct in the softmax layer.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Mixture of Softmax Method\",\n \"publication_ref\": [\n \"b60\"\n ],\n \"table_ref\": [],\n \"text\": \"To overcome the bottleneck, one natural solution is to have multiple hidden states and each hidden state corresponds to a group of possible words (Yang et al., 2018). For example, we can have one hidden state for king and another hidden state for woman.\\nOne major concern of this mixture of softmax (MoS) approach is the computational overhead. MoS needs to compute the final softmax multiple times and merge their resulting distributions. That is, we need to compute the dot products between every hidden state and all the words in the vocabulary, which is expensive especially when the vocabulary size is large. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Multiple Input State Enhancement\",\n \"publication_ref\": [\n \"b5\",\n \"b5\"\n ],\n \"table_ref\": [],\n \"text\": \"In MoS, the multiple hidden states come from the linear projections of the last hidden state. Chang and McCallum (2022) point out that the total degree of freedom among the multiple hidden states is limited by the dimensionality of the hidden state.\\nTo allow LMs to move multiple hidden states more freely, Chang and McCallum (2022) propose to concatenate the projection of a block of hidden state with the last hidden state h M ct so as to increase its dimensionality:\\nq ct = h M ct ⊕ GELU L h (⊕ i,m h M -m c t-i ) ,(2)\\nwhere GELU is the non-linear transformation used in GPT-2 and L h is a linear transformation that allows us to consider more hidden states without significantly increasing the model size.\\n⊕ i,m h M -m c t-i\\nis the concatenation of a block of hidden states. We set the block size to be 3x3 in our GPT-2 experiments and 1x3 in our summarization experiments (i.e., considering the last 3 hidden states in the last layer as shown in Figure 3).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Methods\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"To break the softmax bottleneck more efficiently compared to MoS, our overall strategy is simple.\\nIf we can identify a small partition of words that are very likely to become the next word, we can just compute the dot products between a hidden state and the embeddings of these likely words instead of all the words as in MoS. For example, if we can identify king and woman are much more likely to appear than queen and man, we can only compute the dot product between a hidden state and the embeddings of king and woman without being interfered by other words. Specifically, when we compute the next word probability in Equation 1, the logit of the word x given the context c \\nt Logit(x, c t ) = f T ct,S e x if x ∈ S f T ct,V w x O/W ,(3)\\nL f LD L f C\\nFigure 3: Architectures of our method for T5/BART that computes Logit CEP R in Equation 6. In GPT-2, we use same architecture except that we take the 3x3 input hidden state block rather than the 1x3 block and there are no encoder-related components, which are marked by dotted lines.\\nwhere f ct,S = L f S (q ct ) and f ct,V = L f V (q ct ) are the linear projections of the hidden state concatenation q ct in Equation 2. As shown in Table 1, different softmax alternatives have different ways of constructing this set S and use different word embeddings e x .\\nTo simplify our explanation, we will focus on the decoder-only LM (i.e., GPT-2) first and extend our method to encoder-decoder LM (i.e., T5 and BART).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"GPT-2\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We will explain each softmax alternative individually and their connections to previous work such as pointer networks or rerankers.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Pointer Network (P) as Local Word Embedding\",\n \"publication_ref\": [\n \"b38\",\n \"b38\"\n ],\n \"table_ref\": [],\n \"text\": \"Similar to Pointer Sentinel (PS) (Merity et al., 2017), we treat the words in the context differently (S = {x|x ∈ c t }) and let their word embeddings e x come from the previous hidden states:\\ne x = f x,ct,LD = t i=1 1 c i t =x L f LD (q c i t ) t i=1 1 c i t =x ,(4)\\nwhere c i t is the ith input words in the context c t , L f LD is a linear layer, and\\n1 c i t =x = 1 if c i t = x.\\nAs a result, we can use the GPT-2 model to not only predict the hidden state f ct,S = f ct,P D = L f P D (q ct ) and f ct,V but also predict the word embedding of context words e x . Unlike the global word embedding w x , the local word embedding e x is context-dependent, so the LM can break the softmax bottleneck by adjusting the similarity of words based on the context. For example, GPT-2 could increase the similarity between e king and e woman to output the high probability for both words easily.\\nWe call this version of pointer network local decoder (LD) embedding, which has some minor differences compared to PS (Merity et al., 2017) and other variants. For example, we merge their logits while PS merges their probabilities. PS does not do normalization when computing e x . In our experiments, we would show that these pointer network variants all have very similar improvements in modern LMs.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Context Partition (C)\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"To understand the source of the improvements from pointer networks, we simplify their architectures by setting the word embedding e x = w x and the partition S is still the set of context words. Although much simpler, the LM with this context partition method can still break the softmax bottleneck by properly coordinating the hidden state specifically for the context words f ct,S = f ct,C = L f C (q ct ) and the hidden state for other words f ct,V . Compared to the pointer network, one advantage of context partition is that the LM can still leverage the learned global word similarity when estimating the probabilities of context words.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Reranker Partition (R)\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In some cases, the possible next words might not be mentioned in the context. For example, in the context My favorite actor is Ryan [MASK], the next word could be Reynolds, Gosling, or the last names of other Ryan. Hence, using only the context partition does not completely solve the multimodal distribution problem.\\nInspired by the idea of the reranker, we set S to be the top k words with the highest logits f T ct,V w x . In practice, finding an ideal k could be difficult. When k is small, the reranker partition might not include the very likely next word. When k is large, the reranker partition might not be able to separate the output candidates and the interfering words. To alleviate the problem, we can have multiple reranker partitions and use different hidden state embeddings (e.g., f ct,R1 and f ct,R2 ) for different partitions.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Hybrid Approach (CPR)\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Local embeddings in the pointer networks and global embeddings in the context partition are complementary. Using local embeddings is representational powerful while using global embedding can leverage the global similarity of words. Hence, we can combine the two methods by summing their dot products.\\nFor the methods that use different S, we can simply determine an order of computing the dot products and let the later dot products overwrite the existing values. In our experiments, we always use the order illustrated in Figure 3. That is, we compute the logits (Logit CP R (x, c t )) by\\n f T ct,C w x + f T ct,P D f x,ct,LD if x ∈ c t f T ct,R1 w x if x ∈ W (k 1 ) -c t f T ct,R2 w x if x ∈ W (k 2 ) -W (k 1 ) -c t f T ct,V w x O/W ,(5)\\nwhere W (k 2 ) is the top k 2 words with the highest f T ct,V w x and W (k 1 ) is the top k 1 words with the highest max(f T ct,V w x , f T ct,R2 w x ).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"T5 and BART\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In the encoder-decoder architectures, our local decoder embedding, context partition, and reranker partitions are still applicable. Besides, we can leverage the words in the encoder input to further improve the performance.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Encoder Partition (E) and Local Encoder Embedding (P)\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Similar to the context partition, the encoder partition handles the words in the encoder input I differently by setting S = {x|x ∈ I} and using the global word embedding e x = w x .\\nAs in Equation 4, we can also let the hidden states in the last layer pass through another linear layer L f LE () to predict the embeddings of the words in the encoder input. The method is called local encoder (LE) embedding.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Hybrid Approach (CEPR)\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Similar to GPT-2, we combine local encoder embedding and encoder partition for computing the probabilities of the words that are in the encoder context but not in the decoder context. As shown in Figure 3, we compute Logit CEP R (x, c t ) by\\n f T ct,C w x + f T ct,P D f x,ct,LD if x ∈ c t f T ct,E w x + f T ct,P E f x,I,LE if x ∈ I -c t f T ct,R1 w x if x ∈ W (k 1 ) -c t -I f T ct,V w x O/W ,(6)\\nwhich is the same as Equation 5 except that we add the encoder partition and local encoder embedding, and we remove the second reranker partition.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiments\",\n \"publication_ref\": [\n \"b47\"\n ],\n \"table_ref\": [],\n \"text\": \"The pointer network was a popular technique in language modeling (Merity et al., 2017) and summarization (See et al., 2017). Thus, we also focus on these two fundamental applications.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"GPT-2\",\n \"publication_ref\": [\n \"b5\",\n \"b44\"\n ],\n \"table_ref\": [],\n \"text\": \"We follow the setup in Chang and McCallum (2022) to continue training GPT-2 on Wikipedia 2021 and OpenWebText (Radford et al., 2019).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Perplexity Comparison\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_3\",\n \"tab_5\",\n \"tab_6\",\n \"tab_13\"\n ],\n \"text\": \"In Table 2, we first compare their predictions on the next word distribution using the testing data perplexity, which is a standard metric in the LM architecture studies. In the To know how well our method breaks the softmax bottleneck, we implement a word-by-word reranker model on GPT-2, which appends the most likely 100 words to the context when predicting each next word (see Appendix C.3 for more details). In Table 3, we show that our efficient softmax alternative Softmax + CPR:20,100 + Mi achieves significantly lower perplexity. Furthermore, the word-by-word reranker is at least 10 times slower during training. Combining word-by-word reranker with our method only improves the perplexity very 2 Notice that the pointer networks from the previous work were originally designed for RNN. To add them on top of the transformer based LMs and make it more comparable to our methods, we simplify their architectures a little. Please see Appendix C.2 for more details. Table 4: ROUGE-1 F1 (%) of different methods on GPT-2. We compare the scores between the generated text and the reference (i.e., continuation), and between the generation and context. More methods and metrics are reported in Table 8.\\nslightly, which suggests the challenges of further improving LM by breaking softmax bottleneck.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Generated Text Comparison\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_13\"\n ],\n \"text\": \"Next, we would like to understand how the distribution improvement affects the text generation. We sample some contexts in the test set of Wikipedia 2021 and compare the generated text quality of the different models given the contexts. The quality is measured by the ROUGE-1 F1 scores between the generated text and the actual continuation. To know how much the different models copy from the context, we also report the ROUGE-1 scores between the generation and the contexts. 8, we compare methods using more metrics to further support the conclusion.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"The results in\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Qualitative Analysis\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_8\"\n ],\n \"text\": \"In Table 5, we visualize some distributions to explain our improvements. The softmax layer of GPT-2 is unable to properly learn to copy or exclude the word from the input context. For example, Softmax + Mi and MoS + Mi might output \\\"There are plates, keys, scissors, toys, and balloons in front of me, and I pick up the phone\\\", which causes a hallucination problem, while Softmax + CPR:20,100 + Mi and Pointer Sentinel (PS) + Mi can output the mentioned options with similar probability by copying the words in the context. In addition, GPT-2, MoS, and PS + Mi are very likely to output \\\"I like tennis, baseball, golf, basketball, and tennis\\\". This repetition problem happens because the next word should be some words similar to the listed sports names except for the sports that have been mentioned and the softmax layer has difficulties in outputting a donut-shape next word distribution in embedding space. In contrast, Softmax + CPR:20,100 + Mi can learn to exclude the listed sports by putting very negative logits on the context words, which yield the desired donut-shape distribution.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"T5 and BART in Summarization\",\n \"publication_ref\": [\n \"b34\",\n \"b26\",\n \"b43\",\n \"b24\",\n \"b4\"\n ],\n \"table_ref\": [\n \"tab_15\",\n \"tab_16\",\n \"tab_9\"\n ],\n \"text\": \"We In the main paper, we evaluate the quality of summaries using four metrics. ROUGE-1 F1 (Lin, 2004) measures the unigram overlapping between the generated summary and the ground truth summary; CIDEr (Vedantam et al., 2015) adds a tfidf weighting on the n-gram overlapping score to emphasize correct prediction of rare phrases; factCC (Kryscinski et al., 2020) evaluates the factuality of the summary; MAUVE (Pillutla et al., 2021) compares the word distribution of summary and ground truth in a quantized embedding space. To further support our conclusions, we also compare the quality measured by several other metrics and their model sizes in Table 9 andTable 10.\\nThe results are reported in Table 6. Similar to the GPT-2 experiments, the results are generally better as we combine more partitions and local embedding approaches. This demonstrates that we can directly fine-tune the LMs with our softmax alternatives without expensive pretraining.\\nUnlike the GPT-2 experiments, multiple input hidden state enhancement (Mi) is not very effective, so we mainly compare the methods without Mi (i.e., q ct = h M ct , unlike Equation 2). We hypothesize one possible reason is that we haven't pretrained the T5 and BART with our softmax alternatives.\\nOur improvements are larger in smaller models. This is probably because in a smaller word embedding space, there are more likely to be interfering words between the desired next word possibilities. Compared to our methods, the pointer networks perform well in BART-base but usually perform worse in other LMs. We need further investigations in the future to explore the reasons.\\nCompared to ROUGE-1 score, the improvement percentage of CIDEr is overall higher. One major problem of the summarization LMs is that the generated summary contains too many commonly used phrases (King et al., 2022) and our considerably higher CIDEr scores indicate the alleviation of the problem. Our improvement on the factCC is also significant (Cao and Wang, 2021). Finally, our MAUVE improvement percentage on Book-Sum Paragraph dataset could reach around 30% in T5-Small. We hypothesize this is because we often mention the global entity names in the news (e.g., Obama) while the meaning of names in stories (e.g., John) is often defined by the context.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Work\",\n \"publication_ref\": [\n \"b5\",\n \"b5\"\n ],\n \"table_ref\": [],\n \"text\": \"Repetition and hallucination are two common problems in language generation tasks. One common solution for repetition is to avoid outputting the words in the context, which is often called unlike- Our analyses demonstrate that parts of the hallucination and repetition problem come from the softmax bottleneck. The findings provide an explanation for the effectiveness of prior studies such as the above reranker approaches and pointer networks (Li et (Chang and McCallum, 2022) to explain the improvement of a pointer network. Their empirical results also support our conclusion that softmax bottleneck is a major reason that causes the factuality problem of LMs.\\nOur work is motivated and inspired by Chang and McCallum (2022). In their work, they also propose to use different hidden states for different vocabulary partitions, but their partitioning is global and needs to be combined with the mixture of softmax (MoS) approach, which adds a significant overhead compared to the standard softmax layer. Our dynamic partitioning methods not only perform better but greatly reduce the overhead by removing the reliance on MoS.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [\n \"b5\"\n ],\n \"table_ref\": [],\n \"text\": \"Since the transformer becomes the mainstream encoder and decoder for LMs, the output softmax layer seems to be the only reasonable option for computing the word probability distribution. Although being simple and efficient, the softmax layer is inherently limited while the existing solutions are relatively slow (Chang and McCallum, 2022). This work proposes a series of softmax alternatives that can improve the text generation models without increasing the computational costs significantly. Our experiments suggest that the main improvement of the pointer network on top of a transformer comes from breaking the softmax bottleneck. Our results also indicate that the alternatives could alleviate some problems of hallucination, repetition, and too generic generation. Furthermore, all of the proposed alternatives can be applied to the LMs that have already been pretrained using softmax without requiring retraining from scratch. For the practitioner, we recommend using all the partitioning methods together to get the best performance, or using only the simple context partition to keep the architecture simple while getting the majority of the gain.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Acknowledgement\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We thank Nadar Akoury and the anonymous reviewers for their constructive feedback. This work was supported in part by the Center for Data Science and the Center for Intelligent Information Retrieval, in part by the Chan Zuckerberg Initiative under the project Scientific Knowledge Base Construction, in part by the IBM Research AI through the AI Horizons Network, in part using high performance computing equipment obtained under a grant from the Collaborative R&D Fund managed by the Massachusetts Technology Collaborative, and in part by the National Science Foundation (NSF) grant numbers IIS-1922090 and IIS-1763618. Any opinions, findings, conclusions, or recommendations expressed in this material are those of the authors and do not necessarily reflect those of the sponsor.\"\n },\n {\n \"figure_ref\": [\n \"fig_2\",\n \"fig_2\",\n \"fig_2\",\n \"fig_2\"\n ],\n \"heading\": \"Limitations\",\n \"publication_ref\": [\n \"b33\",\n \"b23\"\n ],\n \"table_ref\": [\n \"tab_8\",\n \"tab_8\"\n ],\n \"text\": \"In our experiments, we find that the improvement of our methods tend to be larger in relatively smaller language models. Due to our limited access of computational resources, we are not able to try our methods on larger LMs. To know if a larger LM still suffers from the softmax bottleneck problem, we input the examples we used in Table 5 to GPT-3.5 and report their results in Figure 4.\\nWe find that although GPT-3.5 greatly reduces the chance of hallucination compared to GPT-2, the next word distribution is still not ideal. For example, in Figure 4a, although the incorrect answer queen receives only a small probability, GPT-3.5 puts around 67% probability on woman. Similarly, even though GPT-3.5 is unlikely to hallucinate the sentence: There are plates, keys, scissors, toys, and balloons in front of me, and I pick up the phone as GPT-2, Figure 4b and Figure 4d show that the output distribution is still heavily biased toward one of the options and the most likely next word could change if the order of the options in the context changes. These results suggest that increasing model size indeed alleviates the softmax bottleneck problem but the problem is not completely solved even if a huge hidden state size (12k) and model size (175B) are used (Brown et al., 2020). We expect that adding our methods to the large LMs could rectify the biased distributions as shown in our experiments on smaller LMs (Table 5). Therefore, although improving smaller LMs has already had wide applications in practice, trying our methods on a larger LM is a promising next step, which we haven't been able to do.\\nThe current implementation of our methods also has some room for improvements. Our codes currently contain some unnecessary computation to circumvent the restrictions of PyTorch library, so we should be able to further accelerate it by writing CUDA code. Furthermore, our codes haven't supported the pretraining of BART or T5. We expect that completing the future work could make our method faster and better.\\nSince the focus of this paper is improving the architecture of general transformer decoder, our evaluation of each application is not as comprehensive as the studies for a particular application. For example, although we test our methods using many metrics and the metrics show a consistent trend, there are many other factuality metrics we haven't tried (Li et al., 2022). We also haven't conducted human evaluation to further verify our conclusion because conducting human evaluation properly is challenging (Karpinska et al., 2021) and time-consuming. In addition, if we include more words in a context partition, the performance might be better at the cost of extra computational overhead. We leave the analyses of the tradeoff as future work.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Ethics Statement\",\n \"publication_ref\": [\n \"b29\"\n ],\n \"table_ref\": [],\n \"text\": \"In our experiments, we find that our methods usually copy more words from the context or encoder input. The tendency might have some potential issues. For example, our improvements might be reduced on the languages with more morphology. Furthermore, in some summarization applications, increasing the factuality by increasing the extractiveness might not be ideal (Ladhak et al., 2022;Goyal et al., 2022a).\\nAs described in Section 2.1, one major limita-tion of the popular softmax layer is its global word embeddings. The problem would become more serious when there are more tokens whose meanings are locally defined (e.g., names in the BookSum dataset). Our methods would be more useful in those circumstances and might alleviate some biases described in Finally, our results show that when there are some uncertainties in the next word (e.g., could be king or woman), existing LMs could have some difficulties of copying the words from the context and our methods alleviate the problem. Thus, our methods should also be able to improve the lexically controllable language generation models that put the desired keywords into the context such as Goldfarb-Tarrant et al. ( 2019 \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A Appendix Overview\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In the appendix, we first analyze our methods using more metrics in Appendix B and describe what we learn from the results. Next, we provide some details of our methods and baselines in Appendix C. Finally, we specify some experiment setups and hyperparameters in Appendix D.\"\n },\n {\n \"figure_ref\": [\n \"fig_4\"\n ],\n \"heading\": \"B More Results and Analysis\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this section, we will report more results and provide more detailed analyses accordingly to investigate the advantages of different methods. 2020) demonstrate that the loss decreases linearly as the log of the model size increases. Therefore, a new architecture needs to perform better than the old architecture with a similar model size to verify that the improvement does not come from memorizing more information through the extra parameters. From the loss versus log(model size) curve in Figure 5, we can see that our proposed methods are significantly better than MoS and slightly better than a pointer network baseline as the model becomes larger.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B.1 GPT-2 Experiments\",\n \"publication_ref\": [\n \"b48\",\n \"b8\",\n \"b5\",\n \"b39\",\n \"b5\",\n \"b5\"\n ],\n \"table_ref\": [\n \"tab_13\",\n \"tab_12\"\n ],\n \"text\": \"We use the following metrics to measure the text generated by GPT-2.\\n• ROUGE-1 (R1): The prediction F1 for unigram in the actual continuation.\\n• ROUGE-1 Context (R1C): The prediction F1 for unigram in the context. • ROUGE-1 Proper (R1P): The same as ROUGE-1 except that only the proper nouns are considered. We measure this metric because the correctness of the entity name prediction is critical to the factuality of the generation.\\n• ROUGE-1 Proper Context (R1PC): The same as ROUGE-1 Context (R1C) except that only the proper nouns are considered.\\n• ROUGE-2 (R2): The prediction F1 for bigram in the actual continuation.\\n• Proper Noun Ratio (P Ratio): The average number of proper nouns in the generation divided by the average number of proper nouns in the actual continuation. The LMs usually generate fewer proper nouns compared to the actual continuation (See et al., 2019), so the values are usually lower than 1. The P Ratio closer to 1 is better.\\n• CIDEr (Vedantam et al., 2015): A metric for measuring the quality and specificity of the generation.\\n• NIST (Doddington, 2002): Similar to CIDEr. CIDEr uses tf-idf to weigh the n-gram while NIST measures the information gain.\\nThe results are reported in Table 8. In terms of R1, R2, CIDEr, and NIST, our proposed methods such as Softmax + C + Mi and Softmax + CPR:20,100 + Mi are significantly better than the pointer network baselines PS + Mi and PG + Mi. Comparing with Softmax + CPR:20,100 + Mi, PS + Mi has a significantly higher P Ratio and R1PC but similar R1P. This indicates that PS + Mi copies more proper nouns from the context while there is a similar number of proper nouns that are in actual continuation, so Softmax + CPR:20,100 + Mi actually has a higher accuracy on the proper noun prediction.\\nIn text corpus such as Wikipedia, we do not know the ground truth next word distribution and which context leads to multiple probable next words, so we cannot quantitatively analyze the improvement on the ambiguous contexts. To alleviate the concern, we test our methods on the synthetic dataset constructed by Chang and McCallum (2022). The dataset is built using templates and Google analogy dataset (Mikolov et al., 2013), so we know the ground truth next word distribution. The dataset consists of the ambiguous contexts such as I went to Paris and Germany before, and I love one of the places more, which is, where the next word is either the diagonal words of the parallelogram such as Paris and Germany or the edge words such as Paris and France. For the details of the experimental setup, please refer to Chang and McCallum (2022).\\nIn Table 7, we can see that Softmax + CPR:20,100 + Mi achieves the lowest perplexity in all subsets and outperforms the Softmax + Mi baseline by a large margin, especially in the diagonal subset where the ground truth word embedding distribution has multiple modes. Notice that the performance of MoS + Mi is worse than what reported in Chang and McCallum (2022) probably because we shared the input and output word embeddings.\"\n },\n {\n \"figure_ref\": [\n \"fig_4\",\n \"fig_6\"\n ],\n \"heading\": \"B.2 Summarization\",\n \"publication_ref\": [\n \"b47\"\n ],\n \"table_ref\": [\n \"tab_15\",\n \"tab_16\",\n \"tab_9\"\n ],\n \"text\": \"Compared to Figure 5, Figure 6 shows that our methods improve the loss of T5 in CNN/DM more than GPT-2 in Wikipedia.\\nIn Table 9 and Table 10, we compare the different summarization models by their model size, evaluation losses, inference time, and other metrics which we use in subsection B.1. The pointer network baselines and our methods significantly improve most metrics over the softmax baseline, which is used ubiquitously in nearly all LMs. Although our method generally improves less on the T5-Base model, the percentages of additional parameters and inference time overhead are much smaller. Although our methods tend to improve less in larger language model, we still improve BART Large very significantly in NIST, CIDEr, and MAUVE, and Mi seems to become more effective in BART Large.\\nThe testing set of SAMSUM dataset only has 819 samples, so some metrics such as R1 and R2 are not as stable as other three datasets. PG (See et al., 2017) for T5-Small and T5-Base perform much worse in SAMSUM dataset. We hypothesize that it is because the dialog input in SAMSUM dataset is very different from the pretraining data of T5, which makes training PG unstable.\\nIn most datasets and models, the R Ratio from our method is significantly closer to 1 than the softmax baseline, which means the average number of proper nouns in our summaries is much closer to the average number of proper nouns in the humanwritten summary. For example, in BookSum Paragraph, we improve its R Ratio by 26%, which partially explains our large MAUVE improvement in Table 6. Notice that our methods do not always output more proper nouns. For example, for BART Base in CNN/DM dataset, our methods reduce the R Ratio of the softmax baseline, which is larger than 1. This shows that our methods could learn when we should copy the proper nouns according to the training data.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C Method Details\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We describe some details of our methods and baselines in this section. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C.1 Proposed Methods\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"To allow us to start from existing LMs that are pretrained using softmax, we keep the modified softmax layer initially working almost the same as the original softmax layer. We initialize the linear transformation weights of L f P D (), L f LD (), L f P E (), and L f LE () as 10 -10 • I. The other linear weights L f . () are initialized as the identity matrix I.\\nIn the local decoder embedding method Softmax + P + Mi, the initialization would give the 0 logit to all context words. To solve the issue, we revise Equation 3 a little and compute Logit P (x, c t ) by\\nf T ct,V w x + f T ct,P D f x,ct,LD if x ∈ c t f T ct,V w x O/W .(7)\\nThat is, we initially rely on the original softmax layer to compute all the logits and let the term f T ct,P D f x,ct,LD gradually influences the logits of the context words.\\nIn MoS + CPR:20,100 + Mi, our proposed method only revises the logit in one of the softmax.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C.2 Pointer Network Baselines\",\n \"publication_ref\": [\n \"b15\",\n \"b47\",\n \"b38\"\n ],\n \"table_ref\": [\n \"tab_8\"\n ],\n \"text\": \"The pointer networks are originally designed for RNN, so we are unable to use exactly the same formula proposed in the papers. Nevertheless, we try our best to adapt the pointer networks for the transformer encoder while keeping the gist of the formulas. In all methods, to let the results more comparable to our methods, we use f ct,P E and L f LE to determine the probability of copying the words from the context, and use f T ct,V w x to determine the probability of generating all the words in the vocabulary.\\nIn CopyNet (Gu et al., 2016), we compute the probability of outputting the word x as Notice that CopyNet needs to sum up the exponential of dot products, which often causes overflow problems in GPT-2. We can set b to be a large negative value initially to solve the problem, but its perplexity is much worse than the other two pointer network variants. Thus, we choose to skip the CopyNet in the GPT-2 experiments.\\nIn Pointer Generator (See et al., 2017), we compute the probability of x using\\nP rob(x|I, c t ) = p gen exp f T ct,V w x Z V +(1 -p gen ) |I| j=1 1 I j =x P E (j|I, c t ),(9)\\nwhere P E (j|I, c t ) = exp v T tanh(f c t ,P E +L f LE (h M I j )+b) Z E , p gen = σ(q T h M ct + b ptr ), the normalization term Z V =\\nx∈V exp f T ct,V w x and Z E = |I| j=1 exp v T tanh(f ct,P E + L f LE (h M I j ) + b) . We skip the coverage mechanism in the pointer generator paper to make it more comparable to other methods. In T5 experiments, its training loss is sometimes very large, so we set b ptr as 3 initially to keep the p gen close to 1 (i.e., turn pointer part off initially). In other experiments, we set b ptr = 0.\\nIn Pointer Sentinel (Merity et al., 2017), the probability of x is computed by\\nP rob(x|I, ct) = g exp f T c t ,V wx ZV + |I| j=1 1 I j =x exp f T c t ,P E tanh(L f LE (h M I j )) + b Zp ,(10)\\ng = exp(q T h M c t ) Zp , and Z p exp f T ct,P E tanh(L f LE (h M I j )) + b . In our experiments, we find that the pointer network variants usually have similar performance (except that PG sometimes performs much worse in summarization due to some training stability issues). This suggests that the differences in the pointer network variants often do not influence the performance significantly, which justifies our simplification of the formulas in the original paper and supports our conclusion that the improvement comes from breaking the softmax bottleneck.\\nNotice that in the above pointer network variants, the pointer part can only increase the probability of the context words from the generator part. As a result, it cannot alleviate the repetition problem in the last example of Table 5.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C.3 Word-by-word Reranker Baseline\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We illustrate our word-by-word reranker (wbwR) in Figure 7. The method has two stages. In the first stage, we compute the logits using the projected hidden state f ct,V and retrieve the top k words. At the second stage, we append the top k words to the input context along with the hidden state f ct,R for reranking the context words. 3 We use the same positional embeddings for all candidates to encourage the model to change the ranking of the words. Next, we use the hidden states corresponding to the candidates to compute their local word embeddings as f x,ct,LD . Finally, we re-estimate the probabilities of top k words by\\nf T ct,V w x + f T ct,R f x,ct,LD if x ∈ W (k) f T ct,V w x O/W .(11)\\nTo improve the quality of our top k candidates, the final loss is the addition of the wbwR loss at the second stage and the loss of the original softmax layer that only uses the logits from f T ct,V w x at the first stage. When we combine the wbwR with Softmax + CPR:20,100 + Mi, we simply use Softmax + CPR:20,100 + Mi at the first stage and use the wbwR to overwrite the logits of Softmax + CPR:20,100 + Mi at the second stage.\\nUsing this method, we can update the embeddings of the words that are not in the context and allow the candidates to interact with the input context to determine their probabilities as the classic two-stage reranker while keeping the model size roughly unchanged. Nevertheless, the method can only change the probability of the top k words and its computational overhead and memory requirement prevents us from using a very large k.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Global Word Embeddings\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"GPT-2 encoder ……\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"After debating whether to bow to the king or the woman first, the jester decided on the Unlike the standard GPT-2, we cannot get the probability of all positions in one forward pass because the input contexts are different when computing the probability at each position and the input of the second stage reranker depends on the results of the previous forward pass at the first stage. To speed up, we reuse the computed hidden states and batchify the forward passes.\\nIn our implementation, we first get the top k candidates corresponding to all tokens in the stage1 (just original GPT2) as the input of stage2 reranker. To avoid recalculating the hidden states of the context at stage2, we store the hidden states using the past-key-value in Hugging Face and only compute the hidden states corresponding to the top k candidate tokens at stage2.\\nWe divide the computation of the whole input sequence into several blocks as shown in Figure 8. In each block, we input a batch containing the last few tokens and top k candidates into the GPT-2 while reusing the hidden states of their common contexts from stage1. In this way, we can increase parallelism by increasing the block size if the GPU memory allows it.\\nEven though we spent substantial effort on optimizing the wbwR, the method is still too slow to be practically useful. Even if we use four RTX 8000 (a faster GPU with a larger memory), our wbwR implementation is still around 10 times slower than our proposed Softmax + CPR:20,100 + Mi that uses only one RTX 2080.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"D Experiment Details\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"For the reproducibility, we provide some experimental configuration in this section.\\nPlease see our codes (https: //github.com/iesl/Softmax-CPR) for more details.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"D.1 GPT-2 Experimental Details\",\n \"publication_ref\": [\n \"b5\"\n ],\n \"table_ref\": [],\n \"text\": \"We mostly follow the experimental setup Chang and McCallum (2022) except that we share the input and output word embeddings as in the standard GPT-2 models. As in Chang and McCallum (2022), we use the last 2% of the corpus as the test set and the 2% before that as the validation set. 4 In 4 We do not shuffle the corpus before splitting the datasets. We found that our improvement could be even larger if we In the text completion experiment, we generate 360k continuations with a length of 50 given the prompts in Wikipedia. We first sample 40k sequences in the test data of Wikipedia 2021. Next, we use the first 20, 70, and 120 words in the sequence as our context and let the different models generate the next 50 words as continuations. The references are the actual next 50 words. All the methods use Top-K sampling and K=5.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"D.2 Summarization Experimental Details\",\n \"publication_ref\": [\n \"b27\"\n ],\n \"table_ref\": [],\n \"text\": \"BookSum dataset (Kryściński et al., 2021) includes three summarization tasks: Summarizing a book, a chapter, and a paragraph. We test our methods using the paragraph summarization task due to the input length restriction of BART and T5. The dataset is constructed by automatically aligning the paragraphs in a chapter with the sentences in a chapter summary, which introduces noise to the dataset. Similarly, XSUM uses the first sentence in news instead of manually-written summary as the ground truth reference. The relatively noisier datasets such as XSUM and BookSum Paragraph, and smaller dataset like SAMSUM could test the stability of the methods. The sizes of the summarization datasets could be found in Table 11.\\nshuffle the corpus to let the training data distribution closer to the testing data distribution.\\nWe conduct the summarization experiments based on a summarization example code from Hugging Face 5 . Most of our hyperparameters use the default value in the code. In our preliminary study, our improvement is not sensitive to the hyperparameter choice (e.g., the improvement gap is similar across different numbers of epochs). Thus, we do not tune the hyperparameters for each method or for each dataset unless we cannot reach a low training loss at the end.\\nIn CNN/DM, XSUM, and SAMSUM datasets, We train models for 3 epochs. In BookSum datasets, We trained models for 5 epochs. 6 The learning rate is set to be 5e -05 except for BART Large model in BookSum, where we use 1e -05 to stablize the training of all methods.\\nAll the experiments use batch size 8 and AdamW with betas=(0.9,0.999), epsilon=1e -06, weight-decay=1.2e -6. During the generation, we used Top-K sampling (K=10) as our decoding method. The maximum summary length is set as 128 and maximum input length is 1024. We use warmup for the first 1000 steps in all the experiments, which allows us to change the architecture of T5 and BART more significantly (e.g., using Mi) without having a training stability issue.\\nThe k in the reranker partition and the block size of multiple input hidden states (Mi) is coarsely tuned based on validation performance of CNN/DM. Unlike considering the top 100 words in the open-end text completion using GPT-2, we find that reranking the top 20 words is sufficient for our summarization models, probably because next words are easier to predict in the summarization task.\\nFor our evaluation metrics, we use the default setting for ROUGE 7 and set use_stemmer=True. When reporting the ROUGE scores, we follow the conventions to show their percentages. We use the default setting for MAUVE 8 , CIDER 9 , NIST 10 . For MAUVE, we insert a new line symbol after every sentence as in the original Hugging Face \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"D.3 Computational Environment and Software\",\n \"publication_ref\": [\n \"b59\",\n \"b19\"\n ],\n \"table_ref\": [],\n \"text\": \"We implement our methods by revising the Hugging Face library (Wolf et al., 2020). From Hugging Face, we load the pretrained LMs including GPT-2 Small 12 , GPT-2 Medium 13 , T5-Small 14 , T5-Base 15 , BART Base 16 , and BART Large 17 . We use SpaCy (Honnibal et al., 2020) to detect the proper nouns.\\nFor GPT-2 Medium, T5-Base, and BART Large, we use NVIDIA GeForce RTX 8000 to train the model and for other smaller models, we use NVIDIA GeForce RTX 2080. Most of experiments could be done within one week. In all the inference time experiments, we use NVIDIA GeForce GTX TITAN X, batch size 4 for GPT-2, and batch size 8 for BART and T5.\"\n }\n]"},"abstract":{"kind":"string","value":"Is the output softmax layer, which is adopted by most language models (LMs), always the best way to compute the next word probability? Given so many attention layers in a modern transformer-based LM, are the pointer networks redundant nowadays? In this study, we discover that the answers to both questions are no. This is because the softmax bottleneck sometimes prevents the LMs from predicting the desired distribution and the pointer networks can be used to break the bottleneck efficiently. Based on the finding, we propose several softmax alternatives by simplifying the pointer networks and accelerating the word-by-word rerankers. In GPT-2, our proposals are significantly better and more efficient than mixture of softmax, a state-of-theart softmax alternative. In summarization experiments, without significantly decreasing its training/testing speed, our best method based on T5-Small improves factCC score by 2 points in CNN/DM and XSUM dataset, and improves MAUVE scores by 30% in Book-Sum paragraph-level dataset."},"title":{"kind":"string","value":"Revisiting the Architectures like Pointer Networks to Efficiently Improve the Next Word Distribution, Summarization Factuality, and Beyond"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 2 :2Figure 2: We simplify the pointer network / reranker by using another embedding h ct,S for the words in the context / the top-k likely words.","figure_data":"","figure_id":"fig_0","figure_label":"2","figure_type":"figure"},{"figure_caption":"(a) The example where the next word should be either woman or king (or their synonym such as former and latter). (b) The example where the next word plates, keys, scissors, toys, and balloons should receive similar probabilities. (c) The example where the next word John, Alex, Mary, Kathryn, and Jack should receive similar probabilities. (d) Same as above except that the order of the objects in the context is different.","figure_data":"","figure_id":"fig_1","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 4 :4Figure 4: The next word probabilities outputted by GPT-3.5 (text-davinci-003).","figure_data":"","figure_id":"fig_2","figure_label":"4","figure_type":"figure"},{"figure_caption":"Shwartz et al. (2020) andLadhak et al. (2023). Moreover, the meaning of tokens are also locally defined in many other applications such as variables in code or math problems, the new terminologies in a scientific paper, or the products in a sequential recommendation problem. We believe that our methods could become an efficient alternative of reranker(Cobbe et al., 2021; Welleck et al., 2022) and create impacts in those areas.","figure_data":"","figure_id":"fig_3","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 5 :5Figure 5: The model size versus the model loss in Wikipedia test data after training for 0.4 epochs. The left side points are the results from GPT-2 Small and the right side points come from GPT-2 Medium. The lower curves are better.","figure_data":"","figure_id":"fig_4","figure_label":"5","figure_type":"figure"},{"figure_caption":"Kaplan et al. (2020); Henighan et al. (","figure_data":"","figure_id":"fig_5","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 6 :6Figure 6: The model size versus the model loss in CNN/DM test set. The left side points are the results from T5-Small and the right side points come from T5-Base. The lower curves are better.","figure_data":"","figure_id":"fig_6","figure_label":"6","figure_type":"figure"},{"figure_caption":"P 1 I1rob(x|I, c t ) ∝ exp f T ct,V w x + |I| j=1 j =x exp f T ct,P E L f LE (h M I j ) + b . (8)","figure_data":"","figure_id":"fig_7","figure_label":"1","figure_type":"figure"},{"figure_caption":"Comparison of different methods on top of GPT-2. Wiki and OWT refer to the testing perplexity of Wikipedia 2021 and OpenWebText, respectively. Lower perplexity is better. Time is the inference time of a batch; Mi is the multiple input hidden state enhancement; C is the context partition; R:20,100 is the reranker partition with k Pointer Generator (PG + Mi) (See et al., 2017), and Pointer Sentinel (PS + Mi) (Merity et al., 2017). 2 Their similar performances indicate that the improvement of pointer networks come from breaking the softmax bottleneck. The significantly better performance of PS + Mi compared to PS further supports the finding.","figure_data":"GPT-2 SmallGPT-2 MediumModel NameSizeTime (ms) OWT (↓) Wiki (↓)SizeTime (ms) OWT (↓) Wiki (↓)Softmax (GPT-2)125.0M82.918.9624.28355.9M207.815.8120.12Softmax + Mi130.9M85.618.7424.08366.4M213.815.7120.07Mixture of Softmax (MoS) (Yang et al., 2018) 126.2M130.218.9724.10358.0M262.915.7119.95MoS + Mi (Chang and McCallum, 2022)133.3M133.218.6823.82370.6M268.215.6119.86Pointer Generator (PG) (See et al., 2017)126.2M106.018.6723.70358.0M237.815.7219.95Pointer Sentinel (PS) (Merity et al., 2017)126.2M94.118.7023.79358.0M218.315.7219.95Softmax + R:20 + Mi132.1M90.418.6724.03368.5M203.615.6419.94Softmax + R:20,100 + Mi133.3M101.118.6923.93370.6M228.515.6119.89Softmax + C + Mi132.1M94.818.4823.56368.5M222.715.6019.83Softmax + P + Mi133.3M99.118.5823.66370.6M214.715.6319.90PG + Mi133.3M111.218.4323.43370.6M242.515.6019.89PS + Mi133.3M98.018.4823.53370.6M224.615.6019.87Softmax + CR:20,100 + Mi134.5M113.318.4623.48372.7M234.515.5419.75Softmax + CPR:20,100 + Mi136.8M119.918.4323.42376.9M249.915.5319.71MoS + CPR:20,100 + Mi139.2M165.118.3923.29381.1M300.615.4419.57","figure_id":"tab_3","figure_label":"2","figure_type":"table"},{"figure_caption":"","figure_data":"AllProper NounModel NameRef Context Ref ContextSoftmax + Mi22.9024.047.4914.84MoS + Mi22.8823.987.7015.49PS + Mi22.8525.018.1618.21Softmax + CPR:20,100 + Mi 23.0525.368.1617.92","figure_id":"tab_5","figure_label":"3","figure_type":"table"},{"figure_caption":"","figure_data":"show that different meth-","figure_id":"tab_6","figure_label":"4","figure_type":"table"},{"figure_caption":"Prediction visualization of three input contexts. We show the top five words with the highest prediction probabilities of each model. The reasonable next word predictions are boldfaced.","figure_data":"There are plates, keys, scissors, toys,Choosing between John, Alex,I like tennis, baseball, golf, basketball,Input Contextand balloons in front of me, and IMary, Kathryn, and Jack, I decidedandpick up theto first talk toSoftmax + Mikeys 0.108, pieces 0.045, key 0.036, phone 0.020, balloons 0.019John 0.108, the 0.102, them 0.095, him 0.045, my 0.032tennis 0.089, baseball 0.075, football 0.041, basketball 0.036, I 0.032Mixture of Softmax (MoS) + Mikeys 0.085, phone 0.035, key 0.031, pieces 0.029, balloons 0.016John 0.099, the 0.097, them 0.083, Alex 0.055, Mary 0.040baseball 0.076, basketball 0.062, tennis 0.059, golf 0.037, bad 0.035Pointer Sentinel (PS) + Mikeys 0.091, plates 0.079, scissors 0.050, balloons 0.034, toys 0.033John 0.130, the 0.105, Alex 0.076, them 0.076, Mary 0.037tennis 0.095, golf 0.050, baseball 0.043, I 0.038, other 0.038Softmax + CPR:20,100 + Mikeys 0.077, balloons 0.052, plates 0.036, toys 0.030, pieces 0.030the 0.106, John 0.099, my 0.060, Alex 0.057, them 0.044football 0.075, volleyball 0.058, soccer 0.056, I 0.047, bad 0.038CNN/DMXSUMBookSum ParagraphSAMSUMModel NameR1CIDEr factCC MAUVER1CIDEr factCC MAUVER1CIDEr factCC MAUVER1CIDEr factCC MAUVET5-SmallSoftmax (S)38.255 0.4420.4620.86128.713 0.4460.2540.93916.313 0.0830.4240.32839.472 0.8170.5770.898CopyNet (Gu et al., 2016) 37.990 0.4380.4820.86528.573 0.4420.2740.94016.666 0.0920.4390.40239.525 0.8530.5790.924PG (See et al., 2017)37.913 0.4420.4670.87428.777 0.4500.2570.93116.432 0.0880.4290.37632.451 0.5850.5520.153PS (Merity et al., 2017)38.058 0.4440.4660.85428.442 0.4350.2670.93216.408 0.0900.4360.39538.731 0.8170.5780.865S + R:2037.881 0.4330.4740.87228.557 0.4400.2560.93116.336 0.0860.4310.37039.073 0.7520.5790.847S + E38.137 0.4410.4770.86628.723 0.4440.2720.94216.542 0.0900.4350.39039.056 0.7840.5790.904S + CE38.461 0.4600.4750.87429.155 0.4640.2700.94816.628 0.0930.4360.40340.055 0.8350.5830.943S + CER:2038.346 0.4500.4820.89029.067 0.4590.2760.94216.638 0.0930.4360.40040.505 0.8460.5800.915S + CEPR:2038.807 0.4560.4810.87729.395 0.4740.2730.94216.894 0.0980.4400.41840.127 0.8910.5820.946S + CEPR:20 + Mi38.675 0.4510.4750.87829.348 0.4700.2750.94616.738 0.0960.4380.42640.328 0.8740.5820.932T5-BaseSoftmax (S)40.198 0.5040.4780.90733.571 0.6670.2490.97916.761 0.0960.4240.46744.348 1.0460.5740.986CopyNet (Gu et al., 2016) 39.940 0.5070.4840.90333.557 0.6660.2530.97916.918 0.1010.4300.53144.141 1.0520.5700.973PG (See et al., 2017)39.982 0.4890.4850.91133.605 0.6630.2550.98216.611 0.0950.4230.46337.597 0.7840.5480.140PS (Merity et al., 2017)40.018 0.4950.4830.91433.638 0.6720.2490.98316.905 0.1000.4280.50443.098 1.0080.5750.946S + CEPR:2040.354 0.5110.4870.91933.700 0.6750.2600.98016.997 0.1000.4320.54944.860 1.0640.5730.963S + CEPR:20 + Mi40.510 0.5060.4810.91833.853 0.6830.2630.98316.975 0.1010.4310.54644.488 1.0550.5760.980BART BaseSoftmax (S)39.390 0.4280.4790.90035.675 0.8140.2410.98516.393 0.0940.4140.40445.132 1.1290.5670.966CopyNet (Gu et al., 2016) 39.385 0.4380.4840.90635.515 0.8140.2510.98816.642 0.1000.4220.49544.316 1.1030.5770.970PG (See et al., 2017)39.264 0.4440.4890.90935.653 0.8100.2420.98716.402 0.0940.4140.40245.278 1.1530.5780.977PS (Merity et al., 2017)39.471 0.4590.4900.90635.411 0.8090.2470.98616.718 0.0990.4220.49244.575 1.0840.5730.974S + R:2039.181 0.4340.4750.90535.586 0.8080.2470.98816.419 0.0960.4180.43945.024 1.1540.5720.970S + E39.267 0.4390.4830.90735.698 0.8190.2410.98816.442 0.0970.4150.42944.825 1.1060.5720.981S + CE39.416 0.4420.4810.90835.727 0.8120.2410.98816.555 0.0960.4170.43544.295 1.1160.5720.985S + CER:2039.421 0.4390.4820.90035.576 0.8120.2360.98716.553 0.0960.4180.45445.054 1.1500.5760.988S + CEPR:2039.723 0.4410.4830.90835.732 0.8220.2420.98616.664 0.0980.4200.46744.732 1.1150.5750.974S + CEPR:20 + Mi39.626 0.4420.4820.90735.846 0.8280.2450.98616.597 0.0970.4190.46644.728 1.1320.5740.988BART LargeSoftmax (S)40.749 0.4240.4950.89938.828 0.9210.2630.98817.271 0.1030.4200.46147.384 1.1870.5740.975CopyNet (Gu et al., 2016) 40.622 0.4070.4870.89038.576 0.9200.2580.98917.342 0.1060.4250.51247.911 1.2320.5730.980PG (See et al., 2017)40.766 0.4070.4890.90238.869 0.9440.2560.99017.289 0.1030.4240.47047.737 1.1990.5730.964PS (Merity et al., 2017)40.643 0.4240.5020.90738.886 0.9520.2550.98817.382 0.1050.4260.52748.253 1.2460.5740.986S + CEPR:2040.876 0.4580.5000.92538.991 0.9550.2480.99017.337 0.1060.4230.46747.253 1.2980.5720.976S + CEPR:20 + Mi40.441 0.4630.5000.92738.705 0.9650.2420.99116.995 0.1050.4210.48247.488 1.2710.5710.986","figure_id":"tab_8","figure_label":"5","figure_type":"table"},{"figure_caption":"","figure_data":"lihood training (Welleck et al., 2020; Jiang et al.,Jiang et al., 2022a). Although being effective, the2022b; Su et al., 2022). However, when LM shouldrerankers usually slow down significantly the train-mention some names in the context, this might ex-ing and/or inference speed (as our word-by-wordacerbate the hallucination problem. In contrast, ourreranker baseline) and might occupy extra memorymethod can learn to copy and exclude the words inresources.context as in Table 5.To alleviate the hallucination problem or sat-isfy some constraints, many recent generation mod-els rerank the generated text (Deng et al., 2020;Gabriel et al., 2021; Cobbe et al., 2021; Ravautet al., 2022; Krishna et al., 2022; Glass et al.,2022; An et al., 2022; Arora et al., 2022; Adolphset al., 2022; Meng et al., 2022; Mireshghallah et al.,2022; Kumar et al., 2022; Wan and Bansal, 2022;","figure_id":"tab_9","figure_label":"6","figure_type":"table"},{"figure_caption":"al., 2021; Zhong et al., 2022; Ma et al., 2023). Another example is encouraging the word embeddings to be isotropy (Wang et al., 2020; Su et al., 2022). Their improvement might also come from reducing linear dependency of the candidate word embeddings. Nevertheless, their side effect of breaking the similarity structure in the word embedding space might hurt the generation quality in some cases. Concurrently to our work, Wan et al. (2023) also use the softmax bottleneck theory","figure_data":"","figure_id":"tab_10","figure_label":"","figure_type":"table"},{"figure_caption":") and Lu et al. (2021).","figure_data":"3.20 3.25Softmax Softmax + Mi MoS + Mi Softmax + C + Mi PS + Mi Softmax + CPR:100,20 + Mi3.15loss3.103.053.0018.618.819.0 ln(number of parameters) 19.2 19.419.619.8","figure_id":"tab_11","figure_label":"","figure_type":"table"},{"figure_caption":"Diagonal (e.g., king or woman) Comparing the perplexity of different GPT-2 Small models using the synthetic dataset fromChang and McCallum (2022).","figure_data":"Edge (e.g., king or queen)","figure_id":"tab_12","figure_label":"7","figure_type":"table"},{"figure_caption":"Comparison of the continuation generated by GPT-2 Small in Wikipedia test data. Table 4 is a short summary of this table. The meaning of the metrics is described in Appendix B.1. Higher R1C and R1PC mean copying more words from the context. A higher P Ratio means generating more proper nouns. All ROUGE scores are percentages.","figure_data":"","figure_id":"tab_13","figure_label":"8","figure_type":"table"},{"figure_caption":"Comparison of the summaries generated by different models in the test sets of CNN/DM and XSUM datasets. We also report the number of parameters of each model. From top to bottom, the four sections are the results of T5-Small, T5-Base, BART Base, and BART Large. The meaning of the metrics are described in Appendix B.1. R2 (ROUGE 2-F1) scores are percentages. Within each section, we highlight the smallest loss, the P Ratio that is closest to 1, and highest numbers in the other metrics.","figure_data":"","figure_id":"tab_15","figure_label":"9","figure_type":"table"},{"figure_caption":"Comparison of the summaries generated by different models in the test sets of BookSum and SAM-SUM datasets. We also report the inference time of one samples. The meaning of the metrics are described in Appendix B.1. R2 (ROUGE 2-F1) scores are percentages. Within each section, we highlight the smallest loss, the P Ratio that is closest to 1, and highest numbers in the other metrics.","figure_data":"BookSum ParagraphSAMSUMModelsTime (ms) Loss (↓)R2R1P P Ratio NIST Loss (↓)R2R1P P Ratio NISTT5-SmallSoftmax (S)30.10.6541.673 0.149 0.589 1.3830.38313.806 0.605 0.873 3.945CopyNet (Gu et al., 2016)37.00.6461.722 0.183 0.747 1.4400.38114.210 0.594 0.809 3.965PG (See et al., 2017)43.40.6481.669 0.160 0.631 1.4130.39210.673 0.542 0.711 1.665PS (Merity et al., 2017)37.60.6461.627 0.177 0.700 1.4170.38313.817 0.583 0.794 3.960S + R:2032.90.6521.663 0.159 0.677 1.4030.38013.728 0.598 0.870 3.995S + E33.80.6451.710 0.171 0.673 1.4210.37013.557 0.602 0.892 3.906S + CE34.00.6441.734 0.173 0.680 1.4360.36814.136 0.619 0.892 3.971S + CER:2035.80.6421.710 0.174 0.693 1.4340.36714.281 0.627 0.911 3.968S + CEPR:2038.40.6411.768 0.184 0.725 1.4610.36514.451 0.639 0.909 4.034S + CEPR:20 + Mi41.70.6411.733 0.185 0.721 1.4580.36514.193 0.630 0.922 4.011T5-BaseSoftmax (S)102.40.5871.876 0.160 0.650 1.4430.30817.662 0.672 0.915 4.559CopyNet (Gu et al., 2016)0.5820.744 1.4810.30717.556 0.678 0.901 4.544PG (See et al., 2017)117.70.5851.832 0.159 0.647 1.4340.31714.649 0.611 0.740 1.870PS (Merity et al., 2017)112.00.5821.899 0.176 0.718 1.4650.30817.502 0.660 0.897 4.453S + CEPR:20115.30.5801.842 0.191 0.771 1.4820.30018.082 0.677 0.950 4.553S + CEPR:20 + Mi116.30.5841.860 0.187 0.770 1.4770.30117.617 0.677 0.938 4.521BART BaseSoftmax (S)46.60.6241.807 0.141 0.656 1.4250.32719.379 0.672 0.995 4.546CopyNet (Gu et al., 2016)57.80.6131.866 0.166 0.728 1.4540.32618.227 0.662 0.944 4.535PG (See et al., 2017)64.80.6241.864 0.140 0.668 1.4280.32818.791 0.673 0.963 4.537PS (Merity et al., 2017)57.90.6131.867 0.163 0.723 1.4610.32418.367 0.674 0.951 4.573S + R:2050.50.6271.807 0.154 0.720 1.4300.32619.022 0.671 0.971 4.608S + E54.20.6201.825 0.150 0.688 1.4290.32418.902 0.680 0.970 4.501S + CE56.50.6191.847 0.153 0.685 1.4410.32318.739 0.672 0.949 4.537S + CER:2057.20.6181.834 0.156 0.727 1.4440.32119.267 0.678 0.981 4.561S + CEPR:2058.80.6181.865 0.157 0.742 1.4570.32118.631 0.670 0.992 4.516S + CEPR:20 + Mi63.20.6201.827 0.158 0.733 1.4420.32218.681 0.670 0.987 4.439BART LargeSoftmax (S)143.50.5542.094 0.171 0.722 1.4720.30320.848 0.711 1.006 4.621CopyNet (Gu et al., 2016)168.90.5482.087 0.184 0.762 1.4900.29821.703 0.708 1.026 4.727PG (See et al., 2017)178.30.7312.090 0.174 0.725 1.4790.30121.428 0.706 1.051 4.604PS (Merity et al., 2017)168.50.7262.083 0.184 0.760 1.4930.30022.144 0.710 1.036 4.779S + CEPR:20169.90.5522.069 0.178 0.763 1.5050.30221.326 0.691 1.017 4.595S + CEPR:20 + Mi177.40.5442.024 0.175 0.737 1.5000.29421.244 0.713 0.959 4.746","figure_id":"tab_16","figure_label":"10","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 2 :2Figure 2: We simplify the pointer network / reranker by using another embedding h ct,S for the words in the context / the top-k likely words.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"(a) The example where the next word should be either woman or king (or their synonym such as former and latter). (b) The example where the next word plates, keys, scissors, toys, and balloons should receive similar probabilities. (c) The example where the next word John, Alex, Mary, Kathryn, and Jack should receive similar probabilities. (d) Same as above except that the order of the objects in the context is different.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 4 :4Figure 4: The next word probabilities outputted by GPT-3.5 (text-davinci-003).\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Shwartz et al. (2020) andLadhak et al. (2023). Moreover, the meaning of tokens are also locally defined in many other applications such as variables in code or math problems, the new terminologies in a scientific paper, or the products in a sequential recommendation problem. We believe that our methods could become an efficient alternative of reranker(Cobbe et al., 2021; Welleck et al., 2022) and create impacts in those areas.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 5 :5Figure 5: The model size versus the model loss in Wikipedia test data after training for 0.4 epochs. The left side points are the results from GPT-2 Small and the right side points come from GPT-2 Medium. The lower curves are better.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"5\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Kaplan et al. (2020); Henighan et al. (\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_5\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 6 :6Figure 6: The model size versus the model loss in CNN/DM test set. The left side points are the results from T5-Small and the right side points come from T5-Base. The lower curves are better.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_6\",\n \"figure_label\": \"6\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"P 1 I1rob(x|I, c t ) ∝ exp f T ct,V w x + |I| j=1 j =x exp f T ct,P E L f LE (h M I j ) + b . (8)\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_7\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Comparison of different methods on top of GPT-2. Wiki and OWT refer to the testing perplexity of Wikipedia 2021 and OpenWebText, respectively. Lower perplexity is better. Time is the inference time of a batch; Mi is the multiple input hidden state enhancement; C is the context partition; R:20,100 is the reranker partition with k Pointer Generator (PG + Mi) (See et al., 2017), and Pointer Sentinel (PS + Mi) (Merity et al., 2017). 2 Their similar performances indicate that the improvement of pointer networks come from breaking the softmax bottleneck. The significantly better performance of PS + Mi compared to PS further supports the finding.\",\n \"figure_data\": \"GPT-2 SmallGPT-2 MediumModel NameSizeTime (ms) OWT (↓) Wiki (↓)SizeTime (ms) OWT (↓) Wiki (↓)Softmax (GPT-2)125.0M82.918.9624.28355.9M207.815.8120.12Softmax + Mi130.9M85.618.7424.08366.4M213.815.7120.07Mixture of Softmax (MoS) (Yang et al., 2018) 126.2M130.218.9724.10358.0M262.915.7119.95MoS + Mi (Chang and McCallum, 2022)133.3M133.218.6823.82370.6M268.215.6119.86Pointer Generator (PG) (See et al., 2017)126.2M106.018.6723.70358.0M237.815.7219.95Pointer Sentinel (PS) (Merity et al., 2017)126.2M94.118.7023.79358.0M218.315.7219.95Softmax + R:20 + Mi132.1M90.418.6724.03368.5M203.615.6419.94Softmax + R:20,100 + Mi133.3M101.118.6923.93370.6M228.515.6119.89Softmax + C + Mi132.1M94.818.4823.56368.5M222.715.6019.83Softmax + P + Mi133.3M99.118.5823.66370.6M214.715.6319.90PG + Mi133.3M111.218.4323.43370.6M242.515.6019.89PS + Mi133.3M98.018.4823.53370.6M224.615.6019.87Softmax + CR:20,100 + Mi134.5M113.318.4623.48372.7M234.515.5419.75Softmax + CPR:20,100 + Mi136.8M119.918.4323.42376.9M249.915.5319.71MoS + CPR:20,100 + Mi139.2M165.118.3923.29381.1M300.615.4419.57\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"\",\n \"figure_data\": \"AllProper NounModel NameRef Context Ref ContextSoftmax + Mi22.9024.047.4914.84MoS + Mi22.8823.987.7015.49PS + Mi22.8525.018.1618.21Softmax + CPR:20,100 + Mi 23.0525.368.1617.92\",\n \"figure_id\": \"tab_5\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"\",\n \"figure_data\": \"show that different meth-\",\n \"figure_id\": \"tab_6\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Prediction visualization of three input contexts. We show the top five words with the highest prediction probabilities of each model. The reasonable next word predictions are boldfaced.\",\n \"figure_data\": \"There are plates, keys, scissors, toys,Choosing between John, Alex,I like tennis, baseball, golf, basketball,Input Contextand balloons in front of me, and IMary, Kathryn, and Jack, I decidedandpick up theto first talk toSoftmax + Mikeys 0.108, pieces 0.045, key 0.036, phone 0.020, balloons 0.019John 0.108, the 0.102, them 0.095, him 0.045, my 0.032tennis 0.089, baseball 0.075, football 0.041, basketball 0.036, I 0.032Mixture of Softmax (MoS) + Mikeys 0.085, phone 0.035, key 0.031, pieces 0.029, balloons 0.016John 0.099, the 0.097, them 0.083, Alex 0.055, Mary 0.040baseball 0.076, basketball 0.062, tennis 0.059, golf 0.037, bad 0.035Pointer Sentinel (PS) + Mikeys 0.091, plates 0.079, scissors 0.050, balloons 0.034, toys 0.033John 0.130, the 0.105, Alex 0.076, them 0.076, Mary 0.037tennis 0.095, golf 0.050, baseball 0.043, I 0.038, other 0.038Softmax + CPR:20,100 + Mikeys 0.077, balloons 0.052, plates 0.036, toys 0.030, pieces 0.030the 0.106, John 0.099, my 0.060, Alex 0.057, them 0.044football 0.075, volleyball 0.058, soccer 0.056, I 0.047, bad 0.038CNN/DMXSUMBookSum ParagraphSAMSUMModel NameR1CIDEr factCC MAUVER1CIDEr factCC MAUVER1CIDEr factCC MAUVER1CIDEr factCC MAUVET5-SmallSoftmax (S)38.255 0.4420.4620.86128.713 0.4460.2540.93916.313 0.0830.4240.32839.472 0.8170.5770.898CopyNet (Gu et al., 2016) 37.990 0.4380.4820.86528.573 0.4420.2740.94016.666 0.0920.4390.40239.525 0.8530.5790.924PG (See et al., 2017)37.913 0.4420.4670.87428.777 0.4500.2570.93116.432 0.0880.4290.37632.451 0.5850.5520.153PS (Merity et al., 2017)38.058 0.4440.4660.85428.442 0.4350.2670.93216.408 0.0900.4360.39538.731 0.8170.5780.865S + R:2037.881 0.4330.4740.87228.557 0.4400.2560.93116.336 0.0860.4310.37039.073 0.7520.5790.847S + E38.137 0.4410.4770.86628.723 0.4440.2720.94216.542 0.0900.4350.39039.056 0.7840.5790.904S + CE38.461 0.4600.4750.87429.155 0.4640.2700.94816.628 0.0930.4360.40340.055 0.8350.5830.943S + CER:2038.346 0.4500.4820.89029.067 0.4590.2760.94216.638 0.0930.4360.40040.505 0.8460.5800.915S + CEPR:2038.807 0.4560.4810.87729.395 0.4740.2730.94216.894 0.0980.4400.41840.127 0.8910.5820.946S + CEPR:20 + Mi38.675 0.4510.4750.87829.348 0.4700.2750.94616.738 0.0960.4380.42640.328 0.8740.5820.932T5-BaseSoftmax (S)40.198 0.5040.4780.90733.571 0.6670.2490.97916.761 0.0960.4240.46744.348 1.0460.5740.986CopyNet (Gu et al., 2016) 39.940 0.5070.4840.90333.557 0.6660.2530.97916.918 0.1010.4300.53144.141 1.0520.5700.973PG (See et al., 2017)39.982 0.4890.4850.91133.605 0.6630.2550.98216.611 0.0950.4230.46337.597 0.7840.5480.140PS (Merity et al., 2017)40.018 0.4950.4830.91433.638 0.6720.2490.98316.905 0.1000.4280.50443.098 1.0080.5750.946S + CEPR:2040.354 0.5110.4870.91933.700 0.6750.2600.98016.997 0.1000.4320.54944.860 1.0640.5730.963S + CEPR:20 + Mi40.510 0.5060.4810.91833.853 0.6830.2630.98316.975 0.1010.4310.54644.488 1.0550.5760.980BART BaseSoftmax (S)39.390 0.4280.4790.90035.675 0.8140.2410.98516.393 0.0940.4140.40445.132 1.1290.5670.966CopyNet (Gu et al., 2016) 39.385 0.4380.4840.90635.515 0.8140.2510.98816.642 0.1000.4220.49544.316 1.1030.5770.970PG (See et al., 2017)39.264 0.4440.4890.90935.653 0.8100.2420.98716.402 0.0940.4140.40245.278 1.1530.5780.977PS (Merity et al., 2017)39.471 0.4590.4900.90635.411 0.8090.2470.98616.718 0.0990.4220.49244.575 1.0840.5730.974S + R:2039.181 0.4340.4750.90535.586 0.8080.2470.98816.419 0.0960.4180.43945.024 1.1540.5720.970S + E39.267 0.4390.4830.90735.698 0.8190.2410.98816.442 0.0970.4150.42944.825 1.1060.5720.981S + CE39.416 0.4420.4810.90835.727 0.8120.2410.98816.555 0.0960.4170.43544.295 1.1160.5720.985S + CER:2039.421 0.4390.4820.90035.576 0.8120.2360.98716.553 0.0960.4180.45445.054 1.1500.5760.988S + CEPR:2039.723 0.4410.4830.90835.732 0.8220.2420.98616.664 0.0980.4200.46744.732 1.1150.5750.974S + CEPR:20 + Mi39.626 0.4420.4820.90735.846 0.8280.2450.98616.597 0.0970.4190.46644.728 1.1320.5740.988BART LargeSoftmax (S)40.749 0.4240.4950.89938.828 0.9210.2630.98817.271 0.1030.4200.46147.384 1.1870.5740.975CopyNet (Gu et al., 2016) 40.622 0.4070.4870.89038.576 0.9200.2580.98917.342 0.1060.4250.51247.911 1.2320.5730.980PG (See et al., 2017)40.766 0.4070.4890.90238.869 0.9440.2560.99017.289 0.1030.4240.47047.737 1.1990.5730.964PS (Merity et al., 2017)40.643 0.4240.5020.90738.886 0.9520.2550.98817.382 0.1050.4260.52748.253 1.2460.5740.986S + CEPR:2040.876 0.4580.5000.92538.991 0.9550.2480.99017.337 0.1060.4230.46747.253 1.2980.5720.976S + CEPR:20 + Mi40.441 0.4630.5000.92738.705 0.9650.2420.99116.995 0.1050.4210.48247.488 1.2710.5710.986\",\n \"figure_id\": \"tab_8\",\n \"figure_label\": \"5\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"\",\n \"figure_data\": \"lihood training (Welleck et al., 2020; Jiang et al.,Jiang et al., 2022a). Although being effective, the2022b; Su et al., 2022). However, when LM shouldrerankers usually slow down significantly the train-mention some names in the context, this might ex-ing and/or inference speed (as our word-by-wordacerbate the hallucination problem. In contrast, ourreranker baseline) and might occupy extra memorymethod can learn to copy and exclude the words inresources.context as in Table 5.To alleviate the hallucination problem or sat-isfy some constraints, many recent generation mod-els rerank the generated text (Deng et al., 2020;Gabriel et al., 2021; Cobbe et al., 2021; Ravautet al., 2022; Krishna et al., 2022; Glass et al.,2022; An et al., 2022; Arora et al., 2022; Adolphset al., 2022; Meng et al., 2022; Mireshghallah et al.,2022; Kumar et al., 2022; Wan and Bansal, 2022;\",\n \"figure_id\": \"tab_9\",\n \"figure_label\": \"6\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"al., 2021; Zhong et al., 2022; Ma et al., 2023). Another example is encouraging the word embeddings to be isotropy (Wang et al., 2020; Su et al., 2022). Their improvement might also come from reducing linear dependency of the candidate word embeddings. Nevertheless, their side effect of breaking the similarity structure in the word embedding space might hurt the generation quality in some cases. Concurrently to our work, Wan et al. (2023) also use the softmax bottleneck theory\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_10\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \") and Lu et al. (2021).\",\n \"figure_data\": \"3.20 3.25Softmax Softmax + Mi MoS + Mi Softmax + C + Mi PS + Mi Softmax + CPR:100,20 + Mi3.15loss3.103.053.0018.618.819.0 ln(number of parameters) 19.2 19.419.619.8\",\n \"figure_id\": \"tab_11\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Diagonal (e.g., king or woman) Comparing the perplexity of different GPT-2 Small models using the synthetic dataset fromChang and McCallum (2022).\",\n \"figure_data\": \"Edge (e.g., king or queen)\",\n \"figure_id\": \"tab_12\",\n \"figure_label\": \"7\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Comparison of the continuation generated by GPT-2 Small in Wikipedia test data. Table 4 is a short summary of this table. The meaning of the metrics is described in Appendix B.1. Higher R1C and R1PC mean copying more words from the context. A higher P Ratio means generating more proper nouns. All ROUGE scores are percentages.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_13\",\n \"figure_label\": \"8\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Comparison of the summaries generated by different models in the test sets of CNN/DM and XSUM datasets. We also report the number of parameters of each model. From top to bottom, the four sections are the results of T5-Small, T5-Base, BART Base, and BART Large. The meaning of the metrics are described in Appendix B.1. R2 (ROUGE 2-F1) scores are percentages. Within each section, we highlight the smallest loss, the P Ratio that is closest to 1, and highest numbers in the other metrics.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_15\",\n \"figure_label\": \"9\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Comparison of the summaries generated by different models in the test sets of BookSum and SAM-SUM datasets. We also report the inference time of one samples. The meaning of the metrics are described in Appendix B.1. R2 (ROUGE 2-F1) scores are percentages. Within each section, we highlight the smallest loss, the P Ratio that is closest to 1, and highest numbers in the other metrics.\",\n \"figure_data\": \"BookSum ParagraphSAMSUMModelsTime (ms) Loss (↓)R2R1P P Ratio NIST Loss (↓)R2R1P P Ratio NISTT5-SmallSoftmax (S)30.10.6541.673 0.149 0.589 1.3830.38313.806 0.605 0.873 3.945CopyNet (Gu et al., 2016)37.00.6461.722 0.183 0.747 1.4400.38114.210 0.594 0.809 3.965PG (See et al., 2017)43.40.6481.669 0.160 0.631 1.4130.39210.673 0.542 0.711 1.665PS (Merity et al., 2017)37.60.6461.627 0.177 0.700 1.4170.38313.817 0.583 0.794 3.960S + R:2032.90.6521.663 0.159 0.677 1.4030.38013.728 0.598 0.870 3.995S + E33.80.6451.710 0.171 0.673 1.4210.37013.557 0.602 0.892 3.906S + CE34.00.6441.734 0.173 0.680 1.4360.36814.136 0.619 0.892 3.971S + CER:2035.80.6421.710 0.174 0.693 1.4340.36714.281 0.627 0.911 3.968S + CEPR:2038.40.6411.768 0.184 0.725 1.4610.36514.451 0.639 0.909 4.034S + CEPR:20 + Mi41.70.6411.733 0.185 0.721 1.4580.36514.193 0.630 0.922 4.011T5-BaseSoftmax (S)102.40.5871.876 0.160 0.650 1.4430.30817.662 0.672 0.915 4.559CopyNet (Gu et al., 2016)0.5820.744 1.4810.30717.556 0.678 0.901 4.544PG (See et al., 2017)117.70.5851.832 0.159 0.647 1.4340.31714.649 0.611 0.740 1.870PS (Merity et al., 2017)112.00.5821.899 0.176 0.718 1.4650.30817.502 0.660 0.897 4.453S + CEPR:20115.30.5801.842 0.191 0.771 1.4820.30018.082 0.677 0.950 4.553S + CEPR:20 + Mi116.30.5841.860 0.187 0.770 1.4770.30117.617 0.677 0.938 4.521BART BaseSoftmax (S)46.60.6241.807 0.141 0.656 1.4250.32719.379 0.672 0.995 4.546CopyNet (Gu et al., 2016)57.80.6131.866 0.166 0.728 1.4540.32618.227 0.662 0.944 4.535PG (See et al., 2017)64.80.6241.864 0.140 0.668 1.4280.32818.791 0.673 0.963 4.537PS (Merity et al., 2017)57.90.6131.867 0.163 0.723 1.4610.32418.367 0.674 0.951 4.573S + R:2050.50.6271.807 0.154 0.720 1.4300.32619.022 0.671 0.971 4.608S + E54.20.6201.825 0.150 0.688 1.4290.32418.902 0.680 0.970 4.501S + CE56.50.6191.847 0.153 0.685 1.4410.32318.739 0.672 0.949 4.537S + CER:2057.20.6181.834 0.156 0.727 1.4440.32119.267 0.678 0.981 4.561S + CEPR:2058.80.6181.865 0.157 0.742 1.4570.32118.631 0.670 0.992 4.516S + CEPR:20 + Mi63.20.6201.827 0.158 0.733 1.4420.32218.681 0.670 0.987 4.439BART LargeSoftmax (S)143.50.5542.094 0.171 0.722 1.4720.30320.848 0.711 1.006 4.621CopyNet (Gu et al., 2016)168.90.5482.087 0.184 0.762 1.4900.29821.703 0.708 1.026 4.727PG (See et al., 2017)178.30.7312.090 0.174 0.725 1.4790.30121.428 0.706 1.051 4.604PS (Merity et al., 2017)168.50.7262.083 0.184 0.760 1.4930.30022.144 0.710 1.036 4.779S + CEPR:20169.90.5522.069 0.178 0.763 1.5050.30221.326 0.691 1.017 4.595S + CEPR:20 + Mi177.40.5442.024 0.175 0.737 1.5000.29421.244 0.713 0.959 4.746\",\n \"figure_id\": \"tab_16\",\n \"figure_label\": \"10\",\n \"figure_type\": \"table\"\n }\n]"},"authors":{"kind":"string","value":"Haw-Shiuan Chang; Zonghai Yao; Alolika Gon; Hong Yu; Andrew Mccallum; ‡ Cics"},"references":{"kind":"list like","value":[{"authors":"Leonard Adolphs; Tianyu Gao; Jing Xu; Kurt Shuster; Sainbayar Sukhbaatar; Jason Weston","journal":"","ref_id":"b0","title":"The cringe loss: Learning what language not to model","year":"2022"},{"authors":"Chenxin An; Jiangtao Feng; Kai Lv; Lingpeng Kong; Xipeng Qiu; Xuanjing Huang","journal":"","ref_id":"b1","title":"Cont: Contrastive neural text generation","year":"2022"},{"authors":"Kushal Arora; Kurt Shuster; Sainbayar Sukhbaatar; Jason Weston","journal":"","ref_id":"b2","title":"Director: Generator-classifiers for supervised language modeling","year":"2022"},{"authors":"B Tom; Benjamin Brown; Nick Mann; Melanie Ryder; Jared Subbiah; Prafulla Kaplan; Arvind Dhariwal; Pranav Neelakantan; Girish Shyam; Amanda Sastry; Sandhini Askell; Ariel Agarwal; Gretchen Herbert-Voss; Tom Krueger; Rewon Henighan; Aditya Child; Daniel M Ramesh; Jeffrey Ziegler; Clemens Wu; Christopher Winter; Mark Hesse; Eric Chen; Mateusz Sigler; Scott Litwin; Benjamin Gray; Jack Chess; Christopher Clark; Sam Berner; Alec Mc-Candlish; Ilya Radford; Dario Sutskever; Amodei","journal":"","ref_id":"b3","title":"Language models are few-shot learners","year":"2020-12-06"},{"authors":"Shuyang Cao; Lu Wang","journal":"Association for Computational Linguistics","ref_id":"b4","title":"CLIFF: Contrastive learning for improving faithfulness and factuality in abstractive summarization","year":"2021"},{"authors":"Haw-Shiuan Chang; Andrew Mccallum","journal":"Association for Computational Linguistics","ref_id":"b5","title":"Softmax bottleneck makes language models unable to represent multi-mode word distributions","year":"2022"},{"authors":"Karl Cobbe; Vineet Kosaraju; Mohammad Bavarian; Mark Chen; Heewoo Jun; Lukasz Kaiser; Matthias Plappert; Jerry Tworek; Jacob Hilton; Reiichiro Nakano","journal":"","ref_id":"b6","title":"Training verifiers to solve math word problems","year":"2021"},{"authors":"Yuntian Deng; Anton Bakhtin; Myle Ott; Arthur Szlam; Marc'aurelio Ranzato","journal":"","ref_id":"b7","title":"Residual energybased models for text generation","year":"2020-04-26"},{"authors":"George Doddington","journal":"","ref_id":"b8","title":"Automatic evaluation of machine translation quality using n-gram cooccurrence statistics","year":"2002"},{"authors":"Saadia Gabriel; Antoine Bosselut; Jeff Da; Ari Holtzman; Jan Buys; Kyle Lo; Asli Celikyilmaz; Yejin Choi","journal":"","ref_id":"b9","title":"Discourse understanding and factual consistency in abstractive summarization","year":"2021"},{"authors":"Michael Glass; Gaetano Rossiello; Md Faisal; Mahbub Chowdhury; Ankita Naik; Pengshan Cai; Alfio Gliozzo","journal":"Association for Computational Linguistics","ref_id":"b10","title":"Re2G: Retrieve, rerank, generate","year":"2022"},{"authors":"Bogdan Gliwa; Iwona Mochol; Maciej Biesek; Aleksander Wawer","journal":"Association for Computational Linguistics","ref_id":"b11","title":"SAMSum corpus: A human-annotated dialogue dataset for abstractive summarization","year":"2019"},{"authors":"Seraphina Goldfarb-Tarrant; Haining Feng; Nanyun Peng","journal":"Association for Computational Linguistics","ref_id":"b12","title":"Plan, write, and revise: an interactive system for open-domain story generation","year":"2019"},{"authors":"Tanya Goyal; Junyi ; Jessy Li; Greg Durrett","journal":"","ref_id":"b13","title":"News summarization and evaluation in the era of gpt-3","year":"2022"},{"authors":"Tanya Goyal; Junyi ; Jessy Li; Greg Durrett","journal":"","ref_id":"b14","title":"Snac: Coherence error detection for narrative summarization","year":"2022"},{"authors":"Jiatao Gu; Zhengdong Lu; Hang Li; O K Victor; Li","journal":"Association for Computational Linguistics","ref_id":"b15","title":"Incorporating copying mechanism in sequence-to-sequence learning","year":"2016"},{"authors":"Jian Guan; Zhenyu Yang; Rongsheng Zhang; Zhipeng Hu; Minlie Huang","journal":"","ref_id":"b16","title":"Generating coherent narratives by learning dynamic and discrete entity states with a contrastive framework","year":"2022"},{"authors":"Tom Henighan; Jared Kaplan; Mor Katz; Mark Chen; Christopher Hesse; Jacob Jackson; Heewoo Jun; Prafulla Tom B Brown; Scott Dhariwal; Gray","journal":"","ref_id":"b17","title":"Scaling laws for autoregressive generative modeling","year":"2020"},{"authors":"Sepp Hochreiter; Jürgen Schmidhuber","journal":"Neural computation","ref_id":"b18","title":"Long short-term memory","year":"1997"},{"authors":"Matthew Honnibal; Ines Montani; Sofie Van Landeghem; Adriane Boyd","journal":"","ref_id":"b19","title":"spaCy: Industrial-strength Natural Language Processing in Python","year":"2020"},{"authors":"Dongfu Jiang; Bill Yuchen Lin; Xiang Ren","journal":"","ref_id":"b20","title":"Pairreranker: Pairwise reranking for natural language generation","year":"2022"},{"authors":"Shaojie Jiang; Ruqing Zhang; Svitlana Vakulenko; Maarten De Rijke","journal":"","ref_id":"b21","title":"A simple contrastive learning objective for alleviating neural text degeneration","year":"2022"},{"authors":"Jared Kaplan; Sam Mccandlish; Tom Henighan; Tom B Brown; Benjamin Chess; Rewon Child; Scott Gray; Alec Radford; Jeffrey Wu; Dario Amodei","journal":"","ref_id":"b22","title":"Scaling laws for neural language models","year":"2020"},{"authors":"Marzena Karpinska; Nader Akoury; Mohit Iyyer","journal":"Association for Computational Linguistics","ref_id":"b23","title":"The perils of using Mechanical Turk to evaluate open-ended text generation","year":"2021"},{"authors":"Daniel King; Zejiang Shen; Nishant Subramani; Iz Daniel S Weld; Doug Beltagy; Downey","journal":"","ref_id":"b24","title":"Don't say what you don't know: Improving the consistency of abstractive summarization by constraining beam search","year":"2022"},{"authors":"Kalpesh Krishna; Yapei Chang; John Wieting; Mohit Iyyer","journal":"","ref_id":"b25","title":"Rankgen: Improving text generation with large ranking models","year":"2022"},{"authors":"Wojciech Kryscinski; Bryan Mccann; Caiming Xiong; Richard Socher","journal":"Association for Computational Linguistics","ref_id":"b26","title":"Evaluating the factual consistency of abstractive text summarization","year":"2020"},{"authors":"Wojciech Kryściński; Nazneen Rajani; Divyansh Agarwal; Caiming Xiong; Dragomir Radev","journal":"","ref_id":"b27","title":"BookSum: A collection of datasets for long-form narrative summarization","year":"2021"},{"authors":"Sachin Kumar; Biswajit Paria; Yulia Tsvetkov","journal":"","ref_id":"b28","title":"Gradient-based constrained sampling from language models","year":"2022"},{"authors":"Faisal Ladhak; Esin Durmus; He He; Claire Cardie; Kathleen Mckeown","journal":"Association for Computational Linguistics","ref_id":"b29","title":"Faithful or extractive? on mitigating the faithfulness-abstractiveness tradeoff in abstractive summarization","year":"2022"},{"authors":"Faisal Ladhak; Esin Durmus; Mirac Suzgun; Tianyi Zhang; Dan Jurafsky; Kathleen Mckeown; Tatsunori B Hashimoto","journal":"","ref_id":"b30","title":"When do pre-training biases propagate to downstream tasks? a case study in text summarization","year":"2023"},{"authors":"Mike Lewis; Yinhan Liu; Naman Goyal; Marjan Ghazvininejad; Abdelrahman Mohamed; Omer Levy; Veselin Stoyanov; Luke Zettlemoyer","journal":"Association for Computational Linguistics","ref_id":"b31","title":"BART: Denoising sequence-to-sequence pretraining for natural language generation, translation, and comprehension","year":"2020"},{"authors":"Haoran Li; Song Xu; Peng Yuan; Yujia Wang; Youzheng Wu; Xiaodong He; Bowen Zhou","journal":"","ref_id":"b32","title":"Learn to copy from the copying history: Correlational copy network for abstractive summarization","year":"2021"},{"authors":"Wei Li; Wenhao Wu; Moye Chen; Jiachen Liu; Xinyan Xiao; Hua Wu","journal":"","ref_id":"b33","title":"Faithfulness in natural language generation: A systematic survey of analysis, evaluation and optimization methods","year":"2022"},{"authors":"Chin-Yew Lin","journal":"Association for Computational Linguistics","ref_id":"b34","title":"ROUGE: A package for automatic evaluation of summaries","year":"2004"},{"authors":"Ximing Lu; Peter West; Rowan Zellers; Le Ronan; Chandra Bras; Yejin Bhagavatula; Choi","journal":"Association for Computational Linguistics","ref_id":"b35","title":"Neuro-Logic decoding: (un)supervised neural text generation with predicate logic constraints","year":"2021"},{"authors":"Yeyun Xinbei ; Ma; Pengcheng Gong; Hai He; Nan Zhao; Duan","journal":"","ref_id":"b36","title":"Prom: A phrase-level copying mechanism with pre-training for abstractive summarization","year":"2023"},{"authors":"Tao Meng; Sidi Lu; Nanyun Peng; Kai-Wei Chang","journal":"","ref_id":"b37","title":"Controllable text generation with neurally-decomposed oracle","year":"2022"},{"authors":"Stephen Merity; Caiming Xiong; James Bradbury; Richard Socher","journal":"","ref_id":"b38","title":"Pointer sentinel mixture models","year":"2017-04-24"},{"authors":"Tomás Mikolov; Ilya Sutskever; Kai Chen; Gregory S Corrado; Jeffrey Dean","journal":"","ref_id":"b39","title":"Distributed representations of words and phrases and their compositionality","year":"2013-12-05"},{"authors":"Fatemehsadat Mireshghallah; Kartik Goyal; Taylor Berg-Kirkpatrick","journal":"Association for Computational Linguistics","ref_id":"b40","title":"Mix and match: Learningfree controllable text generationusing energy language models","year":"2022"},{"authors":"Shashi Narayan; Shay B Cohen; Mirella Lapata","journal":"Association for Computational Linguistics","ref_id":"b41","title":"Don't give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization","year":"2018"},{"authors":"Pinelopi Papalampidi; Kris Cao; Tomas Kocisky","journal":"","ref_id":"b42","title":"Towards coherent and consistent use of entities in narrative generation","year":"2022"},{"authors":"Krishna Pillutla; Swabha Swayamdipta; Rowan Zellers; John Thickstun; Sean Welleck; Yejin Choi; Zaid Harchaoui","journal":"Advances in Neural Information Processing Systems","ref_id":"b43","title":"MAUVE: Measuring the gap between neural text and human text using divergence frontiers","year":"2021"},{"authors":"Alec Radford; Jeffrey Wu; Rewon Child; David Luan; Dario Amodei; Ilya Sutskever","journal":"","ref_id":"b44","title":"Language models are unsupervised multitask learners","year":"2019"},{"authors":"Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu","journal":"Journal of Machine Learning Research","ref_id":"b45","title":"Exploring the limits of transfer learning with a unified text-totext transformer","year":"2020"},{"authors":"Shafiq Mathieu Ravaut; Nancy Joty; Chen","journal":"Association for Computational Linguistics","ref_id":"b46","title":"SummaReranker: A multi-task mixture-of-experts re-ranking framework for abstractive summarization","year":"2022"},{"authors":"Abigail See; Peter J Liu; Christopher D Manning","journal":"Association for Computational Linguistics","ref_id":"b47","title":"Get to the point: Summarization with pointergenerator networks","year":"2017"},{"authors":"Abigail See; Aneesh Pappu; Rohun Saxena; Akhila Yerukola; Christopher D Manning","journal":"Association for Computational Linguistics","ref_id":"b48","title":"Do massively pretrained language models make better storytellers?","year":"2019"},{"authors":"Kurt Shuster; Jack Urbanek; Arthur Szlam; Jason Weston","journal":"Association for Computational Linguistics","ref_id":"b49","title":"Am I me or you? state-of-the-art dialogue models cannot maintain an identity","year":"2022"},{"authors":"Vered Shwartz; Rachel Rudinger; Oyvind Tafjord","journal":"Association for Computational Linguistics","ref_id":"b50","title":"you are grounded!\": Latent name artifacts in pre-trained language models","year":"2020"},{"authors":"Yixuan Su; Tian Lan; Yan Wang; Dani Yogatama; Lingpeng Kong; Nigel Collier","journal":"","ref_id":"b51","title":"A contrastive framework for neural text generation","year":"2022"},{"authors":"Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Lukasz Kaiser; Illia Polosukhin","journal":"","ref_id":"b52","title":"Attention is all you need","year":"2017-09"},{"authors":"C Lawrence Ramakrishna Vedantam; Devi Zitnick; Parikh","journal":"IEEE Computer Society","ref_id":"b53","title":"Cider: Consensus-based image description evaluation","year":"2015-06-07"},{"authors":"David Wan; Mohit Bansal","journal":"","ref_id":"b54","title":"Factpegasus: Factuality-aware pre-training and fine-tuning for abstractive summarization","year":"2022"},{"authors":"David Wan; Shiyue Zhang; Mohit Bansal","journal":"","ref_id":"b55","title":"Histalign: Improving context dependency in language generation by aligning with history","year":"2023"},{"authors":"Lingxiao Wang; Jing Huang; Kevin Huang; Ziniu Hu; Guangtao Wang; Quanquan Gu","journal":"","ref_id":"b56","title":"Improving neural language generation with spectrum control","year":"2020-04-26"},{"authors":"Sean Welleck; Ilia Kulikov; Stephen Roller; Emily Dinan; Kyunghyun Cho; Jason Weston","journal":"","ref_id":"b57","title":"Neural text generation with unlikelihood training","year":"2020-04-26"},{"authors":"Sean Welleck; Jiacheng Liu; Ximing Lu; Hannaneh Hajishirzi; Yejin Choi","journal":"","ref_id":"b58","title":"Naturalprover: Grounded mathematical proof generation with language models","year":"2022"},{"authors":"Thomas Wolf; Lysandre Debut; Victor Sanh; Julien Chaumond; Clement Delangue; Anthony Moi; Pierric Cistac; Tim Rault; Remi Louf; Morgan Funtowicz; Joe Davison; Sam Shleifer; Clara Patrick Von Platen; Yacine Ma; Julien Jernite; Canwen Plu; Teven Xu; Sylvain Le Scao; Mariama Gugger; Quentin Drame; Alexander Lhoest; Rush","journal":"","ref_id":"b59","title":"Transformers: State-of-the-art natural language processing","year":"2020"},{"authors":"Zhilin Yang; Zihang Dai; Ruslan Salakhutdinov; William W Cohen","journal":"","ref_id":"b60","title":"Breaking the softmax bottleneck: A high-rank RNN language model","year":"2018-04-30"},{"authors":"Haopeng Zhang; Semih Yavuz; Wojciech Kryscinski; Kazuma Hashimoto; Yingbo Zhou","journal":"Association for Computational Linguistics","ref_id":"b61","title":"Improving the faithfulness of abstractive summarization via entity coverage control","year":"2022"},{"authors":"Zexuan Zhong; Tao Lei; Danqi Chen","journal":"Association for Computational Linguistics","ref_id":"b62","title":"Training language models with memory augmentation","year":"2022-12-07"}],"string":"[\n {\n \"authors\": \"Leonard Adolphs; Tianyu Gao; Jing Xu; Kurt Shuster; Sainbayar Sukhbaatar; Jason Weston\",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"The cringe loss: Learning what language not to model\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Chenxin An; Jiangtao Feng; Kai Lv; Lingpeng Kong; Xipeng Qiu; Xuanjing Huang\",\n \"journal\": \"\",\n \"ref_id\": \"b1\",\n \"title\": \"Cont: Contrastive neural text generation\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Kushal Arora; Kurt Shuster; Sainbayar Sukhbaatar; Jason Weston\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"Director: Generator-classifiers for supervised language modeling\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"B Tom; Benjamin Brown; Nick Mann; Melanie Ryder; Jared Subbiah; Prafulla Kaplan; Arvind Dhariwal; Pranav Neelakantan; Girish Shyam; Amanda Sastry; Sandhini Askell; Ariel Agarwal; Gretchen Herbert-Voss; Tom Krueger; Rewon Henighan; Aditya Child; Daniel M Ramesh; Jeffrey Ziegler; Clemens Wu; Christopher Winter; Mark Hesse; Eric Chen; Mateusz Sigler; Scott Litwin; Benjamin Gray; Jack Chess; Christopher Clark; Sam Berner; Alec Mc-Candlish; Ilya Radford; Dario Sutskever; Amodei\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"Language models are few-shot learners\",\n \"year\": \"2020-12-06\"\n },\n {\n \"authors\": \"Shuyang Cao; Lu Wang\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b4\",\n \"title\": \"CLIFF: Contrastive learning for improving faithfulness and factuality in abstractive summarization\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Haw-Shiuan Chang; Andrew Mccallum\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b5\",\n \"title\": \"Softmax bottleneck makes language models unable to represent multi-mode word distributions\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Karl Cobbe; Vineet Kosaraju; Mohammad Bavarian; Mark Chen; Heewoo Jun; Lukasz Kaiser; Matthias Plappert; Jerry Tworek; Jacob Hilton; Reiichiro Nakano\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"Training verifiers to solve math word problems\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Yuntian Deng; Anton Bakhtin; Myle Ott; Arthur Szlam; Marc'aurelio Ranzato\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"Residual energybased models for text generation\",\n \"year\": \"2020-04-26\"\n },\n {\n \"authors\": \"George Doddington\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"Automatic evaluation of machine translation quality using n-gram cooccurrence statistics\",\n \"year\": \"2002\"\n },\n {\n \"authors\": \"Saadia Gabriel; Antoine Bosselut; Jeff Da; Ari Holtzman; Jan Buys; Kyle Lo; Asli Celikyilmaz; Yejin Choi\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"Discourse understanding and factual consistency in abstractive summarization\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Michael Glass; Gaetano Rossiello; Md Faisal; Mahbub Chowdhury; Ankita Naik; Pengshan Cai; Alfio Gliozzo\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b10\",\n \"title\": \"Re2G: Retrieve, rerank, generate\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Bogdan Gliwa; Iwona Mochol; Maciej Biesek; Aleksander Wawer\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b11\",\n \"title\": \"SAMSum corpus: A human-annotated dialogue dataset for abstractive summarization\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Seraphina Goldfarb-Tarrant; Haining Feng; Nanyun Peng\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b12\",\n \"title\": \"Plan, write, and revise: an interactive system for open-domain story generation\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Tanya Goyal; Junyi ; Jessy Li; Greg Durrett\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"News summarization and evaluation in the era of gpt-3\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Tanya Goyal; Junyi ; Jessy Li; Greg Durrett\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"Snac: Coherence error detection for narrative summarization\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jiatao Gu; Zhengdong Lu; Hang Li; O K Victor; Li\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b15\",\n \"title\": \"Incorporating copying mechanism in sequence-to-sequence learning\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Jian Guan; Zhenyu Yang; Rongsheng Zhang; Zhipeng Hu; Minlie Huang\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"Generating coherent narratives by learning dynamic and discrete entity states with a contrastive framework\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Tom Henighan; Jared Kaplan; Mor Katz; Mark Chen; Christopher Hesse; Jacob Jackson; Heewoo Jun; Prafulla Tom B Brown; Scott Dhariwal; Gray\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"Scaling laws for autoregressive generative modeling\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Sepp Hochreiter; Jürgen Schmidhuber\",\n \"journal\": \"Neural computation\",\n \"ref_id\": \"b18\",\n \"title\": \"Long short-term memory\",\n \"year\": \"1997\"\n },\n {\n \"authors\": \"Matthew Honnibal; Ines Montani; Sofie Van Landeghem; Adriane Boyd\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"spaCy: Industrial-strength Natural Language Processing in Python\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Dongfu Jiang; Bill Yuchen Lin; Xiang Ren\",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"Pairreranker: Pairwise reranking for natural language generation\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Shaojie Jiang; Ruqing Zhang; Svitlana Vakulenko; Maarten De Rijke\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"A simple contrastive learning objective for alleviating neural text degeneration\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jared Kaplan; Sam Mccandlish; Tom Henighan; Tom B Brown; Benjamin Chess; Rewon Child; Scott Gray; Alec Radford; Jeffrey Wu; Dario Amodei\",\n \"journal\": \"\",\n \"ref_id\": \"b22\",\n \"title\": \"Scaling laws for neural language models\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Marzena Karpinska; Nader Akoury; Mohit Iyyer\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b23\",\n \"title\": \"The perils of using Mechanical Turk to evaluate open-ended text generation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Daniel King; Zejiang Shen; Nishant Subramani; Iz Daniel S Weld; Doug Beltagy; Downey\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Don't say what you don't know: Improving the consistency of abstractive summarization by constraining beam search\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Kalpesh Krishna; Yapei Chang; John Wieting; Mohit Iyyer\",\n \"journal\": \"\",\n \"ref_id\": \"b25\",\n \"title\": \"Rankgen: Improving text generation with large ranking models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Wojciech Kryscinski; Bryan Mccann; Caiming Xiong; Richard Socher\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b26\",\n \"title\": \"Evaluating the factual consistency of abstractive text summarization\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Wojciech Kryściński; Nazneen Rajani; Divyansh Agarwal; Caiming Xiong; Dragomir Radev\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"BookSum: A collection of datasets for long-form narrative summarization\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Sachin Kumar; Biswajit Paria; Yulia Tsvetkov\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"Gradient-based constrained sampling from language models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Faisal Ladhak; Esin Durmus; He He; Claire Cardie; Kathleen Mckeown\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b29\",\n \"title\": \"Faithful or extractive? on mitigating the faithfulness-abstractiveness tradeoff in abstractive summarization\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Faisal Ladhak; Esin Durmus; Mirac Suzgun; Tianyi Zhang; Dan Jurafsky; Kathleen Mckeown; Tatsunori B Hashimoto\",\n \"journal\": \"\",\n \"ref_id\": \"b30\",\n \"title\": \"When do pre-training biases propagate to downstream tasks? a case study in text summarization\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Mike Lewis; Yinhan Liu; Naman Goyal; Marjan Ghazvininejad; Abdelrahman Mohamed; Omer Levy; Veselin Stoyanov; Luke Zettlemoyer\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b31\",\n \"title\": \"BART: Denoising sequence-to-sequence pretraining for natural language generation, translation, and comprehension\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Haoran Li; Song Xu; Peng Yuan; Yujia Wang; Youzheng Wu; Xiaodong He; Bowen Zhou\",\n \"journal\": \"\",\n \"ref_id\": \"b32\",\n \"title\": \"Learn to copy from the copying history: Correlational copy network for abstractive summarization\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Wei Li; Wenhao Wu; Moye Chen; Jiachen Liu; Xinyan Xiao; Hua Wu\",\n \"journal\": \"\",\n \"ref_id\": \"b33\",\n \"title\": \"Faithfulness in natural language generation: A systematic survey of analysis, evaluation and optimization methods\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Chin-Yew Lin\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b34\",\n \"title\": \"ROUGE: A package for automatic evaluation of summaries\",\n \"year\": \"2004\"\n },\n {\n \"authors\": \"Ximing Lu; Peter West; Rowan Zellers; Le Ronan; Chandra Bras; Yejin Bhagavatula; Choi\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b35\",\n \"title\": \"Neuro-Logic decoding: (un)supervised neural text generation with predicate logic constraints\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Yeyun Xinbei ; Ma; Pengcheng Gong; Hai He; Nan Zhao; Duan\",\n \"journal\": \"\",\n \"ref_id\": \"b36\",\n \"title\": \"Prom: A phrase-level copying mechanism with pre-training for abstractive summarization\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Tao Meng; Sidi Lu; Nanyun Peng; Kai-Wei Chang\",\n \"journal\": \"\",\n \"ref_id\": \"b37\",\n \"title\": \"Controllable text generation with neurally-decomposed oracle\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Stephen Merity; Caiming Xiong; James Bradbury; Richard Socher\",\n \"journal\": \"\",\n \"ref_id\": \"b38\",\n \"title\": \"Pointer sentinel mixture models\",\n \"year\": \"2017-04-24\"\n },\n {\n \"authors\": \"Tomás Mikolov; Ilya Sutskever; Kai Chen; Gregory S Corrado; Jeffrey Dean\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"Distributed representations of words and phrases and their compositionality\",\n \"year\": \"2013-12-05\"\n },\n {\n \"authors\": \"Fatemehsadat Mireshghallah; Kartik Goyal; Taylor Berg-Kirkpatrick\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b40\",\n \"title\": \"Mix and match: Learningfree controllable text generationusing energy language models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Shashi Narayan; Shay B Cohen; Mirella Lapata\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b41\",\n \"title\": \"Don't give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Pinelopi Papalampidi; Kris Cao; Tomas Kocisky\",\n \"journal\": \"\",\n \"ref_id\": \"b42\",\n \"title\": \"Towards coherent and consistent use of entities in narrative generation\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Krishna Pillutla; Swabha Swayamdipta; Rowan Zellers; John Thickstun; Sean Welleck; Yejin Choi; Zaid Harchaoui\",\n \"journal\": \"Advances in Neural Information Processing Systems\",\n \"ref_id\": \"b43\",\n \"title\": \"MAUVE: Measuring the gap between neural text and human text using divergence frontiers\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Alec Radford; Jeffrey Wu; Rewon Child; David Luan; Dario Amodei; Ilya Sutskever\",\n \"journal\": \"\",\n \"ref_id\": \"b44\",\n \"title\": \"Language models are unsupervised multitask learners\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu\",\n \"journal\": \"Journal of Machine Learning Research\",\n \"ref_id\": \"b45\",\n \"title\": \"Exploring the limits of transfer learning with a unified text-totext transformer\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Shafiq Mathieu Ravaut; Nancy Joty; Chen\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b46\",\n \"title\": \"SummaReranker: A multi-task mixture-of-experts re-ranking framework for abstractive summarization\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Abigail See; Peter J Liu; Christopher D Manning\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b47\",\n \"title\": \"Get to the point: Summarization with pointergenerator networks\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Abigail See; Aneesh Pappu; Rohun Saxena; Akhila Yerukola; Christopher D Manning\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b48\",\n \"title\": \"Do massively pretrained language models make better storytellers?\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Kurt Shuster; Jack Urbanek; Arthur Szlam; Jason Weston\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b49\",\n \"title\": \"Am I me or you? state-of-the-art dialogue models cannot maintain an identity\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Vered Shwartz; Rachel Rudinger; Oyvind Tafjord\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b50\",\n \"title\": \"you are grounded!\\\": Latent name artifacts in pre-trained language models\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Yixuan Su; Tian Lan; Yan Wang; Dani Yogatama; Lingpeng Kong; Nigel Collier\",\n \"journal\": \"\",\n \"ref_id\": \"b51\",\n \"title\": \"A contrastive framework for neural text generation\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Lukasz Kaiser; Illia Polosukhin\",\n \"journal\": \"\",\n \"ref_id\": \"b52\",\n \"title\": \"Attention is all you need\",\n \"year\": \"2017-09\"\n },\n {\n \"authors\": \"C Lawrence Ramakrishna Vedantam; Devi Zitnick; Parikh\",\n \"journal\": \"IEEE Computer Society\",\n \"ref_id\": \"b53\",\n \"title\": \"Cider: Consensus-based image description evaluation\",\n \"year\": \"2015-06-07\"\n },\n {\n \"authors\": \"David Wan; Mohit Bansal\",\n \"journal\": \"\",\n \"ref_id\": \"b54\",\n \"title\": \"Factpegasus: Factuality-aware pre-training and fine-tuning for abstractive summarization\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"David Wan; Shiyue Zhang; Mohit Bansal\",\n \"journal\": \"\",\n \"ref_id\": \"b55\",\n \"title\": \"Histalign: Improving context dependency in language generation by aligning with history\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Lingxiao Wang; Jing Huang; Kevin Huang; Ziniu Hu; Guangtao Wang; Quanquan Gu\",\n \"journal\": \"\",\n \"ref_id\": \"b56\",\n \"title\": \"Improving neural language generation with spectrum control\",\n \"year\": \"2020-04-26\"\n },\n {\n \"authors\": \"Sean Welleck; Ilia Kulikov; Stephen Roller; Emily Dinan; Kyunghyun Cho; Jason Weston\",\n \"journal\": \"\",\n \"ref_id\": \"b57\",\n \"title\": \"Neural text generation with unlikelihood training\",\n \"year\": \"2020-04-26\"\n },\n {\n \"authors\": \"Sean Welleck; Jiacheng Liu; Ximing Lu; Hannaneh Hajishirzi; Yejin Choi\",\n \"journal\": \"\",\n \"ref_id\": \"b58\",\n \"title\": \"Naturalprover: Grounded mathematical proof generation with language models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Thomas Wolf; Lysandre Debut; Victor Sanh; Julien Chaumond; Clement Delangue; Anthony Moi; Pierric Cistac; Tim Rault; Remi Louf; Morgan Funtowicz; Joe Davison; Sam Shleifer; Clara Patrick Von Platen; Yacine Ma; Julien Jernite; Canwen Plu; Teven Xu; Sylvain Le Scao; Mariama Gugger; Quentin Drame; Alexander Lhoest; Rush\",\n \"journal\": \"\",\n \"ref_id\": \"b59\",\n \"title\": \"Transformers: State-of-the-art natural language processing\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Zhilin Yang; Zihang Dai; Ruslan Salakhutdinov; William W Cohen\",\n \"journal\": \"\",\n \"ref_id\": \"b60\",\n \"title\": \"Breaking the softmax bottleneck: A high-rank RNN language model\",\n \"year\": \"2018-04-30\"\n },\n {\n \"authors\": \"Haopeng Zhang; Semih Yavuz; Wojciech Kryscinski; Kazuma Hashimoto; Yingbo Zhou\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b61\",\n \"title\": \"Improving the faithfulness of abstractive summarization via entity coverage control\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Zexuan Zhong; Tao Lei; Danqi Chen\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b62\",\n \"title\": \"Training language models with memory augmentation\",\n \"year\": \"2022-12-07\"\n }\n]"},"formulas":{"kind":"list like","value":[{"formula_coordinates":[3,98.94,305.41,145.85,18.06],"formula_id":"formula_0","formula_text":"P M (x|c t ) = exp(Logit(x, c t ))"},{"formula_coordinates":[3,316.45,361.48,207.96,15.02],"formula_id":"formula_2","formula_text":"q ct = h M ct ⊕ GELU L h (⊕ i,m h M -m c t-i ) ,(2)"},{"formula_coordinates":[3,471.86,426.15,52.05,14.47],"formula_id":"formula_3","formula_text":"⊕ i,m h M -m c t-i"},{"formula_coordinates":[3,335.83,732.87,188.58,43.63],"formula_id":"formula_4","formula_text":"t Logit(x, c t ) = f T ct,S e x if x ∈ S f T ct,V w x O/W ,(3)"},{"formula_coordinates":[4,252.34,194.84,41.13,25.83],"formula_id":"formula_5","formula_text":"L f LD L f C"},{"formula_coordinates":[4,84.79,672,204.34,35.52],"formula_id":"formula_6","formula_text":"e x = f x,ct,LD = t i=1 1 c i t =x L f LD (q c i t ) t i=1 1 c i t =x ,(4)"},{"formula_coordinates":[4,184.9,733.35,92.98,16.34],"formula_id":"formula_7","formula_text":"1 c i t =x = 1 if c i t = x."},{"formula_coordinates":[5,79.39,568.46,209.74,56.98],"formula_id":"formula_8","formula_text":" f T ct,C w x + f T ct,P D f x,ct,LD if x ∈ c t f T ct,R1 w x if x ∈ W (k 1 ) -c t f T ct,R2 w x if x ∈ W (k 2 ) -W (k 1 ) -c t f T ct,V w x O/W ,(5)"},{"formula_coordinates":[5,311.98,327.04,212.43,56.98],"formula_id":"formula_9","formula_text":" f T ct,C w x + f T ct,P D f x,ct,LD if x ∈ c t f T ct,E w x + f T ct,P E f x,I,LE if x ∈ I -c t f T ct,R1 w x if x ∈ W (k 1 ) -c t -I f T ct,V w x O/W ,(6)"},{"formula_coordinates":[18,94.68,713.67,194.46,28.45],"formula_id":"formula_10","formula_text":"f T ct,V w x + f T ct,P D f x,ct,LD if x ∈ c t f T ct,V w x O/W .(7)"},{"formula_coordinates":[19,97.9,326.43,191.23,70.27],"formula_id":"formula_11","formula_text":"P rob(x|I, c t ) = p gen exp f T ct,V w x Z V +(1 -p gen ) |I| j=1 1 I j =x P E (j|I, c t ),(9)"},{"formula_coordinates":[19,82.6,624.52,206.53,56.91],"formula_id":"formula_12","formula_text":"P rob(x|I, ct) = g exp f T c t ,V wx ZV + |I| j=1 1 I j =x exp f T c t ,P E tanh(L f LE (h M I j )) + b Zp ,(10)"},{"formula_coordinates":[19,321.85,450.99,202.56,28.45],"formula_id":"formula_13","formula_text":"f T ct,V w x + f T ct,R f x,ct,LD if x ∈ W (k) f T ct,V w x O/W .(11)"}],"string":"[\n {\n \"formula_coordinates\": [\n 3,\n 98.94,\n 305.41,\n 145.85,\n 18.06\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"P M (x|c t ) = exp(Logit(x, c t ))\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 316.45,\n 361.48,\n 207.96,\n 15.02\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"q ct = h M ct ⊕ GELU L h (⊕ i,m h M -m c t-i ) ,(2)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 471.86,\n 426.15,\n 52.05,\n 14.47\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"⊕ i,m h M -m c t-i\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 335.83,\n 732.87,\n 188.58,\n 43.63\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"t Logit(x, c t ) = f T ct,S e x if x ∈ S f T ct,V w x O/W ,(3)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 252.34,\n 194.84,\n 41.13,\n 25.83\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"L f LD L f C\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 84.79,\n 672,\n 204.34,\n 35.52\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"e x = f x,ct,LD = t i=1 1 c i t =x L f LD (q c i t ) t i=1 1 c i t =x ,(4)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 184.9,\n 733.35,\n 92.98,\n 16.34\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"1 c i t =x = 1 if c i t = x.\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 79.39,\n 568.46,\n 209.74,\n 56.98\n ],\n \"formula_id\": \"formula_8\",\n \"formula_text\": \" f T ct,C w x + f T ct,P D f x,ct,LD if x ∈ c t f T ct,R1 w x if x ∈ W (k 1 ) -c t f T ct,R2 w x if x ∈ W (k 2 ) -W (k 1 ) -c t f T ct,V w x O/W ,(5)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 311.98,\n 327.04,\n 212.43,\n 56.98\n ],\n \"formula_id\": \"formula_9\",\n \"formula_text\": \" f T ct,C w x + f T ct,P D f x,ct,LD if x ∈ c t f T ct,E w x + f T ct,P E f x,I,LE if x ∈ I -c t f T ct,R1 w x if x ∈ W (k 1 ) -c t -I f T ct,V w x O/W ,(6)\"\n },\n {\n \"formula_coordinates\": [\n 18,\n 94.68,\n 713.67,\n 194.46,\n 28.45\n ],\n \"formula_id\": \"formula_10\",\n \"formula_text\": \"f T ct,V w x + f T ct,P D f x,ct,LD if x ∈ c t f T ct,V w x O/W .(7)\"\n },\n {\n \"formula_coordinates\": [\n 19,\n 97.9,\n 326.43,\n 191.23,\n 70.27\n ],\n \"formula_id\": \"formula_11\",\n \"formula_text\": \"P rob(x|I, c t ) = p gen exp f T ct,V w x Z V +(1 -p gen ) |I| j=1 1 I j =x P E (j|I, c t ),(9)\"\n },\n {\n \"formula_coordinates\": [\n 19,\n 82.6,\n 624.52,\n 206.53,\n 56.91\n ],\n \"formula_id\": \"formula_12\",\n \"formula_text\": \"P rob(x|I, ct) = g exp f T c t ,V wx ZV + |I| j=1 1 I j =x exp f T c t ,P E tanh(L f LE (h M I j )) + b Zp ,(10)\"\n },\n {\n \"formula_coordinates\": [\n 19,\n 321.85,\n 450.99,\n 202.56,\n 28.45\n ],\n \"formula_id\": \"formula_13\",\n \"formula_text\": \"f T ct,V w x + f T ct,R f x,ct,LD if x ∈ W (k) f T ct,V w x O/W .(11)\"\n }\n]"}}},{"rowIdx":392,"cells":{"doi":{"kind":"string","value":"10.3115/v1/D14-1082"},"pub_date":{"kind":"string","value":"2023-05-21"},"sections":{"kind":"list like","value":[{"figure_ref":["fig_1"],"heading":"Introduction","publication_ref":["b10","b17","b8","b17","b10","b2","b3","b9","b5"],"table_ref":[],"text":"Fine-grained entity typing (FET) is an important task in text analysis. Assigning fine-grained semantic types to parsed entity mention spans based on the local context enables effective and structured analysis of unstructured text data, such as entity linking (Ling and Weld, 2012;Onoe and Durrett, 2020), relation extraction (Koch et al., 2014), and coreference resolution (Onoe and Durrett, 2020). Example 1. Given a sentence S 1 : \"Sammy Sosa got a standing ovation at Wrigley Field.\" and a parsed entity mention span \"Sammy Sosa\" in the sentence, an FET method aims to assign it not only the coarse-grained type \"Person\" but also the finegrained types \"Athlete\" or \"Player\".\nFET on large text corpora is challenging due to (1) the high cost to obtain a large amount of humanannotated training data, especially in dynamic and emerging domains, and (2) inaccurate annotations due to (i) different annotators marking concepts at different granularity (e.g., person vs. politician vs. president), and (ii) contextual subtlety on finegrained types (e.g., Boston vs. Detroit could be two sport teams, instead of two cities). Most existing methods utilize weak or distant supervision to automatically generate training data for the FET tasks. There are three major approaches to obtaining weakly-labeled training data to tackle these challenges. The first is to automatically match the mentions in text with the concepts in some existing knowledge bases (e.g., Wikipedia) (Ling and Weld, 2012). The typical workflow is to first detect entity mentions from a corpus, map these mentions to knowledge base (KB) entities of target types, and then leverage the confidently mapped types as labeled data to infer the final type. The second is to directly obtain the head words of nominal mentions as its fine-grained type (Choi et al., 2018). It leverages the head words of the entity mention to consolidate context-aware types derived from the KB matching. However, both approaches suffer from the label sparsity and context-agnostic problems, resulting in the inability to generate high-quality training data for FET.\nThe third approach is to probe the pre-trained language models through the use of masked patterns and entailment templates. Leveraging masked language model (MLM) prompting to generate rich and context-aware weak supervisions for finegrained entity typing is a recent trend, aiming to reduce expensive human labor (Dai et al., 2021;Li et al., 2022).For example, given a sentence that con-tains a mention, a short piece of text that contains a \"[MASK]\" token is added to generate candidate entity types. This method conducts labeling in a context-aware manner and greatly enriches the fine-grained types labeled for each mention. However, such a process can still generate tokens unsuitable for typing (e.g., {Teams, Thing} for \"Wrigley Field\" in S 1 ) or a mixture of rough and fine-grained types (e.g., {Location, Building, Stadium}). Due to the lack of further hierarchical knowledge of the generated tokens/types, such problems cannot be resolved automatically.\nIn this study, we vision that an ontology structure, which provides a semantics-rich, hierarchical structure, may help select the best results generated by multiple PLM models. We propose an zero-shot, ontology-guided, fine-grained entity typing (FET) method, ONTOTYPE, that leverages an input ontology and the power of MLM prompting and Natural Language Inference (NLI). We first ensemble multiple Hearst patterns to perform MLM prompting, reducing the noise in the initial candidate type generation. Since the generated candidate labels for a given mention are likely a mixture of fine and coarse-grained labels, or tokens unsuitable for typing, we propose to automatically match the generated candidate labels to a coarse-grained type in our type ontology and then rank and select a coarsegrained type with a pre-trained entailment model under the local context. Such a type resolution process will progress deeper to finer levels, based on the same principle of entailment score-based type selection, following the type ontology, until the finest possible label can be consolidated.\nExample 2. For sentence S 1 in Ex. 1, candidate type generation (Step 1) ensembles prompting results of multiple Hearst patterns and generates a set of candidate labels {\"Stadium\", \"Venues\", \"Locations\", \"Games\"} for \"Wrigley Field\" (Fig. 2). Using a given ontology structure (Fig. 1), this set of types is resolved to the course-grained type \"Location\" by leveraging a pre-trained entailment model and the local context (Fig. 3). Note that without the ontology type structure, \"Stadium\", \"Venues\", and \"Locations\" are rivals; but with the structure, \"Stadium\" and \"Venues\" are fine-grained \"Locations\" and they support each other at multilevel resolution. By the same principle, the type resolution proceeds deeper to finer-grained levels, along the type ontology, from \"Location\" to \"Building\" and further down to \"Stadium\" for \"Wrigley Field\", leading to the most accurate fine-grained type (Fig. 4). Our contributions are summarized as follows.\n1. A fully zero-shot, ontology-guided, fine-grained typing method, ONTOTYPE, is proposed 2. ONTOTYPE improves fine-grained entity typing (FET) by leveraging candidate labels generated and refined with three information sources: (i) pre-trained language models, (ii) a fine-grained type ontology, and (iii) head words 3. Experiments on the Ontonotes, FIGER, and NYT datasets (Gillick et al., 2014) using their associated ontological structures show that ON-TOTYPE clearly outperforms existing zero-shot named entity typing methods and even rivals supervised methods. Our error analysis shows that refinement of ontology structures will further improve fine-grained entity typing."},{"figure_ref":[],"heading":"Related Work","publication_ref":["b23","b10","b6","b2","b10","b24","b2","b3","b22","b25","b28","b19","b13","b4","b3","b11","b9","b3","b9","b21"],"table_ref":[],"text":"Supervised FET. Fine-grained entity typing benefits various downstream tasks and has received extensive attention in natural language research. Recent studies focus on different contexts from the phrase level (Yao et al., 2013) to considering specific entity mentions in the sentence or document level (Ling and Weld, 2012;Gupta et al., 2017;Choi et al., 2018). Entity typing has been generally studied under supervised learning settings (Ling and Weld, 2012;Yosef et al., 2012;Choi et al., 2018;Dai et al., 2021). These studies require annotated data to achieve high performance and lack the flexibility to identify newly defined entities.\nWeakly supervised or zero-shot FET. To handle such difficulties, the zero-shot learning setting has been introduced for named entity typing (Xia et al., 2018). Several studies (Yuan and Downey, 2018;Zhou et al., 2018) address the problems by grounding the mentions with Wikipedia entries from the assembled related pages. These methods achieve good performance but also require a lot of resources. Other studies explore the pre-trained semantic word-level embeddings (Ren et al., 2016;Ma et al., 2016) or extract raw embeddings without auxiliary information and utilize end-to-end neural networks (Zhang et al., 2020b). However, they still suffer from low accuracy and inefficiency in zeroshot settings. As a result, ONTOTYPE turns to the weak supervision from pre-trained language models, such as BERT (Devlin et al., 2019), that have substantial knowledge in language understanding.\nLeveraging pre-trained language models. Some recent studies (e.g., (Zhang et al., 2020a;Dai et al., 2021;Liu et al., 2021;Li et al., 2022)) leverage pre-trained language models and prompting templates to obtain knowledge for entity mentions in given sentences. (Dai et al., 2021) improves ultrafine entity typing with BERT Masked Language Model (MLM). However, they use a single prompt for each mention to generate labels for weak supervision, which may generate erroneous types. (Li et al., 2022) improves ultra-fine entity typing by treating the task of predicting an entity type as an NLI task. However, they ignore rich semantic information contained in a type ontology during the typing process. Our method ensembles multiple MLM prompting and NLI results to reduce noises in candidate type generation. Further, we utilize the fine-grained type ontology structure as guidance to progressively resolve candidate labels from coarse to fine under the local context. ChemNER (Wang et al., 2021) leverages a type ontology structure to guide fine-grained Chemistry entity recognition. It relies on an existing knowledge base whereas this study leverages the pre-trained language models as rich and context-aware weak supervision."},{"figure_ref":["fig_1"],"heading":"Methodology","publication_ref":["b12"],"table_ref":[],"text":"We propose ONTOTYPE, a zero-shot, ontologyguided, fine-grained entity typing method using pre-trained language models and a fine-grained ontology structure. Given an input sentence and a set of pre-identified mentions in the sentence, ONTO-TYPE consists of the following steps: (1) generating a set of candidate labels for each input mention with both head word parsing and an ensemble of MLM prompting (Fig. 2); (2) resolving the coarsegrained types by matching and ranking the generated labels to the type ontology using an entailment model (Fig. 3); and (3) progressively refining the fine-grained types along the type ontology following the principle of entailment score-based type selection (Fig. 4). We utilize the inherent structure of the fine-grained type ontology and a pre-trained entailment model (RoBERTa model pre-trained on the MNLI dataset (Liu et al., 2019)) to guide our fine-grained entity typing."},{"figure_ref":[],"heading":"Problem Definition","publication_ref":[],"table_ref":[],"text":"The input to our proposed ONTOTYPE framework is a text corpus D and input fine-grained type ontology O. In this study, we assume our input text corpus D includes a set of pre-identified entity mentions. An entity mention, e, is a token span in the text document that refers to a real-world entity. Given a sentence S and a parsed entity mention e ∈ S, the fine-grained entity typing (FET) task is to identify one or more types t from the label space T (provided in a structured ontology G) for the entity mention e.\nAs an example of our FET task, in the sentence S 1 of Example 1, the entity mention to be typed is e 1 : \"Wrigley Field\". It should be labeled progressively deeper as \"Location→Building→Stadium\" as opposed to other labels like \"Organization\", \"Person\", or \"School\".\nFine-Grained Type Ontology Structure. The structure of the type ontology is fundamental to the ONTOTYPE algorithm. In this work, we leverage the fine-grained type ontologies provided in the OntoNotes and FIGER datasets. These ontologies are structured as forests or a disjoint union of trees. Each tree consists of a type hierarchy stemming from a handful of \"root\" coarse-grained types that include but are not limited to \"Organization\", \"Person\" and \"Location\". Each tree imposes the following structural conditions: (1) Each type has a singular parent type; (2) each type can have an infinite (potentially) number of children; and (3) each type is connected by a directional edge indicating a hypernym-hyponym type relationship between the parent and child nodes. In our experiments, we find that general types such as \"things\" lead to inaccurate labels as pre-trained entailment models often rank vague types higher than specific types. Note that in the given OntoNotes type ontology (Fig. 1), Building and Country are sibling types since they share the same parent type Location. Such a hypernym-hyponym or \"is-a\" relationship is critical for guiding fine-grained type selection."},{"figure_ref":["fig_1","fig_1"],"heading":"Candidate Type Generation","publication_ref":["b2","b3","b2","b1","b4","b7","b20"],"table_ref":["tab_0"],"text":"To generate a set of candidate labels for each input mention, we leverage two main information sources: (1) head word parsing, and (2) an ensemble of MLM prompting. Head words and hypernyms can serve as powerful context-aware type indicators that can be leveraged as weak supervision sources for entity typing (Choi et al., 2018;Dai et al., 2021). Our method adopts both information sources to first select a candidate label to guide our fine-grained type refinement. First, we generate and parse the head words for the given entity mention. Second, we generate candidate entity types with the use of ensembled MLM prompting1 . The generated candidate labels are used as input to the following steps of ontology-guided type resolution.\nHead Word Parsing. As discussed by (Choi et al., 2018), the input text often contains cues that explicitly match a mention to its type, in the form of the mention's head word. Thus, given the pre-identified entity mentions in the input sentence, we first identify the head word of the input mention. We utilize the Stanford Dependency Parser (Chen and Manning, 2014) to extract the head word of the entity that we are interested in typing. Consider the sentence \"Governor Arnold Schwarzenegger gives a speech at Mission Serve's service project on Veterans Day 2010\" with \"Governor Arnold Schwarzenegger\" as an entity mention. We can describe \"Arnold Schwarzenegger\" with numerous types including actor, father, and governor. However with the use of the head word \"Governor\", we consolidate the fine-grained type of governor given this input sentence. Thus, ONTOTYPE leverages the head words of the input entity, if any, to select an initial context-sensitive coarse-grained type and guide the selection of the final fine-grained type.\nEnsembled MLM Prompting. While head words can provide strong type indicators, they do not always provide sufficient information to consolidate a high-level type. In some cases, head words cannot directly provide accurate type information as they are not directly present in the input type ontology. Furthermore, head words are not always available in the input sentences. Thus, with the parsed entity mention span as input, we propose to leverage context-aware hypernyms as initial type labels for the target mention. With an ensembled cloze prompting method, ONTOTYPE generates candidate labels of the mention and performs an initial high-level typing on the input mention. Specifically, we leverage a BERT masked language model (Devlin et al., 2019) and artificial Hearst patterns (Hearst, 1992) to generate context-aware hypernyms that serve as labels for the mentions. We first modify the input sentence by inserting a Hearst pattern and [MASK] token into the sentence. Then we use the BERT MLM to generate candidate labels for the target mention under the local context. For example, in Fig. 2, we first insert Hearst patterns such as \"and the other [MASK]\" in the input sentence and then use the BERT MLM model to generate candidate labels such as \"Venues\", \"Teams\", and \"Stadiums\".\nWe evaluated the quality of hypernyms generated with the 44 patterns proposed in (Seitner et al., 2016) on the Ontonotes Development Dataset. When generating hypernyms, we expect high-quality candidates to be semantically equivalent to concepts contained in our input type ontology. Overall, we found that the four patterns in Table 1 provided the highest quality hypernyms under a simple direct matching to types in the OntoNotes ontology. However, based on the syntax and grammar of the sentence, hearst patterns can generate tokens that are unsuitable to serve as fine-grained entity types. For example, with a single prompt, the MLM can generate \"Famous\", \"Actors\", \"Celebrities\" and \"People\" as the most probable words. Note that \"Famous\" is unsuitable to serve as a fine-grained entity type. To reduce the noises caused by the use of a single Hearst pattern, we ensemble n Hearst Patterns to consolidate the most commonly generated candidate labels. For each pattern in the pattern list P , we collect the top k most probable tokens from the probability distribution predicted by the BERT MLM. Then, we aggregate the tokens and identify the set of labels that have the largest overlap. We perform the voting ensemble as follows:\ny = count (m) {H 1 (x), H 2 (x), ..., H n (x)} (1)\nwhere H n (x) is the candidate type generated by the n th hearst pattern and count (m) is the function that selects all candidate labels generated at least m times. In our experiments, we use m = n 2 + 1. In addition, we find that the number of ensembled patterns is not sensitive as ensembling ensures we retain the most confident candidate labels. Example 3: As shown in Fig. 2, by ensembling the results from prompting with several Hearst patterns, the quality types for e 1 \"Stadiums, Venues, Locations, Games\" retain but the noisy types \"Things\" and \"Teams\" are removed."},{"figure_ref":[],"heading":"High-Level Type Resolution","publication_ref":["b14"],"table_ref":[],"text":"Given the generated candidate labels and head words for each mention in the sentence, we seek to resolve the concrete type for this candidate type set at the high levels of the type ontology. Specifically, we first align the generated candidate labels to several high-level types in the type ontology, and then select the most accurate high-level types with a pre-trained entailment language model.\nCandidate Type Alignment. Following the previous step of candidate type generation, we combine the generated candidate labels from both the parsed head words and the ensembled cloze prompting to form a candidate type set. These labels are generally noisy and may not exist directly in our type ontology. However, we observe that the generated candidates will usually cluster closely around a high-level concept in the ontology. Thus, we perform our high-level type alignment with a cosinesimilarity-based word matching. We use Word2Vec 2 (Mikolov et al., 2013) embeddings to construct our type embeddings for the cosine-similarity-based type alignment. We construct a verbalizer by selecting at least l seman-2 https://code.google.com/archive/p/word2vec/ tically related words for each coarse type. For a high-level type of \"Organization\", we might include seed types such as \"Corporation\", \"University\", \"Firm\", \"Business\", and \"Government\" in its verbalizer. This verbalizer is then utilized to systematically map the MLM hypernym prediction to the most relevant type. In our experiments, we provide at least five seed types S for each type node c in the first level of the type ontology. Increasing the number of seed types increases the coverage and confidence of the verbalizer. Since the most commonly generated hypernyms for each concept exists in the input ontologies, ONTOTYPE is not so sensitive to the number of seed types collected. With the seed types, we construct a node embedding N by taking the average of word embeddings from both the first-level type and its corresponding seed types provided by human,\nN = s i ∈S emb(s i ) + emb(c) |S| + 1 . (2\n)\nwhere emb(•) indicates the Word2Vec embeddings. Then we rank each generated candidate type to a first-level type on our type ontology by calculating the cosine similarity between the embeddings of the generated candidate labels b and that of the first-level types T , score(b, T ) rank = cosine(emb(b), emb(T )).\n(3)\nFinally, the first-level type that has the highest similarity is selected as the aligned high-level type."},{"figure_ref":["fig_1"],"heading":"⋮ Figure 3: Candidate Type Generation","publication_ref":["b12"],"table_ref":[],"text":"Example 4: In Figures 2 and3, the consolidated candidate labels \"Locations\", \"Stadiums\", \"Venues\", and \"Games\" are closely related to the high-level type \"location\" in our ontology. By performing the cosine-similarity-based matching, \"Location\" is identified as e 1 's high-level type over \"Organization\" or \"Person\". Thus, we finally select the most similar high-level type given by the head word, generated candidate labels, and the entailment model.\nHigh-Level Type Selection After the previous step of candidate type alignment, we obtain several high-level types for each entity mention in the sentence. Given these types, we seek to select the most accurate high-level type for each entity mention under the local context. We observe that the task of selecting the most appropriate entity type can be viewed as a Natural Language Inference (NLI) task. Thus, we treat the input sentence as the premise in NLI and generate the hypothesis using a pre-defined declarative template. To resolve the type of the input mention we utilize the following template: \"In this sentence, [MENTION] is a [TYPE]\". ONTOTYPE then ranks each type in the first level of the ontology with the entailment score from a RoBERTa NLI model (Liu et al., 2019).\nExample 5: In Fig. 3, we have aligned a majority of the generated candidate type set to the Location seed types. The NLI model further ranks Location over Organization or Person. By utilizing the information in conjunction, e 1 is solidly aligned to the high-level type \"Location\"."},{"figure_ref":[],"heading":"Fine-Grained Type Resolution","publication_ref":[],"table_ref":[],"text":"Given the high-level types of the entity mentions, ONTOTYPE further leverages the ontology structure to progressively resolve the fine-grained label.\nFollowing the same principle of entailment-based type selection for the high-level types, we utilize the entailment model to automatically select the most accurate fine-grained entity types along our type ontology. Specifically, we first examine the child types of the previously determined higherlevel types and then select the child type with the highest ranked score as the fine-grained type.\nIn addition to the entailment model, we also utilize the candidate type set to refine our fine-grained type selection. If the parsed head word is present in our type ontology, the entailment scores of that parent and its child types are weighted higher. Similarly, if the generated candidate labels are in our type ontology, the entailment scores of those parents and their child types are also weighted higher. Finally, we select the fine-grained type for each mention with the highest weighted entailment score.\nWe calculate the ranking score for the entities at the current level of the ontology as follows:\nrank(type) = σ entail + σ cand(4)\nWe first leverage our NLI pre-trained model to find the entailment score σ entail for each entity type. Then if a type in the candidate type set is a descendent to the entity type, we add a weight σ cand . We repeat this entailment-based selection process recursively along the type ontology to select the best fine-grained type.\nDefinition (Entity Type Granularity Parameter θ):\nWe assume there is a scalar of θ indicating how granular or specific the final selected entity type label should be. The smaller θ is, the more granular entity type labels are consolidated as the final label. Thus, if the child types do not have a sufficient gain of at least θ in ranking score over the parent type at a certain level, we stop the recursion and select the parent type as the final fine-grained type. We conduct a parameter study to explore the sensitivity of ONTOTYPE to the parameter θ in Appendix A.3\n⋮ Figure 4: Fine-Grained Type Refinement\nExample 6: In Fig. 4, we consider all descendent types of \"location\" as potential fine-grained entity types for e 1 . To begin, ONTOTYPE generates the hypotheses and ranks all child types of \"location\". With the resultant entailment score rankings of these sibling types, we consolidate and select the \"building\" type as the highest-ranked label. Then, a similar process is done at a deeper level of the ontology to select the final type of \"stadium\"."},{"figure_ref":[],"heading":"Experiments","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Datasets","publication_ref":["b5","b10"],"table_ref":[],"text":"We compare the performance of ONTOTYPE and several baseline methods on three benchmark FET datasets: NYT, Ontonotes (Gillick et al., 2014) and FIGER (Ling and Weld, 2012). The basic statistics of the datasets are shown in Table 3: Results (%) on Three Test Sets (Some slots in supervised methods marked \"-\" due to no fully annotated training data).\nOntoNotes and FIGER datasets, we leverage the included type ontologies while the NYT dataset leverages the input FIGER ontology. All NER test sets are annotated using the ontologies as a set of type labels. Thus, each entity mention is labeled with a fine-grained label represented as a path within the ontology. The ontologies have a maximum depth of three and contain four to six high-level types (e.g., Location, Person, and Organization)."},{"figure_ref":[],"heading":"Baselines","publication_ref":[],"table_ref":[],"text":"ONTOTYPE is a zero-shot FET method that does not require human annotation as supervision. We compare ONTOTYPE with two groups of baseline methods for FET: supervised methods and zeroshot methods. Details of the parameter settings for each model is listed in Appendix A.2. We also conduct parameter and ablation studies listed in Appendices A.3 and A.4 respectively."},{"figure_ref":[],"heading":"Supervised FET Methods","publication_ref":["b19"],"table_ref":[],"text":"We compare the performance of ONTOTYPE with three supervised FET methods as discussed below.\nAFET (Ren et al., 2016) MZET (Zhang et al., 2020b): This zero-shot baseline leverages the semantic meaning and the hierarchical structure into the type representation. The method leverages the knowledge from seen types to label the zero-shot types through the use of a memory component."},{"figure_ref":[],"heading":"Main Results","publication_ref":["b28"],"table_ref":[],"text":"Table 3 strates the benefit of leveraging the knowledge embedded in pre-trained language models as a form of minimal supervision to identify entity mention types. ZOE leverages a type ontology and maps a given mention to type-compatible Wikipedia Entries. As a result, ZOE relies on surface-level information from the mention string. OntoType ensembles contextual information from various PLMs to consolidate the final entity type. Given a sentence: \"The biggest cause for concern for McGuff is the bruised hamstring Regina Rogers suffered against (Utah) last Saturday\", ZOE incorrectly utilizes the surface string to label \"Utah\" as a location. With PLMs, OntoType recognizes teams or opponents as candidates and finally consolidates \"sports_team\". On the FIGER dataset, ONTOTYPE achieves 0.3 absolute Macro-F1 improvement over state-of-the-art zero-shot fine-grained entity typing method ZOE (Zhou et al., 2018). However, our method trails ZOE in strict accuracy and Micro-F1. In the FIGER dataset, predictions are made based on both the surface-level information and the contextual information in the sentence. The major advantage of ONTOTYPE is to accurately capture the contextual information to provide a fine-grained entity type. However, ONTOTYPE does not involve a mechanism to capture the surface-level informa-tion of an entity mention. We discuss this issue further in our error analysis (Appendix A.6)."},{"figure_ref":[],"heading":"Comparative Case Study","publication_ref":[],"table_ref":["tab_6"],"text":"Table 4 presents a sentence from a recent news article with tagged mentions and predicted entity types. We find that methods like MZET sometimes predict incompatible types due to incorrect or misleading surface information. For example, when typing \"The White House\", MZET and ZOE leverage the surface-level information from large KBs to identify the mention as a location. However, given the local context, the White House clearly refers to the U.S. government. When considering the mention \"US President Joe Biden\", our method utilizes the type information from the candidate type set (Officials, Leaders, Politicians, Individuals) to explore the \"Person\" branch within the ontology. We search the \"Politician\" branch for the best fine-grained context-aware type to label our given entity mention. Thus, with the assistance of the pre-trained language models, we can incorporate contextual information to derive more context-aware types."},{"figure_ref":[],"heading":"Error Analysis","publication_ref":[],"table_ref":[],"text":"The structure of the fine-grained type ontology is important to the performance of ONTOTYPE. Input ontologies should utilize hypernym-type relationships where each parent type is a generalization of the child type. The provided FIGER ontology contains logical inconsistencies in how various types of relations are organized. Specifically, the FIGER ontology considers the parent of the \"Road\" type to be \"Transportation\" rather than \"Location\". This logical inconsistency leads to erroneous typing from our method. We further examine our method's errors and evaluate ONTOTYPE using a modified type ontology in Appendix A.6."},{"figure_ref":[],"heading":"Large Language Models","publication_ref":["b18"],"table_ref":[],"text":"Traditional language models (LM) aim at assigning a Probability P (w) to an input sequence of n words s = [w 1 , w 2 , ..., w n ]. Recent works include large language models (LLMs) such as ChatGPT and GPT-4 (OpenAI, 2023), which are pre-trained on with a large text corpora enables them to generate contextualized word representation to achieve state of the art performance on various NLP tasks including language translation, text summarization, and question-answering. These recent works provide a richer knowledge base for ONTOTYPE's candidate type generation. As a result, we can similarly leverage models like T-5 (Raffel et al., 2020) or GPT-3 (Brown et al., 2020) to obtain high-quality candidate types through hypernym prompting methods.\nExample 6: Given a sentence S 1 : \"Sammy Sosa got a standing ovation at Wrigley Field.\" and a parsed entity mention span \"Sammy Sosa\", Chat-GPT generates \"Athletes\", \"Players\", \"Teammates\", \"Stars\" and \"Figures\" while T-5 generates \"Athletes\", \"Players\", \"Teammates\" and \"Individuals\". Overall, we find that even LLMs can leverage the surrounding context and learned \"common-sense\" knowledge to generate high quality candidate types. ONTOTYPE provides a structured approach to provide guidance for LLMs like ChatGPT and GPT-4 which are extremely susceptible to hallucinations. By grounding our types to mention hypernyms and a fine-grained type ontology, we can confidently extract the most fine-grained type for the given entity mention span."},{"figure_ref":[],"heading":"Conclusions","publication_ref":[],"table_ref":[],"text":"We propose ONTOTYPE, which leverages weak supervision setting of pre-trained language model prompting. We use head words and MLM cloze prompting for fine-grained candidate label generation. Then we automatically match the generated fine-grained types to our type ontology with an inference method to select the most appropriate finegrained types under the local context. Extensive experiments on real-world datasets show that ON-TOTYPE is highly effective and substantially outperforms the state-of-the-art zero-shot FET methods. In the future, we plan to further refine and enrich a type ontology which will enable us to incorporate more type information for even better performance."},{"figure_ref":[],"heading":"Limitations","publication_ref":[],"table_ref":[],"text":"ONTOTYPE utilizes pre-trained language models to perform context-aware fine-grained entity typing leveraging the rich structural information from type ontologies, head words and cloze prompting. However, this method still has a few limitations. Since the pre-trained language models utilized are not fine-tuned, some candidate labels and fine-grained types generated from the MLM and NLI models are noisy and unsuitable to serve as fine-grained types. One future improvement of ONTOTYPE is to finetune the MLM and NLI models based on the local corpus and type ontology. Furthermore, ONTO-TYPE ignores the surface-level information of the entity mentions. Prior distantly-supervised methods have been able to leverage the surface-level information through entity linking to the knowledge bases. Another future improvement of ONTOTYPE is to incorporate these rich information sources to perform typing on the entity mentions with the additional information from the knowledge bases."},{"figure_ref":[],"heading":"Ethical Considerations","publication_ref":["b15"],"table_ref":[],"text":"Fine-grained entity typing is an important task for many downstream applications. We are not aware of any ongoing research on potential biases in finegrained entity typing. However, bias has been reported and explored in the use of Masked Language Models (Nangia et al., 2020), which is leveraged to generate ONTOTYPE's candidate types. In addition, we present a newly annotated NYT dataset which we will release for use in future studies. Finally, we also leverage the existing Ontonotes and FIGER test data sets for baseline comparisons."},{"figure_ref":[],"heading":"A Appendix","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"A.1 Efficiency & Scalability Study","publication_ref":[],"table_ref":["tab_8"],"text":"We conduct an efficiency and scalability study to gauge the impact of larger scale data sets on our method and include the results in Table 5. "},{"figure_ref":[],"heading":"Number of Mentions","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"A.2 Parameter Settings","publication_ref":["b19","b2","b3","b28","b25","b16","b4","b12","b14"],"table_ref":[],"text":"For AFET (Ren et al., 2016), UFET (Choi et al., 2018), and MLMET (Dai et al., 2021), we utilize the default parameters as detailed in their studies.\nFor the zero-shot methods, ZOE (Zhou et al., 2018), OTyper (Yuan and Downey, 2018), DZET (Obeidat et al., 2019) and MZET (Zhang et al., 2020b), we leverage the default parameters as detailed in their respective studies. ONTOTYPE leverages a pre-trained BERT (Devlin et al., 2019) (BERT-base, uncased) and pre-trained RoBERTa fine-tuned on the MNLI dataset (Liu et al., 2019) available in the HuggingFace Library. In addition, ONTOTYPE utilizes Word2Vec 3 (Mikolov et al., 2013) embeddings to construct our type embeddings."},{"figure_ref":[],"heading":"A.3 Study of Sensitivity to Parameters","publication_ref":[],"table_ref":[],"text":"A potential concern with the experimental setup can be overtly high sensitivity of ONTOTYPE to the Entity Type Specificity parameter θ. For all experiments in Table 3, we leverage the same θ value of 0.3. Additionally, from the plot in Figure 5, we clearly see that F1 scores are not drastically sensitive to theta with standard deviations of 0.3631 and 0.6262 on FIGER and OntoNotes respectively."},{"figure_ref":[],"heading":"A.4 Ablation Study","publication_ref":["b19","b2","b3","b9","b28","b25","b16"],"table_ref":[],"text":"We also include the prediction results of our ablation models to demonstrate how the NLI module contributes to the final type assignment. We utilize a simple type mapping to evaluate our ensembled BERT Cloze Prompting module. Figure 6 shows the results of ablation studies on the test set of the Ontonotes dataset. We compared the ONTOTYPE full model with various ablations and variations. We find that direct prompting\\NLI performs the (Ren et al., 2016) 55.3 66.4 69.3 UFET (Choi et al., 2018) ---BERT-MLMET (Dai et al., 2021) ---LITE (Li et al., 2022) 66.2 74.7 80.1\nZero-Shot ZOE (Zhou et al., 2018) 58.8 71.3 74.8 OTyper (Yuan and Downey, 2018) 47.2 67.2 69.1 DZET (Obeidat et al., 2019) 28.5 56.0 55.1 MZET (Zhang et al., 2020b) 31.9 57.9 55.5 ONTOTYPE (Ours) 49.1 67.4 75.1 ONTOTYPE + Modified FIGER Type Ontology (Ours) 51.1 68.9 77.2\nTable 7: Results (%) on the FIGER test set with Provided and Modified Type Ontology ontologies where the parent type is a generalization of the child type. The included FIGER type ontology considers the parent type of the \"Road\" type to be \"Other\\Transportation\" and the parent type of \"Building\" to be \"Other\". While \"Other\" can be considered a valid generalization, it is extremely broad and the coarse-grained type of \"Location\" serves as a stronger parent for both finegrained types. Thus, we modify the included ontology by reorganizing the fine-grained types under parent types that have stronger generalizations (e.g. Building & Road under Location rather than Other). ONTOTYPE achieves 2.1 absolute Ma-F1 improvement given this modified type ontology. Thus, we find that ONTOTYPE can be further improved with correct type ontologies to leverage the inherent hierarchical relationships between coarse and fine-grained types.\nA.6 Analysis of Typing Errors ONTOTYPE, though having high performance, still generates nontrivial errors. We analyze the reasons behind the errors on the Ontonotes and FIGER test sets and categorize them into three types. Note the Gold Type for each mention is highlighted in blue.\nError Type 1: Lack of Knowledge Base Sentence E1: He was a caseworker in Minnesota [\\Location\\Region] but left the job because he found himself perpetually sick from the environments in which he worked. In E1, ONTOTYPE incorrectly types Minnesota as a Country rather than the gold type of Region. In this context, we might consider State to be a more reasonable type to predict. With a simple KB-matching method, we would be able to capture the surface-level information to type it as a state/province. ONTOTYPE relies on neither human annotations nor knowledge base. Obviously, introducing a knowledge base and KB-matching mechanism will further improve its performance.\nError Type 2: Incomplete Fine-Grained Type Ontology Sentence E2: Valley Federal Savings & Loan Association said Imperial Corp. of America withdrew from regulators its application to buy five Valley Federal branches [\\Location\\Structure], leaving the transaction in limbo. In E2, ONTOTYPE generates \\Other. Even when ONTOTYPE is able to generate high-quality candidate labels, it can sometimes fail to align to the correct entity type due to an incomplete type set. In this example, our Candidate Type set generated by prompting consists of Asset, Property, Facility, Bank, and Branches. Clearly, the best fine-grained type should be Asset or Property (rather than the Gold Type: Location or Structure). Since the provided OntoNotes ontology does not include such fine-grained types, ONTOTYPE is unable to generate the correct answer. Clearly, a refined ontology will further improve its performance.\nError Type 3: Incapability to Type Nested Entities Sentence E3: The 33-year-old Billings [\\Loca-tion\\City] native enlisted as a military veterinarian.\nIn E3, ONTOTYPE identifies Billings as \\Person. This mistake can be caused by confusing the type of the whole entity \"Billings native\" with the type of the nested entity \"Billings\". PLM is good at generating candidate types for the whole entity based on its contextual structure, whereas the knowledge about a nested entity like \"Billings\" (in front of head word \"native\") can be more easily derived from a knowledge-base or from some nested entity type patterns. We will leave the issue of typing nested entities to future work. The Kremlin's \\Location\\Building effort to shield the Russian \\Person population from the impact of its invasion of Ukraine \\Location\\Country are wearing thin as the war \\Other heads into its second year.\nTrailing two games to one in the NBA Finals \\Other\\Event and facing the daunting task of trying to beat the Boston Celtics \\Organiza-tion\\Company in the hostile environment of TD Garden \\Location\\Building on Friday night, the Warriors knew they needed to summon one of the best efforts of their dynastic run in order to even the best-of-seven series."},{"figure_ref":[],"heading":"ZOE","publication_ref":[],"table_ref":[],"text":"The Kremlin's \\Location\\Building effort to shield the Russian \\Person\\Ethnicity population from the impact of its invasion of Ukraine \\Location\\Country are wearing thin as the war \\Other\\Event heads into its second year.\nTrailing two games to one in the NBA Finals \\Other\\Event and facing the daunting task of trying to beat the Boston Celtics \\Organiza-tion\\Sports_Team in the hostile environment of TD Garden \\Location\\Building\\Sports_Facility on Friday night, the Warriors knew they needed to summon one of the best efforts of their dynastic run in order to even the best-of-seven series. ONTOTYPE The Kremlin's \\Organization\\Government effort to shield the Russian \\Person\\Ethnicity population from the impact of its invasion of Ukraine \\Location\\Country are wearing thin as the war \\Other\\Event\\Conflict heads into its second year.\nTrailing two games to one in the NBA Finals \\Other\\Event\\Finals and facing the daunting task of trying to beat the Boston Celtics \\Organization\\Sports_Team\\Basketball_Team in the hostile environment of TD Garden \\Loca-tion\\Building\\Sports_Facility on Friday night, the Warriors knew they needed to summon one of the best efforts of their dynastic run in order to even the best-of-seven series. "},{"figure_ref":[],"heading":"A.7 Additional Comparative Case Study","publication_ref":[],"table_ref":["tab_10"],"text":"Table 8 presents further sentences from two recent news articles with tagged mentions and predicted entity types. In the first sentence, ONTO-TYPE learns the best label for the entity mention span \"the war\" through the use of our NLI typing method. ZOE is able to leverage similar entries in a knowledge base to align the event type correctly. However, the candidate types generated by MLM prompting allow ONTOTYPE to traverse the ontology structure and find the most fine-grained type that fits the given entity mention. As before, MZET and ZOE heavily leverage the surface level information from large KB to identify \"The Kremlin\" as a fortress or physical building in Moscow. However, the local context makes it clear that \"The Kremlin\" refers to the Russian government. With the aid of pre-trained language models, ONTOTYPE is able to leverage the local context to consolidate the most appropriate fine-grained types.\nIn order to mitigate an over-reliance on surfacelevel information, ZOE attempts to match the concepts that best align given the local context as a constraint. This enables ZOE to recognize that TD Garden refers to a stadium or sports facility. When ONTOTYPE types TD Garden, we learn the most fine-grained label through the use of local context and pre-trained language models. The type indicators (Venues, Places, Sites, Locations) allow us to consolidate the high-level type of Location. However, by leveraging the surrounding contextual clues and knowledge learned during the training, NLI models enable ONTOTYPE to establish that TD Garden refers to a building or specifically a sports facility.\nFinally, ONTOTYPE is able to derive the most fine-grained and accurate types by leveraging the inherent structure of the input type ontology. Consider the mention span of \"The NBA Finals\". The predicted candidate labels and head words (Finals, Playoffs, Semifinals, Rounds) indicate that the type is related to an Event. Thus, with the assistance of the pre-trained language models and head words, we select the most context-aware type of \\Other\\Event. Thus with these information sources and the ontology structure as guidance, ONTOTYPE can refine and select the final type of \\Other\\Event\\Finals."},{"figure_ref":[],"heading":"A.8 Evaluation Metrics","publication_ref":["b3","b10","b13"],"table_ref":[],"text":"Following the prior FET studies ( (Dai et al., 2021), (Ling and Weld, 2012), (Ma et al., 2016)), we evaluate our methods and the baselines using three evaluation metrics: Strict Accuracy (Acc), Micro-F1 (Mi-F1), and Macro-F1 . Accuracy. Given a set of entity mentions M , we denote the set of ground truths and predicted types as t M and t M respectively. Given σ as an indicator function, strict accuracy is defined as\nAcc = Σ m∈M σ(t m == t m ) M\nMacro-F1. Macro-F1 is calculated using Macro-Precision (P ma ) and Macro-Recall (R ma ) where\nP ma = 1 |M | Σ m∈M |t m ∩ t m | t m R ma = 1 |M | Σ m∈M |t m ∩ t m | t m\nMicro-F1. Micro-F1 is calculated using Micro-Precision (P mi ) and Micro-Recall (R mi ) where\nP mi = Σ m∈M |t m ∩ t m | Σ m∈M t m R mi = Σ m∈M |t m ∩ t m | Σ m∈M t m\nMacro-F1 and Micro-F1 are calculated using the F1 score formula with their respective granular precision and recall scores. "}],"string":"[\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b10\",\n \"b17\",\n \"b8\",\n \"b17\",\n \"b10\",\n \"b2\",\n \"b3\",\n \"b9\",\n \"b5\"\n ],\n \"table_ref\": [],\n \"text\": \"Fine-grained entity typing (FET) is an important task in text analysis. Assigning fine-grained semantic types to parsed entity mention spans based on the local context enables effective and structured analysis of unstructured text data, such as entity linking (Ling and Weld, 2012;Onoe and Durrett, 2020), relation extraction (Koch et al., 2014), and coreference resolution (Onoe and Durrett, 2020). Example 1. Given a sentence S 1 : \\\"Sammy Sosa got a standing ovation at Wrigley Field.\\\" and a parsed entity mention span \\\"Sammy Sosa\\\" in the sentence, an FET method aims to assign it not only the coarse-grained type \\\"Person\\\" but also the finegrained types \\\"Athlete\\\" or \\\"Player\\\".\\nFET on large text corpora is challenging due to (1) the high cost to obtain a large amount of humanannotated training data, especially in dynamic and emerging domains, and (2) inaccurate annotations due to (i) different annotators marking concepts at different granularity (e.g., person vs. politician vs. president), and (ii) contextual subtlety on finegrained types (e.g., Boston vs. Detroit could be two sport teams, instead of two cities). Most existing methods utilize weak or distant supervision to automatically generate training data for the FET tasks. There are three major approaches to obtaining weakly-labeled training data to tackle these challenges. The first is to automatically match the mentions in text with the concepts in some existing knowledge bases (e.g., Wikipedia) (Ling and Weld, 2012). The typical workflow is to first detect entity mentions from a corpus, map these mentions to knowledge base (KB) entities of target types, and then leverage the confidently mapped types as labeled data to infer the final type. The second is to directly obtain the head words of nominal mentions as its fine-grained type (Choi et al., 2018). It leverages the head words of the entity mention to consolidate context-aware types derived from the KB matching. However, both approaches suffer from the label sparsity and context-agnostic problems, resulting in the inability to generate high-quality training data for FET.\\nThe third approach is to probe the pre-trained language models through the use of masked patterns and entailment templates. Leveraging masked language model (MLM) prompting to generate rich and context-aware weak supervisions for finegrained entity typing is a recent trend, aiming to reduce expensive human labor (Dai et al., 2021;Li et al., 2022).For example, given a sentence that con-tains a mention, a short piece of text that contains a \\\"[MASK]\\\" token is added to generate candidate entity types. This method conducts labeling in a context-aware manner and greatly enriches the fine-grained types labeled for each mention. However, such a process can still generate tokens unsuitable for typing (e.g., {Teams, Thing} for \\\"Wrigley Field\\\" in S 1 ) or a mixture of rough and fine-grained types (e.g., {Location, Building, Stadium}). Due to the lack of further hierarchical knowledge of the generated tokens/types, such problems cannot be resolved automatically.\\nIn this study, we vision that an ontology structure, which provides a semantics-rich, hierarchical structure, may help select the best results generated by multiple PLM models. We propose an zero-shot, ontology-guided, fine-grained entity typing (FET) method, ONTOTYPE, that leverages an input ontology and the power of MLM prompting and Natural Language Inference (NLI). We first ensemble multiple Hearst patterns to perform MLM prompting, reducing the noise in the initial candidate type generation. Since the generated candidate labels for a given mention are likely a mixture of fine and coarse-grained labels, or tokens unsuitable for typing, we propose to automatically match the generated candidate labels to a coarse-grained type in our type ontology and then rank and select a coarsegrained type with a pre-trained entailment model under the local context. Such a type resolution process will progress deeper to finer levels, based on the same principle of entailment score-based type selection, following the type ontology, until the finest possible label can be consolidated.\\nExample 2. For sentence S 1 in Ex. 1, candidate type generation (Step 1) ensembles prompting results of multiple Hearst patterns and generates a set of candidate labels {\\\"Stadium\\\", \\\"Venues\\\", \\\"Locations\\\", \\\"Games\\\"} for \\\"Wrigley Field\\\" (Fig. 2). Using a given ontology structure (Fig. 1), this set of types is resolved to the course-grained type \\\"Location\\\" by leveraging a pre-trained entailment model and the local context (Fig. 3). Note that without the ontology type structure, \\\"Stadium\\\", \\\"Venues\\\", and \\\"Locations\\\" are rivals; but with the structure, \\\"Stadium\\\" and \\\"Venues\\\" are fine-grained \\\"Locations\\\" and they support each other at multilevel resolution. By the same principle, the type resolution proceeds deeper to finer-grained levels, along the type ontology, from \\\"Location\\\" to \\\"Building\\\" and further down to \\\"Stadium\\\" for \\\"Wrigley Field\\\", leading to the most accurate fine-grained type (Fig. 4). Our contributions are summarized as follows.\\n1. A fully zero-shot, ontology-guided, fine-grained typing method, ONTOTYPE, is proposed 2. ONTOTYPE improves fine-grained entity typing (FET) by leveraging candidate labels generated and refined with three information sources: (i) pre-trained language models, (ii) a fine-grained type ontology, and (iii) head words 3. Experiments on the Ontonotes, FIGER, and NYT datasets (Gillick et al., 2014) using their associated ontological structures show that ON-TOTYPE clearly outperforms existing zero-shot named entity typing methods and even rivals supervised methods. Our error analysis shows that refinement of ontology structures will further improve fine-grained entity typing.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Work\",\n \"publication_ref\": [\n \"b23\",\n \"b10\",\n \"b6\",\n \"b2\",\n \"b10\",\n \"b24\",\n \"b2\",\n \"b3\",\n \"b22\",\n \"b25\",\n \"b28\",\n \"b19\",\n \"b13\",\n \"b4\",\n \"b3\",\n \"b11\",\n \"b9\",\n \"b3\",\n \"b9\",\n \"b21\"\n ],\n \"table_ref\": [],\n \"text\": \"Supervised FET. Fine-grained entity typing benefits various downstream tasks and has received extensive attention in natural language research. Recent studies focus on different contexts from the phrase level (Yao et al., 2013) to considering specific entity mentions in the sentence or document level (Ling and Weld, 2012;Gupta et al., 2017;Choi et al., 2018). Entity typing has been generally studied under supervised learning settings (Ling and Weld, 2012;Yosef et al., 2012;Choi et al., 2018;Dai et al., 2021). These studies require annotated data to achieve high performance and lack the flexibility to identify newly defined entities.\\nWeakly supervised or zero-shot FET. To handle such difficulties, the zero-shot learning setting has been introduced for named entity typing (Xia et al., 2018). Several studies (Yuan and Downey, 2018;Zhou et al., 2018) address the problems by grounding the mentions with Wikipedia entries from the assembled related pages. These methods achieve good performance but also require a lot of resources. Other studies explore the pre-trained semantic word-level embeddings (Ren et al., 2016;Ma et al., 2016) or extract raw embeddings without auxiliary information and utilize end-to-end neural networks (Zhang et al., 2020b). However, they still suffer from low accuracy and inefficiency in zeroshot settings. As a result, ONTOTYPE turns to the weak supervision from pre-trained language models, such as BERT (Devlin et al., 2019), that have substantial knowledge in language understanding.\\nLeveraging pre-trained language models. Some recent studies (e.g., (Zhang et al., 2020a;Dai et al., 2021;Liu et al., 2021;Li et al., 2022)) leverage pre-trained language models and prompting templates to obtain knowledge for entity mentions in given sentences. (Dai et al., 2021) improves ultrafine entity typing with BERT Masked Language Model (MLM). However, they use a single prompt for each mention to generate labels for weak supervision, which may generate erroneous types. (Li et al., 2022) improves ultra-fine entity typing by treating the task of predicting an entity type as an NLI task. However, they ignore rich semantic information contained in a type ontology during the typing process. Our method ensembles multiple MLM prompting and NLI results to reduce noises in candidate type generation. Further, we utilize the fine-grained type ontology structure as guidance to progressively resolve candidate labels from coarse to fine under the local context. ChemNER (Wang et al., 2021) leverages a type ontology structure to guide fine-grained Chemistry entity recognition. It relies on an existing knowledge base whereas this study leverages the pre-trained language models as rich and context-aware weak supervision.\"\n },\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"Methodology\",\n \"publication_ref\": [\n \"b12\"\n ],\n \"table_ref\": [],\n \"text\": \"We propose ONTOTYPE, a zero-shot, ontologyguided, fine-grained entity typing method using pre-trained language models and a fine-grained ontology structure. Given an input sentence and a set of pre-identified mentions in the sentence, ONTO-TYPE consists of the following steps: (1) generating a set of candidate labels for each input mention with both head word parsing and an ensemble of MLM prompting (Fig. 2); (2) resolving the coarsegrained types by matching and ranking the generated labels to the type ontology using an entailment model (Fig. 3); and (3) progressively refining the fine-grained types along the type ontology following the principle of entailment score-based type selection (Fig. 4). We utilize the inherent structure of the fine-grained type ontology and a pre-trained entailment model (RoBERTa model pre-trained on the MNLI dataset (Liu et al., 2019)) to guide our fine-grained entity typing.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Problem Definition\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The input to our proposed ONTOTYPE framework is a text corpus D and input fine-grained type ontology O. In this study, we assume our input text corpus D includes a set of pre-identified entity mentions. An entity mention, e, is a token span in the text document that refers to a real-world entity. Given a sentence S and a parsed entity mention e ∈ S, the fine-grained entity typing (FET) task is to identify one or more types t from the label space T (provided in a structured ontology G) for the entity mention e.\\nAs an example of our FET task, in the sentence S 1 of Example 1, the entity mention to be typed is e 1 : \\\"Wrigley Field\\\". It should be labeled progressively deeper as \\\"Location→Building→Stadium\\\" as opposed to other labels like \\\"Organization\\\", \\\"Person\\\", or \\\"School\\\".\\nFine-Grained Type Ontology Structure. The structure of the type ontology is fundamental to the ONTOTYPE algorithm. In this work, we leverage the fine-grained type ontologies provided in the OntoNotes and FIGER datasets. These ontologies are structured as forests or a disjoint union of trees. Each tree consists of a type hierarchy stemming from a handful of \\\"root\\\" coarse-grained types that include but are not limited to \\\"Organization\\\", \\\"Person\\\" and \\\"Location\\\". Each tree imposes the following structural conditions: (1) Each type has a singular parent type; (2) each type can have an infinite (potentially) number of children; and (3) each type is connected by a directional edge indicating a hypernym-hyponym type relationship between the parent and child nodes. In our experiments, we find that general types such as \\\"things\\\" lead to inaccurate labels as pre-trained entailment models often rank vague types higher than specific types. Note that in the given OntoNotes type ontology (Fig. 1), Building and Country are sibling types since they share the same parent type Location. Such a hypernym-hyponym or \\\"is-a\\\" relationship is critical for guiding fine-grained type selection.\"\n },\n {\n \"figure_ref\": [\n \"fig_1\",\n \"fig_1\"\n ],\n \"heading\": \"Candidate Type Generation\",\n \"publication_ref\": [\n \"b2\",\n \"b3\",\n \"b2\",\n \"b1\",\n \"b4\",\n \"b7\",\n \"b20\"\n ],\n \"table_ref\": [\n \"tab_0\"\n ],\n \"text\": \"To generate a set of candidate labels for each input mention, we leverage two main information sources: (1) head word parsing, and (2) an ensemble of MLM prompting. Head words and hypernyms can serve as powerful context-aware type indicators that can be leveraged as weak supervision sources for entity typing (Choi et al., 2018;Dai et al., 2021). Our method adopts both information sources to first select a candidate label to guide our fine-grained type refinement. First, we generate and parse the head words for the given entity mention. Second, we generate candidate entity types with the use of ensembled MLM prompting1 . The generated candidate labels are used as input to the following steps of ontology-guided type resolution.\\nHead Word Parsing. As discussed by (Choi et al., 2018), the input text often contains cues that explicitly match a mention to its type, in the form of the mention's head word. Thus, given the pre-identified entity mentions in the input sentence, we first identify the head word of the input mention. We utilize the Stanford Dependency Parser (Chen and Manning, 2014) to extract the head word of the entity that we are interested in typing. Consider the sentence \\\"Governor Arnold Schwarzenegger gives a speech at Mission Serve's service project on Veterans Day 2010\\\" with \\\"Governor Arnold Schwarzenegger\\\" as an entity mention. We can describe \\\"Arnold Schwarzenegger\\\" with numerous types including actor, father, and governor. However with the use of the head word \\\"Governor\\\", we consolidate the fine-grained type of governor given this input sentence. Thus, ONTOTYPE leverages the head words of the input entity, if any, to select an initial context-sensitive coarse-grained type and guide the selection of the final fine-grained type.\\nEnsembled MLM Prompting. While head words can provide strong type indicators, they do not always provide sufficient information to consolidate a high-level type. In some cases, head words cannot directly provide accurate type information as they are not directly present in the input type ontology. Furthermore, head words are not always available in the input sentences. Thus, with the parsed entity mention span as input, we propose to leverage context-aware hypernyms as initial type labels for the target mention. With an ensembled cloze prompting method, ONTOTYPE generates candidate labels of the mention and performs an initial high-level typing on the input mention. Specifically, we leverage a BERT masked language model (Devlin et al., 2019) and artificial Hearst patterns (Hearst, 1992) to generate context-aware hypernyms that serve as labels for the mentions. We first modify the input sentence by inserting a Hearst pattern and [MASK] token into the sentence. Then we use the BERT MLM to generate candidate labels for the target mention under the local context. For example, in Fig. 2, we first insert Hearst patterns such as \\\"and the other [MASK]\\\" in the input sentence and then use the BERT MLM model to generate candidate labels such as \\\"Venues\\\", \\\"Teams\\\", and \\\"Stadiums\\\".\\nWe evaluated the quality of hypernyms generated with the 44 patterns proposed in (Seitner et al., 2016) on the Ontonotes Development Dataset. When generating hypernyms, we expect high-quality candidates to be semantically equivalent to concepts contained in our input type ontology. Overall, we found that the four patterns in Table 1 provided the highest quality hypernyms under a simple direct matching to types in the OntoNotes ontology. However, based on the syntax and grammar of the sentence, hearst patterns can generate tokens that are unsuitable to serve as fine-grained entity types. For example, with a single prompt, the MLM can generate \\\"Famous\\\", \\\"Actors\\\", \\\"Celebrities\\\" and \\\"People\\\" as the most probable words. Note that \\\"Famous\\\" is unsuitable to serve as a fine-grained entity type. To reduce the noises caused by the use of a single Hearst pattern, we ensemble n Hearst Patterns to consolidate the most commonly generated candidate labels. For each pattern in the pattern list P , we collect the top k most probable tokens from the probability distribution predicted by the BERT MLM. Then, we aggregate the tokens and identify the set of labels that have the largest overlap. We perform the voting ensemble as follows:\\ny = count (m) {H 1 (x), H 2 (x), ..., H n (x)} (1)\\nwhere H n (x) is the candidate type generated by the n th hearst pattern and count (m) is the function that selects all candidate labels generated at least m times. In our experiments, we use m = n 2 + 1. In addition, we find that the number of ensembled patterns is not sensitive as ensembling ensures we retain the most confident candidate labels. Example 3: As shown in Fig. 2, by ensembling the results from prompting with several Hearst patterns, the quality types for e 1 \\\"Stadiums, Venues, Locations, Games\\\" retain but the noisy types \\\"Things\\\" and \\\"Teams\\\" are removed.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"High-Level Type Resolution\",\n \"publication_ref\": [\n \"b14\"\n ],\n \"table_ref\": [],\n \"text\": \"Given the generated candidate labels and head words for each mention in the sentence, we seek to resolve the concrete type for this candidate type set at the high levels of the type ontology. Specifically, we first align the generated candidate labels to several high-level types in the type ontology, and then select the most accurate high-level types with a pre-trained entailment language model.\\nCandidate Type Alignment. Following the previous step of candidate type generation, we combine the generated candidate labels from both the parsed head words and the ensembled cloze prompting to form a candidate type set. These labels are generally noisy and may not exist directly in our type ontology. However, we observe that the generated candidates will usually cluster closely around a high-level concept in the ontology. Thus, we perform our high-level type alignment with a cosinesimilarity-based word matching. We use Word2Vec 2 (Mikolov et al., 2013) embeddings to construct our type embeddings for the cosine-similarity-based type alignment. We construct a verbalizer by selecting at least l seman-2 https://code.google.com/archive/p/word2vec/ tically related words for each coarse type. For a high-level type of \\\"Organization\\\", we might include seed types such as \\\"Corporation\\\", \\\"University\\\", \\\"Firm\\\", \\\"Business\\\", and \\\"Government\\\" in its verbalizer. This verbalizer is then utilized to systematically map the MLM hypernym prediction to the most relevant type. In our experiments, we provide at least five seed types S for each type node c in the first level of the type ontology. Increasing the number of seed types increases the coverage and confidence of the verbalizer. Since the most commonly generated hypernyms for each concept exists in the input ontologies, ONTOTYPE is not so sensitive to the number of seed types collected. With the seed types, we construct a node embedding N by taking the average of word embeddings from both the first-level type and its corresponding seed types provided by human,\\nN = s i ∈S emb(s i ) + emb(c) |S| + 1 . (2\\n)\\nwhere emb(•) indicates the Word2Vec embeddings. Then we rank each generated candidate type to a first-level type on our type ontology by calculating the cosine similarity between the embeddings of the generated candidate labels b and that of the first-level types T , score(b, T ) rank = cosine(emb(b), emb(T )).\\n(3)\\nFinally, the first-level type that has the highest similarity is selected as the aligned high-level type.\"\n },\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"⋮ Figure 3: Candidate Type Generation\",\n \"publication_ref\": [\n \"b12\"\n ],\n \"table_ref\": [],\n \"text\": \"Example 4: In Figures 2 and3, the consolidated candidate labels \\\"Locations\\\", \\\"Stadiums\\\", \\\"Venues\\\", and \\\"Games\\\" are closely related to the high-level type \\\"location\\\" in our ontology. By performing the cosine-similarity-based matching, \\\"Location\\\" is identified as e 1 's high-level type over \\\"Organization\\\" or \\\"Person\\\". Thus, we finally select the most similar high-level type given by the head word, generated candidate labels, and the entailment model.\\nHigh-Level Type Selection After the previous step of candidate type alignment, we obtain several high-level types for each entity mention in the sentence. Given these types, we seek to select the most accurate high-level type for each entity mention under the local context. We observe that the task of selecting the most appropriate entity type can be viewed as a Natural Language Inference (NLI) task. Thus, we treat the input sentence as the premise in NLI and generate the hypothesis using a pre-defined declarative template. To resolve the type of the input mention we utilize the following template: \\\"In this sentence, [MENTION] is a [TYPE]\\\". ONTOTYPE then ranks each type in the first level of the ontology with the entailment score from a RoBERTa NLI model (Liu et al., 2019).\\nExample 5: In Fig. 3, we have aligned a majority of the generated candidate type set to the Location seed types. The NLI model further ranks Location over Organization or Person. By utilizing the information in conjunction, e 1 is solidly aligned to the high-level type \\\"Location\\\".\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Fine-Grained Type Resolution\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Given the high-level types of the entity mentions, ONTOTYPE further leverages the ontology structure to progressively resolve the fine-grained label.\\nFollowing the same principle of entailment-based type selection for the high-level types, we utilize the entailment model to automatically select the most accurate fine-grained entity types along our type ontology. Specifically, we first examine the child types of the previously determined higherlevel types and then select the child type with the highest ranked score as the fine-grained type.\\nIn addition to the entailment model, we also utilize the candidate type set to refine our fine-grained type selection. If the parsed head word is present in our type ontology, the entailment scores of that parent and its child types are weighted higher. Similarly, if the generated candidate labels are in our type ontology, the entailment scores of those parents and their child types are also weighted higher. Finally, we select the fine-grained type for each mention with the highest weighted entailment score.\\nWe calculate the ranking score for the entities at the current level of the ontology as follows:\\nrank(type) = σ entail + σ cand(4)\\nWe first leverage our NLI pre-trained model to find the entailment score σ entail for each entity type. Then if a type in the candidate type set is a descendent to the entity type, we add a weight σ cand . We repeat this entailment-based selection process recursively along the type ontology to select the best fine-grained type.\\nDefinition (Entity Type Granularity Parameter θ):\\nWe assume there is a scalar of θ indicating how granular or specific the final selected entity type label should be. The smaller θ is, the more granular entity type labels are consolidated as the final label. Thus, if the child types do not have a sufficient gain of at least θ in ranking score over the parent type at a certain level, we stop the recursion and select the parent type as the final fine-grained type. We conduct a parameter study to explore the sensitivity of ONTOTYPE to the parameter θ in Appendix A.3\\n⋮ Figure 4: Fine-Grained Type Refinement\\nExample 6: In Fig. 4, we consider all descendent types of \\\"location\\\" as potential fine-grained entity types for e 1 . To begin, ONTOTYPE generates the hypotheses and ranks all child types of \\\"location\\\". With the resultant entailment score rankings of these sibling types, we consolidate and select the \\\"building\\\" type as the highest-ranked label. Then, a similar process is done at a deeper level of the ontology to select the final type of \\\"stadium\\\".\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiments\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Datasets\",\n \"publication_ref\": [\n \"b5\",\n \"b10\"\n ],\n \"table_ref\": [],\n \"text\": \"We compare the performance of ONTOTYPE and several baseline methods on three benchmark FET datasets: NYT, Ontonotes (Gillick et al., 2014) and FIGER (Ling and Weld, 2012). The basic statistics of the datasets are shown in Table 3: Results (%) on Three Test Sets (Some slots in supervised methods marked \\\"-\\\" due to no fully annotated training data).\\nOntoNotes and FIGER datasets, we leverage the included type ontologies while the NYT dataset leverages the input FIGER ontology. All NER test sets are annotated using the ontologies as a set of type labels. Thus, each entity mention is labeled with a fine-grained label represented as a path within the ontology. The ontologies have a maximum depth of three and contain four to six high-level types (e.g., Location, Person, and Organization).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Baselines\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"ONTOTYPE is a zero-shot FET method that does not require human annotation as supervision. We compare ONTOTYPE with two groups of baseline methods for FET: supervised methods and zeroshot methods. Details of the parameter settings for each model is listed in Appendix A.2. We also conduct parameter and ablation studies listed in Appendices A.3 and A.4 respectively.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Supervised FET Methods\",\n \"publication_ref\": [\n \"b19\"\n ],\n \"table_ref\": [],\n \"text\": \"We compare the performance of ONTOTYPE with three supervised FET methods as discussed below.\\nAFET (Ren et al., 2016) MZET (Zhang et al., 2020b): This zero-shot baseline leverages the semantic meaning and the hierarchical structure into the type representation. The method leverages the knowledge from seen types to label the zero-shot types through the use of a memory component.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Main Results\",\n \"publication_ref\": [\n \"b28\"\n ],\n \"table_ref\": [],\n \"text\": \"Table 3 strates the benefit of leveraging the knowledge embedded in pre-trained language models as a form of minimal supervision to identify entity mention types. ZOE leverages a type ontology and maps a given mention to type-compatible Wikipedia Entries. As a result, ZOE relies on surface-level information from the mention string. OntoType ensembles contextual information from various PLMs to consolidate the final entity type. Given a sentence: \\\"The biggest cause for concern for McGuff is the bruised hamstring Regina Rogers suffered against (Utah) last Saturday\\\", ZOE incorrectly utilizes the surface string to label \\\"Utah\\\" as a location. With PLMs, OntoType recognizes teams or opponents as candidates and finally consolidates \\\"sports_team\\\". On the FIGER dataset, ONTOTYPE achieves 0.3 absolute Macro-F1 improvement over state-of-the-art zero-shot fine-grained entity typing method ZOE (Zhou et al., 2018). However, our method trails ZOE in strict accuracy and Micro-F1. In the FIGER dataset, predictions are made based on both the surface-level information and the contextual information in the sentence. The major advantage of ONTOTYPE is to accurately capture the contextual information to provide a fine-grained entity type. However, ONTOTYPE does not involve a mechanism to capture the surface-level informa-tion of an entity mention. We discuss this issue further in our error analysis (Appendix A.6).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Comparative Case Study\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_6\"\n ],\n \"text\": \"Table 4 presents a sentence from a recent news article with tagged mentions and predicted entity types. We find that methods like MZET sometimes predict incompatible types due to incorrect or misleading surface information. For example, when typing \\\"The White House\\\", MZET and ZOE leverage the surface-level information from large KBs to identify the mention as a location. However, given the local context, the White House clearly refers to the U.S. government. When considering the mention \\\"US President Joe Biden\\\", our method utilizes the type information from the candidate type set (Officials, Leaders, Politicians, Individuals) to explore the \\\"Person\\\" branch within the ontology. We search the \\\"Politician\\\" branch for the best fine-grained context-aware type to label our given entity mention. Thus, with the assistance of the pre-trained language models, we can incorporate contextual information to derive more context-aware types.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Error Analysis\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The structure of the fine-grained type ontology is important to the performance of ONTOTYPE. Input ontologies should utilize hypernym-type relationships where each parent type is a generalization of the child type. The provided FIGER ontology contains logical inconsistencies in how various types of relations are organized. Specifically, the FIGER ontology considers the parent of the \\\"Road\\\" type to be \\\"Transportation\\\" rather than \\\"Location\\\". This logical inconsistency leads to erroneous typing from our method. We further examine our method's errors and evaluate ONTOTYPE using a modified type ontology in Appendix A.6.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Large Language Models\",\n \"publication_ref\": [\n \"b18\"\n ],\n \"table_ref\": [],\n \"text\": \"Traditional language models (LM) aim at assigning a Probability P (w) to an input sequence of n words s = [w 1 , w 2 , ..., w n ]. Recent works include large language models (LLMs) such as ChatGPT and GPT-4 (OpenAI, 2023), which are pre-trained on with a large text corpora enables them to generate contextualized word representation to achieve state of the art performance on various NLP tasks including language translation, text summarization, and question-answering. These recent works provide a richer knowledge base for ONTOTYPE's candidate type generation. As a result, we can similarly leverage models like T-5 (Raffel et al., 2020) or GPT-3 (Brown et al., 2020) to obtain high-quality candidate types through hypernym prompting methods.\\nExample 6: Given a sentence S 1 : \\\"Sammy Sosa got a standing ovation at Wrigley Field.\\\" and a parsed entity mention span \\\"Sammy Sosa\\\", Chat-GPT generates \\\"Athletes\\\", \\\"Players\\\", \\\"Teammates\\\", \\\"Stars\\\" and \\\"Figures\\\" while T-5 generates \\\"Athletes\\\", \\\"Players\\\", \\\"Teammates\\\" and \\\"Individuals\\\". Overall, we find that even LLMs can leverage the surrounding context and learned \\\"common-sense\\\" knowledge to generate high quality candidate types. ONTOTYPE provides a structured approach to provide guidance for LLMs like ChatGPT and GPT-4 which are extremely susceptible to hallucinations. By grounding our types to mention hypernyms and a fine-grained type ontology, we can confidently extract the most fine-grained type for the given entity mention span.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusions\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We propose ONTOTYPE, which leverages weak supervision setting of pre-trained language model prompting. We use head words and MLM cloze prompting for fine-grained candidate label generation. Then we automatically match the generated fine-grained types to our type ontology with an inference method to select the most appropriate finegrained types under the local context. Extensive experiments on real-world datasets show that ON-TOTYPE is highly effective and substantially outperforms the state-of-the-art zero-shot FET methods. In the future, we plan to further refine and enrich a type ontology which will enable us to incorporate more type information for even better performance.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Limitations\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"ONTOTYPE utilizes pre-trained language models to perform context-aware fine-grained entity typing leveraging the rich structural information from type ontologies, head words and cloze prompting. However, this method still has a few limitations. Since the pre-trained language models utilized are not fine-tuned, some candidate labels and fine-grained types generated from the MLM and NLI models are noisy and unsuitable to serve as fine-grained types. One future improvement of ONTOTYPE is to finetune the MLM and NLI models based on the local corpus and type ontology. Furthermore, ONTO-TYPE ignores the surface-level information of the entity mentions. Prior distantly-supervised methods have been able to leverage the surface-level information through entity linking to the knowledge bases. Another future improvement of ONTOTYPE is to incorporate these rich information sources to perform typing on the entity mentions with the additional information from the knowledge bases.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Ethical Considerations\",\n \"publication_ref\": [\n \"b15\"\n ],\n \"table_ref\": [],\n \"text\": \"Fine-grained entity typing is an important task for many downstream applications. We are not aware of any ongoing research on potential biases in finegrained entity typing. However, bias has been reported and explored in the use of Masked Language Models (Nangia et al., 2020), which is leveraged to generate ONTOTYPE's candidate types. In addition, we present a newly annotated NYT dataset which we will release for use in future studies. Finally, we also leverage the existing Ontonotes and FIGER test data sets for baseline comparisons.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A Appendix\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.1 Efficiency & Scalability Study\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_8\"\n ],\n \"text\": \"We conduct an efficiency and scalability study to gauge the impact of larger scale data sets on our method and include the results in Table 5. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Number of Mentions\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.2 Parameter Settings\",\n \"publication_ref\": [\n \"b19\",\n \"b2\",\n \"b3\",\n \"b28\",\n \"b25\",\n \"b16\",\n \"b4\",\n \"b12\",\n \"b14\"\n ],\n \"table_ref\": [],\n \"text\": \"For AFET (Ren et al., 2016), UFET (Choi et al., 2018), and MLMET (Dai et al., 2021), we utilize the default parameters as detailed in their studies.\\nFor the zero-shot methods, ZOE (Zhou et al., 2018), OTyper (Yuan and Downey, 2018), DZET (Obeidat et al., 2019) and MZET (Zhang et al., 2020b), we leverage the default parameters as detailed in their respective studies. ONTOTYPE leverages a pre-trained BERT (Devlin et al., 2019) (BERT-base, uncased) and pre-trained RoBERTa fine-tuned on the MNLI dataset (Liu et al., 2019) available in the HuggingFace Library. In addition, ONTOTYPE utilizes Word2Vec 3 (Mikolov et al., 2013) embeddings to construct our type embeddings.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.3 Study of Sensitivity to Parameters\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"A potential concern with the experimental setup can be overtly high sensitivity of ONTOTYPE to the Entity Type Specificity parameter θ. For all experiments in Table 3, we leverage the same θ value of 0.3. Additionally, from the plot in Figure 5, we clearly see that F1 scores are not drastically sensitive to theta with standard deviations of 0.3631 and 0.6262 on FIGER and OntoNotes respectively.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.4 Ablation Study\",\n \"publication_ref\": [\n \"b19\",\n \"b2\",\n \"b3\",\n \"b9\",\n \"b28\",\n \"b25\",\n \"b16\"\n ],\n \"table_ref\": [],\n \"text\": \"We also include the prediction results of our ablation models to demonstrate how the NLI module contributes to the final type assignment. We utilize a simple type mapping to evaluate our ensembled BERT Cloze Prompting module. Figure 6 shows the results of ablation studies on the test set of the Ontonotes dataset. We compared the ONTOTYPE full model with various ablations and variations. We find that direct prompting\\\\NLI performs the (Ren et al., 2016) 55.3 66.4 69.3 UFET (Choi et al., 2018) ---BERT-MLMET (Dai et al., 2021) ---LITE (Li et al., 2022) 66.2 74.7 80.1\\nZero-Shot ZOE (Zhou et al., 2018) 58.8 71.3 74.8 OTyper (Yuan and Downey, 2018) 47.2 67.2 69.1 DZET (Obeidat et al., 2019) 28.5 56.0 55.1 MZET (Zhang et al., 2020b) 31.9 57.9 55.5 ONTOTYPE (Ours) 49.1 67.4 75.1 ONTOTYPE + Modified FIGER Type Ontology (Ours) 51.1 68.9 77.2\\nTable 7: Results (%) on the FIGER test set with Provided and Modified Type Ontology ontologies where the parent type is a generalization of the child type. The included FIGER type ontology considers the parent type of the \\\"Road\\\" type to be \\\"Other\\\\Transportation\\\" and the parent type of \\\"Building\\\" to be \\\"Other\\\". While \\\"Other\\\" can be considered a valid generalization, it is extremely broad and the coarse-grained type of \\\"Location\\\" serves as a stronger parent for both finegrained types. Thus, we modify the included ontology by reorganizing the fine-grained types under parent types that have stronger generalizations (e.g. Building & Road under Location rather than Other). ONTOTYPE achieves 2.1 absolute Ma-F1 improvement given this modified type ontology. Thus, we find that ONTOTYPE can be further improved with correct type ontologies to leverage the inherent hierarchical relationships between coarse and fine-grained types.\\nA.6 Analysis of Typing Errors ONTOTYPE, though having high performance, still generates nontrivial errors. We analyze the reasons behind the errors on the Ontonotes and FIGER test sets and categorize them into three types. Note the Gold Type for each mention is highlighted in blue.\\nError Type 1: Lack of Knowledge Base Sentence E1: He was a caseworker in Minnesota [\\\\Location\\\\Region] but left the job because he found himself perpetually sick from the environments in which he worked. In E1, ONTOTYPE incorrectly types Minnesota as a Country rather than the gold type of Region. In this context, we might consider State to be a more reasonable type to predict. With a simple KB-matching method, we would be able to capture the surface-level information to type it as a state/province. ONTOTYPE relies on neither human annotations nor knowledge base. Obviously, introducing a knowledge base and KB-matching mechanism will further improve its performance.\\nError Type 2: Incomplete Fine-Grained Type Ontology Sentence E2: Valley Federal Savings & Loan Association said Imperial Corp. of America withdrew from regulators its application to buy five Valley Federal branches [\\\\Location\\\\Structure], leaving the transaction in limbo. In E2, ONTOTYPE generates \\\\Other. Even when ONTOTYPE is able to generate high-quality candidate labels, it can sometimes fail to align to the correct entity type due to an incomplete type set. In this example, our Candidate Type set generated by prompting consists of Asset, Property, Facility, Bank, and Branches. Clearly, the best fine-grained type should be Asset or Property (rather than the Gold Type: Location or Structure). Since the provided OntoNotes ontology does not include such fine-grained types, ONTOTYPE is unable to generate the correct answer. Clearly, a refined ontology will further improve its performance.\\nError Type 3: Incapability to Type Nested Entities Sentence E3: The 33-year-old Billings [\\\\Loca-tion\\\\City] native enlisted as a military veterinarian.\\nIn E3, ONTOTYPE identifies Billings as \\\\Person. This mistake can be caused by confusing the type of the whole entity \\\"Billings native\\\" with the type of the nested entity \\\"Billings\\\". PLM is good at generating candidate types for the whole entity based on its contextual structure, whereas the knowledge about a nested entity like \\\"Billings\\\" (in front of head word \\\"native\\\") can be more easily derived from a knowledge-base or from some nested entity type patterns. We will leave the issue of typing nested entities to future work. The Kremlin's \\\\Location\\\\Building effort to shield the Russian \\\\Person population from the impact of its invasion of Ukraine \\\\Location\\\\Country are wearing thin as the war \\\\Other heads into its second year.\\nTrailing two games to one in the NBA Finals \\\\Other\\\\Event and facing the daunting task of trying to beat the Boston Celtics \\\\Organiza-tion\\\\Company in the hostile environment of TD Garden \\\\Location\\\\Building on Friday night, the Warriors knew they needed to summon one of the best efforts of their dynastic run in order to even the best-of-seven series.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"ZOE\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The Kremlin's \\\\Location\\\\Building effort to shield the Russian \\\\Person\\\\Ethnicity population from the impact of its invasion of Ukraine \\\\Location\\\\Country are wearing thin as the war \\\\Other\\\\Event heads into its second year.\\nTrailing two games to one in the NBA Finals \\\\Other\\\\Event and facing the daunting task of trying to beat the Boston Celtics \\\\Organiza-tion\\\\Sports_Team in the hostile environment of TD Garden \\\\Location\\\\Building\\\\Sports_Facility on Friday night, the Warriors knew they needed to summon one of the best efforts of their dynastic run in order to even the best-of-seven series. ONTOTYPE The Kremlin's \\\\Organization\\\\Government effort to shield the Russian \\\\Person\\\\Ethnicity population from the impact of its invasion of Ukraine \\\\Location\\\\Country are wearing thin as the war \\\\Other\\\\Event\\\\Conflict heads into its second year.\\nTrailing two games to one in the NBA Finals \\\\Other\\\\Event\\\\Finals and facing the daunting task of trying to beat the Boston Celtics \\\\Organization\\\\Sports_Team\\\\Basketball_Team in the hostile environment of TD Garden \\\\Loca-tion\\\\Building\\\\Sports_Facility on Friday night, the Warriors knew they needed to summon one of the best efforts of their dynastic run in order to even the best-of-seven series. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.7 Additional Comparative Case Study\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_10\"\n ],\n \"text\": \"Table 8 presents further sentences from two recent news articles with tagged mentions and predicted entity types. In the first sentence, ONTO-TYPE learns the best label for the entity mention span \\\"the war\\\" through the use of our NLI typing method. ZOE is able to leverage similar entries in a knowledge base to align the event type correctly. However, the candidate types generated by MLM prompting allow ONTOTYPE to traverse the ontology structure and find the most fine-grained type that fits the given entity mention. As before, MZET and ZOE heavily leverage the surface level information from large KB to identify \\\"The Kremlin\\\" as a fortress or physical building in Moscow. However, the local context makes it clear that \\\"The Kremlin\\\" refers to the Russian government. With the aid of pre-trained language models, ONTOTYPE is able to leverage the local context to consolidate the most appropriate fine-grained types.\\nIn order to mitigate an over-reliance on surfacelevel information, ZOE attempts to match the concepts that best align given the local context as a constraint. This enables ZOE to recognize that TD Garden refers to a stadium or sports facility. When ONTOTYPE types TD Garden, we learn the most fine-grained label through the use of local context and pre-trained language models. The type indicators (Venues, Places, Sites, Locations) allow us to consolidate the high-level type of Location. However, by leveraging the surrounding contextual clues and knowledge learned during the training, NLI models enable ONTOTYPE to establish that TD Garden refers to a building or specifically a sports facility.\\nFinally, ONTOTYPE is able to derive the most fine-grained and accurate types by leveraging the inherent structure of the input type ontology. Consider the mention span of \\\"The NBA Finals\\\". The predicted candidate labels and head words (Finals, Playoffs, Semifinals, Rounds) indicate that the type is related to an Event. Thus, with the assistance of the pre-trained language models and head words, we select the most context-aware type of \\\\Other\\\\Event. Thus with these information sources and the ontology structure as guidance, ONTOTYPE can refine and select the final type of \\\\Other\\\\Event\\\\Finals.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.8 Evaluation Metrics\",\n \"publication_ref\": [\n \"b3\",\n \"b10\",\n \"b13\"\n ],\n \"table_ref\": [],\n \"text\": \"Following the prior FET studies ( (Dai et al., 2021), (Ling and Weld, 2012), (Ma et al., 2016)), we evaluate our methods and the baselines using three evaluation metrics: Strict Accuracy (Acc), Micro-F1 (Mi-F1), and Macro-F1 . Accuracy. Given a set of entity mentions M , we denote the set of ground truths and predicted types as t M and t M respectively. Given σ as an indicator function, strict accuracy is defined as\\nAcc = Σ m∈M σ(t m == t m ) M\\nMacro-F1. Macro-F1 is calculated using Macro-Precision (P ma ) and Macro-Recall (R ma ) where\\nP ma = 1 |M | Σ m∈M |t m ∩ t m | t m R ma = 1 |M | Σ m∈M |t m ∩ t m | t m\\nMicro-F1. Micro-F1 is calculated using Micro-Precision (P mi ) and Micro-Recall (R mi ) where\\nP mi = Σ m∈M |t m ∩ t m | Σ m∈M t m R mi = Σ m∈M |t m ∩ t m | Σ m∈M t m\\nMacro-F1 and Micro-F1 are calculated using the F1 score formula with their respective granular precision and recall scores. \"\n }\n]"},"abstract":{"kind":"string","value":"Fine-grained entity typing (FET), which assigns entities in text with context-sensitive, fine-grained semantic types, will play an important role in natural language understanding. A supervised FET method, which typically relies on human-annotated corpora for training, is costly and difficult to scale. Recent studies leverage pre-trained language models (PLMs) to generate rich and context-aware weak supervision for FET. However, a PLM may still generate a mixture of rough and fine-grained types, or tokens unsuitable for typing. In this study, we vision that an ontology provides a semantics-rich, hierarchical structure, which will help select the best results generated by multiple PLM models and head words. Specifically, we propose a novel zero-shot, ontologyguided FET method, ONTOTYPE, which follows a type ontological structure, from coarse to fine, ensembles multiple PLM prompting results to generate a set of type candidates, and refines its type resolution, under the local context with a natural language inference model. Our experiments on the Ontonotes, FIGER, and NYT datasets using their associated ontological structures demonstrate that our method outperforms the state-of-the-art zero-shot fine-grained entity typing methods. Our error analysis shows that refinement of the existing ontology structures will further improve fine-grained entity typing."},"title":{"kind":"string","value":"ONTOTYPE: Ontology-Guided Zero-Shot Fine-Grained Entity Typing with Weak Supervision from Pre-Trained Language Models"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 1: OntoNotes Type Ontology","figure_data":"","figure_id":"fig_0","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 2 :2Figure 2: Candidate Type Generation","figure_data":"","figure_id":"fig_1","figure_label":"2","figure_type":"figure"},{"figure_caption":"Performance of Hearst Patterns with simple type mapping on Ontonotes Dataset.","figure_data":"Hearst PatternPrec Rec F1[MASK] such as53.372.4 61.4such [MASK] as47.968.7 56.5and some other [MASK] 48.866.6 56.4and the other [MASK]47.668.3 56.1","figure_id":"tab_0","figure_label":"1","figure_type":"table"},{"figure_caption":"Dataset Statistics","figure_data":"","figure_id":"tab_2","figure_label":"2","figure_type":"table"},{"figure_caption":"For the","figure_data":"SettingsModelNYT Acc Mi-F1 Ma-F1 Acc Mi-F1 Ma-F1 Acc FIGEROntonotes Mi-F1 Ma-F1AFET (Ren et al., 2016)---55.3 66.469.355.164.771.1SupervisedUFET (Choi et al., 2018)------59.571.876.8BERT-MLMET (Dai et al., 2021)------67.44 80.3585.44LITE (Li et al., 2022)---66.2 74.780.168.281.486.6ZOE (Zhou et al., 2018)62.1 73.776.958.8 71.374.850.760.866.9Zero-ShotOTyper (Yuan and Downey, 2018) 46.4 65.767.347.2 67.269.131.836.039.1DZET (Obeidat et al., 2019)27.3 53.151.628.5 56.055.123.128.127.6MZET (Zhang et al., 2020b)30.7 58.256.731.9 57.955.533.743.742.3ONTOTYPE + Original Ontology69.6 78.482.849.1 67.475.165.773.481.5","figure_id":"tab_3","figure_label":"2","figure_type":"table"},{"figure_caption":"Examples showing how ONTOTYPE and other Zero-Shot FET methods perform on recent news articles with a modified FIGER type ontology. Entity mentions are bolded and predicted labels are in color.","figure_data":"","figure_id":"tab_6","figure_label":"4","figure_type":"table"},{"figure_caption":"Results of the efficiency & scalability study.","figure_data":"","figure_id":"tab_8","figure_label":"5","figure_type":"table"},{"figure_caption":"Examples showing how ONTOTYPE performs on recent news articles with a modified FIGER type ontology. Entity mentions are bolded in the given sentences. The predicted labels are in color.","figure_data":"","figure_id":"tab_10","figure_label":"8","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 1: OntoNotes Type Ontology\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 2 :2Figure 2: Candidate Type Generation\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Performance of Hearst Patterns with simple type mapping on Ontonotes Dataset.\",\n \"figure_data\": \"Hearst PatternPrec Rec F1[MASK] such as53.372.4 61.4such [MASK] as47.968.7 56.5and some other [MASK] 48.866.6 56.4and the other [MASK]47.668.3 56.1\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Dataset Statistics\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"For the\",\n \"figure_data\": \"SettingsModelNYT Acc Mi-F1 Ma-F1 Acc Mi-F1 Ma-F1 Acc FIGEROntonotes Mi-F1 Ma-F1AFET (Ren et al., 2016)---55.3 66.469.355.164.771.1SupervisedUFET (Choi et al., 2018)------59.571.876.8BERT-MLMET (Dai et al., 2021)------67.44 80.3585.44LITE (Li et al., 2022)---66.2 74.780.168.281.486.6ZOE (Zhou et al., 2018)62.1 73.776.958.8 71.374.850.760.866.9Zero-ShotOTyper (Yuan and Downey, 2018) 46.4 65.767.347.2 67.269.131.836.039.1DZET (Obeidat et al., 2019)27.3 53.151.628.5 56.055.123.128.127.6MZET (Zhang et al., 2020b)30.7 58.256.731.9 57.955.533.743.742.3ONTOTYPE + Original Ontology69.6 78.482.849.1 67.475.165.773.481.5\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Examples showing how ONTOTYPE and other Zero-Shot FET methods perform on recent news articles with a modified FIGER type ontology. Entity mentions are bolded and predicted labels are in color.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_6\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Results of the efficiency & scalability study.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_8\",\n \"figure_label\": \"5\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Examples showing how ONTOTYPE performs on recent news articles with a modified FIGER type ontology. Entity mentions are bolded in the given sentences. The predicted labels are in color.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_10\",\n \"figure_label\": \"8\",\n \"figure_type\": \"table\"\n }\n]"},"authors":{"kind":"string","value":"Tanay Komarlu; Minhao Jiang; Xuan Wang; Jiawei Han"},"references":{"kind":"list like","value":[{"authors":"B Tom; Benjamin Brown; Nick Mann; Melanie Ryder; Jared Subbiah; Prafulla Kaplan; Arvind Dhariwal; Pranav Neelakantan; Girish Shyam; Amanda Sastry; Sandhini Askell; Ariel Agarwal; Gretchen Herbert-Voss; Tom Krueger; Rewon Henighan; Aditya Child; Daniel M Ramesh; Jeffrey Ziegler; Clemens Wu; Christopher Winter; Mark Hesse; Eric Chen; Mateusz Sigler; Scott Litwin; Benjamin Gray; Jack Chess; Christopher Clark; Sam Berner; Alec Mc-Candlish; Ilya Radford; Dario Sutskever; Amodei","journal":"","ref_id":"b0","title":"Language Models are Few-Shot Learners","year":"2020"},{"authors":"Danqi Chen; Christopher Manning","journal":"","ref_id":"b1","title":"A Fast and Accurate Dependency Parser using Neural Networks","year":"2014"},{"authors":"Eunsol Choi; Omer Levy; Yejin Choi; Luke Zettlemoyer","journal":"Association for Computational Linguistics","ref_id":"b2","title":"Ultra-Fine Entity Typing","year":"2018"},{"authors":"Hongliang Dai; Yangqiu Song; Haixun Wang","journal":"","ref_id":"b3","title":"Ultra-Fine Entity Typing with Weak Supervision from a Masked Language Model","year":"2021"},{"authors":"Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova","journal":"Association for Computational Linguistics","ref_id":"b4","title":"BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding","year":"2019"},{"authors":"Daniel Gillick; Nevena Lazic; Kuzman Ganchev; Jesse Kirchner; David Huynh","journal":"","ref_id":"b5","title":"Context-Dependent Fine-Grained Entity Type Tagging","year":"2014"},{"authors":"Nitish Gupta; Sameer Singh; Dan Roth","journal":"","ref_id":"b6","title":"Entity Linking via Joint Encoding of Types, Descriptions, and Context","year":"2017"},{"authors":"Marti A Hearst","journal":"","ref_id":"b7","title":"Automatic Acquisition of Hyponyms from Large Text Corpora","year":"1992"},{"authors":"Mitchell Koch; John Gilmer; Stephen Soderland; Daniel S Weld","journal":"","ref_id":"b8","title":"Type-aware distantly supervised relation extraction with linked arguments","year":"2014"},{"authors":"Bangzheng Li; Wenpeng Yin; Muhao Chen","journal":"Transactions of the Association for Computational Linguistics","ref_id":"b9","title":"Ultra-fine Entity Typing with Indirect Supervision from Natural Language Inference","year":"2022"},{"authors":"Xiao Ling; Daniel S Weld","journal":"AAAI Press","ref_id":"b10","title":"Fine-Grained Entity Recognition","year":"2012"},{"authors":"Pengfei Liu; Weizhe Yuan; Jinlan Fu; Zhengbao Jiang; Hiroaki Hayashi; Graham Neubig","journal":"","ref_id":"b11","title":"Pretrain, Prompt, and Predict: A Systematic Survey of Prompting Methods in Natural Language Processing","year":"2021"},{"authors":"Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov","journal":"","ref_id":"b12","title":"RoBERTa: A Robustly Optimized BERT Pretraining Approach","year":"2019"},{"authors":"Yukun Ma; Erik Cambria; Sa Gao","journal":"","ref_id":"b13","title":"Label Embedding for Zero-shot Fine-grained Named Entity Typing","year":"2016"},{"authors":"Tomas Mikolov; Ilya Sutskever; Kai Chen; Greg S Corrado; Jeff Dean","journal":"Curran Associates, Inc","ref_id":"b14","title":"Distributed Representations of Words and Phrases and their Compositionality","year":"2013"},{"authors":"Nikita Nangia; Clara Vania; Rasika Bhalerao; Samuel R Bowman","journal":"","ref_id":"b15","title":"CrowS-Pairs: A Challenge Dataset for Measuring Social Biases in Masked Language Models","year":"1953"},{"authors":"Rasha Obeidat; Xiaoli Fern; Hamed Shahbazi; Prasad Tadepalli","journal":"Association for Computational Linguistics","ref_id":"b16","title":"Description-Based Zeroshot Fine-Grained Entity Typing","year":"2019"},{"authors":"Yasumasa Onoe; Greg Durrett","journal":"OpenAI","ref_id":"b17","title":"Interpretable Entity Representations through Large-Scale Typing","year":"2020"},{"authors":"Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu","journal":"","ref_id":"b18","title":"Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer","year":"2020"},{"authors":"Xiang Ren; Wenqi He; Meng Qu; Lifu Huang; Ji Heng; Jiawei Han","journal":"Association for Computational Linguistics","ref_id":"b19","title":"AFET: Automatic Fine-Grained Entity Typing by Hierarchical Partial-Label Embedding","year":"2016"},{"authors":"Julian Seitner; Christian Bizer; Kai Eckert; Stefano Faralli; Robert Meusel; Heiko Paulheim; Simone Paolo; Ponzetto ","journal":"","ref_id":"b20","title":"A Large DataBase of Hypernymy Relations Extracted from the Web","year":"2016"},{"authors":"Xuan Wang; Vivian Hu; Xiangchen Song; Shweta Garg; Jinfeng Xiao; Jiawei Han","journal":"","ref_id":"b21","title":"CHEMNER: Fine-Grained Chemistry Named Entity Recognition with Ontology-Guided Distant Supervision","year":"2021"},{"authors":"Congying Xia; Chenwei Zhang; Xiaohui Yan; Yi Chang; Philip Yu","journal":"","ref_id":"b22","title":"Zero-shot User Intent Detection via Capsule Neural Networks","year":"2018"},{"authors":"Limin Yao; Sebastian Riedel; Andrew Mccallum","journal":"Association for Computing Machinery","ref_id":"b23","title":"Universal Schema for Entity Type Prediction (AKBC '13)","year":"2013"},{"authors":"Mohamed Amir Yosef; Sandro Bauer; Johannes Hoffart; Marc Spaniol; Gerhard Weikum","journal":"","ref_id":"b24","title":"HYENA: Hierarchical Type Classification for Entity Names","year":"2012"},{"authors":"Zheng Yuan; Doug Downey","journal":"AAAI Press","ref_id":"b25","title":"OTyper: A Neural Architecture for Open Named Entity Typing","year":"2018"},{"authors":"Tao Zhang; Congying Xia; Chun-Ta Lu; Philip Yu","journal":"","ref_id":"b26","title":"MZET: Memory Augmented Zero-Shot Fine-grained Named Entity Typing","year":"2020"},{"authors":"Yunyi Zhang; Jiaming Shen; Jingbo Shang; Jiawei Han","journal":"","ref_id":"b27","title":"Empower Entity Set Expansion via Language Model Probing","year":"2020"},{"authors":"Ben Zhou; Daniel Khashabi; Chen-Tse Tsai; Dan Roth","journal":"","ref_id":"b28","title":"Zero-Shot Open Entity Typing as Type-Compatible Grounding","year":"2018"}],"string":"[\n {\n \"authors\": \"B Tom; Benjamin Brown; Nick Mann; Melanie Ryder; Jared Subbiah; Prafulla Kaplan; Arvind Dhariwal; Pranav Neelakantan; Girish Shyam; Amanda Sastry; Sandhini Askell; Ariel Agarwal; Gretchen Herbert-Voss; Tom Krueger; Rewon Henighan; Aditya Child; Daniel M Ramesh; Jeffrey Ziegler; Clemens Wu; Christopher Winter; Mark Hesse; Eric Chen; Mateusz Sigler; Scott Litwin; Benjamin Gray; Jack Chess; Christopher Clark; Sam Berner; Alec Mc-Candlish; Ilya Radford; Dario Sutskever; Amodei\",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"Language Models are Few-Shot Learners\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Danqi Chen; Christopher Manning\",\n \"journal\": \"\",\n \"ref_id\": \"b1\",\n \"title\": \"A Fast and Accurate Dependency Parser using Neural Networks\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"Eunsol Choi; Omer Levy; Yejin Choi; Luke Zettlemoyer\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b2\",\n \"title\": \"Ultra-Fine Entity Typing\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Hongliang Dai; Yangqiu Song; Haixun Wang\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"Ultra-Fine Entity Typing with Weak Supervision from a Masked Language Model\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b4\",\n \"title\": \"BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Daniel Gillick; Nevena Lazic; Kuzman Ganchev; Jesse Kirchner; David Huynh\",\n \"journal\": \"\",\n \"ref_id\": \"b5\",\n \"title\": \"Context-Dependent Fine-Grained Entity Type Tagging\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"Nitish Gupta; Sameer Singh; Dan Roth\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"Entity Linking via Joint Encoding of Types, Descriptions, and Context\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Marti A Hearst\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"Automatic Acquisition of Hyponyms from Large Text Corpora\",\n \"year\": \"1992\"\n },\n {\n \"authors\": \"Mitchell Koch; John Gilmer; Stephen Soderland; Daniel S Weld\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"Type-aware distantly supervised relation extraction with linked arguments\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"Bangzheng Li; Wenpeng Yin; Muhao Chen\",\n \"journal\": \"Transactions of the Association for Computational Linguistics\",\n \"ref_id\": \"b9\",\n \"title\": \"Ultra-fine Entity Typing with Indirect Supervision from Natural Language Inference\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Xiao Ling; Daniel S Weld\",\n \"journal\": \"AAAI Press\",\n \"ref_id\": \"b10\",\n \"title\": \"Fine-Grained Entity Recognition\",\n \"year\": \"2012\"\n },\n {\n \"authors\": \"Pengfei Liu; Weizhe Yuan; Jinlan Fu; Zhengbao Jiang; Hiroaki Hayashi; Graham Neubig\",\n \"journal\": \"\",\n \"ref_id\": \"b11\",\n \"title\": \"Pretrain, Prompt, and Predict: A Systematic Survey of Prompting Methods in Natural Language Processing\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"RoBERTa: A Robustly Optimized BERT Pretraining Approach\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Yukun Ma; Erik Cambria; Sa Gao\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"Label Embedding for Zero-shot Fine-grained Named Entity Typing\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Tomas Mikolov; Ilya Sutskever; Kai Chen; Greg S Corrado; Jeff Dean\",\n \"journal\": \"Curran Associates, Inc\",\n \"ref_id\": \"b14\",\n \"title\": \"Distributed Representations of Words and Phrases and their Compositionality\",\n \"year\": \"2013\"\n },\n {\n \"authors\": \"Nikita Nangia; Clara Vania; Rasika Bhalerao; Samuel R Bowman\",\n \"journal\": \"\",\n \"ref_id\": \"b15\",\n \"title\": \"CrowS-Pairs: A Challenge Dataset for Measuring Social Biases in Masked Language Models\",\n \"year\": \"1953\"\n },\n {\n \"authors\": \"Rasha Obeidat; Xiaoli Fern; Hamed Shahbazi; Prasad Tadepalli\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b16\",\n \"title\": \"Description-Based Zeroshot Fine-Grained Entity Typing\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Yasumasa Onoe; Greg Durrett\",\n \"journal\": \"OpenAI\",\n \"ref_id\": \"b17\",\n \"title\": \"Interpretable Entity Representations through Large-Scale Typing\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b18\",\n \"title\": \"Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Xiang Ren; Wenqi He; Meng Qu; Lifu Huang; Ji Heng; Jiawei Han\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b19\",\n \"title\": \"AFET: Automatic Fine-Grained Entity Typing by Hierarchical Partial-Label Embedding\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Julian Seitner; Christian Bizer; Kai Eckert; Stefano Faralli; Robert Meusel; Heiko Paulheim; Simone Paolo; Ponzetto \",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"A Large DataBase of Hypernymy Relations Extracted from the Web\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Xuan Wang; Vivian Hu; Xiangchen Song; Shweta Garg; Jinfeng Xiao; Jiawei Han\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"CHEMNER: Fine-Grained Chemistry Named Entity Recognition with Ontology-Guided Distant Supervision\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Congying Xia; Chenwei Zhang; Xiaohui Yan; Yi Chang; Philip Yu\",\n \"journal\": \"\",\n \"ref_id\": \"b22\",\n \"title\": \"Zero-shot User Intent Detection via Capsule Neural Networks\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Limin Yao; Sebastian Riedel; Andrew Mccallum\",\n \"journal\": \"Association for Computing Machinery\",\n \"ref_id\": \"b23\",\n \"title\": \"Universal Schema for Entity Type Prediction (AKBC '13)\",\n \"year\": \"2013\"\n },\n {\n \"authors\": \"Mohamed Amir Yosef; Sandro Bauer; Johannes Hoffart; Marc Spaniol; Gerhard Weikum\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"HYENA: Hierarchical Type Classification for Entity Names\",\n \"year\": \"2012\"\n },\n {\n \"authors\": \"Zheng Yuan; Doug Downey\",\n \"journal\": \"AAAI Press\",\n \"ref_id\": \"b25\",\n \"title\": \"OTyper: A Neural Architecture for Open Named Entity Typing\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Tao Zhang; Congying Xia; Chun-Ta Lu; Philip Yu\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"MZET: Memory Augmented Zero-Shot Fine-grained Named Entity Typing\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Yunyi Zhang; Jiaming Shen; Jingbo Shang; Jiawei Han\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"Empower Entity Set Expansion via Language Model Probing\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Ben Zhou; Daniel Khashabi; Chen-Tse Tsai; Dan Roth\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"Zero-Shot Open Entity Typing as Type-Compatible Grounding\",\n \"year\": \"2018\"\n }\n]"},"formulas":{"kind":"list like","value":[{"formula_coordinates":[4,317.61,656.33,206.8,11.22],"formula_id":"formula_0","formula_text":"y = count (m) {H 1 (x), H 2 (x), ..., H n (x)} (1)"},{"formula_coordinates":[5,343.64,327.85,176.53,25.14],"formula_id":"formula_1","formula_text":"N = s i ∈S emb(s i ) + emb(c) |S| + 1 . (2"},{"formula_coordinates":[5,520.17,336.28,4.24,9.46],"formula_id":"formula_2","formula_text":")"},{"formula_coordinates":[6,113.29,730.83,175.85,10.77],"formula_id":"formula_3","formula_text":"rank(type) = σ entail + σ cand(4)"},{"formula_coordinates":[6,333.75,455.3,163.06,67.23],"formula_id":"formula_4","formula_text":"⋮ Figure 4: Fine-Grained Type Refinement"},{"formula_coordinates":[15,116.47,96.27,125.86,26.44],"formula_id":"formula_5","formula_text":"Acc = Σ m∈M σ(t m == t m ) M"},{"formula_coordinates":[15,114.77,167.9,129.27,66.64],"formula_id":"formula_6","formula_text":"P ma = 1 |M | Σ m∈M |t m ∩ t m | t m R ma = 1 |M | Σ m∈M |t m ∩ t m | t m"},{"formula_coordinates":[15,125.69,276.92,107.43,66.93],"formula_id":"formula_7","formula_text":"P mi = Σ m∈M |t m ∩ t m | Σ m∈M t m R mi = Σ m∈M |t m ∩ t m | Σ m∈M t m"}],"string":"[\n {\n \"formula_coordinates\": [\n 4,\n 317.61,\n 656.33,\n 206.8,\n 11.22\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"y = count (m) {H 1 (x), H 2 (x), ..., H n (x)} (1)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 343.64,\n 327.85,\n 176.53,\n 25.14\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"N = s i ∈S emb(s i ) + emb(c) |S| + 1 . (2\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 520.17,\n 336.28,\n 4.24,\n 9.46\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \")\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 113.29,\n 730.83,\n 175.85,\n 10.77\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"rank(type) = σ entail + σ cand(4)\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 333.75,\n 455.3,\n 163.06,\n 67.23\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"⋮ Figure 4: Fine-Grained Type Refinement\"\n },\n {\n \"formula_coordinates\": [\n 15,\n 116.47,\n 96.27,\n 125.86,\n 26.44\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"Acc = Σ m∈M σ(t m == t m ) M\"\n },\n {\n \"formula_coordinates\": [\n 15,\n 114.77,\n 167.9,\n 129.27,\n 66.64\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"P ma = 1 |M | Σ m∈M |t m ∩ t m | t m R ma = 1 |M | Σ m∈M |t m ∩ t m | t m\"\n },\n {\n \"formula_coordinates\": [\n 15,\n 125.69,\n 276.92,\n 107.43,\n 66.93\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"P mi = Σ m∈M |t m ∩ t m | Σ m∈M t m R mi = Σ m∈M |t m ∩ t m | Σ m∈M t m\"\n }\n]"}}},{"rowIdx":393,"cells":{"doi":{"kind":"string","value":"10.1145/nnnnnnn.nnnnnnn"},"pub_date":{"kind":"string","value":"2023-05-21"},"sections":{"kind":"list like","value":[{"figure_ref":[],"heading":"INTRODUCTION","publication_ref":["b22","b16","b28","b12","b16","b22","b28"],"table_ref":[],"text":"Nowadays, the ever-expanding new breeds of content, e.g., pictures, live streams, and short videos, drive the recommender system to drift from the traditional homogeneous form into an integrated form. Integrated recommender systems (IRSs) aim to simultaneously recommend heterogeneous items from multiple sources/channels in a row. This integrated form greatly expands users' choices on different types of content thereby satisfying users' diversified preferences on both item-level and channel-level. Therefore, IRS has nowadays been widely deployed in various online platforms such as the homepage feeds in Kuaishou [23], XiaohongShu [17], Taobao [29], and AliExpress [13]. In these products, users continuously slide down to browse and interact with heterogeneous items in a sequential manner, as shown in Figure 1.\nThough attractive, integrated feed recommendation faces more challenges than conventional recommendation with homogeneous items. First, real-world applications usually impose upper or lower exposure guarantees on different channels, such as lower constraints for sponsored/new content (e.g., ads and cold-start items) or upper constraints for individual channels to ensure diversity. These constraints lead to greater difficulty in the joint ranking of heterogeneous items. Second, heterogeneous items from multiple channels usually have different features and ranking strategies. Hence, it is difficult to directly compare items from different channels for joint ranking. Third, users' interests on different channels have a great impact on their behaviors, such that traditional user-item prediction models need to evolve into user-channel-item prediction models by considering both intra-channel and inter-channel information and their correlation with user interests. Finally, in feed products, users tend to review a large number of items in a row such that the previously viewed items have a great impact on the users' behavior towards the next item [17,23,29]. Therefore, it is of vital importance to consider the influence from page context when determining the item order in the return list.\nAlthough integrated feed recommendation has been widely deployed in practice, there are still few works focusing on the above challenges systematically. In this paper, we propose a general framework named Multi-channel Integrated Recommendation with Exposure Constraints (MIREC) to deal with the multi-channel integrated recommendation task under resource constraints in feed products. MIREC consists of two layers: an allocation-layer which optimizes the exposure of different channels from a global view over all user requests, and a ranking-layer which determines the optimal item layout of a given user request from a local view. These two layers operate in an iterative manner to make online decisions along with the arrival of user requests. The main contributions are as follows.\n• This work formulates the integrated recommendation task with exposure constraints as a binary online linear programming problem and proposes a two-layer framework named MIREC to obtain the optimal solution. We also describe a practical system architecture for its implementation in industrial platforms. • This work proposes an efficient multi-channel allocation algorithm to obtain the optimal exposure assignment of different channels over a fixed time horizon. The proposed algorithm is able to reach an optimal solution with linear complexity w.r.t. the number of constraints. We also prove that this algorithm admits a regret bound of O ( √ 𝑇 ) towards the global optimal point under certain assumptions.\n• This work proposes a series of collaborative models to determine the optimal layout of heterogeneous items on a page, with joint modeling of user interests, cross-channel correlation, and page context. This aligns more with the browsing nature of feed products than existing models. • This work conducts extensive experiments on both offline datasets and online A/B tests to verify the superiority of our proposed method.\nMIREC has been implemented on the homepage of Taobao to serve the main traffic. It brings 3% lift on user clicks, 1.56% lift on purchase, and 1.42% lift on stay time. It now serves hundreds of millions of users towards billions of items every day."},{"figure_ref":[],"heading":"RELATED WORK","publication_ref":["b9","b38","b27","b40","b3","b0","b26","b15","b19","b21","b31","b10","b14","b2","b8","b24","b33","b34","b1","b17","b39","b13","b5","b23","b12","b32","b36","b20"],"table_ref":[],"text":"Re-ranking Methods. The main objective of re-ranking methods is to consider the mutual influence among a list of items in order to refine the prediction results produced by point-wise ranking models. Three prevalent models are commonly adopted in the existing literature: RNN-based methods, attention-based methods, and evaluator-generator-based methods. The first two methods feed an initial ranking list produced by point-wise models (e.g., Wide&Deep [10], DIN [39] and SIM [28]) into RNN-based (e.g., MiDNN [41], Seq2Slate [4] and DLCM [1]) or attention-based structure (e.g., PRM [27], PFRN [16], PEAR [20], and Raiss [22]) sequentially and output the encoded vector at each subsequent layer to model the mutual influences among items. The evaluator-generatorbased methods (e.g., SEG [32] and GRN [11]), use a generator to generate feasible permutations and use an evaluator to evaluate their list-wise utility to determine the optimal permutation. However, most re-ranking methods mainly focus on capturing the mutual influence among homogeneous items provided by one channel, instead of heterogeneous items provided by multiple channels. Moreover, they only optimize the item order at a single time slot, instead of considering a cumulative utility over a broad time horizon under resource constraints.\nOnline Allocation Methods. The online allocation problem with resource constraints has been mostly studied in online convex optimization [15]. The primal-dual methods [3,9,25,34,35] avoid taking expensive projection iterations by penalizing the violation of constraints through duality. The BwK methods [2,18] determines an optimal action from a finite set of possible actions and then optimize the policy of decision-making according to the observed rewards and costs over a fixed period of time. Several recent works studied the practical performance of online allocation in advertising recommendations. For example, PDOA [40] adopts the primal-dual framework by optimizing the dual prices with online gradient descent to eliminate the online max-min problem's regret. However, it assumes that the utility and cost values can be optimally estimated and only verify the performance through offline simulations. MSBCB [14] and HCA2E [6] proposed a two-level optimization framework based on BwK methods, where the high-level determines whether to present ads on the given request while the low-level searches the optimal position to insert ads. Most related works on online convex optimization focus on theoretical analysis (e.g., regret bound) instead of real-world applications. Other related works on advertising mainly consider binary content, i.e., ads or non-ads, instead of heterogeneous content. Directly extending them to deal with multi-channel recommendations in IRSs may lead to sub-optimal results.\nIntegrated Recommendation. The integrated recommendation is a newly emerged but rapidly developing domain driven by practical problems [24]. Integrated recommendation methods need to consider both intra-channel and inter-channel features within the heterogeneous content and provide recommendation results continuously along with user arrivals. Recently, DHANR [13] proposed a hierarchical self-attention structure to consider the cross-channel interactions. HRL-Rec [33] decomposed the integrated re-ranking problem into two subtasks: source selection and item ranking, and use hierarchical reinforcement learning to solve the problem. DEAR [37] proposed to interpolate ads and organic items by deep Q-networks. Cross-DQN [21] also adopt a reinforcement learning solution with a cross-channel attention unit. However, many integrated methods only focus on ranking at a single time slot instead of over a continuous time horizon. The joint consideration of both integrated ranking and online allocation of limited resources still remains to be explored."},{"figure_ref":[],"heading":"PROBLEM FORMULATION","publication_ref":["b11","b5","b25"],"table_ref":[],"text":"This section formulates the integrated recommendation task with exposure constraints as a binary online linear programming problem. Specifically, we consider a generic IRS setting where user requests arrive sequentially during a finite time horizon. For each request, the IRS needs to rank a list of heterogeneous items provided by multiple channels. The aim is to maximize the overall utilities (e.g., the sum of clicks and pays) of all channels over the entire time horizon, subject to multiple resource constraints.\nFormally, the request of user 𝑢 triggered at time 𝑡 is described as 𝑒 𝑡 = (𝑢, 𝑓 , 𝑔, X 𝑡 ), where 𝑓 ∈ R + is a non-negative utility function, 𝑔 ∈ R + is a non-negative resource consumption function, and X 𝑡 ⊂ R 𝑑 + is a compact set denoting all possible item layouts for decisionmaking. For each request 𝑒 𝑡 , the IRS needs to choose a number of 𝑁 heterogeneous items from a candidate set 𝐼 𝑡 and place them into 𝑁 slots to form a complete page and return it to the user. This action 𝒙 𝑡 ∈ X 𝑡 can be represented as a decision matrix 𝒙 𝑡 ∈ [0, 1] 𝑁 ×|𝐼 𝑡 | , where each entry 𝑥 𝑡,𝑛,𝑖 is indexed by a slot 𝑛 and a card index 𝑖. Once the user finished viewing the current page, a new user request will be triggered to ask the platform to return to the next page. Therefore, this decision-making process will be performed repeatedly. Moreover, in real-world applications, the item layouts need to satisfy the following constraints:\nX 𝑡 = 𝑖 ∈𝐼 𝑡 𝑥 𝑡,𝑛,𝑖 = 1, ∀𝑡 ∈ T , ∀𝑛 ∈ N 𝑛 𝑥 𝑡,𝑛,𝑖 ≤ 1, ∀𝑡 ∈ T , ∀𝑖 ∈ 𝐼,(1)\nwhere the upper constraint restricts that each slot must be assigned to one item and the lower constraint restricts that each item can be assigned to at most one slot.\nAfter executing an action 𝒙 𝑡 at request 𝑒 𝑡 , the IRS consumes a resource cost 𝑔(𝒙 𝑡 ) and obtains an utility 𝑓 (𝒙 𝑡 ). In IRS, the utility function 𝑓 (𝒙 𝑡 ) is defined according to the concerned metrics. For example, it can be defined as a combination of stay time, adds to cart, and favorites to encourage user engagement, or defined as a combination of clicks and purchases to encourage user conversion. On the other hand, the consumption function 𝑔(𝒙 𝑡 ) is defined based on the concerned resource constraints. For example, the platform may need to allocate a certain amount of exposure to new channels in order to support the growth of new content [12]. Meanwhile, a too large proportion of exposures on one specific channel will damage the recommendation diversity thereby harming user experience [6,26]. In this case, the IRS needs to guarantee both a lower exposure limit and an upper exposure limit for the heterogeneous items from different channels.\nIn this paper, we focus on the exposure constraints in practical systems which lead to the following optimization problem\nP 0 : OPT(S) = max 𝒙 𝑡 ∈X 𝑡 ∑︁ 𝑇 𝑡 =1 𝑓 (𝒙 𝑡 )(2)\ns.t. 𝐶 1 : ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≤ 𝐺 max 𝑚,𝑡ℎ 𝑁 (S), ∀𝑚 ∈ M,(3)\n𝐶 2 : ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≥ 𝐺 min 𝑚,𝑡ℎ 𝑁 (S), ∀𝑚 ∈ M,(4)\nwhere 𝑁 (S) denotes the total available exposures to allocate over the entire time horizon, 𝐺 max 𝑚,𝑡ℎ and 𝐺 min 𝑚,𝑡ℎ denote the proportion of upper exposure limits and lower exposure limits for each channel 𝑚 ∈ M, respectively, and 𝑔 𝑚 (𝒙 𝑡 ) denotes the consumed exposures of cards from channel 𝑚 after executing 𝒙 𝑡 at request 𝑒 𝑡 . Although this paper mainly focuses on the exposure guarantee, the above formulation is generally applicable to other problems with different resource constraints, e.g., the number of coupons to allocate."},{"figure_ref":[],"heading":"METHODOLOGY 4.1 Framework Overview","publication_ref":[],"table_ref":[],"text":"Directly solving problem P 0 is challenging. On one hand, the estimation accuracy of utility 𝑓 (𝒙 𝑡 ) and consumption 𝑔(𝒙 𝑡 ) suffer influence from multi-factors, including the user's personal interest, the page context, and the cross-correlation between different channels. On the other hand, the determination of each 𝒙 𝑡 needs to consider the cumulative exposures over the entire time horizon due to the exposure guarantees. Therefore, the optimization of exposure allocation must be performed from a global view over the entire timeline instead of a single time slot.\nTo this end, we propose the MIREC framework which solves P 0 through online primal-dual iterations. Specifically, MIREC consists of two layers, i.e., the allocation-layer and the ranking-layer, which correspond to the dual and primal problem of P 0 , respectively. The allocation-layer optimizes dual variables to control the cumulative exposure of different channels on all user requests from a global view to guarantee the exposure limits. Meanwhile, the rankinglayer optimizes the item layout under fixed dual variables given by the allocation-layer, with the aim to maximize the instant utility on a single user request from a local view. These two layers operate in an iterative manner along with the arrival of online user requests to determine the optimal item layout at user requests continuously. The general workflow of MIREC is shown in Figure . 2.\nFor the allocation-layer, we propose a simple but efficient Mirrordescent based Multi-channel Exposure Allocation (ME2A) algorithm to adaptively balance the utility gain and the exposure cost of presenting heterogeneous items from different channels. The proposed M2EA algorithm has a closed form solution that can be computed in linear time and admits a regret bound of O ( √ 𝑇 ) towards the global optimal point under certain assumptions.\nFor the ranking-layer, we propose a personalized cross-channel ranking (PCR) model and a context-aware reranking (CAR) model to jointly determine the optimal item layout on a given user request, with fixed dual parameters given by the allocation-layer. In particular, PCR gives point-wise estimation of the quality of candidate items by joint modeling the influence from user interests, intra-channel information, and inter-channel correlations. Afterward, CAR refines the point-wise estimation generated by PCR into context-aware estimation by making use of both the context information and the high-level knowledge extracted from PCR."},{"figure_ref":[],"heading":"Global: Online Exposure Allocation","publication_ref":[],"table_ref":[],"text":"In this section, we introduce the primal-dual formulation of P 0 and propose the ME2A algorithm to obtain the optimal solution for online systems. The complete algorithm is presented in Algorithm 1. \n+ ∑︁ 𝑚 𝜆 𝑚 ∑︁ 𝑡 𝑔 𝑚 (𝒙 𝑡 ) -𝐺 min 𝑚 .(5)\nHere 𝝁 ≥ 0 and 𝝀 ≥ 0 are the introduced dual parameters, 𝐺 max 𝑚 and 𝐺 min 𝑚 are short for 𝐺 max 𝑚,𝑡ℎ 𝑁 (S) and 𝐺 min 𝑚,𝑡ℎ 𝑁 (S), respectively. Note that the parameters 𝝁 and 𝝀 are related to the violation of exposure consumption over the upper bound limit and the lower bound limit, respectively, which are mutually exclusive. In particular, if one of them is positive, the other one must be zero; otherwise, both of them are zero. Hence, it is viable to only introduce one real number dual variable 𝝁 ∈ R 1×𝑀 to replace 𝝁 and 𝝀 in the dual problem, which simplifies (5) into\nmin 𝝁 𝐷 (𝝁) = ∑︁ 𝑡 𝑓 (𝒙 𝑡 ) - ∑︁ 𝑚 [𝜇 𝑚 ] + ∑︁ 𝑡 𝑔 𝑚 (𝒙 𝑡 ) -𝐺 max 𝑚 (6a) + ∑︁ 𝑚 [-𝜇 𝑚 ] + ∑︁ 𝑡 𝑔 𝑚 (𝒙 𝑡 ) -𝐺 min 𝑚 (6b) = ∑︁ 𝑡 𝑓 (𝒙 𝑡 ) - ∑︁ 𝑚 𝜇 𝑚 𝑔 𝑚 (𝒙 𝑡 ) + ∑︁ 𝑚 [𝜇 𝑚 ] + 𝐺 max 𝑚 -[-𝜇 𝑚 ] + 𝐺 min 𝑚 ,(6c)\nwhere\n[𝜇 𝑚 ] + = max{𝜇 𝑚 , 0}."},{"figure_ref":[],"heading":"Dual","publication_ref":["b4","b9"],"table_ref":[],"text":"Optimization for Exposure Constraints. The dual problem in (6) can be solved optimally via primal-dual updates. Specifically, given a user request 𝑒 𝑡 = (𝑢, 𝑓 , 𝑏, X 𝑡 ), we assume that the utility 𝑓 (𝒙 𝑡 ) and cost 𝝁 𝑇 𝑡 𝑔(𝒙 𝑡 ) under different item layout 𝒙 𝑡 can be properly estimated by the models at the ranking layer. As such, the optimal item layout 𝒙 𝑡 under a fixed dual variable 𝝁 𝑡 can be obtained by solving the following primal problem:\nP 1 : x𝑡 = arg max 𝒙 𝑡 ∈X 𝑓 (𝒙 𝑡 ) -𝝁 𝑇 𝑡 𝑔(𝒙 𝑡 ) .(7)\nThis is the focus of the ranking-layer to be discussed layer. After the optimal item layout 𝒙 𝑡 at user request 𝑒 𝑡 is properly determined, the next step is to update the dual variable 𝝁 𝑡 to adjust the exposure of different channels in future user requests. Specifically, the remained exposure resource of different channels after the presentation of 𝒙 𝑡 at 𝑒 𝑡 is updated by\n𝐺 max 𝑚,𝑡 +1 = 𝐺 max 𝑚,𝑡 -𝑔 𝑚 (𝒙 𝑡 ), ∀𝑚 ∈ M.(8)\nThe sub-gradient of the dual function in ( 6) can be obtained via Danskin's theorem [5] by 9) where 1(𝑥 ∈ 𝐴) is an indicator function which equals to one if 𝑥 ∈ 𝐴 otherwise zero. As such, the dual variable 𝝁 𝑡 can be updated based on the mirror-descent method as\n∇𝜇 𝑚,𝑡 = -𝑔 𝑚 (𝒙 𝑡 ) +𝐺 max 𝑚,𝑡 +1 • 1(𝜇 𝑚,𝑡 ≥ 0) +𝐺 min 𝑚 • 1(𝜇 𝑚,𝑡 ≤ 0), (\n𝜇 𝑚,𝑡 +1 = arg min 𝜇 𝑚 ∈R 𝜇 𝑚 ∇𝜇 𝑚,𝑡 + 1 𝜂 𝑉 ℎ (𝜇 𝑚 , 𝜇 𝑚,𝑡 ),(10)\nwhere 𝑉 ℎ (𝑥, 𝑦) = ℎ(𝑥)ℎ(𝑦)∇ℎ(𝑦) 𝑇 (𝑥-𝑦) is the Bregman divergence based on reference function ℎ(•) and 𝜂 ∈ R is a fixed step-size. Note that this mirror descent step can be computed in linear time since (10) admits a closed-form solution. For example, if we use ℎ(𝝁) = 1 2 ∥𝝁 ∥ 2 as the reference function, the dual update in (10) becomes\n𝝁 𝑡 +1 = [𝝁 𝑡 -𝜂∇𝝁 𝑡 ] + ,(11)\nwhich recovers the online projected gradient descent method. Moreover, in order to guarantee the upper exposure constraints, one needs to examine the violation of upper limits of each channel before the determination of 𝒙 𝑡 at each user request. If the sum of exposures of a specific channel exceeds its upper bound, one needs to remove all candidate items from this channel to forbid allocate more exposures when determining 𝒙 𝑡 . We present the optimality of this proposed ME2A algorithm and its feasibility to guarantee exposure constraints of different channels as follows. Detailed proofs are deferred to the appendix."},{"figure_ref":[],"heading":"Algorithm 1","publication_ref":[],"table_ref":[],"text":"The proposed ME2A algorithm of MIREC Receive request 𝑒 𝑡 = (𝑢, 𝑓 , 𝑏, X 𝑡 )."},{"figure_ref":[],"heading":"6:","publication_ref":[],"table_ref":[],"text":"Update the candidate set 𝐼 𝑡 provided by multi-channels."},{"figure_ref":[],"heading":"7:","publication_ref":["b6"],"table_ref":[],"text":"Determine the optimal item list 𝒙 𝑡 by solving the primal problem in (7) at the ranking-layer."},{"figure_ref":[],"heading":"8:","publication_ref":[],"table_ref":[],"text":"Update the remained exposure resource via (8)."},{"figure_ref":[],"heading":"9:","publication_ref":["b8","b10","b2","b24","b39","b9"],"table_ref":[],"text":"Obtain the sub-gradient of the dual variable via (9). Update the dual variable based on mirror descent via (11). 11: end for 4.2.3 Optimality. It is viable to prove that Algorithm 1 is asymptotically optimal and admits a regret bound scales as O ( √ 𝑇 ) when the user requests arrive from an i.i.d unknown distribution. This assumption is reasonable when the number of requests is numerous [3,25,40]. Specifically, we denote Algorithm 1 as 𝜋 and the overall utility over all user requests in set S under the running of 𝜋 as 𝑅(𝜋 |S) = 𝑇 𝑡 =1 𝑓 (𝒙 𝑡 ). The regret of model 𝜋 is defined as the worst-case difference over S between the expected performance of the global optimal solution and the model 𝜋:\nRegret(𝜋 |S) = sup E S [OPT(S) -𝑅(𝜋 |S)] ,(12)\nwhere OPT(S) denotes the optimal utilities one can obtain under the request set S. The regret bound can be given as follows.\nTheorem 1. Suppose that the requests come from an i.i.d model with unknown distribution. Then, Regret(𝜋 |S) ≤ 𝐶 1 +𝐶 2 𝜂𝑇 + 𝐶 3 𝜂 with 𝜂 > 0 holds for any 𝑇 ≥ 1. Here 𝐶 1 , 𝐶 2 and 𝐶 3 are constant values depending on the numerical bounds of the utility 𝑓 , the consumption 𝑔, and terms from the dual iterates in Eq. (10).\nFrom Theorem 1, we obtain Regret(𝜋 |S) ≤ 𝑂 ( √ 𝑇 ) when using a step-size 𝜂 ∝ 𝑐/"},{"figure_ref":[],"heading":"√","publication_ref":[],"table_ref":[],"text":"𝑇 with any constant 𝑐 > 0. We defer the proof and detailed definitions of 𝐶 1 , 𝐶 2 , and 𝐶 3 into the appendix."},{"figure_ref":[],"heading":"Exposure Feasibility.","publication_ref":["b9"],"table_ref":[],"text":"In Algorithm 1, if the upper exposure limit of a specific channel is violated, we will forbid the exposure of any item from this channel when determining the item list 𝒙 𝑡 . Therefore, the exposure can never be overspent. On the other hand, the lower exposure limits are soft-restricted by adaptively adjusting the dual variable 𝝁. This may cause exposure underspent. However, it is viable to prove that the violation of the lower exposure limit of any channel also admits a convergence rate of 𝑂 ( √ 𝑇 ). In other words, even if the violations on lower exposure limits may occur, their growth is considerably smaller than 𝑇 . Proposition 1. Suppose the requests come from an i.i.d model with unknown distribution. Then, it holds for any 𝑇 ≥ 1 and any\nchannel 𝑚 ∈ M that 𝐺 min 𝑚 -E 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≤ 𝐶 4 + 𝐶 5\n𝜂 , where 𝐶 4 and 𝐶 5 are constant values depends on the numerical bounds of utility 𝑓 , consumption 𝑔, and terms from the dual iterates (10).\nProposition 1 states that when using 𝜂 ∝ 𝑐/"},{"figure_ref":[],"heading":"√","publication_ref":[],"table_ref":[],"text":"𝑇 with 𝑐 > 0, the exposure underspend of any channel is bounded by 𝑂 ( √ 𝑇 ). We defer the proof and definitions of 𝐶 4 and 𝐶 5 into the appendix."},{"figure_ref":[],"heading":"Local: Context-Aware Integrated Ranking","publication_ref":["b5","b6"],"table_ref":[],"text":"Different from the allocation-layer which optimizes an objective with accumulative utilities over the entire time horizon as defined in (6), the ranking-layer focus on maximizing the utilities on a single time slot. This corresponds to the primal problem given in (7):\nP 1 : x𝑡 = arg max 𝒙 𝑡 ∈X 𝑓 (𝒙 𝑡 ) -𝝁 𝑇 𝑡 𝑔(𝒙 𝑡 ) .\nIn other words, the allocation layer adjusts the exposure of items from different channels by optimizing the dual parameter 𝝁 𝑡 from a global view of all user requests. While the ranking-layer determines the optimal item list 𝒙 𝑡 under a fixed dual variable 𝝁 𝑡 from a local view of a given user request 𝑒 𝑡 .\nThere are two common characteristics that are strongly related to the estimation of 𝑓 (𝒙 𝑡 ) and 𝑔(𝒙 𝑡 ) in the integrated recommendation. First, users' preference on different channels has a great impact on the utility (e.g., prefer to click or not) and exposure (e.g., prefer to view or not) estimations. Therefore, it is of vital importance to consider both intra-channel and inter-channel correlations with reference to user interests during the estimation. Second, in feed products, users tend to review a large number of items in a row such that the previously viewed items have a great impact on users' behavior towards the next item. Therefore, it is necessary to consider page context when determining the item order.\nTherefore, we propose two models to deal with the above two challenges, respectively. First, we propose PCR model to deal with the joint modeling of user interests and inter/intra-channel correlation. It gives a point-wise estimation of the utility/exposure value of presenting each candidate item. Second, we propose CAR model to refine the point-wise estimation from PCR into contextaware estimation by considering both context information and the high-level knowledge obtained from PCR. It is also responsible for selecting optimal items from a set of candidate items to generate the final return list. In real-world systems, for each user request, we only need to run PCR once to get the point-wise scores, and then run CAR multiple times to generate the return list. Next, we mainly focus on the estimation of 𝑓 (𝑥), the estimation of 𝑔(𝑥) can be performed in a similar way by changing the learning goals."},{"figure_ref":["fig_0"],"heading":"Personalized","publication_ref":["b30","b26","b12","b6","b27","b37","b38","b19","b18"],"table_ref":[],"text":"Cross-Channel Ranking Model. PCR takes four types of features as input, i.e., the user profile feature 𝑋 𝑢 , the user behavior sequences 𝑋 𝑏 , the candidate items provided by each channel 𝑋 𝑙 , and the item-level features of target item 𝑋 𝑖 . As shown in Figure 2, We use an embedding layer to transform these features into dense embedding vectors, denoted as 𝐸 𝑢 , 𝐸 𝑏 , 𝐸 𝑙 and 𝐸 𝑖 , respectively. These embedding vectors are then fed into three components, i.e., the intra-channel encoding layer, interest-aware evolution layer, and inter-channel encoding layer in order, which are described below. Intra-Channel Encoding layer. This layer aims at extracting the mutual influence of item pairs and other extra information within the channel. We adopt the well-known multi-head attention [31] as the basic learning unit for intra-channel encoding. This is due to that the self-attention mechanism is able to directly capture the mutual influences between any two items, and is robust to far distance within the sequence. Formally, the formulation of this attention-based encoding can be written as\n𝑉 𝑚 𝑙 = [ℎ𝑒𝑎𝑑 1 , ℎ𝑒𝑎𝑑 2 , ..., ℎ𝑒𝑎𝑑 ℎ ]𝑊 𝑂 ,(13a)\nℎ𝑒𝑎𝑑 𝑖 = Softmax (𝐸 𝑏 𝑊 𝑄 )(𝐸 𝑏 𝑊 𝐾 ) 𝑇 √︁ 𝑑 ℎ /ℎ (𝐸 𝑏 𝑊 𝑉 ) ,(13b)\nwhere 𝑊 𝑂 ∈ R 𝑑 ℎ ×𝑑 ℎ denotes the learnable parameters for each head with 𝑑 ℎ being the length of projected embedding vector after attention, 𝑊 𝑄 ,𝑊 𝐾 ,𝑊 𝑉 ∈ R 𝑑×𝑑 ℎ /ℎ are the vectors of query, key and value with 𝑑 being the length of original embedding vector and ℎ being the number of heads, 𝑉 𝑚 𝑙 represents the encoded candidate items of each channel 𝑚 ∈ M. Interest-Aware Evolution Layer. Existing works such as PRM [27] and DHANR [13] directly apply the self-attention mechanism to model the inter-dependencies among items and channels without considering user's recent behavior. However, the interests hidden in user's behavior items usually have a great impact on the prediction accuracy in recommendation tasks [7,28,38,39]. The recently proposed PEAR [20] firstly models the dependency between the candidate item list and the user's historical behaviors based on a transformer-like structure, which, however, suffers from two limitations. First, directly mixing the raw item-level features from user behavior items may introduce redundant or noisy information to degrade the learning performance. Second, each user may exhibit multiple interest points, such that it is beneficial to reinforce the interest related to the target item before feature-crossing to avoid drifting. Therefore, we first reinforce the interest vector according to the correlation between behavior items and the target item as\n𝑉 𝑈 = 𝑓 (𝐸 𝑏 ; 𝐸 𝑖 ) = ∑︁ 𝐵 𝑖=1 𝐴(𝑏 𝑖 , 𝐸 𝑖 )𝑏 𝑖 = ∑︁ 𝐵 𝑖=1 𝑤 𝑖 𝑏 𝑖 , (14\n)\nwhere 𝐵 is the length of behavior sequence, 𝑏 𝑖 is the behavior item, 𝑉 𝑈 denotes the user representation feature with respect to 𝐸 𝑖 , and 𝐴(•) is a feed-forward network whose output is the activation weight 𝑤 𝑖 . Then, we make use of this reinforced interest vector to extract useful information from the candidate items of different channels. Formally, for each channel 𝑚, given 𝑉 𝑚 𝑙 and 𝑉 𝑈 as inputs, we use scaled dot-product attention formulated as follows:\n𝐻 𝑖 𝑠 = Softmax (𝑉 𝑚 𝑙 𝑊 𝑞 ) [𝑉 𝑈 𝑊 𝑘1 , 𝑉 𝑚 𝑙 𝑊 𝑘2 ] 𝑇 √︁ 𝑑 ℎ [𝑉 𝑈 𝑊 𝑣1 , 𝑉 𝑚 𝑙 𝑊 𝑣2 ],(15)\nwhere 𝑊 𝑘1 ,𝑊 𝑣1 ∈ R 𝑑×𝑑 ℎ and 𝑊 𝑘2 ,𝑊 𝑞 ,𝑊 𝑣2 ∈ R 𝑑 ℎ ×𝑑 ℎ are all learnable parameters, [•] denotes the concatenation operation. After the above operations, we successfully merged the information from the candidate item lists and the user's historical behaviors into a series of evolved embedding vectors for further processing. Inter-Channel Encoding Layer. Previous layers mainly extract intra-channel correlations. We now focus on the modeling of interchannel correlation. First, we feed the embedding vector of each channel and the target item embedding into the MLP layer with softmax function to obtain the importance weights on each channel that is related to the target item:\n𝑊 𝐶𝐻 = Softmax(MLP[𝐻 𝑚 𝑠 , 𝐸 𝑖 ]),(16)\nwhere 𝑊 𝐶𝐻 ∈ R 1×𝑚 is the importance weights, and 𝐻 𝑚 𝑠 denotes the concatenation of all channels' output from the Interest-Aware Evolution Layer. Second, we perform multi-head self-attention on the evolved embedding vector 𝐻 𝑚 𝑠 of each channel 𝑚 ∈ M to obtain the mixed embedding H𝑚 𝑠 which contains rich inter-channel information. Then, we perform the weighted sum on the mixed embedding of all channels based on𝑊 𝐶𝐻 to get the final representation of multi-channel modeling:\n𝑉 𝐿 = 𝑊 𝐶𝐻 • [ H𝑚 𝑠 ] 𝑇 , 𝑚 ∈ M,(17)\nwhere [ H𝑚 𝑠 ] represents the concatenation of the mixed embeddings of all channels.\nFinally, we concatenate all vectors as input and feed it into the MLP layers with a sigmoid function to predict the utility of presenting a given target item to a given target user as We maintain two types of context information in CAR, i.e., the context of previous items and the context of remaining candidate items. Specifically, when selecting the 𝑘-th item in a page, we represent the context of previously presented 𝑘 -1 items ℎ 𝑝𝑟𝑒 by mean-pooling over their embeddings. Meanwhile, we represent the context of all candidate items ℎ 𝑐𝑎𝑛 by mean-pooling over the embeddings of all remained candidates. These two context vectors are updated and repeated along with the item selection process. Furthermore, we perform a series of embedding crossing operations between the target item embedding 𝑒 𝑖 and the context embeddings to model the influence from page context. In specific, the operations in the context of previous items can be formulated as follows:\n𝑌 𝑃𝐶𝑅 = Sigmoid(Concat(𝐸 𝑢 , 𝐸 𝑖 , 𝑉 𝐿 ))(18\n𝐻 𝑝𝑡 = Concat(ℎ 𝑝𝑟𝑒 ⊕ 𝑒 𝑖 , ℎ 𝑝𝑟𝑒 ⊗ 𝑒 𝑖 , ℎ 𝑝𝑟𝑒 ⊖ 𝑒 𝑖 ),(19)\nwhere ⊕, ⊗, and ⊖ denote the addition, subtraction, and dot product between embedding vectors, respectively. The same goes for 𝐻 𝑐𝑡 by replacing ℎ 𝑝𝑟𝑒 in (19) with the context of candidate items ℎ 𝑐𝑎𝑛 . Additionally, we also perform embedding-crossing between the context embeddings and the high-level knowledge 𝑉 𝐿 from PCR to obtain another two vectors, i.e., 𝐻 𝑣 𝑐𝑡 and 𝐻 𝑣 𝑝𝑟𝑒 . Finally, for each candidate item 𝑖 ∈ 𝐼 𝑐𝑎𝑛𝑑 , we predict the contextaware utility score by feeding these embedding vectors along with the point-wise utility score 𝑌 𝑃𝐶𝑅 from PCR and user profile features 𝐸 𝑢 into one MLP layer as\n𝐻 all = Concat(𝐻 𝑝𝑡 , 𝐻 𝑐𝑡 , 𝐻 𝑣 𝑐𝑡 , 𝐻 𝑣 𝑝𝑟𝑒 , ℎ 𝑝𝑟𝑒 , ℎ 𝑐𝑎𝑛 , 𝐸 𝑢 , 𝑌 𝑃𝐶𝑅 ), (20a) 𝑌 𝐶𝐴𝑅 =𝜎 (MLP(𝐻 all )),(20b)\nwhere 𝜎 represents the sigmoid activation function. After scoring all candidate items, we choose the item with the highest score as the optimal item for slot 𝑘, and update the context vectors and the remained candidate items accordingly. This process will be repeated 𝐾 times to generate a return item list of length 𝐾. Note that the above operations in CAR only involve linear computations, such that this item selection process is cost-efficient in online systems. Both PCR and CAR can be trained with the commonly used crossentropy loss as in other ranking models, the learning objective can be given as\n𝐽 = ∑︁ 𝑒 𝑡 ∈ D 𝑦 𝑒 𝑡 𝑢,𝑖 log ŷ𝑒 𝑡 𝑢,𝑖 + (1 -𝑦 𝑒 𝑡 𝑢,𝑖 ) log(1 -ŷ𝑒 𝑡 𝑢,𝑖 ) ,(21)\nwhere D denotes the training dataset, 𝑦 𝑒 𝑡 𝑢,𝑖 is the real user-item recommendation label (equals 1 or 0) between user 𝑢 and item 𝑖 at request 𝑒 𝑡 , and ŷ𝑒 𝑡 𝑢,𝑖 is the predicted label given by 𝑌 𝑃𝐶𝑅 or 𝑌 𝐶𝐴𝑅 . In our experiments, when predicting the utility function 𝑓 with user clicks, 𝑦 𝑒 𝑡 𝑢,𝑖 refers to the click label; when predicting the cost function 𝑔 with exposure constraints, 𝑦 𝑒 𝑡 𝑢,𝑖 refers to the exposure label between user 𝑢 and item 𝑖 at request 𝑒 𝑡 , i.e., whether user 𝑢 has seen item 𝑖 at request 𝑒 𝑡 . One can readily change the learning objective according to actual demands."},{"figure_ref":[],"heading":"Online Implementation","publication_ref":[],"table_ref":[],"text":"In this section, we introduce the online implementation of our proposed MIREC model in the homepage feed of Taobao. The presented system architecture is able to handle 120, 000 QPS at traffic peak and respond within 20 milliseconds in general. It now serves the main traffic of Taobao to provide services to hundreds of millions of users towards billions of items in Taobao every day.\nFigure . 3 gives a general architecture to implement our proposed MIREC model in real-world IRS. Each time a user request is triggered from the device, the upstream RS of each channel will run its own recommendation models to determine the top items to return. The Integrated Recommendation Controller uses the top items from all channels as the candidates. It retrieves user/item features from a feature center in real-time and ranks candidate items by solving based on our proposed LCPR and PCR model. Meanwhile, the dual variable 𝜇 is estimated by an Exposure Controller. This module monitors the completeness of exposure guarantees based on the real-time exposures collected from user logs and updates the dual variable to adjust the exposures on different channels periodically."},{"figure_ref":[],"heading":"EXPERIMENTAL RESULTS","publication_ref":["b7"],"table_ref":[],"text":"This section conducts extensive experiments on both offline datasets and real-world applications with the goal of answering the following research questions: Q1: Does our proposed PCR and CAR outperform other baseline models in integrated ranking tasks? Q2: Does our proposed MIREC framework outperform other methods in integrated recommendation tasks with exposure constraints? Q3: How does MIREC perform in real-world applications?\n5.1 Experimental Setup 5.1.1 Datasets. We use one public dataset named MicroVideo-1.7M and one industrial dataset named Taobao for experiments. The public available MicroVideo-1.7M dataset1 released by [8] contains 12, 737, 619 interactions that 10, 986 users have made on 1, 704, 880 micro-videos. This dataset provides rich user behavior data and timestamps to evaluate the performance on both interest modeling and context-aware reranking. The Taobao dataset is an industrial private dataset that contains users' behaviors and feedback logs from multiple channels in the homepage feed of Taobao Mobile App. It is one of the largest feed scenarios for online merchandise in China. The feed provides items in form of the streams, videos, pictures, etc, from various channels. Users can slide to view more items in a row. This dataset contains about ten billion interactions that one hundred million of users have made on sixty million items. We also conduct online A/B tests on the platform Taobao to examine the performance of MIREC in real-world applications."},{"figure_ref":[],"heading":"Comparing","publication_ref":["b38","b38","b0","b26","b19","b29","b12","b35","b5","b5"],"table_ref":[],"text":"Methods. We compare MIREC with two mainstreams of baselines. The first steam of baselines are the methods for ranking tasks with different goals on user interest modeling (i.e., DIN and DIEN), re-ranking (i.e., DLCM, PRM, and PEAR), or multichannel recommendation (i.e., STAR and DHANR). Specifically, DIN [39] is a widely used benchmark for sequential user data modeling in point-wise CTR predictions, which models short behavior sequences with target attention. DIEN [39] combines GRUs and attention to capture temporal interests from users' historical behaviors with respect to the target item. DLCM [1] uses gated recurrent units (GRU) to sequentially encode the top-ranked items with their feature vectors. PRM [27]: directly optimizes the whole recommendation list by employing a Transformer structure to efficiently encode the information of all items in the list. PEAR [20] not only captures feature-level and item-level interactions but also models item contexts from both the candidate list and the historical clicked item list. STAR [30] trains a unified model to serve all channels simultaneously, which consists of shared centered parameters and channel-specific parameters. DHANR [13] proposes a hierarchical self-attention structure to consider cross-channel interactions.\nThe second stream of baselines is the online allocation methods which have been successfully applied in industrial applications for online resource allocation. Fixed is the fixed-positions strategy, where the positions of recommended items and ads are manually pre-determined for every request. 𝛽-WPO: is based on the Whole-Page Optimization (WPO) [36]. WPO ranks recommended and ad candidates jointly according to the predefined ranking scores. Similar to [6], we introduce an adjustable variable 𝛽 to control the proportion of different channels on each request to satisfy the resource constraint. In general, 𝛽-WPO can be regarded as a heuristic list merging algorithm. Each list from one channel is assigned a priority weight. The algorithm merges the top items of each list based on both their ranking scores and the priority weights into a final return list. HCA2E [6]: proposed a two-level optimization framework based on BwK methods. The high-level determines whether to present ads on the page while the low-level searches the optimal position to insert ads heuristically."},{"figure_ref":[],"heading":"Metrics.","publication_ref":["b18"],"table_ref":[],"text":"For offline experiments, we use user clicks to measure the utility function 𝑓 . For online experiments, we consider a joint measurement of user click, purchase, and stay-time for utility function 𝑓 . For all experiments, we use the exposure of items to measure cost function 𝑔. In this case, we compare the performance of integrated ranking in offline evaluation using the widely used Area Under ROC (AUC) and normalized discounted cumulative gain (nDCG) [19]. Remark that nDCG@K refers to the performance of top-k recommended items in the return list. The online performance is evaluated by CLICK, Click-Through-Rate (CTR), Gross Merchandise Volume (GMV), and Stay Time. Here, CLICK refers to the total number of clicked items. CTR is defined as CLICK/PV with PV denoting the total number of impressed items. CTR measures users' willingness to click and is therefore a widely used metric in practical applications. GMV is a term used in online retailing to indicate a total sales monetary-value for merchandise sold over a certain period of time. Stay Time denotes the time period of users' average stay time in the product, averaged on all users.\n5.1.4 Reproducibility. Our source codes have been made public to ensure reproducibility 2 . In all experiments, we use the validation set to tune the hyper-parameters to generate the best performance for different methods. The learning rate is searched from 10 -4 to 10 -2 . The L2 regularization term is searched from 10 -4 to 1. All models use Adam as the optimizer."},{"figure_ref":[],"heading":"Offline Evaluation","publication_ref":[],"table_ref":["tab_3"],"text":"5.2.1 Q1: Performance on Integrated Ranking. We first compare the performance with the first stream of baselines on item ranking. The results are shown in Table 1, which leads to the following findings. First, the re-ranking methods perform generally better than the point-wise user interest methods, indicating that modeling mutual influence among the input ranking list is of vital importance for the ranking. Therefore, it is essential to consider the influence from page context in feed recommendations. Second, the multi-channel methods perform better than the reranking methods, which verifies that exploiting the distinction and mutual influence among different channels has a great impact on integrated recommendations. Besides, we also notice that DHANR performs better than STAR, which may be due to that DHANR considers both the correlation among different channels and the influence from the candidate list. Finally, our proposed MIREC model achieves superior performance than other competitors on all datasets, verifying the effectiveness of joint modeling the cross-channel information, user interest, context information, and candidate list."},{"figure_ref":[],"heading":"Ablation Study.","publication_ref":["b19"],"table_ref":[],"text":"The results in Table . 2 investigates the impact of each component of MIREC on item quality estimation. Specifically, PCR * replaces the attention mechanism for user behaviors in the Merged-Sequence Evolution layer with a self-attention mechanism which is in accordance with PEAR [20]. PCR outperforms PCR * indicates that the tailored attention mechanism in PCR can filter out noisy or redundant information from historical behaviors to benefit the subsequent modeling of bi-sequence interaction. PCR † removes the scaled dot-product attention mechanism (i.e. there is no explicit interaction between initial lists and user behaviors) and achieved a worse performance. This demonstrates the necessity of this direct modeling between sequences, directly guiding the reordering of the initial lists. PCR w/o IntraCE removes the Intra-Channel Encoding module, i.e., directly feeding the embeddings of the initial item lists into subsequent layers for learning. The result shows that PCR achieves superior performance than PCR w/o IntraCE, verifying that it is of vital importance to model the mutual information inside each channel for final prediction. PCR w/o InterCE removes the Inter-Channel Encoding, which also leads to worse performance. It verifies that without considering the relationship and distinction between different channels will degrade the model performance considerably. Moreover, the joint learning of PCR and CAR performs better than only using PCR. This verifies that the modeling of page-wise context information can improve prediction accuracy effectively."},{"figure_ref":[],"heading":"Q2:","publication_ref":[],"table_ref":["tab_5","tab_5"],"text":"Performance with Exposure Constraints. To the best of our knowledge, there does not exist publicly available datasets which has rich user logs and multi-channel features to examine the joint performance of integrated recommendation and exposure allocation. Therefore, we only perform experiments on Taobao dataset, +1.42% using the complete platform logs. In this experiment, we assume that the IRS needs to allocate exposures to satisfy the exposure guarantees of four distinct channels. The aim is to maximize the overall user-click utility of all channels. The compared fixed, WPO, and HCA2E methods all use point-wise scores to be consistent with their original proposals. For HCA2E, we use their proposed heuristic search method to determine the final order of the item list. The results are shown in Table 3, which are averaged on multiple runs to give a fair comparison. The simulated time horizon is one complete day with more than one billion user requests from real productive environment. We evaluate the performance using two different sets of lower bounds: 1) Channel 1=55%, Channel 2 = 20%, Channel 3 = 15%, Channel 4 = 10%; 2) Channel 1=70%, Channel 2 = 15%, Channel 3 = 10%, Channel 4 = 5%. The parameters of all comparing methods are carefully tuned to satisfy the exposure constraints. The completeness in Table 3 shows that all methods can control the violation of constraints to a low-level. HCA2E and our proposed MIREC perform slightly better than the fixed method and the WPO method. Noticeably, our proposed method outperforms other comparing methods considerably in terms of CTR enhancement, which verifies that the joint use of the allocation and estimation algorithm can bring a remarkable improvement in practical environments."},{"figure_ref":[],"heading":"Online Evaluation","publication_ref":[],"table_ref":["tab_6"],"text":"MIREC has been fully deployed in the homepage feed of Taobao named Guess-you-like to serve the main traffic. Guess-you-like is one of the largest merchandise feed recommendation platform in China, which serves more than hundreds of millions of users toward billions of items every day. We deploy MIREC at the integrated recommendation stage in Guess-you-like platform, which takes hundreds of candidate items provided by multiple channels as input and outputs the final item list to return to the user. The online performance is compared against our previous baseline which is similar as a combination of 𝛽-WPO and HCA2E. In particular, the baseline uses a point-wise model for item quality estimation and uses a PID-based feedback control to automatically adjust parameter 𝛽 to guarantee the exposures for different channels. For each user request, the baseline also runs an MDP-based search method to determine the optimal card layout based on the estimated scores, which is similar as the heuristic search method in HCA2E. The overall performance in Table 4 is averaged over two consecutive weeks. The results show that compared with the baseline method, MIREC brings an improvement of 3.00% for CLICK, 1.75% for CTR, 1.56% for GMV, and 1.42% for stay time. Compared with the fixed method, MIREC brings an improvement of 3.00% for CLICK, 1.75% for CTR, 1.56% for GMV, and 1.42% for stay time. These improvements indicate that our framework is able to increase user's willingness to stay and interact with the recommended items in practical applications. It is noteworthy that 1% improvement on CLICK in Guess-you-like brings millions of clicks every day. Figure . 4 shows a detailed comparison of the exposure allocation results of a specific channel, where the items from this channel have a generally lower CTR than others. Each line in Figure . 4(a) represents the robustness of long-term exposure guarantee of this channel within two consecutive weeks. It is clear that compared with the baseline, our proposed MIREC is more robust to alleviate daily exposure fluctuations. The distribution of exposures on different positions in the feed is given in Figure . 4(b). The result shows that our proposed framework tends to put lower-quality items backward to increase the overall utilities of all channels. Consequently, as shown in Figure . 4(c), the averaged CTR of all channels on each position can be improved remarkably. This verifies that MIREC is superior in optimizing the item layout from a global perspective."},{"figure_ref":[],"heading":"A PROOF OF REGRET BOUND","publication_ref":["b2","b24","b2","b24","b2","b24","b24","b24","b29","b33","b24","b24"],"table_ref":[],"text":"Our proof shares the same spirit as that of Theorem 1 in [3,25]. The difference is that [3] does not consider a lower resource limit while [25] develops proof with an additional learnable parameter within 𝑓 (𝒙 𝑡 ) and 𝑔(𝒙 𝑡 ). Therefore, we here develop a separate proof that is consistent with our formulation. We directly refer to a few propositions in [3,25] as preliminaries for simplicity. It is noteworthy that developing new proof of the online revenue maximization problem is not the main focus of this paper.\nRecall that the integrated recommendation problem is\nP 0 : OPT(S) = max 𝒙 𝑡 ∈X ∑︁ 𝑇 𝑡 =1 𝑓 (𝒙 𝑡 )(22)\ns.t. 𝐶 1 : ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≤ 𝐺 max 𝑚,𝑡ℎ 𝑁 (S), ∀𝑚 ∈ M,(23)\n𝐶 2 : ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≥ 𝐺 min 𝑚,𝑡ℎ 𝑁 (S), ∀𝑚 ∈ M.(24)\nSince 𝑁 (S) denotes the sum exposures over the entire time horizon from 𝑡 = 1 to 𝑇 , we can replace the the upper exposure limit 𝐺 max 𝑚,𝑡ℎ 𝑁 (S) and lower exposure limit 𝐺 min 𝑚,𝑡ℎ 𝑁 (S) with 𝑇𝐺 𝑚 and 𝛼𝑇𝐺 𝑚 for simplicity, respectively, where 𝐺 𝑚 , 𝛼 ∈ [0, 1] are constants. As such, problem P 0 can be reformulated as\nP 1 : OPT(S) = max 𝒙 𝑡 ∈X ∑︁ 𝑇 𝑡 =1 𝑓 (𝒙 𝑡 )(25) s\n.t. 𝛼𝑇𝐺 𝑚 ≤ ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≤ 𝑇𝐺 𝑚 , ∀𝑚 ∈ M.(26)\nBefore our analysis, we define constants f , ḡ > 0, 𝐺 > 0 and Ḡ > 0 such that sup 𝒙 ∈X 𝑓 (𝒙) ≤ f , sup 𝒙 ∈X 𝑔(𝒙) ≤ ḡ, 𝐺 := min 𝑚 ∈M 𝐺 𝑚 and Ḡ := max 𝑚 ∈M 𝐺 𝑚 . Also, 𝜃 refers to the strongly-convexity parameter of the reference function ℎ(•).\nFirst, we bound the dual iterates as follows.\nAssumption A.1. There exists a constant 𝐶 ℎ > 0 such that the dual iterates\n𝜇 𝑡 satisfy E[||∇ℎ(𝜇 𝑡 )|| ∞ ] ≤ 𝐶 ℎ , ∀𝑡 ∈ [𝑇 ].\nRemark 1. Note that, when choosing the reference function ℎ(𝜆) := 1 2 ∥𝜆∥ 2 , Assumption A.1 can be omited according to Proposition 3 in [25].\nDenote the online Algorithm 1 as 𝜋 which makes a real-time decision 𝒙 𝑡 at time 𝑡. Define the stopping time 𝜏 𝜋 ≤ 𝑇 as the minimum between 𝑇 and the smallest time 𝑡 such that there exists 𝑚 ∈ M with 𝜏 𝜋 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) + ḡ > 𝑇𝐺 𝑘 . In other words, 𝜏 𝜋 refers to the first time the violation of one resource constraint happens. We bound the averaged gap between 𝑇 and 𝜏 𝜋 as follows.\nProposition 2. Suppose that Assumption A.1 holds, using a constant step-size 𝜂 > 0 in Algorithm 1 yields\nE [𝑇 -𝜏 𝜋 ] ≤ ḡ 𝐺 + 𝐶 ℎ + ∥∇ℎ(𝜆 1 )∥ ∞ 𝜂𝐺 .(27)\nProof. According to Step. 9 in Algorithm 1 we have\n∇𝜇 𝑘,𝑡 = -𝑔 𝑘 (𝒙 𝑡 ) + 𝐺 𝑘 (1(𝝁 𝑘 ≥ 0) + 𝛼 𝑘 1(𝝁 𝑘 < 0)) , ≤ -𝑔 𝑘 (𝒙 𝑡 ) + 𝐺 𝑘 , ∀𝑘 ∈ [𝑚].(28)\nAssume that the stopping time 𝜏 𝜋 is activated due to the violation of constraint on 𝑘-th channel, we have\n𝜏 𝜋 ∑︁ 𝑡 =1 ∇𝜇 𝑘,𝑡 ≤𝐺 𝑘 𝜏 𝜋 - 𝜏 𝜋 ∑︁ 𝑡 =1 𝑔 𝑘 (𝒙 𝑡 ) ≤ 𝐺 𝑘 𝜏 𝜋 -𝑇𝐺 𝑘 + ḡ,(29)\nwhich leads to\n𝑇 -𝜏 𝜋 ≤ 1 𝐺 𝑘 ḡ - 𝜏 𝜋 ∑︁ 𝑡 =1 ∇𝜇 𝑘,𝑡 .(30)\nAccording to Proposition 6 in [25], the gradients of mirror descent satisfy ∇ℎ 𝑘 (𝜇 𝑡 +1 𝑘 ) ≥ ∇ℎ 𝑘 (𝜇 𝑡 𝑘 ) -𝜂∇𝜇 𝑡 𝑘,𝑡 , ∀𝑡 ≤ 𝜏 𝜋 , such that\n-𝜏 𝜋 𝑡 =1 ∇𝜇 𝑘,𝑡 ≤ 1 𝜂 ∇ℎ 𝑘 (𝜇 𝜏 𝜋 +1 𝑘 ) -∇ℎ 𝑘 (𝜇 1 𝑘 ) .\nCombing with the inequality in (30), we obtain\nE [𝑇 -𝜏 𝜋 ] ≤ ḡ 𝐺 𝑘 + E ∇ℎ 𝑘 (𝜇 𝜏 𝜋 +1 𝑘 ) -∇ℎ 𝑘 (𝜇 1 𝑘 ) 𝜂𝐺 𝑘 (31) ≤ ḡ 𝐺 + 𝐶 ℎ + ∥∇ℎ(𝜆 1 )∥ ∞ 𝜂𝐺 ,(32)\nas required. ■ Let us denote the random variable 𝛾 𝑡 to be the type of the request at period 𝑡, which can determine the sample of the request. 𝐺 𝑘 ([𝝁 𝑘 ] + -𝛼 𝑘 [-𝝁 𝑘 ] + ) , (34) where 𝜑 (𝝁) = 𝑓 * (𝝁) -𝝁 𝑇 𝑔(𝒙 𝑡 ). Considering that 𝒙 𝑡 is an optimal solution of 𝜑 (𝝁 𝑡 ) not of 𝜑 (𝝁), we have 𝑓 (𝒙 𝑡 ) -𝝁 𝑇 𝑔(𝒙 𝑡 ) ≤ 𝜑 (𝝁). Then, by taking 𝝁 = [0, 0, . . . , 0], and summing from one to 𝜏 𝜋 , we obtain \nCombining ( 37) and ( 38), we complete the proof of Proposition 3. ■ Before providing more details on the proof of regret bound, we introduce a new benchmark of problem P 0 as in [25] due to that problem P 0 may be infeasible due the presence of both lower and upper exposure constraints. Specifically, we define OPT(S, 𝜆).\nThis benchmark is an interpolate value between the expected optimal value of problem P 0 and a deterministic problem which replaces the varying utility values 𝑓 (𝒙 𝑡 ) and cost values 𝑔(𝒙 𝑡 ) with their expected values. For this benchmark, we have \nwhere the inequality uses the fact that OPT(S) ≤ 𝐷 (𝝁 |S) according to Proposition 1 in [25]and that OPT(S) ≤ 𝑇 f . Combining all findings together, we have \nwhere the second inequality comes from Proposition 2."}],"string":"[\n {\n \"figure_ref\": [],\n \"heading\": \"INTRODUCTION\",\n \"publication_ref\": [\n \"b22\",\n \"b16\",\n \"b28\",\n \"b12\",\n \"b16\",\n \"b22\",\n \"b28\"\n ],\n \"table_ref\": [],\n \"text\": \"Nowadays, the ever-expanding new breeds of content, e.g., pictures, live streams, and short videos, drive the recommender system to drift from the traditional homogeneous form into an integrated form. Integrated recommender systems (IRSs) aim to simultaneously recommend heterogeneous items from multiple sources/channels in a row. This integrated form greatly expands users' choices on different types of content thereby satisfying users' diversified preferences on both item-level and channel-level. Therefore, IRS has nowadays been widely deployed in various online platforms such as the homepage feeds in Kuaishou [23], XiaohongShu [17], Taobao [29], and AliExpress [13]. In these products, users continuously slide down to browse and interact with heterogeneous items in a sequential manner, as shown in Figure 1.\\nThough attractive, integrated feed recommendation faces more challenges than conventional recommendation with homogeneous items. First, real-world applications usually impose upper or lower exposure guarantees on different channels, such as lower constraints for sponsored/new content (e.g., ads and cold-start items) or upper constraints for individual channels to ensure diversity. These constraints lead to greater difficulty in the joint ranking of heterogeneous items. Second, heterogeneous items from multiple channels usually have different features and ranking strategies. Hence, it is difficult to directly compare items from different channels for joint ranking. Third, users' interests on different channels have a great impact on their behaviors, such that traditional user-item prediction models need to evolve into user-channel-item prediction models by considering both intra-channel and inter-channel information and their correlation with user interests. Finally, in feed products, users tend to review a large number of items in a row such that the previously viewed items have a great impact on the users' behavior towards the next item [17,23,29]. Therefore, it is of vital importance to consider the influence from page context when determining the item order in the return list.\\nAlthough integrated feed recommendation has been widely deployed in practice, there are still few works focusing on the above challenges systematically. In this paper, we propose a general framework named Multi-channel Integrated Recommendation with Exposure Constraints (MIREC) to deal with the multi-channel integrated recommendation task under resource constraints in feed products. MIREC consists of two layers: an allocation-layer which optimizes the exposure of different channels from a global view over all user requests, and a ranking-layer which determines the optimal item layout of a given user request from a local view. These two layers operate in an iterative manner to make online decisions along with the arrival of user requests. The main contributions are as follows.\\n• This work formulates the integrated recommendation task with exposure constraints as a binary online linear programming problem and proposes a two-layer framework named MIREC to obtain the optimal solution. We also describe a practical system architecture for its implementation in industrial platforms. • This work proposes an efficient multi-channel allocation algorithm to obtain the optimal exposure assignment of different channels over a fixed time horizon. The proposed algorithm is able to reach an optimal solution with linear complexity w.r.t. the number of constraints. We also prove that this algorithm admits a regret bound of O ( √ 𝑇 ) towards the global optimal point under certain assumptions.\\n• This work proposes a series of collaborative models to determine the optimal layout of heterogeneous items on a page, with joint modeling of user interests, cross-channel correlation, and page context. This aligns more with the browsing nature of feed products than existing models. • This work conducts extensive experiments on both offline datasets and online A/B tests to verify the superiority of our proposed method.\\nMIREC has been implemented on the homepage of Taobao to serve the main traffic. It brings 3% lift on user clicks, 1.56% lift on purchase, and 1.42% lift on stay time. It now serves hundreds of millions of users towards billions of items every day.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"RELATED WORK\",\n \"publication_ref\": [\n \"b9\",\n \"b38\",\n \"b27\",\n \"b40\",\n \"b3\",\n \"b0\",\n \"b26\",\n \"b15\",\n \"b19\",\n \"b21\",\n \"b31\",\n \"b10\",\n \"b14\",\n \"b2\",\n \"b8\",\n \"b24\",\n \"b33\",\n \"b34\",\n \"b1\",\n \"b17\",\n \"b39\",\n \"b13\",\n \"b5\",\n \"b23\",\n \"b12\",\n \"b32\",\n \"b36\",\n \"b20\"\n ],\n \"table_ref\": [],\n \"text\": \"Re-ranking Methods. The main objective of re-ranking methods is to consider the mutual influence among a list of items in order to refine the prediction results produced by point-wise ranking models. Three prevalent models are commonly adopted in the existing literature: RNN-based methods, attention-based methods, and evaluator-generator-based methods. The first two methods feed an initial ranking list produced by point-wise models (e.g., Wide&Deep [10], DIN [39] and SIM [28]) into RNN-based (e.g., MiDNN [41], Seq2Slate [4] and DLCM [1]) or attention-based structure (e.g., PRM [27], PFRN [16], PEAR [20], and Raiss [22]) sequentially and output the encoded vector at each subsequent layer to model the mutual influences among items. The evaluator-generatorbased methods (e.g., SEG [32] and GRN [11]), use a generator to generate feasible permutations and use an evaluator to evaluate their list-wise utility to determine the optimal permutation. However, most re-ranking methods mainly focus on capturing the mutual influence among homogeneous items provided by one channel, instead of heterogeneous items provided by multiple channels. Moreover, they only optimize the item order at a single time slot, instead of considering a cumulative utility over a broad time horizon under resource constraints.\\nOnline Allocation Methods. The online allocation problem with resource constraints has been mostly studied in online convex optimization [15]. The primal-dual methods [3,9,25,34,35] avoid taking expensive projection iterations by penalizing the violation of constraints through duality. The BwK methods [2,18] determines an optimal action from a finite set of possible actions and then optimize the policy of decision-making according to the observed rewards and costs over a fixed period of time. Several recent works studied the practical performance of online allocation in advertising recommendations. For example, PDOA [40] adopts the primal-dual framework by optimizing the dual prices with online gradient descent to eliminate the online max-min problem's regret. However, it assumes that the utility and cost values can be optimally estimated and only verify the performance through offline simulations. MSBCB [14] and HCA2E [6] proposed a two-level optimization framework based on BwK methods, where the high-level determines whether to present ads on the given request while the low-level searches the optimal position to insert ads. Most related works on online convex optimization focus on theoretical analysis (e.g., regret bound) instead of real-world applications. Other related works on advertising mainly consider binary content, i.e., ads or non-ads, instead of heterogeneous content. Directly extending them to deal with multi-channel recommendations in IRSs may lead to sub-optimal results.\\nIntegrated Recommendation. The integrated recommendation is a newly emerged but rapidly developing domain driven by practical problems [24]. Integrated recommendation methods need to consider both intra-channel and inter-channel features within the heterogeneous content and provide recommendation results continuously along with user arrivals. Recently, DHANR [13] proposed a hierarchical self-attention structure to consider the cross-channel interactions. HRL-Rec [33] decomposed the integrated re-ranking problem into two subtasks: source selection and item ranking, and use hierarchical reinforcement learning to solve the problem. DEAR [37] proposed to interpolate ads and organic items by deep Q-networks. Cross-DQN [21] also adopt a reinforcement learning solution with a cross-channel attention unit. However, many integrated methods only focus on ranking at a single time slot instead of over a continuous time horizon. The joint consideration of both integrated ranking and online allocation of limited resources still remains to be explored.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"PROBLEM FORMULATION\",\n \"publication_ref\": [\n \"b11\",\n \"b5\",\n \"b25\"\n ],\n \"table_ref\": [],\n \"text\": \"This section formulates the integrated recommendation task with exposure constraints as a binary online linear programming problem. Specifically, we consider a generic IRS setting where user requests arrive sequentially during a finite time horizon. For each request, the IRS needs to rank a list of heterogeneous items provided by multiple channels. The aim is to maximize the overall utilities (e.g., the sum of clicks and pays) of all channels over the entire time horizon, subject to multiple resource constraints.\\nFormally, the request of user 𝑢 triggered at time 𝑡 is described as 𝑒 𝑡 = (𝑢, 𝑓 , 𝑔, X 𝑡 ), where 𝑓 ∈ R + is a non-negative utility function, 𝑔 ∈ R + is a non-negative resource consumption function, and X 𝑡 ⊂ R 𝑑 + is a compact set denoting all possible item layouts for decisionmaking. For each request 𝑒 𝑡 , the IRS needs to choose a number of 𝑁 heterogeneous items from a candidate set 𝐼 𝑡 and place them into 𝑁 slots to form a complete page and return it to the user. This action 𝒙 𝑡 ∈ X 𝑡 can be represented as a decision matrix 𝒙 𝑡 ∈ [0, 1] 𝑁 ×|𝐼 𝑡 | , where each entry 𝑥 𝑡,𝑛,𝑖 is indexed by a slot 𝑛 and a card index 𝑖. Once the user finished viewing the current page, a new user request will be triggered to ask the platform to return to the next page. Therefore, this decision-making process will be performed repeatedly. Moreover, in real-world applications, the item layouts need to satisfy the following constraints:\\nX 𝑡 = 𝑖 ∈𝐼 𝑡 𝑥 𝑡,𝑛,𝑖 = 1, ∀𝑡 ∈ T , ∀𝑛 ∈ N 𝑛 𝑥 𝑡,𝑛,𝑖 ≤ 1, ∀𝑡 ∈ T , ∀𝑖 ∈ 𝐼,(1)\\nwhere the upper constraint restricts that each slot must be assigned to one item and the lower constraint restricts that each item can be assigned to at most one slot.\\nAfter executing an action 𝒙 𝑡 at request 𝑒 𝑡 , the IRS consumes a resource cost 𝑔(𝒙 𝑡 ) and obtains an utility 𝑓 (𝒙 𝑡 ). In IRS, the utility function 𝑓 (𝒙 𝑡 ) is defined according to the concerned metrics. For example, it can be defined as a combination of stay time, adds to cart, and favorites to encourage user engagement, or defined as a combination of clicks and purchases to encourage user conversion. On the other hand, the consumption function 𝑔(𝒙 𝑡 ) is defined based on the concerned resource constraints. For example, the platform may need to allocate a certain amount of exposure to new channels in order to support the growth of new content [12]. Meanwhile, a too large proportion of exposures on one specific channel will damage the recommendation diversity thereby harming user experience [6,26]. In this case, the IRS needs to guarantee both a lower exposure limit and an upper exposure limit for the heterogeneous items from different channels.\\nIn this paper, we focus on the exposure constraints in practical systems which lead to the following optimization problem\\nP 0 : OPT(S) = max 𝒙 𝑡 ∈X 𝑡 ∑︁ 𝑇 𝑡 =1 𝑓 (𝒙 𝑡 )(2)\\ns.t. 𝐶 1 : ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≤ 𝐺 max 𝑚,𝑡ℎ 𝑁 (S), ∀𝑚 ∈ M,(3)\\n𝐶 2 : ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≥ 𝐺 min 𝑚,𝑡ℎ 𝑁 (S), ∀𝑚 ∈ M,(4)\\nwhere 𝑁 (S) denotes the total available exposures to allocate over the entire time horizon, 𝐺 max 𝑚,𝑡ℎ and 𝐺 min 𝑚,𝑡ℎ denote the proportion of upper exposure limits and lower exposure limits for each channel 𝑚 ∈ M, respectively, and 𝑔 𝑚 (𝒙 𝑡 ) denotes the consumed exposures of cards from channel 𝑚 after executing 𝒙 𝑡 at request 𝑒 𝑡 . Although this paper mainly focuses on the exposure guarantee, the above formulation is generally applicable to other problems with different resource constraints, e.g., the number of coupons to allocate.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"METHODOLOGY 4.1 Framework Overview\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Directly solving problem P 0 is challenging. On one hand, the estimation accuracy of utility 𝑓 (𝒙 𝑡 ) and consumption 𝑔(𝒙 𝑡 ) suffer influence from multi-factors, including the user's personal interest, the page context, and the cross-correlation between different channels. On the other hand, the determination of each 𝒙 𝑡 needs to consider the cumulative exposures over the entire time horizon due to the exposure guarantees. Therefore, the optimization of exposure allocation must be performed from a global view over the entire timeline instead of a single time slot.\\nTo this end, we propose the MIREC framework which solves P 0 through online primal-dual iterations. Specifically, MIREC consists of two layers, i.e., the allocation-layer and the ranking-layer, which correspond to the dual and primal problem of P 0 , respectively. The allocation-layer optimizes dual variables to control the cumulative exposure of different channels on all user requests from a global view to guarantee the exposure limits. Meanwhile, the rankinglayer optimizes the item layout under fixed dual variables given by the allocation-layer, with the aim to maximize the instant utility on a single user request from a local view. These two layers operate in an iterative manner along with the arrival of online user requests to determine the optimal item layout at user requests continuously. The general workflow of MIREC is shown in Figure . 2.\\nFor the allocation-layer, we propose a simple but efficient Mirrordescent based Multi-channel Exposure Allocation (ME2A) algorithm to adaptively balance the utility gain and the exposure cost of presenting heterogeneous items from different channels. The proposed M2EA algorithm has a closed form solution that can be computed in linear time and admits a regret bound of O ( √ 𝑇 ) towards the global optimal point under certain assumptions.\\nFor the ranking-layer, we propose a personalized cross-channel ranking (PCR) model and a context-aware reranking (CAR) model to jointly determine the optimal item layout on a given user request, with fixed dual parameters given by the allocation-layer. In particular, PCR gives point-wise estimation of the quality of candidate items by joint modeling the influence from user interests, intra-channel information, and inter-channel correlations. Afterward, CAR refines the point-wise estimation generated by PCR into context-aware estimation by making use of both the context information and the high-level knowledge extracted from PCR.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Global: Online Exposure Allocation\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this section, we introduce the primal-dual formulation of P 0 and propose the ME2A algorithm to obtain the optimal solution for online systems. The complete algorithm is presented in Algorithm 1. \\n+ ∑︁ 𝑚 𝜆 𝑚 ∑︁ 𝑡 𝑔 𝑚 (𝒙 𝑡 ) -𝐺 min 𝑚 .(5)\\nHere 𝝁 ≥ 0 and 𝝀 ≥ 0 are the introduced dual parameters, 𝐺 max 𝑚 and 𝐺 min 𝑚 are short for 𝐺 max 𝑚,𝑡ℎ 𝑁 (S) and 𝐺 min 𝑚,𝑡ℎ 𝑁 (S), respectively. Note that the parameters 𝝁 and 𝝀 are related to the violation of exposure consumption over the upper bound limit and the lower bound limit, respectively, which are mutually exclusive. In particular, if one of them is positive, the other one must be zero; otherwise, both of them are zero. Hence, it is viable to only introduce one real number dual variable 𝝁 ∈ R 1×𝑀 to replace 𝝁 and 𝝀 in the dual problem, which simplifies (5) into\\nmin 𝝁 𝐷 (𝝁) = ∑︁ 𝑡 𝑓 (𝒙 𝑡 ) - ∑︁ 𝑚 [𝜇 𝑚 ] + ∑︁ 𝑡 𝑔 𝑚 (𝒙 𝑡 ) -𝐺 max 𝑚 (6a) + ∑︁ 𝑚 [-𝜇 𝑚 ] + ∑︁ 𝑡 𝑔 𝑚 (𝒙 𝑡 ) -𝐺 min 𝑚 (6b) = ∑︁ 𝑡 𝑓 (𝒙 𝑡 ) - ∑︁ 𝑚 𝜇 𝑚 𝑔 𝑚 (𝒙 𝑡 ) + ∑︁ 𝑚 [𝜇 𝑚 ] + 𝐺 max 𝑚 -[-𝜇 𝑚 ] + 𝐺 min 𝑚 ,(6c)\\nwhere\\n[𝜇 𝑚 ] + = max{𝜇 𝑚 , 0}.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Dual\",\n \"publication_ref\": [\n \"b4\",\n \"b9\"\n ],\n \"table_ref\": [],\n \"text\": \"Optimization for Exposure Constraints. The dual problem in (6) can be solved optimally via primal-dual updates. Specifically, given a user request 𝑒 𝑡 = (𝑢, 𝑓 , 𝑏, X 𝑡 ), we assume that the utility 𝑓 (𝒙 𝑡 ) and cost 𝝁 𝑇 𝑡 𝑔(𝒙 𝑡 ) under different item layout 𝒙 𝑡 can be properly estimated by the models at the ranking layer. As such, the optimal item layout 𝒙 𝑡 under a fixed dual variable 𝝁 𝑡 can be obtained by solving the following primal problem:\\nP 1 : x𝑡 = arg max 𝒙 𝑡 ∈X 𝑓 (𝒙 𝑡 ) -𝝁 𝑇 𝑡 𝑔(𝒙 𝑡 ) .(7)\\nThis is the focus of the ranking-layer to be discussed layer. After the optimal item layout 𝒙 𝑡 at user request 𝑒 𝑡 is properly determined, the next step is to update the dual variable 𝝁 𝑡 to adjust the exposure of different channels in future user requests. Specifically, the remained exposure resource of different channels after the presentation of 𝒙 𝑡 at 𝑒 𝑡 is updated by\\n𝐺 max 𝑚,𝑡 +1 = 𝐺 max 𝑚,𝑡 -𝑔 𝑚 (𝒙 𝑡 ), ∀𝑚 ∈ M.(8)\\nThe sub-gradient of the dual function in ( 6) can be obtained via Danskin's theorem [5] by 9) where 1(𝑥 ∈ 𝐴) is an indicator function which equals to one if 𝑥 ∈ 𝐴 otherwise zero. As such, the dual variable 𝝁 𝑡 can be updated based on the mirror-descent method as\\n∇𝜇 𝑚,𝑡 = -𝑔 𝑚 (𝒙 𝑡 ) +𝐺 max 𝑚,𝑡 +1 • 1(𝜇 𝑚,𝑡 ≥ 0) +𝐺 min 𝑚 • 1(𝜇 𝑚,𝑡 ≤ 0), (\\n𝜇 𝑚,𝑡 +1 = arg min 𝜇 𝑚 ∈R 𝜇 𝑚 ∇𝜇 𝑚,𝑡 + 1 𝜂 𝑉 ℎ (𝜇 𝑚 , 𝜇 𝑚,𝑡 ),(10)\\nwhere 𝑉 ℎ (𝑥, 𝑦) = ℎ(𝑥)ℎ(𝑦)∇ℎ(𝑦) 𝑇 (𝑥-𝑦) is the Bregman divergence based on reference function ℎ(•) and 𝜂 ∈ R is a fixed step-size. Note that this mirror descent step can be computed in linear time since (10) admits a closed-form solution. For example, if we use ℎ(𝝁) = 1 2 ∥𝝁 ∥ 2 as the reference function, the dual update in (10) becomes\\n𝝁 𝑡 +1 = [𝝁 𝑡 -𝜂∇𝝁 𝑡 ] + ,(11)\\nwhich recovers the online projected gradient descent method. Moreover, in order to guarantee the upper exposure constraints, one needs to examine the violation of upper limits of each channel before the determination of 𝒙 𝑡 at each user request. If the sum of exposures of a specific channel exceeds its upper bound, one needs to remove all candidate items from this channel to forbid allocate more exposures when determining 𝒙 𝑡 . We present the optimality of this proposed ME2A algorithm and its feasibility to guarantee exposure constraints of different channels as follows. Detailed proofs are deferred to the appendix.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Algorithm 1\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The proposed ME2A algorithm of MIREC Receive request 𝑒 𝑡 = (𝑢, 𝑓 , 𝑏, X 𝑡 ).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"6:\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Update the candidate set 𝐼 𝑡 provided by multi-channels.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"7:\",\n \"publication_ref\": [\n \"b6\"\n ],\n \"table_ref\": [],\n \"text\": \"Determine the optimal item list 𝒙 𝑡 by solving the primal problem in (7) at the ranking-layer.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"8:\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Update the remained exposure resource via (8).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"9:\",\n \"publication_ref\": [\n \"b8\",\n \"b10\",\n \"b2\",\n \"b24\",\n \"b39\",\n \"b9\"\n ],\n \"table_ref\": [],\n \"text\": \"Obtain the sub-gradient of the dual variable via (9). Update the dual variable based on mirror descent via (11). 11: end for 4.2.3 Optimality. It is viable to prove that Algorithm 1 is asymptotically optimal and admits a regret bound scales as O ( √ 𝑇 ) when the user requests arrive from an i.i.d unknown distribution. This assumption is reasonable when the number of requests is numerous [3,25,40]. Specifically, we denote Algorithm 1 as 𝜋 and the overall utility over all user requests in set S under the running of 𝜋 as 𝑅(𝜋 |S) = 𝑇 𝑡 =1 𝑓 (𝒙 𝑡 ). The regret of model 𝜋 is defined as the worst-case difference over S between the expected performance of the global optimal solution and the model 𝜋:\\nRegret(𝜋 |S) = sup E S [OPT(S) -𝑅(𝜋 |S)] ,(12)\\nwhere OPT(S) denotes the optimal utilities one can obtain under the request set S. The regret bound can be given as follows.\\nTheorem 1. Suppose that the requests come from an i.i.d model with unknown distribution. Then, Regret(𝜋 |S) ≤ 𝐶 1 +𝐶 2 𝜂𝑇 + 𝐶 3 𝜂 with 𝜂 > 0 holds for any 𝑇 ≥ 1. Here 𝐶 1 , 𝐶 2 and 𝐶 3 are constant values depending on the numerical bounds of the utility 𝑓 , the consumption 𝑔, and terms from the dual iterates in Eq. (10).\\nFrom Theorem 1, we obtain Regret(𝜋 |S) ≤ 𝑂 ( √ 𝑇 ) when using a step-size 𝜂 ∝ 𝑐/\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"√\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"𝑇 with any constant 𝑐 > 0. We defer the proof and detailed definitions of 𝐶 1 , 𝐶 2 , and 𝐶 3 into the appendix.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Exposure Feasibility.\",\n \"publication_ref\": [\n \"b9\"\n ],\n \"table_ref\": [],\n \"text\": \"In Algorithm 1, if the upper exposure limit of a specific channel is violated, we will forbid the exposure of any item from this channel when determining the item list 𝒙 𝑡 . Therefore, the exposure can never be overspent. On the other hand, the lower exposure limits are soft-restricted by adaptively adjusting the dual variable 𝝁. This may cause exposure underspent. However, it is viable to prove that the violation of the lower exposure limit of any channel also admits a convergence rate of 𝑂 ( √ 𝑇 ). In other words, even if the violations on lower exposure limits may occur, their growth is considerably smaller than 𝑇 . Proposition 1. Suppose the requests come from an i.i.d model with unknown distribution. Then, it holds for any 𝑇 ≥ 1 and any\\nchannel 𝑚 ∈ M that 𝐺 min 𝑚 -E 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≤ 𝐶 4 + 𝐶 5\\n𝜂 , where 𝐶 4 and 𝐶 5 are constant values depends on the numerical bounds of utility 𝑓 , consumption 𝑔, and terms from the dual iterates (10).\\nProposition 1 states that when using 𝜂 ∝ 𝑐/\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"√\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"𝑇 with 𝑐 > 0, the exposure underspend of any channel is bounded by 𝑂 ( √ 𝑇 ). We defer the proof and definitions of 𝐶 4 and 𝐶 5 into the appendix.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Local: Context-Aware Integrated Ranking\",\n \"publication_ref\": [\n \"b5\",\n \"b6\"\n ],\n \"table_ref\": [],\n \"text\": \"Different from the allocation-layer which optimizes an objective with accumulative utilities over the entire time horizon as defined in (6), the ranking-layer focus on maximizing the utilities on a single time slot. This corresponds to the primal problem given in (7):\\nP 1 : x𝑡 = arg max 𝒙 𝑡 ∈X 𝑓 (𝒙 𝑡 ) -𝝁 𝑇 𝑡 𝑔(𝒙 𝑡 ) .\\nIn other words, the allocation layer adjusts the exposure of items from different channels by optimizing the dual parameter 𝝁 𝑡 from a global view of all user requests. While the ranking-layer determines the optimal item list 𝒙 𝑡 under a fixed dual variable 𝝁 𝑡 from a local view of a given user request 𝑒 𝑡 .\\nThere are two common characteristics that are strongly related to the estimation of 𝑓 (𝒙 𝑡 ) and 𝑔(𝒙 𝑡 ) in the integrated recommendation. First, users' preference on different channels has a great impact on the utility (e.g., prefer to click or not) and exposure (e.g., prefer to view or not) estimations. Therefore, it is of vital importance to consider both intra-channel and inter-channel correlations with reference to user interests during the estimation. Second, in feed products, users tend to review a large number of items in a row such that the previously viewed items have a great impact on users' behavior towards the next item. Therefore, it is necessary to consider page context when determining the item order.\\nTherefore, we propose two models to deal with the above two challenges, respectively. First, we propose PCR model to deal with the joint modeling of user interests and inter/intra-channel correlation. It gives a point-wise estimation of the utility/exposure value of presenting each candidate item. Second, we propose CAR model to refine the point-wise estimation from PCR into contextaware estimation by considering both context information and the high-level knowledge obtained from PCR. It is also responsible for selecting optimal items from a set of candidate items to generate the final return list. In real-world systems, for each user request, we only need to run PCR once to get the point-wise scores, and then run CAR multiple times to generate the return list. Next, we mainly focus on the estimation of 𝑓 (𝑥), the estimation of 𝑔(𝑥) can be performed in a similar way by changing the learning goals.\"\n },\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"Personalized\",\n \"publication_ref\": [\n \"b30\",\n \"b26\",\n \"b12\",\n \"b6\",\n \"b27\",\n \"b37\",\n \"b38\",\n \"b19\",\n \"b18\"\n ],\n \"table_ref\": [],\n \"text\": \"Cross-Channel Ranking Model. PCR takes four types of features as input, i.e., the user profile feature 𝑋 𝑢 , the user behavior sequences 𝑋 𝑏 , the candidate items provided by each channel 𝑋 𝑙 , and the item-level features of target item 𝑋 𝑖 . As shown in Figure 2, We use an embedding layer to transform these features into dense embedding vectors, denoted as 𝐸 𝑢 , 𝐸 𝑏 , 𝐸 𝑙 and 𝐸 𝑖 , respectively. These embedding vectors are then fed into three components, i.e., the intra-channel encoding layer, interest-aware evolution layer, and inter-channel encoding layer in order, which are described below. Intra-Channel Encoding layer. This layer aims at extracting the mutual influence of item pairs and other extra information within the channel. We adopt the well-known multi-head attention [31] as the basic learning unit for intra-channel encoding. This is due to that the self-attention mechanism is able to directly capture the mutual influences between any two items, and is robust to far distance within the sequence. Formally, the formulation of this attention-based encoding can be written as\\n𝑉 𝑚 𝑙 = [ℎ𝑒𝑎𝑑 1 , ℎ𝑒𝑎𝑑 2 , ..., ℎ𝑒𝑎𝑑 ℎ ]𝑊 𝑂 ,(13a)\\nℎ𝑒𝑎𝑑 𝑖 = Softmax (𝐸 𝑏 𝑊 𝑄 )(𝐸 𝑏 𝑊 𝐾 ) 𝑇 √︁ 𝑑 ℎ /ℎ (𝐸 𝑏 𝑊 𝑉 ) ,(13b)\\nwhere 𝑊 𝑂 ∈ R 𝑑 ℎ ×𝑑 ℎ denotes the learnable parameters for each head with 𝑑 ℎ being the length of projected embedding vector after attention, 𝑊 𝑄 ,𝑊 𝐾 ,𝑊 𝑉 ∈ R 𝑑×𝑑 ℎ /ℎ are the vectors of query, key and value with 𝑑 being the length of original embedding vector and ℎ being the number of heads, 𝑉 𝑚 𝑙 represents the encoded candidate items of each channel 𝑚 ∈ M. Interest-Aware Evolution Layer. Existing works such as PRM [27] and DHANR [13] directly apply the self-attention mechanism to model the inter-dependencies among items and channels without considering user's recent behavior. However, the interests hidden in user's behavior items usually have a great impact on the prediction accuracy in recommendation tasks [7,28,38,39]. The recently proposed PEAR [20] firstly models the dependency between the candidate item list and the user's historical behaviors based on a transformer-like structure, which, however, suffers from two limitations. First, directly mixing the raw item-level features from user behavior items may introduce redundant or noisy information to degrade the learning performance. Second, each user may exhibit multiple interest points, such that it is beneficial to reinforce the interest related to the target item before feature-crossing to avoid drifting. Therefore, we first reinforce the interest vector according to the correlation between behavior items and the target item as\\n𝑉 𝑈 = 𝑓 (𝐸 𝑏 ; 𝐸 𝑖 ) = ∑︁ 𝐵 𝑖=1 𝐴(𝑏 𝑖 , 𝐸 𝑖 )𝑏 𝑖 = ∑︁ 𝐵 𝑖=1 𝑤 𝑖 𝑏 𝑖 , (14\\n)\\nwhere 𝐵 is the length of behavior sequence, 𝑏 𝑖 is the behavior item, 𝑉 𝑈 denotes the user representation feature with respect to 𝐸 𝑖 , and 𝐴(•) is a feed-forward network whose output is the activation weight 𝑤 𝑖 . Then, we make use of this reinforced interest vector to extract useful information from the candidate items of different channels. Formally, for each channel 𝑚, given 𝑉 𝑚 𝑙 and 𝑉 𝑈 as inputs, we use scaled dot-product attention formulated as follows:\\n𝐻 𝑖 𝑠 = Softmax (𝑉 𝑚 𝑙 𝑊 𝑞 ) [𝑉 𝑈 𝑊 𝑘1 , 𝑉 𝑚 𝑙 𝑊 𝑘2 ] 𝑇 √︁ 𝑑 ℎ [𝑉 𝑈 𝑊 𝑣1 , 𝑉 𝑚 𝑙 𝑊 𝑣2 ],(15)\\nwhere 𝑊 𝑘1 ,𝑊 𝑣1 ∈ R 𝑑×𝑑 ℎ and 𝑊 𝑘2 ,𝑊 𝑞 ,𝑊 𝑣2 ∈ R 𝑑 ℎ ×𝑑 ℎ are all learnable parameters, [•] denotes the concatenation operation. After the above operations, we successfully merged the information from the candidate item lists and the user's historical behaviors into a series of evolved embedding vectors for further processing. Inter-Channel Encoding Layer. Previous layers mainly extract intra-channel correlations. We now focus on the modeling of interchannel correlation. First, we feed the embedding vector of each channel and the target item embedding into the MLP layer with softmax function to obtain the importance weights on each channel that is related to the target item:\\n𝑊 𝐶𝐻 = Softmax(MLP[𝐻 𝑚 𝑠 , 𝐸 𝑖 ]),(16)\\nwhere 𝑊 𝐶𝐻 ∈ R 1×𝑚 is the importance weights, and 𝐻 𝑚 𝑠 denotes the concatenation of all channels' output from the Interest-Aware Evolution Layer. Second, we perform multi-head self-attention on the evolved embedding vector 𝐻 𝑚 𝑠 of each channel 𝑚 ∈ M to obtain the mixed embedding H𝑚 𝑠 which contains rich inter-channel information. Then, we perform the weighted sum on the mixed embedding of all channels based on𝑊 𝐶𝐻 to get the final representation of multi-channel modeling:\\n𝑉 𝐿 = 𝑊 𝐶𝐻 • [ H𝑚 𝑠 ] 𝑇 , 𝑚 ∈ M,(17)\\nwhere [ H𝑚 𝑠 ] represents the concatenation of the mixed embeddings of all channels.\\nFinally, we concatenate all vectors as input and feed it into the MLP layers with a sigmoid function to predict the utility of presenting a given target item to a given target user as We maintain two types of context information in CAR, i.e., the context of previous items and the context of remaining candidate items. Specifically, when selecting the 𝑘-th item in a page, we represent the context of previously presented 𝑘 -1 items ℎ 𝑝𝑟𝑒 by mean-pooling over their embeddings. Meanwhile, we represent the context of all candidate items ℎ 𝑐𝑎𝑛 by mean-pooling over the embeddings of all remained candidates. These two context vectors are updated and repeated along with the item selection process. Furthermore, we perform a series of embedding crossing operations between the target item embedding 𝑒 𝑖 and the context embeddings to model the influence from page context. In specific, the operations in the context of previous items can be formulated as follows:\\n𝑌 𝑃𝐶𝑅 = Sigmoid(Concat(𝐸 𝑢 , 𝐸 𝑖 , 𝑉 𝐿 ))(18\\n𝐻 𝑝𝑡 = Concat(ℎ 𝑝𝑟𝑒 ⊕ 𝑒 𝑖 , ℎ 𝑝𝑟𝑒 ⊗ 𝑒 𝑖 , ℎ 𝑝𝑟𝑒 ⊖ 𝑒 𝑖 ),(19)\\nwhere ⊕, ⊗, and ⊖ denote the addition, subtraction, and dot product between embedding vectors, respectively. The same goes for 𝐻 𝑐𝑡 by replacing ℎ 𝑝𝑟𝑒 in (19) with the context of candidate items ℎ 𝑐𝑎𝑛 . Additionally, we also perform embedding-crossing between the context embeddings and the high-level knowledge 𝑉 𝐿 from PCR to obtain another two vectors, i.e., 𝐻 𝑣 𝑐𝑡 and 𝐻 𝑣 𝑝𝑟𝑒 . Finally, for each candidate item 𝑖 ∈ 𝐼 𝑐𝑎𝑛𝑑 , we predict the contextaware utility score by feeding these embedding vectors along with the point-wise utility score 𝑌 𝑃𝐶𝑅 from PCR and user profile features 𝐸 𝑢 into one MLP layer as\\n𝐻 all = Concat(𝐻 𝑝𝑡 , 𝐻 𝑐𝑡 , 𝐻 𝑣 𝑐𝑡 , 𝐻 𝑣 𝑝𝑟𝑒 , ℎ 𝑝𝑟𝑒 , ℎ 𝑐𝑎𝑛 , 𝐸 𝑢 , 𝑌 𝑃𝐶𝑅 ), (20a) 𝑌 𝐶𝐴𝑅 =𝜎 (MLP(𝐻 all )),(20b)\\nwhere 𝜎 represents the sigmoid activation function. After scoring all candidate items, we choose the item with the highest score as the optimal item for slot 𝑘, and update the context vectors and the remained candidate items accordingly. This process will be repeated 𝐾 times to generate a return item list of length 𝐾. Note that the above operations in CAR only involve linear computations, such that this item selection process is cost-efficient in online systems. Both PCR and CAR can be trained with the commonly used crossentropy loss as in other ranking models, the learning objective can be given as\\n𝐽 = ∑︁ 𝑒 𝑡 ∈ D 𝑦 𝑒 𝑡 𝑢,𝑖 log ŷ𝑒 𝑡 𝑢,𝑖 + (1 -𝑦 𝑒 𝑡 𝑢,𝑖 ) log(1 -ŷ𝑒 𝑡 𝑢,𝑖 ) ,(21)\\nwhere D denotes the training dataset, 𝑦 𝑒 𝑡 𝑢,𝑖 is the real user-item recommendation label (equals 1 or 0) between user 𝑢 and item 𝑖 at request 𝑒 𝑡 , and ŷ𝑒 𝑡 𝑢,𝑖 is the predicted label given by 𝑌 𝑃𝐶𝑅 or 𝑌 𝐶𝐴𝑅 . In our experiments, when predicting the utility function 𝑓 with user clicks, 𝑦 𝑒 𝑡 𝑢,𝑖 refers to the click label; when predicting the cost function 𝑔 with exposure constraints, 𝑦 𝑒 𝑡 𝑢,𝑖 refers to the exposure label between user 𝑢 and item 𝑖 at request 𝑒 𝑡 , i.e., whether user 𝑢 has seen item 𝑖 at request 𝑒 𝑡 . One can readily change the learning objective according to actual demands.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Online Implementation\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this section, we introduce the online implementation of our proposed MIREC model in the homepage feed of Taobao. The presented system architecture is able to handle 120, 000 QPS at traffic peak and respond within 20 milliseconds in general. It now serves the main traffic of Taobao to provide services to hundreds of millions of users towards billions of items in Taobao every day.\\nFigure . 3 gives a general architecture to implement our proposed MIREC model in real-world IRS. Each time a user request is triggered from the device, the upstream RS of each channel will run its own recommendation models to determine the top items to return. The Integrated Recommendation Controller uses the top items from all channels as the candidates. It retrieves user/item features from a feature center in real-time and ranks candidate items by solving based on our proposed LCPR and PCR model. Meanwhile, the dual variable 𝜇 is estimated by an Exposure Controller. This module monitors the completeness of exposure guarantees based on the real-time exposures collected from user logs and updates the dual variable to adjust the exposures on different channels periodically.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"EXPERIMENTAL RESULTS\",\n \"publication_ref\": [\n \"b7\"\n ],\n \"table_ref\": [],\n \"text\": \"This section conducts extensive experiments on both offline datasets and real-world applications with the goal of answering the following research questions: Q1: Does our proposed PCR and CAR outperform other baseline models in integrated ranking tasks? Q2: Does our proposed MIREC framework outperform other methods in integrated recommendation tasks with exposure constraints? Q3: How does MIREC perform in real-world applications?\\n5.1 Experimental Setup 5.1.1 Datasets. We use one public dataset named MicroVideo-1.7M and one industrial dataset named Taobao for experiments. The public available MicroVideo-1.7M dataset1 released by [8] contains 12, 737, 619 interactions that 10, 986 users have made on 1, 704, 880 micro-videos. This dataset provides rich user behavior data and timestamps to evaluate the performance on both interest modeling and context-aware reranking. The Taobao dataset is an industrial private dataset that contains users' behaviors and feedback logs from multiple channels in the homepage feed of Taobao Mobile App. It is one of the largest feed scenarios for online merchandise in China. The feed provides items in form of the streams, videos, pictures, etc, from various channels. Users can slide to view more items in a row. This dataset contains about ten billion interactions that one hundred million of users have made on sixty million items. We also conduct online A/B tests on the platform Taobao to examine the performance of MIREC in real-world applications.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Comparing\",\n \"publication_ref\": [\n \"b38\",\n \"b38\",\n \"b0\",\n \"b26\",\n \"b19\",\n \"b29\",\n \"b12\",\n \"b35\",\n \"b5\",\n \"b5\"\n ],\n \"table_ref\": [],\n \"text\": \"Methods. We compare MIREC with two mainstreams of baselines. The first steam of baselines are the methods for ranking tasks with different goals on user interest modeling (i.e., DIN and DIEN), re-ranking (i.e., DLCM, PRM, and PEAR), or multichannel recommendation (i.e., STAR and DHANR). Specifically, DIN [39] is a widely used benchmark for sequential user data modeling in point-wise CTR predictions, which models short behavior sequences with target attention. DIEN [39] combines GRUs and attention to capture temporal interests from users' historical behaviors with respect to the target item. DLCM [1] uses gated recurrent units (GRU) to sequentially encode the top-ranked items with their feature vectors. PRM [27]: directly optimizes the whole recommendation list by employing a Transformer structure to efficiently encode the information of all items in the list. PEAR [20] not only captures feature-level and item-level interactions but also models item contexts from both the candidate list and the historical clicked item list. STAR [30] trains a unified model to serve all channels simultaneously, which consists of shared centered parameters and channel-specific parameters. DHANR [13] proposes a hierarchical self-attention structure to consider cross-channel interactions.\\nThe second stream of baselines is the online allocation methods which have been successfully applied in industrial applications for online resource allocation. Fixed is the fixed-positions strategy, where the positions of recommended items and ads are manually pre-determined for every request. 𝛽-WPO: is based on the Whole-Page Optimization (WPO) [36]. WPO ranks recommended and ad candidates jointly according to the predefined ranking scores. Similar to [6], we introduce an adjustable variable 𝛽 to control the proportion of different channels on each request to satisfy the resource constraint. In general, 𝛽-WPO can be regarded as a heuristic list merging algorithm. Each list from one channel is assigned a priority weight. The algorithm merges the top items of each list based on both their ranking scores and the priority weights into a final return list. HCA2E [6]: proposed a two-level optimization framework based on BwK methods. The high-level determines whether to present ads on the page while the low-level searches the optimal position to insert ads heuristically.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Metrics.\",\n \"publication_ref\": [\n \"b18\"\n ],\n \"table_ref\": [],\n \"text\": \"For offline experiments, we use user clicks to measure the utility function 𝑓 . For online experiments, we consider a joint measurement of user click, purchase, and stay-time for utility function 𝑓 . For all experiments, we use the exposure of items to measure cost function 𝑔. In this case, we compare the performance of integrated ranking in offline evaluation using the widely used Area Under ROC (AUC) and normalized discounted cumulative gain (nDCG) [19]. Remark that nDCG@K refers to the performance of top-k recommended items in the return list. The online performance is evaluated by CLICK, Click-Through-Rate (CTR), Gross Merchandise Volume (GMV), and Stay Time. Here, CLICK refers to the total number of clicked items. CTR is defined as CLICK/PV with PV denoting the total number of impressed items. CTR measures users' willingness to click and is therefore a widely used metric in practical applications. GMV is a term used in online retailing to indicate a total sales monetary-value for merchandise sold over a certain period of time. Stay Time denotes the time period of users' average stay time in the product, averaged on all users.\\n5.1.4 Reproducibility. Our source codes have been made public to ensure reproducibility 2 . In all experiments, we use the validation set to tune the hyper-parameters to generate the best performance for different methods. The learning rate is searched from 10 -4 to 10 -2 . The L2 regularization term is searched from 10 -4 to 1. All models use Adam as the optimizer.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Offline Evaluation\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_3\"\n ],\n \"text\": \"5.2.1 Q1: Performance on Integrated Ranking. We first compare the performance with the first stream of baselines on item ranking. The results are shown in Table 1, which leads to the following findings. First, the re-ranking methods perform generally better than the point-wise user interest methods, indicating that modeling mutual influence among the input ranking list is of vital importance for the ranking. Therefore, it is essential to consider the influence from page context in feed recommendations. Second, the multi-channel methods perform better than the reranking methods, which verifies that exploiting the distinction and mutual influence among different channels has a great impact on integrated recommendations. Besides, we also notice that DHANR performs better than STAR, which may be due to that DHANR considers both the correlation among different channels and the influence from the candidate list. Finally, our proposed MIREC model achieves superior performance than other competitors on all datasets, verifying the effectiveness of joint modeling the cross-channel information, user interest, context information, and candidate list.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Ablation Study.\",\n \"publication_ref\": [\n \"b19\"\n ],\n \"table_ref\": [],\n \"text\": \"The results in Table . 2 investigates the impact of each component of MIREC on item quality estimation. Specifically, PCR * replaces the attention mechanism for user behaviors in the Merged-Sequence Evolution layer with a self-attention mechanism which is in accordance with PEAR [20]. PCR outperforms PCR * indicates that the tailored attention mechanism in PCR can filter out noisy or redundant information from historical behaviors to benefit the subsequent modeling of bi-sequence interaction. PCR † removes the scaled dot-product attention mechanism (i.e. there is no explicit interaction between initial lists and user behaviors) and achieved a worse performance. This demonstrates the necessity of this direct modeling between sequences, directly guiding the reordering of the initial lists. PCR w/o IntraCE removes the Intra-Channel Encoding module, i.e., directly feeding the embeddings of the initial item lists into subsequent layers for learning. The result shows that PCR achieves superior performance than PCR w/o IntraCE, verifying that it is of vital importance to model the mutual information inside each channel for final prediction. PCR w/o InterCE removes the Inter-Channel Encoding, which also leads to worse performance. It verifies that without considering the relationship and distinction between different channels will degrade the model performance considerably. Moreover, the joint learning of PCR and CAR performs better than only using PCR. This verifies that the modeling of page-wise context information can improve prediction accuracy effectively.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Q2:\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_5\",\n \"tab_5\"\n ],\n \"text\": \"Performance with Exposure Constraints. To the best of our knowledge, there does not exist publicly available datasets which has rich user logs and multi-channel features to examine the joint performance of integrated recommendation and exposure allocation. Therefore, we only perform experiments on Taobao dataset, +1.42% using the complete platform logs. In this experiment, we assume that the IRS needs to allocate exposures to satisfy the exposure guarantees of four distinct channels. The aim is to maximize the overall user-click utility of all channels. The compared fixed, WPO, and HCA2E methods all use point-wise scores to be consistent with their original proposals. For HCA2E, we use their proposed heuristic search method to determine the final order of the item list. The results are shown in Table 3, which are averaged on multiple runs to give a fair comparison. The simulated time horizon is one complete day with more than one billion user requests from real productive environment. We evaluate the performance using two different sets of lower bounds: 1) Channel 1=55%, Channel 2 = 20%, Channel 3 = 15%, Channel 4 = 10%; 2) Channel 1=70%, Channel 2 = 15%, Channel 3 = 10%, Channel 4 = 5%. The parameters of all comparing methods are carefully tuned to satisfy the exposure constraints. The completeness in Table 3 shows that all methods can control the violation of constraints to a low-level. HCA2E and our proposed MIREC perform slightly better than the fixed method and the WPO method. Noticeably, our proposed method outperforms other comparing methods considerably in terms of CTR enhancement, which verifies that the joint use of the allocation and estimation algorithm can bring a remarkable improvement in practical environments.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Online Evaluation\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_6\"\n ],\n \"text\": \"MIREC has been fully deployed in the homepage feed of Taobao named Guess-you-like to serve the main traffic. Guess-you-like is one of the largest merchandise feed recommendation platform in China, which serves more than hundreds of millions of users toward billions of items every day. We deploy MIREC at the integrated recommendation stage in Guess-you-like platform, which takes hundreds of candidate items provided by multiple channels as input and outputs the final item list to return to the user. The online performance is compared against our previous baseline which is similar as a combination of 𝛽-WPO and HCA2E. In particular, the baseline uses a point-wise model for item quality estimation and uses a PID-based feedback control to automatically adjust parameter 𝛽 to guarantee the exposures for different channels. For each user request, the baseline also runs an MDP-based search method to determine the optimal card layout based on the estimated scores, which is similar as the heuristic search method in HCA2E. The overall performance in Table 4 is averaged over two consecutive weeks. The results show that compared with the baseline method, MIREC brings an improvement of 3.00% for CLICK, 1.75% for CTR, 1.56% for GMV, and 1.42% for stay time. Compared with the fixed method, MIREC brings an improvement of 3.00% for CLICK, 1.75% for CTR, 1.56% for GMV, and 1.42% for stay time. These improvements indicate that our framework is able to increase user's willingness to stay and interact with the recommended items in practical applications. It is noteworthy that 1% improvement on CLICK in Guess-you-like brings millions of clicks every day. Figure . 4 shows a detailed comparison of the exposure allocation results of a specific channel, where the items from this channel have a generally lower CTR than others. Each line in Figure . 4(a) represents the robustness of long-term exposure guarantee of this channel within two consecutive weeks. It is clear that compared with the baseline, our proposed MIREC is more robust to alleviate daily exposure fluctuations. The distribution of exposures on different positions in the feed is given in Figure . 4(b). The result shows that our proposed framework tends to put lower-quality items backward to increase the overall utilities of all channels. Consequently, as shown in Figure . 4(c), the averaged CTR of all channels on each position can be improved remarkably. This verifies that MIREC is superior in optimizing the item layout from a global perspective.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A PROOF OF REGRET BOUND\",\n \"publication_ref\": [\n \"b2\",\n \"b24\",\n \"b2\",\n \"b24\",\n \"b2\",\n \"b24\",\n \"b24\",\n \"b24\",\n \"b29\",\n \"b33\",\n \"b24\",\n \"b24\"\n ],\n \"table_ref\": [],\n \"text\": \"Our proof shares the same spirit as that of Theorem 1 in [3,25]. The difference is that [3] does not consider a lower resource limit while [25] develops proof with an additional learnable parameter within 𝑓 (𝒙 𝑡 ) and 𝑔(𝒙 𝑡 ). Therefore, we here develop a separate proof that is consistent with our formulation. We directly refer to a few propositions in [3,25] as preliminaries for simplicity. It is noteworthy that developing new proof of the online revenue maximization problem is not the main focus of this paper.\\nRecall that the integrated recommendation problem is\\nP 0 : OPT(S) = max 𝒙 𝑡 ∈X ∑︁ 𝑇 𝑡 =1 𝑓 (𝒙 𝑡 )(22)\\ns.t. 𝐶 1 : ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≤ 𝐺 max 𝑚,𝑡ℎ 𝑁 (S), ∀𝑚 ∈ M,(23)\\n𝐶 2 : ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≥ 𝐺 min 𝑚,𝑡ℎ 𝑁 (S), ∀𝑚 ∈ M.(24)\\nSince 𝑁 (S) denotes the sum exposures over the entire time horizon from 𝑡 = 1 to 𝑇 , we can replace the the upper exposure limit 𝐺 max 𝑚,𝑡ℎ 𝑁 (S) and lower exposure limit 𝐺 min 𝑚,𝑡ℎ 𝑁 (S) with 𝑇𝐺 𝑚 and 𝛼𝑇𝐺 𝑚 for simplicity, respectively, where 𝐺 𝑚 , 𝛼 ∈ [0, 1] are constants. As such, problem P 0 can be reformulated as\\nP 1 : OPT(S) = max 𝒙 𝑡 ∈X ∑︁ 𝑇 𝑡 =1 𝑓 (𝒙 𝑡 )(25) s\\n.t. 𝛼𝑇𝐺 𝑚 ≤ ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≤ 𝑇𝐺 𝑚 , ∀𝑚 ∈ M.(26)\\nBefore our analysis, we define constants f , ḡ > 0, 𝐺 > 0 and Ḡ > 0 such that sup 𝒙 ∈X 𝑓 (𝒙) ≤ f , sup 𝒙 ∈X 𝑔(𝒙) ≤ ḡ, 𝐺 := min 𝑚 ∈M 𝐺 𝑚 and Ḡ := max 𝑚 ∈M 𝐺 𝑚 . Also, 𝜃 refers to the strongly-convexity parameter of the reference function ℎ(•).\\nFirst, we bound the dual iterates as follows.\\nAssumption A.1. There exists a constant 𝐶 ℎ > 0 such that the dual iterates\\n𝜇 𝑡 satisfy E[||∇ℎ(𝜇 𝑡 )|| ∞ ] ≤ 𝐶 ℎ , ∀𝑡 ∈ [𝑇 ].\\nRemark 1. Note that, when choosing the reference function ℎ(𝜆) := 1 2 ∥𝜆∥ 2 , Assumption A.1 can be omited according to Proposition 3 in [25].\\nDenote the online Algorithm 1 as 𝜋 which makes a real-time decision 𝒙 𝑡 at time 𝑡. Define the stopping time 𝜏 𝜋 ≤ 𝑇 as the minimum between 𝑇 and the smallest time 𝑡 such that there exists 𝑚 ∈ M with 𝜏 𝜋 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) + ḡ > 𝑇𝐺 𝑘 . In other words, 𝜏 𝜋 refers to the first time the violation of one resource constraint happens. We bound the averaged gap between 𝑇 and 𝜏 𝜋 as follows.\\nProposition 2. Suppose that Assumption A.1 holds, using a constant step-size 𝜂 > 0 in Algorithm 1 yields\\nE [𝑇 -𝜏 𝜋 ] ≤ ḡ 𝐺 + 𝐶 ℎ + ∥∇ℎ(𝜆 1 )∥ ∞ 𝜂𝐺 .(27)\\nProof. According to Step. 9 in Algorithm 1 we have\\n∇𝜇 𝑘,𝑡 = -𝑔 𝑘 (𝒙 𝑡 ) + 𝐺 𝑘 (1(𝝁 𝑘 ≥ 0) + 𝛼 𝑘 1(𝝁 𝑘 < 0)) , ≤ -𝑔 𝑘 (𝒙 𝑡 ) + 𝐺 𝑘 , ∀𝑘 ∈ [𝑚].(28)\\nAssume that the stopping time 𝜏 𝜋 is activated due to the violation of constraint on 𝑘-th channel, we have\\n𝜏 𝜋 ∑︁ 𝑡 =1 ∇𝜇 𝑘,𝑡 ≤𝐺 𝑘 𝜏 𝜋 - 𝜏 𝜋 ∑︁ 𝑡 =1 𝑔 𝑘 (𝒙 𝑡 ) ≤ 𝐺 𝑘 𝜏 𝜋 -𝑇𝐺 𝑘 + ḡ,(29)\\nwhich leads to\\n𝑇 -𝜏 𝜋 ≤ 1 𝐺 𝑘 ḡ - 𝜏 𝜋 ∑︁ 𝑡 =1 ∇𝜇 𝑘,𝑡 .(30)\\nAccording to Proposition 6 in [25], the gradients of mirror descent satisfy ∇ℎ 𝑘 (𝜇 𝑡 +1 𝑘 ) ≥ ∇ℎ 𝑘 (𝜇 𝑡 𝑘 ) -𝜂∇𝜇 𝑡 𝑘,𝑡 , ∀𝑡 ≤ 𝜏 𝜋 , such that\\n-𝜏 𝜋 𝑡 =1 ∇𝜇 𝑘,𝑡 ≤ 1 𝜂 ∇ℎ 𝑘 (𝜇 𝜏 𝜋 +1 𝑘 ) -∇ℎ 𝑘 (𝜇 1 𝑘 ) .\\nCombing with the inequality in (30), we obtain\\nE [𝑇 -𝜏 𝜋 ] ≤ ḡ 𝐺 𝑘 + E ∇ℎ 𝑘 (𝜇 𝜏 𝜋 +1 𝑘 ) -∇ℎ 𝑘 (𝜇 1 𝑘 ) 𝜂𝐺 𝑘 (31) ≤ ḡ 𝐺 + 𝐶 ℎ + ∥∇ℎ(𝜆 1 )∥ ∞ 𝜂𝐺 ,(32)\\nas required. ■ Let us denote the random variable 𝛾 𝑡 to be the type of the request at period 𝑡, which can determine the sample of the request. 𝐺 𝑘 ([𝝁 𝑘 ] + -𝛼 𝑘 [-𝝁 𝑘 ] + ) , (34) where 𝜑 (𝝁) = 𝑓 * (𝝁) -𝝁 𝑇 𝑔(𝒙 𝑡 ). Considering that 𝒙 𝑡 is an optimal solution of 𝜑 (𝝁 𝑡 ) not of 𝜑 (𝝁), we have 𝑓 (𝒙 𝑡 ) -𝝁 𝑇 𝑔(𝒙 𝑡 ) ≤ 𝜑 (𝝁). Then, by taking 𝝁 = [0, 0, . . . , 0], and summing from one to 𝜏 𝜋 , we obtain \\nCombining ( 37) and ( 38), we complete the proof of Proposition 3. ■ Before providing more details on the proof of regret bound, we introduce a new benchmark of problem P 0 as in [25] due to that problem P 0 may be infeasible due the presence of both lower and upper exposure constraints. Specifically, we define OPT(S, 𝜆).\\nThis benchmark is an interpolate value between the expected optimal value of problem P 0 and a deterministic problem which replaces the varying utility values 𝑓 (𝒙 𝑡 ) and cost values 𝑔(𝒙 𝑡 ) with their expected values. For this benchmark, we have \\nwhere the inequality uses the fact that OPT(S) ≤ 𝐷 (𝝁 |S) according to Proposition 1 in [25]and that OPT(S) ≤ 𝑇 f . Combining all findings together, we have \\nwhere the second inequality comes from Proposition 2.\"\n }\n]"},"abstract":{"kind":"string","value":"Integrated recommendation, which aims at jointly recommending heterogeneous items from different channels in a main feed, has been widely applied to various online platforms. Though attractive, integrated recommendation requires the ranking methods to migrate from conventional user-item models to the new user-channelitem paradigm in order to better capture users' preferences on both item and channel levels. Moreover, practical feed recommendation systems usually impose exposure constraints on different channels to ensure user experience. This leads to greater difficulty in the joint ranking of heterogeneous items. In this paper, we investigate the integrated recommendation task with exposure constraints in practical recommender systems. Our contribution is forth-fold. First, we formulate this task as a binary online linear programming problem and propose a two-layer framework named Multi-channel Integrated Recommendation with Exposure Constraints (MIREC) to obtain the optimal solution. Second, we propose an efficient online allocation algorithm to determine the optimal exposure assignment of different channels from a global view of all user requests over the entire time horizon. We prove that this algorithm reaches the optimal point under a regret bound of O ( √ 𝑇 ) with linear complexity. Third, we propose a series of collaborative models to determine the optimal layout of heterogeneous items at each user request. The joint modeling of user interests, cross-channel correlation, and page context in our models aligns more with the browsing nature of feed products than existing models. Finally, we conduct extensive experiments on both offline datasets and online A/B tests to verify the effectiveness of MIREC. The proposed framework has now been implemented on the homepage of Taobao to serve the main traffic, providing service to hundreds of millions of users towards billions of items every day."},"title":{"kind":"string","value":"Multi-channel Integrated Recommendation with Exposure Constraints"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 2 :2Figure 2: An overview of the MIREC framework.","figure_data":"","figure_id":"fig_0","figure_label":"2","figure_type":"figure"},{"figure_caption":"10:","figure_data":"","figure_id":"fig_1","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 3 :3Figure 3: Online system architecture.","figure_data":"","figure_id":"fig_2","figure_label":"3","figure_type":"figure"},{"figure_caption":"CTR on different positions.","figure_data":"","figure_id":"fig_3","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 4 :4Figure 4: Online Performance Analysis.","figure_data":"","figure_id":"fig_4","figure_label":"4","figure_type":"figure"},{"figure_caption":"Proposition 3 .3Using a constant step-size rule 𝜂 > 0 for 𝑡 > 1 in Algorithm 1, it holdsE 𝜏 𝜋 𝐷 ( μ𝜏 𝜋 ) -𝜏 𝜋 ∑︁ 𝑡 =1 𝑓 (𝒙 𝑡 ) ≤ 2( ḡ2 + Ḡ2 ) 𝜃 𝜂E[𝜏 𝜋 ] + 1 𝜂 𝑉 ℎ (𝝁, 𝝁 1 ), (33)whereμ𝜏 𝜋 = 𝜏𝜋 𝑡 =1 𝝁 𝑡 𝜏 𝜋 .Proof. According to the definition of ∇𝝁 𝑡 and the subgradient inequality, we have(∇𝝁 𝑡 ) 𝑇 (𝝁 𝑡 -𝝁) ≥ 𝐷 (𝝁 𝑡 ) -𝐷 (𝝁) ≥ 𝐷 (𝝁 𝑡 ) -E 𝛾 𝑡 [𝜑 (𝝁)] + ∑︁ 𝑘 ∈ [𝐾 ]","figure_data":"","figure_id":"fig_5","figure_label":"3","figure_type":"figure"},{"figure_caption":"F(𝒙 𝑡 , 𝜆) = (1 -𝜆)𝑓 (𝒙 𝑡 ) + 𝜆E S [𝑓 (𝒙 𝑡 )] G(𝒙 𝑡 , 𝜆) = (1 -𝜆)𝑔(𝒙 𝑡 ) + 𝜆E S [𝑔(𝒙 𝑡 )],where 𝜆 ∈ [0, 1] is the interpolation parameter. We defineOPT(S, 𝜆) = E S 𝑇 max 𝒙 𝑡 ,𝑡 ∈ [𝑇 ] 𝑇 𝑡 =1 𝐹 (𝒙 𝑡 , 𝜆) s.t. 𝑇 𝛼 ⊙ 𝐺 ≤ 𝑇 𝑡 =1 𝐺 (𝒙 𝑡 , 𝜆) ≤ 𝑇𝐺whereS 𝑇 := S × • • • × S isa product distribution of length 𝑇 . Now we give the definition of the new benchmark as OPT(S) := max 𝜆 ∈ [0,1]","figure_data":"","figure_id":"fig_6","figure_label":"","figure_type":"figure"},{"figure_caption":"ES [OPT(S)]= 𝜏 𝜋 𝑇 E S [OPT(S)] + 𝑇 -𝜏 𝜋 𝑇 E S [OPT(S)] ≤ 𝜏 𝜋 D (𝝁 𝜏 𝜋 |S) + (𝑇 -𝜏 𝜋 ) f ,","figure_data":"","figure_id":"fig_7","figure_label":"","figure_type":"figure"},{"figure_caption":"(Regret (𝜋 |S) = E S [OPT(S) -𝑅(𝜋 |S)](41a)≤ E S 𝜏 𝜋 D (𝝁 𝜏 𝜋 |S) + (𝑇 -𝜏 𝜋 ) f -𝜏 𝜋 ∑︁ 𝑡 =1 𝑓 (𝒙 𝑡 ) (41b) = E S 𝜏 𝜋 D (𝝁 𝜏 𝜋 |S) -𝜏 𝜋 ∑︁ 𝑡 =1 𝑓 (𝒙 𝑡 ) + E S [𝑇 -𝜏 𝜋 ] f (41c) ≤ 2( ḡ2 + Ḡ2 ) 𝜃 𝜂E[𝜏 𝜋 ] + 1 𝜂 𝑉 ℎ (𝝁, 𝝁 1 ) + ḡ 𝐺 + 𝐶 ℎ + ∥∇ℎ(𝜆 1 )∥ ∞ 𝜂𝐺 ,(41d)where the first inequality is from(40) and the second inequality is from Proposition 2 and Proposition 3. Therefore, the constants inTheorem 1 are 𝐶 1 = ḡ 𝐺 , 𝐶 2 = 2( C2 + b2 ) 𝜃 𝜂, and 𝐶 3 = 𝑉 ℎ (𝝁, 𝝁 1 ) + ḡ 𝐺 + 𝐶 ℎ + ∥ ∇ℎ (𝜆 1 ) ∥ ∞ 𝐺, respectively. Moreover, recall Remark 1, we choose ℎ(𝜆) := 1 2 ∥𝜆∥ 2 and dual iterates are bounded. Hence, we complete the proof of Theorem 1. ■B PROOF OF COST FEASIBILITYProposition 1 shows that a solution obtained using Algorithm 1 can not overspend, but may underspend. Based on the definition of subgradient ∇𝜇 𝑡 𝑘 , we have ∇ℎ 𝑘 (𝜇 1 ) -∇ℎ 𝑘 (𝜇 𝜏 𝜋 +1 ) 𝐺 𝑘 (1(𝜇 𝑘 ≥ 0) +𝛼 𝑘 1(𝜇 𝑘 < 0)) -𝑔 𝑘 (𝒙 𝑡 )) (42) Now, given that 1(𝜇 ≥ 0) + 𝛼 𝑘 1(𝜇 < 0) ≥ 𝛼 𝑘 for any 𝜇 ∈ R and that 𝜏 𝜋 ≤ 𝑇 by definition, we have𝜏 𝐴 ∑︁ 𝑡 =1 𝐺 𝑘 (1(𝜇 𝑘 ≥ 0) +𝛼 𝑘 1(𝜇 𝑘 < 0)) + (𝑇 -𝜏 𝐴 )𝛼 𝑘 𝑏 𝑘 ≥𝑇 𝛼 𝑘 𝑏 𝑘 . (43) Combining (42) and (43) and taking expectation, we get 𝑇 𝛼 𝑘 𝐺 𝑘 -E[","figure_data":"","figure_id":"fig_8","figure_label":"","figure_type":"figure"},{"figure_caption":"","figure_data":"Inter-channel Encoding Layer Interest-Aware Evolution LayerLinear Linear V KLinear Linear V K MHSALinear Linear V KMLP & Softmax Weighted Sum QQPCR EmbsConcate & Flatten MLP & Sigmoid (} Concate & MLP Target Item Prev Context Cand Context PCR Embs PCR Scores (c) CAR Model (b) PCR CAR Point-wise Scores Context-aware Scores | ) ( ) ( max{ arg t t t t t I x g x f x Optimal Item Selection Ranking-layerEncoding Layer Intra-channelMHSAMHSAMHSALinear Linear V KConcate&LinearConstraints Update Allocation-layerDuals Update Real-time Duals...Real-timeMulti-ChannelFinalExposuresCandidate ItemsItem ListCandidate Seq Channel-1Candidate Seq Channel-2Candidate Seq ... Channel-N Embedding Layer Behavior Seq UserItem TargetInfo ChannelProfile UserStreaming Data CenterCH-1 Upstream RSs ... CH-NReportRequestMHSAMulti-head Self-attentionScaled Dot-productEmbedding VectorWeightsEmbedding CrossingDot ProductConcateUser Device","figure_id":"tab_0","figure_label":"","figure_type":"table"},{"figure_caption":") 4.3.2 Context-Aware Refinement Model. In this section, we propose the CAR model to refine the point-wise utility scores estimated by PCR into context-aware utility scores. Given a candidate item set 𝐼 𝑐𝑎𝑛𝑑 with size 𝑁 , the aim of CAR is to optimally choose 𝐾 items from 𝐼 𝑐𝑎𝑛𝑑 and allocate them to the 𝐾 slots in a page based on the learning results from PCR.","figure_data":"","figure_id":"tab_2","figure_label":"","figure_type":"table"},{"figure_caption":"Comparison of ranking performance (bold: best; underline: runner-up).","figure_data":"DatasetMethodAUCLogloss NDCG@20 NDCG@30DIN0.68310.59220.54030.6535DIEN0.68420.59090.54080.6537DLCM0.68720.58980.55820.6698MicroVideo 3PRM0.69790.58720.55910.6708PEAR0.70210.58210.56320.6745Ours0.7084 0.57870.56670.6826DIN0.76810.49820.52030.6481DIEN0.76920.49710.52020.6479DLCM0.76990.49650.52090.6482TaobaoPRM0.77220.49410.52110.6489PEAR0.77480.49190.52320.6511STAR0.77380.49310.52190.6492DHANR 0.77530.49230.52430.6513Ours0.7791 0.48990.52750.6545","figure_id":"tab_3","figure_label":"1","figure_type":"table"},{"figure_caption":"Ablation study of the ranking components.","figure_data":"AUCLogloss NDCG@20 NDCG@30PCR *0.77580.49330.52220.6511PCR †0.77610.49320.52460.6513PCR w/o IntraCE0.77630.49280.52470.6516PCR w/o InterCE0.77560.49350.52390.6509PCR0.77780.49130.52620.6531PCR+CAR (propsed) 0.77910.48990.52750.6545","figure_id":"tab_4","figure_label":"2","figure_type":"table"},{"figure_caption":"Joint performance of allocation and ranking (bold: best; dagger: baseline).","figure_data":"Exposure completenessExp. Settings MethodCTRCTR LiftCH1CH2CH3CH4Fixed0.44% 1.35% 0.20% 0.50% 5.54% †-WPO0.15% 0.15% 0.13% 0.30% 6.09%+9.93%Setting-1HCA2E 0.24% 0.95% 0.33% 0.10% 6.34%+14.44%Ours0.02% 0.65% 0.67% 0.20% 6.56% * +18.41% *Fixed0.10% 0.53% 0.10% 0.20% 5.91% †-WPO0.16% 0.27% 1.40% 0.20% 6.28%+6.26%Setting-2HCA2E 0.19% 0.53% 0.30% 0.40% 6.53%+10.49%Ours0.17% 0.53% 0.40% 0.20% 6.76% * +14.38% *","figure_id":"tab_5","figure_label":"3","figure_type":"table"},{"figure_caption":"Results of online A/B tests.","figure_data":"CLICKCTRGMV Stay TimeOurs vs Fixed+4.02% +2.15% +1.98%+2.01%Ours vs Baseline +3.00% +1.75% +1.56%","figure_id":"tab_6","figure_label":"4","figure_type":"table"},{"figure_caption":"𝛾 𝑡 [𝑓 (𝒙 𝑡 )] ≥𝜏 𝜋 𝐷 ( μ𝜏 𝜋 ) -𝜏𝜋 𝑡 =1 𝝁 𝑡 𝜏 𝜋and the inequality is based on the fact that the dual function is convex. In this paper, we adopt 𝜃 -strongly convex function as the relation function ℎ(•) in mirror descents. According to Step. 2 of Proposition 8 in[25], we have 𝛾 𝑡 [𝑓 (𝒙 𝑡 )] = E","figure_data":"According to Step. 3 of Proposition 8 in [25], we haveE𝜏 𝜋 ∑︁E 𝜏 𝜋 ∑︁𝑡 =1𝑡 =1𝜏 𝜋𝜏 𝜋𝜏 𝜋∑︁ 𝑡 =1∑︁ 𝑡 =1 E 𝜏 𝜋 𝐷 (𝝁 𝑡 ) -∑︁ 𝑡 =1∑︁E 𝛾 𝑡 [𝑓 (𝒙 𝑡 ))],(35)𝑡 =1where μ𝜏 𝜋 =E𝑡 =1 𝜏 𝜋 ∑︁(∇𝝁 𝑡 ) 𝑇 (𝝁 𝑡 -𝝁) ≤𝜃 2( ḡ2 + Ḡ2 )𝜂E[𝜏 𝐴 ] +𝜂 𝑉 ℎ (𝜆, 𝜆 1 ). (36)Combining (36) with (35), we get(37)","figure_id":"tab_7","figure_label":"","figure_type":"table"},{"figure_caption":"≤ ∇ℎ 𝑘 (𝜇 1 ) -E[∇ℎ 𝑘 (𝜇 𝜏 𝜋 +1 )] 𝜂 + E[𝑇 -𝜏 𝐴 ]𝛼 𝑘 𝑏 𝑘","figure_data":"𝜏 𝜋∑︁𝑔 𝑘 (𝒙 𝑡 )](44a)𝑡 =1(44b)≤∥∇ℎ(𝜆 1 )∥ ∞ + 𝐶 ℎ𝐺 + 𝛼 𝑘 𝑏 𝑘+𝛼 𝑘 𝑏 𝑘ḡ,𝜂𝐺𝐺","figure_id":"tab_8","figure_label":"","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 2 :2Figure 2: An overview of the MIREC framework.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"10:\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 :3Figure 3: Online system architecture.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"CTR on different positions.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 4 :4Figure 4: Online Performance Analysis.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Proposition 3 .3Using a constant step-size rule 𝜂 > 0 for 𝑡 > 1 in Algorithm 1, it holdsE 𝜏 𝜋 𝐷 ( μ𝜏 𝜋 ) -𝜏 𝜋 ∑︁ 𝑡 =1 𝑓 (𝒙 𝑡 ) ≤ 2( ḡ2 + Ḡ2 ) 𝜃 𝜂E[𝜏 𝜋 ] + 1 𝜂 𝑉 ℎ (𝝁, 𝝁 1 ), (33)whereμ𝜏 𝜋 = 𝜏𝜋 𝑡 =1 𝝁 𝑡 𝜏 𝜋 .Proof. According to the definition of ∇𝝁 𝑡 and the subgradient inequality, we have(∇𝝁 𝑡 ) 𝑇 (𝝁 𝑡 -𝝁) ≥ 𝐷 (𝝁 𝑡 ) -𝐷 (𝝁) ≥ 𝐷 (𝝁 𝑡 ) -E 𝛾 𝑡 [𝜑 (𝝁)] + ∑︁ 𝑘 ∈ [𝐾 ]\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_5\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"F(𝒙 𝑡 , 𝜆) = (1 -𝜆)𝑓 (𝒙 𝑡 ) + 𝜆E S [𝑓 (𝒙 𝑡 )] G(𝒙 𝑡 , 𝜆) = (1 -𝜆)𝑔(𝒙 𝑡 ) + 𝜆E S [𝑔(𝒙 𝑡 )],where 𝜆 ∈ [0, 1] is the interpolation parameter. We defineOPT(S, 𝜆) = E S 𝑇 max 𝒙 𝑡 ,𝑡 ∈ [𝑇 ] 𝑇 𝑡 =1 𝐹 (𝒙 𝑡 , 𝜆) s.t. 𝑇 𝛼 ⊙ 𝐺 ≤ 𝑇 𝑡 =1 𝐺 (𝒙 𝑡 , 𝜆) ≤ 𝑇𝐺whereS 𝑇 := S × • • • × S isa product distribution of length 𝑇 . Now we give the definition of the new benchmark as OPT(S) := max 𝜆 ∈ [0,1]\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_6\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"ES [OPT(S)]= 𝜏 𝜋 𝑇 E S [OPT(S)] + 𝑇 -𝜏 𝜋 𝑇 E S [OPT(S)] ≤ 𝜏 𝜋 D (𝝁 𝜏 𝜋 |S) + (𝑇 -𝜏 𝜋 ) f ,\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_7\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"(Regret (𝜋 |S) = E S [OPT(S) -𝑅(𝜋 |S)](41a)≤ E S 𝜏 𝜋 D (𝝁 𝜏 𝜋 |S) + (𝑇 -𝜏 𝜋 ) f -𝜏 𝜋 ∑︁ 𝑡 =1 𝑓 (𝒙 𝑡 ) (41b) = E S 𝜏 𝜋 D (𝝁 𝜏 𝜋 |S) -𝜏 𝜋 ∑︁ 𝑡 =1 𝑓 (𝒙 𝑡 ) + E S [𝑇 -𝜏 𝜋 ] f (41c) ≤ 2( ḡ2 + Ḡ2 ) 𝜃 𝜂E[𝜏 𝜋 ] + 1 𝜂 𝑉 ℎ (𝝁, 𝝁 1 ) + ḡ 𝐺 + 𝐶 ℎ + ∥∇ℎ(𝜆 1 )∥ ∞ 𝜂𝐺 ,(41d)where the first inequality is from(40) and the second inequality is from Proposition 2 and Proposition 3. Therefore, the constants inTheorem 1 are 𝐶 1 = ḡ 𝐺 , 𝐶 2 = 2( C2 + b2 ) 𝜃 𝜂, and 𝐶 3 = 𝑉 ℎ (𝝁, 𝝁 1 ) + ḡ 𝐺 + 𝐶 ℎ + ∥ ∇ℎ (𝜆 1 ) ∥ ∞ 𝐺, respectively. Moreover, recall Remark 1, we choose ℎ(𝜆) := 1 2 ∥𝜆∥ 2 and dual iterates are bounded. Hence, we complete the proof of Theorem 1. ■B PROOF OF COST FEASIBILITYProposition 1 shows that a solution obtained using Algorithm 1 can not overspend, but may underspend. Based on the definition of subgradient ∇𝜇 𝑡 𝑘 , we have ∇ℎ 𝑘 (𝜇 1 ) -∇ℎ 𝑘 (𝜇 𝜏 𝜋 +1 ) 𝐺 𝑘 (1(𝜇 𝑘 ≥ 0) +𝛼 𝑘 1(𝜇 𝑘 < 0)) -𝑔 𝑘 (𝒙 𝑡 )) (42) Now, given that 1(𝜇 ≥ 0) + 𝛼 𝑘 1(𝜇 < 0) ≥ 𝛼 𝑘 for any 𝜇 ∈ R and that 𝜏 𝜋 ≤ 𝑇 by definition, we have𝜏 𝐴 ∑︁ 𝑡 =1 𝐺 𝑘 (1(𝜇 𝑘 ≥ 0) +𝛼 𝑘 1(𝜇 𝑘 < 0)) + (𝑇 -𝜏 𝐴 )𝛼 𝑘 𝑏 𝑘 ≥𝑇 𝛼 𝑘 𝑏 𝑘 . (43) Combining (42) and (43) and taking expectation, we get 𝑇 𝛼 𝑘 𝐺 𝑘 -E[\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_8\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"\",\n \"figure_data\": \"Inter-channel Encoding Layer Interest-Aware Evolution LayerLinear Linear V KLinear Linear V K MHSALinear Linear V KMLP & Softmax Weighted Sum QQPCR EmbsConcate & Flatten MLP & Sigmoid (} Concate & MLP Target Item Prev Context Cand Context PCR Embs PCR Scores (c) CAR Model (b) PCR CAR Point-wise Scores Context-aware Scores | ) ( ) ( max{ arg t t t t t I x g x f x Optimal Item Selection Ranking-layerEncoding Layer Intra-channelMHSAMHSAMHSALinear Linear V KConcate&LinearConstraints Update Allocation-layerDuals Update Real-time Duals...Real-timeMulti-ChannelFinalExposuresCandidate ItemsItem ListCandidate Seq Channel-1Candidate Seq Channel-2Candidate Seq ... Channel-N Embedding Layer Behavior Seq UserItem TargetInfo ChannelProfile UserStreaming Data CenterCH-1 Upstream RSs ... CH-NReportRequestMHSAMulti-head Self-attentionScaled Dot-productEmbedding VectorWeightsEmbedding CrossingDot ProductConcateUser Device\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \") 4.3.2 Context-Aware Refinement Model. In this section, we propose the CAR model to refine the point-wise utility scores estimated by PCR into context-aware utility scores. Given a candidate item set 𝐼 𝑐𝑎𝑛𝑑 with size 𝑁 , the aim of CAR is to optimally choose 𝐾 items from 𝐼 𝑐𝑎𝑛𝑑 and allocate them to the 𝐾 slots in a page based on the learning results from PCR.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Comparison of ranking performance (bold: best; underline: runner-up).\",\n \"figure_data\": \"DatasetMethodAUCLogloss NDCG@20 NDCG@30DIN0.68310.59220.54030.6535DIEN0.68420.59090.54080.6537DLCM0.68720.58980.55820.6698MicroVideo 3PRM0.69790.58720.55910.6708PEAR0.70210.58210.56320.6745Ours0.7084 0.57870.56670.6826DIN0.76810.49820.52030.6481DIEN0.76920.49710.52020.6479DLCM0.76990.49650.52090.6482TaobaoPRM0.77220.49410.52110.6489PEAR0.77480.49190.52320.6511STAR0.77380.49310.52190.6492DHANR 0.77530.49230.52430.6513Ours0.7791 0.48990.52750.6545\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Ablation study of the ranking components.\",\n \"figure_data\": \"AUCLogloss NDCG@20 NDCG@30PCR *0.77580.49330.52220.6511PCR †0.77610.49320.52460.6513PCR w/o IntraCE0.77630.49280.52470.6516PCR w/o InterCE0.77560.49350.52390.6509PCR0.77780.49130.52620.6531PCR+CAR (propsed) 0.77910.48990.52750.6545\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Joint performance of allocation and ranking (bold: best; dagger: baseline).\",\n \"figure_data\": \"Exposure completenessExp. Settings MethodCTRCTR LiftCH1CH2CH3CH4Fixed0.44% 1.35% 0.20% 0.50% 5.54% †-WPO0.15% 0.15% 0.13% 0.30% 6.09%+9.93%Setting-1HCA2E 0.24% 0.95% 0.33% 0.10% 6.34%+14.44%Ours0.02% 0.65% 0.67% 0.20% 6.56% * +18.41% *Fixed0.10% 0.53% 0.10% 0.20% 5.91% †-WPO0.16% 0.27% 1.40% 0.20% 6.28%+6.26%Setting-2HCA2E 0.19% 0.53% 0.30% 0.40% 6.53%+10.49%Ours0.17% 0.53% 0.40% 0.20% 6.76% * +14.38% *\",\n \"figure_id\": \"tab_5\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Results of online A/B tests.\",\n \"figure_data\": \"CLICKCTRGMV Stay TimeOurs vs Fixed+4.02% +2.15% +1.98%+2.01%Ours vs Baseline +3.00% +1.75% +1.56%\",\n \"figure_id\": \"tab_6\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"𝛾 𝑡 [𝑓 (𝒙 𝑡 )] ≥𝜏 𝜋 𝐷 ( μ𝜏 𝜋 ) -𝜏𝜋 𝑡 =1 𝝁 𝑡 𝜏 𝜋and the inequality is based on the fact that the dual function is convex. In this paper, we adopt 𝜃 -strongly convex function as the relation function ℎ(•) in mirror descents. According to Step. 2 of Proposition 8 in[25], we have 𝛾 𝑡 [𝑓 (𝒙 𝑡 )] = E\",\n \"figure_data\": \"According to Step. 3 of Proposition 8 in [25], we haveE𝜏 𝜋 ∑︁E 𝜏 𝜋 ∑︁𝑡 =1𝑡 =1𝜏 𝜋𝜏 𝜋𝜏 𝜋∑︁ 𝑡 =1∑︁ 𝑡 =1 E 𝜏 𝜋 𝐷 (𝝁 𝑡 ) -∑︁ 𝑡 =1∑︁E 𝛾 𝑡 [𝑓 (𝒙 𝑡 ))],(35)𝑡 =1where μ𝜏 𝜋 =E𝑡 =1 𝜏 𝜋 ∑︁(∇𝝁 𝑡 ) 𝑇 (𝝁 𝑡 -𝝁) ≤𝜃 2( ḡ2 + Ḡ2 )𝜂E[𝜏 𝐴 ] +𝜂 𝑉 ℎ (𝜆, 𝜆 1 ). (36)Combining (36) with (35), we get(37)\",\n \"figure_id\": \"tab_7\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"≤ ∇ℎ 𝑘 (𝜇 1 ) -E[∇ℎ 𝑘 (𝜇 𝜏 𝜋 +1 )] 𝜂 + E[𝑇 -𝜏 𝐴 ]𝛼 𝑘 𝑏 𝑘\",\n \"figure_data\": \"𝜏 𝜋∑︁𝑔 𝑘 (𝒙 𝑡 )](44a)𝑡 =1(44b)≤∥∇ℎ(𝜆 1 )∥ ∞ + 𝐶 ℎ𝐺 + 𝛼 𝑘 𝑏 𝑘+𝛼 𝑘 𝑏 𝑘ḡ,𝜂𝐺𝐺\",\n \"figure_id\": \"tab_8\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n }\n]"},"authors":{"kind":"string","value":"Yue Xu; Qijie Shen; Dimin Wang; Hao Chen; Lixiang Lai; Tao Zhuang"},"references":{"kind":"list like","value":[{"authors":"Qingyao Ai; Keping Bi; Jiafeng Guo; Bruce Croft","journal":"","ref_id":"b0","title":"Learning a deep listwise context model for ranking refinement","year":"2018"},{"authors":"Ashwinkumar Badanidiyuru; Robert Kleinberg; Aleksandrs Slivkins","journal":"Journal of the ACM (JACM)","ref_id":"b1","title":"Bandits with knapsacks","year":"2018"},{"authors":"Haihao Santiago R Balseiro; Vahab Lu; Mirrokni","journal":"Operations Research","ref_id":"b2","title":"The best of many worlds: Dual mirror descent for online allocation problems","year":"2022"},{"authors":"Irwan Bello; Sayali Kulkarni; Sagar Jain; Craig Boutilier; Ed Chi; Elad Eban; Xiyang Luo; Alan Mackey; Ofer Meshi","journal":"","ref_id":"b3","title":"Seq2slate: Re-ranking and slate optimization with rnns","year":"2018"},{"authors":"Dimitri P Bertsekas","journal":"Journal of the Operational Research Society","ref_id":"b4","title":"Nonlinear programming","year":"1997"},{"authors":"Dagui Chen; Qi Yan; Chunjie Chen; Zhenzhe Zheng; Yangsu Liu; Zhenjia Ma; Chuan Yu; Jian Xu; Bo Zheng","journal":"","ref_id":"b5","title":"Hierarchically constrained adaptive ad exposure in feeds","year":"2022"},{"authors":"Qiwei Chen; Yue Xu; Changhua Pei; Shanshan Lv; Tao Zhuang; Junfeng Ge","journal":"","ref_id":"b6","title":"Efficient long sequential user data modeling for click-through rate prediction","year":"2022"},{"authors":"Xusong Chen; Dong Liu; Zheng-Jun Zha; Wengang Zhou; Zhiwei Xiong; Yan Li","journal":"MM","ref_id":"b7","title":"Temporal hierarchical attention at category-and item-level for micro-video click-through prediction","year":"2018"},{"authors":"Yuwei Chen; Zengde Deng; Yinzhi Zhou; Zaiyi Chen; Yujie Chen; Haoyuan Hu","journal":"IEEE","ref_id":"b8","title":"An online algorithm for chance constrained resource allocation","year":"2023"},{"authors":"Heng-Tze Cheng; Levent Koc; Jeremiah Harmsen; Tal Shaked; Tushar Chandra; Hrishi Aradhye; Glen Anderson; Greg Corrado; Wei Chai; Mustafa Ispir","journal":"","ref_id":"b9","title":"Wide & deep learning for recommender systems","year":"2016"},{"authors":"Yufei Feng; Binbin Hu; Yu Gong; Fei Sun; Qingwen Liu; Wenwu Ou","journal":"","ref_id":"b10","title":"GRN: Generative rerank network for context-wise recommendation","year":"2021"},{"authors":"Jyotirmoy Gope; Sanjay Kumar; Jain ","journal":"IEEE","ref_id":"b11","title":"A survey on solving cold start problem in recommender systems","year":"2017"},{"authors":"Qi Hao; Tianze Luo; Guangda Huzhang","journal":"","ref_id":"b12","title":"Re-ranking with constraints on diversified exposures for homepage recommender system","year":"2021"},{"authors":"Xiaotian Hao; Zhaoqing Peng; Yi Ma; Guan Wang; Junqi Jin; Jianye Hao; Shan Chen; Rongquan Bai; Mingzhou Xie; Miao Xu","journal":"","ref_id":"b13","title":"Dynamic knapsack optimization towards efficient multi-channel sequential advertising","year":"2020"},{"authors":"Elad Hazan","journal":"Foundations and Trends® in Optimization","ref_id":"b14","title":"Introduction to online convex optimization","year":"2016"},{"authors":"Jinhong Huang; Yang Li; Shan Sun; Bufeng Zhang; Jin Huang","journal":"","ref_id":"b15","title":"Personalized flight itinerary ranking at fliggy","year":"2020"},{"authors":"Yanhua Huang; Weikun Wang; Lei Zhang; Ruiwen Xu","journal":"","ref_id":"b16","title":"Sliding spectrum decomposition for diversified recommendation","year":"2021"},{"authors":"Nicole Immorlica; Karthik Abinav Sankararaman; Robert Schapire; Aleksandrs Slivkins","journal":"","ref_id":"b17","title":"Adversarial bandits with knapsacks","year":"2019"},{"authors":"Kalervo Järvelin; Jaana Kekäläinen","journal":"","ref_id":"b18","title":"IR evaluation methods for retrieving highly relevant documents","year":"2017"},{"authors":"Yi Li; Jieming Zhu; Weiwen Liu; Liangcai Su; Guohao Cai; Qi Zhang; Ruiming Tang; Xi Xiao; Xiuqiang He","journal":"","ref_id":"b19","title":"Pear: Personalized re-ranking with contextualized transformer for recommendation","year":"2022"},{"authors":"Guogang Liao; Ze Wang; Xiaoxu Wu; Xiaowen Shi; Chuheng Zhang; Yongkang Wang; Xingxing Wang; Dong Wang","journal":"","ref_id":"b20","title":"Cross DQN: Cross deep Q network for ads allocation in feed","year":"2022"},{"authors":"Zhuoyi Lin; Sheng Zang; Rundong Wang; Zhu Sun; Chi Xu; Chee-Keong Kwoh","journal":"","ref_id":"b21","title":"Attention over self-attention: Intention-aware re-ranking with dynamic transformer encoders for recommendation","year":"2022"},{"authors":"Zihan Lin; Hui Wang; Jingshu Mao; Wayne Xin Zhao; Cheng Wang; Peng Jiang; Ji-Rong Wen","journal":"","ref_id":"b22","title":"Feature-aware diversified re-ranking with disentangled representations for relevant recommendation","year":"2022"},{"authors":"Weiwen Liu; Yunjia Xi; Jiarui Qin; Fei Sun; Bo Chen; Weinan Zhang; Rui Zhang; Ruiming Tang","journal":"","ref_id":"b23","title":"Neural re-ranking in multi-stage recommender systems: A review","year":"2020"},{"authors":"Alfonso Lobos; Paul Grigas; Zheng Wen","journal":"","ref_id":"b24","title":"Joint online learning and decisionmaking via dual mirror descent","year":"2021"},{"authors":"Xingyu Lu; Qintong Wu; Wenliang Zhong","journal":"PMLR","ref_id":"b25","title":"Multi-slots online matching with high entropy","year":"2022"},{"authors":"Changhua Pei; Yi Zhang; Yongfeng Zhang; Fei Sun; Xiao Lin; Hanxiao Sun; Jian Wu; Peng Jiang; Junfeng Ge; Wenwu Ou","journal":"","ref_id":"b26","title":"Personalized re-ranking for recommendation","year":"2019"},{"authors":"Qi Pi; Weijie Bian; Guorui Zhou; Xiaoqiang Zhu; Kun Gai","journal":"","ref_id":"b27","title":"Practice on long sequential user behavior modeling for click-through rate prediction","year":"2019"},{"authors":"Xufeng Qian; Yue Xu; Fuyu Lv; Shengyu Zhang; Ziwen Jiang; Qingwen Liu; Xiaoyi Zeng; Tat-Seng Chua; Fei Wu","journal":"","ref_id":"b28","title":"Intelligent request strategy design in recommender system","year":"2022"},{"authors":"Xiang-Rong Sheng; Liqin Zhao; Guorui Zhou; Xinyao Ding; Binding Dai; Qiang Luo; Siran Yang; Jingshan Lv; Chi Zhang; Hongbo Deng","journal":"","ref_id":"b29","title":"One model to serve all: Star topology adaptive recommender for multi-domain ctr prediction","year":"2021"},{"authors":"Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Łukasz Kaiser; Illia Polosukhin","journal":"Advances in neural information processing systems","ref_id":"b30","title":"Attention is all you need","year":"2017"},{"authors":"Fan Wang; Xiaomin Fang; Lihang Liu; Yaxue Chen; Jiucheng Tao; Zhiming Peng; Cihang Jin; Hao Tian","journal":"","ref_id":"b31","title":"Sequential evaluation and generation framework for combinatorial recommender system","year":"2019"},{"authors":"Ruobing Xie; Shaoliang Zhang; Rui Wang; Feng Xia; Leyu Lin","journal":"","ref_id":"b32","title":"Hierarchical reinforcement learning for integrated recommendation","year":"2021"},{"authors":"Biao Yuan; Zengde Deng; Na Geng; Yujie Chen; Haoyuan Hu","journal":"INFORMS Journal on Applied Analytics","ref_id":"b33","title":"Practice summary: Cainiao optimizes the fulfillment routes of parcels","year":"2023"},{"authors":"Jianjun Yuan; Andrew Lamperski","journal":"","ref_id":"b34","title":"Online convex optimization for cumulative constraints","year":"2018"},{"authors":"Weiru Zhang; Chao Wei; Xiaonan Meng; Yi Hu; Hao Wang","journal":"","ref_id":"b35","title":"The whole-page optimization via dynamic ad allocation","year":"2018"},{"authors":"Xiangyu Zhao; Changsheng Gu; Haoshenglun Zhang; Xiwang Yang; Xiaobing Liu; Jiliang Tang; Hui Liu","journal":"","ref_id":"b36","title":"Dear: Deep reinforcement learning for online advertising impression in recommender systems","year":"2021"},{"authors":"Guorui Zhou; Na Mou; Ying Fan; Qi Pi; Weijie Bian; Chang Zhou; Xiaoqiang Zhu; Kun Gai","journal":"","ref_id":"b37","title":"Deep interest evolution network for click-through rate prediction","year":"2019"},{"authors":"Guorui Zhou; Xiaoqiang Zhu; Chenru Song; Ying Fan; Han Zhu; Xiao Ma; Yanghui Yan; Junqi Jin; Han Li; Kun Gai","journal":"","ref_id":"b38","title":"Deep interest network for click-through rate prediction","year":"2018"},{"authors":"Yu-Hang Zhou; Peng Hu; Chen Liang; Huan Xu; Guangda Huzhang; Yinfu Feng; Qing Da; Xinshang Wang; An-Xiang Zeng","journal":"","ref_id":"b39","title":"A primal-dual online algorithm for online matching problem in dynamic environments","year":"2021"},{"authors":"Tao Zhuang; Wenwu Ou; Zhirong Wang","journal":"","ref_id":"b40","title":"Globally optimized mutual influence aware ranking in e-commerce search","year":"2018"}],"string":"[\n {\n \"authors\": \"Qingyao Ai; Keping Bi; Jiafeng Guo; Bruce Croft\",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"Learning a deep listwise context model for ranking refinement\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Ashwinkumar Badanidiyuru; Robert Kleinberg; Aleksandrs Slivkins\",\n \"journal\": \"Journal of the ACM (JACM)\",\n \"ref_id\": \"b1\",\n \"title\": \"Bandits with knapsacks\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Haihao Santiago R Balseiro; Vahab Lu; Mirrokni\",\n \"journal\": \"Operations Research\",\n \"ref_id\": \"b2\",\n \"title\": \"The best of many worlds: Dual mirror descent for online allocation problems\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Irwan Bello; Sayali Kulkarni; Sagar Jain; Craig Boutilier; Ed Chi; Elad Eban; Xiyang Luo; Alan Mackey; Ofer Meshi\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"Seq2slate: Re-ranking and slate optimization with rnns\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Dimitri P Bertsekas\",\n \"journal\": \"Journal of the Operational Research Society\",\n \"ref_id\": \"b4\",\n \"title\": \"Nonlinear programming\",\n \"year\": \"1997\"\n },\n {\n \"authors\": \"Dagui Chen; Qi Yan; Chunjie Chen; Zhenzhe Zheng; Yangsu Liu; Zhenjia Ma; Chuan Yu; Jian Xu; Bo Zheng\",\n \"journal\": \"\",\n \"ref_id\": \"b5\",\n \"title\": \"Hierarchically constrained adaptive ad exposure in feeds\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Qiwei Chen; Yue Xu; Changhua Pei; Shanshan Lv; Tao Zhuang; Junfeng Ge\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"Efficient long sequential user data modeling for click-through rate prediction\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Xusong Chen; Dong Liu; Zheng-Jun Zha; Wengang Zhou; Zhiwei Xiong; Yan Li\",\n \"journal\": \"MM\",\n \"ref_id\": \"b7\",\n \"title\": \"Temporal hierarchical attention at category-and item-level for micro-video click-through prediction\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Yuwei Chen; Zengde Deng; Yinzhi Zhou; Zaiyi Chen; Yujie Chen; Haoyuan Hu\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b8\",\n \"title\": \"An online algorithm for chance constrained resource allocation\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Heng-Tze Cheng; Levent Koc; Jeremiah Harmsen; Tal Shaked; Tushar Chandra; Hrishi Aradhye; Glen Anderson; Greg Corrado; Wei Chai; Mustafa Ispir\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"Wide & deep learning for recommender systems\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Yufei Feng; Binbin Hu; Yu Gong; Fei Sun; Qingwen Liu; Wenwu Ou\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"GRN: Generative rerank network for context-wise recommendation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Jyotirmoy Gope; Sanjay Kumar; Jain \",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b11\",\n \"title\": \"A survey on solving cold start problem in recommender systems\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Qi Hao; Tianze Luo; Guangda Huzhang\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"Re-ranking with constraints on diversified exposures for homepage recommender system\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Xiaotian Hao; Zhaoqing Peng; Yi Ma; Guan Wang; Junqi Jin; Jianye Hao; Shan Chen; Rongquan Bai; Mingzhou Xie; Miao Xu\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"Dynamic knapsack optimization towards efficient multi-channel sequential advertising\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Elad Hazan\",\n \"journal\": \"Foundations and Trends® in Optimization\",\n \"ref_id\": \"b14\",\n \"title\": \"Introduction to online convex optimization\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Jinhong Huang; Yang Li; Shan Sun; Bufeng Zhang; Jin Huang\",\n \"journal\": \"\",\n \"ref_id\": \"b15\",\n \"title\": \"Personalized flight itinerary ranking at fliggy\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Yanhua Huang; Weikun Wang; Lei Zhang; Ruiwen Xu\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"Sliding spectrum decomposition for diversified recommendation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Nicole Immorlica; Karthik Abinav Sankararaman; Robert Schapire; Aleksandrs Slivkins\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"Adversarial bandits with knapsacks\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Kalervo Järvelin; Jaana Kekäläinen\",\n \"journal\": \"\",\n \"ref_id\": \"b18\",\n \"title\": \"IR evaluation methods for retrieving highly relevant documents\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Yi Li; Jieming Zhu; Weiwen Liu; Liangcai Su; Guohao Cai; Qi Zhang; Ruiming Tang; Xi Xiao; Xiuqiang He\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"Pear: Personalized re-ranking with contextualized transformer for recommendation\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Guogang Liao; Ze Wang; Xiaoxu Wu; Xiaowen Shi; Chuheng Zhang; Yongkang Wang; Xingxing Wang; Dong Wang\",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"Cross DQN: Cross deep Q network for ads allocation in feed\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Zhuoyi Lin; Sheng Zang; Rundong Wang; Zhu Sun; Chi Xu; Chee-Keong Kwoh\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"Attention over self-attention: Intention-aware re-ranking with dynamic transformer encoders for recommendation\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Zihan Lin; Hui Wang; Jingshu Mao; Wayne Xin Zhao; Cheng Wang; Peng Jiang; Ji-Rong Wen\",\n \"journal\": \"\",\n \"ref_id\": \"b22\",\n \"title\": \"Feature-aware diversified re-ranking with disentangled representations for relevant recommendation\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Weiwen Liu; Yunjia Xi; Jiarui Qin; Fei Sun; Bo Chen; Weinan Zhang; Rui Zhang; Ruiming Tang\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"Neural re-ranking in multi-stage recommender systems: A review\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Alfonso Lobos; Paul Grigas; Zheng Wen\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Joint online learning and decisionmaking via dual mirror descent\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Xingyu Lu; Qintong Wu; Wenliang Zhong\",\n \"journal\": \"PMLR\",\n \"ref_id\": \"b25\",\n \"title\": \"Multi-slots online matching with high entropy\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Changhua Pei; Yi Zhang; Yongfeng Zhang; Fei Sun; Xiao Lin; Hanxiao Sun; Jian Wu; Peng Jiang; Junfeng Ge; Wenwu Ou\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"Personalized re-ranking for recommendation\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Qi Pi; Weijie Bian; Guorui Zhou; Xiaoqiang Zhu; Kun Gai\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"Practice on long sequential user behavior modeling for click-through rate prediction\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Xufeng Qian; Yue Xu; Fuyu Lv; Shengyu Zhang; Ziwen Jiang; Qingwen Liu; Xiaoyi Zeng; Tat-Seng Chua; Fei Wu\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"Intelligent request strategy design in recommender system\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Xiang-Rong Sheng; Liqin Zhao; Guorui Zhou; Xinyao Ding; Binding Dai; Qiang Luo; Siran Yang; Jingshan Lv; Chi Zhang; Hongbo Deng\",\n \"journal\": \"\",\n \"ref_id\": \"b29\",\n \"title\": \"One model to serve all: Star topology adaptive recommender for multi-domain ctr prediction\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Łukasz Kaiser; Illia Polosukhin\",\n \"journal\": \"Advances in neural information processing systems\",\n \"ref_id\": \"b30\",\n \"title\": \"Attention is all you need\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Fan Wang; Xiaomin Fang; Lihang Liu; Yaxue Chen; Jiucheng Tao; Zhiming Peng; Cihang Jin; Hao Tian\",\n \"journal\": \"\",\n \"ref_id\": \"b31\",\n \"title\": \"Sequential evaluation and generation framework for combinatorial recommender system\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Ruobing Xie; Shaoliang Zhang; Rui Wang; Feng Xia; Leyu Lin\",\n \"journal\": \"\",\n \"ref_id\": \"b32\",\n \"title\": \"Hierarchical reinforcement learning for integrated recommendation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Biao Yuan; Zengde Deng; Na Geng; Yujie Chen; Haoyuan Hu\",\n \"journal\": \"INFORMS Journal on Applied Analytics\",\n \"ref_id\": \"b33\",\n \"title\": \"Practice summary: Cainiao optimizes the fulfillment routes of parcels\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Jianjun Yuan; Andrew Lamperski\",\n \"journal\": \"\",\n \"ref_id\": \"b34\",\n \"title\": \"Online convex optimization for cumulative constraints\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Weiru Zhang; Chao Wei; Xiaonan Meng; Yi Hu; Hao Wang\",\n \"journal\": \"\",\n \"ref_id\": \"b35\",\n \"title\": \"The whole-page optimization via dynamic ad allocation\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Xiangyu Zhao; Changsheng Gu; Haoshenglun Zhang; Xiwang Yang; Xiaobing Liu; Jiliang Tang; Hui Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b36\",\n \"title\": \"Dear: Deep reinforcement learning for online advertising impression in recommender systems\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Guorui Zhou; Na Mou; Ying Fan; Qi Pi; Weijie Bian; Chang Zhou; Xiaoqiang Zhu; Kun Gai\",\n \"journal\": \"\",\n \"ref_id\": \"b37\",\n \"title\": \"Deep interest evolution network for click-through rate prediction\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Guorui Zhou; Xiaoqiang Zhu; Chenru Song; Ying Fan; Han Zhu; Xiao Ma; Yanghui Yan; Junqi Jin; Han Li; Kun Gai\",\n \"journal\": \"\",\n \"ref_id\": \"b38\",\n \"title\": \"Deep interest network for click-through rate prediction\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Yu-Hang Zhou; Peng Hu; Chen Liang; Huan Xu; Guangda Huzhang; Yinfu Feng; Qing Da; Xinshang Wang; An-Xiang Zeng\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"A primal-dual online algorithm for online matching problem in dynamic environments\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Tao Zhuang; Wenwu Ou; Zhirong Wang\",\n \"journal\": \"\",\n \"ref_id\": \"b40\",\n \"title\": \"Globally optimized mutual influence aware ranking in e-commerce search\",\n \"year\": \"2018\"\n }\n]"},"formulas":{"kind":"list like","value":[{"formula_coordinates":[3,94.41,394.83,199.64,19.41],"formula_id":"formula_0","formula_text":"X 𝑡 = 𝑖 ∈𝐼 𝑡 𝑥 𝑡,𝑛,𝑖 = 1, ∀𝑡 ∈ T , ∀𝑛 ∈ N 𝑛 𝑥 𝑡,𝑛,𝑖 ≤ 1, ∀𝑡 ∈ T , ∀𝑖 ∈ 𝐼,(1)"},{"formula_coordinates":[3,83.4,649.45,210.64,17.89],"formula_id":"formula_1","formula_text":"P 0 : OPT(S) = max 𝒙 𝑡 ∈X 𝑡 ∑︁ 𝑇 𝑡 =1 𝑓 (𝒙 𝑡 )(2)"},{"formula_coordinates":[3,87.11,672.85,206.94,15.89],"formula_id":"formula_2","formula_text":"s.t. 𝐶 1 : ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≤ 𝐺 max 𝑚,𝑡ℎ 𝑁 (S), ∀𝑚 ∈ M,(3)"},{"formula_coordinates":[3,106.4,692.7,187.64,15.89],"formula_id":"formula_3","formula_text":"𝐶 2 : ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≥ 𝐺 min 𝑚,𝑡ℎ 𝑁 (S), ∀𝑚 ∈ M,(4)"},{"formula_coordinates":[4,119.42,361.16,174.62,39.44],"formula_id":"formula_4","formula_text":"+ ∑︁ 𝑚 𝜆 𝑚 ∑︁ 𝑡 𝑔 𝑚 (𝒙 𝑡 ) -𝐺 min 𝑚 .(5)"},{"formula_coordinates":[4,62.85,509.68,231.19,78.43],"formula_id":"formula_5","formula_text":"min 𝝁 𝐷 (𝝁) = ∑︁ 𝑡 𝑓 (𝒙 𝑡 ) - ∑︁ 𝑚 [𝜇 𝑚 ] + ∑︁ 𝑡 𝑔 𝑚 (𝒙 𝑡 ) -𝐺 max 𝑚 (6a) + ∑︁ 𝑚 [-𝜇 𝑚 ] + ∑︁ 𝑡 𝑔 𝑚 (𝒙 𝑡 ) -𝐺 min 𝑚 (6b) = ∑︁ 𝑡 𝑓 (𝒙 𝑡 ) - ∑︁ 𝑚 𝜇 𝑚 𝑔 𝑚 (𝒙 𝑡 ) + ∑︁ 𝑚 [𝜇 𝑚 ] + 𝐺 max 𝑚 -[-𝜇 𝑚 ] + 𝐺 min 𝑚 ,(6c)"},{"formula_coordinates":[4,79.05,596.72,77.01,8.79],"formula_id":"formula_6","formula_text":"[𝜇 𝑚 ] + = max{𝜇 𝑚 , 0}."},{"formula_coordinates":[4,103.58,695.94,190.46,15.08],"formula_id":"formula_7","formula_text":"P 1 : x𝑡 = arg max 𝒙 𝑡 ∈X 𝑓 (𝒙 𝑡 ) -𝝁 𝑇 𝑡 𝑔(𝒙 𝑡 ) .(7)"},{"formula_coordinates":[4,373.66,400.35,184.54,10.98],"formula_id":"formula_8","formula_text":"𝐺 max 𝑚,𝑡 +1 = 𝐺 max 𝑚,𝑡 -𝑔 𝑚 (𝒙 𝑡 ), ∀𝑚 ∈ M.(8)"},{"formula_coordinates":[4,322.53,442.11,229.33,12.25],"formula_id":"formula_9","formula_text":"∇𝜇 𝑚,𝑡 = -𝑔 𝑚 (𝒙 𝑡 ) +𝐺 max 𝑚,𝑡 +1 • 1(𝜇 𝑚,𝑡 ≥ 0) +𝐺 min 𝑚 • 1(𝜇 𝑚,𝑡 ≤ 0), ("},{"formula_coordinates":[4,356.03,495.43,202.17,17.62],"formula_id":"formula_10","formula_text":"𝜇 𝑚,𝑡 +1 = arg min 𝜇 𝑚 ∈R 𝜇 𝑚 ∇𝜇 𝑚,𝑡 + 1 𝜂 𝑉 ℎ (𝜇 𝑚 , 𝜇 𝑚,𝑡 ),(10)"},{"formula_coordinates":[4,398.84,588.64,159.36,8.79],"formula_id":"formula_11","formula_text":"𝝁 𝑡 +1 = [𝝁 𝑡 -𝜂∇𝝁 𝑡 ] + ,(11)"},{"formula_coordinates":[5,90.65,378.9,203.4,9.43],"formula_id":"formula_12","formula_text":"Regret(𝜋 |S) = sup E S [OPT(S) -𝑅(𝜋 |S)] ,(12)"},{"formula_coordinates":[5,53.8,675.97,212.23,11.72],"formula_id":"formula_13","formula_text":"channel 𝑚 ∈ M that 𝐺 min 𝑚 -E 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≤ 𝐶 4 + 𝐶 5"},{"formula_coordinates":[5,367.74,197.82,140.75,15.08],"formula_id":"formula_14","formula_text":"P 1 : x𝑡 = arg max 𝒙 𝑡 ∈X 𝑓 (𝒙 𝑡 ) -𝝁 𝑇 𝑡 𝑔(𝒙 𝑡 ) ."},{"formula_coordinates":[6,86.07,135.89,207.98,10.42],"formula_id":"formula_15","formula_text":"𝑉 𝑚 𝑙 = [ℎ𝑒𝑎𝑑 1 , ℎ𝑒𝑎𝑑 2 , ..., ℎ𝑒𝑎𝑑 ℎ ]𝑊 𝑂 ,(13a)"},{"formula_coordinates":[6,78.94,152.44,215.11,23.78],"formula_id":"formula_16","formula_text":"ℎ𝑒𝑎𝑑 𝑖 = Softmax (𝐸 𝑏 𝑊 𝑄 )(𝐸 𝑏 𝑊 𝐾 ) 𝑇 √︁ 𝑑 ℎ /ℎ (𝐸 𝑏 𝑊 𝑉 ) ,(13b)"},{"formula_coordinates":[6,83.55,433.83,207.08,14.69],"formula_id":"formula_17","formula_text":"𝑉 𝑈 = 𝑓 (𝐸 𝑏 ; 𝐸 𝑖 ) = ∑︁ 𝐵 𝑖=1 𝐴(𝑏 𝑖 , 𝐸 𝑖 )𝑏 𝑖 = ∑︁ 𝐵 𝑖=1 𝑤 𝑖 𝑏 𝑖 , (14"},{"formula_coordinates":[6,290.62,439.44,3.42,4.09],"formula_id":"formula_18","formula_text":")"},{"formula_coordinates":[6,59.33,535.71,234.72,32.96],"formula_id":"formula_19","formula_text":"𝐻 𝑖 𝑠 = Softmax (𝑉 𝑚 𝑙 𝑊 𝑞 ) [𝑉 𝑈 𝑊 𝑘1 , 𝑉 𝑚 𝑙 𝑊 𝑘2 ] 𝑇 √︁ 𝑑 ℎ [𝑉 𝑈 𝑊 𝑣1 , 𝑉 𝑚 𝑙 𝑊 𝑣2 ],(15)"},{"formula_coordinates":[6,114.79,700.48,179.25,9.16],"formula_id":"formula_20","formula_text":"𝑊 𝐶𝐻 = Softmax(MLP[𝐻 𝑚 𝑠 , 𝐸 𝑖 ]),(16)"},{"formula_coordinates":[6,386.35,179.89,171.85,9.16],"formula_id":"formula_21","formula_text":"𝑉 𝐿 = 𝑊 𝐶𝐻 • [ H𝑚 𝑠 ] 𝑇 , 𝑚 ∈ M,(17)"},{"formula_coordinates":[6,371.35,255.18,183.43,8.03],"formula_id":"formula_22","formula_text":"𝑌 𝑃𝐶𝑅 = Sigmoid(Concat(𝐸 𝑢 , 𝐸 𝑖 , 𝑉 𝐿 ))(18"},{"formula_coordinates":[6,355.65,476.43,202.55,7.93],"formula_id":"formula_23","formula_text":"𝐻 𝑝𝑡 = Concat(ℎ 𝑝𝑟𝑒 ⊕ 𝑒 𝑖 , ℎ 𝑝𝑟𝑒 ⊗ 𝑒 𝑖 , ℎ 𝑝𝑟𝑒 ⊖ 𝑒 𝑖 ),(19)"},{"formula_coordinates":[6,327.06,605.41,231.15,22.71],"formula_id":"formula_24","formula_text":"𝐻 all = Concat(𝐻 𝑝𝑡 , 𝐻 𝑐𝑡 , 𝐻 𝑣 𝑐𝑡 , 𝐻 𝑣 𝑝𝑟𝑒 , ℎ 𝑝𝑟𝑒 , ℎ 𝑐𝑎𝑛 , 𝐸 𝑢 , 𝑌 𝑃𝐶𝑅 ), (20a) 𝑌 𝐶𝐴𝑅 =𝜎 (MLP(𝐻 all )),(20b)"},{"formula_coordinates":[7,73.99,280.04,220.06,15.76],"formula_id":"formula_25","formula_text":"𝐽 = ∑︁ 𝑒 𝑡 ∈ D 𝑦 𝑒 𝑡 𝑢,𝑖 log ŷ𝑒 𝑡 𝑢,𝑖 + (1 -𝑦 𝑒 𝑡 𝑢,𝑖 ) log(1 -ŷ𝑒 𝑡 𝑢,𝑖 ) ,(21)"},{"formula_coordinates":[11,83.16,203.43,210.88,17.89],"formula_id":"formula_26","formula_text":"P 0 : OPT(S) = max 𝒙 𝑡 ∈X ∑︁ 𝑇 𝑡 =1 𝑓 (𝒙 𝑡 )(22)"},{"formula_coordinates":[11,86.87,226.84,207.17,15.89],"formula_id":"formula_27","formula_text":"s.t. 𝐶 1 : ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≤ 𝐺 max 𝑚,𝑡ℎ 𝑁 (S), ∀𝑚 ∈ M,(23)"},{"formula_coordinates":[11,106.17,246.68,187.88,15.89],"formula_id":"formula_28","formula_text":"𝐶 2 : ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≥ 𝐺 min 𝑚,𝑡ℎ 𝑁 (S), ∀𝑚 ∈ M.(24)"},{"formula_coordinates":[11,86.25,326.24,207.8,33.1],"formula_id":"formula_29","formula_text":"P 1 : OPT(S) = max 𝒙 𝑡 ∈X ∑︁ 𝑇 𝑡 =1 𝑓 (𝒙 𝑡 )(25) s"},{"formula_coordinates":[11,92.63,349.65,201.41,15.89],"formula_id":"formula_30","formula_text":".t. 𝛼𝑇𝐺 𝑚 ≤ ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≤ 𝑇𝐺 𝑚 , ∀𝑚 ∈ M.(26)"},{"formula_coordinates":[11,100.11,439.94,148.82,9.18],"formula_id":"formula_31","formula_text":"𝜇 𝑡 satisfy E[||∇ℎ(𝜇 𝑡 )|| ∞ ] ≤ 𝐶 ℎ , ∀𝑡 ∈ [𝑇 ]."},{"formula_coordinates":[11,107.58,587.15,186.47,18.88],"formula_id":"formula_32","formula_text":"E [𝑇 -𝜏 𝜋 ] ≤ ḡ 𝐺 + 𝐶 ℎ + ∥∇ℎ(𝜆 1 )∥ ∞ 𝜂𝐺 .(27)"},{"formula_coordinates":[11,74.11,626.56,219.94,24.19],"formula_id":"formula_33","formula_text":"∇𝜇 𝑘,𝑡 = -𝑔 𝑘 (𝒙 𝑡 ) + 𝐺 𝑘 (1(𝝁 𝑘 ≥ 0) + 𝛼 𝑘 1(𝝁 𝑘 < 0)) , ≤ -𝑔 𝑘 (𝒙 𝑡 ) + 𝐺 𝑘 , ∀𝑘 ∈ [𝑚].(28)"},{"formula_coordinates":[11,84.19,681.37,209.85,26.33],"formula_id":"formula_34","formula_text":"𝜏 𝜋 ∑︁ 𝑡 =1 ∇𝜇 𝑘,𝑡 ≤𝐺 𝑘 𝜏 𝜋 - 𝜏 𝜋 ∑︁ 𝑡 =1 𝑔 𝑘 (𝒙 𝑡 ) ≤ 𝐺 𝑘 𝜏 𝜋 -𝑇𝐺 𝑘 + ḡ,(29)"},{"formula_coordinates":[11,384.19,101.83,174.01,26.33],"formula_id":"formula_35","formula_text":"𝑇 -𝜏 𝜋 ≤ 1 𝐺 𝑘 ḡ - 𝜏 𝜋 ∑︁ 𝑡 =1 ∇𝜇 𝑘,𝑡 .(30)"},{"formula_coordinates":[11,317.96,158.71,160.14,14.08],"formula_id":"formula_36","formula_text":"-𝜏 𝜋 𝑡 =1 ∇𝜇 𝑘,𝑡 ≤ 1 𝜂 ∇ℎ 𝑘 (𝜇 𝜏 𝜋 +1 𝑘 ) -∇ℎ 𝑘 (𝜇 1 𝑘 ) ."},{"formula_coordinates":[11,351.57,189.55,206.63,52.52],"formula_id":"formula_37","formula_text":"E [𝑇 -𝜏 𝜋 ] ≤ ḡ 𝐺 𝑘 + E ∇ℎ 𝑘 (𝜇 𝜏 𝜋 +1 𝑘 ) -∇ℎ 𝑘 (𝜇 1 𝑘 ) 𝜂𝐺 𝑘 (31) ≤ ḡ 𝐺 + 𝐶 ℎ + ∥∇ℎ(𝜆 1 )∥ ∞ 𝜂𝐺 ,(32)"}],"string":"[\n {\n \"formula_coordinates\": [\n 3,\n 94.41,\n 394.83,\n 199.64,\n 19.41\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"X 𝑡 = 𝑖 ∈𝐼 𝑡 𝑥 𝑡,𝑛,𝑖 = 1, ∀𝑡 ∈ T , ∀𝑛 ∈ N 𝑛 𝑥 𝑡,𝑛,𝑖 ≤ 1, ∀𝑡 ∈ T , ∀𝑖 ∈ 𝐼,(1)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 83.4,\n 649.45,\n 210.64,\n 17.89\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"P 0 : OPT(S) = max 𝒙 𝑡 ∈X 𝑡 ∑︁ 𝑇 𝑡 =1 𝑓 (𝒙 𝑡 )(2)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 87.11,\n 672.85,\n 206.94,\n 15.89\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"s.t. 𝐶 1 : ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≤ 𝐺 max 𝑚,𝑡ℎ 𝑁 (S), ∀𝑚 ∈ M,(3)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 106.4,\n 692.7,\n 187.64,\n 15.89\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"𝐶 2 : ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≥ 𝐺 min 𝑚,𝑡ℎ 𝑁 (S), ∀𝑚 ∈ M,(4)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 119.42,\n 361.16,\n 174.62,\n 39.44\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"+ ∑︁ 𝑚 𝜆 𝑚 ∑︁ 𝑡 𝑔 𝑚 (𝒙 𝑡 ) -𝐺 min 𝑚 .(5)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 62.85,\n 509.68,\n 231.19,\n 78.43\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"min 𝝁 𝐷 (𝝁) = ∑︁ 𝑡 𝑓 (𝒙 𝑡 ) - ∑︁ 𝑚 [𝜇 𝑚 ] + ∑︁ 𝑡 𝑔 𝑚 (𝒙 𝑡 ) -𝐺 max 𝑚 (6a) + ∑︁ 𝑚 [-𝜇 𝑚 ] + ∑︁ 𝑡 𝑔 𝑚 (𝒙 𝑡 ) -𝐺 min 𝑚 (6b) = ∑︁ 𝑡 𝑓 (𝒙 𝑡 ) - ∑︁ 𝑚 𝜇 𝑚 𝑔 𝑚 (𝒙 𝑡 ) + ∑︁ 𝑚 [𝜇 𝑚 ] + 𝐺 max 𝑚 -[-𝜇 𝑚 ] + 𝐺 min 𝑚 ,(6c)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 79.05,\n 596.72,\n 77.01,\n 8.79\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"[𝜇 𝑚 ] + = max{𝜇 𝑚 , 0}.\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 103.58,\n 695.94,\n 190.46,\n 15.08\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"P 1 : x𝑡 = arg max 𝒙 𝑡 ∈X 𝑓 (𝒙 𝑡 ) -𝝁 𝑇 𝑡 𝑔(𝒙 𝑡 ) .(7)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 373.66,\n 400.35,\n 184.54,\n 10.98\n ],\n \"formula_id\": \"formula_8\",\n \"formula_text\": \"𝐺 max 𝑚,𝑡 +1 = 𝐺 max 𝑚,𝑡 -𝑔 𝑚 (𝒙 𝑡 ), ∀𝑚 ∈ M.(8)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 322.53,\n 442.11,\n 229.33,\n 12.25\n ],\n \"formula_id\": \"formula_9\",\n \"formula_text\": \"∇𝜇 𝑚,𝑡 = -𝑔 𝑚 (𝒙 𝑡 ) +𝐺 max 𝑚,𝑡 +1 • 1(𝜇 𝑚,𝑡 ≥ 0) +𝐺 min 𝑚 • 1(𝜇 𝑚,𝑡 ≤ 0), (\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 356.03,\n 495.43,\n 202.17,\n 17.62\n ],\n \"formula_id\": \"formula_10\",\n \"formula_text\": \"𝜇 𝑚,𝑡 +1 = arg min 𝜇 𝑚 ∈R 𝜇 𝑚 ∇𝜇 𝑚,𝑡 + 1 𝜂 𝑉 ℎ (𝜇 𝑚 , 𝜇 𝑚,𝑡 ),(10)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 398.84,\n 588.64,\n 159.36,\n 8.79\n ],\n \"formula_id\": \"formula_11\",\n \"formula_text\": \"𝝁 𝑡 +1 = [𝝁 𝑡 -𝜂∇𝝁 𝑡 ] + ,(11)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 90.65,\n 378.9,\n 203.4,\n 9.43\n ],\n \"formula_id\": \"formula_12\",\n \"formula_text\": \"Regret(𝜋 |S) = sup E S [OPT(S) -𝑅(𝜋 |S)] ,(12)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 53.8,\n 675.97,\n 212.23,\n 11.72\n ],\n \"formula_id\": \"formula_13\",\n \"formula_text\": \"channel 𝑚 ∈ M that 𝐺 min 𝑚 -E 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≤ 𝐶 4 + 𝐶 5\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 367.74,\n 197.82,\n 140.75,\n 15.08\n ],\n \"formula_id\": \"formula_14\",\n \"formula_text\": \"P 1 : x𝑡 = arg max 𝒙 𝑡 ∈X 𝑓 (𝒙 𝑡 ) -𝝁 𝑇 𝑡 𝑔(𝒙 𝑡 ) .\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 86.07,\n 135.89,\n 207.98,\n 10.42\n ],\n \"formula_id\": \"formula_15\",\n \"formula_text\": \"𝑉 𝑚 𝑙 = [ℎ𝑒𝑎𝑑 1 , ℎ𝑒𝑎𝑑 2 , ..., ℎ𝑒𝑎𝑑 ℎ ]𝑊 𝑂 ,(13a)\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 78.94,\n 152.44,\n 215.11,\n 23.78\n ],\n \"formula_id\": \"formula_16\",\n \"formula_text\": \"ℎ𝑒𝑎𝑑 𝑖 = Softmax (𝐸 𝑏 𝑊 𝑄 )(𝐸 𝑏 𝑊 𝐾 ) 𝑇 √︁ 𝑑 ℎ /ℎ (𝐸 𝑏 𝑊 𝑉 ) ,(13b)\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 83.55,\n 433.83,\n 207.08,\n 14.69\n ],\n \"formula_id\": \"formula_17\",\n \"formula_text\": \"𝑉 𝑈 = 𝑓 (𝐸 𝑏 ; 𝐸 𝑖 ) = ∑︁ 𝐵 𝑖=1 𝐴(𝑏 𝑖 , 𝐸 𝑖 )𝑏 𝑖 = ∑︁ 𝐵 𝑖=1 𝑤 𝑖 𝑏 𝑖 , (14\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 290.62,\n 439.44,\n 3.42,\n 4.09\n ],\n \"formula_id\": \"formula_18\",\n \"formula_text\": \")\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 59.33,\n 535.71,\n 234.72,\n 32.96\n ],\n \"formula_id\": \"formula_19\",\n \"formula_text\": \"𝐻 𝑖 𝑠 = Softmax (𝑉 𝑚 𝑙 𝑊 𝑞 ) [𝑉 𝑈 𝑊 𝑘1 , 𝑉 𝑚 𝑙 𝑊 𝑘2 ] 𝑇 √︁ 𝑑 ℎ [𝑉 𝑈 𝑊 𝑣1 , 𝑉 𝑚 𝑙 𝑊 𝑣2 ],(15)\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 114.79,\n 700.48,\n 179.25,\n 9.16\n ],\n \"formula_id\": \"formula_20\",\n \"formula_text\": \"𝑊 𝐶𝐻 = Softmax(MLP[𝐻 𝑚 𝑠 , 𝐸 𝑖 ]),(16)\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 386.35,\n 179.89,\n 171.85,\n 9.16\n ],\n \"formula_id\": \"formula_21\",\n \"formula_text\": \"𝑉 𝐿 = 𝑊 𝐶𝐻 • [ H𝑚 𝑠 ] 𝑇 , 𝑚 ∈ M,(17)\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 371.35,\n 255.18,\n 183.43,\n 8.03\n ],\n \"formula_id\": \"formula_22\",\n \"formula_text\": \"𝑌 𝑃𝐶𝑅 = Sigmoid(Concat(𝐸 𝑢 , 𝐸 𝑖 , 𝑉 𝐿 ))(18\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 355.65,\n 476.43,\n 202.55,\n 7.93\n ],\n \"formula_id\": \"formula_23\",\n \"formula_text\": \"𝐻 𝑝𝑡 = Concat(ℎ 𝑝𝑟𝑒 ⊕ 𝑒 𝑖 , ℎ 𝑝𝑟𝑒 ⊗ 𝑒 𝑖 , ℎ 𝑝𝑟𝑒 ⊖ 𝑒 𝑖 ),(19)\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 327.06,\n 605.41,\n 231.15,\n 22.71\n ],\n \"formula_id\": \"formula_24\",\n \"formula_text\": \"𝐻 all = Concat(𝐻 𝑝𝑡 , 𝐻 𝑐𝑡 , 𝐻 𝑣 𝑐𝑡 , 𝐻 𝑣 𝑝𝑟𝑒 , ℎ 𝑝𝑟𝑒 , ℎ 𝑐𝑎𝑛 , 𝐸 𝑢 , 𝑌 𝑃𝐶𝑅 ), (20a) 𝑌 𝐶𝐴𝑅 =𝜎 (MLP(𝐻 all )),(20b)\"\n },\n {\n \"formula_coordinates\": [\n 7,\n 73.99,\n 280.04,\n 220.06,\n 15.76\n ],\n \"formula_id\": \"formula_25\",\n \"formula_text\": \"𝐽 = ∑︁ 𝑒 𝑡 ∈ D 𝑦 𝑒 𝑡 𝑢,𝑖 log ŷ𝑒 𝑡 𝑢,𝑖 + (1 -𝑦 𝑒 𝑡 𝑢,𝑖 ) log(1 -ŷ𝑒 𝑡 𝑢,𝑖 ) ,(21)\"\n },\n {\n \"formula_coordinates\": [\n 11,\n 83.16,\n 203.43,\n 210.88,\n 17.89\n ],\n \"formula_id\": \"formula_26\",\n \"formula_text\": \"P 0 : OPT(S) = max 𝒙 𝑡 ∈X ∑︁ 𝑇 𝑡 =1 𝑓 (𝒙 𝑡 )(22)\"\n },\n {\n \"formula_coordinates\": [\n 11,\n 86.87,\n 226.84,\n 207.17,\n 15.89\n ],\n \"formula_id\": \"formula_27\",\n \"formula_text\": \"s.t. 𝐶 1 : ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≤ 𝐺 max 𝑚,𝑡ℎ 𝑁 (S), ∀𝑚 ∈ M,(23)\"\n },\n {\n \"formula_coordinates\": [\n 11,\n 106.17,\n 246.68,\n 187.88,\n 15.89\n ],\n \"formula_id\": \"formula_28\",\n \"formula_text\": \"𝐶 2 : ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≥ 𝐺 min 𝑚,𝑡ℎ 𝑁 (S), ∀𝑚 ∈ M.(24)\"\n },\n {\n \"formula_coordinates\": [\n 11,\n 86.25,\n 326.24,\n 207.8,\n 33.1\n ],\n \"formula_id\": \"formula_29\",\n \"formula_text\": \"P 1 : OPT(S) = max 𝒙 𝑡 ∈X ∑︁ 𝑇 𝑡 =1 𝑓 (𝒙 𝑡 )(25) s\"\n },\n {\n \"formula_coordinates\": [\n 11,\n 92.63,\n 349.65,\n 201.41,\n 15.89\n ],\n \"formula_id\": \"formula_30\",\n \"formula_text\": \".t. 𝛼𝑇𝐺 𝑚 ≤ ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≤ 𝑇𝐺 𝑚 , ∀𝑚 ∈ M.(26)\"\n },\n {\n \"formula_coordinates\": [\n 11,\n 100.11,\n 439.94,\n 148.82,\n 9.18\n ],\n \"formula_id\": \"formula_31\",\n \"formula_text\": \"𝜇 𝑡 satisfy E[||∇ℎ(𝜇 𝑡 )|| ∞ ] ≤ 𝐶 ℎ , ∀𝑡 ∈ [𝑇 ].\"\n },\n {\n \"formula_coordinates\": [\n 11,\n 107.58,\n 587.15,\n 186.47,\n 18.88\n ],\n \"formula_id\": \"formula_32\",\n \"formula_text\": \"E [𝑇 -𝜏 𝜋 ] ≤ ḡ 𝐺 + 𝐶 ℎ + ∥∇ℎ(𝜆 1 )∥ ∞ 𝜂𝐺 .(27)\"\n },\n {\n \"formula_coordinates\": [\n 11,\n 74.11,\n 626.56,\n 219.94,\n 24.19\n ],\n \"formula_id\": \"formula_33\",\n \"formula_text\": \"∇𝜇 𝑘,𝑡 = -𝑔 𝑘 (𝒙 𝑡 ) + 𝐺 𝑘 (1(𝝁 𝑘 ≥ 0) + 𝛼 𝑘 1(𝝁 𝑘 < 0)) , ≤ -𝑔 𝑘 (𝒙 𝑡 ) + 𝐺 𝑘 , ∀𝑘 ∈ [𝑚].(28)\"\n },\n {\n \"formula_coordinates\": [\n 11,\n 84.19,\n 681.37,\n 209.85,\n 26.33\n ],\n \"formula_id\": \"formula_34\",\n \"formula_text\": \"𝜏 𝜋 ∑︁ 𝑡 =1 ∇𝜇 𝑘,𝑡 ≤𝐺 𝑘 𝜏 𝜋 - 𝜏 𝜋 ∑︁ 𝑡 =1 𝑔 𝑘 (𝒙 𝑡 ) ≤ 𝐺 𝑘 𝜏 𝜋 -𝑇𝐺 𝑘 + ḡ,(29)\"\n },\n {\n \"formula_coordinates\": [\n 11,\n 384.19,\n 101.83,\n 174.01,\n 26.33\n ],\n \"formula_id\": \"formula_35\",\n \"formula_text\": \"𝑇 -𝜏 𝜋 ≤ 1 𝐺 𝑘 ḡ - 𝜏 𝜋 ∑︁ 𝑡 =1 ∇𝜇 𝑘,𝑡 .(30)\"\n },\n {\n \"formula_coordinates\": [\n 11,\n 317.96,\n 158.71,\n 160.14,\n 14.08\n ],\n \"formula_id\": \"formula_36\",\n \"formula_text\": \"-𝜏 𝜋 𝑡 =1 ∇𝜇 𝑘,𝑡 ≤ 1 𝜂 ∇ℎ 𝑘 (𝜇 𝜏 𝜋 +1 𝑘 ) -∇ℎ 𝑘 (𝜇 1 𝑘 ) .\"\n },\n {\n \"formula_coordinates\": [\n 11,\n 351.57,\n 189.55,\n 206.63,\n 52.52\n ],\n \"formula_id\": \"formula_37\",\n \"formula_text\": \"E [𝑇 -𝜏 𝜋 ] ≤ ḡ 𝐺 𝑘 + E ∇ℎ 𝑘 (𝜇 𝜏 𝜋 +1 𝑘 ) -∇ℎ 𝑘 (𝜇 1 𝑘 ) 𝜂𝐺 𝑘 (31) ≤ ḡ 𝐺 + 𝐶 ℎ + ∥∇ℎ(𝜆 1 )∥ ∞ 𝜂𝐺 ,(32)\"\n }\n]"}}},{"rowIdx":394,"cells":{"doi":{"kind":"string","value":"10.1161/CIR.0000000000001039"},"pub_date":{"kind":"string","value":""},"sections":{"kind":"list like","value":[{"figure_ref":["fig_0"],"heading":"Introduction","publication_ref":["b0","b1","b2","b3","b4","b5","b6","b7","b8","b9","b10","b11","b11","b12"],"table_ref":[],"text":"Coronary artery disease (CAD) has long been the leading cause of death in the United States [1] and is rapidly becoming the top killer in other counties in the world [2,3]. CAD is characterized by the accumulation of cholesterol-rich plaque within the inner layers of the coronary artery wall. This plaque buildup leads to varying degrees of stenosis, or narrowing, in the arteries. As a consequence, blood flow to the myocardium, the heart muscle, becomes restricted, resulting in myocardial ischemia -a condition marked by insufficient oxygen and nutrient supply to the heart. As the blood supply-demand mismatch worsens, symptoms of CAD such as chest pain (angina), shortness of breath, and others may appear. Moreover, instability in the areas of stenosis can lead to acute occlusion of the coronary artery, resulting in a heart attack. Currently, the treatment for significant coronary artery stenosis involves percutaneous coronary intervention (PCI) or coronary artery bypass grafting (CABG) along with aggressive medical therapies [4].\nInvasive coronary angiography (ICA) indeed remains the gold standard for diagnosing CAD [5]. It is an established imaging technique that plays a crucial role in both the diagnosis and treatment of heart conditions, particularly CAD. During an ICA procedure, a thin, flexible tube called a catheter is inserted into an artery, usually in the groin or arm, and guiding it through the blood vessels to the heart. Once the catheter is in place, a special dye is injected into the coronary arteries, which makes them visible. These images support cardiologists identifying any blockages or narrowing in the coronary arteries. Depending on the findings, the cardiologists may use the same catheter to perform treatments such as angioplasty or stent placement to open blocked arteries and improve blood flow to the heart. While ICA is highly effective in providing detailed images of the coronary arteries, it is important to acknowledge the limitations associated with subjective visual assessment. Human interpretation of the angiograms can introduce variability and subjectivity into the diagnosis, potentially leading to unreliable assessments [6].\nThe coronary vascular tree is indeed complex and contains two major systems: the left coronary artery (LCA) and the right coronary artery (RCA) systems. The LCA is more clinically relevant, given it provides most of the blood supply to the left ventricle. LCA system and codominance are associated with modestly increased PCI in-hospital mortality in patients with stable CAD [7]. According to the 15-segments model defined by the American Heart Association [8], the LCA system can be further subdivided into three main coronary arteries: the left anterior descending (LAD) artery, the left circumflex (LCX) artery, and the left main artery (LMA), which act as the main blood suppliers of the myocardium. In normal coronary anatomy, the LMA bifurcates into two main branches: LAD and LCX. The LAD artery supplies blood to the front and part of the left side of the myocardium muscle while the LCX courses around the left side of the myocardium and supplies blood to the left atrium and part of the left ventricle [9]. The length of LAD varies between 10 to 13 cm and gives rise to the diagonal branches (D), which further contribute to the blood supply of the myocardium. Similarly, the length of the LCX artery varies from 5 to 8 cm and gives rise to obtuse marginal (OM) branches, extending its reach to specific areas of the left ventricle [10].\nHowever, it is challenging to identify individual coronary artery using ICAs because of the morphological similarity among different segments [11] and loss of partial high-frequency detail information due to the projection of the 3D vascular tree into a 2D plane during ICA imaging acquisition [12]. This loss of information limits the ability to precisely discern the intricate structures and boundaries of the coronary arteries, making their identification more challenging. In addition, the coronary arteries not only span over a long distance but also show similar semantic features with each other [12], making it challenging to associate them with the exact branches. This causes the confusion in judgement of whether a coronary segment belongs to main branch or side branch. ICA image understanding is a complex task due to several factors that degrade the visual quality of the images. Furthermore, ICA images suffer from various factors that degrade their visual quality. The contrast degradation occurs as the contrast dye dissipates, leading to reduced visibility of the coronary arteries. Additionally, spatial blurring and overlapping caused by nonvessel tissues and structures further obscure the vessel boundaries and hinder accurate interpretation.\nThe limitations of existing coronary artery semantic labeling methods that rely solely on position and imaging features are evident when processing complex coronary vasculature from different view angles of ICA images [13]. The morphological similarity among different types of arteries poses a challenge for pixel-intensity-based models to accurately discern each arterial segment and generate meaningful semantic segmentation. To address these challenges, we propose an innovative approach that leverages the concept of graphs to represent the coronary arteries and their connections. By converting the arteries into graph structures, we can incorporate non-traditional features, such as node degrees and graph-theory distances, in addition to the pixel-derived features, to enhance the accuracy of semantic segmentation. A key aspect of our method is the use of graph matching techniques, which aim to find the semantic correspondence between arteries in the labeled graphs. By comparing the graph structures and identifying similar patterns, we can establish meaningful associations between arterial segments and assign appropriate semantic labels.\nIn this paper, we propose a novel algorithm to perform coronary artery semantic labeling using ICA images. We propose an edge attention graph matching network (EAGMN) to build the semantic correspondence between coronary arterial segments from ICAs. The problem of semantic segmentation is converted into a problem that classifies the type of an unlabeled arterial segment by searching for a labelled arterial segment with the maximal similarity in a database of arterial graphs generated from a large number of ICAs. In detail, the individual graph of the coronary artery is generated according to coronary artery binary segmentation result. The node in the individual graph represents a segment of the coronary artery and the edge indicates the connectivity between arterial segments according to the physical connection of the vascular tree. Then an association graph is constructed from two individual graphs, where each vertex is built from two nodes in the individual graph, representing the node-to-node correspondence of two arterial segments from two individual graphs. Thus, the coronary artery semantic labeling task is converted into vertex classification task using the generated association graph. EAGMN incorporates an encoder module responsible for embedding vertex and edge features using a graph attention convolution network. This module extracts meaningful representations from the association graph by considering the interactions and relationships between the vertices and edges. Additionally, the EAGMN employs a decoder module that facilitates the readout of the feature representations generated by the encoder. By examining the positive vertices in the association graph, which indicate matched nodes from the individual graphs, the EAGMN accomplishes the semantic labeling task by assigning appropriate labels to the coronary arterial segments. The workflow of the proposed EAGMN for coronary artery semantic labeling is shown in Figure 1. 1) The utilization of graph matching for coronary artery semantic labeling: This paper introduces a novel approach that employs graph matching techniques to establish the node-to-node correspondence between labeled and unlabeled coronary arterial segments. This enables accurate and reliable semantic labeling of the coronary artery.\n2) The EAGMN utilizes an edge attention mechanism to dynamically aggregate features of adjacency edges, which in turn update vertex features for vertex classification using the association graph.\n3) We demonstrate the robustness of the proposed model by conducting experiments using corrupted ICAs, which simulates challenging real-world scenarios."},{"figure_ref":[],"heading":"4)","publication_ref":[],"table_ref":[],"text":"We employ ZORRO to provide interpretability and explainability of the graph matching model, shedding light on the decision-making process. This paper is organized as follows. Section 1 introduces the background, challenges of coronary artery semantic labeling and the highlights of the paper. Section 2 reviews existing algorithms on coronary artery semantic labeling and state-of-the-art graph matching algorithms. In Section 3, the proposed EAGMN is described in detail. The enrolled subjects, implementation details, experimental results, and discussion are presented in Section 4. Section 5 illustrates the conclusion."},{"figure_ref":[],"heading":"Related work","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Coronary artery semantic labeling","publication_ref":["b14","b5","b11","b15","b16","b17","b12"],"table_ref":[],"text":"Coronary artery semantic labeling can be categorized into pixel-to-pixel/voxel-to-voxel based semantic segmentation methods and segment identification based semantic labeling methods. Pixel-to-pixel based methods focus on achieving dense pixel-level or voxel-level classification by mapping each individual pixel or voxel to its corresponding semantic category. On the other hand, segment identification based methods take a different approach by classifying the entire arterial segment as a whole into a specific semantic category.\nPixel-to-pixel based methods are straightforward because the algorithm assigns unique labels to each pixel in the arterial images. Jun et al. proposed a T-Net for main artery segmentation using ICAs and T-Net achieved a Dice coefficient similarity of 0.8377 using 4700 ICAs for LAD, LCX and RCA segmentation [15]. Xian et al. proposed a U-Net based residual attention network, which integrates attention mechanism and stacks multiple attention modules for main arteries segmentation. The model achieved an F1 score of 0.921 using an dataset with 3,200 ICAs [6]. Zhang et al. proposed the progressive perception learning framework to capture the long-distance semantic relationship, enhance the foreground and suppress the background pixels, and highlight the boundary details for main artery semantic segmentation [12]. The model was validated using a dataset containing 1085 subjects with ICAs and a Dice coefficient similarity of 0.9585 was achieved. Although the aforementioned studies have demonstrated impressive performance in segmenting the main arteries, they rely on separate deep learning models for extracting each type of artery and they cannot label side branches, such as D and OM arteries. Focusing solely on the main branches, such as LMA, LAD, and LCX, may not provide adequate support for downstream CAD analysis.\nSegment identification based semantic labeling methods usually contain a vascular tree binary segmentation step and a segment classification step. Cao et al. proposed an automatic coronary artery semantic labeling algorithm based on blood flow and logical rules using coronary computed tomography angiography (CCTA) images [16]. Wu et al. extracted the arterial features according to the spatial locations and directions and organized the arterial segments as tree-structured sequential data. Then, a bi-directional tree-structured long short-term memory network (LSTM) was employed to classify each segment to perform coronary artery semantic labeling [17]. Yang et al. integrated the imaging features extracted by convolutional LSTM and position features as the feature embedding and employed a partial-residual graph convolution network (GCN) for coronary artery semantic labeling [18]. Existing coronary artery semantic labeling methods have demonstrated remarkable performance when applied to 3D CCTA data. However, their direct applicability to 2D ICA images is limited due to the loss of partial high-frequency detail information resulting from the projection of the 3D vascular tree into a 2D plane. Consequently, the labeling and segmentation of coronary arteries in 2D ICA images may suffer from inaccuracies. Our previous work performed coronary artery semantic labeling by learning the semantic correspondence of arterial branches from different individual graphs using graph matching [13]. The information from neighbors is uniformly averaged during the graph matching. However, different types of arteries contribute unequally which induced a lower accuracy. It is crucial to develop methods specifically designed for 2D ICA images to enhance the accuracy and reliability of coronary artery disease diagnosis and treatment."},{"figure_ref":[],"heading":"Graph matching","publication_ref":["b18","b19","b20","b21","b22","b23","b24"],"table_ref":[],"text":"Graph-matching is a combinatorial optimization problem based on graph structure, which is NP-hard or practically intractable for large-scale settings. The graph matching aims to establish a meaningful node-tonode and edge-to-edge correspondence between different graphs, which is often formulated as a graph edit problem [19], subgraph isomorphism problem and quadradic assignment problem [20]. Traditional graph matching algorithms primarily emphasize combinatorial matching techniques, which involve comparing and matching the structural components of graphs. On the other hand, learning-based methods take a different approach by incorporating feature extraction and affinity learning. These methods are particularly advantageous when dealing with large-scale and high-dimensional data, as they enable the identification of meaningful patterns and relationships within graphs [21]. With the advert development of graph neural network (GNN), it has been embedded into the learning-based graph matching problem because GNN models the structured information and helps transform the graph matching problem into a linear assignment task [22]. Nowak et al. employed GNN to extract structured features of nodes and employed Sinkhorn network as a differentiable layer to solve the linear assignment problem [23]. Wang et al. employed the GNN for both intra-graph node embedding and cross-graph node embedding iteratively for graph matching [24]. Later, Wang et al. directly employed the association graph induced affinity matrix for embedding learning and the Koopmans-Beckmann's QAP was adopted for affinity learning [25]. While existing graph matching algorithms have demonstrated promising performance in various domains, such as nature images, their application to medical images with topological features is still relatively limited and understudied. Given the significance of topological features in medical image analysis, there is a growing need for dedicated research efforts to develop graph matching algorithms specifically tailored to medical imaging. These algorithms should consider the complex interplay between structural elements and leverage the intrinsic knowledge of anatomical connectivity to achieve more accurate and meaningful matching results."},{"figure_ref":[],"heading":"Methodology","publication_ref":[],"table_ref":[],"text":"The approach outlined in this study centers on segment identification-based coronary artery semantic labeling. Our proposed EAGMN aims to establish the node-to-node correspondence between two individual graphs, taking into account the edge information. In this framework, each node represents a coronary arterial segment. Consequently, the task of coronary artery semantic labeling is transformed into the problem of finding one-to-one or one-to-zero mappings for arterial segments from two distinct vascular trees. Our method consists of two main steps: vascular tree binary segmentation and segment classification."},{"figure_ref":["fig_1"],"heading":"Vascular tree extraction and graph generation","publication_ref":["b13","b25","b26"],"table_ref":[],"text":"Our previous work, Feature Pyramid U-Net++ (FP-U-Net++) [14], is used to extract the vascular tree. To analyze the arterial anatomy and perform the semantic labeling, it is essential to extract the arterial centerline from the extracted vascular tree. The arterial graph generation includes the centerline extraction and arterial segment separation. Centerline extraction is the process of removing the redundant foreground pixels using binary image while preserving the connectivity and the topology of the vascular tree. We adopt the hit-and-mass transformation algorithm [26], which is based on morphological thinning, to extract the arterial centerline. The edge linking algorithm is employed to identify the end points and bifurcation points within the arterial segments. The end points denote the termination of an arterial segment, while the bifurcation points represent the junctions where the arterial segment branches out into sub-branches [27]. Based on the detected end points and bifurcation points, the arterial centerline is divided into distinct centerline segments. Each centerline segment is then associated with an arterial segment, and the semantic labeling involves assigning a label to each arterial segment. An example of the individual graph generation pipeline is shown in Figure 2.\nThe individual coronary arterial graph is constructed based on the extracted arterial segments, centerline segments, end points, and bifurcation points. The graph's connectivity, represented by its edges, is determined by the connections between arterial segments. Each node in the arterial graph is formed by the endpoints and the bifurcation points, and the semantic labels should be assigned to each edge. However, in the context of the graph matching neural network, the focus is on establishing the correspondence between nodes. Therefore, in practical terms, we interchange the concepts of nodes and edges in the individual graph. In this adjusted perspective, each node represents an arterial segment, and each edge represents an endpoint or a bifurcation point in the arterial centerline, which preserves the connectivity of the arterial tree. "},{"figure_ref":["fig_2"],"heading":"Graph matching for coronary artery semantic labeling","publication_ref":["b24","b27","b10","b28","b29","b10","b30","b31","b32","b33","b34"],"table_ref":["tab_7"],"text":"Graph matching aims to establish a meaningful node-to-node correspondence between individual graphs. Given two undirected attributed individual graphs 𝒢 1 = (𝕍 1 , 𝔼 1 , 𝒱 1 , ℰ 1 ) and 𝒢 2 = (𝕍 2 , 𝔼 2 , 𝒱 2 , ℰ 2 ), where |𝕍 1 | = 𝑛 1 and |𝕍 2 | = 𝑛 2 , and |𝔼 1 | = 𝑛 𝑒1 and |𝔼 2 | = 𝑛 𝑒2 . We aim to find the node correspondence between them by considering the node-to-node affinities and edge-to-edge affinities. Without loss of the generality, we assume 𝑛 1 ≤ 𝑛 2 . The node-to-node correspondence can be represented by an assignment matrix 𝑀 ∈ {0,1} 𝑛 1 ×𝑛 2 that is defined in Eq. 1.\n𝑀 𝑖𝑗 = { 1 𝑖𝑓 𝑉 𝑖 𝑚𝑎𝑡𝑐ℎ𝑒𝑠 𝑉 𝑗 0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒(1)\nInspired by the success of applying association graphs in combinatorial optimization [25,28], we transform the graph matching problem between two individual graphs 𝒢 1 and 𝒢 2 into a vertex binary classification problem using an attributed undirected association graph 𝒢 𝐴 = (𝕍 𝐴 , 𝔼 𝐴 , 𝒱 𝐴 , ℰ 𝐴 ). 𝒢 𝐴 contains all candidate correspondence between nodes from two individual graphs. Each vertex of 𝒢 𝐴 is built by the nodes of the individual graphs that 𝕍 𝐴 = {(𝑉 1 1 , 𝑉 2 have the same semantic labels, then the vertex 𝑉 𝑖𝑎 ∈ 𝒢 𝐴 is a positive vertex, indicating that these two nodes are matched and these two arterial segments are matched; and verse visa. Our association-graph based edge attention graph matching network for coronary artery semantic labeling contains 6 modules, as shown in Figure 3. 1) Feature extraction in individual graphs. According to the extracted coronary artery segment, segment centerline, endpoints, bifurcation points and topology of the vascular graph, we extract the imaging features, arterial position features, and topology features for each node in individual graph. For the imaging features, we extract the texture features [11,29] according to masked ICA images for the artery using PyRadiomics [30], including first-ordered features, shape-based features, and gray-level based features. To measure the absolute position of the arterial segment relative to the entire vascular tree, we design 20 hand-crafted position features. These features capture various spatial attributes, such as the distance from the segment to the root of the tree, and the relative position within the tree hierarchy [11]. For the topology feature, we use the degree of the two endpoints of the artery segment as the features. The detailed feature information is shown in Table S1. In total, 121 features are extracted. Without loss of the generality, the dimension of node feature is denoted as 𝑑 and 𝑣 𝑖 ∈ ℝ 𝑑 .\n2) Feature extraction in association graph. The vertex is generated according to the nodes from the individual graphs. Then, the vertex feature 𝑣 𝑖𝑎 is generated by concatenating features from features of node 𝑉 𝑖 ∈ 𝒢 1 and features of node 𝑉 𝑎 ∈ 𝒢 2 , as defined in Eq. 2.\n𝑣 𝑖𝑎 = [𝑣 𝑖 , 𝑣 𝑎 ] ∈ ℝ 2𝑑 s.t. 𝑖 ∈ [1, ⋯ , 𝑛 1 ] and 𝑗 ∈ [1, ⋯ , 𝑛 2 ](2)\nwhere [⋅] denotes feature concatenation. In the same manner, the feature of edge 𝐸 𝑖𝑎,𝑗𝑏 in 𝒢 𝐴 is constructed by concatenating the features of edge 𝐸 𝑖𝑗 in 𝒢 1 and the features of edge 𝐸 𝑎𝑏 in 𝒢 2 , as defined in Eq. 3.\n𝑒 𝑖𝑎,𝑗𝑏 = [𝑒 𝑖𝑗 1 , 𝑒 𝑎𝑏 2 ] ∈ ℝ 4𝑑 s.t. 𝑖, 𝑗 ∈ [1, ⋯ , 𝑛 1 ] and 𝑎, 𝑏 ∈ [1, ⋯ , 𝑛 2 ](3)\nwhere\n𝑒 𝑖𝑗 𝑔 = [𝑣 𝑖 , 𝑣 𝑗 ] s.t. 𝑔 ∈ [1,2].\n3) Feature embedding. In the feature embedding stage, we aim to transform the input features into a lowerdimensional representation which captures the essential information for the task of coronary artery semantic labeling. This process involves mapping the input features into a feature space where relationships and patterns relevant to the task can be more easily learned and utilized by the model. Using the constructed attribute association graph 𝒢 𝐴 , two multi-layer perceptron (MLP) based encoders are employed to perform vertex and edge feature embedding, as defined in Eq. 4.\n𝑣 𝑖𝑎 𝑒𝑚𝑏 = 𝑓 𝑣 𝑒𝑚𝑏 (𝑣 𝑖𝑎 ) 𝑒 𝑖𝑎,𝑗𝑏 𝑒𝑚𝑏 = 𝑓 𝑒 𝑒𝑚𝑏 (𝑒 𝑖𝑎,𝑗𝑏 )(4)\nwhere 𝑓 𝑣 𝑒𝑚𝑏 and 𝑓 𝑒 𝑒𝑚𝑏 are MLPs with layer-wise instance normalization [31].\n4) Attention score calculation for dynamic feature aggregation. GCNs are a type of neural networks specifically designed to operate on graph-structured data. The key idea of GCN is to perform convolution on graphs by aggregating information from neighborhood nodes and edges [32]. One of the popular GCNs is graph sample and aggregated (GraphSAGE) GCN [33], which aggregates information from a node's local neighborhood using various aggregation functions, such as mean, max or concatenation. The information from neighbors is uniformly averaged because GCN treats the neighbors equally. However, in coronary artery semantic labeling task, different types of arteries contribute unequally to the central artery segment. Thus, we employed graph attention network (GAT) which incorporates attention mechanisms to improve information aggregation using graph-structured data [34,35]. In detail, an edge attention network is employed to calculate the edge attention score dynamically according to the embedded features in Eq. 4.\nIn detail, the two MLP based encoders are employed to further embed the vertex and edge features, as defined in Eq. 5.\n𝑣̅ 𝑖𝑎 𝑒𝑚𝑏 = 𝑔 𝑣 𝑒𝑚𝑏 (𝑣 𝑖𝑎 𝑒𝑚𝑏 ) 𝑒̅ 𝑖𝑎,𝑗𝑏 𝑒𝑚𝑏 = 𝑔 𝑒 𝑒𝑚𝑏 (𝑒 𝑖𝑎,𝑗𝑏 𝑒𝑚𝑏 )(5)\nwhere 𝑔 𝑣 𝑒𝑚𝑏 and 𝑔 𝑒 𝑒𝑚𝑏 are MLPs with layer-wise instance normalization. Then, a GCN layer is employed to perform information aggregation. The GCN layer contains an edge convolution layer and a vertex convolution layer. The edge convolution layer aggregates the information of the connected vertices and updates edge features according to the aggregated features and the previous edge features. Formally, the edge convolution layer contains an aggregation function and an update function, as defined in Eq. 6. "},{"figure_ref":[],"heading":"𝑒̅ 𝑖𝑎","publication_ref":[],"table_ref":[],"text":"where 𝑔 𝑒 𝑔𝑎𝑡 is an MLP encoder which aggregates features based on the embedded vertex features in Eq. 5\nof the connected two vertices of edge 𝐸 𝑖𝑎,𝑗𝑏 ; and 𝑔 𝑒 𝑢 is another MLP encoder which updates edge features according to updated features and the edge features defined in Eq. 5. The purpose of this aggregation process is to enhance the representation of the current edge by considering the characteristics and interactions of the connected vertices. This helps to capture more comprehensive and discriminative information about the graph structure and facilitates the subsequent vertex classification task.\nThe vertex convolution layer aggregates the information of the connected edges and updates its attributes according to the aggregated features and the embedded features in Eq. 5. Formally, node convolution layer contains an information aggregation layer and an update function, as defined in Eq. 7.\n𝑣̅ \nwhere 𝑓 𝑐𝑙𝑓 is an MLP decoder with ReLU activation functions and 𝑦 ̂𝑖𝑎 ∈ ℝ represents the predicted probability of the vertex 𝑉 𝑖𝑎 ."},{"figure_ref":[],"heading":"Training and testing","publication_ref":["b7","b35","b36"],"table_ref":[],"text":"To train the proposed EAGMN, we need to prepare pairs of individual graphs, and then generate the association graph based on these selected pairs. In our study, we initially manually annotated the coronary arterial segments using ICA images. Subsequently, we employed the method described in Section 3.1 of our paper to generate the labeled individual graphs. According to the coronary artery anatomy defined by the American Heart Association [8], two cardiologists assigned semantic labels to each arterial segment. Then, the node correspondences between arterial segments are automatically identified and the ground truth of assignment matrix 𝑀 is generated and 𝑦 𝑖𝑎 = 1 if two arterial segments are matched.\nHowever, it should be noted that the main arterial branches, such as LAD and LCX, are often divided into multiple segments due to the presence of side branches like D and OM. As a result, the individual graph representation of these main branches consists of multiple nodes, all of which share the same semantic labels. This introduces a unique challenge in the graph matching task, as it transforms into a one-to-many or many-to-many mapping problem. Unlike the simpler one-to-one or one-to-zero mappings, the presence of multiple possible mappings between nodes significantly increases the complexity of the search space [36,37]. To simplify the complexity of the graph matching task, our approach focuses on one-to-one graph matching. We handle the issue of multiple nodes representing the separated main branches by assigning semantic labels with increasing indices along the blood flow. By assigning incremental semantic labels, we ensure that each node in the individual graph representing a main branch segment has a unique label. This allows us to establish a one-to-one correspondence between the nodes in the individual graphs, simplifying the graph matching process. For example, if LCX and OM1 exist, then LCX is separated into two arterial segments with the semantic labels of LCX1 and LCX2. Then, each arterial segment in 𝒢 1 is only matched with one exact segment in 𝒢 2 .\nFor model training, a batch of individual graph pairs is selected, and the association graphs are used as the input of the EAGMN. In our study, we specifically focus on ICAs obtained from two different view angles: left anterior oblique (LAO) and right anterior oblique (RAO). We acknowledge that the anatomy and visual characteristics of the coronary arteries can differ between these two view angles. To ensure consistency and accuracy in our analysis, we select individual graphs exclusively from ICAs captured using the same view angle. This allows us to maintain a more homogenous dataset for our analysis and ensure reliable results in coronary artery semantic labeling. The mean squared error (MSE) between the predicted vertex classification probability and the ground truth is used to optimize the model, as defined in Eq. 12.\n𝐿 = 𝑀𝑆𝐸(𝑀, 𝑀 ̂) = ∑ ∑(𝑀 𝑖𝑎 -𝑀 ̂𝑖𝑎 ) 2 𝑛 2 𝑎=1 𝑛 1 𝑖=1(12)\nwhere 𝑀 𝑖𝑎 = 𝑦 𝑖𝑎 is the ground truth and 𝑀 ̂𝑖𝑎 = 𝑦 ̂𝑖𝑎 is the predicted vertex class. An Adam optimizer with the learning rate of 0.0001 was used to optimize the model. The training algorithm is shown in Algorithm 1."},{"figure_ref":[],"heading":"Algorithm 1.","publication_ref":[],"table_ref":[],"text":"The training process of our proposed EAGMN.\nDuring the model testing phase, we employ a voting strategy to assign semantic labels to artery branches in the tested ICAs. Our proposed EAGMN establishes the node-to-node semantic correspondence between the individual graphs. During the testing, 𝒢 1 is generated using ICAs in the testing set and 𝒢 2 is generated using ICAs in the template set, where template set contains a set of representative ICAs selected by experienced cardiologists. We simulate the learning procedure by graph matching that the cardiologist learns the coronary artery anatomy by comparing the anatomy of the testing case with the reference cases in the template set. By applying the EAGMN to the association graph generated by the tested ICA and template ICA, the artery segments are matched, and the semantic labeling is achieved. Given the complex and diverse structure of coronary artery anatomy, clinical decisions are often made based on multiple ICA frames. Therefore, we perform graph matching between the tested ICA and every ICA in the template set. Each arterial segment of the tested ICA is matched to multiple arterial segments in the template set, and a majority voting strategy is employed to assign the label according to the labels of the matched arterial segments in the template set. This approach helps ensure robustness and accuracy in the final semantic labeling process. The testing algorithm is shown in Algorithm 2. 4. Embed vertex and edge features 𝑣̅ 𝑖𝑎 𝑒𝑚𝑏 and 𝑒̅ 𝑖𝑎,𝑗𝑏 𝑒𝑚𝑏 with Eq. 5."},{"figure_ref":[],"heading":"Input","publication_ref":[],"table_ref":[],"text":"For 𝑢 = 1 ⋯ 𝑁 𝑎𝑡𝑡 do 5. Update edge features 𝑣̅ 𝑖𝑎 𝑔𝑎𝑡 using Eq. 6 and vertex features 𝑣̅ 𝑖𝑎 𝑔𝑎𝑡 with Eq. 7.\n6. Calculate edge attention score 𝜃 𝑖𝑎,𝑗𝑏 with Eq. 8.\nFor 𝑣 = 1 ⋯ 𝑁 𝑚𝑝 do 7. Update edge features using 𝑒 𝑖𝑎,𝑗𝑏 𝑔𝑐𝑛 with Eq. 9 and vertex features 𝑣 𝑖𝑎 𝑔𝑐𝑛 with Eq. 10 8. Perform vertex classification with Eq. 11. 9. Optimize EAGMN using Eq .12.\nAlgorithm 2. The testing process of our proposed EAGMN."},{"figure_ref":[],"heading":"Evaluation","publication_ref":["b15"],"table_ref":[],"text":"The × 𝑛 𝑐 (16) where 𝑇𝑃 𝑐 , 𝑇𝑁 𝑐 , 𝐹𝑃 𝑐 and 𝐹𝑁 𝑐 represent the true positive, true negative, false positive, and false negative arterial segments, respectively. 𝐶 is the total number of classes. 𝑛 𝐶 is the number of arterial segments in class 𝐶, and 𝑛 is the total number of arterial segments."},{"figure_ref":[],"heading":"Model interpretation","publication_ref":["b37","b38"],"table_ref":[],"text":"Interpreting the decisions made by EAGMN is crucial for building trust and confidence in its predictions, especially in clinical practice. Existing approaches for explaining both node and edge features in GNN focus on gradient-based approaches and perturbation-based methods [38]. Gradient-based methods suffer from saturation problems in that the model output changes minimally with respect to any input change; thus, we adopt a perturbation-based method, ZORRO [39], which compares the output variations with respect to different input perturbations to calculate the node and feature importance for our graph matching network. Given a generated association graph, ZORRO iteratively and recursively adds the important features and nodes according to the fidelity score, where the fidelity score measures the difference between the raw"},{"figure_ref":[],"heading":"Input:","publication_ref":["b38"],"table_ref":[],"text":"𝐷 𝑡𝑒 = {𝒢 𝑖 = (𝕍 𝑖 , 𝔼 model predictions and the predictions after masking out the important features and nodes [39]. Formally, the fidelity score is defined as\n𝐹(𝕍 𝑠 , 𝔽 𝑠 ) = 1 𝑛 1 𝑛 2 ∑ ∑ 𝕀(𝑦 𝑖𝑎 𝑠 = 𝑦 ̂𝑖𝑎 ) 𝑛 2 𝑎=1 𝑛 1 𝑖=1 (17\n)\nwhere 𝕀 is the indicator function. 𝑦 ̂𝑖𝑎 is the original prediction of the model and 𝑦 𝑖𝑎 𝑠 is the prediction using the masked nodes by 𝕍 𝑠 and the masked features by 𝔽 𝑠 . 𝕍 𝑠 and 𝔽 𝑠 are binary vectors in which 1\nindicates the node/feature is selected and vice versa. An explanation for GNN-based model indicates that using the masked features and the masked nodes in graph, the fidelity score measured by the new prediction and the original prediction reached 𝜏, i.e. 𝜏 ≥ 𝐹(𝕍 𝑠 , 𝔽 𝑠 ).\nExplaining feature importance. In section 3.1, we manually designed 121 hand-crafted features. The vertex features are generated by the concatenation of the nodes in the individual graph and identifying the concatenated features for each vertex cannot explain the feature importance for the node in the individual graph. To explain the feature importance, a unified feature mask is applied to the feature of every node in the individual graphs, then the masked features are concatenated as the vertex features using Eq. 2. ZORRO masks out all features at the beginning of the explanation and gradually adding features with the highest fidelity scores until the fidelity score is reached at 𝜏. By applying ZORRO to all individual pairs generated by the testing set and the template set, the frequency of the selected important features indicates the feature importance used for graph matching.\nExplaining node importance. For the original ZORRO, neighbor nodes are removed first and then gradually added to test the importance of the node by the improvement of the fidelity scores. The aim of explaining node importance is to interpret the importance of different arterial segments for graph matching. During the testing phase, the individual graph 𝒢 1 is the tested case and 𝒢 2 is the template ICA graph. However, our graph matching algorithm is applied to the association graph, examining the importance of vertex in the association graph cannot reflect the node importance. Thus, we modify the ZORRO to explain the node importance that at the beginning of the explanation, all nodes in 𝒢 1 is retained and all nodes in 𝒢 2 are removed. ZORRO iteratively adds the node from 𝒢 2 with the highest improvement of the fidelity score.\nAccording to the improvement of the fidelity score, the importance of the node is obtained. The node importance in our approach reflects the significance of each arterial segment in the template ICAs for both graph matching and semantic labeling. By assigning importance scores to the nodes, we can capture the relevance and influence of each arterial segment in the template set. This allows us to prioritize and weight the contributions of different segments during the graph matching process."},{"figure_ref":[],"heading":"Experiments and discussions","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Dataset","publication_ref":["b12"],"table_ref":[],"text":"In this study, we manually annotated 204 and 59 ICAs from site 1 at The First Affiliated Hospital of Nanjing Medical University and site 2 at Chang Bing Show Chwan Memorial Hospital, respectively. In total, this retrospective study enrolled 263 ICA images. The detailed description of the image acquisition was illustrated in our previous work [13]. For each patient, a frame that was used for anatomical structure analysis in clinical practice was selected from the view video for semantic labeling. In this study, we only focus on semantic labeling for the main branches of LMA, LAD, and LCX, and the side branches of D and OM."},{"figure_ref":[],"heading":"Implementation details","publication_ref":["b39"],"table_ref":[],"text":"We implemented our EAGMN using TensorFlow and GraphNets [40]. The vertex feature embedding module 𝑓 𝑣 𝑒𝑚𝑏 , the edge feature embedding module 𝑓 𝑒 𝑒𝑚𝑏 , the graph convolution module 𝑓 𝑒 𝑔𝑐𝑛 , 𝑓 𝑒 𝑢 , 𝑓 𝑣 𝑔𝑐𝑛 and 𝑓 𝑣 𝑢 were implemented using two MLP layers with layer-wise instance normalization and the number of hidden units were set as 64. For the attention score calculation, the vertex feature embedding module 𝑔 𝑣 𝑒𝑚𝑏 , the edge feature embedding module 𝑔 𝑒 𝑒𝑚𝑏 , the graph attention module 𝑔 𝑒 𝑔𝑎𝑡 , 𝑔 𝑒 𝑢 , 𝑔 𝑣 𝑔𝑎𝑡 and 𝑔 𝑣 𝑢 were also implemented using two MLP layers with the 64 hidden unites. The edge attention score readout module 𝑔 𝑒 𝑜𝑢𝑡 and the vertex classification layer 𝑓 𝑐𝑙𝑓 were implemented using two MLP layers with ReLU activation functions. The number of message-passing steps for graph attention score calculation, 𝑁 𝑎𝑡𝑡 , was set as 3 and the number of message-passing steps for graph convolution, 𝑁 𝑚𝑝 , was set as 2."},{"figure_ref":[],"heading":"Coronary artery semantic labeling performance","publication_ref":["b10","b16","b17","b23","b24","b12","b23","b24","b12"],"table_ref":["tab_6","tab_8","tab_8"],"text":"Our ICA dataset contains 263 annotated ICA graphs, and 79 of them were selected as the template set. The remining 184 ICAs were used for five-fold cross validation, which indicates that each fold 147 ICAs were used as the training set and 37 ICAs were used as the testing set. We trained our EAGMN for 𝑁 = 100,000 steps. The averaged performance on the five-folds were reported. The achieved performance of our EAGMN for coronary artery semantic labeling is shown in Table 1. We also compared the EAGMN with three existing segment identification based semantic labeling methods.\nIn addition, we implemented 3 graph matching neural networks and applied them into coronary artery semantic labeling task using the same training and testing algorithms developed in this study.\n• SVM [11]: In our previous work, a support vector machine was employed to classify coronary artery segments using the 20 position features and 2 topology features as described in section 3.1.\n• BiTreeLSTM [17]: The bidirectional tree LSTM (BiTreeLSTM) was initially applied for coronary artery semantic labeling using CCTA according to the spatial locations and directions of arteries in 3D. We adopted the same network architecture but extracted the spatial locations and directions of coronary arteries in 2D. • CPR-GCN [18]: CPR-GCN was proposed to perform coronary artery semantic labeling using CCTA with 3D convolution LSTM for arterial imaging feature extraction and GCN for artery semantic labeling. We adopted the same network architecture but extracted the arterial imaging features using 2D convolution LSTM. • IPCA [24]: Iterative Permutation loss and Cross-graph Affinity (IPCA) for graph matching employed the GNN for both intra-graph node embedding and cross-graph node embedding iteratively for graph matching using individual graphs. • NGM [25]: Neural Graph Matching (NGM) network employed the association graph induced affinity matrix for embedding learning. In our EAGMN, the connectivity of the association was generated by the connectivity of the individual graphs; however, NGM adopted the Koopmans-Beckmann's QAP for affinity learning to calculate the assignment matrix. • AGMN [13]: Our previous work adopted the association graph for graph matching without the graph attention module. For each baseline, we performed five-fold cross validation, and the performance comparisons of the coronary artery semantic labeling are illustrated in Table 2. According to Table 2, the proposed EAGMN achieved the highest accuracy of coronary artery semantic labeling for 0.8653. However, the BiTreeLSTM baseline achieved stable performance in labeling LMA branches. For BiTreeLSTM and CPR-GCN models, they were initially designed for coronary artery semantic labeling using CCTA datasets. Though they showed impressive performance with an ACC greater than 0.9, they may not be suitable for coronary artery semantic labeling using ICAs. Understanding the relationships between arteries and their spatial orientation requires a three-dimensional perspective; however, coronary artery labeling using 2D images is challenging and these two methods cannot achieve satisfactory performance. In detail, the CPR-GCN achieved a weighted ACC of 0.4581 while the BiTreeLSTM achieved an ACC of 0.7492, which were significantly lower than the graph matching based methods. The results indicate that the complex branching patterns and variations in anatomy make it more challenging to understand coronary artery anatomy, and learning the ICA anatomy using individual graphs is difficult.\nFor the graph matching based methods, we use IPCA, NGM and AGMN as baselines. IPCA [24] first performs feature embedding using individual graphs, and then performs cross-graph feature embeddings using GCNs. Compared to association graph-based graph matching algorithms, individual graph-based matching methods are computationally efficient and suitable for simpler graphs but may struggle with complex graph structures. IPCA achieved an ACC of 0.8039, which was lower than the association graphbased methods. NGM [25] first embeds features of nodes of the individual graphs to calculate the affinity matrix, and uses the calculated affinity matrix as the adjacency matrix for the association graph. Then, NGM performs the feature embedding according to the node features for vertex classification.\nOn the contrary, AGMN [13] and our proposed EAGMN concatenate the node features of the individual graphs and perform the node embedding using the concatenated features. The adjacency matrix of the association graph is generated by the connectivity of the individual graphs rather than the affinity matrix used by NGM. AGMN and EAGMN achieved the ACC of 0.8264 and 0.8653, which were higher than that of NGM. Experimental results indicate that the physical connectivity of the coronary artery plays an important role in graph matching, where the physical connectivity refers to the actual anatomical connections between different arterial segments within the coronary artery network. The physical connectivity of the coronary artery provides valuable cues and constraints that guide the matching process. It helps ensure that the matched arterial segments are not only similar in appearance or position but also consistent with the underlying anatomical structure. Compared to AGMN, our proposed EAGMN employs the edge attention mechanisms to aggregate information from the adjacent edges dynamically. By employing edge attention mechanisms, EAGMN can adaptively prioritize and weight the importance of different edges based on their relevance to the graph matching task. This dynamic attention mechanism allows the model to focus on the most informative and discriminative edges while suppressing the influence of less relevant edges. Experimental results indicate that EAGMN improved the ACC by 0.0389 for all types of arteries compared to that of AGMN, which outperformed other baseline models significantly."},{"figure_ref":["fig_3","fig_3"],"heading":"Robustness test","publication_ref":[],"table_ref":[],"text":"In clinical practice, due to the imaging quality, artery overlapping and human subjective annotation, the algorithm proposed in section 3.1 may not generate a perfect coronary arterial graph. To reflect the robustness of the proposed EAGMN and test the clinical flexibility, we performed the robustness test using the corrupted individual graphs with random arterial segment dropping. Note that the graph-based approaches use the topology information and the connected graphs as the input, thus, we only randomly removed the arterial branches with at least one endpoint during the robustness test. In addition, the LMA was retained in all experiments due to the importance of this in-let artery segmentation for the coronary artery anatomy. We set the probability of 5%, 7.5%, 10%, 12.5%, 15%, 17.5% and 20% to randomly remove the arteries, and the averaged performance for the robustness test among baselines and EAGMN of ICAs in testing set is shown in Figure 4. According to Figure 4, the proposed EAGMN was robust since the achieved ACCs were greater than 0.81 during these robust testing experiments, even with 20% missing arterial segments. The BiTreeLSTM and CPR-GCN showed a minor performance dropping, which indicated that these two methods did not fully utilize the graph structure of coronary artery. In comparison to the graph matching-based methods, such as IPCA, NGM, and AGMN, our proposed EAGMN consistently demonstrated superior performance in terms of ACC, PREC, REC, and F1 across various missing probabilities. These results highlight the robustness of the proposed EAGMN in handling missing data scenarios."},{"figure_ref":["fig_4","fig_5","fig_5"],"heading":"Result explanation","publication_ref":[],"table_ref":[],"text":"We adopted 37 ICAs as the testing set and 70 ICAs as the template set. Each ICA in testing set was used as 𝒢 1 to perform graph matching with every ICA as 𝒢 2 in template set. Note that number of nodes in 𝒢 1 is smaller than 𝒢 2 , and 1148 individual graph matching pairs were generated. We applied ZORRO to explain each graph matching pair to iteratively and recursively calculate the feature and node importance for graph matching. The threshold of the fidelity score used to measure the new prediction and the original prediction, 𝜏, was set as 0.8. We ranked the frequency of the selected important features to measure the feature importance, as shown in Figure 5. For explaining the node importance, we visualized the graph matching results and the fidelity score improvement when adding the node in the 𝒢 2 into the node mask. Example results are shown in Figure 6.\nIn Figure 6, the top four pairs of ICA graphs demonstrate a successful and accurate matching with only minor errors. However, the bottom two pairs do not exhibit a satisfactory level of matching. Upon analyzing the fidelity scores depicted in the pseudo color visualization, we observed that the LMA branch plays a crucial role in the identification of coronary arteries. The successful identification of the LMA branch, along with the first segment of LAD (LAD1) and the first segment of LCX (LCX1), is of paramount importance for the semantic labeling of the entire vascular graph.\nIn contrast, for the bottom two pairs of graphs, the significance of the LMA branch was diminished compared to the top four pairs, leading to a degradation in performance. The analysis of the node importance further confirms that our EAGMN closely mimics the approach of cardiologists in identifying coronary arteries. It starts by focusing on the main LMA branch and subsequently considers the two most vital subbranches, namely LAD1 and LCX1, in order to achieve accurate semantic labeling. Overall, these findings highlight the importance of the LMA branch and the initial segments of LAD and LCX in the graph matching process. They demonstrate that our EAGMN model effectively emulates the decisionmaking process followed by cardiologists when identifying and labeling coronary arteries. "},{"figure_ref":[],"heading":"Conclusion","publication_ref":[],"table_ref":[],"text":"In this paper, we present an edge attention graph matching network for coronary artery semantic labeling. By performing the graph matching between individual ICA generated graphs, the relationship between the arterial segments is obtained and the unlabeled coronary arterial segments are labeled by template ICAs. Experimental results showed that our approach is powerful and robust. By employing the ZORRO, we explained the graph matching results and improved the interpretability of coronary artery semantic labeling using our proposed approach."},{"figure_ref":[],"heading":"Acknowledgement","publication_ref":[],"table_ref":[],"text":"This research was supported in part by a research seed fund from Michigan Technological University Health Research Institute and an NIH grant (U19AG055373)."}],"string":"[\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b0\",\n \"b1\",\n \"b2\",\n \"b3\",\n \"b4\",\n \"b5\",\n \"b6\",\n \"b7\",\n \"b8\",\n \"b9\",\n \"b10\",\n \"b11\",\n \"b11\",\n \"b12\"\n ],\n \"table_ref\": [],\n \"text\": \"Coronary artery disease (CAD) has long been the leading cause of death in the United States [1] and is rapidly becoming the top killer in other counties in the world [2,3]. CAD is characterized by the accumulation of cholesterol-rich plaque within the inner layers of the coronary artery wall. This plaque buildup leads to varying degrees of stenosis, or narrowing, in the arteries. As a consequence, blood flow to the myocardium, the heart muscle, becomes restricted, resulting in myocardial ischemia -a condition marked by insufficient oxygen and nutrient supply to the heart. As the blood supply-demand mismatch worsens, symptoms of CAD such as chest pain (angina), shortness of breath, and others may appear. Moreover, instability in the areas of stenosis can lead to acute occlusion of the coronary artery, resulting in a heart attack. Currently, the treatment for significant coronary artery stenosis involves percutaneous coronary intervention (PCI) or coronary artery bypass grafting (CABG) along with aggressive medical therapies [4].\\nInvasive coronary angiography (ICA) indeed remains the gold standard for diagnosing CAD [5]. It is an established imaging technique that plays a crucial role in both the diagnosis and treatment of heart conditions, particularly CAD. During an ICA procedure, a thin, flexible tube called a catheter is inserted into an artery, usually in the groin or arm, and guiding it through the blood vessels to the heart. Once the catheter is in place, a special dye is injected into the coronary arteries, which makes them visible. These images support cardiologists identifying any blockages or narrowing in the coronary arteries. Depending on the findings, the cardiologists may use the same catheter to perform treatments such as angioplasty or stent placement to open blocked arteries and improve blood flow to the heart. While ICA is highly effective in providing detailed images of the coronary arteries, it is important to acknowledge the limitations associated with subjective visual assessment. Human interpretation of the angiograms can introduce variability and subjectivity into the diagnosis, potentially leading to unreliable assessments [6].\\nThe coronary vascular tree is indeed complex and contains two major systems: the left coronary artery (LCA) and the right coronary artery (RCA) systems. The LCA is more clinically relevant, given it provides most of the blood supply to the left ventricle. LCA system and codominance are associated with modestly increased PCI in-hospital mortality in patients with stable CAD [7]. According to the 15-segments model defined by the American Heart Association [8], the LCA system can be further subdivided into three main coronary arteries: the left anterior descending (LAD) artery, the left circumflex (LCX) artery, and the left main artery (LMA), which act as the main blood suppliers of the myocardium. In normal coronary anatomy, the LMA bifurcates into two main branches: LAD and LCX. The LAD artery supplies blood to the front and part of the left side of the myocardium muscle while the LCX courses around the left side of the myocardium and supplies blood to the left atrium and part of the left ventricle [9]. The length of LAD varies between 10 to 13 cm and gives rise to the diagonal branches (D), which further contribute to the blood supply of the myocardium. Similarly, the length of the LCX artery varies from 5 to 8 cm and gives rise to obtuse marginal (OM) branches, extending its reach to specific areas of the left ventricle [10].\\nHowever, it is challenging to identify individual coronary artery using ICAs because of the morphological similarity among different segments [11] and loss of partial high-frequency detail information due to the projection of the 3D vascular tree into a 2D plane during ICA imaging acquisition [12]. This loss of information limits the ability to precisely discern the intricate structures and boundaries of the coronary arteries, making their identification more challenging. In addition, the coronary arteries not only span over a long distance but also show similar semantic features with each other [12], making it challenging to associate them with the exact branches. This causes the confusion in judgement of whether a coronary segment belongs to main branch or side branch. ICA image understanding is a complex task due to several factors that degrade the visual quality of the images. Furthermore, ICA images suffer from various factors that degrade their visual quality. The contrast degradation occurs as the contrast dye dissipates, leading to reduced visibility of the coronary arteries. Additionally, spatial blurring and overlapping caused by nonvessel tissues and structures further obscure the vessel boundaries and hinder accurate interpretation.\\nThe limitations of existing coronary artery semantic labeling methods that rely solely on position and imaging features are evident when processing complex coronary vasculature from different view angles of ICA images [13]. The morphological similarity among different types of arteries poses a challenge for pixel-intensity-based models to accurately discern each arterial segment and generate meaningful semantic segmentation. To address these challenges, we propose an innovative approach that leverages the concept of graphs to represent the coronary arteries and their connections. By converting the arteries into graph structures, we can incorporate non-traditional features, such as node degrees and graph-theory distances, in addition to the pixel-derived features, to enhance the accuracy of semantic segmentation. A key aspect of our method is the use of graph matching techniques, which aim to find the semantic correspondence between arteries in the labeled graphs. By comparing the graph structures and identifying similar patterns, we can establish meaningful associations between arterial segments and assign appropriate semantic labels.\\nIn this paper, we propose a novel algorithm to perform coronary artery semantic labeling using ICA images. We propose an edge attention graph matching network (EAGMN) to build the semantic correspondence between coronary arterial segments from ICAs. The problem of semantic segmentation is converted into a problem that classifies the type of an unlabeled arterial segment by searching for a labelled arterial segment with the maximal similarity in a database of arterial graphs generated from a large number of ICAs. In detail, the individual graph of the coronary artery is generated according to coronary artery binary segmentation result. The node in the individual graph represents a segment of the coronary artery and the edge indicates the connectivity between arterial segments according to the physical connection of the vascular tree. Then an association graph is constructed from two individual graphs, where each vertex is built from two nodes in the individual graph, representing the node-to-node correspondence of two arterial segments from two individual graphs. Thus, the coronary artery semantic labeling task is converted into vertex classification task using the generated association graph. EAGMN incorporates an encoder module responsible for embedding vertex and edge features using a graph attention convolution network. This module extracts meaningful representations from the association graph by considering the interactions and relationships between the vertices and edges. Additionally, the EAGMN employs a decoder module that facilitates the readout of the feature representations generated by the encoder. By examining the positive vertices in the association graph, which indicate matched nodes from the individual graphs, the EAGMN accomplishes the semantic labeling task by assigning appropriate labels to the coronary arterial segments. The workflow of the proposed EAGMN for coronary artery semantic labeling is shown in Figure 1. 1) The utilization of graph matching for coronary artery semantic labeling: This paper introduces a novel approach that employs graph matching techniques to establish the node-to-node correspondence between labeled and unlabeled coronary arterial segments. This enables accurate and reliable semantic labeling of the coronary artery.\\n2) The EAGMN utilizes an edge attention mechanism to dynamically aggregate features of adjacency edges, which in turn update vertex features for vertex classification using the association graph.\\n3) We demonstrate the robustness of the proposed model by conducting experiments using corrupted ICAs, which simulates challenging real-world scenarios.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"4)\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We employ ZORRO to provide interpretability and explainability of the graph matching model, shedding light on the decision-making process. This paper is organized as follows. Section 1 introduces the background, challenges of coronary artery semantic labeling and the highlights of the paper. Section 2 reviews existing algorithms on coronary artery semantic labeling and state-of-the-art graph matching algorithms. In Section 3, the proposed EAGMN is described in detail. The enrolled subjects, implementation details, experimental results, and discussion are presented in Section 4. Section 5 illustrates the conclusion.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related work\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Coronary artery semantic labeling\",\n \"publication_ref\": [\n \"b14\",\n \"b5\",\n \"b11\",\n \"b15\",\n \"b16\",\n \"b17\",\n \"b12\"\n ],\n \"table_ref\": [],\n \"text\": \"Coronary artery semantic labeling can be categorized into pixel-to-pixel/voxel-to-voxel based semantic segmentation methods and segment identification based semantic labeling methods. Pixel-to-pixel based methods focus on achieving dense pixel-level or voxel-level classification by mapping each individual pixel or voxel to its corresponding semantic category. On the other hand, segment identification based methods take a different approach by classifying the entire arterial segment as a whole into a specific semantic category.\\nPixel-to-pixel based methods are straightforward because the algorithm assigns unique labels to each pixel in the arterial images. Jun et al. proposed a T-Net for main artery segmentation using ICAs and T-Net achieved a Dice coefficient similarity of 0.8377 using 4700 ICAs for LAD, LCX and RCA segmentation [15]. Xian et al. proposed a U-Net based residual attention network, which integrates attention mechanism and stacks multiple attention modules for main arteries segmentation. The model achieved an F1 score of 0.921 using an dataset with 3,200 ICAs [6]. Zhang et al. proposed the progressive perception learning framework to capture the long-distance semantic relationship, enhance the foreground and suppress the background pixels, and highlight the boundary details for main artery semantic segmentation [12]. The model was validated using a dataset containing 1085 subjects with ICAs and a Dice coefficient similarity of 0.9585 was achieved. Although the aforementioned studies have demonstrated impressive performance in segmenting the main arteries, they rely on separate deep learning models for extracting each type of artery and they cannot label side branches, such as D and OM arteries. Focusing solely on the main branches, such as LMA, LAD, and LCX, may not provide adequate support for downstream CAD analysis.\\nSegment identification based semantic labeling methods usually contain a vascular tree binary segmentation step and a segment classification step. Cao et al. proposed an automatic coronary artery semantic labeling algorithm based on blood flow and logical rules using coronary computed tomography angiography (CCTA) images [16]. Wu et al. extracted the arterial features according to the spatial locations and directions and organized the arterial segments as tree-structured sequential data. Then, a bi-directional tree-structured long short-term memory network (LSTM) was employed to classify each segment to perform coronary artery semantic labeling [17]. Yang et al. integrated the imaging features extracted by convolutional LSTM and position features as the feature embedding and employed a partial-residual graph convolution network (GCN) for coronary artery semantic labeling [18]. Existing coronary artery semantic labeling methods have demonstrated remarkable performance when applied to 3D CCTA data. However, their direct applicability to 2D ICA images is limited due to the loss of partial high-frequency detail information resulting from the projection of the 3D vascular tree into a 2D plane. Consequently, the labeling and segmentation of coronary arteries in 2D ICA images may suffer from inaccuracies. Our previous work performed coronary artery semantic labeling by learning the semantic correspondence of arterial branches from different individual graphs using graph matching [13]. The information from neighbors is uniformly averaged during the graph matching. However, different types of arteries contribute unequally which induced a lower accuracy. It is crucial to develop methods specifically designed for 2D ICA images to enhance the accuracy and reliability of coronary artery disease diagnosis and treatment.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Graph matching\",\n \"publication_ref\": [\n \"b18\",\n \"b19\",\n \"b20\",\n \"b21\",\n \"b22\",\n \"b23\",\n \"b24\"\n ],\n \"table_ref\": [],\n \"text\": \"Graph-matching is a combinatorial optimization problem based on graph structure, which is NP-hard or practically intractable for large-scale settings. The graph matching aims to establish a meaningful node-tonode and edge-to-edge correspondence between different graphs, which is often formulated as a graph edit problem [19], subgraph isomorphism problem and quadradic assignment problem [20]. Traditional graph matching algorithms primarily emphasize combinatorial matching techniques, which involve comparing and matching the structural components of graphs. On the other hand, learning-based methods take a different approach by incorporating feature extraction and affinity learning. These methods are particularly advantageous when dealing with large-scale and high-dimensional data, as they enable the identification of meaningful patterns and relationships within graphs [21]. With the advert development of graph neural network (GNN), it has been embedded into the learning-based graph matching problem because GNN models the structured information and helps transform the graph matching problem into a linear assignment task [22]. Nowak et al. employed GNN to extract structured features of nodes and employed Sinkhorn network as a differentiable layer to solve the linear assignment problem [23]. Wang et al. employed the GNN for both intra-graph node embedding and cross-graph node embedding iteratively for graph matching [24]. Later, Wang et al. directly employed the association graph induced affinity matrix for embedding learning and the Koopmans-Beckmann's QAP was adopted for affinity learning [25]. While existing graph matching algorithms have demonstrated promising performance in various domains, such as nature images, their application to medical images with topological features is still relatively limited and understudied. Given the significance of topological features in medical image analysis, there is a growing need for dedicated research efforts to develop graph matching algorithms specifically tailored to medical imaging. These algorithms should consider the complex interplay between structural elements and leverage the intrinsic knowledge of anatomical connectivity to achieve more accurate and meaningful matching results.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Methodology\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The approach outlined in this study centers on segment identification-based coronary artery semantic labeling. Our proposed EAGMN aims to establish the node-to-node correspondence between two individual graphs, taking into account the edge information. In this framework, each node represents a coronary arterial segment. Consequently, the task of coronary artery semantic labeling is transformed into the problem of finding one-to-one or one-to-zero mappings for arterial segments from two distinct vascular trees. Our method consists of two main steps: vascular tree binary segmentation and segment classification.\"\n },\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"Vascular tree extraction and graph generation\",\n \"publication_ref\": [\n \"b13\",\n \"b25\",\n \"b26\"\n ],\n \"table_ref\": [],\n \"text\": \"Our previous work, Feature Pyramid U-Net++ (FP-U-Net++) [14], is used to extract the vascular tree. To analyze the arterial anatomy and perform the semantic labeling, it is essential to extract the arterial centerline from the extracted vascular tree. The arterial graph generation includes the centerline extraction and arterial segment separation. Centerline extraction is the process of removing the redundant foreground pixels using binary image while preserving the connectivity and the topology of the vascular tree. We adopt the hit-and-mass transformation algorithm [26], which is based on morphological thinning, to extract the arterial centerline. The edge linking algorithm is employed to identify the end points and bifurcation points within the arterial segments. The end points denote the termination of an arterial segment, while the bifurcation points represent the junctions where the arterial segment branches out into sub-branches [27]. Based on the detected end points and bifurcation points, the arterial centerline is divided into distinct centerline segments. Each centerline segment is then associated with an arterial segment, and the semantic labeling involves assigning a label to each arterial segment. An example of the individual graph generation pipeline is shown in Figure 2.\\nThe individual coronary arterial graph is constructed based on the extracted arterial segments, centerline segments, end points, and bifurcation points. The graph's connectivity, represented by its edges, is determined by the connections between arterial segments. Each node in the arterial graph is formed by the endpoints and the bifurcation points, and the semantic labels should be assigned to each edge. However, in the context of the graph matching neural network, the focus is on establishing the correspondence between nodes. Therefore, in practical terms, we interchange the concepts of nodes and edges in the individual graph. In this adjusted perspective, each node represents an arterial segment, and each edge represents an endpoint or a bifurcation point in the arterial centerline, which preserves the connectivity of the arterial tree. \"\n },\n {\n \"figure_ref\": [\n \"fig_2\"\n ],\n \"heading\": \"Graph matching for coronary artery semantic labeling\",\n \"publication_ref\": [\n \"b24\",\n \"b27\",\n \"b10\",\n \"b28\",\n \"b29\",\n \"b10\",\n \"b30\",\n \"b31\",\n \"b32\",\n \"b33\",\n \"b34\"\n ],\n \"table_ref\": [\n \"tab_7\"\n ],\n \"text\": \"Graph matching aims to establish a meaningful node-to-node correspondence between individual graphs. Given two undirected attributed individual graphs 𝒢 1 = (𝕍 1 , 𝔼 1 , 𝒱 1 , ℰ 1 ) and 𝒢 2 = (𝕍 2 , 𝔼 2 , 𝒱 2 , ℰ 2 ), where |𝕍 1 | = 𝑛 1 and |𝕍 2 | = 𝑛 2 , and |𝔼 1 | = 𝑛 𝑒1 and |𝔼 2 | = 𝑛 𝑒2 . We aim to find the node correspondence between them by considering the node-to-node affinities and edge-to-edge affinities. Without loss of the generality, we assume 𝑛 1 ≤ 𝑛 2 . The node-to-node correspondence can be represented by an assignment matrix 𝑀 ∈ {0,1} 𝑛 1 ×𝑛 2 that is defined in Eq. 1.\\n𝑀 𝑖𝑗 = { 1 𝑖𝑓 𝑉 𝑖 𝑚𝑎𝑡𝑐ℎ𝑒𝑠 𝑉 𝑗 0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒(1)\\nInspired by the success of applying association graphs in combinatorial optimization [25,28], we transform the graph matching problem between two individual graphs 𝒢 1 and 𝒢 2 into a vertex binary classification problem using an attributed undirected association graph 𝒢 𝐴 = (𝕍 𝐴 , 𝔼 𝐴 , 𝒱 𝐴 , ℰ 𝐴 ). 𝒢 𝐴 contains all candidate correspondence between nodes from two individual graphs. Each vertex of 𝒢 𝐴 is built by the nodes of the individual graphs that 𝕍 𝐴 = {(𝑉 1 1 , 𝑉 2 have the same semantic labels, then the vertex 𝑉 𝑖𝑎 ∈ 𝒢 𝐴 is a positive vertex, indicating that these two nodes are matched and these two arterial segments are matched; and verse visa. Our association-graph based edge attention graph matching network for coronary artery semantic labeling contains 6 modules, as shown in Figure 3. 1) Feature extraction in individual graphs. According to the extracted coronary artery segment, segment centerline, endpoints, bifurcation points and topology of the vascular graph, we extract the imaging features, arterial position features, and topology features for each node in individual graph. For the imaging features, we extract the texture features [11,29] according to masked ICA images for the artery using PyRadiomics [30], including first-ordered features, shape-based features, and gray-level based features. To measure the absolute position of the arterial segment relative to the entire vascular tree, we design 20 hand-crafted position features. These features capture various spatial attributes, such as the distance from the segment to the root of the tree, and the relative position within the tree hierarchy [11]. For the topology feature, we use the degree of the two endpoints of the artery segment as the features. The detailed feature information is shown in Table S1. In total, 121 features are extracted. Without loss of the generality, the dimension of node feature is denoted as 𝑑 and 𝑣 𝑖 ∈ ℝ 𝑑 .\\n2) Feature extraction in association graph. The vertex is generated according to the nodes from the individual graphs. Then, the vertex feature 𝑣 𝑖𝑎 is generated by concatenating features from features of node 𝑉 𝑖 ∈ 𝒢 1 and features of node 𝑉 𝑎 ∈ 𝒢 2 , as defined in Eq. 2.\\n𝑣 𝑖𝑎 = [𝑣 𝑖 , 𝑣 𝑎 ] ∈ ℝ 2𝑑 s.t. 𝑖 ∈ [1, ⋯ , 𝑛 1 ] and 𝑗 ∈ [1, ⋯ , 𝑛 2 ](2)\\nwhere [⋅] denotes feature concatenation. In the same manner, the feature of edge 𝐸 𝑖𝑎,𝑗𝑏 in 𝒢 𝐴 is constructed by concatenating the features of edge 𝐸 𝑖𝑗 in 𝒢 1 and the features of edge 𝐸 𝑎𝑏 in 𝒢 2 , as defined in Eq. 3.\\n𝑒 𝑖𝑎,𝑗𝑏 = [𝑒 𝑖𝑗 1 , 𝑒 𝑎𝑏 2 ] ∈ ℝ 4𝑑 s.t. 𝑖, 𝑗 ∈ [1, ⋯ , 𝑛 1 ] and 𝑎, 𝑏 ∈ [1, ⋯ , 𝑛 2 ](3)\\nwhere\\n𝑒 𝑖𝑗 𝑔 = [𝑣 𝑖 , 𝑣 𝑗 ] s.t. 𝑔 ∈ [1,2].\\n3) Feature embedding. In the feature embedding stage, we aim to transform the input features into a lowerdimensional representation which captures the essential information for the task of coronary artery semantic labeling. This process involves mapping the input features into a feature space where relationships and patterns relevant to the task can be more easily learned and utilized by the model. Using the constructed attribute association graph 𝒢 𝐴 , two multi-layer perceptron (MLP) based encoders are employed to perform vertex and edge feature embedding, as defined in Eq. 4.\\n𝑣 𝑖𝑎 𝑒𝑚𝑏 = 𝑓 𝑣 𝑒𝑚𝑏 (𝑣 𝑖𝑎 ) 𝑒 𝑖𝑎,𝑗𝑏 𝑒𝑚𝑏 = 𝑓 𝑒 𝑒𝑚𝑏 (𝑒 𝑖𝑎,𝑗𝑏 )(4)\\nwhere 𝑓 𝑣 𝑒𝑚𝑏 and 𝑓 𝑒 𝑒𝑚𝑏 are MLPs with layer-wise instance normalization [31].\\n4) Attention score calculation for dynamic feature aggregation. GCNs are a type of neural networks specifically designed to operate on graph-structured data. The key idea of GCN is to perform convolution on graphs by aggregating information from neighborhood nodes and edges [32]. One of the popular GCNs is graph sample and aggregated (GraphSAGE) GCN [33], which aggregates information from a node's local neighborhood using various aggregation functions, such as mean, max or concatenation. The information from neighbors is uniformly averaged because GCN treats the neighbors equally. However, in coronary artery semantic labeling task, different types of arteries contribute unequally to the central artery segment. Thus, we employed graph attention network (GAT) which incorporates attention mechanisms to improve information aggregation using graph-structured data [34,35]. In detail, an edge attention network is employed to calculate the edge attention score dynamically according to the embedded features in Eq. 4.\\nIn detail, the two MLP based encoders are employed to further embed the vertex and edge features, as defined in Eq. 5.\\n𝑣̅ 𝑖𝑎 𝑒𝑚𝑏 = 𝑔 𝑣 𝑒𝑚𝑏 (𝑣 𝑖𝑎 𝑒𝑚𝑏 ) 𝑒̅ 𝑖𝑎,𝑗𝑏 𝑒𝑚𝑏 = 𝑔 𝑒 𝑒𝑚𝑏 (𝑒 𝑖𝑎,𝑗𝑏 𝑒𝑚𝑏 )(5)\\nwhere 𝑔 𝑣 𝑒𝑚𝑏 and 𝑔 𝑒 𝑒𝑚𝑏 are MLPs with layer-wise instance normalization. Then, a GCN layer is employed to perform information aggregation. The GCN layer contains an edge convolution layer and a vertex convolution layer. The edge convolution layer aggregates the information of the connected vertices and updates edge features according to the aggregated features and the previous edge features. Formally, the edge convolution layer contains an aggregation function and an update function, as defined in Eq. 6. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"𝑒̅ 𝑖𝑎\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"where 𝑔 𝑒 𝑔𝑎𝑡 is an MLP encoder which aggregates features based on the embedded vertex features in Eq. 5\\nof the connected two vertices of edge 𝐸 𝑖𝑎,𝑗𝑏 ; and 𝑔 𝑒 𝑢 is another MLP encoder which updates edge features according to updated features and the edge features defined in Eq. 5. The purpose of this aggregation process is to enhance the representation of the current edge by considering the characteristics and interactions of the connected vertices. This helps to capture more comprehensive and discriminative information about the graph structure and facilitates the subsequent vertex classification task.\\nThe vertex convolution layer aggregates the information of the connected edges and updates its attributes according to the aggregated features and the embedded features in Eq. 5. Formally, node convolution layer contains an information aggregation layer and an update function, as defined in Eq. 7.\\n𝑣̅ \\nwhere 𝑓 𝑐𝑙𝑓 is an MLP decoder with ReLU activation functions and 𝑦 ̂𝑖𝑎 ∈ ℝ represents the predicted probability of the vertex 𝑉 𝑖𝑎 .\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Training and testing\",\n \"publication_ref\": [\n \"b7\",\n \"b35\",\n \"b36\"\n ],\n \"table_ref\": [],\n \"text\": \"To train the proposed EAGMN, we need to prepare pairs of individual graphs, and then generate the association graph based on these selected pairs. In our study, we initially manually annotated the coronary arterial segments using ICA images. Subsequently, we employed the method described in Section 3.1 of our paper to generate the labeled individual graphs. According to the coronary artery anatomy defined by the American Heart Association [8], two cardiologists assigned semantic labels to each arterial segment. Then, the node correspondences between arterial segments are automatically identified and the ground truth of assignment matrix 𝑀 is generated and 𝑦 𝑖𝑎 = 1 if two arterial segments are matched.\\nHowever, it should be noted that the main arterial branches, such as LAD and LCX, are often divided into multiple segments due to the presence of side branches like D and OM. As a result, the individual graph representation of these main branches consists of multiple nodes, all of which share the same semantic labels. This introduces a unique challenge in the graph matching task, as it transforms into a one-to-many or many-to-many mapping problem. Unlike the simpler one-to-one or one-to-zero mappings, the presence of multiple possible mappings between nodes significantly increases the complexity of the search space [36,37]. To simplify the complexity of the graph matching task, our approach focuses on one-to-one graph matching. We handle the issue of multiple nodes representing the separated main branches by assigning semantic labels with increasing indices along the blood flow. By assigning incremental semantic labels, we ensure that each node in the individual graph representing a main branch segment has a unique label. This allows us to establish a one-to-one correspondence between the nodes in the individual graphs, simplifying the graph matching process. For example, if LCX and OM1 exist, then LCX is separated into two arterial segments with the semantic labels of LCX1 and LCX2. Then, each arterial segment in 𝒢 1 is only matched with one exact segment in 𝒢 2 .\\nFor model training, a batch of individual graph pairs is selected, and the association graphs are used as the input of the EAGMN. In our study, we specifically focus on ICAs obtained from two different view angles: left anterior oblique (LAO) and right anterior oblique (RAO). We acknowledge that the anatomy and visual characteristics of the coronary arteries can differ between these two view angles. To ensure consistency and accuracy in our analysis, we select individual graphs exclusively from ICAs captured using the same view angle. This allows us to maintain a more homogenous dataset for our analysis and ensure reliable results in coronary artery semantic labeling. The mean squared error (MSE) between the predicted vertex classification probability and the ground truth is used to optimize the model, as defined in Eq. 12.\\n𝐿 = 𝑀𝑆𝐸(𝑀, 𝑀 ̂) = ∑ ∑(𝑀 𝑖𝑎 -𝑀 ̂𝑖𝑎 ) 2 𝑛 2 𝑎=1 𝑛 1 𝑖=1(12)\\nwhere 𝑀 𝑖𝑎 = 𝑦 𝑖𝑎 is the ground truth and 𝑀 ̂𝑖𝑎 = 𝑦 ̂𝑖𝑎 is the predicted vertex class. An Adam optimizer with the learning rate of 0.0001 was used to optimize the model. The training algorithm is shown in Algorithm 1.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Algorithm 1.\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The training process of our proposed EAGMN.\\nDuring the model testing phase, we employ a voting strategy to assign semantic labels to artery branches in the tested ICAs. Our proposed EAGMN establishes the node-to-node semantic correspondence between the individual graphs. During the testing, 𝒢 1 is generated using ICAs in the testing set and 𝒢 2 is generated using ICAs in the template set, where template set contains a set of representative ICAs selected by experienced cardiologists. We simulate the learning procedure by graph matching that the cardiologist learns the coronary artery anatomy by comparing the anatomy of the testing case with the reference cases in the template set. By applying the EAGMN to the association graph generated by the tested ICA and template ICA, the artery segments are matched, and the semantic labeling is achieved. Given the complex and diverse structure of coronary artery anatomy, clinical decisions are often made based on multiple ICA frames. Therefore, we perform graph matching between the tested ICA and every ICA in the template set. Each arterial segment of the tested ICA is matched to multiple arterial segments in the template set, and a majority voting strategy is employed to assign the label according to the labels of the matched arterial segments in the template set. This approach helps ensure robustness and accuracy in the final semantic labeling process. The testing algorithm is shown in Algorithm 2. 4. Embed vertex and edge features 𝑣̅ 𝑖𝑎 𝑒𝑚𝑏 and 𝑒̅ 𝑖𝑎,𝑗𝑏 𝑒𝑚𝑏 with Eq. 5.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Input\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"For 𝑢 = 1 ⋯ 𝑁 𝑎𝑡𝑡 do 5. Update edge features 𝑣̅ 𝑖𝑎 𝑔𝑎𝑡 using Eq. 6 and vertex features 𝑣̅ 𝑖𝑎 𝑔𝑎𝑡 with Eq. 7.\\n6. Calculate edge attention score 𝜃 𝑖𝑎,𝑗𝑏 with Eq. 8.\\nFor 𝑣 = 1 ⋯ 𝑁 𝑚𝑝 do 7. Update edge features using 𝑒 𝑖𝑎,𝑗𝑏 𝑔𝑐𝑛 with Eq. 9 and vertex features 𝑣 𝑖𝑎 𝑔𝑐𝑛 with Eq. 10 8. Perform vertex classification with Eq. 11. 9. Optimize EAGMN using Eq .12.\\nAlgorithm 2. The testing process of our proposed EAGMN.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Evaluation\",\n \"publication_ref\": [\n \"b15\"\n ],\n \"table_ref\": [],\n \"text\": \"The × 𝑛 𝑐 (16) where 𝑇𝑃 𝑐 , 𝑇𝑁 𝑐 , 𝐹𝑃 𝑐 and 𝐹𝑁 𝑐 represent the true positive, true negative, false positive, and false negative arterial segments, respectively. 𝐶 is the total number of classes. 𝑛 𝐶 is the number of arterial segments in class 𝐶, and 𝑛 is the total number of arterial segments.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Model interpretation\",\n \"publication_ref\": [\n \"b37\",\n \"b38\"\n ],\n \"table_ref\": [],\n \"text\": \"Interpreting the decisions made by EAGMN is crucial for building trust and confidence in its predictions, especially in clinical practice. Existing approaches for explaining both node and edge features in GNN focus on gradient-based approaches and perturbation-based methods [38]. Gradient-based methods suffer from saturation problems in that the model output changes minimally with respect to any input change; thus, we adopt a perturbation-based method, ZORRO [39], which compares the output variations with respect to different input perturbations to calculate the node and feature importance for our graph matching network. Given a generated association graph, ZORRO iteratively and recursively adds the important features and nodes according to the fidelity score, where the fidelity score measures the difference between the raw\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Input:\",\n \"publication_ref\": [\n \"b38\"\n ],\n \"table_ref\": [],\n \"text\": \"𝐷 𝑡𝑒 = {𝒢 𝑖 = (𝕍 𝑖 , 𝔼 model predictions and the predictions after masking out the important features and nodes [39]. Formally, the fidelity score is defined as\\n𝐹(𝕍 𝑠 , 𝔽 𝑠 ) = 1 𝑛 1 𝑛 2 ∑ ∑ 𝕀(𝑦 𝑖𝑎 𝑠 = 𝑦 ̂𝑖𝑎 ) 𝑛 2 𝑎=1 𝑛 1 𝑖=1 (17\\n)\\nwhere 𝕀 is the indicator function. 𝑦 ̂𝑖𝑎 is the original prediction of the model and 𝑦 𝑖𝑎 𝑠 is the prediction using the masked nodes by 𝕍 𝑠 and the masked features by 𝔽 𝑠 . 𝕍 𝑠 and 𝔽 𝑠 are binary vectors in which 1\\nindicates the node/feature is selected and vice versa. An explanation for GNN-based model indicates that using the masked features and the masked nodes in graph, the fidelity score measured by the new prediction and the original prediction reached 𝜏, i.e. 𝜏 ≥ 𝐹(𝕍 𝑠 , 𝔽 𝑠 ).\\nExplaining feature importance. In section 3.1, we manually designed 121 hand-crafted features. The vertex features are generated by the concatenation of the nodes in the individual graph and identifying the concatenated features for each vertex cannot explain the feature importance for the node in the individual graph. To explain the feature importance, a unified feature mask is applied to the feature of every node in the individual graphs, then the masked features are concatenated as the vertex features using Eq. 2. ZORRO masks out all features at the beginning of the explanation and gradually adding features with the highest fidelity scores until the fidelity score is reached at 𝜏. By applying ZORRO to all individual pairs generated by the testing set and the template set, the frequency of the selected important features indicates the feature importance used for graph matching.\\nExplaining node importance. For the original ZORRO, neighbor nodes are removed first and then gradually added to test the importance of the node by the improvement of the fidelity scores. The aim of explaining node importance is to interpret the importance of different arterial segments for graph matching. During the testing phase, the individual graph 𝒢 1 is the tested case and 𝒢 2 is the template ICA graph. However, our graph matching algorithm is applied to the association graph, examining the importance of vertex in the association graph cannot reflect the node importance. Thus, we modify the ZORRO to explain the node importance that at the beginning of the explanation, all nodes in 𝒢 1 is retained and all nodes in 𝒢 2 are removed. ZORRO iteratively adds the node from 𝒢 2 with the highest improvement of the fidelity score.\\nAccording to the improvement of the fidelity score, the importance of the node is obtained. The node importance in our approach reflects the significance of each arterial segment in the template ICAs for both graph matching and semantic labeling. By assigning importance scores to the nodes, we can capture the relevance and influence of each arterial segment in the template set. This allows us to prioritize and weight the contributions of different segments during the graph matching process.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiments and discussions\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Dataset\",\n \"publication_ref\": [\n \"b12\"\n ],\n \"table_ref\": [],\n \"text\": \"In this study, we manually annotated 204 and 59 ICAs from site 1 at The First Affiliated Hospital of Nanjing Medical University and site 2 at Chang Bing Show Chwan Memorial Hospital, respectively. In total, this retrospective study enrolled 263 ICA images. The detailed description of the image acquisition was illustrated in our previous work [13]. For each patient, a frame that was used for anatomical structure analysis in clinical practice was selected from the view video for semantic labeling. In this study, we only focus on semantic labeling for the main branches of LMA, LAD, and LCX, and the side branches of D and OM.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Implementation details\",\n \"publication_ref\": [\n \"b39\"\n ],\n \"table_ref\": [],\n \"text\": \"We implemented our EAGMN using TensorFlow and GraphNets [40]. The vertex feature embedding module 𝑓 𝑣 𝑒𝑚𝑏 , the edge feature embedding module 𝑓 𝑒 𝑒𝑚𝑏 , the graph convolution module 𝑓 𝑒 𝑔𝑐𝑛 , 𝑓 𝑒 𝑢 , 𝑓 𝑣 𝑔𝑐𝑛 and 𝑓 𝑣 𝑢 were implemented using two MLP layers with layer-wise instance normalization and the number of hidden units were set as 64. For the attention score calculation, the vertex feature embedding module 𝑔 𝑣 𝑒𝑚𝑏 , the edge feature embedding module 𝑔 𝑒 𝑒𝑚𝑏 , the graph attention module 𝑔 𝑒 𝑔𝑎𝑡 , 𝑔 𝑒 𝑢 , 𝑔 𝑣 𝑔𝑎𝑡 and 𝑔 𝑣 𝑢 were also implemented using two MLP layers with the 64 hidden unites. The edge attention score readout module 𝑔 𝑒 𝑜𝑢𝑡 and the vertex classification layer 𝑓 𝑐𝑙𝑓 were implemented using two MLP layers with ReLU activation functions. The number of message-passing steps for graph attention score calculation, 𝑁 𝑎𝑡𝑡 , was set as 3 and the number of message-passing steps for graph convolution, 𝑁 𝑚𝑝 , was set as 2.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Coronary artery semantic labeling performance\",\n \"publication_ref\": [\n \"b10\",\n \"b16\",\n \"b17\",\n \"b23\",\n \"b24\",\n \"b12\",\n \"b23\",\n \"b24\",\n \"b12\"\n ],\n \"table_ref\": [\n \"tab_6\",\n \"tab_8\",\n \"tab_8\"\n ],\n \"text\": \"Our ICA dataset contains 263 annotated ICA graphs, and 79 of them were selected as the template set. The remining 184 ICAs were used for five-fold cross validation, which indicates that each fold 147 ICAs were used as the training set and 37 ICAs were used as the testing set. We trained our EAGMN for 𝑁 = 100,000 steps. The averaged performance on the five-folds were reported. The achieved performance of our EAGMN for coronary artery semantic labeling is shown in Table 1. We also compared the EAGMN with three existing segment identification based semantic labeling methods.\\nIn addition, we implemented 3 graph matching neural networks and applied them into coronary artery semantic labeling task using the same training and testing algorithms developed in this study.\\n• SVM [11]: In our previous work, a support vector machine was employed to classify coronary artery segments using the 20 position features and 2 topology features as described in section 3.1.\\n• BiTreeLSTM [17]: The bidirectional tree LSTM (BiTreeLSTM) was initially applied for coronary artery semantic labeling using CCTA according to the spatial locations and directions of arteries in 3D. We adopted the same network architecture but extracted the spatial locations and directions of coronary arteries in 2D. • CPR-GCN [18]: CPR-GCN was proposed to perform coronary artery semantic labeling using CCTA with 3D convolution LSTM for arterial imaging feature extraction and GCN for artery semantic labeling. We adopted the same network architecture but extracted the arterial imaging features using 2D convolution LSTM. • IPCA [24]: Iterative Permutation loss and Cross-graph Affinity (IPCA) for graph matching employed the GNN for both intra-graph node embedding and cross-graph node embedding iteratively for graph matching using individual graphs. • NGM [25]: Neural Graph Matching (NGM) network employed the association graph induced affinity matrix for embedding learning. In our EAGMN, the connectivity of the association was generated by the connectivity of the individual graphs; however, NGM adopted the Koopmans-Beckmann's QAP for affinity learning to calculate the assignment matrix. • AGMN [13]: Our previous work adopted the association graph for graph matching without the graph attention module. For each baseline, we performed five-fold cross validation, and the performance comparisons of the coronary artery semantic labeling are illustrated in Table 2. According to Table 2, the proposed EAGMN achieved the highest accuracy of coronary artery semantic labeling for 0.8653. However, the BiTreeLSTM baseline achieved stable performance in labeling LMA branches. For BiTreeLSTM and CPR-GCN models, they were initially designed for coronary artery semantic labeling using CCTA datasets. Though they showed impressive performance with an ACC greater than 0.9, they may not be suitable for coronary artery semantic labeling using ICAs. Understanding the relationships between arteries and their spatial orientation requires a three-dimensional perspective; however, coronary artery labeling using 2D images is challenging and these two methods cannot achieve satisfactory performance. In detail, the CPR-GCN achieved a weighted ACC of 0.4581 while the BiTreeLSTM achieved an ACC of 0.7492, which were significantly lower than the graph matching based methods. The results indicate that the complex branching patterns and variations in anatomy make it more challenging to understand coronary artery anatomy, and learning the ICA anatomy using individual graphs is difficult.\\nFor the graph matching based methods, we use IPCA, NGM and AGMN as baselines. IPCA [24] first performs feature embedding using individual graphs, and then performs cross-graph feature embeddings using GCNs. Compared to association graph-based graph matching algorithms, individual graph-based matching methods are computationally efficient and suitable for simpler graphs but may struggle with complex graph structures. IPCA achieved an ACC of 0.8039, which was lower than the association graphbased methods. NGM [25] first embeds features of nodes of the individual graphs to calculate the affinity matrix, and uses the calculated affinity matrix as the adjacency matrix for the association graph. Then, NGM performs the feature embedding according to the node features for vertex classification.\\nOn the contrary, AGMN [13] and our proposed EAGMN concatenate the node features of the individual graphs and perform the node embedding using the concatenated features. The adjacency matrix of the association graph is generated by the connectivity of the individual graphs rather than the affinity matrix used by NGM. AGMN and EAGMN achieved the ACC of 0.8264 and 0.8653, which were higher than that of NGM. Experimental results indicate that the physical connectivity of the coronary artery plays an important role in graph matching, where the physical connectivity refers to the actual anatomical connections between different arterial segments within the coronary artery network. The physical connectivity of the coronary artery provides valuable cues and constraints that guide the matching process. It helps ensure that the matched arterial segments are not only similar in appearance or position but also consistent with the underlying anatomical structure. Compared to AGMN, our proposed EAGMN employs the edge attention mechanisms to aggregate information from the adjacent edges dynamically. By employing edge attention mechanisms, EAGMN can adaptively prioritize and weight the importance of different edges based on their relevance to the graph matching task. This dynamic attention mechanism allows the model to focus on the most informative and discriminative edges while suppressing the influence of less relevant edges. Experimental results indicate that EAGMN improved the ACC by 0.0389 for all types of arteries compared to that of AGMN, which outperformed other baseline models significantly.\"\n },\n {\n \"figure_ref\": [\n \"fig_3\",\n \"fig_3\"\n ],\n \"heading\": \"Robustness test\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In clinical practice, due to the imaging quality, artery overlapping and human subjective annotation, the algorithm proposed in section 3.1 may not generate a perfect coronary arterial graph. To reflect the robustness of the proposed EAGMN and test the clinical flexibility, we performed the robustness test using the corrupted individual graphs with random arterial segment dropping. Note that the graph-based approaches use the topology information and the connected graphs as the input, thus, we only randomly removed the arterial branches with at least one endpoint during the robustness test. In addition, the LMA was retained in all experiments due to the importance of this in-let artery segmentation for the coronary artery anatomy. We set the probability of 5%, 7.5%, 10%, 12.5%, 15%, 17.5% and 20% to randomly remove the arteries, and the averaged performance for the robustness test among baselines and EAGMN of ICAs in testing set is shown in Figure 4. According to Figure 4, the proposed EAGMN was robust since the achieved ACCs were greater than 0.81 during these robust testing experiments, even with 20% missing arterial segments. The BiTreeLSTM and CPR-GCN showed a minor performance dropping, which indicated that these two methods did not fully utilize the graph structure of coronary artery. In comparison to the graph matching-based methods, such as IPCA, NGM, and AGMN, our proposed EAGMN consistently demonstrated superior performance in terms of ACC, PREC, REC, and F1 across various missing probabilities. These results highlight the robustness of the proposed EAGMN in handling missing data scenarios.\"\n },\n {\n \"figure_ref\": [\n \"fig_4\",\n \"fig_5\",\n \"fig_5\"\n ],\n \"heading\": \"Result explanation\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We adopted 37 ICAs as the testing set and 70 ICAs as the template set. Each ICA in testing set was used as 𝒢 1 to perform graph matching with every ICA as 𝒢 2 in template set. Note that number of nodes in 𝒢 1 is smaller than 𝒢 2 , and 1148 individual graph matching pairs were generated. We applied ZORRO to explain each graph matching pair to iteratively and recursively calculate the feature and node importance for graph matching. The threshold of the fidelity score used to measure the new prediction and the original prediction, 𝜏, was set as 0.8. We ranked the frequency of the selected important features to measure the feature importance, as shown in Figure 5. For explaining the node importance, we visualized the graph matching results and the fidelity score improvement when adding the node in the 𝒢 2 into the node mask. Example results are shown in Figure 6.\\nIn Figure 6, the top four pairs of ICA graphs demonstrate a successful and accurate matching with only minor errors. However, the bottom two pairs do not exhibit a satisfactory level of matching. Upon analyzing the fidelity scores depicted in the pseudo color visualization, we observed that the LMA branch plays a crucial role in the identification of coronary arteries. The successful identification of the LMA branch, along with the first segment of LAD (LAD1) and the first segment of LCX (LCX1), is of paramount importance for the semantic labeling of the entire vascular graph.\\nIn contrast, for the bottom two pairs of graphs, the significance of the LMA branch was diminished compared to the top four pairs, leading to a degradation in performance. The analysis of the node importance further confirms that our EAGMN closely mimics the approach of cardiologists in identifying coronary arteries. It starts by focusing on the main LMA branch and subsequently considers the two most vital subbranches, namely LAD1 and LCX1, in order to achieve accurate semantic labeling. Overall, these findings highlight the importance of the LMA branch and the initial segments of LAD and LCX in the graph matching process. They demonstrate that our EAGMN model effectively emulates the decisionmaking process followed by cardiologists when identifying and labeling coronary arteries. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this paper, we present an edge attention graph matching network for coronary artery semantic labeling. By performing the graph matching between individual ICA generated graphs, the relationship between the arterial segments is obtained and the unlabeled coronary arterial segments are labeled by template ICAs. Experimental results showed that our approach is powerful and robust. By employing the ZORRO, we explained the graph matching results and improved the interpretability of coronary artery semantic labeling using our proposed approach.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Acknowledgement\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This research was supported in part by a research seed fund from Michigan Technological University Health Research Institute and an NIH grant (U19AG055373).\"\n }\n]"},"abstract":{"kind":"string","value":"Coronary artery disease (CAD) is one of the primary causes leading deaths worldwide. The presence of atherosclerotic lesions in coronary arteries is the underlying pathophysiological basis of CAD, and accurate extraction of individual arterial branches using invasive coronary angiography (ICA) is crucial for stenosis detection and CAD diagnosis. However, deep-learning-based models face challenges in generating semantic segmentation for coronary arteries due to the morphological similarity among different types of arteries. To address this challenge, we propose an innovative approach called the Edge Attention Graph Matching Network (EAGMN) for coronary artery semantic labeling. Inspired by the learning process of interventional cardiologists in interpreting ICA images, our model compares arterial branches between two individual graphs generated from different ICAs. We begin with extracting individual graphs based on the vascular tree obtained from the ICA. Each node in the individual graph represents an arterial segment, and the EAGMN aims to learn the similarity between nodes from the two individual graphs. By converting the coronary artery semantic segmentation task into a graph node similarity comparison task, identifying the node-to-node correspondence would assign semantic labels for each arterial branch. More specifically, The EAGMN utilizes the association graph constructed from the two individual graphs as input. A graph attention module is employed for feature embedding and aggregation, while a decoder generates the linear assignment for node-to-node semantic mapping. Based on the learned node-to-node relationships, unlabeled coronary arterial segments are classified using the labeled coronary arterial segments, thereby achieving semantic labeling. A dataset with 263 labeled ICAs is used to train and validate the EAGMN. Experimental results indicate the EAGMN achieved a weighted accuracy of 0.8653, a weighted precision of 0.8656, a weighted recall of 0.8653 and a weighted F1-score of 0.8643. Furthermore, we employ ZORRO to provide interpretability and explainability of the graph matching for artery semantic labeling. These findings highlight the potential of the EAGMN for accurate and efficient coronary artery semantic labeling using ICAs. By leveraging the inherent characteristics of ICAs and incorporating graph matching techniques, our proposed model provides a promising solution for improving CAD diagnosis and treatment."},"title":{"kind":"string","value":"Coronary Artery Semantic Labeling using Edge Attention Graph Matching Network"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 1 .1Figure 1. Workflow of the proposed EAGMN for coronary artery semantic labeling using EAGMN. (a) Coronary artery binary segmentation by our previous work, feature pyramid U-Net++ [14], and individual graph generation. (b) Coronary artery semantic labeling using EAGMN by comparing the artery-to-artery correspondence between the unlabeled artery in 𝒢 1 and labeled arteries in 𝒢 2 . The positive vertex in yellow and '1' in assignment matrix represent two arterial branches are matched and the type of the unlabeled artery is assigned. The highlights of this paper are shown below:","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"Figure 2 .2Figure 2. Pipeline of the individual graph generation. (a) Original selected ICA frame; (b) Vascular tree generated by FP-U-Net++; (c) Centerline extraction and key points detection; (d) Artery segments with semantic labels; (e) Connectivity of the individual graph; (f) Switched individual graph that each node represents an arterial segment with semantic label and each edge represents the connectivity of the arteries. Formally, the individual graph represented by an attributed undirected graph as 𝒢 = (𝕍, 𝔼, 𝒱, ℰ), where • 𝕍 = {𝑉 𝑖 } 𝑠. 𝑡. 𝑖 ∈ [1, ⋯ , 𝑛] represents node set, and |𝕍| = 𝑛. • 𝔼 = {𝐸 𝑖𝑗 } 𝑠. 𝑡. 𝑖, 𝑗 ∈ [1, ⋯ , 𝑛] indicates the set of edges and |𝔼| = 𝑛 𝑒 . • 𝒱 = {𝑣 𝑖 } 𝑠. 𝑡. 𝑖 ∈ [1, ⋯ , 𝑛] indicates the attribute vectors associated with each node. • ℰ = {𝑒 𝑖𝑗 } s.t. 𝑖, 𝑗 ∈ [1, ⋯ , 𝑛] indicates the attribute vectors associated with each edge.","figure_data":"","figure_id":"fig_1","figure_label":"2","figure_type":"figure"},{"figure_caption":"Figure 3 .3Figure 3. Architecture of EAGMN for coronary artery semantic labeling. (a) Individual graph generation and feature extraction; (b) Association graph generation and feature extraction; (c) Vertex and edge feature embedding; (d) Edge attention score calculation using GCN; (e) Feature representation learning according to edge attention score derived in (d); (f) Vertex classification according to the learned feature representations.","figure_data":"","figure_id":"fig_2","figure_label":"3","figure_type":"figure"},{"figure_caption":"Figure 4 .4Figure 4. Robustness test of the coronary artery semantic labeling among baselines and the proposed EAGMN. The horizontal axis indicates the probability of dropping an artery segment and the vertical axis represents its corresponding performance.","figure_data":"","figure_id":"fig_3","figure_label":"4","figure_type":"figure"},{"figure_caption":"Figure 5 .5Figure 5. Feature importance explained by modified ZORRO. The vertical axis represents the name of the hand-crafted features, and the horizontal axis represents the frequency of the selected features when the explaination was reached. According to Figure 5. The top 15 features with the most importance are topology-based features, i.e. p2_degree and p1_degree, and position-based features. The p1_degree and p2_degree represent the degree of the endpoint of the left and right end of the arterial segment. The results indicate that the topology is an extremely important factor in arterial identification. The position features capture the absolute and relative position of the arterial segment in relation to the overall vascular tree, which also provide concrete information for artery semantic labeling. This information helps in distinguishing and identifying the individual arterial segments based on their unique spatial locations.","figure_data":"","figure_id":"fig_4","figure_label":"5","figure_type":"figure"},{"figure_caption":"Figure 6 .6Figure 6. Visualization of graph matching results and the improvement of the fidelity score when adding the arterial segment in template graph for graph matching. The green line indicates a correct match, and the red line represents an error.","figure_data":"","figure_id":"fig_5","figure_label":"6","figure_type":"figure"},{"figure_caption":"𝑛 2 ]) from 𝒢 2 exist, then edge 𝐸 𝑖𝑎,𝑗𝑏 is constructed in 𝒢 𝐴 . As such, |𝔼 𝐴 | = 2 × 𝑛 𝑒1 × 𝑛 𝑒2 . If the node 𝑉 𝑖 ∈ 𝒢 1 and node 𝑉 𝑎 ∈ 𝒢","figure_data":"1 2 ), (𝑉 1 1 , 𝑉 2 2 ), ⋯ , (𝑉 𝑛 1 1 , 𝑉 𝑛 2","figure_id":"tab_0","figure_label":"","figure_type":"table"},{"figure_caption":"where 𝑔 𝑣 𝑔𝑎𝑡 is an MLP encoder which aggregates the updated edge embeddings in Eq. 6 and 𝐴(𝑉 𝑖𝑎 )represents the set of the connected vertices of vertex 𝑉 𝑖𝑎 . The summation indicates a non-parametric function for feature aggregation. 𝑔 𝑣 𝑢 is another MLP encoder which updates vertex features according to aggregated edge features and node embedding 𝑣̅ 𝑖𝑎 𝑒𝑚𝑏 defined in Eq. 5.By iteratively applying Eqs. 6 and 7 for 𝑁 𝑎𝑡𝑡 times, the edge features are updated according to both the local and global edge and vertex features. Then, a readout module is employed to calculate the edge attention score, as defined in Eq. 8. 𝑔 𝑒 𝑜𝑢𝑡 is an MLP decoder to calculate the attention score of edge 𝐸 𝑖𝑗,𝑎𝑏 ∈ 𝒢 𝐴 , which is denoted as 𝜃 𝑖𝑎,𝑗𝑏 ∈ ℝ. The readout module takes the aggregated features from the previous step and applies further transformations to capture the importance or relevance of each edge. The edge attention score plays a crucial role in guiding the subsequent steps of the model. It helps prioritize the edges that contribute the most to the overall graph structure and assists in making accurate and meaningful vertex classification predictions.5) Feature representation learning using GCN. The feature representation learning module employs a GCN and the calculated graph attention score in Eq. 8. Similarly, this GCN also contains an edge convolution layer and a vertex convolution layer. Formally, edge convolution layer and vertex convolution layer for feature representation learning are defined in Eqs. 9 and 10.","figure_data":"𝑣̅ 𝑖𝑎 𝑔𝑎𝑡 ← 𝑔 𝑣 𝑢 ([𝑣̅ 𝑖𝑎 𝑔𝑎𝑡 , 𝑣̅ 𝑖𝑎 𝑒𝑚𝑏 ])𝜃 𝑖𝑎,𝑗𝑏 = 𝑔 𝑒 𝑜𝑢𝑡 (𝑒̅ 𝑖𝑎,𝑗𝑏 𝑔𝑎𝑡 )(8)𝑔𝑐𝑛 = 𝑓 𝑒 𝑔𝑐𝑛 ([𝑣 𝑖𝑎 𝑒𝑚𝑏 , 𝑣 𝑗𝑏 𝑒𝑚𝑏 ]) where the 𝑒 𝑖𝑎,𝑗𝑏 𝑒 𝑖𝑎,𝑗𝑏 𝑔𝑐𝑛 ← 𝑓 𝑒 𝑢 ([exp (-𝜃 𝑖𝑎,𝑗𝑏 ) ⋅ 𝑒 𝑖𝑎,𝑗𝑏 𝑔𝑐𝑛 , exp (-𝜃 𝑖𝑎,𝑗𝑏 ) ⋅ 𝑒 𝑖𝑎,𝑗𝑏 𝑒𝑚𝑏 ])(9)𝑣 𝑖𝑎 𝑔𝑐𝑛 =∑𝑓 𝑣 𝑔𝑐𝑛 (exp (-𝜃 𝑖𝑎,𝑗𝑏 ) ⋅ 𝑒 𝑖𝑎,𝑗𝑏 𝑔𝑐𝑛 )∀𝑉 𝑗𝑏 ∈𝐴(𝑉 𝑖𝑎 )(10)𝑣 𝑖𝑎 𝑔𝑐𝑛 ← 𝑓 𝑣 𝑢 ([𝑣 𝑖𝑎 𝑔𝑐𝑛 , 𝑣 𝑖𝑎 𝑒𝑚𝑏 ])where 𝑓 𝑒 𝑔𝑐𝑛 , 𝑓 𝑒 𝑢 , 𝑓 𝑣 𝑔𝑐𝑛 and 𝑓 𝑣𝑦 ̂𝑖𝑎 = 𝑓 𝑐𝑙𝑓 (𝑣 𝑖𝑎 𝑔𝑐𝑛 )𝑖𝑎 𝑔𝑎𝑡 =∑𝑔 𝑣 𝑔𝑎𝑡 (𝑒̅ 𝑖𝑎,𝑗𝑏 𝑔𝑎𝑡 )(7)∀𝑉 𝑗𝑏 ∈𝐴(𝑉 𝑖𝑎 )","figure_id":"tab_2","figure_label":"","figure_type":"table"},{"figure_caption":"𝐷 𝑡𝑟 = {𝒢 𝑖 = (𝕍 𝑖 , 𝔼 𝑖 , 𝒱 𝑖 , ℰ 𝑖 )} 𝑖=1 𝑛 𝑡𝑟 : training set contains 𝑛 𝑡𝑟 individual graphs with labeled arteries 𝑁 𝑎𝑡𝑡 : number of message-passing steps for graph attention score calculation 𝑁 𝑚𝑝 : number of message-passing steps for graph convolution 𝑁: number of the training iterations Random select two individual graphs 𝒢 1 and 𝒢 2 from the same view angles. 2. Construct association graph 𝒢 𝐴 using 𝒢 1 and 𝒢 2 with Eqs. 2 and 3. 3. Perform feature embedding for 𝑣 𝑖𝑎 𝑒𝑚𝑏 and 𝑒 𝑖𝑎,𝑗𝑏 𝑒𝑚𝑏 with Eq. 4 .","figure_data":"Output:Φ: Trained EAGMNFor 𝑖𝑡𝑒𝑟 = 1 ⋯ 𝑁 do1.","figure_id":"tab_3","figure_label":"","figure_type":"table"},{"figure_caption":"𝑖 , 𝒱 𝑖 , ℰ 𝑖 )} 𝑖=1 𝑛 𝑡𝑒 : testing set contains 𝑛 𝑡𝑒 individual graphs without labeled arteries. 𝐷 𝑡𝑝 = {𝒢 𝑗 = (𝕍 𝑗 , 𝔼 𝑗 , 𝒱 𝑗 , ℰ 𝑗 )} 𝑗=1 𝑛 𝑡𝑝 : template set contains 𝑛 𝑡𝑝 individual graphs with labeled arteries. Labels for each segment of 𝑛 𝑡𝑒 ICA graphs in 𝐷 𝑡𝑒 For 𝑖 = 1 ⋯ 𝑛 𝑡𝑒 do For 𝑖 = 𝑗 ⋯ 𝑛 𝑡𝑝 do If |𝕍 𝑖 | ≤ |𝕍 𝑗 | and 𝒢 𝑖 and 𝒢 𝑗 are from the same view angles, then 1. Construct association graph 𝒢 𝐴 using 𝒢 𝑖 and 𝒢 𝑗 with Eqs. 2 and 3 2. Calculate assignment matrix 𝑀 ̂𝑖𝑗 = Φ(𝒢 𝐴 ) 3. Assign labels to 𝒢 𝑖 according to the majority voting among set of the 𝑀 ̂𝑖𝑗 , s.t. 𝑗 ∈ [1, ⋯ 𝑛 𝑡𝑝 ]","figure_data":"Φ: trained AGAMNOutput:","figure_id":"tab_5","figure_label":"","figure_type":"table"},{"figure_caption":"Achieved performance for coronary artery semantic labeling using the proposed EAGMN.","figure_data":"Artery typeACCPRECRECF1-scoreLMA1.0000±0.00000.9841±0.01301.0000±0.00000.9919±0.0066LAD0.8905±0.06610.8692±0.04910.8905±0.06610.8795±0.0563LCX0.8831±0.06520.8444±0.04050.8831±0.06520.8629±0.0497D0.8142±0.07140.8335±0.08210.8142±0.07140.8233±0.0741OM0.7597±0.04080.8472±0.07690.7597±0.04080.7999±0.0505weighted average0.8653±0.04880.8656±0.04990.8653±0.04880.8643±0.0485As illustrated in","figure_id":"tab_6","figure_label":"1","figure_type":"table"},{"figure_caption":"our EAGMN achieved an average accuracy of 0.8653, which indicating 86.53% of the arterial segments in the testing set were correctly classified according to the trained EAGMN and the selected ICAs in the template set.","figure_data":"","figure_id":"tab_7","figure_label":"1","figure_type":"table"},{"figure_caption":"Comparison of coronary artery semantic labeling between the proposed EAGMN and existing methods using our ICA dataset. The achieved highest performance is shown in bold.","figure_data":"MethodMetricLMALADLCXDOMmeanML0.9925±0.01510.6331±0.05400.6388±0.03900.6147±0.04100.5907±0.04190.6651±0.0080BiTreeLSTM1.0000±0.00000.8845±0.01500.9871±0.01200.0000±0.00000.5981±0.01650.7492±0.0085CPR-GCN0.5361±0.29960.5319±0.12390.5072±0.14470.0624±0.09530.5341±0.30450.4581±0.0536IPCAACC0.9820±0.02200.8222±0.06990.7837±0.06340.7924±0.08570.7140±0.06760.8039±0.0537NGM0.9953±0.00930.8316±0.05860.8154±0.07390.7629±0.08700.7105±0.10360.8130±0.0636AGMN0.9956±0.00890.8432±0.03060.8046±0.04520.7956±0.04120.7565±0.08250.8264±0.0302EAGMN1.0000±0.00000.8905±0.06610.8831±0.06520.8142±0.07140.7597±0.04080.8653±0.0488ML0.9778±0.00710.6586±0.01740.6375±0.03780.5554±0.01010.6278±0.02610.6679±0.0081BiTreeLSTM1.0000±0.00000.8562±0.01900.5853±0.00990.0000±0.00000.9808±0.01220.6927±0.0074CPR-GCN0.6208±0.32400.5675±0.05400.3964±0.05700.2802±0.37270.3821±0.01390.4463±0.1075IPCAPREC0.9779±0.01350.7865±0.06340.8008±0.05520.7233±0.05720.7932±0.06730.8046±0.0545NGM0.9910±0.01100.8269±0.07060.7985±0.06690.7546±0.06880.7492±0.08660.8122±0.0644AGMN0.9911±0.01090.8476±0.04810.8256±0.03070.7536±0.04930.7613±0.03190.8276±0.0298EAGMN0.9841±0.01300.8692±0.04910.8444±0.04050.8335±0.08210.8472±0.07690.8656±0.0499ML0.9925±0.01510.6331±0.05400.6388±0.03900.6147±0.04100.5907±0.04190.6651±0.0080BiTreeLSTM1.0000±0.00000.8845±0.01500.9871±0.01200.0000±0.00000.5981±0.01650.7492±0.0085CPR-GCN0.5361±0.29960.5319±0.12390.5072±0.14470.0624±0.09530.5341±0.30450.4581±0.0536IPCAREC0.9820±0.02200.8222±0.06990.7837±0.06340.7924±0.08570.7140±0.06760.8039±0.0537NGM0.9953±0.00930.8316±0.05860.8154±0.07390.7629±0.08700.7105±0.10360.8130±0.0636AGMN0.9956±0.00890.8432±0.03060.8046±0.04520.7956±0.04120.7565±0.08250.8264±0.0302EAGMN1.0000±0.00000.8905±0.06610.8831±0.06520.8142±0.07140.7597±0.04080.8653±0.0488ML0.9850±0.00760.6437±0.02130.6360±0.00870.5832±0.02340.6071±0.01830.6646±0.0077BiTreeLSTM1.0000±0.00000.8699±0.01010.7348±0.00930.0000±0.00000.7429±0.01410.6967±0.0085CPR-GCN0.5698±0.30260.5455±0.08990.4353±0.06320.0742±0.09570.3924±0.16600.4192±0.0661IPCAF10.9798±0.01330.8038±0.06540.7921±0.05910.7559±0.06950.7510±0.06450.8033±0.0538NGM0.9932±0.00920.8290±0.06370.8067±0.06990.7583±0.07600.7289±0.09460.8123±0.0640AGMN0.9933±0.00890.8452±0.03860.8143±0.03100.7736±0.04240.7569±0.05080.8262±0.0301EAGMN0.9919±0.00660.8795±0.05630.8629±0.04970.8233±0.07410.7999±0.05050.8643±0.0485","figure_id":"tab_8","figure_label":"2","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 1 .1Figure 1. Workflow of the proposed EAGMN for coronary artery semantic labeling using EAGMN. (a) Coronary artery binary segmentation by our previous work, feature pyramid U-Net++ [14], and individual graph generation. (b) Coronary artery semantic labeling using EAGMN by comparing the artery-to-artery correspondence between the unlabeled artery in 𝒢 1 and labeled arteries in 𝒢 2 . The positive vertex in yellow and '1' in assignment matrix represent two arterial branches are matched and the type of the unlabeled artery is assigned. The highlights of this paper are shown below:\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 2 .2Figure 2. Pipeline of the individual graph generation. (a) Original selected ICA frame; (b) Vascular tree generated by FP-U-Net++; (c) Centerline extraction and key points detection; (d) Artery segments with semantic labels; (e) Connectivity of the individual graph; (f) Switched individual graph that each node represents an arterial segment with semantic label and each edge represents the connectivity of the arteries. Formally, the individual graph represented by an attributed undirected graph as 𝒢 = (𝕍, 𝔼, 𝒱, ℰ), where • 𝕍 = {𝑉 𝑖 } 𝑠. 𝑡. 𝑖 ∈ [1, ⋯ , 𝑛] represents node set, and |𝕍| = 𝑛. • 𝔼 = {𝐸 𝑖𝑗 } 𝑠. 𝑡. 𝑖, 𝑗 ∈ [1, ⋯ , 𝑛] indicates the set of edges and |𝔼| = 𝑛 𝑒 . • 𝒱 = {𝑣 𝑖 } 𝑠. 𝑡. 𝑖 ∈ [1, ⋯ , 𝑛] indicates the attribute vectors associated with each node. • ℰ = {𝑒 𝑖𝑗 } s.t. 𝑖, 𝑗 ∈ [1, ⋯ , 𝑛] indicates the attribute vectors associated with each edge.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 .3Figure 3. Architecture of EAGMN for coronary artery semantic labeling. (a) Individual graph generation and feature extraction; (b) Association graph generation and feature extraction; (c) Vertex and edge feature embedding; (d) Edge attention score calculation using GCN; (e) Feature representation learning according to edge attention score derived in (d); (f) Vertex classification according to the learned feature representations.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 4 .4Figure 4. Robustness test of the coronary artery semantic labeling among baselines and the proposed EAGMN. The horizontal axis indicates the probability of dropping an artery segment and the vertical axis represents its corresponding performance.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 5 .5Figure 5. Feature importance explained by modified ZORRO. The vertical axis represents the name of the hand-crafted features, and the horizontal axis represents the frequency of the selected features when the explaination was reached. According to Figure 5. The top 15 features with the most importance are topology-based features, i.e. p2_degree and p1_degree, and position-based features. The p1_degree and p2_degree represent the degree of the endpoint of the left and right end of the arterial segment. The results indicate that the topology is an extremely important factor in arterial identification. The position features capture the absolute and relative position of the arterial segment in relation to the overall vascular tree, which also provide concrete information for artery semantic labeling. This information helps in distinguishing and identifying the individual arterial segments based on their unique spatial locations.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"5\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 6 .6Figure 6. Visualization of graph matching results and the improvement of the fidelity score when adding the arterial segment in template graph for graph matching. The green line indicates a correct match, and the red line represents an error.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_5\",\n \"figure_label\": \"6\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"𝑛 2 ]) from 𝒢 2 exist, then edge 𝐸 𝑖𝑎,𝑗𝑏 is constructed in 𝒢 𝐴 . As such, |𝔼 𝐴 | = 2 × 𝑛 𝑒1 × 𝑛 𝑒2 . If the node 𝑉 𝑖 ∈ 𝒢 1 and node 𝑉 𝑎 ∈ 𝒢\",\n \"figure_data\": \"1 2 ), (𝑉 1 1 , 𝑉 2 2 ), ⋯ , (𝑉 𝑛 1 1 , 𝑉 𝑛 2\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"where 𝑔 𝑣 𝑔𝑎𝑡 is an MLP encoder which aggregates the updated edge embeddings in Eq. 6 and 𝐴(𝑉 𝑖𝑎 )represents the set of the connected vertices of vertex 𝑉 𝑖𝑎 . The summation indicates a non-parametric function for feature aggregation. 𝑔 𝑣 𝑢 is another MLP encoder which updates vertex features according to aggregated edge features and node embedding 𝑣̅ 𝑖𝑎 𝑒𝑚𝑏 defined in Eq. 5.By iteratively applying Eqs. 6 and 7 for 𝑁 𝑎𝑡𝑡 times, the edge features are updated according to both the local and global edge and vertex features. Then, a readout module is employed to calculate the edge attention score, as defined in Eq. 8. 𝑔 𝑒 𝑜𝑢𝑡 is an MLP decoder to calculate the attention score of edge 𝐸 𝑖𝑗,𝑎𝑏 ∈ 𝒢 𝐴 , which is denoted as 𝜃 𝑖𝑎,𝑗𝑏 ∈ ℝ. The readout module takes the aggregated features from the previous step and applies further transformations to capture the importance or relevance of each edge. The edge attention score plays a crucial role in guiding the subsequent steps of the model. It helps prioritize the edges that contribute the most to the overall graph structure and assists in making accurate and meaningful vertex classification predictions.5) Feature representation learning using GCN. The feature representation learning module employs a GCN and the calculated graph attention score in Eq. 8. Similarly, this GCN also contains an edge convolution layer and a vertex convolution layer. Formally, edge convolution layer and vertex convolution layer for feature representation learning are defined in Eqs. 9 and 10.\",\n \"figure_data\": \"𝑣̅ 𝑖𝑎 𝑔𝑎𝑡 ← 𝑔 𝑣 𝑢 ([𝑣̅ 𝑖𝑎 𝑔𝑎𝑡 , 𝑣̅ 𝑖𝑎 𝑒𝑚𝑏 ])𝜃 𝑖𝑎,𝑗𝑏 = 𝑔 𝑒 𝑜𝑢𝑡 (𝑒̅ 𝑖𝑎,𝑗𝑏 𝑔𝑎𝑡 )(8)𝑔𝑐𝑛 = 𝑓 𝑒 𝑔𝑐𝑛 ([𝑣 𝑖𝑎 𝑒𝑚𝑏 , 𝑣 𝑗𝑏 𝑒𝑚𝑏 ]) where the 𝑒 𝑖𝑎,𝑗𝑏 𝑒 𝑖𝑎,𝑗𝑏 𝑔𝑐𝑛 ← 𝑓 𝑒 𝑢 ([exp (-𝜃 𝑖𝑎,𝑗𝑏 ) ⋅ 𝑒 𝑖𝑎,𝑗𝑏 𝑔𝑐𝑛 , exp (-𝜃 𝑖𝑎,𝑗𝑏 ) ⋅ 𝑒 𝑖𝑎,𝑗𝑏 𝑒𝑚𝑏 ])(9)𝑣 𝑖𝑎 𝑔𝑐𝑛 =∑𝑓 𝑣 𝑔𝑐𝑛 (exp (-𝜃 𝑖𝑎,𝑗𝑏 ) ⋅ 𝑒 𝑖𝑎,𝑗𝑏 𝑔𝑐𝑛 )∀𝑉 𝑗𝑏 ∈𝐴(𝑉 𝑖𝑎 )(10)𝑣 𝑖𝑎 𝑔𝑐𝑛 ← 𝑓 𝑣 𝑢 ([𝑣 𝑖𝑎 𝑔𝑐𝑛 , 𝑣 𝑖𝑎 𝑒𝑚𝑏 ])where 𝑓 𝑒 𝑔𝑐𝑛 , 𝑓 𝑒 𝑢 , 𝑓 𝑣 𝑔𝑐𝑛 and 𝑓 𝑣𝑦 ̂𝑖𝑎 = 𝑓 𝑐𝑙𝑓 (𝑣 𝑖𝑎 𝑔𝑐𝑛 )𝑖𝑎 𝑔𝑎𝑡 =∑𝑔 𝑣 𝑔𝑎𝑡 (𝑒̅ 𝑖𝑎,𝑗𝑏 𝑔𝑎𝑡 )(7)∀𝑉 𝑗𝑏 ∈𝐴(𝑉 𝑖𝑎 )\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"𝐷 𝑡𝑟 = {𝒢 𝑖 = (𝕍 𝑖 , 𝔼 𝑖 , 𝒱 𝑖 , ℰ 𝑖 )} 𝑖=1 𝑛 𝑡𝑟 : training set contains 𝑛 𝑡𝑟 individual graphs with labeled arteries 𝑁 𝑎𝑡𝑡 : number of message-passing steps for graph attention score calculation 𝑁 𝑚𝑝 : number of message-passing steps for graph convolution 𝑁: number of the training iterations Random select two individual graphs 𝒢 1 and 𝒢 2 from the same view angles. 2. Construct association graph 𝒢 𝐴 using 𝒢 1 and 𝒢 2 with Eqs. 2 and 3. 3. Perform feature embedding for 𝑣 𝑖𝑎 𝑒𝑚𝑏 and 𝑒 𝑖𝑎,𝑗𝑏 𝑒𝑚𝑏 with Eq. 4 .\",\n \"figure_data\": \"Output:Φ: Trained EAGMNFor 𝑖𝑡𝑒𝑟 = 1 ⋯ 𝑁 do1.\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"𝑖 , 𝒱 𝑖 , ℰ 𝑖 )} 𝑖=1 𝑛 𝑡𝑒 : testing set contains 𝑛 𝑡𝑒 individual graphs without labeled arteries. 𝐷 𝑡𝑝 = {𝒢 𝑗 = (𝕍 𝑗 , 𝔼 𝑗 , 𝒱 𝑗 , ℰ 𝑗 )} 𝑗=1 𝑛 𝑡𝑝 : template set contains 𝑛 𝑡𝑝 individual graphs with labeled arteries. Labels for each segment of 𝑛 𝑡𝑒 ICA graphs in 𝐷 𝑡𝑒 For 𝑖 = 1 ⋯ 𝑛 𝑡𝑒 do For 𝑖 = 𝑗 ⋯ 𝑛 𝑡𝑝 do If |𝕍 𝑖 | ≤ |𝕍 𝑗 | and 𝒢 𝑖 and 𝒢 𝑗 are from the same view angles, then 1. Construct association graph 𝒢 𝐴 using 𝒢 𝑖 and 𝒢 𝑗 with Eqs. 2 and 3 2. Calculate assignment matrix 𝑀 ̂𝑖𝑗 = Φ(𝒢 𝐴 ) 3. Assign labels to 𝒢 𝑖 according to the majority voting among set of the 𝑀 ̂𝑖𝑗 , s.t. 𝑗 ∈ [1, ⋯ 𝑛 𝑡𝑝 ]\",\n \"figure_data\": \"Φ: trained AGAMNOutput:\",\n \"figure_id\": \"tab_5\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Achieved performance for coronary artery semantic labeling using the proposed EAGMN.\",\n \"figure_data\": \"Artery typeACCPRECRECF1-scoreLMA1.0000±0.00000.9841±0.01301.0000±0.00000.9919±0.0066LAD0.8905±0.06610.8692±0.04910.8905±0.06610.8795±0.0563LCX0.8831±0.06520.8444±0.04050.8831±0.06520.8629±0.0497D0.8142±0.07140.8335±0.08210.8142±0.07140.8233±0.0741OM0.7597±0.04080.8472±0.07690.7597±0.04080.7999±0.0505weighted average0.8653±0.04880.8656±0.04990.8653±0.04880.8643±0.0485As illustrated in\",\n \"figure_id\": \"tab_6\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"our EAGMN achieved an average accuracy of 0.8653, which indicating 86.53% of the arterial segments in the testing set were correctly classified according to the trained EAGMN and the selected ICAs in the template set.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_7\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Comparison of coronary artery semantic labeling between the proposed EAGMN and existing methods using our ICA dataset. The achieved highest performance is shown in bold.\",\n \"figure_data\": \"MethodMetricLMALADLCXDOMmeanML0.9925±0.01510.6331±0.05400.6388±0.03900.6147±0.04100.5907±0.04190.6651±0.0080BiTreeLSTM1.0000±0.00000.8845±0.01500.9871±0.01200.0000±0.00000.5981±0.01650.7492±0.0085CPR-GCN0.5361±0.29960.5319±0.12390.5072±0.14470.0624±0.09530.5341±0.30450.4581±0.0536IPCAACC0.9820±0.02200.8222±0.06990.7837±0.06340.7924±0.08570.7140±0.06760.8039±0.0537NGM0.9953±0.00930.8316±0.05860.8154±0.07390.7629±0.08700.7105±0.10360.8130±0.0636AGMN0.9956±0.00890.8432±0.03060.8046±0.04520.7956±0.04120.7565±0.08250.8264±0.0302EAGMN1.0000±0.00000.8905±0.06610.8831±0.06520.8142±0.07140.7597±0.04080.8653±0.0488ML0.9778±0.00710.6586±0.01740.6375±0.03780.5554±0.01010.6278±0.02610.6679±0.0081BiTreeLSTM1.0000±0.00000.8562±0.01900.5853±0.00990.0000±0.00000.9808±0.01220.6927±0.0074CPR-GCN0.6208±0.32400.5675±0.05400.3964±0.05700.2802±0.37270.3821±0.01390.4463±0.1075IPCAPREC0.9779±0.01350.7865±0.06340.8008±0.05520.7233±0.05720.7932±0.06730.8046±0.0545NGM0.9910±0.01100.8269±0.07060.7985±0.06690.7546±0.06880.7492±0.08660.8122±0.0644AGMN0.9911±0.01090.8476±0.04810.8256±0.03070.7536±0.04930.7613±0.03190.8276±0.0298EAGMN0.9841±0.01300.8692±0.04910.8444±0.04050.8335±0.08210.8472±0.07690.8656±0.0499ML0.9925±0.01510.6331±0.05400.6388±0.03900.6147±0.04100.5907±0.04190.6651±0.0080BiTreeLSTM1.0000±0.00000.8845±0.01500.9871±0.01200.0000±0.00000.5981±0.01650.7492±0.0085CPR-GCN0.5361±0.29960.5319±0.12390.5072±0.14470.0624±0.09530.5341±0.30450.4581±0.0536IPCAREC0.9820±0.02200.8222±0.06990.7837±0.06340.7924±0.08570.7140±0.06760.8039±0.0537NGM0.9953±0.00930.8316±0.05860.8154±0.07390.7629±0.08700.7105±0.10360.8130±0.0636AGMN0.9956±0.00890.8432±0.03060.8046±0.04520.7956±0.04120.7565±0.08250.8264±0.0302EAGMN1.0000±0.00000.8905±0.06610.8831±0.06520.8142±0.07140.7597±0.04080.8653±0.0488ML0.9850±0.00760.6437±0.02130.6360±0.00870.5832±0.02340.6071±0.01830.6646±0.0077BiTreeLSTM1.0000±0.00000.8699±0.01010.7348±0.00930.0000±0.00000.7429±0.01410.6967±0.0085CPR-GCN0.5698±0.30260.5455±0.08990.4353±0.06320.0742±0.09570.3924±0.16600.4192±0.0661IPCAF10.9798±0.01330.8038±0.06540.7921±0.05910.7559±0.06950.7510±0.06450.8033±0.0538NGM0.9932±0.00920.8290±0.06370.8067±0.06990.7583±0.07600.7289±0.09460.8123±0.0640AGMN0.9933±0.00890.8452±0.03860.8143±0.03100.7736±0.04240.7569±0.05080.8262±0.0301EAGMN0.9919±0.00660.8795±0.05630.8629±0.04970.8233±0.07410.7999±0.05050.8643±0.0485\",\n \"figure_id\": \"tab_8\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n }\n]"},"authors":{"kind":"string","value":"Chen Zhao; Zhihui Xu; Guang-Uei Hung; Weihua Zhou"},"references":{"kind":"list like","value":[{"authors":"J S Lawton; J E Tamis-Holland; S Bangalore; E R Bates; T M Beckie; J M Bischoff; J A Bittl; M G Cohen; J M Dimaio; C W Don; S E Fremes; M F Gaudino; Z D Goldberger; M C Grant; J B Jaswal; P A Kurlansky; R Mehran; T S Metkus; L C Nnacheta; S V Rao; F W Sellke; G Sharma; C M Yong; B A Zwischenberger","journal":"Circulation","ref_id":"b0","title":"ACC/AHA/SCAI Guideline for Coronary Artery Revascularization: Executive Summary: A Report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines","year":"2021"},{"authors":"W E Boden; R A O'rourke; K K Teo; P M Hartigan; D J Maron; W J Kostuk; M Knudtson; M Dada; P Casperson; C L Harris; B R Chaitman; L Shaw; G Gosselin; S Nawaz; L M Title; G Gau; A S Blaustein; D C Booth; E R Bates; J A Spertus; D S Berman; G B J Mancini; W S Weintraub","journal":"N Engl J Med","ref_id":"b1","title":"Optimal Medical Therapy with or without PCI for Stable Coronary Disease","year":"2007"},{"authors":"The Discharge Trial Group; P Maurovich-Horvat; M Bosserdt; K F Kofoed; N Rieckmann; T Benedek; P Donnelly; J Rodriguez-Palomares; A Erglis; C Štěchovský; G Šakalyte; N Čemerlić Adić; M Gutberlet; J D Dodd; I Diez; G Davis; E Zimmermann; C Kępka; R Vidakovic; M Francone; M Ilnicka-Suckiel; F Plank; J Knuuti; R Faria; S Schröder; C Berry; L Saba; B Ruzsics; C Kubiak; I Gutierrez-Ibarluzea; K Schultz Hansen; J Müller-Nordhorn; B Merkely; A D Knudsen; I Benedek; C Orr; F Xavier; L Valente; V Zvaigzne; L Suchánek; F Zajančkauskiene; M Adić; M Woinke; I Hensey; E Lecumberri; M Thwaite; M Laule; A N Kruk; M Neskovic; D Mancone; G Kuśmierz; M Feuchtner; V Pietilä; T Gama Ribeiro; C Drosch; G Delles; M Matta; B Fisher; L Szilveszter; M Larsen; S Ratiu; B Kelly; A Garcia Del Blanco; Z D Rubio; B Drobni; I Jurlander; S Rodean; H Regan; M Cuéllar Calabria; T Boussoussou; R Engstrøm; A E Hodas; R Napp; S Haase; L M Feger; K Serna-Higuita; H Neumann; M Dreger; V Rief; M Wieske; P Estrella; M Martus; Dewey","journal":"N Engl J Med","ref_id":"b2","title":"CT or Invasive Coronary Angiography in Stable Chest Pain","year":"2022"},{"authors":"C Spadaccio; U Benedetto","journal":"Ann Cardiothorac Surg","ref_id":"b3","title":"Coronary artery bypass grafting (CABG) vs. percutaneous coronary intervention (PCI) in the treatment of multivessel coronary disease: quo vadis? -a review of the evidences on coronary artery disease","year":"2018"},{"authors":"Z Li; Y Zhang; G Liu; H Shao; W Li; X Tang","journal":"Biomedical Signal Processing and Control","ref_id":"b4","title":"A robust coronary artery identification and centerline extraction method in angiographies","year":"2015"},{"authors":"Z Xian; X Wang; S Yan; D Yang; J Chen; C Peng","journal":"Mathematical Problems in Engineering","ref_id":"b5","title":"Main Coronary Vessel Segmentation Using Deep Learning in Smart Medical","year":"2020"},{"authors":"N I Parikh; E F Honeycutt; M T Roe; M Neely; E J Rosenthal; M A Mittleman; J P Carrozza; K K L Ho","journal":"Circ: Cardiovascular Quality and Outcomes","ref_id":"b6","title":"Left and Codominant Coronary Artery Circulations Are Associated With Higher In-Hospital Mortality Among Patients Undergoing Percutaneous Coronary Intervention for Acute Coronary Syndromes: Report From the National Cardiovascular Database Cath Percutaneous Coronary Intervention (CathPCI) Registry","year":"2012"},{"authors":"W Austen; J Edwards; R Frye; G Gensini; V Gott; L Griffith; D Mcgoon; M Murphy; B Roe","journal":"American Heart Association, Circulation","ref_id":"b7","title":"A reporting system on patients evaluated for coronary artery disease. Report of the Ad Hoc Committee for Grading of Coronary Artery Disease, Council on Cardiovascular Surgery","year":"1975"},{"authors":"S Kastellanos; K Aznaouridis; C Vlachopoulos; E Tsiamis; E Oikonomou; D Tousoulis","journal":"WJC","ref_id":"b8","title":"Overview of coronary artery variants, aberrations and anomalies","year":"2018"},{"authors":"G Yang; A Broersen; R Petr; P Kitslaar; M A De Graaf; J J Bax; J H C Reiber; J Dijkstra","journal":"Computing in Cardiology","ref_id":"b9","title":"Automatic coronary artery tree labeling in coronary computed tomographic angiography datasets","year":"2011"},{"authors":"C Zhao; R Bober; H Tang; J Tang; M Dong; C Zhang; Z He; M Esposito; Z Xu; W Zhou","journal":"Journal of Advances in Applied & Computational Mathematics","ref_id":"b10","title":"Semantic Segmentation to Extract Coronary Arteries in Invasive Coronary Angiograms","year":"2022"},{"authors":"H Zhang; Z Gao; D Zhang; W K Hau; H Zhang","journal":"IEEE Transactions on Medical Imaging","ref_id":"b11","title":"Progressive Perception Learning for Main Coronary Segmentation in X-ray Angiography","year":"2022"},{"authors":"C Zhao; Z Xu; J Jiang; M Esposito; D Pienta; G.-U Hung; W Zhou","journal":"","ref_id":"b12","title":"AGMN: Association Graph-based Graph Matching Network for Coronary Artery Semantic Labeling on Invasive Coronary Angiograms","year":"2023-02-10"},{"authors":"C Zhao; A Vij; S Malhotra; J Tang; H Tang; D Pienta; Z Xu; W Zhou","journal":"Computers in Biology and Medicine","ref_id":"b13","title":"Automatic extraction and stenosis evaluation of coronary arteries in invasive coronary angiograms","year":"2021"},{"authors":"T J Jun; J Kweon; Y.-H Kim; D Kim; T-Net ","journal":"Neural Networks","ref_id":"b14","title":"Nested encoder-decoder architecture for the main vessel segmentation in coronary angiography","year":"2020"},{"authors":"Q Cao; A Broersen; M A De Graaf; P H Kitslaar; G Yang; A J Scholte; B P F Lelieveldt; J H C Reiber; J Dijkstra","journal":"Int J Cardiovasc Imaging","ref_id":"b15","title":"Automatic identification of coronary tree anatomy in coronary computed tomography angiography","year":"2017"},{"authors":"D Wu; X Wang; J Bai; X Xu; B Ouyang; Y Li; H Zhang; Q Song; K Cao; Y Yin","journal":"Int J CARS","ref_id":"b16","title":"Automated anatomical labeling of coronary arteries via bidirectional tree LSTMs","year":"2019"},{"authors":"H Yang; X Zhen; Y Chi; L Zhang; X.-S Hua","journal":"IEEE","ref_id":"b17","title":"CPR-GCN: Conditional Partial-Residual Graph Convolutional Network in Automated Anatomical Labeling of Coronary Arteries","year":"2020"},{"authors":"H Bunke","journal":"Pattern Recognition Letters","ref_id":"b18","title":"On a relation between graph edit distance and maximum common subgraph","year":"1997"},{"authors":"E L Lawler","journal":"Management Science","ref_id":"b19","title":"The quadratic assignment problem","year":"1963"},{"authors":"J Yan; S Yang; E Hancock","journal":"","ref_id":"b20","title":"Learning for Graph Matching and Related Combinatorial Optimization Problems","year":"2020"},{"authors":"H Dai; E B Khalil; Y Zhang; B Dilkina; L Song","journal":"","ref_id":"b21","title":"Learning Combinatorial Optimization Algorithms over Graphs","year":"2018-05-09"},{"authors":"A Nowak; S Villar; A S Bandeira; J Bruna","journal":"IEEE","ref_id":"b22","title":"Revised Note on Learning Quadratic Assignment with Graph Neural Networks","year":"2018"},{"authors":"R Wang; J Yan; X Yang","journal":"IEEE Trans. Pattern Anal. Mach. Intell","ref_id":"b23","title":"Combinatorial Learning of Robust Deep Graph Matching: an Embedding based Approach","year":"2020"},{"authors":"R Wang; J Yan; X Yang","journal":"IEEE Transactions on Pattern Analysis and Machine Intelligence","ref_id":"b24","title":"Neural Graph Matching Network: Learning Lawler's Quadratic Assignment Problem With Extension to Hypergraph and Multiple-Graph Matching","year":"2022"},{"authors":"E R Dougherty","journal":"SPIE, Optical Engineering Press","ref_id":"b25","title":"An introduction to morphological image processing","year":"1992"},{"authors":"J Xie; Y Zhao; Y Liu; P Su; Y Zhao; J Cheng; Y Zheng; J Liu","journal":"","ref_id":"b26","title":"Topology reconstruction of treelike structure in images via structural similarity measure and dominant set clustering","year":"2019"},{"authors":"T Wang; H Liu; Y Li; Y Jin; X Hou; H Ling","journal":"IEEE","ref_id":"b27","title":"Learning Combinatorial Solver for Graph Matching","year":"2020"},{"authors":"C Zhao; Y Xu; Z He; J Tang; Y Zhang; J Han; Y Shi; W Zhou","journal":"Pattern Recognition","ref_id":"b28","title":"Lung segmentation and automatic detection of COVID-19 using radiomic features from chest CT images","year":"2021"},{"authors":"J J M Van Griethuysen; A Fedorov; C Parmar; A Hosny; N Aucoin; V Narayan; R G H Beets-Tan; J.-C Fillion-Robin; S Pieper; H J W L Aerts","journal":"Cancer Res","ref_id":"b29","title":"Computational Radiomics System to Decode the Radiographic Phenotype","year":"2017"},{"authors":"D Ulyanov; A Vedaldi; V Lempitsky","journal":"","ref_id":"b30","title":"Instance normalization: The missing ingredient for fast stylization","year":"2016"},{"authors":"J Gilmer; S S Schoenholz; P F Riley; O Vinyals; G E Dahl","journal":"","ref_id":"b31","title":"Neural Message Passing for Quantum Chemistry","year":"2017"},{"authors":"W Hamilton; Z Ying; J Leskovec","journal":"Advances in Neural Information Processing Systems","ref_id":"b32","title":"Inductive representation learning on large graphs","year":"2017"},{"authors":"P Veličković; G Cucurull; A Casanova; A Romero; P Lio; Y Bengio","journal":"","ref_id":"b33","title":"Graph attention networks","year":"2017"},{"authors":"J Qu; H Ling; C Zhang; X Lyu; Z Tang","journal":"","ref_id":"b34","title":"Adaptive Edge Attention for Graph Matching with Outliers","year":"2021"},{"authors":"H Zhang; R Yanagi; R Togo; T Ogawa; M Haseyama","journal":"IEEE Access","ref_id":"b35","title":"Cross-Modal Image Retrieval Considering Semantic Relationships With Many-to-Many Correspondence Loss","year":"2023"},{"authors":"M Zaslavskiy; F Bach; J.-P Vert","journal":"","ref_id":"b36","title":"Many-to-Many Graph Matching: a Continuous Relaxation Approach","year":"2010-06-21"},{"authors":"H Yuan; H Yu; S Gui; S Ji","journal":"IEEE Trans. Pattern Anal. Mach. Intell","ref_id":"b37","title":"Explainability in Graph Neural Networks: A Taxonomic Survey","year":"2022"},{"authors":"T Funke; M Khosla; M Rathee; A Anand; Zorro ","journal":"IEEE Trans. Knowl. Data Eng","ref_id":"b38","title":"Valid, Sparse, and Stable Explanations in Graph Neural Networks","year":"2023"},{"authors":"P W Battaglia; J B Hamrick; V Bapst; A Sanchez-Gonzalez; V Zambaldi; M Malinowski; A Tacchetti; D Raposo; A Santoro; R Faulkner; C Gulcehre; F Song; A Ballard; J Gilmer; G Dahl; A Vaswani; K Allen; C Nash; V Langston; C Dyer; N Heess; D Wierstra; P Kohli; M Botvinick; O Vinyals; Y Li; R Pascanu","journal":"","ref_id":"b39","title":"Relational inductive biases, deep learning, and graph networks","year":"2018"}],"string":"[\n {\n \"authors\": \"J S Lawton; J E Tamis-Holland; S Bangalore; E R Bates; T M Beckie; J M Bischoff; J A Bittl; M G Cohen; J M Dimaio; C W Don; S E Fremes; M F Gaudino; Z D Goldberger; M C Grant; J B Jaswal; P A Kurlansky; R Mehran; T S Metkus; L C Nnacheta; S V Rao; F W Sellke; G Sharma; C M Yong; B A Zwischenberger\",\n \"journal\": \"Circulation\",\n \"ref_id\": \"b0\",\n \"title\": \"ACC/AHA/SCAI Guideline for Coronary Artery Revascularization: Executive Summary: A Report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"W E Boden; R A O'rourke; K K Teo; P M Hartigan; D J Maron; W J Kostuk; M Knudtson; M Dada; P Casperson; C L Harris; B R Chaitman; L Shaw; G Gosselin; S Nawaz; L M Title; G Gau; A S Blaustein; D C Booth; E R Bates; J A Spertus; D S Berman; G B J Mancini; W S Weintraub\",\n \"journal\": \"N Engl J Med\",\n \"ref_id\": \"b1\",\n \"title\": \"Optimal Medical Therapy with or without PCI for Stable Coronary Disease\",\n \"year\": \"2007\"\n },\n {\n \"authors\": \"The Discharge Trial Group; P Maurovich-Horvat; M Bosserdt; K F Kofoed; N Rieckmann; T Benedek; P Donnelly; J Rodriguez-Palomares; A Erglis; C Štěchovský; G Šakalyte; N Čemerlić Adić; M Gutberlet; J D Dodd; I Diez; G Davis; E Zimmermann; C Kępka; R Vidakovic; M Francone; M Ilnicka-Suckiel; F Plank; J Knuuti; R Faria; S Schröder; C Berry; L Saba; B Ruzsics; C Kubiak; I Gutierrez-Ibarluzea; K Schultz Hansen; J Müller-Nordhorn; B Merkely; A D Knudsen; I Benedek; C Orr; F Xavier; L Valente; V Zvaigzne; L Suchánek; F Zajančkauskiene; M Adić; M Woinke; I Hensey; E Lecumberri; M Thwaite; M Laule; A N Kruk; M Neskovic; D Mancone; G Kuśmierz; M Feuchtner; V Pietilä; T Gama Ribeiro; C Drosch; G Delles; M Matta; B Fisher; L Szilveszter; M Larsen; S Ratiu; B Kelly; A Garcia Del Blanco; Z D Rubio; B Drobni; I Jurlander; S Rodean; H Regan; M Cuéllar Calabria; T Boussoussou; R Engstrøm; A E Hodas; R Napp; S Haase; L M Feger; K Serna-Higuita; H Neumann; M Dreger; V Rief; M Wieske; P Estrella; M Martus; Dewey\",\n \"journal\": \"N Engl J Med\",\n \"ref_id\": \"b2\",\n \"title\": \"CT or Invasive Coronary Angiography in Stable Chest Pain\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"C Spadaccio; U Benedetto\",\n \"journal\": \"Ann Cardiothorac Surg\",\n \"ref_id\": \"b3\",\n \"title\": \"Coronary artery bypass grafting (CABG) vs. percutaneous coronary intervention (PCI) in the treatment of multivessel coronary disease: quo vadis? -a review of the evidences on coronary artery disease\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Z Li; Y Zhang; G Liu; H Shao; W Li; X Tang\",\n \"journal\": \"Biomedical Signal Processing and Control\",\n \"ref_id\": \"b4\",\n \"title\": \"A robust coronary artery identification and centerline extraction method in angiographies\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Z Xian; X Wang; S Yan; D Yang; J Chen; C Peng\",\n \"journal\": \"Mathematical Problems in Engineering\",\n \"ref_id\": \"b5\",\n \"title\": \"Main Coronary Vessel Segmentation Using Deep Learning in Smart Medical\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"N I Parikh; E F Honeycutt; M T Roe; M Neely; E J Rosenthal; M A Mittleman; J P Carrozza; K K L Ho\",\n \"journal\": \"Circ: Cardiovascular Quality and Outcomes\",\n \"ref_id\": \"b6\",\n \"title\": \"Left and Codominant Coronary Artery Circulations Are Associated With Higher In-Hospital Mortality Among Patients Undergoing Percutaneous Coronary Intervention for Acute Coronary Syndromes: Report From the National Cardiovascular Database Cath Percutaneous Coronary Intervention (CathPCI) Registry\",\n \"year\": \"2012\"\n },\n {\n \"authors\": \"W Austen; J Edwards; R Frye; G Gensini; V Gott; L Griffith; D Mcgoon; M Murphy; B Roe\",\n \"journal\": \"American Heart Association, Circulation\",\n \"ref_id\": \"b7\",\n \"title\": \"A reporting system on patients evaluated for coronary artery disease. Report of the Ad Hoc Committee for Grading of Coronary Artery Disease, Council on Cardiovascular Surgery\",\n \"year\": \"1975\"\n },\n {\n \"authors\": \"S Kastellanos; K Aznaouridis; C Vlachopoulos; E Tsiamis; E Oikonomou; D Tousoulis\",\n \"journal\": \"WJC\",\n \"ref_id\": \"b8\",\n \"title\": \"Overview of coronary artery variants, aberrations and anomalies\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"G Yang; A Broersen; R Petr; P Kitslaar; M A De Graaf; J J Bax; J H C Reiber; J Dijkstra\",\n \"journal\": \"Computing in Cardiology\",\n \"ref_id\": \"b9\",\n \"title\": \"Automatic coronary artery tree labeling in coronary computed tomographic angiography datasets\",\n \"year\": \"2011\"\n },\n {\n \"authors\": \"C Zhao; R Bober; H Tang; J Tang; M Dong; C Zhang; Z He; M Esposito; Z Xu; W Zhou\",\n \"journal\": \"Journal of Advances in Applied & Computational Mathematics\",\n \"ref_id\": \"b10\",\n \"title\": \"Semantic Segmentation to Extract Coronary Arteries in Invasive Coronary Angiograms\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"H Zhang; Z Gao; D Zhang; W K Hau; H Zhang\",\n \"journal\": \"IEEE Transactions on Medical Imaging\",\n \"ref_id\": \"b11\",\n \"title\": \"Progressive Perception Learning for Main Coronary Segmentation in X-ray Angiography\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"C Zhao; Z Xu; J Jiang; M Esposito; D Pienta; G.-U Hung; W Zhou\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"AGMN: Association Graph-based Graph Matching Network for Coronary Artery Semantic Labeling on Invasive Coronary Angiograms\",\n \"year\": \"2023-02-10\"\n },\n {\n \"authors\": \"C Zhao; A Vij; S Malhotra; J Tang; H Tang; D Pienta; Z Xu; W Zhou\",\n \"journal\": \"Computers in Biology and Medicine\",\n \"ref_id\": \"b13\",\n \"title\": \"Automatic extraction and stenosis evaluation of coronary arteries in invasive coronary angiograms\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"T J Jun; J Kweon; Y.-H Kim; D Kim; T-Net \",\n \"journal\": \"Neural Networks\",\n \"ref_id\": \"b14\",\n \"title\": \"Nested encoder-decoder architecture for the main vessel segmentation in coronary angiography\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Q Cao; A Broersen; M A De Graaf; P H Kitslaar; G Yang; A J Scholte; B P F Lelieveldt; J H C Reiber; J Dijkstra\",\n \"journal\": \"Int J Cardiovasc Imaging\",\n \"ref_id\": \"b15\",\n \"title\": \"Automatic identification of coronary tree anatomy in coronary computed tomography angiography\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"D Wu; X Wang; J Bai; X Xu; B Ouyang; Y Li; H Zhang; Q Song; K Cao; Y Yin\",\n \"journal\": \"Int J CARS\",\n \"ref_id\": \"b16\",\n \"title\": \"Automated anatomical labeling of coronary arteries via bidirectional tree LSTMs\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"H Yang; X Zhen; Y Chi; L Zhang; X.-S Hua\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b17\",\n \"title\": \"CPR-GCN: Conditional Partial-Residual Graph Convolutional Network in Automated Anatomical Labeling of Coronary Arteries\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"H Bunke\",\n \"journal\": \"Pattern Recognition Letters\",\n \"ref_id\": \"b18\",\n \"title\": \"On a relation between graph edit distance and maximum common subgraph\",\n \"year\": \"1997\"\n },\n {\n \"authors\": \"E L Lawler\",\n \"journal\": \"Management Science\",\n \"ref_id\": \"b19\",\n \"title\": \"The quadratic assignment problem\",\n \"year\": \"1963\"\n },\n {\n \"authors\": \"J Yan; S Yang; E Hancock\",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"Learning for Graph Matching and Related Combinatorial Optimization Problems\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"H Dai; E B Khalil; Y Zhang; B Dilkina; L Song\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"Learning Combinatorial Optimization Algorithms over Graphs\",\n \"year\": \"2018-05-09\"\n },\n {\n \"authors\": \"A Nowak; S Villar; A S Bandeira; J Bruna\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b22\",\n \"title\": \"Revised Note on Learning Quadratic Assignment with Graph Neural Networks\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"R Wang; J Yan; X Yang\",\n \"journal\": \"IEEE Trans. Pattern Anal. Mach. Intell\",\n \"ref_id\": \"b23\",\n \"title\": \"Combinatorial Learning of Robust Deep Graph Matching: an Embedding based Approach\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"R Wang; J Yan; X Yang\",\n \"journal\": \"IEEE Transactions on Pattern Analysis and Machine Intelligence\",\n \"ref_id\": \"b24\",\n \"title\": \"Neural Graph Matching Network: Learning Lawler's Quadratic Assignment Problem With Extension to Hypergraph and Multiple-Graph Matching\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"E R Dougherty\",\n \"journal\": \"SPIE, Optical Engineering Press\",\n \"ref_id\": \"b25\",\n \"title\": \"An introduction to morphological image processing\",\n \"year\": \"1992\"\n },\n {\n \"authors\": \"J Xie; Y Zhao; Y Liu; P Su; Y Zhao; J Cheng; Y Zheng; J Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"Topology reconstruction of treelike structure in images via structural similarity measure and dominant set clustering\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"T Wang; H Liu; Y Li; Y Jin; X Hou; H Ling\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b27\",\n \"title\": \"Learning Combinatorial Solver for Graph Matching\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"C Zhao; Y Xu; Z He; J Tang; Y Zhang; J Han; Y Shi; W Zhou\",\n \"journal\": \"Pattern Recognition\",\n \"ref_id\": \"b28\",\n \"title\": \"Lung segmentation and automatic detection of COVID-19 using radiomic features from chest CT images\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"J J M Van Griethuysen; A Fedorov; C Parmar; A Hosny; N Aucoin; V Narayan; R G H Beets-Tan; J.-C Fillion-Robin; S Pieper; H J W L Aerts\",\n \"journal\": \"Cancer Res\",\n \"ref_id\": \"b29\",\n \"title\": \"Computational Radiomics System to Decode the Radiographic Phenotype\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"D Ulyanov; A Vedaldi; V Lempitsky\",\n \"journal\": \"\",\n \"ref_id\": \"b30\",\n \"title\": \"Instance normalization: The missing ingredient for fast stylization\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"J Gilmer; S S Schoenholz; P F Riley; O Vinyals; G E Dahl\",\n \"journal\": \"\",\n \"ref_id\": \"b31\",\n \"title\": \"Neural Message Passing for Quantum Chemistry\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"W Hamilton; Z Ying; J Leskovec\",\n \"journal\": \"Advances in Neural Information Processing Systems\",\n \"ref_id\": \"b32\",\n \"title\": \"Inductive representation learning on large graphs\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"P Veličković; G Cucurull; A Casanova; A Romero; P Lio; Y Bengio\",\n \"journal\": \"\",\n \"ref_id\": \"b33\",\n \"title\": \"Graph attention networks\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"J Qu; H Ling; C Zhang; X Lyu; Z Tang\",\n \"journal\": \"\",\n \"ref_id\": \"b34\",\n \"title\": \"Adaptive Edge Attention for Graph Matching with Outliers\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"H Zhang; R Yanagi; R Togo; T Ogawa; M Haseyama\",\n \"journal\": \"IEEE Access\",\n \"ref_id\": \"b35\",\n \"title\": \"Cross-Modal Image Retrieval Considering Semantic Relationships With Many-to-Many Correspondence Loss\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"M Zaslavskiy; F Bach; J.-P Vert\",\n \"journal\": \"\",\n \"ref_id\": \"b36\",\n \"title\": \"Many-to-Many Graph Matching: a Continuous Relaxation Approach\",\n \"year\": \"2010-06-21\"\n },\n {\n \"authors\": \"H Yuan; H Yu; S Gui; S Ji\",\n \"journal\": \"IEEE Trans. Pattern Anal. Mach. Intell\",\n \"ref_id\": \"b37\",\n \"title\": \"Explainability in Graph Neural Networks: A Taxonomic Survey\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"T Funke; M Khosla; M Rathee; A Anand; Zorro \",\n \"journal\": \"IEEE Trans. Knowl. Data Eng\",\n \"ref_id\": \"b38\",\n \"title\": \"Valid, Sparse, and Stable Explanations in Graph Neural Networks\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"P W Battaglia; J B Hamrick; V Bapst; A Sanchez-Gonzalez; V Zambaldi; M Malinowski; A Tacchetti; D Raposo; A Santoro; R Faulkner; C Gulcehre; F Song; A Ballard; J Gilmer; G Dahl; A Vaswani; K Allen; C Nash; V Langston; C Dyer; N Heess; D Wierstra; P Kohli; M Botvinick; O Vinyals; Y Li; R Pascanu\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"Relational inductive biases, deep learning, and graph networks\",\n \"year\": \"2018\"\n }\n]"},"formulas":{"kind":"list like","value":[{"formula_coordinates":[8,228.29,596.31,304.46,26.04],"formula_id":"formula_0","formula_text":"𝑀 𝑖𝑗 = { 1 𝑖𝑓 𝑉 𝑖 𝑚𝑎𝑡𝑐ℎ𝑒𝑠 𝑉 𝑗 0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒(1)"},{"formula_coordinates":[9,166.37,589.16,367.82,14.4],"formula_id":"formula_1","formula_text":"𝑣 𝑖𝑎 = [𝑣 𝑖 , 𝑣 𝑎 ] ∈ ℝ 2𝑑 s.t. 𝑖 ∈ [1, ⋯ , 𝑛 1 ] and 𝑗 ∈ [1, ⋯ , 𝑛 2 ](2)"},{"formula_coordinates":[9,147.62,644.72,387.17,15.12],"formula_id":"formula_2","formula_text":"𝑒 𝑖𝑎,𝑗𝑏 = [𝑒 𝑖𝑗 1 , 𝑒 𝑎𝑏 2 ] ∈ ℝ 4𝑑 s.t. 𝑖, 𝑗 ∈ [1, ⋯ , 𝑛 1 ] and 𝑎, 𝑏 ∈ [1, ⋯ , 𝑛 2 ](3)"},{"formula_coordinates":[9,101.66,665,121.35,17.27],"formula_id":"formula_3","formula_text":"𝑒 𝑖𝑗 𝑔 = [𝑣 𝑖 , 𝑣 𝑗 ] s.t. 𝑔 ∈ [1,2]."},{"formula_coordinates":[10,246.53,129.84,288.26,29.52],"formula_id":"formula_4","formula_text":"𝑣 𝑖𝑎 𝑒𝑚𝑏 = 𝑓 𝑣 𝑒𝑚𝑏 (𝑣 𝑖𝑎 ) 𝑒 𝑖𝑎,𝑗𝑏 𝑒𝑚𝑏 = 𝑓 𝑒 𝑒𝑚𝑏 (𝑒 𝑖𝑎,𝑗𝑏 )(4)"},{"formula_coordinates":[10,246.17,348.99,288.62,36.24],"formula_id":"formula_5","formula_text":"𝑣̅ 𝑖𝑎 𝑒𝑚𝑏 = 𝑔 𝑣 𝑒𝑚𝑏 (𝑣 𝑖𝑎 𝑒𝑚𝑏 ) 𝑒̅ 𝑖𝑎,𝑗𝑏 𝑒𝑚𝑏 = 𝑔 𝑒 𝑒𝑚𝑏 (𝑒 𝑖𝑎,𝑗𝑏 𝑒𝑚𝑏 )(5)"},{"formula_coordinates":[12,197.69,425.57,335.42,40.2],"formula_id":"formula_8","formula_text":"𝐿 = 𝑀𝑆𝐸(𝑀, 𝑀 ̂) = ∑ ∑(𝑀 𝑖𝑎 -𝑀 ̂𝑖𝑎 ) 2 𝑛 2 𝑎=1 𝑛 1 𝑖=1(12)"},{"formula_coordinates":[15,204.89,108.72,324.81,40.2],"formula_id":"formula_9","formula_text":"𝐹(𝕍 𝑠 , 𝔽 𝑠 ) = 1 𝑛 1 𝑛 2 ∑ ∑ 𝕀(𝑦 𝑖𝑎 𝑠 = 𝑦 ̂𝑖𝑎 ) 𝑛 2 𝑎=1 𝑛 1 𝑖=1 (17"},{"formula_coordinates":[15,529.7,124.6,4.61,9.94],"formula_id":"formula_10","formula_text":")"}],"string":"[\n {\n \"formula_coordinates\": [\n 8,\n 228.29,\n 596.31,\n 304.46,\n 26.04\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"𝑀 𝑖𝑗 = { 1 𝑖𝑓 𝑉 𝑖 𝑚𝑎𝑡𝑐ℎ𝑒𝑠 𝑉 𝑗 0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒(1)\"\n },\n {\n \"formula_coordinates\": [\n 9,\n 166.37,\n 589.16,\n 367.82,\n 14.4\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"𝑣 𝑖𝑎 = [𝑣 𝑖 , 𝑣 𝑎 ] ∈ ℝ 2𝑑 s.t. 𝑖 ∈ [1, ⋯ , 𝑛 1 ] and 𝑗 ∈ [1, ⋯ , 𝑛 2 ](2)\"\n },\n {\n \"formula_coordinates\": [\n 9,\n 147.62,\n 644.72,\n 387.17,\n 15.12\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"𝑒 𝑖𝑎,𝑗𝑏 = [𝑒 𝑖𝑗 1 , 𝑒 𝑎𝑏 2 ] ∈ ℝ 4𝑑 s.t. 𝑖, 𝑗 ∈ [1, ⋯ , 𝑛 1 ] and 𝑎, 𝑏 ∈ [1, ⋯ , 𝑛 2 ](3)\"\n },\n {\n \"formula_coordinates\": [\n 9,\n 101.66,\n 665,\n 121.35,\n 17.27\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"𝑒 𝑖𝑗 𝑔 = [𝑣 𝑖 , 𝑣 𝑗 ] s.t. 𝑔 ∈ [1,2].\"\n },\n {\n \"formula_coordinates\": [\n 10,\n 246.53,\n 129.84,\n 288.26,\n 29.52\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"𝑣 𝑖𝑎 𝑒𝑚𝑏 = 𝑓 𝑣 𝑒𝑚𝑏 (𝑣 𝑖𝑎 ) 𝑒 𝑖𝑎,𝑗𝑏 𝑒𝑚𝑏 = 𝑓 𝑒 𝑒𝑚𝑏 (𝑒 𝑖𝑎,𝑗𝑏 )(4)\"\n },\n {\n \"formula_coordinates\": [\n 10,\n 246.17,\n 348.99,\n 288.62,\n 36.24\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"𝑣̅ 𝑖𝑎 𝑒𝑚𝑏 = 𝑔 𝑣 𝑒𝑚𝑏 (𝑣 𝑖𝑎 𝑒𝑚𝑏 ) 𝑒̅ 𝑖𝑎,𝑗𝑏 𝑒𝑚𝑏 = 𝑔 𝑒 𝑒𝑚𝑏 (𝑒 𝑖𝑎,𝑗𝑏 𝑒𝑚𝑏 )(5)\"\n },\n {\n \"formula_coordinates\": [\n 12,\n 197.69,\n 425.57,\n 335.42,\n 40.2\n ],\n \"formula_id\": \"formula_8\",\n \"formula_text\": \"𝐿 = 𝑀𝑆𝐸(𝑀, 𝑀 ̂) = ∑ ∑(𝑀 𝑖𝑎 -𝑀 ̂𝑖𝑎 ) 2 𝑛 2 𝑎=1 𝑛 1 𝑖=1(12)\"\n },\n {\n \"formula_coordinates\": [\n 15,\n 204.89,\n 108.72,\n 324.81,\n 40.2\n ],\n \"formula_id\": \"formula_9\",\n \"formula_text\": \"𝐹(𝕍 𝑠 , 𝔽 𝑠 ) = 1 𝑛 1 𝑛 2 ∑ ∑ 𝕀(𝑦 𝑖𝑎 𝑠 = 𝑦 ̂𝑖𝑎 ) 𝑛 2 𝑎=1 𝑛 1 𝑖=1 (17\"\n },\n {\n \"formula_coordinates\": [\n 15,\n 529.7,\n 124.6,\n 4.61,\n 9.94\n ],\n \"formula_id\": \"formula_10\",\n \"formula_text\": \")\"\n }\n]"}}},{"rowIdx":395,"cells":{"doi":{"kind":"string","value":""},"pub_date":{"kind":"string","value":""},"sections":{"kind":"list like","value":[{"figure_ref":[],"heading":"Introduction","publication_ref":["b0","b1","b2","b3","b4","b5","b6"],"table_ref":[],"text":"Nowadays, router devices are so important with Internet Service Provider (ISP) in core network. Juniper router that is used in SCTV core network enables a wide range of business and residential applications and services (e.g., high-speed transport, Virtual Private Network services, high-speed Internet) [1]. To detect and prevent abnormal operations of Juniper router devices, there are some solutions such as monitoring systems, checking syslog server [2]. Abnormal operations of devices may be cause of routing errors in the network, chassis errors, Distributed Denial-of-service (DDos) attack or some fail processes in router devices. These abnormal operations of router devices are quite reported to syslog server if they are configured. It is possible to detect abnormal operation by inspecting manually these logs. Anomaly detection has been a practical research topic and has great importance in many application domains in university and in industry (e.g., [3], [4], [5]). For conventional small systems, engineers manually define rules or check system logs to detect anomalies based on their domain knowledge. Additionally, they can use regular expression match or keyword searches (e.g., \"error\", \"fail\"). However, the extremely large sizes of log data generated (up to million logs) by hundreds of devices make manual analysis impossible.\nAs a result, automated log analysis methods for anomaly detection of Juniper router devices based on machine learning systems are highly in demand. However, we found no resources about this problem. This paper is the first work about anomaly detection using One-Class SVM for logs of Juniper router devices. We present a new way to get features from log data of Juniper router devices based on importance characteristics of log messages. The standard support vector machine (SVM) [6] is the machine learning method to classify two class or multiple class of data. But the log data of anomaly detection are very special, the abnormal logs are much less than the normal logs. Therefore, the standard SVM does not work well on our task, we use One-Class SVM described in [7] to handle the logs classification."},{"figure_ref":[],"heading":"Collecting Log Data, Preprocessing and Feature Extraction","publication_ref":[],"table_ref":[],"text":"Collecting log data for training and testing is necessary because we use the real logs from real router Juniper devices. Logs must be made clean before they can be used for feature extraction."},{"figure_ref":["fig_0"],"heading":"Collecting Log Data and Text Preprocessing","publication_ref":[],"table_ref":[],"text":"Juniper router devices routinely generate logs to record device states and runtime information, each including a content indicating what has happened and a timestamp. These devices are configured to send logs to syslog server that always receives data on corresponding port. The log messages have some common characteristics although the format of them is not fixed length. This valuable information could be utilized for anomaly detection and other purposes; thereby logs are collected first and saved as a file in syslog server for further usage. For example, Fig. 1 depicts some log messages of Juniper router devices. As the example above, a timestamp is the beginning of each message, a name of Juniper router device with the same format and other characteristics: the log messages at syslog server are written in English and comprise digits, lower, upper case letters and many special characters. The raw logs are pre-processed by Python program. Our text normalization procedures are given below: Remove timestamps: Remove whole timestamp before each message (both date and time). Remove router device's name: Remove device's name. Remove digits, special characters: Remove any special character, including punctuations, all digits. Replace continuous spaces with a single space: The length of log message depends on how many spaces in log message, so that replacing continuous spaces with a single space is necessary. Lower cases: Replace all upper case letters by lower case.\nAlthough timestamp information, which is the time the event happened, is very important according to the SCTV engineers, we still remove it from the log message. After anomaly detection, the abnormal original log messages (including timestamp information) will be sent to the engineers."},{"figure_ref":[],"heading":"Feature Extraction","publication_ref":["b7","b8","b9"],"table_ref":[],"text":"Some methods used to perform feature extraction for text classification [8][9][10] are not effective with logs of Juniper router devices. In our works, we use three characteristics of each log message to be three elements of the feature vector: the length of log message, the number of different words and the sum of TF-IDF in log message.\nThe Length of Log.\nThe length of log message is a significant characteristic. In the process of manually defining abnormal logs, the length of logs shows that the logs with irregular length have a higher probability of being abnormal logs. We denote i S as the length of log i . We use the spaces of each log message to calculate i S ."},{"figure_ref":[],"heading":"The Number of Different Words in Log Message.","publication_ref":["b10","b11","b12"],"table_ref":[],"text":"The number of different words in log message: the number of words which are different from dictionary of each log message. Bag-of-Words (BoW) model and the length of log are used to calculate this characteristic. Some of new document representation methods [11][12][13] are developed based on BoW.\nThe dictionary is built based on normal logs that we classify from original logs. To get the number of words that are in dictionary, BoW model is the accordant and simple classical model for our purpose. Based on BoW, each vector represents a log message; each element denotes the normalized number of occurrence of a word in log. The words that do not appear in dictionary are not counted by BoW model.\nBased on the dictionary and BoW model, each log is converted from text to a vector space. The log data become the matrix that is given by: where m is the length of dictionary, n is the size of log data. Each row of the matrix A represents a log message in log data. The number of words which are different from dictionary of each log message is given by the formula:\n1 m i i ij j L S a (2)\nwhere i S is length of the i-th log."},{"figure_ref":[],"heading":"The Sum of TF-IDF.","publication_ref":["b13","b14","b15"],"table_ref":[],"text":"Term frequency-inverse document frequency (TF-IDF) is one of the most famous algorithms used in document mining research; it is used for calculating the weight of each word. The word frequency means the number of time a term is repeated in a log message, and Inverse Document Frequency is an algorithm used to calculate the inverse probability of finding a word in log data [14]. Some improvements of feature extraction based on TF-IDF are mentioned in [15,16]. TF-IDF Formula:\nlog ij ij j N g tf df (3)\nwhere ij g is the weight of the word j in the log i , N is the total number of log mes- sages, ij tf is the frequency of the word j in the log i , j df is the number of logs con- taining the word j . The sum of TF-IDF in log is the third element of feature vector. Equation 4is the summary equation used to calculate the sum of TF-IDF:\n1 h i ij j Gg (4)\nwhere i G is the sum of TF-IDF of log i , h is the number of words in log i . The feature vector of log i :"},{"figure_ref":[],"heading":" ","publication_ref":[],"table_ref":[],"text":",,\ni i i i x S L G (5)\nThe feature vectors are the input data of One-Class SVM model, which will be discussed below."},{"figure_ref":[],"heading":"One-Class SVM","publication_ref":["b5","b6"],"table_ref":[],"text":"One famous machine learning method is the support vector machine (SVM), which was invented by Vladimir Vapnik and Alexey Ya. Chervonenkis in 1963, and is widely applied for pattern classification and data analyzing. However, Vladimir Vapnik and Corinna Cortes proposed the current standard incarnation in 1993 [6].\nThe labels associated with training data are based to group anomaly detection techniques for log data into two broad categories: one-class and multi-class anomaly detection technique. The labels in log messages of Juniper router devices are grouped into two types: abnormal log and normal log, so that one-class anomaly detection technique was chosen for our purpose.\nFor the case of one-class classification, Scholkopf et al. proposed a maximum margin based classifier that is an adaptation of the Support Vector Machine algorithm [7]. The data are separated from the origin by a separating hyperplane ,0 wz with maximum margin (where w and are respectively the normal vector of the hyper- plane and the distance from the hyperplane to the origin).\nThe maximum margin from the origin is found by solving the below quadratic optimization problem:\n 2 ,, 1 1 min 2 subject to ( ) , 0. i i w i i i w vl w x (6)\nwhere i are so-called slack variables that are used to model the separation errors. The 0,1 v is a parameter that adjusts the balance between maximizing the distance from the origin and the region created by the hyperplane containing most of the data. () is a non-linear projection is evaluated through a kernel function that is used as a mapping from the original feature space to a possibly higher dimensional feature space:\n ( , ) ( ) ( k x y x y . In our works, we consider the kernel Radial Basis Function (RBF), linear kernel, polynomial kernel and sigmoid kernel. They are expressed respectively by these following equations ( 7)-( 10): x y e\n (7) , T linear k x y x y C (8) ( , ) d T polynomial k x y x y C (9)\n \n( , ) tanh\nT sigmoid k x y x y C (10)"},{"figure_ref":[],"heading":"Experimental Results","publication_ref":["b16","b17","b18"],"table_ref":[],"text":"In this section, we descript the log datasets, the evaluate method, perform an experimental evaluation and comparison of our anomaly detection method for log data of Juniper router devices based on One-Class SVM model. We have chosen Python programming language for our convenience purpose. Python is a powerful interpreted and popular language and supports some power libraries for data science, machine learning (numpy, matplotlib, scikit-learn, etc) [17][18][19]."},{"figure_ref":["fig_3","fig_3"],"heading":"Log Datasets","publication_ref":[],"table_ref":["tab_0"],"text":"Publicly available production logs are scarce data, especially log data of router devices because companies and ISPs rarely publish them to community due to confidential issues. So that we collected log data from real Juniper router devices in core network of SCTV (Saigontourist Cable Television Company). Logs from devices were sent through the internet network to syslog server and saved as txt files. These files can be read directly and easily by most popular programs. The log data contains 12907 log messages and 266 abnormal log messages, which are manually labeled by us and the SCTV experts. Using this log dataset, we evaluated the performance and compared the results of the models. More statistical information of the log dataset is provided in Table 1. After the raw logs are collected using Python script, they are called from our main Python program. Feature extraction is applied to these log messages, each log is represented by a feature vector include: the length of log message, the number of words which are different from dictionary and the sum of TF-IDF value in log message. We choose 60% of log messages as the training data and the remainders as the testing data.\nFigure 2 shows the training data and testing data on 3-dimensional feature space. Blue points are the normal data, red points are abnormal data. Due to the fact that the log data come from over a hundred router Juniper devices, we receive a lot of similar log messages. There are some differences between them such as the timestamp, device's name, etc. So a lot of log messages become the same after they are passed the preprocess step that we mentioned above. Because of that, the data points in the figure may look less than reality. Although the visualizations in figure 2 shows that the data are naturally well-separated after they are converted to feature vectors, we still propose the One-Class SVM method to separate the normal log data and abnormal log data automatically, and make sure the model still works efficiently in case there are more abnormal log data appeared when the hardware system is extended."},{"figure_ref":[],"heading":"Method Evaluation","publication_ref":["b19"],"table_ref":[],"text":"The Precision, Recall and F-measure, which are the most commonly used metrics, are used to evaluate the accuracy of One-Class SVM anomaly detection method using different kernels (Radial Basis Function (RBF), linear kernel, polynomial kernel and sigmoid kernel) as we have already the ground truth for the log data. As shown below, Precision shows the percentage of true anomalies among all anomalies detected, Recall measures the percentage of how many real anomalies are detected, and F-measure denotes the harmonic mean of precision and recall [20]. They are expressed respectively by the following equations ( 11) -( 13):"},{"figure_ref":[],"heading":"Anomalies detected Precision","publication_ref":[],"table_ref":[],"text":"Anomalies reported (11)"},{"figure_ref":[],"heading":"Anomalies detected Recall","publication_ref":[],"table_ref":[],"text":"All anomalies\n (12) 2 Precision Recall F measure Precision Recall (13)"},{"figure_ref":["fig_4"],"heading":"Result Evaluation","publication_ref":[],"table_ref":[],"text":"Based on evaluate methods that we mentioned, we show the results which evaluate of each model applied on training data and testing data. Figure 3 shows the accuracy of anomaly detection on training log data. Three models (with linear, polynomial, sigmoid kernels) have no good performance on training data with the F-measure close to 0.75. We can observe that recall measures of these models are low (close to 0.65). We give priority to minimize the number of abnormal log messages which the model predicted wrongly, One-Class SVM method with RBG kernel is the best model in our case. Recall measure of this model is equal to 1, it shows that the model predicted rightly all abnormal logs. One-Class SVM RBF kernel model's performance on training data is better than others models. However, their accuracy on testing data varies with different kernels. When applying these models for testing data, three models (with linear, polynomial, sigmoid kernels) become unacceptable. These kernels are not suitable for our purpose. We can observe that One-Class SVM methods with RBF and polynomial kernel achieve high Precision (over 0.95), which implies that normal instances and abnormal instances are well separated by using our feature representation. As we observe on "},{"figure_ref":[],"heading":"Conclusion","publication_ref":[],"table_ref":[],"text":"Logs are widely utilized to detection anomalies in Juniper router device systems. However, traditional anomaly detection that depends heavily on manual log inspection becomes impossible due to the limit of human ability and the sharp increase of log size.\nTo reduce manual effort, automated log analysis and anomaly detection methods have been widely studied in recent years.\nIn this paper, we successfully created the new feature extraction for log data of Juniper router devices and used the One-Class SVM model with different kernels for anomaly detection. We also compared their accuracy and efficiency on training and testing real log datasets. We find that the One-Class SVM model with RBF kernel has the best accuracy in terms of precision, recall and F-measure."},{"figure_ref":[],"heading":"Acknowledgments","publication_ref":[],"table_ref":[],"text":"This work was partially supported by the Ministry of Science and Technology, Taiwan, under grant MOST 107-2221-E-006-222. In addition, this work got the encouragement from Posts and Telecommunications Institute of Technology, Vietnam."}],"string":"[\n {\n \"figure_ref\": [],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b0\",\n \"b1\",\n \"b2\",\n \"b3\",\n \"b4\",\n \"b5\",\n \"b6\"\n ],\n \"table_ref\": [],\n \"text\": \"Nowadays, router devices are so important with Internet Service Provider (ISP) in core network. Juniper router that is used in SCTV core network enables a wide range of business and residential applications and services (e.g., high-speed transport, Virtual Private Network services, high-speed Internet) [1]. To detect and prevent abnormal operations of Juniper router devices, there are some solutions such as monitoring systems, checking syslog server [2]. Abnormal operations of devices may be cause of routing errors in the network, chassis errors, Distributed Denial-of-service (DDos) attack or some fail processes in router devices. These abnormal operations of router devices are quite reported to syslog server if they are configured. It is possible to detect abnormal operation by inspecting manually these logs. Anomaly detection has been a practical research topic and has great importance in many application domains in university and in industry (e.g., [3], [4], [5]). For conventional small systems, engineers manually define rules or check system logs to detect anomalies based on their domain knowledge. Additionally, they can use regular expression match or keyword searches (e.g., \\\"error\\\", \\\"fail\\\"). However, the extremely large sizes of log data generated (up to million logs) by hundreds of devices make manual analysis impossible.\\nAs a result, automated log analysis methods for anomaly detection of Juniper router devices based on machine learning systems are highly in demand. However, we found no resources about this problem. This paper is the first work about anomaly detection using One-Class SVM for logs of Juniper router devices. We present a new way to get features from log data of Juniper router devices based on importance characteristics of log messages. The standard support vector machine (SVM) [6] is the machine learning method to classify two class or multiple class of data. But the log data of anomaly detection are very special, the abnormal logs are much less than the normal logs. Therefore, the standard SVM does not work well on our task, we use One-Class SVM described in [7] to handle the logs classification.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Collecting Log Data, Preprocessing and Feature Extraction\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Collecting log data for training and testing is necessary because we use the real logs from real router Juniper devices. Logs must be made clean before they can be used for feature extraction.\"\n },\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"Collecting Log Data and Text Preprocessing\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Juniper router devices routinely generate logs to record device states and runtime information, each including a content indicating what has happened and a timestamp. These devices are configured to send logs to syslog server that always receives data on corresponding port. The log messages have some common characteristics although the format of them is not fixed length. This valuable information could be utilized for anomaly detection and other purposes; thereby logs are collected first and saved as a file in syslog server for further usage. For example, Fig. 1 depicts some log messages of Juniper router devices. As the example above, a timestamp is the beginning of each message, a name of Juniper router device with the same format and other characteristics: the log messages at syslog server are written in English and comprise digits, lower, upper case letters and many special characters. The raw logs are pre-processed by Python program. Our text normalization procedures are given below: Remove timestamps: Remove whole timestamp before each message (both date and time). Remove router device's name: Remove device's name. Remove digits, special characters: Remove any special character, including punctuations, all digits. Replace continuous spaces with a single space: The length of log message depends on how many spaces in log message, so that replacing continuous spaces with a single space is necessary. Lower cases: Replace all upper case letters by lower case.\\nAlthough timestamp information, which is the time the event happened, is very important according to the SCTV engineers, we still remove it from the log message. After anomaly detection, the abnormal original log messages (including timestamp information) will be sent to the engineers.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Feature Extraction\",\n \"publication_ref\": [\n \"b7\",\n \"b8\",\n \"b9\"\n ],\n \"table_ref\": [],\n \"text\": \"Some methods used to perform feature extraction for text classification [8][9][10] are not effective with logs of Juniper router devices. In our works, we use three characteristics of each log message to be three elements of the feature vector: the length of log message, the number of different words and the sum of TF-IDF in log message.\\nThe Length of Log.\\nThe length of log message is a significant characteristic. In the process of manually defining abnormal logs, the length of logs shows that the logs with irregular length have a higher probability of being abnormal logs. We denote i S as the length of log i . We use the spaces of each log message to calculate i S .\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"The Number of Different Words in Log Message.\",\n \"publication_ref\": [\n \"b10\",\n \"b11\",\n \"b12\"\n ],\n \"table_ref\": [],\n \"text\": \"The number of different words in log message: the number of words which are different from dictionary of each log message. Bag-of-Words (BoW) model and the length of log are used to calculate this characteristic. Some of new document representation methods [11][12][13] are developed based on BoW.\\nThe dictionary is built based on normal logs that we classify from original logs. To get the number of words that are in dictionary, BoW model is the accordant and simple classical model for our purpose. Based on BoW, each vector represents a log message; each element denotes the normalized number of occurrence of a word in log. The words that do not appear in dictionary are not counted by BoW model.\\nBased on the dictionary and BoW model, each log is converted from text to a vector space. The log data become the matrix that is given by: where m is the length of dictionary, n is the size of log data. Each row of the matrix A represents a log message in log data. The number of words which are different from dictionary of each log message is given by the formula:\\n1 m i i ij j L S a (2)\\nwhere i S is length of the i-th log.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"The Sum of TF-IDF.\",\n \"publication_ref\": [\n \"b13\",\n \"b14\",\n \"b15\"\n ],\n \"table_ref\": [],\n \"text\": \"Term frequency-inverse document frequency (TF-IDF) is one of the most famous algorithms used in document mining research; it is used for calculating the weight of each word. The word frequency means the number of time a term is repeated in a log message, and Inverse Document Frequency is an algorithm used to calculate the inverse probability of finding a word in log data [14]. Some improvements of feature extraction based on TF-IDF are mentioned in [15,16]. TF-IDF Formula:\\nlog ij ij j N g tf df (3)\\nwhere ij g is the weight of the word j in the log i , N is the total number of log mes- sages, ij tf is the frequency of the word j in the log i , j df is the number of logs con- taining the word j . The sum of TF-IDF in log is the third element of feature vector. Equation 4is the summary equation used to calculate the sum of TF-IDF:\\n1 h i ij j Gg (4)\\nwhere i G is the sum of TF-IDF of log i , h is the number of words in log i . The feature vector of log i :\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \" \",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \",,\\ni i i i x S L G (5)\\nThe feature vectors are the input data of One-Class SVM model, which will be discussed below.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"One-Class SVM\",\n \"publication_ref\": [\n \"b5\",\n \"b6\"\n ],\n \"table_ref\": [],\n \"text\": \"One famous machine learning method is the support vector machine (SVM), which was invented by Vladimir Vapnik and Alexey Ya. Chervonenkis in 1963, and is widely applied for pattern classification and data analyzing. However, Vladimir Vapnik and Corinna Cortes proposed the current standard incarnation in 1993 [6].\\nThe labels associated with training data are based to group anomaly detection techniques for log data into two broad categories: one-class and multi-class anomaly detection technique. The labels in log messages of Juniper router devices are grouped into two types: abnormal log and normal log, so that one-class anomaly detection technique was chosen for our purpose.\\nFor the case of one-class classification, Scholkopf et al. proposed a maximum margin based classifier that is an adaptation of the Support Vector Machine algorithm [7]. The data are separated from the origin by a separating hyperplane ,0 wz with maximum margin (where w and are respectively the normal vector of the hyper- plane and the distance from the hyperplane to the origin).\\nThe maximum margin from the origin is found by solving the below quadratic optimization problem:\\n 2 ,, 1 1 min 2 subject to ( ) , 0. i i w i i i w vl w x (6)\\nwhere i are so-called slack variables that are used to model the separation errors. The 0,1 v is a parameter that adjusts the balance between maximizing the distance from the origin and the region created by the hyperplane containing most of the data. () is a non-linear projection is evaluated through a kernel function that is used as a mapping from the original feature space to a possibly higher dimensional feature space:\\n ( , ) ( ) ( k x y x y . In our works, we consider the kernel Radial Basis Function (RBF), linear kernel, polynomial kernel and sigmoid kernel. They are expressed respectively by these following equations ( 7)-( 10): x y e\\n (7) , T linear k x y x y C (8) ( , ) d T polynomial k x y x y C (9)\\n \\n( , ) tanh\\nT sigmoid k x y x y C (10)\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experimental Results\",\n \"publication_ref\": [\n \"b16\",\n \"b17\",\n \"b18\"\n ],\n \"table_ref\": [],\n \"text\": \"In this section, we descript the log datasets, the evaluate method, perform an experimental evaluation and comparison of our anomaly detection method for log data of Juniper router devices based on One-Class SVM model. We have chosen Python programming language for our convenience purpose. Python is a powerful interpreted and popular language and supports some power libraries for data science, machine learning (numpy, matplotlib, scikit-learn, etc) [17][18][19].\"\n },\n {\n \"figure_ref\": [\n \"fig_3\",\n \"fig_3\"\n ],\n \"heading\": \"Log Datasets\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_0\"\n ],\n \"text\": \"Publicly available production logs are scarce data, especially log data of router devices because companies and ISPs rarely publish them to community due to confidential issues. So that we collected log data from real Juniper router devices in core network of SCTV (Saigontourist Cable Television Company). Logs from devices were sent through the internet network to syslog server and saved as txt files. These files can be read directly and easily by most popular programs. The log data contains 12907 log messages and 266 abnormal log messages, which are manually labeled by us and the SCTV experts. Using this log dataset, we evaluated the performance and compared the results of the models. More statistical information of the log dataset is provided in Table 1. After the raw logs are collected using Python script, they are called from our main Python program. Feature extraction is applied to these log messages, each log is represented by a feature vector include: the length of log message, the number of words which are different from dictionary and the sum of TF-IDF value in log message. We choose 60% of log messages as the training data and the remainders as the testing data.\\nFigure 2 shows the training data and testing data on 3-dimensional feature space. Blue points are the normal data, red points are abnormal data. Due to the fact that the log data come from over a hundred router Juniper devices, we receive a lot of similar log messages. There are some differences between them such as the timestamp, device's name, etc. So a lot of log messages become the same after they are passed the preprocess step that we mentioned above. Because of that, the data points in the figure may look less than reality. Although the visualizations in figure 2 shows that the data are naturally well-separated after they are converted to feature vectors, we still propose the One-Class SVM method to separate the normal log data and abnormal log data automatically, and make sure the model still works efficiently in case there are more abnormal log data appeared when the hardware system is extended.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Method Evaluation\",\n \"publication_ref\": [\n \"b19\"\n ],\n \"table_ref\": [],\n \"text\": \"The Precision, Recall and F-measure, which are the most commonly used metrics, are used to evaluate the accuracy of One-Class SVM anomaly detection method using different kernels (Radial Basis Function (RBF), linear kernel, polynomial kernel and sigmoid kernel) as we have already the ground truth for the log data. As shown below, Precision shows the percentage of true anomalies among all anomalies detected, Recall measures the percentage of how many real anomalies are detected, and F-measure denotes the harmonic mean of precision and recall [20]. They are expressed respectively by the following equations ( 11) -( 13):\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Anomalies detected Precision\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Anomalies reported (11)\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Anomalies detected Recall\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"All anomalies\\n (12) 2 Precision Recall F measure Precision Recall (13)\"\n },\n {\n \"figure_ref\": [\n \"fig_4\"\n ],\n \"heading\": \"Result Evaluation\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Based on evaluate methods that we mentioned, we show the results which evaluate of each model applied on training data and testing data. Figure 3 shows the accuracy of anomaly detection on training log data. Three models (with linear, polynomial, sigmoid kernels) have no good performance on training data with the F-measure close to 0.75. We can observe that recall measures of these models are low (close to 0.65). We give priority to minimize the number of abnormal log messages which the model predicted wrongly, One-Class SVM method with RBG kernel is the best model in our case. Recall measure of this model is equal to 1, it shows that the model predicted rightly all abnormal logs. One-Class SVM RBF kernel model's performance on training data is better than others models. However, their accuracy on testing data varies with different kernels. When applying these models for testing data, three models (with linear, polynomial, sigmoid kernels) become unacceptable. These kernels are not suitable for our purpose. We can observe that One-Class SVM methods with RBF and polynomial kernel achieve high Precision (over 0.95), which implies that normal instances and abnormal instances are well separated by using our feature representation. As we observe on \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Logs are widely utilized to detection anomalies in Juniper router device systems. However, traditional anomaly detection that depends heavily on manual log inspection becomes impossible due to the limit of human ability and the sharp increase of log size.\\nTo reduce manual effort, automated log analysis and anomaly detection methods have been widely studied in recent years.\\nIn this paper, we successfully created the new feature extraction for log data of Juniper router devices and used the One-Class SVM model with different kernels for anomaly detection. We also compared their accuracy and efficiency on training and testing real log datasets. We find that the One-Class SVM model with RBF kernel has the best accuracy in terms of precision, recall and F-measure.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Acknowledgments\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This work was partially supported by the Ministry of Science and Technology, Taiwan, under grant MOST 107-2221-E-006-222. In addition, this work got the encouragement from Posts and Telecommunications Institute of Technology, Vietnam.\"\n }\n]"},"abstract":{"kind":"string","value":"The article deals with anomaly detection of Juniper router logs. Abnormal Juniper router logs include logs that are usually different from the normal operation, and they often reflect the abnormal operation of router devices. To prevent router devices from being damaged and help administrator to grasp the situation of error quickly, detecting abnormal operation soon is very important. In this work, we present a new way to get important features from log data of Juniper router devices and use machine learning method (basing on One-Class SVM model) for anomaly detection. One-Class SVM model requires some knowledge and comprehension about logs of Juniper router devices so that it can analyze, interpret, and test the knowledge acquired. We collect log data from a lot of real Juniper router devices and classify them based on our knowledge. Before these logs are used for training and testing the One-Class SVM model, the feature extraction phase for these data was carried out. Finally, with the proposed method, the system errors of the routers were dectected quickly and accurately. This may help our company to reduce the operation cost for the router systems."},"title":{"kind":"string","value":"Anomaly Detection Using One-Class SVM for Logs of Juniper Router Devices"},"figures":{"kind":"list like","value":[{"figure_caption":"Fig. 1 .1Fig. 1. Some log messages of Juniper router devices.","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"Fig. 2 .2Fig. 2. Training data and testing data on 3-dimensional feature space.","figure_data":"","figure_id":"fig_3","figure_label":"2","figure_type":"figure"},{"figure_caption":"Fig. 3 .3Fig. 3. Training accuracy of One-Class SVM method.","figure_data":"","figure_id":"fig_4","figure_label":"3","figure_type":"figure"},{"figure_caption":"Fig. 4 .4Fig. 4. Testing accuracy of One-Class SVM method.","figure_data":"","figure_id":"fig_5","figure_label":"4","figure_type":"figure"},{"figure_caption":"Figure 4 ,Fig. 5 .45Fig. 5. The decision boundary of One-Class SVM model with RBF kernel.","figure_data":"","figure_id":"fig_6","figure_label":"45","figure_type":"figure"},{"figure_caption":"","figure_data":"","figure_id":"","figure_label":"","figure_type":"figure"},{"figure_caption":"Summary of log datasets.","figure_data":"DevicesData sizeNumber of messages AnomaliesRouter Juniper1,7 Mb12 907266","figure_id":"tab_0","figure_label":"1","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Fig. 1 .1Fig. 1. Some log messages of Juniper router devices.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 2 .2Fig. 2. Training data and testing data on 3-dimensional feature space.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 3 .3Fig. 3. Training accuracy of One-Class SVM method.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 4 .4Fig. 4. Testing accuracy of One-Class SVM method.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_5\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 4 ,Fig. 5 .45Fig. 5. The decision boundary of One-Class SVM model with RBF kernel.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_6\",\n \"figure_label\": \"45\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"\",\n \"figure_data\": \"\",\n \"figure_id\": \"\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Summary of log datasets.\",\n \"figure_data\": \"DevicesData sizeNumber of messages AnomaliesRouter Juniper1,7 Mb12 907266\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n }\n]"},"authors":{"kind":"string","value":"Tat-Bao-Thien Nguyen; Teh-Lu Liao; Tuan-Anh Vu"},"references":{"kind":"list like","value":[{"authors":"","journal":"","ref_id":"b0","title":"MX960 3D Universal Edge Router Hardware Guide","year":"2018-01-22"},{"authors":"","journal":"","ref_id":"b1","title":"Network Performance Monitor Getting Started Guide","year":"2018-01-25"},{"authors":"Q Lin; J G Lou; H Zhang","journal":"","ref_id":"b2","title":"Log Clustering Based Problem Identification for Online Service Systems","year":"2016"},{"authors":"M Macit; E Delibaş; B Karanlık","journal":"","ref_id":"b3","title":"Real time distributed analysis of MPLS network logs for anomaly detection","year":"2016"},{"authors":"Y Gao; Y Ma; D Li","journal":"","ref_id":"b4","title":"Anomaly detection of malicious users' behaviors for web applications based on web logs","year":"2017"},{"authors":"C Cortes; V Vapnik","journal":"Journal Machine Learning","ref_id":"b5","title":"Support-Vector Networks","year":"1995"},{"authors":"B Scholkopf; J Platt; J Shawe-Taylor; A J Smola; R Williamson","journal":"Neural Computation","ref_id":"b6","title":"Estimating the support of a high-dimentional distribution","year":"2001"},{"authors":"C Y Chang; S J Lee; C C Lai","journal":"","ref_id":"b7","title":"Weighted word2vec based on the distance of words","year":"2017"},{"authors":"W Tian; J Li; H Li","journal":"","ref_id":"b8","title":"A Method of Feature Selection Based on Word2Vec in Text Categorization","year":"2018"},{"authors":"T P Van; T M Thanh","journal":"","ref_id":"b9","title":"Vietnamese news classification based on BoW with keywords extraction and neural network","year":"2017"},{"authors":"R Zhao; K Mao","journal":"IEEE Transactions on Fuzzy Systems","ref_id":"b10","title":"Fuzzy Bag-of-Words Model for Document Representation","year":"2018"},{"authors":"L Wu; S C H Hoi; N Yu","journal":"IEEE Transactions on Image Processing","ref_id":"b11","title":"Semantics-Preserving Bag-of-Words Models and Applications","year":"2010"},{"authors":"A Alahmadi; A Joorabchi; A E Mahdi","journal":"","ref_id":"b12","title":"A new text representation scheme combining Bag-of-Words and Bag-of-Concepts approaches for automatic text classification","year":"2013"},{"authors":"S M H Dadgar; M S Araghi; M M Farahani","journal":"","ref_id":"b13","title":"A novel text mining approach based on TF-IDF and Support Vector Machine for news classification","year":"2016"},{"authors":"X Huang; Q Wu","journal":"","ref_id":"b14","title":"Micro-blog commercial word extraction based on improved TF-IDF algorithm","year":"2013"},{"authors":"A Guo; T Yang","journal":"","ref_id":"b15","title":"Research and improvement of feature words weight based on TFIDF algorithm","year":"2016"},{"authors":"F Dubosson; S Bromuri; M Schumacher","journal":"","ref_id":"b16","title":"A Python Framework for Exhaustive Machine Learning Algorithms and Features Evaluations","year":"2016"},{"authors":"E Patterson; R Mcburney; H Schmidt","journal":"IBM Journal of Research and Development","ref_id":"b17","title":"Dataflow representation of data analyses: Toward a platform for collaborative data science","year":"2017"},{"authors":"C P Hwang; M S Chen; C M Shih","journal":"","ref_id":"b18","title":"Apply Scikit-Learn in Python to Analyze Driver Behavior Based on OBD Data","year":"2018"},{"authors":"M Sokolova; G Lapalme","journal":"Information Processing and Management","ref_id":"b19","title":"A systematic analysis of performance measures for classification tasks","year":"2009"}],"string":"[\n {\n \"authors\": \"\",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"MX960 3D Universal Edge Router Hardware Guide\",\n \"year\": \"2018-01-22\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"\",\n \"ref_id\": \"b1\",\n \"title\": \"Network Performance Monitor Getting Started Guide\",\n \"year\": \"2018-01-25\"\n },\n {\n \"authors\": \"Q Lin; J G Lou; H Zhang\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"Log Clustering Based Problem Identification for Online Service Systems\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"M Macit; E Delibaş; B Karanlık\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"Real time distributed analysis of MPLS network logs for anomaly detection\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Y Gao; Y Ma; D Li\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"Anomaly detection of malicious users' behaviors for web applications based on web logs\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"C Cortes; V Vapnik\",\n \"journal\": \"Journal Machine Learning\",\n \"ref_id\": \"b5\",\n \"title\": \"Support-Vector Networks\",\n \"year\": \"1995\"\n },\n {\n \"authors\": \"B Scholkopf; J Platt; J Shawe-Taylor; A J Smola; R Williamson\",\n \"journal\": \"Neural Computation\",\n \"ref_id\": \"b6\",\n \"title\": \"Estimating the support of a high-dimentional distribution\",\n \"year\": \"2001\"\n },\n {\n \"authors\": \"C Y Chang; S J Lee; C C Lai\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"Weighted word2vec based on the distance of words\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"W Tian; J Li; H Li\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"A Method of Feature Selection Based on Word2Vec in Text Categorization\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"T P Van; T M Thanh\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"Vietnamese news classification based on BoW with keywords extraction and neural network\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"R Zhao; K Mao\",\n \"journal\": \"IEEE Transactions on Fuzzy Systems\",\n \"ref_id\": \"b10\",\n \"title\": \"Fuzzy Bag-of-Words Model for Document Representation\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"L Wu; S C H Hoi; N Yu\",\n \"journal\": \"IEEE Transactions on Image Processing\",\n \"ref_id\": \"b11\",\n \"title\": \"Semantics-Preserving Bag-of-Words Models and Applications\",\n \"year\": \"2010\"\n },\n {\n \"authors\": \"A Alahmadi; A Joorabchi; A E Mahdi\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"A new text representation scheme combining Bag-of-Words and Bag-of-Concepts approaches for automatic text classification\",\n \"year\": \"2013\"\n },\n {\n \"authors\": \"S M H Dadgar; M S Araghi; M M Farahani\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"A novel text mining approach based on TF-IDF and Support Vector Machine for news classification\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"X Huang; Q Wu\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"Micro-blog commercial word extraction based on improved TF-IDF algorithm\",\n \"year\": \"2013\"\n },\n {\n \"authors\": \"A Guo; T Yang\",\n \"journal\": \"\",\n \"ref_id\": \"b15\",\n \"title\": \"Research and improvement of feature words weight based on TFIDF algorithm\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"F Dubosson; S Bromuri; M Schumacher\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"A Python Framework for Exhaustive Machine Learning Algorithms and Features Evaluations\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"E Patterson; R Mcburney; H Schmidt\",\n \"journal\": \"IBM Journal of Research and Development\",\n \"ref_id\": \"b17\",\n \"title\": \"Dataflow representation of data analyses: Toward a platform for collaborative data science\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"C P Hwang; M S Chen; C M Shih\",\n \"journal\": \"\",\n \"ref_id\": \"b18\",\n \"title\": \"Apply Scikit-Learn in Python to Analyze Driver Behavior Based on OBD Data\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"M Sokolova; G Lapalme\",\n \"journal\": \"Information Processing and Management\",\n \"ref_id\": \"b19\",\n \"title\": \"A systematic analysis of performance measures for classification tasks\",\n \"year\": \"2009\"\n }\n]"},"formulas":{"kind":"list like","value":[{"formula_coordinates":[4,261.91,468.63,208.79,19.33],"formula_id":"formula_0","formula_text":"1 m i i ij j L S a (2)"},{"formula_coordinates":[4,251.02,640.75,219.67,29.21],"formula_id":"formula_1","formula_text":"log ij ij j N g tf df (3)"},{"formula_coordinates":[5,267.75,232.74,202.94,20.83],"formula_id":"formula_2","formula_text":"1 h i ij j Gg (4)"},{"formula_coordinates":[5,260.01,303.34,210.68,10.78],"formula_id":"formula_3","formula_text":"i i i i x S L G (5)"},{"formula_coordinates":[5,225.5,603.34,245.2,32.26],"formula_id":"formula_4","formula_text":" 2 ,, 1 1 min 2 subject to ( ) , 0. i i w i i i w vl w x (6)"},{"formula_coordinates":[6,232.06,224.16,238.64,83.98],"formula_id":"formula_5","formula_text":" (7) , T linear k x y x y C (8) ( , ) d T polynomial k x y x y C (9)"},{"formula_coordinates":[6,229.04,318.06,241.66,17.02],"formula_id":"formula_6","formula_text":"T sigmoid k x y x y C (10)"},{"formula_coordinates":[8,215.21,194.46,255.49,52.68],"formula_id":"formula_7","formula_text":" (12) 2 Precision Recall F measure Precision Recall (13)"}],"string":"[\n {\n \"formula_coordinates\": [\n 4,\n 261.91,\n 468.63,\n 208.79,\n 19.33\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"1 m i i ij j L S a (2)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 251.02,\n 640.75,\n 219.67,\n 29.21\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"log ij ij j N g tf df (3)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 267.75,\n 232.74,\n 202.94,\n 20.83\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"1 h i ij j Gg (4)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 260.01,\n 303.34,\n 210.68,\n 10.78\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"i i i i x S L G (5)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 225.5,\n 603.34,\n 245.2,\n 32.26\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \" 2 ,, 1 1 min 2 subject to ( ) , 0. i i w i i i w vl w x (6)\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 232.06,\n 224.16,\n 238.64,\n 83.98\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \" (7) , T linear k x y x y C (8) ( , ) d T polynomial k x y x y C (9)\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 229.04,\n 318.06,\n 241.66,\n 17.02\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"T sigmoid k x y x y C (10)\"\n },\n {\n \"formula_coordinates\": [\n 8,\n 215.21,\n 194.46,\n 255.49,\n 52.68\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \" (12) 2 Precision Recall F measure Precision Recall (13)\"\n }\n]"}}},{"rowIdx":396,"cells":{"doi":{"kind":"string","value":"10.18653/v1/S17-2001"},"pub_date":{"kind":"string","value":"2023-05-21"},"sections":{"kind":"list like","value":[{"figure_ref":[],"heading":"Introduction","publication_ref":["b3","b14","b17","b20","b30","b20","b7","b31","b11","b6","b24","b34","b35","b18","b22","b10","b25","b28","b29","b21","b16","b27"],"table_ref":[],"text":"Pretrained language models (LMs) powered by finetuning have achieved remarkable performance on a wide range of downstream tasks (Devlin et al., 2019;Liu et al., 2019;Radford et al., 2019). Driven by the pursued task-agnostic property, distillation of LMs has witnessed a paradigm shift from taskspecific to task-agnostic distillation (Sanh et al., 2019). Under a teacher-student regime, taskagnostic distillation distils pretrained LMs into ones of small compute on pretraining data so that Figure 1: The failures of prior distillation methods. The setting is to distil a base-scale teacher to a 6-layer student. Either distilling last layer self-attention distributions (Wang et al., 2021) or logits (Sanh et al., 2019) for encoder-decoder LMs yields severe degradation or only marginal gain compared to pretraining from scratch, in contrast to significant improvements for either encoderonly or decoder-only LMs. Note that the lower the perplexity, the better.\nthese small LMs can be applied to tasks by finetuning (Jiao et al., 2020;Wang et al., 2020;Liang et al., 2023). In contrast, task-specific distillation distils finetuned LMs on finetuning data and consumed resource can be even huge when the number of tasks explode (Hinton et al., 2015;Sun et al., 2019;Xia et al., 2022;Yang et al., 2022). Additionally, it is acknowledged that task-agnostic distillation typically brings performance gain over task-specific distillation does (Zhang et al., 2022a).\nDespite so many merits, prior studies mostly lie in distillation of either encoder-only LMs (e.g., BERT, Devlin et al., 2019) or decoder-only LMs (e.g., GPT, Radford et al., 2019) and largely ignore the signifance of task-agnostic distillation of encoder-decoder LMs (e.g., T5, Raffel et al., 2020) given recent advances in task-specific distillation of encoder-decoder LMs though (Shleifer and Rush, 2020;Zhang et al., 2022b;Li et al., 2022;Tao et al., 2022). Frustratingly, we find that existing distillation methods may fail to handle taskagnostic distillation of encoder-decoder LMs since encoder-decoder LMs can behave very differently in comparison with encoder-only and decoder-only LMs (e.g., the use of cross-attention, Vaswani et al., 2017). The failures of prior methods are showcased in Figure 1.\nTo the end, we investigate to, in a task-agnostic style, save the distillation of encoder-decoder LMs from the awkward position. Specifically, we reveal that the key to unlocking the expressiveness of distillation is the interplay between the encoder and the decoder. Therefore, we offer a path named as MINIEND that successfully tackles the distillation of encoder-decoder LMs by alternatively distilling the cross-attention to explicitly fall to both the encoder and the decoder.\nWe check MINIEND on language understanding and abstractive summarization in sense that encoder-decoder LMs are more capable of sequence-to-sequence tasks. For evaluation on language understanding, we take GLUE (Wang et al., 2019) to benchmark the performance. For evaluation on abstractive summarization, we adopt CNN/DailyMail (See et al., 2017) and XSum (Narayan et al., 2018) as two testbeds. The results of both distilling T5 and BART indicate that MINIEND is effective and competitive to other compression options such as quantization. We further scale our method up to the distillation of 3B T5 xlarge with the aid of progressive distillation. The results suggest that distilling large language models (e.g., LLaMA, Touvron et al., 2023) should be promising but can be challenging."},{"figure_ref":[],"heading":"Encoder-Decoder Interplay","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Architecture Perspective","publication_ref":["b28"],"table_ref":[],"text":"Typically, an encoder-decoder LM is composed of an encoder and a decoder, each of which is essentially a stack of transformer layers (Vaswani et al., 2017). Concretely, a transformer layer in the encoder contains a multihead self-attention (MSA) module and a feedforward network (FFN) module. Similarly, a transformer layer in the decoder comprises an MSA module, an FFN module, and additionally a multihead cross-attention (MCA) module that is inserted between the MSA and the FFN modules and accounts for absorption of encoded information from the encoder. Around each of these modules is attached necessarily a layer normaliza-tion and a residual connection.\nMSA and FFN Mathematically, the procedure that a transformer encoder layer consumes an intermediate encoder input X ∈ R n×d containing a n-length sequence of d-dimension vectors from last layer and gives an output to next layer can be depicted as a composition of MSA and FFN:\nMSA(X; W Q , W K , W V ) = A i SelfAttn(X; W Q i , W K i )XW V i W O i , SelfAttn(X; W Q i , W K i ) = Softmax(XW Q i W K i X /d A ), FFN(X; W I , W O ) = I j g(XW I j )W O j ,\nwhere potential details (e.g., linear bias and layer normalization) are omitted. i is used to indicate i-th head parameterized by MCA Likewise, the procedure that a transformer decoder layer processes an intermediate decoder input Z ∈ R m×d based on the final encoder output E ∈ R n×d can be incrementally described as an insertion of MCA:\nW Q i , W K i , W V i ∈ R d×d A , W O i ∈ R d A ×d among A heads,\nMCA(Z, E; W Q , W K , W V ) = A i CrossAttn(Z, E; W Q i , W K i )EW V i W O i , CrossAttn(Z, E; W Q i , W K i ) = Softmax(ZW Q i W K i E /d A ),\nHere, each cross-attention head is parameterized by another set of parameters\nW Q i , W K i , W V i ∈ R d×d A , W O i ∈ R d A ×d .\nInterplay through Architecture In this architectural sense, the decoder is tightly connected to the encoder through MCA modules. In spite that state-of-the-art methods mainly manipulate the decoder during distillation (e.g., logits, Zhang et al., 2022b), the encoder could be learned anyway through the connections offered by MCA modules. However, it is still not clear to what extent the encoder-decoder interplay is significant in the distillation and whether the implicit connections mentioned above are enough for alignment of the interplay."},{"figure_ref":[],"heading":"Gradient Perspective","publication_ref":["b9"],"table_ref":[],"text":"More thoroughly, we take a closer look at the connections between the encoder and the decoder through the lens of gradients.\nWe examine the gradient norms of last layer hidden states of both the encoder and the decoder under two distinguished distillation objectives when distilling from BART (Lewis et al., 2020). The intuition lies in that, in contrast to implicit consideration of the encoder-decoder interplay, a distillation objective explicitly involving the encoder-decoder interplay alignment could behave much differently in terms of gradients if the interplay is central to the distillation of encoder-decoder LMs. And naturally, if suboptimal cases are identified in the implicit objective, we can further highlight that the implicit objective suffers from the limited interplay alignment and the explicit objective can provide a more effective one."},{"figure_ref":["fig_0"],"heading":"Implicit versus Explicit Objective","publication_ref":["b30","b36"],"table_ref":[],"text":"We instantiate the implicit objective as aligning logits and last decoder layer self-attention distributions, and the explicit objective as aligning logits, last decoder layer self-attention distributions, and last decoder layer cross-attention distributions. The core idea of last layer attention distribution alignment is borrowed from MiniLM (Wang et al., 2021). Any alignment can be abstracted as L(S; T , D * ), where D * denotes, with slight abuse of notation, the distribution of the input. As a crucial part, the alignment of self-attention is like the following:\nL SelfAttn (S; T , D Z ) = E Z∼D Z R k=1 KL(Reln(Z; T W Q k ), Reln(Z; S W Q k )) + KL(Reln(Z; T W K k ), Reln(Z; S W K k )) + KL(Reln(Z; T W V k ), Reln(Z; S W V k )), Reln(Z; T W Q k ) = Softmax(Z T W Q k T W Q k Z /d R ),\nwhere S and T are the teacher and the student, and KL stands for kullback-leibler divergence. Particularly, attention heads are first merged from the original A attention heads and then split to R heads for alignment of the number of attention heads.\nT /S W Q k is the redistributed query parameter of the k-th head within totally R heads from the last decoder layer, likewise T /S W K k and T /S W V k are the key and value parameters.\nThe alignment of cross-attention is similar but sort of different in that the keys and the values are aligned in fact from the encoder side, as the following:\nL CrossAttn (S; T , D Z , D E ) = E Z∼D Z ,E∼D E R k=1 KL(Reln(Z; T W Q k ), Reln(Z; S W Q k )) + KL(Reln(E; T W K k ), Reln(E; S W K k )) + KL(Reln(E; T W V k ), Reln(E; S W V k )),\nHere, the notations should be self-contained by referring to previously mentioned ones.\nInterplay through Gradient To recap, the implicit objective is: Contrarily, the explicit objective is derived by adding a cross-attention term as:\nL Logit (S; T , D Z ) + L SelfAttn (S; T , D Z ),\nL Logit (S; T , D Z ) + L SelfAttn (S; T , D Z ) + L CrossAttn (S; T , D Z , D E ),\nPreliminary results are shown in Figure 2, from which we can see that 1) the implicit objective and the explicit objective lead to distinct gradient variations, and 2) the implicit objective exhibits gradient spikes, compared with smooth gradient transitions from the explicit objective, that may result in instability for a nice convergence (Zeng et al., 2022). Thereby, from the gradient perspective, we safely conclude that the encoder-decoder interplay is of importance in the distillation of encoder-decoder LMs and an explicit correspondence to the interplay is superior to an implicit one."},{"figure_ref":["fig_1"],"heading":"MINIEND","publication_ref":[],"table_ref":[],"text":"With aforementioned justifications in mind, we propose a path dubbed as MINIEND that tackles the distillation of encoder-decoder LMs under the guidance of the encoder-decoder interplay alignment. The path can be built in two directions. An overview of these two directions is given in Figure 3."},{"figure_ref":[],"heading":"Decoder Cross-Attention","publication_ref":[],"table_ref":[],"text":"The first is the one used in our pilot study. That said, we should always plus a fraction towards the alignment of output logits and the overall distillation objective is therefore depicted as:\nL(S; T , D Z , D E ) = L Logit (S; T , D Z )+ L SelfAttn (S; T , D Z ) + L CrossAttn (S; T , D Z , D E ),\nThe alignment of logits can be further detailed as:\nL Logit (S; T , D Z ) = E Z∼D Z CE(Z S W E , Z T W E ),\nwhere CE stands for soft cross entropy and T /S W E denotes output embedding.\nEncoder Self-Attention The second is an alternative to the first one where the interplay part is accounted by the last encoder self-attention distributions instead as:\nL(S; T , D Z , D X ) = L Logit (S; T , D Z )+ L SelfAttn (S; T , D Z ) + L EncSelfAttn (S; T , D X ),\nThe rationale of introducing the encoder selfattention alignment abides in that this term together with the decoder self-attention alignment can sufficiently replace the cross-attention term and align the encoder-decoder interplay by aligning both the encoder and the decoder."},{"figure_ref":[],"heading":"Experiments","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Data and Metrics","publication_ref":["b18","b5","b29","b23","b32","b4","b1","b33","b19","b0","b8"],"table_ref":["tab_2"],"text":"Following the pretraining of T5 and BART, we use C4 (Raffel et al., 2020) as the corpus for task-agnostic distillation of T5 and OpenWeb-Text (Gokaslan et al., 2019) for that of BART. They are separately processed to follow the pretraining styles of T5 and BART. That is, C4 is converted to the masked language modeling style and Open-WebText is converted to the denoising style.\nFor evaluation of MINIEND, we mainly take GLUE (Wang et al., 2019) for language understanding. The GLUE benchmark consists of two sequence classification tasks, SST-2 (Socher et al., 2013), i.e., CoLA (Warstadt et al., 2019), and seven sequence-pair classification tasks, i.e., MRPC (Dolan and Brockett, 2005), STS-B (Cer et al., 2017), QQP, MNLI (Williams et al., 2018), QNLI (Rajpurkar et al., 2016), RTE (Bentivogli et al., 2011), WNLI (Levesque et al., 2012). We exclude WNLI and CoLA due to the evaluation 1, where the corpora used for distillation is also attached."},{"figure_ref":[],"heading":"Implementation","publication_ref":["b13","b39","b31"],"table_ref":["tab_3"],"text":"The distillation is carried out on 16 Nvidia A100s. The number of relation heads is set to 32. After the distillation, the finetuning is carried out on one Nvidia A100. For language understanding tasks, T5 is finetuned with simplicity and performance guarantee following EncT5 (Liu et al., 2021) which uses the very first token (i.e., [BOS]) representation from the decoder, while BART is finetued following its original paper which uses the very last token (i.e., [EOS]) representation from the decoder. As for abstractive summarization tasks, both T5 and BART are finetuned in a sequence-to-sequence manner. For fast development, we use greedy search for T5 and beam search for BART only. The beam search setting strictly follows the original paper. In order to achieve higher training efficiency, we utilize fully-sharded data parallel (Zhao et al., 2023) to shard both the teacher and the student across GPUs during the distillation. For all cases, students are always randomly initialized before the distillation following MiniLM (Wang et al., 2020).\nThe details of hyperparameters for distillation and finetuning are shown in Table 2. We will be releasing our code and scripts in the final version for exact reproducibility."},{"figure_ref":[],"heading":"Baselines","publication_ref":["b6","b30","b10"],"table_ref":[],"text":"We name two variants of MINIEND as MINIEND-D and MINIEND-E respectively, where MINIEND-D uses decoder cross-attention for interplay alignment and MINIEND-E uses encoder self-attention instead. As there are no existing work in taskagnostic distillation of encoder-decoder LMs, we mainly compare MINIEND to task-agnostic baselines that are heavily adapted to encoder-decoder LMs and task-specific baselines that may be not super fair for comparison.\nWe compare MINIEND-D and MINIEND-E distilled from T5 to task-agnostic baselines on GLUE, CNN/DailyMail, and XSum: MlmKD (Hinton et al., 2015) that directly distils masked language modeling logits; MiniLM (Wang et al., 2021) that distils last decoder layer attention distributions; MlmKD+MiniLM that is essentially a combination of preceding two. We also compare MINIEND-D and MINIEND-E distilled from T5 to a task- specific baseline that is as far as we know the most comparable one on GLUE: MiniDisc (Zhang et al., 2022a) that exploits a teacher assistant for large compression.\nOn the other hand, we compare MINIEND-D distilled from BART to two recent task-specific baselines on CNN/DailyMail and XSum: LogitKD and DQ-BART (Li et al., 2022) that jointly quantizes and distils from the teacher.\nFor MINIEND above baselines, student structures are denoted either with *L;*H for number of layers and dimension of hidden states in random initialization, or with *% for preserved portion of parameters in pruning initialization."},{"figure_ref":[],"heading":"Main Results","publication_ref":[],"table_ref":["tab_4","tab_5"],"text":"Baselines fail, yet MINIEND triumphs. From results in Table 3 andTable 4, we can tell that baselines fail to handle the distillation of encoder-decoder LMs since they either underperform the baseline pretrained from scratch or out-perform it by only a small margin. For example, MlmKD+MiniLM achieves 84.5 versus 84.6 from T5 in GLUE Score, and 35.8 versus 35.7 from T5 in CNN/DailyMail Rg-1.\nContrarily, MiniEnD can safely escape from performance degradation and bring further performance increment. For example, MINIEND-D reaches 0.1 absolute improvement in GLUE Score, and 0.9 absolute improvement in XSum Rg-1. The improvement in GLUE Score seems to be not very significant, but can be boosted according to the ablation. That is, MINIEND-E w/o L Logit goes up to 85.0, which is notably better than 84.6 from T5 in the average sense. All count, and interplay forms the key. On another note, removing L Logit will consistently produce performance deterioration on CNN/DailyMail and XSum. We conjecture there is a tradeoff of using between using L Logit or not. Namely, the use of L Logit will offer better generative ability but worse discriminative ability, and the removal of it will work reversely. Anyway, either L CrossAttn in MINIEND-D or L SelfEncAttn in MINIEND-E shall be a crucial ingredient as the interplay alignment term is the only difference between MINIEND and MlmKD+MiniLM but results in a considerable performance gap.\nAnd it may be suspected that whether L SelfAttn is still important given that MiniLM is not an ideal choice for the distillation of encoder-decoder LMs. We suggest the use of it in two aspects: 1) MlmKD+MiniLM is better than MlmKD alone; 2) the interplay alignment will witeness a subtle performance drop after the removal of L SelfAttn , say MINIEND-D will decrease from 84.7 to 83.0 in GLUE Score. Quantization has two sides. MINIEND surpasses most of them except DQ-BART. However, we should emphasize that quantized LMs usually perform better but run much slower than distilled LMs do when compression is the same. In our case, DQ-BART uses 8 bit precision and gives rise to a 4× model size reduction which is the same as that of MINIEND. In addition to that, MINIEND is orthogonal to quantization and thus can be enhanced with other quantization schemes."},{"figure_ref":["fig_2"],"heading":"Analyses","publication_ref":["b15","b12"],"table_ref":["tab_6"],"text":"Data Scaling Some would wonder whether the huge amounts of GPU hours due to the large pre-training corpus is necessary. So we inspect the performance variation of MINIEND-D by varying data scale, which is shown in Figure 4.\nThe results generally hint that using a portion of data could hardly approximate the full data performance, though half data can achieve acceptable performance. Therefore, we suggest the use of full data in the distillation.\nModel Scaling Inspired by pioneering work finding a curse that larger teachers induces worse students, we double check the existence of the curse and offer a trial solution to the curse so that we can scale the teacher up to 3B T5 xlarge .\nFrom the results in Table 5, we observe that the curse of capacity gap still exists in our case. With the increase of teacher scale, the student performance decreases. We attempt to apply common solutions the circumvent the curse. The first is to make the student learn from a teacher assistant distilled from the teacher (Mirzadeh et al., 2020). The second is to make the student to learn from a smaller teacher and then from the teacher (Lin et al., 2023). Both two solutions inherit the idea of inserting an additional distillation step thus progressive distillation. We reveal that teacher assistantbased distillation is somewhat useful but not as excepted since T5 xlarge ⇒T5 12L;384H ⇒T5 6L;384H still does not imrpove over T5 xlarge ⇒T5 6L;384H . We claim that distilling large language models like LLaMA can therefore be appealing but challenging."},{"figure_ref":[],"heading":"Conclusions","publication_ref":[],"table_ref":[],"text":"In this paper, we aim to provide a path that successfully tackles the distillation of encoder-decoder LMs, which fails most previous methods in the area. We find through a pilot study that the encoderdecoder interplay is a key component that should be aligned in the distillation so that the distilled encoder-decoder LMs are promising. Based on the idea, we propose two directions that the encoderdecoder interplay alignment can be incorporated and verify their effectiveness on a language understanding benchmark and two abstractive summarization datasets. We further scale the distillation of encoder-decoder LMs to a 3B teacher that requires additional distillation steps. In this sense, we recommend future research to devote more efforts to exploring how large language models can be distilled."},{"figure_ref":[],"heading":"Limitations","publication_ref":["b2","b26"],"table_ref":[],"text":"This paper lacks a validation study on more recently advanced encoder-decoder LMs such as Flan (Chung et al., 2022) and UL2 (Tay et al., 2022) as well as their instruction-tuned version."}],"string":"[\n {\n \"figure_ref\": [],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b3\",\n \"b14\",\n \"b17\",\n \"b20\",\n \"b30\",\n \"b20\",\n \"b7\",\n \"b31\",\n \"b11\",\n \"b6\",\n \"b24\",\n \"b34\",\n \"b35\",\n \"b18\",\n \"b22\",\n \"b10\",\n \"b25\",\n \"b28\",\n \"b29\",\n \"b21\",\n \"b16\",\n \"b27\"\n ],\n \"table_ref\": [],\n \"text\": \"Pretrained language models (LMs) powered by finetuning have achieved remarkable performance on a wide range of downstream tasks (Devlin et al., 2019;Liu et al., 2019;Radford et al., 2019). Driven by the pursued task-agnostic property, distillation of LMs has witnessed a paradigm shift from taskspecific to task-agnostic distillation (Sanh et al., 2019). Under a teacher-student regime, taskagnostic distillation distils pretrained LMs into ones of small compute on pretraining data so that Figure 1: The failures of prior distillation methods. The setting is to distil a base-scale teacher to a 6-layer student. Either distilling last layer self-attention distributions (Wang et al., 2021) or logits (Sanh et al., 2019) for encoder-decoder LMs yields severe degradation or only marginal gain compared to pretraining from scratch, in contrast to significant improvements for either encoderonly or decoder-only LMs. Note that the lower the perplexity, the better.\\nthese small LMs can be applied to tasks by finetuning (Jiao et al., 2020;Wang et al., 2020;Liang et al., 2023). In contrast, task-specific distillation distils finetuned LMs on finetuning data and consumed resource can be even huge when the number of tasks explode (Hinton et al., 2015;Sun et al., 2019;Xia et al., 2022;Yang et al., 2022). Additionally, it is acknowledged that task-agnostic distillation typically brings performance gain over task-specific distillation does (Zhang et al., 2022a).\\nDespite so many merits, prior studies mostly lie in distillation of either encoder-only LMs (e.g., BERT, Devlin et al., 2019) or decoder-only LMs (e.g., GPT, Radford et al., 2019) and largely ignore the signifance of task-agnostic distillation of encoder-decoder LMs (e.g., T5, Raffel et al., 2020) given recent advances in task-specific distillation of encoder-decoder LMs though (Shleifer and Rush, 2020;Zhang et al., 2022b;Li et al., 2022;Tao et al., 2022). Frustratingly, we find that existing distillation methods may fail to handle taskagnostic distillation of encoder-decoder LMs since encoder-decoder LMs can behave very differently in comparison with encoder-only and decoder-only LMs (e.g., the use of cross-attention, Vaswani et al., 2017). The failures of prior methods are showcased in Figure 1.\\nTo the end, we investigate to, in a task-agnostic style, save the distillation of encoder-decoder LMs from the awkward position. Specifically, we reveal that the key to unlocking the expressiveness of distillation is the interplay between the encoder and the decoder. Therefore, we offer a path named as MINIEND that successfully tackles the distillation of encoder-decoder LMs by alternatively distilling the cross-attention to explicitly fall to both the encoder and the decoder.\\nWe check MINIEND on language understanding and abstractive summarization in sense that encoder-decoder LMs are more capable of sequence-to-sequence tasks. For evaluation on language understanding, we take GLUE (Wang et al., 2019) to benchmark the performance. For evaluation on abstractive summarization, we adopt CNN/DailyMail (See et al., 2017) and XSum (Narayan et al., 2018) as two testbeds. The results of both distilling T5 and BART indicate that MINIEND is effective and competitive to other compression options such as quantization. We further scale our method up to the distillation of 3B T5 xlarge with the aid of progressive distillation. The results suggest that distilling large language models (e.g., LLaMA, Touvron et al., 2023) should be promising but can be challenging.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Encoder-Decoder Interplay\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Architecture Perspective\",\n \"publication_ref\": [\n \"b28\"\n ],\n \"table_ref\": [],\n \"text\": \"Typically, an encoder-decoder LM is composed of an encoder and a decoder, each of which is essentially a stack of transformer layers (Vaswani et al., 2017). Concretely, a transformer layer in the encoder contains a multihead self-attention (MSA) module and a feedforward network (FFN) module. Similarly, a transformer layer in the decoder comprises an MSA module, an FFN module, and additionally a multihead cross-attention (MCA) module that is inserted between the MSA and the FFN modules and accounts for absorption of encoded information from the encoder. Around each of these modules is attached necessarily a layer normaliza-tion and a residual connection.\\nMSA and FFN Mathematically, the procedure that a transformer encoder layer consumes an intermediate encoder input X ∈ R n×d containing a n-length sequence of d-dimension vectors from last layer and gives an output to next layer can be depicted as a composition of MSA and FFN:\\nMSA(X; W Q , W K , W V ) = A i SelfAttn(X; W Q i , W K i )XW V i W O i , SelfAttn(X; W Q i , W K i ) = Softmax(XW Q i W K i X /d A ), FFN(X; W I , W O ) = I j g(XW I j )W O j ,\\nwhere potential details (e.g., linear bias and layer normalization) are omitted. i is used to indicate i-th head parameterized by MCA Likewise, the procedure that a transformer decoder layer processes an intermediate decoder input Z ∈ R m×d based on the final encoder output E ∈ R n×d can be incrementally described as an insertion of MCA:\\nW Q i , W K i , W V i ∈ R d×d A , W O i ∈ R d A ×d among A heads,\\nMCA(Z, E; W Q , W K , W V ) = A i CrossAttn(Z, E; W Q i , W K i )EW V i W O i , CrossAttn(Z, E; W Q i , W K i ) = Softmax(ZW Q i W K i E /d A ),\\nHere, each cross-attention head is parameterized by another set of parameters\\nW Q i , W K i , W V i ∈ R d×d A , W O i ∈ R d A ×d .\\nInterplay through Architecture In this architectural sense, the decoder is tightly connected to the encoder through MCA modules. In spite that state-of-the-art methods mainly manipulate the decoder during distillation (e.g., logits, Zhang et al., 2022b), the encoder could be learned anyway through the connections offered by MCA modules. However, it is still not clear to what extent the encoder-decoder interplay is significant in the distillation and whether the implicit connections mentioned above are enough for alignment of the interplay.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Gradient Perspective\",\n \"publication_ref\": [\n \"b9\"\n ],\n \"table_ref\": [],\n \"text\": \"More thoroughly, we take a closer look at the connections between the encoder and the decoder through the lens of gradients.\\nWe examine the gradient norms of last layer hidden states of both the encoder and the decoder under two distinguished distillation objectives when distilling from BART (Lewis et al., 2020). The intuition lies in that, in contrast to implicit consideration of the encoder-decoder interplay, a distillation objective explicitly involving the encoder-decoder interplay alignment could behave much differently in terms of gradients if the interplay is central to the distillation of encoder-decoder LMs. And naturally, if suboptimal cases are identified in the implicit objective, we can further highlight that the implicit objective suffers from the limited interplay alignment and the explicit objective can provide a more effective one.\"\n },\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"Implicit versus Explicit Objective\",\n \"publication_ref\": [\n \"b30\",\n \"b36\"\n ],\n \"table_ref\": [],\n \"text\": \"We instantiate the implicit objective as aligning logits and last decoder layer self-attention distributions, and the explicit objective as aligning logits, last decoder layer self-attention distributions, and last decoder layer cross-attention distributions. The core idea of last layer attention distribution alignment is borrowed from MiniLM (Wang et al., 2021). Any alignment can be abstracted as L(S; T , D * ), where D * denotes, with slight abuse of notation, the distribution of the input. As a crucial part, the alignment of self-attention is like the following:\\nL SelfAttn (S; T , D Z ) = E Z∼D Z R k=1 KL(Reln(Z; T W Q k ), Reln(Z; S W Q k )) + KL(Reln(Z; T W K k ), Reln(Z; S W K k )) + KL(Reln(Z; T W V k ), Reln(Z; S W V k )), Reln(Z; T W Q k ) = Softmax(Z T W Q k T W Q k Z /d R ),\\nwhere S and T are the teacher and the student, and KL stands for kullback-leibler divergence. Particularly, attention heads are first merged from the original A attention heads and then split to R heads for alignment of the number of attention heads.\\nT /S W Q k is the redistributed query parameter of the k-th head within totally R heads from the last decoder layer, likewise T /S W K k and T /S W V k are the key and value parameters.\\nThe alignment of cross-attention is similar but sort of different in that the keys and the values are aligned in fact from the encoder side, as the following:\\nL CrossAttn (S; T , D Z , D E ) = E Z∼D Z ,E∼D E R k=1 KL(Reln(Z; T W Q k ), Reln(Z; S W Q k )) + KL(Reln(E; T W K k ), Reln(E; S W K k )) + KL(Reln(E; T W V k ), Reln(E; S W V k )),\\nHere, the notations should be self-contained by referring to previously mentioned ones.\\nInterplay through Gradient To recap, the implicit objective is: Contrarily, the explicit objective is derived by adding a cross-attention term as:\\nL Logit (S; T , D Z ) + L SelfAttn (S; T , D Z ),\\nL Logit (S; T , D Z ) + L SelfAttn (S; T , D Z ) + L CrossAttn (S; T , D Z , D E ),\\nPreliminary results are shown in Figure 2, from which we can see that 1) the implicit objective and the explicit objective lead to distinct gradient variations, and 2) the implicit objective exhibits gradient spikes, compared with smooth gradient transitions from the explicit objective, that may result in instability for a nice convergence (Zeng et al., 2022). Thereby, from the gradient perspective, we safely conclude that the encoder-decoder interplay is of importance in the distillation of encoder-decoder LMs and an explicit correspondence to the interplay is superior to an implicit one.\"\n },\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"MINIEND\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"With aforementioned justifications in mind, we propose a path dubbed as MINIEND that tackles the distillation of encoder-decoder LMs under the guidance of the encoder-decoder interplay alignment. The path can be built in two directions. An overview of these two directions is given in Figure 3.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Decoder Cross-Attention\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The first is the one used in our pilot study. That said, we should always plus a fraction towards the alignment of output logits and the overall distillation objective is therefore depicted as:\\nL(S; T , D Z , D E ) = L Logit (S; T , D Z )+ L SelfAttn (S; T , D Z ) + L CrossAttn (S; T , D Z , D E ),\\nThe alignment of logits can be further detailed as:\\nL Logit (S; T , D Z ) = E Z∼D Z CE(Z S W E , Z T W E ),\\nwhere CE stands for soft cross entropy and T /S W E denotes output embedding.\\nEncoder Self-Attention The second is an alternative to the first one where the interplay part is accounted by the last encoder self-attention distributions instead as:\\nL(S; T , D Z , D X ) = L Logit (S; T , D Z )+ L SelfAttn (S; T , D Z ) + L EncSelfAttn (S; T , D X ),\\nThe rationale of introducing the encoder selfattention alignment abides in that this term together with the decoder self-attention alignment can sufficiently replace the cross-attention term and align the encoder-decoder interplay by aligning both the encoder and the decoder.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiments\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Data and Metrics\",\n \"publication_ref\": [\n \"b18\",\n \"b5\",\n \"b29\",\n \"b23\",\n \"b32\",\n \"b4\",\n \"b1\",\n \"b33\",\n \"b19\",\n \"b0\",\n \"b8\"\n ],\n \"table_ref\": [\n \"tab_2\"\n ],\n \"text\": \"Following the pretraining of T5 and BART, we use C4 (Raffel et al., 2020) as the corpus for task-agnostic distillation of T5 and OpenWeb-Text (Gokaslan et al., 2019) for that of BART. They are separately processed to follow the pretraining styles of T5 and BART. That is, C4 is converted to the masked language modeling style and Open-WebText is converted to the denoising style.\\nFor evaluation of MINIEND, we mainly take GLUE (Wang et al., 2019) for language understanding. The GLUE benchmark consists of two sequence classification tasks, SST-2 (Socher et al., 2013), i.e., CoLA (Warstadt et al., 2019), and seven sequence-pair classification tasks, i.e., MRPC (Dolan and Brockett, 2005), STS-B (Cer et al., 2017), QQP, MNLI (Williams et al., 2018), QNLI (Rajpurkar et al., 2016), RTE (Bentivogli et al., 2011), WNLI (Levesque et al., 2012). We exclude WNLI and CoLA due to the evaluation 1, where the corpora used for distillation is also attached.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Implementation\",\n \"publication_ref\": [\n \"b13\",\n \"b39\",\n \"b31\"\n ],\n \"table_ref\": [\n \"tab_3\"\n ],\n \"text\": \"The distillation is carried out on 16 Nvidia A100s. The number of relation heads is set to 32. After the distillation, the finetuning is carried out on one Nvidia A100. For language understanding tasks, T5 is finetuned with simplicity and performance guarantee following EncT5 (Liu et al., 2021) which uses the very first token (i.e., [BOS]) representation from the decoder, while BART is finetued following its original paper which uses the very last token (i.e., [EOS]) representation from the decoder. As for abstractive summarization tasks, both T5 and BART are finetuned in a sequence-to-sequence manner. For fast development, we use greedy search for T5 and beam search for BART only. The beam search setting strictly follows the original paper. In order to achieve higher training efficiency, we utilize fully-sharded data parallel (Zhao et al., 2023) to shard both the teacher and the student across GPUs during the distillation. For all cases, students are always randomly initialized before the distillation following MiniLM (Wang et al., 2020).\\nThe details of hyperparameters for distillation and finetuning are shown in Table 2. We will be releasing our code and scripts in the final version for exact reproducibility.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Baselines\",\n \"publication_ref\": [\n \"b6\",\n \"b30\",\n \"b10\"\n ],\n \"table_ref\": [],\n \"text\": \"We name two variants of MINIEND as MINIEND-D and MINIEND-E respectively, where MINIEND-D uses decoder cross-attention for interplay alignment and MINIEND-E uses encoder self-attention instead. As there are no existing work in taskagnostic distillation of encoder-decoder LMs, we mainly compare MINIEND to task-agnostic baselines that are heavily adapted to encoder-decoder LMs and task-specific baselines that may be not super fair for comparison.\\nWe compare MINIEND-D and MINIEND-E distilled from T5 to task-agnostic baselines on GLUE, CNN/DailyMail, and XSum: MlmKD (Hinton et al., 2015) that directly distils masked language modeling logits; MiniLM (Wang et al., 2021) that distils last decoder layer attention distributions; MlmKD+MiniLM that is essentially a combination of preceding two. We also compare MINIEND-D and MINIEND-E distilled from T5 to a task- specific baseline that is as far as we know the most comparable one on GLUE: MiniDisc (Zhang et al., 2022a) that exploits a teacher assistant for large compression.\\nOn the other hand, we compare MINIEND-D distilled from BART to two recent task-specific baselines on CNN/DailyMail and XSum: LogitKD and DQ-BART (Li et al., 2022) that jointly quantizes and distils from the teacher.\\nFor MINIEND above baselines, student structures are denoted either with *L;*H for number of layers and dimension of hidden states in random initialization, or with *% for preserved portion of parameters in pruning initialization.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Main Results\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_4\",\n \"tab_5\"\n ],\n \"text\": \"Baselines fail, yet MINIEND triumphs. From results in Table 3 andTable 4, we can tell that baselines fail to handle the distillation of encoder-decoder LMs since they either underperform the baseline pretrained from scratch or out-perform it by only a small margin. For example, MlmKD+MiniLM achieves 84.5 versus 84.6 from T5 in GLUE Score, and 35.8 versus 35.7 from T5 in CNN/DailyMail Rg-1.\\nContrarily, MiniEnD can safely escape from performance degradation and bring further performance increment. For example, MINIEND-D reaches 0.1 absolute improvement in GLUE Score, and 0.9 absolute improvement in XSum Rg-1. The improvement in GLUE Score seems to be not very significant, but can be boosted according to the ablation. That is, MINIEND-E w/o L Logit goes up to 85.0, which is notably better than 84.6 from T5 in the average sense. All count, and interplay forms the key. On another note, removing L Logit will consistently produce performance deterioration on CNN/DailyMail and XSum. We conjecture there is a tradeoff of using between using L Logit or not. Namely, the use of L Logit will offer better generative ability but worse discriminative ability, and the removal of it will work reversely. Anyway, either L CrossAttn in MINIEND-D or L SelfEncAttn in MINIEND-E shall be a crucial ingredient as the interplay alignment term is the only difference between MINIEND and MlmKD+MiniLM but results in a considerable performance gap.\\nAnd it may be suspected that whether L SelfAttn is still important given that MiniLM is not an ideal choice for the distillation of encoder-decoder LMs. We suggest the use of it in two aspects: 1) MlmKD+MiniLM is better than MlmKD alone; 2) the interplay alignment will witeness a subtle performance drop after the removal of L SelfAttn , say MINIEND-D will decrease from 84.7 to 83.0 in GLUE Score. Quantization has two sides. MINIEND surpasses most of them except DQ-BART. However, we should emphasize that quantized LMs usually perform better but run much slower than distilled LMs do when compression is the same. In our case, DQ-BART uses 8 bit precision and gives rise to a 4× model size reduction which is the same as that of MINIEND. In addition to that, MINIEND is orthogonal to quantization and thus can be enhanced with other quantization schemes.\"\n },\n {\n \"figure_ref\": [\n \"fig_2\"\n ],\n \"heading\": \"Analyses\",\n \"publication_ref\": [\n \"b15\",\n \"b12\"\n ],\n \"table_ref\": [\n \"tab_6\"\n ],\n \"text\": \"Data Scaling Some would wonder whether the huge amounts of GPU hours due to the large pre-training corpus is necessary. So we inspect the performance variation of MINIEND-D by varying data scale, which is shown in Figure 4.\\nThe results generally hint that using a portion of data could hardly approximate the full data performance, though half data can achieve acceptable performance. Therefore, we suggest the use of full data in the distillation.\\nModel Scaling Inspired by pioneering work finding a curse that larger teachers induces worse students, we double check the existence of the curse and offer a trial solution to the curse so that we can scale the teacher up to 3B T5 xlarge .\\nFrom the results in Table 5, we observe that the curse of capacity gap still exists in our case. With the increase of teacher scale, the student performance decreases. We attempt to apply common solutions the circumvent the curse. The first is to make the student learn from a teacher assistant distilled from the teacher (Mirzadeh et al., 2020). The second is to make the student to learn from a smaller teacher and then from the teacher (Lin et al., 2023). Both two solutions inherit the idea of inserting an additional distillation step thus progressive distillation. We reveal that teacher assistantbased distillation is somewhat useful but not as excepted since T5 xlarge ⇒T5 12L;384H ⇒T5 6L;384H still does not imrpove over T5 xlarge ⇒T5 6L;384H . We claim that distilling large language models like LLaMA can therefore be appealing but challenging.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusions\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this paper, we aim to provide a path that successfully tackles the distillation of encoder-decoder LMs, which fails most previous methods in the area. We find through a pilot study that the encoderdecoder interplay is a key component that should be aligned in the distillation so that the distilled encoder-decoder LMs are promising. Based on the idea, we propose two directions that the encoderdecoder interplay alignment can be incorporated and verify their effectiveness on a language understanding benchmark and two abstractive summarization datasets. We further scale the distillation of encoder-decoder LMs to a 3B teacher that requires additional distillation steps. In this sense, we recommend future research to devote more efforts to exploring how large language models can be distilled.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Limitations\",\n \"publication_ref\": [\n \"b2\",\n \"b26\"\n ],\n \"table_ref\": [],\n \"text\": \"This paper lacks a validation study on more recently advanced encoder-decoder LMs such as Flan (Chung et al., 2022) and UL2 (Tay et al., 2022) as well as their instruction-tuned version.\"\n }\n]"},"abstract":{"kind":"string","value":"Finetuning pretrained language models (LMs) have enabled appealing performance on a diverse array of tasks. The intriguing taskagnostic property has driven a shifted focus from task-specific to task-agnostic distillation of LMs. While task-agnostic, computeefficient, performance-preserved LMs can be yielded by task-agnostic distillation, previous studies mainly sit in distillation of either encoder-only LMs (e.g., BERT) or decoderonly ones (e.g., GPT) yet largely neglect that distillation of encoder-decoder LMs (e.g., T5) can posit very distinguished behaviors. Frustratingly, we discover that existing taskagnostic distillation methods can fail to handle the distillation of encoder-decoder LMs. To the demand, we explore a few paths and uncover a path named as MINIEND that successfully tackles the distillation of encoderdecoder LMs in a task-agnostic fashion. We examine MINIEND on language understanding and abstractive summarization. The results showcase that MINIEND is generally effective and is competitive compared to other alternatives. We further scale MINIEND up to distillation of 3B encoder-decoder language models with interpolated distillation. The results imply the opportunities and challenges in distilling large language models (e.g., LLaMA)."},"title":{"kind":"string","value":"Task-agnostic Distillation of Encoder-Decoder Language Models"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 2 :2Figure2: The preliminary results of gradient norms when using the implicit or explicit objective. The implicit objective imposes distinct gradient variations and unexpected gradient spikes during the distillation.","figure_data":"","figure_id":"fig_0","figure_label":"2","figure_type":"figure"},{"figure_caption":"Figure 3 :3Figure 3: The overview of MINIEND. Two directions are proposed to consider the encoder-decoder interplay alignment.","figure_data":"","figure_id":"fig_1","figure_label":"3","figure_type":"figure"},{"figure_caption":"Figure 4 :4Figure 4: The results of data scaling using MINIEND-D.","figure_data":"","figure_id":"fig_2","figure_label":"4","figure_type":"figure"},{"figure_caption":"The data statistics, maximum sequence lengths, and metrics. The maximum decoder sequence lengths of T5 and BART are indicated differently for language understanding tasks since they use different finetuning strategies.","figure_data":"Dataset#Train exam #Dev exam Max enc len Max dec lenMetricC4364.9M-512114-OpenWebText37.8M-512512-SST-267.3K0.9K641 / 64AccuracyMRPC3.7K0.4K1281 / 128F1STS-B7.0K1.5K1281 / 128 Spearman CorrelationQQP364.0K40.0K1281 / 128F1MNLI-m/mm393.0K20.0K1281 / 128AccuracyQNLI105.0K5.5K1281 / 128AccuracyRTE2.5K0.3K1281 / 128AccuracyCNN/DailyMail287.1K13.4K512128F1XSum204.0K11.3K512128F1inconsistency (in other words, MiniLMs get dra-matically worse results while LMs get much betterones as found out in Xia et al., 2022) and use theleft tasks. Following BERT (Devlin et al., 2019),we report Accuracy (Acc) on SST-2, MNLI, QNLI,RTE, Spearman Correlation scores (SpCorr) onSTS-B, and F1 on MRPC, QQP, CoNLL. Aver-age score over tasks from GLUE (GLUE Score) isadditionally computed. Regarding that one of themost promising properties of encoder-decoder LMsis sequence-to-sequence modeling, we addition-ally adopt CNN/DailyMail (See et al., 2017) andXSum (Narayan et al., 2018) for abstractive sum-marization. We report Rouge-{1,2,L} (Rg-{1,2,L})on both of them. Results are reported on develop-ment sets. GFLOPs are also attached as theoreticalspeedup references.The detailed data statistics, maximum sequencelengths, and metrics for datasets we use are shownin Table","figure_id":"tab_2","figure_label":"1","figure_type":"table"},{"figure_caption":"The hyperparameters for both distillation and finetuning. In order to realize the global batch size, necessary gradient accumulations should be used. The beam search setting applies to BART only.","figure_data":"HyperparameterC4Distillation OpenWebTextGLUEFinetuning CNN/DailyMailXSumBatch size10241024{16,32}{16,32}{16,32}OptimizerAdamWAdamWAdamWAdamWAdamWLearning rate3e-43e-4{1e-5,2e-5,3e-5} {1e-4,2e-4,3e-4} {1e-4,2e-4,3e-4}Training epochs15101010Earlystop epochs--555Warmup proportion0.010.010.10.10.1Weight decay0.010.010.010.010.01Number of beams---46Length penalty---2.01.0","figure_id":"tab_3","figure_label":"2","figure_type":"table"},{"figure_caption":"The results on GLUE. The best results are boldfaced.MiniDisc is distilled from T5 xlarge , and owns larger GFLOPs.","figure_data":"MethodGFLOPsSST-2 MRPC STS-B QQP MNLI-m/mm QNLI RTE GLUE Acc F1 SpCorr F1 Acc Acc Acc ScoreT5 base25.41×94.693.090.088.986.7/86.892.974.788.5T5 6L;384H3.1892.290.286.087.381.2/81.788.270.084.6MiniDisc 5% MlmKD 6L;384H MiniLM 6L;384H7.80 3.18 3.183∼8×93.8 92.3 92.189.8 88.7 89.685.3 86.2 85.286.7 87.5 87.082.9/82.7 81.6/82.1 81.2/81.589.2 88.2 88.064.6 67.9 68.684.4 84.3 84.1MlmKD+MiniLM 6L;384H 3.1892.489.286.087.381.7/82.189.167.984.5MINIEND-D 6L;384H3.1892.190.685.887.781.8/82.389.068.684.7w/o L Logit MINIEND-E 6L;384H3.18 3.188×92.2 92.790.1 90.086.6 86.187.6 87.482.2/82.8 81.8/82.189.1 88.868.6 69.384.9 84.8w/o L Logit3.1892.389.986.687.782.5/83.189.269.085.0","figure_id":"tab_4","figure_label":"3","figure_type":"table"},{"figure_caption":"The results on CNN/DailyMail and XSum. The best results are boldfaced.LogitKD is distilled with an asymmetric layer setting, i.e., more encoder layers than decoder layers, for saved performance decline. DQ-BART only quantizes parameter precision to lower one, i.e., 8 bit, but does not reduce parameter amount. Quantization would not give any speedup in GFLOPs though nice reduction in model size.","figure_data":"MethodGFLOPsCNN/DailyMail Rg-1 Rg-2 Rg-L Rg-1 Rg-2 Rg-L XSumT5 base25.41×40.1 19.4 31.534.7 12.4 29.7T5 6L;384H3.1835.7 16.8 28.428.68.924.8MlmKD 6L;384H MiniLM 6L;384H3.18 3.188×36.0 17.0 28.7 35.0 16.5 28.028.9 25.99.2 7.525.0 22.5MlmKD+MiniLM 6L;384H 3.1835.8 17.0 28.729.09.125.1MINIEND-D 6L;384H3.1836.2 17.2 28.929.59.225.4w/o L Logit MINIEND-E 6L;384H3.18 3.188×35.7 17.0 28.6 36.1 17.3 28.927.3 28.98.2 9.123.7 24.9w/o L Logit3.1835.8 17.1 28.727.28.023.6BART base12.71×39.4 18.5 30.636.9 14.7 31.9LogitKD 3/1L;768H DQ-BART 8bit4.23 12.71∼3×38.0 16.0 25.2 42.4 19.3 28.832.9 12.4 26.9 38.2 15.7 30.7MINIEND-D 6L;384H3.184×38.5 18.5 29.733.6 12.9 29.2","figure_id":"tab_5","figure_label":"4","figure_type":"table"},{"figure_caption":"The results of model scaling using MINIEND-D. ⇒ denotes a distillation step, which should be operated sequentially otherwise {} is prioritized.• • • ⇒ • • • ⇒ • • • indicates teacher assistant-based distillation and • • • ⇒ {• • • ⇒ • • • } indicates progressive distillation.","figure_data":"MethodGLUE Score Rg-1 Rg-2 Rg-L Rg-1 Rg-2 Rg-L CNN/DailyMail XSumT5 6L;384H84.635.7 16.8 28.428.68.924.8T5 12L;384H85.037.2 17.9 29.631.2 10.5 27.0T5 base88.540.1 19.4 31.534.7 12.4 29.7T5 large90.740.6 19.4 31.738.2 15.1 32.9T5 xlarge92.040.8 19.7 32.141.1 17.6 35.5T5 base ⇒T5 6L;384H84.736.2 17.2 28.929.59.225.4T5 large ⇒T5 6L;384H84.536.4 17.4 29.029.49.325.3T5 xlarge ⇒T5 6L;384H84.236.1 17.2 28.829.19.125.1T5 xlarge ⇒T5 12L;384H ⇒T5 6L;384H84.636.6 17.5 29.229.29.125.1T5 large ⇒T5 12L;384H85.538.3 18.4 30.432.4 11.2 27.9T5 xlarge ⇒T5 12L;384H85.238.0 18.4 30.332.2 11.1 27.7T5 xlarge ⇒{T5 large ⇒T5 12L;384H }85.838.4 18.5 30.632.9 11.5 28.3in some cases. Nonetheless, we unearth thatprogressive distillation is more promising interms of consistent performance gains whencomparing T5 xlarge ⇒{T5 large ⇒T5 12L;384H } toT5 xlarge ⇒T5 12L;384H","figure_id":"tab_6","figure_label":"5","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 2 :2Figure2: The preliminary results of gradient norms when using the implicit or explicit objective. The implicit objective imposes distinct gradient variations and unexpected gradient spikes during the distillation.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 :3Figure 3: The overview of MINIEND. Two directions are proposed to consider the encoder-decoder interplay alignment.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 4 :4Figure 4: The results of data scaling using MINIEND-D.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"The data statistics, maximum sequence lengths, and metrics. The maximum decoder sequence lengths of T5 and BART are indicated differently for language understanding tasks since they use different finetuning strategies.\",\n \"figure_data\": \"Dataset#Train exam #Dev exam Max enc len Max dec lenMetricC4364.9M-512114-OpenWebText37.8M-512512-SST-267.3K0.9K641 / 64AccuracyMRPC3.7K0.4K1281 / 128F1STS-B7.0K1.5K1281 / 128 Spearman CorrelationQQP364.0K40.0K1281 / 128F1MNLI-m/mm393.0K20.0K1281 / 128AccuracyQNLI105.0K5.5K1281 / 128AccuracyRTE2.5K0.3K1281 / 128AccuracyCNN/DailyMail287.1K13.4K512128F1XSum204.0K11.3K512128F1inconsistency (in other words, MiniLMs get dra-matically worse results while LMs get much betterones as found out in Xia et al., 2022) and use theleft tasks. Following BERT (Devlin et al., 2019),we report Accuracy (Acc) on SST-2, MNLI, QNLI,RTE, Spearman Correlation scores (SpCorr) onSTS-B, and F1 on MRPC, QQP, CoNLL. Aver-age score over tasks from GLUE (GLUE Score) isadditionally computed. Regarding that one of themost promising properties of encoder-decoder LMsis sequence-to-sequence modeling, we addition-ally adopt CNN/DailyMail (See et al., 2017) andXSum (Narayan et al., 2018) for abstractive sum-marization. We report Rouge-{1,2,L} (Rg-{1,2,L})on both of them. Results are reported on develop-ment sets. GFLOPs are also attached as theoreticalspeedup references.The detailed data statistics, maximum sequencelengths, and metrics for datasets we use are shownin Table\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"The hyperparameters for both distillation and finetuning. In order to realize the global batch size, necessary gradient accumulations should be used. The beam search setting applies to BART only.\",\n \"figure_data\": \"HyperparameterC4Distillation OpenWebTextGLUEFinetuning CNN/DailyMailXSumBatch size10241024{16,32}{16,32}{16,32}OptimizerAdamWAdamWAdamWAdamWAdamWLearning rate3e-43e-4{1e-5,2e-5,3e-5} {1e-4,2e-4,3e-4} {1e-4,2e-4,3e-4}Training epochs15101010Earlystop epochs--555Warmup proportion0.010.010.10.10.1Weight decay0.010.010.010.010.01Number of beams---46Length penalty---2.01.0\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"The results on GLUE. The best results are boldfaced.MiniDisc is distilled from T5 xlarge , and owns larger GFLOPs.\",\n \"figure_data\": \"MethodGFLOPsSST-2 MRPC STS-B QQP MNLI-m/mm QNLI RTE GLUE Acc F1 SpCorr F1 Acc Acc Acc ScoreT5 base25.41×94.693.090.088.986.7/86.892.974.788.5T5 6L;384H3.1892.290.286.087.381.2/81.788.270.084.6MiniDisc 5% MlmKD 6L;384H MiniLM 6L;384H7.80 3.18 3.183∼8×93.8 92.3 92.189.8 88.7 89.685.3 86.2 85.286.7 87.5 87.082.9/82.7 81.6/82.1 81.2/81.589.2 88.2 88.064.6 67.9 68.684.4 84.3 84.1MlmKD+MiniLM 6L;384H 3.1892.489.286.087.381.7/82.189.167.984.5MINIEND-D 6L;384H3.1892.190.685.887.781.8/82.389.068.684.7w/o L Logit MINIEND-E 6L;384H3.18 3.188×92.2 92.790.1 90.086.6 86.187.6 87.482.2/82.8 81.8/82.189.1 88.868.6 69.384.9 84.8w/o L Logit3.1892.389.986.687.782.5/83.189.269.085.0\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"The results on CNN/DailyMail and XSum. The best results are boldfaced.LogitKD is distilled with an asymmetric layer setting, i.e., more encoder layers than decoder layers, for saved performance decline. DQ-BART only quantizes parameter precision to lower one, i.e., 8 bit, but does not reduce parameter amount. Quantization would not give any speedup in GFLOPs though nice reduction in model size.\",\n \"figure_data\": \"MethodGFLOPsCNN/DailyMail Rg-1 Rg-2 Rg-L Rg-1 Rg-2 Rg-L XSumT5 base25.41×40.1 19.4 31.534.7 12.4 29.7T5 6L;384H3.1835.7 16.8 28.428.68.924.8MlmKD 6L;384H MiniLM 6L;384H3.18 3.188×36.0 17.0 28.7 35.0 16.5 28.028.9 25.99.2 7.525.0 22.5MlmKD+MiniLM 6L;384H 3.1835.8 17.0 28.729.09.125.1MINIEND-D 6L;384H3.1836.2 17.2 28.929.59.225.4w/o L Logit MINIEND-E 6L;384H3.18 3.188×35.7 17.0 28.6 36.1 17.3 28.927.3 28.98.2 9.123.7 24.9w/o L Logit3.1835.8 17.1 28.727.28.023.6BART base12.71×39.4 18.5 30.636.9 14.7 31.9LogitKD 3/1L;768H DQ-BART 8bit4.23 12.71∼3×38.0 16.0 25.2 42.4 19.3 28.832.9 12.4 26.9 38.2 15.7 30.7MINIEND-D 6L;384H3.184×38.5 18.5 29.733.6 12.9 29.2\",\n \"figure_id\": \"tab_5\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"The results of model scaling using MINIEND-D. ⇒ denotes a distillation step, which should be operated sequentially otherwise {} is prioritized.• • • ⇒ • • • ⇒ • • • indicates teacher assistant-based distillation and • • • ⇒ {• • • ⇒ • • • } indicates progressive distillation.\",\n \"figure_data\": \"MethodGLUE Score Rg-1 Rg-2 Rg-L Rg-1 Rg-2 Rg-L CNN/DailyMail XSumT5 6L;384H84.635.7 16.8 28.428.68.924.8T5 12L;384H85.037.2 17.9 29.631.2 10.5 27.0T5 base88.540.1 19.4 31.534.7 12.4 29.7T5 large90.740.6 19.4 31.738.2 15.1 32.9T5 xlarge92.040.8 19.7 32.141.1 17.6 35.5T5 base ⇒T5 6L;384H84.736.2 17.2 28.929.59.225.4T5 large ⇒T5 6L;384H84.536.4 17.4 29.029.49.325.3T5 xlarge ⇒T5 6L;384H84.236.1 17.2 28.829.19.125.1T5 xlarge ⇒T5 12L;384H ⇒T5 6L;384H84.636.6 17.5 29.229.29.125.1T5 large ⇒T5 12L;384H85.538.3 18.4 30.432.4 11.2 27.9T5 xlarge ⇒T5 12L;384H85.238.0 18.4 30.332.2 11.1 27.7T5 xlarge ⇒{T5 large ⇒T5 12L;384H }85.838.4 18.5 30.632.9 11.5 28.3in some cases. Nonetheless, we unearth thatprogressive distillation is more promising interms of consistent performance gains whencomparing T5 xlarge ⇒{T5 large ⇒T5 12L;384H } toT5 xlarge ⇒T5 12L;384H\",\n \"figure_id\": \"tab_6\",\n \"figure_label\": \"5\",\n \"figure_type\": \"table\"\n }\n]"},"authors":{"kind":"string","value":"Chen Zhang; Yang Yang; Jingang Wang; Dawei Song"},"references":{"kind":"list like","value":[{"authors":"Luisa Bentivogli; Peter Clark; Ido Dagan; Danilo Giampiccolo","journal":"","ref_id":"b0","title":"The seventh PASCAL recognizing textual entailment challenge","year":"2011-11-14"},{"authors":"M Daniel; Mona T Cer; Eneko Diab; Iñigo Agirre; Lucia Lopez-Gazpio; Specia","journal":"","ref_id":"b1","title":"Semeval-2017 task 1: Semantic textual similarity multilingual and crosslingual focused evaluation","year":"2017-08-03"},{"authors":"Chung Hyung Won; Le Hou; Shayne Longpre; Barret Zoph; Yi Tay; William Fedus; Eric Li; Xuezhi Wang; Mostafa Dehghani; Siddhartha Brahma; Albert Webson; Shane Shixiang; Zhuyun Gu; Mirac Dai; Xinyun Suzgun; Aakanksha Chen; Sharan Chowdhery; Gaurav Narang; Adams Mishra; Vincent Y Yu; Yanping Zhao; Andrew M Huang; Hongkun Dai; Slav Yu; Ed H Petrov; Jeff Chi; Jacob Dean; Adam Devlin; Denny Roberts; Quoc V Zhou; Jason Le; Wei","journal":"","ref_id":"b2","title":"Scaling instruction-finetuned language models","year":"2022"},{"authors":"Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova","journal":"Association for Computational Linguistics","ref_id":"b3","title":"BERT: pre-training of deep bidirectional transformers for language understanding","year":"2019-06-02"},{"authors":"William B Dolan; Chris Brockett","journal":"","ref_id":"b4","title":"Automatically constructing a corpus of sentential paraphrases","year":"2005-10"},{"authors":"Aaron Gokaslan; Vanya Cohen; Ellie Pavlick; Stefanie Tellex","journal":"","ref_id":"b5","title":"Openwebtext corpus","year":"2019"},{"authors":"Geoffrey E Hinton; Oriol Vinyals; Jeffrey Dean","journal":"","ref_id":"b6","title":"Distilling the knowledge in a neural network","year":"2015"},{"authors":"Xiaoqi Jiao; Yichun Yin; Lifeng Shang; Xin Jiang; Xiao Chen; Linlin Li; Fang Wang; Qun Liu","journal":"Association for Computational Linguistics","ref_id":"b7","title":"Tinybert: Distilling BERT for natural language understanding","year":"2020-11"},{"authors":"Hector J Levesque; Ernest Davis; Leora Morgenstern","journal":"","ref_id":"b8","title":"The winograd schema challenge","year":"2012-06-10"},{"authors":"Mike Lewis; Yinhan Liu; Naman Goyal; Marjan Ghazvininejad; Abdelrahman Mohamed; Omer Levy; Veselin Stoyanov; Luke Zettlemoyer","journal":"Association for Computational Linguistics","ref_id":"b9","title":"BART: denoising sequence-to-sequence pretraining for natural language generation, translation, and comprehension","year":"2020-07-05"},{"authors":"Zheng Li; Zijian Wang; Ming Tan; Ramesh Nallapati; Parminder Bhatia; Andrew O Arnold; Bing Xiang; Dan Roth","journal":"Association for Computational Linguistics","ref_id":"b10","title":"DQ-BART: efficient sequenceto-sequence model via joint distillation and quantization","year":"2022-05-22"},{"authors":"Chen Liang; Haoming Jiang; Zheng Li; Xianfeng Tang; Bin Yin; Tuo Zhao","journal":"","ref_id":"b11","title":"Homodistil: Homotopic task-agnostic distillation of pre-trained transformers","year":"2023"},{"authors":"Zhenghao Lin; Yeyun Gong; Xiao Liu; Hang Zhang; Chen Lin; Anlei Dong; Jian Jiao; Jingwen Lu; Daxin Jiang; Rangan Majumder; Nan Duan","journal":"ACM","ref_id":"b12","title":"PROD: progressive distillation for dense retrieval","year":"2023-04-30"},{"authors":"Frederick Liu; Siamak Shakeri; Hongkun Yu; Jing Li","journal":"","ref_id":"b13","title":"Enct5: Fine-tuning T5 encoder for nonautoregressive tasks","year":"2021"},{"authors":"Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov","journal":"","ref_id":"b14","title":"Roberta: A robustly optimized BERT pretraining approach","year":"2019"},{"authors":"Seyed-Iman Mirzadeh; Mehrdad Farajtabar; Ang Li; Nir Levine; Akihiro Matsukawa; Hassan Ghasemzadeh","journal":"AAAI Press","ref_id":"b15","title":"Improved knowledge distillation via teacher assistant","year":"2020-02-07"},{"authors":"Shashi Narayan; Shay B Cohen; Mirella Lapata","journal":"Association for Computational Linguistics","ref_id":"b16","title":"Don't give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization","year":"2018-10-31"},{"authors":"Alec Radford; Jeff Wu; Rewon Child; David Luan; Dario Amodei; Ilya Sutskever","journal":"","ref_id":"b17","title":"Language models are unsupervised multitask learners","year":"2019"},{"authors":"Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu","journal":"J. Mach. Learn. Res","ref_id":"b18","title":"Exploring the limits of transfer learning with a unified text-to-text transformer","year":"2020"},{"authors":"Pranav Rajpurkar; Jian Zhang; Konstantin Lopyrev; Percy Liang","journal":"","ref_id":"b19","title":"Squad: 100, 000+ questions for machine comprehension of text","year":"2016-11-01"},{"authors":"Victor Sanh; Lysandre Debut; Julien Chaumond; Thomas Wolf","journal":"","ref_id":"b20","title":"Distilbert, a distilled version of BERT: smaller, faster, cheaper and lighter","year":"2019"},{"authors":"Abigail See; Peter J Liu; Christopher D Manning","journal":"Association for Computational Linguistics","ref_id":"b21","title":"Get to the point: Summarization with pointergenerator networks","year":"2017-07-30"},{"authors":"Sam Shleifer; Alexander M Rush","journal":"","ref_id":"b22","title":"Pre-trained summarization distillation","year":"2020"},{"authors":"Richard Socher; Alex Perelygin; Jean Wu; Jason Chuang; Christopher D Manning; Andrew Y Ng; Christopher Potts","journal":"","ref_id":"b23","title":"Recursive deep models for semantic compositionality over a sentiment treebank","year":"2013-10"},{"authors":"Siqi Sun; Yu Cheng; Zhe Gan; Jingjing Liu","journal":"Association for Computational Linguistics","ref_id":"b24","title":"Patient knowledge distillation for BERT model compression","year":"2019-11-03"},{"authors":"Chaofan Tao; Lu Hou; Wei Zhang; Lifeng Shang; Xin Jiang; Qun Liu; Ping Luo; Ngai Wong","journal":"Association for Computational Linguistics","ref_id":"b25","title":"Compression of generative pre-trained language models via quantization","year":"2022-05-22"},{"authors":"Yi Tay; Mostafa Dehghani; Q Vinh; Xavier Tran; Dara Garcia; Tal Bahri; Huaixiu Schuster; Neil Steven Zheng; Donald Houlsby; Metzler","journal":"","ref_id":"b26","title":"Unifying language learning paradigms","year":"2022"},{"authors":"Hugo Touvron; Thibaut Lavril; Gautier Izacard; Xavier Martinet; Marie-Anne Lachaux; Timothée Lacroix; Baptiste Rozière; Naman Goyal; Eric Hambro; Faisal Azhar; Aurélien Rodriguez; Armand Joulin; Edouard Grave; Guillaume Lample","journal":"","ref_id":"b27","title":"Llama: Open and efficient foundation language models","year":"2023"},{"authors":"Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Lukasz Kaiser; Illia Polosukhin","journal":"","ref_id":"b28","title":"Attention is all you need","year":"2017-09"},{"authors":"Alex Wang; Amanpreet Singh; Julian Michael; Felix Hill; Omer Levy; Samuel R Bowman","journal":"","ref_id":"b29","title":"GLUE: A multi-task benchmark and analysis platform for natural language understanding","year":"2019-05-06"},{"authors":"Wenhui Wang; Hangbo Bao; Shaohan Huang; Li Dong; Furu Wei","journal":"Association for Computational Linguistics","ref_id":"b30","title":"Minilmv2: Multi-head selfattention relation distillation for compressing pretrained transformers","year":"2021-08-01"},{"authors":"Wenhui Wang; Furu Wei; Li Dong; Hangbo Bao; Nan Yang; Ming Zhou","journal":"","ref_id":"b31","title":"Minilm: Deep selfattention distillation for task-agnostic compression of pre-trained transformers","year":"2020-12-06"},{"authors":"Alex Warstadt; Amanpreet Singh; Samuel R Bowman","journal":"Transactions on Association for Computational Linguistics","ref_id":"b32","title":"Neural network acceptability judgments","year":"2019"},{"authors":"Adina Williams; Nikita Nangia; Samuel R Bowman","journal":"","ref_id":"b33","title":"A broad-coverage challenge corpus for sentence understanding through inference","year":"2018-06-01"},{"authors":"Mengzhou Xia; Zexuan Zhong; Danqi Chen","journal":"Association for Computational Linguistics","ref_id":"b34","title":"Structured pruning learns compact and accurate models","year":"2022-05-22"},{"authors":"Yi Yang; Chen Zhang; Dawei Song","journal":"Association for Computational Linguistics","ref_id":"b35","title":"Sparse teachers can be dense with knowledge","year":"2022-12-07"},{"authors":"Aohan Zeng; Xiao Liu; Zhengxiao Du; Zihan Wang; Hanyu Lai; Ming Ding; Zhuoyi Yang; Yifan Xu; Wendi Zheng; Xiao Xia; Weng Lam Tam; Zixuan Ma; Yufei Xue; Jidong Zhai; Wenguang Chen; Peng Zhang; Yuxiao Dong; Jie Tang","journal":"","ref_id":"b36","title":"GLM-130B: an open bilingual pre-trained model","year":"2022"},{"authors":"Chen Zhang; Yang Yang; Qifan Wang; Jiahao Liu; Jingang Wang; Yunsen Xian; Wei Wu; Dawei Song","journal":"","ref_id":"b37","title":"Minidisc: Minimal distillation schedule for language model compression","year":"2022"},{"authors":"Shengqiang Zhang; Xingxing Zhang; Hangbo Bao; Furu Wei","journal":"Association for Computational Linguistics","ref_id":"b38","title":"Attention temperature matters in abstractive summarization distillation","year":"2022-05-22"},{"authors":"Yanli Zhao; Andrew Gu; Rohan Varma; Liang Luo; Chien-Chin Huang; Min Xu; Less Wright; Hamid Shojanazeri; Myle Ott; Sam Shleifer; Alban Desmaison; Can Balioglu; Bernard Nguyen; Geeta Chauhan; Yuchen Hao; Shen Li","journal":"","ref_id":"b39","title":"Pytorch FSDP: experiences on scaling fully sharded data parallel","year":"2023"}],"string":"[\n {\n \"authors\": \"Luisa Bentivogli; Peter Clark; Ido Dagan; Danilo Giampiccolo\",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"The seventh PASCAL recognizing textual entailment challenge\",\n \"year\": \"2011-11-14\"\n },\n {\n \"authors\": \"M Daniel; Mona T Cer; Eneko Diab; Iñigo Agirre; Lucia Lopez-Gazpio; Specia\",\n \"journal\": \"\",\n \"ref_id\": \"b1\",\n \"title\": \"Semeval-2017 task 1: Semantic textual similarity multilingual and crosslingual focused evaluation\",\n \"year\": \"2017-08-03\"\n },\n {\n \"authors\": \"Chung Hyung Won; Le Hou; Shayne Longpre; Barret Zoph; Yi Tay; William Fedus; Eric Li; Xuezhi Wang; Mostafa Dehghani; Siddhartha Brahma; Albert Webson; Shane Shixiang; Zhuyun Gu; Mirac Dai; Xinyun Suzgun; Aakanksha Chen; Sharan Chowdhery; Gaurav Narang; Adams Mishra; Vincent Y Yu; Yanping Zhao; Andrew M Huang; Hongkun Dai; Slav Yu; Ed H Petrov; Jeff Chi; Jacob Dean; Adam Devlin; Denny Roberts; Quoc V Zhou; Jason Le; Wei\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"Scaling instruction-finetuned language models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b3\",\n \"title\": \"BERT: pre-training of deep bidirectional transformers for language understanding\",\n \"year\": \"2019-06-02\"\n },\n {\n \"authors\": \"William B Dolan; Chris Brockett\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"Automatically constructing a corpus of sentential paraphrases\",\n \"year\": \"2005-10\"\n },\n {\n \"authors\": \"Aaron Gokaslan; Vanya Cohen; Ellie Pavlick; Stefanie Tellex\",\n \"journal\": \"\",\n \"ref_id\": \"b5\",\n \"title\": \"Openwebtext corpus\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Geoffrey E Hinton; Oriol Vinyals; Jeffrey Dean\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"Distilling the knowledge in a neural network\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Xiaoqi Jiao; Yichun Yin; Lifeng Shang; Xin Jiang; Xiao Chen; Linlin Li; Fang Wang; Qun Liu\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b7\",\n \"title\": \"Tinybert: Distilling BERT for natural language understanding\",\n \"year\": \"2020-11\"\n },\n {\n \"authors\": \"Hector J Levesque; Ernest Davis; Leora Morgenstern\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"The winograd schema challenge\",\n \"year\": \"2012-06-10\"\n },\n {\n \"authors\": \"Mike Lewis; Yinhan Liu; Naman Goyal; Marjan Ghazvininejad; Abdelrahman Mohamed; Omer Levy; Veselin Stoyanov; Luke Zettlemoyer\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b9\",\n \"title\": \"BART: denoising sequence-to-sequence pretraining for natural language generation, translation, and comprehension\",\n \"year\": \"2020-07-05\"\n },\n {\n \"authors\": \"Zheng Li; Zijian Wang; Ming Tan; Ramesh Nallapati; Parminder Bhatia; Andrew O Arnold; Bing Xiang; Dan Roth\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b10\",\n \"title\": \"DQ-BART: efficient sequenceto-sequence model via joint distillation and quantization\",\n \"year\": \"2022-05-22\"\n },\n {\n \"authors\": \"Chen Liang; Haoming Jiang; Zheng Li; Xianfeng Tang; Bin Yin; Tuo Zhao\",\n \"journal\": \"\",\n \"ref_id\": \"b11\",\n \"title\": \"Homodistil: Homotopic task-agnostic distillation of pre-trained transformers\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Zhenghao Lin; Yeyun Gong; Xiao Liu; Hang Zhang; Chen Lin; Anlei Dong; Jian Jiao; Jingwen Lu; Daxin Jiang; Rangan Majumder; Nan Duan\",\n \"journal\": \"ACM\",\n \"ref_id\": \"b12\",\n \"title\": \"PROD: progressive distillation for dense retrieval\",\n \"year\": \"2023-04-30\"\n },\n {\n \"authors\": \"Frederick Liu; Siamak Shakeri; Hongkun Yu; Jing Li\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"Enct5: Fine-tuning T5 encoder for nonautoregressive tasks\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"Roberta: A robustly optimized BERT pretraining approach\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Seyed-Iman Mirzadeh; Mehrdad Farajtabar; Ang Li; Nir Levine; Akihiro Matsukawa; Hassan Ghasemzadeh\",\n \"journal\": \"AAAI Press\",\n \"ref_id\": \"b15\",\n \"title\": \"Improved knowledge distillation via teacher assistant\",\n \"year\": \"2020-02-07\"\n },\n {\n \"authors\": \"Shashi Narayan; Shay B Cohen; Mirella Lapata\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b16\",\n \"title\": \"Don't give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization\",\n \"year\": \"2018-10-31\"\n },\n {\n \"authors\": \"Alec Radford; Jeff Wu; Rewon Child; David Luan; Dario Amodei; Ilya Sutskever\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"Language models are unsupervised multitask learners\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu\",\n \"journal\": \"J. Mach. Learn. Res\",\n \"ref_id\": \"b18\",\n \"title\": \"Exploring the limits of transfer learning with a unified text-to-text transformer\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Pranav Rajpurkar; Jian Zhang; Konstantin Lopyrev; Percy Liang\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"Squad: 100, 000+ questions for machine comprehension of text\",\n \"year\": \"2016-11-01\"\n },\n {\n \"authors\": \"Victor Sanh; Lysandre Debut; Julien Chaumond; Thomas Wolf\",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"Distilbert, a distilled version of BERT: smaller, faster, cheaper and lighter\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Abigail See; Peter J Liu; Christopher D Manning\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b21\",\n \"title\": \"Get to the point: Summarization with pointergenerator networks\",\n \"year\": \"2017-07-30\"\n },\n {\n \"authors\": \"Sam Shleifer; Alexander M Rush\",\n \"journal\": \"\",\n \"ref_id\": \"b22\",\n \"title\": \"Pre-trained summarization distillation\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Richard Socher; Alex Perelygin; Jean Wu; Jason Chuang; Christopher D Manning; Andrew Y Ng; Christopher Potts\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"Recursive deep models for semantic compositionality over a sentiment treebank\",\n \"year\": \"2013-10\"\n },\n {\n \"authors\": \"Siqi Sun; Yu Cheng; Zhe Gan; Jingjing Liu\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b24\",\n \"title\": \"Patient knowledge distillation for BERT model compression\",\n \"year\": \"2019-11-03\"\n },\n {\n \"authors\": \"Chaofan Tao; Lu Hou; Wei Zhang; Lifeng Shang; Xin Jiang; Qun Liu; Ping Luo; Ngai Wong\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b25\",\n \"title\": \"Compression of generative pre-trained language models via quantization\",\n \"year\": \"2022-05-22\"\n },\n {\n \"authors\": \"Yi Tay; Mostafa Dehghani; Q Vinh; Xavier Tran; Dara Garcia; Tal Bahri; Huaixiu Schuster; Neil Steven Zheng; Donald Houlsby; Metzler\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"Unifying language learning paradigms\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Hugo Touvron; Thibaut Lavril; Gautier Izacard; Xavier Martinet; Marie-Anne Lachaux; Timothée Lacroix; Baptiste Rozière; Naman Goyal; Eric Hambro; Faisal Azhar; Aurélien Rodriguez; Armand Joulin; Edouard Grave; Guillaume Lample\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"Llama: Open and efficient foundation language models\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Lukasz Kaiser; Illia Polosukhin\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"Attention is all you need\",\n \"year\": \"2017-09\"\n },\n {\n \"authors\": \"Alex Wang; Amanpreet Singh; Julian Michael; Felix Hill; Omer Levy; Samuel R Bowman\",\n \"journal\": \"\",\n \"ref_id\": \"b29\",\n \"title\": \"GLUE: A multi-task benchmark and analysis platform for natural language understanding\",\n \"year\": \"2019-05-06\"\n },\n {\n \"authors\": \"Wenhui Wang; Hangbo Bao; Shaohan Huang; Li Dong; Furu Wei\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b30\",\n \"title\": \"Minilmv2: Multi-head selfattention relation distillation for compressing pretrained transformers\",\n \"year\": \"2021-08-01\"\n },\n {\n \"authors\": \"Wenhui Wang; Furu Wei; Li Dong; Hangbo Bao; Nan Yang; Ming Zhou\",\n \"journal\": \"\",\n \"ref_id\": \"b31\",\n \"title\": \"Minilm: Deep selfattention distillation for task-agnostic compression of pre-trained transformers\",\n \"year\": \"2020-12-06\"\n },\n {\n \"authors\": \"Alex Warstadt; Amanpreet Singh; Samuel R Bowman\",\n \"journal\": \"Transactions on Association for Computational Linguistics\",\n \"ref_id\": \"b32\",\n \"title\": \"Neural network acceptability judgments\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Adina Williams; Nikita Nangia; Samuel R Bowman\",\n \"journal\": \"\",\n \"ref_id\": \"b33\",\n \"title\": \"A broad-coverage challenge corpus for sentence understanding through inference\",\n \"year\": \"2018-06-01\"\n },\n {\n \"authors\": \"Mengzhou Xia; Zexuan Zhong; Danqi Chen\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b34\",\n \"title\": \"Structured pruning learns compact and accurate models\",\n \"year\": \"2022-05-22\"\n },\n {\n \"authors\": \"Yi Yang; Chen Zhang; Dawei Song\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b35\",\n \"title\": \"Sparse teachers can be dense with knowledge\",\n \"year\": \"2022-12-07\"\n },\n {\n \"authors\": \"Aohan Zeng; Xiao Liu; Zhengxiao Du; Zihan Wang; Hanyu Lai; Ming Ding; Zhuoyi Yang; Yifan Xu; Wendi Zheng; Xiao Xia; Weng Lam Tam; Zixuan Ma; Yufei Xue; Jidong Zhai; Wenguang Chen; Peng Zhang; Yuxiao Dong; Jie Tang\",\n \"journal\": \"\",\n \"ref_id\": \"b36\",\n \"title\": \"GLM-130B: an open bilingual pre-trained model\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Chen Zhang; Yang Yang; Qifan Wang; Jiahao Liu; Jingang Wang; Yunsen Xian; Wei Wu; Dawei Song\",\n \"journal\": \"\",\n \"ref_id\": \"b37\",\n \"title\": \"Minidisc: Minimal distillation schedule for language model compression\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Shengqiang Zhang; Xingxing Zhang; Hangbo Bao; Furu Wei\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b38\",\n \"title\": \"Attention temperature matters in abstractive summarization distillation\",\n \"year\": \"2022-05-22\"\n },\n {\n \"authors\": \"Yanli Zhao; Andrew Gu; Rohan Varma; Liang Luo; Chien-Chin Huang; Min Xu; Less Wright; Hamid Shojanazeri; Myle Ott; Sam Shleifer; Alban Desmaison; Can Balioglu; Bernard Nguyen; Geeta Chauhan; Yuchen Hao; Shen Li\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"Pytorch FSDP: experiences on scaling fully sharded data parallel\",\n \"year\": \"2023\"\n }\n]"},"formulas":{"kind":"list like","value":[{"formula_coordinates":[2,321.28,182.32,187.99,147.04],"formula_id":"formula_0","formula_text":"MSA(X; W Q , W K , W V ) = A i SelfAttn(X; W Q i , W K i )XW V i W O i , SelfAttn(X; W Q i , W K i ) = Softmax(XW Q i W K i X /d A ), FFN(X; W I , W O ) = I j g(XW I j )W O j ,"},{"formula_coordinates":[2,305.49,363.85,220.28,30.7],"formula_id":"formula_1","formula_text":"W Q i , W K i , W V i ∈ R d×d A , W O i ∈ R d A ×d among A heads,"},{"formula_coordinates":[2,307.26,515.89,216.04,97.94],"formula_id":"formula_2","formula_text":"MCA(Z, E; W Q , W K , W V ) = A i CrossAttn(Z, E; W Q i , W K i )EW V i W O i , CrossAttn(Z, E; W Q i , W K i ) = Softmax(ZW Q i W K i E /d A ),"},{"formula_coordinates":[2,306.14,630.13,218.27,30.09],"formula_id":"formula_3","formula_text":"W Q i , W K i , W V i ∈ R d×d A , W O i ∈ R d A ×d ."},{"formula_coordinates":[3,80.52,579.07,198.96,146.21],"formula_id":"formula_4","formula_text":"L SelfAttn (S; T , D Z ) = E Z∼D Z R k=1 KL(Reln(Z; T W Q k ), Reln(Z; S W Q k )) + KL(Reln(Z; T W K k ), Reln(Z; S W K k )) + KL(Reln(Z; T W V k ), Reln(Z; S W V k )), Reln(Z; T W Q k ) = Softmax(Z T W Q k T W Q k Z /d R ),"},{"formula_coordinates":[3,312.97,595.11,204.45,90.12],"formula_id":"formula_5","formula_text":"L CrossAttn (S; T , D Z , D E ) = E Z∼D Z ,E∼D E R k=1 KL(Reln(Z; T W Q k ), Reln(Z; S W Q k )) + KL(Reln(E; T W K k ), Reln(E; S W K k )) + KL(Reln(E; T W V k ), Reln(E; S W V k )),"},{"formula_coordinates":[3,326.31,760.99,177.94,13.29],"formula_id":"formula_6","formula_text":"L Logit (S; T , D Z ) + L SelfAttn (S; T , D Z ),"},{"formula_coordinates":[4,92.55,282.11,174.91,31.38],"formula_id":"formula_7","formula_text":"L Logit (S; T , D Z ) + L SelfAttn (S; T , D Z ) + L CrossAttn (S; T , D Z , D E ),"},{"formula_coordinates":[4,72.56,683.83,214.87,31.38],"formula_id":"formula_8","formula_text":"L(S; T , D Z , D E ) = L Logit (S; T , D Z )+ L SelfAttn (S; T , D Z ) + L CrossAttn (S; T , D Z , D E ),"},{"formula_coordinates":[4,118.83,744.31,121.34,30.48],"formula_id":"formula_9","formula_text":"L Logit (S; T , D Z ) = E Z∼D Z CE(Z S W E , Z T W E ),"},{"formula_coordinates":[4,313.6,346.46,203.36,31.38],"formula_id":"formula_10","formula_text":"L(S; T , D Z , D X ) = L Logit (S; T , D Z )+ L SelfAttn (S; T , D Z ) + L EncSelfAttn (S; T , D X ),"}],"string":"[\n {\n \"formula_coordinates\": [\n 2,\n 321.28,\n 182.32,\n 187.99,\n 147.04\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"MSA(X; W Q , W K , W V ) = A i SelfAttn(X; W Q i , W K i )XW V i W O i , SelfAttn(X; W Q i , W K i ) = Softmax(XW Q i W K i X /d A ), FFN(X; W I , W O ) = I j g(XW I j )W O j ,\"\n },\n {\n \"formula_coordinates\": [\n 2,\n 305.49,\n 363.85,\n 220.28,\n 30.7\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"W Q i , W K i , W V i ∈ R d×d A , W O i ∈ R d A ×d among A heads,\"\n },\n {\n \"formula_coordinates\": [\n 2,\n 307.26,\n 515.89,\n 216.04,\n 97.94\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"MCA(Z, E; W Q , W K , W V ) = A i CrossAttn(Z, E; W Q i , W K i )EW V i W O i , CrossAttn(Z, E; W Q i , W K i ) = Softmax(ZW Q i W K i E /d A ),\"\n },\n {\n \"formula_coordinates\": [\n 2,\n 306.14,\n 630.13,\n 218.27,\n 30.09\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"W Q i , W K i , W V i ∈ R d×d A , W O i ∈ R d A ×d .\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 80.52,\n 579.07,\n 198.96,\n 146.21\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"L SelfAttn (S; T , D Z ) = E Z∼D Z R k=1 KL(Reln(Z; T W Q k ), Reln(Z; S W Q k )) + KL(Reln(Z; T W K k ), Reln(Z; S W K k )) + KL(Reln(Z; T W V k ), Reln(Z; S W V k )), Reln(Z; T W Q k ) = Softmax(Z T W Q k T W Q k Z /d R ),\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 312.97,\n 595.11,\n 204.45,\n 90.12\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"L CrossAttn (S; T , D Z , D E ) = E Z∼D Z ,E∼D E R k=1 KL(Reln(Z; T W Q k ), Reln(Z; S W Q k )) + KL(Reln(E; T W K k ), Reln(E; S W K k )) + KL(Reln(E; T W V k ), Reln(E; S W V k )),\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 326.31,\n 760.99,\n 177.94,\n 13.29\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"L Logit (S; T , D Z ) + L SelfAttn (S; T , D Z ),\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 92.55,\n 282.11,\n 174.91,\n 31.38\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"L Logit (S; T , D Z ) + L SelfAttn (S; T , D Z ) + L CrossAttn (S; T , D Z , D E ),\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 72.56,\n 683.83,\n 214.87,\n 31.38\n ],\n \"formula_id\": \"formula_8\",\n \"formula_text\": \"L(S; T , D Z , D E ) = L Logit (S; T , D Z )+ L SelfAttn (S; T , D Z ) + L CrossAttn (S; T , D Z , D E ),\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 118.83,\n 744.31,\n 121.34,\n 30.48\n ],\n \"formula_id\": \"formula_9\",\n \"formula_text\": \"L Logit (S; T , D Z ) = E Z∼D Z CE(Z S W E , Z T W E ),\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 313.6,\n 346.46,\n 203.36,\n 31.38\n ],\n \"formula_id\": \"formula_10\",\n \"formula_text\": \"L(S; T , D Z , D X ) = L Logit (S; T , D Z )+ L SelfAttn (S; T , D Z ) + L EncSelfAttn (S; T , D X ),\"\n }\n]"}}},{"rowIdx":397,"cells":{"doi":{"kind":"string","value":"10.3390/s22239384"},"pub_date":{"kind":"string","value":""},"sections":{"kind":"list like","value":[{"figure_ref":["fig_0"],"heading":"Introduction","publication_ref":["b0","b1","b2","b3","b4","b5","b6","b7","b8","b9","b10","b11","b13","b14","b15","b13","b16","b17","b18","b19","b20","b21","b22","b23","b15","b24","b25","b26","b27","b28","b31","b32","b28","b23","b25","b26","b29","b30","b26","b29","b32","b27","b33","b14","b14","b15","b13","b34","b30","b34","b12"],"table_ref":[],"text":"In recent years, object detection with Unmanned Aerial Vehicles (UAVs) has attracted much attention in computer vision research and has provided many benefits in various domains. Such as fire smoke detection [1], military [2], urban surveillance [3], and agriculture [4], [5]. However, it is not easy to accurately detect objects with UAVs that capture object images using the camera from a very high followed by a widely geographic one. Most of the current traditional object detection methods are based only on the sliding-window paradigm and handcrafted features. Like Viola-Jones [6], Histogram of Oriented Gradients (HOG) [7], Scale-Invariant Feature Transform (SIFT) [8], [9], Haar [10], [11], which has made significant progress in the research field of object detection. However, this method takes time and effort to achieve the robustness of feature representation and is still vulnerable to failure when handling variations in data obtained from the UAVs. What is urgently needed by object detection systems with UAVs today is an accurate method capable of processing image data end-to-end. Currently, deep learning [12] is one of the solutions to answer these needs. The CNN can process visual data accurately without the need to go through a separately feature extraction process and has proven can outperform traditional methods in ImageNet Large Scale Visual Recognition Challenge (ILSVRC) [14]. The progress is inseparable from the availability of large-scale data, such as Microsoft Common Objects in Context (COCO) [15], Pascal Visual Object Classes (PASCAL VOC) [16], ImageNet [14], as well as the availability of computing resources, and driven by ongoing research with the proposed various network architectures. Such as VGG [17], GoogLeNet [18], Residual Networks (ResNets) [19], [20], ResNeXt [21], Cross Stage Partial Network (CSPNet) [22], and EfficientNet [23] in the classification tasks which is widely used as a backbone layer for feature extraction in the object detection tasks. Object detection based on deep learning methods generally divided into two: the one-stage detector and the two-stage detector. The two-stage detector method predicts the bounding box through the process of region proposal and then classifies it to detect the class from the object. Such as the Region-based Convolutional Neural Network (R-CNN) proposed by Ross Girshick et al. [24] is the first deep learning based object detection method. R-CNN In the PASCAL VOC 2010 challenge [16] was able to outperform traditional detector methods, such as Deformable Parts Model (DPM) [25], which at that time occupied the first position. This progress is also driven by the development of other popular methods, such as Fast R-CNN [26] and Faster R-CNN [27], which is average have a high prediction accuracy. However, that method is still relatively slow in the detection process. That is deficiency can overcome by one-stage detector methods, such as RetinaNet [28], You Only Look Once (YOLO) [29]- [32] and Single Shot MultiBox Detector (SSD) [33], which are very fast when predicting objects. Such as the YOLO method proposed by Joseph Redmon et al. [29] can predict multiple bounding boxes and class probabilities simultaneously, which makes it very fast during the detection process. However, YOLO in the first version still has several localization errors compared with the region proposal method [24], [26], [27]. So development was also carried out to reduce the shortcomings of previous versions, such as YOLOv2 [30] and YOLOv3 [31]. YOLOv2 uses darknet-19 as backbone layers that consist of 19 convolutional layers and 5 max-pooling. While YOLOv3 is a further development of YOLOv2, which can predict the bounding boxes with multi-scale prediction and uses darknet-53 in the backbone layer. YOLOv3 can produce a balance between accuracy and detection speed. The result of YOLOv3 can get the average precision better than Faster R-CNN [27], YOLOv2 [30], SSD [33], and faster than RetinaNet [28] and Region-based Fully Convolutional Network (R-FCN) [34] on the testing of the COCO dataset [15]. However, the data obtained from the UAVs is not like that of data from COCO [15], PASCAL VOC [16], and ImageNet [14], that dominated by global image objects with large individual objects. UAVs capture object images from a very high camera and produce data with varying perspectives viewing. As illustrated in Figure 1, the image data captured by UAVs is dominated by small object sizes, which makes the image contain less clear features and the density of objects with different illumination levels is also a challenge for detecting objects with the images obtained by UAVs.\nMotivated by the challenge above, we aim to improve the performance of the YOLOv3 method for detecting objects images obtained from UAVs.\nWe added Spatial pyramid pooling (SPP) [35] at the end of the darknet-53 backbone architecture to achieve a more efficient feature extraction process. The details objective and the contribution of this study are explained as follows:\n1) We improved the performance of YOLOv3 [31] by adding SPP [35] on the end layer of the darknet-53 backbone to obtain more efficient feature extraction process in object detection tasks with UAVs. 2) We also show an evaluation study of different versions YOLOv3 method on object detection tasks with UAVs, including YOLOv3 with SPP, YOLOv3, and YOLOv3-tiny which we analyzed with the VisDrone2019-Det dataset [13]."},{"figure_ref":[],"heading":"Research Method","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"YOLOv3","publication_ref":["b28"],"table_ref":[],"text":"You Only Look Once (YOLO) [29] consists of a backbone layer for feature extraction and a head layer for detection. YOLO predicts objects by mapping the image input pixels to 𝑆𝑥𝑆 grid. Each grid cell predicts B bounding box and confidence score, which is described in the following equation,\n𝑐𝑜𝑛𝑓𝑖𝑑𝑒𝑛𝑐𝑒 = P r (Object)*IoU ( truth predict )(1)\nP r (Object) shows the probability of an object inside the bounding box, and 𝐼𝑜𝑈 𝑝𝑟𝑒𝑑𝑖𝑐𝑡 𝑡𝑟𝑢𝑡ℎ shows the Intersection over Union (IoU) of ground truth and box prediction. The confidence will have a value of 0 if there are no objects in the grid cell and a value of 1 if there are objects. The bounding box consists of 5 parameters (𝑥, 𝑦, 𝑤, ℎ, 𝑐𝑜𝑛𝑓𝑖𝑑𝑒𝑛𝑐𝑒), the width and height are represented by 𝑤, ℎ, and 𝑥, 𝑦 represents the center coordinates of the bounding box. In the end, the results of predicted confidence will represent the Intersections over Union (𝐼𝑜𝑈) between the predicted box and the ground truth boxes. At the same time, each grid cell also predicts C conditional class probabilities that described in the following equation,\n𝐶𝑙𝑎𝑠𝑠 𝑝𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦 = 𝑃𝑟(𝐶𝑙𝑎𝑠𝑠 𝑖 |𝑂𝑏𝑗𝑒𝑐𝑡) (2)\nThe predicting process of conditional class probabilities C in each grid cell is conditioned if there are objects in the grid cell. And the testing process will multiply of conditional class probability with the predicted value of the box confidence to get the confidence score class specific in each box. As represented by equation (3), which encodes the probability of the class appearing in the box and also represents how well the predicted box matches the object."},{"figure_ref":["fig_2"],"heading":"𝑃𝑟(𝐶𝑙𝑎𝑠𝑠","publication_ref":["b28","b29","b29","b18","b35"],"table_ref":[],"text":"𝑖 |𝑂𝑏𝑗𝑒𝑐𝑡) * P r (Object) * IoU ( truth predict ) = 𝑃𝑟(𝐶𝑙𝑎𝑠𝑠 𝑖 ) * IoU ( truth predict ) (3)\nYOLOv3 is an improvement over its predecessors [29], [30], which involves different architecture and is more accurate in the detection process. YOLOv3 uses darknet-53 for the feature extraction process, as represented by Figure 3. Darknet-53 uses 3x3 and 1x1 convolutional layers of darknet-19 in YOLOv2 [30], which is organized by residual networks [19]. YOLOv3 predicts bounding boxes with three different scales using ideas from Feature Pyramid Network (FPN) [36], where the final feature map results from the convolutional layers will predict 3D tensors which are coded as bounding boxes, objectness, and class predictions. Each scales predict 3 squares which are represented as 𝑆𝑥𝑆𝑥(3 * (4 + 1 + 80)), where 𝑆𝑥𝑆 represent the size of the feature map, 4 bounding boxes, 1 objectness prediction, and 80 which is illustrated as the total class prediction. "},{"figure_ref":["fig_3"],"heading":"Spatial Pyramid Pooling","publication_ref":["b36","b37","b34","b34"],"table_ref":[],"text":"Spatial Pyramid Pooling (SPP) [37], [38] in CNN was first introduced by [35]. The process of SPP is represented in Figure 4, which receives the input feature map from the convolutional layers. Afterward, in each spatial bin, the pooling layer responds to each filter to produce output (kMvector). M represents the number of bins, k is the number of filters in the last convolutional layer, and the fixed dimensional vector is the input to the fully connected layers. SPP has some extraordinary properties for deep CNN compared with general networks that use pooling sliding windows. Based on research that conducted by He et al. [35] SPP-net is capable of producing the output of fixed-length regardless of the input size and uses multi-level spatial bins. Meanwhile, pooling sliding windows only use single window sizes. In this study, we aim to add SPP to the final layer of darknet-53 in YOLOv3 to improve performance in object detection tasks with UAVs. "},{"figure_ref":["fig_1"],"heading":"Architecture Model","publication_ref":[],"table_ref":[],"text":"In this study, we aim to add an SPP layer to the final darknet-53 layer to improve the performance of YOLOv3 in object detection tasks with data obtained from UAVs. The details of the architecture in this study is represented in Figure 2. The first process is feature extraction from the input image with darknet-53. Then the SPP layer is added to the final darknet-53 layer to improve the feature extraction process. In the end, the results of SPP is a feature map that uses as input into head detection of YOLOv3 for predicting the bounding boxes and class probabilities."},{"figure_ref":[],"heading":"Loss Function","publication_ref":[],"table_ref":[],"text":"In Los Function is used to determine the state of the training model in each iteration to calculate the difference between the value of predicted and the value of ground truth. As represented in equation ( 4), this study split three loss functions: (𝑙 𝑐𝑜𝑜𝑟𝑑 , 𝑙 𝐼𝑜𝑈 , 𝑙 𝑐𝑙𝑎𝑠𝑠 ). The notation of 𝑙 𝑐𝑜𝑜𝑟𝑑 represents the coordinate prediction errors, 𝑙 𝐼𝑜𝑈 is IoU errors, and 𝑙 𝑐𝑙𝑎𝑠𝑠 is the classification errors."},{"figure_ref":[],"heading":"𝐿𝑜𝑠𝑠 = 𝑙 𝑐𝑜𝑜𝑟𝑑 + 𝑙 𝐼𝑜𝑈 + 𝑙 𝑐𝑙𝑎𝑠𝑠 (4)","publication_ref":[],"table_ref":[],"text":"The coordinate prediction error is represented in the following equation,\n𝑙 𝑐𝑜𝑜𝑟𝑑 = 𝜆 𝑐𝑜𝑜𝑟𝑑 ∑ ∑ 𝛪 𝑖𝑗 𝑜𝑏𝑗 [(𝑥 𝑖 -𝑥 ̂𝑖) 2 + (𝑦 𝑖 - 𝐵 𝑗=0 𝑆 2 𝑖=0 𝑦 ̂𝑖) 2 ] + 𝜆 𝑐𝑜𝑜𝑟𝑑 ∑ ∑ 𝛪 𝑖𝑗 𝑜𝑏𝑗 [(√𝑤 𝑖 -√𝑤 ̂𝑖) 2 + 𝐵 𝑗=0 𝑆 2 𝑖=0 (√ℎ 𝑖 -√ℎ ̂𝑖) 2 ] (5)\nWhere 𝜆 𝑐𝑜𝑜𝑟𝑑 is the weight coordinate error, 𝑆 2 is the number of grid cells for each detection layer, and 𝐵 is the number of bounding boxes in each grid cell. (𝑥 𝑖 , 𝑦 𝑖 , 𝑥 ̂𝑖, 𝑦 ̂𝑖) represents the center coordinate of the ground truth and the target object. Whereas (ℎ 𝑖 , 𝑤 𝑖 , ℎ ̂𝑖, 𝑤 ̂𝑖) represents the width and height of the ground truth and the target prediction box. For IoU errors and Classification errors are denoted by equations ( 6) and ( 7) as follows,\n𝑙 𝐼𝑜𝑈 = 𝜆 𝐼𝑜𝑈 ∑ ∑ 𝛪 𝑖𝑗 𝑜𝑏𝑗 [(𝐶 𝑖 -𝐶 ̂𝑖) 2 ] 𝐵 𝑗=0 𝑆 2 𝑖=0 () + 𝜆 𝑛𝑜𝑜𝑏𝑗 ∑ ∑ 𝛪 𝑖𝑗 𝑛𝑜𝑜𝑏𝑗 (𝐶 𝑖 -𝐶 ̂𝑖) 2 𝐵 𝑗=0 𝑆 2 𝑖=0 (6) 𝑙 𝑐𝑙𝑎𝑠𝑠 = 𝜆 𝑐𝑙𝑎𝑠𝑠 ∑ 𝛪 𝑖 𝑜𝑏𝑗 ∑ (𝑝 𝑖 (𝑐) - 𝑐 ∈ 𝑐𝑙𝑎𝑠𝑠𝑒𝑠 𝑆 2 𝑖=0 𝑝̂𝑖(𝑐)) 2 (7)\nThe IoU error indicates the degree of overlap between the ground truth and the prediction box. If the anchor box indicates that there is a target located in grid cells (𝑖, 𝑗), then the value of Ι 𝑖𝑗 𝑜𝑏𝑗 is 1, and otherwise, the value is 0. The notation of 𝜆 𝑛𝑜𝑜𝑏𝑗 represents a belief penalty if the prediction box contains no objects, and also misclassification, which represents classification accuracy. Where 𝑝 𝑖 (𝑐) is the value of true probability and 𝑝̂𝑖(𝑐) is the predicted value of the target."},{"figure_ref":[],"heading":"Experiment","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":["fig_4"],"heading":"Dataset","publication_ref":["b12"],"table_ref":[],"text":"In this study, we used the VisDrone2019-Det dataset [13], which consisted of 10,209 images, 6,471 for training, 548 for validation, and 3,190 for testing. The VisDrone2019-Det dataset consists of ten object categories: pedestrian, person, bicycle, car, van, truck, tricycle, awningtricycle, bus, and motorcycle. As shown in Figure 5. The VisDrone dataset has several objects with different levels of occlusion in each category, which becomes a challenge to detect objects with UAVs. In this study, we use a training set for the training process and evaluate it with set validation. "},{"figure_ref":[],"heading":"Metric Evaluation","publication_ref":["b10"],"table_ref":[],"text":"To evaluate each method, we used the parameters Precision (𝑃), Recall (𝑅), Average Precision (𝐴𝑃), and mean Average Precision (𝑚𝐴𝑃) with 0.5 Intersections over Union (𝐼𝑜𝑈). The details of 𝑃 and 𝑅 parameters are described by the following equation,\nPrecision (P) = TP TP+FP (8) Recall (R) = TP TP+FN (9)\nWhere 𝑇𝑃 is true positive, that is the correct detection of the ground truth bounding box, 𝐹𝑃 is false positive, that is object was detected but misplaced. 𝐹𝑁 is false negative, which means that the basic ground truth of the bounding box was not detected. 𝐴𝑃 and 𝑚𝐴𝑃 parameters are described by the following equation, (11) Where 𝐴𝑃 is the average value of 𝑃 and 𝑅. 𝑚𝐴𝑃 is the average of the 𝐴𝑃 used to measure all class categories in the dataset and is the metric used to measure the accuracy of object detection with UAV."},{"figure_ref":[],"heading":"Experimental Details","publication_ref":["b14"],"table_ref":[],"text":"For the experimental procedures in this study, we use a pre-trained model from the COCO dataset [15]. In the training phase, we use stochastic gradient descent as optimization with momentum 0.9, batch size 16, learning rate 0.01, and training iterations of 50 epochs with an input scale 640x640. The framework used in this study is PyTorch with a Tesla T4 GPU for the training and validation process. "},{"figure_ref":["fig_6"],"heading":"Results and Discussion","publication_ref":[],"table_ref":["tab_0","tab_0","tab_1","tab_2"],"text":"Table 1, shows the results of training with an input scale of 640x640, which can be concluded in several findings. First, the YOLOv3 with SPP obtained a higher mAP of 0.6% than YOLOv3. These results prove that the addition of the SPP architecture to the YOLOv3 can improve the performance of the object detection model. Second, the YOLOv3-tiny obtained the mAP value of 26.6% much lower than YOLOv3 with SPP and 26% from YOLOv3. These results, one of which is influenced by the depth of the network. YOLOv3-tiny is a lightweight model with fewer parameters and depth. So that able to obtain faster detection processing. However, inversely proportional to the obtained accuracy. The details of the detection results in Table 1, are shown in Table 2. When observed from the total of 10 detection classes, YOLOv3 with SPP excels in six classes: pedestrian, bicycle, van, tricycle, awn, and bus compared to YOLOv3, which only excels in four classes: people, cars, trucks, and motorcycles. Whereas the results of YOLOv3-tiny is lower than YOLOv3 with SPP and YOLOv3 from all detections. For one of the results of visualization from the detection is represented in Figure 6..\nTo obtain a more in-depth analysis, we also validate each model with different input scales. Our goal is to find out if the image scale also affects each object detection model. As reported in Table 3. The YOLOv3 with SPP that we propose is still superior to YOLOv3, with an mAP difference of 0.8% on a 960x960 scale, and 0.9% on a 1280x1280 scale. Whereas YOLOv3-tiny is still lower on both scales compared to the results of YOLOv3 with SPP and YOLOv3."},{"figure_ref":[],"heading":"Conclusion","publication_ref":[],"table_ref":[],"text":"This study aims to improve the performance of YOLOv3 in object detection tasks with UAVs by adding an SPP layer at the end of the darknet-53. We trained three different models: YOLOv3 with SPP, YOLOv3, and YOLOv3-tiny with Visdrone2019-Det training set and evaluated them with a validation set at an input scale of 640x640. The results of YOLOv3 with SPP can improve the performance object detection model, with the results of mAP accuracy of 0.6% more height than YOLOv3 and 26.6% than YOLOv3-tiny. The YOLOv3 with SPP also can maintain accuracy at different input scales, which can outperform the results of YOLOv3 with a difference of 0.8% mAP accuracy on a 960x960 input scale and 0.9% on a 1280x1280 scale. Meanwhile, YOLOv3-tiny is still lower on both scales compared to the results of YOLOv3 with SPP and YOLOv3. The results of YOLOv3 with SPP prove that the addition of the SPP layers to YOLOv3 can improve the performance of object detection models with data obtained from UAVs even with different input scales of image."}],"string":"[\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b0\",\n \"b1\",\n \"b2\",\n \"b3\",\n \"b4\",\n \"b5\",\n \"b6\",\n \"b7\",\n \"b8\",\n \"b9\",\n \"b10\",\n \"b11\",\n \"b13\",\n \"b14\",\n \"b15\",\n \"b13\",\n \"b16\",\n \"b17\",\n \"b18\",\n \"b19\",\n \"b20\",\n \"b21\",\n \"b22\",\n \"b23\",\n \"b15\",\n \"b24\",\n \"b25\",\n \"b26\",\n \"b27\",\n \"b28\",\n \"b31\",\n \"b32\",\n \"b28\",\n \"b23\",\n \"b25\",\n \"b26\",\n \"b29\",\n \"b30\",\n \"b26\",\n \"b29\",\n \"b32\",\n \"b27\",\n \"b33\",\n \"b14\",\n \"b14\",\n \"b15\",\n \"b13\",\n \"b34\",\n \"b30\",\n \"b34\",\n \"b12\"\n ],\n \"table_ref\": [],\n \"text\": \"In recent years, object detection with Unmanned Aerial Vehicles (UAVs) has attracted much attention in computer vision research and has provided many benefits in various domains. Such as fire smoke detection [1], military [2], urban surveillance [3], and agriculture [4], [5]. However, it is not easy to accurately detect objects with UAVs that capture object images using the camera from a very high followed by a widely geographic one. Most of the current traditional object detection methods are based only on the sliding-window paradigm and handcrafted features. Like Viola-Jones [6], Histogram of Oriented Gradients (HOG) [7], Scale-Invariant Feature Transform (SIFT) [8], [9], Haar [10], [11], which has made significant progress in the research field of object detection. However, this method takes time and effort to achieve the robustness of feature representation and is still vulnerable to failure when handling variations in data obtained from the UAVs. What is urgently needed by object detection systems with UAVs today is an accurate method capable of processing image data end-to-end. Currently, deep learning [12] is one of the solutions to answer these needs. The CNN can process visual data accurately without the need to go through a separately feature extraction process and has proven can outperform traditional methods in ImageNet Large Scale Visual Recognition Challenge (ILSVRC) [14]. The progress is inseparable from the availability of large-scale data, such as Microsoft Common Objects in Context (COCO) [15], Pascal Visual Object Classes (PASCAL VOC) [16], ImageNet [14], as well as the availability of computing resources, and driven by ongoing research with the proposed various network architectures. Such as VGG [17], GoogLeNet [18], Residual Networks (ResNets) [19], [20], ResNeXt [21], Cross Stage Partial Network (CSPNet) [22], and EfficientNet [23] in the classification tasks which is widely used as a backbone layer for feature extraction in the object detection tasks. Object detection based on deep learning methods generally divided into two: the one-stage detector and the two-stage detector. The two-stage detector method predicts the bounding box through the process of region proposal and then classifies it to detect the class from the object. Such as the Region-based Convolutional Neural Network (R-CNN) proposed by Ross Girshick et al. [24] is the first deep learning based object detection method. R-CNN In the PASCAL VOC 2010 challenge [16] was able to outperform traditional detector methods, such as Deformable Parts Model (DPM) [25], which at that time occupied the first position. This progress is also driven by the development of other popular methods, such as Fast R-CNN [26] and Faster R-CNN [27], which is average have a high prediction accuracy. However, that method is still relatively slow in the detection process. That is deficiency can overcome by one-stage detector methods, such as RetinaNet [28], You Only Look Once (YOLO) [29]- [32] and Single Shot MultiBox Detector (SSD) [33], which are very fast when predicting objects. Such as the YOLO method proposed by Joseph Redmon et al. [29] can predict multiple bounding boxes and class probabilities simultaneously, which makes it very fast during the detection process. However, YOLO in the first version still has several localization errors compared with the region proposal method [24], [26], [27]. So development was also carried out to reduce the shortcomings of previous versions, such as YOLOv2 [30] and YOLOv3 [31]. YOLOv2 uses darknet-19 as backbone layers that consist of 19 convolutional layers and 5 max-pooling. While YOLOv3 is a further development of YOLOv2, which can predict the bounding boxes with multi-scale prediction and uses darknet-53 in the backbone layer. YOLOv3 can produce a balance between accuracy and detection speed. The result of YOLOv3 can get the average precision better than Faster R-CNN [27], YOLOv2 [30], SSD [33], and faster than RetinaNet [28] and Region-based Fully Convolutional Network (R-FCN) [34] on the testing of the COCO dataset [15]. However, the data obtained from the UAVs is not like that of data from COCO [15], PASCAL VOC [16], and ImageNet [14], that dominated by global image objects with large individual objects. UAVs capture object images from a very high camera and produce data with varying perspectives viewing. As illustrated in Figure 1, the image data captured by UAVs is dominated by small object sizes, which makes the image contain less clear features and the density of objects with different illumination levels is also a challenge for detecting objects with the images obtained by UAVs.\\nMotivated by the challenge above, we aim to improve the performance of the YOLOv3 method for detecting objects images obtained from UAVs.\\nWe added Spatial pyramid pooling (SPP) [35] at the end of the darknet-53 backbone architecture to achieve a more efficient feature extraction process. The details objective and the contribution of this study are explained as follows:\\n1) We improved the performance of YOLOv3 [31] by adding SPP [35] on the end layer of the darknet-53 backbone to obtain more efficient feature extraction process in object detection tasks with UAVs. 2) We also show an evaluation study of different versions YOLOv3 method on object detection tasks with UAVs, including YOLOv3 with SPP, YOLOv3, and YOLOv3-tiny which we analyzed with the VisDrone2019-Det dataset [13].\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Research Method\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"YOLOv3\",\n \"publication_ref\": [\n \"b28\"\n ],\n \"table_ref\": [],\n \"text\": \"You Only Look Once (YOLO) [29] consists of a backbone layer for feature extraction and a head layer for detection. YOLO predicts objects by mapping the image input pixels to 𝑆𝑥𝑆 grid. Each grid cell predicts B bounding box and confidence score, which is described in the following equation,\\n𝑐𝑜𝑛𝑓𝑖𝑑𝑒𝑛𝑐𝑒 = P r (Object)*IoU ( truth predict )(1)\\nP r (Object) shows the probability of an object inside the bounding box, and 𝐼𝑜𝑈 𝑝𝑟𝑒𝑑𝑖𝑐𝑡 𝑡𝑟𝑢𝑡ℎ shows the Intersection over Union (IoU) of ground truth and box prediction. The confidence will have a value of 0 if there are no objects in the grid cell and a value of 1 if there are objects. The bounding box consists of 5 parameters (𝑥, 𝑦, 𝑤, ℎ, 𝑐𝑜𝑛𝑓𝑖𝑑𝑒𝑛𝑐𝑒), the width and height are represented by 𝑤, ℎ, and 𝑥, 𝑦 represents the center coordinates of the bounding box. In the end, the results of predicted confidence will represent the Intersections over Union (𝐼𝑜𝑈) between the predicted box and the ground truth boxes. At the same time, each grid cell also predicts C conditional class probabilities that described in the following equation,\\n𝐶𝑙𝑎𝑠𝑠 𝑝𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦 = 𝑃𝑟(𝐶𝑙𝑎𝑠𝑠 𝑖 |𝑂𝑏𝑗𝑒𝑐𝑡) (2)\\nThe predicting process of conditional class probabilities C in each grid cell is conditioned if there are objects in the grid cell. And the testing process will multiply of conditional class probability with the predicted value of the box confidence to get the confidence score class specific in each box. As represented by equation (3), which encodes the probability of the class appearing in the box and also represents how well the predicted box matches the object.\"\n },\n {\n \"figure_ref\": [\n \"fig_2\"\n ],\n \"heading\": \"𝑃𝑟(𝐶𝑙𝑎𝑠𝑠\",\n \"publication_ref\": [\n \"b28\",\n \"b29\",\n \"b29\",\n \"b18\",\n \"b35\"\n ],\n \"table_ref\": [],\n \"text\": \"𝑖 |𝑂𝑏𝑗𝑒𝑐𝑡) * P r (Object) * IoU ( truth predict ) = 𝑃𝑟(𝐶𝑙𝑎𝑠𝑠 𝑖 ) * IoU ( truth predict ) (3)\\nYOLOv3 is an improvement over its predecessors [29], [30], which involves different architecture and is more accurate in the detection process. YOLOv3 uses darknet-53 for the feature extraction process, as represented by Figure 3. Darknet-53 uses 3x3 and 1x1 convolutional layers of darknet-19 in YOLOv2 [30], which is organized by residual networks [19]. YOLOv3 predicts bounding boxes with three different scales using ideas from Feature Pyramid Network (FPN) [36], where the final feature map results from the convolutional layers will predict 3D tensors which are coded as bounding boxes, objectness, and class predictions. Each scales predict 3 squares which are represented as 𝑆𝑥𝑆𝑥(3 * (4 + 1 + 80)), where 𝑆𝑥𝑆 represent the size of the feature map, 4 bounding boxes, 1 objectness prediction, and 80 which is illustrated as the total class prediction. \"\n },\n {\n \"figure_ref\": [\n \"fig_3\"\n ],\n \"heading\": \"Spatial Pyramid Pooling\",\n \"publication_ref\": [\n \"b36\",\n \"b37\",\n \"b34\",\n \"b34\"\n ],\n \"table_ref\": [],\n \"text\": \"Spatial Pyramid Pooling (SPP) [37], [38] in CNN was first introduced by [35]. The process of SPP is represented in Figure 4, which receives the input feature map from the convolutional layers. Afterward, in each spatial bin, the pooling layer responds to each filter to produce output (kMvector). M represents the number of bins, k is the number of filters in the last convolutional layer, and the fixed dimensional vector is the input to the fully connected layers. SPP has some extraordinary properties for deep CNN compared with general networks that use pooling sliding windows. Based on research that conducted by He et al. [35] SPP-net is capable of producing the output of fixed-length regardless of the input size and uses multi-level spatial bins. Meanwhile, pooling sliding windows only use single window sizes. In this study, we aim to add SPP to the final layer of darknet-53 in YOLOv3 to improve performance in object detection tasks with UAVs. \"\n },\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"Architecture Model\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this study, we aim to add an SPP layer to the final darknet-53 layer to improve the performance of YOLOv3 in object detection tasks with data obtained from UAVs. The details of the architecture in this study is represented in Figure 2. The first process is feature extraction from the input image with darknet-53. Then the SPP layer is added to the final darknet-53 layer to improve the feature extraction process. In the end, the results of SPP is a feature map that uses as input into head detection of YOLOv3 for predicting the bounding boxes and class probabilities.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Loss Function\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In Los Function is used to determine the state of the training model in each iteration to calculate the difference between the value of predicted and the value of ground truth. As represented in equation ( 4), this study split three loss functions: (𝑙 𝑐𝑜𝑜𝑟𝑑 , 𝑙 𝐼𝑜𝑈 , 𝑙 𝑐𝑙𝑎𝑠𝑠 ). The notation of 𝑙 𝑐𝑜𝑜𝑟𝑑 represents the coordinate prediction errors, 𝑙 𝐼𝑜𝑈 is IoU errors, and 𝑙 𝑐𝑙𝑎𝑠𝑠 is the classification errors.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"𝐿𝑜𝑠𝑠 = 𝑙 𝑐𝑜𝑜𝑟𝑑 + 𝑙 𝐼𝑜𝑈 + 𝑙 𝑐𝑙𝑎𝑠𝑠 (4)\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The coordinate prediction error is represented in the following equation,\\n𝑙 𝑐𝑜𝑜𝑟𝑑 = 𝜆 𝑐𝑜𝑜𝑟𝑑 ∑ ∑ 𝛪 𝑖𝑗 𝑜𝑏𝑗 [(𝑥 𝑖 -𝑥 ̂𝑖) 2 + (𝑦 𝑖 - 𝐵 𝑗=0 𝑆 2 𝑖=0 𝑦 ̂𝑖) 2 ] + 𝜆 𝑐𝑜𝑜𝑟𝑑 ∑ ∑ 𝛪 𝑖𝑗 𝑜𝑏𝑗 [(√𝑤 𝑖 -√𝑤 ̂𝑖) 2 + 𝐵 𝑗=0 𝑆 2 𝑖=0 (√ℎ 𝑖 -√ℎ ̂𝑖) 2 ] (5)\\nWhere 𝜆 𝑐𝑜𝑜𝑟𝑑 is the weight coordinate error, 𝑆 2 is the number of grid cells for each detection layer, and 𝐵 is the number of bounding boxes in each grid cell. (𝑥 𝑖 , 𝑦 𝑖 , 𝑥 ̂𝑖, 𝑦 ̂𝑖) represents the center coordinate of the ground truth and the target object. Whereas (ℎ 𝑖 , 𝑤 𝑖 , ℎ ̂𝑖, 𝑤 ̂𝑖) represents the width and height of the ground truth and the target prediction box. For IoU errors and Classification errors are denoted by equations ( 6) and ( 7) as follows,\\n𝑙 𝐼𝑜𝑈 = 𝜆 𝐼𝑜𝑈 ∑ ∑ 𝛪 𝑖𝑗 𝑜𝑏𝑗 [(𝐶 𝑖 -𝐶 ̂𝑖) 2 ] 𝐵 𝑗=0 𝑆 2 𝑖=0 () + 𝜆 𝑛𝑜𝑜𝑏𝑗 ∑ ∑ 𝛪 𝑖𝑗 𝑛𝑜𝑜𝑏𝑗 (𝐶 𝑖 -𝐶 ̂𝑖) 2 𝐵 𝑗=0 𝑆 2 𝑖=0 (6) 𝑙 𝑐𝑙𝑎𝑠𝑠 = 𝜆 𝑐𝑙𝑎𝑠𝑠 ∑ 𝛪 𝑖 𝑜𝑏𝑗 ∑ (𝑝 𝑖 (𝑐) - 𝑐 ∈ 𝑐𝑙𝑎𝑠𝑠𝑒𝑠 𝑆 2 𝑖=0 𝑝̂𝑖(𝑐)) 2 (7)\\nThe IoU error indicates the degree of overlap between the ground truth and the prediction box. If the anchor box indicates that there is a target located in grid cells (𝑖, 𝑗), then the value of Ι 𝑖𝑗 𝑜𝑏𝑗 is 1, and otherwise, the value is 0. The notation of 𝜆 𝑛𝑜𝑜𝑏𝑗 represents a belief penalty if the prediction box contains no objects, and also misclassification, which represents classification accuracy. Where 𝑝 𝑖 (𝑐) is the value of true probability and 𝑝̂𝑖(𝑐) is the predicted value of the target.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiment\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [\n \"fig_4\"\n ],\n \"heading\": \"Dataset\",\n \"publication_ref\": [\n \"b12\"\n ],\n \"table_ref\": [],\n \"text\": \"In this study, we used the VisDrone2019-Det dataset [13], which consisted of 10,209 images, 6,471 for training, 548 for validation, and 3,190 for testing. The VisDrone2019-Det dataset consists of ten object categories: pedestrian, person, bicycle, car, van, truck, tricycle, awningtricycle, bus, and motorcycle. As shown in Figure 5. The VisDrone dataset has several objects with different levels of occlusion in each category, which becomes a challenge to detect objects with UAVs. In this study, we use a training set for the training process and evaluate it with set validation. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Metric Evaluation\",\n \"publication_ref\": [\n \"b10\"\n ],\n \"table_ref\": [],\n \"text\": \"To evaluate each method, we used the parameters Precision (𝑃), Recall (𝑅), Average Precision (𝐴𝑃), and mean Average Precision (𝑚𝐴𝑃) with 0.5 Intersections over Union (𝐼𝑜𝑈). The details of 𝑃 and 𝑅 parameters are described by the following equation,\\nPrecision (P) = TP TP+FP (8) Recall (R) = TP TP+FN (9)\\nWhere 𝑇𝑃 is true positive, that is the correct detection of the ground truth bounding box, 𝐹𝑃 is false positive, that is object was detected but misplaced. 𝐹𝑁 is false negative, which means that the basic ground truth of the bounding box was not detected. 𝐴𝑃 and 𝑚𝐴𝑃 parameters are described by the following equation, (11) Where 𝐴𝑃 is the average value of 𝑃 and 𝑅. 𝑚𝐴𝑃 is the average of the 𝐴𝑃 used to measure all class categories in the dataset and is the metric used to measure the accuracy of object detection with UAV.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experimental Details\",\n \"publication_ref\": [\n \"b14\"\n ],\n \"table_ref\": [],\n \"text\": \"For the experimental procedures in this study, we use a pre-trained model from the COCO dataset [15]. In the training phase, we use stochastic gradient descent as optimization with momentum 0.9, batch size 16, learning rate 0.01, and training iterations of 50 epochs with an input scale 640x640. The framework used in this study is PyTorch with a Tesla T4 GPU for the training and validation process. \"\n },\n {\n \"figure_ref\": [\n \"fig_6\"\n ],\n \"heading\": \"Results and Discussion\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_0\",\n \"tab_0\",\n \"tab_1\",\n \"tab_2\"\n ],\n \"text\": \"Table 1, shows the results of training with an input scale of 640x640, which can be concluded in several findings. First, the YOLOv3 with SPP obtained a higher mAP of 0.6% than YOLOv3. These results prove that the addition of the SPP architecture to the YOLOv3 can improve the performance of the object detection model. Second, the YOLOv3-tiny obtained the mAP value of 26.6% much lower than YOLOv3 with SPP and 26% from YOLOv3. These results, one of which is influenced by the depth of the network. YOLOv3-tiny is a lightweight model with fewer parameters and depth. So that able to obtain faster detection processing. However, inversely proportional to the obtained accuracy. The details of the detection results in Table 1, are shown in Table 2. When observed from the total of 10 detection classes, YOLOv3 with SPP excels in six classes: pedestrian, bicycle, van, tricycle, awn, and bus compared to YOLOv3, which only excels in four classes: people, cars, trucks, and motorcycles. Whereas the results of YOLOv3-tiny is lower than YOLOv3 with SPP and YOLOv3 from all detections. For one of the results of visualization from the detection is represented in Figure 6..\\nTo obtain a more in-depth analysis, we also validate each model with different input scales. Our goal is to find out if the image scale also affects each object detection model. As reported in Table 3. The YOLOv3 with SPP that we propose is still superior to YOLOv3, with an mAP difference of 0.8% on a 960x960 scale, and 0.9% on a 1280x1280 scale. Whereas YOLOv3-tiny is still lower on both scales compared to the results of YOLOv3 with SPP and YOLOv3.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This study aims to improve the performance of YOLOv3 in object detection tasks with UAVs by adding an SPP layer at the end of the darknet-53. We trained three different models: YOLOv3 with SPP, YOLOv3, and YOLOv3-tiny with Visdrone2019-Det training set and evaluated them with a validation set at an input scale of 640x640. The results of YOLOv3 with SPP can improve the performance object detection model, with the results of mAP accuracy of 0.6% more height than YOLOv3 and 26.6% than YOLOv3-tiny. The YOLOv3 with SPP also can maintain accuracy at different input scales, which can outperform the results of YOLOv3 with a difference of 0.8% mAP accuracy on a 960x960 input scale and 0.9% on a 1280x1280 scale. Meanwhile, YOLOv3-tiny is still lower on both scales compared to the results of YOLOv3 with SPP and YOLOv3. The results of YOLOv3 with SPP prove that the addition of the SPP layers to YOLOv3 can improve the performance of object detection models with data obtained from UAVs even with different input scales of image.\"\n }\n]"},"abstract":{"kind":"string","value":"Object detection with Unmanned Aerial Vehicles (UAVs) has attracted much attention in the research field of computer vision. However, not easy to accurately detect objects with data obtained from UAVs, which capture images from very high altitudes, making the image dominated by small object sizes, that difficult to detect. Motivated by that challenge, we aim to improve the performance of the one-stage detector YOLOv3 by adding a Spatial Pyramid Pooling (SPP) layer on the end of the backbone darknet-53 to obtain more efficient feature extraction process in object detection tasks with UAVs. We also conducted an evaluation study on different versions of YOLOv3 methods. Includes YOLOv3 with SPP, YOLOv3, and YOLOv3-tiny, which we analyzed with the VisDrone2019-Det dataset. Here we show that YOLOv3 with SPP can get results mAP 0.6% higher than YOLOv3 and 26.6% than YOLOv3-Tiny at 640x640 input scale and is even able to maintain accuracy at different input image scales than other versions of the YOLOv3 method. Those results prove that the addition of SPP layers to YOLOv3 can be an efficient solution for improving the performance of the object detection method with data obtained from UAVs."},"title":{"kind":"string","value":"YOLOv3 with Spatial Pyramid Pooling for Object Detection with Unmanned Aerial Vehicles"},"figures":{"kind":"list like","value":[{"figure_caption":"Fig. 1 .1Fig. 1. Object detection challenges with UAVs: (a) small objects, (b) object density, and (c) differentilluminations[13] ","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"Fig. 2 .2Fig. 2. Architecture method in this study Currently, deep learning methods have become a focus of research in the field of object detection. In particular, deep learning methods based on Convolutional Neural Networks (CNN).The CNN can process visual data accurately without the need to go through a separately feature extraction process and has proven can outperform traditional methods in ImageNet Large Scale Visual Recognition Challenge (ILSVRC)[14]. The progress is inseparable from the availability of large-scale data, such as Microsoft Common Objects in Context (COCO)[15], Pascal Visual Object Classes (PASCAL VOC)[16], ImageNet[14], as well as the availability of computing resources, and driven by ongoing research with the proposed various network architectures. Such as VGG[17], GoogLeNet[18], Residual Networks (ResNets)[19],[20], ResNeXt[21], Cross Stage Partial Network (CSPNet)[22], and EfficientNet[23] in the classification tasks which is widely used as a backbone layer for feature extraction in the object detection tasks. Object detection based on deep learning methods generally divided into two: the one-stage detector and the two-stage detector. The two-stage detector method predicts the bounding box through the process of region proposal and then classifies it to detect the class from the object. Such as the Region-based Convolutional Neural Network (R-CNN) proposed by Ross Girshick et al.[24] is the first deep learning based object detection method. R-CNN In the PASCAL VOC 2010 challenge[16] was able to outperform traditional detector methods, such as Deformable Parts Model (DPM)[25], which at that time occupied the first position. This progress is also driven by the development of other popular methods, such as Fast R-CNN[26] and Faster R-CNN[27], which is average have a high prediction accuracy. However, that method is still relatively slow in the detection process. That is deficiency can overcome by one-stage detector","figure_data":"","figure_id":"fig_1","figure_label":"2","figure_type":"figure"},{"figure_caption":"Fig. 3 .3Fig. 3. Darknet-53 Architecture","figure_data":"","figure_id":"fig_2","figure_label":"3","figure_type":"figure"},{"figure_caption":"Fig. 4 .4Fig. 4. Spatial Pyramid Pooling architecture","figure_data":"","figure_id":"fig_3","figure_label":"4","figure_type":"figure"},{"figure_caption":"Fig. 5 .5Fig. 5. Visdrone dataset with different levels of occlusion","figure_data":"","figure_id":"fig_4","figure_label":"5","figure_type":"figure"},{"figure_caption":"AP = ∑ (𝑅 𝑛+1 -𝑅 𝑛 ) 𝑛 𝑃(𝑅 ̃) 𝑅 ̃:𝑅 ̃≥𝑅 𝑛+1 𝑚𝑎𝑚𝑎𝑥𝑠𝑎𝑥","figure_data":"","figure_id":"fig_5","figure_label":"","figure_type":"figure"},{"figure_caption":"Fig. 6 .6Fig. 6. Detection visualization.","figure_data":"","figure_id":"fig_6","figure_label":"6","figure_type":"figure"},{"figure_caption":"Training results.","figure_data":"ModelPrecisionRecallmAP_50YOLOv350.140.239.7YOLOv3-tiny22.917.913.7YOLOv3-SPP49.341.440.3","figure_id":"tab_0","figure_label":"1","figure_type":"table"},{"figure_caption":"Detection results.","figure_data":"ModelPedestrianPeopleBicycleCarVanTruckTricycleAwnBusMotorYOLOv349.339.816.378.341.838.325.312.249.845.8YOLOv3-tiny1615.62.446.7110.16.52.810.616.6YOLOv3-SPP49.439.417.878.142.837.926.714.350.645.6","figure_id":"tab_1","figure_label":"2","figure_type":"table"},{"figure_caption":"Validation results with different input scales.","figure_data":"ModelInput scalesPrecisionRecallmAP_50YOLOv3960x96048.740.2381280x128049.439.838.2YOLOv3-tiny960x96025.821.115.61280x128025.621.916.1YOLOv3-SPP960x96048.341.338.81280x128047.242.839.1","figure_id":"tab_2","figure_label":"3","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Fig. 1 .1Fig. 1. Object detection challenges with UAVs: (a) small objects, (b) object density, and (c) differentilluminations[13] \",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 2 .2Fig. 2. Architecture method in this study Currently, deep learning methods have become a focus of research in the field of object detection. In particular, deep learning methods based on Convolutional Neural Networks (CNN).The CNN can process visual data accurately without the need to go through a separately feature extraction process and has proven can outperform traditional methods in ImageNet Large Scale Visual Recognition Challenge (ILSVRC)[14]. The progress is inseparable from the availability of large-scale data, such as Microsoft Common Objects in Context (COCO)[15], Pascal Visual Object Classes (PASCAL VOC)[16], ImageNet[14], as well as the availability of computing resources, and driven by ongoing research with the proposed various network architectures. Such as VGG[17], GoogLeNet[18], Residual Networks (ResNets)[19],[20], ResNeXt[21], Cross Stage Partial Network (CSPNet)[22], and EfficientNet[23] in the classification tasks which is widely used as a backbone layer for feature extraction in the object detection tasks. Object detection based on deep learning methods generally divided into two: the one-stage detector and the two-stage detector. The two-stage detector method predicts the bounding box through the process of region proposal and then classifies it to detect the class from the object. Such as the Region-based Convolutional Neural Network (R-CNN) proposed by Ross Girshick et al.[24] is the first deep learning based object detection method. R-CNN In the PASCAL VOC 2010 challenge[16] was able to outperform traditional detector methods, such as Deformable Parts Model (DPM)[25], which at that time occupied the first position. This progress is also driven by the development of other popular methods, such as Fast R-CNN[26] and Faster R-CNN[27], which is average have a high prediction accuracy. However, that method is still relatively slow in the detection process. That is deficiency can overcome by one-stage detector\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 3 .3Fig. 3. Darknet-53 Architecture\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 4 .4Fig. 4. Spatial Pyramid Pooling architecture\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 5 .5Fig. 5. Visdrone dataset with different levels of occlusion\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"5\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"AP = ∑ (𝑅 𝑛+1 -𝑅 𝑛 ) 𝑛 𝑃(𝑅 ̃) 𝑅 ̃:𝑅 ̃≥𝑅 𝑛+1 𝑚𝑎𝑚𝑎𝑥𝑠𝑎𝑥\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_5\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 6 .6Fig. 6. Detection visualization.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_6\",\n \"figure_label\": \"6\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Training results.\",\n \"figure_data\": \"ModelPrecisionRecallmAP_50YOLOv350.140.239.7YOLOv3-tiny22.917.913.7YOLOv3-SPP49.341.440.3\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Detection results.\",\n \"figure_data\": \"ModelPedestrianPeopleBicycleCarVanTruckTricycleAwnBusMotorYOLOv349.339.816.378.341.838.325.312.249.845.8YOLOv3-tiny1615.62.446.7110.16.52.810.616.6YOLOv3-SPP49.439.417.878.142.837.926.714.350.645.6\",\n \"figure_id\": \"tab_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Validation results with different input scales.\",\n \"figure_data\": \"ModelInput scalesPrecisionRecallmAP_50YOLOv3960x96048.740.2381280x128049.439.838.2YOLOv3-tiny960x96025.821.115.61280x128025.621.916.1YOLOv3-SPP960x96048.341.338.81280x128047.242.839.1\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n }\n]"},"authors":{"kind":"string","value":"Wahyu Pebrianto; Panca Mudjirahardjo; Sholeh Hadi Pramono; Raden Arief Rahmadwati; Setyawan"},"references":{"kind":"list like","value":[{"authors":"M Mukhiddinov; A B Abdusalomov; J Cho","journal":"Sensors","ref_id":"b0","title":"A Wildfire Smoke Detection System Using Unmanned Aerial Vehicle Images Based on the Optimized YOLOv5","year":"2022-12"},{"authors":"P Gupta; B Pareek; G Singal; D V Rao","journal":"Multimed. Tools Appl","ref_id":"b1","title":"Edge device based Military Vehicle Detection and Classification from UAV","year":"2022"},{"authors":"W L Leong; N Martinel; S Huang; C Micheloni; G L Foresti; R S H Teo","journal":"J. Intell. Robot. Syst","ref_id":"b2","title":"An Intelligent Auto-Organizing Aerial Robotic Sensor Network System for Urban Surveillance","year":"2021-06"},{"authors":"U R Mogili; B B V L Deepak","journal":"Procedia Comput. Sci","ref_id":"b3","title":"Review on Application of Drone Systems in Precision Agriculture","year":"2018"},{"authors":"L El Hoummaidi; A Larabi; K Alam","journal":"Heliyon","ref_id":"b4","title":"Using unmanned aerial systems and deep learning for agriculture mapping in Dubai","year":"2021"},{"authors":"P Viola; M Jones","journal":"Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit","ref_id":"b5","title":"Rapid object detection using a boosted cascade of simple features","year":"2001"},{"authors":"N Dalal; B Triggs","journal":"","ref_id":"b6","title":"Histograms of oriented gradients for human detection","year":"2005"},{"authors":"D G Lowe","journal":"","ref_id":"b7","title":"Object recognition from scaleinvariant features","year":"1999"},{"authors":"T Nguyen; E.-A Park; J Han; D.-C Park; S.-Y Min","journal":"","ref_id":"b8","title":"Object Detection Using Scale Invariant Feature Transform","year":"2014"},{"authors":"R Lienhart; J Maydt","journal":"","ref_id":"b9","title":"An extended set of Haar-like features for rapid object detection","year":"2002"},{"authors":"L Arreola; G Gudino; G Flores","journal":"","ref_id":"b10","title":"Object recognition and tracking using Haar-like Features Cascade Classifiers: Application to a quad-rotor UAV","year":"2022-05"},{"authors":"Y Lecun; Y Bengio; G Hinton","journal":"Nature","ref_id":"b11","title":"Deep learning","year":"2015-05"},{"authors":"D Du","journal":"","ref_id":"b12","title":"VisDrone-DET2019: The Vision Meets Drone Object Detection in Image Challenge Results","year":"2019-10"},{"authors":"O Russakovsky","journal":"Int. J. Comput. Vis","ref_id":"b13","title":"ImageNet Large Scale Visual Recognition Challenge","year":"2015-12"},{"authors":"T.-Y Lin","journal":"Eccv","ref_id":"b14","title":"Microsoft COCO: Common Objects in Context","year":"2014-06"},{"authors":"M Everingham; L Van Gool; C K I Williams; J Winn; A Zisserman","journal":"Int. J. Comput. Vis","ref_id":"b15","title":"The Pascal Visual Object Classes (VOC) Challenge","year":"2010-06"},{"authors":"K Simonyan; A Zisserman","journal":"","ref_id":"b16","title":"Very Deep Convolutional Networks for Large-Scale Image Recognition","year":"2014-09"},{"authors":"C Szegedy","journal":"","ref_id":"b17","title":"Going deeper with convolutions","year":"2015-06"},{"authors":"K He; X Zhang; S Ren; J Sun","journal":"","ref_id":"b18","title":"Deep Residual Learning for Image Recognition","year":"2016-06"},{"authors":"K He; X Zhang; S Ren; J Sun","journal":"LNCS","ref_id":"b19","title":"Identity Mappings in Deep Residual Networks","year":"2016"},{"authors":"S Xie; R Girshick; P Dollar; Z Tu; K He","journal":"Janua","ref_id":"b20","title":"Aggregated Residual Transformations for Deep Neural Networks","year":"2017-07"},{"authors":"C.-Y Wang; H.-Y. Mark Liao; Y.-H Wu; P.-Y Chen; J.-W Hsieh; I.-H Yeh","journal":"","ref_id":"b21","title":"CSPNet: A New Backbone that can Enhance Learning Capability of CNN","year":"2020-06"},{"authors":"M Tan; Q V Le","journal":"","ref_id":"b22","title":"EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks","year":"2019-05"},{"authors":"R Girshick; J Donahue; T Darrell; J Malik","journal":"","ref_id":"b23","title":"Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation","year":"2014-06"},{"authors":"P F Felzenszwalb; R B Girshick; D Mcallester; D Ramanan","journal":"IEEE Trans. Pattern Anal. Mach. Intell","ref_id":"b24","title":"Object Detection with Discriminatively Trained Part-Based Models","year":"2010-09"},{"authors":"R Girshick","journal":"","ref_id":"b25","title":"Fast R-CNN","year":"2015-12"},{"authors":"S Ren; K He; R Girshick; J Sun","journal":"IEEE Trans. Pattern Anal. Mach. Intell","ref_id":"b26","title":"Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks","year":"2017-06"},{"authors":"T.-Y Lin; P Goyal; R Girshick; K He; P Dollar","journal":"IEEE Trans. Pattern Anal. Mach. Intell","ref_id":"b27","title":"Focal Loss for Dense Object Detection","year":"2020-02"},{"authors":"J Redmon; S Divvala; R Girshick; A Farhadi","journal":"","ref_id":"b28","title":"You Only Look Once: Unified, Real-Time Object Detection","year":"2016-06"},{"authors":"J Redmon; A Farhadi","journal":"","ref_id":"b29","title":"YOLO9000: Better, Faster, Stronger","year":"2017-07"},{"authors":"J Redmon; A Farhadi","journal":"","ref_id":"b30","title":"YOLOv3: An Incremental Improvement","year":"2018-04"},{"authors":"W Pebrianto; P Mudjirahardjo; S H Pramono","journal":"","ref_id":"b31","title":"YOLO Method Analysis and Comparison for Real-Time Human Face Detection","year":"2022-08"},{"authors":"W Liu","journal":"","ref_id":"b32","title":"SSD: Single Shot MultiBox Detector","year":"2016"},{"authors":"J Dai; Y Li; K He; J Sun","journal":"Adv. Neural Inf. Process. Syst","ref_id":"b33","title":"R-FCN: Object detection via region-based fully convolutional networks","year":"2016"},{"authors":"K He; X Zhang; S Ren; J Sun","journal":"IEEE Trans. Pattern Anal. Mach. Intell","ref_id":"b34","title":"Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition","year":"2015"},{"authors":"T.-Y Lin; P Dollár; R Girshick; K He; B Hariharan; S Belongie","journal":"","ref_id":"b35","title":"Feature Pyramid Networks for Object Detection","year":"2016-12"},{"authors":"K Grauman; T Darrell","journal":"","ref_id":"b36","title":"The pyramid match kernel: Discriminative classification with sets of image features","year":"2005"},{"authors":"S Lazebnik; C Schmid; J Ponce","journal":"Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit","ref_id":"b37","title":"Beyond bags of features: Spatial pyramid matching for recognizing natural scene categories","year":"2006"}],"string":"[\n {\n \"authors\": \"M Mukhiddinov; A B Abdusalomov; J Cho\",\n \"journal\": \"Sensors\",\n \"ref_id\": \"b0\",\n \"title\": \"A Wildfire Smoke Detection System Using Unmanned Aerial Vehicle Images Based on the Optimized YOLOv5\",\n \"year\": \"2022-12\"\n },\n {\n \"authors\": \"P Gupta; B Pareek; G Singal; D V Rao\",\n \"journal\": \"Multimed. Tools Appl\",\n \"ref_id\": \"b1\",\n \"title\": \"Edge device based Military Vehicle Detection and Classification from UAV\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"W L Leong; N Martinel; S Huang; C Micheloni; G L Foresti; R S H Teo\",\n \"journal\": \"J. Intell. Robot. Syst\",\n \"ref_id\": \"b2\",\n \"title\": \"An Intelligent Auto-Organizing Aerial Robotic Sensor Network System for Urban Surveillance\",\n \"year\": \"2021-06\"\n },\n {\n \"authors\": \"U R Mogili; B B V L Deepak\",\n \"journal\": \"Procedia Comput. Sci\",\n \"ref_id\": \"b3\",\n \"title\": \"Review on Application of Drone Systems in Precision Agriculture\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"L El Hoummaidi; A Larabi; K Alam\",\n \"journal\": \"Heliyon\",\n \"ref_id\": \"b4\",\n \"title\": \"Using unmanned aerial systems and deep learning for agriculture mapping in Dubai\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"P Viola; M Jones\",\n \"journal\": \"Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit\",\n \"ref_id\": \"b5\",\n \"title\": \"Rapid object detection using a boosted cascade of simple features\",\n \"year\": \"2001\"\n },\n {\n \"authors\": \"N Dalal; B Triggs\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"Histograms of oriented gradients for human detection\",\n \"year\": \"2005\"\n },\n {\n \"authors\": \"D G Lowe\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"Object recognition from scaleinvariant features\",\n \"year\": \"1999\"\n },\n {\n \"authors\": \"T Nguyen; E.-A Park; J Han; D.-C Park; S.-Y Min\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"Object Detection Using Scale Invariant Feature Transform\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"R Lienhart; J Maydt\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"An extended set of Haar-like features for rapid object detection\",\n \"year\": \"2002\"\n },\n {\n \"authors\": \"L Arreola; G Gudino; G Flores\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"Object recognition and tracking using Haar-like Features Cascade Classifiers: Application to a quad-rotor UAV\",\n \"year\": \"2022-05\"\n },\n {\n \"authors\": \"Y Lecun; Y Bengio; G Hinton\",\n \"journal\": \"Nature\",\n \"ref_id\": \"b11\",\n \"title\": \"Deep learning\",\n \"year\": \"2015-05\"\n },\n {\n \"authors\": \"D Du\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"VisDrone-DET2019: The Vision Meets Drone Object Detection in Image Challenge Results\",\n \"year\": \"2019-10\"\n },\n {\n \"authors\": \"O Russakovsky\",\n \"journal\": \"Int. J. Comput. Vis\",\n \"ref_id\": \"b13\",\n \"title\": \"ImageNet Large Scale Visual Recognition Challenge\",\n \"year\": \"2015-12\"\n },\n {\n \"authors\": \"T.-Y Lin\",\n \"journal\": \"Eccv\",\n \"ref_id\": \"b14\",\n \"title\": \"Microsoft COCO: Common Objects in Context\",\n \"year\": \"2014-06\"\n },\n {\n \"authors\": \"M Everingham; L Van Gool; C K I Williams; J Winn; A Zisserman\",\n \"journal\": \"Int. J. Comput. Vis\",\n \"ref_id\": \"b15\",\n \"title\": \"The Pascal Visual Object Classes (VOC) Challenge\",\n \"year\": \"2010-06\"\n },\n {\n \"authors\": \"K Simonyan; A Zisserman\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"Very Deep Convolutional Networks for Large-Scale Image Recognition\",\n \"year\": \"2014-09\"\n },\n {\n \"authors\": \"C Szegedy\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"Going deeper with convolutions\",\n \"year\": \"2015-06\"\n },\n {\n \"authors\": \"K He; X Zhang; S Ren; J Sun\",\n \"journal\": \"\",\n \"ref_id\": \"b18\",\n \"title\": \"Deep Residual Learning for Image Recognition\",\n \"year\": \"2016-06\"\n },\n {\n \"authors\": \"K He; X Zhang; S Ren; J Sun\",\n \"journal\": \"LNCS\",\n \"ref_id\": \"b19\",\n \"title\": \"Identity Mappings in Deep Residual Networks\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"S Xie; R Girshick; P Dollar; Z Tu; K He\",\n \"journal\": \"Janua\",\n \"ref_id\": \"b20\",\n \"title\": \"Aggregated Residual Transformations for Deep Neural Networks\",\n \"year\": \"2017-07\"\n },\n {\n \"authors\": \"C.-Y Wang; H.-Y. Mark Liao; Y.-H Wu; P.-Y Chen; J.-W Hsieh; I.-H Yeh\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"CSPNet: A New Backbone that can Enhance Learning Capability of CNN\",\n \"year\": \"2020-06\"\n },\n {\n \"authors\": \"M Tan; Q V Le\",\n \"journal\": \"\",\n \"ref_id\": \"b22\",\n \"title\": \"EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks\",\n \"year\": \"2019-05\"\n },\n {\n \"authors\": \"R Girshick; J Donahue; T Darrell; J Malik\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation\",\n \"year\": \"2014-06\"\n },\n {\n \"authors\": \"P F Felzenszwalb; R B Girshick; D Mcallester; D Ramanan\",\n \"journal\": \"IEEE Trans. Pattern Anal. Mach. Intell\",\n \"ref_id\": \"b24\",\n \"title\": \"Object Detection with Discriminatively Trained Part-Based Models\",\n \"year\": \"2010-09\"\n },\n {\n \"authors\": \"R Girshick\",\n \"journal\": \"\",\n \"ref_id\": \"b25\",\n \"title\": \"Fast R-CNN\",\n \"year\": \"2015-12\"\n },\n {\n \"authors\": \"S Ren; K He; R Girshick; J Sun\",\n \"journal\": \"IEEE Trans. Pattern Anal. Mach. Intell\",\n \"ref_id\": \"b26\",\n \"title\": \"Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks\",\n \"year\": \"2017-06\"\n },\n {\n \"authors\": \"T.-Y Lin; P Goyal; R Girshick; K He; P Dollar\",\n \"journal\": \"IEEE Trans. Pattern Anal. Mach. Intell\",\n \"ref_id\": \"b27\",\n \"title\": \"Focal Loss for Dense Object Detection\",\n \"year\": \"2020-02\"\n },\n {\n \"authors\": \"J Redmon; S Divvala; R Girshick; A Farhadi\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"You Only Look Once: Unified, Real-Time Object Detection\",\n \"year\": \"2016-06\"\n },\n {\n \"authors\": \"J Redmon; A Farhadi\",\n \"journal\": \"\",\n \"ref_id\": \"b29\",\n \"title\": \"YOLO9000: Better, Faster, Stronger\",\n \"year\": \"2017-07\"\n },\n {\n \"authors\": \"J Redmon; A Farhadi\",\n \"journal\": \"\",\n \"ref_id\": \"b30\",\n \"title\": \"YOLOv3: An Incremental Improvement\",\n \"year\": \"2018-04\"\n },\n {\n \"authors\": \"W Pebrianto; P Mudjirahardjo; S H Pramono\",\n \"journal\": \"\",\n \"ref_id\": \"b31\",\n \"title\": \"YOLO Method Analysis and Comparison for Real-Time Human Face Detection\",\n \"year\": \"2022-08\"\n },\n {\n \"authors\": \"W Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b32\",\n \"title\": \"SSD: Single Shot MultiBox Detector\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"J Dai; Y Li; K He; J Sun\",\n \"journal\": \"Adv. Neural Inf. Process. Syst\",\n \"ref_id\": \"b33\",\n \"title\": \"R-FCN: Object detection via region-based fully convolutional networks\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"K He; X Zhang; S Ren; J Sun\",\n \"journal\": \"IEEE Trans. Pattern Anal. Mach. Intell\",\n \"ref_id\": \"b34\",\n \"title\": \"Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"T.-Y Lin; P Dollár; R Girshick; K He; B Hariharan; S Belongie\",\n \"journal\": \"\",\n \"ref_id\": \"b35\",\n \"title\": \"Feature Pyramid Networks for Object Detection\",\n \"year\": \"2016-12\"\n },\n {\n \"authors\": \"K Grauman; T Darrell\",\n \"journal\": \"\",\n \"ref_id\": \"b36\",\n \"title\": \"The pyramid match kernel: Discriminative classification with sets of image features\",\n \"year\": \"2005\"\n },\n {\n \"authors\": \"S Lazebnik; C Schmid; J Ponce\",\n \"journal\": \"Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit\",\n \"ref_id\": \"b37\",\n \"title\": \"Beyond bags of features: Spatial pyramid matching for recognizing natural scene categories\",\n \"year\": \"2006\"\n }\n]"},"formulas":{"kind":"list like","value":[{"formula_coordinates":[3,96.08,422.05,191.7,14.79],"formula_id":"formula_0","formula_text":"𝑐𝑜𝑛𝑓𝑖𝑑𝑒𝑛𝑐𝑒 = P r (Object)*IoU ( truth predict )(1)"},{"formula_coordinates":[3,96.08,637.78,191.7,11.39],"formula_id":"formula_1","formula_text":"𝐶𝑙𝑎𝑠𝑠 𝑝𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦 = 𝑃𝑟(𝐶𝑙𝑎𝑠𝑠 𝑖 |𝑂𝑏𝑗𝑒𝑐𝑡) (2)"},{"formula_coordinates":[3,350.68,122.57,159.54,32.81],"formula_id":"formula_2","formula_text":"𝑖 |𝑂𝑏𝑗𝑒𝑐𝑡) * P r (Object) * IoU ( truth predict ) = 𝑃𝑟(𝐶𝑙𝑎𝑠𝑠 𝑖 ) * IoU ( truth predict ) (3)"},{"formula_coordinates":[4,308.56,108.49,201.66,63.49],"formula_id":"formula_3","formula_text":"𝑙 𝑐𝑜𝑜𝑟𝑑 = 𝜆 𝑐𝑜𝑜𝑟𝑑 ∑ ∑ 𝛪 𝑖𝑗 𝑜𝑏𝑗 [(𝑥 𝑖 -𝑥 ̂𝑖) 2 + (𝑦 𝑖 - 𝐵 𝑗=0 𝑆 2 𝑖=0 𝑦 ̂𝑖) 2 ] + 𝜆 𝑐𝑜𝑜𝑟𝑑 ∑ ∑ 𝛪 𝑖𝑗 𝑜𝑏𝑗 [(√𝑤 𝑖 -√𝑤 ̂𝑖) 2 + 𝐵 𝑗=0 𝑆 2 𝑖=0 (√ℎ 𝑖 -√ℎ ̂𝑖) 2 ] (5)"},{"formula_coordinates":[4,330.16,306.93,180.06,81.11],"formula_id":"formula_4","formula_text":"𝑙 𝐼𝑜𝑈 = 𝜆 𝐼𝑜𝑈 ∑ ∑ 𝛪 𝑖𝑗 𝑜𝑏𝑗 [(𝐶 𝑖 -𝐶 ̂𝑖) 2 ] 𝐵 𝑗=0 𝑆 2 𝑖=0 () + 𝜆 𝑛𝑜𝑜𝑏𝑗 ∑ ∑ 𝛪 𝑖𝑗 𝑛𝑜𝑜𝑏𝑗 (𝐶 𝑖 -𝐶 ̂𝑖) 2 𝐵 𝑗=0 𝑆 2 𝑖=0 (6) 𝑙 𝑐𝑙𝑎𝑠𝑠 = 𝜆 𝑐𝑙𝑎𝑠𝑠 ∑ 𝛪 𝑖 𝑜𝑏𝑗 ∑ (𝑝 𝑖 (𝑐) - 𝑐 ∈ 𝑐𝑙𝑎𝑠𝑠𝑒𝑠 𝑆 2 𝑖=0 𝑝̂𝑖(𝑐)) 2 (7)"},{"formula_coordinates":[5,127.7,582.04,160.11,47.06],"formula_id":"formula_5","formula_text":"Precision (P) = TP TP+FP (8) Recall (R) = TP TP+FN (9)"}],"string":"[\n {\n \"formula_coordinates\": [\n 3,\n 96.08,\n 422.05,\n 191.7,\n 14.79\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"𝑐𝑜𝑛𝑓𝑖𝑑𝑒𝑛𝑐𝑒 = P r (Object)*IoU ( truth predict )(1)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 96.08,\n 637.78,\n 191.7,\n 11.39\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"𝐶𝑙𝑎𝑠𝑠 𝑝𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦 = 𝑃𝑟(𝐶𝑙𝑎𝑠𝑠 𝑖 |𝑂𝑏𝑗𝑒𝑐𝑡) (2)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 350.68,\n 122.57,\n 159.54,\n 32.81\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"𝑖 |𝑂𝑏𝑗𝑒𝑐𝑡) * P r (Object) * IoU ( truth predict ) = 𝑃𝑟(𝐶𝑙𝑎𝑠𝑠 𝑖 ) * IoU ( truth predict ) (3)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 308.56,\n 108.49,\n 201.66,\n 63.49\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"𝑙 𝑐𝑜𝑜𝑟𝑑 = 𝜆 𝑐𝑜𝑜𝑟𝑑 ∑ ∑ 𝛪 𝑖𝑗 𝑜𝑏𝑗 [(𝑥 𝑖 -𝑥 ̂𝑖) 2 + (𝑦 𝑖 - 𝐵 𝑗=0 𝑆 2 𝑖=0 𝑦 ̂𝑖) 2 ] + 𝜆 𝑐𝑜𝑜𝑟𝑑 ∑ ∑ 𝛪 𝑖𝑗 𝑜𝑏𝑗 [(√𝑤 𝑖 -√𝑤 ̂𝑖) 2 + 𝐵 𝑗=0 𝑆 2 𝑖=0 (√ℎ 𝑖 -√ℎ ̂𝑖) 2 ] (5)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 330.16,\n 306.93,\n 180.06,\n 81.11\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"𝑙 𝐼𝑜𝑈 = 𝜆 𝐼𝑜𝑈 ∑ ∑ 𝛪 𝑖𝑗 𝑜𝑏𝑗 [(𝐶 𝑖 -𝐶 ̂𝑖) 2 ] 𝐵 𝑗=0 𝑆 2 𝑖=0 () + 𝜆 𝑛𝑜𝑜𝑏𝑗 ∑ ∑ 𝛪 𝑖𝑗 𝑛𝑜𝑜𝑏𝑗 (𝐶 𝑖 -𝐶 ̂𝑖) 2 𝐵 𝑗=0 𝑆 2 𝑖=0 (6) 𝑙 𝑐𝑙𝑎𝑠𝑠 = 𝜆 𝑐𝑙𝑎𝑠𝑠 ∑ 𝛪 𝑖 𝑜𝑏𝑗 ∑ (𝑝 𝑖 (𝑐) - 𝑐 ∈ 𝑐𝑙𝑎𝑠𝑠𝑒𝑠 𝑆 2 𝑖=0 𝑝̂𝑖(𝑐)) 2 (7)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 127.7,\n 582.04,\n 160.11,\n 47.06\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"Precision (P) = TP TP+FP (8) Recall (R) = TP TP+FN (9)\"\n }\n]"}}},{"rowIdx":398,"cells":{"doi":{"kind":"string","value":""},"pub_date":{"kind":"string","value":"2023-05-21"},"sections":{"kind":"list like","value":[{"figure_ref":["fig_0","fig_1","fig_1","fig_2"],"heading":"Introduction","publication_ref":["b6","b15","b33","b24","b28","b4","b27","b24","b3","b26","b22","b28","b28","b26"],"table_ref":[],"text":"Transformers, which have gained far-flung fame in natural language processing (NLP) area [8,28], are also attracting increasing attention in lots of computer vision (CV) tasks, such as object detection [4], image classification [9] and many others [13,31], impelling the widespread research on vision transformers (ViTs). There has a natural fit for ViTs to achieve better performance simply by training a larger model on a larger data set. For example, historical records show better performance of a ViT-H model [9] accompanying with astonishing 632M parameters and 162G FLOPs. Such a high model complexity poses a great challenge to deploy models on platforms with short resource supplies. Therefore, both academia and industry call for an ultimate compression of these large models, and the DeiT-Small on ImageNet [17], respectively. Here 8-bit DeiT is quantized with PTQ method [22] and 2/3/4 bit DeiT is trained with QAT method [18]. The binarized ViT is conducted with the baseline method Bi-Real Net [26].\npast years have witnessed some promising techniques such as network pruning [38,5], low-rank decomposition [7], knowledge distillation [12], and quantization [18,19].\nNetwork quantization, which represents weights and activations in a low-bit format, has got great earnestness of many researchers for its reduced memory access costs and increased compute efficiency as well as performance benefit. Using the lower-bit quantized data, in particular to the extreme 1-bit case, requires less data movement, both on-chip and off-chip, and therefore reduces memory bandwidth and saves significant energy. Existing documentary records observe 32× less network size and 58× speedups beneficial from xnor and bit-count logics for 1-bit networks [30]. Earlier attempts [25,22] apply post-training quantization (PTQ) [1,40] directly to ViTs without datadriven fine-tuning, causing sub-optimal performance, in particular to impotent 1-bit ViTs. Therefore, by quantizing while training, quantization-aware training (QAT) methods are more congenial to 1-bit ViTs. Extensive empirical studies [24,20,36,28] have well demonstrated the efficacy of QAT methods in 1-bit convolutional neural networks (CNNs) or BERTs, however, the application to 1-bit ViTs remains not to be fully explored so far.\nIn this paper, we first build a fully-binarized ViT baseline, a straightforward solution constructed upon popular binarized QAT method of Bi-Real Net [26]. Through an empirical study of this baseline, we observe significant performance drops on the ImageNet dataset [17], as shown in Fig. 1. For instance, extending Bi-Real Net to binarize DeiT-Tiny [32] incurs a tremendous performance gap of 52.6% in the Top-1 accuracy compared to the 2-bit quantized counterpart. Similar performance drops occur in DeiT-Small as well. Delving into a deeper analysis, we find that the incompatibility of existing QAT methods mainly stems from the binarized self-attention module in ViTs, where a simple application of existing binarization methods [26] leads to severe attention distortion, as plotted in Fig. 2 (a) and Fig. 2 (b), especially in the diagonal scores of the map which are supposed to be the most attentive.\nIn this paper we dig deeper into this attention distortion problem. Through empirical analysis, we find that this phenomenon is mainly caused by gradient vanishing due to the straight-through-estimator (STE) [2] and non-scaled binarization in self-attention. Meanwhile, a simple distillation utilizing distillation token in DeiT [32] and KL-divergence in ReActNet [24] is ineffective in dismissing the ranking disorder, since it neglects the relative order of the attention map between the binarized ViTs and their real-valued counterpart. To address the aforementioned issues, a fullybinarized ViT (Bi-ViT) is developed by reactivating the vanished gradients through a learnable scaling factor in selfattention and a ranking-aware distillation to further effectively rectify the disordered ranking of attention (see the overview in Fig. 3). In addition, we also provide both empirical and theoretical analysis about how our method can rectify the distorted attention and thus promote the optimization of Bi-ViT. The contributions of our work are summarized as:\n• We identify the bottleneck of a fully-binarized ViT through empirical analyses and formulate the problem in a theoretical perspective. Based on these, we introduce learnable head-wise scaling factor into binarized self-attention to reactivate the vanished gradients.\n• We develop a ranking-aware distillation scheme to eliminate attention distortion. Our distillation method fully utilizes the ranking-aware knowledge from the real-valued teacher to promote the optimization of Bi-ViT.\n• Our Bi-ViT is the first promising way to push the limit of ViT quantization to the fully-binarized version. Extensive experiments on the ImageNet benchmark demonstrate that Bi-ViT surpasses both the baseline and prior binarized methods by a significant margin, achieving a remarkable acceleration rate of up to 61.5×. "},{"figure_ref":[],"heading":"Related Work","publication_ref":["b37","b31","b39","b8","b28","b26"],"table_ref":[],"text":"Vision Transformer. Unlike traditional CNN-based models, ViTs are capable of capturing long-range visual relationships through the self-attention mechanism, and offer a more generalizable paradigm without inductive bias specific to images. The starting ViT [9] views an image as a sequence of 16 × 16 patches and uses a unique class token to predict the classification, yielding promising results. Subsequently, many works, such as DeiT [32] and PVT [35], have improved upon ViT, making it more efficient and applicable to downstream tasks. However, these high-performing ViTs have also accompanied with a significant number of parameters and high computational overhead, limiting their widespread applications. Thus, designing smaller and faster ViTs has become a new trend. LeViT [11] makes progress in faster inference through down-sampling, patch descriptors, and a redesign of the Attention-MLP block.\nDynamicViT [29] proposes a dynamic token sparsification framework to progressively and dynamically prune redundant tokens, achieving a competitive complexity and accuracy trade-off. Evo-ViT [37] proposes a slow-fast updating mechanism that ensures information flow and spatial structure, reducing both the training and inference complexity.\nWhile the aforementioned works focus on efficient model design, this paper aims to boost compression and acceleration through binarization. Network Binarization. BinaryNet is a technique originally proposed to train convolutional neural networks (CNNs) with binary weights. BinaryConnect [6] is the precursor to BinaryNet, where the parameters are binary while the activations remain in full-precision states. Local binary convolution layers (LBC) [16] were introduced to binarize the non-linear activations, and XNOR-Net [30] was introduced to improve convolution efficiency by binarizing the weights and inputs of convolution kernels. Bi-Real Net [26] explores a new variant of residual structure to preserve the information of real activations before the sign function, with a tight approximation to the derivative of the nondifferentiable sign function. Real-to-binary [27] re-scales the feature maps on the channels according to the input before binarized operations and adds an SE-Net [15] like gating module. ReActNet [24] replaces the conventional PReLU and the sign function of the BNNs with RPReLU and RSign with a learnable threshold, thus improving the performance of BNNs. RBONN [36] introduces a recurrent bilinear optimization to address the asynchronous convergence problem for BNNs, which further improves the performance of BNNs. These techniques improve the efficiency and accuracy of binary neural networks (BNNs) and allow them to be applied in practical applications. Majorities of these techniques consider non-scaled binarization in activations, which is beneficial to conventional CNNs while causing gradient mismatch issue for the pecularity of selfattention mechanism in ViTs."},{"figure_ref":[],"heading":"Background","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Multi-Head Self-Attention and Binarization","publication_ref":["b28"],"table_ref":[],"text":"For a multi-head self-attention (MHSA) module, we denote its query, key, and value set as {a {q,k,v} ∈ R h×N ×d }, where h denotes head number, N and d represent the patch and channel numbers of each head. Specifically, N = (W in //W P in ) × (H in //H P in ) where W in and H in are the width and height of the feature, W P in , H P in are the width and height of patch maps respectively. Then, the attention score A and MHSA module output a out are computed as follows [34]:\nA = softmax[(a q • a k )/ √ d], a out = A • a v ,(1)\nwhere softmax(•) represents the softmax operation. Intending to represent query, key, value and attention score, i.e., a q , a k , a v and A, in a 1-bit format, Eq. (1) changes into:\nA = softmax[(b aq • b a k )/ √ d], a out = b A • b av .(2)\nWe follow the common network binarization methods [30] that use the sign function b • = sign(•) in the binary forward pass, and STE\n[2] ∂b• ∂• = 1 |•|≤1\nto compute the gradient for sign function in its backward pass. We omit the non-linear function here for simplicity. For all the projection and linear layers in binarized ViTs, we conduct binarization following [28,26] as\na out = b ain • (α w • b w ) = α w •(b ain •b w ) where α w = {α 1 w , α 2 w , ..., α Cout w } ∈ R Cout +\nis known as the channel-wise scaling factor vector [30] and • represents channel-wise channel-wise multiplication. The matrix multiplication process, i.e., b ain • b w , can be executed by the efficient XNOR and Bit-count instructions on edge devices."},{"figure_ref":[],"heading":"Bottleneck of Fully-Binarized ViTs","publication_ref":[],"table_ref":[],"text":"The high-performing ViTs are built on premise of transformer's supreme ability to model the long-range relationships thanks to the attention mechanism within the MHSA module. Unfortunately, a binarized version of weights and inputs significantly weakens the representation ability. In Module Degradation. By gradually replacing the multilayer perceptron (MLP) and MHSA modules with realvalued weights or activations, we have discovered that maintaining the MLP as \"w1a1\" (all weights and activations in the MLP are binarized) still results in satisfactory performance. For instance, keeping MLP as \"w1a1\" while keeping MHSA as \"w1a32\" obtains 26.3% Top-1 accuracy, which might be acceptable comparing to the 55.2% of realvalued DeiT-Tiny when taking into consideration 47.3× acceleration rates. On the contrast, when maintaining MHSA module as \"w1a1\", we observe a significant drop in performance. To be more specific, even when the MLP was maintained as \"w32a32\", we still observe a significant 50.8% decrease in Top-1 accuracy (from 55.2% to 4.4%). This result indicates that using binarized weights and activations in the MHSA module can have a substantial negative impact on the model's performance, even when other parts retain in real-valued states.\nOperation Degradation. To better understand the impact of fully-binarized ViT's performance, we conduct further analyses by examining the operations within the MHSA module. Specifically, when we maintain the self-attention activations in Eq. ( 1) as real-valued (\"a32\"), we observe only a relatively small decrease in performance from 48.8% to 37.6%. However, when the self-attention activations in Eq. (2) are binarized, significant drops in accuracy occur from 48.8% to 7.6%. This finding highlights the importance of the self-attention process within the MHSA module and suggests more efforts to mitigate the negative impact of binarization on the MHSA module."},{"figure_ref":["fig_5"],"heading":"Gradient Mismatch in Self-Attention","publication_ref":[],"table_ref":[],"text":"With conclusion from the experimental results in Sec. 3.2 that self-attention process, i.e., Eq. (2), is the most critical part causing the performance drops. We attempt to analyze the underlying reasons for this phenomenon from an optimization perspective. For simplicity, we derive the gradient mismatch in a q as an example, and the analysis can be applicable to explain a k as well. We first represent the features before softmax(•) in Eq. (2) as:\np = (b aq • b a k .)/ √ d.(3)\nThe gradient of a hi,n,c q w.r.t. A is formulated as:\n∂A ∂a h i ,n,c q = ∂A ∂p h i ,n,n • ∂p h i ,n,n ∂b h i ,n,c aq • ∂b h i ,n,c aq ∂a h i ,n,c q , (4\n)\nwhere\nh i ∈ R h , n & n ∈ R N , c ∈ R d\nand the gradient of a k is likewise. The explicit form of the first item ∂A ∂p h i ,n,n in Eq. ( 4) is:\n∂A ∂p hi,n,n = ∂ softmax(p hi,n,n ) ∂p hi,n,n = A hi,n,n ⊗ (1 -A hi,n,n ),(5)\nwhere ⊗ denotes Hadamard product. And the second item is formulated as:\n∂p h i ,n,n ∂b h i ,n,c aq = ∂b h i ,n,c aq • b h i ,c,n a k ∂b h i ,n,c aq = b h i ,c,n a k ,(6)\nresult of which is therefore correlated with b a k . The third item is solved through STE [2] as:\n∂b hi,n,c aq ∂a hi,n,c q = 1 |a h i ,n,c q |≤1 .(7)\nCombing Eq. (5)-Eq. ( 7), we have the final gradient form in fully-binarized ViTs as:\n∂A ∂a h i ,n,c q = ∂A ∂p h i ,n,n • ∂p h i ,n,n ∂b h i ,n,c aq • ∂b h i ,n,c aq ∂a h i ,n,c q = A h i ,n,n (1 -A h i ,n,n ) • b h i ,c,n a k • 1 |a h i ,n,c q |≤1 .(8)\nConsidering b hi,n,:\naq = [1, • • • , 1] and •b hi,n ,: a k = [1, • • • , 1] as the extreme condition, b hi,n,: aq • b hi,:,n a k = d. Therefore, a specific element in b aq •b a k is ∈ {-d, • • • , d}.\nWe plot the curve of a specific element in the first item between [-64, 64] in Fig. 5 = 0, likewise for a k . Therefore we formulate the gradient mismatch phenomenon in the aforementioned theoretical analysis. And such gradient mismatch leads to distorted gradient in the optimization of a q & a k and therefore degrades performance of fully-binarized ViTs."},{"figure_ref":[],"heading":"Our Bi-ViT","publication_ref":[],"table_ref":[],"text":"In this section, we propose to dismiss the affect of gradient mismatch mentioned in Sec. 3.3 from perspectives of gradient approximation in Sec. 4.1 and intermediate distillation in Sec. 4.2."},{"figure_ref":["fig_5","fig_5"],"heading":"Learnable Head-wise Scaling Factor","publication_ref":[],"table_ref":[],"text":"As one of the solution to the above mentioned problem, we propose a head-wise scaling factor binarization scheme for the self-attention process, where the scaling factors are learned during training to first modify the gradient clip range in Fig. 5(b). Eq. ( 2) is changed into:\n= softmax(p), p = (α q ⊗ α k ) • (b aq • b a k )/ √ d = α q;k • (b aq • b a k )/ √ d,(9)\nand\nãout = (αA • bA) • (αv • ba v ) = (αA ⊗ αv) • (bA • b av ) = αA;v • (bA • b av ),(10)\nwhere b a• = sign( a• α• ), α q , α k , α v and α A are the headwise learnable scaling factors in binarized MHSA, where\nα {q,k,v,A} = {α 1 {q,k,v,A} , α 2 {q,k,v,A} , • • • , α h {q,k,v,A} } ∈ R h + .\nThe second rows in Eq. ( 9) & Eq. ( 10) are established since the scaling factors are aligned with the head dimension, which is independent with the matrix multiplication operation. Thus,\nα q;k = {α 1 q;k , α 2 q;k , • • • , α h q;k } ∈ R h + and α A;v = {α 1 A;v , α 2 A;v , • • • , α h A;v } ∈ R h + .\nConsequently, the gradient ∂ à ∂a :,n,c q in Eq. ( 8) is further formulated in our Bi-ViT as:\n∂ à ∂a h i ,n,c q = Ãh i ,n,n (1 -Ãh i ,n,n ) ∂ à ∂p h i ,n,n • α h i q;k • b h i ,c,n a k ∂p h i ,n,n ∂b h i ,n,c aq • 1 |a h i ,n,c q |≤αq ∂b h i ,n,c aq ∂a h i ,n,c q . (11\n)\nSince softmax(.) and • are aligned with different dimensions, the value of Eq. ( 5) remains unchanged (softmax(p) = softmax(α q;k • p)). As can be seen, the threshold of gradient clip in Eq. ( 7) changes from 1 into α q , which means that we can surpass the occurance of gradient mismatch by modifying the value of α q . Note that the scaling factor (α q ) is to imitate the magnitude of the latent activations. When p has a large magnitude, i.e., in the circled part of Fig. 5 (a), α q also tends to be larger and a hi,n,c q locates in the field that ∂b h i ,n,c aq ∂a h i ,n,c q > 0. Thus the vanishing gradients are reactivated through the introduced learnable scaling factor."},{"figure_ref":["fig_1","fig_2"],"heading":"Ranking-aware Distillation for Bi-ViT","publication_ref":[],"table_ref":[],"text":"Fig. 2 illustrates a significant difference in the attention map's relative order between Bi-RealNet (a) and its realvalued counterpart (c). This difference could result in a notable decrease in performance. To address this issue during binarized training, a ranking-aware distillation in a teacherstudent framework is introduced:\nL ranking = L l=1 ψ(A T ) -ψ(A S ) 2 ,(12)\nwhere A T and A S represents the attention scores from the real-valued teacher and binarized student. ψ(•) denotes the function for obtaining the ranking, i.e., relative order of an attention score, which is formulated as:\nψ(A :,n,: ) = A :,n,: -A :,n-1,: , if 0 < n ≤ N -1 A :,0,: -A :,N -1,: , otherwise .(13)\nDetailed relative order computation can be seen in the right part of Fig. 3. We implement our Bi-ViT under the teacherstudent framework [32], thus the final objective of our method is formulated as:\nL = L dist + λL ranking ,(14)\nwhere λ is a hyper-parameter to balance these two loss functions."},{"figure_ref":[],"heading":"Experiments","publication_ref":["b16","b5","b28"],"table_ref":[],"text":"In this section, we evaluate the performance of the proposed Bi-ViT model for image classification task using pop-ular DeiT [32] & Swin [23] backbones and object detection task using Mask R-CNN [14] & Cascade [3] Mask R-CNN with Swin-Tiny [23] backbone. To the best of our knowledge, there is no publicly available source codebase on fully-binarized ViTs at this point, so we implement the baseline i.e., Bi-Real Net [26] methods by ourselves."},{"figure_ref":[],"heading":"Datasets and Implementation Details","publication_ref":["b23","b23","b35","b12"],"table_ref":[],"text":"Datasets. The experiments are carried out on the ImageNet ILSVRC12 dataset [17] for image classification task and COCO dataset [21] for object detection task. The ImageNet dataset is challenging due to its large scale and greater diversity. There are 1000 classes and 1.2 million training images, and 50k validation images in it. In our experiments, we use the classic data augmentation method described in [32].\nThe COCO [21] dataset includes images from 80 different categories. All of our COCO dataset experiments are performed on the object detection track of the COCO trainval35k training dataset, which consists of 80k images from the COCO train2014 dataset and 35k images sampled from the COCO val2014 dataset. We report the average precision (AP) for IoUs∈ [0.5:0.05:0.95], designated as mAP@[.5,.95], using COCO's standard evaluation metric. For further analyzing our method, we also report AP 50 , AP 75 , AP s , AP m , and AP l . Experimental settings. In our experiments, we initialize the weights of binarized model with the corresponding pretrained real-valued model. The binarized model is trained for 300 epochs with batch-size 512 and the base learning rate 5e -4. We do not use warm-up scheme. For all the experiments, we apply LAMB [39] optimizer with weight decay set as 0, following DeiT III [33]. Note that we keep the patch embedding (first) layer and the classification (last) layer as real-valued, following [10]. Backbone. We evaluate our binarized method on two popular vision transformer networks: DeiT [32] and Swin Transformer [23]. The DeiT-Tiny, DeiT-Small, DeiT-Base, Swin-Tiny and Swin-Small are adopted as the backbone models, whose Top-1 accuracy on ImageNet dataset are 72.2%, 79.9%, 81.8%, 81.2%, and 83.2% respectively. For a fair comparison, we utilize the official implementation of DeiT and Swin Transformer."},{"figure_ref":["fig_6"],"heading":"Ablation Study","publication_ref":["b28"],"table_ref":[],"text":"Hyper-parameter Selection. We λ of Eq. ( 14) in this part, with experiments conducted on ImageNet [17] dataset. We show the model performance (Top-1 accuracy) with different setups of hyper-parameter λ in Fig. 6, in which the performances increase first and then decrease with the uplift of λ from left to right. Since λ controls the importance of L ranking , we show that the vanilla baseline (λ = 0) performs worse than any versions with Ranking-aware Distillation loss (λ > 0), showing the proposed distillation Effectiveness of components. We conduct the ablative experiments regarding the proposed components on DeiT-Tiny network. Firstly, we compose the baseline network using the binarization method following Bi-Real Net [26]. As shown in the third row of Tab. 1, the baseline networks only obtains 6.6% Top-1 accuracy, which is far from satisfactory. With the introduction of our first novelty, i.e., learnable scaling factor (LSF), the baseline network is boosted by 17.8%, achieving 24.4% Top-1 accuracy. We also observe the other contribution Ranking-aware Disitllation (RD) singly promotes the baseline network by 5.9%, which is also significant on ImageNet dataset. By combining the two main contributions together, we get Bi-ViT, outperforming the vanilla baseline by 22.1%. + Ranking-aware Distillation (RD) 1-1 12.5 +5.9\n+ LSF + RD (Bi-ViT) 1-1 28.7 +22.1"},{"figure_ref":[],"heading":"Results on Image Classification","publication_ref":["b28","b28","b28","b16","b5","b23"],"table_ref":[],"text":"The experimental results are shown in Tab. 2. We compare our method with 1-bit methods including BiB-ERT [28], RBONN [36], and Bi-Real Net [26] based on the same frameworks for the task of image classification with the ImageNet dataset. We also report the classification per- formance of the low-bit training-aware quantization method Q-ViT [18] for further reference. We use model size and OPs following [26] in comparison to other bit-widthe models for further reference. We firstly evaluate the proposed method on DeiT models.\nFor DeiT-Tiny backbone, compared with other binary methods, our Bi-ViT achieves significant performance improvements. For example, our Bi-ViT surpasses the base-line Bi-Real Net [26] by 22.1% Top-1 accuracy, which is significant and meaningful for real-world applications. And it is worth noting that the proposed 1-bit model significantly compresses the DeiT-Tiny by 61.5× on OPs. The proposed method also boosts the performance of baseline by 21.7% with the same architecture and bit-width using DeiT-Small bacobone, a significant improvement on the Im-ageNet dataset. For larger DeiT-B, as shown in Tab. 2, the Table 3. Experiments with Mask R-CNN [14] and Cascade R-CNN [3] using Swin [23] backbones on COCO [21]. \"#Bits\" denotes the bit-width of weights and activations. We report the AP (%) with different IoU threshold and AP for objects in various sizes. The bold denotes the best result with binarized weights and activations."},{"figure_ref":[],"heading":"Framework","publication_ref":[],"table_ref":[],"text":"Backbone Method # Bits Size "},{"figure_ref":[],"heading":"Conclusion","publication_ref":[],"table_ref":[],"text":"In this paper, we present Bi-ViT, an improved version of fully-binarized ViTs that offers a high compression ratio and acceptable performance. Initially, we establish a empirical framework for fully-binarized ViT and analyze the bottlenecks of the baseline. Our empirical analysis shows that attention distortion in MHSA is the primary cause of the significant drop in ViT binarization, which results from gradient vanishing and ranking disorder. To address these issues, we introduce a learnable scaling factor that reactivates vanished gradients, which we illustrate through both theoretical and experimental analysis. Additionally, we propose ranking-aware distillation for Bi-ViT, which rectifies disordered ranking in a teacher-student framework. Our work provides a comprehensive analysis and effective solutions for the crucial issues in ViT full binarization, paving the way for the extreme compression of ViT."}],"string":"[\n {\n \"figure_ref\": [\n \"fig_0\",\n \"fig_1\",\n \"fig_1\",\n \"fig_2\"\n ],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b6\",\n \"b15\",\n \"b33\",\n \"b24\",\n \"b28\",\n \"b4\",\n \"b27\",\n \"b24\",\n \"b3\",\n \"b26\",\n \"b22\",\n \"b28\",\n \"b28\",\n \"b26\"\n ],\n \"table_ref\": [],\n \"text\": \"Transformers, which have gained far-flung fame in natural language processing (NLP) area [8,28], are also attracting increasing attention in lots of computer vision (CV) tasks, such as object detection [4], image classification [9] and many others [13,31], impelling the widespread research on vision transformers (ViTs). There has a natural fit for ViTs to achieve better performance simply by training a larger model on a larger data set. For example, historical records show better performance of a ViT-H model [9] accompanying with astonishing 632M parameters and 162G FLOPs. Such a high model complexity poses a great challenge to deploy models on platforms with short resource supplies. Therefore, both academia and industry call for an ultimate compression of these large models, and the DeiT-Small on ImageNet [17], respectively. Here 8-bit DeiT is quantized with PTQ method [22] and 2/3/4 bit DeiT is trained with QAT method [18]. The binarized ViT is conducted with the baseline method Bi-Real Net [26].\\npast years have witnessed some promising techniques such as network pruning [38,5], low-rank decomposition [7], knowledge distillation [12], and quantization [18,19].\\nNetwork quantization, which represents weights and activations in a low-bit format, has got great earnestness of many researchers for its reduced memory access costs and increased compute efficiency as well as performance benefit. Using the lower-bit quantized data, in particular to the extreme 1-bit case, requires less data movement, both on-chip and off-chip, and therefore reduces memory bandwidth and saves significant energy. Existing documentary records observe 32× less network size and 58× speedups beneficial from xnor and bit-count logics for 1-bit networks [30]. Earlier attempts [25,22] apply post-training quantization (PTQ) [1,40] directly to ViTs without datadriven fine-tuning, causing sub-optimal performance, in particular to impotent 1-bit ViTs. Therefore, by quantizing while training, quantization-aware training (QAT) methods are more congenial to 1-bit ViTs. Extensive empirical studies [24,20,36,28] have well demonstrated the efficacy of QAT methods in 1-bit convolutional neural networks (CNNs) or BERTs, however, the application to 1-bit ViTs remains not to be fully explored so far.\\nIn this paper, we first build a fully-binarized ViT baseline, a straightforward solution constructed upon popular binarized QAT method of Bi-Real Net [26]. Through an empirical study of this baseline, we observe significant performance drops on the ImageNet dataset [17], as shown in Fig. 1. For instance, extending Bi-Real Net to binarize DeiT-Tiny [32] incurs a tremendous performance gap of 52.6% in the Top-1 accuracy compared to the 2-bit quantized counterpart. Similar performance drops occur in DeiT-Small as well. Delving into a deeper analysis, we find that the incompatibility of existing QAT methods mainly stems from the binarized self-attention module in ViTs, where a simple application of existing binarization methods [26] leads to severe attention distortion, as plotted in Fig. 2 (a) and Fig. 2 (b), especially in the diagonal scores of the map which are supposed to be the most attentive.\\nIn this paper we dig deeper into this attention distortion problem. Through empirical analysis, we find that this phenomenon is mainly caused by gradient vanishing due to the straight-through-estimator (STE) [2] and non-scaled binarization in self-attention. Meanwhile, a simple distillation utilizing distillation token in DeiT [32] and KL-divergence in ReActNet [24] is ineffective in dismissing the ranking disorder, since it neglects the relative order of the attention map between the binarized ViTs and their real-valued counterpart. To address the aforementioned issues, a fullybinarized ViT (Bi-ViT) is developed by reactivating the vanished gradients through a learnable scaling factor in selfattention and a ranking-aware distillation to further effectively rectify the disordered ranking of attention (see the overview in Fig. 3). In addition, we also provide both empirical and theoretical analysis about how our method can rectify the distorted attention and thus promote the optimization of Bi-ViT. The contributions of our work are summarized as:\\n• We identify the bottleneck of a fully-binarized ViT through empirical analyses and formulate the problem in a theoretical perspective. Based on these, we introduce learnable head-wise scaling factor into binarized self-attention to reactivate the vanished gradients.\\n• We develop a ranking-aware distillation scheme to eliminate attention distortion. Our distillation method fully utilizes the ranking-aware knowledge from the real-valued teacher to promote the optimization of Bi-ViT.\\n• Our Bi-ViT is the first promising way to push the limit of ViT quantization to the fully-binarized version. Extensive experiments on the ImageNet benchmark demonstrate that Bi-ViT surpasses both the baseline and prior binarized methods by a significant margin, achieving a remarkable acceleration rate of up to 61.5×. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Work\",\n \"publication_ref\": [\n \"b37\",\n \"b31\",\n \"b39\",\n \"b8\",\n \"b28\",\n \"b26\"\n ],\n \"table_ref\": [],\n \"text\": \"Vision Transformer. Unlike traditional CNN-based models, ViTs are capable of capturing long-range visual relationships through the self-attention mechanism, and offer a more generalizable paradigm without inductive bias specific to images. The starting ViT [9] views an image as a sequence of 16 × 16 patches and uses a unique class token to predict the classification, yielding promising results. Subsequently, many works, such as DeiT [32] and PVT [35], have improved upon ViT, making it more efficient and applicable to downstream tasks. However, these high-performing ViTs have also accompanied with a significant number of parameters and high computational overhead, limiting their widespread applications. Thus, designing smaller and faster ViTs has become a new trend. LeViT [11] makes progress in faster inference through down-sampling, patch descriptors, and a redesign of the Attention-MLP block.\\nDynamicViT [29] proposes a dynamic token sparsification framework to progressively and dynamically prune redundant tokens, achieving a competitive complexity and accuracy trade-off. Evo-ViT [37] proposes a slow-fast updating mechanism that ensures information flow and spatial structure, reducing both the training and inference complexity.\\nWhile the aforementioned works focus on efficient model design, this paper aims to boost compression and acceleration through binarization. Network Binarization. BinaryNet is a technique originally proposed to train convolutional neural networks (CNNs) with binary weights. BinaryConnect [6] is the precursor to BinaryNet, where the parameters are binary while the activations remain in full-precision states. Local binary convolution layers (LBC) [16] were introduced to binarize the non-linear activations, and XNOR-Net [30] was introduced to improve convolution efficiency by binarizing the weights and inputs of convolution kernels. Bi-Real Net [26] explores a new variant of residual structure to preserve the information of real activations before the sign function, with a tight approximation to the derivative of the nondifferentiable sign function. Real-to-binary [27] re-scales the feature maps on the channels according to the input before binarized operations and adds an SE-Net [15] like gating module. ReActNet [24] replaces the conventional PReLU and the sign function of the BNNs with RPReLU and RSign with a learnable threshold, thus improving the performance of BNNs. RBONN [36] introduces a recurrent bilinear optimization to address the asynchronous convergence problem for BNNs, which further improves the performance of BNNs. These techniques improve the efficiency and accuracy of binary neural networks (BNNs) and allow them to be applied in practical applications. Majorities of these techniques consider non-scaled binarization in activations, which is beneficial to conventional CNNs while causing gradient mismatch issue for the pecularity of selfattention mechanism in ViTs.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Background\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Multi-Head Self-Attention and Binarization\",\n \"publication_ref\": [\n \"b28\"\n ],\n \"table_ref\": [],\n \"text\": \"For a multi-head self-attention (MHSA) module, we denote its query, key, and value set as {a {q,k,v} ∈ R h×N ×d }, where h denotes head number, N and d represent the patch and channel numbers of each head. Specifically, N = (W in //W P in ) × (H in //H P in ) where W in and H in are the width and height of the feature, W P in , H P in are the width and height of patch maps respectively. Then, the attention score A and MHSA module output a out are computed as follows [34]:\\nA = softmax[(a q • a k )/ √ d], a out = A • a v ,(1)\\nwhere softmax(•) represents the softmax operation. Intending to represent query, key, value and attention score, i.e., a q , a k , a v and A, in a 1-bit format, Eq. (1) changes into:\\nA = softmax[(b aq • b a k )/ √ d], a out = b A • b av .(2)\\nWe follow the common network binarization methods [30] that use the sign function b • = sign(•) in the binary forward pass, and STE\\n[2] ∂b• ∂• = 1 |•|≤1\\nto compute the gradient for sign function in its backward pass. We omit the non-linear function here for simplicity. For all the projection and linear layers in binarized ViTs, we conduct binarization following [28,26] as\\na out = b ain • (α w • b w ) = α w •(b ain •b w ) where α w = {α 1 w , α 2 w , ..., α Cout w } ∈ R Cout +\\nis known as the channel-wise scaling factor vector [30] and • represents channel-wise channel-wise multiplication. The matrix multiplication process, i.e., b ain • b w , can be executed by the efficient XNOR and Bit-count instructions on edge devices.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Bottleneck of Fully-Binarized ViTs\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The high-performing ViTs are built on premise of transformer's supreme ability to model the long-range relationships thanks to the attention mechanism within the MHSA module. Unfortunately, a binarized version of weights and inputs significantly weakens the representation ability. In Module Degradation. By gradually replacing the multilayer perceptron (MLP) and MHSA modules with realvalued weights or activations, we have discovered that maintaining the MLP as \\\"w1a1\\\" (all weights and activations in the MLP are binarized) still results in satisfactory performance. For instance, keeping MLP as \\\"w1a1\\\" while keeping MHSA as \\\"w1a32\\\" obtains 26.3% Top-1 accuracy, which might be acceptable comparing to the 55.2% of realvalued DeiT-Tiny when taking into consideration 47.3× acceleration rates. On the contrast, when maintaining MHSA module as \\\"w1a1\\\", we observe a significant drop in performance. To be more specific, even when the MLP was maintained as \\\"w32a32\\\", we still observe a significant 50.8% decrease in Top-1 accuracy (from 55.2% to 4.4%). This result indicates that using binarized weights and activations in the MHSA module can have a substantial negative impact on the model's performance, even when other parts retain in real-valued states.\\nOperation Degradation. To better understand the impact of fully-binarized ViT's performance, we conduct further analyses by examining the operations within the MHSA module. Specifically, when we maintain the self-attention activations in Eq. ( 1) as real-valued (\\\"a32\\\"), we observe only a relatively small decrease in performance from 48.8% to 37.6%. However, when the self-attention activations in Eq. (2) are binarized, significant drops in accuracy occur from 48.8% to 7.6%. This finding highlights the importance of the self-attention process within the MHSA module and suggests more efforts to mitigate the negative impact of binarization on the MHSA module.\"\n },\n {\n \"figure_ref\": [\n \"fig_5\"\n ],\n \"heading\": \"Gradient Mismatch in Self-Attention\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"With conclusion from the experimental results in Sec. 3.2 that self-attention process, i.e., Eq. (2), is the most critical part causing the performance drops. We attempt to analyze the underlying reasons for this phenomenon from an optimization perspective. For simplicity, we derive the gradient mismatch in a q as an example, and the analysis can be applicable to explain a k as well. We first represent the features before softmax(•) in Eq. (2) as:\\np = (b aq • b a k .)/ √ d.(3)\\nThe gradient of a hi,n,c q w.r.t. A is formulated as:\\n∂A ∂a h i ,n,c q = ∂A ∂p h i ,n,n • ∂p h i ,n,n ∂b h i ,n,c aq • ∂b h i ,n,c aq ∂a h i ,n,c q , (4\\n)\\nwhere\\nh i ∈ R h , n & n ∈ R N , c ∈ R d\\nand the gradient of a k is likewise. The explicit form of the first item ∂A ∂p h i ,n,n in Eq. ( 4) is:\\n∂A ∂p hi,n,n = ∂ softmax(p hi,n,n ) ∂p hi,n,n = A hi,n,n ⊗ (1 -A hi,n,n ),(5)\\nwhere ⊗ denotes Hadamard product. And the second item is formulated as:\\n∂p h i ,n,n ∂b h i ,n,c aq = ∂b h i ,n,c aq • b h i ,c,n a k ∂b h i ,n,c aq = b h i ,c,n a k ,(6)\\nresult of which is therefore correlated with b a k . The third item is solved through STE [2] as:\\n∂b hi,n,c aq ∂a hi,n,c q = 1 |a h i ,n,c q |≤1 .(7)\\nCombing Eq. (5)-Eq. ( 7), we have the final gradient form in fully-binarized ViTs as:\\n∂A ∂a h i ,n,c q = ∂A ∂p h i ,n,n • ∂p h i ,n,n ∂b h i ,n,c aq • ∂b h i ,n,c aq ∂a h i ,n,c q = A h i ,n,n (1 -A h i ,n,n ) • b h i ,c,n a k • 1 |a h i ,n,c q |≤1 .(8)\\nConsidering b hi,n,:\\naq = [1, • • • , 1] and •b hi,n ,: a k = [1, • • • , 1] as the extreme condition, b hi,n,: aq • b hi,:,n a k = d. Therefore, a specific element in b aq •b a k is ∈ {-d, • • • , d}.\\nWe plot the curve of a specific element in the first item between [-64, 64] in Fig. 5 = 0, likewise for a k . Therefore we formulate the gradient mismatch phenomenon in the aforementioned theoretical analysis. And such gradient mismatch leads to distorted gradient in the optimization of a q & a k and therefore degrades performance of fully-binarized ViTs.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Our Bi-ViT\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this section, we propose to dismiss the affect of gradient mismatch mentioned in Sec. 3.3 from perspectives of gradient approximation in Sec. 4.1 and intermediate distillation in Sec. 4.2.\"\n },\n {\n \"figure_ref\": [\n \"fig_5\",\n \"fig_5\"\n ],\n \"heading\": \"Learnable Head-wise Scaling Factor\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"As one of the solution to the above mentioned problem, we propose a head-wise scaling factor binarization scheme for the self-attention process, where the scaling factors are learned during training to first modify the gradient clip range in Fig. 5(b). Eq. ( 2) is changed into:\\n= softmax(p), p = (α q ⊗ α k ) • (b aq • b a k )/ √ d = α q;k • (b aq • b a k )/ √ d,(9)\\nand\\nãout = (αA • bA) • (αv • ba v ) = (αA ⊗ αv) • (bA • b av ) = αA;v • (bA • b av ),(10)\\nwhere b a• = sign( a• α• ), α q , α k , α v and α A are the headwise learnable scaling factors in binarized MHSA, where\\nα {q,k,v,A} = {α 1 {q,k,v,A} , α 2 {q,k,v,A} , • • • , α h {q,k,v,A} } ∈ R h + .\\nThe second rows in Eq. ( 9) & Eq. ( 10) are established since the scaling factors are aligned with the head dimension, which is independent with the matrix multiplication operation. Thus,\\nα q;k = {α 1 q;k , α 2 q;k , • • • , α h q;k } ∈ R h + and α A;v = {α 1 A;v , α 2 A;v , • • • , α h A;v } ∈ R h + .\\nConsequently, the gradient ∂ à ∂a :,n,c q in Eq. ( 8) is further formulated in our Bi-ViT as:\\n∂ à ∂a h i ,n,c q = Ãh i ,n,n (1 -Ãh i ,n,n ) ∂ à ∂p h i ,n,n • α h i q;k • b h i ,c,n a k ∂p h i ,n,n ∂b h i ,n,c aq • 1 |a h i ,n,c q |≤αq ∂b h i ,n,c aq ∂a h i ,n,c q . (11\\n)\\nSince softmax(.) and • are aligned with different dimensions, the value of Eq. ( 5) remains unchanged (softmax(p) = softmax(α q;k • p)). As can be seen, the threshold of gradient clip in Eq. ( 7) changes from 1 into α q , which means that we can surpass the occurance of gradient mismatch by modifying the value of α q . Note that the scaling factor (α q ) is to imitate the magnitude of the latent activations. When p has a large magnitude, i.e., in the circled part of Fig. 5 (a), α q also tends to be larger and a hi,n,c q locates in the field that ∂b h i ,n,c aq ∂a h i ,n,c q > 0. Thus the vanishing gradients are reactivated through the introduced learnable scaling factor.\"\n },\n {\n \"figure_ref\": [\n \"fig_1\",\n \"fig_2\"\n ],\n \"heading\": \"Ranking-aware Distillation for Bi-ViT\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Fig. 2 illustrates a significant difference in the attention map's relative order between Bi-RealNet (a) and its realvalued counterpart (c). This difference could result in a notable decrease in performance. To address this issue during binarized training, a ranking-aware distillation in a teacherstudent framework is introduced:\\nL ranking = L l=1 ψ(A T ) -ψ(A S ) 2 ,(12)\\nwhere A T and A S represents the attention scores from the real-valued teacher and binarized student. ψ(•) denotes the function for obtaining the ranking, i.e., relative order of an attention score, which is formulated as:\\nψ(A :,n,: ) = A :,n,: -A :,n-1,: , if 0 < n ≤ N -1 A :,0,: -A :,N -1,: , otherwise .(13)\\nDetailed relative order computation can be seen in the right part of Fig. 3. We implement our Bi-ViT under the teacherstudent framework [32], thus the final objective of our method is formulated as:\\nL = L dist + λL ranking ,(14)\\nwhere λ is a hyper-parameter to balance these two loss functions.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiments\",\n \"publication_ref\": [\n \"b16\",\n \"b5\",\n \"b28\"\n ],\n \"table_ref\": [],\n \"text\": \"In this section, we evaluate the performance of the proposed Bi-ViT model for image classification task using pop-ular DeiT [32] & Swin [23] backbones and object detection task using Mask R-CNN [14] & Cascade [3] Mask R-CNN with Swin-Tiny [23] backbone. To the best of our knowledge, there is no publicly available source codebase on fully-binarized ViTs at this point, so we implement the baseline i.e., Bi-Real Net [26] methods by ourselves.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Datasets and Implementation Details\",\n \"publication_ref\": [\n \"b23\",\n \"b23\",\n \"b35\",\n \"b12\"\n ],\n \"table_ref\": [],\n \"text\": \"Datasets. The experiments are carried out on the ImageNet ILSVRC12 dataset [17] for image classification task and COCO dataset [21] for object detection task. The ImageNet dataset is challenging due to its large scale and greater diversity. There are 1000 classes and 1.2 million training images, and 50k validation images in it. In our experiments, we use the classic data augmentation method described in [32].\\nThe COCO [21] dataset includes images from 80 different categories. All of our COCO dataset experiments are performed on the object detection track of the COCO trainval35k training dataset, which consists of 80k images from the COCO train2014 dataset and 35k images sampled from the COCO val2014 dataset. We report the average precision (AP) for IoUs∈ [0.5:0.05:0.95], designated as mAP@[.5,.95], using COCO's standard evaluation metric. For further analyzing our method, we also report AP 50 , AP 75 , AP s , AP m , and AP l . Experimental settings. In our experiments, we initialize the weights of binarized model with the corresponding pretrained real-valued model. The binarized model is trained for 300 epochs with batch-size 512 and the base learning rate 5e -4. We do not use warm-up scheme. For all the experiments, we apply LAMB [39] optimizer with weight decay set as 0, following DeiT III [33]. Note that we keep the patch embedding (first) layer and the classification (last) layer as real-valued, following [10]. Backbone. We evaluate our binarized method on two popular vision transformer networks: DeiT [32] and Swin Transformer [23]. The DeiT-Tiny, DeiT-Small, DeiT-Base, Swin-Tiny and Swin-Small are adopted as the backbone models, whose Top-1 accuracy on ImageNet dataset are 72.2%, 79.9%, 81.8%, 81.2%, and 83.2% respectively. For a fair comparison, we utilize the official implementation of DeiT and Swin Transformer.\"\n },\n {\n \"figure_ref\": [\n \"fig_6\"\n ],\n \"heading\": \"Ablation Study\",\n \"publication_ref\": [\n \"b28\"\n ],\n \"table_ref\": [],\n \"text\": \"Hyper-parameter Selection. We λ of Eq. ( 14) in this part, with experiments conducted on ImageNet [17] dataset. We show the model performance (Top-1 accuracy) with different setups of hyper-parameter λ in Fig. 6, in which the performances increase first and then decrease with the uplift of λ from left to right. Since λ controls the importance of L ranking , we show that the vanilla baseline (λ = 0) performs worse than any versions with Ranking-aware Distillation loss (λ > 0), showing the proposed distillation Effectiveness of components. We conduct the ablative experiments regarding the proposed components on DeiT-Tiny network. Firstly, we compose the baseline network using the binarization method following Bi-Real Net [26]. As shown in the third row of Tab. 1, the baseline networks only obtains 6.6% Top-1 accuracy, which is far from satisfactory. With the introduction of our first novelty, i.e., learnable scaling factor (LSF), the baseline network is boosted by 17.8%, achieving 24.4% Top-1 accuracy. We also observe the other contribution Ranking-aware Disitllation (RD) singly promotes the baseline network by 5.9%, which is also significant on ImageNet dataset. By combining the two main contributions together, we get Bi-ViT, outperforming the vanilla baseline by 22.1%. + Ranking-aware Distillation (RD) 1-1 12.5 +5.9\\n+ LSF + RD (Bi-ViT) 1-1 28.7 +22.1\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Results on Image Classification\",\n \"publication_ref\": [\n \"b28\",\n \"b28\",\n \"b28\",\n \"b16\",\n \"b5\",\n \"b23\"\n ],\n \"table_ref\": [],\n \"text\": \"The experimental results are shown in Tab. 2. We compare our method with 1-bit methods including BiB-ERT [28], RBONN [36], and Bi-Real Net [26] based on the same frameworks for the task of image classification with the ImageNet dataset. We also report the classification per- formance of the low-bit training-aware quantization method Q-ViT [18] for further reference. We use model size and OPs following [26] in comparison to other bit-widthe models for further reference. We firstly evaluate the proposed method on DeiT models.\\nFor DeiT-Tiny backbone, compared with other binary methods, our Bi-ViT achieves significant performance improvements. For example, our Bi-ViT surpasses the base-line Bi-Real Net [26] by 22.1% Top-1 accuracy, which is significant and meaningful for real-world applications. And it is worth noting that the proposed 1-bit model significantly compresses the DeiT-Tiny by 61.5× on OPs. The proposed method also boosts the performance of baseline by 21.7% with the same architecture and bit-width using DeiT-Small bacobone, a significant improvement on the Im-ageNet dataset. For larger DeiT-B, as shown in Tab. 2, the Table 3. Experiments with Mask R-CNN [14] and Cascade R-CNN [3] using Swin [23] backbones on COCO [21]. \\\"#Bits\\\" denotes the bit-width of weights and activations. We report the AP (%) with different IoU threshold and AP for objects in various sizes. The bold denotes the best result with binarized weights and activations.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Framework\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Backbone Method # Bits Size \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this paper, we present Bi-ViT, an improved version of fully-binarized ViTs that offers a high compression ratio and acceptable performance. Initially, we establish a empirical framework for fully-binarized ViT and analyze the bottlenecks of the baseline. Our empirical analysis shows that attention distortion in MHSA is the primary cause of the significant drop in ViT binarization, which results from gradient vanishing and ranking disorder. To address these issues, we introduce a learnable scaling factor that reactivates vanished gradients, which we illustrate through both theoretical and experimental analysis. Additionally, we propose ranking-aware distillation for Bi-ViT, which rectifies disordered ranking in a teacher-student framework. Our work provides a comprehensive analysis and effective solutions for the crucial issues in ViT full binarization, paving the way for the extreme compression of ViT.\"\n }\n]"},"abstract":{"kind":"string","value":"Vision transformers (ViTs) quantization offers a promising prospect to facilitate deploying large pre-trained networks on resource-limited devices. Fully-binarized ViTs (Bi-ViT) that pushes the quantization of ViTs to its limit remain largely unexplored and a very challenging task yet, due to their unacceptable performance. Through extensive empirical analyses, we identify the severe drop in ViT binarization is caused by attention distortion in self-attention, which technically stems from the gradient vanishing and ranking disorder. To address these issues, we first introduce a learnable scaling factor to reactivate the vanished gradients and illustrate its effectiveness through theoretical and experimental analyses. We then propose a ranking-aware distillation method to rectify the disordered ranking in a teacher-student framework. Bi-ViT achieves significant improvements over popular DeiT and Swin backbones in terms of Top-1 accuracy and FLOPs. For example, with DeiT-Tiny and Swin-Tiny, our method significantly outperforms baselines by 22.1% and 21.4% respectively, while 61.5× and 56.1× theoretical acceleration in terms of FLOPs compared with real-valued counterparts on ImageNet."},"title":{"kind":"string","value":"Bi-ViT: Pushing the Limit of Vision Transformer Quantization"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 1 .1Figure 1. Performance of real-valued and quantized DeiT [32] with varying bit-widths. We report results with (a) DeiT-Tiny and (b)DeiT-Small on ImageNet [17], respectively. Here 8-bit DeiT is quantized with PTQ method[22] and 2/3/4 bit DeiT is trained with QAT method[18]. The binarized ViT is conducted with the baseline method Bi-Real Net[26].","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"Figure 2 .2Figure 2. Visualization of the attention map before softmax in the first block of DeiT-Tiny [32] on ImageNet [17]. From the left to right, is the baseline method [26], previous binarization method [36], our Bi-ViT and real-valued counterpart.","figure_data":"","figure_id":"fig_1","figure_label":"2","figure_type":"figure"},{"figure_caption":"Figure 3 .3Figure 3. Overview of the proposed Bi-ViT framework. We introduce the learnable scaling factor in an architecture perspective and a ranking-aware distillation scheme incorporated in the optimization process. From left to right, we respectively show the detailed architecture of single block in Bi-ViT and the distillation framework of Bi-ViT.","figure_data":"","figure_id":"fig_2","figure_label":"3","figure_type":"figure"},{"figure_caption":"Figure 4 .4Figure 4. Performance of fully-binarized DeiT-Tiny on Ima-geNet [17] with different binarized/real-valued settings.","figure_data":"","figure_id":"fig_4","figure_label":"4","figure_type":"figure"},{"figure_caption":"Figure 5 .5Figure 5. Illustration of gradient mismatch between Eq. (5) and Eq. (7).","figure_data":"","figure_id":"fig_5","figure_label":"5","figure_type":"figure"},{"figure_caption":"Figure 6 .6Figure 6. Effect of hyper-parameter λ on ImageNet [17].","figure_data":"","figure_id":"fig_6","figure_label":"6","figure_type":"figure"},{"figure_caption":"Evaluating the components of Bi-ViT based on DeiT-Tiny [32] backbone. \"#Bits\" denotes the bit-width of weights and activations. We report the Top-1 (%) accuracy performances.","figure_data":"Method#Bits Top-1 (%)Real-valued32-32 72.1Baseline (Bi-Real Net [26])1-16.6+ Learnable Scaling Factor (LSF)1-124.4 +17.8","figure_id":"tab_0","figure_label":"1","figure_type":"table"},{"figure_caption":"Experiments withDeiT [32] and Swin[23] on ImageNet[17]. \"#Bits\" denotes the bit-width of weights and activations. We report the Top-1 (%) and Top-5 (%) accuracy performances. The bold denotes the best result with binarized weights and activations.","figure_data":"NetworkMethod#BitsSize (MB)OPs (10 8 )Top-1 (%)Top-5 (%)Real-valued32-3222.812.372.291.14-43.01.674.391.7Q-ViT [18]3-32.30.871.591.2DeiT-TinyBiBERT [28]2-21.70.459.0 5.981.8 16.0RBONN [36] Bi-Real Net [26]1-11.00.26.3 6.616.9 17.1Bi-ViT28.7 +22.151.7 +34.6Real-valued32-3288.245.579.995.04-411.45.880.994.9Q-ViT [18]3-38.73.079.094.2DeiT-SmallBiBERT [28]2-26.01.572.1 17.490.3 29.7RBONN [36] Bi-Real Net [26]1-13.40.818.5 19.230.0 30.3Bi-ViT40.9 +21.765.0 +34.7Real-valued32-32346.2174.781.895.64-444.122.083.096.1Q-ViT [18]3-333.411.181.095.1DeiT-BaseBiBERT [28]2-222.75.774.2 24.592.2 36.3RBONN [36] Bi-Real Net [26]1-112.12.926.1 26.538.6 38.8Bi-ViT47.3 +20.872.8 +34.0Real-valued32-32114.244.981.295.54-414.65.882.597.3Q-ViT [18]3-311.23.080.996.1Swin-TinyBiBERT [28]2-210.01.674.7 34.092.5 46.9RBONN [36] Bi-Real Net [26]1-14.20.833.8 34.146.7 46.9Bi-ViT55.5 +21.479.4 +32.5Real-valued32-32199.887.583.296.24-425.311.184.498.3Q-ViT [18]3-319.25.682.797.5Swin-SmallBiBERT [28]2-213.02.976.9 39.494.9 53.0RBONN [36] Bi-Real Net [26]1-16.91.539.0 39.252.7 52.8Bi-ViT60.7 +21.583.9 +31.1","figure_id":"tab_1","figure_label":"2","figure_type":"table"},{"figure_caption":"(MB) AP AP 50 AP 75 AP s AP m AP l","figure_data":"Real-valued32-32 191.343.7 66.647.728.5 47.0 57.34-494.943.3 66.347.128.2 46.5 57.5Q-ViT [18]3-391.440.1 63.543.925.4 42.4 54.9Mask R-CNN [14] Swin-TinyBiBERT [28]2-288.030.2 53.7 8.9 25.033.4 8.615.2 32.0 45.2 1.7 9.0 15.9RBONN [36] Bi-Real Net [26]1-184.59.5 9.925.2 25.28.9 9.21.9 2.19.1 9.116.0 16.4Bi-ViT20.7 38.719.912.0 20.9 27.6Cascade Mask R-CNN [3]Swin-Tiny","figure_id":"tab_2","figure_label":"","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 1 .1Figure 1. Performance of real-valued and quantized DeiT [32] with varying bit-widths. We report results with (a) DeiT-Tiny and (b)DeiT-Small on ImageNet [17], respectively. Here 8-bit DeiT is quantized with PTQ method[22] and 2/3/4 bit DeiT is trained with QAT method[18]. The binarized ViT is conducted with the baseline method Bi-Real Net[26].\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 2 .2Figure 2. Visualization of the attention map before softmax in the first block of DeiT-Tiny [32] on ImageNet [17]. From the left to right, is the baseline method [26], previous binarization method [36], our Bi-ViT and real-valued counterpart.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 .3Figure 3. Overview of the proposed Bi-ViT framework. We introduce the learnable scaling factor in an architecture perspective and a ranking-aware distillation scheme incorporated in the optimization process. From left to right, we respectively show the detailed architecture of single block in Bi-ViT and the distillation framework of Bi-ViT.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 4 .4Figure 4. Performance of fully-binarized DeiT-Tiny on Ima-geNet [17] with different binarized/real-valued settings.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 5 .5Figure 5. Illustration of gradient mismatch between Eq. (5) and Eq. (7).\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_5\",\n \"figure_label\": \"5\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 6 .6Figure 6. Effect of hyper-parameter λ on ImageNet [17].\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_6\",\n \"figure_label\": \"6\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Evaluating the components of Bi-ViT based on DeiT-Tiny [32] backbone. \\\"#Bits\\\" denotes the bit-width of weights and activations. We report the Top-1 (%) accuracy performances.\",\n \"figure_data\": \"Method#Bits Top-1 (%)Real-valued32-32 72.1Baseline (Bi-Real Net [26])1-16.6+ Learnable Scaling Factor (LSF)1-124.4 +17.8\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Experiments withDeiT [32] and Swin[23] on ImageNet[17]. \\\"#Bits\\\" denotes the bit-width of weights and activations. We report the Top-1 (%) and Top-5 (%) accuracy performances. The bold denotes the best result with binarized weights and activations.\",\n \"figure_data\": \"NetworkMethod#BitsSize (MB)OPs (10 8 )Top-1 (%)Top-5 (%)Real-valued32-3222.812.372.291.14-43.01.674.391.7Q-ViT [18]3-32.30.871.591.2DeiT-TinyBiBERT [28]2-21.70.459.0 5.981.8 16.0RBONN [36] Bi-Real Net [26]1-11.00.26.3 6.616.9 17.1Bi-ViT28.7 +22.151.7 +34.6Real-valued32-3288.245.579.995.04-411.45.880.994.9Q-ViT [18]3-38.73.079.094.2DeiT-SmallBiBERT [28]2-26.01.572.1 17.490.3 29.7RBONN [36] Bi-Real Net [26]1-13.40.818.5 19.230.0 30.3Bi-ViT40.9 +21.765.0 +34.7Real-valued32-32346.2174.781.895.64-444.122.083.096.1Q-ViT [18]3-333.411.181.095.1DeiT-BaseBiBERT [28]2-222.75.774.2 24.592.2 36.3RBONN [36] Bi-Real Net [26]1-112.12.926.1 26.538.6 38.8Bi-ViT47.3 +20.872.8 +34.0Real-valued32-32114.244.981.295.54-414.65.882.597.3Q-ViT [18]3-311.23.080.996.1Swin-TinyBiBERT [28]2-210.01.674.7 34.092.5 46.9RBONN [36] Bi-Real Net [26]1-14.20.833.8 34.146.7 46.9Bi-ViT55.5 +21.479.4 +32.5Real-valued32-32199.887.583.296.24-425.311.184.498.3Q-ViT [18]3-319.25.682.797.5Swin-SmallBiBERT [28]2-213.02.976.9 39.494.9 53.0RBONN [36] Bi-Real Net [26]1-16.91.539.0 39.252.7 52.8Bi-ViT60.7 +21.583.9 +31.1\",\n \"figure_id\": \"tab_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"(MB) AP AP 50 AP 75 AP s AP m AP l\",\n \"figure_data\": \"Real-valued32-32 191.343.7 66.647.728.5 47.0 57.34-494.943.3 66.347.128.2 46.5 57.5Q-ViT [18]3-391.440.1 63.543.925.4 42.4 54.9Mask R-CNN [14] Swin-TinyBiBERT [28]2-288.030.2 53.7 8.9 25.033.4 8.615.2 32.0 45.2 1.7 9.0 15.9RBONN [36] Bi-Real Net [26]1-184.59.5 9.925.2 25.28.9 9.21.9 2.19.1 9.116.0 16.4Bi-ViT20.7 38.719.912.0 20.9 27.6Cascade Mask R-CNN [3]Swin-Tiny\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n }\n]"},"authors":{"kind":"string","value":"Yanjing Li; Sheng Xu; Mingbao Lin; Xianbin Cao; Chuanjian Liu; Xiao Sun; Baochang Zhang"},"references":{"kind":"list like","value":[{"authors":"","journal":"Real-valued","ref_id":"b0","title":"","year":""},{"authors":"","journal":"","ref_id":"b1","title":"Bi-Real Net","year":""},{"authors":"","journal":"","ref_id":"b2","title":"To provide a basis for comparison, we also included the performance of 2/3/4-bit Q-ViT in Table 3. Our method outperform BiBERT, RBONN","year":null},{"authors":"Ron Banner; Yury Nahshan; Daniel Soudry","journal":"","ref_id":"b3","title":"Post training 4-bit quantization of convolutional networks for rapiddeployment","year":"2019"},{"authors":"Yoshua Bengio; Nicholas Léonard; Aaron Courville","journal":"","ref_id":"b4","title":"Estimating or propagating gradients through stochastic neurons for conditional computation","year":"2013"},{"authors":"Zhaowei Cai; Nuno Vasconcelos","journal":"","ref_id":"b5","title":"Cascade r-cnn: Delving into high quality object detection","year":"2018"},{"authors":"Nicolas Carion; Francisco Massa; Gabriel Synnaeve; Nicolas Usunier; Alexander Kirillov; Sergey Zagoruyko","journal":"","ref_id":"b6","title":"End-toend object detection with transformers","year":"2020"},{"authors":"Mengzhao Chen; Mingbao Lin; Ke Li; Yunhang Shen; Yongjian Wu; Fei Chao; Rongrong Ji","journal":"","ref_id":"b7","title":"Cf-vit: A general coarse-to-fine method for vision transformer","year":"2023"},{"authors":"Matthieu Courbariaux; Yoshua Bengio; Jean-Pierre David","journal":"","ref_id":"b8","title":"Binaryconnect: Training deep neural networks with binary weights during propagations","year":"2015"},{"authors":"Misha Denil; Babak Shakibi; Laurent Dinh; Marc'aurelio Ranzato; Nando De Freitas","journal":"","ref_id":"b9","title":"Predicting parameters in deep learning","year":"2013"},{"authors":"Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova","journal":"","ref_id":"b10","title":"Bert: Pre-training of deep bidirectional transformers for language understanding","year":"2018"},{"authors":"Alexey Dosovitskiy; Lucas Beyer; Alexander Kolesnikov; Dirk Weissenborn; Xiaohua Zhai; Thomas Unterthiner; Mostafa Dehghani; Matthias Minderer; Georg Heigold; Sylvain Gelly","journal":"","ref_id":"b11","title":"An image is worth 16x16 words: Transformers for image recognition at scale","year":"2020"},{"authors":"Jeffrey L Steven K Esser; Deepika Mckinstry; Rathinakumar Bablani; Dharmendra S Appuswamy; Modha","journal":"","ref_id":"b12","title":"Learned step size quantization","year":"2019"},{"authors":"Benjamin Graham; Alaaeldin El-Nouby; Hugo Touvron; Pierre Stock; Armand Joulin; Hervé Jégou; Matthijs Douze","journal":"","ref_id":"b13","title":"Levit: a vision transformer in convnet's clothing for faster inference","year":"2021"},{"authors":"Zhiwei Hao; Jianyuan Guo; Ding Jia; Kai Han; Yehui Tang; Chao Zhang; Han Hu; Yunhe Wang","journal":"","ref_id":"b14","title":"Learning efficient vision transformers via fine-grained manifold distillation","year":"2021"},{"authors":"Kaiming He; Xinlei Chen; Saining Xie; Yanghao Li; Piotr Dollár; Ross Girshick","journal":"","ref_id":"b15","title":"Masked autoencoders are scalable vision learners","year":"2022"},{"authors":"Kaiming He; Georgia Gkioxari","journal":"","ref_id":"b16","title":"Piotr Dollár, and Ross Girshick","year":"2017"},{"authors":"Jie Hu; Li Shen; Gang Sun","journal":"","ref_id":"b17","title":"Squeeze-and-excitation networks","year":"2018"},{"authors":"Felix Juefei-Xu; Vishnu Naresh Boddeti; Marios Savvides","journal":"","ref_id":"b18","title":"Local binary convolutional neural networks","year":"2017"},{"authors":"Alex Krizhevsky; Ilya Sutskever; Geoffrey E Hinton","journal":"","ref_id":"b19","title":"Imagenet classification with deep convolutional neural networks","year":"2012"},{"authors":"Yanjing Li; Sheng Xu; Baochang Zhang; Xianbin Cao; Peng Gao; Guodong Guo","journal":"","ref_id":"b20","title":"Q-vit: Accurate and fully quantized low-bit vision transformer","year":"2022"},{"authors":"Zhexin Li; Tong Yang; Peisong Wang; Jian Cheng","journal":"","ref_id":"b21","title":"Qvit: Fully differentiable quantization for vision transformer","year":"2022"},{"authors":"Mingbao Lin; Rongrong Ji; Zihan Xu; Baochang Zhang; Yan Wang; Yongjian Wu; Feiyue Huang; Chia-Wen Lin","journal":"","ref_id":"b22","title":"Rotated binary neural network","year":"2020"},{"authors":"Tsung-Yi Lin; Michael Maire; Serge Belongie; James Hays; Pietro Perona; Deva Ramanan; Piotr Dollár; C Lawrence; Zitnick ","journal":"","ref_id":"b23","title":"Microsoft coco: Common objects in context","year":"2014"},{"authors":"Yang Lin; Tianyu Zhang; Peiqin Sun; Zheng Li; Shuchang Zhou","journal":"","ref_id":"b24","title":"Fq-vit: Fully quantized vision transformer without retraining","year":"2022"},{"authors":"Ze Liu; Yutong Lin; Yue Cao; Han Hu; Yixuan Wei; Zheng Zhang; Stephen Lin; Baining Guo","journal":"","ref_id":"b25","title":"Swin transformer: Hierarchical vision transformer using shifted windows","year":"2021"},{"authors":"Zechun Liu; Zhiqiang Shen; Marios Savvides; Kwang-Ting Cheng","journal":"","ref_id":"b26","title":"Reactnet: Towards precise binary neural network with generalized activation functions","year":"2020"},{"authors":"Zhenhua Liu; Yunhe Wang; Kai Han; Wei Zhang; Siwei Ma; Wen Gao","journal":"","ref_id":"b27","title":"Post-training quantization for vision transformer","year":"2021"},{"authors":"Zechun Liu; Baoyuan Wu; Wenhan Luo; Xin Yang; Wei Liu; Kwang-Ting Cheng","journal":"","ref_id":"b28","title":"Bi-real net: Enhancing the performance of 1-bit cnns with improved representational capability and advanced training algorithm","year":"2007"},{"authors":"Brais Martinez; Jing Yang; Adrian Bulat; Georgios Tzimiropoulos","journal":"","ref_id":"b29","title":"Training binary neural networks with real-tobinary convolutions","year":"2020"},{"authors":"Yifu Haotong Qin; Mingyuan Ding; Qinghua Zhang; Aishan Yan; Qingqing Liu; Ziwei Dang; Xianglong Liu; Liu","journal":"","ref_id":"b30","title":"Bibert: Accurate fully binarized bert","year":"2022"},{"authors":"Yongming Rao; Wenliang Zhao; Benlin Liu; Jiwen Lu; Jie Zhou; Cho-Jui Hsieh","journal":"","ref_id":"b31","title":"Dynamicvit: Efficient vision transformers with dynamic token sparsification","year":"2021"},{"authors":"Mohammad Rastegari; Vicente Ordonez; Joseph Redmon; Ali Farhadi","journal":"","ref_id":"b32","title":"Xnor-net: Imagenet classification using binary convolutional neural networks","year":"2016"},{"authors":"Yunjie Tian; Lingxi Xie; Zhaozhi Wang; Longhui Wei; Xiaopeng Zhang; Jianbin Jiao; Yaowei Wang; Qi Tian; Qixiang Ye","journal":"","ref_id":"b33","title":"Integrally pre-trained transformer pyramid networks","year":"2022"},{"authors":"Hugo Touvron; Matthieu Cord; Matthijs Douze; Francisco Massa; Alexandre Sablayrolles; Hervé Jégou","journal":"","ref_id":"b34","title":"Training data-efficient image transformers & distillation through attention","year":"2007"},{"authors":"Hugo Touvron; Matthieu Cord; Hervé Jégou","journal":"","ref_id":"b35","title":"Deit iii: Revenge of the vit","year":"2022"},{"authors":"Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Łukasz Kaiser; Illia Polosukhin","journal":"","ref_id":"b36","title":"Attention is all you need","year":"2017"},{"authors":"Wenhai Wang; Enze Xie; Xiang Li; Deng-Ping Fan; Kaitao Song; Ding Liang; Tong Lu; Ping Luo; Ling Shao","journal":"","ref_id":"b37","title":"Pyramid vision transformer: A versatile backbone for dense prediction without convolutions","year":"2021"},{"authors":"Sheng Xu; Yanjing Li; Tiancheng Wang; Teli Ma; Baochang Zhang; Peng Gao; Yu Qiao; Jinhu Lü; Guodong Guo","journal":"","ref_id":"b38","title":"Recurrent bilinear optimization for binary neural networks","year":"2007"},{"authors":"Yifan Xu; Zhijie Zhang; Mengdan Zhang; Kekai Sheng; Ke Li; Weiming Dong; Liqing Zhang; Changsheng Xu; Xing Sun","journal":"","ref_id":"b39","title":"Evo-vit: Slow-fast token evolution for dynamic vision transformer","year":"2022"},{"authors":"Huanrui Yang; Hongxu Yin; Pavlo Molchanov; Hai Li; Jan Kautz","journal":"","ref_id":"b40","title":"Nvit: Vision transformer compression and parameter redistribution","year":"2021"},{"authors":"Yang You; Jing Li; Sashank Reddi; Jonathan Hseu; Sanjiv Kumar; Srinadh Bhojanapalli; Xiaodan Song; James Demmel; Kurt Keutzer; Cho-Jui Hsieh","journal":"","ref_id":"b41","title":"Large batch optimization for deep learning: Training bert in 76 minutes","year":"2020"},{"authors":"Yunshan Zhong; Mingbao Lin; Mengzhao Chen; Ke Li; Yunhang Shen; Fei Chao; Yongjian Wu; Rongrong Ji","journal":"","ref_id":"b42","title":"Finegrained data distribution alignment for post-training quantization","year":"2022"}],"string":"[\n {\n \"authors\": \"\",\n \"journal\": \"Real-valued\",\n \"ref_id\": \"b0\",\n \"title\": \"\",\n \"year\": \"\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"\",\n \"ref_id\": \"b1\",\n \"title\": \"Bi-Real Net\",\n \"year\": \"\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"To provide a basis for comparison, we also included the performance of 2/3/4-bit Q-ViT in Table 3. Our method outperform BiBERT, RBONN\",\n \"year\": null\n },\n {\n \"authors\": \"Ron Banner; Yury Nahshan; Daniel Soudry\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"Post training 4-bit quantization of convolutional networks for rapiddeployment\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Yoshua Bengio; Nicholas Léonard; Aaron Courville\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"Estimating or propagating gradients through stochastic neurons for conditional computation\",\n \"year\": \"2013\"\n },\n {\n \"authors\": \"Zhaowei Cai; Nuno Vasconcelos\",\n \"journal\": \"\",\n \"ref_id\": \"b5\",\n \"title\": \"Cascade r-cnn: Delving into high quality object detection\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Nicolas Carion; Francisco Massa; Gabriel Synnaeve; Nicolas Usunier; Alexander Kirillov; Sergey Zagoruyko\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"End-toend object detection with transformers\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Mengzhao Chen; Mingbao Lin; Ke Li; Yunhang Shen; Yongjian Wu; Fei Chao; Rongrong Ji\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"Cf-vit: A general coarse-to-fine method for vision transformer\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Matthieu Courbariaux; Yoshua Bengio; Jean-Pierre David\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"Binaryconnect: Training deep neural networks with binary weights during propagations\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Misha Denil; Babak Shakibi; Laurent Dinh; Marc'aurelio Ranzato; Nando De Freitas\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"Predicting parameters in deep learning\",\n \"year\": \"2013\"\n },\n {\n \"authors\": \"Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"Bert: Pre-training of deep bidirectional transformers for language understanding\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Alexey Dosovitskiy; Lucas Beyer; Alexander Kolesnikov; Dirk Weissenborn; Xiaohua Zhai; Thomas Unterthiner; Mostafa Dehghani; Matthias Minderer; Georg Heigold; Sylvain Gelly\",\n \"journal\": \"\",\n \"ref_id\": \"b11\",\n \"title\": \"An image is worth 16x16 words: Transformers for image recognition at scale\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Jeffrey L Steven K Esser; Deepika Mckinstry; Rathinakumar Bablani; Dharmendra S Appuswamy; Modha\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"Learned step size quantization\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Benjamin Graham; Alaaeldin El-Nouby; Hugo Touvron; Pierre Stock; Armand Joulin; Hervé Jégou; Matthijs Douze\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"Levit: a vision transformer in convnet's clothing for faster inference\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Zhiwei Hao; Jianyuan Guo; Ding Jia; Kai Han; Yehui Tang; Chao Zhang; Han Hu; Yunhe Wang\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"Learning efficient vision transformers via fine-grained manifold distillation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Kaiming He; Xinlei Chen; Saining Xie; Yanghao Li; Piotr Dollár; Ross Girshick\",\n \"journal\": \"\",\n \"ref_id\": \"b15\",\n \"title\": \"Masked autoencoders are scalable vision learners\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Kaiming He; Georgia Gkioxari\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"Piotr Dollár, and Ross Girshick\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Jie Hu; Li Shen; Gang Sun\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"Squeeze-and-excitation networks\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Felix Juefei-Xu; Vishnu Naresh Boddeti; Marios Savvides\",\n \"journal\": \"\",\n \"ref_id\": \"b18\",\n \"title\": \"Local binary convolutional neural networks\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Alex Krizhevsky; Ilya Sutskever; Geoffrey E Hinton\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"Imagenet classification with deep convolutional neural networks\",\n \"year\": \"2012\"\n },\n {\n \"authors\": \"Yanjing Li; Sheng Xu; Baochang Zhang; Xianbin Cao; Peng Gao; Guodong Guo\",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"Q-vit: Accurate and fully quantized low-bit vision transformer\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Zhexin Li; Tong Yang; Peisong Wang; Jian Cheng\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"Qvit: Fully differentiable quantization for vision transformer\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Mingbao Lin; Rongrong Ji; Zihan Xu; Baochang Zhang; Yan Wang; Yongjian Wu; Feiyue Huang; Chia-Wen Lin\",\n \"journal\": \"\",\n \"ref_id\": \"b22\",\n \"title\": \"Rotated binary neural network\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Tsung-Yi Lin; Michael Maire; Serge Belongie; James Hays; Pietro Perona; Deva Ramanan; Piotr Dollár; C Lawrence; Zitnick \",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"Microsoft coco: Common objects in context\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"Yang Lin; Tianyu Zhang; Peiqin Sun; Zheng Li; Shuchang Zhou\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Fq-vit: Fully quantized vision transformer without retraining\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Ze Liu; Yutong Lin; Yue Cao; Han Hu; Yixuan Wei; Zheng Zhang; Stephen Lin; Baining Guo\",\n \"journal\": \"\",\n \"ref_id\": \"b25\",\n \"title\": \"Swin transformer: Hierarchical vision transformer using shifted windows\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Zechun Liu; Zhiqiang Shen; Marios Savvides; Kwang-Ting Cheng\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"Reactnet: Towards precise binary neural network with generalized activation functions\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Zhenhua Liu; Yunhe Wang; Kai Han; Wei Zhang; Siwei Ma; Wen Gao\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"Post-training quantization for vision transformer\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Zechun Liu; Baoyuan Wu; Wenhan Luo; Xin Yang; Wei Liu; Kwang-Ting Cheng\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"Bi-real net: Enhancing the performance of 1-bit cnns with improved representational capability and advanced training algorithm\",\n \"year\": \"2007\"\n },\n {\n \"authors\": \"Brais Martinez; Jing Yang; Adrian Bulat; Georgios Tzimiropoulos\",\n \"journal\": \"\",\n \"ref_id\": \"b29\",\n \"title\": \"Training binary neural networks with real-tobinary convolutions\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Yifu Haotong Qin; Mingyuan Ding; Qinghua Zhang; Aishan Yan; Qingqing Liu; Ziwei Dang; Xianglong Liu; Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b30\",\n \"title\": \"Bibert: Accurate fully binarized bert\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Yongming Rao; Wenliang Zhao; Benlin Liu; Jiwen Lu; Jie Zhou; Cho-Jui Hsieh\",\n \"journal\": \"\",\n \"ref_id\": \"b31\",\n \"title\": \"Dynamicvit: Efficient vision transformers with dynamic token sparsification\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Mohammad Rastegari; Vicente Ordonez; Joseph Redmon; Ali Farhadi\",\n \"journal\": \"\",\n \"ref_id\": \"b32\",\n \"title\": \"Xnor-net: Imagenet classification using binary convolutional neural networks\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Yunjie Tian; Lingxi Xie; Zhaozhi Wang; Longhui Wei; Xiaopeng Zhang; Jianbin Jiao; Yaowei Wang; Qi Tian; Qixiang Ye\",\n \"journal\": \"\",\n \"ref_id\": \"b33\",\n \"title\": \"Integrally pre-trained transformer pyramid networks\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Hugo Touvron; Matthieu Cord; Matthijs Douze; Francisco Massa; Alexandre Sablayrolles; Hervé Jégou\",\n \"journal\": \"\",\n \"ref_id\": \"b34\",\n \"title\": \"Training data-efficient image transformers & distillation through attention\",\n \"year\": \"2007\"\n },\n {\n \"authors\": \"Hugo Touvron; Matthieu Cord; Hervé Jégou\",\n \"journal\": \"\",\n \"ref_id\": \"b35\",\n \"title\": \"Deit iii: Revenge of the vit\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Łukasz Kaiser; Illia Polosukhin\",\n \"journal\": \"\",\n \"ref_id\": \"b36\",\n \"title\": \"Attention is all you need\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Wenhai Wang; Enze Xie; Xiang Li; Deng-Ping Fan; Kaitao Song; Ding Liang; Tong Lu; Ping Luo; Ling Shao\",\n \"journal\": \"\",\n \"ref_id\": \"b37\",\n \"title\": \"Pyramid vision transformer: A versatile backbone for dense prediction without convolutions\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Sheng Xu; Yanjing Li; Tiancheng Wang; Teli Ma; Baochang Zhang; Peng Gao; Yu Qiao; Jinhu Lü; Guodong Guo\",\n \"journal\": \"\",\n \"ref_id\": \"b38\",\n \"title\": \"Recurrent bilinear optimization for binary neural networks\",\n \"year\": \"2007\"\n },\n {\n \"authors\": \"Yifan Xu; Zhijie Zhang; Mengdan Zhang; Kekai Sheng; Ke Li; Weiming Dong; Liqing Zhang; Changsheng Xu; Xing Sun\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"Evo-vit: Slow-fast token evolution for dynamic vision transformer\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Huanrui Yang; Hongxu Yin; Pavlo Molchanov; Hai Li; Jan Kautz\",\n \"journal\": \"\",\n \"ref_id\": \"b40\",\n \"title\": \"Nvit: Vision transformer compression and parameter redistribution\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Yang You; Jing Li; Sashank Reddi; Jonathan Hseu; Sanjiv Kumar; Srinadh Bhojanapalli; Xiaodan Song; James Demmel; Kurt Keutzer; Cho-Jui Hsieh\",\n \"journal\": \"\",\n \"ref_id\": \"b41\",\n \"title\": \"Large batch optimization for deep learning: Training bert in 76 minutes\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Yunshan Zhong; Mingbao Lin; Mengzhao Chen; Ke Li; Yunhang Shen; Fei Chao; Yongjian Wu; Rongrong Ji\",\n \"journal\": \"\",\n \"ref_id\": \"b42\",\n \"title\": \"Finegrained data distribution alignment for post-training quantization\",\n \"year\": \"2022\"\n }\n]"},"formulas":{"kind":"list like","value":[{"formula_coordinates":[3,366.84,347.17,178.27,36.13],"formula_id":"formula_0","formula_text":"A = softmax[(a q • a k )/ √ d], a out = A • a v ,(1)"},{"formula_coordinates":[3,362.63,430.27,182.48,36.13],"formula_id":"formula_1","formula_text":"A = softmax[(b aq • b a k )/ √ d], a out = b A • b av .(2)"},{"formula_coordinates":[3,401.07,497.51,64.29,13.64],"formula_id":"formula_2","formula_text":"[2] ∂b• ∂• = 1 |•|≤1"},{"formula_coordinates":[3,308.86,547.34,236.25,22.62],"formula_id":"formula_3","formula_text":"a out = b ain • (α w • b w ) = α w •(b ain •b w ) where α w = {α 1 w , α 2 w , ..., α Cout w } ∈ R Cout +"},{"formula_coordinates":[4,377.37,191.2,167.75,20.27],"formula_id":"formula_4","formula_text":"p = (b aq • b a k .)/ √ d.(3)"},{"formula_coordinates":[4,342.23,245.1,199.4,24.74],"formula_id":"formula_5","formula_text":"∂A ∂a h i ,n,c q = ∂A ∂p h i ,n,n • ∂p h i ,n,n ∂b h i ,n,c aq • ∂b h i ,n,c aq ∂a h i ,n,c q , (4"},{"formula_coordinates":[4,541.63,254.24,3.48,7.77],"formula_id":"formula_6","formula_text":")"},{"formula_coordinates":[4,335.97,281.07,130.08,11.23],"formula_id":"formula_7","formula_text":"h i ∈ R h , n & n ∈ R N , c ∈ R d"},{"formula_coordinates":[4,347.26,329.62,197.85,38.43],"formula_id":"formula_8","formula_text":"∂A ∂p hi,n,n = ∂ softmax(p hi,n,n ) ∂p hi,n,n = A hi,n,n ⊗ (1 -A hi,n,n ),(5)"},{"formula_coordinates":[4,365.18,410.74,179.93,43.06],"formula_id":"formula_9","formula_text":"∂p h i ,n,n ∂b h i ,n,c aq = ∂b h i ,n,c aq • b h i ,c,n a k ∂b h i ,n,c aq = b h i ,c,n a k ,(6)"},{"formula_coordinates":[4,379.33,496.14,165.78,27.74],"formula_id":"formula_10","formula_text":"∂b hi,n,c aq ∂a hi,n,c q = 1 |a h i ,n,c q |≤1 .(7)"},{"formula_coordinates":[4,317.04,567.05,228.07,54.41],"formula_id":"formula_11","formula_text":"∂A ∂a h i ,n,c q = ∂A ∂p h i ,n,n • ∂p h i ,n,n ∂b h i ,n,c aq • ∂b h i ,n,c aq ∂a h i ,n,c q = A h i ,n,n (1 -A h i ,n,n ) • b h i ,c,n a k • 1 |a h i ,n,c q |≤1 .(8)"},{"formula_coordinates":[4,308.86,623.4,236.25,41.56],"formula_id":"formula_12","formula_text":"aq = [1, • • • , 1] and •b hi,n ,: a k = [1, • • • , 1] as the extreme condition, b hi,n,: aq • b hi,:,n a k = d. Therefore, a specific element in b aq •b a k is ∈ {-d, • • • , d}."},{"formula_coordinates":[5,101.31,505.79,185.05,45.98],"formula_id":"formula_13","formula_text":"= softmax(p), p = (α q ⊗ α k ) • (b aq • b a k )/ √ d = α q;k • (b aq • b a k )/ √ d,(9)"},{"formula_coordinates":[5,106.18,570.36,180.18,40.45],"formula_id":"formula_14","formula_text":"ãout = (αA • bA) • (αv • ba v ) = (αA ⊗ αv) • (bA • b av ) = αA;v • (bA • b av ),(10)"},{"formula_coordinates":[5,50.11,638.72,236.25,26.6],"formula_id":"formula_15","formula_text":"α {q,k,v,A} = {α 1 {q,k,v,A} , α 2 {q,k,v,A} , • • • , α h {q,k,v,A} } ∈ R h + ."},{"formula_coordinates":[5,50.11,688.99,236.25,26.11],"formula_id":"formula_16","formula_text":"α q;k = {α 1 q;k , α 2 q;k , • • • , α h q;k } ∈ R h + and α A;v = {α 1 A;v , α 2 A;v , • • • , α h A;v } ∈ R h + ."},{"formula_coordinates":[5,310.06,104.28,252.5,59.27],"formula_id":"formula_17","formula_text":"∂ à ∂a h i ,n,c q = Ãh i ,n,n (1 -Ãh i ,n,n ) ∂ à ∂p h i ,n,n • α h i q;k • b h i ,c,n a k ∂p h i ,n,n ∂b h i ,n,c aq • 1 |a h i ,n,c q |≤αq ∂b h i ,n,c aq ∂a h i ,n,c q . (11"},{"formula_coordinates":[5,541.38,155.78,3.73,7.77],"formula_id":"formula_18","formula_text":")"},{"formula_coordinates":[5,349.59,422.92,195.52,30.55],"formula_id":"formula_19","formula_text":"L ranking = L l=1 ψ(A T ) -ψ(A S ) 2 ,(12)"},{"formula_coordinates":[5,315.14,518.42,229.97,39.48],"formula_id":"formula_20","formula_text":"ψ(A :,n,: ) = A :,n,: -A :,n-1,: , if 0 < n ≤ N -1 A :,0,: -A :,N -1,: , otherwise .(13)"},{"formula_coordinates":[5,377.57,617.26,167.54,9.65],"formula_id":"formula_21","formula_text":"L = L dist + λL ranking ,(14)"}],"string":"[\n {\n \"formula_coordinates\": [\n 3,\n 366.84,\n 347.17,\n 178.27,\n 36.13\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"A = softmax[(a q • a k )/ √ d], a out = A • a v ,(1)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 362.63,\n 430.27,\n 182.48,\n 36.13\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"A = softmax[(b aq • b a k )/ √ d], a out = b A • b av .(2)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 401.07,\n 497.51,\n 64.29,\n 13.64\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"[2] ∂b• ∂• = 1 |•|≤1\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 308.86,\n 547.34,\n 236.25,\n 22.62\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"a out = b ain • (α w • b w ) = α w •(b ain •b w ) where α w = {α 1 w , α 2 w , ..., α Cout w } ∈ R Cout +\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 377.37,\n 191.2,\n 167.75,\n 20.27\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"p = (b aq • b a k .)/ √ d.(3)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 342.23,\n 245.1,\n 199.4,\n 24.74\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"∂A ∂a h i ,n,c q = ∂A ∂p h i ,n,n • ∂p h i ,n,n ∂b h i ,n,c aq • ∂b h i ,n,c aq ∂a h i ,n,c q , (4\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 541.63,\n 254.24,\n 3.48,\n 7.77\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \")\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 335.97,\n 281.07,\n 130.08,\n 11.23\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"h i ∈ R h , n & n ∈ R N , c ∈ R d\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 347.26,\n 329.62,\n 197.85,\n 38.43\n ],\n \"formula_id\": \"formula_8\",\n \"formula_text\": \"∂A ∂p hi,n,n = ∂ softmax(p hi,n,n ) ∂p hi,n,n = A hi,n,n ⊗ (1 -A hi,n,n ),(5)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 365.18,\n 410.74,\n 179.93,\n 43.06\n ],\n \"formula_id\": \"formula_9\",\n \"formula_text\": \"∂p h i ,n,n ∂b h i ,n,c aq = ∂b h i ,n,c aq • b h i ,c,n a k ∂b h i ,n,c aq = b h i ,c,n a k ,(6)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 379.33,\n 496.14,\n 165.78,\n 27.74\n ],\n \"formula_id\": \"formula_10\",\n \"formula_text\": \"∂b hi,n,c aq ∂a hi,n,c q = 1 |a h i ,n,c q |≤1 .(7)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 317.04,\n 567.05,\n 228.07,\n 54.41\n ],\n \"formula_id\": \"formula_11\",\n \"formula_text\": \"∂A ∂a h i ,n,c q = ∂A ∂p h i ,n,n • ∂p h i ,n,n ∂b h i ,n,c aq • ∂b h i ,n,c aq ∂a h i ,n,c q = A h i ,n,n (1 -A h i ,n,n ) • b h i ,c,n a k • 1 |a h i ,n,c q |≤1 .(8)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 308.86,\n 623.4,\n 236.25,\n 41.56\n ],\n \"formula_id\": \"formula_12\",\n \"formula_text\": \"aq = [1, • • • , 1] and •b hi,n ,: a k = [1, • • • , 1] as the extreme condition, b hi,n,: aq • b hi,:,n a k = d. Therefore, a specific element in b aq •b a k is ∈ {-d, • • • , d}.\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 101.31,\n 505.79,\n 185.05,\n 45.98\n ],\n \"formula_id\": \"formula_13\",\n \"formula_text\": \"= softmax(p), p = (α q ⊗ α k ) • (b aq • b a k )/ √ d = α q;k • (b aq • b a k )/ √ d,(9)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 106.18,\n 570.36,\n 180.18,\n 40.45\n ],\n \"formula_id\": \"formula_14\",\n \"formula_text\": \"ãout = (αA • bA) • (αv • ba v ) = (αA ⊗ αv) • (bA • b av ) = αA;v • (bA • b av ),(10)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 50.11,\n 638.72,\n 236.25,\n 26.6\n ],\n \"formula_id\": \"formula_15\",\n \"formula_text\": \"α {q,k,v,A} = {α 1 {q,k,v,A} , α 2 {q,k,v,A} , • • • , α h {q,k,v,A} } ∈ R h + .\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 50.11,\n 688.99,\n 236.25,\n 26.11\n ],\n \"formula_id\": \"formula_16\",\n \"formula_text\": \"α q;k = {α 1 q;k , α 2 q;k , • • • , α h q;k } ∈ R h + and α A;v = {α 1 A;v , α 2 A;v , • • • , α h A;v } ∈ R h + .\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 310.06,\n 104.28,\n 252.5,\n 59.27\n ],\n \"formula_id\": \"formula_17\",\n \"formula_text\": \"∂ à ∂a h i ,n,c q = Ãh i ,n,n (1 -Ãh i ,n,n ) ∂ à ∂p h i ,n,n • α h i q;k • b h i ,c,n a k ∂p h i ,n,n ∂b h i ,n,c aq • 1 |a h i ,n,c q |≤αq ∂b h i ,n,c aq ∂a h i ,n,c q . (11\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 541.38,\n 155.78,\n 3.73,\n 7.77\n ],\n \"formula_id\": \"formula_18\",\n \"formula_text\": \")\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 349.59,\n 422.92,\n 195.52,\n 30.55\n ],\n \"formula_id\": \"formula_19\",\n \"formula_text\": \"L ranking = L l=1 ψ(A T ) -ψ(A S ) 2 ,(12)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 315.14,\n 518.42,\n 229.97,\n 39.48\n ],\n \"formula_id\": \"formula_20\",\n \"formula_text\": \"ψ(A :,n,: ) = A :,n,: -A :,n-1,: , if 0 < n ≤ N -1 A :,0,: -A :,N -1,: , otherwise .(13)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 377.57,\n 617.26,\n 167.54,\n 9.65\n ],\n \"formula_id\": \"formula_21\",\n \"formula_text\": \"L = L dist + λL ranking ,(14)\"\n }\n]"}}},{"rowIdx":399,"cells":{"doi":{"kind":"string","value":""},"pub_date":{"kind":"string","value":"2023-05-21"},"sections":{"kind":"list like","value":[{"figure_ref":["fig_0"],"heading":"Introduction","publication_ref":[],"table_ref":[],"text":"In the past decade, the development of information technology has led to a shift in the main carrier of information from text to images and then to videos. Moreover, with the rise of User-generated Content (UGC), the producer of information has shifted from Occupationally-generated Content (OGC) to UGC. As a result, a large number of videos have emerged on social media platforms and have been widely shared, leading to the increasingly important and challenging problems of video copyright protection. The core challenges of video copy detection are twofold: effective video descriptors and computational costs. Copied videos often involve edited portions, making it difficult for general visual models to differentiate between copied and original content. A powerful model must be capable of discriminating between videos, even when significant editing has taken place. Additionally, the cost of time and resources required to process each query video and identify the most similar reference video is a significant concern, necessitating the development of efficient and cost-effective methods.\nIn this paper, we summarize our proposed method for Meta AI Video Similarity Challenge, which tackle the core challenges of video copy detection through a dual-level detection method. In Fig. 1, there are three typical situations of query videos, including unedited video, copied video with general editing operation and copied video with multiple scenes in each frame. Our proposed dual-level detection method first identify if the video has been edited in video-level. For unedited videos, we use random vectors with small norm as their descriptors. What's more, we replace the bias term of these descriptors with a negative value during score normalization. For edited videos, we notice that it is necessary to deal with the situation that multiple scenes are concatenated along edge. We adopt traditional image processing method to detect and split the scenes in one frame. With our dual-level detection method, we can reduce the storage cost for unedited videos and improve the efficiency and accuracy of retrieval.\nThe main contributions are summarized as follows:\n• We propose a dual-level detection method for Descriptor Track, which detects edited videos at video-level and multiple scenes at frame-level. With the dual-level detection, we can reduce the computational cost and improve the performance.\n• The proposed method achieve outstanding performance on Meta AI Video Similarity Challenge and we got second prize on Descriptor Track. Our ablation study shows the effective of each module. For video with general editing, we extract features of each frame by our pre-trained basic model as descriptors. For video with multiple scenes, we detect and split the scenes in each frame, then use basic model to generate descriptors."},{"figure_ref":[],"heading":"Video Editing Detection","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Method","publication_ref":[],"table_ref":[],"text":"In this section, we first introduce the design of our basic model. Then we explain the details of our proposed dualdetection method including video editing detection, frame scenes detection and score filter normalization."},{"figure_ref":[],"heading":"Basic Model","publication_ref":["b7","b0","b7"],"table_ref":[],"text":"We train a basic model to extract descriptors for video copy detection. There are two potential types of descriptors that can be employed in video copy detection: videolevel features or frame-level features. Given that our objective is not only to discriminate copied videos but also to identify copied portions between query and reference videos, we selected frame-level features as the video descriptor. Therefore, we adopt image transformer [4, 7] as backbone. We follow SSCD [8] to train our basic model in a self-supervised manner. As SSCD uses, we combine SimCLR [1] method with entropy loss [10]. The InfoNCE Loss. We use the InfoNCE loss in SimCLR, which is softmax cross-entropy loss with temprature. The loss function is formulated as follow:\nL InfoNCE = - 1 |P | i,j∈P log exp(cos(z i , z j )/τ ) Σ k =i exp(cos(z i , z k )/τ ) (1)\nWhere P is the set of positive pairs, z i represents the descriptor, τ is the temperature. The Entropy Loss. We follow SSCD using the entropy loss proposed in [10]. The loss function is formulated as:\nL KoLeo = - 1 N N i=1 log(min j =i ||z i -z j ||)(2)\nWhere N is the size of training set.\nThe Final Loss. The final loss function is:\nL = L InfoNCE + λL KoLeo (3\n)\nWhere λ is the weights of Entropy Loss term. Cause our training process of basic model is based on SSCD [8], more details can be found in this paper."},{"figure_ref":[],"heading":"Video Editing Detection","publication_ref":["b8"],"table_ref":[],"text":"To address the issue of high computational costs in video copy detection and provide an efficient solution, we propose a straightforward method to identify edited videos before generating frame-level descriptors. We observe that videos with copies are often edited, incorporating techniques such as blending, blurring, rotations, and other manipulations. This is due to the fact that copied videos must frequently merge multiple clips, necessitating additional editing operations. By filtering out videos that have not been edited among the query videos, we can reduce computational costs. To achieve this goal, we have developed a model capable of discriminating between edited and unedited videos. Since editing operations can be viewed as strong augmentations, we aim to identify videos with such augmentations using a binary classification approach. We utilize CLIP [9] to extract frame features without any post-processing and feed these features into RoBERTa [6]. And we employ the class token to calculate cross entropy. We find that the edited video detection can achieve high accuracy, and using a small value α as threshold can filter most of unedited videos. For the unedited video, we use a random vector with very small value as descriptor. This processing can reduce the storage cost for query videos and speed up searching."},{"figure_ref":["fig_2"],"heading":"Frame Scenes Detection","publication_ref":[],"table_ref":[],"text":"We notice that stacking multiple scenes in one frames is an obvious augmentation of copied videos, and simple traditional image processing method can deal with this situation well. Due to the continuity of the video, the combination of multiple scenes in one frames are limited. As shown in Fig. 3, scenes are usually concatenated along one side and one frame usually has an even number of scenes, most of them have two or four scenes. "},{"figure_ref":[],"heading":"Score Filter Normalization","publication_ref":[],"table_ref":[],"text":"We follow [5, 8] using similarity normalization in our evaluation. It introduce a background image dataset and only queries whose similarity score with reference is much higher than images in background dataset will have high scores. Based on it, we modify the integrated bias to suppress the score of unedited videos. In Sec. 2.2, we use a random vector with very small value as descriptor for unedited video, but scores of these videos are clustered around 0. Because scores of hard positive pairs are clustered around 0 too, we should further suppress the score of unedited videos. Inspired by the integrated bias term of similarity normalization, we can replace it by a negative value and the similarity score with any reference videos will reduce to a negative value."},{"figure_ref":[],"heading":"Experiments","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Implementation Details","publication_ref":[],"table_ref":[],"text":"In video editing detection, we adopt ViT-L-16 of CLIP to extract frame features. The initial weights for RoBERTa is chinese-roberta-wwm-ext [2, 3] in huggingface."},{"figure_ref":[],"heading":"Results","publication_ref":[],"table_ref":[],"text":"Our proposed method achieve outstanding performance on Meta AI Video Similarity Challenge. The results of Phase 1 and Phase 2 on Descriptor Tracks are presented in Tab. 1. On Descriptor Track, our method got second place Phases 1, just 0.021 away from the first place. Although we got the first place on Phase 2, we notice that the performance drop a lot when transfer the model to Phase 2. The reason is that we only ensemble 4 models and our ensemble results are not much better than single model. Without a strong ensemble method, the transfer ability is limited. "},{"figure_ref":[],"heading":"User or teams","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Ablation Study","publication_ref":[],"table_ref":[],"text":"To validate the effective of our proposed method, we split validation set by ourselves and analyze each module on validation set. At the beginning of the competition, we randomly divide queries in training set into 8:2 as offline training set and validation set. And the trend of performance on validation set can reflect the trend on test set. We use single basic model in ablation study because our ensemble method do not improve much. The results are shown in Tab. 2, our basic model can achieve 0.8580 on µAP , it shows that our basic model is a very strong baseline for video copy detection. Then combining frame scenes detection with basic model, the performance increased by 5%. And with the video editing detection and frame scenes detection, the performance achieve 0.9492. "},{"figure_ref":[],"heading":"Conclusion","publication_ref":[],"table_ref":[],"text":"We introduce a dual-detection method for Video Copy Detection in this paper. The video editing detection in video-level can identify unedited videos and use random vectors with small norm and negative bias term as descriptors. The frame scenes detection in frame-level can detect scenes and split them into multiple videos, where video only have one scenes in each frame. Thought the dual-detection method, we got second place on the Descriptor Track of Meta AI Video Similarity Challenge 2022. And our descriptor give strong support to our first-place solution on Matching Track."}],"string":"[\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"Introduction\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In the past decade, the development of information technology has led to a shift in the main carrier of information from text to images and then to videos. Moreover, with the rise of User-generated Content (UGC), the producer of information has shifted from Occupationally-generated Content (OGC) to UGC. As a result, a large number of videos have emerged on social media platforms and have been widely shared, leading to the increasingly important and challenging problems of video copyright protection. The core challenges of video copy detection are twofold: effective video descriptors and computational costs. Copied videos often involve edited portions, making it difficult for general visual models to differentiate between copied and original content. A powerful model must be capable of discriminating between videos, even when significant editing has taken place. Additionally, the cost of time and resources required to process each query video and identify the most similar reference video is a significant concern, necessitating the development of efficient and cost-effective methods.\\nIn this paper, we summarize our proposed method for Meta AI Video Similarity Challenge, which tackle the core challenges of video copy detection through a dual-level detection method. In Fig. 1, there are three typical situations of query videos, including unedited video, copied video with general editing operation and copied video with multiple scenes in each frame. Our proposed dual-level detection method first identify if the video has been edited in video-level. For unedited videos, we use random vectors with small norm as their descriptors. What's more, we replace the bias term of these descriptors with a negative value during score normalization. For edited videos, we notice that it is necessary to deal with the situation that multiple scenes are concatenated along edge. We adopt traditional image processing method to detect and split the scenes in one frame. With our dual-level detection method, we can reduce the storage cost for unedited videos and improve the efficiency and accuracy of retrieval.\\nThe main contributions are summarized as follows:\\n• We propose a dual-level detection method for Descriptor Track, which detects edited videos at video-level and multiple scenes at frame-level. With the dual-level detection, we can reduce the computational cost and improve the performance.\\n• The proposed method achieve outstanding performance on Meta AI Video Similarity Challenge and we got second prize on Descriptor Track. Our ablation study shows the effective of each module. For video with general editing, we extract features of each frame by our pre-trained basic model as descriptors. For video with multiple scenes, we detect and split the scenes in each frame, then use basic model to generate descriptors.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Video Editing Detection\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Method\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this section, we first introduce the design of our basic model. Then we explain the details of our proposed dualdetection method including video editing detection, frame scenes detection and score filter normalization.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Basic Model\",\n \"publication_ref\": [\n \"b7\",\n \"b0\",\n \"b7\"\n ],\n \"table_ref\": [],\n \"text\": \"We train a basic model to extract descriptors for video copy detection. There are two potential types of descriptors that can be employed in video copy detection: videolevel features or frame-level features. Given that our objective is not only to discriminate copied videos but also to identify copied portions between query and reference videos, we selected frame-level features as the video descriptor. Therefore, we adopt image transformer [4, 7] as backbone. We follow SSCD [8] to train our basic model in a self-supervised manner. As SSCD uses, we combine SimCLR [1] method with entropy loss [10]. The InfoNCE Loss. We use the InfoNCE loss in SimCLR, which is softmax cross-entropy loss with temprature. The loss function is formulated as follow:\\nL InfoNCE = - 1 |P | i,j∈P log exp(cos(z i , z j )/τ ) Σ k =i exp(cos(z i , z k )/τ ) (1)\\nWhere P is the set of positive pairs, z i represents the descriptor, τ is the temperature. The Entropy Loss. We follow SSCD using the entropy loss proposed in [10]. The loss function is formulated as:\\nL KoLeo = - 1 N N i=1 log(min j =i ||z i -z j ||)(2)\\nWhere N is the size of training set.\\nThe Final Loss. The final loss function is:\\nL = L InfoNCE + λL KoLeo (3\\n)\\nWhere λ is the weights of Entropy Loss term. Cause our training process of basic model is based on SSCD [8], more details can be found in this paper.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Video Editing Detection\",\n \"publication_ref\": [\n \"b8\"\n ],\n \"table_ref\": [],\n \"text\": \"To address the issue of high computational costs in video copy detection and provide an efficient solution, we propose a straightforward method to identify edited videos before generating frame-level descriptors. We observe that videos with copies are often edited, incorporating techniques such as blending, blurring, rotations, and other manipulations. This is due to the fact that copied videos must frequently merge multiple clips, necessitating additional editing operations. By filtering out videos that have not been edited among the query videos, we can reduce computational costs. To achieve this goal, we have developed a model capable of discriminating between edited and unedited videos. Since editing operations can be viewed as strong augmentations, we aim to identify videos with such augmentations using a binary classification approach. We utilize CLIP [9] to extract frame features without any post-processing and feed these features into RoBERTa [6]. And we employ the class token to calculate cross entropy. We find that the edited video detection can achieve high accuracy, and using a small value α as threshold can filter most of unedited videos. For the unedited video, we use a random vector with very small value as descriptor. This processing can reduce the storage cost for query videos and speed up searching.\"\n },\n {\n \"figure_ref\": [\n \"fig_2\"\n ],\n \"heading\": \"Frame Scenes Detection\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We notice that stacking multiple scenes in one frames is an obvious augmentation of copied videos, and simple traditional image processing method can deal with this situation well. Due to the continuity of the video, the combination of multiple scenes in one frames are limited. As shown in Fig. 3, scenes are usually concatenated along one side and one frame usually has an even number of scenes, most of them have two or four scenes. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Score Filter Normalization\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We follow [5, 8] using similarity normalization in our evaluation. It introduce a background image dataset and only queries whose similarity score with reference is much higher than images in background dataset will have high scores. Based on it, we modify the integrated bias to suppress the score of unedited videos. In Sec. 2.2, we use a random vector with very small value as descriptor for unedited video, but scores of these videos are clustered around 0. Because scores of hard positive pairs are clustered around 0 too, we should further suppress the score of unedited videos. Inspired by the integrated bias term of similarity normalization, we can replace it by a negative value and the similarity score with any reference videos will reduce to a negative value.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiments\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Implementation Details\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In video editing detection, we adopt ViT-L-16 of CLIP to extract frame features. The initial weights for RoBERTa is chinese-roberta-wwm-ext [2, 3] in huggingface.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Our proposed method achieve outstanding performance on Meta AI Video Similarity Challenge. The results of Phase 1 and Phase 2 on Descriptor Tracks are presented in Tab. 1. On Descriptor Track, our method got second place Phases 1, just 0.021 away from the first place. Although we got the first place on Phase 2, we notice that the performance drop a lot when transfer the model to Phase 2. The reason is that we only ensemble 4 models and our ensemble results are not much better than single model. Without a strong ensemble method, the transfer ability is limited. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"User or teams\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Ablation Study\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"To validate the effective of our proposed method, we split validation set by ourselves and analyze each module on validation set. At the beginning of the competition, we randomly divide queries in training set into 8:2 as offline training set and validation set. And the trend of performance on validation set can reflect the trend on test set. We use single basic model in ablation study because our ensemble method do not improve much. The results are shown in Tab. 2, our basic model can achieve 0.8580 on µAP , it shows that our basic model is a very strong baseline for video copy detection. Then combining frame scenes detection with basic model, the performance increased by 5%. And with the video editing detection and frame scenes detection, the performance achieve 0.9492. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We introduce a dual-detection method for Video Copy Detection in this paper. The video editing detection in video-level can identify unedited videos and use random vectors with small norm and negative bias term as descriptors. The frame scenes detection in frame-level can detect scenes and split them into multiple videos, where video only have one scenes in each frame. Thought the dual-detection method, we got second place on the Descriptor Track of Meta AI Video Similarity Challenge 2022. And our descriptor give strong support to our first-place solution on Matching Track.\"\n }\n]"},"abstract":{"kind":"string","value":"Copy Detection has been a crucial problem for social media platforms. Meta AI hold Video Similarity Challenge on CVPR 2023 to push the technology forward. In this paper, we share our winner solutions on both tracks to help progress in this area. For Descriptor Track, we propose a dual-level detection method with Video Editing Detection (VED) and Frame Scenes Detection (FSD) to tackle the core challenges on Video Copy Detection. Experimental results demonstrate the effectiveness and efficiency of our proposed method. Code is available at https://github."},"title":{"kind":"string","value":"A Dual-level Detection Method for Video Copy Detection"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 1 .1Figure 1. Three typical situations of query videos. (a) is an unedited video, which is the most of query videos. (b) is a copied video with general editing operation. (c) is a copied video with multiple scenes in each frame.","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"Figure 2 .2Figure2. Overview of our proposed pipeline. For unedited video, we directly use a random vector with samll norm and negative bias term. For video with general editing, we extract features of each frame by our pre-trained basic model as descriptors. For video with multiple scenes, we detect and split the scenes in each frame, then use basic model to generate descriptors.","figure_data":"","figure_id":"fig_1","figure_label":"2","figure_type":"figure"},{"figure_caption":"Figure 3 .3Figure 3. Multiple scenes in one frame.","figure_data":"","figure_id":"fig_2","figure_label":"3","figure_type":"figure"},{"figure_caption":"Leaderboard results on Descriptor Track. Bold indicates the best result and underline indicates the second best result.","figure_data":"Phase 1 µAP Phase 2 µAPdo something(Ours)0.91760.8717FriendshipFirst0.91970.8514cvl-descriptor0.85340.8362Zihao0.78410.7729","figure_id":"tab_0","figure_label":"1","figure_type":"table"},{"figure_caption":"Ablation study.","figure_data":"MethodµAPBasic model 0.8580+ FSD0.9075+ VED0.9492","figure_id":"tab_1","figure_label":"2","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 1 .1Figure 1. Three typical situations of query videos. (a) is an unedited video, which is the most of query videos. (b) is a copied video with general editing operation. (c) is a copied video with multiple scenes in each frame.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 2 .2Figure2. Overview of our proposed pipeline. For unedited video, we directly use a random vector with samll norm and negative bias term. For video with general editing, we extract features of each frame by our pre-trained basic model as descriptors. For video with multiple scenes, we detect and split the scenes in each frame, then use basic model to generate descriptors.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 .3Figure 3. Multiple scenes in one frame.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Leaderboard results on Descriptor Track. Bold indicates the best result and underline indicates the second best result.\",\n \"figure_data\": \"Phase 1 µAP Phase 2 µAPdo something(Ours)0.91760.8717FriendshipFirst0.91970.8514cvl-descriptor0.85340.8362Zihao0.78410.7729\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Ablation study.\",\n \"figure_data\": \"MethodµAPBasic model 0.8580+ FSD0.9075+ VED0.9492\",\n \"figure_id\": \"tab_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n }\n]"},"authors":{"kind":"string","value":"Tianyi Wang; Feipeng Ma; Zhenhua Liu; Fengyun Rao"},"references":{"kind":"list like","value":[{"authors":"Ting Chen; Simon Kornblith; Mohammad Norouzi; Geoffrey Hinton","journal":"PMLR","ref_id":"b0","title":"A simple framework for contrastive learning of visual representations","year":"2020"},{"authors":"Yiming Cui; Wanxiang Che; Ting Liu; Bing Qin; Shijin Wang; Guoping Hu","journal":"","ref_id":"b1","title":"Revisiting pre-trained models for Chinese natural language processing","year":"2020-11"},{"authors":"Yiming Cui; Wanxiang Che; Ting Liu; Bing Qin; Ziqing Yang; Shijin Wang; Guoping Hu","journal":"","ref_id":"b2","title":"Pre-training with whole word masking for chinese bert","year":"2019"},{"authors":"Alexey Dosovitskiy; Lucas Beyer; Alexander Kolesnikov; Dirk Weissenborn; Xiaohua Zhai; Thomas Unterthiner; Mostafa Dehghani; Matthias Minderer; Georg Heigold; Sylvain Gelly","journal":"","ref_id":"b3","title":"An image is worth 16x16 words: Transformers for image recognition at scale","year":"2020"},{"authors":"Matthijs Douze; Giorgos Tolias; Ed Pizzi; Zoë Papakipos; Lowik Chanussot; Filip Radenovic; Tomas Jenicek; Maxim Maximov; Laura Leal-Taixé; Ismail Elezi","journal":"","ref_id":"b4","title":"The 2021 image similarity dataset and challenge","year":"2021"},{"authors":"Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov","journal":"","ref_id":"b5","title":"Roberta: A robustly optimized bert pretraining approach","year":"2019"},{"authors":"Ze Liu; Han Hu; Yutong Lin; Zhuliang Yao; Zhenda Xie; Yixuan Wei; Jia Ning; Yue Cao; Zheng Zhang; Li Dong; Furu Wei; Baining Guo","journal":"","ref_id":"b6","title":"Swin transformer v2: Scaling up capacity and resolution","year":"2022"},{"authors":"Ed Pizzi; Dutta Sreya; Sugosh Roy; Priya Nagavara Ravindra; Matthijs Goyal; Douze","journal":"","ref_id":"b7","title":"A self-supervised descriptor for image copy detection","year":"2022"},{"authors":"Alec Radford; Jong Wook Kim; Chris Hallacy; Aditya Ramesh; Gabriel Goh; Sandhini Agarwal; Girish Sastry; Amanda Askell; Pamela Mishkin; Jack Clark","journal":"PMLR","ref_id":"b8","title":"Learning transferable visual models from natural language supervision","year":"2021"},{"authors":"Alexandre Sablayrolles; Matthijs Douze; Cordelia Schmid; Hervé Jégou","journal":"","ref_id":"b9","title":"Spreading vectors for similarity search","year":"2018"}],"string":"[\n {\n \"authors\": \"Ting Chen; Simon Kornblith; Mohammad Norouzi; Geoffrey Hinton\",\n \"journal\": \"PMLR\",\n \"ref_id\": \"b0\",\n \"title\": \"A simple framework for contrastive learning of visual representations\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Yiming Cui; Wanxiang Che; Ting Liu; Bing Qin; Shijin Wang; Guoping Hu\",\n \"journal\": \"\",\n \"ref_id\": \"b1\",\n \"title\": \"Revisiting pre-trained models for Chinese natural language processing\",\n \"year\": \"2020-11\"\n },\n {\n \"authors\": \"Yiming Cui; Wanxiang Che; Ting Liu; Bing Qin; Ziqing Yang; Shijin Wang; Guoping Hu\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"Pre-training with whole word masking for chinese bert\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Alexey Dosovitskiy; Lucas Beyer; Alexander Kolesnikov; Dirk Weissenborn; Xiaohua Zhai; Thomas Unterthiner; Mostafa Dehghani; Matthias Minderer; Georg Heigold; Sylvain Gelly\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"An image is worth 16x16 words: Transformers for image recognition at scale\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Matthijs Douze; Giorgos Tolias; Ed Pizzi; Zoë Papakipos; Lowik Chanussot; Filip Radenovic; Tomas Jenicek; Maxim Maximov; Laura Leal-Taixé; Ismail Elezi\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"The 2021 image similarity dataset and challenge\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov\",\n \"journal\": \"\",\n \"ref_id\": \"b5\",\n \"title\": \"Roberta: A robustly optimized bert pretraining approach\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Ze Liu; Han Hu; Yutong Lin; Zhuliang Yao; Zhenda Xie; Yixuan Wei; Jia Ning; Yue Cao; Zheng Zhang; Li Dong; Furu Wei; Baining Guo\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"Swin transformer v2: Scaling up capacity and resolution\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Ed Pizzi; Dutta Sreya; Sugosh Roy; Priya Nagavara Ravindra; Matthijs Goyal; Douze\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"A self-supervised descriptor for image copy detection\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Alec Radford; Jong Wook Kim; Chris Hallacy; Aditya Ramesh; Gabriel Goh; Sandhini Agarwal; Girish Sastry; Amanda Askell; Pamela Mishkin; Jack Clark\",\n \"journal\": \"PMLR\",\n \"ref_id\": \"b8\",\n \"title\": \"Learning transferable visual models from natural language supervision\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Alexandre Sablayrolles; Matthijs Douze; Cordelia Schmid; Hervé Jégou\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"Spreading vectors for similarity search\",\n \"year\": \"2018\"\n }\n]"},"formulas":{"kind":"list like","value":[{"formula_coordinates":[2,60.79,561.08,225.57,37.74],"formula_id":"formula_0","formula_text":"L InfoNCE = - 1 |P | i,j∈P log exp(cos(z i , z j )/τ ) Σ k =i exp(cos(z i , z k )/τ ) (1)"},{"formula_coordinates":[2,88.78,661.4,197.58,30.32],"formula_id":"formula_1","formula_text":"L KoLeo = - 1 N N i=1 log(min j =i ||z i -z j ||)(2)"},{"formula_coordinates":[2,374.11,304.2,167.13,9.65],"formula_id":"formula_2","formula_text":"L = L InfoNCE + λL KoLeo (3"},{"formula_coordinates":[2,541.24,304.52,3.87,8.64],"formula_id":"formula_3","formula_text":")"}],"string":"[\n {\n \"formula_coordinates\": [\n 2,\n 60.79,\n 561.08,\n 225.57,\n 37.74\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"L InfoNCE = - 1 |P | i,j∈P log exp(cos(z i , z j )/τ ) Σ k =i exp(cos(z i , z k )/τ ) (1)\"\n },\n {\n \"formula_coordinates\": [\n 2,\n 88.78,\n 661.4,\n 197.58,\n 30.32\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"L KoLeo = - 1 N N i=1 log(min j =i ||z i -z j ||)(2)\"\n },\n {\n \"formula_coordinates\": [\n 2,\n 374.11,\n 304.2,\n 167.13,\n 9.65\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"L = L InfoNCE + λL KoLeo (3\"\n },\n {\n \"formula_coordinates\": [\n 2,\n 541.24,\n 304.52,\n 3.87,\n 8.64\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \")\"\n }\n]"}}}],"truncated":false,"partial":false},"paginationData":{"pageIndex":3,"numItemsPerPage":100,"numTotalItems":12452,"offset":300,"length":100}},"jwt":"eyJhbGciOiJFZERTQSJ9.eyJyZWFkIjp0cnVlLCJwZXJtaXNzaW9ucyI6eyJyZXBvLmNvbnRlbnQucmVhZCI6dHJ1ZX0sImlhdCI6MTczMTgyNzk1Mywic3ViIjoiL2RhdGFzZXRzL0JhY2hOZ29IL1BhcnNlZEFyeGl2UGFwZXJzXzEyayIsImV4cCI6MTczMTgzMTU1MywiaXNzIjoiaHR0cHM6Ly9odWdnaW5nZmFjZS5jbyJ9.3e1zJF5LvD_2IjTwupahOlX0C-4U-0eYwL2kBQd0AVWDtJn_WRo-KA-h6Nz2zif4ctgfB42-y3P-GPKPwv0xAQ","displayUrls":true},"discussionsStats":{"closed":0,"open":1,"total":1},"fullWidth":true,"hasGatedAccess":true,"hasFullAccess":true,"isEmbedded":false}">
doi
stringlengths 0
570
| pub_date
stringclasses 355
values | sections
listlengths 1
245
| abstract
stringlengths 0
5.25k
| title
stringlengths 0
228
| figures
listlengths 0
130
| authors
stringlengths 0
11.9k
| references
listlengths 0
835
| formulas
listlengths 0
679
|
|---|---|---|---|---|---|---|---|---|
10.48550/arXiv.2303.13560 | [
{
"figure_ref": [
"fig_0"
],
"heading": "",
"publication_ref": [
"b6",
"b7",
"b8",
"b10",
"b11"
],
"table_ref": [],
"text": "ECOS), Ingolstadt, Germany, e-mail: [email protected]. and an enhanced service level, compared to that of the conventional ITS. As illustrated in Fig. 1, road participants -specifically, connected automated vehicles (CAVs)can share information with one another through vehicleto-everything (V2X) networks, which encompass vehicleto-vehicle (V2V), vehicle-to-infrastructure (V2I), vehicle-tonetwork(V2N), and infrastructure-to-network (I2N).\nThe European standards for C-ITS define several types of V2X messages to facilitate decentralized information sharing. Specifically, for cooperative awareness and perception, dedicated message types -the Cooperative Awareness Message (CAM) and the Collective Perception Message (CPM) -are periodically exchanged among CAVs and with roadside infrastructure [7]. By sending and receiving V2X messages, enriched and improved environmental data of road traffic can be made available within vehicular networks.\nIn centralized model training using ML, CAV clients transmit data to a centralized system through vehicle-to-network communications. This process can generate an enormous volume of data, potentially exceeding the network's capacity. Moreover, data collected from CAV clients for ML model training cannot be directly shared due to privacy concerns. Differing from conventional ML, federated learning (FL) trains ML models using data from distributed systems, such as devices or clients, without centralizing the data [8]. In FL, connected clients share a model trained on their local data with a server, which aggregates the local models and updates The deployment of 5G-V2X vehicular networks has further facilitated the use of FL in C-ITS by providing higher data rates and greater reliability for data exchange. This allows for the training of larger ML models for C-ITS applications and services, such as [9]- [11]. Although FL has great potential to preserve privacy and utilize a broader range of data resources [12], the employment of FL in C-ITS has to address major challenges due to heterogeneity in data and networks, which can not only limit the performance but also lead to FL failures. Data heterogeneity. Data across clients is non-iid (nonidentically independently distributed), resulting from various sensor types, combinations, poses, road scenes, traffic scenarios, climate and weather conditions, and more."
},
{
"figure_ref": [],
"heading": "Network heterogeneity. The diverse connection qualities of clients can slow down model sharing and cause communication delays for global model aggregation, which impedes the FL process.",
"publication_ref": [],
"table_ref": [],
"text": "To address these challenges and enhance the application of FL in C-ITS, we propose a novel FL framework. The main idea is to select clients for upcoming communication rounds based on (i) the prediction of connection qualities in the context of road traffic status and (ii) the similarity of local data distribution in clients. "
},
{
"figure_ref": [
"fig_1"
],
"heading": "III. F",
"publication_ref": [
"b19"
],
"table_ref": [],
"text": "Our contextual client selection framework is illustrated in Fig. 2, comprising V2X information sharing, traffic topology prediction, data-level client clustering, and network-level client clustering. V2X message fusion. We first fuse V2X messages. Continuously receiving CAM and CPM enables dynamic road maps with traffic object states. Road-side infrastructure collects and forwards V2X messages to a server via V2I and I2N networks. The server filters and fuses messages, obtaining traffic object states, such as position, speed, and acceleration. Fused results form an road traffic topology graph (RTTG), with each CAV characterized by a node with attributes. The RTTG digitizes C-ITS and recreates vehicular networks virtually. RTTG prediction. We predict future RTTGs. After V2X message fusion, we initialize a prediction instance for each CAV to estimate its trajectory. Predicted trajectories build future RTTGs. Predicted RTTGs integrate with digital C-ITS, providing possible connection quality for each CAV. We simulate networks in digital twin and calculate FL communication latency based on predictive transport scenarios. Data-level client grouping. We cluster clients into groups considering data heterogeneity. Our goal is to group clients with similar data distribution, ensuring each subset represents the whole group's data features. We observe model updates, considering gradient similarity as a data similarity criterion [20]. We group clients based on model parameter similarity. Clients must report gradient updates before a deadline for inclusion in data-level client grouping. After grouping, each subset represents its cluster. Selecting at least one client per cluster ensures satisfactory training performance. Network-level client election. We elect clients in each group based on contextual communication latency. Using predicted RTTG latency, we determine efficient client contributions for upcoming communication rounds. We employ the Fast-𝛾 rule, selecting the 𝛾 clients with the lowest communication delay (0 < 𝛾 < 1) per cluster.\nThrough these stages, representative clients with minimal contextual communication latency are chosen for model aggregation. This process increases FL communication efficiency by optimizing communication rounds and round duration. De-selected clients save computational resources by not training models locally."
},
{
"figure_ref": [],
"heading": "IV. P",
"publication_ref": [],
"table_ref": [],
"text": "We implement and demonstrate FL with our pipeline as well as other four baselines, i.e. greedy, gossip, databased and network-based client selection strategy, on a computer cluster with 4× NVIDIA-A100-PCIE-40GB GPUs and 4× 32-Core-AMD-EPYC-7513 CPUs. The environment is a Linux system with Pytorch 1.8.1 and Cuda 11.1."
},
{
"figure_ref": [],
"heading": "A. Experiment setup",
"publication_ref": [
"b25",
"b26",
"b27"
],
"table_ref": [],
"text": "We conduct the experimental evaluation by training models on three widely used open datasets MNIST [26], CIFAR-10 [27] and SVHN [28] distributed into 100 CAV clients in non-iid setting.\nIn our default non-iid setting, each client owns only 2 out of 10 classes. We compare our pipeline with four other client selection strategies as baselines, i.e., greedy, gossip, data-based and network-based, as described in Sec. II. The learning rate is 0.001 and the batch size is 64. The number of the local epochs is set as 3 for training on MNIST, and 1 for training on CIFAR-10 and SVHN, respectively. Except the greedy strategy (all clients are selected in each communication round), the general selection rate for FL clients is defined as 10%, i.e. around 10 clients are selected in each communication round."
},
{
"figure_ref": [
"fig_3"
],
"heading": "B. Performance results",
"publication_ref": [],
"table_ref": [],
"text": "We show the general performance of FL with contextual client selection for training models on three datasets distributed in 100 vehicle clients with respect to default noniid setting. We train deep learning models with different sizes on MNIST, CIFAR-10 and SVHN as FL tasks. As the experimental results in Fig. 3 show for all three tasks, FL with our contextual client selection can outperform the other four baselines. Generally, the FL with contextual client selection can achieve remarkable higher test accuracy than the other four strategies. Even though the network-based strategy allows the ML-model to be trained to a comparable test accuracy on SVHN, the contextual client selection re-sults showcases much more stable convergence, as the data heterogeneity across CAVs are taken into account.\nWe conduct the experiments with various connection rates and evaluate the performance of FL. We take the required time to reach 0.5 of test accuracy for FL with gossip client selection as a baseline, and evaluate the time reduction rate of FL with other strategies. As the comparison results show in Tab. I, FL with contextual client selection always needs less time than other two strategies at each connection rate. The time reduction rates are robustly over 20× even when only 20% of clients are connected in networks."
},
{
"figure_ref": [],
"heading": "V. C",
"publication_ref": [
"b28",
"b30"
],
"table_ref": [],
"text": "In this work, we reviewed the existing client selection strategies for FL and introduced a novel four-stage V2X-Boosted FL pipeline for C-ITS. The approach tackles both data and network heterogeneity in vehicular networks, boosting communication efficiency by reducing the number of communication rounds and shortening the time required for each round. Compared to other strategies, FL with contextual client selection achieves higher accuracy and more stable convergence performance by leveraging V2X messages disseminated in vehicular networks. Future work will further consider the analytical model of communication networks and conduct more validation in traffic scenario data, such as [29]- [31]. R"
},
{
"figure_ref": [],
"heading": "",
"publication_ref": [],
"table_ref": [],
"text": "This work was supported by the German Federal Ministry for Digital and Transport (BMVI) in the projects \"KIVI -KI im Verkehr Ingolstadt\" and \"5GoIng -5G Innovation Concept Ingolstadt\"."
},
{
"figure_ref": [],
"heading": "VI. A TABLE II C",
"publication_ref": [],
"table_ref": [],
"text": "."
},
{
"figure_ref": [],
"heading": "Greedy Gossip Data-based Network-based Contextual (ours)",
"publication_ref": [],
"table_ref": [],
"text": "All clients should be selected in each FL communication round.\nA random subset of all connected clients is selected in each FL communication round.\nClients are selected according to the similarity of local data distribution.\nClients are selected according to the connection quality and availability in networks.\nClients are selected in consideration of both data and network heterogeneity. We also demonstrate various client selection strategies under different class ratio in each client to evaluate the performance in non-iid settings. We consider a scenario for training a ML-model with 100 CAVs for three minutes. Note that the class ratio 100% indicates iid setting. As Fig. 4 shows, the pure network-based strategy can make the test accuracy of FL higher than others, because the data heterogeneity is not needed to be considered under ideal iid settings. However, in non-iid settings, the contextual client selection enhances FL and leads to a better test accuracy than other strategies. For instances, when the class ratio is 20%, it achieves 2.85× test accuracy compared to a pure data-based and 1.67× to a pure network-based client selection strategy. Even in extremely non-iid setting with only 1 class in each client, the contextual client selection can reach 38% test accuracy within three minutes, while the FL with network-or data-based strategies cannot converge."
}
] | Machine learning (ML) has revolutionized transportation systems, enabling autonomous driving and smart traffic services. Federated learning (FL) overcomes privacy constraints by training ML models in distributed systems, exchanging model parameters instead of raw data. However, the dynamic states of connected vehicles affect the network connection quality and influence the FL performance. To tackle this challenge, we propose a contextual client selection pipeline that uses Vehicle-to-Everything (V2X) messages to select clients based on the predicted communication latency. The pipeline includes: (i) fusing V2X messages, (ii) predicting future traffic topology, (iii) pre-clustering clients based on local data distribution similarity, and (iv) selecting clients with minimal latency for future model aggregation. Experiments show that our pipeline outperforms baselines on various datasets, particularly in noniid settings.Machine learning (ML), a subfield of artificial intelligence, focuses on developing learning algorithms and inference models that enable digital systems to make decisions and predictions in terms of the knowledge learned from data. Over the past years, ML-based approaches exhibited great potential to revolutionize various scientific, engineering, economic, and cultural fields with outstanding technological advancements such as Google AlphaGo and Open AI's Chat-GPT. In the filed of road transportation, ML is possible to empower numerous new applications for realizing Intelligent Transportation System (ITS), e.g., environmental perception, road traffic flow optimization, and trajectory planning, which can significantly enhance the safety and efficiency of transportation systems [1]- [5]. Recently, a new ITS concept referred to as Cooperative Intelligent Transportation System (C-ITS) attacked a lot of interests from both academia and industry [6]. In C-ITS, the cooperation between two or more ITS sub-systems (personal, vehicle, roadside and central) offers better quality | V2X-Boosted Federated Learning for Cooperative Intelligent Transportation Systems with Contextual Client Selection | [
{
"figure_caption": "Fig. 1 .1Fig. 1. An overview of vehicular networks in C-ITS, including vehicleto-vehicle (V2V), vehicle-to-infrastructure (V2I), vehicle-to-network (V2N), and infrastructure-to-network (I2N) communication.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 2 .2Fig. 2. Contextual client selection pipeline: (1) V2X message fusion; (2) Road traffic topology graph (RTTG) prediction; (3) Data-level client grouping; (4) Network-level client selection.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "discuss greedy and gossip client selection in FL, dataand network-based strategies, and FL in vehicular networks considering road traffic features. Greedy and gossip client selection. FL, initially proposed by McMahan et al. [13], suffers from the straggler effect due to varying connection qualities [14]. Greedy client selection includes all clients in each communication round, while gossip (stochastic greedy) selection randomly selects connected clients. Both strategies struggle to avoid the straggler effect. Network-based client selection. Strategies focusing on network quality [15]-[18] reduce the straggler effect but aren't specifically designed for vehicular networks with dynamic connection qualities and high-priority traffic services. Inspired by [19], we optimize client selection by predicting communication latency in vehicular networks. Data-based client selection. Client selection based on data distribution tackles heterogeneity. The approaches in [20]-[25] consider data heterogeneity but overlook network parameters. Our work addresses both data distribution and network quality in vehicular networks. A comparison of the client selection paradigms is shown in Tab. II.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 3 .3Fig. 3. FL training performance (testing accuracy changes over time in seconds) using various client selection strategies on non-iid MNIST (left), CIFAR-10 (middle) and SVHN (right) data distributed in 100 clients.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "",
"figure_data": "R0.5(CR).CRStrategyTime (s)Reduction rate-Gossip3 891.141×Data-based213.5018.23×1.0Network-based620.476.27×Contextuall93.7720.08×Data-based2 446.751.59×0.5Network-based690.295.64×Contextuall79.5421.67×Data-based2 563.201.52×0.2Network-based634.126.14×Contextuall86.4720.87×",
"figure_id": "tab_0",
"figure_label": "I",
"figure_type": "table"
}
] | Rui Song; Lingjuan Lyu; Wei Jiang; Andreas Festag; Alois Knoll | [
{
"authors": "Y Hu",
"journal": "",
"ref_id": "b0",
"title": "Collaboration helps camera overtake LiDAR in 3D detection",
"year": "2023"
},
{
"authors": "R Xu; H Xiang; Z Tu; X Xia; M.-H Yang; J Ma",
"journal": "",
"ref_id": "b1",
"title": "V2X-ViT: Vehicle-to-everything cooperative perception with vision transformer",
"year": "2022"
},
{
"authors": "Y Li; Q Fang; J Bai; S Chen; F Juefei-Xu; C Feng",
"journal": "",
"ref_id": "b2",
"title": "Among us: Adversarially robust collaborative perception by consensus",
"year": "2023"
},
{
"authors": "Z Lei; S Ren; Y Hu; W Zhang; S Chen",
"journal": "EECCV",
"ref_id": "b3",
"title": "Latency-aware collaborative perception",
"year": "2022"
},
{
"authors": "R Xu",
"journal": "",
"ref_id": "b4",
"title": "CoBEVT: Cooperative bird's eye view semantic segmentation with sparse transformers",
"year": "2022"
},
{
"authors": "K Sjöberg; P Andres; T Buburuzan; A Brakemeier",
"journal": "IEEE Vehicular Technology Magazine",
"ref_id": "b5",
"title": "Cooperative Intelligent Transport Systems in Europe: Current deployment status and outlook",
"year": "2017"
},
{
"authors": "G Thandavarayan; M Sepulcre; J Gozalvez",
"journal": "IEEE Transactions on Vehicular Technology",
"ref_id": "b6",
"title": "Generation of cooperative perception messages for connected and automated vehicles",
"year": "2020"
},
{
"authors": "R Yu; P Li",
"journal": "IEEE Network",
"ref_id": "b7",
"title": "Toward resource-efficient federated learning in mobile edge computing",
"year": "2021"
},
{
"authors": "Y Li",
"journal": "",
"ref_id": "b8",
"title": "VoxFormer: Sparse voxel transformer for camerabased 3D semantic scene completion",
"year": "2023"
},
{
"authors": "Y Li; J Zhang; D Ma; Y Wang; C Feng",
"journal": "PMLR",
"ref_id": "b9",
"title": "Multi-robot scene completion: Towards task-agnostic collaborative perception",
"year": "2023"
},
{
"authors": "Y Hu; S Fang; Z Lei; Z Yiqi; C Siheng",
"journal": "",
"ref_id": "b10",
"title": "Where2comm: Communication-efficient collaborative perception via spatial confidence maps",
"year": "2022"
},
{
"authors": "J Posner; L Tseng; M Aloqaily; Y Jararweh",
"journal": "IEEE Network",
"ref_id": "b11",
"title": "Federated learning in vehicular networks: Opportunities and solutions",
"year": "2021"
},
{
"authors": "B Mcmahan; E Moore; D Ramage; S Hampson; B A Arcas",
"journal": "PMLR",
"ref_id": "b12",
"title": "Communication-efficient learning of deep networks from decentralized data",
"year": "2017"
},
{
"authors": "J Jin; J Ren; Y Zhou; L Lyu; J Liu; D Dou",
"journal": "PMLR",
"ref_id": "b13",
"title": "Accelerated federated learning with decoupled adaptive optimization",
"year": "2022"
},
{
"authors": "T Nishio; R Yonetani",
"journal": "",
"ref_id": "b14",
"title": "Client selection for federated learning with heterogeneous resources in mobile edge",
"year": "2019"
},
{
"authors": "S Abdulrahman; H Tout; A Mourad; C Talhi",
"journal": "IEEE Internet of Things Journal",
"ref_id": "b15",
"title": "FedMCCS: Multicriteria client selection model for optimal iot federated learning",
"year": "2020"
},
{
"authors": "M Chahoud; S Otoum; A Mourad",
"journal": "Information Processing & Management",
"ref_id": "b16",
"title": "On the feasibility of federated learning towards on-demand client deployment at the edge",
"year": "2023"
},
{
"authors": "J Xu; H Wang",
"journal": "IEEE Transactions on Wireless Communications",
"ref_id": "b17",
"title": "Client selection and bandwidth allocation in wireless federated learning networks: A long-term perspective",
"year": "2020"
},
{
"authors": "Y Fu",
"journal": "MDPI Electronics",
"ref_id": "b18",
"title": "Digital twin based network latency prediction in vehicular networks",
"year": "2022"
},
{
"authors": "D Yin",
"journal": "PMLR",
"ref_id": "b19",
"title": "Gradient diversity: a key ingredient for scalable distributed learning",
"year": "2018"
},
{
"authors": "W Zhang; X Wang; P Zhou; W Wu; X Zhang",
"journal": "IEEE Access",
"ref_id": "b20",
"title": "Client selection for federated learning with non-IID data in mobile edge computing",
"year": "2021"
},
{
"authors": "Y J Cho; J Wang; G Joshi",
"journal": "PMLR",
"ref_id": "b21",
"title": "Towards understanding biased client selection in federated learning",
"year": "2022"
},
{
"authors": "S K Shyn; D Kim; K Kim",
"journal": "IEEE Access",
"ref_id": "b22",
"title": "Empirical measurement of client contribution for federated learning with data size diversification",
"year": "2022"
},
{
"authors": "R Balakrishnan; T Li; T Zhou; N Himayat; V Smith; J Bilmes",
"journal": "",
"ref_id": "b23",
"title": "Diverse client selection for federated learning via submodular maximization",
"year": "2021"
},
{
"authors": "G Shen; D Gao; D Song; X Zhou; S Pan; W Lou; F Zhou",
"journal": "",
"ref_id": "b24",
"title": "Fast heterogeneous federated learning with hybrid client selection",
"year": "2022"
},
{
"authors": "Y Lecun; C Cortes; C Burges",
"journal": "",
"ref_id": "b25",
"title": "MNIST handwritten digit database",
"year": "2010"
},
{
"authors": "A Krizhevsky",
"journal": "",
"ref_id": "b26",
"title": "Learning multiple layers of features from tiny images",
"year": "2009"
},
{
"authors": "Y Netzer",
"journal": "",
"ref_id": "b27",
"title": "Reading digits in natural images with unsupervised feature learning",
"year": "2011"
},
{
"authors": "R Xu; H Xiang; X Xia; X Han; J Li; J Ma",
"journal": "IEEE",
"ref_id": "b28",
"title": "OPV2V: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication",
"year": "2022"
},
{
"authors": "Y Li",
"journal": "IEEE Robotics and Automation Letters",
"ref_id": "b29",
"title": "V2X-Sim: Multi-agent collaborative perception dataset and benchmark for autonomous driving",
"year": "2022"
},
{
"authors": "R Xu",
"journal": "",
"ref_id": "b30",
"title": "V2v4real: A real-world large-scale dataset for vehicleto-vehicle cooperative perception",
"year": "2023"
}
] | [] |
|
2023-05-19 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b2",
"b37",
"b17",
"b3",
"b27",
"b28",
"b26",
"b30",
"b52",
"b61",
"b62",
"b9",
"b21",
"b24",
"b58",
"b12",
"b43",
"b29",
"b1",
"b68",
"b78",
"b10",
"b38",
"b50",
"b49",
"b53",
"b15",
"b33",
"b39",
"b5",
"b4",
"b18",
"b19",
"b44",
"b48",
"b54",
"b67",
"b79",
"b55",
"b64",
"b25",
"b40",
"b63",
"b32",
"b46",
"b75",
"b80",
"b81",
"b76",
"b47",
"b46",
"b8"
],
"table_ref": [],
"text": "In recent years, 3D content has played significant roles in many applications, such as gaming, robotics, films, and animation. Currently, the most common method of creating 3D assets depends on manual efforts using specialized 3D modeling software like Blender [3] and Maya [38], which is very time-consuming and cost-prohibitive to generate high-quality and diverse 3D shapes. As a result, the need for automatic 3D content generation becomes apparent.\nDuring the past decade, image generation has been widely studied and achieved great success using generative models, including generative adversarial networks (GANs) [18,4,28,29,27], variational autoencoders (VAEs) [31,54,63], autoregressive models [64,10,22], and diffusion models [25,60,13,44,30]. Compared with 2D images, 3D shapes are more complex and have different kinds of representations for geometry and textures. Inspired by the progress in 2D generative models, 3D generative models have become an active research area of computer vision and graphics and have achieved pleasing results in the generation of point clouds [2,70,80], implicit fields [11,39], textures [51,50,55], and shapes [16,34]. In addition, recent works based on neural volume rendering [40] tackle 3D-aware novel view synthesis [6,5,19,20,45,49,56,69,81,57].\nSimilar to 2D image generative models like GANs and diffusion models, modern 3D generative models require large-scale datasets to avoid overfitting and achieve diverse results. Unfortunately, it is not always possible to obtain abundant data under some circumstances. Few-shot generation aims to produce diverse and high-quality generated samples using limited data. Modern few-shot image generation approaches [66,26,41,65,33,47,77,82,83,78] adapt models pre-trained on large-scale source datasets to target domains using a few available training samples to relieve overfitting and produce adapted samples following target distributions. Nevertheless, few-shot 3D shape generation has yet to be studied, constrained by the complexity of 3D shape generation and the limited performance of early 3D shape generative models.\nIn this paper, we make the first attempt to study few-shot 3D shape generation pursuing high-quality and diverse generated shapes using limited data. We follow prior few-shot image generation approaches to adapt pre-trained source models to target domains using limited data. Since 3D shapes contain geometry and texture information, we need to clarify two questions: (i) what to learn from limited training data, and (ii) what to adapt from pre-trained source models to target domains. Naturally, we define two 3D shape domain adaptation setups: (i) geometry and texture adaptation (Setup A): the adapted models are trained to learn the geometry information of target data only and preserve the diversity of geometry and textures from source models, and (ii) geometry adaptation only (Setup B): the adapted models are trained to learn both the geometry and texture information of target data and preserve the diversity of geometry from source models only. Since the adaptation approach under setup A can be directly extended to setup B, we mainly focus on setup A and provide additional analysis and results of setup B in the supplementary.\nWe design a few-shot 3D shape generation approach based on modern 3D shape GANs, which synthesize textured meshes with randomly sampled noises requiring 2D supervision only. Source models directly fine-tuned on limited target data cannot maintain generation diversity and produce results similar to training samples. As shown in Fig. 1, two different source samples become analogous after few-shot domain adaptation, losing diversity of geometry and textures. Therefore, we introduce a pairwise relative distances preservation approach [48,47,9] to keep the probability distributions of geometry and texture pairwise similarities in generated shapes at both feature-level and shape-level during domain adaptation. In this way, the adapted models are guided to learn the common properties of limited training samples instead of replicating them. As a consequence, adapted models maintain similar generation diversity to source models and produce diverse results.\nThe main contributions of our work are concluded as follows:\n• To our knowledge, we are the first to study few-shot 3D shape generation and achieve diverse generated shapes with arbitrary topology and textures.\n• We propose a novel few-shot 3D shape adaptation approach to learn target geometry distributions using 2D silhouettes of extremely limited data (e.g., 10 shapes) while preserving diverse information of geometry and textures learned from source domains.\n• We introduce several metrics to evaluate the quality and diversity of few-shot 3D shape generation and demonstrate the effectiveness of our approach qualitatively and quantitatively.\n2 Related Work"
},
{
"figure_ref": [],
"heading": "3D Generative Models",
"publication_ref": [
"b65",
"b57",
"b34",
"b13",
"b22",
"b1",
"b68",
"b78",
"b10",
"b38",
"b42",
"b35",
"b33",
"b10",
"b50",
"b49",
"b15",
"b6",
"b60"
],
"table_ref": [],
"text": "Early works [67,59,35,14,23] extend 2D image generators to 3D voxel grids directly but fail to produce compelling results with high resolution due to the large computational complexity of 3D convolution networks. Other works explore the generation of alternative 3D shape representations, such as point clouds [2,70,80] and implicit fields [11,39]. Following works generate meshes with arbitrary topology using autoregressive models [43] and GANs [36]. Meshdiffusion [34] first applies diffusion models to generate 3D shapes unconditionally using 3D shapes for supervision. These works produce arbitrary topology only and need post-processing steps to achieve textured meshes which are compatible with modern graphics engines.\nDIBR [11] and Textured3DGAN [51,50] synthesize textured 3D meshes based on input templated meshes, resulting in limited topology. GET3D [16] first proposes a 3D generative model [7,62,53] to achieve arbitrary and diverse generation of 3D geometry structures and textures using 2D images for supervision. The proposed few-shot 3D shape generation approach is implemented with GET3D but is not confined to certain network architectures and can also be applied to other 3D shape generative models using 2D supervision."
},
{
"figure_ref": [],
"heading": "Few-shot Image Generation",
"publication_ref": [
"b64",
"b11",
"b27",
"b69",
"b59",
"b73",
"b77",
"b25",
"b45",
"b40",
"b32",
"b36",
"b63",
"b46",
"b75",
"b80",
"b81",
"b66",
"b0",
"b74",
"b76",
"b82",
"b31",
"b72",
"b71",
"b51"
],
"table_ref": [],
"text": "Few-shot image generation aims to produce high-quality images with great diversity utilizing only a few available training samples. Most modern approaches follow the TGAN [66] method to adapt generative models pre-trained on large source domains, including ImageNet [12], FFHQ [28], and LSUN [71] et al., to target domains with limited data. Augmentation approaches [61,75,79] like ADA [26] help generate more different augmented training samples to relieve overfitting. BSA [46] updates the scale and shift parameters in the generator and fixes the other parameters. FreezeD [41] freezes the high-resolution layers in the discriminator to relieve overfitting. EWC [33] applies elastic weight consolidation to regularize the generator by making it harder to change the critical weights which have higher Fisher information [37] values. MineGAN [65] adds additional networks to shift the distributions of the latent space of GANs by modifying the noise inputs of the generator. CDC [47] proposes a cross-domain consistency loss for generators and patch-level discrimination to build a correspondence between source and target domains. DCL [77] uses contrastive learning to maximize the similarity between the corresponding source and target image pairs and push away the generated samples from training samples for greater diversity. MaskDis [82] proposes to regularize the discriminator using masked features and achieves outstanding visual effects. DDPM-PA [83] first realizes few-shot image generation with diffusion models.\nBesides, other recent works have provided different research perspectives. RSSA [68] proposes a relaxed spatial structural alignment method using compressed latent space derived from inverted GANs [1]. AdAM [76] and RICK [78] achieve improvement in the adaptation of unrelated source/target domains. Research including MTG [84], OSCLIP [32], GDA [74], and DIFA [73] et al. explore single-shot GAN adaptation with the guidance of pre-trained CLIP [52] image encoders. This work first explores few-shot 3D shape generation and shares similar ideas of preserving diverse information provided by source models, achieving the few-shot generation of diverse textured 3D shapes."
},
{
"figure_ref": [],
"heading": "Method",
"publication_ref": [
"b15",
"b56",
"b16",
"b14"
],
"table_ref": [],
"text": "Given 3D generative models pre-trained on large source domains, our approach adapts them to target domains by learning the common geometry properties of limited training data while maintaining the generation diversity of geometry and textures. Directly fine-tuned models tend to replicate training samples instead of producing diverse results since the deep generative networks are vulnerable to overfitting, especially when training data is limited. To this end, we propose to keep the probability distributions of the pairwise relative distances between adapted samples similar to source samples.\nWe employ the 3D shape generative model GET3D [16] to illustrate the proposed approach, as shown in Fig. 2. GET3D realizes arbitrary generation of topology and textures using the combination of geometry and texture generators. Both generators are composed of mapping networks M and synthesis networks S. GET3D utilizes the differentiable surface Figure 2: Overview of the proposed few-shot 3D shape generation approach using Cars → SUVs as an example: We maintain the distributions of pairwise relative distances between the geometry and textures of generated samples at feature-level and shape-level to keep diversity during domain adaptation. Only the silhouettes of few-shot target samples are needed as training data.\nrepresentation DMTet [58] to describe geometry with signed distance fields (SDF) defined on deformation fields [17,15]. The texture generator uses mapped geometry and texture codes as inputs and generates texture fields for explicit meshes obtained by adopting DMTet for surface extraction. GET3D is trained with two 2D discriminators applied to RGB images and silhouettes, respectively. Our approach can be divided into geometry adaptation (Sec. 3.1) and texture adaptation (Sec. 3.2) using source models as reference. Mapping networks of adapted models are fixed during domain adaptation. The silhouettes of target shapes are needed as training data to learn geometry distributions. Our approach is not tied to the network architectures of GET3D and is compatible with other 3D shape GANs using 2D supervision."
},
{
"figure_ref": [],
"heading": "Geometry Adaptation",
"publication_ref": [],
"table_ref": [],
"text": "We aim to guide adapted models to learn the common geometry properties of limited training samples while maintaining geometry diversity similar to source models. We propose to keep the probability distributions of pairwise relative distances between the geometry structures of adapted samples at feature-level and shape-level. We first sample a batch of geometry codes {z n 1 } N 0 following the standard normal distribution N (0, I) and get mapped geometry latent codes {ω n 1 } N 0 using fixed geometry mapping networks M geo . The probability distributions for the i th noise vector z i 1 in the source and target geometry generators at feature-level can be expressed as follows:\np s,l geo,i = sf m( sim(S s,l geo (ω i 1 ), S s,l geo (ω j 1 )) ∀i =j ),(1)\np t,l geo,i = sf m( sim(S t,l geo (ω i 1 ), S t,l geo (ω j 1 )) ∀i =j ),(2)\nwhere sf m and sim represent the softmax function and cosine similarity between activations at the l th layer of the source and target geometry synthesis networks which generate SDF and deformation fields. Then we guide target geometry synthesis networks to keep similar probability distributions to source models during domain adaptation with the feature-level geometry loss:\nL geo (S s geo , S t geo ) = E z i D KL (p t,l geo,i ||p s,l geo,i ),(3)\nwhere D KL represents KL-divergence. Similarly, we use source and target silhouettes in place of the features in geometry synthesis networks to keep the pairwise relative distances of adapted samples at shape-level. For this purpose, we further sample a batch of texture codes {z n 2 } N 0 for shape generation. The probability distributions of shapes generated from the i th noise vectors (z i 1 and z i 2 ) by the source and target generators are given by:\np s mask,i = sf m( sim(M ask(G s (z i 1 , z i 2 )), M ask(G s (z j 1 , z j 2 ))) ∀i =j ),(4)\np t mask,i = sf m( sim(M ask(G t (z i 1 , z i 2 )), M ask(G t (z j 1 , z j 2 ))) ∀i =j ),(5)\nwhere G s and G t are the source and target shape generators, M ask represents the masks of 2D rendered shapes. We have the shape-level mask loss for geometry adaptation as follows:\nL mask (G s , G t ) = E z i 1 ,z i 2 ∼N (0,I) i D KL (p t mask,i ||p s mask,i ).(6)"
},
{
"figure_ref": [],
"heading": "Texture Adaptation",
"publication_ref": [],
"table_ref": [],
"text": "In addition, we also encourage adapted models to preserve the texture information learned from source domains and generate target shapes with diverse textures. We still apply the pairwise relative distances preservation approach to relieve overfitting and keep the generation diversity of textures. Since the generated textures for explicit meshes contain both geometry and texture information, we propose to use textures in regions shared by two generated shapes to compute the pairwise relative distances of textures while alleviating the influence of geometry. In the same way, we use the randomly sampled geometry codes {z n N 0 with fixed geometry and texture mapping networks M geo and M tex , respectively. The shared regions of two generated shapes produced by the source and adapted models are defined as the intersection of the masks of the 2D rendered shapes:\nM s i,j = M ask(G s (z i 1 , z i 2 )) ∧ M ask(G s (z j 1 , z j 2 )) (i = j),(7)\nM t i,j = M ask(G t (z i 1 , z i 2 )) ∧ M ask(G t (z j 1 , z j 2 )) (i = j).(8)\nThe probability distributions for the i th noise vectors (z i 1 and z i 2 ) in the source and target texture generators at feature-level can be expressed as follows:\np s,m tex,i = sf m( sim(S s,m tex (ω i 1 , ω i 2 ) ⊗ M s i,j , S s,m tex (ω j 1 , ω j 2 ) ⊗ M s i,j ) ∀i =j ),(9)\np t,m tex,i = sf m( sim(S t,m tex (ω i 1 , ω i 2 ) ⊗ M t i,j , S t,m tex (ω j 1 , ω j 2 ) ⊗ M t i,j ) ∀i =j ),(10)\nwhere ⊗ and sim represent the element-wise multiplication of tensors and cosine similarity between activations at the m th layer of the source and target texture synthesis networks. For shape-level texture adaptation, we use 2D rendered shapes of RGB formats in place of the features in texture synthesis networks to compute the probability distributions:\np s rgb,i = sf m( sim(RGB(G s (z i 1 , z i 2 )) ⊗ M s i,j , RGB(G s (z j 1 , z j 2 )) ⊗ M s i,j ) ∀i =j ),(11)\np t rgb,i = sf m( sim(RGB(G t (z i 1 , z i 2 )) ⊗ M t i,j , RGB(G t (z j 1 , z j 2 )) ⊗ M t i,j ) ∀i =j ),(12)\nwhere RGB represents the rendered RGB images of generated shapes. We have the feature-level texture loss and shape-level RGB loss for texture adaptation as follows:\nL tex (S s tex , S t tex ) = E z i 1 ,z i 2 ∼N (0,I) m,i D KL (p t,m tex,i ||p s,m tex,i ),(13)\nL rgb (G s , G t ) = E z i 1 ,z i 2 ∼N (0,I) i D KL (p t rgb,i ||p s rgb,i ). (14\n)"
},
{
"figure_ref": [],
"heading": "Overall Optimization Target",
"publication_ref": [],
"table_ref": [],
"text": "Since adapted models are guided to learn the geometry information of training data, we only use the mask discriminator and apply the above-mentioned pairwise relative distances preservation methods to preserve diverse geometry and texture information learned from source domains. In this way, our approach only needs the silhouettes of few-shot target shapes as training data. The overall optimization target L of adapted models is defined as follows:\nL = L(D mask , G t ) + µL reg + µ 1 L geo (S s geo , S t geo ) + µ 2 L mask (G s , G t ) + µ 3 L tex (S s tex , S t tex ) + µ 4 L rgb (G s , G t ).(15)\nHere L(D mask , G t ) and L reg represent the adversarial objective of silhouettes and regularization term of generated SDFs used in GET3D. More details of these two losses are added in Appendix B. µ, µ 1 , µ 2 , µ 3 , µ 4 are hyperparameters set manually to control the regularization levels."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [
"b15",
"b6"
],
"table_ref": [],
"text": "We employ a series of few-shot 3D shape adaptation setups to demonstrate the effectiveness of our approach. We first show the qualitative results in Sec. 4.1. Then we introduce several metrics to evaluate quality and diversity quantitatively in Sec. 4.2. Finally, we ablate our approach in Sec. 4.3.\nBasic Setups Our approach is evaluated with GET3D [16]. The hyperparameter of SDF regularization µ is set as 0.01 for all experiments. We empirically find µ 1 = 2e + 4, µ 2 = 5e + 3, µ 3 = 5e + 3, µ 4 = 1e + 4 to work well for the employed adaptation setups. We conduct experiments with batch size 4 on a single NVIDIA A40 GPU. The learning rates of the generator and discriminator are set as 0.0005. The adapted models are trained for 40K-60K iterations. The resolution of 2D rendered RGB images and silhouettes is 1024×1024. More details of implementation are added in Appendix H.\nDatasets We use ShapeNetCore Cars and Chairs [7] as source datasets and sample several 10-shot shapes as target datasets, including (i) Trucks, (ii) Racing Cars, (iii) Sport Utility Vehicles (SUVs), (iv) Police Cars, (v) Ambulances corresponding to Cars and (vi) Rocking Chairs, (vii) Modern Chairs, (viii) Lawn Chairs corresponding to Chairs. Police Cars and Ambulances are used for the experiments of geometry adaptation (see Appendix C). Other datasets are applied to the experiments of geometry and texture adaptation. The training data are rendered using 24 randomly sampled and evenly distributed camera poses. All the few-shot target datasets are visualized in Appendix E.\nBaselines Since few existing works explore few-shot 3D shape generation, we compare the proposed approach with directly fine-tuned methods (DFTM) and fine-tuned models using fixed texture generators (FreezeT), including fixed texture mapping and texture synthesis networks."
},
{
"figure_ref": [
"fig_1",
"fig_2",
"fig_3",
"fig_4"
],
"heading": "Qualitative Evaluation",
"publication_ref": [],
"table_ref": [],
"text": "We visualize the samples produced by our approach using source models pre-trained on ShapeNetCore Cars and Chairs in Fig. 3 and4, respectively. Our approach only needs the silhouettes of few-shot training samples as target datasets to adapt source models to target domains while maintaining generation diversity of geometry and textures. In Fig. 5, we add generated shapes of different target domains rendered in multiple views. Our approach produces high-quality results different from the few-shot training samples. In addition, we compare the proposed approach with baselines using fixed noise inputs for intuitive comparison in Fig. 6. DFTM models replicate training samples and fail to keep generation diversity. FreezeT also fails to produce diverse textures since the mapped geometry codes influence the fixed texture synthesis networks. As a result, FreezeT models produce textured meshes similar to training samples under the guidance of RGB discriminators. Therefore, we further train FreezeT models without RGB discriminators or using source RGB discriminators. However, these two approaches still fail to preserve the diverse geometry and texture information of source models and cannot produce reasonable shapes. Our approach maintains the pairwise relative distances between generated shapes at feature-level and shape-level. It achieves high-quality and diverse adapted samples sharing geometry and texture information with source samples. "
},
{
"figure_ref": [],
"heading": "Quantitative Evaluation",
"publication_ref": [
"b7",
"b70",
"b46"
],
"table_ref": [
"tab_0"
],
"text": "Evaluation Metrics The generation quality of adapted models represents their capability to learn target geometry distributions. Chamfer distance (CD) [8] is employed to compute the distances of geometry distributions between 5000 adapted samples and target datasets containing relatively abundant target data to obtain reliable results. Besides, we design several metrics based on CD and LPIPS [72] to evaluate the diversity of geometry and textures in adapted samples, respectively. LPIPS measures the perceptual distances between images. Evaluation metrics for generation diversity are computed in two ways: (i) pairwise-distances: we randomly generate 1000 shapes and compute the pairwise distances averaged over them, (ii) intra-distances [47]: we first assign the generated shapes to one of the few-shot training samples with the lowest LPIPS distance and then compute the average pairwise distances within each cluster averaged over all the clusters. LPIPS results are averaged over 8 evenly distributed views of rendered shapes. Adapted models which tend to replicate training samples may achieve fine pairwise distances but only get intra-distances close to 0. Adapted models with great generation diversity achieve large values of both pairwise and intra-distances.\nThe quantitative results of our approach are compared with baselines under several few-shot adaptation setups, as listed in Table 1. Our approach learns target geometry distributions better in terms of CD. Moreover, our approach also performs better on all the benchmarks of generation diversity, indicating its strong capability to produce diverse shapes with different geometry structures and textures."
},
{
"figure_ref": [
"fig_5"
],
"heading": "Ablation Analysis",
"publication_ref": [],
"table_ref": [],
"text": "Our approach is composed of the pairwise relative distances preservation methods applied to geometry and textures at feature-level and shape-level. We provide ablation analysis to show the roles played by each component of our approach. In Fig. 7, we show the qualitative ablation analysis using 10-shot Chairs → Rocking Chairs as an example. Our full approach adapts source samples to target domains while preserving diverse geometry and texture information. Adapted models only using GAN loss with mask discrimination fail to maintain geometry diversity or produce high-quality shapes. Adding fixed source RGB discriminators results in texture degradation. Absence of the feature-level texture loss makes adapted models harder to preserve the texture information learned from source domains. Absence of shape-level RGB loss leads to repetitive textures and discontinuous shapes. As for the feature-level geometry and shape-level mask losses, their absence results in adapted samples sharing similar geometry structures and incomplete shapes. We also add ablations using geometry and mask losses, texture and RGB losses, feature-level losses, and shape-level losses, respectively. None of these approaches generate compelling results with diverse topology and textures. Incomplete geometry structures and low-quality textures can be found in their adapted samples. Moreover, the full approach also achieves quantitative results better than other settings, as shown in Appendix D. "
},
{
"figure_ref": [],
"heading": "Conclusion and Limitations",
"publication_ref": [],
"table_ref": [],
"text": "This paper first explores few-shot 3D shape generation. We introduce a novel domain adaptation approach to produce 3D shapes with diverse topology and textures. The relative distances between generated samples are maintained at both feature-level and shape-level. We only need the silhouettes of few-shot target samples as training data to learn target geometry distributions while keeping diversity. Our approach is implemented based on GET3D to demonstrate its effectiveness. However, it is not constrained by specific network architectures and can be combined with more powerful 3D shape generative models using 2D supervision to produce higher-quality results in the future. Despite the compelling results of our approach, it still has some limitations. Firstly, it sometimes cannot completely preserve the diverse textures of source samples. Besides, it is mainly designed for related source/target domains. Extending our approach to unrelated domain adaptation would be promising. Nevertheless, we believe this work takes a further step towards democratizing 3D content creation by transferring knowledge in available source models to fit target distributions using few-shot data. "
},
{
"figure_ref": [],
"heading": "A Broader Impact",
"publication_ref": [],
"table_ref": [],
"text": "We propose a novel approach for few-shot 3D shape generation, achieving diverse 3D shape generation using limited training data. Our approach is more prone to biases introduced by training data than typical artificial intelligence generative models since it only needs silhouettes of few-shot samples to train adapted models. The proposed approach is applicable to 3D shape generative models and not tailored for sensitive applications like generating human bodies. Therefore, we recommend practitioners to apply abundant caution when dealing with such applications to avoid problems of races, skin tones, or gender identities."
},
{
"figure_ref": [
"fig_6",
"fig_1"
],
"heading": "B More Details of GET3D",
"publication_ref": [
"b15",
"b41",
"b6",
"b60"
],
"table_ref": [],
"text": "GET3D [16] is the first 3D shape generative model to produce textured meshes with arbitrary topology and textures. Here we add more details of the GET3D model. The mapping networks of GET3D are composed of 3D convolutional and fully connected networks. The synthesis networks for SDF and deformation fields are MLPs. As for the texture synthesis networks, GET3D uses generator network structures similar to StyleGAN2 to generate textures using triplane feature maps as inputs. GET3D also follows StyleGAN2 to use the same 2D discriminators and non-saturating GAN objective. Two 2D image discriminators are applied to RGB images and silhouettes, respectively. Given x representing an RGB image or a silhouette, the adversarial objective is defined as:\nL(D x , G t ) = E z∈N [g(D x (R(G t (z))))] + E Ix∈px g(-D x (I x )) + λ||∇D x (I x )|| 2 2 , (16\n)\nwhere g(u) = -log(1 + exp(-u)), p x and R represent the real image distributions and rendering functions for RGB images or silhouettes. In Eq. 15, we employ the discriminator for silhouettes as L(D mask , G t ). The discriminator for RGB images used in GET3D is expressed as L(D rgb , G t ). The regularization loss L reg in Eq. 15 is designed to remove internal floating surfaces since GET3D aims to generate textured meshes without internal structures. L reg is defined as a cross-entropy loss between the SDF values of neighboring vertices [42]:\nL reg = i,j∈Se,i =j H(σ(s i ), sign(s j )) + H(σ(s j ), sign(s i )).(17)\nHere H and σ represent binary cross-entropy loss and sigmoid function. s i , s j are SDF values of neighboring vertices in the set of unique edges S e in the tetrahedral grid. The regularization loss L reg is applied to all the experiments (including ablation analysis) in this paper. GET3D needs multi-view rendered RGB images and silhouettes with corresponding camera distribution parameters as training data. Therefore, it is evaluated with synthetic datasets such as ShapeNetCore [7] and TurboSquid [62]. Future work may extend GET3D to single-view real-world datasets. If so, our approach can be applied to the advanced models to realize few-shot generation of real-world 3D shapes using single-view silhouettes.\nIn Fig. 8, we provide generated samples of the officially released GET3D models trained on ShapNetCore Cars and Chairs datasets. These models are used as source models in our experiments. GET3D generates shapes with arbitrary topology and textures. However, improvement room still exists for better results, such as incomplete textures of tires. As a result, our approach produces some samples with incomplete textures of tires, as shown in Fig. 3. Our approach can be combined with better generative models in the future to achieve better visual effects."
},
{
"figure_ref": [
"fig_7"
],
"heading": "C Geometry Adaptation Only",
"publication_ref": [
"b6",
"b23"
],
"table_ref": [
"tab_1",
"tab_2"
],
"text": "In this section, we add the discussion of geometry adaptation only (Setup B). Source models are trained to learn geometry and textures from limited training data under setup B. Adapted models preserve the diversity of geometry learned from source Table 4: Quantitative ablations of the proposed approach using 10-shot Chairs → Rocking Chairs as an example. The full approach performs the best on both generation quality and diversity. domains. As for textures, we guide adapted models to fit the distributions of training samples. Method The proposed adaptation approach under setup B has two differences compared with setup A (texture and geometry adaptation) discussed in our paper. Firstly, the feature-level texture loss and shape-level RGB loss are no longer needed. Secondly, generators are guided by the RGB discriminator to learn target texture distributions. Therefore, we need RGB images of rendered real samples as inputs for the RGB discriminator. The overall optimization target of adapted models under setup B is defined as follows:\nL = L(D mask , G t ) + L(D rgb , G t ) + µL reg + µ 1 L geo (S s geo , S t geo ) + µ 2 L mask (G s , G t )(18)\nWe follow GET3D to set µ = 0.01 and empirically find µ 1 and µ 2 ranging from 2e+3 to 1e+4 appropriate for the adaptation setups used in our paper.\nExperiments We sample two 10-shot target datasets from ShapeNetCore [7] to evaluate our approach under setup B, including Ambulances and Police Cars in correspondence to the source domain Cars. The basic setups of experiments under setup B are consistent with those under setup A (see Sec. 4). We provide qualitative and quantitative results of our approach to demonstrate its effectiveness under setup B. As shown in Fig. 9, our approach produces ambulances and police cars with diverse topology using few-shot training samples qualitatively. For quantitative evaluation, we further add FID [24] to evaluate the generation quality. FID results are averaged over 24 views of rendered shapes. The quantitative results are listed in Tables 2 and3. Compared with DFTM models, our approach performs better on learning target geometry distributions in terms of CD. As for FID, our approach achieves better results on Cars → Police Cars and gets results close to the DFTM model on Cars → Ambulances. Besides, our approach achieves greater generation diversity in terms of Intra-CD and Intra-LPIPS. DFTM models get better results on Pairwise-CD and results close to our approach on Pairwise-LPIPS but get apparently worse results on intra-distances, indicating that they are overfitting to few-shot training samples and tend to replicate them instead of producing diverse results. We do not include FreezeT models for comparison under setup B since the adapted models are trained to learn the texture information from limited training samples. "
},
{
"figure_ref": [
"fig_8",
"fig_0",
"fig_9",
"fig_1",
"fig_0"
],
"heading": "D Supplementary Ablations",
"publication_ref": [],
"table_ref": [],
"text": "Quantitative Ablations Table 4 shows quantitative ablations of our approach. The full approach achieves the best quantitative results on both generation quality and diversity. Without feature-level geometry loss or shape-level mask loss, adapted models performs worse on geometry diversity in terms of Intra-CD and Pairwise-CD. Similarly, adapted models perform worse on texture diversity in terms of Intra-LPIPS and Pairwise-LPIPS without feature-level texture loss or shape-level RGB loss.\nAblations of Shared Masks In addition, we provide qualitative ablations for the shared masks used for feature-level texture loss and shape-level RGB loss computation in Fig. 10. Absence of shared masks causes geometry structures to bias the domain Figure 12: Ablations of fixed mapping networks during domain adaptation. Without fixed mapping networks, our approach fails to preserve the diverse texture information of source samples and produces blurred textures. adaptation of textures, making the textures of adapted samples more different from source samples. For example, the blue and orange source cars change into yellow-blue and red trucks during the 10-shot domain adaptation. The full approach applies shared masks to relieve the influence of geometry structures and achieves better preservation of the texture information in source models.\nAblations of Hyperparameters We add ablations of the hyperparameters applied to the proposed four adaptation losses. We use different values of hyperparameters and provide qualitative results using 10-shot Cars → SUVs in Fig. 11. Too large values of hyperparameters prevent adapted models from learning target distributions, resulting in results similar to Figure 13: Visualization of the 10-shot 3D shape datasets used in this paper. Using randomly sampled views, we provide one 2D rendered RGB image for each training shape. source samples. Too small values of hyperparameters lead to diversity degradation of geometry and textures. We empirically recommend hyperparameters µ 1 , µ 2 , µ 3 , µ 4 ranging from 2e+3 to 1e+4 for adaptation setups used in this paper.\nAblations of Fixed Mapping Networks As illustrated in Sec. 3, the geometry and texture mapping networks M geo and M tex are fixed during domain adaptation. We propose this design to isolate the geometry and texture adaptation since the texture synthesis networks need the mapped geometry codes as inputs. Without fixed mapping networks, fine-tuned geometry mapping networks would influence the texture adaptation process. We add ablations of fixed mapping networks under different adaptation setups and provide qualitative samples in Fig. 12. The low-quality adapted samples show blurred textures and fail to preserve the diverse texture information of source samples."
},
{
"figure_ref": [
"fig_1"
],
"heading": "E More Details of Datasets",
"publication_ref": [
"b6",
"b7",
"b23"
],
"table_ref": [],
"text": "This paper employs several 10-shot datasets sampled from ShapeNetCore [7] as training data for few-shot 3D shape generation. The rendered datasets of randomly sampled views are shown in Fig. 13. As for the main experiments of our paper, we only need silhouettes of target samples as training data, as shown in Fig. 2. For the experiments of geometry adaptation only (Sec. C), rendered RGB images are also needed to train adapted models.\nWe employ CD [8] and FID [24] as quantitative evaluation metrics for generation quality. Datasets containing relatively "
},
{
"figure_ref": [],
"heading": "F Computational Cost",
"publication_ref": [],
"table_ref": [
"tab_3"
],
"text": "Table 5 shows the computational cost of our approach under two adaptation setups using a single NVIDIA A40 GPU. We also ablate our approach to show the computational cost of each component. The adapted models are trained for about 40K-60K iterations in our experiments, costing about 4.4-6.5 and 3.8-5.7 hours under setup A (geometry and texture adaptation) and setup B (geometry adaptation only), respectively. DFTM under setup B is the same as training GET3D models directly. DFTM under setup A excludes the RGB discriminator. Compared with DFTM, the approach only using GAN loss includes the time cost by source models."
},
{
"figure_ref": [],
"heading": "G Inspiration of Loss Design",
"publication_ref": [
"b47",
"b20",
"b8",
"b46",
"b80",
"b81"
],
"table_ref": [],
"text": "Our approach is composed of feature-level geometry loss, feature-level texture loss, shape-level mask loss, and shape-level RGB loss sharing similar formats to preserve the relative distances between generated shapes. Our approach is mainly inspired by contrastive learning methods [48,21,9]. Similar approaches can be found in recent few-shot image generation approaches [47,82,83] as well. This paper first explores few-shot 3D shape generation and proposes an effective domain adaptation approach by adopting the pairwise relative distances preservation loss for geometry and textures at feature-level and shape-level."
},
{
"figure_ref": [],
"heading": "H More Details of Implementation",
"publication_ref": [
"b15",
"b6"
],
"table_ref": [
"tab_0"
],
"text": "The proposed approach is implemented based on the official code of GET3D [16]. The setups of adapted models are consistent with those of the officially released source models trained on ShapeNetCore Cars and Chairs [7]. The geometry and texture synthesis networks are composed of 2-layers MLP networks. We concatenate the output features of the first layers in the synthesis networks of SDFs and deformation fields for feature-level geometry loss computation since the output features of the second layers have different sizes for SDFs and deformation fields. We also use the features in the synthesis networks of SDFs and deformation fields separately for feature-level geometry loss computation. Unfortunately, it is more time-consuming and fails to produce better results. For feature-level texture loss computation, we use the output features of the second layers in the texture synthesis network, which has the same resolution as the generated shapes. Therefore, we can directly apply the shared masks of generated shapes to the texture features.\nThe weights in target models are initialized to source models. We set the learning rates of the generator and discriminator as 0.0005, which is lower than the learning rates of source models (0.002), to realize more refined adaptation processes. We set the hyperparameters of the proposed losses (µ 1 , µ 2 , µ 3 , µ 4 ) equally for adaptation from Cars and Chairs and achieve high-quality results. Different hyperparameters can be tried to obtain compelling results under other adaptation setups. We train adapted models with batch size 4 on a single NVIDIA A40 GPU (45GB GPU memory). Our approach needs about 20 GB GPU memory for the image resolution of 1024 × 1024. The standard deviations of pairwise-distance and intra-distance results listed in Tables 1,4, and 3 are computed across shape pairs picked from generated samples and 10 clusters (the same number as few-shot training samples), respectively."
},
{
"figure_ref": [
"fig_1",
"fig_2",
"fig_10"
],
"heading": "I More Visualization Results",
"publication_ref": [
"b6"
],
"table_ref": [],
"text": "As supplements to generated samples shown in Fig. 3 and4, we show more examples produced by our approach under several few-shot adaptation setups. Adapted samples obtained with the source models pre-trained on ShapeNetCore Cars and Chairs [7] are shown in Fig. 14 "
}
] | Figure 1: Given pre-trained 3D shape generative models, we propose to adapt them to target domains using a few target samples while preserving diverse geometry and texture information learned from source domains. Compared with directly fine-tuned models which tend to replicate the few-shot target samples, our approach only needs the silhouettes of target samples as training data and achieves diverse generated shapes following target geometry distributions but different from target samples. | Few-shot 3D Shape Generation | [
{
"figure_caption": "1 } N 0 and texture codes {z n 2 } N 0 and get mapped latent codes {ω n 1 } N 0 and {ω n 2 }12",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "12",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: 10-shot generated shapes of our approach on Cars → Trucks, SUVs, and Racing Cars.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: 10-shot generated shapes of our approach on Chairs → Rocking Chairs, Modern Chairs, and Lawn Chairs.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 :5Figure 5: Multi-view rendered shapes produced by our approach on different 10-shot target domains.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Figure 6 :6Figure 6: Visualization samples comparison on 10-shot Cars → SUVs, Cars → Racing Cars, and Chairs → Rocking Chairs. Results of different approaches are synthesized with fixed noise inputs.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "Figure 7 :7Figure 7: Qualitative ablations of our approach using 10-shot Chairs → Rocking Chairs as an example. Results of different approaches are synthesized with fixed noise inputs for intuitive comparison.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "7",
"figure_type": "figure"
},
{
"figure_caption": "Figure 8 :8Figure 8: Generated shapes produced by the source GET3D models trained on ShapeNetCore Cars and Chairs datasets.",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "8",
"figure_type": "figure"
},
{
"figure_caption": "Figure 9 :9Figure 9: 10-shot generated shapes of our approach on Cars → Ambulances and Police Cars.",
"figure_data": "",
"figure_id": "fig_7",
"figure_label": "9",
"figure_type": "figure"
},
{
"figure_caption": "Figure 10 :10Figure 10: Qualitative ablations of shared masks applied to the feature-level texture loss and shape-level RGB loss using 10-shot Cars → Trucks as an example. The generated shapes of different approaches are synthesized with fixed noise inputs for intuitive comparison.",
"figure_data": "",
"figure_id": "fig_8",
"figure_label": "10",
"figure_type": "figure"
},
{
"figure_caption": "Figure 11 :11Figure 11: Qualitative ablations of the hyperparameters applied to the proposed adaptation losses using 10-shot Cars → SUVs as an example.",
"figure_data": "",
"figure_id": "fig_9",
"figure_label": "11",
"figure_type": "figure"
},
{
"figure_caption": "Figure 14 :14Figure 14: Additional 10-shot generated shapes of our approach on Cars → Trucks, SUVs, and Racing Cars.",
"figure_data": "",
"figure_id": "fig_10",
"figure_label": "14",
"figure_type": "figure"
},
{
"figure_caption": "and 15, respectively. ",
"figure_data": "",
"figure_id": "fig_11",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 15 :15Figure 15: Additional 10-shot generated shapes of our approach on Chairs → Rocking Chairs, Modern Chairs, and Lawn Chairs.",
"figure_data": "",
"figure_id": "fig_12",
"figure_label": "15",
"figure_type": "figure"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Quantitative evaluation of our approach. Generated shapes of different approaches are synthesized from fixed noise inputs for fair comparison. CD scores are multiplied by 10 3 . The best results are highlighted in bold. Our approach performs better on both generation quality and diversity.",
"figure_data": "DatasetsApproach CD (↓)Intra-CD (↑)Pairwise-CD (↑) Intra-LPIPS (↑) Pairwise-LPIPS (↑)Cars → SUVsDFTM FreezeT Ours1.401 1.553 1.323 1.323 1.3230.316 ± 0.002 0.240 ± 0.005 0.511 ± 0.006 0.511 ± 0.006 0.511 ± 0.0060.513 ± 0.001 0.326 ± 0.002 0.814 ± 0.007 0.814 ± 0.007 0.814 ± 0.0070.062 ± 0.001 0.055 ± 0.002 0.109 ± 0.026 0.109 ± 0.026 0.109 ± 0.0260.063 ± 0.012 0.060 ± 0.014 0.095 ± 0.022 0.095 ± 0.022 0.095 ± 0.022Cars → TrucksDFTM FreezeT Ours4.014 4.175 3.940 3.940 3.9400.441 ± 0.003 0.412 ± 0.006 1.061 ± 0.014 1.061 ± 0.014 1.061 ± 0.0140.689 ± 0.003 0.766 ± 0.002 1.175 ± 0.004 1.175 ± 0.004 1.175 ± 0.0040.112 ± 0.002 0.120 ± 0.003 0.145 ± 0.022 0.145 ± 0.022 0.145 ± 0.0220.119 ± 0.024 0.128 ± 0.027 0.146 ± 0.033 0.146 ± 0.033 0.146 ± 0.033Chairs →DFTM40.559 4.001 ± 0.005 13.598 ± 0.0130.165 ± 0.0290.141 ± 0.047LawnFreezeT39.422 4.671 ± 0.022 19.269 ± 0.0240.120 ± 0.0320.165 ± 0.040ChairsOurs38.661 38.661 38.661 5.852 ± 0.031 5.852 ± 0.031 5.852 ± 0.031 22.989 ± 0.022 22.989 ± 0.022 22.989 ± 0.0220.278 ± 0.040 0.278 ± 0.040 0.278 ± 0.0400.166 ± 0.054 0.166 ± 0.054 0.166 ± 0.054Chairs →DFTM18.996 7.405 ± 0.022 15.312 ± 0.0110.202 ± 0.0390.203 ± 0.037RockingFreezeT18.503 5.541 ± 0.014 11.977 ± 0.0090.203 ± 0.0460.204 ± 0.036ChairsOurs17.598 17.598 17.598 8.773 ± 0.029 8.773 ± 0.029 8.773 ± 0.029 16.165 ± 0.015 16.165 ± 0.015 16.165 ± 0.0150.289 ± 0.062 0.289 ± 0.062 0.289 ± 0.0620.222 ± 0.063 0.222 ± 0.063 0.222 ± 0.063",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Quantitative evaluation of our approach on generation quality of geometry and textures.",
"figure_data": "DatastesApproach FID (↓) CD (↓)Cars →DFTM101.583 101.583 101.5836.896AmbulancesOurs103.708 5.963 5.963 5.963Cars →DFTM86.8336.440Police CarsOurs74.958 74.958 74.9585.616 5.616 5.616",
"figure_id": "tab_1",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Quantitative evaluation of our approach on generation diversity of geometry and textures.",
"figure_data": "DatastesApproachIntra-CD (↑)Pairwise-CD (↑) Inra-LPIPS (↑) Pairwise-LPIPS (↑)Cars →DFTM0.300 ± 0.0021.027 ± 0.007 1.027 ± 0.007 1.027 ± 0.0070.079 ± 0.0090.083 ± 0.017AmbulancesOurs0.558 ± 0.004 0.558 ± 0.004 0.558 ± 0.0040.638 ± 0.0060.093 ± 0.018 0.093 ± 0.018 0.093 ± 0.0180.086 ± 0.016 0.086 ± 0.016 0.086 ± 0.016Cars →DFTM0.426 ± 0.0030.926 ± 0.008 0.926 ± 0.008 0.926 ± 0.0080.109 ± 0.0020.108 ± 0.017Police CarsOurs0.902 ± 0.005 0.902 ± 0.005 0.902 ± 0.0050.902 ± 0.0060.115 ± 0.009 0.115 ± 0.009 0.115 ± 0.0090.120 ± 0.020 0.120 ± 0.020 0.120 ± 0.020",
"figure_id": "tab_2",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "The time cost of our approach trained for 1K iterations in terms of seconds on a single NVIDIA A40 GPU (image resolution 1024 × 1024, batch size 4). applied for evaluation to obtain reliable results. The few-shot samples are excluded from the relatively abundant datasets to avoid the influence of overfitting. The relatively abundant Trucks, SUVs, Ambulances, Police Cars, Rocking Chairs, and Lawn Chairs datasets contain 40, 369, 73, 133, 87, and 78 samples.",
"figure_data": "SetupsApproachesTime cost for 1K iterationsDFTM228.83GAN loss only272.27GAN loss w/ Texture loss352.80Setup AGAN loss w/ Geometry loss295.28GAN loss w/ RGB loss291.62GAN loss w/ Mask loss279.34Full Approach392.67DFTM281.15GAN loss only316.55Setup BGAN loss w/ Geometry loss344.82GAN loss w/ Mask loss322.51Full Approach340.38abundant data are",
"figure_id": "tab_3",
"figure_label": "5",
"figure_type": "table"
}
] | Jingyuan Zhu; Huimin Ma; Jiansheng Chen; Jian Yuan | [
{
"authors": "R Abdal; Y Qin; P Wonka",
"journal": "",
"ref_id": "b0",
"title": "Image2stylegan++: How to edit the embedded images",
"year": "2020"
},
{
"authors": "P Achlioptas; O Diamanti; I Mitliagkas; L Guibas",
"journal": "PMLR",
"ref_id": "b1",
"title": "Learning representations and generative models for 3d point clouds",
"year": "2018"
},
{
"authors": " Blender",
"journal": "Stichting Blender Foundation",
"ref_id": "b2",
"title": "Blender -a 3d modelling and rendering package",
"year": "2018"
},
{
"authors": "A Brock; J Donahue; K Simonyan",
"journal": "",
"ref_id": "b3",
"title": "Large scale GAN training for high fidelity natural image synthesis",
"year": "2019"
},
{
"authors": "E R Chan; C Z Lin; M A Chan; K Nagano; B Pan; S De Mello; O Gallo; L J Guibas; J Tremblay; S Khamis",
"journal": "",
"ref_id": "b4",
"title": "Efficient geometry-aware 3d generative adversarial networks",
"year": "2022"
},
{
"authors": "E R Chan; M Monteiro; P Kellnhofer; J Wu; G Wetzstein",
"journal": "",
"ref_id": "b5",
"title": "pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis",
"year": "2021"
},
{
"authors": "A X Chang; T Funkhouser; L Guibas; P Hanrahan; Q Huang; Z Li; S Savarese; M Savva; S Song; H Su",
"journal": "",
"ref_id": "b6",
"title": "Shapenet: An information-rich 3d model repository",
"year": "2015"
},
{
"authors": "D.-Y Chen; X.-P Tian; Y.-T Shen; M Ouhyoung",
"journal": "Computer Graphics Forum",
"ref_id": "b7",
"title": "On visual similarity based 3d model retrieval",
"year": "2003"
},
{
"authors": "T Chen; S Kornblith; M Norouzi; G Hinton",
"journal": "PMLR",
"ref_id": "b8",
"title": "A simple framework for contrastive learning of visual representations",
"year": "2020"
},
{
"authors": "X Chen; N Mishra; M Rohaninejad; P Abbeel",
"journal": "PMLR",
"ref_id": "b9",
"title": "Pixelsnail: An improved autoregressive generative model",
"year": "2018"
},
{
"authors": "Z Chen; H Zhang",
"journal": "",
"ref_id": "b10",
"title": "Learning implicit fields for generative shape modeling",
"year": "2019"
},
{
"authors": "J Deng; W Dong; R Socher; L.-J Li; K Li; L Fei-Fei",
"journal": "",
"ref_id": "b11",
"title": "Imagenet: A large-scale hierarchical image database",
"year": "2009"
},
{
"authors": "P Dhariwal; A Nichol",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b12",
"title": "Diffusion models beat gans on image synthesis",
"year": "2021"
},
{
"authors": "M Gadelha; S Maji; R Wang",
"journal": "IEEE",
"ref_id": "b13",
"title": "3d shape induction from 2d views of multiple objects",
"year": "2017"
},
{
"authors": "J Gao; W Chen; T Xiang; A Jacobson; M Mcguire; S Fidler",
"journal": "Advances In Neural Information Processing Systems",
"ref_id": "b14",
"title": "Learning deformable tetrahedral meshes for 3d reconstruction",
"year": "2020"
},
{
"authors": "J Gao; T Shen; Z Wang; W Chen; K Yin; D Li; O Litany; Z Gojcic; S Fidler",
"journal": "Advances In Neural Information Processing Systems",
"ref_id": "b15",
"title": "Get3d: A generative model of high quality 3d textured shapes learned from images",
"year": "2022"
},
{
"authors": "J Gao; Z Wang; J Xuan; S Fidler",
"journal": "Springer",
"ref_id": "b16",
"title": "Beyond fixed grid: Learning geometric image representation with a deformable grid",
"year": "2020"
},
{
"authors": "I Goodfellow; J Pouget-Abadie; M Mirza; B Xu; D Warde-Farley; S Ozair; A Courville; Y Bengio",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b17",
"title": "Generative adversarial nets",
"year": "2014"
},
{
"authors": "J Gu; L Liu; P Wang; C Theobalt",
"journal": "",
"ref_id": "b18",
"title": "Stylenerf: A style-based 3d aware generator for high-resolution image synthesis",
"year": "2022"
},
{
"authors": "Z Hao; A Mallya; S Belongie; M.-Y Liu",
"journal": "",
"ref_id": "b19",
"title": "Gancraft: Unsupervised 3d neural rendering of minecraft worlds",
"year": "2021"
},
{
"authors": "K He; H Fan; Y Wu; S Xie; R Girshick",
"journal": "",
"ref_id": "b20",
"title": "Momentum contrast for unsupervised visual representation learning",
"year": "2020"
},
{
"authors": "T Henighan; J Kaplan; M Katz; M Chen; C Hesse; J Jackson; H Jun; T B Brown; P Dhariwal; S Gray",
"journal": "",
"ref_id": "b21",
"title": "Scaling laws for autoregressive generative modeling",
"year": "2020"
},
{
"authors": "P Henzler; N J Mitra; T Ritschel",
"journal": "",
"ref_id": "b22",
"title": "Escaping plato's cave: 3d shape from adversarial rendering",
"year": "2019"
},
{
"authors": "M Heusel; H Ramsauer; T Unterthiner; B Nessler; S Hochreiter",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b23",
"title": "Gans trained by a two time-scale update rule converge to a local nash equilibrium",
"year": "2017"
},
{
"authors": "J Ho; A Jain; P Abbeel",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b24",
"title": "Denoising diffusion probabilistic models",
"year": "2020"
},
{
"authors": "T Karras; M Aittala; J Hellsten; S Laine; J Lehtinen; T Aila",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b25",
"title": "Training generative adversarial networks with limited data",
"year": "2020"
},
{
"authors": "T Karras; M Aittala; S Laine; E Härkönen; J Hellsten; J Lehtinen; T Aila",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b26",
"title": "Alias-free generative adversarial networks",
"year": "2021"
},
{
"authors": "T Karras; S Laine; T Aila",
"journal": "",
"ref_id": "b27",
"title": "A style-based generator architecture for generative adversarial networks",
"year": "2019"
},
{
"authors": "T Karras; S Laine; M Aittala; J Hellsten; J Lehtinen; T Aila",
"journal": "",
"ref_id": "b28",
"title": "Analyzing and improving the image quality of stylegan",
"year": "2020"
},
{
"authors": "D Kingma; T Salimans; B Poole; J Ho",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b29",
"title": "Variational diffusion models",
"year": "2021"
},
{
"authors": "D P Kingma; M Welling",
"journal": "",
"ref_id": "b30",
"title": "Auto-encoding variational bayes",
"year": "2013"
},
{
"authors": "G Kwon; J C Ye",
"journal": "",
"ref_id": "b31",
"title": "One-shot adaptation of gan in just one clip",
"year": "2022"
},
{
"authors": "Y Li; R Zhang; J Lu; E Shechtman",
"journal": "",
"ref_id": "b32",
"title": "Few-shot image generation with elastic weight consolidation",
"year": "2020"
},
{
"authors": "Z Liu; Y Feng; M J Black; D Nowrouzezahrai; L Paull; W Liu",
"journal": "",
"ref_id": "b33",
"title": "Meshdiffusion: Score-based generative 3d mesh modeling",
"year": "2023"
},
{
"authors": "S Lunz; Y Li; A Fitzgibbon; N Kushman",
"journal": "",
"ref_id": "b34",
"title": "Inverse graphics gan: Learning to generate 3d shapes from unstructured 2d data",
"year": "2020"
},
{
"authors": "A Luo; T Li; W.-H Zhang; T S Lee",
"journal": "",
"ref_id": "b35",
"title": "Surfgen: Adversarial 3d shape synthesis with explicit surface discriminators",
"year": "2021"
},
{
"authors": "A Ly; M Marsman; J Verhagen; R Grasman; E J Wagenmakers",
"journal": "Journal of Mathematical Psychology",
"ref_id": "b36",
"title": "A tutorial on fisher information",
"year": "2017"
},
{
"authors": "Maya ",
"journal": "",
"ref_id": "b37",
"title": "",
"year": "2022-05-19"
},
{
"authors": "L Mescheder; M Oechsle; M Niemeyer; S Nowozin; A Geiger",
"journal": "",
"ref_id": "b38",
"title": "Occupancy networks: Learning 3d reconstruction in function space",
"year": "2019"
},
{
"authors": "B Mildenhall; P P Srinivasan; M Tancik; J T Barron; R Ramamoorthi; R Ng",
"journal": "",
"ref_id": "b39",
"title": "Nerf: Representing scenes as neural radiance fields for view synthesis",
"year": "2020"
},
{
"authors": "S Mo; M Cho; J Shin",
"journal": "",
"ref_id": "b40",
"title": "Freeze the discriminator: A simple baseline for fine-tuning gans",
"year": "2020"
},
{
"authors": "J Munkberg; J Hasselgren; T Shen; J Gao; W Chen; A Evans; T Müller; S Fidler",
"journal": "",
"ref_id": "b41",
"title": "Extracting triangular 3d models, materials, and lighting from images",
"year": "2022"
},
{
"authors": "C Nash; Y Ganin; S A Eslami; P Battaglia",
"journal": "PMLR",
"ref_id": "b42",
"title": "Polygen: An autoregressive generative model of 3d meshes",
"year": "2020"
},
{
"authors": "A Q Nichol; P ",
"journal": "PMLR",
"ref_id": "b43",
"title": "Improved denoising diffusion probabilistic models",
"year": "2021"
},
{
"authors": "M Niemeyer; A Geiger",
"journal": "",
"ref_id": "b44",
"title": "Giraffe: Representing scenes as compositional generative neural feature fields",
"year": "2021"
},
{
"authors": "A Noguchi; T Harada",
"journal": "",
"ref_id": "b45",
"title": "Image generation from small datasets via batch statistics adaptation",
"year": "2019"
},
{
"authors": "U Ojha; Y Li; J Lu; A A Efros; Y J Lee; E Shechtman; R Zhang",
"journal": "",
"ref_id": "b46",
"title": "Few-shot image generation via cross-domain correspondence",
"year": "2021"
},
{
"authors": "A V D Oord; Y Li; O Vinyals",
"journal": "",
"ref_id": "b47",
"title": "Representation learning with contrastive predictive coding",
"year": "2018"
},
{
"authors": "R Or-El; X Luo; M Shan; E Shechtman; J J Park; I Kemelmacher-Shlizerman",
"journal": "",
"ref_id": "b48",
"title": "Stylesdf: High-resolution 3d-consistent image and geometry generation",
"year": "2022"
},
{
"authors": "D Pavllo; J Kohler; T Hofmann; A Lucchi",
"journal": "",
"ref_id": "b49",
"title": "Learning generative models of textured 3d meshes from real-world images",
"year": "2021"
},
{
"authors": "D Pavllo; G Spinks; T Hofmann; M.-F Moens; A Lucchi",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b50",
"title": "Convolutional generation of textured 3d meshes",
"year": "2020"
},
{
"authors": "A Radford; J W Kim; C Hallacy; A Ramesh; G Goh; S Agarwal; G Sastry; A Askell; P Mishkin; J Clark; G Krueger; I Sutskever",
"journal": "PMLR",
"ref_id": "b51",
"title": "Learning transferable visual models from natural language supervision",
"year": "2021"
},
{
"authors": "D J Rezende; S Mohamed; D Wierstra",
"journal": "PMLR",
"ref_id": "b52",
"title": "Stochastic backpropagation and approximate inference in deep generative models",
"year": "2014"
},
{
"authors": "E Richardson; G Metzer; Y Alaluf; R Giryes; D Cohen-Or",
"journal": "",
"ref_id": "b53",
"title": "Texture: Text-guided texturing of 3d shapes",
"year": "2023"
},
{
"authors": "K Schwarz; Y Liao; M Niemeyer; A Geiger",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b54",
"title": "Graf: Generative radiance fields for 3d-aware image synthesis",
"year": "2020"
},
{
"authors": "K Schwarz; A Sauer; M Niemeyer; Y Liao; A Geiger",
"journal": "",
"ref_id": "b55",
"title": "Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids",
"year": "2022"
},
{
"authors": "T Shen; J Gao; K Yin; M.-Y Liu; S Fidler",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b56",
"title": "Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis",
"year": "2021"
},
{
"authors": "E J Smith; D Meger",
"journal": "PMLR",
"ref_id": "b57",
"title": "Improved adversarial systems for 3d object generation and reconstruction",
"year": "2017"
},
{
"authors": "Y Song; S Ermon",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b58",
"title": "Improved techniques for training score-based generative models",
"year": "2020"
},
{
"authors": "N.-T Tran; V.-H Tran; N.-B Nguyen; T.-K Nguyen; N.-M Cheung",
"journal": "IEEE Transactions on Image Processing",
"ref_id": "b59",
"title": "On data augmentation for gan training",
"year": "2021"
},
{
"authors": "",
"journal": "",
"ref_id": "b60",
"title": "Turbosquid",
"year": "2022-05-19"
},
{
"authors": "A Vahdat; J Kautz",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b61",
"title": "Nvae: A deep hierarchical variational autoencoder",
"year": "2020"
},
{
"authors": "A Van Den Oord; N Kalchbrenner; L Espeholt; O Vinyals; A Graves",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b62",
"title": "Conditional image generation with pixelcnn decoders",
"year": "2016"
},
{
"authors": "Y Wang; A Gonzalez-Garcia; D Berga; L Herranz; F S Khan; J V D Weijer",
"journal": "",
"ref_id": "b63",
"title": "Minegan: Effective knowledge transfer from gans to target domains with few images",
"year": "2020"
},
{
"authors": "Y Wang; C Wu; L Herranz; J Van De Weijer; A Gonzalez-Garcia; B Raducanu",
"journal": "",
"ref_id": "b64",
"title": "Transferring gans: Generating images from limited data",
"year": "2018"
},
{
"authors": "J Wu; C Zhang; T Xue; B Freeman; J Tenenbaum",
"journal": "Advances in neural information processing systems",
"ref_id": "b65",
"title": "Learning a probabilistic latent space of object shapes via 3d generativeadversarial modeling",
"year": "2016"
},
{
"authors": "J Xiao; L Li; C Wang; Z.-J Zha; Q Huang",
"journal": "",
"ref_id": "b66",
"title": "Few shot generative model adaption via relaxed spatial structural alignment",
"year": "2022"
},
{
"authors": "Y Xu; S Peng; C Yang; Y Shen; B Zhou",
"journal": "",
"ref_id": "b67",
"title": "3d-aware image synthesis via learning structural and textural representations",
"year": "2022"
},
{
"authors": "G Yang; X Huang; Z Hao; M.-Y Liu; S Belongie; B Hariharan",
"journal": "",
"ref_id": "b68",
"title": "Pointflow: 3d point cloud generation with continuous normalizing flows",
"year": "2019"
},
{
"authors": "F Yu; A Seff; Y Zhang; S Song; T Funkhouser; J Xiao",
"journal": "",
"ref_id": "b69",
"title": "Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop",
"year": "2015"
},
{
"authors": "R Zhang; P Isola; A A Efros; E Shechtman; O Wang",
"journal": "",
"ref_id": "b70",
"title": "The unreasonable effectiveness of deep features as a perceptual metric",
"year": "2018"
},
{
"authors": "Y Zhang; M Yao; Y Wei; Z Ji; J Bai; W Zuo",
"journal": "",
"ref_id": "b71",
"title": "Towards diverse and faithful one-shot adaption of generative adversarial networks",
"year": "2022"
},
{
"authors": "Z Zhang; Y Liu; C Han; T Guo; T Yao; T Mei",
"journal": "",
"ref_id": "b72",
"title": "Generalized one-shot domain adaption of generative adversarial networks",
"year": "2022"
},
{
"authors": "S Zhao; Z Liu; J Lin; J.-Y Zhu; S Han",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b73",
"title": "Differentiable augmentation for data-efficient gan training",
"year": "2020"
},
{
"authors": "Y Zhao; K Chandrasegaran; M Abdollahzadeh; N.-M Cheung",
"journal": "",
"ref_id": "b74",
"title": "Few-shot image generation via adaptation-aware kernel modulation",
"year": "2022"
},
{
"authors": "Y Zhao; H Ding; H Huang; N.-M Cheung",
"journal": "",
"ref_id": "b75",
"title": "A closer look at few-shot image generation",
"year": "2022"
},
{
"authors": "Y Zhao; C Du; M Abdollahzadeh; T Pang; M Lin; S Yan; N.-M Cheung",
"journal": "",
"ref_id": "b76",
"title": "Exploring incompatible knowledge transfer in few-shot image generation",
"year": "2023"
},
{
"authors": "Z Zhao; Z Zhang; T Chen; S Singh; H Zhang",
"journal": "",
"ref_id": "b77",
"title": "Image augmentations for gan training",
"year": "2020"
},
{
"authors": "L Zhou; Y Du; J Wu",
"journal": "",
"ref_id": "b78",
"title": "3d shape generation and completion through point-voxel diffusion",
"year": "2021"
},
{
"authors": "P Zhou; L Xie; B Ni; Q Tian",
"journal": "",
"ref_id": "b79",
"title": "Cips-3d: A 3d-aware generator of gans based on conditionally-independent pixel synthesis",
"year": "2021"
},
{
"authors": "J Zhu; H Ma; J Chen; J Yuan",
"journal": "",
"ref_id": "b80",
"title": "Few-shot image generation via masked discrimination",
"year": "2022"
},
{
"authors": "J Zhu; H Ma; J Chen; J Yuan",
"journal": "",
"ref_id": "b81",
"title": "Few-shot image generation with diffusion models",
"year": "2022"
},
{
"authors": "P Zhu; R Abdal; J Femiani; P Wonka",
"journal": "",
"ref_id": "b82",
"title": "Mind the gap: Domain gap control for single shot domain adaptation for generative adversarial networks",
"year": "2021"
}
] | [
{
"formula_coordinates": [
5,
197.86,
176.47,
347.25,
17.88
],
"formula_id": "formula_0",
"formula_text": "p s,l geo,i = sf m( sim(S s,l geo (ω i 1 ), S s,l geo (ω j 1 )) ∀i =j ),(1)"
},
{
"formula_coordinates": [
5,
197.86,
200.24,
347.25,
17.88
],
"formula_id": "formula_1",
"formula_text": "p t,l geo,i = sf m( sim(S t,l geo (ω i 1 ), S t,l geo (ω j 1 )) ∀i =j ),(2)"
},
{
"formula_coordinates": [
5,
184.06,
278.34,
361.06,
13.68
],
"formula_id": "formula_2",
"formula_text": "L geo (S s geo , S t geo ) = E z i D KL (p t,l geo,i ||p s,l geo,i ),(3)"
},
{
"formula_coordinates": [
5,
154.35,
363.76,
390.76,
17.88
],
"formula_id": "formula_3",
"formula_text": "p s mask,i = sf m( sim(M ask(G s (z i 1 , z i 2 )), M ask(G s (z j 1 , z j 2 ))) ∀i =j ),(4)"
},
{
"formula_coordinates": [
5,
154.35,
387.54,
390.76,
17.88
],
"formula_id": "formula_4",
"formula_text": "p t mask,i = sf m( sim(M ask(G t (z i 1 , z i 2 )), M ask(G t (z j 1 , z j 2 ))) ∀i =j ),(5)"
},
{
"formula_coordinates": [
5,
174.21,
442.52,
370.9,
21.98
],
"formula_id": "formula_5",
"formula_text": "L mask (G s , G t ) = E z i 1 ,z i 2 ∼N (0,I) i D KL (p t mask,i ||p s mask,i ).(6)"
},
{
"formula_coordinates": [
5,
182,
599.81,
363.11,
13.56
],
"formula_id": "formula_6",
"formula_text": "M s i,j = M ask(G s (z i 1 , z i 2 )) ∧ M ask(G s (z j 1 , z j 2 )) (i = j),(7)"
},
{
"formula_coordinates": [
5,
182,
617.01,
363.11,
13.56
],
"formula_id": "formula_7",
"formula_text": "M t i,j = M ask(G t (z i 1 , z i 2 )) ∧ M ask(G t (z j 1 , z j 2 )) (i = j).(8)"
},
{
"formula_coordinates": [
5,
149.85,
669.35,
395.27,
17.88
],
"formula_id": "formula_8",
"formula_text": "p s,m tex,i = sf m( sim(S s,m tex (ω i 1 , ω i 2 ) ⊗ M s i,j , S s,m tex (ω j 1 , ω j 2 ) ⊗ M s i,j ) ∀i =j ),(9)"
},
{
"formula_coordinates": [
5,
149.85,
693.12,
395.27,
17.88
],
"formula_id": "formula_9",
"formula_text": "p t,m tex,i = sf m( sim(S t,m tex (ω i 1 , ω i 2 ) ⊗ M t i,j , S t,m tex (ω j 1 , ω j 2 ) ⊗ M t i,j ) ∀i =j ),(10)"
},
{
"formula_coordinates": [
6,
129.78,
118.35,
415.33,
17.88
],
"formula_id": "formula_10",
"formula_text": "p s rgb,i = sf m( sim(RGB(G s (z i 1 , z i 2 )) ⊗ M s i,j , RGB(G s (z j 1 , z j 2 )) ⊗ M s i,j ) ∀i =j ),(11)"
},
{
"formula_coordinates": [
6,
129.78,
142.12,
415.33,
17.88
],
"formula_id": "formula_11",
"formula_text": "p t rgb,i = sf m( sim(RGB(G t (z i 1 , z i 2 )) ⊗ M t i,j , RGB(G t (z j 1 , z j 2 )) ⊗ M t i,j ) ∀i =j ),(12)"
},
{
"formula_coordinates": [
6,
180.14,
198.67,
364.98,
22.66
],
"formula_id": "formula_12",
"formula_text": "L tex (S s tex , S t tex ) = E z i 1 ,z i 2 ∼N (0,I) m,i D KL (p t,m tex,i ||p s,m tex,i ),(13)"
},
{
"formula_coordinates": [
6,
192.29,
227.89,
348.68,
21.98
],
"formula_id": "formula_13",
"formula_text": "L rgb (G s , G t ) = E z i 1 ,z i 2 ∼N (0,I) i D KL (p t rgb,i ||p s rgb,i ). (14"
},
{
"formula_coordinates": [
6,
540.96,
230.29,
4.15,
8.64
],
"formula_id": "formula_14",
"formula_text": ")"
},
{
"formula_coordinates": [
6,
151.43,
339.75,
393.68,
28.9
],
"formula_id": "formula_15",
"formula_text": "L = L(D mask , G t ) + µL reg + µ 1 L geo (S s geo , S t geo ) + µ 2 L mask (G s , G t ) + µ 3 L tex (S s tex , S t tex ) + µ 4 L rgb (G s , G t ).(15)"
},
{
"formula_coordinates": [
15,
128.15,
668.82,
412.81,
12.69
],
"formula_id": "formula_16",
"formula_text": "L(D x , G t ) = E z∈N [g(D x (R(G t (z))))] + E Ix∈px g(-D x (I x )) + λ||∇D x (I x )|| 2 2 , (16"
},
{
"formula_coordinates": [
15,
540.96,
671.22,
4.15,
8.64
],
"formula_id": "formula_17",
"formula_text": ")"
},
{
"formula_coordinates": [
16,
175.71,
476.23,
369.4,
20.14
],
"formula_id": "formula_18",
"formula_text": "L reg = i,j∈Se,i =j H(σ(s i ), sign(s j )) + H(σ(s j ), sign(s i )).(17)"
},
{
"formula_coordinates": [
17,
120,
515.74,
425.11,
12.69
],
"formula_id": "formula_19",
"formula_text": "L = L(D mask , G t ) + L(D rgb , G t ) + µL reg + µ 1 L geo (S s geo , S t geo ) + µ 2 L mask (G s , G t )(18)"
}
] |
|
10.18653/v1/2021.acl-long.81 | 2023-05-19 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b2",
"b10",
"b21",
"b20",
"b19"
],
"table_ref": [],
"text": "Sentiment Analysis (SA) systems are among the most widely deployed NLP systems, used in hundreds of languages (Chen and Skiena, 2014). It is well-known that English SA models exhibit gender and racial biases (Kiritchenko and Mohammad, 2018;Thelwall, 2018;Sweeney and Najafian, 2020), which are acquired from their training data, training objective, and other system choices (Suresh and Guttag, 2019). Other languages are understudied; though many papers study SA bias in English, few study SA bias in other languages. This may be partly attributable to resource constraints: there are fewer corpora available to audit systems for bias in non-English languages. To remedy this, we create evaluation datasets to evaluate gender and "
},
{
"figure_ref": [
"fig_1"
],
"heading": "その人との会話はむかつかた。",
"publication_ref": [
"b10"
],
"table_ref": [],
"text": "The conversation with that person is annoying.\nThe conversation with that Korean person is annoying. vary a single demographic variable (e.g. race). We measure bias as the difference in scores for the pair. An unbiased model should be invariant to the counterfactual, with a difference of zero.\nracial bias in four languages: Japanese (ja), simplified Chinese (zh), Spanish (es), German (de). Each of these four languages has publicly available data for training SA systems (Keung et al., 2020b), and together they represent three distinct language families. To complement their existing resources with a new resource that measures bias, we use counterfactual evaluation (Figure 1), in which test examples are edited to change a single variable of interestsuch as the race of the subject-extending previous work done in English (Kiritchenko and Mohammad, 2018). We release the evaluation dataset to facilitate further research. 1We demonstrate the value of these evaluation resources by answering the following research questions: (RQ1) What biases do we find in other languages, compared to in English? (RQ2) How does the use of pre-trained models affect bias in SA systems? While pre-trained models are common in NLP, they may import biases not present in task supervision data, since a large pre-training corpus may embody biases not present in the supervision corpus. On the other hand, pre-training might diminish biases that arise from the small sample sizes typical of SA training corpora.\nOur experiments show that both gender and racial bias are present in SA systems for all four languages: when model architecture, data quantity, and domain are held constant, SA systems in other languages display quantitatively more bias than SA systems in English. For RQ2, we find that pre-training also makes SA systems less biased for all languages, in aggregate, though in surprising ways: our non-pre-trained models exhibit extreme changes in behaviour on counterfactual examples, whereas pre-trained models exhibit many small nuanced changes."
},
{
"figure_ref": [],
"heading": "New Counterfactual Evaluation Corpus",
"publication_ref": [
"b14",
"b10",
"b1",
"b22",
"b12",
"b0",
"b6",
"b16",
"b0",
"b10"
],
"table_ref": [
"tab_0",
"tab_1"
],
"text": "Counterfactual (or contrastive) evaluation establishes causal attribution by modifying a single input variable, so that any changes in output can be attributed to that intervention (Pearl, 2009). For example, if our variable of interest is gender, and our original sentence is The conversation with that boy was irritating, then our intervention creates the counterfactual sentence The conversation with that girl was irritating. Importantly, we change no other variables, such as age (boy → woman), register (boy → lady), or relationship (boy → sister). We then evaluate the behavior of our model on many such pairs of original and counterfactual sentences. In a model with no gender bias, sentiment should not change under this intervention. If it does, and does so systematically over many counterfactuals, we conclude that our model is biased.\nTo create counterfactual examples for non-English languages we use template sentences, illustrated in Table 1. Each template has a placeholder for a demographic word, in order to represent the counterfactual; and an emotion word, in order to represent different levels of sentiment polarity.\nThe templates of Kiritchenko and Mohammad (2018) only needed to handle the weak agreement and inflectional morphology of English, so we extend their methodology to handle a variety of grammatical phenomena in other languages. For example, in German we add gender agreement (masculine, feminine, neuter) and noun declension; in Spanish we add gender agreement (masculine, fem-inine, plural of both) and idiomatic verb usage;2 in Japanese we add a distinction between active and passive forms. Chinese requires no special handling since it lacks gender agreement or inflectional morphology.\nIn all languages, we create a gender bias test set by providing contrasting pairs of male/female terms that can fill the placeholder for demographic variable. In German and Japanese we also provide pairs of terms for racial and anti-immigrant bias, which we derive from NGOs, sociology and anthropology resources, and government census data (Buckley, 2006;Weiner, 2009;Muigai, 2010;, FADA). We usually leave the privileged group unmarked to avoid the unnaturalness of markedness (Blodgett et al., 2021). 3 For Spanish anti-immigrant bias, we create pairs of names by using name lists that are strongly associated with migrants or with non-migrants, sourced from Goldfarb-Tarrant et al. (2021), which are based on social science research (Salamanca and Pereira, 2013). We lacked equivalent resources for Chinese, so we test only gender bias. The resulting corpora (Table 2) are comparable to or larger than other common contrastive evaluation benchmarks (Blodgett et al., 2021).\nTo produce the templates, we worked alongside native speakers in Japanese, German, Spanish, and Chinese to translate the English templates of Kiritchenko and Mohammad (2018), often modifying them to prefer naturalness in the target language while preserving sentiment. Our Japanese translator had professional translation experience, while our German, Spanish, and Chinese translators had training in linguistics. While collaborative development and refinement of the translation process required about a week, actual translation took about four hours for each dataset. Further details in A."
},
{
"figure_ref": [],
"heading": "Methodology",
"publication_ref": [
"b13"
],
"table_ref": [],
"text": "For our SA task, we focus on sentiment polarity detection (Pang and Lee, 2007), where the output label represents the sentiment of a text as an ordinal score (shown in parentheses): very negative (1), negative (2), neutral (3), positive (4), or very positive (5).4 "
},
{
"figure_ref": [],
"heading": "Metrics",
"publication_ref": [],
"table_ref": [],
"text": "We measure the mean and variance of the differences in sentiment score between each pair of counterfactual sentences. Formally, each corpus consists of n sentences, S = {s i ...s n }, and a demographic variable A = {a, b} where a is the privileged class (male or privileged) and b is the minoritised class (female or racial minority). The sentiment classifier produces a score R for each sentence, and our aggregate measure of bias is:\n1 N n i=0 R(s i | A = a) -R(s i | A = b)\nValues greater than zero indicate bias against the minoritised group, values less than zero indicate bias against the privileged group, and zero indicates no bias. Scores are discrete integers ranging from 1 to 5, so the range of possible values is -4 to 4. Our counterfactual evaluation process enables us to examine bias behaviour more granularly as well. We generate confusion matrices of privileged vs. minoritised scores such that an unbiased model would have all scores along the diagonal. This enables us to distinguish between many minor changes in sentiment or fewer large changes, which are otherwise obscured by aggregate metrics as described above.\nIn results we shade 3% of total range for easier visual inspection. This is an arbitrary choice: 'no bias' differs by application and values within the shaded range may still be unacceptable. Intuitively, this corresponds to models being maximally biased for three of every hundred examples, or making minor biased errors for twelve of every hundred."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [
"b3",
"b17"
],
"table_ref": [],
"text": "We want to answer the questions: what biases arise in SA systems in each of these languages (RQ1)? Does pre-training improve or worsen biases (RQ2)? To answer these questions, we measure the bias of a baseline SVM classification model to a model based on a pre-trained transformer model. We compare standard and distilled transformer models; distilled models are often used in practice since they are better suited to the computational constraints of real-world systems.\nOur baseline (no pre-training) models are bagof-words linear kernel support vector machines (SVMs) trained on the supervision data in each language. Our pre-trained (mono-T) models are pre-trained bert-base (Devlin et al., 2018) for each language. We randomly initialise a linear classification layer and simultaneously train the classifier and fine-tune the language model on the same supervision data. Our distilled (distil-mono-T) models are identical, but based on distilbert-base (Sanh et al., 2019).\nWe train each model five times with different random seeds (or five separate runs for the baseline) and then ensemble by taking their majority vote, a standard procedure to reduce variance. All models converge to performance on par with SotA on this task and data. Training details and F1 scores on the SA task are reported in Appendix B and C.\nTraining data For each model, we use the language appropriate subset of the Multilingual Amazon Reviews Corpus (MARC; Keung et al., 2020a), which contains 200 word reviews in English, Japanese, German, French, Chinese and Spanish, with discrete sentiment labels ranging from 1-5, balanced across labels."
},
{
"figure_ref": [],
"heading": "Results",
"publication_ref": [
"b21"
],
"table_ref": [],
"text": "The baseline models are most biased for both gender and race in all languages (Figure 2), though not always against minoritised groups: systems are often biased against the male demographic, consistent with previous work on SA (Thelwall, 2018). 5 5 Because this task is sentiment analysis, it is more possible to get bias against a male demographic than if the task were, say, biography classification. For the latter, the male demographic is associated with prestige roles (and thus generally bias is anti-female), but for sentiment analysis, male demographics can be associated with negative characteristics (violence, aggression, if a model is stereotyping) as well as with competence, so a few works have found female subjects to sometimes have more positive sentiment, depending on Figure 2 also shows that English models tend to be less biased than the other languages.\nAnalyzing the granular differences (Figure 3) reveals interesting behaviour not captured by aggregate metrics: much of the bias exhibited by the baselines arises from consistently flipping specific labels in the counterfactual, while bias exhibited by pre-trained models is more varied. 6 For example, the Japanese baseline exhibits racial bias by frequently changing neutral labels to very negative labels, whereas in the mono-T model the change under the counterfactual is expressed as many less extreme changes. The model is still biased overall: though the changes are more varied, in aggregate they associate racial minorities with more negative sentiment. The German baseline model is more extreme: when the demographic variable changes from privileged to minoritised, the model changes its prediction from very positive to very negative. The German mono-T model also makes biased choices, though more moderately (neutral to negative) and there is more 'counter-bias' in the upper triangle, which lessens overall bias."
},
{
"figure_ref": [],
"heading": "Related Work and Conclusion",
"publication_ref": [
"b5",
"b7",
"b0",
"b24",
"b11",
"b18"
],
"table_ref": [],
"text": "Counterfactual evaluation is frequently used in bias research on classification tasks (Garg et al., 2019), and sometimes even on generation tasks (Huang et al., 2020). There have also been works exposing common pitfalls in the design of counterfactuals (Blodgett et al., 2021;Zhang et al., 2021;Krishna et al., 2022). Anyone expanding or replicating our counterfactual evaluation work should consult these as prerequisites. The contemporary work of context. 6 We show Japanese and German for illustration; the trend is present in all languages. All graphs are in Appendix D. Seshadri et al. (2022) find many ways that other templates for bias evaluation can be brittle, so future work should take this into account and take measures to ensure robustness, such as testing with multiple paraphrases of the templates.\nWe have laid the groundwork for investigating bias in sentiment analysis beyond English. We created resources, presented an evaluation procedure, and used it to do the first analysis of bias in SA in a simulated low-resource setting across multiple languages. We showed that using pre-trained models produces much less biased models than using baseline SVMs. We also showed that pre-trained models have very different patterns of bias; a type of analysis that is enabled by the counterfactual design of our corpus. We invite the NLP community to use the data and methods from this work to continue analysis of languages beyond English."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [],
"table_ref": [],
"text": "Like all bias tests, these experiments have positive predictive power: they can find the biases they test for, but they cannot eliminate the possibility of there being biases that the tests overlook.\nOur Japanese, German, Spanish, and Chinese translators were from Japan, Germany, Spain, and mainland China, respectively. Hence, their translations may reflect their native dialects of these languages. While these dialects are consistent with the corresponding training datasets in these languages, this fact may limit conclusions that we or others can draw about SA in other dialects of these languages, such as Central and South American dialects of Spanish, or Chinese (Traditional)."
},
{
"figure_ref": [],
"heading": "Ethics Statement",
"publication_ref": [],
"table_ref": [],
"text": "Because of the aforementioned limitation regarding positive predictive power, there is always a risk with research on social biases that it can give practitioners a false sense of security. It is absolutely possible to evaluate on our corpus and get no bias, and still end up causing harm to racial or gender demographics, since they do not cover all biases or all domains. This should be kept in mind whenever applying this research."
},
{
"figure_ref": [],
"heading": "A Benchmark Dataset Creation",
"publication_ref": [
"b0",
"b0",
"b10",
"b0",
"b6",
"b16",
"b10"
],
"table_ref": [],
"text": "We followed the recommendations of Blodgett et al. (2021) to ensure the validity of our datasets. Many of the pitfalls enumerated in their work do not apply to our dataset, as we are measuring sentiment, rather than stereotypes, but we took care to avoid those that do apply. These are:\nMarkedness. In most cases we contrast the minority group, e.g. Turkish people with the unmarked group, e.g. people. Using a marked privileged group-white people, straight people, etcis in most cases uncommon and occurs in only particular settings, which threatens the validity of the contrastive test (Blodgett et al., 2021). We do make a few exceptions and mark privileged groups. We do mark them for gender bias, since gender is explicitly marked in language more than other demographic traits (e.g. we contrast woman with man, not with person). We also sometimes use first names as proxies for demographics such as race, class, and immigration status (in Spanish and English) and in these cases the privileged group is another name.\nNaturalistic Text. Some of the sentences in the original Kiritchenko and Mohammad (2018) would be valid grammatical sentences if translated directly into other languages, but would not sound natural. For example, reflexive pronouns (himself, herself) aren't used the same way in Chinese as in English, so in translating the English template <person subject> found himself/herself in a/an <emotional situation word> situation. we instead used the Chinese template <person subject> 经 历 了 一 件<emotional situation word> 的 事., which means <person subject> was in a <emotional situation word> situation. These small changes preserve the same rough semantics, and more importantly preserve naturalness.\nIndirect Demographic Identification. Blodgett et al. (2021) caution against the use of proper names or other proxies as a stand in for a demographic group, because their reliability for this use is untested. We would add that names are difficult to use in a contrastive pair where we need to change only one demographic variable, because names indicate many bits of demographic information at once: race, gender, class, place of birth, period of birth, etc. We intentionally avoid this by using identity terms (Turk, Korean, etc) most of the time, which do sometimes conflate race and country of origin, but are otherwise the most precise option. We use proper names only in Spanish based on the work of Goldfarb-Tarrant et al. (2021) and Salamanca and Pereira (2013), who show that there is data backing up the migrant vs. non-migrant names. Even so, there is some conflation between migrant status and socioeconomic class in that set of names: we consider that acceptable for our purposes. There are also names as a proxy for African-Americans in English, as the dataset is from Kiritchenko and Mohammad (2018) and that is what they use.\nBasic Consistency A few other applicable pitfalls, which Blodgett et al. ( 2021) capture under the heading 'Basic Control and Consistency' we avoid organically by our template based construction, e.g. differences in sentence length between sentences A and B, are a possible confound, but by construction we contrast only one word in a pair and the sentence is otherwise unperturbed.\nOnce we had designed our translation process, we did a multi-step qualitative evaluation. After we had settled on the first version of the three sets of templates, demographic terms, and emotion words in each language, we worked with the native speaker to iterate and make sure there were no accidental unnatural sentences or grammatical errors. We generated a few examples for each template + emotion + demographic combination, manually reviewed 200 examples per language, and then made corrections to the templates, words and the rules for combining them. We then repeated this exact process a second time after the adjustments."
},
{
"figure_ref": [],
"heading": "B Model Implementation Details",
"publication_ref": [],
"table_ref": [],
"text": "Monolingual transformer models have 110 million parameters (± 1 million) and vocabularies of 30-32k with 768D embeddings. We train the monolingual models with the same training settings as preferred in Keung et al. (2020a), and allow the pretrained weights to fine-tune along with the newly initialised classification layer."
},
{
"figure_ref": [
"fig_4"
],
"heading": "C Model Performance",
"publication_ref": [],
"table_ref": [
"tab_2"
],
"text": "Performance at convergence for models in each language is given in Table 3.\nWe determined convergence by examining loss curves and selecting the model where training loss was flat, and validation had not yet increased. We did not use early-stopping, as we wanted to save many model checkpoints in order to study the training dynamics of bias, including after convergence when the model was overtrained. However, we found no clear trends in how bias changed over the course of training, so for this study we used only one model, at convergence, per language. We hope that by releasing all model checkpoints (15 per language), other researchers may be able to expand our work into the training dynamics of bias. Figure 4: All confusion matrices for experiments in this paper. Higher colour saturation in the lower triangle is bias against the minoritised group, in the upper triangle is bias against the privileged group. Saturations are not normalised across all languages and models; this is not a proxy for aggregate comparative bias, it shows the pattern across sentiment scores."
},
{
"figure_ref": [],
"heading": "Standard",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Acknowledgements",
"publication_ref": [],
"table_ref": [],
"text": "We thank Björn Ross for many comments and helping shape the draft, Lluís Màrquez for helping manage the project at Amazon, and the Amazon Barcelona Search team for their enthusiastic support of the project."
}
] | Sentiment analysis (SA) systems are used in many products and hundreds of languages. Gender and racial biases are well-studied in English SA systems, but understudied in other languages, with few resources for such studies. To remedy this, we build a counterfactual evaluation corpus for gender and racial/migrant bias in four languages. We demonstrate its usefulness by answering a simple but important question that an engineer might need to answer when deploying a system: What biases do systems import from pre-trained models when compared to a baseline with no pre-training? Our evaluation corpus, by virtue of being counterfactual, not only reveals which models have less bias, but also pinpoints changes in model bias behaviour, which enables more targeted mitigation strategies. We release our code and evaluation corpora to facilitate future research. | Bias Beyond English: Counterfactual Tests for Bias in Sentiment Analysis in Four Languages | [
{
"figure_caption": "Figure 1 :1Figure 1: We create corpora and then do counterfactual evaluation to evaluate how bias is transferred from training data. Counterfactual pairs (e.g. sentences a, b) vary a single demographic variable (e.g. race). We measure bias as the difference in scores for the pair. An unbiased model should be invariant to the counterfactual, with a difference of zero.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :Figure 3 :23Figure2: Aggregate bias metrics for baseline (blue), pretrained mono-T (orange), and pretrained distil mono-T (green) models. Mean and variance of differences in the sentiment label under each counterfactual pair, one graph per language and type of bias tested. Higher numbers indicate greater bias against the minoritized group. The dashed line at zero indicates no bias, the shaded region corresponds to 3% of total range (see 3.1). Spanish (es) distilled model is intentionally missing for lack of comparable pretrained model.German",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "23",
"figure_type": "figure"
},
{
"figure_caption": "DFull set of confusion matrices comparing baseline and monolingual models.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 44Figure 4 contains all confusion matrices for all languages, of which we displayed a subset in the body of this work.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "The conversation with <person object> was <emotional situation word>. The conversation with [him\\her] was irritating. ja <person> との会は <emotion word passive>た[彼\\彼女] との会は イライラさた。 zh 跟 <person> 的谈话很 <emotional situation word>.跟 [他\\她] 的谈话很 令人生气. de Das Gespräch mit <person dat. object> war <emotional situation word>. Das Gespräch mit [ihm\\ihr] war irritierend. es La conversación con <person> fue <emotional situation word female>.La conversación con [él\\ella] fue irritante. Example sentence templates for each language and their counterfactual words that, when filled in, create a contrastive pair; in this case, for gender bias. For illustration, all five examples are translations of the same sentence.",
"figure_data": "Template",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "tab_1",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "F1 at convergence and steps at convergence for standard size, distilled, and baseline models. Performance is measured on the MARC data.",
"figure_data": "DistilledBaselineF1Steps F1Steps F1ja 0.62 44370 0.61 60436 0.38zh 0.56 35190 0.53 43750 0.42de 0.63 36720 0.63 52621 0.51es 0.61 41310-0.48en 0.65 27050 0.65 44285 0.53",
"figure_id": "tab_2",
"figure_label": "3",
"figure_type": "table"
}
] | Seraphina Goldfarb-Tarrant; Adam Lopez; Roi Blanco; Diego Marcheggiani | [
{
"authors": "Lin Su; Gilsinia Blodgett; Alexandra Lopez; Robert Olteanu; Hanna Sim; Wallach",
"journal": "",
"ref_id": "b0",
"title": "Stereotyping Norwegian salmon: An inventory of pitfalls in fairness benchmark datasets",
"year": "2021"
},
{
"authors": "Sandra Buckley",
"journal": "Routledge",
"ref_id": "b1",
"title": "Encyclopedia of contemporary Japanese culture",
"year": "2006"
},
{
"authors": "Yanqing Chen; Steven Skiena",
"journal": "Association for Computational Linguistics",
"ref_id": "b2",
"title": "Building sentiment lexicons for all major languages",
"year": "2014"
},
{
"authors": "Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova",
"journal": "",
"ref_id": "b3",
"title": "BERT: pre-training of deep bidirectional transformers for language understanding",
"year": "2018"
},
{
"authors": "",
"journal": "The Federal Anti-Discrimination Agency (FADA)",
"ref_id": "b4",
"title": "Equal rights, equal opportunities: Annual report of the federal anti-discrimination agency",
"year": "2020"
},
{
"authors": "Sahaj Garg; Vincent Perot; Nicole Limtiaco; Ankur Taly; Ed H Chi; Alex Beutel",
"journal": "Association for Computing Machinery",
"ref_id": "b5",
"title": "Counterfactual fairness in text classification through robustness",
"year": "2019"
},
{
"authors": "Seraphina Goldfarb-Tarrant; Rebecca Marchant; Ricardo Muñoz Sánchez; Mugdha Pandya; Adam Lopez",
"journal": "Association for Computational Linguistics",
"ref_id": "b6",
"title": "Intrinsic bias metrics do not correlate with application bias",
"year": "2021"
},
{
"authors": "Po-Sen Huang; Huan Zhang; Ray Jiang; Robert Stanforth; Johannes Welbl; Jack Rae; Vishal Maini; Dani Yogatama; Pushmeet Kohli",
"journal": "Association for Computational Linguistics",
"ref_id": "b7",
"title": "Reducing sentiment bias in language models via counterfactual evaluation",
"year": "2020"
},
{
"authors": "Phillip Keung; Yichao Lu; György Szarvas; Noah A Smith",
"journal": "",
"ref_id": "b8",
"title": "a. The multilingual Amazon reviews corpus",
"year": "2020"
},
{
"authors": "Phillip Keung; Julian Salazar; Yichao Lu; Noah A Smith",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b9",
"title": "Unsupervised bitext mining and translation via self-trained contextual embeddings",
"year": "2020"
},
{
"authors": "Svetlana Kiritchenko; Saif Mohammad",
"journal": "Association for Computational Linguistics",
"ref_id": "b10",
"title": "Examining gender and race bias in two hundred sentiment analysis systems",
"year": "2018"
},
{
"authors": "Satyapriya Krishna; Rahul Gupta; Apurv Verma; Jwala Dhamala; Yada Pruksachatkun; Kai-Wei Chang",
"journal": "Association for Computational Linguistics",
"ref_id": "b11",
"title": "Measuring fairness of text classifiers via prediction sensitivity",
"year": "2022"
},
{
"authors": "Githu Muigai",
"journal": "",
"ref_id": "b12",
"title": "Report of the special rapporteur on contemporary forms of racism, racial discrimination, xenophobia and related intolerance, githu muigai, on his mission to germany",
"year": "2009-07-01"
},
{
"authors": "Bo Pang; Lillian Lee",
"journal": "Found. Trends Inf. Retr",
"ref_id": "b13",
"title": "Opinion mining and sentiment analysis",
"year": "2007"
},
{
"authors": "Judea Pearl",
"journal": "Statistics Surveys",
"ref_id": "b14",
"title": "Causal inference in statistics: An overview",
"year": "2009"
},
{
"authors": "Soujanya Poria; Devamanyu Hazarika; Navonil Majumder; Rada Mihalcea",
"journal": "",
"ref_id": "b15",
"title": "Beneath the tip of the iceberg: Current challenges and new directions in sentiment analysis research",
"year": "2020"
},
{
"authors": "Gastã Salamanca; Lidia Pereira",
"journal": "SUJETOS DE NIVEL EDUCA-CIONAL SUPERIOR. Universum (Talca)",
"ref_id": "b16",
"title": "PRESTI-GIO Y ESTIGMATIZACIÃ\"N DE 60 NOMBRES PROPIOS EN 40",
"year": "2013"
},
{
"authors": "Victor Sanh; Lysandre Debut; Julien Chaumond; Thomas Wolf",
"journal": "",
"ref_id": "b17",
"title": "Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter",
"year": "2019"
},
{
"authors": "Preethi Seshadri; Pouya Pezeshkpour; Sameer Singh",
"journal": "",
"ref_id": "b18",
"title": "Quantifying social biases using templates is unreliable",
"year": "2022"
},
{
"authors": "Harini Suresh; John V Guttag",
"journal": "",
"ref_id": "b19",
"title": "A framework for understanding unintended consequences of machine learning",
"year": "2019"
},
{
"authors": "Chris Sweeney; Maryam Najafian",
"journal": "Association for Computing Machinery",
"ref_id": "b20",
"title": "Reducing sentiment polarity for demographic attributes in word embeddings using adversarial learning",
"year": "2020"
},
{
"authors": "Mike Thelwall",
"journal": "Online Information Review",
"ref_id": "b21",
"title": "Gender bias in sentiment analysis",
"year": "2018"
},
{
"authors": "Michael Weiner",
"journal": "",
"ref_id": "b22",
"title": "Japan's minorities: the illusion of homogeneity",
"year": "2009"
},
{
"authors": "Francis Taylor",
"journal": "",
"ref_id": "b23",
"title": "",
"year": ""
},
{
"authors": "Chong Zhang; Jieyu Zhao; Huan Zhang; Kai-Wei Chang; Cho-Jui Hsieh",
"journal": "Association for Computational Linguistics",
"ref_id": "b24",
"title": "Double perturbation: On the robustness of robustness and counterfactual bias evaluation",
"year": "2021"
}
] | [
{
"formula_coordinates": [
3,
96.27,
532.46,
168.66,
33.71
],
"formula_id": "formula_0",
"formula_text": "1 N n i=0 R(s i | A = a) -R(s i | A = b)"
}
] |
2023-08-17 | [
{
"figure_ref": [
"fig_0",
"fig_0",
"fig_1"
],
"heading": "Introduction",
"publication_ref": [
"b40",
"b7",
"b28",
"b16",
"b28",
"b37"
],
"table_ref": [],
"text": "Image composition aims to synthesize foreground objects from one image into another, which is a common task in image editing. However, human eyes could clearly dis- tinguish synthetic images due to the visual inconsistency between foreground and background in composited images. In attempting to solve the photo-unrealistic problem, image harmonization is proposed to adjust the foreground objects based on the illumination and color tone in background environment, which plays an important role in image editing.\nTraditional image harmonization approaches are mainly based on low-level feature matching, which is only effective for specific scenes. Recently, numerous learning-based methods have achieved remarkable progress by addressing image harmonization as a generation task. Existing learning-based methods could be categorized from two angles, i.e., local-translation and region-matching. 1) The former employs a convolutional encoder-decoder to learn a foreground pixel-to-pixel translation [41,8]. But a shallow CNN only captures limited surrounding background. As shown in Figure 1a, these approaches harmonize the current pixel with local references, which is insufficient for harmonization as inner foreground pixels could not attach background reference. Besides, related long-distance references are effective in some cases. 2) The latter region matching methods [29,7] distinguish foreground and background regions as two styles or domains. As shown in Figure 1b, they tackle harmonization as a matching problem with a unified view of these two regions by statistics components or discriminators. Though these approaches harmonize images with a broader reference range, they totally neglect the spatial variations in two regions. Hang et al. [17] begin to notice this problem and add attention-based references in region matching method [29]. But they still separate two regions independently and harmonize foreground by unified matching without considering foreground spatial variations.\nTo further illustrate existing problems, we provide two common harmonization cases in Figure 2. In the first case, a small foreground object appears in the background with obvious color changes. The region-matching method Rain-Net [7] provides a poor color correction result while the local method iS 2 AM [38] could tackle this case well, which indicates that the unified view of background will blend the overall complex color conditions. In the second case, related long-distance references exist in the background, while the local method could only attach insufficient adjacent information. Region-matching method RainNet could obtain whole background blue tone by matching, but it still excessively harmonizes the house due to the unified view. These two cases indicate that local reference is insufficient, but region-matching methods could not model longdistance reference well and will cause unbalanced harmonization problems by rough matching.\nTo solve this problem, we rethink essential proximity priors in image harmonization, i.e., when we paste an object into background, the color or light is related to location and will be influenced by its neighboring first. Moreover, the effective long-distance information in background changes with pasted locations, which requires us to learn adaptive references for each part. Inspired by this observation, we propose a novel Global-aware Kernel Network (GKNet) to integrate local harmony modulation and longdistance background references, including harmony kernel prediction and harmony kernel modulation. For harmony kernel prediction, we propose a novel global-aware kernel prediction method including Long-distance Reference Extractor (LRE) to obtain long-distance references and Ker-nel Prediction Blocks (KPB) to predict multi-level adaptive kernels with selected long-distance references by Selective Correlation Fusion (SCF). For kernel modulation, we propose to model local harmony operation by predicted globalaware kernels and multi-level features. Focusing on features in kernel region, kernel modulation is significant in alleviating unbalanced region-matching errors in complex scenes.\nTo summarize, we make following contributions: "
},
{
"figure_ref": [],
"heading": "Related work 2.1. Image Harmonization",
"publication_ref": [
"b35",
"b38",
"b24",
"b40",
"b37",
"b40",
"b7",
"b28",
"b7",
"b14",
"b16",
"b28",
"b15",
"b3"
],
"table_ref": [],
"text": "Traditional image harmonization works have focused on finding a better method for low-level appearances matching between foreground and background regions in images, which includes color statistics [36,35,45], gradient information [20,34,40], and multi-scale statistical features [39,25]. However, traditional methods could only be effective in specific scenes. With the advanced generative ability of deep learning, Tsai et al. [41] firstly propose a learning-based encoder-decoder network assisted by a semantic branch. In observation that semantic information is effective in image harmonization, Soffiuk et al. [38] also add additional pre-train semantic model to baseline DIH [41] and S 2 AM [8]. Inspired by domain transfer, Cong et al.\n[7] adopt a verification discriminator to distinguish foreground and background domains. Similarly, Ling et al. [29] also treat the composited image as two independent parts and apply style transfer idea to match mean-variance statistics. To focus on harmonize foreground region, some methods add attention mechanisms. Cun et al. [8] add a spatial-separated attention module. Guo et al. [15] for the first time introduce Transformer architecture to image harmonization. Hang et al. [17] add background attention calculation to the style transfer block [29], and they also incorporated the idea of contrast learning. Besides, Guoet al. [16] decompose image into reflectance and illumination by autoencoder for separate harmonization based on Retinex theory. Some high-resolution methods [24,44] frame image harmonization as an image-level problem to learn white-box arguments. However, the above methods neglect spatial proximity prior and could not model longdistance references well. Instead in this paper, we design a better local modulation method combined with selected long-distance references to alleviate this problem."
},
{
"figure_ref": [],
"heading": "Dynamic Filtering in Image Editing",
"publication_ref": [
"b22",
"b5",
"b8",
"b0",
"b41",
"b25",
"b1"
],
"table_ref": [],
"text": "The input-dependent dynamic filtering first proposed by Jia [21] et al. aims to learn position-specific filters on pixel inputs and apply the generated kernels to another input, which has been widely used in numerous vision tasks [33,23,46,9]. This method also shows effectiveness in image editing tasks, such as denoising [1,32,42], shadow removing [13], deraining [14], image inpainting [26], and blur synthesis [2]. However, most above methods apply dynamic filtering at image-level filter prediction and utilization. We propose to learn a multi-level global-aware kernel with long-term context references for harmonization."
},
{
"figure_ref": [],
"heading": "Feature Fusion",
"publication_ref": [
"b17",
"b49",
"b47",
"b11",
"b29",
"b14"
],
"table_ref": [],
"text": "Feature fusion is to combine features from different layers or branches, which is an omnipresent part of modern neural networks and has been studied extensively. Most previous works [18,37,28] for feature fusion focus on the pathways structure design, applying two linear classic methods of summation or concatenation. Recently, benefit from the successful use of Transformer in computer vision [4,11,31,50,49,48,3,4,12,43,22,30,15], some works [19, 27, 52, 10, 47, 51, 5] apply attention mechanism to present nonlinear approaches for feature fusion. As global information is significant in image harmonization, we design a dynamic weighted fusion method to effectively fuse long-distance reference into kernel prediction."
},
{
"figure_ref": [],
"heading": "Method",
"publication_ref": [],
"table_ref": [],
"text": "Given a composited image I c with its corresponding binary mask M indicating the region to be harmonized, our goal is to learn a network G that outputs harmonized image I h , which could be formulated as I h = G(I c , M). To make this composited image I c look natural, we train our model G to adjust the foreground region I f in a supervised manner with paired real image I. In this paper, we also define the background region as I b , and then the composition process could be formulated as\nI c = I f • M + (1 -M) • I b\n, where • denotes element-wise multiplication."
},
{
"figure_ref": [
"fig_2"
],
"heading": "Overview of Our Network",
"publication_ref": [
"b7",
"b37",
"b28",
"b7"
],
"table_ref": [],
"text": "As shown in Figure 3, we design a novel network architecture for image harmonization tasks to allow our network to pay attention to short-distance and long-distance information simultaneously. Following the standard designs in image harmonization works [8,38,29] we use simple U-Net [37] with attention blocks [8] as the basic structure. We also take composited RGB image I c ∈ R 3×H×W concatenated with foreground region mask M ∈ R 1×H×W as input.\nMotivated by proximity prior in image harmonization, we propose Global-aware Kernel Network (GKNet) to learn global-aware harmony kernel for image harmonization, which consists of two branches, harmony kernel prediction and harmony kernel modulation. Firstly, as longdistance reference is crucial for harmonization task, we design global-aware kernel prediction branch to predict harmony kernel with context modeling, which contains a transformer-based Long-term Reference Extractor (LRE) to extract global reference and Kernel Prediction Block (KPB) to predict harmony kernels. In order to incorporate relevant long-term references for local harmonization, a novel Selective Correlation Fusion (SCF) is proposed to select more effective references in backgrounds. Secondly, we design a multi-level harmony kernel modulation in decoder layers to employ the predicted global-aware kernels. The mechanism between global-aware harmony kernel prediction and harmony kernel modulation finally achieves local-global interaction for image harmonization."
},
{
"figure_ref": [
"fig_2",
"fig_3"
],
"heading": "Harmony Kernel Prediction",
"publication_ref": [
"b0"
],
"table_ref": [],
"text": "Inspired by recent works in image editing [1,14,32], we propose to apply dynamic kernels to force the current pixel harmonized with surrounding regions adaptively. This approach effectively makes up for the lack of consideration of proximity prior in previous harmonization works. However, basic dynamic kernels for image editing tasks such as denoising and deraining are applied with a fixed size at imagelevel. To predict more proper adaptive kernels for image harmonization, we analyze the following: 1) Global modeling is necessary for image harmonization as long-distance references may appear in the background. Hence, we design a novel global-aware harmony kernel prediction branch with LRE to extract global information and KPB to predict global-aware kernels with fusion module SCF. 2) Fixed-size kernels applied at image-level could not handle the scale problem well. e.g. The pixels inside the large foreground mask can hardly obtain any real background information, while predicting large kernels will bring high computation costs and breaks the intention of proximity prior. Besides, image-level dynamic kernels pay more attention to detailed structure, while in image harmonization we also need to adapt multiple scene variations to harmonize foregrounds at semantic level. In order to adapt to multi-scene and multiscale problems, we propose to predict kernels in multi-level structures.\nLong-distance Reference Extractor. In order to obtain long-term context, we employ l-transformer layers [11] as our global information extractor. We feed the deepest feature map F 1 E from CNN encoder into transformer layers. With the down-sampling feature map in low-resolution of ( w r , h r ), we treat each pixel as a token to generate embeddings. With the multi-head attention mechanism, we obtain global interactive feature F global ∈ R C×HW after ltransformer layers. After reshaping and post-convolution layer, we obtain the long-term reference feature F LR .\nKernel Prediction Block. To adapt diverse foreground scales and background scenarios, we apply our local operation kernel modulation in multiple decoder levels. Nevertheless, deep-level features contain more semantic information, and shallow features contain more details, we need to predict corresponding adaptive kernels for different level harmony kernel modulation. Thus, our designed globalaware harmony kernel prediction branch is in a multi-level structure to increasingly predict a series of kernels.\nIn Figure 3, we show our proposed KPB structure from predicting kernels to harmony kernel modulation operation. The operation in lth KPBlock can be formulated as\nK l , F l KPB = KPBlock(F l E , F l-1 KPB ),(1)\nwhere KPBlock(•) is the KPBlock to predict K l . For each KPB, we take F l-1 KPB ∈ R C l ×H l ×W l transferred from (l -1)th KPB (For the deepest KPB, we input F LR ) and the lth encoder layer feature F l E as input (We denote F 1 E as the deepest Specifically, as shown in Figure 4, we take the (l -1)th layer feature F l-1 KPB ∈ R C l-1 ×H l-1 ×W l-1 and the lth encoder feature F l E ∈ R C l ×H l ×W l as input and extract attention vector α KPB , α E ∈ R C l by 3 × 3 convolutions and MLP. Subsequently, the attention vector is divided into n groups with length m as αKPB , αE ∈ R n×m . Thus, the channel-wise attention relation A ∈ R n×n can be calculated by matrix product\nA = αKPB ⊙ αT E .(2)\nAfter that, we calculate the selective factor α KPB , α E ∈ R C l through N × N convolutions with splitting. Then, we obtain the selective attention weights S l ∈ {S l E , S l-1 KPB } for each features by α ∈ {α KPB , α E } and α ∈ {α KPB , α E }, which can be formulated as\nS l = σ(α + b • FC(α)),(3)\nwhere b is a learnable parameter, σ is sigmoid function.\nBased on the attention weight vector, then the shallow and deep information are interacted as follows:\nFl KPB = S l E •Conv(F l E )+Upsample(S l-1 KPB •Conv(F l-1 KPB )),(4)\nwhere • denotes to element-wise multiplication."
},
{
"figure_ref": [
"fig_2"
],
"heading": "Harmony Kernel Modulation",
"publication_ref": [],
"table_ref": [],
"text": "The global-aware adaptive kernel obtained from the harmony kernel prediction branch is then utilized in harmony kernel modulation. In this section, we illustrate the regional harmony operation method kernel modulation in decoder layers, which converts the previous overall treatment of background and foreground into a local reference. As mentioned in Section 3.2, we apply multi-level kernel modulation in decoder layer to adapt scales and scenario variation problem. Figure 3 shows our proposed kernel modulation in lth decoder layers, which could be formulated as\nFl D = F l D ⊛ K l ,(5)\nwhere ⊛ denotes the harmony kernel modulation, F l D ∈ R C×H×W is the deep feature extracted from the lth layer in decoder, and the Fl D ∈ R C×H×W is its corresponding harmony kernel modulation result feature in the lth layer. The tensor K l ∈ R C×N 2 ×H×W represents the kernels with size of N for harmony kernel modulation in the lth feature layer, which we obtain from KPB. For the kernel modulation in each pixel, we can expand the above equation as\nFl D [p] = q∈N p K l p [p -q]F l D [q],(6)\nwhere p and q are the coordinates of pixels in the image, K l p is the kernel for filtering the element p of F l D via its surrounding elements, i.e.,N p . As we illustrate in Eq. 5, K l contains all element-wise kernels,i.e., K l p ∈ R C×N×N for filtering operations. After the kernel modulation in decoder layers, we finally obtain the harmonization result. In this paper, we define the U-Net decoder layer as ϕ D (•), then we can formulate our harmonization process with kernel modulation as Îh\n= ϕ L D (• • • ϕ 1 D (F 1 D ⊛ K 1 )) • M + (1 -M) • I c ."
},
{
"figure_ref": [],
"heading": "Objective function",
"publication_ref": [
"b37"
],
"table_ref": [],
"text": "In the training phase, we only employ foregroundnormalized MSE loss as our objective function. Compared with normal MSE loss, it reduces the impact of copying background area:\nL rec = h,w Î -I 2 2 max A min , h,w M h,w ,(7)\nwhere A min is a hyperparameter to keep the loss function stable as there might be some too small foreground objects.\nIn this paper, we set A min = 100 as suggested in [38]."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Implementation Details",
"publication_ref": [],
"table_ref": [],
"text": "We conduct the image harmonization experiment at resolution 256 × 256 on the benchmark dataset iHarmony4 [7]. The initial learning rate is set to 10 -4 , and the models are trained for 120 epochs with a batch size of 16 on four 2080Ti GPUs. For optimizer, we adopt an Adam optimizer with β 1 = 0.9, β 2 = 0.999 and ϵ = 10 -8 . It takes about two days for training. Our proposed model is implemented by PyTorch, and more detailed network architectures could be found in the supplementary file."
},
{
"figure_ref": [],
"heading": "Datasets and Metrics",
"publication_ref": [
"b37"
],
"table_ref": [],
"text": "Datasets. To evaluate our proposed method for image harmonization, we conduct our experiments on the benchmark dataset iHarmony4 [7], which consists of 4 sub-datasets: HCOCO, HAdobe5K, HFlickr, and Hday2night, including 73147 pairs of synthesized composite images with their corresponding foreground mask and ground truth image. Evaluation Metrics. Following the standard setups in image harmonization, we use the peak signal-to-noise ratio (PSNR) and Mean Squared Error (MSE) as evaluation metrics. Furthermore, it is more accurate to only calculate the difference in the foreground region with the metric foreground MSE (fMSE) introduced by [38]. The metric MSE is calculated for the whole image, while there is no changes for pixels in the background region in harmonization task. Without considering the foreground ratio, the average MSE and PSNR results of the dataset will be more responsive to the performance of large-scale targets. In this paper, we argue that fMSE is more suitable for harmonization task."
},
{
"figure_ref": [],
"heading": "Comparison with SOTAs",
"publication_ref": [
"b37",
"b40"
],
"table_ref": [
"tab_1",
"tab_2",
"tab_2"
],
"text": "Quantitative comparison. As results shown in Table 1, we compare our method with other state-of-the-art image harmonization methods on iHarmony4 [7]. Following [7,38], we also evaluate the model performance on different ratios of foreground by splitting the test images into three groups,i.e., 0% ∼5%, 5%∼15%, and 15%∼100%. We provide these results in Table 2. Observing the quantitative experiment results above, we can summarize the following conclusions: 1) Our method achieves SOTA results of all evaluation metrics on average iharmony4 datasets. More specifically, our method achieves 0.78dB↑ improvement in PSNR, 1.43↓ in MSE, and 28.42↓ in fMSE compared with suboptimal methods. 2) Our method obtains the best fMSE scores on all sub-datasets, meaning that the foreground regions generated by our method are more natural and closer to real images. 3) As shown in Table 2, our model performs well on each foreground ratio, especially on 0%∼5%, which indicates that our method has a strong ability to handle variable foreground scales. It also proves that our proposed global-aware adaptive kernel could process global-to-local impact with excellence. Qualitative comparison. We further provide qualitative results on iHarmony4 datasets in Figure 5. It could be observed that our GKNet generates more visually coherent images of the foreground and background than composited images, which are also closer to ground truth. 1) The first two rows of examples show that our method can effectively handle the small object harmonization problem in complex scenes with proximity prior consideration.\n2) The last two row examples show that our regional modulation with multi-level structure performs well in large foreground cases, which handles a wide range of enormous contrast. For more detailed descriptions, please refer to the caption in Figure 5. In contrast, our method presents more photorealistic results. More visual comparison results on iHarmony4 datasets could be seen in supplementary materials. real composition datasets [41] can be seen in Figure 6. For real composition cases, evaluation metrics are impossible to calculate as there are no ground truth images. Hence, we only show qualitative results and human study results here. More visual comparison results on real datasets could be seen in supplementary materials. "
},
{
"figure_ref": [
"fig_6",
"fig_7"
],
"heading": "Ablation Studies",
"publication_ref": [],
"table_ref": [],
"text": "Effectiveness of network components. We further conduct quantitative experiments to verify the effectiveness of components in GKNet. Note that modules in GKNet have dependencies, we can only show gradually added ablation studies. As the results in Interpretability of Global-aware Harmony Kernel. To further illustrate the effectiveness of the adaptive harmony kernels, we cluster the per-pixel adaptive harmony kernels predicted from KPB by K-means. As shown in Figure 8, the clusters show strong spatial structure, which indicates that our predicted dynamic kernel can make the structural adjustment to harmonize the foreground. Moreover, in some cases, the change of kernel classes is related to the object mask. This exhibits that our harmony kernels are predicted dynamically to deal with visual inconsistency in different spatial locations (e.g. fore-/-background or edges). Interpretability of LRE and SCF Module. We visualize the attention mechanism in LRE and SCF for interpreting global-local interaction. In Figure 9, we visualize feature maps in decoder layers to illustrate our channel-wise attention module SCF. Visual feature maps with attention weights show that harmony kernels are predicted with more attention on related background area and less attention on irrelevant background area or foreground. In Figure 10, we visualize the attention maps of LRE, focusing on an example point in foreground. As the long-distance information in predicted harmony kernels is brought by LRE, the visualized attention maps in different heads indicate two points:\n1) The kernels for local operation have a global perceptive field with long-distance information. 2) Different heads pay attention to different reference parts, i.e. relevant background reference, foreground content, overall tone, etc."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "This paper proposes an effective network GKNet to learn global-aware harmony kernels for image harmonization, including harmony kernel prediction and harmony kernel modulation branches. For harmony kernel prediction, we propose LRE to extract long-term references and KPB to predict global-aware kernels. To better fuse long-term context, we design SCF to select relevant references. For harmony kernel modulation, we employ the predicted kernels for harmonization with location awareness. Extensive experiments demonstrate that our proposed algorithm outperforms the state-of-the-art algorithms on the iHarmony4 dataset and real image composition datasets."
},
{
"figure_ref": [],
"heading": "",
"publication_ref": [],
"table_ref": [],
"text": "Acknowledgements. This work was supported by a Grant from The National Natural Science Foundation of China (No. 2021YFB2012300)."
}
] | Image harmonization aims to solve the visual inconsistency problem in composited images by adaptively adjusting the foreground pixels with the background as references. Existing methods employ local color transformation or region matching between foreground and background, which neglects powerful proximity prior and independently distinguishes fore-/back-ground as a whole part for harmonization. As a result, they still show a limited performance across varied foreground objects and scenes. To address this issue, we propose a novel Global-aware Kernel Network (GKNet) to harmonize local regions with comprehensive consideration of long-distance background references. Specifically, GKNet includes two parts, i.e., harmony kernel prediction and harmony kernel modulation branches. The former includes a Long-distance Reference Extractor (LRE) to obtain long-distance context and Kernel Prediction Blocks (KPB) to predict multi-level harmony kernels by fusing global information with local features. To achieve this goal, a novel Selective Correlation Fusion (SCF) module is proposed to better select relevant long-distance background references for local harmonization. The latter employs the predicted kernels to harmonize foreground regions with local and global awareness. Abundant experiments demonstrate the superiority of our method for image harmonization over state-of-the-art methods, e.g., achieving 39.53dB PSNR that surpasses the best counterpart by +0.78dB ↑; decreasing fMSE/MSE by 11.5%↓/6.7%↓ compared with the SoTA method. Code will be available at here. | Learning Global-aware Kernel for Image Harmonization | [
{
"figure_caption": "Figure 1 .1Figure 1. Comparison of background reference methods in harmonization. Blue/Red region represent foreground/background, respectively, and white/red arrows refer to interaction/injection, respectively. (a) Local-translation methods reference nearby pixels. (b) Region-matching methods transfer reference with a unified view of fore-/back-ground region. (c) Our method interacts longdistance reference and injects it with short-distance consideration.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 .2Figure 2. Left: Two challenging samples in image harmonization. Mask in column one and Red boxes represents the foreground. Right: Performance comparison with SOTA methods in terms of PSNR and model size. The circle size represents the floating-point number.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 .3Figure 3. The overview of our proposed GKNet, which consists of harmony kernel prediction branch and harmony kernel modulation branch. As shown in gray box, the harmony kernel prediction branch is combined with a Long-term Reference Extractor (LRE) and multilevel Kernel Prediction Blocks (KPB). As shown in yellow box, we propose Selective Correlation Fusion (SCF) module in KPB for better long-distance references. Given a composited image I c with corresponding foreground mask M, we extract deep features F l E from encoder ϕ E . Then, harmony kernel prediction branch utilizes the deepest feature map and {F l E } to predict multi-level dynamic harmony kernels {K l } increasingly. The predicted global-aware kernels are employed for harmony kernel modulation in decoder ϕ D .",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 .4Figure 4. Schematic diagram of SCF. The module takes the (l-1)th layer feature F l-1 KPB and encoder feature F l E as input and outputs the correlation-aware fusion feature Fl KPB . layer in encoder), which outputs feature F l KPB for next KP-Block and the global-aware kernels K l = Conv(F l KPB ) after post-convolutions. Selective Correlation Fusion. Long-distance reference F LR obtained from LRE is then injected into the deepest KPBlock to model global information for local harmonization. The standard way of feature fusion, like concatenation or addition, equally treats low-level and high-level features. To efficiently model long-term information for local harmonization, we propose SCF to select relevant global information by interacting encoder features F E and long-distance references based on channel-wise attention mechanism.Specifically, as shown in Figure4, we take the (l -1)th layer feature F l-1 KPB ∈ R C l-1 ×H l-1 ×W l-1 and the lth encoder feature F l E ∈ R C l ×H l ×W l as input and extract attention vector α KPB , α E ∈ R C l by 3 × 3 convolutions and MLP. Subsequently, the attention vector is divided into n groups with length m as αKPB , αE ∈ R n×m . Thus, the channel-wise attention relation A ∈ R n×n can be calculated by matrix product",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 .Figure 6 .56Figure 5. Qualitative comparisons with SOTA methods on iHarmony4 [7]. Mask in column one and Red box in Input represents the foreground. Case 1: The background color shows spatial variation, while only our method captures practical color reference by predicted kernel and local harmony modulation. Case 2:The harmony result obtained by our method is better in the detailed structure like duckweed center due to target harmony kernels for each foreground part. Case 3 & 4: For large foregrounds, our approach could also achieve better results, which preserves more original details and is closer to ground truth.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "56",
"figure_type": "figure"
},
{
"figure_caption": "Figure 7 .Figure 8 .78Figure 7. Qualitative ablation study of our approach. Red boxes in input image mark foreground.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "78",
"figure_type": "figure"
},
{
"figure_caption": "Figure 9 .9Figure 9. Feature maps in SCF. Red number in upper right corner of feature map represents channel-wise attention weights.",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "9",
"figure_type": "figure"
},
{
"figure_caption": "Figure 10 .10Figure 10. Cross-attention from example point (10,10). We show four attention maps for different heads, which proves LRE can match model related long-distance references.",
"figure_data": "",
"figure_id": "fig_7",
"figure_label": "10",
"figure_type": "figure"
},
{
"figure_caption": "Quantitative comparisons across four sub-datasets of iHarmony4[7]. ↑ indicates the higher the better, and ↓ indicates the lower the better. We compute fMSE for better reflection on harmonization tasks. Best results are in bold and the suboptimal results are in underline.",
"figure_data": "HCOCOHAdobe5kHFlickrHday2nightALLMethodPSNR↑ MSE↓ fMSE↓ PSNR↑ MSE↓ fMSE↓ PSNR↑ MSE↓ fMSE↓ PSNR↑ MSE↓ fMSE↓ PSNR↑ MSE↓ fMSE↓Composite33.9469.37 996.5928.16345.54 2051.6128.32264.35 1574.3734.01109.65 1409.9831.63172.47 1376.42DIH [41]34.6951.85 798.9932.2892.65593.0329.55163.38 1099.1334.6282.34 1129.4033.4176.77773.18S 2 AM [8]35.4741.07 542.0633.7763.40404.6230.03143.45 785.6535.6950.87835.0634.3559.67594.67DoveNet [7]35.8336.72 551.0134.3452.32380.3930.21133.14 827.0335.2751.95 1075.7134.7652.33532.62IIH [16]37.1624.92 416.3835.2043.02284.2131.34105.13 716.6035.9655.53797.0435.9038.71400.29RAINNet [29]37.0829.52 501.1736.2243.35317.3531.64110.59 688.4034.8357.40916.4836.1240.29469.60iDIH-HRNet [38]39.1616.48 266.1938.0821.88173.9633.1369.67443.6537.7240.59590.9738.1924.44264.96D-HT [15]38.7616.89 299.3036.8838.53265.1133.1374.51515.4537.1053.01704.4237.5530.30320.78Harmonizer [24]38.7717.34 298.4237.6421.89170.0533.6364.81434.0637.5633.14542.0737.8424.26280.51DCCF [44]39.7214.55 267.7938.2420.20171.0133.7266.20440.8438.1851.40629.6738.6022.64265.41SCS-Co [17]39.8813.58 245.5438.2921.01165.4834.2255.83393.7237.8341.75606.8038.7521.33248.86Ours40.3212.95 222.3139.9717.84138.2234.4557.58372.9038.4742.76546.0639.5319.90220.44",
"figure_id": "tab_1",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Quantitative comparisons on different ratios of foreground based on iHarmony4 by MSE and fMSE metrics. The best results are in bold and the suboptimal results are in underline.",
"figure_data": "0%∼5%5%∼15%15%∼100%MethodVenueMSE↓ fMSE↓ MSE↓ fMSE↓ MSE↓ fMSE↓DIHCVPR'1718.92 799.1764.23 725.86 228.86 768.89S 2 AMTIP'2015.09 623.1148.33 540.54 117.62 592.83DoveNet CVPR'2014.03 591.8844.90 504.42 152.07 505.82RainNet CVPR'2111.66 550.3832.05 378.69 117.41 389.80iS 2 AMWACV'216.73 294.7618.03 204.6963.02 207.82GKNetOurs5.36 244.0617.46 200.3457.31 188.75",
"figure_id": "tab_2",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Comparisions on Real Datasets. Experiments results on",
"figure_data": "MaskInputRainNet [29]D-HT [15]Harmonizer [24]OursReal(GT)",
"figure_id": "tab_3",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "B-T scores comparison on real composite images.Following[7, 6,16,17], we conduct our human study by inviting 50 volunteers to compare 24750 pairwise results. The pairwise results are obtained from 99 real composited images, with 25 results for each pair of different methods on average. We also use the Bradley-Terry model (B-T model) to calculate the global ranking score. Our method achieves the best results as shown in Table3.",
"figure_data": "MethodComposite DoveNet[7] RainNet[29] iDIH-HRNet[38] OursB-T Score ↑0.4160.6860.9721.5321.944",
"figure_id": "tab_4",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "our full model obtains the highest performance on all metrics when KPB, LRE, and SCF work together. Table 4 also illustrates the effectiveness of each component. Moreover, to further illustrate our global-local interaction method for harmonization, we show",
"figure_data": "",
"figure_id": "tab_5",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Quantitative ablation study of our approach with different components on iHarmony4[7]Baseline KPB LRE SCF MSE ↓ PSNR ↑ fMSE ↓",
"figure_data": "✓✗✗✗27.2737.83280.56✓✓✗✗21.4238.68235.72✓✓✓✗20.5039.30229.63✓✓✓✓19.9039.53220.44InputGTBaselineKPBLREFull Model",
"figure_id": "tab_6",
"figure_label": "4",
"figure_type": "table"
}
] | Xintian Shen; Jiangning Zhang; Jun Chen; Shipeng Bai; Yue Han; Yabiao Wang; Chengjie Wang; Yong Liu | [
{
"authors": "Steve Bako; Thijs Vogels; Brian Mcwilliams; Mark Meyer; Jan Novák; Alex Harvill; Pradeep Sen; Tony Derose; Fabrice Rousselle",
"journal": "ACM Trans. Graph",
"ref_id": "b0",
"title": "Kernel-predicting convolutional networks for denoising monte carlo renderings",
"year": "2017"
},
{
"authors": "Tim Brooks; Jonathan T Barron",
"journal": "",
"ref_id": "b1",
"title": "Learning to synthesize motion blur",
"year": "2019"
},
{
"authors": "Nicolas Carion; Francisco Massa; Gabriel Synnaeve; Nicolas Usunier; Alexander Kirillov; Sergey Zagoruyko",
"journal": "Springer",
"ref_id": "b2",
"title": "Endto-end object detection with transformers",
"year": "2020"
},
{
"authors": "Mark Chen; Alec Radford; Rewon Child; Jeffrey Wu; Heewoo Jun; David Luan; Ilya Sutskever",
"journal": "PMLR",
"ref_id": "b3",
"title": "Generative pretraining from pixels",
"year": "2020"
},
{
"authors": "Xuhai Chen; Jiangning Zhang; Chao Xu; Yabiao Wang; Chengjie Wang; Yong Liu",
"journal": "",
"ref_id": "b4",
"title": "Better\" cmos\" produces clearer images: Learning space-variant blur estimation for blind image super-resolution",
"year": "2023"
},
{
"authors": "Wenyan Cong; Li Niu; Jianfu Zhang; Jing Liang; Liqing Zhang",
"journal": "",
"ref_id": "b5",
"title": "Bargainnet: Background-guided domain translation for image harmonization",
"year": "2021"
},
{
"authors": "Wenyan Cong; Jianfu Zhang; Li Niu; Liu Liu; Zhixin Ling; Weiyuan Li; Liqing Zhang",
"journal": "",
"ref_id": "b6",
"title": "Dovenet: Deep image harmonization via domain verification",
"year": "2008"
},
{
"authors": "Xiaodong Cun; Chi-Man Pun",
"journal": "IEEE TIP",
"ref_id": "b7",
"title": "Improving the Harmony of the Composite Image by Spatial-Separated Attention Module",
"year": "2020"
},
{
"authors": "Jifeng Dai; Haozhi Qi; Yuwen Xiong; Yi Li; Guodong Zhang; Han Hu; Yichen Wei",
"journal": "",
"ref_id": "b8",
"title": "Deformable convolutional networks",
"year": "2017"
},
{
"authors": "Yimian Dai; Fabian Gieseke; Stefan Oehmcke; Yiquan Wu; Kobus Barnard",
"journal": "",
"ref_id": "b9",
"title": "Attentional feature fusion",
"year": "2021"
},
{
"authors": "Alexey Dosovitskiy; Lucas Beyer; Alexander Kolesnikov; Dirk Weissenborn; Xiaohua Zhai; Thomas Unterthiner; Mostafa Dehghani; Matthias Minderer; Georg Heigold; Sylvain Gelly",
"journal": "",
"ref_id": "b10",
"title": "An image is worth 16x16 words: Transformers for image recognition at scale",
"year": "2020"
},
{
"authors": "Patrick Esser; Robin Rombach; Bjorn Ommer",
"journal": "",
"ref_id": "b11",
"title": "Taming transformers for high-resolution image synthesis",
"year": "2021"
},
{
"authors": "Lan Fu; Changqing Zhou; Qing Guo; Felix Juefei-Xu; Hongkai Yu; Wei Feng; Yang Liu; Song Wang",
"journal": "",
"ref_id": "b12",
"title": "Autoexposure fusion for single-image shadow removal",
"year": "2021"
},
{
"authors": "Qing Guo; Jingyang Sun; Felix Juefei-Xu; Lei Ma; Xiaofei Xie; Wei Feng; Yang Liu; Jianjun Zhao",
"journal": "AAAI",
"ref_id": "b13",
"title": "Efficientderain: Learning pixel-wise dilation filtering for highefficiency single-image deraining",
"year": "2021"
},
{
"authors": "Zonghui Guo; Dongsheng Guo; Haiyong Zheng; Zhaorui Gu; Bing Zheng; Junyu Dong",
"journal": "",
"ref_id": "b14",
"title": "Image harmonization with transformer",
"year": "2007"
},
{
"authors": "Zonghui Guo; Haiyong Zheng; Yufeng Jiang; Zhaorui Gu; Bing Zheng",
"journal": "",
"ref_id": "b15",
"title": "Intrinsic image harmonization",
"year": "2007"
},
{
"authors": "Yucheng Hang; Bin Xia; Wenming Yang; Qingmin Liao",
"journal": "",
"ref_id": "b16",
"title": "Scs-co: Self-consistent style contrastive learning for image harmonization",
"year": "2007"
},
{
"authors": "Kaiming He; Xiangyu Zhang; Shaoqing Ren; Jian Sun",
"journal": "",
"ref_id": "b17",
"title": "Deep residual learning for image recognition",
"year": "2016"
},
{
"authors": "Jie Hu; Li Shen; Gang Sun",
"journal": "",
"ref_id": "b18",
"title": "Squeeze-and-excitation networks",
"year": "2018"
},
{
"authors": "Jiaya Jia; Jian Sun; Chi-Keung Tang; Heung-Yeung Shum",
"journal": "ACM Transactions on graphics (TOG)",
"ref_id": "b19",
"title": "Drag-and-drop pasting",
"year": "2006"
},
{
"authors": "Xu Jia; Bert De Brabandere; Tinne Tuytelaars; Luc V Gool",
"journal": "",
"ref_id": "b20",
"title": "Dynamic filter networks",
"year": "2016"
},
{
"authors": "Yifan Jiang; Shiyu Chang; Zhangyang Wang",
"journal": "",
"ref_id": "b21",
"title": "Transgan: Two transformers can make one strong gan",
"year": ""
},
{
"authors": "Younghyun Jo; Seoung Wug Oh; Jaeyeon Kang; Seon Joo Kim",
"journal": "",
"ref_id": "b22",
"title": "Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation",
"year": "2018"
},
{
"authors": "Zhanghan Ke; Chunyi Sun; Lei Zhu; Ke Xu; Rynson W H Lau",
"journal": "",
"ref_id": "b23",
"title": "Harmonizer: Learning to perform white-box image and video harmonization",
"year": "2022"
},
{
"authors": "Jean-Francois Lalonde; Alexei A Efros",
"journal": "IEEE",
"ref_id": "b24",
"title": "Using color compatibility for assessing image realism",
"year": "2007"
},
{
"authors": "Xiaoguang Li; Qing Guo; Di Lin; Ping Li; Wei Feng; Song Wang",
"journal": "",
"ref_id": "b25",
"title": "Misf: Multi-level interactive siamese filtering for high-fidelity image inpainting",
"year": "2022"
},
{
"authors": "Xiang Li; Wenhai Wang; Xiaolin Hu; Jian Yang",
"journal": "",
"ref_id": "b26",
"title": "Selective kernel networks",
"year": "2019"
},
{
"authors": "Tsung-Yi Lin; Piotr Dollár; Ross Girshick; Kaiming He; Bharath Hariharan; Serge Belongie",
"journal": "",
"ref_id": "b27",
"title": "Feature pyramid networks for object detection",
"year": "2017"
},
{
"authors": "Jun Ling; Han Xue; Li Song; Rong Xie; Xiao Gu",
"journal": "",
"ref_id": "b28",
"title": "Region-aware adaptive instance normalization for image harmonization",
"year": "2007"
},
{
"authors": "Rui Liu; Hanming Deng; Yangyi Huang; Xiaoyu Shi; Lewei Lu; Wenxiu Sun; Xiaogang Wang; Jifeng Dai; Hongsheng Li",
"journal": "",
"ref_id": "b29",
"title": "Fuseformer: Fusing fine-grained information in transformers for video inpainting",
"year": "2021"
},
{
"authors": "Ze Liu; Yutong Lin; Yue Cao; Han Hu; Yixuan Wei; Zheng Zhang; Stephen Lin; Baining Guo",
"journal": "",
"ref_id": "b30",
"title": "Swin transformer: Hierarchical vision transformer using shifted windows",
"year": "2021"
},
{
"authors": "Ben Mildenhall; Jonathan T Barron; Jiawen Chen; Dillon Sharlet; Ren Ng; Robert Carroll",
"journal": "",
"ref_id": "b31",
"title": "Burst denoising with kernel prediction networks",
"year": "2018"
},
{
"authors": "Simon Niklaus; Long Mai; Feng Liu",
"journal": "",
"ref_id": "b32",
"title": "Video frame interpolation via adaptive convolution",
"year": "2017"
},
{
"authors": "Patrick Pérez; Michel Gangnet; Andrew Blake",
"journal": "ACM",
"ref_id": "b33",
"title": "Poisson image editing",
"year": "2003"
},
{
"authors": "Pitié Franc; Anil Kokaram",
"journal": "IPSJ Transactions on Computer Vision and Applications",
"ref_id": "b34",
"title": "The linear mongekantorovitch linear colour mapping for example-based colour transfer",
"year": "2007"
},
{
"authors": "Erik Reinhard; Michael Adhikhmin; Bruce Gooch; Peter Shirley",
"journal": "IEEE Computer graphics and applications",
"ref_id": "b35",
"title": "Color transfer between images",
"year": "2001"
},
{
"authors": "Olaf Ronneberger; Philipp Fischer; Thomas Brox",
"journal": "Springer",
"ref_id": "b36",
"title": "Unet: Convolutional networks for biomedical image segmentation",
"year": "2015"
},
{
"authors": "Konstantin Sofiiuk; Polina Popenova; Anton Konushin",
"journal": "",
"ref_id": "b37",
"title": "Foreground-aware semantic representations for image harmonization",
"year": "2007"
},
{
"authors": "Kalyan Sunkavalli; Micah K Johnson; Wojciech Matusik; Hanspeter Pfister",
"journal": "ACM Transactions on Graphics (TOG)",
"ref_id": "b38",
"title": "Multi-scale image harmonization",
"year": "2010"
},
{
"authors": "Micah K Michael W Tao; Sylvain Johnson; Paris",
"journal": "Springer",
"ref_id": "b39",
"title": "Errortolerant image compositing",
"year": "2010"
},
{
"authors": "Yi-Hsuan Tsai; Xiaohui Shen; Zhe Lin; Kalyan Sunkavalli; Xin Lu; Ming-Hsuan Yang",
"journal": "",
"ref_id": "b40",
"title": "Deep image harmonization",
"year": "2007"
},
{
"authors": "Thijs Vogels; Fabrice Rousselle; Brian Mcwilliams; Gerhard Röthlin; Alex Harvill; David Adler; Mark Meyer; Jan Novák",
"journal": "ACM Transactions on Graphics (TOG)",
"ref_id": "b41",
"title": "Denoising with kernel prediction and asymmetric loss functions",
"year": "2018"
},
{
"authors": "Ziyu Wan; Jingbo Zhang; Dongdong Chen; Jing Liao",
"journal": "",
"ref_id": "b42",
"title": "High-fidelity pluralistic image completion with transformers",
"year": "2021"
},
{
"authors": "Ben Xue; Shenghui Ran; Quan Chen; Rongfei Jia; Binqiang Zhao; Xing Tang",
"journal": "",
"ref_id": "b43",
"title": "Dccf: Deep comprehensible color filter learning framework for high-resolution image harmonization",
"year": "2022"
},
{
"authors": "Su Xue; Aseem Agarwala; Julie Dorsey; Holly Rushmeier",
"journal": "ACM Transactions on graphics (TOG)",
"ref_id": "b44",
"title": "Understanding and improving the realism of image composites",
"year": "2012"
},
{
"authors": "Brandon Yang; Gabriel Bender; Quoc V Le; Jiquan Ngiam",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b45",
"title": "Condconv: Conditionally parameterized convolutions for efficient inference",
"year": "2019"
},
{
"authors": "Hang Zhang; Chongruo Wu; Zhongyue Zhang; Yi Zhu; Haibin Lin; Zhi Zhang; Yue Sun; Tong He; Jonas Mueller; R Manmatha",
"journal": "",
"ref_id": "b46",
"title": "Resnest: Split-attention networks",
"year": "2022"
},
{
"authors": "Jiangning Zhang; Xiangtai Li; Jian Li; Liang Liu; Zhucun Xue; Boshen Zhang; Zhengkai Jiang; Tianxin Huang; Yabiao Wang; Chengjie Wang",
"journal": "",
"ref_id": "b47",
"title": "Rethinking mobile block for efficient neural models",
"year": "2023"
},
{
"authors": "Jiangning Zhang; Xiangtai Li; Yabiao Wang; Chengjie Wang; Yibo Yang; Yong Liu; Dacheng Tao",
"journal": "",
"ref_id": "b48",
"title": "Eatformer: improving vision transformer inspired by evolutionary algorithm",
"year": "2022"
},
{
"authors": "Jiangning Zhang; Chao Xu; Jian Li; Wenzhou Chen; Yabiao Wang; Ying Tai; Shuo Chen; Chengjie Wang; Feiyue Huang; Yong Liu",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b49",
"title": "Analogous to evolutionary algorithm: Designing a unified sequence model",
"year": "2021"
},
{
"authors": "Jiangning Zhang; Chao Xu; Jian Li; Yue Han; Yabiao Wang; Ying Tai; Yong Liu",
"journal": "",
"ref_id": "b50",
"title": "Scsnet: an efficient paradigm for learning simultaneously image colorization and superresolution",
"year": "2022"
},
{
"authors": "Zhenli Zhang; Xiangyu Zhang; Chao Peng; Xiangyang Xue; Jian Sun",
"journal": "",
"ref_id": "b51",
"title": "Exfuse: Enhancing feature fusion for semantic segmentation",
"year": "2018"
}
] | [
{
"formula_coordinates": [
4,
146,
295.46,
104.37,
10.4
],
"formula_id": "formula_0",
"formula_text": "I c = I f • M + (1 -M) • I b"
},
{
"formula_coordinates": [
4,
362.45,
644.74,
182.66,
12.94
],
"formula_id": "formula_1",
"formula_text": "K l , F l KPB = KPBlock(F l E , F l-1 KPB ),(1)"
},
{
"formula_coordinates": [
5,
136.25,
532.74,
150.12,
11.58
],
"formula_id": "formula_2",
"formula_text": "A = αKPB ⊙ αT E .(2)"
},
{
"formula_coordinates": [
5,
122.41,
618.88,
163.96,
11.26
],
"formula_id": "formula_3",
"formula_text": "S l = σ(α + b • FC(α)),(3)"
},
{
"formula_coordinates": [
5,
56.48,
682.28,
229.88,
13.07
],
"formula_id": "formula_4",
"formula_text": "Fl KPB = S l E •Conv(F l E )+Upsample(S l-1 KPB •Conv(F l-1 KPB )),(4)"
},
{
"formula_coordinates": [
5,
398.91,
216.44,
146.2,
13.13
],
"formula_id": "formula_5",
"formula_text": "Fl D = F l D ⊛ K l ,(5)"
},
{
"formula_coordinates": [
5,
367.66,
337.34,
177.45,
22.55
],
"formula_id": "formula_6",
"formula_text": "Fl D [p] = q∈N p K l p [p -q]F l D [q],(6)"
},
{
"formula_coordinates": [
5,
360.14,
463.21,
169.76,
12.93
],
"formula_id": "formula_7",
"formula_text": "= ϕ L D (• • • ϕ 1 D (F 1 D ⊛ K 1 )) • M + (1 -M) • I c ."
},
{
"formula_coordinates": [
5,
369.01,
552.72,
176.1,
48.95
],
"formula_id": "formula_8",
"formula_text": "L rec = h,w Î -I 2 2 max A min , h,w M h,w ,(7)"
}
] |
|
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b22",
"b19",
"b13",
"b9",
"b8",
"b7"
],
"table_ref": [],
"text": "While deep learning models continue to provide impressive performance for computer vision and Natural Language Processing (NLP), their practical use in some real-life problems (e.g. credit scoring) remains limited due to legislation with one of the important points: the interpretability of the used models (e.g., GDPR: Article 22 in Europe). Since the promising results of the transformer architecture on machine translation tasks [23], many efforts have been made to improve models' accuracy for tabular modeling using the attention mechanism [20,14,10,9]. The principal motivation behind our work is to push forward this effort by proposing an interpretable attention model for tabular learning. To achieve this goal, we found it necessary to develop a representation learning block or layer that (i) preserves the initial feature space (i.e., R p -→ R p ), and (ii) reduces to the maximum some extra steps (e.g., residual connection, LayerNorm) that make the overall architecture less interpretable. In this paper, we present a new attention-based representation learning block for tabular arXiv:2305.11684v1 [cs.LG] 19 May 2023 data, called Self-Reinforcement Attention (SRA). We summarize our contributions as follows:\n-SRA is a pre-processing block that allows to learn intelligible representation by weighting each raw feature with a positive alignment (i.e., an attention score). The obtained representation, called \"reinforced vector\", is then passed to an aggregation model to provide the decision boundary. It is based on the previous work about Epigenetics algorithms [8].\n-SRA provides a feature score that facilitates interpretations. This score is then used as a coefficient to amplify or reduce some characteristics according to the context of the observations (here the spatial information), allowing to: (i) take into account possible interactions without the need to artificially add additional features or terms, (ii) identify important features for the final prediction.\n-Our experiments on synthetic and benchmark imbalanced datasets show a promising performance while preserving intrinsic interpretability of the resulting architecture.\nThe rest of the paper is organized as follows: Section 2 presents a brief discussion of state-of-the-art works. Section 3 describes the SRA block and its architecture. The experimental setup, the discussion of the obtained results and the limitations are presented in Section 4. Finally, Section 5 concludes the paper and highlights some perspectives."
},
{
"figure_ref": [],
"heading": "Related work",
"publication_ref": [
"b3",
"b5",
"b12",
"b0",
"b4",
"b17",
"b2",
"b0",
"b17",
"b19",
"b13",
"b9",
"b8",
"b22",
"b19",
"b13",
"b9",
"b8"
],
"table_ref": [],
"text": "Classical models. For many tabular or heterogeneous data classification problems, treebased ensemble models are still widely used and preferred as they generate high performance results while not being greedy in terms of computational resources. Random Forest [4] is one of the well-known models that combines the prediction of several independent trees through majority voting or average. Tree based models are also used in a sequential manner (Gradient Boosting Machine or GBM) where the newest trees are built to minimize the errors of the previous ones. XGBoost [6] and LightGBM [13] are two fast implementations of GBM that empower the regularization in the boosting mechanism and are considered as the state-of-the-art for many problems or competitions. As for linear models, these are commonly used in settings where models' interpretability is strongly required. For these models, linear relations between features are expected otherwise continuous variables are discretized aiming to take into account non-linearity; otherwise other features are added as interaction terms.\nDeep learning models and feature attribution. Many deep learning models were designed to provide local explanations to their predictions. Among these models, we mention Neural Additive Models (NAM) [1] which present a neural implementation of the classic Generalized additive models (GAM). In addition to the marginal feature contribution, NAM provides the shape function which by visualization can help understand the global contribution of a given feature. NODE-G 2 AM [5] is an improvement of NAM built on top of a NODE architecture [18] to take into account pairwise interactions among features. Among the drawbacks of these deep learning architectures is that they apply one sub-network per feature (or pair of features) which can be very resource consuming, especially for high dimensional problems, and the management of higher order interactions is not guaranteed. In TabNet [3], a multiplicative sparse mask vector is used for instance-wise feature selection and the masks are aggregated across many stages to compute features' importance. Matrix multiplication trick is used to reconstruct initial feature dimension which is lost with the passage in the feature transformer module. Contrary to TabNet, our proposed SRA model preserves the initial feature dimension and uses the dot-product to compute its attention weights.\nAttention-based models for tabular data classification. Compared to classical models, deep learning models have several advantages in various settings, mainly (i) in continuous learning especially in the presence of concept drift (e.g., using transfer learning or fine tuning), (ii) in multimodal problems (e.g., encode tabular information with text, image, etc.), and (iii) in multitask settings [1]. These reasons motivated many researchers to improve the traditional fully connected MultiLayer Perceptron (MLP) by imitating tree models [18] or to use the attention mechanism [20,14,10,9]. One common feature of these attention-based architectures is that each feature is considered as a token and is embedded similarly as in [23]. This consideration, although favoring the expressiveness of the resulting architecture, leads to a significant increase in the number of learnable parameters and complicates the explanation of local predictions, especially when the number of transformer layers or stages exceeds 1. The SRA-based models proposed in our work use less parameters compared to [20,14,10,9]. Also they are accurate in comparison to state-of-art models while being self-explainable, in the sense that they provide an intrinsic explanation of their predictions.\n3 Self-Reinforcement Attention (SRA)"
},
{
"figure_ref": [
"fig_1"
],
"heading": "SRA Model",
"publication_ref": [
"b19",
"b13",
"b9",
"b8"
],
"table_ref": [],
"text": "The challenge in most supervised tabular learning problems using attention mechanism [20,14,10,9] is to estimate the output ŷ = f θ (x) given the feature vector x = (x 1 , . \no = a x(1)\nwhere is the element-wise multiplication.\nIf we instantiate our supervised aggregation model (Figure 1b) as a linear transformation, then the SRA model can be formalized as follows: β i a i x i represents the contribution (the prediction importance) of the feature x i to the output, β = (β 1 , β 2 , ..., β p ) is the linear regression coefficients and a i is interpreted as the amplification (or the correction) that the feature x i received from other features or itself due to the interactions. We call this SRA model (instantiation) as SRALinear. g represents the link function (e.g., usually g(µ) = log( µ 1-µ ) for binary classification and g = Identity for regression tasks).\ng(ŷ) = β • o = β 1 o 1 + ... + β i o i + ... + β p o p = β 1 a 1 x 1 + ... + β i a i x i + ... + β p a p x p(2)"
},
{
"figure_ref": [
"fig_1"
],
"heading": "SRA block",
"publication_ref": [
"b19",
"b9"
],
"table_ref": [],
"text": "Given the input vector x = (x 1 , ...x i , ..., x p ) ∈ R p , the SRA block encodes it into\np keys in K = [k 1 , k 2 , ..., k i , ..., k p ] T with k i = (k 1 i , ..., k d k i ) ∈ R d k using the key\nencoder and queries matrix Q = [q 1 , q 2 , ..., q i , ..., q p ] T with q i = (q 1 i , ..., q\nd k i ) ∈ R d k\nusing the query encoder (see Figure 1a and the pseudocode provided in Algorithm 1). The matrix of queries (Q) and keys (K) are generated by two separate fully connected feed-forward networks (F F N ) namely QueryEncoder and KeyEncoder.\nThe KeyEncoder (resp. QueryEncoder) produces directly p keys (resp. queries) using a single F F N instead of using p independent F F N s per feature as in [20,10]. This embedding should be particularly useful for heterogeneous (tabular) data especially in the presence of strong features' interactions and at the same time alleviate the need of using several attention blocks (layers) or extra processing which could affect the interpretable of the attention coefficients. Furthermore, with a Sigmoid activation function, all elements k j i of K (resp. q j i of Q) are scalar numbers bounded in [0, 1]. The keys in K are compared to the queries Q component by component, allowing to quantify the alignment of different transformations of the same input calculating the attention weights a = (a 1 , .., a i , ..., a p ) as follows :\na i = q i • k i d k for i ∈ 1, • • • , p(3)\nWe further use the scaling by d k in order to reduce the magnitude of the dot-product and to get dimension-free attention coefficients\na i ∈ [0, 1].\nWe propose this attention estimation to produce a concise explanation of the decision process. Indeed, considering the potential internal conflict between the input components (due to the interactions), the attention weights vector a may enhance or reduce some components (of the input vector) at strategic and specific positions. \nQ = self.KeyEncoder(x) # Q is (b, p, d k ) K = self.QueryEncoder(x) # K is (b, p, d k ) QK = Q * K * self.scale # scale= 1/d k , QK is (b, p, d k ) a = QK.sum(axis = -1) # a is (b, p) return a\n4 Experiments"
},
{
"figure_ref": [],
"heading": "Experimental setup",
"publication_ref": [
"b20",
"b6"
],
"table_ref": [
"tab_1"
],
"text": "Our motivation when building the SRA is to combine interpretability and performance in a single model with a focus on imbalanced classification tasks. Typically, the interpretability of models is assessed separately from their performance, which can make it challenging to gauge the effectiveness of our SRA solution. Nonetheless, we believe it is appropriate to measure the value of our SRA solution by comparing it to both comprehensible and fully complex benchmarks using the following criteria:\n-Intelligibility: Are the representations learned by SRA understandable? Are the explanations provided by SRA models faithful? -Effectiveness: Are SRA-based models accurate compared to state-of-the-art models?\nDatasets. As we focus particularly on finance as an application domain, we considered three UCI datasets (Default of Credit Card Clients, Bank Marketing, and Adult Income) and four Kaggle datasets (Credit Card Fraud, Bank Churn Modelling, Blastchar, and Telco Churn) and the Heloc Fico dataset for our experiments. All of these datasets are used for binary classification tasks, and the number of features (both numerical and categorical) ranges from 10 to 63. The percentage of positive class instances varies between 0.17% and 48% (see Table 1 for further details). Unless otherwise specified, all categorical inputs are one-hot encoded, and numerical inputs are scaled using the mean and standard deviation to accelerate the convergence of the algorithms. Model setup.\n-Choice of the query and key encoder: we use the same architecture for the key and query encoders which is a two hidden layers fully connected neural network of dimension {d 1 , d 2 } with,\nd 1 = p × (d k /4) and d 2 = p × (d k /2), d k ≥ 4.\n-Regularization: to increase the generalization power, we used regularization in the SRA block. Specifically, we used dropout [21] in both the key and query encoders during the training. Also, we used weight decay (L 2 penalization) to empower the smoothness in the embeddings (of the key and query).\nEvaluation measures. We evaluate the models using 5-stratified fold cross validation (80% for the training) and report the mean and standard deviation of the Area Under the ROC curve (AUCROC) on the validation set. Particularly for highly imbalanced datasets (e.g., the Credit Card Fraud dataset), we optimize and report the Average Precision or Precision-Recall (AUCPR). In fact, AUCPR gives a more informative picture of an algorithm's performance than AUCROC in highly imbalanced data settings and algorithms that optimize AUCPR are guaranteed to optimize AUCROC [7]."
},
{
"figure_ref": [
"fig_5"
],
"heading": "Intelligibility of SRA",
"publication_ref": [
"b11",
"b21",
"b14",
"b1",
"b16",
"b1",
"b1",
"b14",
"b10",
"b15",
"b9",
"b2",
"b0",
"b4",
"b5",
"b16",
"b18",
"b19",
"b13",
"b9",
"b8"
],
"table_ref": [
"tab_2",
"tab_3"
],
"text": "One interesting property of an SRA-based model is that it provides interpretable information about its behavior. In this section, we explore some of these interpretable aspects through visualizations and the ability to identify relevant features. We focus in this section on its combination with the linear model. The SRALinear model (Equation 2) has two interesting properties:\n1. Each feature x i appears in the equation as in a classical linear model and β i a i x i is its contribution to the output. 2. Faithfulness: the attention coefficients are clearly correlated to model's outputs. This is actually a desirable property for considering attention as an explanation of predictions [12].\nHow the raw data is reinforced using the SRA block. To illustrate how the raw data is reinforced in practice, we use 2D toy datasets with the objective of facilitating the visualization. We consider first the following function: Through multiplication, values of x 2 are significantly reduced (e.g., to 0) when needed (i.e., o 2 ∼ 0 when x 1 < 0, x 2 < 0), which makes the classes easy to separate with the downstream linear model. We included another synthetic dataset, the 2D chainLink [22], as depicted in Fig3. By applying SRA coefficients to this dataset, we acquired a new data representation that enables easy separation of the classes, as shown in Fig3b. Even without knowledge of the true data generating process, it is apparent that all the purple observations have been moved strategically so that a simple rule, o 2 > 0, can effectively isolate nearly all the yellow observations of interest. For a more detailed depiction of the reinforced vectors, please refer to the supplementary materials provided in Section A.1.\nF 1 (x) = 5x 1 -\nCan the SRALinear find important features. In order to interpret machine learning models, it is essential to perform feature attribution, which involves identifying which variables contributed to a high output score. For classical state-of-the-art models like XGBoost, post-hoc explanation tools such as TreeSHAP and LIME are often used to provide individual prediction explanations. However, these tools can introduce their own biases [15,2]. In this investigation, we aim to assess SRALinear's ability to identify crucial features in comparison to that of Logistic Regression, a self-explanatory model, and XGBoost coupled with TreeSHAP [17]. As TreeSHAP calculates the exact Shapley value and is computationally efficient, it is particularly well-suited for treebased models. For this purpose, we generate two synthetic datasets with 5 features x = (x 1 , x 2 , x 3 , x 4 , x 5 ) of size 30000 and 60000, respectively based on the Gaussian distribution (0 mean and variance 1) as follows:\ny = (5x 1 -5x 2 )1 x5≤0 + (5x 3 -5x 4 )1 x5>0(5)\ny = 1 if (x 1 +2.5) 2 +x 2 2 < 1 or (x 1 -2.5) 2 +(x 2 -1.5) 2 < 1 and 0 otherwise (6) The example called Synthetic 1 (Equation 5) is borrowed from [2] . It is interesting for this work because it highlights the interactions between the features. The goal is to design a model that can achieve perfect accuracy by using only the features x 1 and x 2 , or alternatively, depending on the sign of x 5 , using only the features x 3 and x 4 . To evaluate the model's performance, we compute the True Positive Rate (TPR) using a test set consisting of 20% of the data points, with the remaining 80% used for training. We restrict our analysis to those data points with x 5 ≤ 0, which comprise 3750 instances. Specifically, we assess the ability of SRALinear to identify the two most important features among (x 1 , x 2 , x 3 , x 4 ). Regarding the example that we called Synthetic 2 (Equation 6), it is rather attribution tools friendly as features are independent (although there is a non-linearity). Only x 1 and x 2 are relevant to predict the class 1. We consider all data points from class 1 in test set (695 data points) and try to find the two most important features among (x 1 , x 2 , x 3 , x 4 , x 5 ).\nAs shown in Table 2, SRALinear is able to accurately detect the most relevant features with a high True Positive Rate (TPR) of over 99%. As expected, TreeSHAP (combined with XGBoost) is able to accurately detect the two most relevant features for Synthetic 2, but struggles with the Synthetic 1 dataset, achieving a TPR of approximately 75%. Knowing that XGBoost has a perfect performance on this dataset (R 2 > 99%), we argue that the incorrect attributions are due to the interpretability tool, which fails to provide the important features in XGBoost's decision. For brevity, we encourage interested readers to refer to [2,15,11] for more details on attribution methods and variable interactions. Regarding the Linear models (Linear regression and Logistic Regression), although being highly explainable, they detect important variables only with moderate accuracy. This is due to bias when using linear models for handing nonlinear data (R 2 = 50 %, AUCROC = 74%). From these two synthetic examples, we show that there are two possible biases when using feature attribution: (i) the first is due to underfitting (e.g., using linear models to fit complex data); (ii) the second is due to post-hoc interpretabilty tools used to explain full complexity models. In the context mentioned above, the SRALinear model appears to be a good compromise for both the feature attribution and accuracy aspects.\nLimitations of SRA based explanations. The SRA model, as proposed, should not be used directly as a global feature selector but rather after identifying all relevant variables. This is because the feature importance measure provided by SRALinear is 'the local prediction importance' and not 'the local feature importance' (cf. Equation 2). Although these two terms are usually used interchangeably in the literature of feature attribution methods, there are some nuances [16]. Specifically, the feature that is important to a prediction is automatically relevant, but the inverse is not always true, especially when there are interactions. Regarding the SRALinear model, an illustrative example is the synthetic 1 dataset (Equation 5). For this dataset, a perfect SRALinear model will always give zero prediction importance to the feature x 5 as it cannot be used as main effect or feature (although it can be used to reduce the contribution of other features in the attention vector). Thus, based solely on the prediction importance, one may be tempted to delete the feature x 5 to create a model with fewer variables. However with further analysis (e.g., visualizing or computing the gradient β i a i x i vs x 5 ) we can notice that x 5 must be kept. An shown in Fig 4, an important information needs to be known about x 5 ; which is its sign. When x 5 < 0 (resp. x 5 > 0), the prediction contribution (or importance) of x 3 is close to 0 (resp. the prediction contribution of x 1 is close to 0). A similar visualization would lead to the same finding for the contributions of x 2 and x 4 , indicating that x 5 is indeed relevant to the model. Dropping it would result in a drastic reduction in SRALinear's performance, as it would behave like a simple linear regression. In this section, we discuss the effectiveness of considering the SRA block by comparing the accuracy achieved by SRALinear model (Equation 2) on benchmark datasets relatively to baseline models (interpretable and non-counterparts).\nBaseline models. We compare quantitatively the performance of the proposed SRA models (Equation 2) with the following baselines:\n-Logistic Regression (LR): It is a highly interpretable model obtained by simple linear combination of features followed by Sigmoid activation for binary classification problems. -MultiLayer Perceptron (MLP): it is a full complexity architecture that can model non-linear effects and interactions; making it not directly interpretable by humans. We consider two (2) hidden layers MLP model of dimensions {4 × p, 2 × p} as in [10]. p is the input feature dimension. -TabNet [3]: it is a deep learning model that provides local explanation of its predictions without imposing a limit on the order of interactions between the variables in contrast to [1,5].\n-XGBoost [6]: Despite the need for feature attribution tools such as TreeSHAP [17] and LIME [19] to explain its local predictions, XGBoost remains a favorite and leading state-of-the-art model for several real-life use cases and tabular learning competitions. It is selected for comparison with the intention of measuring the performance that may be lost by preferring an intrinsically interpretable model. It is also to be noted that we do not compare directy to some attention-based models, such as [20,14,10,9] as they are more motivated by performance than interpretability and XGBoost can give an idea of the upper bound that these models can reach in most cases.\nEvaluation of the Accuracy. As shown in Table 3, SRALinear achieved the best performance in 6/8 cases among self-explainable models (over TabNet, LR). Furthermore, the obtained performance is often close (for 6/8 benckmark datasets) to the one of the overall best performing model which is XGBoost. These results confirms the effectiveness of SRA block particularly when observing the difference of performances between the Logistic Regression (LR) and SRALinear which ranges from +0.09 for the Adult-Income dataset to +10.05 AUC for the Bank Churn dataset. We recall that LR model is the resulting architecture when removing the SRA block or setting attention weights to 1 (cf. "
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "We presented a novel attention mechanism for tabular learning named Self-Reinforcement Attention (SRA), a deep learning based representation learning block to produce reinforced version from raw input data through element-wise multiplication. We demonstrated the effectiveness and the benefits of SRA with both synthetic and benchmark imbalanced classification datasets. We also showed that the SRA models are intelligible in sense that they provides an intrinsic attribution for feature, which can be further used for global model behavior understanding. Our experimental results confirms the proposed model as a promising solution for self-explainable models in tabular learning settings without the need to 'sacrificing the accuracy'. Overall, we recommend to the interested user to check as much as possible the agreement of the SRA based explanations with their data knowledge since these are not causalities. The SRA block as proposed can be further enriched especially to deal with complex tasks. In this concern, we are currently working on how to use several heads and layers, similar to what is often done in attention-based architectures. Also, studying empirically the local stability of SRA explanations is an important direction of future research as well as incorporating data knowledge in the training phase (e.g. use monotonic constraints with respect to some features)."
},
{
"figure_ref": [],
"heading": "A Additional experimentals results",
"publication_ref": [],
"table_ref": [],
"text": "A.1 How the raw data is reinforced using the SRA block. "
}
] | Apart from the high accuracy of machine learning models, what interests many researchers in real-life problems (e.g., fraud detection, credit scoring) is to find hidden patterns in data; particularly when dealing with their challenging imbalanced characteristics. Interpretability is also a key requirement that needs to accompany the used machine learning model. In this concern, often, intrinsically interpretable models are preferred to complex ones, which are in most cases black-box models. Also, linear models are used in some high-risk fields to handle tabular data, even if performance must be sacrificed. In this paper, we introduce Self-Reinforcement Attention (SRA), a novel attention mechanism that provides a relevance of features as a weight vector which is used to learn an intelligible representation. This weight is then used to reinforce or reduce some components of the raw input through element-wise vector multiplication. Our results on synthetic and real-world imbalanced data show that our proposed SRA block is effective in end-to-end combination with baseline models. | Self-Reinforcement Attention Mechanism For Tabular Learning | [
{
"figure_caption": "Fig. 1 :1Fig. 1: SRA architecture.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Algorithm 1 :1PyTorch-style forward pass pseudocode of the SRA Block # b is batch size, p the number of features def forward(self, x):",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "5x 2 1 Fig. 2 :12Fig. 2: Illustration of the reinforcement process on 7500 synthetic data points with 0 mean, unity variance Gaussian distribution. The yellow color is used for the class of interest. The green color a possible decision boundary to separate the two classes.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "12",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 3 :3Fig.3: Illustration of the reinforcement process on the chainLink 2D[22] with 1000 datapoints",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "(a) Prediction importance of x1 vs x5 (b) Prediction importance of x3 vs x5",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 4 :4Fig. 4: synthetic 1: relevance analysis of the feature x 5",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Fig 1).",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 5 :Fig. 6 :Fig. 7 :Fig. 8 :Fig. 9 :56789Fig. 5: Illustration on Five sphere with 250 datapoints",
"figure_data": "",
"figure_id": "fig_7",
"figure_label": "56789",
"figure_type": "figure"
},
{
"figure_caption": ".., x p ) ∈ R p . The parametric model f θ is learned using the training data D = Given the raw input x, the SRA block produces an attention vector a = (a 1 , ..., a i , ...a p ). Thereafter the attention vector is used to learn an intelligible representation o = (o 1 , ..., o",
"figure_data": "{(x i , y i )} n i=1 with y i ∈ {0, 1} for binary classification or y i ∈ R for regression tasks.Our proposed SRA model f θ (Fig 1b) contains a SRA block (Fig 1a) which is a novelattention mechanism layer denoted as a function a(.).",
"figure_id": "tab_0",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Benchmark datasets",
"figure_data": "Datasets# Datapoints # features # Categorical features Positive Class (%)Bank Churn1000010220.37Credit Default3000023322.16Bank Marketing4521116911.70Adult Income3016214824.89Credit Card Fraud2848072900.17Blastchar7043191626.54Telco Churn6646963020.92Heloc Fico1045923047.81",
"figure_id": "tab_1",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Relevance feature discovry capacity. The True Positive Rate (TPR) (%) is used as metric. R 2 (the higher the better) is to evaluate the test performance of Synthetic 1 and AUCROC (%) is used for Synthetic 2 dataset.",
"figure_data": "DatasetsModelsTPR Test performanceLinear Regression51.2850.00Synthetic 1XGBoost+TreeSHAP 75.4799.21SRALinear99.7799.67Logistic Regression (LR) 66.8373.77Synthetic 2XGBoost+TreeSHAP 98.6399.99SRALinear99.8699.72",
"figure_id": "tab_2",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Accuracy of the SRALinear model. Mean and standard deviation AUC (%), reported from 5-stratified cross validation. Bold highlights the best performance when comparing self-explainable (LR, TabNet, SRALinear) models and italic is used for the overall best performing model.",
"figure_data": "DatasetsLRTabNetSRALinearMLPXGBoostBankChurn 76.93 (1.56) 86.99 (0.79) 86.98 (0.46) 87.08 (0.73) 86.82 (0.79)CreditDefault 72.53 (0.49) 77.85 (1.03) 77.55 (0.56) 78.24 (0.78) 78.56 (0.69)BankMarketing 90.79 (0.49) 92.74 (0.70) 93.33 (0.50) 93.44 (0.41) 93.82 (0.38)AdultIncome 90.50 (0.41) 90.46 (0.52) 91.07 (0.42) 91.45 (0.38) 92.63 (0.37)CreditCardFraud 77.08 (2.59) 81.09 (3.92) 86.58 (2.81) 85.69 (2.53) 86.54 (2.19)Blastchar84.54 (1.48) 83.53 (1.45) 84.63 (1.51) 84.63 (1.52) 84.89 (1.21)TelcoChurn 88.95 (0.29) 90.45(0.33) 90.52 (0.31) 90.54 (0.28) 91.13 (0.37)HelocFico78.26 (0.52) 79.39 (0.57) 79.43 (0.41) 79.50 (0.46) 79.75 (0.74)4.3 The effectiveness of the SRA block.",
"figure_id": "tab_3",
"figure_label": "3",
"figure_type": "table"
}
] | Kodjo Mawuena Amekoe; Mohamed Djallel Dilmi; Hanene Azzag; Mustapha Lebbah; Zaineb Chelly Dagdia; Gregoire Jaffre | [
{
"authors": "R Agarwal; L Melnick; N Frosst; X Zhang; B Lengerich; R Caruana; G E Hinton",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b0",
"title": "Neural additive models: Interpretable machine learning with neural nets",
"year": "2021"
},
{
"authors": "S I Amoukou; T Salaün; N Brunel",
"journal": "PMLR",
"ref_id": "b1",
"title": "Accurate shapley values for explaining tree-based models",
"year": "2022"
},
{
"authors": "S Ö Arik; T Pfister",
"journal": "",
"ref_id": "b2",
"title": "Tabnet: Attentive interpretable tabular learning",
"year": "2021"
},
{
"authors": "L Breiman",
"journal": "Machine learning",
"ref_id": "b3",
"title": "Random forests",
"year": "2001"
},
{
"authors": "C H Chang; R Caruana; A Goldenberg",
"journal": "",
"ref_id": "b4",
"title": "Node-gam: Neural generalized additive model for interpretable deep learning",
"year": "2021"
},
{
"authors": "T Chen; C Guestrin",
"journal": "",
"ref_id": "b5",
"title": "Xgboost: A scalable tree boosting system",
"year": "2016"
},
{
"authors": "J Davis; M Goadrich",
"journal": "",
"ref_id": "b6",
"title": "The relationship between precision-recall and roc curves",
"year": "2006"
},
{
"authors": "M D Dilmi; H Azzag; M Lebbah",
"journal": "",
"ref_id": "b7",
"title": "Epigenetics algorithms: Self-reinforcement-attention mechanism to regulate chromosomes expression",
"year": "2023"
},
{
"authors": "Y Gorishniy; I Rubachev; V Khrulkov; A Babenko",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b8",
"title": "Revisiting deep learning models for tabular data",
"year": "2021"
},
{
"authors": "X Huang; A Khetan; M Cvitkovic; Z Karnin",
"journal": "",
"ref_id": "b9",
"title": "Tabtransformer: Tabular data modeling using contextual embeddings",
"year": "2020"
},
{
"authors": "X Huang; J Marques-Silva",
"journal": "",
"ref_id": "b10",
"title": "The inadequacy of shapley values for explainability",
"year": "2023"
},
{
"authors": "S Jain; B C Wallace",
"journal": "",
"ref_id": "b11",
"title": "Attention is not explanation",
"year": "2019"
},
{
"authors": "G Ke; Q Meng; T Finley; T Wang; W Chen; W Ma; Q Ye; T Y Liu",
"journal": "Advances in neural information processing systems",
"ref_id": "b12",
"title": "Lightgbm: A highly efficient gradient boosting decision tree",
"year": "2017"
},
{
"authors": "J Kossen; N Band; C Lyle; A N Gomez; T Rainforth; Y Gal",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b13",
"title": "Self-attention between datapoints: Going beyond individual input-output pairs in deep learning",
"year": "2021"
},
{
"authors": "I E Kumar; S Venkatasubramanian; C Scheidegger; S Friedler",
"journal": "PMLR",
"ref_id": "b14",
"title": "Problems with shapleyvalue-based explanations as feature importance measures",
"year": "2020"
},
{
"authors": "I Lemhadri; H H Li; T Hastie",
"journal": "",
"ref_id": "b15",
"title": "Rbx: Region-based explanations of prediction models",
"year": "2022"
},
{
"authors": "S M Lundberg; G Erion; H Chen; A Degrave; J M Prutkin; B Nair; R Katz; J Himmelfarb; N Bansal; S I Lee",
"journal": "Nature machine intelligence",
"ref_id": "b16",
"title": "From local explanations to global understanding with explainable ai for trees",
"year": "2020"
},
{
"authors": "S Popov; S Morozov; A Babenko",
"journal": "",
"ref_id": "b17",
"title": "Neural oblivious decision ensembles for deep learning on tabular data",
"year": "2019"
},
{
"authors": "M T Ribeiro; S Singh; C Guestrin",
"journal": "",
"ref_id": "b18",
"title": "why should i trust you?\" explaining the predictions of any classifier",
"year": "2016"
},
{
"authors": "G Somepalli; M Goldblum; A Schwarzschild; C B Bruss; T Goldstein",
"journal": "",
"ref_id": "b19",
"title": "Saint: Improved neural networks for tabular data via row attention and contrastive pre-training",
"year": "2021"
},
{
"authors": "N Srivastava; G Hinton; A Krizhevsky; I Sutskever; R Salakhutdinov",
"journal": "The journal of machine learning research",
"ref_id": "b20",
"title": "Dropout: a simple way to prevent neural networks from overfitting",
"year": "2014"
},
{
"authors": "A Ultsch",
"journal": "",
"ref_id": "b21",
"title": "Clustering wih som: U* c",
"year": "2005-01"
},
{
"authors": "A Vaswani; N Shazeer; N Parmar; J Uszkoreit; L Jones; A N Gomez; Ł Kaiser; I Polosukhin",
"journal": "Advances in neural information processing systems",
"ref_id": "b22",
"title": "Attention is all you need",
"year": "2017"
}
] | [
{
"formula_coordinates": [
3,
286.28,
566.28,
194.32,
8.99
],
"formula_id": "formula_0",
"formula_text": "o = a x(1)"
},
{
"formula_coordinates": [
3,
215.85,
627.59,
264.74,
39.57
],
"formula_id": "formula_1",
"formula_text": "g(ŷ) = β • o = β 1 o 1 + ... + β i o i + ... + β p o p = β 1 a 1 x 1 + ... + β i a i x i + ... + β p a p x p(2)"
},
{
"formula_coordinates": [
4,
134.77,
497.03,
345.83,
13.38
],
"formula_id": "formula_2",
"formula_text": "p keys in K = [k 1 , k 2 , ..., k i , ..., k p ] T with k i = (k 1 i , ..., k d k i ) ∈ R d k using the key"
},
{
"formula_coordinates": [
4,
437.82,
510.21,
41.69,
13.38
],
"formula_id": "formula_3",
"formula_text": "d k i ) ∈ R d k"
},
{
"formula_coordinates": [
5,
242.12,
139.46,
238.48,
23.25
],
"formula_id": "formula_4",
"formula_text": "a i = q i • k i d k for i ∈ 1, • • • , p(3)"
},
{
"formula_coordinates": [
5,
324.88,
184.16,
43.18,
9.65
],
"formula_id": "formula_5",
"formula_text": "a i ∈ [0, 1]."
},
{
"formula_coordinates": [
5,
165.05,
290.96,
251.66,
51.28
],
"formula_id": "formula_6",
"formula_text": "Q = self.KeyEncoder(x) # Q is (b, p, d k ) K = self.QueryEncoder(x) # K is (b, p, d k ) QK = Q * K * self.scale # scale= 1/d k , QK is (b, p, d k ) a = QK.sum(axis = -1) # a is (b, p) return a"
},
{
"formula_coordinates": [
6,
254.59,
297.34,
191.04,
9.65
],
"formula_id": "formula_7",
"formula_text": "d 1 = p × (d k /4) and d 2 = p × (d k /2), d k ≥ 4."
},
{
"formula_coordinates": [
6,
145.32,
656.12,
62.71,
9.65
],
"formula_id": "formula_8",
"formula_text": "F 1 (x) = 5x 1 -"
},
{
"formula_coordinates": [
8,
217.01,
385.82,
263.59,
9.65
],
"formula_id": "formula_9",
"formula_text": "y = (5x 1 -5x 2 )1 x5≤0 + (5x 3 -5x 4 )1 x5>0(5)"
}
] |
||
10.18653/v1/2022.findings-naacl.55 | 2023-05-31 | [
{
"figure_ref": [
"fig_1",
"fig_2"
],
"heading": "Introduction",
"publication_ref": [
"b12",
"b15",
"b11",
"b18"
],
"table_ref": [],
"text": "People often express their information needs with multiple preferences or constraints. Queries corresponding to such needs typically implicitly express set operations such as intersection, difference, and union. For example, a movie-goer might be looking for a science-fiction film from the 90s which does not feature aliens and a reader might be interested in a historical fiction novel set in France. Similarly, a botanist attempting to identify a species based on their recollection might search for shrubs that are evergreen and found in Panama. Further, if the set of entities that satisfy the constraints is relatively small, a reader may like to see and explore an exhaustive list of these entities. In addition, to verify and trust a system's recommendations, users benefit from being shown evidence from trusted sources (Lamm et al., 2021).\nAddressing such queries has been primarily studied in the context of question answering with structured knowledge bases (KBs), where query constraints are grounded to predefined predicates and symbolically executed. However, KBs can be incomplete and expensive to curate and maintain. Meanwhile, advances in information retrieval may enable developing systems that can address such queries without relying on structured KBs, by matching query constraints directly to supporting evidence in text documents. However, queries that combine multiple constraints with implicit set operations are not well represented in existing retrieval benchmarks such as MSMarco (Nguyen et al., 2016) and Natural Questions (Kwiatkowski et al., 2019). Also, such datasets do not focus on retrieving an exhaustive document set, instead limiting annotation to the top few results of a baseline information retrieval system.\nTo analyze retrieval system performance on such queries, we present QUEST, a dataset with natural language queries from four domains, that are mapped to relatively comprehensive sets of entities corresponding to Wikipedia pages. We use categories and their mapping to entities in Wikipedia as a building block for our dataset construction approach, but do not allow access to this semistructured data source at inference time, to simulate text-based retrieval. Wikipedia categories represent a broad set of natural language descriptions of entity properties and often correspond to selective information need queries that could be plausibly issued by a search engine user. The relationship between property names and document text is often subtle and requires sophisticated reasoning to determine, representing the natural language inference challenge inherent in the task.\nOur dataset construction process is outlined in Figure 1. The base queries are semi-automatically generated using Wikipedia category names. To construct complex queries, we sample category names and compose them by using pre-defined templates (for example, A ∩ B \\ C). Next, we ask crowdworkers to paraphrase these automatically generated queries, while ensuring that the paraphrased queries are fluent and clearly describe what a user could be looking for. These are then validated for naturalness and fluency by a different set of crowdworkers, and filtered according to those criteria. Finally, for a large subset of the data, we collect scalar relevance labels based on the entity documents and fine-grained textual attributions mapping query constraints to spans of document text. Such annotation could aid the development of systems that can make precise inferences from trusted sources.\nPerforming well on this dataset requires systems that can match query constraints with corresponding evidence in documents and handle set operations implicitly specified by the query (see Figure 2), while also efficiently scaling to large collections of entities. We evaluate several retrieval systems by finetuning pretrained models on our dataset. Systems are trained to retrieve multidocument sets given a query. We find that current dual encoder and cross-attention models up to the size of T5-Large (Raffel et al., 2020) are largely not effective at performing retrieval for queries with set operations. Queries with conjunctions and negations prove to be especially challenging for models and systems are further challenged with combinations of set operations. Our error analysis reveals that non-relevant false positive entities are often caused by the model ignoring negated constraints, or ignoring the conjunctive constraints in a query."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b2",
"b28",
"b23",
"b8",
"b13",
"b26",
"b26",
"b22",
"b3",
"b15",
"b11",
"b24",
"b27",
"b9",
"b10",
"b14",
"b0",
"b14",
"b28",
"b29"
],
"table_ref": [],
"text": "Previous work in question answering and information retrieval has focused on QA over knowledge bases as well as open-domain QA and retrieval over a set of entities or documents. We highlight how these relate to our work below.\nKnowledge Base QA Several datasets have been proposed for question answering over knowledge bases (Berant et al., 2013;Yih et al., 2016;Talmor and Berant, 2018;Keysers et al., 2020;Gu et al., 2021, inter alia). These benchmarks require retrieval of a set of entities that exist as nodes or relations in an accompanying knowledge base. Questions are optionally supplemented with logical forms. Lan et al. (2021) provide a comprehensive survey of complex KBQA datasets.\nPrevious work has simultaneously noted that large curated KBs are incomplete (Watanabe et al., 2017). Notably, KBQA systems operate over a constrained answer schema, which limits the types of queries they can handle. Further, these schema are expensive to construct and maintain. For this reason, our work focuses on a setting where we do not assume access to a KB. We note that KBQA datasets have also been adapted to settings where a KB is incomplete or unavailable (Watanabe et al., 2017;Sun et al., 2019). This was done by either removing some subset of the data from the KB or ignoring the KB entirely. A key difference from these datasets is also that we do not focus on multihop reasoning over multiple documents. Instead, the relevance of an entity can be determined solely based on its document.\nOpen-Domain QA and Retrieval Many opendomain QA benchmarks, which consider QA over unstructured text corpora, have been proposed in prior work. Some of these, such as TREC (Craswell et al., 2020), MSMarco (Nguyen et al., 2016) and Natural Questions (Kwiatkowski et al., 2019) are constructed using \"found data\", using real user queries on search engines. Thakur et al. (2021) present a benchmark where they consider many such existing datasets. Datasets such as Hot-potQA (Yang et al., 2018), and MultiRC (Khashabi et al., 2018) have focused on multi-hop question answering. Other work has explored e-commerce datasets (for example, (Kong et al., 2022)), but these have not been released publicly. Notably, the focus of these datasets differs from ours as we focus on queries that contain implicit set operations over exhaustive answer sets. Such queries are not well represented in existing datasets because they occur in the tail of the query distributions considered.\nMulti-Answer Retrieval Related work (Min et al., 2021;Amouyal et al., 2022) also studies the problem of multi-answer retrieval, where systems are required to predict multiple distinct answers for a query. Min et al. (2021) adapt existing datasets (for example, WebQuestionsSP (Yih et al., 2016)) to study this setting and propose a new metric, MRecall@K, to evaluate exhaustive recall of multiple answers. We also consider the problem of multi-answer set retrieval, but consider queries that implicitly contain set constraints.\nIn concurrent work, RomQA (Zhong et al., 2022) proposes an open-domain QA dataset, focusing on combinations of constraints extracted from Wikidata. RomQA shares our motivation to enable answering queries with multiple constraints, which have possibly large answer sets. To make attribution to evidence feasible without human annotation, RomQA focuses on questions whose component constraints can be verified from single entity-linked sentences from Wikipedia abstracts, annotated with relations automatically through distant supervision, with high precision but possibly low recall (T-Rex corpus). In QUEST, we broaden the scope of queryevidence matching operations by allowing for attribution through more global, document-level inference. To make human annotation for attribution feasible, we limit the answer set size and the evidence for an answer to a single document."
},
{
"figure_ref": [],
"heading": "Dataset Generation",
"publication_ref": [],
"table_ref": [],
"text": "QUEST consists of 3357 queries paired with up to 20 corresponding entities. Each entity has an associated document derived from its Wikipedia page. The dataset is divided into 1307 queries for training, 323 for validation, and 1727 for testing.\nThe task for a system is to return the correct set of entities for a given query. Additionally, as the collection contains 325,505 entities, the task requires retrieval systems that can scale efficiently. We do not allow systems to access additional information outside of the text descriptions of entities at inference time. Category labels are omitted from all entity documents."
},
{
"figure_ref": [],
"heading": "Atomic Queries",
"publication_ref": [],
"table_ref": [],
"text": "The base atomic queries (i.e., queries without any introduced set operations) in our dataset are derived from Wikipedia category names2 . These are handcurated natural language labels assigned to groups of related documents in Wikipedia3 . Category assignments to documents allow us to automatically determine the set of answer entities for queries with high precision and relatively high recall. We compute transitive closures of all relevant categories to determine their answer sets.\nHowever, repurposing these categories for constructing queries poses challenges: 1) lack of evi- dence in documents: documents may not contain sufficient evidence for judging their relevance to a category, potentially providing noisy signal for relevance attributable to the document text, 2) low recall: entities may be missing from categories to which they belong. For about half of the dataset, we crowdsource relevance labels and attribution based on document text, and investigate recall through manual error analysis ( §5).\nWe select four domains to represent some diversity in queries: films, books, animals and plants. Focusing on four rather than all possible domains enables higher quality control. The former two model a general search scenario, while the latter two model a scientific search scenario."
},
{
"figure_ref": [],
"heading": "Introducing set operations",
"publication_ref": [],
"table_ref": [],
"text": "To construct queries with set operations, we define templates that represent plausible combinations of atomic queries. Denoting atomic queries as A, B and C, our templates and corresponding examples from different domains are listed in Table 1. Templates were constructed by composing three basic set operations (intersection, union and difference). They were chosen to ensure unambiguous interpretations of resulting queries by omitting those combinations of set operations that are non-associative.\nBelow we describe the logic behind sampling atomic queries (i.e., A, B, C) for composing com-plex queries, with different set operations. In all cases, we ensure that answer sets contain between 2-20 entities so that crowdsourcing relevance judgements is feasible. We sample 200 queries per template and domain, for a total of 4200 initial queries. The dataset is split into train + validation (80-20 split) and testing equally. In each of these sets, we sampled an equal number of queries per template. Intersection. The intersection operation for a template A∩B is particularly interesting and potentially challenging when both A and B have large answer sets but their intersection is small. We require the minimum answer set sizes of each A and B to be fairly large (>50 entities), while their intersection to be small (2-20 entities). Difference. Similar to intersection, we require the answer sets for both A and B to be substantial (>50 entities), but also place maximum size constraints on both A (<200 entities) and B (<10000 entities) as very large categories tend to suffer from recall issues in Wikipedia. We also limit the intersection of A and B (see reasoning in Appendix B). Union. For the union operation, we require both A and B to be well-represented through the entities in the answer set for their union A ∪ B. Hence, we require both A and B to have at least 3 entities. Further, we require their intersection to be non-zero but less than 1/3rd of their union. This is so that A and B are somewhat related queries. For all other templates that contain compositions of the above set operations, we apply the same constraints recursively. For example, for A∩B \\C, we sample atomic queries A and B for the intersection operation, then sample C based on the relationship between A ∩ B and C."
},
{
"figure_ref": [],
"heading": "Annotation Tasks",
"publication_ref": [],
"table_ref": [],
"text": "Automatically generating queries based on templates results in queries that are not always fluent and coherent. Further, entities mapped to a query may not actually be relevant and don't always have attributable evidence for judging their relevance. We conduct crowdsourcing to tackle these issues.\nThe annotation tasks aim at ensuring that 1) queries are fluent, unambiguous and contain diverse natural language logical connectives, (2) entities are verified as being relevant or non-relevant and (3) relevance judgements are attributed to document text for each relevant entity. Crowdsourcing is performed in three stages, described below. More annotation details and the annotation interfaces can be found in Appendix C."
},
{
"figure_ref": [],
"heading": "Paraphrasing",
"publication_ref": [],
"table_ref": [],
"text": "Crowdworkers were asked to paraphrase a templatically generated query so that the paraphrased query is fluent, expresses all constraints in the original query, and clearly describes what a user could be looking for. This annotation was done by one worker per query."
},
{
"figure_ref": [],
"heading": "Validation",
"publication_ref": [],
"table_ref": [],
"text": "This stage is aimed at validating the queries we obtain from the paraphrasing stage. Crowdworkers were given queries from the first stage and asked to label whether the query is 1) fluent, 2) equivalent to the original templatic query in meaning, and 3) rate its naturalness (how likely it is to be issued by a real user). This annotation was done by 3 workers per query. We excluded those queries which were rated as not fluent, unnatural or having a different meaning than the original query, based on a ma-jority vote. Based on the validation, we removed around around 11% of the queries from stage 1."
},
{
"figure_ref": [],
"heading": "Relevance Labeling",
"publication_ref": [
"b19"
],
"table_ref": [],
"text": "Next, crowdworkers were asked to provide relevance judgements for the automatically determined answer sets of queries. Specifically, they were given a query and associated entities/documents, and asked to label their relevance on a scale of 0-3 (definitely not relevant, likely not relevant, likely relevant, definitely relevant). They were asked to ensure that relevance should mostly be inferred from the document, but they could use some background knowledge and do minimal research.\nWe also asked them to provide attributions for document relevance. Specifically, we ask them to first label whether the document provides sufficient evidence for the relevance of the entity (complete/partial/no). Then, for different phrases in the query (determined by the annotator), we ask them to mark sentence(s) in the document that indicate its relevance. The attribution annotation is broadly inspired by Rashkin et al. (2021). For negated constraints, we ask annotators to mark attributable sentences if they provide counter-evidence. Since this annotation was time-intensive, we collected these annotations for two domains (films and books). We found that relevance labeling was especially difficult for the plants and animals domains, as they required more specialized scientific knowledge. In our pilot study prior to larger scale data collection, we collected 3 relevance ratings from different annotators for 905 query and document pairs from the films domain. In 61.4% of cases, all 3 raters judged the document to be \"Definitely relevant\" or \"Likely relevant\" or all 3 raters judged the document to be \"Definitely not relevant\" or \"Likely not relevant\". The Fleiss' kappa metric on this data was found to be K=0.43. We excluded all entities which were marked as likely or definitely not relevant to a query based on the document text from its answer set. Around 23.7% of query-document pairs from stage 2 were excluded. "
},
{
"figure_ref": [],
"heading": "Dataset Statistics",
"publication_ref": [],
"table_ref": [
"tab_1"
],
"text": "Basic dataset statistics are reported in Table 2. The dataset contains more entities from the films domain, because this domain is more populated in Wikipedia. The average length of queries is 8.6 words and the average document length is 452 words. Documents from the films and books domains are longer on average, as they often contain plots and storylines. Around ∼69% of entities have complete evidence and ∼30% have partial evidence. Evidence was labeled as partial when not all phrases in the query had explicit evidence in the document (i.e., they may require background knowledge or reasoning). There are on average 33.2 words attributed for each entity with the maximum attribution text span ranging up to length 1837 words. Finally, the average answer set size is 10.5 entities."
},
{
"figure_ref": [],
"heading": "Additional Training Examples",
"publication_ref": [],
"table_ref": [],
"text": "Beyond the annotated data, we generated additional synthetic examples for training. We found including such examples improved model performance, and we include these examples for the experiments in §4. To generate these examples, we sample 5000 atomic queries from all domains, ensuring that they do not already appear as sub-queries in any of the queries in QUEST and use their corresponding entities in Wikipedia as their relevant entity set."
},
{
"figure_ref": [],
"heading": "Experimental Setup",
"publication_ref": [],
"table_ref": [],
"text": "We evaluate modern retrieval systems to establish baseline performances. We also perform extensive error analysis to understand patterns of model errors and the quality of the labels in QUEST."
},
{
"figure_ref": [],
"heading": "Task Definition",
"publication_ref": [],
"table_ref": [],
"text": "We consider a corpus, E, that contains entities across all domains in the dataset. Each entity is accompanied with a document based on its Wikipedia page. An example in our dataset consists of a query, x, and an annotated set of relevant entities, y ⊂ E. As described in §3, for all examples |y| < 20. Our task is to develop a system that, given E and a query x, predicts a set of relevant entities, ŷ ⊂ E."
},
{
"figure_ref": [],
"heading": "Evaluation",
"publication_ref": [],
"table_ref": [],
"text": "Our primary evaluation metric is average F 1 , which averages per-example F 1 scores. We compute F 1 for each example by comparing the predicted set of entities, ŷ, with the annotated set, y."
},
{
"figure_ref": [
"fig_3"
],
"heading": "Baseline Systems",
"publication_ref": [
"b20",
"b16",
"b18",
"b7",
"b17"
],
"table_ref": [],
"text": "We evaluated several combinations of retrievers and classifiers, as shown in Figure 3. For the retriever component, we consider a sparse BM25 retriever (Robertson et al., 2009) and a dense dual encoder retriever (denoted DE). Following Ni et al. (2022), we initialize our dual encoder from a T5 (Raffel et al., 2020) encoder and train with an in-batch sampled softmax loss (Henderson et al., 2017). Once we have a candidate set, we need to determine a set of relevant entities. To classify relevance of each candidate document for the given query, we consider a cross-attention model which consists of a T5 encoder and decoder. 4 We train the cross-attention classifier using a binary cross-entropy loss with negative examples based on non-relevant documents in top 1,000 documents retrieved by BM25 and random non-relevant documents (similarly to Nogueira and Cho (2019)). As cross-attention classification for a large number of candidates is computationally expensive, we restrict BM25 and the dual encoder to retrieve 100 candidates which are then considered by the crossattention classifier. As our T5-based dual encoder can only efficiently accommodate up to 512 tokens, 4 Scores from BM25 and dual encoders trained with a softmax loss are not normalized to provide relevance probabilities for documents. We found that naively applying a global threshold to these scores to produce answer sets did not perform as well as using a classifier trained with a binary cross-entropy loss to predict document relevance. we truncate document text. We discuss the impact of this and alternatives in §5. Further, since T5 was pre-trained on Wikipedia, we investigate the impact of memorization in Appendix D. Additional details and hyperparameter settings are in Appendix A."
},
{
"figure_ref": [],
"heading": "Manual Error Annotation",
"publication_ref": [],
"table_ref": [],
"text": "For the best overall system, we sampled errors and manually annotated 1145 query-document pairs from the validation set. For the retriever, we sampled relevant documents not included in the top-100 candidate set and non-relevant documents ranked higher than relevant ones. For the classifier, we sampled false positive and false negative errors made in the top-100 candidate set. This annotation process included judgements of document relevance (to assess agreement with the annotations in the dataset) and whether the document (and the truncated version considered by the dual encoder or classifier) contained sufficient evidence to reasonably determine relevance. We also annotated relevance for each constraint within a query. We discuss these results in §5."
},
{
"figure_ref": [],
"heading": "Results and Analysis",
"publication_ref": [
"b14",
"b21",
"b6",
"b1"
],
"table_ref": [
"tab_2",
"tab_2",
"tab_3"
],
"text": "We report the performance of our baseline systems on the test set in Table 3. In this section, we summarize the key findings from our analysis of these results and the error annotation described in §4.4.\nDual encoders outperform BM25. As shown in Table 3, the best overall system uses a T5-Large Dual Encoder instead of BM25 for retrieval. The performance difference is even more significant when comparing recall of Dual Encoders and BM25 directly. We report average recall (average per-example recall of the full set of relevant documents) and MRecall (Min et al., 2021) (the percentage of examples where the candidate set contains all relevant documents), over various candidate set sizes in Table 4. Retrieval and classification are both challenging. As we consider only the top-100 candidates from the retriever, the retriever's recall@100 sets an upper bound on the recall of the overall system. Recall@100 is only 0.476 for the T5-Large Dual Encoder, and the overall recall is further reduced by the T5-Large classifier to 0.368, despite achieving only 0.165 precision. This suggests that there is room for improvement from both stages to improve overall scores. As performance improves for larger T5 sizes for both retrieval and classification, further model scaling could be beneficial. Models struggle with intersection and difference. We also analyzed results across different templates and domains, as shown in Table 5. Different constraints lead to varying distributions over answer set sizes and the atomic categories used. Therefore, it can be difficult to interpret differences in F1 scores across templates. Nevertheless, we found the queries with set union have the highest average F1 scores. Queries with set intersection have the lowest average F1 scores, and queries with set difference also appear to be challenging.\nTo analyze why queries with conjunction and negation are challenging, we labeled the relevance of individual query constraints ( §4.4), where a system incorrectly judges relevance of a non-relevant document. The results are summarized in Table 6. For a majority of false positive errors involving intersection, at least one constraint is satisfied. This could be interpreted as models incorrectly treating intersection as union when determining relevance. Similarly, for a majority of examples with set difference, the negated constraint is not satisfied. This suggests that the systems are not sufficiently sensitive to negations. There is significant headroom to improve both precision and recall. As part of our manual error analysis ( §4.4), we made our own judgements of relevance and measured agreement with the relevance annotations in QUEST. As this analysis focused on cases where our best system disagreed with the relevance labels in the dataset, we would expect agreement on these cases to be significantly lower than on randomly selected query-document pairs in the dataset. Therefore, it provides a focused way to judge the headroom and annotation quality of the dataset.\nFor false negative errors, we judged 91.1% of the entities to be relevant for the films and books domains, and 81.4% for plants and animals. Notably, we collected relevance labels for the films and books domains and removed some entities based on these labels, as described in §3, which likely explains the higher agreement for false negatives from these domains. This indicates significant headroom for improving recall as defined by QUEST, especially for the domains where we collected relevance labels.\nFor false positive errors, we judged 28.8% of the entities to be relevant, showing a larger disagreement with the relevance labels in the dataset. This is primarily due to entities not included in the entity sets derived from the Wikipedia category taxonomy (97.7%), rather than entities removed due to relevance labeling. This is a difficult issue to fully resolve, as it is not feasible to exhaustively label relevance for all entities to correct for recall issues in the Wikipedia category taxonomy. Future work can use pooling to continually grow the set Table 6: Analysis of false positive errors from the T5-Large classifier and cases where a non-relevant document was ranked ahead of a relevant one for the T5-Large dual encoder. For queries with conjunction, we determined the percentage of cases where 1, 2, or 3 constraints in the template were not satisfied by the predicted document (# Constraints). For queries with negation, we measured the percentage of cases where the negated constraint (Neg.) was not satisfied.\nof relevant documents (Sparck Jones and Van Rijsbergen, 1975). Despite this, our analysis suggests there is significant headroom for improving precision, as we judged a large majority of the false positive predictions to be non-relevant.\nTruncating document text usually provides sufficient context. In our experiments, we truncate document text to 512 tokens for the dual encoder, and 384 tokens for the classifier to allow for the document and query to be concatenated. Based on our error analysis ( §4.4), out of the documents with sufficient evidence to judge relevance, evidence occurred in this truncated context 93.2% of the time for the dual encoder, and 96.1% of the time for the classifier. This may explain the relative success of this simple baseline for handling long documents. We also evaluated alternative strategies but these performed worse in preliminary experiments5 . Future work can evaluate efficient transformer variants (Guo et al., 2022;Beltagy et al., 2020)."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [
"b25"
],
"table_ref": [],
"text": "We present QUEST, a new benchmark of queries which contain implicit set operations with corresponding sets of relevant entity documents. Our experiments indicate that such queries present a challenge for modern retrieval systems. Future work could consider approaches that have better inductive biases for handling set operations in natural language expressions (for example, Vilnis et al. (2018)). The attributions in QUEST can be leveraged for building systems that can provide finegrained attributions at inference time. The potential of pretrained generative LMs and multi-evidence aggregation methods to answer set-seeking selective queries, while providing attribution to sources, can also be investigated."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [
"b29"
],
"table_ref": [],
"text": "Naturalness. Since our dataset relies on the Wikipedia category names and semi-automatically generated compositions, it does not represent an unbiased sample from a natural distribution of real search queries that contain implicit set operations. Further, we limit attention to non-ambiguous queries and do not address the additional challenges that arise due to ambiguity in real search scenarios. However, the queries in our dataset were judged to plausibly correspond to real user search needs and system improvements measured on QUEST should correlate with improvements on at least a fraction of natural search engine queries with set operations.\nRecall. We also note that because Wikipedia categories have imperfect recall of all relevant entities (that contain sufficient evidence in their documents), systems may be incorrectly penalised for predicted relevant entities assessed as false positive. We quantify this in section 5. We have also limited the trusted source for an entity to its Wikipedia document but entities with insufficient textual evidence in their documents may still be relevant. Ideally, multiple trusted sources could be taken into account and evidence could be aggregated to make relevance decisions. RomQA (Zhong et al., 2022) takes a step in this latter direction although the evidence attribution is not manually verified.\nAnswer Set Sizes. To ensure that relevance labels are correct and verifiable, we seek the help of crowdworkers. However, this meant that we needed to restrict the answer set sizes to 20 for the queries in our dataset, to make annotation feasible. On one hand, this is realistic for a search scenario because users may only be interested in a limited set of results. On the other hand, our dataset does not model a scenario where the answer set sizes are much larger."
},
{
"figure_ref": [],
"heading": "A Experimental Details and Hyperparameters",
"publication_ref": [],
"table_ref": [],
"text": "All models were fine-tuned starting from T5 1.1 checkpoints 6 . We fine-tune T5 models on 32 Cloud TPU v3 cores 7 . Fine-tuning takes less than 8 hours for all models.\nDual Encoder. We used the t5x_retrieval library 8 for implementing dual encoder models. We tuned some parameters based on results on the validation set. Relevant hyperparameters for training the dual encoder are:\n• Learning Rate: 1e-3\n• Warmup Steps: 1500\n• Finetuning Steps: 15000\n• Batch Size: 512\n• Max Query Length: 64\n• Max Candidate Length: 512\nClassifier.\nFor negative examples, we sampled 250 random non-relevant documents and sampled 250 non-relevant documents from the top-1000 documents retrieved by BM25. We also replicated each positive example 50 times. We found an approximately even number of positive and negative examples lead to better performance than training with a large class imbalance. We found a combination of random negatives and negatives from BM25 performed better than using only either individual type of negative examples. Additionally, selecting negative examples from BM25 performed better than selecting negative examples from the T5-Large dual encoder.\nFor the T5 input we concatenated the query and truncated document text. The T5 output is the string \"relevant\" or \"not relevant\". To classify document relevance at inference time, we applied a threshold to the probability assigned to the \"relevant\" label, which we tuned on the validation set. When classifying BM25 candidates we used a threshold of 0.9 and when classifying the dual encoder candidates we used a threshold of 0.95.\nOther relevant hyperparameters for training the classifier are: Discussion While recall is simply equal to r A , precision is a more complicated function of r B and r ∩ , and can be very low for large values of r ∩ . Intuitively, if subtracting B from  removes most of Â, then the precision of the resulting set will be dominated by the relevant entities missing from B. This motivates limiting the intersection of the two sets used to construct queries involving set intersection. For example, if r B = 0.95, then with r ∩ < 0.8, we can ensure p > 0.83."
},
{
"figure_ref": [
"fig_7",
"fig_8"
],
"heading": "C Annotation Details",
"publication_ref": [],
"table_ref": [],
"text": "The annotation tasks in QUEST were carried out by participants who were paid contractors. They are based in Austin, TX and either have a bachelor's degree (55%) or equivalent work experience (45%). They were paid by the hour for their work and were recruited from a vendor who screened them for knowledge of US English. They were informed of how their work would be used and could opt out. They received a standard contracted wage, which complies with living wage laws in their country of employment. The annotation interfaces presented to the annotators are shown in Figures 4, 5 and6."
},
{
"figure_ref": [],
"heading": "D Impact of Memorization of Pre-training Data",
"publication_ref": [],
"table_ref": [],
"text": "Since the T5 checkpoints we use to initialize our models were pre-trained on the C4 corpus (which includes Wikipedia), we investigate whether these models have memorized aspects of the Wikipedia category graph. We compare recall of the T5-based dual encoder model for Wikipedia documents that were created prior to the pre-training date of the T5 checkpoint compared with documents that were added after pre-training. We report these in Table 7, along with the recalls for the same sets of documents with a BM25 retriever, for a baseline Avg. Recall@100 Retriever Before After BM25 0.183 0.050 T5-Large DE 0.466 0.171 Table 7: Average recall@100 on the subsets of documents created before vs after T5 pre-training.\ncomparison. We note that the ratio of scores between the documents added before pre-training to documents added after pre-training is similar for both systems, which suggests factors other than memorization may explain the difference. For example, the documents created before vs. after the pre-training date have average lengths of 759.7 vs. 441.2 words, respectively. "
},
{
"figure_ref": [],
"heading": "Acknowledgements",
"publication_ref": [],
"table_ref": [],
"text": "We would like to thank Isabel Kraus-Liang, Mahesh Maddinala, Andrew Smith, Daphne Domansi, and all the annotators for their work. We would also like to thank Mark Yatskar, Dan Roth, Zhuyun Dai, Jianmo Ni, William Cohen, Andrew McCallum, Shib Sankar Dasgupta and Nicholas Fitzgerald for useful discussions."
}
] | Formulating selective information needs results in queries that implicitly specify set operations, such as intersection, union, and difference. For instance, one might search for "shorebirds that are not sandpipers" or "science-fiction films shot in England". To study the ability of retrieval systems to meet such information needs, we construct QUEST, a dataset of 3357 natural language queries with implicit set operations, that map to a set of entities corresponding to Wikipedia documents. The dataset challenges models to match multiple constraints mentioned in queries with corresponding evidence in documents and correctly perform various set operations. The dataset is constructed semi-automatically using Wikipedia category names. Queries are automatically composed from individual categories, then paraphrased and further validated for naturalness and fluency by crowdworkers. Crowdworkers also assess the relevance of entities based on their documents and highlight attribution of query constraints to spans of document text. We analyze several modern retrieval systems, finding that they often struggle on such queries. Queries involving negation and conjunction are particularly challenging and systems are further challenged with combinations of these operations. | QUEST: A Retrieval Dataset of Entity-Seeking Queries with Implicit Set Operations | [
{
"figure_caption": "Birds of Venezuelan Andes -Birds of ColombiaBirds found in the Venezuelan Andes but not in Colombia",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 1 :1Figure 1: The dataset construction process for QUEST. First, (1) we sample Wikipedia category names and find their corresponding set of relevant entities. (2) Then, we compose a query with set operations and have this query paraphrased by crowdworkers. (3) These queries are then validated for fluency and naturalness. (4) Finally, crowdworkers mark the entities' relevance by highlighting attributable spans in their documents.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: An example of a query and relevant entity from QUEST. The attribution for different query constraints can come from different parts of the document.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure3: We compare several systems consisting of a retriever for efficiently selecting a set of candidates from the document corpus and a document relevance classifier for determining the final predicted document set.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "For simplicity, we also assume that the overlap between  and B is such that|  ∩ B| = r A * |A ∩ B| and |  ∩ B| = r A * r B * |A ∩ B|.Derivation What is the recall (r) and precision (p) of  \\ B relative to A \\ B as a function of r A , r B , and r ∩ ?First, we derive this function for recall: 9r = |(A \\ B) ∩ (  \\ B)| |(A \\ B)| r = |(  \\ B)| |(A \\ B)| r = | Â| -|  ∩ B| |A| -|A ∩ B| r = r A * |A| -r A * r ∩ * |A| |A| -(r ∩ * |A|) r = r A * (1 -r ∩ ) * |A| (1 -r ∩ ) * |A| r = r AAnd for precision:p = |(A \\ B) ∩ (  \\ B)| |(  \\ B)|9 We note some useful properties of pairs of sets X and Y :X \\ Y = X ∩ Y c , |X \\ Y | = |X| -|X ∩ Y |, if X ⊂ Y then X ∩ Y = X, and if X ⊂ Y then Y c ⊂ X c . p = |(  \\ B)| |(  \\ B)| p = | Â| -|  ∩ B| | Â| -|  ∩ B| p = r A * |A| -r A * r ∩ * |A| r A * |A| -r A * r B * r ∩ * |A| p = r A * (1 -r ∩ ) * |A| r A * (1 -r B * r ∩ ) * |A| p = (1 -r ∩ ) (1 -r B * r ∩ )",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: Annotation interface for the paraphrasing stage.",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 :5Figure 5: Annotation interface for the validation stage.",
"figure_data": "",
"figure_id": "fig_7",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Figure 6 :6Figure 6: Annotation interface for the relevance labeling stage.",
"figure_data": "",
"figure_id": "fig_8",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "Trees from the Northwestern US that can't be found in Canada 61 A ∪ B ∪ C Moths or Insects or Arthropods of Guadeloupe 121 A ∩ B ∩ C Plants the Arctic, the United Kingdom, and the Caucasus have in common 123 A ∩ B \\ C",
"figure_data": "DomainTemplateExampleNum. QueriesABiographical Italian bandits films125A ∪ BDutch crime comedy or romantic comedy films135A ∩ BItalian crime films set in the 1970's143FilmsA \\ BIndian sport films that are not about cricket126A ∪ B ∪ CDutch or Swiss war films, or war films from 1945122A ∩ B ∩ C2020's drama films shot in cleveland124A ∩ B \\ CEpic films about Christianity not set in Israel121A2004 German novels125A ∪ B1925 Russian novels or Novels by Ivan Bunin125A ∩ B1991 Novels set in Iceland133BooksA \\ BNovels set in the 1900s not based on real events123A ∪ B ∪ CNovels set in Nanjing, Hebei, or Jiangsu125A ∩ B ∩ CEnglish language Harper & Brothers Children's fiction books124A ∩ B \\ CNovels that take place in Vietnam that aren't about war115Aplants only from Gabon115A ∪ BTrees of Manitoba or Subarctic America125A ∩ BShrubs used in traditional Native American medicine135PlantsA \\ BOrchids of Indonesia and Malaysia but not Thailand122Awhat are the Rodents of Cambodia115A ∪ BAnimals from Cuba or Jamaica that are extinct121A ∩ BNeogene mammals of Africa that are Odd-toed ungulates111AnimalsA \\ B",
"figure_id": "tab_0",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Statistics of examples in QUEST across different domains.",
"figure_data": "Films Books Plants AnimalsAllNum. Queries8968708027893357Num. Entities146368 50784 8367244681325505Avg. Query Len.8.687.938.949.098.64Avg. Doc. Len.532.2655.3 258.1293.1452.2Avg. Ans. Set Size8.88.612.212.610.5",
"figure_id": "tab_1",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Average Precision, Recall, and F1 of baseline systems evaluated on the test set.",
"figure_data": "Retriever (K=100) Classifier Avg. Precision Avg. Recall Avg. F1BM25T5-Base0.1680.1600.141BM25T5-Large0.1780.1680.150T5-Large DET5-Base0.1530.3540.176T5-Large DET5-Large0.1650.3680.192Avg. Recall@[email protected] 0.153 0.197 0.395 0.020 0.030 0.037 0.087T5-Base DE 0.255 0.372 0.455 0.726 0.045 0.088 0.127 0.360T5-Large DE 0.265 0.386 0.476 0.757 0.047 0.100 0.142 0.408",
"figure_id": "tab_2",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Average Recall and MRecall of various retrievers.",
"figure_data": "",
"figure_id": "tab_3",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "and Assumptions Let us assume we have two sets derived from the Wikipedia category graph,  and B. The Wikipedia category graph can be missing some relevant entities, such that  ⊂ A and B ⊂ B, where A and B are interpreted as the hypothetical sets containing all relevant entities. We quantify the degree of missing entities by denoting recall as r A and r B , such that | Â| = r A * |A| and | B| = r B * |B|. We quantify the fraction of elements in A that are also in B as r ∩ , such that |A ∩ B| = r ∩ * |A|.",
"figure_data": "• Learning Rate: 1e-3• Warmup Steps: 1000• Finetuning Steps: 10000• Batch Size: 1024• Max Source Length: 512• Max Target Length: 16B Set Difference and Recall",
"figure_id": "tab_6",
"figure_label": "",
"figure_type": "table"
}
] | Chaitanya Malaviya; Peter Shaw; Ming-Wei Chang; Kenton Lee; Kristina Toutanova | [
{
"authors": "Samuel Joseph Amouyal; Ohad Rubin; Ori Yoran; Tomer Wolfson; Jonathan Herzig; Jonathan Berant",
"journal": "",
"ref_id": "b0",
"title": "Qampari:: An open-domain question answering benchmark for questions with many answers from multiple paragraphs",
"year": "2022"
},
{
"authors": "Iz Beltagy; Matthew E Peters; Arman Cohan",
"journal": "",
"ref_id": "b1",
"title": "Longformer: The long-document transformer",
"year": "2020"
},
{
"authors": "Jonathan Berant; Andrew Chou; Roy Frostig; Percy Liang",
"journal": "Association for Computational Linguistics",
"ref_id": "b2",
"title": "Semantic parsing on Freebase from question-answer pairs",
"year": "2013"
},
{
"authors": "Nick Craswell; Mitra Bhaskar; Emine Yilmaz; Daniel Campos; Ellen M Voorhees",
"journal": "",
"ref_id": "b3",
"title": "Overview of the TREC 2019 deep learning track",
"year": "2020"
},
{
"authors": "Zhuyun Dai; Jamie Callan",
"journal": "",
"ref_id": "b4",
"title": "Deeper text understanding for ir with contextual neural language modeling",
"year": "2019"
},
{
"authors": "Yu Gu; Sue Kase; Michelle Vanni; Brian Sadler; Percy Liang; Xifeng Yan; Yu Su",
"journal": "",
"ref_id": "b5",
"title": "Beyond iid: three levels of generalization for question answering on knowledge bases",
"year": "2021"
},
{
"authors": "Mandy Guo; Joshua Ainslie; David Uthus; Santiago Ontanon; Jianmo Ni; Yun-Hsuan Sung; Yinfei Yang",
"journal": "Association for Computational Linguistics",
"ref_id": "b6",
"title": "LongT5: Efficient text-to-text transformer for long sequences",
"year": "2022"
},
{
"authors": "Matthew Henderson; Rami Al-Rfou; Brian Strope; Yun-Hsuan Sung; László Lukács; Ruiqi Guo; Sanjiv Kumar; Balint Miklos; Ray Kurzweil",
"journal": "",
"ref_id": "b7",
"title": "Efficient natural language response suggestion for smart reply",
"year": "2017"
},
{
"authors": "Daniel Keysers; Nathanael Schärli; Nathan Scales; Hylke Buisman; Daniel Furrer; Sergii Kashubin; Nikola Momchev; Danila Sinopalnikov; Lukasz Stafiniak; Tibor Tihon; Dmitry Tsarkov; Xiao Wang; Marc Van Zee; Olivier Bousquet",
"journal": "",
"ref_id": "b8",
"title": "Measuring compositional generalization: A comprehensive method on realistic data",
"year": "2020-04-26"
},
{
"authors": "Daniel Khashabi; Snigdha Chaturvedi; Michael Roth; Shyam Upadhyay; Dan Roth",
"journal": "",
"ref_id": "b9",
"title": "Looking beyond the surface: A challenge set for reading comprehension over multiple sentences",
"year": "2018"
},
{
"authors": "Weize Kong; Swaraj Khadanga; Cheng Li; Shaleen Gupta; Mingyang Zhang; Wensong Xu; Mike Bendersky",
"journal": "",
"ref_id": "b10",
"title": "Multi-aspect dense retrieval",
"year": "2022"
},
{
"authors": "Tom Kwiatkowski; Jennimaria Palomaki; Olivia Redfield; Michael Collins; Ankur Parikh; Chris Alberti; Danielle Epstein; Illia Polosukhin; Jacob Devlin; Kenton Lee; Kristina Toutanova; Llion Jones; Matthew Kelcey; Ming-Wei Chang; Andrew M Dai; Jakob Uszkoreit; Quoc Le; Slav Petrov",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b11",
"title": "Natural questions: A benchmark for question answering research",
"year": "2019"
},
{
"authors": "Matthew Lamm; Jennimaria Palomaki; Chris Alberti; Daniel Andor; Eunsol Choi; Livio Baldini Soares; Michael Collins",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b12",
"title": "QED: A Framework and Dataset for Explanations in Question Answering",
"year": "2021"
},
{
"authors": "Yunshi Lan; Gaole He; Jinhao Jiang; Jing Jiang; Wayne Xin Zhao; Ji-Rong Wen",
"journal": "",
"ref_id": "b13",
"title": "A survey on complex knowledge base question answering: Methods, challenges and solutions",
"year": "2021"
},
{
"authors": "Sewon Min; Kenton Lee; Ming-Wei Chang; Kristina Toutanova; Hannaneh Hajishirzi",
"journal": "Association for Computational Linguistics",
"ref_id": "b14",
"title": "Joint passage ranking for diverse multi-answer retrieval",
"year": "2021"
},
{
"authors": "Tri Nguyen; Mir Rosenberg; Xia Song; Jianfeng Gao; Saurabh Tiwary; Rangan Majumder; Li Deng",
"journal": "",
"ref_id": "b15",
"title": "Ms marco: A human generated machine reading comprehension dataset",
"year": "2016"
},
{
"authors": "Jianmo Ni; Chen Qu; Jing Lu; Zhuyun Dai; Gustavo Hernandez Abrego; Ji Ma; Vincent Zhao; Yi Luan; Keith Hall; Ming-Wei Chang; Yinfei Yang",
"journal": "",
"ref_id": "b16",
"title": "Large dual encoders are generalizable retrievers",
"year": "2022"
},
{
"authors": "Rodrigo Nogueira; Kyunghyun Cho",
"journal": "",
"ref_id": "b17",
"title": "Passage re-ranking with bert",
"year": "2019"
},
{
"authors": "Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu",
"journal": "Journal of Machine Learning Research",
"ref_id": "b18",
"title": "Exploring the limits of transfer learning with a unified text-to-text transformer",
"year": "2020"
},
{
"authors": "Vitaly Hannah Rashkin; Matthew Nikolaev; Lora Lamm; Michael Aroyo; Dipanjan Collins; Slav Das; Gaurav Petrov; Iulia Singh Tomar; David Turc; Reitter",
"journal": "",
"ref_id": "b19",
"title": "Measuring attribution in natural language generation models",
"year": "2021"
},
{
"authors": "Stephen Robertson; Hugo Zaragoza",
"journal": "Foundations and Trends® in Information Retrieval",
"ref_id": "b20",
"title": "The probabilistic relevance framework: Bm25 and beyond",
"year": "2009"
},
{
"authors": "K Sparck Jones; C J Van Rijsbergen",
"journal": "",
"ref_id": "b21",
"title": "Report on the need for and provision of an ideal information retrieval test collection",
"year": "1975"
},
{
"authors": "Haitian Sun; Tania Bedrax-Weiss; William Cohen",
"journal": "Association for Computational Linguistics",
"ref_id": "b22",
"title": "PullNet: Open domain question answering with iterative retrieval on knowledge bases and text",
"year": "2019"
},
{
"authors": "Alon Talmor; Jonathan Berant",
"journal": "Association for Computational Linguistics",
"ref_id": "b23",
"title": "The web as a knowledge-base for answering complex questions",
"year": "2018"
},
{
"authors": "Nandan Thakur; Nils Reimers; Andreas Rücklé; Abhishek Srivastava; Iryna Gurevych",
"journal": "",
"ref_id": "b24",
"title": "BEIR: A heterogeneous benchmark for zero-shot evaluation of information retrieval models",
"year": "2021"
},
{
"authors": "Luke Vilnis; Xiang Li; Shikhar Murty; Andrew Mccallum",
"journal": "Association for Computational Linguistics",
"ref_id": "b25",
"title": "Probabilistic embedding of knowledge graphs with box lattice measures",
"year": "2018"
},
{
"authors": "Yusuke Watanabe; Bhuwan Dhingra; Ruslan Salakhutdinov",
"journal": "",
"ref_id": "b26",
"title": "Question answering from unstructured text by retrieval and comprehension",
"year": "2017"
},
{
"authors": "Zhilin Yang; Peng Qi; Saizheng Zhang; Yoshua Bengio; William Cohen; Ruslan Salakhutdinov; Christopher D Manning",
"journal": "Association for Computational Linguistics",
"ref_id": "b27",
"title": "HotpotQA: A dataset for diverse, explainable multi-hop question answering",
"year": "2018"
},
{
"authors": "Wen-Tau Yih; Matthew Richardson; Chris Meek; Ming-Wei Chang; Jina Suh",
"journal": "Association for Computational Linguistics",
"ref_id": "b28",
"title": "The value of semantic parse labeling for knowledge base question answering",
"year": "2016"
},
{
"authors": "Victor Zhong; Weijia Shi; Wen-Tau Yih; Luke Zettlemoyer",
"journal": "",
"ref_id": "b29",
"title": "RoMQA: A benchmark for robust, multi-evidence, multi-answer question answering",
"year": "2022"
}
] | [] |
2023-05-19 | [
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b57",
"b47",
"b48",
"b38",
"b11",
"b21",
"b49",
"b57",
"b43",
"b3",
"b48",
"b58",
"b24",
"b1",
"b50",
"b9",
"b17",
"b19",
"b42",
"b14",
"b7",
"b44",
"b13",
"b57",
"b2",
"b16",
"b8",
"b27",
"b5",
"b46"
],
"table_ref": [],
"text": "The explosive growth of online text data has highlighted the importance of text content recommendation in numerous domains, including e-commerce, news recommendation, and social media. Text-based collaborative filtering (TCF) has emerged as a critical technology that provides personalized recommendations to users based on textual data, such as product descriptions, reviews, or news articles [58,48,49]. The goal of TCF is to accurately capture the user's preferences and interests from textual data and provide tailored recommendations that align with their needs. TCF typically utilizes language models (LMs) as text encoders, integrated into a recommender architecture using collaborative filtering techniques [39,12,22] to generate user-item matching scores (see Figure 1). TCF's promising results have made it the mainstream approach for text-based recommendation [50,58,44].\nBy using LMs as item encoders, TCF naturally benefits from the latest advances in the field of natural language processing (NLP). Particularly, in recent years, large LMs such as GPT-3 [4] and Clearly, the above questions are essential for directing research on the mainstream TCF paradigm. However, despite many TCF algorithms has been proposed in literature [49,59,25,2,51,10], none of them have explicitly addressed the above questions. Therefore, rather than introducing a new algorithm as before, we aim to decipher the classic TCF models via a series of audacious experiments that require immense computational resources. Specifically, we explore the below sub-questions.\nQ1: How does the recommender system's performance respond to the continuous increase in the item encoder's size? Is the performance limits attainable at the scale of hundreds of billions? To answer it, we conduct the empirical study by progressively expanding the size of the text encoder from 100 million to 175 billion on three recommendation datasets, using the two most representative recommender architectures: a simple two-tower CTR model DSSM [18] and a state-of-the-art sequential model SASRec [20] with Transformer [43] as the backbone.\nQ2: Can super-large LMs, such as GPT-3 with 175-billion parameters, generate universal text representations? Developing universal foundation models is an ambitious goal in the field of NLP. Many studies have shown that the representations learned by these massive LMs are applicable to various NLP tasks. Unlike them, we use the user-centered recommendation as the downstream task to explore the universality of a 175-billion LM pre-trained on non-recommendation data.\nQ3: Can recommender models with a 175-billion parameter LM as the item encoder easily beat the simplest ID embedding based models (IDCF), especially for warm item recommendation? IDCF is a prevailing recommendation paradigm that has dominated the recommender system (RS) community for more than a decade. It produces high-quality recommendations without relying on any item content information. However, recent studies [15,8,45,14] suggest that ID features are the key barrier to achieving foundation or transferable recommender models since they are in general not shareable in practice. [58] discovered that to compete with IDCF, TCF has to retrain its text encoder on the recommendation dataset, otherwise there is still a big accuracy gap. However, their study only explored some medium-sized text encoders with around 100 million parameters. What would happen if we use the 175-billion GPT-3 as the text encoder? Would it be necessary to retrain GPT-3 to outperform IDCF for non-cold or warm item recommendation tasks? Q4: How close is the TCF paradigm to a universal recommender model? In addition to performance, a frequently mentioned advantage of TCF is its potential transferability, which could enable cross-domain and cross-platform recommendations and establish a universal foundation model [3,17] for the recommender system field. Thus, we aim to investigate the cross-domain recommendation capability of TCF with a text encoder of 175-billion parameters.\nQ5: Will the classic TCF paradigm be replaced by a recent prompt engineering based recommendation method that utilizes ChatGPT (called ChatGPT4Rec)? With the emergence of ChatGPT, a series of recent work [9,28,6,47] have leveraged the ChatGPT API and prompt to generate recommendations directly, eliminating the need for training and resulting in a highly efficient approach. An interesting question to explore is whether ChatGPT4Rec can outperform the traditional TCF paradigm in the typical recommendation setting and challenge the established TCF paradigm."
},
{
"figure_ref": [],
"heading": "Background",
"publication_ref": [
"b31",
"b20",
"b32",
"b15",
"b28",
"b23",
"b52",
"b18",
"b35",
"b33",
"b34",
"b3",
"b16",
"b48",
"b48",
"b47",
"b58",
"b57",
"b49",
"b40",
"b12",
"b55",
"b41",
"b7",
"b1",
"b48",
"b57",
"b51",
"b57",
"b9",
"b25"
],
"table_ref": [],
"text": "LMs for Text. In recent years, significant advancements have been made in the development of LMs, with several landmark breakthroughs that have helped to shape the field of NLP. Word2vec, developed in 2013, revolutionized NLP by providing a scalable and efficient way of learning word embeddings from large text corpora. Since then, numerous improvements have been made to word representation models, such as GloVe [32], TextCNN [21], ELMo [33], and ULMFiT [16], etc. In 2018, the Bidirectional Encoder Representations from Transformers (BERT) model demonstrated state-of-the-art performance on a range of NLP tasks by introducing a pre-training approach based on a masked language modeling objective. BERT and its variants (RoBERTa [29], ALBERT [24], XLNet [53], TinyBERT [19], T5 [36], etc.) have become a dominant paradigm in the NLP community in recent years. More recently, ChatGPT, a conversational AI model based on the GPT-3 architecture, has gained significant attention due to its remarkable performance in various language tasks. Along this line, several other notable works have contributed to the advancement of LMs, including the Transformer architecture and the GPT series of models [34,35,4]. These advancements have not only improved the accuracy of NLP models but also opened up new avenues for research and applications in a wide range of domains outside of NLP.\nLMs for Recommender Systems. Over the past decade, language models have been widely used in item recommendation tasks [17,49], with two main lines of research in this area. The first involves using LMs to represent textual items [49,48,59,58,50], while the second involves using LMs as user encoders or recommendation backbones, such as SASRec, BERT4Rec [41], GRU4Rec [13], NextItNet [56], and Caser [42]. In this paper, we focus primarily on the first line of research. Among the various item encoders, lightweight word2vec and medium-sized BERT are the two most popular. The literature on this topic can further be classified into two categories: applying pre-extracted textual features (equivalent to free text encoder) [8,2,49] and end-to-end (E2E) training of text encoders [58,52]. While E2E training typically achieves better results than using a frozen text encoder, the latter approach is much more computationally efficient than E2E training [58].\nIn addition, with the enormous success of ChatGPT, many recent studies have started using prompt [10,26] techniques to guide ChatGPT in achieving personalized recommendations. This approach directly employs the ChatGPT API, without the need for separately training a model. However, this approach has some significant limitations. For example, when the number of candidate items reaches tens of thousands or even millions, ChatGPT may not be able to effectively recall and rank them."
},
{
"figure_ref": [],
"heading": "Preliminary",
"publication_ref": [
"b37",
"b53",
"b49"
],
"table_ref": [],
"text": "We introduce some basic notations and describe two typical recommender paradigms: IDCF & TCF.\nDefinition. We define the set of users as U = {u 1 , u 2 , ..., u m } and the set of items as V = {v 1 , v 2 , ..., v n }. The user-item interactions are represented by a binary matrix R = {r uv }, where r uv ∈ {0, 1} indicates whether user u has interacted with item v.\nIn the standard collaborative filtering (CF) setup, we represent each user by a vector θ u ∈ R k and each item by a vector β v ∈ R k . The predicted interaction score between user u and item v is computed as ruv = θ T u β v . To obtain the user and item vectors, we typically optimize a loss function l(r uv , θ T u β v ), where l can either be a pairwise BPR [38] loss or a cross-entropy classification loss [54].\nIn the popular ID-based CF (IDCF) models, θ u and β v , also known as userID and itemID embeddings, can be learned by backpropagating from the user-item interaction data. Following this path, various advanced recommender models have been developed. For instance, if we use a deep neural network to output the user vector θ u and the item vector β v , denoted by g(u i ) and h(v i ) respectively, the scoring function becomes ruv = g(u i ) • h(v i ), which is known as the two-tower DSSM model. Alternatively, Datasets. We evaluate TCF with LLM as text encoders on three real-world text datasets: the MIND news clicking recommendation dataset released by Microsoft [50], the HM clothing purchase dataset from the H&M2 platform, and the Bili3 comment dataset from an online video recommendation platform. For the MIND dataset, we represent items using their news article titles, while for the HM and Bili datasets, we utilize the corresponding titles and descriptions of products or videos to represent the items. In all three datasets, each positive user-item interaction is either a click, purchase, or comment, which served as implicit indicators of user preference.\nDue to memory issues for some E2E training experiments, we constructed interaction sequences for each user by selecting their latest 23 items. We remove users with less than 5 interactions, simply because we do not consider cold user settings. After the basic pre-processing, we randomly selected 200,000 users (and their interactions) from both MIND and HM datasets, and 50,000 users from Bili."
},
{
"figure_ref": [],
"heading": "Models and Training.",
"publication_ref": [
"b56",
"b10",
"b4",
"b26",
"b62",
"b57",
"b53",
"b59",
"b57",
"b22"
],
"table_ref": [],
"text": "To support our main arguments, we selected two representative recommendation architectures for evaluation: the two-tower DSSM model as an example of CTR (click-through rate) prediction models, and the SASRec model as an example of Transformer-style sequential models. Note we do not study other CTR prediction models, as they generally belong to the same category as DSSM, with the main difference being that many CTR models use single-tower backbone networks [57,11,5]. This distinction is not expected to significantly affect our conclusions [27,63,58].\nDuring training, we utilize the popular batch softmax loss [54], which is widely adopted in industrial systems. For text encoders, we evaluated nine different sizes of GPT models as the text encoder, ranging from 125 million to 175 billion parameters. These GPT models were re-implemented by Meta AI and are interchangeably referred to as OPT [60]. As for hyper-parameters, we first perform a grid search for standard IDCF as a reference, After determining the optimal hyper-parameters for IDCF, we search them for TCF around these optimal values. We report details in Appendix B. Evaluation. We evaluate the performance of all models using two popular top-K ranking metrics, namely HR@10 (Hit Ratio) and NDCG@10 (Normalized Discounted Cumulative Gain) [58]. NDCG@10 is reported in Appendix C for saving space. The latest user-item interaction was used for evaluation, while the second-to-last interaction was used for hyper-parameter searching, and all other interactions were used for training. All items in the pool are used for evaluation, suggested by [23]."
},
{
"figure_ref": [
"fig_2",
"fig_4"
],
"heading": "Q1: Has the TCF paradigm hit a performance ceiling?",
"publication_ref": [
"b57",
"b12",
"b40",
"b61",
"b57",
"b57",
"b57",
"b57",
"b14",
"b36",
"b44",
"b13",
"b7",
"b39",
"b7",
"b7",
"b9",
"b45",
"b25",
"b60",
"b8",
"b27",
"b5",
"b46",
"b27"
],
"table_ref": [
"tab_6",
"tab_3",
"tab_3",
"tab_5",
"tab_6",
"tab_4",
"tab_7",
"tab_7",
"tab_7"
],
"text": "To answer Q1, we conduct experiments by increasing the size of text encoders in the TCF models, ranging from 125 million (125M) to 175 billion (175B) parameters. We use SASRec & DSSM as recommender backbones. The results are given in Figure 2. All LMs are frozen in this study.\nAs shown, TCF models generally improve their performance by increasing the size of their text encoders. For instance, with the SASRec as the backbone, TCF improved the recommendation accuracy from 19.07 to 20.24 on MIND, from 9.37 to 11.11 on HM, and from 4.45 to 7.05 on Bili, resulting in improvements of 6.1%, 18.6%, and 58.4%, respectively. Similar observations can also be made for the DSSM backbone. Furthermore, based on the observed performance trend, we can conclude that the TCF models' performance has not yet converged when increasing the size of their text encoders, such as from 13B to 175B. These results suggest that (answer to Q1) the TCF model with a 175B parameter LM may not have reached its performance ceiling. In other words, if we had an even larger LM as the text encoder, TCF's performance could potentially be further improved. This is a highly desirable property because it indicates that using more powerful LMs (if developed in the future) as text encoders can result in higher recommendation accuracy.\nInterestingly, we find that the TCF model with the 350M parameter LM shows the worst results for all three datasets and with both DSSM and SASRec backbones, despite not being the smallest text encoder. This could happen because the scaling relationship between text encoder size and performance is not necessarily strictly linear. However, by examining the pre-training code and official documentation, we discovered that the 350M-parameter OPT was implemented with several differences compared to all other versions. 4 This provides an explanation for our results. Additionally, beyond the discussion scope of this paper, we also observe that TCF with the SASRec backbone largely outperforms that with the DSSM backbone. A similar finding has also been reported in much previous literature [58,13,41,62]. One possible reason for this is that representing users using their interacted items is more effective than using solely the userID feature. Another reason could be that Figure 3: TCF with retrained LM vs frozen LM (y-axis: HR@10(%)), where only the top two layers are retrained. The 175B LM is not retrained due to its ultra-high computational cost.\nthe SASRec architecture, based on the sequence-to-sequence (seq2seq) training approach, is more powerful than the DSSM architecture, which predicts the <user, item> pair. We wonder whether a language model with 175B parameters possess a degree of universality in text encoding. Unlike the traditional NLP tasks, we examine this property using text recommendation as a downstream task. Assuming that a k-dimensional text representation β v encoded by the 175B parameter LM is an ideal universal representation, any application involving text representation can directly choose a subset or the entire set of features from β v by providing a weight vector w that represents the importance of these elements, i.e., y = w T β v . For example, in a basic matrix factorization setting, w could represent user preference weight to item features, i.e. w = θ u . If all factors of user preference can be observed by the features in β v , we only need to learn their linear relationship. Moreover, for a perfect universal vector β v , using a frozen representation should be just as effective as fine-tuning it on a new dataset, or even superior to fine-tuning. From an optimization perspective, using the frozen representation requires fewer training parameters than fine-tuning, and the training process is generally easier if desired item features have been fixed in advance.\nBased on the above analysis, we only need to compare the frozen item representation with the finetuned item representation for this study. It should be noted that previous studies such as [58] have investigated this issue, but they only examined text encoders with a size of 100 million parameters.\nGiven that the frozen representation by a 175B LM is much more powerful (see Table 5), it remains unclear whether their findings hold when the encoder is scaled up by a factor of 1000.\nAs shown in Figure 3, TCF models (both SASRec and DSSM) outperform their frozen versions when the text encoders are retrained on the recommendation dataset. Surprisingly, TCF with a fine-tuned 125M LM is even more powerful than the same model with a frozen 175B LM. This result potentially suggests that (answer to Q2) even the item representation learned by an extremely large LM (e.g., GPT-3) may not result in a universal representation, at least not for the text Another key insight is that although large LMs have revolutionized so many NLP problems, there is still a significant domain gap between recommender systems and NLPnamely, inferring user preferences looks more challenging. We suspect that the text representation even extracted from the strongest and largest LM developed in the future may not perfectly adapt to the recommender system dataset. Retraining the LM on the target recommendation data appears to be necessary for optimal results. However, from a positive perspective, since large LMs have not yet reached the performance limit, if future more powerful LMs are developed, the performance of frozen text representation may become more close to fine-tuning. For instance, we observe that SASRec with a 175B LM (compared to the 125M LM) is already very close in performance to the fine-tuned 66B LM, with relative accuracy gaps of 3.92%, 16%, 13.5% on HM, and Bili, respectively. This is a promising discovery since fine-tuning such a large LM is very challenging in practical scenarios. 5Note although we did not retrain all parameters of the largest LM, we did evaluate the performance using medium-sized LMs (e.g., 1.3B & 13B) by optimizing all layers and the top two layers, which are comparable.\nIt is worth noting that the above conclusions are based on the assumption that user-item interaction feedback serves as the gold standard for the recommendation task, but this may not always be the case in practice. As a limitation, this study does not address this issue, as the entire theory of modern recommender systems is currently based on this assumption.\n7 Q3: Can IDCF be easily surpassed by TCF with a 175B parameter LM?\nTCF is a classical paradigm for text-based recommender systems, while IDCF is the dominant paradigm in the entire field of recommender systems. Can TCF models with a 175B parameter LM easily beat IDCF models with free and learnable item vectors? While many prior studies have reported that TCF models achieved state-of-the-art results, few have explicitly compared their models with corresponding IDCF counterparts under the same backbone networks and experimental settings (including samplers and loss functions). Moreover, many of them focus on cold item setting, with fewer studies explicitly examining regular (with both cold and warm items) or popular item settings.\nRecently, [58] discovered that TCF can be comparable to IDCF by jointly training a 100-million parameter LM, but frozen representations still significantly underperformed. Therefore, a natural question is whether our conclusions would differ if we use a 175B parameter LM as the item encoder?\nAs shown in Table 2, we observe that even with the 175B parameter LM and fine-tuned 66B parameter LM, TCF is still substantially inferior to IDCF when using DSSM as the backbone. These results are consistent with [58]. As explained, the DSSM architecture and training approach are not very friendly to TCF, and both IDCF and TCF with DSSM perform worse than the seq2seq-based SASRec model. In contrast, we find that TCF with the SASRec backbone performs comparably to IDCF on MIND and Bili datasets, even when the LM encoder is frozen, as shown in Table 2 and4. This is a significant advancement, as no previous study has explicitly claimed that TCF by freezing a NLP encoder can attain performance comparable to its IDCF counterparts for warm or popular item recommendation. 6This is probably because item text encoders in prior literature, such as BERT and word2vec, are inadequate in generating effective text representations comparable to IDCF, see Table 5. The reason for the weaker performance of TCF on HM is that textual information alone is insufficient to fully represent the product item, as factors such as price and quality are also critical in enticing user clicks and purchases on HM. However, in the case of news recommendation, we can generally assume that users are primarily drawn to the textual content (i.e., titles) of items, although this may not always be the case. That is the reason we believe TCF with frozen text encoders performs on par with IDCF is surprising as IDCF can implicitly learn latent factors beyond textual features but feature representation pre-extracted from a NLP encoder cannot. Furthermore, we notice that SASRec with a fine-tuned text encoder can clearly outperform IDCF on all three datasets. However, as mentioned, such end-to-end training using a text encoder is computationally expensive, despite its effectiveness.\nThe answer to Q3 is that, for text-centric recommendation, TCF with the SASRec backbone and utilizing a 175B-parameter frozen LM can achieve similar performance to standard IDCF, even for popular item recommendation. However, even by retraining a super-large LM item encoder, TCF with a DSSM7 backbone has little chance to compete with its corresponding IDCF. The simple IDCF still remains a highly competitive approach in the warm item recommendation setting. If the computation can be reduced, joint training of a powerful sequential recommender model (i.e., SASRec) with its text encoder can lead to markedly better results than IDCF.\n8 Q4: How close is the TCF paradigm to a universal recommender model?\nIn this paper, we are particularly interested in comparing with the dominant IDCF paradigm. This is because ID features (including userIDs and itemIDs) are considered as a primary obstacle to the transferable or foundation recommender models due to their non-sharability [58,15,37,45,14,8,40]. We argue that to achieve foundation models in recommender systems may require satisfying two conditions (see Figure 4): (1) abandoning userID and itemID features, and (2) achieving effective transferability across domains and platforms. Based on the above results, we conclude that for textcentric recommender systems, TCF-based sequential recommender models can basically substitute IDCF methods. However, regarding (2), it remains uncertain whether TCF has impressive transfer learning ability, especially when its item representations are extracted from a extremely large LM.\nTaking inspiration from the remarkable success of zero-shot learning in NLP, our goal is to assess the zero-shot transfer learning capability of TCF, considering that items with text features may be inherently transferable. Following [8], we first pre-train a SASRec-based TCF model with the 175B parameter frozen LM as item encoder in a large-scale text recommendation dataset 8 . We then directly evaluate the pre-trained model in the testing set of MIND, HM and QB 9 . The results, presented in Table 3, indicate that while TCF models outperform random item recommendation by achieving an accuracy improvement of 6-40x, they still fall notably short of TCF models that have been retrained on the new data. We note that user behaviors in the source Bili dataset may differ significantly from those of the e-commerce recommendation HM and new recommendation MIND datasets. However, it should be similar to that of QB as they are similar types of item recommendations.\nOur finding is consistent to that reported in [8]: (answer to Q4) while TCF models with large LMs do exhibit a certain degree of transfer learning capability, they still fall significantly short of being a universal recommender model, as we had initially envisioned. For a universal recommender system model, not only should item representations be transferable, but also the matching relationship between users and items needs to be transferable. However, the matching relationship is closely related to the exposure strategy of the specific recommender system. Therefore, compared to NLP and computer vision (CV), the transferability of recommender system models is even more challenging. This also explains why, to date, there have been no pre-trained models in recommender systems that have achieved the same level of fame and recognition as BERT and ChatGPT in the NLP field. However, this does not necessarily indicate that TCF have no potential to become a universal recommender model. It will require the collective effort of the entire recommender system community, such as utilizing highly diverse and extremely large pre-training data along with more advanced training and transfer learning techniques.\n9 Q5: ChatGPT4Rec vs TCF.\nBeyond the TCF paradigm, building text recommender models by leveraging prompt strategies is also becoming increasingly popular [10,46,26,61]. Recently, due to the tremendous success of ChatGPT, a number of preprint papers have explored the use of prompt engineering with ChatGPT for recommender systems [9,28,6,47]. Here we aim to study whether prompt-based techniques on ChatGPT, referred to as ChatGPT4Rec10 , can outperform the classical TCF paradigm.\nWe randomly selected 1024 users from the testing sets of MIND, HM, and Bili, and created two tasks for ChatGPT. In the first task (Task 1 in Table 6), ChatGPT was asked to select the most preferred item from four candidates (one ground truth and three randomly selected items), given the user's historical interactions as a condition. The second task (Task 2 in Table 6) was to ask ChatGPT to rank the top-10 preferred items from 100 candidates (one ground truth and 99 randomly selected items, excluding all historical interactions), also provided with the user's historical interactions as input. We begin by asking ChatGPT if it understands the request, in order to ensure the quality of the prompts. Both the prompts and its answer are included in Appendix D. The results are given in Table 6, which illustrate ChatGPT's poor performance compared to TCF in typical recommendation settings. Similar bad results have also been reported in [28]. Despite that, we believe with more finely-tuned prompts, ChatGPT may have the potential for certain recommendation scenarios. Another major drawback of ChatGPT is that it cannot generate recommendations from an item pool with millions of items due to limited memory. Thus, the answer to Q5 is that based on its current performance and limitations, ChatGPT is unable to substitute the classical TCF paradigm."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "This paper does not describe a new text recommender algorithm. Instead, it extensively explores the performance limits and several core issues of the prevailing text-based collaborative filtering (TCF) techniques. From a positive perspective, TCF has not yet reached its performance ceiling. With the further advancement of the representation capacity of NLP large models, TCF is expected to achieve better performance. However, it is regrettable that even with an item encoder of tens of billions of parameters, it still needs to be re-adapted to new data for optimal recommendations. Plus, the current cutting-edge TCF models do not exhibit strong transferability that was anticipated, indicating that building large foundation recommender models may be an even more daunting task than in NLP and CV fields. Nonetheless, TCF with text encoders of 175 billion parameters is already a significant leap forward, as it fundamentally challenges the dominant ID-based CF paradigm, which is considered the biggest obstacle to developing foundation recommender models, although not the only one."
},
{
"figure_ref": [],
"heading": "B Hyper-parameter tuning",
"publication_ref": [
"b29"
],
"table_ref": [],
"text": "Before tuning hyper-parameters for TCF, we grid search IDCF on each dataset as a reference. Specifically, we search for learning rates within the range of {1e-3, 1e-4, 1e-5, 5e-5} and hidden dimensions from {64, 128, 256, 512, 1024} for both DSSM and SASRec; we search batch size within {64, 128, 256, 512} for SASRec and {1024, 2048, 4096} for DSSM; we set a fixed dropout rate of 0.1, and tune the weight decay within {0.01, 0.1}; we search the number of Transformer layers in SASRec within {1, 2, 3, 4}, and the number of attention heads within {2, 4, 8}. After determining the optimal hyper-parameters for IDCF, we search the TCF around these optimal values with the frozen text encoder (using the 125M variant) by the same stride. To ensure a fair comparison of the scaling effect, we employ the same hyper-parameters for all TCF models with different sizes of frozen text encoder (i.e., pre-extracted features). For TCF models with E2E learning of text encoders, we kept the optimal hyper-parameters the same as those with frozen encoder, except for the learning rates.\nWe separately tune the learning rate, as larger text encoders typically require a smaller learning rate. The details are given below. We utilize the AdamW optimizer [30] for all models. C More results on NDCG@10 6: TCF with retrained LM vs frozen LM (y-axis: NDCG@10(%)), where only the top two layers are retrained. The 175B LM is not retrained due to its ultra-high computational cost."
},
{
"figure_ref": [
"fig_5",
"fig_6"
],
"heading": "D The prompts of ChatGPT4Rec",
"publication_ref": [],
"table_ref": [],
"text": "The output by ChatGPT in Figure 7 indicates that ChatGPT fully understands the recommendation request. Figure 8, 9 and 10 are prompts for ChatGPT on MIND, HM, and Bili respectively.\nDo you understand our request, is it clear? \"Please return a list of the top 10 most relevant video titles from the unordered candidate pool as a personalized recommendation, based on the user's viewed history, formatted as a JSON array. Do not include any titles from the user's viewed history in the recommendations. Please note that the candidate pool is not in any particular order, any ranking based on the order of items in the pool should be assumed.\nUser viewed history: 1.MC Lucky Gun PVP! When you fire a gun in \"Desert Ash\" and duel with lucky cubes! 2.Minecraft Kep1er WADADA Dance Relay Open ... 12.Central Conservatory of Music professors perform famous songs! The roof beautifully."
},
{
"figure_ref": [],
"heading": "Unordered candidate news title pool (no order should be assumed):",
"publication_ref": [],
"table_ref": [
"tab_7"
],
"text": "-Times Youth League Ma Jiaqi -Teach you to make homemade strawberry ice cream,with delicate taste and no ice residue, ... -Pea Shooter: There's something wrong with the zombies here today! Top 10 news titles from candidates pool to recommend to the user as JSON array (from highest to lowest priority, titles only)：\".\nYes, I understand your request. You would like to receive a list of the top 10 most relevant video titles as personalized recommendations for a user. The recommendations should be based on the user's viewed history, and the titles should not include any videos that the user has already viewed. The candidate pool is an unordered list, and any ranking based on the order of items in the pool should be assumed.To clarify, you have provided a sample user viewed history and an unordered candidate video pool. You are requesting the top 10 video titles from the candidate pool to be recommended to the user. The recommendations should be returned as a JSON array, sorted from highest to lowest priority based on relevance. Only the titles should be included in the JSON array.Is that correct? [ \"He Never Smoked a Day in His Life. At 28, He Was Diagnosed With Lung Cancer.\", \"Vladimir Tarasenko to undergo surgery, out for at least 5 months\", \"Another anti-Muslim group wants to hold event at Trump's Palm Beach resort, Mar-a-Lago\", \"Unique friendship continues to grow 1 year after Camp Fire\", \"Deadline for public input on proposed HUD rule nears\", \"Former President Jimmy Carter recovering after brain surgery\", \"Twitter reacts to the return of Markelle Fultz\", \"Why aren't more women getting mammograms?\", \"Social Security: 5 smart ways to get more benefits\", \"For men with breast cancer, early detection is critical\" ] (b) Example of Task 2 in Table 6) for MIND "
},
{
"figure_ref": [],
"heading": " ",
"publication_ref": [],
"table_ref": [],
"text": "6\n) for HM Please return a list of the top 10 most relevant product titles from the unordered candidate pool as a personalized recommendation, based on the user's purchase history, formatted as a JSON array. Do not include any titles from the user's purchase history in the recommendations. Please note that the candidate pool is not in any particular order, any ranking based on the order of items in the pool should not be assumed! 6) for Bili Please return a list of the top 10 most relevant video titles from the unordered candidate pool as a personalized recommendation, based on the user's viewed history, formatted as a JSON array. Do not include any titles from the user's viewed history in the recommendations. Please note that the candidate pool is not in any particular order, any ranking based on the order of items in the pool should not be assumed! User viewed history: 1. MC Lucky Gun PVP! When you fire a gun in \"Desert Ash\" and duel with lucky cubes! minecraft 2. Kep1er WADADA Dance Relay Open "
}
] | Text-based collaborative filtering (TCF) has become the mainstream approach for text and news recommendation, utilizing text encoders, also known as language models (LMs), to represent items. However, existing TCF models primarily focus on using small or medium-sized LMs. It remains uncertain what impact replacing the item encoder with one of the largest and most powerful LMs, such as the 175billion parameter , would have on recommendation performance. Can we expect unprecedented results? To this end, we conduct an extensive series of experiments aimed at exploring the performance limits of the TCF paradigm. Specifically, we increase the size of item encoders from one hundred million to one hundred billion to reveal the scaling limits of the TCF paradigm. We then examine whether these extremely large LMs could enable a universal item representation for the recommendation task. Furthermore, we compare the performance of the TCF paradigm utilizing the most powerful LMs to the currently dominant ID embeddingbased paradigm and investigate the transferability of this TCF paradigm. Finally, we compare TCF with the recently popularized prompt-based recommendation using ChatGPT 1 . Our research findings have not only yielded positive results but also uncovered some surprising and previously unknown negative outcomes, which can inspire deeper reflection and innovative thinking regarding text-based recommender systems. Codes & datasets will be released for further research. | Exploring the Upper Limits of Text-Based Collaborative Filtering Using Large Language Models: Discoveries and Insights | [
{
"figure_caption": "Figure 1 :1Figure 1: TCF with SASRec and DSSM as recommender backbones. The DTL block is the dense dimension transformation layers. Item or text encoder used in this study can be 175B parameters.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: TCF's performance (y-axis: HR@10(%)) with 9 text encoders of increasing size (x-axis). SASRec (upper three subfigures) and DSSM (bottom three subfigures) are used as the backbone.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: Route to foundation recommender models (FRM). The cross indicates that the IDCF paradigm have no chance to achieve FRM, the tick indicates that for text-centric RS, TCF can basically replace IDCF, and the question mark indicates that whether the TCF paradigm can achieve the widely recognized FRM remains still unknown.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Figure 7 :7Figure 7: Verifying that ChatGPT understands the request.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "7",
"figure_type": "figure"
},
{
"figure_caption": "Figure 8 :8Figure 8: Prompt for MIND",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "8",
"figure_type": "figure"
},
{
"figure_caption": "Dataset characteristics",
"figure_data": "Dataset#User#Item #Interaction Item ExampleMIND 200,000 54,2462,920,730 Eagles fans rooting guide for Week 7. (News Title)HM200,000 85,0193,160,543 Solid. White. Ladieswear. (Product Description)Bili50,000 22,377723,071 Spoofs: Japanese guys fight kacoko. (Video Title)if we represent a user by a sequence of k items that she has interacted with, the scoring function isruv = G(v 1 , v 2 , ..., v k ) T β v , where G(•) is a sequential network, such as SASRec & BERT4Rec.By utilizing a text encoder f (v i ) to output item representation vectors from the description text,instead of relying on itemID embedding features, the IDCF model can be converted into the TCFmodel, as depicted in Figure 1. Clearly, the only difference between TCF and the typical IDCFmodel is in the item representation part. In contrast to IDCF, TCF has the advantage of being able toutilize both item textual content features and user-item interaction feedback data. In theory, the textencoder f (v i ) can take the form of any language model, such as a shallow-layer word2vec model,a medium-sized BERT model, or a super-large GPT-3 model. The text encoder f (v i ) can be eitherfrozen or trained with the whole recommender model in an end-to-end (E2E) fashion.However, in practice, due to computational costs, most real-world recommender systems adopt atwo-stage approach where offline features are extracted beforehand (i.e., a frozen item encoder) andthen incorporated into the recommender model for training. This is because joint or E2E training oftext encoders usually requires significant computing power and training time.",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Accuracy comparison (HR@10) of IDCF and TCF using the DSSM & SASRec backbones. FR is TCF using frozen LM, while FT is TCF using fine-tuned LM.",
"figure_data": "DataSASRecDSSMIDCF 175B FR 66B FT IDCF 175B FR 66B FTMIND 20.05 20.24 21.07 3.99 2.83 3.27HM12.02 11.24 13.29 6.79 2.09 2.35Bili7.01 6.88 8.15 2.27 2.00 2.01",
"figure_id": "tab_3",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Zero-shot recommendation accuracy (HR@10). 175B zero means zeroshot accuracy of TCF with 175B LM. 'train' is to retrain TCF on these data.",
"figure_data": "ModelMIND HMQBRandom 0.020.010.18175B zero 0.130.394.30175B train 20.2411.11 29.90",
"figure_id": "tab_4",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Warm item recommendation (HR@10). 20 means items < 20 interactions are removed. TCF 175B uses the pre-extracted features from the 175B LM. Only the SASRec backbone is reported.",
"figure_data": "DataMINDHMBili#Interaction205020020502002050200IDCF20.56 20.87 23.04 13.02 14.38 18.07 7.899.03 15.58TCF 175B20.59 21.20 22.85 12.03 12.68 16.06 7.768.96 15.47recommendation task.",
"figure_id": "tab_5",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "TCF's results (HR@10) with representative text encoders in the last 10 years. Text encoders are frozen and the SASRec backbone is used. Advances in NLP benefit RS.",
"figure_data": "ModelDate MIND HMBiliword2vec 2013 15.21 8.082.66BERT large 2018 18.99 9.683.56T5 XXL2019 19.56 9.214.81OPT 175B 2022 20.24 11.11 7.05",
"figure_id": "tab_6",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "ChatGPT4Rec vs TCF. FR & FT means freezing and fine-tuning LM respectively.",
"figure_data": "DataTask 1-HR@1Task 2-HR@10Random ChatGPT TCF 175BFR TCF 66BFTRandom ChatGPT TCF 175BFR TCF 66BFTMIND 25.0025.6896.4896.5810.009.8697.0797.9HM25.0029.5988.1890.6310.0012.2183.7990.33Bili25.0024.5177.6481.0510.008.5070.8073.34",
"figure_id": "tab_7",
"figure_label": "6",
"figure_type": "table"
},
{
"figure_caption": "Optimal hyper-parameters for IDCF, including learning rate (lr), embedding size (k), batch size (bs), the number of Transformer layers (l), the number of attention heads (h), and weight decay (wd). The dimension of feed forward layer in Transformer block is 4 × k.",
"figure_data": "DataSASRecDSSMlrk bs l h wdlrkbsl h wdMIND 1e-4 512 64 2 2 0.1 1e-5 256 4096 2 2 0.1HM1e-3 128 128 2 2 0.1 1e-4 1024 1024 2 2 0.1Bili1e-3 128 256 2 2 0.1 1e-3 1024 1024 2 2 0.1",
"figure_id": "tab_8",
"figure_label": "8",
"figure_type": "table"
},
{
"figure_caption": "Optimal hyper-parameters for TCF with frozen text encoder.",
"figure_data": "DataSASRecDSSMlrk bs l h wdlrkbsl h wdMIND 1e-4 512 64 2 2 0.1 1e-5 256 4096 2 2 0.1HM1e-4 512 64 2 2 0.1 1e-3 1024 1024 2 2 0.1Bili1e-3 128 64 2 2 0.1 1e-3 512 1024 2 2 0.1",
"figure_id": "tab_9",
"figure_label": "9",
"figure_type": "table"
},
{
"figure_caption": "The learning rate of item encoder for TCF with E2E learning. The search range is suggested by the original paper of OPT.",
"figure_data": "DataSASRecDSSM125M1.3B13B66B125M1.3B13B66BMIND1e-41e-48e-53e-51e-41e-41e-41e-4HM1e-41e-41e-48e-51e-41e-41e-41e-4Bili1e-41e-43e-53e-51e-41e-41e-41e-4",
"figure_id": "tab_10",
"figure_label": "10",
"figure_type": "table"
},
{
"figure_caption": "Warm item recommendation (NDCG@10). 20 means items < 20 interactions are removed. TCF 175B uses the pre-extracted features from the 175B LM. Only SASRec backbone is reported.",
"figure_data": "DataMINDHMBili#Inter.205020020502002050200IDCF11.36 11.47 12.71 8.479.35 12.07 4.415.018.30TCF 175B11.38 11.61 12.56 7.447.90 10.33 4.344.847.97",
"figure_id": "tab_11",
"figure_label": "11",
"figure_type": "table"
},
{
"figure_caption": "Accuracy (NDCG@10) comparison of IDCF and TCF using DSSM and SASRec. FR represents using frozen LM, while FT represents using fine-tuned LM.",
"figure_data": "DataMetricSASRecDSSMIDTCF 175BFRTCF 66BFTIDTCF 175BFRTCF 66BFTMIND NDCG@10 11.0611.0911.771.721.421.58HMNDCG@10 7.766.918.204.191.081.22BiliNDCG@10 3.933.774.561.121.011.06",
"figure_id": "tab_12",
"figure_label": "12",
"figure_type": "table"
},
{
"figure_caption": "Zero-shot recommendation accuracy (NDCG@10). 175B zero means zero-shot accuracy of TCF with 175B LM. 'train' is to retrain TCF on these data.",
"figure_data": "ModelDate MINDHMBiliWord2vec 20137.524.811.30BERT large201810.456.011.83T5 XXL201910.725.502.54OPT 175B202211.176.883.95",
"figure_id": "tab_13",
"figure_label": "13",
"figure_type": "table"
}
] | Ruyu Li; Wenhao Deng; Yu Cheng; Zheng Yuan; Jiaqi Zhang; Fajie Yuan | [
{
"authors": "Rachith Aiyappa; Jisun An; Haewoon Kwak; Yong-Yeol Ahn",
"journal": "",
"ref_id": "b0",
"title": "Can we trust the evaluation on chatgpt?",
"year": "2023"
},
{
"authors": "Qiwei Bi; Jian Li; Lifeng Shang; Xin Jiang; Qun Liu; Hanfang Yang",
"journal": "",
"ref_id": "b1",
"title": "Mtrec: Multi-task learning over bert for news recommendation",
"year": "2022"
},
{
"authors": "Rishi Bommasani; Drew A Hudson; Ehsan Adeli; Russ Altman; Simran Arora; Sydney Von Arx; Jeannette Michael S Bernstein; Antoine Bohg; Emma Bosselut; Brunskill",
"journal": "",
"ref_id": "b2",
"title": "On the opportunities and risks of foundation models",
"year": "2021"
},
{
"authors": "Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell",
"journal": "Advances in neural information processing systems",
"ref_id": "b3",
"title": "Language models are few-shot learners",
"year": "2020"
},
{
"authors": "Paul Covington; Jay Adams; Emre Sargin",
"journal": "",
"ref_id": "b4",
"title": "Deep neural networks for youtube recommendations",
"year": "2016"
},
{
"authors": "Sunhao Dai; Ninglu Shao; Haiyuan Zhao; Weijie Yu; Zihua Si; Chen Xu; Zhongxiang Sun; Xiao Zhang; Jun Xu",
"journal": "",
"ref_id": "b5",
"title": "Uncovering chatgpt's capabilities in recommender systems",
"year": "2023"
},
{
"authors": "Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova",
"journal": "",
"ref_id": "b6",
"title": "Bert: Pre-training of deep bidirectional transformers for language understanding",
"year": "2018"
},
{
"authors": "Yifei Hao Ding; Anoop Ma; Yuyang Deoras; Hao Wang; Wang",
"journal": "",
"ref_id": "b7",
"title": "Zero-shot recommender systems",
"year": "2021"
},
{
"authors": "Yunfan Gao; Tao Sheng; Youlin Xiang; Yun Xiong; Haofen Wang; Jiawei Zhang",
"journal": "",
"ref_id": "b8",
"title": "Chat-rec: Towards interactive and explainable llms-augmented recommender system",
"year": "2023"
},
{
"authors": "Shijie Geng; Shuchang Liu; Zuohui Fu; Yingqiang Ge; Yongfeng Zhang",
"journal": "",
"ref_id": "b9",
"title": "Recommendation as language processing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5)",
"year": "2022"
},
{
"authors": "Huifeng Guo; Ruiming Tang; Yunming Ye; Zhenguo Li; Xiuqiang He",
"journal": "",
"ref_id": "b10",
"title": "Deepfm: a factorization-machine based neural network for ctr prediction",
"year": "2017"
},
{
"authors": "Xiangnan He; Lizi Liao; Hanwang Zhang; Liqiang Nie; Xia Hu; Tat-Seng Chua",
"journal": "",
"ref_id": "b11",
"title": "Neural collaborative filtering",
"year": "2017"
},
{
"authors": "Balázs Hidasi; Alexandros Karatzoglou; Linas Baltrunas; Domonkos Tikk",
"journal": "",
"ref_id": "b12",
"title": "Session-based recommendations with recurrent neural networks",
"year": "2015"
},
{
"authors": "Yupeng Hou; Zhankui He; Julian Mcauley; Wayne Xin Zhao",
"journal": "",
"ref_id": "b13",
"title": "Learning vector-quantized item representation for transferable sequential recommenders",
"year": "2023"
},
{
"authors": "Yupeng Hou; Shanlei Mu; Wayne Xin Zhao; Yaliang Li; Bolin Ding; Ji-Rong Wen",
"journal": "",
"ref_id": "b14",
"title": "Towards universal sequence representation learning for recommender systems",
"year": "2022"
},
{
"authors": "Jeremy Howard; Sebastian Ruder",
"journal": "",
"ref_id": "b15",
"title": "Universal language model fine-tuning for text classification",
"year": "2018"
},
{
"authors": "Wenyue Hua; Shuyuan Xu; Yingqiang Ge; Yongfeng Zhang",
"journal": "",
"ref_id": "b16",
"title": "How to index item ids for recommendation foundation models",
"year": "2023"
},
{
"authors": "Po-Sen Huang; Xiaodong He; Jianfeng Gao; Li Deng; Alex Acero; Larry Heck",
"journal": "",
"ref_id": "b17",
"title": "Learning deep structured semantic models for web search using clickthrough data",
"year": "2013"
},
{
"authors": "Xiaoqi Jiao; Yichun Yin; Lifeng Shang; Xin Jiang; Xiao Chen; Linlin Li; Fang Wang; Qun Liu",
"journal": "",
"ref_id": "b18",
"title": "Tinybert: Distilling bert for natural language understanding",
"year": "2019"
},
{
"authors": "Wang-Cheng Kang; Julian Mcauley",
"journal": "IEEE",
"ref_id": "b19",
"title": "Self-attentive sequential recommendation",
"year": "2018"
},
{
"authors": "Yoon Kim",
"journal": "",
"ref_id": "b20",
"title": "Convolutional neural networks for sentence classification",
"year": "2015"
},
{
"authors": "Yehuda Koren; Robert Bell; Chris Volinsky",
"journal": "Computer",
"ref_id": "b21",
"title": "Matrix factorization techniques for recommender systems",
"year": "2009"
},
{
"authors": "Walid Krichene; Steffen Rendle",
"journal": "",
"ref_id": "b22",
"title": "On sampled metrics for item recommendation",
"year": "2020"
},
{
"authors": "Zhenzhong Lan; Mingda Chen; Sebastian Goodman; Kevin Gimpel; Piyush Sharma; Radu Soricut",
"journal": "",
"ref_id": "b23",
"title": "Albert: A lite bert for self-supervised learning of language representations",
"year": "2019"
},
{
"authors": "Jian Li; Jieming Zhu; Qiwei Bi; Guohao Cai; Lifeng Shang; Zhenhua Dong; Xin Jiang; Qun Liu",
"journal": "",
"ref_id": "b24",
"title": "Miner: Multi-interest matching network for news recommendation",
"year": "2022"
},
{
"authors": "Lei Li; Yongfeng Zhang; Li Chen",
"journal": "ACM Transactions on Information Systems",
"ref_id": "b25",
"title": "Personalized prompt learning for explainable recommendation",
"year": "2023"
},
{
"authors": "Xiangyang Li; Bo Chen; Huifeng Guo; Jingjie Li; Chenxu Zhu; Xiang Long; Sujian Li; Yichao Wang; Wei Guo; Longxia Mao",
"journal": "",
"ref_id": "b26",
"title": "Inttower: the next generation of two-tower model for pre-ranking system",
"year": "2022"
},
{
"authors": "Junling Liu; Chao Liu; Renjie Lv; Kang Zhou; Yan Zhang",
"journal": "",
"ref_id": "b27",
"title": "Is chatgpt a good recommender? a preliminary study",
"year": "2023"
},
{
"authors": "Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov",
"journal": "",
"ref_id": "b28",
"title": "Roberta: A robustly optimized bert pretraining approach",
"year": "2019"
},
{
"authors": "Ilya Loshchilov; Frank Hutter",
"journal": "",
"ref_id": "b29",
"title": "Decoupled weight decay regularization",
"year": "2017"
},
{
"authors": "Tomas Mikolov; Ilya Sutskever; Kai Chen; Greg S Corrado; Jeff Dean",
"journal": "Advances in neural information processing systems",
"ref_id": "b30",
"title": "Distributed representations of words and phrases and their compositionality",
"year": "2013"
},
{
"authors": "Jeffrey Pennington; Richard Socher; Christopher D Manning",
"journal": "",
"ref_id": "b31",
"title": "Glove: Global vectors for word representation",
"year": "2014"
},
{
"authors": "Mark Matthew E Peters; Mohit Neumann; Matt Iyyer; Christopher Gardner; Kenton Clark; Luke Lee; Zettlemoyer",
"journal": "",
"ref_id": "b32",
"title": "Deep contextualized word representations",
"year": "2018"
},
{
"authors": "Alec Radford; Karthik Narasimhan; Tim Salimans; Ilya Sutskever",
"journal": "",
"ref_id": "b33",
"title": "Improving language understanding by generative pre-training",
"year": "2018"
},
{
"authors": "Alec Radford; Jeffrey Wu; Rewon Child; David Luan; Dario Amodei; Ilya Sutskever",
"journal": "OpenAI blog",
"ref_id": "b34",
"title": "Language models are unsupervised multitask learners",
"year": "2019"
},
{
"authors": "Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu",
"journal": "The Journal of Machine Learning Research",
"ref_id": "b35",
"title": "Exploring the limits of transfer learning with a unified text-to-text transformer",
"year": "2020"
},
{
"authors": "Shashank Rajput; Nikhil Mehta; Anima Singh; Trung Raghunandan H Keshavan; Lukasz Vu; Lichan Heldt; Yi Hong; Tay; Jonah Vinh Q Tran; Samost",
"journal": "",
"ref_id": "b36",
"title": "Recommender systems with generative retrieval",
"year": "2023"
},
{
"authors": "Steffen Rendle; Christoph Freudenthaler; Zeno Gantner; Lars Schmidt-Thieme",
"journal": "",
"ref_id": "b37",
"title": "Bpr: Bayesian personalized ranking from implicit feedback",
"year": "2012"
},
{
"authors": "Steffen Rendle; Christoph Freudenthaler; Lars Schmidt-Thieme",
"journal": "",
"ref_id": "b38",
"title": "Factorizing personalized markov chains for next-basket recommendation",
"year": "2010"
},
{
"authors": "Kyuyong Shin; Hanock Kwak; Kyung-Min Kim; Minkyu Kim; Young-Jin Park; Jisu Jeong; Seungjae Jung",
"journal": "",
"ref_id": "b39",
"title": "One4all user representation for recommender systems in e-commerce",
"year": "2021"
},
{
"authors": "Fei Sun; Jun Liu; Jian Wu; Changhua Pei; Xiao Lin; Wenwu Ou; Peng Jiang",
"journal": "",
"ref_id": "b40",
"title": "Bert4rec: Sequential recommendation with bidirectional encoder representations from transformer",
"year": "2019"
},
{
"authors": "Jiaxi Tang; Ke Wang",
"journal": "",
"ref_id": "b41",
"title": "Personalized top-n sequential recommendation via convolutional sequence embedding",
"year": "2018"
},
{
"authors": "Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Lukasz Kaiser; Illia Polosukhin",
"journal": "",
"ref_id": "b42",
"title": "Attention is all you need",
"year": "2017"
},
{
"authors": "Congcong Wang; Paul Nulty; David Lillis",
"journal": "",
"ref_id": "b43",
"title": "A comparative study on word embeddings in deep learning for text classification",
"year": "2020"
},
{
"authors": "Jie Wang; Fajie Yuan; Mingyue Cheng; M Joemon; Chenyun Jose; Beibei Yu; Zhijin Kong; Bo Wang; Zang Hu; Li",
"journal": "",
"ref_id": "b44",
"title": "Transrec: Learning transferable recommendation from mixture-of-modality feedback",
"year": "2022"
},
{
"authors": "Lei Wang; Ee-Peng Lim",
"journal": "",
"ref_id": "b45",
"title": "Zero-shot next-item recommendation using large pretrained language models",
"year": "2023"
},
{
"authors": "Wenjie Wang; Xinyu Lin; Fuli Feng; Xiangnan He; Tat-Seng Chua",
"journal": "",
"ref_id": "b46",
"title": "Generative recommendation: Towards next-generation recommender paradigm",
"year": "2023"
},
{
"authors": "Chuhan Wu; Fangzhao Wu; Suyu Ge; Tao Qi; Yongfeng Huang; Xing Xie",
"journal": "",
"ref_id": "b47",
"title": "Neural news recommendation with multi-head self-attention",
"year": "2019"
},
{
"authors": "Chuhan Wu; Fangzhao Wu; Tao Qi; Yongfeng Huang",
"journal": "",
"ref_id": "b48",
"title": "Empowering news recommendation with pre-trained language models",
"year": "2021"
},
{
"authors": "Fangzhao Wu; Ying Qiao; Jiun-Hung Chen; Chuhan Wu; Tao Qi; Jianxun Lian; Danyang Liu; Xing Xie; Jianfeng Gao; Winnie Wu; Ming Zhou",
"journal": "",
"ref_id": "b49",
"title": "Mind: A large-scale dataset for news recommendation",
"year": ""
},
{
"authors": "Shitao Xiao; Zheng Liu; Yingxia Shao; Tao Di; Bhuvan Middha; Fangzhao Wu; Xing Xie",
"journal": "",
"ref_id": "b50",
"title": "Training large-scale news recommenders with pretrained language models in the loop",
"year": "2022"
},
{
"authors": "Yoonseok Yang; Seok Kyu; Minsam Kim; Juneyoung Kim; Park",
"journal": "",
"ref_id": "b51",
"title": "Gram: Fast finetuning of pre-trained language models for content-based collaborative filtering",
"year": "2022"
},
{
"authors": "Zhilin Yang; Zihang Dai; Yiming Yang; Jaime Carbonell; Ruslan Salakhutdinov; Quoc V Le",
"journal": "",
"ref_id": "b52",
"title": "Xlnet: Generalized autoregressive pretraining for language understanding",
"year": "2019"
},
{
"authors": "Xinyang Yi; Ji Yang; Lichan Hong; Derek Zhiyuan Cheng; Lukasz Heldt; Aditee Kumthekar; Zhe Zhao; Li Wei; Ed Chi",
"journal": "",
"ref_id": "b53",
"title": "Sampling-bias-corrected neural modeling for large corpus item recommendations",
"year": "2019"
},
{
"authors": "Fajie Yuan; Xiangnan He; Haochuan Jiang; Guibing Guo; Jian Xiong; Zhezhao Xu; Yilin Xiong",
"journal": "",
"ref_id": "b54",
"title": "Future data helps training: Modeling future contexts for session-based recommendation",
"year": "2020"
},
{
"authors": "Fajie Yuan; Alexandros Karatzoglou; Ioannis Arapakis; Joemon M Jose; Xiangnan He",
"journal": "",
"ref_id": "b55",
"title": "A simple convolutional generative network for next item recommendation",
"year": "2019"
},
{
"authors": "Guanghu Yuan; Fajie Yuan; Yudong Li; Beibei Kong; Shujie Li; Lei Chen; Min Yang; Chenyun Yu; Bo Hu; Zang Li",
"journal": "",
"ref_id": "b56",
"title": "Tenrec: A large-scale multipurpose benchmark dataset for recommender systems",
"year": "2022"
},
{
"authors": "Zheng Yuan; Fajie Yuan; Yu Song; Youhua Li; Junchen Fu; Fei Yang; Yunzhu Pan; Yongxin Ni",
"journal": "",
"ref_id": "b57",
"title": "Where to go next for recommender systems? id-vs. modality-based recommender models revisited",
"year": "2023"
},
{
"authors": "Qi Zhang; Jingjie Li; Qinglin Jia; Chuyuan Wang; Jieming Zhu; Zhaowei Wang; Xiuqiang He",
"journal": "",
"ref_id": "b58",
"title": "Unbert: User-news matching bert for news recommendation",
"year": "2021"
},
{
"authors": "Susan Zhang; Stephen Roller; Naman Goyal; Mikel Artetxe; Moya Chen; Shuohui Chen; Christopher Dewan; Mona Diab; Xian Li; Xi Victoria Lin",
"journal": "",
"ref_id": "b59",
"title": "Opt: Open pre-trained transformer language models",
"year": "2022"
},
{
"authors": "Yuhui Zhang; Hao Ding; Zeren Shui; Yifei Ma; James Zou; Anoop Deoras; Hao Wang",
"journal": "",
"ref_id": "b60",
"title": "Language models as recommender systems: Evaluations and limitations",
"year": "2021"
},
{
"authors": "Kun Zhou; Hui Wang; Wayne Xin Zhao; Yutao Zhu; Sirui Wang; Fuzheng Zhang; Zhongyuan Wang; Ji-Rong Wen",
"journal": "",
"ref_id": "b61",
"title": "S3-rec: Self-supervised learning for sequential recommendation with mutual information maximization",
"year": "2020"
},
{
"authors": "Jieming Zhu; Quanyu Dai; Liangcai Su; Rong Ma; Jinyang Liu; Guohao Cai; Xi Xiao; Rui Zhang",
"journal": "",
"ref_id": "b62",
"title": "Bars: Towards open benchmarking for recommender systems",
"year": "2022"
},
{
"authors": "",
"journal": "",
"ref_id": "b63",
"title": "A Text Encoder details Table 7: List of Large LMs and their details Name Model Size Parameters Architecture Source BERT Large 340M Encoder-only",
"year": ""
}
] | [] |
|
2023-05-19 | [
{
"figure_ref": [],
"heading": "INTRODUCTION",
"publication_ref": [
"b0",
"b1",
"b2",
"b3",
"b4",
"b6",
"b7",
"b8",
"b9",
"b10",
"b8",
"b11",
"b15",
"b20",
"b21",
"b22",
"b23",
"b12",
"b11",
"b14",
"b15",
"b16",
"b17",
"b18",
"b19",
"b13",
"b24",
"b10",
"b12",
"b25",
"b30",
"b10",
"b26",
"b27",
"b31",
"b32",
"b11",
"b12",
"b28",
"b29",
"b29",
"b33",
"b11",
"b34",
"b30",
"b35",
"b36",
"b12",
"b31",
"b37"
],
"table_ref": [],
"text": "R IGID point cloud registration is a core and fundamental problem in the field of 3D vision and robotics with a wide range of applications, such as autonomous driving [1], [2], 3D reconstruction [3], [4], and simultaneous localization and mapping (SLAM) [5]- [7]. Given the source and target point clouds in different coordinate systems, it aims to estimate the 6 degrees of freedom (DOF) transformation in SE(3) to align the two point clouds best. The 6-DOF transformation includes both 3-DOF rotation in SO(3) and 3-DOF translation in R 3 .\nDespite decades of research, rigid point could registration is still an active and challenging problem since it has chicken-and-egg property [8]. Specifically, the registration problem comprises two mutually interlocked sub-problems: pose and correspondence estimations. If one sub-problem is solved, another sub-problem will be solved accordingly. Commonly, existing registration methods are classified into two categories based on the requirement of correspondences or not, which are correspondence-based registration (e.g., fast global registration, FGR [9]) and simultaneous pose and correspondence registration (SPCR) (e.g., iterative closest point, ICP [10]). The widely used ICP is a local optimization method for SPCR, which means that it is highly dependent on the initialization of transformation and thus prone to fall into local minima, as shown in Fig. 12). The global methods for SPCR, however, deliver relatively low efficiency (e.g., globally optimal ICP, Go-ICP [11]), as shown in Fig. 123). Thus the correspondence-based registration approaches are gradually attracting attention since they are initializationfree and more efficient [9], [12]. In this paper, we focus on the correspondence-based registration problem, while, interestingly, our approach can also be extended to solve the challenging SPCR problem.\nCurrent 3D feature matching approaches have made satisfactory development. However, outlier correspondences are still inevitable either for handcrafted or learning-based descriptors [16], [21], [22]. Several paradigms have been extensively developed to implement robust registration, of which the consensus maximization is inherently robust to outliers without smoothing or trimming to change the objective function [23], [24]. Random sample consensus (RANSAC) is the most popular heuristic method for solving the consensus maximization problem of correspondence-Correspondence-based (FPFH descriptor) (a-1) Initial (94.20% outlier) (a-2) GORE [13] (a-3) TEASER [12] (a-4) Ours -E R = 0.041 Fig. 1: The proposed method can efficiently address the rigid registration problem in different scenarios with high outlier rates or low overlap rates. For the correspondence-based registration problem, the input correspondences are generated by the traditional descriptor FPFH [15] and the learning-based descriptor FCGF [16]. The input point clouds are selected from (a) Bremen dataset [17], (b) ETH dataset [18], (c) KITTI dataset [19], and (d) Bunny dataset [20], respectively. The source point cloud is green, the target point cloud is yellow, and the aligned point cloud is blue. Compared with state-ofthe-art correspondence-based methods, the proposed method achieves significant performance in terms of robustness and efficiency. Besides, the proposed method also can solve the SPCR problem efficiently and robustly.\n• E R = 0.069 • E R = 0.004 • - E t = 0.055m E t =\nbased registration. However, RANSAC and its variants are non-deterministic and only generate satisfactory solutions with a certain probability due to the random sampling mechanism [14], [25].\nMore recently, many global and deterministic methods based on the branch and bound (BnB) framework have been applied to solve the point cloud registration problem with optimality guarantees [11], [13], [26]- [31]. However, the computational complexity of BnB optimization is exponential to the dimension of the solution domain. Most studies address the issue by jointly searching for the optimal solution in SE(3) [11], [27], [28]. In order to improve the algorithm efficiency, one direction is utilizing the known gravity directions measured by inertial measurement units (IMUs) to reduce the dimension of the parameter space to 4-dimensional [32], [33]. Another direction for reducing the problem dimension is to decompose the original problem into two 3-DOF sub-problems by leveraging the geometric properties [12], [13], [29], [30]. Typically, two unique features are employed for pose decoupling, i.e., the rotation invariant features (RIFs) [30], [34] and the translation invariant measurements (TIMs) [12], [35]. Nonetheless, the pairwise features make the number of input data increase squarely. Furthermore, a more efficient strategy is proposed based on the rotation decomposition, which decouples 6-DOF transformation into i) (2+1)-DOF, i.e., 2-DOF rotation axis and 1-DOF of translation along the axis, and ii) (1+2)-DOF, i.e., the remaining 1-DOF rotation and 2-DOF translation [31].\nIn contrast, we propose a novel efficient and determin-istic search strategy based on residual projections for the rigid registration problem, in which a novel pose decoupling strategy is introduced. Specifically, we decouple the 6-DOF original problem into three 2-DOF rotation search subproblems by projecting the residuals based on the Chebyshev distance, i.e., L ∞ residual [36], [37], on the coordinate axes. We define the consensus maximization objective function for each sub-problem and apply a BnB-based optimization method to search for the solution globally and deterministically. We derive a novel polynomial-time upper bound for our objective function based on the interval stabbing technology [13], [32], [38]. The proposed method sequentially searches for three 2-DOF rotation matrix components by BnB. Meanwhile, the translation projections on the coordinate axes are implicitly estimated by interval stabbing. After solving these three sub-problems, we can finally obtain the optimal 6-DOF transformation. Compared with existing methods, the parameter space of the proposed method only is 2-dimensional, thus enhancing the computational efficiency, as shown in Fig. 1.\nNotably, in contrast to existing BnB-based approaches, the proposed method requires no initialization of the translation domain, which is challenging to be accurately determined in different practical scenarios. Therefore, it avoids the problems that would arise when the translation domain is not initialized correctly. Notably, we can also partially verify if the total solution of the decomposed sub-problems is near the globally-optimal solution of the original problem by checking whether the solutions of the three sub-problems are orthogonal. Because the rotation matrix inherently is orthogonal.\nThe main contributions of this paper can be summarized as follows:"
},
{
"figure_ref": [],
"heading": "•",
"publication_ref": [],
"table_ref": [],
"text": "We propose a novel pose decoupling strategy based on the L ∞ residual projections. Compared with existing methods, our approach searches for the optimal solution in the low-dimensional solution domain, thereby improving search efficiency."
},
{
"figure_ref": [],
"heading": "•",
"publication_ref": [],
"table_ref": [],
"text": "We propose a novel BnB-based efficient and deterministic search method for each decoupled subproblem. Specifically, we define the inlier set maximization objective function and derive the upper bound for our objective based on the interval stabbing technology.\n• Due to its significant robustness, the proposed method can be extended to solve the challenging SPCR problem. We adapt the proposed upper bound to the extended objective function by interval merging technology.\nThe rest of this paper is organized as follows: The next section addresses the related work in two directions and discusses the innovations of the proposed method. Section 3 illustrates the problem formulation of our proposed method. Section 4 demonstrates the principle and details of our method. Section 5 presents extensive experimental results on both synthetic and real-world datasets. Finally, Section 6 gives a conclusion."
},
{
"figure_ref": [],
"heading": "RELATED WORK",
"publication_ref": [],
"table_ref": [],
"text": "As discussed in the previous section, there are two paradigms for the rigid point cloud registration problem regarding whether putative correspondences are given, i.e., correspondence-based registration and SPCR."
},
{
"figure_ref": [],
"heading": "Correspondence-Based Registration",
"publication_ref": [
"b38",
"b14",
"b15",
"b39",
"b13",
"b40",
"b42",
"b13",
"b24",
"b12",
"b30",
"b32",
"b12",
"b31",
"b30",
"b8",
"b11",
"b43"
],
"table_ref": [],
"text": "The correspondence-based registration comprises two steps: i) extract the 3D key-points and build putative correspondences by 3D feature descriptors, and ii) estimate the 6-DOF transformation based on the given correspondences. When the correspondences are correct, the registration problem has the elegant closed-form solution [39], although the non-convexity of SO(3). However, outliers in putative correspondences are inevitable in practical applications either for handcrafted or learning-based 3D descriptors [15], [16]. Therefore, robust registration techniques are indispensable. Consensus maximization is one of the most popular paradigms to address the robust registration problem, of which heuristic RANSAC [40] is the most representative. During each iteration, RANSAC employs a minimal solver to calculate the 3-DOF rotation and the 3-DOF translation separately. However, RANSAC only works efficiently under the conditions of low outlier rates. Recently, several RANSAC-based variants [14], [41]- [43] are proposed by introducing novel sampling strategies or local optimization methods. For instance, Graph-cut RANSAC [14] (GC-RANSAC) introduced the graph-cut algorithm to improve the local optimization performance. Nonetheless, RANSACbased methods are non-deterministic and generate a correct solution only with a certain probability due to the essence of random sampling [25].\nGiven this context, numerous deterministic and robust correspondence-based registration methods have been proposed, most of which rely on the globally-optimal BnB framework [13], [31]- [33]. The fundamental concept underlying BnB involves the iterative alternation between solution domain segmentation (branch) and sub-branch bounds computation (bound) until the globally-optimal solution is obtained. Parra and Chin [13] proposed a guaranteed outlier removal (GORE) method, which leverages geometrical bounds to prune outliers and guarantees that eliminated correspondences are not the inlier. GORE converts the 6-DOF registration problem to a 3DOF rotational registration problem and then utilizes BnB to maximize the inlier set. Similar to GORE, Cai et al. [32] presented a deterministic pre-processing method to prune outliers for the 4-DOF terrestrial LiDAR registration problem. More recently, Chen et al. [31] introduced an efficient decomposition scheme for the 6-DOF rigid registration, which decouples the 6-DOF original problem into a (2+1)-DOF sub-problem and a (1+2)-DOF sub-problem. These two 3-DOF sub-problems are then sequentially solved by BnB-based search methods. On the other hand, a representative method of the deterministic Mestimation paradigm is FGR [9]. This method formulates the registration problem by the Geman-McClure objective function and then combines graduated non-convexity (GNC) to solve the problem. Although the high efficiency of this method, it is easy to generate incorrect solutions at a high outlier rate. Combining the ideas of outlier removal and Mestimation, Yang et al. [12] proposed a certifiable and deterministic approach, i.e., Truncated least squares Estimation And SEmidefinite Relaxation (TEASER). TEASER leverages TIMs to decouple the 6-DOF transformation search problem into a 3-DOF rotation search sub-problem followed by a 3-DOF translation search sub-problem. Meanwhile, TEASER allows outlier pruning by maximum clique method [44], which, however requires quadratic memory space (O(N 2 )). As mentioned before, existing decoupling-based methods typically decompose the original 6-DOF problem into two 3-DOF sub-problems. Motivated by this observation, our study aims to develop a novel pose decoupling strategy to search the optimal parameters in low-dimensional space by exploring the geometric properties of the rigid point cloud registration problem."
},
{
"figure_ref": [],
"heading": "Simultaneous Pose and Correspondence Registration",
"publication_ref": [
"b9",
"b44",
"b45",
"b10",
"b26",
"b29",
"b10",
"b27",
"b26",
"b28",
"b29"
],
"table_ref": [],
"text": "The SPCR problem is more challenging since the transformation and correspondences need to be estimated simultaneously. A typical algorithm is ICP [10], an expectationmaximization (EM) type method. However, ICP is susceptible to local minima and is greatly influenced by the initialization of the transformation. Another series of noise-robust methods represent the point cloud as Gaussian mixture models (GMMs) to build robust objective functions based on the probability density [45], [46]. Although all these methods can efficiently converge to an optimum when they are initialized well, they can not provide any optimality guarantees.\nAnother line for the SPCR problem is to estimate the globally-optimal solution without initialization, which is commonly based on the BnB framework [11], [27]- [30]. Go-ICP [11] is the first practical globally-optimal approach for the 6-DOF SPCR problem that employs the nested BnB search structure to minimize the objective function based on the L 2 residual. Parra et al. [28] formulated the registration as a consensus maximization problem and proposed a BnBbased method with a tighter bound than Go-ICP. Campbell et al. [27] proposed a more efficient and robust BnB-based approach to minimize the GMM-based objective function. However, both these methods jointly search for the globallyoptimal solution over the 6-dimensional parameter space, leading to relatively high computational costs. One common direction to improve efficiency is reducing the dimension of the solution domain by decoupling the transformation. For instance, Straub et al. [29] proposed a decoupling method based on surface normal distributions that decomposes the 6-DOF registration problem into the separate 3-DOF rotation and 3-DOF translation sub-problems. Liu et al. [30] introduced the RIFs to enable sequential estimations of the 3-DOF translation and the 3-DOF rotation instead of the joint 6-DOF transformation estimation. By contrast, the proposed method not only has a novel decoupling strategy based on residual projections, but also can be extended to solve the SPCR problem by adjusting the objective and bound functions. The detailed formulation will be given in Section 4.3. Fig. 2: A toy 2D registration example to demonstrate L ∞ residual projection. Specifically, {(p i , q i )} 3 i=1 is the set of input correspondences, the red line segments represent the projections of the residual on the coordinate axes X and Y , i.e., r T j p 1 + t j -q j 1 , j = X, Y . The inlier constraint for L ∞ residual indicates that, (p 1 , q 1 ) is an inlier only if both residual projections on the coordinate axes are not larger than the inlier threshold."
},
{
"figure_ref": [],
"heading": "PROBLEM FORMULATION",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Inlier Set Maximization",
"publication_ref": [
"b10",
"b12",
"b27",
"b30",
"b31",
"b29",
"b35",
"b36"
],
"table_ref": [],
"text": "Given the source point cloud P and the target point cloud Q, a set of putative correspondences\nK = {(p i , q i )} N i=1\nis extracted by matching points between P and Q, where p i , q i ∈ R 3 , and N is the correspondences number. The proposed method aims to estimate the rigid transformation between the source and target point clouds. Specifically, the 6-DOF transformation matrix T ∈ SE( 3) is formed by the 3-DOF rotation matrix R ∈ SO(3) and the 3-DOF translation vector t ∈ R 3 . The rotation matrix R is an orthogonal matrix in which the columns and rows are orthogonal vectors, i.e., RR T = I. Formally, we adopt the inlier set maximization formulation for the robust registration problem:\nT * = arg max T ∈SE(3) C (T (P), Q) , (1\n)\nwhere C is the objective function for calculating the cardinality of the inlier set. Different from existing approaches [11], [13], [28], [31], [32] that commonly employ the L 2 residual to measure the alignment, we apply the Chebyshev distance, i.e., L ∞ residual [30], [36], [37], to build the robust objective function. Therefore, considering the presence of noise, we estimate the rotation and translation that maximize the objective:\nE(R, t|K, ) = N i=1 I ( Rp i + t -q i ∞ ≤ ) ,(2)\nwhere I(•) is the indicator function that returns 1 if the input condition is true and 0 otherwise, • ∞ denotes the L ∞norm, and is the inlier threshold."
},
{
"figure_ref": [],
"heading": "Residual Projections and Pose Decoupling",
"publication_ref": [
"b5"
],
"table_ref": [],
"text": "Mathematically, we apply the following definitions to derive the residual projections. Firstly, we denote the rotation matrix as\nR r X1 r X2 r X3 r Y 1 r Y 2 r Y 3 r Z1 r Z2 r Z3 = r X r Y r Z T . (3\n)\nwhere r j = [r j1 , r j2 , r j3 ] T , j = X, Y, Z, is the transpose of each row of the rotation matrix. The translation vector is\nt [t X , t Y , t Z ] T .(4)\nGiven the definitions of R and t, according to the definition of Chebyshev distance, the inlier constraint in the objective function (2) can be rewritten as\nRp i + t -q i ∞ ≤ (5a) ⇔ r T X r T Y r T Z p i + t X t Y t Z - q X i q Y i q Z i ∞ ≤ (5b) ⇔ max r T X p i + t X -q X i , r T Y p i + t Y -q Y i , r T Z p i + t Z -q Z i ≤ (5c) ⇔ r T X p i + t X -q X i ≤ , r T Y p i + t Y -q Y i ≤ , r T Z p i + t Z -q Z i ≤ (5d) ⇔ I r T X p i + t X -q X i ≤ = 1, I r T Y p i + t Y -q Y i ≤ = 1, I r T Z p i + t Z -q Z i ≤ = 1(5e)\nwhere q i q X i , q Y i , q Z i T , and r T j p i + t j -q j i , j = X, Y, Z, are projections of the i-th residual on the coordinate axes, as shown in Fig. 2. Then we can set I r T j p i + t j -q j i ≤ = L j i . Therefore, the objective function ( 2) can be reformulated as\nE(R, t|K, ) = N i=1 I L X i ∧ L Y i ∧ L Z i , (6\n)\nwhere ∧ is the logical AND operation. Geometrically, the objective function (6) indicates that, given an arbitrary correspondence (p i , q i ) and the inlier threshold , only when the residual projections on the X, Y , and Z coordinate axes are not larger than , (p i , q i ) is an inlier, as shown in Fig. 2. Notably, these three conditions are equally independent. Accordingly, we may reduce the original constraint in Eq. ( 6) as three separate constraints, i.e., L X i , L Y i , and L Z i . In this way, the original search problem for the transformation in SE(3) can be decoupled into three sub-problems. The new inlier set maximization objective for each subproblem can be\nE j (r j , t j |K, ) = N i=1 I r T j p i + t j -q j i ≤ , j = X, Y, Z.(7)\nIn other words, we reformulate the L ∞ residual-based objective function in the form of residual projections. Then we decompose the joint constraint into three independent constraints to decouple the original registration problem into three sub-problems, i.e., max\nE X (r X , t X |K, ), max E Y (r Y , t Y |K, ), and max E Z (r Z , t Z |K, ).\nIn the next section, we will introduce a step-wise search strategy to solve these three sub-problems."
},
{
"figure_ref": [],
"heading": "STEP-WISE SEARCH STRATEGY BASED ON BRANCH AND BOUND",
"publication_ref": [],
"table_ref": [],
"text": "Branch and bound (BnB) is an algorithm framework for global optimization. To design the BnB-based algorithm, two main aspects need to be addressed, i) how to parameterize and branch the solution domain, and ii) how to efficiently calculate the upper and lower bounds. Then the BnB-based algorithm can recursively divide the solution domain into smaller spaces and prune the sub-branches by upper and lower bounds until convergence."
},
{
"figure_ref": [],
"heading": "Parametrization of Solution Domain",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Rotation",
"publication_ref": [
"b6",
"b46",
"b47",
"b47",
"b46"
],
"table_ref": [],
"text": "For each sub-problem of the objective function (7), the unknown-but-sought vector r j (denoted by r in this section) is in a unit sphere (denoted by S 2 ). Then we divide the unit sphere into two unit hemispheres (S 2+ and S 2-) to represent the parameter spaces of the \"positive\" vector r and the \"negative\" vector -r. The \"upper\" hemisphere is defined as\nS 2+ = r|r T r = 1, r 3 ≥ 0 ,(8)\nwhere r [r 1 , r 2 , r 3 ] T is a unit vector in R 3 . Geometrically, since these two hemispheres are centrally symmetric, the \"lower\" hemisphere is S 2-which can be seen as -S 2+ . In order to parametrize S 2+ and S 2-minimally, we introduce the exponential mapping [47], [48] technology to map a 3dimensional hemisphere to a 2-dimensional disk efficiently. Specifically, given a vector r ∈ S 2+ , it can be represented by a corresponding point d ∈ R 2 in the 2D disk, i.e., r T = sin(γ) dT , cos(γ) , and\nd = γ d(9)\nwhere\nγ ∈ [0, π/2], dT = d/ d is a unit vector in R 2 .\nNotably, the range of γ corresponds to r 3 ≥ 0, and its maximum corresponds to the radius of the 2D disk, i.e., π/2, as shown in Fig. 3. For a vector -r ∈ S 2-, we define another exponential mapping method,\n-r T = -sin(γ) dT , cos(γ) .(10)\nAccordingly, the total solution domain (unit sphere) is mapped as two identical 2D disks, which represent the parameter spaces of r and -r, respectively. Compared to the unit sphere representation within three parameters and a unit-norm constraint, the exponential mapping is a more compact representation within only two parameters [48]. Meanwhile, for ease of operation, a circumscribed square of the disk domain is initialized as the domain of r in the proposed BnB algorithm, and the domain of -r is relaxed in the same way. Further, we introduce the following lemma [47] about the exponential mapping between S 2+ and R 2 . Lemma 1. r a , r b ∈ S 2+ are two vectors in the unit hemisphere, and d a , d b ∈ R 2 are corresponding points in the 2D disk. Then we have\n∠(r a , r b ) ≤ d a -d b . (11\n)\nFig. 3: The solution domain before and after exponential mapping, and the pipeline of our proposed BnB algorithm. The original solution domain of the vector r is a unit sphere in the 3D Euclidean space. The exponential mapping method maps the unit sphere to two identical 2D disks, representing the solution domains of r and -r, respectively. We can only branch one 2D-disk domain during each iteration, followed by the calculation of upper and lower bounds for each sub-branch. The proposed BnB algorithm converges until the optimal solution r * is found, and the optimal t * is found by interval stabbing at the same time. In the visualization results of interval stabbing, the black line segments are the candidate intervals of each correspondence, and the red line segments are the intervals crossed by the blue probe with the max-stabbing number.\nThe probe position is the max-stabbing position.\nAccording to Lemma 1, we can obtain the following proposition.\nProposition 1. Given a sub-branch of the square-shaped domain D, its center is d c ∈ R 2 and half-side length is σ. For ∀d ∈ D, we have\n∠ (r, r c ) ≤ d -d c ≤ √ 2σ,(12)\nwhere r and r c correspond to d and d c , respectively.\nDefining α max ∠ (r, r c ), we can obtain α ≤ √ 2σ with Proposition 1, as shown in Fig. 3. Geometrically, Proposition 1 indicates that one square-shaped sub-branch of the 2D disk domain is relaxed to a spherical patch of the 3D unit sphere. In addition, Lemma 1 and Proposition 1 hold for both hemispheres S 2+ and S 2-. In this study, we apply Proposition 1 as one of the fundamental parts to derive our proposed bounds."
},
{
"figure_ref": [],
"heading": "Translation",
"publication_ref": [
"b10",
"b29",
"b30"
],
"table_ref": [],
"text": "Estimating the translation component t j ∈ R, j = X, Y, Z in the objective function ( 7) is a 1-dimensional problem. The translation is unconstrained, and it is not easy to estimate a suitable solution domain accurately in advance for various practical scenarios. Existing BnB-based approaches [11], [30], [31] commonly initialize the translation domain as a redundant space and search it exhaustively, leading to a significant decrease in efficiency. Meanwhile, if the translation domain is not initialized correctly, the algorithm may not find the optimal (correct) solution since the optimal solution may be excluded from the initial search domain.\nIn our study, we propose an interval stabbing-based method to estimate the translation components {t X , t Y , t Z } without any prior information on the size of the translation domain, which can effectively reduce the total parameter space and improve the algorithm efficiency. It also avoids the problems that may arise when the translation initialization is incorrect. The proposed method will be described thoroughly in Section 4.2."
},
{
"figure_ref": [],
"heading": "Interval Stabbing and Bounds",
"publication_ref": [
"b6"
],
"table_ref": [],
"text": "We first introduce the following lemma to derive the bounds for the objective function (7)."
},
{
"figure_ref": [],
"heading": "Lemma 2. Given an arbitrary consensus maximization objective",
"publication_ref": [
"b23",
"b48",
"b31",
"b31",
"b48",
"b6",
"b6",
"b6",
"b6",
"b49",
"b29",
"b30"
],
"table_ref": [],
"text": "F (x|A) = M i=1 F i (x, a i ),\nwhere x is the variable to be calculated, A = {a i } M i=1 is the set of input measurements, and F i (x, a i ) is an indicator function with a certain constraint. Then we have\nmax x F (x|A) = max x M i=1 F i (x, a i ) ≤ M i=1 max x F i (x, a i ) . (13\n)\nProof. For the i-th input measurement a i , we can obtain\nF i (x, a i ) ≤ max x F i (x, a i ) ≤ 1. Therefore, it is obvious that the maximum of M i=1 F i (x, a i ) is not bigger than the sum of max x F i (x, a i ).\nIn this study, the upper and lower bounds are proposed as follows:\nProposition 2 (Upper bound for S 2+ ). Given a sub-branch of the square-shaped domain D, whose center is\nd c ∈ R 2 (corre- sponds to r c j ∈ S 2+\n) and half-side length is σ, the upper bound can be set as\nE + (D) = max tj N i=1 I t j ∈ t i- j , t i+ j ,(14a)\nt i- j = --p i cos max ∠ r c j , p i - √ 2σ, 0 + q j i , (14b) t i+ j = -p i cos min ∠ r c j , p i + √ 2σ, π + q j i .(14c)\nProof. First, we rewrite the maximum of the objective function (7) as,\nmax rj ,tj N i=1 L j i = max tj max rj N i=1 L j i .(15)\nTherefore, according to Lemma 2, we have\nmax tj max rj N i=1 L j i ≤ max tj N i=1 max rj I r T j p i + t j -q j i ≤ .(16\n) Additionally, given a sub-branch D, according to the triangle inequality in spherical geometry [24] and Proposition 1, we have\n∠ (r j , p i ) ≤ ∠ r c j , p i + ∠ r c j , r j (17a) ≤ ∠ r c j , p i + α (17b) ≤ ∠ r c j , p i + √ 2σ,(17c)\nand\n∠ (r j , p i ) ≥ ∠ r c j , p i -∠ r c j , r j (18a) ≥ ∠ r c j , p i -α (18b) ≥ ∠ r c j , p i - √ 2σ.(18c)\nThus, according to r T j p i = r j p i cos ∠ (r j , p i ) and r j = 1, we have\nr T j p i ∈ p i cos min ∠ r c j , p i + √ 2σ, π , p i cos max ∠ r c j , p i - √ 2σ, 0 .(19)\nThen, given a sub-branch D, whose center is d c (corresponds to r c j ) and half-side length is σ, we can observe that,\nmax rj I r T j p i + t j -q j i ≤ (20a) = max rj I --r T j p i + q j i ≤ t j ≤ -r T j p i + q j i (20b) ≤I t j ∈ t i- j , t i+ j ,(20c)\nwhere\nt i- j = --p i cos max ∠ r c j , p i - √ 2σ, 0 + q j i ,(21a)\nt i+ j = -p i cos min ∠ r c j , p i + √ 2σ, π + q j i .(21b)\nThen,\nN i=1 max rj I r T j p i + t j -q j i ≤ ≤ N i=1 I t j ∈ t i- j , t i+ j (22) Finally, we have max rj ,tj N i=1 L j i ≤ max tj N i=1 I t j ∈ t i- j , t i+ j (23\n)\nTherefore, Proposition 2 is proved.\nProposition 3 (Upper bound for S 2-). Given a sub-branch of the square-shaped domain D, whose center is\nd c ∈ R 2 (corre- sponds to -r c j ∈ S 2-\n) and half-side length is σ, the upper bound can be set as\nE -(D) = max tj N i=1 I t j ∈ t i- j , t i+ j ,(24a)\nt i- j = -+ p i cos min ∠ r c j , p i + √ 2σ, π + q j i ,(24b)\nt i+ j = + p i cos max ∠ r c j , p i - √ 2σ, 0 + q j i .(24c)\nProof. The proof is similar to Proposition 2, which is simple enough that we omit it.\nAlthough the upper bounds in Proposition 2 and Proposition 3 are theoretically provided, we still need to find an appropriate method to compute them. Mathematically, the calculation of the upper bounds is a typical interval stabbing problem [49]. As shown in Fig. 3, the interval stabbing problem aims to find a probe (i.e., the blue line segment) that stabs the maximum number of intervals. There has been a deterministic and polynomial-time algorithm [32] to solve the interval stabbing problem. More details are given in [32], [49].\nBy utilizing the interval stabbing technology to compute the upper bounds, the proposed BnB-based method only needs to search a 2-dimensional solution domain, thereby improving the algorithm efficiency. Meanwhile, the translation projections {t X , t Y , t Z } are implicitly estimated by interval stabbing without requiring the initialization of the translation domain. In other words, the interval stabbing approach returns not only the max-stabbing number (i.e., the upper bound), but also the max-stabbing position (i.e., the estimation of t j ).\nTo sum up, considering the total solution domain S 2+ and S 2-, we have the following proposition.\nProposition 4 (Upper bound for S 2 ). Given a sub-branch of the square-shaped domain D, whose center is d c ∈ R 2 and halfside length is σ, the upper bound of the objective function (7), can be set as\nE(D) = max E + (D), E -(D) .(25)\nProof. The maximum of these two upper bounds is not smaller than the maximum of the objective function (7).\nTherefore, E(D) is the final upper bound of the objective function (7).\nProposition 5 (Lower bound for S 2 ). Given a sub-branch of the square-shaped domain D, whose center is d c ∈ R 2 and halfside length is σ, the lower bound of the objective function ( 7) can be set as\nE(D) = max E + (D), E -(D) ,(26a)\nE + (D) = N i=1 I r c j T p i + t + j -q j i ≤ ,(26b)\nE -(D) = N i=1 I -r c j T p i + t - j -q j i ≤ ,(26c)\nwhere t + j is the max-stabbing position of the upper bound for S 2+ , and t j is the max-stabbing position of the upper bound for S 2-. Proof. The maximum of the objective function in the given sub-branch D should be no less than any objective value at a specific point. Therefore, E(D) is the lower bound of the objective function (7).\nBased on the upper and lower bounds in Proposition 4 and Proposition 5, the proposed BnB-based algorithm for the 2-DOF rotation matrix components search sub-problem is outlined in Algorithm 1. We employ the depth-first search strategy [50] to implement the proposed BnB algorithm. As we indicated in Section 4.1, although the initial solution domain S 2 is mapped to two identical 2D-disks, only one disk domain is branched, since the bounds of S 2+ and S 2-can be computed separately in the same disk domain, as shown in Fig. 3. During each iteration, the branch with maximal upper bound is partitioned into four sub-branches since the current parameter space is only 2-dimensional. Then the branch list is updated, and the upper and lower bounds for each sub-branch are estimated. The sub-branches that do not have a better solution than the best-so-far solution are eliminated. As the number of iterations increases, the gap between the upper and lower bounds gradually decreases. Until the gap reduces to zero, the BnB algorithm obtains the optimal solutions r * j and t * j . As shown in [30], [31], existing methods usually solve sub-problems sequentially. However, we can solve three sub-problems in an arbitrary order with Algorithm 1 and obtain the final optimal solution R * and t * ."
},
{
"figure_ref": [],
"heading": "Simultaneous Pose and Correspondence Registration",
"publication_ref": [
"b23",
"b33",
"b26",
"b48"
],
"table_ref": [],
"text": "In this section, we extend our proposed correspondencebased registration method to address the challenging simultaneous pose and correspondence registration (SPCR) problem. The SPCR problem is much more complicated than the correspondence-based problem. Formally, given the source point cloud P = {p i } M i=1 and the target point cloud Q = {q k } N k=1 , there are M × N candidate correspondences totally. Similar to [24], [34], we define an inlier set maximization objective function for the SPCR problem as 10\nS j (r j , t j |P, Q, ) = M i=1 max k I r T j p i + t j -q j k ≤ ,(27)\nUpdate U = maxE(D k ), D k ∈ ξ; 11 Update L = max {L, E(D k )} with D k ∈ ξ, if E(D k ) > L, set r * j = δ(D k ) and t * j = η(D k ); 12 Eliminate D k from ξ if E(D k ) < L, D k ∈ ξ; 13 end\nwhere j = X, Y, Z. This problem formulation means that for each point p i we attempt to seek a \"closest\" point q k that contributes a maximum of 1 to the objective function. The upper and lower bounds for the SPCR objective (27) are slightly different from those of correspondence-based registration, given by the following propositions. In addition, the optimization of objective ( 27) is also based on BnB.\nProposition 6 (SPCR Upper bound for S 2 ). Given a subbranch of the square-shaped domain D, whose center is d c ∈ R 2 and half-side length is σ, the SPCR upper bound for S 2+ can be set as\nE + SP CR (D) = max tj M i=1 max k I t j ∈ t ik- j , t ik+ j ,(28a)\nt ik- j = --p i cos max ∠ r c j , p i - √ 2σ, 0 + q j k ,(28b)\nt ik+ j = -p i cos min ∠ r c j , p i + √ 2σ, π + q j k . (28c\n)\nThe SPCR upper bound for S 2-can be set as\nE - SP CR (D) = max tj M i=1 max k I t j ∈ t ik- j , t ik+ j ,(29a)\nt ik- j = -+ p i cos min ∠ r c j , p i + √ 2σ, π + q j k ,(29b)\nt ik+ j = + p i cos max ∠ r c j , p i - √ 2σ, 0 + q j k . (29c\n)\nThe final SPCR upper bound for S 2 can be set as\nE SP CR (D) = max E + SP CR (D), E - SP CR (D) .(30)\nProof. The proof is similar to the proofs of Proposition 2, Proposition 3, and Proposition 4, hence we omit it.\nAccording to Proposition 6, since there are N intervals for each point p i , we cannot directly employ the interval stabbing for these M × N intervals. Therefore, the interval merging technology [49] can be employed as a preprocessing method before applying the interval stabbing algorithm to calculate bounds (28a) and (29a). In other words, after interval merging, the max-stabbing probe can only stab through at most one interval for each point p i , i.e., it contributes a maximum of 1 to the upper bound functions (28a) and (29a). We then develop Algorithm 2 to achieve interval merging. As shown in Algorithm 2, interval merging is executed one time for each point p i , and then a total of M times for point cloud P. An example of the visualization results of interval merging and stabbing is given in Fig. 4. Similarly, when computing the SPCR lower bound, we employ another indicator function to solve this \"multi-interval\" problem, as shown in Eq. (31b) and Eq. (31d) of the following proposition.\nProposition 7 (SPCR Lower bound for S 2 ). Given a subbranch of the square-shaped domain D, whose center is d c ∈ R 2 and half-side length is σ, the SPCR lower bound can be set as\nE SP CR (D) = max E + SP CR (D), E - SP CR (D) ,(31a)\nE + SP CR (D) = M i=1 I M j+ i > 0 ,(31b)\nM j+ i = N k=1 I r c j T p i + t + j -q j k ≤ ,(31c)\nE - SP CR (D) = M i=1 I M j- i > 0 ,(31d)\nM j- i = N k=1 I -r c j T p i + t - j -q j k ≤ ,(31e)\nwhere t + j is the max-stabbing position of the SPCR upper bound for S 2+ , and t j is the max-stabbing position of the SPCR upper bound for S 2-.\nProof. The proof is similar to the proofs of Proposition 5, hence we omit it.\nTo improve the total efficiency, we can only solve the first sub-problem (i.e., S X (r X , t X |P, Q, )) using the extended BnB-based SPCR approach. Then we can solve the second sub-problem (i.e., E Y (r Y , t Y |K, )) and third sub-problem (i.e., E Z (r Z , t Z |K, )) by Algorithm 1. This is because we can obtain the candidate inlier correspondences after solving the first SPCR sub-problem, which are implicitly determined by the residual projection constraint. Notably, partial outliers occasionally satisfy this constraint and cannot be removed. However, the proposed correspondence-based method can be directly applied to address the two remaining sub-problems robustly. , where N i is the number of merged intervals for p i ; 17 end"
},
{
"figure_ref": [],
"heading": "EXPERIMENTS",
"publication_ref": [],
"table_ref": [],
"text": "This section presents a comprehensive comparison of the proposed method with state-of-the-art correspondencebased methods on both synthetic and real-world datasets. Additionally, we evaluate the extended method against existing SPCR methods specifically on synthetic data. We implement the proposed method in Matlab 2019b and conduct all experiments on a laptop with an i7-9750H CPU and 16GB RAM."
},
{
"figure_ref": [],
"heading": "Experimental Setting",
"publication_ref": [
"b12",
"b39",
"b11",
"b8",
"b13",
"b12",
"b11",
"b39",
"b10",
"b9",
"b44",
"b45",
"b11",
"b30",
"b51",
"b51"
],
"table_ref": [],
"text": "We denote the proposed method as Ours. The compared methods for correspondence-based registration are as follows,\n• GORE [13]: A guaranteed outlier removal registration method based on BnB and pose decoupling. It is implemented in C++.\n• RANSAC [40]: A typical consensus maximization registration approach implemented in Matlab. The maximum number of iterations is set to 10 4 .\n• TEASER [12]: A certifiable decoupling-based registration method with a robust cost function. It is implemented in C++.\n• FGR [9]: A fast registration method with a robust cost function. It is implemented in C++.\n• GC-RANSAC [14]: A variant of RANSAC-based registration method with improvements in local optimization. It is implemented in C++, and the maximum number of iterations is set to 10 4 . GORE [13] >1 hour TEASER [12] out-of-memory RANSAC [40] 0 Besides, the compared methods for SPCR are as follows,\n• GO-ICP [11]: A 6-DOF global optimal registration method based on BnB. It is implemented in C++.\n• GO-ICPT: A variant of GO-ICP with outlier trimming and the trimming fraction is set to 10%.\n• ICP [10]: A typical EM-type method implemented by pcregistericp function in MATLAB.\n• CPD [45]: A robust GMM-based registration approach implemented in C.\n• GMMReg [46]: A robust and general GMM-based registration method implemented in C.\nSimilar to [12], [31], [52], the evaluation metrics for point cloud registration in this study include 1) rotation error E R , 2) translation error E t , 3) running time, 4) success rate SR, and 5) F 1-score. The error definitions are as follows:\nE R = arccos T r(R -1 gt R * ) -1 2 , (32a\n)\nE t = t gt -t * , (32b\n)\nwhere t gt and R gt are the ground truth, t * and R * are the estimated solutions, and T r(•) is the trace of a matrix. The successful cases must satisfy the predefined threshold for E R and E t . Besides, the definition of F 1-score is given in [52]."
},
{
"figure_ref": [],
"heading": "Synthetic Data Experiments",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we conduct various experiments on synthetic data to compare the performance of the proposed method with state-of-the-art correspondence-based and correspondence-free registration methods."
},
{
"figure_ref": [],
"heading": "Data generation",
"publication_ref": [],
"table_ref": [],
"text": "Firstly, we randomly generate the source point cloud P in the cube [-100, 100] 3 . The source point cloud is transformed by a random rotation R gt ∈ SO(3) and a random translation t gt ∈ [-100, 100] 3 to generate the target point cloud Q.\nThen a portion of the points in the target point cloud is replaced by arbitrarily generated points to simulate outliers. The outlier rate η is the ratio of these replaced points to all points. Besides, zero-mean Gaussian noise with standard deviation σ is added to the target point cloud. Notably, the inlier threshold in each synthetic data experiment is set according to the standard deviation of the noise."
},
{
"figure_ref": [],
"heading": "Efficiency and accuracy experiments",
"publication_ref": [
"b30",
"b30"
],
"table_ref": [],
"text": "This section presents three sets of experiments comparing the efficiency and accuracy of Ours with GORE, RANSAC, TEASER, FGR, GC-RANSAC, and TR-DE. Rotation errors, translation errors, and time costs are recorded for each experiment group. The first group focuses on experiments with a regular number of correspondences. We randomly generate N = {1000, 2000, . . . , 5000} correspondences with a noise level of σ = 0.5 and an outlier rate of η = 0.5.\nThe experiment is repeated 50 times for each setting, and the average results are depicted in Fig. 5. It is worth noting that results are not reported when the running time exceeds 1800 seconds. Among the deterministic methods, GORE and TEASER exhibit relatively high accuracy. However, their time costs increase significantly as the number of correspondences grows, with TEASER being the fastest in this regard. FGR, on the other hand, demonstrates occasional unsuccessful results but shows high efficiency. RANSAC and GC-RANSAC suffer from lower accuracy due to sampling uncertainty. Nevertheless, they exhibit relatively high efficiency at the regular outlier rate (η = 0.5). In contrast, Ours outperforms all other methods in terms of both efficiency and accuracy. Notably, when N reaches 4000, Ours is approximately 10 4 times faster than GORE and TEASER. This may be explained by the reason that even after outlier rejection, a significant number of candidate inlier correspondences are still retained when dealing with a large number of correspondences. Consequently, the optimization process for GORE and TEASER becomes slower.\nSince the code of TR-DE [31] is not released publicly, we set the same experimental conditions as TR-DE to compare the performance, which is the second group of experiments. Specifically, the source point cloud is randomly generated within the unit cube, and the experiment is conducted with N = {2000, 2500, . . . , 4000}, σ = 0.005, and, η = 0.55. We also conduct 200 independent trials for each setting and record the average experiment results, as shown in Fig. 6. We use the gray rectangular region to approximately represent the results of TR-DE given in [31]. We can observe that, Ours is about 10 times faster than TR-DE while keeping comparable accuracy.\nTo further investigate the potential efficiency advantages of Ours, we conduct the third group of experiments, specifically focusing on extremely high numbers of correspondences: N = {10k, 20k, 50k, 100k, 200k, 500k} (where k denotes one thousand). The remaining settings are consistent with those of the first group. Table 1 presents the average rotation errors, average translation errors, and average time costs of each method. GORE's running time exceeds one hour starting from N = 10k, thus its results are not reported. Furthermore, TEASER demands a substantial amount of memory space, which renders it unable to operate efficiently under such extreme experimental conditions. As N increases to 500k, RANSAC yields numerous unsatisfactory solutions and incurs a time cost of up to 315.7s. Additionally, GC-RANSAC fails to converge to the correct result after N reaches 50k due to early termination. In comparison to FGR, Ours delivers more accurate rotation estimates but slightly less accurate translation estimates. However, experimental results indicate that the number of correspondences has a relatively minor impact on the efficiency of our method. For instance, when the number of correspondences increases from 10k to 500k, Ours is approximately 8 to 20 times faster than RANSAC and roughly 4 times faster than FGR. Overall, the proposed method exhibits superior efficiency while maintaining competitive accuracy compared to state-of-the-art approaches."
},
{
"figure_ref": [
"fig_5"
],
"heading": "Robustness experiments",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we conduct a group of controlled experiments to compare the robustness of Ours with GORE, RANSAC, TEASER, FGR, and GC-RANSAC. We randomly generate N = 2000 correspondences with varying outlier rates (η = {0.1, 0.2, . . . , 0.8}) and a noise level of σ = 0.5. The average rotation errors, average translation errors, and average time costs for each method are reported in Fig. 7. Results beyond a running time of 1800 seconds are not recorded in this group of experiments. Comparing the registration errors demonstrates that Ours, GORE, RANSAC, and TEASER are robust against up to 80% outlier rates. RANSAC has relatively higher registration errors compared to Ours, GORE, and TEASER. Moreover, the running time of RANSAC increases significantly with an increase in the outlier rate. In contrast, both GORE and TEASER display a significant decrease in running times as the outlier rate increases, due to a corresponding reduction in the number of inliers. This indicates that, for GORE and TEASER, the time required for outlier removal is considerably smaller compared to the time spent on the optimization part. Consequently, they exhibit lower efficiency at regular outlier rates (e.g., η ≤ 0.5). On the other hand, despite the high efficiency exhibited by both the deterministic FGR and the non-deterministic GC-RANSAC, they do not perform well when confronted with high outlier rates (e.g., η ≥ 0.7). In contrast, Ours stands out as one of the fastest and most robust methods."
},
{
"figure_ref": [],
"heading": "Challenging SPCR experiments",
"publication_ref": [
"b19",
"b10",
"b11"
],
"table_ref": [],
"text": "In this section, we evaluate the performance of our extended simultaneous pose and correspondence registration (SPCR) method against GoICP, GoICPT, ICP, CPD, and GMMReg using the Bunny dataset [20]. The Bunny dataset consists of 35947 points, and is pre-normalized to fit within the cube [-1, 1] 3 , as required by GoICP [11]. Similar to [12], we down-sample the Bunny dataset to M = 100 points, which serve as the source point cloud P. To generate the target point cloud Q, we apply a random rotation and translation to the source point cloud. Additionally, we randomly remove a certain proportion of points from Q to simulate partial overlap between P and Q. The visualization results for a pair of synthetic data are shown in Fig. 1(d-1), where the bolded points represent the down-sampled point clouds. Furthermore, we add zero-mean Gaussian noise with σ = 0.001 to the source point cloud P. The registration experiment is repeated 50 times for each overlap rate in ρ = {0.9, 0.8, . . . , 0.4}.\nThe registration errors and average running times for each approach are presented in Fig. 8. During repeated experiments, the local methods, ICP, CPD, and exhibit a tendency to converge to local optima, resulting in incorrect results. However, their efficiency remains a notable advantage. In contrast, the global methods GoICP and its variant GoICPT demonstrate greater robustness compared to these local methods. In particular, GoICPT with a 10% trimming ratio achieves significantly higher accuracy at an overlap rate of 0.1. Nevertheless, these global methods suffer from relatively slow running times, which increase more rapidly than our proposed method. Consequently, when the overlap ratio is low (e.g., ρ ≤ 0.7), Ours is faster than GoICP and GoICPT. As a global method, Ours also falls short in terms of efficiency compared to the local methods. However, Ours is more robust than local methods ICP, CPD, and GMMReg. Furthermore, as depicted in Fig. 1(d), Ours exhibits greater robustness than ICP and higher efficiency than GoICP on a randomly generated pair of Bunny data (ρ = 0.6). These experiments illustrate the potential practi- cality of our proposed approach in addressing the challenging SPCR problem and its strength in terms of robustness and efficiency."
},
{
"figure_ref": [],
"heading": "Real-World Data Experiments",
"publication_ref": [
"b16",
"b17",
"b18"
],
"table_ref": [],
"text": "To assess the performance of the proposed method on real-world data, we utilize the Bremen dataset [17], ETH dataset [18], and KITTI dataset [19] in this section. These datasets present challenging outdoor LiDAR scenarios, with the former two captured using terrestrial LiDAR and the latter collected from onboard LiDAR."
},
{
"figure_ref": [
"fig_6"
],
"heading": "Bremen dataset experiments",
"publication_ref": [
"b16",
"b3",
"b6",
"b52",
"b53",
"b14"
],
"table_ref": [
"tab_1"
],
"text": "The Bremen dataset [17] is a large-scale outdoor dataset with 13 LiDAR scans. Similar to [4], [7], we initially down-sample the scans using the voxel grid algorithm [53]. Subsequently, we extract ISS [54] key-points and calculate FPFH [15] descriptors for each key-point. Through K-nearest neighbor search, we generate the set of putative correspondences K.\nThe ground-truth pose for each scan is provided within the dataset. Since the proposed method is only for pairwise registration, we construct 12 scan pairs to register all scans. Table 2 provides detailed information for each scan pair from the Bremen dataset, including the number of points, number of key-points, number of correspondences, and outlier rate. The down-sampling resolution for the Bremen dataset is set to 0.15m, which also determines the inlier threshold. With several thousand correspondences, the outlier rate ranges from approximately 90% to 99% for the Bremen dataset. We employ the proposed method (Ours), as well as GORE, RANSAC, TEASER, FGR, and GC-RANSAC, to register these scan pairs.\nWe compare the rotation error, translation error, and running time of each method for each scan pair, as shown in Fig. 9. Notably, when dealing with the registration of the s2-s0 scan pair, all compared methods, except for Ours, fail due to the exceptionally high outlier rate (99.64%). GORE and TEASER demonstrate successful alignment for the remaining scan pairs with relatively high accuracy. Despite this, GORE exhibits the highest time cost among all methods, even when the number of correspondences is small or the outlier rate is low, consistent with the findings from synthetic data experiments. For instance, in the case of the s10-s9 pair, which only has a 90.67% outlier rate, GORE requires over 3 hours for alignment, while TEASER takes up to 26.84 seconds. In contrast, Ours achieves registration in a mere 0.342 seconds. Furthermore, Fig. 1(a) shows another registration case for the s8-s7 scan pair, where Ours not only achieves superior accuracy but also is approximately 10 3 times faster than GORE and about 4 times faster than TEASER.\nOn the other hand, RANSAC demonstrates unstable performance, occasionally generating unsatisfactory solutions with significant registration errors, as observed in pairs s1-s0, s2-s0, s4-s2, and s9-s7. Moreover, RANSAC is also timeconsuming in these practical scenarios with high outlier rates and a large number of correspondences. FGR, while fast for all scan pairs, often converges to erroneous results. Although GC-RANSAC outperforms RANSAC in terms of stability and efficiency, it still struggles to successfully register all scan pairs. In contrast, Ours exhibits remarkable robustness, achieving a 100% registration success rate on the Bremen dataset. Furthermore, Ours shows higher efficiency compared to GORE and TEASER, which exhibit similar levels of robustness and accuracy as Ours."
},
{
"figure_ref": [
"fig_8",
"fig_10"
],
"heading": "ETH dataset experiments",
"publication_ref": [
"b17",
"b6"
],
"table_ref": [
"tab_2"
],
"text": "The ETH dataset [18] is a challenging large-scale LiDAR dataset that encompasses five distinct scenarios: Arch, Courtyard, Facade, Office, and Trees. The average overlap rates of these scenarios are 30 -40%, 40 -70%, 60 -70%, >80%, ≈ 50%, respectively, as reported in [7]. To ensure the generality of the registration algorithm, we select two scan pairs from each scenario for our registration experiments. The ETH dataset provides ground truth information regarding the relative pose, enabling accurate evaluation. We follow the same data preparation strategy outlined in Section 5.3.1 to establish the initial correspondence set K. For the ETH dataset, the down-sampling resolution and the inlier threshold are both set to 0.1m. Detailed information about the ETH dataset, including the number of points, number of key-points, number of correspondences, and outlier rate, can be found in Table 3. The outlier rate in the ETH dataset ranges from approximately 86% to 99%, with the number of correspondences varying from around 1k to 15k. To evaluate the registration performance, we compare Ours, GORE, RANSAC, TEASER, FGR, and GC-RANSAC using a total of 10 scan pairs from the ETH dataset.\nFig. 11 reports the rotation error, translation error, and running time for each method evaluated on the ETH dataset. Ours, GORE, and TEASER achieve remarkable robustness over all five scenes, successfully registering all scan pairs. GORE exhibits better accuracy overall compared to Ours, although it is time-consuming. Nevertheless, the registration errors achieved by Ours are still acceptable for practical applications. While the overall accuracy of TEASER is lower than that of Ours and GORE, its time cost increases significantly when dealing with a large number of inliers. results of the proposed method for the remaining four scenarios are provided in Fig. 12. Similar to the results in Section 5.3.1, FGR and GC-RANSAC demonstrate relatively high efficiency but exhibit instability when registering scan pairs with high outlier rates. RANSAC is not only timeconsuming on the ETH dataset, but also prone to producing incorrect registration results. In summary, benefiting from the pose decoupling strategy based on residual projections, the proposed registration method is more efficient than the state-of-the-art methods while ensuring superior robustness."
},
{
"figure_ref": [],
"heading": "KITTI dataset experiments",
"publication_ref": [
"b15",
"b30",
"b51",
"b18",
"b15",
"b50",
"b51"
],
"table_ref": [
"tab_3"
],
"text": "Following the data preparation strategy in [16], [31], [52], we evaluate the performance of the proposed method on the KITTI dataset [19]. The initial correspondences are generated using the learning-based descriptor FCGF [16], and the inlier threshold is set to 0.6m. For successful registration, we set the thresholds for rotation error (E R ) and translation error (E t ) to 5 • and 0.6m, respectively. In addition to comparing the performance of Ours against traditional methods such as RANSAC, TEASER, FGR, GC-RANSAC, and TR-DE, we also compare it with learningbased methods, including DGR [51] and PointDSC [52].\nNotably, the learning-based descriptor FCGF outperforms traditional descriptors, resulting in a relatively low outlier rate for FCGF-based correspondences (approximately 58.7% on average). Consequently, GORE is significantly slow on the KITTI dataset, and therefore, we do not report its results. As shown in Table 4, the success rate of all methods is over 95% due to the low outlier rate of FCGF-based correspondences. Among these methods, Ours achieves the highest success rate of 98.20% as well as TR-DE. Although Ours is not the most efficient method, it ranks second in terms of efficiency among all methods. For instance, Ours is approximately 5 times faster than the BnB-based TR-DE, about 4 times faster than the non-deterministic RANSAC, and approximately 50 times faster than the deterministic TEASER. It is worth mentioning that the most efficient method is the learning-based registration method PointDSC. However, learning-based methods often require additional training procedures and may perform well only on the datasets they were trained on. Additionally, Ours exhibits the best rotation accuracy and F 1-score. Fig. 1(c) provides an example of registering a selected pair from the KITTI dataset, where Ours has better accuracy and efficiency than FGR and GC-RANSAC. In general, compared to the stateof-the-art methods, including learning-based methods, Ours demonstrates competitive performance in efficiency and even better performance in robustness. This demonstrates the effectiveness of both the pose decoupling strategy and the BnB-based search method in our proposed method."
},
{
"figure_ref": [],
"heading": "CONCLUSION",
"publication_ref": [],
"table_ref": [],
"text": "In this paper, we present an efficient and deterministic point cloud registration method, leveraging a novel pose decoupling strategy. By utilizing L ∞ residual projections, we successfully decouple the initial 6-DOF problem into three 2-DOF sub-problems, resulting in improved efficiency. Furthermore, we introduce a step-wise search strategy based on branch and bound for these sub-problems. Specifically, we define the inlier set maximization objective function and derive the novel upper bound based on the interval stabbing technology. Interestingly, thanks to its significant robustness, our proposed method can be extended to solve the challenging SPCR problem by introducing the interval merging technology. Extensive experiments conducted on both synthetic and real-world datasets demonstrate the superior performance of our proposed method in terms of efficiency and robustness when compared to state-of-the-art approaches."
},
{
"figure_ref": [],
"heading": "ACKNOWLEDGMENT",
"publication_ref": [],
"table_ref": [],
"text": "This research was supported by the Federal Ministry for Digital and Transport, Germany as part of the Providentia++ research project (01MM19008A)."
}
] | Estimating the rigid transformation between two LiDAR scans through putative 3D correspondences is a typical point cloud registration paradigm. Current 3D feature matching approaches commonly lead to numerous outlier correspondences, making outlier-robust registration techniques indispensable. Many recent studies have adopted the branch and bound (BnB) optimization framework to solve the correspondence-based point cloud registration problem globally and deterministically. Nonetheless, BnB-based methods are time-consuming to search the entire 6-dimensional parameter space, since their computational complexity is exponential to the dimension of the solution domain. In order to enhance algorithm efficiency, existing works attempt to decouple the 6 degrees of freedom (DOF) original problem into two 3-DOF sub-problems, thereby reducing the dimension of the parameter space. In contrast, our proposed approach introduces a novel pose decoupling strategy based on residual projections, effectively decomposing the raw problem into three 2-DOF rotation search sub-problems. Subsequently, we employ a novel BnB-based search method to solve these sub-problems, achieving efficient and deterministic registration. Furthermore, our method can be adapted to address the challenging problem of simultaneous pose and correspondence registration (SPCR). Through extensive experiments conducted on synthetic and real-world datasets, we demonstrate that our proposed method outperforms state-of-the-art methods in terms of efficiency, while simultaneously ensuring robustness. | Efficient and Deterministic Search Strategy Based on Residual Projections for Point Cloud Registration | [
{
"figure_caption": "-E R = 0.055 • E R = 0.197 • E R = 0.104 • -E t = 0.018m E t = 0.031m E t = 0.018m -time = 687.7s time = 5.917s time = 1.436sCorrespondence-based(FCGF descriptor) (c-1) Initial (73.73% outlier) (c-2) FGR [9] (c-3) GC-RANSAC [14] (c-4) Ours -E R = 0.220 • E R = 0.294 • E R = 0.212 • -E t = 0.521m E t = 0.510m E t = 0.374m -time = 1.258s time = 1.576s time = 0.562sCorrespondence-free (d-1) Initial (60% overlap) (d-2) ICP [10] (d-3) GoICP [11] (d-4) Ours -E R = 173.7 • E R = 2.092 • E R = 0.061 • -E t = 0.240m E t = 0.017m E t = 0.004m -time = 0.069s time = 80.40s time = 2.524s",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "38 Fig. 4 :384Fig. 4: The visualization results for the 24-th iteration of a representative SPCR test on synthetic data. The final SPCR upper bound for S 2 is 38. The black line segments are the intervals after interval merging, and the red line segments are the intervals crossed by the blue probe with the maxstabbing number.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "384",
"figure_type": "figure"
},
{
"figure_caption": "Algorithm 1 : 8 Subdivide 9 Insert189BnB for 2-DOF rotation matrix components search sub-problem Input: The set of correspondences K = {(p i , q i )} N i=1 , inlier threshold . Output: Optimal solution r * j ∈ S 2 and t * j ∈ R. 1 Initialize the solution domain D 0 ; 2 Initialize the list of sub-branches ξ = {D 0 }; 3 Initialize the lower bound L = 0, and the upper bound U = N ; 4 Define function δ(D) returns the center of sub-branch D corresponding to the upper bound; 5 Define function η(D) returns the max-stabbing position of sub-branch D corresponding to the upper bound; 6 while U -L > 0 do 7 Select a sub-branch D with the maximal upper bound from ξ, i.e., D = arg max E(D k ), D k ∈ ξ; D into four sub-branches {D 1 , . . . , D 4 }; {D 1 , . . . , D 4 } into ξ and eliminate D from ξ;",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "189",
"figure_type": "figure"
},
{
"figure_caption": "Algorithm 2 : 4 Sort5 6 Initialize a 1 =2461Interval merging for SPCR bounds calculation Merged intervals ψ = {[a l , b l ]} N l=1 , where N = M i=1 N i . 1 Initialize the index i = 1; 2 Initialize the list of merged intervals ψ = ∅; 3 while i ≤ M do Initialize the index k = 1, l = 1; t i1- j , b 1 = t i1+ j ; intervals into list ψ, i.e., ψ = ψ ∪ [a l , b l ] Ni l=1",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "2461",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 5 :Fig. 6 :56Fig. 5: Controlled experiments with N = {1000, 2000, . . . , 5000}. The results include average rotation errors, average translation errors, and average running times.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "56",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 7 :7Fig. 7: Controlled experiments with η = {0.1, 0.2, . . . , 0.8}. The results include average rotation errors, average translation errors, and average running times.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "7",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 9 :9Fig.9: Experiment results on the Bremen dataset[17] with the FPFH[15] descriptor. The results include rotation errors, translation errors, and running times.",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "9",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 10 :10Fig. 10: Complete registration results of Ours on the Bremen dataset [17], where different scans are indicated by different colors. The pair-wise point cloud registration is conducted for all 12 scan pairs.",
"figure_data": "",
"figure_id": "fig_7",
"figure_label": "10",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 11 :11Fig.11: Experiment results on the ETH dataset[18] with the FPFH[15] descriptor. The results include rotation errors, translation errors, and running times.",
"figure_data": "",
"figure_id": "fig_8",
"figure_label": "11",
"figure_type": "figure"
},
{
"figure_caption": "For instance, Ours is approximately 280 times faster than TEASER in aligning the scan pair Courtyard1 with an outlier rate of 86.55%, and about 42 times faster than TEASER in aligning the scan pair Courtyard2 with an outlier rate of 90.62%. Another registration case for the scan pair Arch1 is illustrated in Fig. 1(b), where Ours achieves the lowest translation error and is roughly 480 times faster than GORE and about 4 times faster than TEASER. The visualization",
"figure_data": "",
"figure_id": "fig_9",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 12 :12Fig. 12: Registration results of the proposed method on the ETH dataset [18], including four scan pairs: (a) Courtyard1, (b) Facade1, (c) Office1, and (d) Trees1. The aligned source point cloud is blue, and the target point cloud is yellow.",
"figure_data": "",
"figure_id": "fig_10",
"figure_label": "12",
"figure_type": "figure"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "",
"figure_data": ".121|0.152|4.5060.086|0.185|6.7320.073|0.153|12.510.079|0.137|30.130.080|0.143|92.374.961|4.883|315.7GC-RANSAC [14]0.111|0.176|2.6360.100|0.200|9.976122.4|141.3|20.64139.3|142.0|20.64126.3|138.4|20.64133.4|148.5|20.64FGR [9]0.021|0.010|1.5400.024|0.013|2.4770.037|0.022|6.3460.031|0.018|12.130.024|0.018|23.970.047|0.025|68.74Ours 0.016|0• DGR [51]: A learning-based outlier rejection methodfor point cloud registration. It is implemented inPython.•PointDSC [52]: A learning-based outlier rejectionmethod for point cloud registration. It is imple-mented in Python.",
"figure_id": "tab_0",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Detailed information about the Bremen Dataset.",
"figure_data": "Scan pairNumber of points (10 6 )Number of key-pointsNumber of correspondencesOutlier rates1-s016.16-15.90 30328-30290500198.54%s2-s015.25-15.90 39368-30290630399.64%s3-s215.03-15.25 43856-39368819495.08%s4-s218.05-15.25 26581-39368539397.59%s5-s418.76-18.05 20023-26581476891.19%s6-s520.33-18.76 9423-20023184097.34%s7-s618.47-20.33 16608-9423255493.46%s8-s715.85-18.47 19599-16608393494.20%s9-s716.29-18.47 32281-16608429197.48%s10-s915.18-16.29 36689-32281866290.67%s11-s914.61-16.29 37187-32281756396.50%s12-s10 15.76-15.18 36084-36689821492.62%",
"figure_id": "tab_1",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Detailed information about the ETH Dataset.",
"figure_data": "Scan pairNumber of points (10 6 )Number of key-pointsNumber of correspondencesOutlier rateArch123.56-30.90 19007-122541261798.45%Arch230.90-29.45 12254-132861169998.77%Courtyard1 12.71-12.15 9634-121251532586.55%Courtyard2 12.15-16.75 12125-4081806990.62%Facade125.08-15.25 1586-2810190197.16%Facade215.25-15.79 2810-2215236896.92%Office110.73-10.69 1348-1277127997.65%Office210.69-10.75 1277-1486135598.97%Trees119.63-19.60 10883-10898954399.41%Trees220.39-20.48 12542-125221125397.87%",
"figure_id": "tab_2",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Experiment results on the KITTI dataset[19] with FCGF[16] descriptors. Bolded and underlined fonts indicate the first two best values.",
"figure_data": "s)",
"figure_id": "tab_3",
"figure_label": "4",
"figure_type": "table"
}
] | Xinyi Li; Yinlong Liu; Hu Cao; Xueli Liu; Feihu Zhang; Alois Knoll | [
{
"authors": "M Zhao; L Ma; X Jia; D.-M Yan; T Huang",
"journal": "IEEE Transactions on Image Processing",
"ref_id": "b0",
"title": "Graphreg: Dynamical point cloud registration with geometry-aware graph signal processing",
"year": "2022"
},
{
"authors": "X Li; Y Liu; V Lakshminarasimhan; H Cao; F Zhang; A Knoll",
"journal": "IEEE Transactions on Intelligent Transportation Systems",
"ref_id": "b1",
"title": "Globally optimal robust radar calibration in intelligent transportation systems",
"year": "2023"
},
{
"authors": "G Blais; M D Levine",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b2",
"title": "Registering multiview range data to create 3d computer objects",
"year": "1995"
},
{
"authors": "J Li",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b3",
"title": "A practical o(n2) outlier removal method for correspondence-based point cloud registration",
"year": "2022"
},
{
"authors": "D Cattaneo; M Vaghi; A Valada",
"journal": "IEEE Transactions on Robotics",
"ref_id": "b4",
"title": "Lcdnet: Deep loop closure detection and point cloud registration for lidar slam",
"year": "2022"
},
{
"authors": "J Zhang; S Singh",
"journal": "IEEE",
"ref_id": "b5",
"title": "Visual-lidar odometry and mapping: Lowdrift, robust, and fast",
"year": "2015"
},
{
"authors": "J Li; P Shi; Q Hu; Y Zhang",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b6",
"title": "Qgore: Quadratic-time guaranteed outlier removal for point cloud registration",
"year": "2023"
},
{
"authors": "H Li; R Hartley",
"journal": "IEEE",
"ref_id": "b7",
"title": "The 3d-3d registration problem revisited",
"year": "2007"
},
{
"authors": "Q.-Y Zhou; J Park; V Koltun",
"journal": "Springer",
"ref_id": "b8",
"title": "Fast global registration",
"year": "2016"
},
{
"authors": "P Besl; N D Mckay",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b9",
"title": "A method for registration of 3-d shapes",
"year": "1992"
},
{
"authors": "J Yang; H Li; D Campbell; Y Jia",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b10",
"title": "Go-icp: A globally optimal solution to 3d icp point-set registration",
"year": "2016"
},
{
"authors": "H Yang; J Shi; L Carlone",
"journal": "IEEE Transactions on Robotics",
"ref_id": "b11",
"title": "Teaser: Fast and certifiable point cloud registration",
"year": "2020"
},
{
"authors": "A P Bustos; T.-J Chin",
"journal": "IEEE transactions on pattern analysis and machine intelligence",
"ref_id": "b12",
"title": "Guaranteed outlier removal for point cloud registration with correspondences",
"year": "2017"
},
{
"authors": "D Barath; J Matas",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b13",
"title": "Graph-cut ransac: Local optimization on spatially coherent structures",
"year": "2021"
},
{
"authors": "R B Rusu; N Blodow; M Beetz",
"journal": "IEEE",
"ref_id": "b14",
"title": "Fast point feature histograms (fpfh) for 3d registration",
"year": "2009"
},
{
"authors": "C Choy; J Park; V Koltun",
"journal": "",
"ref_id": "b15",
"title": "Fully convolutional geometric features",
"year": "2019"
},
{
"authors": "D Borrmann; J Elseberg; A N Üchter",
"journal": "Springer",
"ref_id": "b16",
"title": "Thermal 3d mapping of building fac ¸ades",
"year": "2012"
},
{
"authors": "P W Theiler; J D Wegner; K Schindler",
"journal": "ISPRS journal of photogrammetry and remote sensing",
"ref_id": "b17",
"title": "Keypoint-based 4-points congruent sets-automated marker-less registration of laser scans",
"year": "2014"
},
{
"authors": "A Geiger; P Lenz; R Urtasun",
"journal": "IEEE",
"ref_id": "b18",
"title": "Are we ready for autonomous driving? the kitti vision benchmark suite",
"year": "2012"
},
{
"authors": "B Curless; M Levoy",
"journal": "",
"ref_id": "b19",
"title": "A volumetric method for building complex models from range images",
"year": "1996"
},
{
"authors": "S Huang; Z Gojcic; M Usvyatsov; A Wieser; K Schindler",
"journal": "",
"ref_id": "b20",
"title": "Predator: Registration of 3d point clouds with low overlap",
"year": "2021"
},
{
"authors": "L Yan; P Wei; H Xie; J Dai; H Wu; M Huang",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b21",
"title": "A new outlier removal strategy based on reliability of correspondence graph for fast point cloud registration",
"year": "2022"
},
{
"authors": "H Li",
"journal": "",
"ref_id": "b22",
"title": "Consensus set maximization with guaranteed global optimality for robust geometry estimation",
"year": "2009"
},
{
"authors": "D Campbell; L Petersson; L Kneip; H Li",
"journal": "IEEE transactions on pattern analysis and machine intelligence",
"ref_id": "b23",
"title": "Globally-optimal inlier set maximisation for camera pose and correspondence estimation",
"year": "2018"
},
{
"authors": "H Le; T.-J Chin; A Eriksson; T.-T Do; D Suter",
"journal": "IEEE transactions on pattern analysis and machine intelligence",
"ref_id": "b24",
"title": "Deterministic approximate methods for maximum consensus robust fitting",
"year": "2019"
},
{
"authors": "R I Hartley; F Kahl",
"journal": "International Journal of Computer Vision",
"ref_id": "b25",
"title": "Global optimization through rotation space search",
"year": "2009"
},
{
"authors": "D Campbell; L Petersson",
"journal": "",
"ref_id": "b26",
"title": "Gogma: Globally-optimal gaussian mixture alignment",
"year": "2016"
},
{
"authors": "A P Bustos; T.-J Chin; A Eriksson; H Li; D Suter",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b27",
"title": "Fast rotation search with stereographic projections for 3d registration",
"year": "2016"
},
{
"authors": "J Straub; T Campbell; J P How; J W Fisher",
"journal": "",
"ref_id": "b28",
"title": "Efficient global point cloud alignment using bayesian nonparametric mixtures",
"year": "2017"
},
{
"authors": "Y Liu; C Wang; Z Song; M Wang",
"journal": "",
"ref_id": "b29",
"title": "Efficient global point cloud registration by matching rotation invariant features through translation search",
"year": "2018"
},
{
"authors": "W Chen; H Li; Q Nie; Y.-H Liu",
"journal": "",
"ref_id": "b30",
"title": "Deterministic point cloud registration via novel transformation decomposition",
"year": "2022"
},
{
"authors": "Z Cai; T.-J Chin; A P Bustos; K Schindler",
"journal": "ISPRS journal of photogrammetry and remote sensing",
"ref_id": "b31",
"title": "Practical optimal registration of terrestrial lidar scan pairs",
"year": "2019"
},
{
"authors": "X Li; Y Liu; Y Xia; V Lakshminarasimhan; H Cao; F Zhang; U Stilla; A Knoll",
"journal": "ISPRS Journal of Photogrammetry and Remote Sensing",
"ref_id": "b32",
"title": "Fast and deterministic (3+1)dof point set registration with gravity prior",
"year": "2023"
},
{
"authors": "C Wang; Y Liu; Y Wang; X Li; M Wang",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b33",
"title": "Efficient and outlier-robust simultaneous pose and correspondence determination by branch-and-bound and transformation decomposition",
"year": "2021"
},
{
"authors": "Y Jiao; Y Wang; X Ding; M Wang; R Xiong",
"journal": "IEEE Transactions on Intelligent Transportation Systems",
"ref_id": "b34",
"title": "Deterministic optimality for robust vehicle localization using visual measurements",
"year": "2022"
},
{
"authors": "K Sim; R Hartley",
"journal": "IEEE",
"ref_id": "b35",
"title": "Removing outliers using the L∞ norm",
"year": "2006"
},
{
"authors": "F Kahl; R Hartley",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b36",
"title": "Multiple-view geometry under the L∞-norm",
"year": "2008"
},
{
"authors": "L Peng; M C Tsakiris; R Vidal",
"journal": "",
"ref_id": "b37",
"title": "Arcs: Accurate rotation and correspondence search",
"year": "2022"
},
{
"authors": "B K Horn",
"journal": "Josa a",
"ref_id": "b38",
"title": "Closed-form solution of absolute orientation using unit quaternions",
"year": "1987"
},
{
"authors": "M A Fischler; R C Bolles",
"journal": "Communications of the ACM",
"ref_id": "b39",
"title": "Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography",
"year": "1981"
},
{
"authors": "J Li; Q Hu; M Ai",
"journal": "ISPRS Journal of Photogrammetry and Remote Sensing",
"ref_id": "b40",
"title": "Gesac: Robust graph enhanced sample consensus for point cloud registration",
"year": "2020"
},
{
"authors": "L Sun",
"journal": "IEEE Robotics and Automation Letters",
"ref_id": "b41",
"title": "Ransic: Fast and highly robust estimation for rotation search and point cloud registration using invariant compatibility",
"year": "2021"
},
{
"authors": "D Barath; J Noskova; J Matas",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b42",
"title": "Marginalizing sample consensus",
"year": "2021"
},
{
"authors": "Á P Bustos; T Chin; F Neumann; T Friedrich; M Katzmann",
"journal": "CoRR",
"ref_id": "b43",
"title": "A practical maximum clique algorithm for matching with pairwise constraints",
"year": "2019"
},
{
"authors": "A Myronenko; X Song",
"journal": "IEEE transactions on pattern analysis and machine intelligence",
"ref_id": "b44",
"title": "Point set registration: Coherent point drift",
"year": "2010"
},
{
"authors": "B Jian; B C Vemuri",
"journal": "IEEE transactions on pattern analysis and machine intelligence",
"ref_id": "b45",
"title": "Robust point set registration using gaussian mixture models",
"year": "2010"
},
{
"authors": "Y Liu; G Chen; A Knoll",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b46",
"title": "Globally optimal vertical direction estimation in atlanta world",
"year": "2020"
},
{
"authors": "Y Liu; Y Wang; M Wang; G Chen; A Knoll; Z Song",
"journal": "International Journal of Computer Vision",
"ref_id": "b47",
"title": "Globally optimal linear model fitting with unit-norm constraint",
"year": "2022"
},
{
"authors": "M De Berg; M Van Kreveld; M Overmars; O Schwarzkopf; M Berg; M Van Kreveld; M Overmars; O Schwarzkopf",
"journal": "Springer",
"ref_id": "b48",
"title": "Computational Geometry: Introduction",
"year": "1997"
},
{
"authors": "K Mehlhorn; P Sanders; P Sanders",
"journal": "Springer",
"ref_id": "b49",
"title": "Algorithms and data structures: The basic toolbox",
"year": "2008"
},
{
"authors": "C Choy; W Dong; V Koltun",
"journal": "",
"ref_id": "b50",
"title": "Deep global registration",
"year": "2020"
},
{
"authors": "X Bai; Z Luo; L Zhou; H Chen; L Li; Z Hu; H Fu; C.-L Tai",
"journal": "",
"ref_id": "b51",
"title": "Pointdsc: Robust point cloud registration using deep spatial consistency",
"year": "2021"
},
{
"authors": "R B Rusu; S Cousins",
"journal": "IEEE",
"ref_id": "b52",
"title": "3d is here: Point cloud library (pcl)",
"year": "2011"
},
{
"authors": "Y Zhong",
"journal": "IEEE",
"ref_id": "b53",
"title": "Intrinsic shape signatures: A shape descriptor for 3d object recognition",
"year": "2009"
}
] | [
{
"formula_coordinates": [
2,
144.44,
117.3,
400.51,
20.78
],
"formula_id": "formula_0",
"formula_text": "• E R = 0.069 • E R = 0.004 • - E t = 0.055m E t ="
},
{
"formula_coordinates": [
4,
482.42,
317.53,
81.08,
14.11
],
"formula_id": "formula_1",
"formula_text": "K = {(p i , q i )} N i=1"
},
{
"formula_coordinates": [
4,
379.34,
452.08,
180.97,
18.98
],
"formula_id": "formula_2",
"formula_text": "T * = arg max T ∈SE(3) C (T (P), Q) , (1"
},
{
"formula_coordinates": [
4,
560.31,
454.5,
3.69,
9.14
],
"formula_id": "formula_3",
"formula_text": ")"
},
{
"formula_coordinates": [
4,
343.44,
577.19,
220.56,
29.41
],
"formula_id": "formula_4",
"formula_text": "E(R, t|K, ) = N i=1 I ( Rp i + t -q i ∞ ≤ ) ,(2)"
},
{
"formula_coordinates": [
4,
342.6,
715.54,
217.7,
33.79
],
"formula_id": "formula_5",
"formula_text": "R r X1 r X2 r X3 r Y 1 r Y 2 r Y 3 r Z1 r Z2 r Z3 = r X r Y r Z T . (3"
},
{
"formula_coordinates": [
4,
560.31,
728.67,
3.69,
9.14
],
"formula_id": "formula_6",
"formula_text": ")"
},
{
"formula_coordinates": [
5,
137.13,
72.51,
162.87,
12.62
],
"formula_id": "formula_7",
"formula_text": "t [t X , t Y , t Z ] T .(4)"
},
{
"formula_coordinates": [
5,
95.17,
136.61,
204.83,
226.83
],
"formula_id": "formula_8",
"formula_text": "Rp i + t -q i ∞ ≤ (5a) ⇔ r T X r T Y r T Z p i + t X t Y t Z - q X i q Y i q Z i ∞ ≤ (5b) ⇔ max r T X p i + t X -q X i , r T Y p i + t Y -q Y i , r T Z p i + t Z -q Z i ≤ (5c) ⇔ r T X p i + t X -q X i ≤ , r T Y p i + t Y -q Y i ≤ , r T Z p i + t Z -q Z i ≤ (5d) ⇔ I r T X p i + t X -q X i ≤ = 1, I r T Y p i + t Y -q Y i ≤ = 1, I r T Z p i + t Z -q Z i ≤ = 1(5e)"
},
{
"formula_coordinates": [
5,
91.31,
445.79,
205,
29.41
],
"formula_id": "formula_9",
"formula_text": "E(R, t|K, ) = N i=1 I L X i ∧ L Y i ∧ L Z i , (6"
},
{
"formula_coordinates": [
5,
296.31,
456.08,
3.69,
9.14
],
"formula_id": "formula_10",
"formula_text": ")"
},
{
"formula_coordinates": [
5,
49.67,
639,
250.33,
38.61
],
"formula_id": "formula_11",
"formula_text": "E j (r j , t j |K, ) = N i=1 I r T j p i + t j -q j i ≤ , j = X, Y, Z.(7)"
},
{
"formula_coordinates": [
5,
48,
725.83,
252,
21.19
],
"formula_id": "formula_12",
"formula_text": "E X (r X , t X |K, ), max E Y (r Y , t Y |K, ), and max E Z (r Z , t Z |K, )."
},
{
"formula_coordinates": [
5,
377.06,
326.73,
186.94,
11.72
],
"formula_id": "formula_13",
"formula_text": "S 2+ = r|r T r = 1, r 3 ≥ 0 ,(8)"
},
{
"formula_coordinates": [
5,
494.62,
448.27,
69.38,
12.11
],
"formula_id": "formula_14",
"formula_text": "d = γ d(9)"
},
{
"formula_coordinates": [
5,
342.49,
470.3,
221.51,
12.11
],
"formula_id": "formula_15",
"formula_text": "γ ∈ [0, π/2], dT = d/ d is a unit vector in R 2 ."
},
{
"formula_coordinates": [
5,
375.61,
536.59,
188.4,
12.11
],
"formula_id": "formula_16",
"formula_text": "-r T = -sin(γ) dT , cos(γ) .(10)"
},
{
"formula_coordinates": [
5,
388.39,
737.34,
171.65,
9.68
],
"formula_id": "formula_17",
"formula_text": "∠(r a , r b ) ≤ d a -d b . (11"
},
{
"formula_coordinates": [
5,
560.04,
737.72,
3.96,
9.14
],
"formula_id": "formula_18",
"formula_text": ")"
},
{
"formula_coordinates": [
6,
112.81,
354.32,
187.19,
18.2
],
"formula_id": "formula_19",
"formula_text": "∠ (r, r c ) ≤ d -d c ≤ √ 2σ,(12)"
},
{
"formula_coordinates": [
6,
330.15,
349.55,
114.63,
13.65
],
"formula_id": "formula_20",
"formula_text": "F (x|A) = M i=1 F i (x, a i ),"
},
{
"formula_coordinates": [
6,
323.32,
404.2,
236.72,
38.61
],
"formula_id": "formula_21",
"formula_text": "max x F (x|A) = max x M i=1 F i (x, a i ) ≤ M i=1 max x F i (x, a i ) . (13"
},
{
"formula_coordinates": [
6,
560.04,
433.67,
3.96,
9.14
],
"formula_id": "formula_22",
"formula_text": ")"
},
{
"formula_coordinates": [
6,
312,
465.81,
252,
33.22
],
"formula_id": "formula_23",
"formula_text": "F i (x, a i ) ≤ max x F i (x, a i ) ≤ 1. Therefore, it is obvious that the maximum of M i=1 F i (x, a i ) is not bigger than the sum of max x F i (x, a i )."
},
{
"formula_coordinates": [
6,
312,
549.14,
252,
23.79
],
"formula_id": "formula_24",
"formula_text": "d c ∈ R 2 (corre- sponds to r c j ∈ S 2+"
},
{
"formula_coordinates": [
6,
320.27,
590.09,
243.73,
29.41
],
"formula_id": "formula_25",
"formula_text": "E + (D) = max tj N i=1 I t j ∈ t i- j , t i+ j ,(14a)"
},
{
"formula_coordinates": [
6,
320.27,
617.5,
243.73,
63.87
],
"formula_id": "formula_26",
"formula_text": "t i- j = --p i cos max ∠ r c j , p i - √ 2σ, 0 + q j i , (14b) t i+ j = -p i cos min ∠ r c j , p i + √ 2σ, π + q j i .(14c)"
},
{
"formula_coordinates": [
6,
374.05,
720.93,
189.95,
29.41
],
"formula_id": "formula_27",
"formula_text": "max rj ,tj N i=1 L j i = max tj max rj N i=1 L j i .(15)"
},
{
"formula_coordinates": [
7,
53.36,
67.35,
242.94,
38.61
],
"formula_id": "formula_28",
"formula_text": "max tj max rj N i=1 L j i ≤ max tj N i=1 max rj I r T j p i + t j -q j i ≤ .(16"
},
{
"formula_coordinates": [
7,
100.85,
153.67,
199.15,
43.81
],
"formula_id": "formula_29",
"formula_text": "∠ (r j , p i ) ≤ ∠ r c j , p i + ∠ r c j , r j (17a) ≤ ∠ r c j , p i + α (17b) ≤ ∠ r c j , p i + √ 2σ,(17c)"
},
{
"formula_coordinates": [
7,
100.85,
231.74,
199.15,
43.81
],
"formula_id": "formula_30",
"formula_text": "∠ (r j , p i ) ≥ ∠ r c j , p i -∠ r c j , r j (18a) ≥ ∠ r c j , p i -α (18b) ≥ ∠ r c j , p i - √ 2σ.(18c)"
},
{
"formula_coordinates": [
7,
61.37,
318.13,
238.63,
40.16
],
"formula_id": "formula_31",
"formula_text": "r T j p i ∈ p i cos min ∠ r c j , p i + √ 2σ, π , p i cos max ∠ r c j , p i - √ 2σ, 0 .(19)"
},
{
"formula_coordinates": [
7,
60.77,
408.65,
239.23,
58.69
],
"formula_id": "formula_32",
"formula_text": "max rj I r T j p i + t j -q j i ≤ (20a) = max rj I --r T j p i + q j i ≤ t j ≤ -r T j p i + q j i (20b) ≤I t j ∈ t i- j , t i+ j ,(20c)"
},
{
"formula_coordinates": [
7,
56.27,
497.32,
243.73,
32.34
],
"formula_id": "formula_33",
"formula_text": "t i- j = --p i cos max ∠ r c j , p i - √ 2σ, 0 + q j i ,(21a)"
},
{
"formula_coordinates": [
7,
56.38,
528.85,
243.62,
32.34
],
"formula_id": "formula_34",
"formula_text": "t i+ j = -p i cos min ∠ r c j , p i + √ 2σ, π + q j i .(21b)"
},
{
"formula_coordinates": [
7,
48,
595.18,
252,
92.46
],
"formula_id": "formula_35",
"formula_text": "N i=1 max rj I r T j p i + t j -q j i ≤ ≤ N i=1 I t j ∈ t i- j , t i+ j (22) Finally, we have max rj ,tj N i=1 L j i ≤ max tj N i=1 I t j ∈ t i- j , t i+ j (23"
},
{
"formula_coordinates": [
7,
296.04,
668.52,
3.96,
9.14
],
"formula_id": "formula_36",
"formula_text": ")"
},
{
"formula_coordinates": [
7,
236.32,
43.4,
159.76,
703.26
],
"formula_id": "formula_37",
"formula_text": "d c ∈ R 2 (corre- sponds to -r c j ∈ S 2-"
},
{
"formula_coordinates": [
7,
320.71,
71.85,
243.29,
29.41
],
"formula_id": "formula_38",
"formula_text": "E -(D) = max tj N i=1 I t j ∈ t i- j , t i+ j ,(24a)"
},
{
"formula_coordinates": [
7,
320.71,
99.26,
243.29,
32.34
],
"formula_id": "formula_39",
"formula_text": "t i- j = -+ p i cos min ∠ r c j , p i + √ 2σ, π + q j i ,(24b)"
},
{
"formula_coordinates": [
7,
320.71,
130.79,
243.29,
32.34
],
"formula_id": "formula_40",
"formula_text": "t i+ j = + p i cos max ∠ r c j , p i - √ 2σ, 0 + q j i .(24c)"
},
{
"formula_coordinates": [
7,
370.02,
505.1,
193.98,
13.69
],
"formula_id": "formula_41",
"formula_text": "E(D) = max E + (D), E -(D) .(25)"
},
{
"formula_coordinates": [
7,
332.32,
634.84,
231.68,
11.79
],
"formula_id": "formula_42",
"formula_text": "E(D) = max E + (D), E -(D) ,(26a)"
},
{
"formula_coordinates": [
7,
332.32,
651.25,
231.68,
29.41
],
"formula_id": "formula_43",
"formula_text": "E + (D) = N i=1 I r c j T p i + t + j -q j i ≤ ,(26b)"
},
{
"formula_coordinates": [
7,
332.32,
684.1,
231.68,
29.41
],
"formula_id": "formula_44",
"formula_text": "E -(D) = N i=1 I -r c j T p i + t - j -q j i ≤ ,(26c)"
},
{
"formula_coordinates": [
8,
60.24,
708.24,
239.76,
38.61
],
"formula_id": "formula_45",
"formula_text": "S j (r j , t j |P, Q, ) = M i=1 max k I r T j p i + t j -q j k ≤ ,(27)"
},
{
"formula_coordinates": [
8,
314.9,
295.84,
215.44,
56.53
],
"formula_id": "formula_46",
"formula_text": "Update U = maxE(D k ), D k ∈ ξ; 11 Update L = max {L, E(D k )} with D k ∈ ξ, if E(D k ) > L, set r * j = δ(D k ) and t * j = η(D k ); 12 Eliminate D k from ξ if E(D k ) < L, D k ∈ ξ; 13 end"
},
{
"formula_coordinates": [
8,
317.9,
525.74,
246.1,
29.41
],
"formula_id": "formula_47",
"formula_text": "E + SP CR (D) = max tj M i=1 max k I t j ∈ t ik- j , t ik+ j ,(28a)"
},
{
"formula_coordinates": [
8,
317.9,
553.16,
246.1,
32.34
],
"formula_id": "formula_48",
"formula_text": "t ik- j = --p i cos max ∠ r c j , p i - √ 2σ, 0 + q j k ,(28b)"
},
{
"formula_coordinates": [
8,
317.9,
584.68,
242.09,
32.34
],
"formula_id": "formula_49",
"formula_text": "t ik+ j = -p i cos min ∠ r c j , p i + √ 2σ, π + q j k . (28c"
},
{
"formula_coordinates": [
8,
559.99,
607.88,
4.01,
9.14
],
"formula_id": "formula_50",
"formula_text": ")"
},
{
"formula_coordinates": [
8,
318.34,
652.11,
245.66,
29.41
],
"formula_id": "formula_51",
"formula_text": "E - SP CR (D) = max tj M i=1 max k I t j ∈ t ik- j , t ik+ j ,(29a)"
},
{
"formula_coordinates": [
8,
318.34,
679.53,
245.66,
32.34
],
"formula_id": "formula_52",
"formula_text": "t ik- j = -+ p i cos min ∠ r c j , p i + √ 2σ, π + q j k ,(29b)"
},
{
"formula_coordinates": [
8,
318.34,
711.05,
241.65,
32.34
],
"formula_id": "formula_53",
"formula_text": "t ik+ j = + p i cos max ∠ r c j , p i - √ 2σ, 0 + q j k . (29c"
},
{
"formula_coordinates": [
8,
559.99,
734.25,
4.01,
9.14
],
"formula_id": "formula_54",
"formula_text": ")"
},
{
"formula_coordinates": [
9,
68.54,
62.89,
231.46,
14.82
],
"formula_id": "formula_55",
"formula_text": "E SP CR (D) = max E + SP CR (D), E - SP CR (D) .(30)"
},
{
"formula_coordinates": [
9,
66.96,
375.76,
233.04,
13.23
],
"formula_id": "formula_56",
"formula_text": "E SP CR (D) = max E + SP CR (D), E - SP CR (D) ,(31a)"
},
{
"formula_coordinates": [
9,
66.96,
392.16,
233.04,
29.41
],
"formula_id": "formula_57",
"formula_text": "E + SP CR (D) = M i=1 I M j+ i > 0 ,(31b)"
},
{
"formula_coordinates": [
9,
66.96,
425.01,
233.04,
29.64
],
"formula_id": "formula_58",
"formula_text": "M j+ i = N k=1 I r c j T p i + t + j -q j k ≤ ,(31c)"
},
{
"formula_coordinates": [
9,
66.96,
458.09,
233.04,
29.41
],
"formula_id": "formula_59",
"formula_text": "E - SP CR (D) = M i=1 I M j- i > 0 ,(31d)"
},
{
"formula_coordinates": [
9,
66.96,
490.94,
233.04,
29.64
],
"formula_id": "formula_60",
"formula_text": "M j- i = N k=1 I -r c j T p i + t - j -q j k ≤ ,(31e)"
},
{
"formula_coordinates": [
10,
364.98,
628.21,
194.9,
25.35
],
"formula_id": "formula_61",
"formula_text": "E R = arccos T r(R -1 gt R * ) -1 2 , (32a"
},
{
"formula_coordinates": [
10,
559.88,
638.33,
4.12,
9.14
],
"formula_id": "formula_62",
"formula_text": ")"
},
{
"formula_coordinates": [
10,
364.98,
658.64,
194.8,
11.72
],
"formula_id": "formula_63",
"formula_text": "E t = t gt -t * , (32b"
},
{
"formula_coordinates": [
10,
559.78,
661.05,
4.22,
9.14
],
"formula_id": "formula_64",
"formula_text": ")"
}
] |
|
2023-05-19 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b29",
"b42",
"b6",
"b23",
"b43",
"b13",
"b31",
"b0",
"b29",
"b42",
"b32",
"b3"
],
"table_ref": [],
"text": "The vision community has witnessed the rapid progress of deep generative models, pushing image generation quality to an unprecedented level. As a fundamental task, generating realistic images from arbitrary inputs (e.g., class labels) can empower humans to create rich and diverse visual content and bring numerous real-world applications. Unify-Figure 1. Illustration of our motivation. (a) Existing fixed-length coding ignores information densities, which results in insufficiency in dense information regions like region ② and redundancy in sparse information regions like region ①, generating poor details and inconsistent structure. Our information-density-based variable-length coding encodes accurately and produces rich details and consistent structure. (b) Comparison of existing unnatural raster-scan autoregressive generation order and our natural and more effective coarse-to-fine autoregressive generation order. crete codes, where each code represents a local visual pattern, while the second stage learns to generate codes of local regions and then restores to images. The importance lies in that the local details could be well encoded in the first stage and thus the second stage could effectively focus on global structure modeling, leading to better generation quality and scalability. Existing models mainly focus on the second stage to better generate codes for improving generation quality, such as raster-scan autoregression [11,30,43], bi-direction [7,24,44], or diffusion [5,14,32]. Only a few works aim to improve the fundamental code representation itself in the first stage, including perceptual and adversarial loss for context-rich codebook [13], residual quantization [23], and more expressive transformer backbone [42], etc. Their commonality is that they all focus on encoding more information of all image regions together.\nHowever, existing fundamental encoding works inherently fail to effectively encode image information for an accurate and compact code representation, because they ignore the naturally different information densities of different image regions and encode fixed-size regions into fixed-length codes. As a result, they suffer from two limitations: (1) insufficient coding for important regions with dense information, which fails to encode all necessary information for faithful reconstruction and therefore degrades the realism of local details in both stages. (2) redundant coding for unimportant ones with sparse information, bringing huge redundant codes that mislead the second stage to focus on the redundancy and therefore significantly hinder the global structure modeling on important ones. As shown in Figure 1(a), the fixed-length codes result in large reconstruction errors in important cheetah regions and produce poor local details (e.g., face, hair) in both stages. Meanwhile, the fixed-length codes are overwhelmed for unimportant background regions, which misleads the second stage to generate redundant background and inconsistent cheetah structure. Moreover, as shown in Figure 1(b), since all regions are encoded into fixed-length codes, there is no way for the second stage to distinguish their varying importance and thus results in an unnatural raster-scan order [13] for existing autoregressive models [11,23,30,42,43], which fails to consider the image content for an effective generation.\nTo address this problem, inspired by the classical information coding theorems [18,33,34] and their dynamic coding principle, we propose information-density-based variable-length coding for an accurate and compact code representation to improve generation quality and speed. Moreover, we further propose a natural coarse-to-fine autoregressive model for a more effective generation. Specifically, we propose a novel two-stage generation framework: (1) Dynamic-Quantization VAE (DQ-VAE) which first constructs hierarchical image representations of multiple candidate granularities for each region, and then uses a novel Dynamic Grained Coding module to assign the most suitable granularity for each region under the constraint of a proposed budget loss, matching the percentage of each granularity to the desired expectation holistically. (2) DQ-Transformer which thereby generates images autoregressively from coarse-grained (smooth regions with fewer codes) to fine-grained (details regions with more codes) to more effectively achieve consistent structures. Considering the distribution of different granularities varying, DQ-Transformer models the position and content of codes in each granularity alternately through a novel stacked-transformer architecture. To effectively teach the difference between different granularities, we further design shared-content and non-shared-position input layers.\nOur main contributions are summarized as follows:\nConceptual contribution. We point to the inherent insufficiency and redundancy in existing fixed-length coding since they ignore information density. For the first time, we propose information-density-based variablelength coding for accurate & compact code representations.\nTechnical contribution.\n(1) We propose DQ-VAE to dynamically assign variable-length codes to regions based on their different information densities through a novel Dynamic Grained Coding module and budget loss. (2) We propose DQ-Transformer to generate images autoregressively from coarse-grained to fine-grained for the first time, which models the position and content of codes alternately in each granularity by stacked-transformer architecture with shared-content and non-shared position input layers design.\nExperimental contribution. Comprehensive experiments on various generations validate our superiority, e.g., we achieve 7.4% quality improvement and faster speed compared to existing state-of-the-art autoregressive model on unconditional generation, and 17.3% quality improvement compared to existing million-level parameters stateof-the-art models on class-conditional generation."
},
{
"figure_ref": [],
"heading": "Related Works",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Vector Quantization for Image Generation",
"publication_ref": [
"b31",
"b29",
"b42",
"b0",
"b13",
"b31",
"b44",
"b6",
"b23",
"b43",
"b13",
"b29",
"b30",
"b36",
"b42",
"b43"
],
"table_ref": [],
"text": "Existing VQ-based models follow a two-stage paradigm that first learns a codebook to encode images into discrete space and then models the underlying distribution in this discrete space. The VQ-based paradigm has attracted increasing interest and is adopted by most milestone generative models, such as latent diffusion [32], DALL-E [30], Parti [43], etc. Most works focus on the second stage for better learning in the discrete space, such as discrete diffusion [1,14,32,36,45], bidirection [5,7,24,44], and the most popular raster-scan autoregression [11,13,14,23,30,31,37,43]. Only a few works aim to improve the fundamental encoding, e.g., VQGAN [13] introduces perceptual and adversarial loss for a context-rich codebook. [23] intro- duces residual-quantization. [42] proposes a more expressive transformer backbone. Recently, [44] proposes to insert spatially variant information. However, existing fixedlength coding ignores information density and is thus limited by insufficiency and redundancy. For the first time, we propose information-density-based variable-length coding and a more natural coarse-to-fine autoregression."
},
{
"figure_ref": [],
"heading": "Dynamic Network",
"publication_ref": [
"b14",
"b38",
"b5",
"b16",
"b24",
"b4",
"b40"
],
"table_ref": [],
"text": "Designing dynamic architectures is an effective approach for efficient deep learning and yields better representation power and generality [15]. Literately, current research can be mainly categorized into three directions, i.e., dynamic depth for network early exiting [4] or layer skipping [39], dynamic width for skipping neurons [3] or channels [26] and dynamic routing for multi-branch structure networks [17,25,35,41]. Our work belongs in the last direction. To the best of our knowledge, the dynamic network has never been studied in VQ-based generation and we present the first work to realize the variable-length coding of classical information coding theorems through the dynamic network."
},
{
"figure_ref": [
"fig_0"
],
"heading": "Methodology",
"publication_ref": [],
"table_ref": [],
"text": "Our overall two-stage framework is depicted in Figure 2. In the following, we will first briefly revisit the formulation of VQ and then describe our proposed method in detail."
},
{
"figure_ref": [],
"heading": "Preliminary",
"publication_ref": [
"b36"
],
"table_ref": [],
"text": "Vector Quantization (VQ) [37] denotes the technique that learns a codebook to encode images into discrete code representations. Formally, the codebook is defined as\nC := {(k, e(k))} k∈[K]\n, where K is the codebook size and n z is the dimension of codes. An image X ∈ R H0×W0×3 is first encoded into grid features Z = E(X) ∈ R H×W ×nz by the encoder E, where (H, W ) = (H 0 /f, W 0 /f ) and f is the downsampling factor. For each vector z ∈ R nz in Z, the quantization operation Q(•) replaces it with the code that has the closest euclidean distance with it in the codebook C:\nQ(z; C) = arg min k∈[K] ||z -e k || 2 2 .(1)\nHere, Q(z; C) is the quantized code. z q = e(Q(z; C)) is the quantized vector. Therefore, the quantized encoded fea-tures are Z q ∈ R H×W ×nz . The decoder D is used to reconstruct the original image by X = D(Z q ). Here each code roughly represents a fixed f 2 size visual pattern and each image region is represented by the same length of codes without distinguishing their different information densities. As a result, existing works suffer from both insufficiency in important regions and redundancy in unimportant ones."
},
{
"figure_ref": [
"fig_1"
],
"heading": "Stage 1:Dynamic-Quantization VAE(DQ-VAE)",
"publication_ref": [
"b45"
],
"table_ref": [],
"text": "Different from existing works that adopt a fixed downsampling factor f to represent image regions as fixed-length codes, DQ-VAE first defines a set of candidates:\nF = {f 1 , f 2 , ..., f K }, where f 1 < f 2 < ... < f K , (2)\nand encodes images into hierarchical features\nZ = {Z 1 , Z 2 , ..., Z K } through a hierarchical encoder E h , where Z i ∈ R Hi×Wi×nz and (H i , W i ) = (H 0 /f i , W 0 /f i ) for each i ∈ {1, 2, ..., K}.\nThe image region's size is set as the maximum downsampling factor, i.e., S = f K , and therefore each S 2 size image region now has multiple granularity representations containing different numbers of features. Then the Dynamic Grained Coding (DGC) module assigns the most suitable granularity for each region and results in multi-grained representations, which are further quantized by VQ. To deal with the irregular code map that different regions have different numbers of codes, we further propose a simple but effective nearest-neighbor replication, that is, in each region the quantized codes are replicated to the code number of the finest granularity if the finest granularity is not assigned for it, resulting in a regular code map that could be conveniently decoded by the convolutional decoder D.\nDynamic Grained Coding (DGC) module. As illustrated in Figure 3, given the encoded hierarchical image features Z = {Z 1 , Z 2 , ..., Z K }, we implement a discrete gating network with Gumbel-Softmax technique [19] to determine the granularity for each image region. Specifically, each granularity feature is first normed by groupnormalization to stabilize training and then pooled to the size of the coarsest granularity feature by average-pooling, except the coarsest granularity (i.e., f K ) feature itself. The pooled features are denoted as\nZ ′ = {Z ′ 1 , Z ′ 2 , ..., Z ′ K } and Z ′ i ∈ R Hs×Ws×nz for i ∈ {1, 2, ..., K}, where (H s , W s ) = (H 0 /f K , W 0 /f K ).\nThe gating logits G are generated as:\nG = (Z ′ 1 ∥Z ′ 2 ∥...∥Z ′ K )W g ∈ R Hs×Ws×k , (3\n)\nwhere ∥ is the concatenation operation along the channel dimension and W g ∈ R (K×nz)×K is the learnable weight.\nFor each region (i, j), its gating logits g i,j ∈ R K is used to decide the granularity by calculating the gating index: To enable the end-to-end training of this discrete process, inspired by [40,46], the determined decisions in Eq.( 4) are replaced with the stochastic sampling process. Mathematically, given a categorical distribution with unnormalized log probabilities, discrete gating indices can be yielded with noise samples drawn from a standard Gumbel distribution:\nθ i,j = arg max k (g i,j,k ) ∈ {1, 2, ..., K}.(4)\nθi,j = arg max k (g i,j,k + n k ), where n k ∼ Gumbel(0,1). (5)\nTo enable the back-propagation of the above hard decision, we adopt the Gumbel-Softmax technique [19] to give a continuous and differentiable approximation by replacing the argmax with a softmax operation. The soft gating score p i,j for each region is then selected by the gating indices:\np i,j = exp((g i,j,θi,j + n θi,j ))/τ K k exp((g i,j,k + n k )/τ ) ∈ [0, 1],(6)\nwhere the temperature τ = 1. We use a straight-through estimator for the gradients of gating logits, which are obtained through the soft gating score p i,j during the backward pass.\nThe above stochastic process is only adopted during training and no random sampling is required during inference. Budget Loss. We adopt the training loss of VQGAN [13] as L vanilla , which includes reconstruction loss (l 1 loss, perceptual loss, adversarial loss) and quantization loss. In the absence of a budget constraint, the DGC module typically prefers to assign the finest granularity for all image regions, which is in contrast to our purpose. Therefore, we further propose a budget loss to match the percentage of each granularity to our desired expectation. Specifically, we denote the desired ratio of each granularity k as r k and K k r k = 1. For an image sample whose current assigned ratio of each granularity k is r ′ k , we define budget loss as:\nL budget = K-1 k (r k -r ′ k ) 2 , (7\n)\nwhere we only calculate on K -1 granularities since the ratio of the last granularity is determined by 1 -\nK-1 k\nr k . The final loss for DQ-VAE is defined as:\nL stage1 = L vanilla + λL budget ,(8)\nwhere λ is a loss balance hyper-parameter. The expected ratio of each granularity is holistic on the dataset level. Therefore, since important regions contribute more to the reconstruction quality, the variable-length coding is realized from two aspects, i.e., inter-dynamic, longer code sequence for complex images while shorter code sequence for easy ones; intra-dynamic, for each image, more codes for important regions while fewer codes for unimportant ones."
},
{
"figure_ref": [
"fig_0"
],
"heading": "Stage 2: DQ-Transformer",
"publication_ref": [
"b37"
],
"table_ref": [],
"text": "Different images share different perceptually important regions and different complexities. Therefore, DQ-VAE encodes images as the code sequence of variable lengths and the distribution of each granularity region in images is also completely different. Though learning this dynamic underlying prior is very challenging, it also opens a promising potential for autoregressive image generation, that is, a natural and more effective coarse-grained to fine-grained generation order since DQ-VAE naturally divides coarse regions (smooth regions with fewer codes) apart from fine regions (details regions with more codes). Imagine image generation as a jigsaw puzzle problem, it is more effective and efficient that we first fill in the large and easy pieces (coarse regions) and then fill in the small and difficult ones (fine regions). With this motivation, DQ-Transformer first constructs the codes' content and position sequence in each granularity separately and then concatenates them in a coarse-to-fine manner to autoregressively predict the next code's position and content through the stacked Position-Transformer and Content-Transformer. The distinction of different granularities is realized by the shared content, non-shared-position, and granularity input layers designs.\nTraining sequence construction. As illustrated in stage 2(a) in Figure 2, the sequence of each granularity is constructed separately. As for the content sequence, each index is the quantized code index. As for the position sequence, each index is the position of the corresponding code index in the position map of current granularity. We add a special <sos> code at the beginning of all content and position sequences to indicate the start of the sequence, and another special <eos> code at the end of them to indicate the end of the sequence. To enable batch training and sampling, we use a special <pad> code to pad all samples to the same length in each granularity. Finally, we concatenate all granularities' content and position sequences in a coarse-to-fine manner, which we denote as C and P , respectively.\nPosition-Transformer. We first learn to predict the next code position conditioned on all previous codes and their positions. The input of Position-Transformer consists of four parts: (1) content embedding which is calculated from C by a shared-content layer for all granularities, (2) position embedding which is calculated from P by non-sharedposition layers for each granularity separately, (3) granular-ity embedding which is used for distinguishing each granularity, and (4) a learned absolute position embedding for making the network aware of the absolute position of the sequence, which is the same as most transformer-architecture [13,36,38]. After processing by Position-Transformer, the output hidden vector H p encodes both code and their position information and is used for next position predicting. The negative log-likelihood (NLL) loss for the next code position autoregressive training is:\nL position = E(-log p(P l |P <l , C <l ))(9)\nContent-Transformer. We then learn to predict the next code's content conditioned on all previous codes and the position of current code. Specifically, The input of Content-Transformer is two parts: (1) the output of Position-Transformer H p and (2) the ground-truth information of the current position which also is calculated by the non-sharedposition layers. For example, if the input position sequence for Position-Transformer is P [0:-2] , then the input groundtruth position sequence for Content-Transformer is P [1:-1] . The negative log-likelihood (NLL) loss for the next code's content autoregressive training is:\nL content = E(-log p(C l |P ≤l , C <l ))(10)\nTraining & Inference. During training, the total loss for DQ-Transformer is defined as:\nL stage2 = L position + L content .(11)\nOur proposed DQ-Transformer is a general visual generative model which could be easily extended to various other generation tasks. As for the class-conditional generation, we replace the <sos> code in the content sequence of each granularity with the class-label code. During inference, we could also autoregressively generate images from coarsegrained to fine-grained, as illustrated in Algorithm 1, where we take the unconditional generation as an example and other conditional generations can be derived accordingly."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [
"b43",
"b15",
"b15"
],
"table_ref": [],
"text": "Benchmarks. We evaluate our method on unconditional FFHQ [20] benchmark and class-conditional ImageNet [9] benchmark with 256 × 256 image resolution.\nMetrics. Following previous works [13,23,44], the standard Fréchet Inception Distance (FID) [16] is adopted for evaluating the generation and reconstruction quality (denoted as rFID). rFID is calculated over the entire test set. Inception Score (IS) [16] is also adopted for class-conditional generation on the ImageNet benchmark.\nImplementation. DQ-VAE follows the architecture of VQGAN [13] except for the lightweight DRC module, which is trained with the codebook size K = 1024 and λ = The initial empty position (code) sequence P (C). Output: The generated image I.\n1:\nfor each k ∈ [1, K] do 2:\n// sample each granularity in a coarse-to-fine order 3:\nP = concat(P, <sos>), C = concat(C, <sos>) 4:\nwhile NOT all samples have sampled <eos> do // if sampled <eos>, the following will only can be <pad> for this sample in current granularity "
},
{
"figure_ref": [],
"heading": "Comparison with state-of-the-art methods",
"publication_ref": [],
"table_ref": [
"tab_0",
"tab_3",
"tab_1",
"tab_2"
],
"text": "The main results are reported on dual granularities of F = {8, 16}, and the ratio r f =8 = 0.5 (640 average length).\nUnconditional generation. As shown in Table 1, our model outperform all existing autoregressive state-of-theart models including the strongest large-scale ViT-VQGAN [42] by a 7.4% quality improvement. We compare with other types of state-of-the-art models in Table 4 and also achieve top-level performance. The qualitative results of unconditional generation are shown on the left of Figure 4.\nClass-conditional generation. into Million/Billion according to whether they can be trained under normal computing resources (i.e., 24G memory 3090). We first compare with all million-level parameters state-of-the-art in Table 2. Our model with 355M parameters already outperforms all autoregressive and diffusion models. Moreover, our model with 655M outperforms GAN-based and bi-direct state-of-the-art, which demonstrates our effectiveness and scalability. We further compare with large-scale billion-level state-of-the-art in Table 3, where we achieve top-level performance with fewer parameters. The qualitative results of class-conditional generation are shown on the right of Figure 4."
},
{
"figure_ref": [
"fig_5",
"fig_5"
],
"heading": "Ablations & Analysis",
"publication_ref": [
"b0"
],
"table_ref": [
"tab_5",
"tab_7",
"tab_8"
],
"text": "Analysis on DQ-VAE. We first demonstrate that our variable-length coding has better reconstruction compared to the existing fixed-length one in Table 5. We take VQ-Figure 5. Visualization of the variable-length coding of our DQ-VAE, where our coding map exactly matches the error map of VQGAN and therefore leads to better reconstruction quality, i.e., the information-dense regions where VQGAN has higher reconstruction error are assigned to more codes, while information-sparse regions where VQGAN has lower reconstruction error are assigned to few codes. Here F denotes the granularity candidates set. \"ratio\" denotes the ratio of each granularity. We show that variable-length coding could bring better reconstruction compared to fixed-length coding on the same code length.\nr f =32 = 4 × r f =8 ,(12)\nunimportant ones since they are less informative. The phenomenon also reveals that existing fixed-length coding is both insufficient in important regions and redundant in unimportant ones.\n(2) When we improperly increase r f =8 , we get a larger r f =32 which will inevitably assign some important regions with fewer codes and thus degrade the reconstruction quality. (3) Moreover, DQ-VAE's adaptive assignment significantly outperforms the random one (ours 4.08 vs. random's 7.32) which demonstrates that DQ-VAE could distinguish important regions from unimportant ones.\nWe then analyze the impact of different ratio percentages in Table 6, where DQ-VAE adopts dual granularities of F = {8, 16}. We show that: (1) The mean code length of each ratio matches the expectation well, which validates our proposed budget loss. (2) The results are consistent with the Pareto principle, which is also known as 20/80 laws. To be specific, when increasing r f =8 from 0 to 0.3, we get 1.44 FID improvement while only a slight codebook usage drop, which indicates that the first 30% percentage important regions contribute the most valid information of images and existing fixed-length coding is insufficient in them. Mean- while, when increasing r f =8 from 0.7 to 1.0, we only get a subtle 0.21 FID improvement but a significant 9.6% codebook usage drop, which indicates that the last 30% percentage unimportant regions contribute little valid information of images but most redundancy. The experimental results strongly support our motivations for variable-length coding to get rid of insufficiency and redundancy simultaneously. We visualize our variable-length coding on ImageNet in Figure 5, where DQ-VAE adopts dual granularities of F = {8, 16} and r f =8 = 0.3. The error map is calculated by l 1 loss of each 16 2 size region between images and VQGAN (f = 16) reconstructions. The red regions in our coding map are assigned to f = 8 (4 codes) while the blue ones are assigned to f = 16 (1 code). We show that our coding map matches VQGAN's error map, i.e., important regions are assigned to more codes and unimportant ones are assigned to few codes, leading to better reconstruction quality.\nAnalysis on the effectiveness of DQ-Transformer. We first validate our input layers designs in Table 7. The nonshared-position and granularity layers are very important since they distinguish different granularities. Without these designs, DQ-Transformer fails to know which granularity of code should be generated next, and thus performs worse.\nWe then analyze the generation quality of different ratios in Figure 6 left. The generation speed of autoregressive models mostly depends on their code length. The Pareto curve shows that the generation quality (FID) saturates when r f =8 reaches 0.5. The experimental phenomenon reveals that a proper ratio is important for the unity of a high generation quality and fast generation speed since it guarantees effective coding in both important regions and unim- Analysis on the efficiency of DQ-Transformer. We compare our generation speed to the existing state-of-the-art autoregressive model ViT-VQGAN [42] according to different batch sizes in Figure 6 right. The generation speeds are evaluated on a single RTX-3090 GPU and the setup of ViT-VQGAN is implemented the same as its original paper. Our model achieves a much faster generation speed for all batch sizes which validates the efficiency brought by our accurate and compact code representation."
},
{
"figure_ref": [],
"heading": "Conclusion & Future Direction",
"publication_ref": [
"b29",
"b42",
"b43",
"b13",
"b31",
"b6",
"b26"
],
"table_ref": [],
"text": "In this study, we point out that the existing fixed-length coding ignores the naturally different information densities of image regions and is inherently limited by insufficiency and redundancy, which degrades generation quality and speed. Moreover, the fixed-length coding brings an unnatural raster-scan autoregression. We thereby propose a novel two-stage generation framework: (1) DQ-VAE which dynamically assigns variable-length codes to regions based on their information densities for an accurate and compact code representation. (2) DQ-Transformer which then models the position and content of codes alternately, generating images autoregressively in a more natural and effective coarse-to-fine order for the first time. To effectively teach the difference between different granularities, we further design shared-content, non-shared-position, and granularity input layers. Comprehensive experiments on various image generations validate our effectiveness and efficiency.\nFuture Direction. VQ is the foundation for modern autoregressive [11,23,30,43,44], discrete diffusion [14,32], and bidirectional [7] generation, and even pretraining [2,27]. Our study validates the effectiveness and efficiency of the variable-length coding for autoregressive generation, but its great potential for diffusion, bi-direction, and pretraining is worth further exploration in the future."
},
{
"figure_ref": [],
"heading": "Acknowledgments",
"publication_ref": [],
"table_ref": [],
"text": "This work is supported by National Natural Science Foundation of China under Grant 62222212 and Science Fund for Creative Research Groups under Grant 62121002."
}
] | Existing vector quantization (VQ) based autoregressive models follow a two-stage generation paradigm that first learns a codebook to encode images as discrete codes, and then completes generation based on the learned codebook. However, they encode fixed-size image regions into fixed-length codes and ignore their naturally different information densities, which results in insufficiency in important regions and redundancy in unimportant ones, and finally degrades the generation quality and speed. Moreover, the fixed-length coding leads to an unnatural rasterscan autoregressive generation. To address the problem, we propose a novel two-stage framework: (1) Dynamic-Quantization VAE (DQ-VAE) which encodes image regions into variable-length codes based on their information densities for an accurate & compact code representation. (2) DQ-Transformer which thereby generates images autoregressively from coarse-grained (smooth regions with fewer codes) to fine-grained (details regions with more codes) by modeling the position and content of codes in each granularity alternately, through a novel stackedtransformer architecture and shared-content, non-shared position input layers designs. Comprehensive experiments on various generation tasks validate our superiorities in both effectiveness and efficiency. Code will be released at https : / / github . com / CrossmodalGroup / DynamicVectorQuantization. * Zhendong Mao is the corresponding author. Error map: l 1 loss of each 32 2 region between original images and reconstructions, higher (redder) worse. Existing examples are taken from [13]. ing the realism of local details and the consistency of global structure is the eternal pursuit for all image generations. Recently, vector quantization (VQ) [37] has been a foundation for various types of generative models as evidenced by numerous large-scale diffusion models like LDM [32], autoregressive models like DALL-E [30], etc. These models follow a two-stage generation paradigm, i.e., the first stage learns a codebook by VQ to encode images as dis- | Towards Accurate Image Coding: Improved Autoregressive Image Generation with Dynamic Vector Quantization | [
{
"figure_caption": "Figure 2 .2Figure 2. The overview of our proposed two-stage framework. (1) DQ-VAE dynamically assigns variable-length codes for each image region through Dynamic Grained Coding (DGC) module. (2) DQ-Transformer models the position and content of codes alternately by the stacked Position-Transformer and Content-Transformer, generating images autoregressively from coarse-grained to fine-grained. To effectively teach the difference between granularities, we further design shared-content, non-shared-position, and granularity input layers.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 .3Figure 3. Illustration of our Dynamic Grained Coding module.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 .Algorithm 141Figure 4. Qualitative results. Left: Our unconditional generation on FFHQ. Right: Our class-conditional generation on ImageNet.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "41",
"figure_type": "figure"
},
{
"figure_caption": "indexes to avoid repeat 6: sample next code position Pi ∈ R B 7:if P i,b ==<eos>, for b ∈ [1, B]then8: P >i,b = <pad> 9:",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "C= concat(C, Ci), P = concat(P, Pi) 13: end while 14: end for 15: return decoded image I from P and C 10. DQ-Transformer adopts a stack of causal self-attention blocks [38] and is trained with two different settings, i.e., DQ-Transformer b (base) with 6 layers Position-Transformer and 18 layers Content-Transformer of a total 308M parameters, and DQ-Transformer l (large) with 6 layers Position-Transformer and 42 layers Content-Transformer of a total 608M parameters to demonstrate our scalability. All models are trained with eight RTX-3090 GPUs. Top-k and top-p sampling are used to report the best performance. More details can be found in the supplementary.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 6 .6Figure 6. Left: The Pareto curves of the different ratios between generation quality (FID) and generation speed (code length) on FFHQ. Right: The speed comparison between large-scale ViT-VQGAN [42] and our DQ-Transformer(base) according to different batch sizes on FFHQ.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "The comparison is split Comparison of unconditional autoregressive generation on FFHQ. L is coding length. #Params splits in (VAE+autoregressive model).",
"figure_data": "ModelL#ParamsFID↓VQGAN ( ′ 21) [13]256(72.1+307)M11.4DCT ( ′ 21) [28]>1024738M13.06ViT-VQGAN ( ′ 22) [42]1024(599+1697)M5.3RQ-VAE ( ′ 22) [23]256(100+355)M10.38Mo-VQGAN ( ′ 22) ) [44]1024(82.7+307)M8.52DQ-Transformer b640(47.5+308)M4.91TypeModelL#ParamsFID↓IS↑GANBigGAN-deep [6]-160M6.95198.2diffusion[29]-280M12.26-diffusionADM [10]-554M10.94 101.0bi-directMaskGIT [7]1024227M6.18182.1ARMVQGAN* [13]256379M17.575ARMDCT [28]>1024738M36.5-ARMRQ-VAE [23]256480M15.7286.8ARMRQ-VAE [23]256821M13.11 104.3ARMMo-VQGAN [44]1024389M7.12138.3ARMDQ-Transformer b640355M7.34152.8ARMDQ-Transformer l640655M5.11178.2",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Comparison of class-conditional generation with million-level parameters on ImageNet. L is coding length. ARM denotes for autoregressive model. * denotes for our reproduction.",
"figure_data": "",
"figure_id": "tab_1",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Comparison between our million-level model and large-scale billion-level big models of class-conditional generation on ImageNet.",
"figure_data": "TypeModelL#ParamsFID↓IS↑DiffusionImageBART [12]-3.5B21.1961.6ARMVQ-VAE-2 [31]512013.5B31.1145ARMVQGAN [13]2561.4B15.7878.3ARMViT-VQGAN [36]10242.2B4.17175.1ARMRQ-VAE [23]2563.8B7.55134ARMDQ-Transformer b640355M7.34152.8ARMDQ-Transformer l640655M5.11178.2Model TypeModelFID↓GANBigGAN [6]12.4GANStyleGAN2 [21]3.8VAEVDVAE [8]28.5DiffusionImageBART [12]9.57DiffusionUDM [22]5.54Autoregressive DQ-Transformer b4.91",
"figure_id": "tab_2",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Comparison with other types of state-of-the-art on unconditional FFHQ, where we further improve the quality of autoregressive models.GAN [13] of f = 16 as the baseline, and DQ-VAE adopts triple granularities of F = {8, 16, 32} and subject to:",
"figure_data": "",
"figure_id": "tab_3",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Ablations of the proposed variable-length coding on ImageNet.",
"figure_data": "",
"figure_id": "tab_5",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "Ablations of different granularity ratios of DQ-VAE with F={8, 16} on FFHQ. Here r f =8 denotes the ratio of f = 8. \"mean\" and \"var\" denote the mean and variance of dynamic coding length. The codebook usage is calculated as the percentage of used codes over the entire test set.",
"figure_data": "ContentPositionGranularity Absolute positionFID↓sharednon-shared4.91non-shared non-shared5.54sharedshared18.28sharednon-shared16.87sharednon-shared5.06",
"figure_id": "tab_7",
"figure_label": "6",
"figure_type": "table"
},
{
"figure_caption": "Ablations of DQ-Transformer input designs on FFHQ. Here \"granularity\" denotes for DQ-Transformer's granularity layer.",
"figure_data": "",
"figure_id": "tab_8",
"figure_label": "7",
"figure_type": "table"
}
] | Mengqi Huang; Zhendong Mao; Zhuowei Chen; Yongdong Zhang | [
{
"authors": "Jacob Austin; Jonathan Daniel D Johnson; Daniel Ho; Rianne Tarlow; Van Den; Berg",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b0",
"title": "Structured denoising diffusion models in discrete state-spaces",
"year": "2021"
},
{
"authors": "Hangbo Bao; Li Dong; Furu Wei",
"journal": "",
"ref_id": "b1",
"title": "Beit: Bert pre-training of image transformers",
"year": "2021"
},
{
"authors": "Emmanuel Bengio; Pierre-Luc Bacon; Joelle Pineau; Doina Precup",
"journal": "",
"ref_id": "b2",
"title": "Conditional computation in neural networks for faster models",
"year": "2015"
},
{
"authors": "Tolga Bolukbasi; Joseph Wang; Ofer Dekel; Venkatesh Saligrama",
"journal": "PMLR",
"ref_id": "b3",
"title": "Adaptive neural networks for efficient inference",
"year": "2017"
},
{
"authors": "Sam Bond-Taylor; Peter Hessey; Hiroshi Sasaki; Toby P Breckon; Chris G Willcocks",
"journal": "",
"ref_id": "b4",
"title": "Unleashing transformers: Parallel token prediction with discrete absorbing diffusion for fast high-resolution image generation from vectorquantized codes",
"year": "2021"
},
{
"authors": "Andrew Brock; Jeff Donahue; Karen Simonyan",
"journal": "",
"ref_id": "b5",
"title": "Large scale gan training for high fidelity natural image synthesis",
"year": "2018"
},
{
"authors": "Huiwen Chang; Han Zhang; Lu Jiang; Ce Liu; William T Freeman",
"journal": "",
"ref_id": "b6",
"title": "Maskgit: Masked generative image transformer",
"year": "2022"
},
{
"authors": "Rewon Child",
"journal": "",
"ref_id": "b7",
"title": "Very deep vaes generalize autoregressive models and can outperform them on images",
"year": "2020"
},
{
"authors": "Jia Deng; Wei Dong; Richard Socher; Li-Jia Li; Kai Li; Li Fei-Fei",
"journal": "Ieee",
"ref_id": "b8",
"title": "Imagenet: A large-scale hierarchical image database",
"year": "2009"
},
{
"authors": "Prafulla Dhariwal; Alexander Nichol",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b9",
"title": "Diffusion models beat gans on image synthesis",
"year": "2021"
},
{
"authors": "Ming Ding; Zhuoyi Yang; Wenyi Hong; Wendi Zheng; Chang Zhou; Da Yin; Junyang Lin; Xu Zou; Zhou Shao; Hongxia Yang",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b10",
"title": "Cogview: Mastering text-to-image generation via transformers",
"year": "2021"
},
{
"authors": "Patrick Esser; Robin Rombach; Andreas Blattmann; Bjorn Ommer",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b11",
"title": "Imagebart: Bidirectional context with multinomial diffusion for autoregressive image synthesis",
"year": "2021"
},
{
"authors": "Patrick Esser; Robin Rombach; Bjorn Ommer",
"journal": "",
"ref_id": "b12",
"title": "Taming transformers for high-resolution image synthesis",
"year": "2007"
},
{
"authors": "Shuyang Gu; Dong Chen; Jianmin Bao; Fang Wen; Bo Zhang; Dongdong Chen; Lu Yuan; Baining Guo",
"journal": "",
"ref_id": "b13",
"title": "Vector quantized diffusion model for text-to-image synthesis",
"year": "2022"
},
{
"authors": "Yizeng Han; Gao Huang; Shiji Song; Le Yang; Honghui Wang; Yulin Wang",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b14",
"title": "Dynamic neural networks: A survey",
"year": "2021"
},
{
"authors": "Martin Heusel; Hubert Ramsauer; Thomas Unterthiner; Bernhard Nessler; Sepp Hochreiter",
"journal": "Advances in neural information processing systems",
"ref_id": "b15",
"title": "Gans trained by a two time-scale update rule converge to a local nash equilibrium",
"year": "2017"
},
{
"authors": "Mengqi Huang; Zhendong Mao; Penghui Wang; Quan Wang; Yongdong Zhang",
"journal": "",
"ref_id": "b16",
"title": "Dse-gan: Dynamic semantic evolution generative adversarial network for text-to-image generation",
"year": "2022"
},
{
"authors": "A David; Huffman",
"journal": "Proceedings of the IRE",
"ref_id": "b17",
"title": "A method for the construction of minimum-redundancy codes",
"year": "1952"
},
{
"authors": "Eric Jang; Shixiang Gu; Ben Poole",
"journal": "",
"ref_id": "b18",
"title": "Categorical reparameterization with gumbel-softmax",
"year": "2016"
},
{
"authors": "Tero Karras; Samuli Laine; Timo Aila",
"journal": "",
"ref_id": "b19",
"title": "A style-based generator architecture for generative adversarial networks",
"year": "2019"
},
{
"authors": "Tero Karras; Samuli Laine; Miika Aittala; Janne Hellsten; Jaakko Lehtinen; Timo Aila",
"journal": "",
"ref_id": "b20",
"title": "Analyzing and improving the image quality of stylegan",
"year": "2020"
},
{
"authors": "Dongjun Kim; Seungjae Shin; Kyungwoo Song; Wanmo Kang; Il-Chul Moon",
"journal": "",
"ref_id": "b21",
"title": "Score matching model for unbounded data score",
"year": "2021"
},
{
"authors": "Doyup Lee; Chiheon Kim; Saehoon Kim; Minsu Cho; Wook-Shin Han",
"journal": "",
"ref_id": "b22",
"title": "Autoregressive image generation using residual quantization",
"year": "2022"
},
{
"authors": "Doyup Lee; Chiheon Kim; Saehoon Kim; Minsu Cho; Wook-Shin Han",
"journal": "",
"ref_id": "b23",
"title": "Draft-and-revise: Effective image generation with contextual rq-transformer",
"year": "2022"
},
{
"authors": "Yanwei Li; Lin Song; Yukang Chen; Zeming Li; Xiangyu Zhang; Xingang Wang; Jian Sun",
"journal": "",
"ref_id": "b24",
"title": "Learning dynamic routing for semantic segmentation",
"year": "2020"
},
{
"authors": "Zhuang Liu; Jianguo Li; Zhiqiang Shen; Gao Huang; Shoumeng Yan; Changshui Zhang",
"journal": "",
"ref_id": "b25",
"title": "Learning efficient convolutional networks through network slimming",
"year": "2017"
},
{
"authors": "Chengzhi Mao; Lu Jiang; Mostafa Dehghani; Carl Vondrick; Rahul Sukthankar; Irfan Essa",
"journal": "",
"ref_id": "b26",
"title": "Discrete representations strengthen vision transformer robustness",
"year": "2021"
},
{
"authors": "Charlie Nash; Jacob Menick; Sander Dieleman; Peter W Battaglia",
"journal": "",
"ref_id": "b27",
"title": "Generating images with sparse representations",
"year": "2021"
},
{
"authors": "Alexander Quinn; Nichol ; Prafulla Dhariwal",
"journal": "PMLR",
"ref_id": "b28",
"title": "Improved denoising diffusion probabilistic models",
"year": "2021"
},
{
"authors": "Aditya Ramesh; Mikhail Pavlov; Gabriel Goh; Scott Gray; Chelsea Voss; Alec Radford; Mark Chen; Ilya Sutskever",
"journal": "PMLR",
"ref_id": "b29",
"title": "Zero-shot text-to-image generation",
"year": "2021"
},
{
"authors": "Ali Razavi; Aaron Van Den Oord; Oriol Vinyals",
"journal": "Advances in neural information processing systems",
"ref_id": "b30",
"title": "Generating diverse high-fidelity images with vq-vae-2",
"year": "2019"
},
{
"authors": "Robin Rombach; Andreas Blattmann; Dominik Lorenz; Patrick Esser; Björn Ommer",
"journal": "",
"ref_id": "b31",
"title": "High-resolution image synthesis with latent diffusion models",
"year": "2022"
},
{
"authors": "Claude Elwood; Shannon ",
"journal": "The Bell system technical journal",
"ref_id": "b32",
"title": "A mathematical theory of communication",
"year": "1948"
},
{
"authors": "Claude E Shannon",
"journal": "IRE Nat. Conv. Rec",
"ref_id": "b33",
"title": "Coding theorems for a discrete source with a fidelity criterion",
"year": "1959"
},
{
"authors": "Lin Song; Songyang Zhang; Songtao Liu; Zeming Li; Xuming He; Hongbin Sun; Jian Sun; Nanning Zheng",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b34",
"title": "Dynamic grained encoder for vision transformers",
"year": "2021"
},
{
"authors": "Zhicong Tang; Shuyang Gu; Jianmin Bao; Dong Chen; Fang Wen",
"journal": "",
"ref_id": "b35",
"title": "Improved vector quantized diffusion models",
"year": "2022"
},
{
"authors": "Aaron Van Den; Oriol Oord; Vinyals",
"journal": "Advances in neural information processing systems",
"ref_id": "b36",
"title": "Neural discrete representation learning",
"year": "2017"
},
{
"authors": "Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Łukasz Kaiser; Illia Polosukhin",
"journal": "Advances in neural information processing systems",
"ref_id": "b37",
"title": "Attention is all you need",
"year": "2017"
},
{
"authors": "Andreas Veit; Serge Belongie",
"journal": "",
"ref_id": "b38",
"title": "Convolutional networks with adaptive inference graphs",
"year": "2018"
},
{
"authors": "Zhenda Xie; Zheng Zhang; Xizhou Zhu; Gao Huang; Stephen Lin",
"journal": "Springer",
"ref_id": "b39",
"title": "Spatially adaptive inference with stochastic feature sampling and interpolation",
"year": "2020"
},
{
"authors": "Le Yang; Yizeng Han; Xi Chen; Shiji Song; Jifeng Dai; Gao Huang",
"journal": "",
"ref_id": "b40",
"title": "Resolution adaptive networks for efficient inference",
"year": "2020"
},
{
"authors": "Jiahui Yu; Xin Li; Jing Yu Koh; Han Zhang; Ruoming Pang; James Qin; Alexander Ku; Yuanzhong Xu; Jason Baldridge; Yonghui Wu",
"journal": "",
"ref_id": "b41",
"title": "Vector-quantized image modeling with improved vqgan",
"year": "2021"
},
{
"authors": "Jiahui Yu; Yuanzhong Xu; Jing Yu Koh; Thang Luong; Gunjan Baid; Zirui Wang; Vijay Vasudevan; Alexander Ku; Yinfei Yang; Burcu Karagol Ayan",
"journal": "",
"ref_id": "b42",
"title": "Scaling autoregressive models for content-rich text-to-image generation",
"year": "2022"
},
{
"authors": "Chuanxia Zheng; Long Tung Vuong; Jianfei Cai; Dinh Phung",
"journal": "",
"ref_id": "b43",
"title": "Movq: Modulating quantized vectors for highfidelity image generation",
"year": "2022"
},
{
"authors": "Ye Zhu; Yu Wu; Kyle Olszewski; Jian Ren; Sergey Tulyakov; Yan Yan",
"journal": "",
"ref_id": "b44",
"title": "Discrete contrastive diffusion for cross-modal and conditional generation",
"year": "2022"
},
{
"authors": "Yichen Zhu; Yuqin Zhu; Jie Du; Yi Wang; Zhicai Ou; Feifei Feng; Jian Tang",
"journal": "",
"ref_id": "b45",
"title": "Make a long image short: Adaptive token length for vision transformers",
"year": "2021"
}
] | [
{
"formula_coordinates": [
3,
308.86,
559.32,
236.25,
21.91
],
"formula_id": "formula_0",
"formula_text": "C := {(k, e(k))} k∈[K]"
},
{
"formula_coordinates": [
3,
361.48,
665.56,
183.63,
17.12
],
"formula_id": "formula_1",
"formula_text": "Q(z; C) = arg min k∈[K] ||z -e k || 2 2 .(1)"
},
{
"formula_coordinates": [
4,
64.42,
241.48,
221.94,
9.65
],
"formula_id": "formula_2",
"formula_text": "F = {f 1 , f 2 , ..., f K }, where f 1 < f 2 < ... < f K , (2)"
},
{
"formula_coordinates": [
4,
50.11,
262.62,
236.25,
44.85
],
"formula_id": "formula_3",
"formula_text": "Z = {Z 1 , Z 2 , ..., Z K } through a hierarchical encoder E h , where Z i ∈ R Hi×Wi×nz and (H i , W i ) = (H 0 /f i , W 0 /f i ) for each i ∈ {1, 2, ..., K}."
},
{
"formula_coordinates": [
4,
50.11,
570.27,
236.25,
37.96
],
"formula_id": "formula_4",
"formula_text": "Z ′ = {Z ′ 1 , Z ′ 2 , ..., Z ′ K } and Z ′ i ∈ R Hs×Ws×nz for i ∈ {1, 2, ..., K}, where (H s , W s ) = (H 0 /f K , W 0 /f K )."
},
{
"formula_coordinates": [
4,
83.7,
618.76,
198.79,
14.34
],
"formula_id": "formula_5",
"formula_text": "G = (Z ′ 1 ∥Z ′ 2 ∥...∥Z ′ K )W g ∈ R Hs×Ws×k , (3"
},
{
"formula_coordinates": [
4,
282.49,
622.8,
3.87,
8.64
],
"formula_id": "formula_6",
"formula_text": ")"
},
{
"formula_coordinates": [
4,
90.54,
699.74,
195.83,
16.6
],
"formula_id": "formula_7",
"formula_text": "θ i,j = arg max k (g i,j,k ) ∈ {1, 2, ..., K}.(4)"
},
{
"formula_coordinates": [
4,
319.7,
287.02,
225.41,
14.67
],
"formula_id": "formula_8",
"formula_text": "θi,j = arg max k (g i,j,k + n k ), where n k ∼ Gumbel(0,1). (5)"
},
{
"formula_coordinates": [
4,
342.23,
381.04,
202.88,
27.17
],
"formula_id": "formula_9",
"formula_text": "p i,j = exp((g i,j,θi,j + n θi,j ))/τ K k exp((g i,j,k + n k )/τ ) ∈ [0, 1],(6)"
},
{
"formula_coordinates": [
4,
373.35,
617.39,
167.89,
30.55
],
"formula_id": "formula_10",
"formula_text": "L budget = K-1 k (r k -r ′ k ) 2 , (7"
},
{
"formula_coordinates": [
4,
541.24,
628.12,
3.87,
8.64
],
"formula_id": "formula_11",
"formula_text": ")"
},
{
"formula_coordinates": [
4,
513.69,
664.33,
17.38,
14.11
],
"formula_id": "formula_12",
"formula_text": "K-1 k"
},
{
"formula_coordinates": [
4,
362.91,
704.2,
182.2,
9.65
],
"formula_id": "formula_13",
"formula_text": "L stage1 = L vanilla + λL budget ,(8)"
},
{
"formula_coordinates": [
5,
352.1,
195.59,
193.02,
9.65
],
"formula_id": "formula_14",
"formula_text": "L position = E(-log p(P l |P <l , C <l ))(9)"
},
{
"formula_coordinates": [
5,
352.82,
358.74,
192.3,
9.65
],
"formula_id": "formula_15",
"formula_text": "L content = E(-log p(C l |P ≤l , C <l ))(10)"
},
{
"formula_coordinates": [
5,
362.11,
414.3,
183.01,
9.65
],
"formula_id": "formula_16",
"formula_text": "L stage2 = L position + L content .(11)"
},
{
"formula_coordinates": [
6,
54.67,
252.82,
93.12,
18.87
],
"formula_id": "formula_17",
"formula_text": "for each k ∈ [1, K] do 2:"
},
{
"formula_coordinates": [
6,
54.67,
274.8,
203.26,
18.81
],
"formula_id": "formula_18",
"formula_text": "P = concat(P, <sos>), C = concat(C, <sos>) 4:"
},
{
"formula_coordinates": [
7,
129.9,
612.26,
156.46,
9.65
],
"formula_id": "formula_19",
"formula_text": "r f =32 = 4 × r f =8 ,(12)"
}
] |
|
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b15",
"b23",
"b17",
"b26",
"b3",
"b24",
"b7",
"b7",
"b6",
"b9"
],
"table_ref": [],
"text": "Question answering systems devote to answering various questions with the evidence located in the structured knowledge base (e.g., table) (Pasupat and Liang, 2015;Yu et al., 2018) or unstructured texts (Rajpurkar et al., 2016). Considering that many questions need to utilize multiple sources of knowledge jointly in real-world applications, the hybrid form of question answering over texts and tables (TextTableQA) has been proposed and attracted more and more attention (Chen et al., 2020b,a; Zhu et al., 2021;Chen et al., 2021;Zhao et al., 2022;Wang et al., 2022a). Fact reasoning (Chen et al., 2020a,b) is a critical question type of TextTableQA. It requires jointly using multiple evidence from tables and texts to reasoning the answers with different operations, such as correlation (e.g., multi-hop) and aggregation (e.g., comparison). Hyperlinks among some table cells and linked passages are essential resources to establish their relationship and support the retrieval and reasoning for multi-hop questions. As shown in Figure 1, answering a complex question Q1 requires jointly reasoning from textual evidence (P1) to table evidence ([R2,Place]) and then to other table evidence ([R2,Athlete]).\nExisting methods consist of two main stages: retriever and reader (Chen et al., 2020b;Feng et al., 2022). The retriever filters out the cells and passages with high relevance to the question, and then the reader extracts a span from the retrieval results as the final answer. However, current methods with two stages still have three limitations as follows.\n1) Noisy labeling for training retriever. Existing retrieval methods usually ignore the weakly supervised answer annotation (Chen et al., 2020b;Wang et al., 2022b;Feng et al., 2022). For the Q2 of Figure 1, we cannot know the specific location of the hybrid evidence, only given the final answer \"1960\". Therefore, there is a lot of pseudo-true evidence labeled (Marked in green) automatically by string matching, which introduces a lot of evidence noise.\n2) Insufficient utilization of heterogeneous information. After retrieval, existing methods selected a particular cell or passage for reading to extract the final answer (Chen et al., 2020b;Wang et al., 2022b). As for Q1 in Figure 1, previous models were more likely to choose P1 or the coordinates [R2,Place] to extract the answer. However, these methods seldomly used the hybrid information of table schema and cell-passage hyperlinks, which is the key factor in answering multi-hop questions.\n3) Deficient ability for different reasoning operations. Previous methods (Eisenschlos et al., 2021;Kumar et al., 2021;Wang et al., 2022b) mainly used an extraction module to obtain answers, which cannot support knowledge reasoning that requires comparison, calculation, and other operations.\nIn this paper, we propose a three-stage approach S 3 HQA to solve the above problems. (1) Retriever with Refinement Training, we propose a two-step training method, splitting the training data into two parts, so that the noise in the retrieval phase can be alleviated. (2) Hybrid Selector has been proposed and selects supporting facts with different granularity and resources depending on the question type. By considering the hybrid data of tables and text, this paper proposes a hybrid selection algorithm that can effectively utilize the heterogeneous information of tables and passages. (3) Generationbased reasoner utilizes a generation-based model for addressing different question types. The model allows better aggregation of information on the input side, which not only have better multi-hop reasoning capabilities but also be able to handle comparison and counting questions. Furthermore, we are the first to use the LLM in-context learning approach for table-text hybrid question-answering tasks.\nWe evaluate our proposed model on the challenging TextTableQA benchmark HybridQA. The empirical results show that our approach outperforms all the existing models2 ."
},
{
"figure_ref": [],
"heading": "Our Approach",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Problem Definition",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Given a natural language question",
"publication_ref": [],
"table_ref": [],
"text": "Q = {q i } |Q| i=1 and a table T with H, R , H indicates the table headers, and R = {r i } |R| i=1 indicates the rows with number |R|. Each row r i is consists of N cells r i = {c ij } N j=1\n. The header's number is also N . Some cells have a linked passage P ij . Our goal aims to generate the answer A with model Θ, which is a span from table cells or linked passage or a derivation result of counting questions."
},
{
"figure_ref": [],
"heading": "Retriever with Refinement Training",
"publication_ref": [
"b9",
"b6"
],
"table_ref": [],
"text": "The retriever aims to perform initial filtering of heterogeneous resources. However, accurately labeling the location of answers consumes high labeling costs. For TextTableQA data, the answer A usually appears in multiple locations, which makes it difficult for us to generate precise retrieval la-bels. We use a two-step training method, with a row-based retriever and a passage-based retriever for each step.\nInspired by (Kumar et al., 2021), the retrieval has two steps. First, we divide the data D into two folds according to the string matching labels G i . Specifically, for a question-answer instance, the answer A appears one time as D 1 , and the instance whose answer A appears multiple times as D 2 . Take the example in Figure 1, Q1, Q3 belongs to D 1 while Q2 belongs to D 2 . The data is organized in the form of\n[CLS]q 1 q 2 ...q |Q| [SEP]c i1 c i2 ...c iN [SEP] or [CLS]q 1 q 2 ...q |Q| [SEP]p ij [SEP].\nIn the first step, we only use D 1 to train a model Θ 1 , which data are noiseless. Then in the second step, we use the trained weight Θ 1 to train the model Θ 2 . For the input x, the loss function is:\nL(Θ 2 , x, R) = z∈R -q(z) log p Θ 1 (z|x)\nwhere q(z) = p Θ 1 (z|x, z ∈ R) is the probability distribution given by the model restricted to candidate rows R containing the answer span, taken here as a constant with zero gradients (Eisenschlos et al., 2021).\nMeanwhile, we use a passage-based retriever to enhance the performance of a row-based retriever (PassageFilter). Specifically, we use the passage-based retriever to obtain a prediction score of passage relevance. Based on this score, we reorder the input of the row-based retriever. It avoids the limitation on input sequence length imposed by the pre-trained model."
},
{
"figure_ref": [],
"heading": "Hybrid Selector",
"publication_ref": [],
"table_ref": [],
"text": "This module needs to combine the results of the two granularity retrievers. As for this task, we consider the question type and the relationships between the table and linked passages essential. As shown in Figure 2, the hybrid selector chooses the appropriate data source from the two retrieval results depending on question types.\nSpecifically, for general bridge multi-hop questions, we use a single row and its linked passage. While for comparison/count questions, we consider multiple rows and further filter the related sentences, delete the linked paragraphs with the low scores. This not only enables the generation module to obtain accurate information, but also prevents the introduction of a large amount of unrelated information. The selector algorithm outputs a mixed sequence with high relevance based on the relationship between the question, the table, and the passages. The algorithm is shown in Algorithm 1.\nAlgorithm 1 Hybrid Selector Algorithm.\nInput: question Q, table rows R, linked passages P, rowbased retriever ΘR, passage-based retriever ΘP , selector target row count NS Output: generator input S Get the row/passage ordered list by relevant scores\n1: OR ← sort(ΘR(Q, R)) 2: OP ← sort(ΘP (Q, P)) 3: p type ← Classif ication(Q) 4: if p type = bridge then 5: if OP [0] in OR[0] then 6: S ← Q + OR[0] 7: else 8: S ← Q + OR[0] + OP [0] 9:\nend if 10: else 11:\nOPC ← P[len(OP )//2 :] 12:\nS ← Q + OR[0 : NS] -OPC 13: end if 14: return S"
},
{
"figure_ref": [],
"heading": "Generation-based Reasoner",
"publication_ref": [],
"table_ref": [],
"text": "The results of the selector take into account both two granularity. Unlike the previous approaches, which were based on a span extraction module, we use a generation-based model for answer prediction."
},
{
"figure_ref": [],
"heading": "Row-wise generator",
"publication_ref": [],
"table_ref": [],
"text": "To generate an accurate answer string A = (a 1 , a 2 , ..., a n ) given the question Q and selection evidence S, we perform lexical analysis to identify the question type, such as counting or comparison, by looking for certain keywords or comparative adjectives. We utilize two special tags Count and Compare , which indicates the question types. We then use the results of the passage retriever to rank the passages in order of their relevance, eliminating the impact of model input length limitations. Finally, we train a Seq2Seq language model with parameters Θ, using the input sequence Q, S and the previous outputs a <i to optimize the product of the probabilities of the output sequence a 1 , a 2 , ..., a n :\nA = argmax n i=1 P (a i |a <i , Q, S; Θ)"
},
{
"figure_ref": [],
"heading": "LLM prompting generator",
"publication_ref": [
"b5",
"b21"
],
"table_ref": [],
"text": "With the emergence of large language models, In-Context Learning (Dong et al., 2022) and Chain-of-Thought prompting (Wei et al., 2022) have become two particularly popular research topics in this field.\nIn this paper, we introduce a prompting strategy for multi-hop TextTableQA.\nWe utilize selection evidence S and apply LLMbased prompting. We conducted experiments on both vanilla prompting and chain-of-thought prompting in zero-shot and few-shot scenarios."
},
{
"figure_ref": [],
"heading": "Experiment",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Experiment Setup",
"publication_ref": [],
"table_ref": [],
"text": "Datasets We conduct experiments on Hy-bridQA (Chen et al., 2020b). The detailed statistics are shown in Appendix A. For evaluation, we followed the official evaluation to report exact match accuracy and F1 score. Implementation details The implementation details are shown in Appendix B. The experimental results are the average of five times results."
},
{
"figure_ref": [],
"heading": "Fully-supervised Results",
"publication_ref": [
"b7"
],
"table_ref": [
"tab_1"
],
"text": "Table 1 shows the comparison results between our models with previous typical approaches on both development and test sets. It shows that our proposed S 3 HQA works significantly better than the baselines in terms of EM and F1 on HybridQA. The results indicate that S 3 HQA is an effective model for multi-hop question answering over tabular and textual data. Specifically, it can effectively handle multi-hop reasoning and make full use of heterogeneous information.\nHowever, we found that our approach was outperformed by the DEHG model (Feng et al., 2022) in terms of F1 score on the Dev set. We speculate that this might be because the DEHG approach uses their own Open Information Extraction (OIE) tool."
},
{
"figure_ref": [],
"heading": "Model",
"publication_ref": [],
"table_ref": [],
"text": "Dev EM F1 Zero-shot prompt GPT3.5 direct 33.1 50.5 GPT3.5 CoT 52.9 66.6 Few-shot prompt (2-shot) GPT3.5 direct 57.1 68.8 GPT3.5 CoT 60.3 72.1 "
},
{
"figure_ref": [],
"heading": "LLM-prompting Results",
"publication_ref": [],
"table_ref": [
"tab_2",
"tab_1"
],
"text": "We present our zero-shot and few-shot results in Table 2. \"Direct\" refers to a simple prompting method where only the question, context, and answer are provided to the model without any additional reasoning process. In contrast, \"CoT\" involves a human-authored Chain-of-Thought reasoning process that provides a more structured and logical way of prompting the model. The experiments demonstrate that in-context learning used to prompt large language models can achieve promising results. Specifically, utilizing the Chain-of-Thought prompt method can significantly enhance the model's performance.\nHowever, it's worth noting that there is still a performance gap compared to fine-tuning the model on the full dataset (Table 1). Fine-tuning allows the model to learn more specific information about the TextTableQA task, resulting in better performance. Nevertheless, our results show that the LLM-prompting method can be a useful alternative to fine-tuning, especially when there is a limited amount of labeled data available."
},
{
"figure_ref": [],
"heading": "Ablation Studies",
"publication_ref": [
"b9"
],
"table_ref": [
"tab_3"
],
"text": "We conduct ablation studies on the test set. We validate the effects of three modules: retriever with refinement training, hybrid selector, and generation-based reasoner. The retriever performs initial filtering of heterogeneous resources; Selectors combined with hyperlinks further identify the exact evidence needed to answer multi-hop questions; and the reasoner uses the selection evidence to obtain the final answer. Effect of proposed retriever. As shown in the Table 3, under the setting of using the BERTbase-uncased model, sing the BERT-base-uncased model setting, the retriever with refinement training achieved 87.2. When we use Deberta-base, the top1 retrieval performance improved by 0.8%. For w/o refinement training, we use the entire data directly for training, the top1 recall drops about 3.2%. For w/o PassageFilter, we remove the mechanism, the top1 recall drops about 3.2%. For Vanilla-Retriever, we use the row-based retriever (Kumar et al., 2021) and remove all our mechanisms, the top1 score drops about 5.3%. This shows that our model can solve the weakly supervised data noise problem well."
},
{
"figure_ref": [],
"heading": "Model",
"publication_ref": [
"b4"
],
"table_ref": [
"tab_4"
],
"text": "Effect of hybrid selector. As shown in the Table 4, we removed the selector of S 3 HQA and replaced it with the previous cell-based selector (Wang et al., 2022b). This method directly uses the top1 result of the row retriever as input to the generator. w/o hybrid selector shows that the EM drops 2.9% and F1 drops 1.6%, which proves the effectiveness of our selector approach.\nEffect of reasoner. As shown in the Table 4, we design two baselines. BERT-large reader (Chen et al., 2020b;Wang et al., 2022b) uses BERT (Devlin et al., 2018) as encoder and solves this task by predicting the start/end tokens. w/o special tags deletes the special tags. Both the two experiments demonstrate our S 3 HQA reasoner performs the best for HybridQA task."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b14",
"b6",
"b9",
"b7",
"b26",
"b24",
"b25",
"b11",
"b13",
"b22"
],
"table_ref": [],
"text": "The TextTableQA task (Wang et al., 2022a) has attracted more and more attention. As for multi-hop type dataset, previous work used pipeline approach (Chen et al., 2020b), unsupervised approach (Pan et al., 2021), multigranularity (Wang et al., 2022b), table pre-trained language model (Eisenschlos et al., 2021), multiinstance learning (Kumar et al., 2021) and graph neural network (Feng et al., 2022) to solve this task. As for numerical reasoning task, which is quite different from multi-hop type dataset, there is also a lot of work (Zhu et al., 2021;Zhao et al., 2022;Zhou et al., 2022;Lei et al., 2022;Li et al., 2022;Wei et al., 2023) to look at these types of questions. Unlike these methods, our proposed three-stage model S 3 HQA can alleviate noises from weakly supervised and solve different types of multi-hop TextTableQA questions by handling the relationship between tables and text."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "This paper proposes a three-stage model consisting of retriever, selector, and reasoner, which can effectively address multi-hop TextTableQA. The proposed method solves three drawbacks of the previous methods: noisy labeling for training retriever, insufficient utilization of heterogeneous information, and deficient ability for reasoning. It achieves new state-of-the-art performance on the widely used benchmark HybridQA. In future work, we will design more interpretable TextTableQA models to predict the explicit reasoning path."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [],
"table_ref": [],
"text": "Since the multi-hop TextTableQA problem has only one dataset HybridQA, our model has experimented on only one dataset. This may lead to a lack of generalizability of our model. Transparency and interpretability are important in multi-hop question answering. While our model achieves the best results, the model does not fully predict the reasoning path explicitly and can only predict the row-level path and passage-level path. In future work, we will design more interpretable TextTableQA models."
},
{
"figure_ref": [],
"heading": "Acknowledgements",
"publication_ref": [],
"table_ref": [],
"text": "This work was supported by the National Key R&D Program of China (2022ZD0160503) and the National Natural Science Foundation of China (No.U1936207, No.61976211). This work was supported by the Strategic Priority Research Program of Chinese Academy of Sciences (No.XDA27020100), the Youth Innovation Promotion Association CAS, Yunnan Provincial Major Science and Technology Special Plan Projects (No.202202AD080004) and CCF-DiDi GAIA Collaborative Research Funds for Young Scholars."
},
{
"figure_ref": [],
"heading": "A HybridQA Dataset",
"publication_ref": [],
"table_ref": [],
"text": "HybridQA is a large-scale, complex, and multihop TextTableQA benchmark. Tables and texts are crawled from Wikipedia. Each row in the table describes several attributes of an instance. Each "
},
{
"figure_ref": [],
"heading": "B Implementation Details",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "B.1 Fully-supervised Setting",
"publication_ref": [
"b16",
"b0",
"b4",
"b8",
"b12"
],
"table_ref": [],
"text": "We utilize PyTorch (Paszke et al., 2019) to implement our proposed model. During pre-processing, the input of questions, tables and passages are tokenized and lemmatized with the NLTK (Bird, 2006) toolkit. We conducted the experiments on a single NVIDIA GeForce RTX 3090.\nIn the retriever stage, we use BERT-baseuncased (Devlin et al., 2018) and Deberta-base (He et al., 2020) to obtain the initial representations. For the first step, batch size is 1, epoch number is 5, learning rate is 7e-6 (selected from 1e-5, 7e-6, 5e-6). The training process may take around 10 hours. For the second step, we use a smaller learning rate 2e-6 (selected from 5e-6, 3e-6, 2e-6), epoch number is 5. The training process may take around 8 hours. In the selector stage, target row count N S is 3. In the generator stage, we use BART-large language model (Lewis et al., 2020), the learning rate is 1e-5 (selected from 5e-5, 1e-5, 5e-6), batch size is 8, epoch number is 10, beam size is 3 and max generate length is 20."
},
{
"figure_ref": [],
"heading": "B.2 LLM-prompting Setting",
"publication_ref": [],
"table_ref": [],
"text": "We use the OpenAI GPT-3.5 (text-davinci-003) API model with the setting temperature = 0 in our experiments. For the few-shot setting, we use 2 shots. To elicit the LLM's capability to perform multi-hop reasoning, we use the text \"Read the following table and text information, answer a question. Let's think step by step.\" as our prompt."
}
] | Answering multi-hop questions over hybrid factual knowledge from the given text and table (TextTableQA) is a challenging task. Existing models mainly adopt a retriever-reader framework, which have several deficiencies, such as noisy labeling in training retriever, insufficient utilization of heterogeneous information over text and table, and deficient ability for different reasoning operations. In this paper, we propose a three-stage TextTableQA framework S 3 HQA, which comprises of retriever, selector, and reasoner. We use a retriever with refinement training to solve the noisy labeling problem. Then, a hybrid selector considers the linked relationships between heterogeneous data to select the most relevant factual knowledge. For the final stage, instead of adapting a reading comprehension module like in previous methods, we employ a generation-based reasoner to obtain answers. This includes two approaches: a row-wise generator and an LLM prompting generator (first time used in this task). The experimental results demonstrate that our method achieves competitive results in the few-shot setting. When trained on the full dataset, our approach outperforms all baseline methods, ranking first on the HybridQA leaderboard. | S 3 HQA: A Three-Stage Approach for Multi-hop Text-Table Hybrid Question Answering | [
{
"figure_caption": "Figure 1: The examples of HybridQA.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure2: An overview of S 3 HQA framework. The retrieval stage is divided into two steps. The hybrid selector considers the linked relationships between heterogeneous data to select the most relevant factual knowledge.",
"figure_data": "First Step RetrievalSecond Step RetrievalRow-based Θ 1 Passage-basedGenerate pseudo labelΘ 2Row-based Passage-basedR 1 R 2 R 3 R 4P 1 P 3 P 6P 2 P 4 P 7P 5 …………Q1:Who is the athlete in a cityQ3:Who is the higher scoring athletelocated on the Mississippi River?from the cities of Eugene and Walnut?D 1D 3Hybrid SelectorRow ranks R 2 R 1 R 3 R 4 …… Passage ranks P 3 P 2 P 5 P 4 P 6 …… P 3 R 2 Delete low relevance passages Row ranks R 1 R 4 R 2 R 3 …… Passage ranks P 6 P 1 P 2 P 7 …… P 6 R 4P 5P 4R 1 R 2P 1DD 1D 2Passage token Special token Table tokenGeneration-based ReasonorAnswer in R 2Answer in P 1Divided by |G i |Answer",
"figure_id": "tab_0",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Performance of our model and related work on the HybridQA dataset.",
"figure_data": "TablePassageTotalDevTestDevTestDevTestEMF1EMF1EMF1EMF1EMF1EMF1Unsupervised-QG (Pan et al., 2021)--------25.7 30.5--HYBRIDER (Chen et al., 2020b)54.3 61.4 56.2 63.3 39.1 45.7 37.5 44.4 44.0 50.7 43.8 50.6DocHopper (Sun et al., 2021)--------47.7 55.0 46.3 53.3MuGER 2 (Wang et al., 2022b)60.9 69.2 58.7 66.6 56.9 68.9 57.1 68.6 57.1 67.3 56.3 66.2POINTR (Eisenschlos et al., 2021)68.6 74.2 66.9 72.3 62.8 71.9 62.8 71.9 63.4 71.0 62.8 70.2DEHG (Feng et al., 2022)--------65.2 76.3 63.9 75.5MITQA (Kumar et al., 2021)68.1 73.3 68.5 74.4 66.7 75.6 64.3 73.3 65.5 72.7 64.3 71.9MAFiD (Lee et al., 2023)69.4 75.2 68.5 74.9 66.5 75.5 65.7 75.3 66.2 74.1 65.4 73.6S 3 HQA70.3 75.3 70.6 76.3 69.9 78.2 68.7 77.8 68.4 75.3 67.9 75.5Human----------88.2 93.5",
"figure_id": "tab_1",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Performance Comparison of LLM-Prompting Method on Zero-Shot and Few-Shot Scenarios for Hy-bridQA Dataset.",
"figure_data": "",
"figure_id": "tab_2",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Ablation study of retrieval results. DB and BE denote models based on Deberta-base(He et al., 2020) and BERT-base-uncased(Devlin et al., 2018), respectively",
"figure_data": "Top1S 3 HQA-Retriever DB S 3 HQA-Retriever BE88.0 87.3w/o Refinement training 84.1w/o PassageFilter85.3Vanilla-Retriever BE82.0ModelEM F1S 3 HQA67.9 76.5w/o hybrid selector 65.0 74.9w/o special tags67.2 76.0BERT-large reader 66.8 75.8",
"figure_id": "tab_3",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Ablation study of S 3 HQA.",
"figure_data": "",
"figure_id": "tab_4",
"figure_label": "4",
"figure_type": "table"
}
] | Fangyu Lei; Xiang Li; Yifan Wei; Shizhu He; Yiming Huang; Jun Zhao; Kang Liu; Rafer Johnson; Philip Mulkey Memphis; Chuan-Kwang Yang | [
{
"authors": "Steven Bird",
"journal": "",
"ref_id": "b0",
"title": "Nltk: the natural language toolkit",
"year": "2006"
},
{
"authors": "Wenhu Chen; Ming-Wei Chang; Eva Schlinger; William Yang; Wang ; William W Cohen ; A",
"journal": "",
"ref_id": "b1",
"title": "Open question answering over tables and text",
"year": "2020"
},
{
"authors": "Wenhu Chen; Hanwen Zha; Zhiyu Chen; Wenhan Xiong; Hong Wang; William Yang; Wang ",
"journal": "",
"ref_id": "b2",
"title": "Hybridqa: A dataset of multi-hop question answering over tabular and textual data",
"year": "2020"
},
{
"authors": "Zhiyu Chen; Wenhu Chen; Charese Smiley; Sameena Shah; Iana Borova; Dylan Langdon; Reema Moussa; Matt Beane; Ting-Hao Huang; Bryan R Routledge",
"journal": "",
"ref_id": "b3",
"title": "Finqa: A dataset of numerical reasoning over financial data",
"year": "2021"
},
{
"authors": "Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova",
"journal": "",
"ref_id": "b4",
"title": "Bert: Pre-training of deep bidirectional transformers for language understanding",
"year": "2018"
},
{
"authors": "Qingxiu Dong; Lei Li; Damai Dai; Ce Zheng; Zhiyong Wu; Baobao Chang; Xu Sun; Jingjing Xu; Zhifang Sui",
"journal": "",
"ref_id": "b5",
"title": "A survey for in-context learning",
"year": "2022"
},
{
"authors": "Julian Eisenschlos; Maharshi Gor; Thomas Mueller; William Cohen",
"journal": "",
"ref_id": "b6",
"title": "Mate: Multi-view attention for table transformer efficiency",
"year": "2021"
},
{
"authors": "Yue Feng; Zhen Han; Mingming Sun; Ping Li",
"journal": "",
"ref_id": "b7",
"title": "Multi-hop open-domain question answering over structured and unstructured knowledge",
"year": "2022"
},
{
"authors": "Pengcheng He; Xiaodong Liu; Jianfeng Gao; Weizhu Chen",
"journal": "",
"ref_id": "b8",
"title": "Deberta: Decoding-enhanced bert with disentangled attention",
"year": "2020"
},
{
"authors": "Vishwajeet Kumar; Saneem Chemmengath; Yash Gupta; Jaydeep Sen; Samarth Bharadwaj; Soumen Chakrabarti",
"journal": "",
"ref_id": "b9",
"title": "Multi-instance training for question answering across table and linked text",
"year": "2021"
},
{
"authors": "Sung-Min Lee; Eunhwan Park; Daeryong Seo; Donghyeon Jeon; Inho Kang; Seung-Hoon Na",
"journal": "",
"ref_id": "b10",
"title": "Mafid: Moving average equipped fusion-indecoder for question answering over tabular and textual data",
"year": "2023"
},
{
"authors": "Fangyu Lei; Shizhu He; Xiang Li; Jun Zhao; Kang Liu",
"journal": "",
"ref_id": "b11",
"title": "Answering numerical reasoning questions in table-text hybrid contents with graph-based encoder and tree-based decoder",
"year": "2022"
},
{
"authors": "Mike Lewis; Yinhan Liu; Naman Goyal; Marjan Ghazvininejad; Abdelrahman Mohamed; Omer Levy; Veselin Stoyanov; Luke Zettlemoyer",
"journal": "",
"ref_id": "b12",
"title": "Bart: Denoising sequence-to-sequence pretraining for natural language generation, translation, and comprehension",
"year": "2020"
},
{
"authors": "Xiao Li; Yin Zhu; Sichen Liu; Jiangzhou Ju; Yuzhong Qu; Gong Cheng",
"journal": "",
"ref_id": "b13",
"title": "Dyrren: A dynamic retriever-reranker-generator model for numerical reasoning over tabular and textual data",
"year": "2022"
},
{
"authors": "Liangming Pan; Wenhu Chen; Wenhan Xiong; Min-Yen Kan; William Yang; Wang ",
"journal": "",
"ref_id": "b14",
"title": "Unsupervised multi-hop question answering by question generation",
"year": "2021"
},
{
"authors": "Panupong Pasupat; Percy Liang",
"journal": "",
"ref_id": "b15",
"title": "Compositional semantic parsing on semi-structured tables",
"year": "2015"
},
{
"authors": "Adam Paszke; Sam Gross; Francisco Massa; Adam Lerer; James Bradbury; Gregory Chanan; Trevor Killeen; Zeming Lin; Natalia Gimelshein; Luca Antiga",
"journal": "Advances in neural information processing systems",
"ref_id": "b16",
"title": "Pytorch: An imperative style, high-performance deep learning library",
"year": "2019"
},
{
"authors": "Pranav Rajpurkar; Jian Zhang; Konstantin Lopyrev; Percy Liang",
"journal": "",
"ref_id": "b17",
"title": "Squad: 100,000+ questions for machine comprehension of text",
"year": "2016"
},
{
"authors": "Haitian Sun; William W Cohen; Ruslan Salakhutdinov",
"journal": "",
"ref_id": "b18",
"title": "End-to-end multihop retrieval for compositional question answering over long documents",
"year": "2021"
},
{
"authors": "Dingzirui Wang; Longxu Dou; Wanxiang Che",
"journal": "",
"ref_id": "b19",
"title": "A survey on table-and-text hybridqa: Concepts, methods, challenges and future directions",
"year": "2022"
},
{
"authors": "Yingyao Wang; Junwei Bao; Chaoqun Duan; Youzheng Wu; Xiaodong He; Tiejun Zhao",
"journal": "Association for Computational Linguistics",
"ref_id": "b20",
"title": "MuGER2: Multi-granularity evidence retrieval and reasoning for hybrid question answering",
"year": "2022"
},
{
"authors": "Jason Wei; Xuezhi Wang; Dale Schuurmans; Maarten Bosma; Ed Chi; Quoc Le; Denny Zhou",
"journal": "",
"ref_id": "b21",
"title": "Chain of thought prompting elicits reasoning in large language models",
"year": "2022"
},
{
"authors": "Yifan Wei; Fangyu Lei; Yuanzhe Zhang; Jun Zhao; Kang Liu",
"journal": "",
"ref_id": "b22",
"title": "Multi-view graph representation learning for answering hybrid numerical reasoning question",
"year": "2023"
},
{
"authors": "Tao Yu; Rui Zhang; Kai Yang; Michihiro Yasunaga; Dongxu Wang; Zifan Li; James Ma; Irene Li; Qingning Yao; Shanelle Roman",
"journal": "",
"ref_id": "b23",
"title": "Spider: A large-scale human-labeled dataset for complex and cross-domain semantic parsing and text-to-sql task",
"year": "2018"
},
{
"authors": "Yilun Zhao; Yunxiang Li; Chenying Li; Rui Zhang",
"journal": "",
"ref_id": "b24",
"title": "Multihiertt: Numerical reasoning over multi hierarchical tabular and textual data",
"year": "2022"
},
{
"authors": "Yongwei Zhou; Junwei Bao; Chaoqun Duan; Youzheng Wu; Xiaodong He; Tiejun Zhao",
"journal": "",
"ref_id": "b25",
"title": "Unirpg: Unified discrete reasoning over table and text as program generation",
"year": "2022"
},
{
"authors": "Fengbin Zhu; Wenqiang Lei; Youcheng Huang; Chao Wang; Shuo Zhang; Jiancheng Lv; Fuli Feng; Tat-Seng Chua",
"journal": "",
"ref_id": "b26",
"title": "Tat-qa: A question answering benchmark on a hybrid of tabular and textual content in finance",
"year": "2021"
}
] | [
{
"formula_coordinates": [
2,
306.14,
509.43,
218.27,
72.52
],
"formula_id": "formula_0",
"formula_text": "Q = {q i } |Q| i=1 and a table T with H, R , H indicates the table headers, and R = {r i } |R| i=1 indicates the rows with number |R|. Each row r i is consists of N cells r i = {c ij } N j=1"
},
{
"formula_coordinates": [
3,
70.87,
223.59,
218.45,
24.77
],
"formula_id": "formula_1",
"formula_text": "[CLS]q 1 q 2 ...q |Q| [SEP]c i1 c i2 ...c iN [SEP] or [CLS]q 1 q 2 ...q |Q| [SEP]p ij [SEP]."
},
{
"formula_coordinates": [
3,
93.07,
321.07,
173.85,
22.26
],
"formula_id": "formula_2",
"formula_text": "L(Θ 2 , x, R) = z∈R -q(z) log p Θ 1 (z|x)"
},
{
"formula_coordinates": [
3,
309.93,
191.16,
131.1,
87.76
],
"formula_id": "formula_3",
"formula_text": "1: OR ← sort(ΘR(Q, R)) 2: OP ← sort(ΘP (Q, P)) 3: p type ← Classif ication(Q) 4: if p type = bridge then 5: if OP [0] in OR[0] then 6: S ← Q + OR[0] 7: else 8: S ← Q + OR[0] + OP [0] 9:"
},
{
"formula_coordinates": [
3,
333.64,
670.58,
163.28,
33.71
],
"formula_id": "formula_4",
"formula_text": "A = argmax n i=1 P (a i |a <i , Q, S; Θ)"
}
] |
||
2023-05-19 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b46",
"b16",
"b10",
"b19",
"b25",
"b17",
"b32",
"b23",
"b24",
"b12",
"b29",
"b27",
"b38",
"b13",
"b33",
"b16",
"b15",
"b46",
"b18",
"b7",
"b46",
"b18",
"b7",
"b23",
"b38",
"b16",
"b17",
"b12",
"b27",
"b12",
"b27",
"b29",
"b12",
"b12",
"b16"
],
"table_ref": [],
"text": "In many online learning problems, the decision constraint sets are often high-dimensional and complicated, rendering optimization over such sets challenging. In these cases, traditional projectionbased methods, such as Online Gradient Descent (OGD) (Zinkevich, 2003), often suffer heavy computational costs due to the time-consuming or even intractable projection operations. To address this limitation, projection-free online methods, which replace projections with less expensive computations (e.g., linear optimizations) and thus can be implemented efficiently in many cases of interest, have drawn considerable attention in the online learning community (Hazan and Kale, 2012;Garber and Hazan, 2016;Huang et al., 2016;Levy and Krause, 2019;Hazan and Minasyan, 2020;Molinaro, 2020;Kalhan et al., 2021;Wan and Zhang, 2021;Wan et al., 2021;Kretzu and Garber, 2021;Garber and Kretzu, 2022;Mhammedi, 2022;Lu et al., 2023;Wan et al., 2023;Garber and Kretzu, 2023).\nThe studies of projection-free online methods follow the framework of Online Convex Optimization (OCO), which can be regarded as a repeated game between a learner against an adversary (Shalev-Shwartz, 2012). At round t, the learner chooses an action x t from a convex domain set K, and then suffers an instantaneous loss f t (x t ), where the convex loss function f t (•) : K → R is chosen by the adversary. The majority of existing projection-free methods, e.g., Online Frank-Wolfe (OFW) (Hazan and Kale, 2012), minimize the static regret:\nRegret T = T t=1 f t (x t ) -min x∈K T t=1 f t (x),(1)\nwhich benchmarks the cumulative loss of the online method against that of the best fixed action in hindsight. However, in real-world scenarios such as online recommendation and online traffic scheduling (Hazan, 2016), this static metric is unsuitable as the environments are non-stationary and the best action is drifting over time. To tackle this issue, two novel performance metrics: dynamic regret and adaptive regret, are proposed independently (Zinkevich, 2003;Hazan and Seshadhri, 2007;Daniely et al., 2015).\nThe dynamic regret stems from Zinkevich (2003), who defines\nD-Regret T (u 1 , • • • , u T ) = T t=1 f t (x t ) - T t=1 f t (u t ),(2)\nwhere u 1 , • • • , u T ∈ K are any possible comparators. Unfortunately, obtaining a sublinear dynamic regret with arbitrarily varying sequences is impossible. As a result, to establish a meaningful bound, it is common to introduce some regularities of the comparator sequence, such as the path-length\nP T = T t=2 u t-1 -u t 2 .\nThe adaptive regret is originally introduced by Hazan and Seshadhri (2007), and further strengthened by Daniely et al. (2015). Formally, it is defined as\nSA-Regret T (τ ) = max [s,s+τ -1]⊆[T ] s+τ -1 t=s f t (x t ) -min x∈K s+τ -1 t=s f t (x) ,(3)\nwhich is the maximum static regret over any interval with the length τ . Since in different intervals the best actions can be different, (3) essentially measures the performance of the online method against changing comparators. In the literature, only a few projection-free online methods (Kalhan et al., 2021;Wan et al., 2021Wan et al., , 2023) ) have investigated dynamic regret minimization, but all of them focus on the worst case of (2), where u t ∈ arg min u∈K f t (u) is a minimizer of f t (•). However, the worst-case dynamic regret is too pessimistic, and cannot recover the static regret bound of previous methods (Hazan and Kale, 2012;Hazan and Minasyan, 2020). Besides, there exist two studies (Garber and Kretzu, 2022;Lu et al., 2023) that propose projection-free methods for adaptive regret minimization. However, Garber and Kretzu (2022) only consider a weak form of (3) which does not respect short intervals well, and the method of Lu et al. (2023) could be time-consuming in many popular domains, e.g., bounded trace norm matrices and matroid polytopes (Mhammedi, 2022).\nIn this paper, we choose (2) and (3) as the performance metrics, and propose two novel methods for non-stationary projection-free online learning. Specifically, in the dynamic regret minimization, we first establish a novel dynamic regret bound of O(T 3/4 (1 + P T )) for an existing projection-free variant of Online Gradient Descent, termed as BOGD IP (Garber and Kretzu, 2022). 1 Then, we improve the upper bound to O(T 3/4 (1 + P T ) 1/4 ) by proposing a two-layer method named POLD, which maintains multiple BOGD IP algorithms with different step sizes, and tracks the best one on the fly by a meta algorithm. In the adaptive regret minimization, we propose a novel projectionfree method named POLA, which attains an Õ(τ 3/4 ) adaptive regret bound for any interval with the length τ . The key idea is to construct a set of intervals dynamically, run a BOGD IP algorithm that aims to minimize the static regret for each interval, and combine them by a meta algorithm. Moreover, we show that our POLA can also minimize the dynamic regret, and ensures an O(T 3/4 (1 + P T ) 1/4 ) bound. Notably, although POLA can achieve the same dynamic regret bound as POLD, the latter one is still valuable in the sense that it employs a clearer structure and a simpler meta algorithm, rendering it much easier to comprehend and implement.\nContributions. We summarize the contributions of this work below.\n• For dynamic regret, we first provide a novel analysis for BOGD IP (Garber and Kretzu, 2022), and establish an O(T 3/4 (1+P T )) dynamic regret. Then, we improve this bound to O(T 3/4 (1+ P T ) 1/4 ) by proposing a two-layer method named POLD. Note that the obtained bounds can recover the previous O(T 3/4 ) static regret (Hazan and Kale, 2012) by setting P T = 0. To the best of our knowledge, these are the first general-case dynamic regret bounds in projectionfree online learning. • For adaptive regret, based on BOGD IP , we propose a novel projection-free method named POLA and obtain an Õ(τ 3/4 ) adaptive regret which nearly matches previous static results. Moreover, we show that POLA can also ensure an O(T 3/4 (1+P T ) 1/4 ) dynamic regret bound.\nIn other words, it can minimize dynamic regret and adaptive regret simultaneously. • We conduct experiments on practical problems to verify our theoretical findings in dynamic regret and adaptive regret minimization. Empirical results demonstrate the advantage of proposed methods."
},
{
"figure_ref": [],
"heading": "Related work",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we briefly review related work in dynamic regret and adaptive regret."
},
{
"figure_ref": [],
"heading": "Dynamic regret",
"publication_ref": [
"b46",
"b46",
"b30",
"b35",
"b2",
"b3",
"b4",
"b20",
"b31",
"b40",
"b1",
"b23",
"b38",
"b23"
],
"table_ref": [],
"text": "In the literature, dynamic regret has two different forms. One is the general case (2) introduced by Zinkevich (2003), who defines it as the difference between the cumulative loss of the online method and that of any possible comparator sequence. In this seminal work, Zinkevich (2003) establishes the first general-case bound of O( √ T (1 + P T )) for OGD. Later, Zhang et al. (2018a) improve \n) smooth & convex LO WD-R O( √ T (1 + F T + √ D T )) Wan et al. (2023) smooth & convex LO WD-R O( T (1 + F T )) Wan et al. (2021) convex LO WD-R O(max{T 2/3 F 1/3 T , √ T }) strongly convex LO WD-R O(max{ √ T F T log T , log T }) BOGD IP (this work) convex LO D-R O(T 3/4 (1 + P T )) POLD (this work) convex LO D-R O(T 3/4 (1 + P T ) 1/4 ) POLA (this work) convex LO D-R O(T 3/4 (1 + P T ) 1/4 ) Garber and Kretzu (2022) convex LO A-R O(T 3/4 ) Lu et al. (2023) convex MO SA-R Õ( √ τ ) POLA (this work) convex LO SA-R Õ(τ 3/4 )\nthe upper bound to O( T (1 + P T )), motivated by the strategy of maintaining multiple step sizes in MetaGrad (van Erven and Koolen, 2016;Mhammedi et al., 2019;van Erven et al., 2021). In recent years, several studies have further investigated the general-case dynamic regret by leveraging the curvature of loss functions, such as exponential concavity (Baby and Wang, 2021) and strong convexity (Baby and Wang, 2022).\nThe other is the worst case of (2), which specializes the comparators as the minimizers of loss functions (Besbes et al., 2015;Jadbabaie et al., 2015;Mokhtari et al., 2016;Yang et al., 2016;Baby and Wang, 2019):\nD-Regret T (u * 1 , • • • , u * T ) = T t=1 f t (x t ) - T t=1 f t (u * t )(4)\nwhere u * t ∈ arg min u∈K f t (u) is a minimizer of f t (•). However, as pointed out by Zhang et al. (2018a), the worst-case dynamic regret (4) is too pessimistic and could lead to overfitting in the stationary problems.\nIn projection-free online learning, Kalhan et al. (2021) and Wan et al. (2021Wan et al. ( , 2023) ) have investigated the dynamic regret recently, but they only consider the worst-case formulation (4). Specifically, for smooth and convex losses, Kalhan et al. (2021) \nestablish an O( √ T (1 + F T + √ D T )) worst-case bound, where F T = T t=2 sup x∈K |f t (x) -f t-1 (x)| and D T = T t=2 ∇f t (x t ) - ∇f t-1 (x t-1 ) 2 2 .\nFor convex losses and strongly convex losses, Wan et al. (2021) develop the O(max{T 2/3 F 1/3 T ,"
},
{
"figure_ref": [],
"heading": "√",
"publication_ref": [
"b38",
"b23",
"b16"
],
"table_ref": [],
"text": "T }) and O(max{ √ T F T log T , log T }) worst-case bounds, respectively. Very recently, Wan et al. (2023) refine the analysis of Kalhan et al. (2021), achieving an improved O( T (1 + F T )) bound. However, due to the weakness of (4), their bounds can be very loose for any other comparators, and cannot recover the static regret of existing methods, e.g., O(T 3/4 ) for convex losses (Hazan and Kale, 2012)."
},
{
"figure_ref": [],
"heading": "Adaptive regret",
"publication_ref": [
"b26",
"b14",
"b28",
"b0",
"b18",
"b7",
"b43",
"b18",
"b15",
"b7",
"b39",
"b14",
"b12",
"b27",
"b16",
"b29",
"b29",
"b27"
],
"table_ref": [
"tab_0"
],
"text": "Prior work in adaptive regret minimization mainly focus on the setting of Prediction with Expert Advice (PEA) (Littlestone and Warmuth, 1994;Freund et al., 1997;György et al., 2012;Luo and Schapire, 2015;Adamskiy et al., 2016), and OCO (Hazan and Seshadhri, 2007;Daniely et al., 2015;Jun et al., 2017a,b;Zhang et al., 2019). In this section, we specifically introduce the related work of the latter one. Hazan and Seshadhri (2007) first introduce the notion of adaptive regret, but in a weak form:\nA-Regret T = max [s,e]⊆[T ] e t=s f t (x t ) -min x∈K e t=s f t (x) ,(5)\nwhich is the maximum static regret over any contiguous interval. To minimize (5), they propose Follow the Leading History (FLH) with an O(d log 2 T ) weak adaptive regret bound for exponentially concave losses where d denotes the dimensionality. However, (5) could be dominated by long intervals and hence, cannot respect short intervals well. For example, one may obtain an O( √ T ) weak adaptive regret for OGD, but this is vacuous for the intervals with length o( √ T ) (Hazan, 2016). For this reason, Daniely et al. (2015) put forth the (strongly) adaptive regret (3), and design a two-layer algorithm named Strongly Adaptive Online Learner (SAOL). The basic idea is first to construct a set of Geometric Covering (GC) intervals and for each interval, run an OGD algorithm that can obtain the optimal static regret. Then, SAOL combines the actions of these OGD algorithms by a meta algorithm. We observe that the technique of constructing GC intervals can be traced back to the prior studies (Willems and Krom, 1997;György et al., 2012).\nIn projection-free online learning, Garber and Kretzu (2022) study the weak version of adaptive regret (5), and propose a projection-free extension of OGD named BOGD IP with an O(T 3/4 ) bound. Unfortunately, due to the limitation of (5), their bound does not respect short intervals well. Very recently, following the framework of SAOL, Lu et al. (2023) propose a novel two-layer method to minimize (3). Different from previous projection-free algorithms, e.g., OFW (Hazan and Kale, 2012), their method circumvents the projections with membership operations (Mhammedi, 2022). However, such operations could be inefficient in many practical scenarios, e.g., bounded trace norm matrices and matroid polytopes (Mhammedi, 2022). Besides, in each round, their method need to perform O(log T ) membership operations for each expert algorithm, which brings heavy computational costs when T is large.\nSummary. While a few work have investigated non-stationary projection-free online learning (see Table 1 for details), they are still unsatisfactory in the following aspects:\n• In the dynamic regret minimization, there is no study optimizing the general-case form (2), which is more challenging since it needs to build a universal guarantee over any comparator sequences. • In the adaptive regret minimization, although Lu et al. (2023) have established bounds for\n(3), their method is based on the membership operations, instead of the more popular linear optimizations."
},
{
"figure_ref": [],
"heading": "Main results",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we first introduce the basic assumptions. Then, we present our proposed methods as well as their theoretical guarantees in dynamic regret and adaptive regret minimization."
},
{
"figure_ref": [],
"heading": "Assumptions",
"publication_ref": [
"b33",
"b15"
],
"table_ref": [],
"text": "Similar to previous studies on OCO, we adopt the following standard assumptions (Shalev-Shwartz, 2012;Hazan, 2016).\nAssumption 1 The convex decision set K contains the origin 0, and belongs to an Euclidean ball RB with the diameter D = 2R, i.e.,\n∀x, x ∈ K, x -x 2 ≤ D.(6)\nAssumption 2 At each round t, the loss function\nf t (•) is G-Lipschitz over K, i.e., ∀x, y ∈ K, |f t (x) -f t (y)| ≤ G x -y 2 . (7\n)\nAssumption 3 At each round t, the loss function\nf t (•) is convex over K, i.e., ∀x, y ∈ K, f t (y) ≥ f t (x) + ∇f t (x) (y -x). (8\n)\nAssumption 4 At each round t, the loss function value\nf t (x) belongs to [0, 1] for any x ∈ K, i.e., ∀x ∈ K, 0 ≤ f t (x) ≤ 1.(9)"
},
{
"figure_ref": [],
"heading": "Projection-free dynamic regret",
"publication_ref": [
"b12",
"b12",
"b11",
"b17",
"b12",
"b46"
],
"table_ref": [],
"text": "We first revisit BOGD IP (Garber and Kretzu, 2022), of which the key idea is to replace the projection operation with an infeasible projection oracle O IP , defined as following.\nDefinition 1 Let O IP be an infeasible projection oracle over K ⊆ RB, and be the error tolerance.\nThen, for any input points (x 0 , y 0 ) ∈ K × R d , the infeasible projection oracle returns\nx, ỹ = O IP (K, , x 0 , y 0 ), where (x, ỹ) ∈ K × RB, and x -ỹ 2 ≤ √ 3 and ∀z ∈ K, ỹ -z 2 ≤ y 0 -z 2 .\nRemark: O IP can be implemented efficiently by solving linear optimizations. We briefly introduce this implementation in Appendix A, and refer interested readers to Garber and Kretzu (2022) for a deeper comprehension. Besides, BOGD IP utilizes the blocking technique (Garber and Kretzu, 2020;Hazan and Minasyan, 2020), which divides the time horizon T into equally-sized blocks and only conducts updating at the end of each block. In other words, for each block m, BOGD IP maintains (x m , ỹm ) ∈ K × RB, and updates them at the last round of block m. To be precise, BOGD IP first performs gradient descent on ỹm with the step size η:\ny m+1 = ỹm -η mK r=(m-1)K+1 ∇f r (x m ), (10\n)\nwhere K is the block size and mK r=(m-1)K+1 ∇f r (x m ) is the sum of all gradients during the block m. Then, BOGD IP invokes O IP to obtain x m+1 and ỹm+1 for the next block:\nx m+1 , ỹm+1 = O IP (K, , x m , y m+1 ). (11\n)\nAlgorithm 1 Blocked Online Gradient Descent with Infeasible Projections (BOGD IP ) Input: Number of rounds T , domain set K, step size η, infeasible projection oracle O IP Initialization: Choose arbitrary point x 1 ∈ K and set ỹ1 = x 1 , m = 1, block size K = η -2/3 and error tolerance = η 2/3 .\n1: for t = 1 to T do 2: Submit x t = x m , observe f t (x t\n) and obtain ∇f t (x t )\n3:\nif t mod K = 0 then 4:\nUpdate y m+1 according to (10) 5:\nSet x m+1 , ỹm+1 according to (11), and m = t/K + 1 6:\nend if 7: end for With appropriate parameters, we can prove that BOGD IP requires O(T 1/2 ) invocations of O IP , and each invocation solves O(T 1/2 ) linear optimizations. As a result, there are at most O(T ) linear optimizations for the time horizon T . We summarize the detailed procedure in Algorithm 1.\nIn the prior study, Garber and Kretzu (2022) have investigated the weak adaptive regret (5). Different from them, we focus on minimizing the general-case dynamic regret (2) and establish an O(T 3/4 (1 + P T )) bound for BOGD IP as shown in Theorem 2. The intuition lies in that BOGD IP is a projection-free variant of OGD, which is very suitable for dynamic regret minimization (Zinkevich, 2003).\nTheorem 2 Let η = T -3/4 , K = η -2/3 = T 1/2 and = η 2/3 = T -1/2 . Under Assumptions 1, 2 and 3, Algorithm 1 guarantees D-Regret T (u 1 , • • • , u T ) ≤ O η 1/3 T + η -1 (1 + P T ) = O T 3/4 (1 + P T ) ."
},
{
"figure_ref": [],
"heading": "Moreover, the overall number of solving linear optimizations is O(T ).",
"publication_ref": [
"b16",
"b12"
],
"table_ref": [],
"text": "Remark: Our result is the first general-case dynamic regret bound in projection-free online learning, and can automatically adapt to the nature of environments. For example, when the comparators are fixed (i.e., P T = 0), our dynamic regret degenerates to O(T 3/4 ), which matches the static regret bound of Hazan and Kale (2012). To be specific, we have the following corollary, which can also be derived from Theorem 3 of Garber and Kretzu (2022) Corollary 3 Under Assumptions 1, 2 3, Algorithm 1 with the same parameter setting in Theorem 2 guarantees a static regret bound of Regret T ≤ O(T 3/4 ).\n(12)"
},
{
"figure_ref": [],
"heading": "Improved projection-free dynamic regret",
"publication_ref": [
"b34"
],
"table_ref": [],
"text": "Note that the linear dependency on P T in Theorem 2 is too loose and the obtained bound can be vacuous with P T = Ω(T 1/4 ). To address this issue, we propose a two-layer method, termed as Projection-free Online Learning with Dynamic Regret (POLD), with a tighter bound of O(T 3/4 (1+ P T ) 1/4 ). To help understanding, we first briefly introduce the motivation behind POLD. -3/4 (1 + PT ) 3/4 ) and achieve a tighter O(T 3/4 (1+ PT ) 1/4 ) bound. This indicates that if the path-length is known, we can actually tune the step size to obtain an improved bound. To deal with the uncertainty of the path-length, we adopt the strategy of maintaining multiple step sizes (van Erven and Koolen, 2016;Zhang et al., 2018a), and leverage the two-layer structure: running multiple BOGD IP algorithms with different step sizes and combining them by a meta algorithm. In the following, we describe the detailed procedure.\nFirst, we create a set of step sizes\nH = η i = 2 i-1 7D 2 2G 2 T 3/4 i = 1, • • • , N ,(13)\nwhere N = 3 4 log 2 (1 + 4T /7) + 1. Then, we activate a set of experts {E i | η i ∈ H}, each of which is an instance of BOGD IP with the step size η i chosen from H. For each expert E i , we initiate its weight w i 1 = C i(i+1) where C = 1 + 1 N . Next, inspired by the Hedge algorithm (Freund and Schapire, 1997), we combine the actions of experts in a weighted-average fashion. Concretely, in each round t, POLD receives the action x i t from expert E i , and computes the weighted average action:\nx t = i∈H w i t x i t ,(14)\nwhere w i t is the weight assigned to E i . After that, POLD updates the weight of E i by\nw i t+1 = w i t e -αft(x i t ) µ∈H w µ t e -αft(x µ t ) ,(15)\nwhere α denotes the learning rate of the meta algorithm. Finally, POLD reveals the function f t (•) to all experts so that they can update their actions for the next round. We summarize all the procedure in Algorithm 2, and present the following theorem.\nTheorem 4 Let α = 8/T and H be defined as (13). Under Assumptions 1, 2, 3 and 4, Algorithm 2 guarantees\nD-Regret T (u 1 , • • • , u T ) ≤ O T 3/4 (1 + P T ) 1/4 .\nAlgorithm 3 Projection-free Online Learning with Adaptive Regret (POLA) 1: for t = 1 to T do 2:\nfor I ∈ C t do 3:\nCreate an expert E I which runs BOGD IP from an arbitrary initial point with η = |I| -3/4\n4:\nFor the expert E I , set R t-1,I = C t-1,I = 0 5:\nAdd expert E I to the set of active experts A t 6:\nend for 7:\nFrom A t , remove all experts who end at the round t 8:\nReceive the action x t,I of each expert E I ∈ A t and calculate its weight w t,I according to (17) 9:\nSubmit x t defined in ( 18) and then receive f t (•) 10:\nFor each\nE I ∈ A t , update R t,I = R t-1,I + f t (x t ) -f t (x t,I ) C t,I = C t-1,I + |f t (x t ) -f t (x t,I )| 11:\nSend f t (•) to each expert E I ∈ A t 12: end for t 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Remark: Compared with the upper bound in Theorem 2, the dependence on the path-length is reduced from P T to P 1/4 T .\n• • • I 0 [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] • • • I 1 [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ • • • I 2 [ ] [ ] [ ] [ ] [ • • • I 3 [ ] [ • • •"
},
{
"figure_ref": [
"fig_0"
],
"heading": "Projection-free adaptive regret",
"publication_ref": [
"b18",
"b7",
"b7",
"b28",
"b43",
"b12",
"b27",
"b12",
"b27",
"b29",
"b44",
"b6",
"b44"
],
"table_ref": [],
"text": "As mentioned before, besides the dynamic regret (2), there do exist another metric called (strongly) adaptive regret (3) in the non-stationary environments. In this section, we proceed to investigate minimizing (3) and present Projection-free Online Learning with Adaptive Regret (POLA). Following existing studies on adaptive regret (Hazan and Seshadhri, 2007;Daniely et al., 2015), POLA contains three parts: an expert algorithm, a set of intervals, and a meta algorithm. In the following, we specify them separately. First, we take BOGD IP as the expert algorithm, since it is projection-free and ensures an O(|I| 3/4 ) static regret for a given interval I as shown in Corollary 3. Then, we build the GC intervals (Daniely et al., 2015) shown in Figure 1:\nI = k∈N∪{0} I k , I k = [i • 2 k , (i + 1) • 2 k -1] : i ∈ N . (16\n)\nFor each interval I, we maintain an instance of BOGD IP , denoted as the expert E I , to minimize the static regret over that interval. According to Corollary 3, we set the step size η = |I| -3/4 to obtain the O(|I| 3/4 ) static regret bound over the interval I.\nNext, to track the best expert on the fly, we choose AdaNormalHedge (Luo and Schapire, 2015) as the meta algorithm since it naturally supports the setting that the number of experts varies over time (Zhang et al., 2019). The key ingredient of AdaNormalHedge is the potential function:\nΦ(R, C) = exp ([R] 2 + /3C)\n, where [x] + = max(0, x), Φ(0, 0) = 1 and R, C are two variables maintained by each expert. Based on Φ(R, C), we can compute the weight for each expert according to the following weight function:\nw(R, C) = 1 2 (Φ(R + 1, C + 1) -Φ(R + 1, C -1)) .\nPutting all pieces together, we obtain our projection-free POLA for adaptive regret minimization. Below, we describe the detailed procedure, which is also summarized in Algorithm 3.\nFor brevity, we denote the set of all active experts as A t for the round t, and the set of intervals that start from the round t as C t = {I | I ∈ I, t ∈ I, (t -1) / ∈ I}. In Step 3, we create an instance of BOGD IP as the expert E I for each I ∈ C t , and initiate it from an arbitrary initial point with the step size η = |I| -3/4 . In Step 4, we set the variables R t-1,I = C t-1,I = 0 for E I , where R t-1,I = t-1 u=min I f t (x t ) -f t (x t,I ) denotes the regret of E I up to round t -1, and\nC t-1,I = t-1 u=min I |f t (x t ) -f t (x t,I )\n| denotes the sum of the absolute value of instantaneous regrets, and min I denotes the beginning round of I. In Step 5, the new expert E I is added to A t . Then, we remove all experts from A t , who end at the round t (Step 7). After receiving the action x t,I from E I , we update its corresponding weight as following:\nw t,I = w(R t-1,I , C t-1,I ) E I ∈At w(R t-1,I , C t-1,I ) . (17\n)\nIn\nStep 9, we submit the weighted action\nx t = E I ∈At w t,I x t,I ,(18)\nand receive the loss function f t (•). In Step 10, for each E I ∈ A t , we compute its corresponding variables R t,I and C t,I . At the end, we reveal f t (•) to all active experts, so that they can update their actions for the next round (Step 11). We present the adaptive regret bound of POLA below.\nTheorem 5 Under Assumptions 1, 2, 3 and 4, Algorithm 3 guarantees\nSA-Regret T (τ ) ≤ O( τ log T + τ 3/4 ) = Õ τ 3/4 .\nRemark: Compared to existing methods (Garber and Kretzu, 2022;Lu et al., 2023) for adaptive regret minimization, POLA has following advantages.\n• POLA enjoys an Õ(τ 3/4 ) strongly adaptive regret, and thus can still perform well on short intervals. In contrast, Garber and Kretzu (2022) minimize the weak adaptive regret (5), which only promises a performance guarantee on long intervals. • For each expert, POLA performs only O(1) linear optimizations per round on average, whereas Lu et al. (2023) require a significantly higher number of O(log T ) membership operations. Moreover, their operations could be inefficient compared to linear optimizations in many popular domains. For example, the trace norm constraints K = {X| X * ≤ δ, X ⊂ R m×n } incurs a membership operations cost of O(mn 2 ) while the linear optimization cost is O(nnz(X)), where nnz(X) denotes the number of non-zero entries (Mhammedi, 2022). Moreover, we note that previous studies on projection-based online learning (Zhang et al., 2020;Cutkosky, 2020) have shown that it is possible to design a single algorithm to minimize dynamic regret and adaptive regret simultaneously. In particular, our POLA shares a similar twolayer structure with the method of Zhang et al. (2020), inspiring us to investigate the performance of POLA for dynamic regret minimization. The following theorem shows that POLA also enjoys an O(T 3/4 (1 + P T ) 1/4 ) dynamic regret bound.\nTheorem 6 Under Assumptions 1, 2, 3 and 4, Algorithm 3 guarantees\nD-Regret T (u 1 , • • • , u T ) ≤ O T 3/4 (1 + P T ) 1/4 .\nRemark: Although POLA achieves the same dynamic regret bound as POLD, this does not imply that the latter one is insignificant. Compared with POLA, POLD employs a simpler meta algorithm and does not need to construct GC intervals, making it much easier to comprehend and implement."
},
{
"figure_ref": [
"fig_1",
"fig_2"
],
"heading": "Experiments",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we present numerical experiments to support our theoretical results in dynamic regret and adaptive regret minimization. All experiments are conducted on a machine equipped with the Intel Xeon E5-2620 CPU and 32G memory, and each of them is repeated five times with different random seeds. We present experimental results (mean and standard deviation) in Figures 2 and3."
},
{
"figure_ref": [
"fig_1"
],
"heading": "Dynamic regret minimization",
"publication_ref": [],
"table_ref": [],
"text": "Setup. To evaluate our methods (i.e. BOGD IP , POLD and POLA) in dynamic regret minimization, we study the problem of online matrix completion, of which the goal is to produce a matrix X from the trace norm ball in an online fashion to approximate the target matrix M ∈ R m×n . Specifically, in each round t, the learner receive a sampled data (i, j) with the value M ij from the entry set OB of M . Then, the learner chooses X from the trace norm ball K = {X| X * ≤ δ, X ⊂ R m×n } where δ is the parameter, and suffers the online loss\nf t (X) = (i,j)∈OB |X ij -M ij |.\nWe conduct the experiments with δ = 10 4 on the public dataset: MovieLens 100K 2 , which contains 100000 ratings from 943 users on 1682 movies. Following Wan et al. (2021), we slightly modify the dataset to simulate the non-stationary environments. Concretely, we generate an extended datasets\n{(i k , j k , M i k j k )} 300000 k=1\nby merging three copies of MovieLens 100K. For entries corresponding to k = 100001, • • • , 200000, we negate the original values M i k j k to obtain -M i k j k . For simplicity, we divide the extended datasets into T = 3000 partitions. In this way, the target matrix M drifts every 1000 rounds.\nContenders. We compare our methods with the projection-free algorithm: Multi-OCG (Wan et al., 2021), and the projection-based algorithm: Ader (Zhang et al., 2018b). All parameters of each method are set according to the theoretical suggestions. For instance, the learning rate of the i-th expert is set as η i = c 2 i-1 -1/2 in Multi-OCG, and η i = c2 i-1 T -1/2 in Ader, and η i = c2 i-1 T -3/4 in POLD, where c is the hyper-parameters selected from {10 -1 , 10 0 , • • • , 10 6 }.\nResults. We report the average instantaneous loss, the cumulative loss and the runtime (in seconds) against the number of rounds for each method in Figure 2. As evident from the results, projection-free methods are significantly more efficient compared to the projection-based approach (i.e. Ader), albeit with a slight compromise on cumulative loss. This observation is reasonable in the sense that (i) the cost of linear optimization over the trace norm ball is O(nnz(X)) whereas projection operation suffers a much higher O(mn 2 ) cost; (ii) our methods ensure an O(T 3/4 (1 + P T ) 1/4 ) bound against the O( T (1 + P T )) bound of Ader. Moreover, owing to the inherent advantage in minimizing the general-case dynamic regret, our methods yield a lower cumulative loss compared to the projection-free contender Multi-OCG."
},
{
"figure_ref": [
"fig_2"
],
"heading": "Adaptive regret minimization",
"publication_ref": [
"b5",
"b27"
],
"table_ref": [],
"text": "Setup. To evaluate our method (i.e., POLA) in adaptive regret minimization, we consider the problem of online multiclass classification. In each round t, the learner is presented a sampled data (e t , l t ) with e t ∈ R d being the feature and l t ∈ C = {1, • • • , h} being the corresponding class label. Then, the learner is required to choose a decision matrix X = [x 1 , • • • , x h ] ∈ K, where K = {X| X * ≤ δ, X ⊂ R h×d } denotes the trace norm ball with the parameter δ, and predict the class label as arg max j∈C x j e t . Next, the learner incurs a convex multivariate logistic loss\nf t (X) = log 1 + j =lt exp x j e t -x lt e t .\nWe perform experiments with δ = 102 on the public shuttle dataset (Chang and Lin, 2011), which contains 43500 data belonging to 7 classes. For bevity, the dataset is divided equally into 4350 partitions, and we flip the original features by multiplying -1 every 1000 partitions to simulate the non-stationary environments.\nContenders. To verify the performance benefit of POLA by solving linear optimizations, we choose Projection-free Adaptive via a Membership Oracle (PAMO) (Lu et al., 2023), which employs membership oracle in lieu of the projection operation, as the contender. All parameters of each method are set according to the theoretical suggestions. For instance, the learning rate of the i-th expert is set as\nη i = c 2 i-1 -1/2 in PAMO and η i = c 2 i-1 -3/4 in POLA, where c is the hyper-parameters selected from {2 -4 , 2 -3 , • • • , 2 4 }.\nResults. We plot the average instantaneous loss, the cumulative loss and the runtime (in seconds) of each algorithm in Figure 3. It can be clearly seen that POLA is significantly faster than PAMO, despite a slight sacrifice of cumulative loss. This is reasonable since each invocation of membership oracle over the trace norm ball K requires O(h 2 d) costs, while linear oracle suffers only O(nnz(X)). Therefore, as mentioned previously, each expert in PAMO suffers a total of O(h 2 d log T ) computational costs per round, compared to O(nnz(X)) in POLA."
},
{
"figure_ref": [],
"heading": "Conclusion and future work",
"publication_ref": [
"b12",
"b45"
],
"table_ref": [],
"text": "In this paper, we investigate non-stationary projection-free online learning with dynamic regret and adaptive regret guarantees. Specifically, in the dynamic regret minimization, we provide a novel dynamic regret analysis for BOGD IP (Garber and Kretzu, 2022), and establish the first O(T 3/4 (1 + P T )) general-case dynamic regret. Then, we improve this bound to O(T 3/4 (1 + P T ) 1/4 ) by proposing POLD, which runs a set of BOGD IP algorithms with different step sizes in parallel and tracks the best one on the fly. In the adaptive regret minimization, we present our method POLA with an Õ(τ 3/4 ) strongly adaptive regret bound. The essential idea is to construct the GC intervals, maintain an instance of BOGD IP to minimize the static regret for each interval, and then combine actions of instances by a meta algorithm. Furthermore, we show that POLA can also minimize the dynamic regret and achieve the same bound as that of POLD. Empirical studies on dynamic regret and adaptive regret minimization have verified our theoretical findings.\nCurrently, both POLD and POLA need to maintain O(log T ) experts, which leads to O(log T ) linear optimizations per round. Therefore, a natural question arises: is it possible to further reduce the number of linear optimizations in each round, i.e., from O(log T ) to O(1)? We note that in nonstationary projection-based online learning, O(log T ) projection operations can indeed be reduced to O(1) (Zhao et al., 2022). But in the projection-free setting, it seems highly non-trivial and we leave it as a future work.\nAlgorithm 4 Subroutine of O IP (Algorithm 3 in Garber and Kretzu ( 2022)) Input: Domain set K, error tolerance , initial point x 1 ∈ K, target point y\n1: for i = 1, • • • do 2: Compute v i = arg min x∈K x i -y, x 3: if (x i -y) (x i -v i ) ≤ or x i -y 2 2 ≤ 3 then 4: Return x ← x i 5:\nend if"
},
{
"figure_ref": [],
"heading": "6:",
"publication_ref": [
"b12",
"b46"
],
"table_ref": [],
"text": "Set Garber and Kretzu (2022)) First, we divide the upper bound of dynamic regret into three terms. Let K be the block size, and t = (m -1)K + k be the k-th round of the block m. By utilizing the convexity of f t (•), we have\nδ i = arg min δ∈[0,1] { x i + δ(v i -x i ) -y 2 2 } 7: Set x i+1 = x i + δ i (v i -x i ) 8: end for Algorithm 5 Infeasible Projection Oracle, O IP (Algorithm 4 in\nInput: Domain set K, error tolerance , initial point x 0 ∈ K, initial point y 0 1: y 1 = y 0 / max{1, y 2 /R} 2: if x 0 -y 0 2 2 ≤ 3 then 3: Return (x, ỹ) ← (x 0 , y 1 ) 4: end if 5: for i = 1, • • • do 6: x i ← Algorithm 4 (K, , x i-1 , y i ) 7: if x i -y i 2 2 > 3 then 8: y i+1 = y i -γ(y i -x i ) 9:\nT t=1 f t (x t ) - T t=1 f t (u t ) = T /K m=1 K k=1 [f t (x m ) -f t (u t )] (8) ≤ T /K m=1 K k=1 ∇f t (x m ), x m -ỹm :=A + T /K m=1 K k=1 ∇f t (x m ), ỹm -u s(m) :=B + T /K m=1 K k=1 ∇f t (x m ), u s(m) -u t :=C ,(20)\nwhere we assume T /K is an integer without loss of generality, and denote u s(m) as the comparator of the first round s(m) = (m -1)K + 1 in block m for brevity.\nThen, we analyze above three terms separately. To upper bound term A of (20), we introduce the following lemma.\nLemma 9 Let O IP be the infeasible projection oracle over the domain set K ⊆ RB, and be the error tolerance. To compute the action for the block m, BOGD IP invokes O IP as following\nx m , ỹm = O IP (K, , x m-1 , y m ),(21)\nand obtains (x m , ỹm ) ∈ K × RB with\nx m -ỹm 2 ≤ √ 3 . (22\n)\nBy choosing proper parameters, each invocation of O IP requires solving at most O(T 1/2 ) linear optimizations.\nBy exploiting Lemma 9 and the Lipschitz continuity of f t , the first term A of ( 20) can be upper bounded as follows.\nA = T /K m=1 K k=1 ∇f t (x m ), x m -ỹm (7),(22) ≤ GT √ 3 .(23)\nThen, we upper bound the second term B of (20). To this end, we first denote P T as the pathlength of the sequence of u s(1) , • • • , u s(T /K) , i.e.,\nP T = T /K m=2 u s(m-1) -u s(m) 2 ,(24)\nand introduce the following lemma.\nLemma 10 Let P T = T /K m=2 u s(m-1) -u s(m) 2 and P T = T t=2 u t-1 -u t 2 , then we have\nP T ≤ P T .(25)\nInspired by the dynamic regret analysis for OGD in Zinkevich (2003), we can obtain the upper bound of B with respect to P T as shown below.\nLemma 11 Let P T = T /K m=2 u s(m-1) -u s(m) 2 , and K be the block size and η be the step size. Under Assumptions 1 and 2, we have\nB = T /K m=1 K k=1 ∇f t (x m ), ỹm -u s(m) ≤ 7 4η D 2 + η 2 KT G 2 + D η P T .(26)\nBy applying Lemma 10 and Lemma 11, we have\nB = T /K m=1 K k=1 ∇f t (x m ), ỹm -u s(m) ≤ 7 4η D 2 + η 2 KT G 2 + D η P T .(27)\nNext, we proceed to upper bound the third term C of (20). To simplify the notation, we denote the local path-length during the block m as\nP T (m) = K k=2 u (m-1)K+k-1 -u (m-1)K+k 2 . (28\n)\nAnd for any u (m-1)K+k (1\n≤ k ≤ K) in block m, we have u s(m) -u t 2 = u s(m) -u (m-1)K+k 2 = u (m-1)K+1 -u (m-1)K+k 2 ≤ P T (m),(29)\nwhere t = (m -1)K + k.\nMoreover, the sum of all local path-length P T (m) is upper bounded by the global path-length P T , i.e.,\nT /K m=1 P T (m) ≤ P T .(30)\nTherefore, substituting ( 7), ( 29) and ( 30) into the third term C of (20), we have\nC = T /K m=1 K k=1 ∇f t (x m ), u s(m) -u t (7) ≤ G T /K m=1 K k=1 u s(m) -u t 2(29)\n≤ KG\nT /K m=1 P T (m)(30)\n≤ KGP T ,(31)\nwhere we denote t = (m -1)K + k for brevity. Combining ( 23), ( 27), (31), we have\nD-Regret T (u 1 , • • • , u T ) ≤ GT √ 3 + 7 4η D 2 + η 2 KT G 2 + D η + KG P T = ( √ 3G + 7 4 D 2 + 1 2 G 2 + DP T )T 3/4 + GT 1/2 P T = O T 3/4 (1 + P T ) ,(32)\nwhere\nη = T -3/4 , K = η -2/3 = T 1/2 , and = η 2/3 = T -1/2 .\nAs shown in Lemma 9, each invocation of O IP requires solving at most O(T 1/2 ) linear optimizations with above parameter choice. Therefore, the total number of solving linear optimizations is N LOO = O(T ), since BOGD IP only uses T 1/2 calls to O IP with the block size K = T 1/2 ."
},
{
"figure_ref": [],
"heading": "B.2. Proof of Lemma 9",
"publication_ref": [
"b12"
],
"table_ref": [],
"text": "The proof can be found in Garber and Kretzu (2022), but for the sake of completeness, we present it in detail here.\nIn BOGD IP , at the end of block m -1, we invoke O IP as following:\nx m , ỹm = O IP (K, , x m-1 , y m ).(33)\nAccording to Lemma 8, we obtain that\nx m -ỹm 2 ≤ √ 3 , ỹ ∈ RB.\nAnd the number of calls to O IP is\nn IP = max x m -y m+1 2 2 ( x m -y m+1 2 2 -) 4 2 + 1, 1 .(34)\nAccording to Lemma 7, each invocation on O IP solves at most 27D 2"
},
{
"figure_ref": [],
"heading": "4",
"publication_ref": [],
"table_ref": [],
"text": "-2 linear optimizations. Hence, the number of solving linear optimizations is\nn LO = max x m -y m+1 2 2 ( x m -y m+1 2 2 -) 4 2 + 1, 1 • 27D 2 4 -2(35)\nAnd by the update step y m+1 = ỹm -η mK r=(m-1)K+1 ∇f r (x m ), we have\nx m -y m+1 2 = x m -ỹm + η mK r=(m-1)K+1 ∇f r (x m ) 2 (7) ≤ x m -ỹm 2 + ηKG (19) ≤ √ 3 + ηKG.(36)\nHence, according to (a + b) 2 ≤ 2a 2 + 2b 2 , we have\nx m -y m+1 2 2 ≤ ( √ 3 + ηKG) 2 ≤ 6 + 2η 2 K 2 G 2 . (37\n)\nTherefore, the number of solving linear optimizations is at most\nn LO ≤ (6 + 2η 2 K 2 G 2 )(6 + 2η 2 K 2 G 2 -) 4 2 + 1 • 27D 2 4 -2 ≤ 8.5 + 5.5 η 2 K 2 G 2 + η 4 K 4 G 4 2 27D 2 4 = O(T 1/2 ),(38)\nwhere\nη = T -3/4 , K = η -2/3 = T 1/2 and = η 2/3 = T -1/2 ."
},
{
"figure_ref": [],
"heading": "B.3. Proof of Lemma 10",
"publication_ref": [
"b46",
"b46"
],
"table_ref": [],
"text": "By the definition of P T , we have\nP T = T /K m=2 u s(m-1) -u s(m) 2 = u s(1) -u s(2) 2 + u s(2) -u s(3) 2 + • • • + u s(T /K-1) -u s(T /K) 2 . (39\n)\nAccording to triangle inequality, we have\nu s(1) -u s(2) 2 = u 1 -u K+1 2 ≤ u 1 -u 2 2 + • • • + u K -u K+1 2 u s(2) -u s(3) 2 = u K+1 -u 2K+1 2 ≤ u K+1 -u K+2 2 + • • • + u 2K -u 2K+1 2 • • • u s(T /K-1) -u s(T /K) 2 = u T -2K+1 -u T -K+1 2 ≤ u T -2K+1 -u T -2K+2 2 + • • • + u T -K -u T -K+1 2 . (40\n)\nSumming both side of (40), we have\nP T ≤ T -K+1 t=2 u t-1 -u t 2 ≤ T t=2 u t-1 -u t 2 = P T (41) B.4. Proof of Lemma 11\nThe analysis is inspired by Zinkevich (2003), but we consider a more general case since Lemma 11 can nearly reduce to Theorem 2 in Zinkevich (2003) by setting the block size K = 1.\nFor brevity, we denote t = (m -1)K + k and u s(m) as the first action in the block m where s(m) = (m -1)K + 1. In the block m, we have y m+1 = ỹm -η mK r=(m-1)K+1 ∇f r (x m ) = ỹm -η K k=1 ∇f t (x m ). Hence, we can prove that\nK k=1 ∇f t (x m ), ỹm -u s(m) = 1 η ỹm -y m+1 , ỹm -u s(m) = 1 2η ỹm -u s(m) 2 2 -y m+1 -u s(m) 2 2 + ỹt -y t+1 2 2 = 1 2η ỹm -u s(m) 2 2 -y m+1 -u s(m) 2 2 + η 2 K k=1 ∇f t (x m ) 2 2 (7) ≤ 1 2η ỹm -u s(m) 2 2 -ỹm+1 -u s(m) 2 2 + η 2 K 2 G 2 = 1 2η ỹm 2 2 -ỹm+1 2 2 + 1 η ỹm+1 -ỹm , u s(m) + η 2 K 2 G 2 . (42\n)\nLet P T be the path-length of the first action u s(m) in block m, i.e.,\nP T = T /K m=2 u s(m-1) -u s(m) 2 . (43\n)\nSumming up both side from m = 1 to T /K, we have\nB = T /K m=1 K k=1 ∇f t (x m ), ỹm -u s(m) ≤ 1 2η ỹ1 2 2 + 1 η T /K m=1 ỹm+1 -ỹm , u s(m) + η 2 KT G 2 ≤ 1 2η ỹ1 2 2 + 1 η ỹ T /K+1 u s(T /K) -ỹ 1 u 1 + 1 η T /K m=2 u s(m-1) -u s(m) , ỹt + η 2 KT G 2 ≤ 7 4η D 2 + η 2 KT G 2 + D η P T(44)\nThe last inequality is due to that ∀t, ỹt , u t ∈ RB (D = 2R)."
},
{
"figure_ref": [],
"heading": "B.5. Proof of Theorem 4",
"publication_ref": [],
"table_ref": [],
"text": "The key is to divide the dynamic regret into two terms, and upper bound them separately. First, we show that the dynamic regret can be decomposed as follows.\nT t=1\nf t (x t ) - T t=1 f t (u t ) = T t=1 f t (x t ) - T t=1 f t (x k t ) :=A + T t=1 f t (x k t ) - T t=1 f t (u t ) :=B ,(45)\nwhere k = 3 4 log 2 (1 + 4P T 7D ) + 1. To upper bound the term A of (45), we introduce the following lemma.\nLemma 12 (Lemma 1 in Zhang et al. (2018a)) Let C = 1 + 1 N and w i 1 = C i(i+1) for any expert i. Under Assumption 4, Algorithm 2 satisfies\nT t=1 f t (x t ) - T t=1 f t (x i t ) ≤ √ 2T 4 [1 + 2 ln(i + 1)].(46)\nAnd, the term A of (45) can be bounded as following\nA = T t=1 f t (x t ) - T t=1 f t (x k t ) ≤ √ 2T 4 [1 + 2 ln(k + 1)],(47)\nwhere k = 3 4 log 2 (1 + 4P T 7D ) + 1. Then, we proceed to upper bound the term B of (45). According to Theorem 2, for any expert i, we have\nT t=1 f t (x i t ) - T t=1 f t (u t ) ≤ GT √ 3 i + 7 4η i D 2 + η i 2 K i T G 2 + D η i + K i G P T ,(48)\nwhere i = η 2/3 i and K i = η -2/3 i . According to previous analysis, given a certain path-length PT , we can actually choose a certain step size η = O(T -3/4 (1+ PT ) 3/4 ), and achieve an improved bound. In the following, we introduce how to search this target step size η by maintaining multiple experts.\nTo facilitate computations, we assign the target step size as:\nη = 7D 2 + 4DP T 2G 2 T 3/4 .(49)\nNext, we show that η can actually be found by using multiple experts with different η ∈ H.\nBy the definition of P T , we have\n0 ≤ P T = T t=2 u t-1 -u t 2 ≤ T D.(50)\nHence, we have\n7D 2 2G 2 T 3/4 ≤ η ≤ 7D 2 + 4D 2 T 2G 2 T 3/4 .(51)\nIt can be verified that\nmin H ≤ 7D 2G 2 T 3/4 and 7D 2 + 4D 2 T 2G 2 T 3/4 ≤ max H.(52)\nThus, there exists expert k that satisfies\nη k = 2 k-1 7D 2 2G 2 T 3/4 ≤ η ≤ 2η k ,(53)\nwhere k = 3 4 log 2 (1 + 4P T 7D ) + 1. For expert k, we have\nB = T t=1 f t (x k t ) - T t=1 f t (u t ) ≤ GT √ 3 k + 7 4η k D 2 + η k 2 K k T G 2 + D η k + K k G P T = GT 3η 2/3 k + 7 4η k D 2 + η k 2 η -2/3 k T G 2 + D η k + η -2/3 k G P T (53) ≤ GT 3η 2/3 + 7 2η D 2 + 1 2 η1/3 T G 2 + 2D η + 2 2/3 η-2/3 G P T (49) ≤ (3G 3/2 + 2G 1/2 )T 3/4 (7D 2 + 4DP T ) 1/4 + 4G 2 T 1/2 P T (7D 2 + 4DP T ) 1/2 (54)\nTherefore, the dynamic regret of Algorithm 2 is upper bounded by\nT t=1 f t (x t ) - T t=1 f t (u t ) = T t=1 f t (x t ) - T t=1 f t (x k t ) + T t=1 f t (x k t ) - T t=1 f t (u t ) ≤ √ 2T 4 [1 + 2 ln(k + 1)] + (3G 3/2 + 2G 1/2 )T 3/4 (7D 2 + 4DP T ) 1/4 + 4G 2 T 1/2 P T (7D 2 + 4DP T ) 1/2 = O T 3/4 (1 + P T ) 1/4 . (55\n)\nwhere k = 3 4 log 2 (1 + 4P T 7D ) + 1."
},
{
"figure_ref": [],
"heading": "B.6. Proof of Theorem 5",
"publication_ref": [
"b7"
],
"table_ref": [],
"text": "The analysis is divided into two parts. First, we upper bound the strongly adaptive regret of Algorithm 3 over any interval J = [i, j] ∈ I. Then, we extend the regret bound over any interval\nI = [s, s + τ -1] ⊆ [T ]. For J = [i, j] ∈ I, we have j u=i f u (x u,J ) -min x∈K j u=i f u (x)(32)\n≤ G|J| √ 3 + 7 4η D 2 + η 2 K|J|G 2 ≤ √ 3G + 7 4 D 2 + 1 2 G 2 |J| 3/4 ,(56)\nwhere the first inequality is due to (32) with the static comparator (i.e., P T = 0) and the second inequality is due to η = |J| -3/4 , = η 2/3 = |J| -1/2 , K = η -2/3 = |J| -1/2 . First, we present the regret bound of meta algorithm (POLA) with respect to experts (BOGD IP ).\nLemma 13 (Lemma 8 in Zhang et al. ( 2020)) Under Assumption 4, for any interval J = [i, j] ∈ I, Algorithm 3 guarantees\nj u=i f u (x u ) - j u=i f u (x u,J ) ≤ 3c(j)|J|,(57)\nwhere c(j) ≤ 1 + ln j + ln(1 + log 2 j) + ln 5+3 ln(1+j) 2 .\nThen, combining (57) and ( 56), we can obtain that\nj u=i f u (x u ) -min x∈K j u=i f u (x) ≤ 3c(j)|J| 1/2 + √ 3G + 7 4 D 2 + 1 2 G 2 |J| 3/4 .(58)\nNow, we proceed to extend the regret bound over J to any interval\nI = [s, s + τ -1] ⊆ [T ].\nThe key of the extension is that I = [s, s + τ -1] can be partitioned into two sequences of intervals in GC, as shown below.\nLemma 14 (Lemma 1.2 of Daniely et al. (2015)) Any interval I = [s, s + τ -1] ⊆ [T ] can be into two sequences of disjoint and consecutive intervals,\nI -p , • • • , I 0 ∈ I and I 1 , • • • , I q ∈ I, which satisfy ∀i ≥ 1, |I -i |/|I -i+1 | ≤ 1/2 and ∀i ≥ 2, |I i |/|I i-1 | ≤ 1/2(59)\nAccording to Lemma 14, for any fixed x ∈ K, we can obtain that s+τ -1\nt=s f t (x t ) - s+τ -1 t=s f t (x) = q i=-p t∈I i f t (x t ) - t∈I i f t (x) (58) ≤ q i=-p 3c(s + τ -1)|I i | 1/2 + √ 3G + 7 4 D 2 + 1 2 G 2 |I i | 3/4 (59) ≤ 2 3c(s + τ -1) ∞ i=0 (2 -i τ ) 1/2 + 2 √ 3G + 7 4 D 2 + 1 2 G 2 ∞ i=0 (2 -i τ ) 3/4 ≤ 8 3c(s + τ -1)τ 1/2 + 6 √ 3G + 7 4 D 2 + 1 2 G 2 τ 3/4 .(60)\nTherefore, the strongly adaptive regret bound of Algorithm 3 is\nSA-R(T, τ ) = max [s,s+τ -1]⊆[T ] s+τ -1 t=s f t (x t ) -min x∈K s+τ -1 t=s f t (x) ≤ 8 3c(T )τ 1/2 + 6 √ 3G + 7 4 D 2 + 1 2 G 2 τ 3/4 = Õ τ 3/4 ,(61)\nwhere c(T ) ≤ 1 + ln T + ln(1 + log 2 T ) + ln 5+3 ln(1+T ) 2 ."
},
{
"figure_ref": [],
"heading": "B.7. Proof of Theorem 6",
"publication_ref": [
"b44"
],
"table_ref": [],
"text": "The analysis is similar to Zhang et al. (2020), and the key is to prove the dynamic regret bound on the experts\n{E I 1 k , E I 2 k , • • • } over several interval sets {I 1 k , I 2 k , • • • } ⊆ I. Specifically, due to P T = T t=2 u t-1 -u t 2 ∈ [0, DT ],(62)\nwe divide the path-length P T into two cases: P T ∈ [0, D] and P T ∈ (D, DT ], and establish the dynamic regret in the two cases separately."
},
{
"figure_ref": [],
"heading": "B.7.1. CASE 1: P",
"publication_ref": [],
"table_ref": [],
"text": "T ∈ [0, D] Note that I = I 0 ∪ I 1 ∪ • • • ∪ I α .(63)\nFor brevity, we denote α = log 2 T and I 1 j = [2 j , 2 j+1 -1] as the first interval of I j ⊆ {I 0 , • • • , I α }. We upper bound the dynamic regret over I 1 j (j = 0, • • • , α), as shown below\nT t=1 f t (x t ) - T t=1 f t (u t ) = α-1 j=0 2 j+1 -1 t=2 j f t (x t ) -f t (u t ) + T t=2 α f t (x t ) -f t (u t ) = α-1 j=0 2 j+1 -1 t=2 j f t (x t ) -f t (x t,I 1 j ) + T t=2 α f t (x t ) -f t (x t,I 1 α ) :=A + α-1 j=0 2 j+1 -1 t=2 j f t (x t,I 1 j ) -f t (u t ) + T t=2 α f t (x t,I 1 α ) -f t (u t ) :=B(64)\nWe proceed to bound the term A. By using Lemma 13, we have\nA (57) ≤ α-1 j=0 3c(2 j+1 -1)2 j-1 + 3c(T )(T -2 α ) ≤ 27c(T )2 α + 3c(T )(T -2 α ) ≤ 30c(T )T ,(65)\nwhere the second inequality is due to ∀j ≤ α -1, 2 j+1 -1 ≤ T , and the third inequality is due to Cauchy-Schwarz Inequality.\nFor simplicity, we denote P i:j = j t=i+1 u t-1 -u t 2 and the path-length P T can be decomposed as following: G 2 ) 2 j 3/4 + D 2 j 3/4 P 2 j :(2 j+1 -1) + G 2 j 1/2 P 2 j :(2 j+1 -1)\nP T = P\n+ ( √ 3G + 7 4 D 2 + 1 2 G 2 ) (T -2 α ) 3/4 + D (T -2 α ) 3/4 P 2 α :T + G (T -2 α ) 1/2 P 2 α :T ≤ 3( √ 3G + 7 4 D 2 + 1 2\nG 2 ) (2 α ) 3/4 + D (2 α ) 3/4 P 1:(2 α -1) + G (2 α ) 1/2 P 1:(2 α -1)\n+ ( √ 3G + 7 4 D 2 + 1 2 G 2 ) (T -2 α ) 3/4 + D (T -2 α ) 3/4 P 2 α :T + G (T -2 α ) 1/2 P 2 α :T ≤ 4(2 √ 3G + 7 2 D 2 + G 2 )T 3/4 + 2D 7/4 T 3/4 P 1/4 T + 2GD 1/2 T 1/2 P 1/2 T ,(67)\nwhere the last inequality is due to T 2 ≤ 2 α ≤ T and P T ≤ D. We complete the proof by combining ( 65) and (67)."
},
{
"figure_ref": [],
"heading": "B.7.2. CASE 2: P T ∈ (D, DT ]",
"publication_ref": [],
"table_ref": [],
"text": "For this case, we divide (D, DT ] as following\n(D2 0 , D2 1 ] δ 1 , (D2 1 , D2 2 ] δ 2 , • • • , (D2 i-1 , D2 i ] δ i , • • • , (D2 s-1 , D2 s ] δs ,(68)\nwhere s = log T . And for the interval δ i = (D2 i-1 , D2 i ], we analyze the dynamic regret bound over I 0 , • • • , I s-i . Specifically, we consider the first interval in I 0 , • • • , I s-i-1 , i.e., I 1 0 , • • • , I 1 s-i-1 and all intervals in I s-i , i.e., I 1 s-i , • • • , I u s-i , • • • , I m s-i (m = T /2 s-i -1). For brevity, the beginning of I u s-i is denoted as s u = u•2 s-i and the end of I u s-i is denoted as e u = (u+1)•2 s-i -1. Besides, by the fact that D2 i-1 ≤ P T ≤ D2 i , and s = log T , (69) we have\n2 s ≤ T, 2 s-i ≤ T, P T D ≤ T 2 s-i ≤ 2P T D . (70\n)\nSimilar to the case 1, the dynamic regret can be decomposed as below. (71)\nTo simplify the notation, we denote (71) as the sum of term A and term B, where term A is defined as \nA := s-i-1 j=0 2 j+1 -1 t=2 j f t (\nIn the following, we first upper bound the term A. By using Lemma 13, we have G 2 ) 2 s-i 3/4 + D 2 s-i 3/4 P s u :e u + G 2 s-i 1/2 P s u :e u + 2( √ 3G + 7 4 D 2 + 1 2 G 2 ) 2 s-i 3/4 + D 2 s-i 3/4 P s m :T + G 2 s-i 1/2 P s m :T ≤ 2 √ 3G + 7 4 D 2 + 1 2 G 2 (m + 1) 2 s-i 3/4 + D 2 s-i 3/4 P T + G 2 s-i 1/2 P T .\n(76)"
},
{
"figure_ref": [],
"heading": "Appendix A. Infeasible projection oracle",
"publication_ref": [
"b12",
"b12",
"b12"
],
"table_ref": [],
"text": "For the sake of completeness, we provide a brief description of infeasible projection oracle O IP , which can also be found in Garber and Kretzu (2022).\nDifferent from traditional projection operations, the infeasible projections consider projecting the original point back to an infeasible point, which could be still at the outside of the domain K, but is sufficiently close to K. Garber and Kretzu (2022) show that such a transform is effective, and can be efficiently achieved by calling to the infeasible projection oracle O IP over the domain K ⊆ RB with the error tolerance as follows.\nx, ỹ = O IP (K, , x 0 , y 0 ). Specifically, given the inputs Garber and Kretzu (2022) provide an implementation of O IP by solving linear optimizations, which repeats following two steps: (i) computing the feasible point x ∈ K that is sufficiently close to the infeasible point ỹ; (ii) \"pulling\" ỹ close to K based on the feasible point x ∈ K. In the following, we sketch this procedure.\nFirst, we briefly introduce how to obtain the action x ∈ K, which is also summarized in Algorithm 4. We compute the feasible action x from the initial point x 1 iteratively. This procedure can be viewed as a variant of Frank-Wolfe algorithm with line-search, and the loss function is\nStep 2 and Step 6-7). The main different lies in Step 3-5, which indicates the stop condition and takes responsibility for iterations. We can prove that the output x ∈ K ensures following theoretical guarantees.\nLemma 7 (Lemma 6 in Garber and Kretzu (2022)) Under Assumption 1, for a fixed error tolerance , Algorithm 4 stops after at most 27D 2 /4 -2 iterations, and the output x guarantees:\nThen, we proceed to compute an infeasible point ỹ after obtained x ∈ K (Algorithm 5). Specifically, we compute ỹ iteratively, and the loop stops until xỹ 2 2 ≤ 3 . In Step 8, the intermediate point y is updated along the direction yx with the step size γ. Notably, we can prove that the output ỹ gets closer to K during this iteration, which is summarized in following lemma.\nLemma 8 (Lemma 7 in Garber and Kretzu (2022)) Under Assumption 1, for a fixed error tolerance and γ = 2 x 0 -y 0 2 2 , Algorithm 5 stops after at most max{ x 0 -y 0 2 2 ( x 0 -y 0 2 2 -)/4 2 + 1, 1} iterations, and returns (x, ỹ) ∈ K × RB (D = 2R) which satisfy\nIt is noteworthy that the infeasible projection oracle is actually implemented based on the linear optimization over K (see Step 2 in Algorithm 4)."
},
{
"figure_ref": [],
"heading": "Note the fact that",
"publication_ref": [],
"table_ref": [],
"text": "Substituting ( 77) into (76), we have\nWe complete the proof by combining ( 74) and ( 78)."
}
] | Projection-free online learning has drawn increasing interest due to its efficiency in solving highdimensional problems with complicated constraints. However, most existing projection-free online methods focus on minimizing the static regret, which unfortunately fails to capture the challenge of changing environments. In this paper, we investigate non-stationary projection-free online learning, and choose dynamic regret and adaptive regret to measure the performance. Specifically, we first provide a novel dynamic regret analysis for an existing projection-free method named BOGD IP , and establish an O(T 3/4 (1 + P T )) dynamic regret bound, where P T denotes the path-length of the comparator sequence. Then, we improve the upper bound to O(T 3/4 (1 + P T ) 1/4 ) by running multiple BOGD IP algorithms with different step sizes in parallel, and tracking the best one on the fly. Our results are the first general-case dynamic regret bounds for projection-free online learning, and can recover the existing O(T 3/4 ) static regret by setting P T = 0. Furthermore, we propose a projection-free method to attain an Õ(τ 3/4 ) adaptive regret bound for any interval with length τ , which nearly matches the static regret over that interval. The essential idea is to maintain a set of BOGD IP algorithms dynamically, and combine them by a meta algorithm. Moreover, we demonstrate that it is also equipped with an O(T 3/4 (1 + P T ) 1/4 ) dynamic regret bound. Finally, empirical studies verify our theoretical findings. | Non-stationary Projection-free Online Learning with Dynamic and Adaptive Regret Guarantees | [
{
"figure_caption": "Figure 1 :1Figure 1: Geometric Covering (GC) intervals. In the figure, each interval is denoted by [ ].",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: Experimental results for dynamic regret minimization.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: Experimental results for adaptive regret minimization.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "2 0 :(2 1 -1) + • • • + P 2 j :(2 j+1 -1) + • • • + P 2 α :T (66)Then, we proceed to bound the term B.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "G+u )(e u -s u + 1) + 3c(T )(T -s m + 1)≤ 12c(T )s 1 + m-1 u=1 12c(T )(e u -s u + 1) + 12c(T )(T -s m + 1)≤ 12(m + 1)c(T )T ≤ 5 c(T )T 1 2 j 3/4 + D 2 j 3/4 P 2 j :(2 j+1 -1) + G 2 j 1/2 P 2 j :(2 j+1 -1) (e u -s u + 1) 3/4 + D (e u -s u + 1) 3/4 P s u :e u + (e u -s u + 1) 1/2 P s u :e u + ( (T -s m + 1) 3/4 + D (T -s m + 1) 3/4 P s m :T + G (T -s m + 1) 1/2 P s m :T 2 s-i 3/4 + D 2 s-i 3/4 P 1:e 0 + G 2 s-i 1/2 P 1:e 0",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Summary of existing methods in non-stationary projection-free online learning. Abbreviations: linear optimization → LO, membership operation → MO, worst-case dynamic",
"figure_data": "regret (4) → WD-R, general-case dynamic regret (2) → D-R, weak adaptive regret (5) →A-R, strongly adaptive regret (3) → SA-R.MethodLossOperation MetricBoundKalhan et al. (2021",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Algorithm 2 Projection-free Online Learning with Dynamic Regret (POLD) Input: A learning rate α, a set H containing step size η i for each expert E i Initialization: Activate a set of experts {E i | η i ∈ H} by invoking BOGD IP for each η i ∈ H.Let us consider a given sequence ũ1 , • • • , ũT ∈ K with the path-length PT = T t=2 ũt-1ũt 2 . According to Theorem 2, we can choose the step size η = O(T",
"figure_data": "1: For each expert E i , set w i 1 = C i(i+1) where C = 1 + 1 N2: for t = 1 to T do3:Receive x i t from each expert E i4:Compute x t according to (14)5:Submit x t , and update the weight w i t+1 for each expert E i according to (15)6:Send f t (•) to each expert E i7: end for",
"figure_id": "tab_1",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "x t,I 1",
"figure_data": "and term B is defined asB :=s-i-1 j=0 2 j+1 -1 t=2 jf t (x t ) -f t (x t,I 1 j) +u=1 m-1t=s u e uf t (x t ) -f t (x t,I m s-i )T+f t (x t ) -f t (x t,I m s-i ) ,t=s m+T) -f t (u t ) f (72) m-1 e u + f t (x t,I m s-i ) -f t (u t ) j u=1 t=s ut=s m",
"figure_id": "tab_3",
"figure_label": "",
"figure_type": "table"
}
] | Yibo Wang; Wenhao Yang; Wei Jiang; Shiyin Lu; Bing Wang; Haihong Tang; Yuanyu Wan; Lijun Zhang | [
{
"authors": "Dmitry Adamskiy; M Wouter; Alexey Koolen; Vladimir Chernov; Vovk",
"journal": "Journal of Machine Learning Research",
"ref_id": "b0",
"title": "A closer look at adaptive regret",
"year": "2016"
},
{
"authors": "Dheeraj Baby; Yu-Xiang Wang",
"journal": "",
"ref_id": "b1",
"title": "Online forecasting of total-variation-bounded sequences",
"year": "2019"
},
{
"authors": "Dheeraj Baby; Yu-Xiang Wang",
"journal": "",
"ref_id": "b2",
"title": "Optimal dynamic regret in exp-concave online learning",
"year": "2021"
},
{
"authors": "Dheeraj Baby; Yu-Xiang Wang",
"journal": "",
"ref_id": "b3",
"title": "Optimal dynamic regret in proper online learning with strongly convex losses and beyond",
"year": "2022"
},
{
"authors": "Omar Besbes; Yonatan Gur; Assaf J Zeevi",
"journal": "Operations Research",
"ref_id": "b4",
"title": "Non-stationary stochastic optimization",
"year": "2015"
},
{
"authors": "Chih-Chung Chang; Chih-Jen Lin",
"journal": "ACM Transactions on Intelligent Systems and Technology",
"ref_id": "b5",
"title": "Libsvm: A library for support vector machines",
"year": "2011"
},
{
"authors": "Ashok Cutkosky",
"journal": "",
"ref_id": "b6",
"title": "Parameter-free, dynamic, and strongly-adaptive online learning",
"year": "2020"
},
{
"authors": "Amit Daniely; Alon Gonen; Shai Shalev-Shwartz",
"journal": "",
"ref_id": "b7",
"title": "Strongly adaptive online learning",
"year": "2015"
},
{
"authors": "Yoav Freund; Robert E Schapire",
"journal": "Journal of Computer and System Sciences",
"ref_id": "b8",
"title": "A decision-theoretic generalization of on-line learning and an application to boosting",
"year": "1997"
},
{
"authors": "Yoav Freund; Robert E Schapire; Yoram Singer; Manfred K Warmuth",
"journal": "",
"ref_id": "b9",
"title": "Using and combining predictors that specialize",
"year": "1997"
},
{
"authors": "Dan Garber; Elad Hazan",
"journal": "SIAM Journal on Optimization",
"ref_id": "b10",
"title": "A linearly convergent conditional gradient algorithm with applications to online and stochastic optimization",
"year": "2016"
},
{
"authors": "Dan Garber; Ben Kretzu",
"journal": "",
"ref_id": "b11",
"title": "Improved regret bounds for projection-free bandit convex optimization",
"year": "2020"
},
{
"authors": "Dan Garber; Ben Kretzu",
"journal": "",
"ref_id": "b12",
"title": "New projection-free algorithms for online convex optimization with adaptive regret guarantees",
"year": "2022"
},
{
"authors": "Dan Garber; Ben Kretzu",
"journal": "",
"ref_id": "b13",
"title": "Projection-free online exp-concave optimization",
"year": "2023"
},
{
"authors": "András György; Tamás Linder; Gábor Lugosi",
"journal": "IEEE Transactions on Information Theory",
"ref_id": "b14",
"title": "Efficient tracking of large classes of experts",
"year": "2012"
},
{
"authors": "Elad Hazan",
"journal": "Foundations and Trends in Optimization",
"ref_id": "b15",
"title": "Introduction to online convex optimization",
"year": "2016"
},
{
"authors": "Elad Hazan; Satyen Kale",
"journal": "",
"ref_id": "b16",
"title": "Projection-free online learning",
"year": "2012"
},
{
"authors": "Elad Hazan; Edgar Minasyan",
"journal": "",
"ref_id": "b17",
"title": "Faster projection-free online learning",
"year": "2020"
},
{
"authors": "Elad Hazan; C Seshadhri",
"journal": "Electronic Colloquium on Computational Complexity",
"ref_id": "b18",
"title": "Adaptive algorithms for online decision problems",
"year": "2007"
},
{
"authors": "Ruitong Huang; Tor Lattimore; András György; Csaba Szepesvari",
"journal": "",
"ref_id": "b19",
"title": "Following the leader and fast rates in linear prediction: Curved constraint sets and other regularities",
"year": "2016"
},
{
"authors": "Ali Jadbabaie; Alexander Rakhlin; Shahin Shahrampour; Karthik Sridharan",
"journal": "",
"ref_id": "b20",
"title": "Online optimization: Competing with dynamic comparators",
"year": "2015"
},
{
"authors": "Kwang-Sung Jun; Francesco Orabona; Rebecca Willett; Stephen Wright",
"journal": "",
"ref_id": "b21",
"title": "Improved strongly adaptive online learning using coin betting",
"year": "2017"
},
{
"authors": "Kwang-Sung Jun; Francesco Orabona; Stephen Wright; Rebecca Willett",
"journal": "Electronic Journal of Statistics",
"ref_id": "b22",
"title": "Online learning for changing environments using coin betting",
"year": "2017"
},
{
"authors": "S Deepak; Amrit Singh Kalhan; Alec Bedi; Ketan Koppel; Hamed Rajawat; Hassani; K Abhishek; Adrish Gupta; Banerjee",
"journal": "IEEE Transactions on Signal Processing",
"ref_id": "b23",
"title": "Dynamic online learning via frank-wolfe algorithm",
"year": "2021"
},
{
"authors": "Ben Kretzu; Dan Garber",
"journal": "",
"ref_id": "b24",
"title": "Revisiting projection-free online learning: the strongly convex case",
"year": "2021"
},
{
"authors": "Kfir Levy; Andreas Krause",
"journal": "",
"ref_id": "b25",
"title": "Projection free online learning over smooth sets",
"year": "2019"
},
{
"authors": "Nick Littlestone; Manfred K Warmuth",
"journal": "Information and Computation",
"ref_id": "b26",
"title": "The weighted majority algorithm",
"year": "1994"
},
{
"authors": "Zhou Lu; Nataly Brukhim; Paula Gradu; Elad Hazan",
"journal": "",
"ref_id": "b27",
"title": "Projection-free adaptive regret with membership oracles",
"year": "2023"
},
{
"authors": "Haipeng Luo; Robert E Schapire",
"journal": "",
"ref_id": "b28",
"title": "Achieving all with no parameters: Adanormalhedge",
"year": "2015"
},
{
"authors": "Zakaria Mhammedi",
"journal": "",
"ref_id": "b29",
"title": "Efficient projection-free online convex optimization with membership oracle",
"year": "2022"
},
{
"authors": "Zakaria Mhammedi; M Wouter; Tim Koolen; Van Erven",
"journal": "",
"ref_id": "b30",
"title": "Lipschitz adaptivity with multiple learning rates in online learning",
"year": "2019"
},
{
"authors": "Aryan Mokhtari; Shahin Shahrampour; Ali Jadbabaie; Alejandro Ribeiro",
"journal": "",
"ref_id": "b31",
"title": "Online optimization in dynamic environments: Improved regret rates for strongly convex problems",
"year": "2016"
},
{
"authors": "Marco Molinaro",
"journal": "",
"ref_id": "b32",
"title": "Curvature of feasible sets in offline and online optimization",
"year": "2020"
},
{
"authors": "Shai Shalev-Shwartz",
"journal": "Foundations and Trends in Machine Learning",
"ref_id": "b33",
"title": "Online learning and online convex optimization",
"year": "2012"
},
{
"authors": "Tim Van Erven; Wouter M Koolen",
"journal": "",
"ref_id": "b34",
"title": "Metagrad: Multiple learning rates in online learning",
"year": "2016"
},
{
"authors": "Tim Van Erven; M Wouter; Dirk Koolen; Van Der Hoeven",
"journal": "Journal of Machine Learning Research",
"ref_id": "b35",
"title": "Metagrad: adaptation using multiple learning rates in online learning",
"year": "2021"
},
{
"authors": "Yuanyu Wan; Lijun Zhang",
"journal": "",
"ref_id": "b36",
"title": "Projection-free online learning over strongly convex set",
"year": "2021"
},
{
"authors": "Yuanyu Wan; Bo Xue; Lijun Zhang",
"journal": "",
"ref_id": "b37",
"title": "Projection-free online learning in dynamic environments",
"year": "2021"
},
{
"authors": "Yuanyu Wan; Lijun Zhang; Mingli Song",
"journal": "",
"ref_id": "b38",
"title": "Improved dynamic regret for online frank-wolfe",
"year": "2023"
},
{
"authors": "Frans Willems; Marco Krom",
"journal": "",
"ref_id": "b39",
"title": "Live-and-die coding for binary piecewise i.i.d. sources",
"year": "1997"
},
{
"authors": "Tianbao Yang; Lijun Zhang; Rong Jin; Jinfeng Yi",
"journal": "",
"ref_id": "b40",
"title": "Tracking slowly moving clairvoyant: Optimal dynamic regret of online learning with true and noisy gradient",
"year": "2016"
},
{
"authors": "Lijun Zhang; Shiyin Lu; Zhi-Hua Zhou",
"journal": "",
"ref_id": "b41",
"title": "Adaptive online learning in dynamic environments",
"year": "2018"
},
{
"authors": "Lijun Zhang; Tianbao Yang; Rong Jin; Zhi-Hua Zhou",
"journal": "",
"ref_id": "b42",
"title": "Dynamic regret of strongly adaptive methods",
"year": "2018"
},
{
"authors": "Lijun Zhang; Tie-Yan Liu; Zhi-Hua Zhou",
"journal": "",
"ref_id": "b43",
"title": "Adaptive regret of convex and smooth functions",
"year": "2019"
},
{
"authors": "Lijun Zhang; Shiyin Lu; Tianbao Yang",
"journal": "",
"ref_id": "b44",
"title": "Minimizing dynamic regret and adaptive regret simultaneously",
"year": "2020"
},
{
"authors": "Peng Zhao; Yan-Feng Xie; Lijun Zhang; Zhi-Hua Zhou",
"journal": "",
"ref_id": "b45",
"title": "Efficient methods for non-stationary online learning",
"year": "2022"
},
{
"authors": "Martin Zinkevich",
"journal": "",
"ref_id": "b46",
"title": "Online convex programming and generalized infinitesimal gradient ascent",
"year": "2003"
}
] | [
{
"formula_coordinates": [
2,
219.43,
226.29,
302.57,
33.58
],
"formula_id": "formula_0",
"formula_text": "Regret T = T t=1 f t (x t ) -min x∈K T t=1 f t (x),(1)"
},
{
"formula_coordinates": [
2,
191.81,
376.68,
330.19,
33.58
],
"formula_id": "formula_1",
"formula_text": "D-Regret T (u 1 , • • • , u T ) = T t=1 f t (x t ) - T t=1 f t (u t ),(2)"
},
{
"formula_coordinates": [
2,
251.2,
472.73,
109.61,
33.58
],
"formula_id": "formula_2",
"formula_text": "P T = T t=2 u t-1 -u t 2 ."
},
{
"formula_coordinates": [
2,
155.36,
555.06,
366.64,
33.35
],
"formula_id": "formula_3",
"formula_text": "SA-Regret T (τ ) = max [s,s+τ -1]⊆[T ] s+τ -1 t=s f t (x t ) -min x∈K s+τ -1 t=s f t (x) ,(3)"
},
{
"formula_coordinates": [
4,
100.63,
185.1,
408.01,
147.27
],
"formula_id": "formula_4",
"formula_text": ") smooth & convex LO WD-R O( √ T (1 + F T + √ D T )) Wan et al. (2023) smooth & convex LO WD-R O( T (1 + F T )) Wan et al. (2021) convex LO WD-R O(max{T 2/3 F 1/3 T , √ T }) strongly convex LO WD-R O(max{ √ T F T log T , log T }) BOGD IP (this work) convex LO D-R O(T 3/4 (1 + P T )) POLD (this work) convex LO D-R O(T 3/4 (1 + P T ) 1/4 ) POLA (this work) convex LO D-R O(T 3/4 (1 + P T ) 1/4 ) Garber and Kretzu (2022) convex LO A-R O(T 3/4 ) Lu et al. (2023) convex MO SA-R Õ( √ τ ) POLA (this work) convex LO SA-R Õ(τ 3/4 )"
},
{
"formula_coordinates": [
4,
192.74,
479.8,
329.26,
33.58
],
"formula_id": "formula_5",
"formula_text": "D-Regret T (u * 1 , • • • , u * T ) = T t=1 f t (x t ) - T t=1 f t (u * t )(4)"
},
{
"formula_coordinates": [
4,
90,
589.23,
432,
48.25
],
"formula_id": "formula_6",
"formula_text": "establish an O( √ T (1 + F T + √ D T )) worst-case bound, where F T = T t=2 sup x∈K |f t (x) -f t-1 (x)| and D T = T t=2 ∇f t (x t ) - ∇f t-1 (x t-1 ) 2 2 ."
},
{
"formula_coordinates": [
5,
186.92,
201.1,
335.08,
33.35
],
"formula_id": "formula_7",
"formula_text": "A-Regret T = max [s,e]⊆[T ] e t=s f t (x t ) -min x∈K e t=s f t (x) ,(5)"
},
{
"formula_coordinates": [
6,
243.24,
187.99,
278.76,
10.67
],
"formula_id": "formula_8",
"formula_text": "∀x, x ∈ K, x -x 2 ≤ D.(6)"
},
{
"formula_coordinates": [
6,
212.45,
212.14,
305.31,
33.76
],
"formula_id": "formula_9",
"formula_text": "f t (•) is G-Lipschitz over K, i.e., ∀x, y ∈ K, |f t (x) -f t (y)| ≤ G x -y 2 . (7"
},
{
"formula_coordinates": [
6,
517.76,
235.62,
4.24,
9.46
],
"formula_id": "formula_10",
"formula_text": ")"
},
{
"formula_coordinates": [
6,
203.26,
259.39,
314.5,
33.76
],
"formula_id": "formula_11",
"formula_text": "f t (•) is convex over K, i.e., ∀x, y ∈ K, f t (y) ≥ f t (x) + ∇f t (x) (y -x). (8"
},
{
"formula_coordinates": [
6,
517.76,
282.86,
4.24,
9.46
],
"formula_id": "formula_12",
"formula_text": ")"
},
{
"formula_coordinates": [
6,
251.45,
306.6,
270.55,
33.79
],
"formula_id": "formula_13",
"formula_text": "f t (x) belongs to [0, 1] for any x ∈ K, i.e., ∀x ∈ K, 0 ≤ f t (x) ≤ 1.(9)"
},
{
"formula_coordinates": [
6,
90,
448.22,
371.82,
33.91
],
"formula_id": "formula_14",
"formula_text": "x, ỹ = O IP (K, , x 0 , y 0 ), where (x, ỹ) ∈ K × RB, and x -ỹ 2 ≤ √ 3 and ∀z ∈ K, ỹ -z 2 ≤ y 0 -z 2 ."
},
{
"formula_coordinates": [
6,
217.73,
609.18,
299.73,
34.42
],
"formula_id": "formula_15",
"formula_text": "y m+1 = ỹm -η mK r=(m-1)K+1 ∇f r (x m ), (10"
},
{
"formula_coordinates": [
6,
517.46,
621.16,
4.54,
9.46
],
"formula_id": "formula_16",
"formula_text": ")"
},
{
"formula_coordinates": [
6,
223.24,
695.26,
294.22,
10.96
],
"formula_id": "formula_17",
"formula_text": "x m+1 , ỹm+1 = O IP (K, , x m , y m+1 ). (11"
},
{
"formula_coordinates": [
6,
517.46,
695.76,
4.54,
9.46
],
"formula_id": "formula_18",
"formula_text": ")"
},
{
"formula_coordinates": [
7,
96.12,
150.67,
161.69,
24.26
],
"formula_id": "formula_19",
"formula_text": "1: for t = 1 to T do 2: Submit x t = x m , observe f t (x t"
},
{
"formula_coordinates": [
7,
96.12,
177.77,
124.6,
23.01
],
"formula_id": "formula_20",
"formula_text": "if t mod K = 0 then 4:"
},
{
"formula_coordinates": [
7,
90,
389.17,
432,
54.4
],
"formula_id": "formula_21",
"formula_text": "Theorem 2 Let η = T -3/4 , K = η -2/3 = T 1/2 and = η 2/3 = T -1/2 . Under Assumptions 1, 2 and 3, Algorithm 1 guarantees D-Regret T (u 1 , • • • , u T ) ≤ O η 1/3 T + η -1 (1 + P T ) = O T 3/4 (1 + P T ) ."
},
{
"formula_coordinates": [
8,
194.45,
382.74,
327.55,
28.36
],
"formula_id": "formula_22",
"formula_text": "H = η i = 2 i-1 7D 2 2G 2 T 3/4 i = 1, • • • , N ,(13)"
},
{
"formula_coordinates": [
8,
272.51,
506.64,
249.5,
24.76
],
"formula_id": "formula_23",
"formula_text": "x t = i∈H w i t x i t ,(14)"
},
{
"formula_coordinates": [
8,
241.31,
560.83,
280.69,
31.83
],
"formula_id": "formula_24",
"formula_text": "w i t+1 = w i t e -αft(x i t ) µ∈H w µ t e -αft(x µ t ) ,(15)"
},
{
"formula_coordinates": [
8,
193.37,
691.3,
225.26,
14.35
],
"formula_id": "formula_25",
"formula_text": "D-Regret T (u 1 , • • • , u T ) ≤ O T 3/4 (1 + P T ) 1/4 ."
},
{
"formula_coordinates": [
9,
91.63,
243.47,
307.27,
73.93
],
"formula_id": "formula_26",
"formula_text": "E I ∈ A t , update R t,I = R t-1,I + f t (x t ) -f t (x t,I ) C t,I = C t-1,I + |f t (x t ) -f t (x t,I )| 11:"
},
{
"formula_coordinates": [
9,
92.31,
351.03,
425.55,
64.83
],
"formula_id": "formula_27",
"formula_text": "• • • I 0 [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] • • • I 1 [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ • • • I 2 [ ] [ ] [ ] [ ] [ • • • I 3 [ ] [ • • •"
},
{
"formula_coordinates": [
9,
173.63,
681.51,
343.83,
25.29
],
"formula_id": "formula_28",
"formula_text": "I = k∈N∪{0} I k , I k = [i • 2 k , (i + 1) • 2 k -1] : i ∈ N . (16"
},
{
"formula_coordinates": [
9,
517.46,
684.36,
4.54,
9.46
],
"formula_id": "formula_29",
"formula_text": ")"
},
{
"formula_coordinates": [
10,
90,
174.32,
121.74,
13.65
],
"formula_id": "formula_30",
"formula_text": "Φ(R, C) = exp ([R] 2 + /3C)"
},
{
"formula_coordinates": [
10,
185.64,
226.53,
240.72,
24.43
],
"formula_id": "formula_31",
"formula_text": "w(R, C) = 1 2 (Φ(R + 1, C + 1) -Φ(R + 1, C -1)) ."
},
{
"formula_coordinates": [
10,
90,
355.64,
174.19,
15.24
],
"formula_id": "formula_32",
"formula_text": "C t-1,I = t-1 u=min I |f t (x t ) -f t (x t,I )"
},
{
"formula_coordinates": [
10,
227.21,
424.5,
290.25,
28.15
],
"formula_id": "formula_33",
"formula_text": "w t,I = w(R t-1,I , C t-1,I ) E I ∈At w(R t-1,I , C t-1,I ) . (17"
},
{
"formula_coordinates": [
10,
517.46,
432.34,
4.54,
9.46
],
"formula_id": "formula_34",
"formula_text": ")"
},
{
"formula_coordinates": [
10,
260.69,
492.81,
261.31,
23.25
],
"formula_id": "formula_35",
"formula_text": "x t = E I ∈At w t,I x t,I ,(18)"
},
{
"formula_coordinates": [
10,
187.45,
608.94,
237.11,
14.61
],
"formula_id": "formula_36",
"formula_text": "SA-Regret T (τ ) ≤ O( τ log T + τ 3/4 ) = Õ τ 3/4 ."
},
{
"formula_coordinates": [
11,
193.37,
459.84,
225.26,
14.35
],
"formula_id": "formula_37",
"formula_text": "D-Regret T (u 1 , • • • , u T ) ≤ O T 3/4 (1 + P T ) 1/4 ."
},
{
"formula_coordinates": [
12,
238.5,
329.05,
135.01,
22.79
],
"formula_id": "formula_38",
"formula_text": "f t (X) = (i,j)∈OB |X ij -M ij |."
},
{
"formula_coordinates": [
12,
90,
417.85,
99.52,
14.27
],
"formula_id": "formula_39",
"formula_text": "{(i k , j k , M i k j k )} 300000 k=1"
},
{
"formula_coordinates": [
13,
201.25,
207.25,
209.49,
37.26
],
"formula_id": "formula_40",
"formula_text": "f t (X) = log 1 + j =lt exp x j e t -x lt e t ."
},
{
"formula_coordinates": [
13,
90,
376.03,
432,
27.47
],
"formula_id": "formula_41",
"formula_text": "η i = c 2 i-1 -1/2 in PAMO and η i = c 2 i-1 -3/4 in POLA, where c is the hyper-parameters selected from {2 -4 , 2 -3 , • • • , 2 4 }."
},
{
"formula_coordinates": [
19,
96.12,
123.57,
247.31,
63.66
],
"formula_id": "formula_42",
"formula_text": "1: for i = 1, • • • do 2: Compute v i = arg min x∈K x i -y, x 3: if (x i -y) (x i -v i ) ≤ or x i -y 2 2 ≤ 3 then 4: Return x ← x i 5:"
},
{
"formula_coordinates": [
19,
90,
189.44,
274.71,
68.12
],
"formula_id": "formula_43",
"formula_text": "δ i = arg min δ∈[0,1] { x i + δ(v i -x i ) -y 2 2 } 7: Set x i+1 = x i + δ i (v i -x i ) 8: end for Algorithm 5 Infeasible Projection Oracle, O IP (Algorithm 4 in"
},
{
"formula_coordinates": [
19,
90,
260.88,
331.8,
133.59
],
"formula_id": "formula_44",
"formula_text": "Input: Domain set K, error tolerance , initial point x 0 ∈ K, initial point y 0 1: y 1 = y 0 / max{1, y 2 /R} 2: if x 0 -y 0 2 2 ≤ 3 then 3: Return (x, ỹ) ← (x 0 , y 1 ) 4: end if 5: for i = 1, • • • do 6: x i ← Algorithm 4 (K, , x i-1 , y i ) 7: if x i -y i 2 2 > 3 then 8: y i+1 = y i -γ(y i -x i ) 9:"
},
{
"formula_coordinates": [
19,
105.67,
561.96,
416.33,
105.76
],
"formula_id": "formula_45",
"formula_text": "T t=1 f t (x t ) - T t=1 f t (u t ) = T /K m=1 K k=1 [f t (x m ) -f t (u t )] (8) ≤ T /K m=1 K k=1 ∇f t (x m ), x m -ỹm :=A + T /K m=1 K k=1 ∇f t (x m ), ỹm -u s(m) :=B + T /K m=1 K k=1 ∇f t (x m ), u s(m) -u t :=C ,(20)"
},
{
"formula_coordinates": [
20,
233.84,
169.7,
288.16,
10.78
],
"formula_id": "formula_46",
"formula_text": "x m , ỹm = O IP (K, , x m-1 , y m ),(21)"
},
{
"formula_coordinates": [
20,
264.08,
210.34,
253.37,
20.19
],
"formula_id": "formula_47",
"formula_text": "x m -ỹm 2 ≤ √ 3 . (22"
},
{
"formula_coordinates": [
20,
517.46,
220.24,
4.54,
9.46
],
"formula_id": "formula_48",
"formula_text": ")"
},
{
"formula_coordinates": [
20,
192.57,
316.91,
329.43,
35
],
"formula_id": "formula_49",
"formula_text": "A = T /K m=1 K k=1 ∇f t (x m ), x m -ỹm (7),(22) ≤ GT √ 3 .(23)"
},
{
"formula_coordinates": [
20,
234.99,
404.33,
287.02,
34.6
],
"formula_id": "formula_50",
"formula_text": "P T = T /K m=2 u s(m-1) -u s(m) 2 ,(24)"
},
{
"formula_coordinates": [
20,
283.6,
500.86,
238.4,
11.69
],
"formula_id": "formula_51",
"formula_text": "P T ≤ P T .(25)"
},
{
"formula_coordinates": [
20,
152.27,
602.72,
369.73,
35
],
"formula_id": "formula_52",
"formula_text": "B = T /K m=1 K k=1 ∇f t (x m ), ỹm -u s(m) ≤ 7 4η D 2 + η 2 KT G 2 + D η P T .(26)"
},
{
"formula_coordinates": [
20,
152.27,
673.79,
369.73,
35
],
"formula_id": "formula_53",
"formula_text": "B = T /K m=1 K k=1 ∇f t (x m ), ỹm -u s(m) ≤ 7 4η D 2 + η 2 KT G 2 + D η P T .(27)"
},
{
"formula_coordinates": [
21,
201.41,
131.45,
316.04,
33.98
],
"formula_id": "formula_54",
"formula_text": "P T (m) = K k=2 u (m-1)K+k-1 -u (m-1)K+k 2 . (28"
},
{
"formula_coordinates": [
21,
517.46,
143.43,
4.54,
9.46
],
"formula_id": "formula_55",
"formula_text": ")"
},
{
"formula_coordinates": [
21,
112.29,
178.34,
409.71,
37.8
],
"formula_id": "formula_56",
"formula_text": "≤ k ≤ K) in block m, we have u s(m) -u t 2 = u s(m) -u (m-1)K+k 2 = u (m-1)K+1 -u (m-1)K+k 2 ≤ P T (m),(29)"
},
{
"formula_coordinates": [
21,
265.42,
271.45,
256.58,
34.6
],
"formula_id": "formula_57",
"formula_text": "T /K m=1 P T (m) ≤ P T .(30)"
},
{
"formula_coordinates": [
21,
164.41,
337.92,
160.73,
75.16
],
"formula_id": "formula_58",
"formula_text": "C = T /K m=1 K k=1 ∇f t (x m ), u s(m) -u t (7) ≤ G T /K m=1 K k=1 u s(m) -u t 2(29)"
},
{
"formula_coordinates": [
21,
341.19,
378.08,
68.11,
34.6
],
"formula_id": "formula_59",
"formula_text": "T /K m=1 P T (m)(30)"
},
{
"formula_coordinates": [
21,
398.41,
370.09,
123.59,
31.33
],
"formula_id": "formula_60",
"formula_text": "≤ KGP T ,(31)"
},
{
"formula_coordinates": [
21,
136.41,
461.01,
385.59,
73.01
],
"formula_id": "formula_61",
"formula_text": "D-Regret T (u 1 , • • • , u T ) ≤ GT √ 3 + 7 4η D 2 + η 2 KT G 2 + D η + KG P T = ( √ 3G + 7 4 D 2 + 1 2 G 2 + DP T )T 3/4 + GT 1/2 P T = O T 3/4 (1 + P T ) ,(32)"
},
{
"formula_coordinates": [
21,
119.38,
551.05,
246.17,
11.76
],
"formula_id": "formula_62",
"formula_text": "η = T -3/4 , K = η -2/3 = T 1/2 , and = η 2/3 = T -1/2 ."
},
{
"formula_coordinates": [
21,
234.06,
695.26,
287.94,
10.96
],
"formula_id": "formula_63",
"formula_text": "x m , ỹm = O IP (K, , x m-1 , y m ).(33)"
},
{
"formula_coordinates": [
22,
238.21,
112.83,
141.04,
20.19
],
"formula_id": "formula_64",
"formula_text": "x m -ỹm 2 ≤ √ 3 , ỹ ∈ RB."
},
{
"formula_coordinates": [
22,
172.12,
174.66,
349.88,
26.38
],
"formula_id": "formula_65",
"formula_text": "n IP = max x m -y m+1 2 2 ( x m -y m+1 2 2 -) 4 2 + 1, 1 .(34)"
},
{
"formula_coordinates": [
22,
139.39,
259.3,
382.61,
26.38
],
"formula_id": "formula_66",
"formula_text": "n LO = max x m -y m+1 2 2 ( x m -y m+1 2 2 -) 4 2 + 1, 1 • 27D 2 4 -2(35)"
},
{
"formula_coordinates": [
22,
180.51,
337.53,
341.49,
59.75
],
"formula_id": "formula_67",
"formula_text": "x m -y m+1 2 = x m -ỹm + η mK r=(m-1)K+1 ∇f r (x m ) 2 (7) ≤ x m -ỹm 2 + ηKG (19) ≤ √ 3 + ηKG.(36)"
},
{
"formula_coordinates": [
22,
192.26,
433.74,
325.2,
21.25
],
"formula_id": "formula_68",
"formula_text": "x m -y m+1 2 2 ≤ ( √ 3 + ηKG) 2 ≤ 6 + 2η 2 K 2 G 2 . (37"
},
{
"formula_coordinates": [
22,
517.46,
443.64,
4.54,
9.46
],
"formula_id": "formula_69",
"formula_text": ")"
},
{
"formula_coordinates": [
22,
155.5,
496.29,
366.5,
57.21
],
"formula_id": "formula_70",
"formula_text": "n LO ≤ (6 + 2η 2 K 2 G 2 )(6 + 2η 2 K 2 G 2 -) 4 2 + 1 • 27D 2 4 -2 ≤ 8.5 + 5.5 η 2 K 2 G 2 + η 4 K 4 G 4 2 27D 2 4 = O(T 1/2 ),(38)"
},
{
"formula_coordinates": [
22,
119.38,
570.26,
243.45,
11.76
],
"formula_id": "formula_71",
"formula_text": "η = T -3/4 , K = η -2/3 = T 1/2 and = η 2/3 = T -1/2 ."
},
{
"formula_coordinates": [
22,
134.98,
656.51,
382.47,
51.38
],
"formula_id": "formula_72",
"formula_text": "P T = T /K m=2 u s(m-1) -u s(m) 2 = u s(1) -u s(2) 2 + u s(2) -u s(3) 2 + • • • + u s(T /K-1) -u s(T /K) 2 . (39"
},
{
"formula_coordinates": [
22,
517.46,
677.24,
4.54,
9.46
],
"formula_id": "formula_73",
"formula_text": ")"
},
{
"formula_coordinates": [
23,
113.04,
124.83,
404.42,
113.38
],
"formula_id": "formula_74",
"formula_text": "u s(1) -u s(2) 2 = u 1 -u K+1 2 ≤ u 1 -u 2 2 + • • • + u K -u K+1 2 u s(2) -u s(3) 2 = u K+1 -u 2K+1 2 ≤ u K+1 -u K+2 2 + • • • + u 2K -u 2K+1 2 • • • u s(T /K-1) -u s(T /K) 2 = u T -2K+1 -u T -K+1 2 ≤ u T -2K+1 -u T -2K+2 2 + • • • + u T -K -u T -K+1 2 . (40"
},
{
"formula_coordinates": [
23,
517.46,
176.56,
4.54,
9.46
],
"formula_id": "formula_75",
"formula_text": ")"
},
{
"formula_coordinates": [
23,
90,
287.22,
432,
71.53
],
"formula_id": "formula_76",
"formula_text": "P T ≤ T -K+1 t=2 u t-1 -u t 2 ≤ T t=2 u t-1 -u t 2 = P T (41) B.4. Proof of Lemma 11"
},
{
"formula_coordinates": [
23,
158.02,
464.87,
359.44,
157.8
],
"formula_id": "formula_77",
"formula_text": "K k=1 ∇f t (x m ), ỹm -u s(m) = 1 η ỹm -y m+1 , ỹm -u s(m) = 1 2η ỹm -u s(m) 2 2 -y m+1 -u s(m) 2 2 + ỹt -y t+1 2 2 = 1 2η ỹm -u s(m) 2 2 -y m+1 -u s(m) 2 2 + η 2 K k=1 ∇f t (x m ) 2 2 (7) ≤ 1 2η ỹm -u s(m) 2 2 -ỹm+1 -u s(m) 2 2 + η 2 K 2 G 2 = 1 2η ỹm 2 2 -ỹm+1 2 2 + 1 η ỹm+1 -ỹm , u s(m) + η 2 K 2 G 2 . (42"
},
{
"formula_coordinates": [
23,
517.46,
538.86,
4.54,
9.46
],
"formula_id": "formula_78",
"formula_text": ")"
},
{
"formula_coordinates": [
23,
234.99,
674.19,
282.47,
34.6
],
"formula_id": "formula_79",
"formula_text": "P T = T /K m=2 u s(m-1) -u s(m) 2 . (43"
},
{
"formula_coordinates": [
23,
517.46,
687.19,
4.54,
9.46
],
"formula_id": "formula_80",
"formula_text": ")"
},
{
"formula_coordinates": [
24,
102.25,
118.41,
419.75,
155.32
],
"formula_id": "formula_81",
"formula_text": "B = T /K m=1 K k=1 ∇f t (x m ), ỹm -u s(m) ≤ 1 2η ỹ1 2 2 + 1 η T /K m=1 ỹm+1 -ỹm , u s(m) + η 2 KT G 2 ≤ 1 2η ỹ1 2 2 + 1 η ỹ T /K+1 u s(T /K) -ỹ 1 u 1 + 1 η T /K m=2 u s(m-1) -u s(m) , ỹt + η 2 KT G 2 ≤ 7 4η D 2 + η 2 KT G 2 + D η P T(44)"
},
{
"formula_coordinates": [
24,
151.41,
372.47,
370.59,
48.88
],
"formula_id": "formula_82",
"formula_text": "f t (x t ) - T t=1 f t (u t ) = T t=1 f t (x t ) - T t=1 f t (x k t ) :=A + T t=1 f t (x k t ) - T t=1 f t (u t ) :=B ,(45)"
},
{
"formula_coordinates": [
24,
198.72,
504.06,
323.28,
38.57
],
"formula_id": "formula_83",
"formula_text": "T t=1 f t (x t ) - T t=1 f t (x i t ) ≤ √ 2T 4 [1 + 2 ln(i + 1)].(46)"
},
{
"formula_coordinates": [
24,
183.6,
572.65,
338.4,
38.57
],
"formula_id": "formula_84",
"formula_text": "A = T t=1 f t (x t ) - T t=1 f t (x k t ) ≤ √ 2T 4 [1 + 2 ln(k + 1)],(47)"
},
{
"formula_coordinates": [
24,
130.01,
675.22,
391.99,
33.58
],
"formula_id": "formula_85",
"formula_text": "T t=1 f t (x i t ) - T t=1 f t (u t ) ≤ GT √ 3 i + 7 4η i D 2 + η i 2 K i T G 2 + D η i + K i G P T ,(48)"
},
{
"formula_coordinates": [
25,
247.73,
177.63,
274.27,
28.36
],
"formula_id": "formula_86",
"formula_text": "η = 7D 2 + 4DP T 2G 2 T 3/4 .(49)"
},
{
"formula_coordinates": [
25,
225.31,
264.72,
296.69,
33.58
],
"formula_id": "formula_87",
"formula_text": "0 ≤ P T = T t=2 u t-1 -u t 2 ≤ T D.(50)"
},
{
"formula_coordinates": [
25,
221.15,
327.51,
300.85,
28.36
],
"formula_id": "formula_88",
"formula_text": "7D 2 2G 2 T 3/4 ≤ η ≤ 7D 2 + 4D 2 T 2G 2 T 3/4 .(51)"
},
{
"formula_coordinates": [
25,
164.75,
393.18,
357.25,
28.36
],
"formula_id": "formula_89",
"formula_text": "min H ≤ 7D 2G 2 T 3/4 and 7D 2 + 4D 2 T 2G 2 T 3/4 ≤ max H.(52)"
},
{
"formula_coordinates": [
25,
225.74,
465.09,
296.26,
28.36
],
"formula_id": "formula_90",
"formula_text": "η k = 2 k-1 7D 2 2G 2 T 3/4 ≤ η ≤ 2η k ,(53)"
},
{
"formula_coordinates": [
25,
132.51,
553.69,
389.5,
155.15
],
"formula_id": "formula_91",
"formula_text": "B = T t=1 f t (x k t ) - T t=1 f t (u t ) ≤ GT √ 3 k + 7 4η k D 2 + η k 2 K k T G 2 + D η k + K k G P T = GT 3η 2/3 k + 7 4η k D 2 + η k 2 η -2/3 k T G 2 + D η k + η -2/3 k G P T (53) ≤ GT 3η 2/3 + 7 2η D 2 + 1 2 η1/3 T G 2 + 2D η + 2 2/3 η-2/3 G P T (49) ≤ (3G 3/2 + 2G 1/2 )T 3/4 (7D 2 + 4DP T ) 1/4 + 4G 2 T 1/2 P T (7D 2 + 4DP T ) 1/2 (54)"
},
{
"formula_coordinates": [
26,
101.2,
116.48,
416.26,
128.03
],
"formula_id": "formula_92",
"formula_text": "T t=1 f t (x t ) - T t=1 f t (u t ) = T t=1 f t (x t ) - T t=1 f t (x k t ) + T t=1 f t (x k t ) - T t=1 f t (u t ) ≤ √ 2T 4 [1 + 2 ln(k + 1)] + (3G 3/2 + 2G 1/2 )T 3/4 (7D 2 + 4DP T ) 1/4 + 4G 2 T 1/2 P T (7D 2 + 4DP T ) 1/2 = O T 3/4 (1 + P T ) 1/4 . (55"
},
{
"formula_coordinates": [
26,
517.46,
235.05,
4.54,
9.46
],
"formula_id": "formula_93",
"formula_text": ")"
},
{
"formula_coordinates": [
26,
90,
334.37,
219.39,
70.14
],
"formula_id": "formula_94",
"formula_text": "I = [s, s + τ -1] ⊆ [T ]. For J = [i, j] ∈ I, we have j u=i f u (x u,J ) -min x∈K j u=i f u (x)(32)"
},
{
"formula_coordinates": [
26,
296.11,
372.87,
225.89,
60.59
],
"formula_id": "formula_95",
"formula_text": "≤ G|J| √ 3 + 7 4η D 2 + η 2 K|J|G 2 ≤ √ 3G + 7 4 D 2 + 1 2 G 2 |J| 3/4 ,(56)"
},
{
"formula_coordinates": [
26,
215.54,
529.78,
306.46,
34.29
],
"formula_id": "formula_96",
"formula_text": "j u=i f u (x u ) - j u=i f u (x u,J ) ≤ 3c(j)|J|,(57)"
},
{
"formula_coordinates": [
26,
130.29,
622.54,
391.72,
34.29
],
"formula_id": "formula_97",
"formula_text": "j u=i f u (x u ) -min x∈K j u=i f u (x) ≤ 3c(j)|J| 1/2 + √ 3G + 7 4 D 2 + 1 2 G 2 |J| 3/4 .(58)"
},
{
"formula_coordinates": [
26,
405.98,
668.31,
116.02,
9.57
],
"formula_id": "formula_98",
"formula_text": "I = [s, s + τ -1] ⊆ [T ]."
},
{
"formula_coordinates": [
27,
90,
107.92,
432,
50.44
],
"formula_id": "formula_99",
"formula_text": "I -p , • • • , I 0 ∈ I and I 1 , • • • , I q ∈ I, which satisfy ∀i ≥ 1, |I -i |/|I -i+1 | ≤ 1/2 and ∀i ≥ 2, |I i |/|I i-1 | ≤ 1/2(59)"
},
{
"formula_coordinates": [
27,
134.78,
196.97,
387.22,
180.33
],
"formula_id": "formula_100",
"formula_text": "t=s f t (x t ) - s+τ -1 t=s f t (x) = q i=-p t∈I i f t (x t ) - t∈I i f t (x) (58) ≤ q i=-p 3c(s + τ -1)|I i | 1/2 + √ 3G + 7 4 D 2 + 1 2 G 2 |I i | 3/4 (59) ≤ 2 3c(s + τ -1) ∞ i=0 (2 -i τ ) 1/2 + 2 √ 3G + 7 4 D 2 + 1 2 G 2 ∞ i=0 (2 -i τ ) 3/4 ≤ 8 3c(s + τ -1)τ 1/2 + 6 √ 3G + 7 4 D 2 + 1 2 G 2 τ 3/4 .(60)"
},
{
"formula_coordinates": [
27,
135.44,
413.22,
386.56,
62.29
],
"formula_id": "formula_101",
"formula_text": "SA-R(T, τ ) = max [s,s+τ -1]⊆[T ] s+τ -1 t=s f t (x t ) -min x∈K s+τ -1 t=s f t (x) ≤ 8 3c(T )τ 1/2 + 6 √ 3G + 7 4 D 2 + 1 2 G 2 τ 3/4 = Õ τ 3/4 ,(61)"
},
{
"formula_coordinates": [
27,
106.94,
552.02,
415.06,
61.36
],
"formula_id": "formula_102",
"formula_text": "{E I 1 k , E I 2 k , • • • } over several interval sets {I 1 k , I 2 k , • • • } ⊆ I. Specifically, due to P T = T t=2 u t-1 -u t 2 ∈ [0, DT ],(62)"
},
{
"formula_coordinates": [
27,
90,
662.08,
432,
43.96
],
"formula_id": "formula_103",
"formula_text": "T ∈ [0, D] Note that I = I 0 ∪ I 1 ∪ • • • ∪ I α .(63)"
},
{
"formula_coordinates": [
28,
104.23,
129.88,
417.77,
169.66
],
"formula_id": "formula_104",
"formula_text": "T t=1 f t (x t ) - T t=1 f t (u t ) = α-1 j=0 2 j+1 -1 t=2 j f t (x t ) -f t (u t ) + T t=2 α f t (x t ) -f t (u t ) = α-1 j=0 2 j+1 -1 t=2 j f t (x t ) -f t (x t,I 1 j ) + T t=2 α f t (x t ) -f t (x t,I 1 α ) :=A + α-1 j=0 2 j+1 -1 t=2 j f t (x t,I 1 j ) -f t (u t ) + T t=2 α f t (x t,I 1 α ) -f t (u t ) :=B(64)"
},
{
"formula_coordinates": [
28,
192.42,
330.24,
329.58,
70.42
],
"formula_id": "formula_105",
"formula_text": "A (57) ≤ α-1 j=0 3c(2 j+1 -1)2 j-1 + 3c(T )(T -2 α ) ≤ 27c(T )2 α + 3c(T )(T -2 α ) ≤ 30c(T )T ,(65)"
},
{
"formula_coordinates": [
28,
191.37,
475.98,
35.16,
10.68
],
"formula_id": "formula_106",
"formula_text": "P T = P"
},
{
"formula_coordinates": [
28,
106.11,
553.71,
378.15,
52.49
],
"formula_id": "formula_107",
"formula_text": "+ ( √ 3G + 7 4 D 2 + 1 2 G 2 ) (T -2 α ) 3/4 + D (T -2 α ) 3/4 P 2 α :T + G (T -2 α ) 1/2 P 2 α :T ≤ 3( √ 3G + 7 4 D 2 + 1 2"
},
{
"formula_coordinates": [
28,
106.11,
605.46,
415.9,
63.21
],
"formula_id": "formula_108",
"formula_text": "+ ( √ 3G + 7 4 D 2 + 1 2 G 2 ) (T -2 α ) 3/4 + D (T -2 α ) 3/4 P 2 α :T + G (T -2 α ) 1/2 P 2 α :T ≤ 4(2 √ 3G + 7 2 D 2 + G 2 )T 3/4 + 2D 7/4 T 3/4 P 1/4 T + 2GD 1/2 T 1/2 P 1/2 T ,(67)"
},
{
"formula_coordinates": [
29,
138.16,
140.01,
383.84,
29.35
],
"formula_id": "formula_109",
"formula_text": "(D2 0 , D2 1 ] δ 1 , (D2 1 , D2 2 ] δ 2 , • • • , (D2 i-1 , D2 i ] δ i , • • • , (D2 s-1 , D2 s ] δs ,(68)"
},
{
"formula_coordinates": [
29,
208.16,
306.09,
309.3,
24.43
],
"formula_id": "formula_110",
"formula_text": "2 s ≤ T, 2 s-i ≤ T, P T D ≤ T 2 s-i ≤ 2P T D . (70"
},
{
"formula_coordinates": [
29,
517.46,
313.82,
4.54,
9.46
],
"formula_id": "formula_111",
"formula_text": ")"
},
{
"formula_coordinates": [
29,
115.87,
630.66,
112.34,
38.39
],
"formula_id": "formula_112",
"formula_text": "A := s-i-1 j=0 2 j+1 -1 t=2 j f t ("
}
] |
|
2023-05-19 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b44",
"b23",
"b33",
"b36",
"b38",
"b38",
"b23",
"b22"
],
"table_ref": [],
"text": "Video saliency detection is the task of estimating human eye fixations when perceiving dynamic scenes. The problem of modeling human attention has gained more and more interest over the recent years due to its important contribution in a variety of applications such as video summarization, video compression, virtual reality and robotics. In parallel, the development of Deep Learning techniques and especially Convolutional Neural Networks (CNNs) has helped achieve remarkable results in various computer vision problems such as image segmentation, classification, saliency estimation etc.\nVideo saliency estimation is considered to be a more challenging task compared to static image saliency prediction. That is because in videos we need to frames RGB RGBD depth Fig. 1. Frames with their eye-tracking data from a Hollywood movie, along with the frames' estimated depth. The third row depicts saliency maps produced by RGB-only saliency network, while the last row is the output of our proposed ViDaS RGB-D network, which succeeds in better predicting human attention.\naccurately extract both spatial and temporal features and fuse them effectively in order to obtain a final saliency map. Previously introduced methods widely depend on LSTMs and Convolutional Neural Networks [45] in order to compute the spatio-temporal features and are mostly focusing on the visual information provided by the various eye-tracking datasets.\nHuman attention and perception is however deeply affected by many different cues that can be present in a video scene and awake various human senses. Recent research and studies [24,34,37,39] indicate that such cues can be the depth information as well as the audio. As illustrated by Fig. 1 incorporating such information can assist in locating salient regions.\nDepth information is inseparably connected to visual stimuli since the human brain has the ability to detect the different objects present in a video scene or image and estimate their relative distance. This cue is naturally perceived and processed by the brain and this led us to the idea of integrating depth and RGB information in a single network to assess whether it could improve saliency prediction in videos. Moreover, nowdays depth can be easily captured either by depth sensors, that have started appearing in many everyday life devices such as mobile phones, or by employing modern deep learning methods. Fig. 1 highlights and motivates the problem we investigate. A person is coming towards another person in a movies scene. This information is clearly captured by depth, and is mirrored in the results, where RGB alone could not predict attention as well as the combination of RGB and Depth does.\nWe propose ViDaS, a video depth-aware saliency network that efficiently combines RGB and Depth information in order to predict saliency in video scenes. Our approach extends and improves the RGB variant of an existing state-of-the-art video saliency network proposed in [39]. We include a second stream that takes as an input the produced depth maps of the corresponding RGB frames given as input to the first stream and extracts multi-scale features. The features obtained from both streams, are fused effectively together in order to obtain a final saliency map. Our problem differs significantly from salient object detection, since it is not restricted to specific salient objects, but predicts human attention in a more general aspect. These two problems not only have different objectives, but also different ground truth data and evaluation metrics.\nTo the best of our knowledge, there is no eye-tracking dataset containing depth information for video sequences except for [24], where a limited data collection using a Kinect camera took place, but it is not publicly available. Therefore, we use a robust state-of-the-art depth estimation network [23] capable of accurately predicting both indoor and outdoor scenes, in order to extract depth from the 2D RGB frames of the eye-tracking video databases. We investigate 3 different methods for depth extraction in order to assess depth contribution in the network's performance. Experimentation is carried out in 9 different databases with a large variety of video content, including sports, movies, usermade videos, documentaries, meeting scenes, etc. For comparison purposes, we employ different training setups, including our RGB-only variant, and we compare our performance with 11 different state-of-the-art models. Results indicate that depth successfully contributes to saliency modeling and is a useful modality to employ for attention modeling. ViDaS RGB-D performance accross the various databases and unseen datasets indicates that our model is capable of modeling saliency \"in-the-wild\"."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "RGB Saliency",
"publication_ref": [
"b18",
"b39",
"b32",
"b14",
"b15",
"b5",
"b25",
"b42",
"b43",
"b31",
"b41",
"b28",
"b9",
"b0",
"b19",
"b36",
"b38",
"b16",
"b10"
],
"table_ref": [],
"text": "Early CNN video saliency estimation approaches have been based on the adaptation of pretrained CNN models initially proposed for visual action recognition tasks [19,40]. Later, in [33] shallow and deep CNNs were trained for saliency prediction while in [15,16] training was performed by optimizing common saliency evaluation metrics. Long-Short Term Memory (LSTMs) and Generative Adversarial Networks (GANs) have also been developed for saliency prediction, e.g. LSTMs for spatial-only saliency in static images [6] and spatio-temporal in videos [26,43,44], as well as GANs in [32]. Multi-level saliency information from different layers through skip connections has been employed in [42]. More recently, the TASED method [29] employs a 3D fully-convolutional network with temporal aggregation, based on the assumption that the saliency map of any frame can be predicted by considering a limited number of past frames. Also, the authors of [10] essentially unify spatial and spatio-temporal saliency, i.e. image and video saliency into a joint saliency network by introducing four novel domain adaptation techniques. In order to improve saliency estimation in videos, some approaches have employed multiple modalities by combining them in multi-stream networks. For example, RGB/Optical Flow (OF) have been both employed in [1] and more recently in [20], RGB/Audio in [37,39]. Another multi-stream example is multiple subnets, such as objectness/motion [17] or saliency/gaze [11] pathways."
},
{
"figure_ref": [],
"heading": "RGB-D Saliency",
"publication_ref": [
"b46",
"b8",
"b21",
"b4",
"b33",
"b47",
"b40",
"b23",
"b23",
"b12",
"b20",
"b22",
"b24",
"b22"
],
"table_ref": [],
"text": "Depth has been employed in a variety of computer vision related problems. However, depth-aware saliency estimation in videos in the context of general attention modeling has not been explored as much as in the specific context of salient object detection (SOD). According to a recent survey [47], more than 100 models have used depth along with RGB frames (RGB-D) for SOD, starting back in 2012 [9,22] and continuing till today with deep learning models [5,34,48]. However, SOD tasks concentrate solely on finding salient objects in a video scene. On the other hand, saliency estimation for attention modeling is a different problem, because it focuses on modeling human attention in a video scene by predicting fixation points. Attention might be captured by objects, but it is not limited to well-defined structures. For example, it might be captured by more abstract visual cues, color difference or salient regions within an object. A few past works only have incorporated depth into a visual saliency model [41]. Especially concerning deep learning methods, there are even fewer [24]. In this work [24], RGB, Depth and Optical Flow are used to produce saliency maps employing generative CNNs. Here the architecture is much more naive, with the depth information integrated in the training process by simply including the depthmaps in the same stream as the RGB frames and processing them together in the various spatio-temporal scales. This method was trained and evaluated using a rather small and restricted RGBD video dataset consisting of only 54 videos. Each of these videos necessarily contains multiple levels of depth, which does not allow investigating the behavior and accuracy of the method in the wild, where depth levels in a scene could possibly be fewer. Also this method incorporates the optical flow in the training process, which not only adds computational cost, but also fails to investigate the possible benefit of using depth alone. Talking about depth, existing eye-tracking databases do not contain depth images. However, several methods exist (lightweight networks, depth cameras integrated even in mobile phones, disparity based methods in videos) that enable robust depth extraction from RGB frames at a low computational cost [13,21,23,25]. In [23], a robust depth extraction method has been developed by training and testing on different large datasets (zero-shot cross-training)."
},
{
"figure_ref": [],
"heading": "Proposed Method",
"publication_ref": [],
"table_ref": [],
"text": "The proposed method follows a 3D fully-convolutional encoder-decoder architecture. It consists of two identical spatio-temporal visual streams (encoders), that compute RGB and Depth saliency features, two fully-convolutional spatial decoder modules that perform an effective fusion of these multi-scale features and the appropriate loss function. The method is explained in detail in the following sections."
},
{
"figure_ref": [
"fig_0",
"fig_1",
"fig_1"
],
"heading": "Encoder",
"publication_ref": [
"b45",
"b3",
"b41"
],
"table_ref": [],
"text": "First, we present the backbone network of our model, displayed in Fig. 2 that is used in order to extract the multi-scale spatio-temporal features from both the RGB frames and the Depth. The architecture of the Encoder extends the 3D version of ResNet50, initially proposed for action classification. It consists of 4 3-D fully convolutional blocks that calculate spatio-temporal features in different scales of the input frames represented by X 1 , X 2 , X 3 , X 4 . Each output X m of the convolutional blocks is first refined using an attention mechanism before it continues to the next block. For that purpose we are using the Deeply Supervised Attention Module (DSAM) depicted in Fig. 3 Deep supervision has been formerly used in edge detection [46], object segmentation [4] and static saliency [42]. The role of DSAM in our model is triple: It is used for enhancing spatial feature representations, for providing the multilevel activation maps A m that will be used to calculate the loss, and finally for providing the multi-level, 64-channel saliency feature maps S m that will be used as an input to the Decoders in order to later obtain a final saliency map. Thus, DSAM parameters W m am are trained along with all the other trainable parameters of the network.\nFigure 3 displays the DSAM module architecture at level m. It includes an average pooling in the temporal dimension in order to obtain a 2D representation of the feature maps. The output of the temporal average pooling is then directed to two different paths inside the module. The one path consists of just one spatial convolutional layer that provides the 64-feature saliency maps S m . The other path consists of two convolutional layers that finally calculate a single activation map A m . A spatial softmax operation applied at the activation map A m yields the attention map M m :\nM m (x, y) = exp(A m (x, y)) x y exp(A m (x, y))(1)\nFinally the activation map A m is up-sampled to the initial dimensions of the input frames using a transposed convolutional layer. The output X m of the corresponding m-level convolutional block of the visual stream is then elementwise multiplied with the attention map M m and added to its initial value in order to enhance its most salient regions, providing the input for the next convolutional block Xm :\nXm = (1 + M m ) ⊙ X m , m = 1, ..., 4(2)\nwhere ⊙ denotes the element-wise multiplication."
},
{
"figure_ref": [
"fig_2"
],
"heading": "Decoder",
"publication_ref": [],
"table_ref": [],
"text": "For each stream, the outputs of its 4 DSAM modules are passed as an input to the Decoder in order to obtain the final saliency map. The Decoder implements a U-Net-like architecture, gradually fusing smaller scale features produced deeper in the network with features calculated at earlier layers. The architecture is illustrated in Fig. 4. The Decoder module consists of three 2D fully-convolutional blocks that are used in order to effectively fuse the multi-scale features produced by the backbone network. The first convolutional block of the Decoder takes as an input the outputs S 3 , S 4 of the two last DSAM modules of the encoder. After that, each convolutional block at level l takes as input the output S m of the corresponding DSAM module and the result of the previous block D l-1 of the Decoder. To produce the final result, the blocks first contain a 2D bilinear upsampling layer U in order to match the input's spatial dimensions. The two inputs S m , D l-1 are then concatenated and fused together using a 2D convolutional layer C l , followed by a Batch Normalization layer BN l to avoid the problem of exploding gradients. No activation function is applied. Thus, the outputs of the convolutional blocks of the Decoder can be estimated by:\nD l = BN l (C l (S 3 , U(S 4 ))), l = 1(3)\nD l = BN l (C l (S m , U(D l-1 ))), l = 2, 3(4)"
},
{
"figure_ref": [],
"heading": "RGB-D Fusion",
"publication_ref": [],
"table_ref": [],
"text": "The overall proposed architecture consists of both RGB and Depth streams. The two streams do not interfere with each other until the output D rgb , D d of each Decoder is calculated. That way, each stream can be trained end-to-end concentrated on its separate task which is to learn the appropriate representations from the provided input. By applying a last-layer fusion of RGB and Depth features, we observed that the network could learn when Depth features are beneficial per input case, compared to other approaches we followed. That being said, the fusion between RGB and depth features is done after the decoding is finished.\nThe two outputs D rgb , D b of the two Decoders are up-sampled using bilinear interpolation layer U to match the input's dimensions. Finally the up-sampled feature maps are fused using 2D convolutional layers to produce the final saliency map S rgb-d of the network:\nS rgb-d = F(U(D rgb ), U(D d ))(5)\nwhere F(•) denotes the concatenation and convolutional fusion of the Decoder outputs."
},
{
"figure_ref": [],
"heading": "Saliency Loss",
"publication_ref": [],
"table_ref": [],
"text": "For training our model, we implemented a custom loss function L that is calculated by combining three different losses. To compute these three losses and eventually the final loss, we use the ground truth saliency map Y , compared not only with the output saliency map S rgb-d of the network, but also with the 4 multi-scale activation maps A m of each of the RGB and the Depth stream, denoted as (A m rgb ) and (A m d ) respectively:\nL = L sal (S rgb-d , Y ) + (1 -ϵ) 4 m=1 L rgb (A m rgb , Y )+ (1 -ϵ) 4 m=1 L d (A m d , Y )(6)\nwhere ϵ is a decaying parameter equal to currentEpoch #totalEpochs . For the losses L sal , L rgb and L d , we calculate three different metrics. We first calculate the cross entropy loss between the generated maps M , where\nM = S rgb-d , A m rgb , A m d , m = 1, ...,4\nand the continuous ground truth saliency map Y c that is obtained by a convolving the binary fixation map Y b of the eye-tracking data with a gaussian kernel:\nL CE (M, Y c ) = - x,y Y c (x, y) ⊙ log M (x, y) +(1 -Y c (x, y)) ⊙ (1 -log M (x, y))(7)\nThe second metric calculated is the linear Correlation Coefficient (CC) between the map M and the continuous ground truth saliency map Y c . The CC metric treats the predicted and the ground truth maps as random variables and uses their covariance cov and standard deviation ρ to calculate their correlation:\nL CC (M, Y c ) = - cov(M (x, y), Y c (x, y)) ρ(M (x, y)) • ρ(Y c (x, y))(8)\nThe last metric calculated for the saliency loss, is the Normalized Scanpath Saliency (NSS) metric between the map M and the binary fixation map Y b :\nL N SS (M, Y b ) = - 1 N b x,y M (x, y) ⊙ Y b (x, y)(9)\nwhere M (x, y) = M (x,y)-M (x,y) ρ(M (x,y))\n, the normalized map M to zero-mean and unit standard deviation and N b = x,y Y b (x, y), the total number of discrete fixation points in the binary ground truth saliency map. The three initial losses can be written as:\nL sal (S rgb-d , Y ) = w 1 L CE (S rgb-d , Y c )+ w 2 L CC (S rgb-d , Y c ) + w 3 L N SS (S rgb-d , Y b ) (10) L rgb (A m rgb , Y ) = w 1 L CE (A m rgb , Y c )+ w 2 L CC (A m rgb , Y c ) + w 3 L N SS (A m rgb , Y b )(11)\nL d (A m d , Y ) = w 1 L CE (A m d , Y c )+ w 2 L CC (A m d , Y c ) + w 3 L N SS (A m d , Y b )(12)\nwhere w 1 , w 2 , w 3 are the weights used to get the weighted sum of the three losses."
},
{
"figure_ref": [],
"heading": "Implementation",
"publication_ref": [
"b13"
],
"table_ref": [],
"text": "Our implementation and experimentation with the visual network uses as backbone the 3D ResNet-50 architecture [14] that has showed competitive performance against other deeper architectures for action recognition tasks, in terms of performance and computational budget. As starting point for the trainable parameters W rgb , W d of the RGB and the Depth stream respectively, we used the weights from the pretrained model in the Kinetics 400 database. Training:\nFor training we employ stochastic gradient descent with momentum 0.9, while we assign a weight decay of 1e-5 for regularization. We have also employed effective batch sizes of 128 samples, and multi-step learning rate. The layers of the DSAM and the Decoder modules are trained using an initial learning rate of 0.0001 while the backbone streams are trained using an initial learning rate of 0.001. The network is trained for 60 epochs. The input data is spatially resized to 112x112 and a sliding window of 16 frames is applied, with the final prediction of the network corresponding to the medium frame. For data augmentation we use random horizontal flipping to the input frames, depthmaps and corresponding ground truth saliency maps with a probability P = 0.5. No other transformations are performed. The weights w 1 , w 2 , w 3 for the saliency loss are selected as 0.1, 2, 1 respectively, after experimentation."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Datasets",
"publication_ref": [
"b27",
"b34",
"b35",
"b27"
],
"table_ref": [],
"text": "For training and evaluation of the proposed saliency network, 9 different datasets are employed: DHF1K, Hollywood2 action recognition tasks, but was later adopted for video saliency tasks, after eyetracking data were collected [28]. Train and test sets consist of 3100 and 3559 different, non overlapping clips respectively, each one viewed by 12 persons. UCF-Sports: UCF-Sports [35,36] similarly to Hollywood2, also contains short clips initially collected for action recognition. Later, eyetracking data from 19 viewers have been recorded [28]. "
},
{
"figure_ref": [
"fig_3"
],
"heading": "Experimental Results",
"publication_ref": [
"b1",
"b27",
"b42",
"b38",
"b2",
"b22",
"b38",
"b22",
"b24",
"b12"
],
"table_ref": [
"tab_0",
"tab_0"
],
"text": "Training has been performed in several different setups by combining data from one or more datasets: For DHF1K, Hollywood2, UCF-Sports and DIEM, the standard splits from literature have been employed [2,28,43]. For the other 5 databases, where there is no particular split, the approach adopted in [39] has been followed (3 non overlapping splits per database and average among splits).\nFor the evaluation of ViDaS network, we perform an ablation study in order to assess the importance of several parameters and fusion modules, and the contribution of depth in comparison to RGB. Additionally, we compare 3 different depth extraction methods. We then pick our best RGB-D method and compare it to 11 state-of-the-art visual saliency methods (using their publicly available codes and models, re-evaluated), in all 9 databases on the same test data. We also test our RGB-only variant. For all state-of-the-art models, we have employed their best model and re-evaluated the model on the test splits of all databases. During evaluation, we assess ViDaS performance on the various datasets for several training setups. Five widely-used saliency evaluation metrics are employed [3]: CC, NSS, AUC-Judd (AUC-J), shuffled AUC (sAUC) and SIM (similarity). For sAUC we select the negative samples from the union of all viewers' fixations across all other frames in the current video except for the currently processed frame. Ablation study: Regarding Table 1, ablation study employs 6 datasets: DIEM, AVAD, Coutrot1 and 2, SumMe and ETMD, which are relatively small but with diverse content. Depth extraction is part of the ablation study, and except for the last two rows of Table 1, in all other variants depth extraction has been conducted using MIDAS [23]. Methods: Number 16 or 64 refers to the number of feature maps S m coming from the DSAM modules. \"SC\" refers to Simple Concatenation of the DSAM output maps S m without a decoder, following the approach of [39]. \"MF\" refers to multiscale fusion, i.e. the integration of the Decoder. RGB-Depth Fusion:The \"ADD\", \"CON\", \"CLL\" refer to the different fusion schemes between RGB and Depth. For \"ADD\" and \"CON\" a simple addition and concatenation respectively of RGB and Depth maps is performed at every scale, and a single common Decoder module was employed for both streams. For \"CLL\", each stream is processed individually using its own Decoder module and without any interaction in the different spatio-temporal scales, except for the last layer, where the outputs of each Decoder are concatenated and fused through two convolutional layers. Depth extraction: Depth is extracted using 3 different depth extraction methods for comparisons and for assessing if a detailed or coarse depth estimation is closer to the concept of saliency. \"MID\", \"MEG\" and \"DIL\" refer to the 3 different depth extraction methods that were investigated, MIDAS [23], Megadepth [25], and DILATED [13].\nOur ablation results indicate that multiscale fusion increased the performance for the RGB-only model. The choice of 64 feature maps along with multiscale fusion further boosted the performance of the RGB model, as well as Depth only model, which in all cases performs worse than the RGB-only model, indicating that Depth can be employed as an additional modality for saliency estimation, but it could not perform equally well on its own. The experimentation with the several fusion schemes indicates that concatenation on the last layer (i.e. late fusion) is more effective and leads to the best performance of our model, probably because each stream learns independently the most it can learn, and the two are combined in the end to produce a single attention map. Lastly, among the three different depth extraction methods, MIDAS leads to our best results. MIDAS produced the most detailed, fine depth estimations among the three, indicating that saliency might be sensitive to depth information. For the rest of the paper, by ViDaS [STD] we refer to RGBD64MF CLL MID, and by ViDaS [ST] to our RGB-only variant. An example comparison of the two versions is depicted in Fig. 5 along with the original frames and the ground truth saliency maps. Also the NSS curve is depicted over time. The particular frame has many levels of depth and indeed the RGB-D version captures saliency better than the RGB-only version.\nComparison to state-of-the-art: Extensive comparisons with 11 different state-of-the-art saliency methods on 9 different databases are depicted for the five metrics per database, in Tables 2, 3 and 4. The models were not retrained, since in some cases the code is not available, but the published pretrained models were RGB-D ViDaS version performs consistently better than the RGB-only version, endowing the model with robustness and smoother estimations across time. An interesting finding is that ViDaS model performance is not degraded even on unseen datasets, and in such cases the existence of depth makes a even bigger difference. For example, in AVAD, SumMe, Coutrot2 datasets, the best performance is achieved by training on these datasets, but the model trained on DHF1K or UHD still achieves a competitive performance, whereas other methods like UNISAL, SALEMA and TASED that exhibited good performance in seen datasets (e.g. DHF1K), when tested in unseen data (e.g. Coutrot2), their performance is not consistently good. To sum up, that ViDaS RGB-D network can well generalize into unseen datasets, without a large compromise in performance, confirming its potential for modeling saliency \"in-the-wild\"."
},
{
"figure_ref": [],
"heading": "Conclusions",
"publication_ref": [],
"table_ref": [],
"text": "We presented ViDaS, a novel video depth-aware saliency network that efficiently predicts fixations in videos, by combining an RGB and a Depth stream in order to produce a single saliency map. Network performance has been extensively evaluated in various datasets with highly diverse content. Results for 5 different metrics in 9 different databases and comparison with 11 state-of-the-art methods indicate that depth can endow an RGB network with robustness and performance boost. Our RGB-D method achieves the best or competitive performance in all cases. Also, its better performance in unseen datasets indicate its appropriateness for estimating saliency \"in-the-wild\"."
},
{
"figure_ref": [],
"heading": "Method",
"publication_ref": [],
"table_ref": [],
"text": "Method used to consistently evaluate the results in the same way. Since all of the other methods have been trained on combinations of SALICON, DHF1K, Hollywood-2 and UCF-Sports datasets, we included 5 different training set-ups (on DHF1K only, on Hollywood2, on UCF, on a combination of these three denoted by UHD and on the rest 6 databases, denoted by STAViS as depicted in Tables 2,3, and 4) to enable fair comparisons, as well as the performance assessment of the model in unseen or seen datasets. For each training setup, we have also trained our RGB-only variant. Overall, ViDaS RGB-D network achieves the best or competitive performance on the various datasets. In Table 2 UNISAL performs better than ViDaS in some metrics, perhaps because it has been pretrained on saliency datasets (SALICON, DHF1K), that seems to boost performance, whereas ViDaS uses the pretrained weights of Kinetics400 (which is an action recognition dataset) as a starting point. Also some methods have tuned their parameters to these most widely used datasets, whereas our method was trained in a more robust way. TASED method that performs better in DHF1K, employs a 32-frame temporal length, compared to ours which is 16. For visualization purposes, in Fig. 6 sample frames are presented with their corresponding eyetracking data, ground truth saliency maps, and the corresponding saliency maps from our proposed ViDaS RGB-D network and other state-of-the-art methods: ACLNet, TASED, Unisal, SALEMA and STAViS. It can easily be observed that our results are closer to the ground truth, especially when frames have several levels of depth."
},
{
"figure_ref": [],
"heading": "Discussion",
"publication_ref": [],
"table_ref": [],
"text": "Results in the various datasets indicate that depth as a complimentary modality for saliency estimation indeed boosts performance, especially when frames contain many levels of depth, e.g. Coutrot1 and UCF sports. In almost all cases, the"
}
] | We introduce ViDaS, a two-stream, fully convolutional Video, Depth-Aware Saliency network to address the problem of attention modeling "in-the-wild", via saliency prediction in videos. Contrary to existing visual saliency approaches using only RGB frames as input, our network employs also depth as an additional modality. The network consists of two visual streams, one for the RGB frames, and one for the depth frames. Both streams follow an encoder-decoder approach and are fused to obtain a final saliency map. The network is trained end-to-end and is evaluated in a variety of different databases with eye-tracking data, containing a wide range of video content. Although the publicly available datasets do not contain depth, we estimate it using three different state-of-the-art methods, to enable comparisons and a deeper insight. Our method outperforms in most cases state-of-the-art models and our RGB-only variant, which indicates that depth can be beneficial to accurately estimating saliency in videos displayed on a 2D screen. Depth has been widely used to assist salient object detection problems, where it has been proven to be very beneficial. Our problem though differs significantly from salient object detection, since it is not restricted to specific salient objects, but predicts human attention in a more general aspect. These two problems not only have different objectives, but also different ground truth data and evaluation metrics. To our best knowledge, this is the first competitive deep learning video saliency estimation approach that combines both RGB and Depth features to address the general problem of saliency estimation "in-the-wild". The code will be publicly released. | ViDaS: Video Depth-aware Saliency Network | [
{
"figure_caption": "Fig. 2 .2Fig. 2. ViDaS architecture. The network consists of two identical streams, computing RGB and Depth saliency features respectively. The output saliency feature maps from the different scales for RGB and Depth each pass through a Decoder and are fused in the last network layer in order to produce a single saliency map.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 3 .3Fig. 3. Deeply Supervised Attention Module (DSAM) enhances the global network's representations and provides the multi-level saliency maps for spatio-temporal saliency.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 4 .4Fig. 4. Architecture of the Decoder module. The Decoder is used as the prediction part of the network, which fuses the multi-scale spatio-temporal features to obtain the final saliency map.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 5 .5Fig. 5. Sample frames from Coutrot1 database with their eye-tracking data, the corresponding ground truth, RGB-only, and RGB-D saliency maps as produced by ViDaS (RGB and RGB-D). Also NSS curve over time for the two approaches.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 6 .6Fig. 6. Sample frames from several databases with their eye-tracking data, the corresponding ground truth, ViDaS RGB-D network and several other state-of-the-art methods.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "Ablation study: Different fusion schemes, feature size and depth extraction methods are investigated.",
"figure_data": ", UCF-Sports, DIEM, AVAD, Coutrot1, Coutrot2,SumMe, and ETMD. These databases consist of various types of videos, rang-ing from very structured small videos to completely unstructured, user-madeYoutube videos. A short description for each database follows.DHF1K: DHF1K [43] is a large dataset with high content diversity and variablelength (from 400 frames to 1200 frames). It includes 1000 videos, out of which700 are publicly annotated, and 300 are withheld for testing purposes.Hollywood-2: Hollywood2 [27] contains a collection of short clips with actionsperformed in Hollywood movies. The dataset was initially developed for human",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "UCF-sports has been split in train and test set of 104 and 48 non overlapping clips respectively. Both Hollywood2 and UCF-Sports contain shots much shorter and smaller than a DHF1K video sample, ranging from 40 frames to just a single frame per shot. AVAD: AVAD database[30] contains 45 short clips of 5-10 sec duration with several audiovisual scenes, e.g. dancing, guitar playing, bird signing, etc. Eyetracking data from 16 participants have been recorded.",
"figure_data": "Coutrot databases: Coutrot databases [7,8] are split in Coutrot1 and Coutrot2:Coutrot1 contains 60 clips with dynamic natural scenes split in 4 visual cate-gories: one/several moving objects, landscapes, and faces. Eye-tracking data from72 participants have been recorded. Coutrot2 contains 15 clips of 4 persons in ameeting and the corresponding eye-tracking data from 40 persons.DIEM: DIEM database [31] consists of 84 movies of all sorts, sourced frompublicly accessible repositories, including advertisements, documentaries, gametrailers, movie trailers, music videos, news clips, and time-lapse footage. Eyemovement data from 42 participants were recorded.SumMe: SumMe database [12, 38] contains 25 unstructured videos, i.e. mostlyuser-made videos, from public sources. Audiovisual eye-tracking data have beencollected [38] from 10 viewers.ETMD: ETMD database [18,38] contains 12 videos from six different hollywoodmovies. Audiovisual eye-tracking data have been collected [38] from 10 viewers,recorded via an Eyelink eye-tracker.",
"figure_id": "tab_1",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Evaluation results for saliency in DIEM, Coutrot1 and Coutrot2 databases. The proposed method's (ViDaS [STD] and the RGB-only variant [ST]) results are depicted for different training setups. [STD] stands for spatio-temporal plus depth, [STA] for spatio-temporal plus audio, [ST] for spatio-temporal visual models while [S] denotes spatial only models.",
"figure_data": "",
"figure_id": "tab_2",
"figure_label": "3",
"figure_type": "table"
}
] | Ioanna Diamanti; Antigoni Tsiami; Petros Koutras; Petros Maragos | [
{
"authors": "C Bak; A Kocak; E Erdem; A Erdem",
"journal": "IEEE Trans. Multimedia",
"ref_id": "b0",
"title": "Spatio-temporal saliency networks for dynamic saliency prediction",
"year": "2018"
},
{
"authors": "A Borji; D N Sihite; L Itti",
"journal": "IEEE Trans. Image Process",
"ref_id": "b1",
"title": "Quantitative analysis of human-model agreement in visual saliency modeling: A comparative study",
"year": "2013"
},
{
"authors": "Z Bylinskii; T Judd; A Oliva; A Torralba; F Durand",
"journal": "",
"ref_id": "b2",
"title": "What do different evaluation metrics tell us about saliency models?",
"year": "2016"
},
{
"authors": "S Caelles; K K Maninis; J Pont-Tuset; L Leal-Taixé; D Cremers; L Van Gool",
"journal": "",
"ref_id": "b3",
"title": "One-shot video object segmentation",
"year": "2017"
},
{
"authors": "Q Chen; Z Liu; Y Zhang; K Fu; Q Zhao; H Du",
"journal": "",
"ref_id": "b4",
"title": "Rgb-d salient object detection via 3d convolutional neural networks",
"year": "2021"
},
{
"authors": "M Cornia; L Baraldi; G Serra; R Cucchiara",
"journal": "IEEE Trans. Image Process",
"ref_id": "b5",
"title": "Predicting Human Eye Fixations via an LSTM-based Saliency Attentive Model",
"year": "2018"
},
{
"authors": "A Coutrot; N Guyader",
"journal": "Journal of Vision",
"ref_id": "b6",
"title": "How saliency, faces, and sound influence gaze in dynamic social scenes",
"year": "2014"
},
{
"authors": "A Coutrot; N Guyader",
"journal": "Springer",
"ref_id": "b7",
"title": "Multimodal saliency models for videos",
"year": "2016"
},
{
"authors": "K Desingh; K M Krishna; D Rajan; C Jawahar",
"journal": "",
"ref_id": "b8",
"title": "Depth really matters: Improving visual salient region detection with depth",
"year": "2013"
},
{
"authors": "R Droste; J Jiao; J A Noble",
"journal": "",
"ref_id": "b9",
"title": "Unified Image and Video Saliency Modeling",
"year": "2020"
},
{
"authors": "S Gorji; J J Clark",
"journal": "",
"ref_id": "b10",
"title": "Going from image to video saliency: Augmenting image salience with dynamic attentional push",
"year": "2018"
},
{
"authors": "M Gygli; H Grabner; H Riemenschneider; L Van Gool",
"journal": "",
"ref_id": "b11",
"title": "Creating summaries from user videos",
"year": "2014"
},
{
"authors": "Z Hao; Y Li; S You; F Lu",
"journal": "IEEE",
"ref_id": "b12",
"title": "Detail preserving depth estimation from a single image using attention guided networks",
"year": "2018"
},
{
"authors": "K Hara; H Kataoka; Y Satoh",
"journal": "",
"ref_id": "b13",
"title": "Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet",
"year": "2018"
},
{
"authors": "X Huang; C Shen; X Boix; Q Zhao",
"journal": "",
"ref_id": "b14",
"title": "Salicon: Reducing the semantic gap in saliency prediction by adapting deep neural networks",
"year": "2015"
},
{
"authors": "S Jetley; N Murray; E Vig",
"journal": "",
"ref_id": "b15",
"title": "End-to-end saliency mapping via probability distribution prediction",
"year": "2016"
},
{
"authors": "L Jiang; M Xu; T Liu; M Qiao; Z Wang",
"journal": "",
"ref_id": "b16",
"title": "Deepvs: A deep learning based video saliency prediction approach",
"year": "2018"
},
{
"authors": "P Koutras; P Maragos",
"journal": "Signal Process.: Image Communication",
"ref_id": "b17",
"title": "A perceptually based spatio-temporal computational framework for visual saliency estimation",
"year": "2015"
},
{
"authors": "M Kümmerer; L Theis; M Bethge",
"journal": "",
"ref_id": "b18",
"title": "Deep gaze I: Boosting saliency prediction with feature maps trained on imagenet",
"year": "2015"
},
{
"authors": "Q Lai; W Wang; H Sun; J Shen",
"journal": "IEEE Trans. Image Process",
"ref_id": "b19",
"title": "Video saliency prediction using spatiotemporal residual attentive networks",
"year": "2019"
},
{
"authors": "I Laina; C Rupprecht; V Belagiannis; F Tombari; N Navab",
"journal": "IEEE",
"ref_id": "b20",
"title": "Deeper depth prediction with fully convolutional residual networks",
"year": "2016"
},
{
"authors": "C Lang; T V Nguyen; H Katti; K Yadati; M Kankanhalli; S Yan",
"journal": "Springer",
"ref_id": "b21",
"title": "Depth matters: Influence of depth cues on visual saliency",
"year": "2012"
},
{
"authors": "K Lasinger; R Ranftl; K Schindler; V Koltun",
"journal": "IEEE Trans. Pattern Anal. Mach. Intell",
"ref_id": "b22",
"title": "Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer",
"year": "2020"
},
{
"authors": "G Leifman; D Rudoy; T Swedish; E Bayro-Corrochano; R Raskar",
"journal": "",
"ref_id": "b23",
"title": "Learning gaze transitions from depth to improve video saliency estimation",
"year": "2017"
},
{
"authors": "Z Li; N Snavely",
"journal": "",
"ref_id": "b24",
"title": "Megadepth: Learning single-view depth prediction from internet photos",
"year": "2018"
},
{
"authors": "P Linardos; E Mohedano; J J Nieto; K Mcguinness; X Giro-I Nieto; N E O'connor",
"journal": "",
"ref_id": "b25",
"title": "Simple vs complex temporal recurrences for video saliency prediction",
"year": "2019"
},
{
"authors": "M Marszalek; I Laptev; C Schmid",
"journal": "IEEE",
"ref_id": "b26",
"title": "Actions in context",
"year": "2009"
},
{
"authors": "S Mathe; C Sminchisescu",
"journal": "IEEE Trans. Pattern Anal. Mach. Intell. (PAMI)",
"ref_id": "b27",
"title": "Actions in the eye: Dynamic gaze datasets and learnt saliency models for visual recognition",
"year": "2014"
},
{
"authors": "K Min; J J Corso",
"journal": "",
"ref_id": "b28",
"title": "Tased-net: Temporally-aggregating spatial encoder-decoder network for video saliency detection",
"year": "2019"
},
{
"authors": "X Min; G Zhai; K Gu; X Yang",
"journal": "ACM Transactions on Multimedia Computing, Communications, and Applications",
"ref_id": "b29",
"title": "Fixation prediction through multimodal analysis",
"year": "2017"
},
{
"authors": "P K Mital; T J Smith; R Hill; J M Henderson",
"journal": "Cognitive Computation",
"ref_id": "b30",
"title": "Clustering of gaze during dynamic scene viewing is predicted by motion",
"year": "2011"
},
{
"authors": "J Pan; C Canton; K Mcguinness; N E O'connor; J Torres; E Sayrol; X Giro-I Nieto",
"journal": "",
"ref_id": "b31",
"title": "Salgan: Visual saliency prediction with generative adversarial networks",
"year": "2017"
},
{
"authors": "J Pan; E Sayrol; X Giro-I Nieto; K Mcguinness; N E O'connor",
"journal": "",
"ref_id": "b32",
"title": "Shallow and deep convolutional networks for saliency prediction",
"year": "2016"
},
{
"authors": "Y Piao; W Ji; J Li; M Zhang; H Lu",
"journal": "",
"ref_id": "b33",
"title": "Depth-induced multi-scale recurrent attention network for saliency detection",
"year": "2019-10"
},
{
"authors": "M D Rodriguez; J Ahmed; M Shah",
"journal": "IEEE",
"ref_id": "b34",
"title": "Action mach a spatio-temporal maximum average correlation height filter for action recognition",
"year": "2008"
},
{
"authors": "K Soomro; A R Zamir",
"journal": "Springer",
"ref_id": "b35",
"title": "Action recognition in realistic sports videos",
"year": "2014"
},
{
"authors": "H R Tavakoli; A Borji; E Rahtu; J Kannala",
"journal": "",
"ref_id": "b36",
"title": "Dave: A deep audio-visual embedding for dynamic saliency prediction",
"year": "2019"
},
{
"authors": "A Tsiami; P Koutras; A Katsamanis; A Vatakis; P Maragos",
"journal": "Signal Processing: Image Communication",
"ref_id": "b37",
"title": "A behaviorally inspired fusion approach for computational audiovisual saliency modeling",
"year": "2019"
},
{
"authors": "A Tsiami; P Koutras; P Maragos",
"journal": "",
"ref_id": "b38",
"title": "Stavis: Spatio-temporal audiovisual saliency network",
"year": "2020-06"
},
{
"authors": "E Vig; M Dorr; D Cox",
"journal": "",
"ref_id": "b39",
"title": "Large-scale optimization of hierarchical features for saliency prediction in natural images",
"year": "2014"
},
{
"authors": "J Wang; M P Da Silva; P Le Callet; V Ricordel",
"journal": "IEEE Trans. Image Process",
"ref_id": "b40",
"title": "Computational model of stereoscopic 3d visual saliency",
"year": "2013"
},
{
"authors": "W Wang; J Shen",
"journal": "IEEE Trans. Image Process",
"ref_id": "b41",
"title": "Deep visual attention prediction",
"year": "2018"
},
{
"authors": "W Wang; J Shen; F Guo; M M Cheng; A Borji",
"journal": "",
"ref_id": "b42",
"title": "Revisiting video saliency: A large-scale benchmark and a new model",
"year": "2018"
},
{
"authors": "W Wang; J Shen; J Xie; M M Cheng; H Ling; A Borji",
"journal": "IEEE Trans. Pattern Anal. Mach. Intell. (PAMI)",
"ref_id": "b43",
"title": "Revisiting video saliency prediction in the deep learning era",
"year": "2019"
},
{
"authors": "X Wu; Z Wu; J Zhang; L Ju; S Wang",
"journal": "",
"ref_id": "b44",
"title": "Salsac: a video saliency prediction model with shuffled attentions and correlation-based convlstm",
"year": "2020"
},
{
"authors": "S Xie; Z Tu",
"journal": "",
"ref_id": "b45",
"title": "Holistically-nested edge detection",
"year": "2015"
},
{
"authors": "T Zhou; D P Fan; M M Cheng; J Shen; L Shao",
"journal": "Computational Visual Media",
"ref_id": "b46",
"title": "Rgb-d salient object detection: A survey",
"year": "2021"
},
{
"authors": "C Zhu; X Cai; K Huang; T H Li; G Li",
"journal": "IEEE",
"ref_id": "b47",
"title": "Pdnet: Prior-model guided depthenhanced network for salient object detection",
"year": "2019"
}
] | [
{
"formula_coordinates": [
5,
231.49,
605.73,
249.11,
26.29
],
"formula_id": "formula_0",
"formula_text": "M m (x, y) = exp(A m (x, y)) x y exp(A m (x, y))(1)"
},
{
"formula_coordinates": [
6,
232.8,
309.82,
247.79,
11.26
],
"formula_id": "formula_1",
"formula_text": "Xm = (1 + M m ) ⊙ X m , m = 1, ..., 4(2)"
},
{
"formula_coordinates": [
6,
248.45,
575.83,
232.15,
11.72
],
"formula_id": "formula_2",
"formula_text": "D l = BN l (C l (S 3 , U(S 4 ))), l = 1(3)"
},
{
"formula_coordinates": [
6,
225.31,
590.78,
255.28,
11.72
],
"formula_id": "formula_3",
"formula_text": "D l = BN l (C l (S m , U(D l-1 ))), l = 2, 3(4)"
},
{
"formula_coordinates": [
7,
245.61,
402.87,
234.98,
9.65
],
"formula_id": "formula_4",
"formula_text": "S rgb-d = F(U(D rgb ), U(D d ))(5)"
},
{
"formula_coordinates": [
7,
200.25,
555.08,
280.34,
64.76
],
"formula_id": "formula_5",
"formula_text": "L = L sal (S rgb-d , Y ) + (1 -ϵ) 4 m=1 L rgb (A m rgb , Y )+ (1 -ϵ) 4 m=1 L d (A m d , Y )(6)"
},
{
"formula_coordinates": [
7,
328.79,
654.54,
151.8,
12.55
],
"formula_id": "formula_6",
"formula_text": "M = S rgb-d , A m rgb , A m d , m = 1, ...,4"
},
{
"formula_coordinates": [
8,
215.53,
153.5,
265.06,
35.82
],
"formula_id": "formula_7",
"formula_text": "L CE (M, Y c ) = - x,y Y c (x, y) ⊙ log M (x, y) +(1 -Y c (x, y)) ⊙ (1 -log M (x, y))(7)"
},
{
"formula_coordinates": [
8,
220.06,
250.62,
260.53,
23.22
],
"formula_id": "formula_8",
"formula_text": "L CC (M, Y c ) = - cov(M (x, y), Y c (x, y)) ρ(M (x, y)) • ρ(Y c (x, y))(8)"
},
{
"formula_coordinates": [
8,
213.07,
309.1,
267.52,
26.35
],
"formula_id": "formula_9",
"formula_text": "L N SS (M, Y b ) = - 1 N b x,y M (x, y) ⊙ Y b (x, y)(9)"
},
{
"formula_coordinates": [
8,
217.09,
402.16,
263.5,
75.48
],
"formula_id": "formula_10",
"formula_text": "L sal (S rgb-d , Y ) = w 1 L CE (S rgb-d , Y c )+ w 2 L CC (S rgb-d , Y c ) + w 3 L N SS (S rgb-d , Y b ) (10) L rgb (A m rgb , Y ) = w 1 L CE (A m rgb , Y c )+ w 2 L CC (A m rgb , Y c ) + w 3 L N SS (A m rgb , Y b )(11)"
},
{
"formula_coordinates": [
8,
230.58,
498.2,
250.01,
27.64
],
"formula_id": "formula_11",
"formula_text": "L d (A m d , Y ) = w 1 L CE (A m d , Y c )+ w 2 L CC (A m d , Y c ) + w 3 L N SS (A m d , Y b )(12)"
}
] |
|
2023-05-19 | [
{
"figure_ref": [
"fig_0",
"fig_0",
"fig_1"
],
"heading": "Introduction",
"publication_ref": [
"b36",
"b9",
"b30",
"b34",
"b39",
"b24",
"b2",
"b2",
"b35",
"b3"
],
"table_ref": [],
"text": "Deep neural networks (DNNs) have been widely utilized in a variety of visual recognition problems [6,7,21,28] by virtue of the large-scale, high-quality, and annotated datasets. DNNs usually require the training dataset to be artificially balanced and have sufficient samples of each class. Unfortunately, from a practical perspective, object frequency usually follows a power law and typically ex-* Yiu-ming Cheung is the Corresponding Author. hibits a long-tailed distribution. Naive learning on such data is prone to undesirable bias towards the head classes which occupy the majority of the training samples [37]. Since tail classes have few training samples that cannot cover the real distribution in embedding space, their spatial span is severely compressed by head classes. In addition, a vast number of head class samples generate overwhelming discouraging gradients for tail classes. Thus, the learning of a classifier is biased towards the head classes. As a result, directly training on long-tailed data brings two key problems: 1) the distorted embedding space, and 2) the biased classifier.\nIn the literature, most of the recently proposed approaches focus on addressing the second problem only, i.e., the biased classifier. For example, Menon et al. [17] and Hong et al. [8] applied post-adjust strategy to the trained model to calibrate the class boundary. Nevertheless, the distorted embedding cannot be adjusted with the post-hoc calibration, which is not conducive to further improving the model performance. Most recently, the two-stage decoupling methods [2, 10,31,35,40] have been proposed to obtain good embeddings in the first stage and then re-balance the classifier in the second stage. These methods obtain the representation by cross-entropy (CE) loss, which, however, leads to a severely uneven distributed embedding space. We implement a toy experiment to illustrate the distortion of the embedding space as shown in Fig. 1, where t-SNE [25] is utilized to visualize the features of a long-tailed subset from CIFAR-10 dataset. We can observe that the tail class occupies a much small spatial span than the head class. This is because the tail class with fewer samples cannot cover the ground truth distribution. Moreover, Fig. 1 also shows that there are obscure regions (i.e., the grey area) between different classes. Softmax saturation [3] is one of the factors of these obscure regions because it leads to insufficient training. These obscure regions have a severe effect on the tail classes but little on the head classes. Since tail class samples clustered around the class boundary aggravate their spatial squeezing, while the head class samples with enough variety can already cover the true distribution.\nSoftmax saturation refers to the inopportune early gradients vanishing produced by the softmax [3,36], which weakens the validity of training samples and impedes model training. However, from another perspective, the seemingly harmful softmax saturation has the ability to balance the valid samples of different classes and thus help calibrate the distortion of embedding space. Specifically, we disturb the logit of different classes with different amplitudes. We name the disturbed logit as Gaussian clouded logit (GCL) and the amplitude of the disturbance as cloud size, because we set the disturbance to a Gaussian distribution. The tail classes have few training samples and thus the training samples of them should be more valid. We therefore disturb the logit of tail classes with large relative cloud sizes to reduce the softmax saturation. In this way, tail class samples can provide more gradients without overfitting and thus indirectly affect their embedding space. In addition, a large cloud size of the tail class logit corresponds to the large cloud size on feature in the direction of the class anchor. Therefore, tail classes can have large margins towards the class boundary, so as to alleviate the severe uneven distribution between the head and tail classes. Conversely, the head classes are set to small cloud sizes, so that they can be automatically filtered out during training. Eventually, as shown in Fig. 2, the tail class samples can be pushed more away from the class boundary so as the distortion of the embedding space can be calibrated.\nTo address the biased classifier, we re-balance the training data with a class-wise sampling strategy. As training with GCL makes the validity of different classes vary, the so-called \"effectiveness\" [4] of them are different. Existing class-wise balanced sampling strategies will lead to exces- Extensive experiments on multiple commonly used longtailed recognition benchmark datasets demonstrate that the proposed GCL surpasses the recently proposed counterparts. In summary, the key contributions of our work are three-fold:\n• We propose the GCL adjustment loss function, which utilizes softmax saturation to balance the sample validity of different classes. An evenly distributed embedding can be obtained with the proposed GCL.\n• We propose a simple but effective class-based effective number (CBEN) sampling strategy for re-balancing the classifier to avoid repeat training of tail classes. This sampling strategy can further boost the performance of GCL.\n• Extensive experiments on popular long-tailed datasets demonstrate that the proposed method outperforms the state-of-the-art counterparts."
},
{
"figure_ref": [],
"heading": "Related Works",
"publication_ref": [],
"table_ref": [],
"text": "Long-tailed classification is one of the long-standing research problems in machine learning. Several kinds of approaches have been proposed to address it. This section briefly introduces the most related three regimes, namely loss modification, logit adjustment, and decoupling representation."
},
{
"figure_ref": [],
"heading": "Loss Modification",
"publication_ref": [
"b14",
"b14",
"b3",
"b8",
"b10",
"b12"
],
"table_ref": [],
"text": "Modifying the loss function through re-weighting is the most natural method. Sample-wise re-weighting meth-ods [15,20] attempt to make the model pay more attention to the difficult samples by introducing fine-gained coefficients in the loss for imbalanced learning. For example, focal loss [15] introduces a tunable focusing parameter, which is negatively correlated with the predicted probability of the target class. This focusing parameter helps the model training focus on hard samples and prevents the numerous easy negatives from overwhelming. Class-wise reweighting methods [4,9,11,23] assign the standard CE loss with category-specific parameters that are inversely proportional to the class frequencies. For example, Tan et al. [23] proposed equalization loss, which utilizes a weight term to randomly ignore the discouraging gradients of head class samples. These methods can alleviate the data imbalance to a certain extent. However, the classification difficulty of a sample is not directly related to its corresponding class size. Further, another side effect of assigning higher weights to difficult samples/tail classes is overly focusing on harmful samples (e.g., noisy data or mislabeled data) [13]."
},
{
"figure_ref": [],
"heading": "Logit Adjustment",
"publication_ref": [
"b34",
"b0"
],
"table_ref": [],
"text": "Logit adjustment assigns relatively large margins for tail classes. Most recently, Menon et al. [17] have proposed a logit adjustment (LA) method which is consistent with minimizing the balanced error. The logit shifting in LA of different classes is based on label frequencies of training data. Differently, LADE [8] calibrates the logit to the test set using the label distribution of test data, so that the test set can also be imbalanced. Tang et al.\n[24] adopted causal intervention to remove the \"bad\" SGD momentum and keep the \"good\" one to avoid the harmful causal effect for tail prediction. DisAlign [35] adjusts the logit by calibrating the model prediction to a reference distribution of classes that favors the balanced prediction. These methods well adjust the model logits through post-hoc shifting but without considering calibrating the embedding space. Another type of approach [1,2] addresses long-tailed data by leaving large relative margins for tail classes during training. For example, label-distribution-aware margin (LDAM) loss proposed by Cao et al. [2] utilizes Rademacher complexity to theoretically prove that the margin should be inversely proportional to a quarter power of label frequencies. The hard margin on target logit helps make the intro-class samples more compact, but does not truly enlarge the tail class span in embedding space."
},
{
"figure_ref": [],
"heading": "Decoupling Representation",
"publication_ref": [
"b9",
"b30",
"b34",
"b34",
"b39",
"b39"
],
"table_ref": [],
"text": "Many recent works have focused on improving the long-tailed visual recognition performance by decoupling the representation and classifier. Most recently, LDAM-DRW [2] has been proposed, which learns features in the first stage and adopts the deferred re-weighting (DRW) to fine-tune the decision boundary in the second stage. It significantly improves the long-tailed prediction accuracy, but the theoretical explanation of DRW is not clear. After that, Kang et al. [10] precisely pointed out that the learning process of representation and classifier can be decoupled into two separate stages. The representation learning is conducted on the original long-tailed data in the first stage and the classifier learning is performed on classbalanced re-sampling data in the second stage. A lot of works [31,32,35,39] have further refined this strategy. For example, Zhang et al. [35] proposed an adaptive calibration function to calibrate the predicted logits of different classes into a balanced class prior in the second stage. Zhong et al. [39] proposed label distribution-based soft label to deal with different degrees of over-confidence for classes and can improve the classifier learning in the second stage. Another alternative direction is proposed by Zhou et al. [40], which splits the network structure into two branches that focus on learning the representation of head and tail classes, respectively. This method incorporates feature mixup [27] into a cumulative learning strategy and also achieves the state-of-the-art results. Following [40], Wang et al. [30] introduced contrastive learning into this bilateral-branch network to further improve the long-tailed classification performance."
},
{
"figure_ref": [],
"heading": "Proposed Approach: GCL",
"publication_ref": [],
"table_ref": [],
"text": "The key idea of our proposed GCL is to utilize the softmax saturation to automatically balance the valid samples of head and tail classes. The theoretical motivation and the formulation of the loss function of the proposed approach are presented as follows."
},
{
"figure_ref": [
"fig_0"
],
"heading": "Motivation",
"publication_ref": [
"b2",
"b4",
"b35"
],
"table_ref": [],
"text": "Fig. 1 shows that the obscure region among different classes, especially the tail class, is large. One important factor of this obscure region is the softmax saturation in CE loss [3]. Suppose {x, y} ∈ T represents a sample {x, y} from the training set T with the total N samples in C classes, and y ∈ {1, . . . , C} is the ground truth label. The softmax loss function for the input image x can be written as:\nL(x) = -log p y , with p y = e zy C j=1 e zj ,(1)\nwhere z j represents the predicted logit of class j. We use the subscript y (j ̸ = y) to represent the target class. That is, z y indicates the target logit and z j (j ̸ = y) is the non-target logit.\nIn backward propagation, the gradients on z j is calculated by: ∂L ∂z j = p j -1, j = y p j , j ̸ = y.\n(2)\nWithout loss of generality, we use the binary classification as an example. Supposing x is from class 1, the gradients on z 1 is then calculated by:\n∂L ∂z 1 = - 1 1 + e z1-z2 .(3)\nEq. ( 3) indicates that the gradient of the target class rapidly approaches zero with the increase of the logit difference. Softmax can only slightly separate various classes, and lacks the power to evenly distribute each class in the embedded space. Therefore, there are many overlapping areas among the classes. In particular, under the circumstances of long-tailed classification, the tail class features are not sufficient to cover the real distribution in embedding space. The early gradient vanish caused by softmax saturation exacerbates the squeezing of their embedding space. A straightforward approach is to introduce hard margin [2, 5,36]. However, the hard margin will cause the samples to shrink towards the class anchor and easy to overfit tail classes, which cannot evenly distribute the embedding space well. Fortunately, softmax saturation can help filter out the head class samples and make the tail class samples fully participate in training. In this way, the tail classes can be pushed away from the head classes and indirectly enlarge their embedding space."
},
{
"figure_ref": [],
"heading": "Embedding Space Calibration",
"publication_ref": [
"b4",
"b28"
],
"table_ref": [],
"text": "Suppose the features of different class samples satisfy Gaussian distribution. We can obtain a disturbed feature f cld of the input by Gaussian sampling, which is represented as:\nf cld ≜ f + δE,(4)\nwhere f ∈ R D is the feature obtained from the embedding layer with the dimension of D. E ∼ N (u, Σ) is the disturbance sampled from Gaussian distribution, and the mean vector and covariance matrix are represented by u ∈ R D and Σ ∈ R D× D , respectively. δ > 0 is a parameter that is used to adjust the amplitude of disturbance. In addition, δ should be a small number because a large disturbance will mislead the model. This disturbed feature is the input of the classifier. We use\nW = {w 1 , w 2 , • • • , w C } ∈ R D×C to\nrepresent the weight matrix of the classifier, where w j represents the anchor vector of class j in the classifier. Then, the corresponding disturbed logit z cld j of class j is calculated by:\nz cld j = w T j f cld + b j = w T j f + b j + w T j (δE) = z j + δ(w T j E).(5)\nAs the range of z cld j is enlarged with random Gaussian disturbances, we call it Gaussian clouded logit, and δ(w T j E) is the clouded term. Please note that the clouded term has the different degrees of influence on the final predicted results based on different predicted logits. It has a relatively small impact on z cld j when the original logit z j is large. On the contrary, it will play a key role for z cld j when z j is small. As a result, we need to normalize the effect caused by different predicted logits and maintain the consistency of the influence of the clouded term. Inspired by [5,28,29], we normalize the clouded logits based on cosine distance. In this way, the norm of the feature and the class anchor can be normalized to the fixed numbers. We use s 1 and s 2 to represent these two numbers. The normalized clouded logit is named clouded cosine logit, which is calculated by:\nzcld j = s 1 w T j • s 2 f cld ∥w T j ∥∥f cld ∥ = s • ( w T j f ∥w T j ∥∥f + δE∥ + δ w T j E ∥w T j ∥∥f + δE∥ ) ,(6)\nwhere s = s 1 • s 2 is a constant. In the first term of Eq. ( 6), ∥f + δE∥ ≈ ∥f ∥ because δ is a small number. In the second term, the norm of feature is normalized to s 1 . Thus, zcld j can be simplified as:\nzcld j ≈ s • ( w T j f ∥w T j ∥∥f ∥ + δ s 1 I j E),(7)\nwhere I j is the identity vector that has the same direction as w T j . In order to simplify the calculation, we make the clouded cosine logit still satisfy the Gaussian distribution. Thus, we introduce a constant σ and set the covariance matrix Σ = σI, where I ∈ R D× D is the identity matrix. Then, I j E is the projection of the noise sampled by Gaussian in the direction of the anchor vector of class j. We denote its magnitude by ε j . Therefore, zcld j can be calculated by:\nzcld j = s • (z j + δ s1 ε j ) ⇔ s • (z j + δ j ε) ,(8)\nwhere zj = cos θ j is the cosine distance, and θ j is the angle between f and w j . δ j is the logit cloud size that depends on different classes.\nTo achieve the two goals mentioned in Sec. 3.1, i.e., 1) encourage tail class samples to participate more in training; 2) enlarge the embedding space for the tail classes, the size of logit cloud should be negatively correlated with the number of training samples. For the most frequent class, the diversity of training samples is sufficient and we set its logit cloud size to zero, while utilizing larger cloud sizes for tail classes. The merits of this large relative cloud size of tail classes are three-fold: 1) reduce the softmax saturation and thereby increase the training degree of tail classes;\n2) different values are sampled randomly from the Gaussian cloud so as to avoid overfitting; 3) enlarge the margin of class boundary for tail classes and can calibrate the distortion of the embedding space. We therefore empirically set the cloud size for class j as:\nδ j = log n max -log n j ,(9)\nwhere n max is the sample numbers of the most frequent class. We experimentally verify the effectiveness of this cloud size adjustment strategy in Sec. 4.5.2 . The Gaussian clouded logit difference ∆ y j between the target and non-target classes is:\n∆ y j = z cld y -z cld j = z y -z j + ε(δ y -δ j ) . (10\n)\nIf ε > 0, ∆ y j for tail classes will be increased. However, our goal is to reduce the logit difference to alleviate the softmax saturation for tail classes. In addition, a reduced logit corresponds to the feature that is relatively far from the class anchor. If the relatively distant feature can be predicted correctly, the closer one will be able to assign the right label. Therefore, we require ε to be negative. Subsequently, the clouded cosine logit can be written in the following form:\nzcld j = s • (z j -δ j ∥ε∥).(11)\nTaking the clouded cosine logit into the original softmax, we can obtain the loss function of GCL:\nL GCL = - 1 N i log e zcld y i j e zcld j .(12)"
},
{
"figure_ref": [],
"heading": "Classifier Re-balance",
"publication_ref": [],
"table_ref": [],
"text": "The gradients derived in Eq. (2) demonstrate that the sample of the target class y punishes the classifier weights w j of non-target class j, j ̸ = y w.r.t. p j . The head classes have enormously greater training instances than tail classes. Therefore, the classifier weights of tail classes receive much more penalty than positive signals during training. Consequently, the classifier will bias towards the head classes, and the predicted logits of the tail classes will be seriously suppressed, resulting in low classification accuracy of the tail classes. A straightforward approach is to use the re-sampled data to re-train the classifier. We apply the classifier re-training (cRT), which was adopted by Kang et al. Calculate sampling rate: The sampling probability ρ j of a sample from class j is calculated by:\nβ j ← b × δj -δmax δmax-δmin + a; ρ j ← 1-β n j j 1-βj ; ρ j ← ρj i ρi ;\nρ j = 1 -β j 1 -β nj j . (13\n)\nSince the sum of the sampling probability for all data needs to be 1, we normalize ρ j by ρ j ← ρj i ρi . β j reflects the validity of different class samples. The class samples with large cloud size participate more in training. Therefore, β j is positively correlated with cloud size δ j . We set β j as:\nβ j = b × δ j -δ min δ max -δ min + a,(14)\nso that β j can be in the region [a, a + b], where a and b are the range hyper-parameters.\nThe overall training procedure of the proposed method is summarized in Algorithm 1."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Datasets",
"publication_ref": [
"b25",
"b40"
],
"table_ref": [],
"text": "We use five benchmarks: long-tailed CIFAR datasets that include CIFAR-10-LT and CIFAR-100-LT, long-tailed ImageNet-2012 (ImageNet-LT), iNaturalist 2018 [26] and long-tailed Places-2 (Places-LT). The original version of CIFAR-10/100 [14], ImageNet-2012 [22] and Places-2 [41] are all balanced datasets. We follow Cao et al. iNaturalist 2018. The 2018 version of iNaturalist is a real-world fine-grained dataset for classification and detection, which exhibits extremely imbalanced distribution. It contains 437.5K training images and 24.4K validation images from 8,142 categories. We follow the official splits of training and validation sets in the experiments."
},
{
"figure_ref": [],
"heading": "Experimental Setting",
"publication_ref": [
"b18",
"b9",
"b9"
],
"table_ref": [],
"text": "The pre-setting parameters in the first stage were the Gaussian distribution parameters (µ, σ 2 ) and the region [a, b] of sample validity β j . We know that zi ∈ [-1, 1], thus the maximum feature cloud size cannot exceed 1. Since Gaussian distribution has a probability of about 99.7% falling in [µ -3σ, µ + 3σ], we set µ = 0 and σ = 1 3 . We further clamped the ε to [-1, 1] to prevent its amplitude from exceeding 1. We set β j ∈ [0.999, 0.9999], i.e. a = 0.999 and b = 0.0009. Moreover, we normalized δ i , i = {1, 2, • • • , C} by δ i ≜ δ i /δ max to ensure that the maximum value of δ i did not exceed 1. Similar with Zhong et al.\n[39], the mixup [33] strategy was also adopted in our experiments.\nWe utilized PyTorch [19] to implement all the back-bones. SGD optimizer with momentum of 0.9 and the multi-step learning rate schedule were adopted. All the models were trained from scratch except ResNet-152 that was pre-trained on the original balanced version of ImageNet-2012. For the first stage, we selected ResNet-32 as the backbone network and followed the setting in Cao et al.\n[2] for CIFAR-10/100-LT. For the large-scale dataset, namely ImageNet-LT, iNaturalist 2018, and Places-LT, we mainly followed Kang et al. [10] except the learning rate schedule. For the second stage, i.e., re-balancing the classifier, we followed Kang et al. [10] for all datasets."
},
{
"figure_ref": [],
"heading": "Competing Methods",
"publication_ref": [
"b11",
"b33",
"b37",
"b39",
"b9",
"b34",
"b9"
],
"table_ref": [],
"text": "To verify the effectiveness of the proposed method, we have conducted extensive experiments to compare with the previous methods, including the following two groups:\nBaseline Methods. We implemented vanilla training with cross-entropy (CE) loss as one of our baseline methods. Many visual recognition works [12,18,34,38] have shown the efficacy of mixup, CE loss cooperated with mixup was therefore also compared.\nState-of-the-art Methods. The recently proposed representation learning method, namely OLTR [16] and logit adjustment method, namely De-confound-TDE inference [24] were compared. We also compared with the two-stage methods including LDAM-DRW [2] and MisLAS [39], which both achieve satisfactory classification accuracy on the aforementioned long-tailed datasets. For CIFAR-10/100-LT datasets, we made comparison with BBN [40] and contrastive learning [30]. For the large-scale datasets, we compared with the most recently proposed two-stage methods, including decoupling [10], logit adjustment [17] and DisAlign [35]. For a fair comparison, we additionally conducted the comparison experiment with the two-stage strategy which added classifier re-training (cRT) [10] to CE loss + mixup on all datasets."
},
{
"figure_ref": [],
"heading": "Comparison Results",
"publication_ref": [],
"table_ref": [],
"text": "Comparative studies have been conducted to show the efficacy of the proposed GCL. The results are presented in Tab. 1 and Tab. 2. We use top-1 accuracy on test sets as the performance metric. For the results from those papers that have yet to release the code or relevant hyper-parameters, we directly quote their results from the original papers."
},
{
"figure_ref": [],
"heading": "Experimental Results on CIFAR-10/100-LT",
"publication_ref": [],
"table_ref": [],
"text": "The results on CIFAR-10/100-LT datasets are summarized in Tab. 1. We can observe that our proposed GCL outperforms the previous methods by notable margins with all imbalanced ratios. Especially for the largest one, i.e., γ = 200, the proposed approach has obvious improvement. We get 79.03% and 44.88% in top-1 classification accuracy "
},
{
"figure_ref": [],
"heading": "Model Validation and Analysis",
"publication_ref": [],
"table_ref": [],
"text": "We conduct a series of ablation studies to further analyze the proposed method."
},
{
"figure_ref": [
"fig_6",
"fig_6",
"fig_6"
],
"heading": "The Role of Gaussian Clouded Logit",
"publication_ref": [],
"table_ref": [],
"text": "In order to obtain additional insight, we utilized t-SNE projection of the embedding for visualization. Since the loss functions of baseline and MisLAS are both CE loss and MisLAS performed the second-best in most cases we have tried so far, we visualized CE loss embedding for comparison. The embeddings were calculated from the samples in CIFAR-10-LT with γ = 100. Fig. 3 shows the visualization results on the training and test set. From the result of the training set (Fig. 3a), we can see that the embeddings obtained via GCL of different classes are more scattered. Therefore, the GCL embedding of each class is much easier to separate. The results of the test set shown in Fig. 3b justify the efficacy of our proposed approach. The obscure region of CE loss embedding is larger than that of GCL em- bedding. Good embedding helps improve the model performance. We only re-fine the classifier with the simple cRT without any other complicated technologies, but the classification accuracy can be improved a lot."
},
{
"figure_ref": [],
"heading": "Cloud Size Adjustment Strategy",
"publication_ref": [],
"table_ref": [],
"text": "We explored several different cloud size adjustment strategies (AS), which included cosine form (cos.), power difference (pow. diff.) with different exponents (e:1/3 and e:1/4), and logarithmic difference (log. diff.). For a fair comparison, the sampler and re-training strategy were selected as CBEN and cRT, respectively. Tab. 3 shows the results. We choose the log. diff. strategy according to Tab. 3."
},
{
"figure_ref": [],
"heading": "Classifier Re-balance Strategies",
"publication_ref": [
"b9",
"b9"
],
"table_ref": [],
"text": "We compared different strategies of data re-sampling and the classifier re-training to better analyze our proposed method. The re-sampling strategy (sam.) included: instance balance (IB) [10], class balance (CB) [10], class "
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "In this paper, we have found that softmax saturation reduces sample validity, which has different effects on head and tail classes. This implies that, from another perspective, softmax saturation can be utilized to automatically adjust the training sample validity of different classes. Subsequently, we have proposed the GCL. The tail class logits are set to relatively large cloud sizes to encourage more tail class samples to participate in training as well as leave large margins, which help obtain evenly distributed embedding space. The effectiveness of different classes is varied via GCL. Then, the simple but effective CBEN sampling strategy incorporated with cRT for classifier balancing has been proposed, which can further boost the model performance. Extensive experiments on various benchmark datasets have demonstrated that the proposed GCL has superior performance compared to the existing state-of-the-art methods."
},
{
"figure_ref": [],
"heading": "",
"publication_ref": [],
"table_ref": [],
"text": "Acknowledgment This work was supported in part by NSFC/RGC JRS Grant: N HKBU214/21, ORP of Zhejiang Lab: 2021KB0AB03, GRF Grant: 12201321, NSFC Grants: 62002302 and 61672444, NSF of Fujian Province: 2020J01005, HKBU Grants: RC-FNRA-IG/18-19/SCI/03."
}
] | Long-tailed data is still a big challenge for deep neural networks, even though they have achieved great success on balanced data. We observe that vanilla training on longtailed data with cross-entropy loss makes the instance-rich head classes severely squeeze the spatial distribution of the tail classes, which leads to difficulty in classifying tail class samples. Furthermore, the original cross-entropy loss can only propagate gradient short-lively because the gradient in softmax form rapidly approaches zero as the logit difference increases. This phenomenon is called softmax saturation. It is unfavorable for training on balanced data, but can be utilized to adjust the validity of the samples in long-tailed data, thereby solving the distorted embedding space of long-tailed problems. To this end, this paper proposes the Gaussian clouded logit adjustment by Gaussian perturbation of different class logits with varied amplitude. We define the amplitude of perturbation as cloud size and set relatively large cloud sizes to tail classes. The large cloud size can reduce the softmax saturation and thereby making tail class samples more active as well as enlarging the embedding space. To alleviate the bias in a classifier, we therefore propose the class-based effective number sampling strategy with classifier re-training. Extensive experiments on benchmark datasets validate the superior performance of the proposed method. | Long-tailed Visual Recognition via Gaussian Clouded Logit Adjustment | [
{
"figure_caption": "Figure 1 .1Figure 1. t-SNE visualization of the distorted embedding space. (Color for the best view.) The embeddings are calculated with ResNet-32 on a subset with four classes of CIFAR-10-LT. We randomly select four classes with the training numbers 500, 200, 100, and 50, respectively. The gray areas show the obscure regions between different classes.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 .2Figure 2. An overview of GCL. (Color for the best view.) The tail class logit is assigned to a larger sample cloud size than the head class, which corresponds to a large relative cloud size of the feature in the direction of the tail class anchor. In this way, the distortion of the embedding space can be calibrated well.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "[10] and Wang et al.[31]. As the GCL loss enables different class samples to participate in training to different degrees, the effectiveness of different class samples is varied. Class-balanced sampling will lead to repeat training for tail classes. Drawing on the effective number proposed by Cui et al.[4], we propose the class-based effective number (CBEN) sampling to avoid excessive training of tail classes.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Algorithm 1 : 4 5 1 b61451Gaussian clouded logit Input: Training dataset T ; Output: Predicted labels; 1 Initialize the model parameters ω of the CNN network ϕ((x, y); ω) randomly ; 2 for iter = 1 to I 0 do 3 Sample a batch samples B from the original long-tailed data T with batch size b;Obtain the logit cloud size:δ j ← log n max -log n j ;Calculate the loss by Eq. (12): L((x, y); ω) = (x,y)∈B L GCL (x, y);Update model parameters: ω = ω -α∇ ω L((x, y); ω). 7 end 8 for iter = I 0 + 1 to I 0 + I 1 do 9",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "1451",
"figure_type": "figure"
},
{
"figure_caption": "10 1 b1Sample a batch samples B ′ with the sampling probability ρ j and the batch size b; 11 Calculate the loss by Eq. (12): L((x, y); ω) = (x,y)∈B ′ L GCL (x, y); 12 Update classifier parameters ω cls (representation parameters are frozen): ω cls = ω cls -α∇ ω cls L((x, y); ω cls ). 13 end",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "[2] and Cui et al.[4] to create long-tailed versions of CIFAR-10/100 and use the long-tailed versions of ImageNet-2012 and Places-2 produced byLiu et al. [16].CIFAR-10/100-LT. The original CIFAR-10 and CIFAR-100 consist of 10 and 100 classes, respectively. They both have 60,000 color images of size 32×32. 50,000 of them are used for training and the remaining images are for validation. Following [2, 4], we down-sampling training samples per class with the exponential function n i = n oi ×µ i , where i is the class index (0-indexed), n oi is the number of training samples in original CIFAR and µ ∈ (0, 1). The validation sets are kept unchanged. The imbalance ratio γ is defined as the ratio of the sample size of the most and the least frequent classes, i.e. γ = max (n i )/ min (n i ), i = 0, 1, ..., C -1. γ is set at its common values, i.e. γ = 50, 100 and 200, in our experiments.ImageNet-LT and Places-LT. The balanced versions of ImageNet-2012 and Places-2 are large-scale real-world datasets for classification and localization. We follow Liu et al.'s work [16] to construct the long-tailed version of these two datasets by truncating a subset with the Pareto distribution with the power value α = 6 from the balanced versions. The original balanced validation sets remain unchanged. Overall, ImageNet-LT has 115.8K training images from 1,000 categories with γ = 1, 280/5. Places-LT contains 62.5K training images from 365 categories with γ = 4, 980/5.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 .3Figure 3. Visualization of the embedding via t-SNE from CIFAR-10-LT with γ = 100, where backbone network is ResNet-32. (Color for the best view.)",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "balance with effective number (EN)[4], and our proposed class-based effective number(CBEN).For a fair comparison, the re-training strategies for all samplers were cRT. Tab. 4 shows the effectiveness of CBEN. For the selection of classifier re-training strategy, we first trained the backbone without any classifier re-training technology. Then, we fixed the representation and re-balance the classifier with learnable weight scaling (LWS) [10], τ -normalization (τ -nor.) [10], and cRT, respectively. Tab. 5 presents the top-1 accuracy of CIFAR-10-LT with γ = 100. We can observe that, even without any classifier re-training technique, our approach can still beat most state-of-the-arts including twostage methods. For example, our GCL without classifier retraining suppresses BBN by 0.7%. Further, cRT performs the best among the classifier re-training strategies, which improves the top-1 accuracy by 1.64%. From Tab. 4 and Tab. 5, we can observe that IB+cRT degrades model performance, which indicates that training the classifier with IB may lead to classifier overfitting.",
"figure_data": "",
"figure_id": "fig_7",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Comparison results on CIFAR-10/100-LT in terms of top-1 accuracy (%), where the best and the second-best results are shown in underline bold and bold, respectively. *indicates that the results are quoted from the corresponding references. The other results are obtained by re-implementing with the official codes.",
"figure_data": "DatasetCIFAR-10-LTCIFAR-100-LTBackbone NetResNet-32Imbalance ratio2001005020010050CE loss65.6870.7074.8134.8438.4343.9CE loss + mixup [33] (2018)65.8472.9679.4835.8440.0145.16LDAM-DRW [2] (2019)73.5277.0381.0338.9142.0447.62De-confound-TDE * [24] (2020)-80.6083.60-44.1550.31CE loss + mixup + cRT [10] (2020)73.0679.1584.2141.7345.1250.86BBN [40] (2020)73.4779.8281.1837.2142.5647.02Contrastive learning * [30] (2021)-81.4085.36-46.7251.87MisLAS [39] (2021)77.3182.0685.1642.3347.5052.62GCL79.0382.6885.4644.8848.7153.55",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Comparison results on ImageNet-LT, iNaturalist 2018 and Places-LT in terms of top-1 accuracy (%), where the best and the second-best results are shown in underline bold and bold, respectively. *indicates that the results are quoted from the corresponding references. The other results are obtained by re-implementing with the official codes.",
"figure_data": "DatasetImageNet-LTiNaturalist 2018Places-LTBackbone NetResNet-50ResNet-50ResNet-152CE loss44.5163.8027.13CE loss + mixup [33] (2018)45.6665.7729.51LDAM-DRW [2] * (2019)48.8068.00-OLTR * [16] (2019)--35.9Decoupling [10] (2020)47.7069.4937.62CE loss + mixup + cRT [10] (2020)51.6870.1638.51Logit adjustment * [17](2021)51.1166.36-DisAlign * [35] (2021)52.9170.0639.30MisLAS [39] (2021)52.1171.5740.15GCL54.8872.0140.64for CIFAR-10-LT and CIFAR-100-LT with γ = 200, whichsurpasses the second best method, i.e., MisLAS by a signif-icant margin of 1.72% and 2.55%, respectively.4.4.2 Experimental Results on Large-scale LatasetsThe results on three large-scale long-tailed datasets, i.e.,ImageNet-LT, iNaturalist 2018, and Place-LT, are reportedin Tab. 2. Our approach is superior to prior art on alldatasets. On ImageNet-LT, our method achieves 54.88%top-1 accuracy, outperforming DisAlign by a large marginat 1.97% and MisLAS at 2.77%, respectively. On iNatu-ralist 2018, the proposed approach achieves 72.01% top-1 accuracy, which outperforms the second-best method by0.44%. On Place-LT, our method achieves 40.64% top-1classification accuracy, with a performance gain at 0.49%over MisLAS. Although the performance gain comparedwith MisLAS on iNaturalist 2018 and Place-LT is not ashigh as other datasets, our method does not require hyper-parameters searching for different datasets, and thus it isrelatively easy to implement.",
"figure_id": "tab_1",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Ablation experiment of different cloud size adjustment strategies (AS) on CIFAR-10-LT with γ = 100.",
"figure_data": "ASExpressionAcc.(%)cos.cos(n j /n max • π/2)79.21pow. diff. (e:1/3)n1/3 max -n1/3 j80.80pow. diff. (e:1/4)n1/4 max -n1/4 j82.31log. diff.log n max -log n j82.68",
"figure_id": "tab_2",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Ablation experiment of different re-sampling strategy on CIFAR-10-LT with γ = 100.",
"figure_data": "Sam.RT Acc.(%)IBcRT80.41CBcRT82.43ENcRT82.47CBEN cRT82.68",
"figure_id": "tab_3",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Ablation experiment of different re-training strategies on CIFAR-10-LT with γ = 100.",
"figure_data": "Sam.RTAcc.(%)-w/o RT 80.52CBEN LWS82.25CBEN τ -nor.82.16CBEN cRT82.68",
"figure_id": "tab_4",
"figure_label": "5",
"figure_type": "table"
}
] | Mengke Li; Yiu-Ming Cheung; Yang Lu | [
{
"authors": "Dong Cao; Xiangyu Zhu; Xingyu Huang; Jianzhu Guo; Zhen Lei",
"journal": "",
"ref_id": "b0",
"title": "Domain balancing: Face recognition on longtailed domains",
"year": "2020"
},
{
"authors": "Kaidi Cao; Colin Wei; Adrien Gaidon; Nikos Arechiga; Tengyu Ma",
"journal": "NeurIPS",
"ref_id": "b1",
"title": "Learning imbalanced datasets with labeldistribution-aware margin loss",
"year": "2019"
},
{
"authors": "Binghui Chen; Weihong Deng; Junping Du",
"journal": "",
"ref_id": "b2",
"title": "Noisy softmax: Improving the generalization ability of dcnn via postponing the early softmax saturation",
"year": "2017"
},
{
"authors": "Yin Cui; Menglin Jia; Tsung-Yi Lin; Yang Song; Serge Belongie",
"journal": "",
"ref_id": "b3",
"title": "Class-balanced loss based on effective number of samples",
"year": "2019"
},
{
"authors": "Jiankang Deng; Jia Guo; Niannan Xue; Stefanos Zafeiriou",
"journal": "",
"ref_id": "b4",
"title": "Arcface: Additive angular margin loss for deep face recognition",
"year": "2019"
},
{
"authors": "Kaiming He; Georgia Gkioxari; Piotr Dollár",
"journal": "",
"ref_id": "b5",
"title": "Ross Girshick",
"year": "2017"
},
{
"authors": "Kaiming He; Xiangyu Zhang; Shaoqing Ren; Jian Sun",
"journal": "",
"ref_id": "b6",
"title": "Deep residual learning for image recognition",
"year": "2016"
},
{
"authors": "Youngkyu Hong; Seungju Han; Kwanghee Choi; Seokjun Seo; Beomsu Kim; Buru Chang",
"journal": "",
"ref_id": "b7",
"title": "Disentangling label distribution for long-tailed visual recognition",
"year": "2021"
},
{
"authors": "Chen Huang; Yining Li; Chen Change Loy; Xiaoou Tang",
"journal": "",
"ref_id": "b8",
"title": "Learning deep representation for imbalanced classification",
"year": "2016"
},
{
"authors": "Bingyi Kang; Saining Xie; Marcus Rohrbach; Zhicheng Yan; Albert Gordo; Jiashi Feng; Yannis Kalantidis",
"journal": "ICLR",
"ref_id": "b9",
"title": "Decoupling representation and classifier for long-tailed recognition",
"year": "2020"
},
{
"authors": "Salman H Khan; Munawar Hayat; Mohammed Bennamoun; Ferdous ; Ahmed Sohel; Roberto Togneri",
"journal": "IEEE TNNLS",
"ref_id": "b10",
"title": "Cost-sensitive learning of deep feature representations from imbalanced data",
"year": "2018"
},
{
"authors": "Jang-Hyun Kim; Wonho Choo; Hosan Jeong; Hyun Oh Song",
"journal": "ICLR",
"ref_id": "b11",
"title": "Co-mixup: Saliency guided joint mixup with supermodular diversity",
"year": "2021"
},
{
"authors": "Pang Wei; Koh ; Percy Liang",
"journal": "",
"ref_id": "b12",
"title": "Understanding black-box predictions via influence functions",
"year": "2017"
},
{
"authors": "Alex Krizhevsky; Geoffrey Hinton",
"journal": "",
"ref_id": "b13",
"title": "Learning multiple layers of features from tiny images",
"year": "2009"
},
{
"authors": "Tsung-Yi Lin; Priya Goyal; Ross B Girshick; Kaiming He; Piotr Dollár",
"journal": "IEEE TPAMI",
"ref_id": "b14",
"title": "Focal loss for dense object detection",
"year": "2020"
},
{
"authors": "Ziwei Liu; Zhongqi Miao; Xiaohang Zhan; Jiayun Wang; Boqing Gong; Stella X Yu",
"journal": "",
"ref_id": "b15",
"title": "Large-scale long-tailed recognition in an open world",
"year": "2019"
},
{
"authors": "Aditya Krishna Menon; Sadeep Jayasumana; Ankit Singh Rawat; Himanshu Jain; Andreas Veit; Sanjiv Kumar",
"journal": "ICLR",
"ref_id": "b16",
"title": "Long-tail learning via logit adjustment",
"year": "2021"
},
{
"authors": "Tianyu Pang; Kun Xu; Jun Zhu",
"journal": "ICLR",
"ref_id": "b17",
"title": "Mixup inference: Better exploiting mixup to defend adversarial attacks",
"year": "2020"
},
{
"authors": "Adam Paszke; Sam Gross; Francisco Massa; Adam Lerer; James Bradbury; Gregory Chanan; Trevor Killeen; Zeming Lin; Natalia Gimelshein; Luca Antiga",
"journal": "",
"ref_id": "b18",
"title": "Pytorch: An imperative style, high-performance deep learning library",
"year": "2019"
},
{
"authors": "Mengye Ren; Wenyuan Zeng; Bin Yang; Raquel Urtasun",
"journal": "",
"ref_id": "b19",
"title": "Learning to reweight examples for robust deep learning",
"year": "2018"
},
{
"authors": "Kaiming Shaoqing Ren; Ross He; Jian Girshick; Sun",
"journal": "IEEE TPAMI",
"ref_id": "b20",
"title": "Faster r-cnn: towards real-time object detection with region proposal networks",
"year": "2016"
},
{
"authors": "Olga Russakovsky; Jia Deng; Hao Su; Jonathan Krause; Sanjeev Satheesh; Sean Ma; Zhiheng Huang; Andrej Karpathy; Aditya Khosla; Michael Bernstein; Alexander C Berg; Li Fei-Fei",
"journal": "IJCV",
"ref_id": "b21",
"title": "Imagenet large scale visual recognition challenge",
"year": "2015"
},
{
"authors": "Jingru Tan; Changbao Wang; Buyu Li; Quanquan Li; Wanli Ouyang; Changqing Yin; Junjie Yan",
"journal": "",
"ref_id": "b22",
"title": "Equalization loss for long-tailed object recognition",
"year": "2020"
},
{
"authors": "Kaihua Tang; Jianqiang Huang; Hanwang Zhang",
"journal": "NeurIPS",
"ref_id": "b23",
"title": "Longtailed classification by keeping the good and removing the bad momentum causal effect",
"year": "2020"
},
{
"authors": "Laurens Van Der Maaten; Geoffrey Hinton",
"journal": "JMLR",
"ref_id": "b24",
"title": "Visualizing data using t-sne",
"year": "2008"
},
{
"authors": "Grant Van Horn; Oisin Mac Aodha; Yang Song; Yin Cui; Chen Sun; Alex Shepard; Hartwig Adam; Pietro Perona; Serge J Belongie",
"journal": "",
"ref_id": "b25",
"title": "The inaturalist species classification and detection dataset",
"year": "2018"
},
{
"authors": "Vikas Verma; Alex Lamb; Christopher Beckham; Amir Najafi; Ioannis Mitliagkas; David Lopez-Paz; Yoshua Bengio",
"journal": "",
"ref_id": "b26",
"title": "Manifold mixup: Better representations by interpolating hidden states",
"year": "2019"
},
{
"authors": "Feng Wang; Xiang Xiang; Jian Cheng; Alan Loddon; Yuille ",
"journal": "ACM MM",
"ref_id": "b27",
"title": "Normface: L 2 hypersphere embedding for face verification",
"year": "2017"
},
{
"authors": "Hao Wang; Yitong Wang; Zheng Zhou; Xing Ji; Dihong Gong; Jingchao Zhou; Zhifeng Li; Wei Liu",
"journal": "",
"ref_id": "b28",
"title": "Cosface: Large margin cosine loss for deep face recognition",
"year": "2018"
},
{
"authors": "Peng Wang; Kai Han; Xiu-Shen Wei; Lei Zhang; Lei Wang",
"journal": "",
"ref_id": "b29",
"title": "Contrastive learning based hybrid networks for longtailed image classification",
"year": "2021"
},
{
"authors": "Tao Wang; Yu Li; Bingyi Kang; Junnan Li; Junhao Liew; Sheng Tang; Steven Hoi; Jiashi Feng",
"journal": "",
"ref_id": "b30",
"title": "The devil is in classification: A simple framework for long-tail instance segmentation",
"year": "2020"
},
{
"authors": "Xin Wang; Thomas E Huang; Joseph Gonzalez; Darrell Trevor; Fisher Yu",
"journal": "",
"ref_id": "b31",
"title": "Frustratingly simple few-shot object detection",
"year": "2020"
},
{
"authors": "Hongyi Zhang; Moustapha Cisse; David Yann N Dauphin; Lopez-Paz",
"journal": "ICLR",
"ref_id": "b32",
"title": "mixup: Beyond empirical risk minimization",
"year": "2018"
},
{
"authors": "Linjun Zhang; Zhun Deng; Kenji Kawaguchi; Amirata Ghorbani; James Zou",
"journal": "ICLR",
"ref_id": "b33",
"title": "How does mixup help with robustness and generalization?",
"year": "2021"
},
{
"authors": "Songyang Zhang; Zeming Li; Shipeng Yan; Xuming He; Jian Sun",
"journal": "",
"ref_id": "b34",
"title": "Distribution alignment: A unified framework for long-tail visual recognition",
"year": "2021"
},
{
"authors": "Wanping Zhang; Yongru Chen; Wenming Yang; Guijin Wang; Jing-Hao Xue; Qingmin Liao",
"journal": "IEEE TNNLS",
"ref_id": "b35",
"title": "Class-variant margin normalized softmax loss for deep face recognition",
"year": "2021"
},
{
"authors": "Yifan Zhang; Bingyi Kang; Bryan Hooi; Shuicheng Yan; Jiashi Feng",
"journal": "",
"ref_id": "b36",
"title": "Deep long-tailed learning: A survey",
"year": "2021"
},
{
"authors": "Yongshun Zhang; Xiu-Shen Wei; Boyan Zhou; Jianxin Wu",
"journal": "",
"ref_id": "b37",
"title": "Bag of tricks for long-tailed visual recognition with deep convolutional neural networks",
"year": "2021"
},
{
"authors": "Zhisheng Zhong; Jiequan Cui; Shu Liu; Jiaya Jia",
"journal": "",
"ref_id": "b38",
"title": "Improving calibration for long-tailed recognition",
"year": "2021"
},
{
"authors": "Boyan Zhou; Quan Cui; Xiu-Shen Wei; Zhao-Min Chen",
"journal": "",
"ref_id": "b39",
"title": "BBN: Bilateral-branch network with cumulative learning for long-tailed visual recognition",
"year": "2020"
},
{
"authors": "Bolei Zhou; Agata Lapedriza; Aditya Khosla; Aude Oliva; Antonio Torralba",
"journal": "IEEE TPAMI",
"ref_id": "b40",
"title": "Places: A 10 million image database for scene recognition",
"year": "2017"
}
] | [
{
"formula_coordinates": [
3,
346.07,
579.81,
199.04,
28.14
],
"formula_id": "formula_0",
"formula_text": "L(x) = -log p y , with p y = e zy C j=1 e zj ,(1)"
},
{
"formula_coordinates": [
4,
125.71,
118.13,
160.65,
23.22
],
"formula_id": "formula_1",
"formula_text": "∂L ∂z 1 = - 1 1 + e z1-z2 .(3)"
},
{
"formula_coordinates": [
4,
137.99,
448.38,
148.38,
11.37
],
"formula_id": "formula_2",
"formula_text": "f cld ≜ f + δE,(4)"
},
{
"formula_coordinates": [
4,
126.07,
561.87,
160.3,
11.23
],
"formula_id": "formula_3",
"formula_text": "W = {w 1 , w 2 , • • • , w C } ∈ R D×C to"
},
{
"formula_coordinates": [
4,
105.58,
630.29,
180.78,
48.94
],
"formula_id": "formula_4",
"formula_text": "z cld j = w T j f cld + b j = w T j f + b j + w T j (δE) = z j + δ(w T j E).(5)"
},
{
"formula_coordinates": [
4,
323.63,
249.6,
221.49,
70.82
],
"formula_id": "formula_5",
"formula_text": "zcld j = s 1 w T j • s 2 f cld ∥w T j ∥∥f cld ∥ = s • ( w T j f ∥w T j ∥∥f + δE∥ + δ w T j E ∥w T j ∥∥f + δE∥ ) ,(6)"
},
{
"formula_coordinates": [
4,
361.12,
378.94,
183.99,
27.44
],
"formula_id": "formula_6",
"formula_text": "zcld j ≈ s • ( w T j f ∥w T j ∥∥f ∥ + δ s 1 I j E),(7)"
},
{
"formula_coordinates": [
4,
378.29,
522.76,
166.82,
30.31
],
"formula_id": "formula_7",
"formula_text": "zcld j = s • (z j + δ s1 ε j ) ⇔ s • (z j + δ j ε) ,(8)"
},
{
"formula_coordinates": [
5,
119,
144.26,
167.36,
9.65
],
"formula_id": "formula_8",
"formula_text": "δ j = log n max -log n j ,(9)"
},
{
"formula_coordinates": [
5,
104.45,
232.74,
177.76,
29.52
],
"formula_id": "formula_9",
"formula_text": "∆ y j = z cld y -z cld j = z y -z j + ε(δ y -δ j ) . (10"
},
{
"formula_coordinates": [
5,
282.21,
243.83,
4.15,
8.64
],
"formula_id": "formula_10",
"formula_text": ")"
},
{
"formula_coordinates": [
5,
120.96,
378.3,
165.4,
10.62
],
"formula_id": "formula_11",
"formula_text": "zcld j = s • (z j -δ j ∥ε∥).(11)"
},
{
"formula_coordinates": [
5,
102.95,
430.51,
183.41,
29.59
],
"formula_id": "formula_12",
"formula_text": "L GCL = - 1 N i log e zcld y i j e zcld j .(12)"
},
{
"formula_coordinates": [
5,
341.14,
285.69,
162.11,
36.12
],
"formula_id": "formula_13",
"formula_text": "β j ← b × δj -δmax δmax-δmin + a; ρ j ← 1-β n j j 1-βj ; ρ j ← ρj i ρi ;"
},
{
"formula_coordinates": [
5,
396.86,
469.11,
144.1,
24.98
],
"formula_id": "formula_14",
"formula_text": "ρ j = 1 -β j 1 -β nj j . (13"
},
{
"formula_coordinates": [
5,
540.96,
476.17,
4.15,
8.64
],
"formula_id": "formula_15",
"formula_text": ")"
},
{
"formula_coordinates": [
5,
369.53,
566.31,
175.58,
23.23
],
"formula_id": "formula_16",
"formula_text": "β j = b × δ j -δ min δ max -δ min + a,(14)"
}
] |
|
10.7289/v5d21vjd | 2023-05-19 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b101",
"b206",
"b110",
"b110",
"b138",
"b275",
"b302",
"b314",
"b303",
"b135",
"b213",
"b368",
"b358",
"b253",
"b303",
"b234",
"b401",
"b326",
"b184",
"b295",
"b223"
],
"table_ref": [],
"text": "Plankton, including phytoplankton and zooplankton, is a fundamental component of aquatic ecosystems. They form the basis of the food web and are essential for global biogeochemical cycles (Arrigo, 2005;Hays et al., 2005). In order to improve management practices of aquatic ecosystems, it is essential to understand functioning of planktonic communities and how they are affected by anthropogenic and climate changes.\nStudying and monitoring plankton is hindered by their microscopic size, fast turnover rates and close interaction with the multiscale hydrodynamics (Benfield et al., 2007). Recent advances in plankton imaging systems have led to their popularization and integration into monitoring programs, collectively accumulating information on plankton systems and simultaneously gathering massive amounts of image data (Benfield et al., 2007;Cowen and Guigand, 2008;Lombard et al., 2019;Olson and Sosik, 2007;Picheral et al., 2010). The major constraint to the use of these datasets lies in the expert annotation of plankton images, which is expensive, time-consuming, and error-prone. To fully benefit from the technological development and to properly explore the gathered information, there is a clear need for automated analysis methods. During recent years, significant research effort has been put into exploring and developing automated methods for performing plankton recognition based on computer vision techniques and machine learning methods (e.g. Lumini and Nanni, 2019a;Orenstein and Beijbom, 2017).\nThe research on automatic plankton image recognition has matured from early works based on hand-engineered image features combined with traditional classifiers such as support vector machine (SVM) (Cortes and Vapnik, 1995) and random decision forest (RDF) (Ho, 1995) (see e.g. Tang et al., 1998;Sosik and Olson, 2007) to feature learning-based approaches utilizing deep learning and especially convolutional neural networks (CNNs) (Lee et al., 2016;Orenstein and Beijbom, 2017;Lumini and Nanni, 2019a;Kloster et al., 2020). Various custom methods and modifications to general-purpose techniques have been proposed to address the special characteristics of plankton image data. However, despite the high recognition accuracies reported in the literature, these methods have not been widely adapted to the operational use. The methods that are utilized typically follow rather simple approaches and do not fully exploit the latest advances in computer vision and machine learning. Deploying deep learning based methods for new environments requires large amounts of training data and expert knowledge while publicly available feature engineering based plankton recognition libraries are accessible for non-experts. Some survey papers on more general microorganism recognition, as well as utilizing machine learning for marine ecology already exist. Zhang et al. (2022) presented a review of machine learning approaches for microorganism image analysis including history, trends, and applications. The paper covers the segmentation, clustering, and classification of various types of microorganism data. Rani et al. (2021) described and compared existing microorganism recognition methods. While the challenges are briefly discussed, the discussion remains on a general level and does not go deeply into the solutions. Li et al. (2019a) provided a review on microorganism recognition for various different application domains with the focus on traditional feature engineering approaches. The survey by Goodwin et al. (2022) covers an even larger scope by addressing the utilization of deep learning methods in marine research. A similar survey was provided by Mittal et al. (2022), who presented existing methods on underwater image classification including fish, plankton, coral reefs, seagrass, and submarines. Irisson et al. (2022) provided a plankton recognition review from the application (aquatic research) point-of-view. They present a rather compact survey of the machine learning methods but provide several insights on utilizing machine learning in solving various application-related research questions. Luo et al. (2021b) considered plankton analysis using imaging flow cytometry. In addition to the different imaging technologies, also automatic image analysis methods are reviewed. Those earlier surveys either have considerably wider scope considering var-ious machine learning tasks and organisms, and therefore, not focusing on challenges specific to plankton recognition, or a more narrow scope concentrating on certain technologies for plankton imaging, and thus, lacking a comprehensive review on plankton recognition in general.\nIn contrast to earlier surveys, we focus on the challenges that researchers commonly face when developing plankton recognition methods and on existing solutions to them. The main goals of this survey are 1) to provide an extensive guide on the available methods to address the challenging characteristics of plankton image data, and 2) to enumerate the challenges that remain unsolved, and which are the most beneficial directions for the future research on the topic. We identify and list the most notable challenges in automatic plankton recognition and provide detailed descriptions of the solutions found in the plankton recognition literature for each challenge. To the best of our knowledge, this paper is the first comprehensive survey focusing exclusively on plankton recognition and the specific challenges related to it.\nThe rest of the paper is organized as follows. In Section 2, the plankton imaging, i.e., imaging instruments and existing image datasets are reviewed. In Section 3, automatic plankton recognition including feature engineering and CNNs are discussed. In Section 4, the most notable challenges in plankton recognition are identified. In Section 5, the existing solutions for each challenge are described. Finally, the paper concludes with a direction for future research in Section 6."
},
{
"figure_ref": [],
"heading": "Plankton imaging",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Imaging instruments",
"publication_ref": [
"b110",
"b275",
"b302",
"b352",
"b129",
"b275",
"b138",
"b190",
"b186",
"b147",
"b314",
"b347",
"b275",
"b121",
"b161",
"b163",
"b162",
"b163",
"b299",
"b381",
"b291",
"b389",
"b381",
"b389",
"b275",
"b358",
"b352",
"b352",
"b352",
"b186",
"b314",
"b138",
"b190",
"b147",
"b347"
],
"table_ref": [],
"text": "A fundamental understanding of how plankton species composition is regulated requires frequent and sustained observations. As plankton communities are diverse and dynamic, monitoring plankton is challenging. Different types of plankton imaging and analysis systems have been developed to identify and enumerate living (plankton) and non-living particles in natural waters (Benfield et al., 2007). Instruments designed for monitoring plankton communities are briefly discussed next (see review by Lombard et al. (2019) for more detailed information). The specifications of the imaging instruments are summarized in Table 1.\nImaging flow cytometry (IFC) combines fluidics, optical characterization and the imaging of cells/colonies. The Imaging FlowCytobot (IFCB) (Olson and Sosik, 2007) and the CytoSense/Cytobuoy (Dubelaar et al., 1999), as well as simpler flow systems such as the FlowCam (Sieracki et al., 1998) and the ZooCAM (Colas et al., 2018) are among the imaging devices most frequently used within aquatic research. The IFCB is a fully automated, submersible instrument with built-in design features that routinely operate during deployments imaging each particle triggering the camera. The CytoSense, available either as a bench top or submersible versions, records forward scatter (FSC), side scatter (SSC) and multiple fluorescence signals of each particle, additionally it can image a subset of the analysed particles. Unlike the IFCB and CytoSense, the FlowCam does not have sheath fluid and it is not an automated in situ instrument. Particle detection in IFCB and CytoSense is triggered by one of the optical sensors (scatter or fluorescence), while FlowCam captures images of a field of view at regular intervals where particles can be identified (autotrigger mode). If the FlowCam is equipped with a laser, particle imaging can be triggered by fluorescence properties, such as the presence of chlorophyll-a. The imaging resolution of the IFCB and CytoSense is targeted for a size range of approximately from larger nanoplankton to smaller mesoplankton. The targeted size range for the FlowCam vary according to the combination of flowcell and objective used and instrument versions for imaging of smaller and larger objects and organisms, FlowCam-Nano and FlowCam-Macro, respectively are currently available and image capture is based on autotrigger. The ZooCAM uses an imaging principle similar to that of FlowCam autotrigger.\nFor obtaining quantitative information from plankton larger than 100 µm, larger volumes of water are needed to be examined than is possible with IFC (Lombard et al., 2019). For imaging of larger particles different types of instruments have been developed utilizing slightly distinct techniques. There are many commercially available instruments such as the In-situ Ichthyoplankton Imaging System (ISIIS) (Cowen and Guigand, 2008), Continuous Plankton Imaging and Classification Sensor (CPICS) (Grossmann et al., 2015), ZooScan (Gorsky et al., 2010), Video Plankton Recorder (VPR) (Davis et al., 2005), Underwater Vision Profiler (UVP) (Picheral et al., 2010), and Lightframe On-sight Keyspecies Investigation (LOKI) (Schulz et al., 2010) which are mostly in situ imaging systems and their operational principles as well as capabilities are reviewed by Lombard et al. (2019). Some instruments have been developed through research purposes but are not commercially available such as the ZooCAM and Prince William Sound Plankton Camera (PWSPC) (Campbell et al., 2020). Some of the more recent imaging instruments include the SPC (Scripps Plankton Camera) system (Orenstein et al., 2020b), a submersible Digital Holographic Camera (DHC) instrument for temporal and spatial plankton measurements (Dyomin et al., 2020(Dyomin et al., , 2019)), and its modification, the miniDHC (Dyomin et al., 2021(Dyomin et al., , 2019)). Also HOLOCAM (Nayak et al., 2018), HoloSea (Walcutt et al., 2020;MacNeil et al., 2021), and LISST-Holo are utilized for underwater microscopy using digital holographic imaging (DHI). SPC utilizes an underwater dark-field imaging microscope combined with an onboard computer that allows real-time processing of the images, while the four latter instruments produce 3-D holograms of the imaged volume. The core principal of DHI is in the optical interference phenomenon. A coherent light source, typically a laser, produces the optical interference pattern between undeviated portion of the beam and light diffracted by the object which is recorded on the sensor, and then holograms are reconstructed with pre-/post-processed computer-based algorithms (Watson, 2018). The main reasons of emerging DHI microscopy are a wide depth-of-field and field-ofview, i.e., larger sampling volume, and mechanically simpler optical configuration compared to lens-based devices (Walcutt et al., 2020;Watson, 2018).\nTable 1: Plankton imaging instruments. For more detailed information about plankton imaging and existing instruments, see Lombard et al. (2019).\nImaging instrument Environment Particle size range Image type In situ onboard laboratory Imaging FlowCytobot (Sosik and Olson, 2007) 10 -150µm monochrome\nCytoSense (Dubelaar et al., 1999) 1 -800µm monochrome FlowCam Nano (Sieracki et al., 1998) 300nm -30µm monochrome FlowCam (2x-10x) (Sieracki et al., 1998) 3 -1000µm * monochrome/color FlowCam Macro (Sieracki et al., 1998) 300µm -5mm monochrome/color ZooScan (Gorsky et al., 2010) 150µm -100mm monochrome LISST-Holo2 25 -2500µm hologram UVP-5 (Picheral et al., 2010) 60µm -20mm monochrome UVP-6LP 60µm -20mm monochrome ISIIS (Cowen and Guigand, 2008) 60µm -130mm monochrome CPICS (Grossmann et al., 2015) 30µm -20mm color VPR (Davis et al., 2005) 30µm -50mm monochrome video LOKI (Schulz et al., 2010) 50µm -20mm monochrome"
},
{
"figure_ref": [
"fig_0"
],
"heading": "Publicly available image datasets",
"publication_ref": [
"b166",
"b137",
"b304",
"b359",
"b231",
"b313",
"b360",
"b316"
],
"table_ref": [
"tab_0"
],
"text": "Publicly available image datasets are crucial on the development of the automatic plankton recognition methods since the most labor intensive part of the process is to create large training and testing datasets. The available datasets are also important for the traceability and comparability of the developed methods. There are several publicly available datasets to be utilized in the research for developing the machine learning methods of plankton recognition. The details of the publicly available and commonly used datasets are summarized in Table 2, and example images from the datasets are shown in Fig. 1. The most frequently used datasets are ZooScanNet (Elineau et al., 2018), Kaggle-Plankton (PlanktonSet-1.0) (Cowen et al., 2015), WHOI-Plankton (Orenstein et al., 2015;Sosik et al., 2021) and their manifold task specific subsets. They all comprise grayscale images collected with a single plankton imaging instrument. UVP5/MC dataset (Kiko and Simon-Martin, 2020) consists of data collected in the EcoTaxa application (Picheral et al., 2017). A part of the UVP5/MC dataset has been annotated by an expert and part with an automated tool. More recently collected datasets include PMID2019 (Li et al., 2019b), miniPPlankton (Sun et al., 2020), DYB-PlanktonNet (Li et al., 2021a), Lake-Zooplankton (Kyathanahally et al., 2021a), and the one collected by Plonus et al. (2021b). They are acquired with modern imaging instruments and characterized by the presence of color and a higher resolution. SYKE-plankton IFCB 2022 (Kraft et al., 2022c) and SYKE-plankton IFCB Utö 2021 (Kraft et al., 2022a) datasets consist of IFCB images of phytoplankton collected from the Baltic Sea. There are also references to some older commonly used plankton datasets that are not available any more. One example is Automatic Diatom Identification And Classification (ADIAC) database (Du Buf et al., 1999)."
},
{
"figure_ref": [],
"heading": "Automatic Plankton Recognition",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Feature engineering",
"publication_ref": [
"b114",
"b118",
"b168",
"b189",
"b358",
"b400",
"b104",
"b111",
"b135",
"b213",
"b368",
"b137",
"b304",
"b166",
"b145",
"b117",
"b125",
"b132",
"b358",
"b400",
"b297",
"b365",
"b168",
"b407",
"b333",
"b358",
"b400",
"b216",
"b371",
"b272",
"b405",
"b404",
"b230",
"b114",
"b367",
"b273",
"b273",
"b334",
"b112",
"b358",
"b232",
"b288",
"b241",
"b367",
"b394",
"b400",
"b217",
"b272",
"b350",
"b390",
"b301",
"b348",
"b123",
"b222",
"b64",
"b118",
"b144",
"b113",
"b277",
"b373",
"b105",
"b123",
"b266",
"b407",
"b236",
"b64",
"b378",
"b407",
"b358",
"b380",
"b407",
"b367",
"b263",
"b123",
"b228"
],
"table_ref": [],
"text": "A traditional solution for image classification including plankton recognition is to divide the problem into two steps: image feature extraction and classification (Blaschko et al., 2005;Bueno et al., 2017;Ellen et al., 2015;Grosjean et al., 2004;Sosik and Olson, 2007;Zetsche et al., 2014;Barsanti et al., 2021). Ideally, image features form a lower-dimensional representation of the image content that contains relevant information for the classification. The main challenge is to design and select good features that are both general and provide good discrimination between the classes. As a result of feature extraction, the obtained feature vectors are used to train a classifier that can then classify unseen images. The most commonly used classifiers for plankton recognition are support vector machine (SVM) (Bernhard et al., 1992;Cortes and Vapnik, 1995) and random decision forest (RDF) (Ho, 1995). SVM in its most simplistic form is a binary linear classifier that works by mapping the data points in the feature space in such way that the margin between two classes is maximised. It can be extended to multi-class case, for example, by utilizing multiple binary classifiers and to non-linear classification by using a kernel trick. The RDF is a widely used classification method that is based on the observation that combining several classifiers to form an ensemble typically provides better classification performance than any of the individual classifiers. In a typical RDF, a large number of decision tree classifiers are constructed and the final classification is obtained by computing the mode of individual classifications. This way, the typical problem of overfitting in the case of decision trees is avoided.\nThe first work on automatic plankton image classification was presented by Tang et al. (1998). The image data were produced using a video plankton (Cowen et al., 2015); (b) WHOI-Plankton (Orenstein et al., 2015); (c) PMID2019 (Li et al., 2019b); (d) ZooScan (Elineau et al., 2018); (d) DYB-PlanktonNet (Li et al., 2021a); (f) SYKE 2022 (Kraft et al., 2022c). recorder (VPR) (Davis et al., 1992) and the proposed method combined texture and shape information of plankton images in a descriptor that is the combination of traditional invariant moment features and Fourier boundary descriptors with gray-scale morphological granulometries. It should be noted that some papers on automatic plankton recognition based on non-image data have been published even earlier. For example, Boddy et al. (1994) utilized light scatter and fluorescence data obtained by flow cytometry to train an artificial neural network (ANN) to classify plankton species.\nFinding good image features is essential for any plankton classification system (Cheng et al., 2018;Corgnati et al., 2016). Various feature extraction technologies have been proposed and put into practice for different underwater imaging environments (Sosik and Olson, 2007;Zetsche et al., 2014). Frequently used plankton features include texture features (e.g. Mosleh et al., 2012), geometric and shape features (e.g. Tan et al., 2014), color features (e.g. Ellen et al., 2015), local features (e.g. Zheng et al., 2017), and model-based features (e.g. Rivas-Villar et al., 2021). Tables A.4 and A.5 in Appendix A categorize and summarize various features used for plankton recognition.\nThe most commonly used image feature type in plankton recognition is shape features (see e.g. Sosik and Olson, 2007;Zetsche et al., 2014) that characterize either the contour or binary mask of the object (plankton). In their simplest form geometric features are numerical descriptors of generic geometric aspects such as major and minor axis length, perimeter, equivalent spherical diameter and area of an object computed from binarized image. Another common approach is to utilize image moments to describe the shape. Both Hu moments (Hu, 1962;Thiel et al., 1995;Liu et al., 2021;Zhao et al., 2005Zhao et al., , 2010) ) and Zernike moments (Khotanzad and Hong, 1990;Blaschko et al., 2005) have been proposed for plankton recognition. Also, various advanced features quantifying the shape of the contour have been proposed for plankton data. These include boundary smoothness (e.g. Tang et al., 2006;Liu and Watson, 2020), affine curvature descriptors (Liu and Watson, 2020), Freeman contour code features (Rodenacker et al., 2006), and elliptical Fourier descriptors (Sánchez et al., 2019a;Beszteri et al., 2018). Further geometric features applied for plankton recognition include symmetry measures (e.g. Hausdorff distance (Guo et al., 2021c;Sosik and Olson, 2007)) and granulometries (Kingman, 1975) utilizing morphological operations (Luo et al., 2005;Kramer, 2005;Tang et al., 2006;Wu and Sheu, 1998).\nOther frequently used type of features in plankton recognition systems are texture features that quantify spatial distribution of intensity or color values in local image regions. While shape features consider only the boundary of plankton, texture features describe the region inside the boundary. The simplest texture features commonly applied in plankton recognition are first-order statistical descriptors that compute simple statistical values directly from the intensity values (see e.g. Lisin, 2006;Zetsche et al., 2014;Guo et al., 2021c). These are sometimes called color features and include, for example, mean intensity, variance of intensity, as well as, skewness and kurtosis that quantify the shape of the color or intensity histogram. The first order statistics only provide information on how the intensity or color values are distributed in the image. To obtain further spatial information on texture, various second-order statistical descriptors have been proposed. The most common second-order statistical descriptor used in plankton recognition is the co-occurrence matrices (Hu and Davis, 2005;Liu et al., 2021;Shan et al., 2020;Wei et al., 2022), that describe the statistics of pixel color pairs occurring with certain distance from each other in the image. More advanced texture features proposed for plankton recognition include Local Binary Patterns (LBP) (Ojala et al., 2002;Schulze et al., 2013;Chang et al., 2016;Lisin, 2006), and Gabor descriptors (Idrissa and Acheroy, 2002;Sánchez et al., 2019b;Bueno et al., 2017).\nThe third widely utilized group of image features is local features that typically combine the feature detectors and descriptors. Feature detectors search the image for characteristic interest points or regions that contain useful information for the task, i.e. plankton recognition. Local feature descriptors then quantify these regions. General-purpose feature descriptors that have been applied for plankton images include Histogram of Oriented Gradient (HOG) (Dalal and Triggs, 2005;Bi et al., 2015;Guo et al., 2021c), Scale Invariant Feature Transform (SIFT) (Lowe, 2004;Tsechpenakis et al., 2007), Speeded Up Robust Features (SURF) (Bay et al., 2006;Chang et al., 2016), Inner-Distance shape context (IDSC) (Ling and Jacobs, 2007;Zheng et al., 2017), and Phase congruency descriptors (PCD) (Kovesi, 2000;Sánchez et al., 2019b;Verikas et al., 2012).\nFeature engineering-based methods for plankton recognition usually combine features from different groups to obtain more representative feature vectors. For example, Zheng et al. (2017) used geometric features (e.g. size and shape measurements, such as area, circularity, elongation, convex rate), color features (e.g. sum, mean, standard deviation of color values), texture features (e.g. Gabor descriptors and Local Binary Pattern (LBP)) and local features (e.g. HOG and SIFT). Sosik and Olson (2007) applied simple geometry features, shape and symmetry features, as well as texture features including co-occurrence matrices for phytoplankton recognition. Wacquet et al. (2018) extracted 26 features including basic shape features, advanced morphological features, and color features.\nTypical plankton recognition systems further apply additional feature selection (see e.g. Zheng et al., 2017) or dimensional reduction steps to construct compact feature representations. In feature selection, the large set of initial features are ranked based on how representative or informative they are, and the least informative features are discarded. For example, Tang et al. (2006) proposed normalized multilevel dominant eigenvector estimation (NMDEE) technique to select a best feature set for plankton recognition. In dimensional reduction, principle component analysis (PCA) or similar technique is applied to reduce the length of the extracted feature vector while preserving maximum amount of information. For example, Li et al. (2013) and Chang et al. (2016) utilized PCA as a part of the plankton recognition system.\nAlthough feature-engineering-based techniques have been applied with promising results, they require discrete parts, i.e., feature extraction, selection, and training a classifier. Due to the difficulty of finding general features that provide high classification accuracy over different datasets, feature engineering based plankton recognition methods are often ad-hoc solutions tuned for a single imaging instrument and provide limited accuracy. Moreover, based on previous works (Al-Barazanchi et al., 2015b;Khalid et al., 2014), it typically requires extensive work to integrate a new class to the existing system. Each new class requires intensive work to find new features that could represent the new class. Depending on the quality of feature design, providing a suitable framework for the accurate, rapid and simplified classification of plankton species is not always possible."
},
{
"figure_ref": [],
"heading": "Convolutional neural networks",
"publication_ref": [
"b370",
"b191",
"b304",
"b303",
"b294",
"b410"
],
"table_ref": [
"tab_1"
],
"text": "Recently, CNNs have replaced traditional feature engineering techniques in various computer vision applications. The notable difference is that the image features are learnt from the data instead of manually designing them. CNN (LeCun et al., 2015) is a type of neural network model for image processing inspired by the animal visual cortex. The key component of CNNs are the convolutional layers that consist of neurons each processing data only for their receptive field. Due to the shared-weight architecture, these neurons fundamentally perform the convolution operation to the input with a filter defined by the weights of the neurons. This makes it possible to learn the feature extraction filters (weights) through backpropagation. A typical CNN involves repetitions of several convolution layers and a pooling layer, followed by a set of fully connected layers. The convolution and pooling layers perform feature extraction and the fully connected layers perform the higher-level reasoning and map the extracted features into final output. An example of CNN structure is shown in Fig. 2. In the recent years CNN-based approaches have become dominant in various image analysis tasks providing state-of-the-art performance, for example, in image classification, object localization, and image segmentation tasks (Teuwen and Moriakov, 2020). One reason why CNNs have become more popular is that they have been shown to outperform the traditional approach utilizing feature engineering multiple times and the architectural components have been studied with care (Gu et al., 2018). For example, Zheng and Wang (2015) compared a CNN-based plankton image classifier to traditional classifiers such as a multi-layer perceptron (MLP) model utilizing hand-engineered features. The results showed that CNN outperformed the earlier methods. In various experiments (Orenstein et al., 2015;Orenstein and Beijbom, 2017;Guo et al., 2021c), CNNs have demonstrated higher plankton recognition accuracy than RDF combined with hand-selected features. The preliminary experiments done by Mitra et al. (2019) suggest that CNN can even surpass the human in plankton recognition accuracy. However, in some special cases, if the computation time is heavily restricted (e.g. embedded systems), feature-engineering based approaches might still be preferable (see e.g. Zimmerman et al., 2020). 1 9 8 9 1 9 9 4 1 9 9 7 1 9 9 8 2 0 0 1 2 0 0 2 2 0 0 3 2 0 0 4 2 0 0 5 2 0 0 6 2 0 0 7 2 0 0 8 2 0 0 9 2 0 1 0 2 0 1 1 2 0 1 2 2 0 1 3 2 0 1 4 2 0 1 5 2 0 1 6 2 0 1 7 2 0 1 8 2 0 1 9 2 0 2 0 2 0 2 1 2 0 2 The number of papers per year "
},
{
"figure_ref": [],
"heading": "Feature engineering CNN",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "CNN architectures",
"publication_ref": [
"b243",
"b219",
"b209",
"b353",
"b362",
"b220",
"b353",
"b199",
"b64",
"b363",
"b234",
"b96",
"b260",
"b397",
"b340",
"b396",
"b321",
"b291",
"b402",
"b341",
"b224",
"b303",
"b64"
],
"table_ref": [],
"text": "Numerous CNN architectures have been suggested for plankton recognition. These include various common CNN developed for generic image recognition. For example, Lumini and Nanni (2019a) compared AlexNet (Krizhevsky et al., 2012), DenseNet (Huang et al., 2017), ResNet (He et al., 2016), VGGNet (Simonyan and Zisserman, 2014), GoogleNet (Szegedy et al., 2015), and SqueezeNet (Iandola et al., 2016). DenseNet produced the best classification results with ZooScan, Kaggle-Plankton and WHOI datasets. Liu et al. (2018a) evaluated AlexNet, VGG16 (Simonyan and Zisserman, 2014), GoogleNet, PyramidNet (Han et al., 2017) and ResNet. The results suggest that PyramidNet provided improvement on accuracy on a WHOI-Plankton dataset. Sánchez et al. (2019b) performed a comparison of ResNet, AlexNet, VGGNet, SqueezeNet, DenseNet, and InceptionV3 (Szegedy et al., 2016) on a dataset consisting of 1085 diatom images of 14 different classes and DenseNet, ResNet and VGG provided the highest accuracy. Kloster et al. (2020) tested extensively various CNN architectures with fine-tuning. Notably, relatively shallow VGG-16 model outperformed more modern architectures. Table A .6 in Appendix A gives a summary of different architectures that have been utilized in plankton recognition.\nThere are also CNN architectures developed specifically for plankton recognition. Al-Barazanchi et al. (2018) proposed a shallow VGGNet-based architecture for the task. Dai et al. (2016a) proposed a CNN architecture called ZooplanktoNet that was characterized by the ability to capture more general and representative features than previous predefined feature extraction algorithms. It was strongly inspired by AlexNet and VGGNet. A comparative experiment with different CNN architectures including AlexNet, VGGNet and GoogleNet was carried out and ZooplanktoNet was found to outperform other architectures on zooplankton classification. Li et al. (2019c) proposed tiny attention network (TANet) consisting of three main parts: a reduction module, self-attention operation, and group convolution. The reduction module was utilized to reduce the information loss caused by pooling operation, self-attention was used to improve the feature learning ability and the group convolution was applied to compress the model size. One of the benefits of the TANet model is its small size which allows real-time classification on mobile devices. Yan et al. (2017) proposed another light CNN architecture for plankton recognition was proposed by utilizing smaller filter size and less fully-connected layers. Luo et al. (2021a) presented a custom architecture MCellNet derived from MobileNetV2 (Sandler et al., 2018). The model was shown to outperform MobileNetV2 on plankton data on both accuracy and computation time. Xu et al. (2022) developed a CNN for classifying algae based on ResNet and SeNet architectures.\nCustom architectures have also been developed for holographic microscopy images as existing image recognition models cannot be directly applied to raw digital holographic microscopy data. A straightforward approach is to first reconstruct images and then utilize any common image recognition architecture (see e.g. Qiao et al., 2021;MacNeil et al., 2021). This, however, leads to long processing times as the reconstruction stage is computationally heavy. It has been shown that by using a custom architecture CNNs can be successfully applied to the raw digital holographic data and the reconstruction step can be avoided (Guo et al., 2021a;Zhang et al., 2021). Also, simulated holograms have been proposed for training and testing simultaneous detection and classification of plankton (Scherrer et al., 2021).\nVarious works have suggested to use CNNs only for the feature extraction and utilize other classifiers, such as SVM or RDF for the final classification step. Jindal and Mundra (2015) suggested to use output of the first fully-connected layer of two CNNs (ClassyFireNet and GoogLeNet) as image features and fed it to RDF for plankton recognition. Similar approach was evaluated by Orenstein and Beijbom (2017) who utilized AlexNet to extract features for an RDF-based classifier. Sánchez et al. (2019b) compared both approaches: fine-tuned CNN for classification, and CNN for feature extraction. Based on the experiments with various CNN architectures fine-tuned CNN outperformed the approach where CNN was used as feature extractor.\nOther commonly used approach is to combine multiple CNNs into ensemble to improve the accuracy. This so called ensemble learning is based on the assumption that limited performance of an individual recognition model can be compensated by utilizing additional models more capable of classifying different sets of classes. Kuang (2015) proposed various approaches for model ensemble. These include averaging softmax probabilities and applying principal component analysis for concatenated CNN features before softmax classifier. Lumini and Nanni (2019a) "
},
{
"figure_ref": [],
"heading": "Hybrid methods",
"publication_ref": [
"b303",
"b226",
"b280",
"b333",
"b139",
"b169",
"b235",
"b215",
"b281",
"b188",
"b127"
],
"table_ref": [],
"text": "Multiple methods that aim to combine the feature engineering approach with CNNs have been proposed. One approach is to utilize a separate classifier (e.g. RDF) as above. This way CNN features can be simply supplemented with selected hand-engineered features before classification (see e.g. Orenstein and Beijbom, 2017;Keçeli et al., 2017). Similarly, ensembles of classifiers can be utilized to combine handcrafted feature based classification and CNNs. For example, in the method proposed by Lumini and Nanni (2019a); Lumini et al. (2020) individual classifiers utilized in the ensembles included various CNNs applied to both original images and preprocessed (filtered) images. The preprocessing techniques included various filters commonly used to compute local features, such as gradient, LBP and wavelets. Rivas-Villar et al. (2021) combined color and texture features with deep CNN features. Both RDF and SVM were tested for classification. Dai et al. (2016b) proposed a multi-stream CNN for plankton classification. In addition to the original image, global feature image representing the shape and local feature image representing the edge information were used as input. All three images were processed with a CNN through separate streams. Similar approach was proposed in paper by Cui et al. (2018), where the original image, shape image, and texture image were processed in streams before feature concatenation. Concatenated feature maps were processed with one more convolutional and pooling layer, a set of fully connected layers and softmax layer. A related approach was proposed by Ellen et al. (2019) who utilized non-image information (metadata) in the CNN-based plankton classification. Various architectures to fuse Metadata with CNN-based image features were proposed consisting of a set of convolutional and pooling layers for the image and fully-connected layers for the metadata before feature concatenation and common fully-connected layers for the classification.\nAlso various other modifications to baseline CNN classifiers exist. Kosov et al. (2018) proposed Conditional Random Field model to utilize spatial relations among pixel-based CNN classification results and global features for microorganism detection and recognition. Liu et al. (2018b) proposed to include squeeze-and-excitation block (Hu et al., 2018) to deep pyramidal residual network to increase the plankton recognition accuracy. Luo et al. (2018) took into account the fact that typical plankton images contain a large amount of background pixels without useful information and applied spatially sparse convolutional neural networks originally developed for handwriting recognition (Graham, 2014). Cheng et al. (2020) proposed to combine two CNNs, one applied to normal Cartesian coordinate image and one to the same image transformed into Polar representation. This way rotational invariance was obtained in addition to the translation invariance of the baseline CNN."
},
{
"figure_ref": [],
"heading": "Transformers",
"publication_ref": [
"b153",
"b247"
],
"table_ref": [],
"text": "In addition to CNNs, also other feature learning approaches have been proposed for plankton recognition. One of the most promising approach is Vision Transformers (ViTs) (Dosovitskiy et al., 2021), that works by dividing the image into patches resulting in a sequence of vectors (tokens) that are fed to the model. The architecture allows the model to measure relationships between pairs of image patches making it possible to learn to identify the most informative regions in an image via self-attention. Kyathanahally et al. (2022) applied ensembles of Data-efficient image Transformers (DeiTs) for various ecological image datasets including four publicly available plankton datasets and provided state-of-the-art performance."
},
{
"figure_ref": [],
"heading": "Plankton detection",
"publication_ref": [
"b296",
"b180",
"b329",
"b312",
"b354",
"b387",
"b332",
"b271",
"b330",
"b177",
"b235"
],
"table_ref": [],
"text": "Depending on the imaging instrument, there is sometimes a need to first detect the plankton particles in the images (Moniruzzaman et al., 2017). Modern CNN-based object detection methods such as R-CNN (Girshick et al., 2014), YOLO (Redmon et al., 2016), and their modifications perform the detection and recognition simultaneously, providing end-to-end methods for plankton recognition. For example, Pedraza et al. (2018) applied R-CNN to detect and classify diatoms in microscopy images, and Soh et al. (2018) used YOLO to detect and recognize plankton. Wang et al. (2022b) compared multiple CNN-based object detection methods including Faster R-CNN (Ren et al., 2017), SSD (Liu et al., 2016), YOLOv3 (Redmon and Farhadi, 2018) and YOLOX (Ge et al., 2021) on imaging flow cytometer data. YOLOX achieved the best accuracy. Li et al. (2021b,c) proposed an improved YOLOv3-based model for plankton detection. The proposed model contains two YOLOv3 networks fused with DenseNet architecture. Kosov et al. (2018) applied CNN-based images, features and conditional random fields for plankton localization and segmentation."
},
{
"figure_ref": [],
"heading": "Comparison",
"publication_ref": [
"b343",
"b102"
],
"table_ref": [
"tab_1"
],
"text": "Many papers utilize in-house datasets and most publicly available datasets do not provide standardized evaluation protocol meaning that different papers utilize different train-test splits and performance metrics. This makes comparison of the performance of different solutions challenging before the principles of making the science findable, accessible, interoperable, reusable (FAIR) are fully adopted (Schoening et al., 2022). Table 3 summarizes some published results obtained on publicly available datasets. However, the provided accuracies are not directly comparable due to the reasons mentioned above. One notable comparison of plankton recognition methods is The National Data Science Bowl (Aurelia et al., 2014) from 2015. The winning team used an ensemble of over 40 convolutional neural networks."
},
{
"figure_ref": [
"fig_5"
],
"heading": "Challenges in plankton recognition",
"publication_ref": [
"b290",
"b304"
],
"table_ref": [],
"text": "Based on the literature on automatic plankton recognition various challenges can be identified. The most notable challenges are as follows: Image classification with datasets that suffer from a greatly imbalanced class distribution is a challenging task in the computer vision field. Data of plankton species naturally exhibit an imbalance in their class distribution, with some plankton species occurring naturally more commonly than others. This results in highly biased datasets and makes it difficult to learn to recognize rare species, having a serious impact on the performance of classifiers. Furthermore, with highly unbalanced datasets the overall classification accuracy (e.g. percentage of images that were correctly classified) provides little information about the classes with a small number of samples which may bias the evaluation of the goodness of the classification methods. 3. Visual differences between certain classes are small.\nCertain plankton species, especially those that are taxonomically close to each other, resemble each other visually, which renders the recognition task a fine-grained classification problem. Limitations in the amount of training data make it challenging to ensure that the recognition model learns the subtle differences between the classes reducing the recognition accuracy. 4. Imaging instruments vary between datasets.\nIf two datasets have been obtained with different imaging instruments producing visually different images (domain shift) the classification model trained on one dataset does not provide sufficient classification accuracy on the other dataset when applied directly. This makes it challenging to develop general-purpose classifiers that could be applied to new datasets limiting the applicability of the existing publicly available large image datasets. There is a need for approaches that allow the adaptation of the trained models to new imaging instruments. 5. Training sets do not contain all the classes that can be captured.\nWhen deploying a recognition model in operational use, it should be able to handle images from the classes that were not present in the training phase. Different datasets often have different sets of plankton species due to, for example, the geographical distance between the imaging locations or the particle size range of the imaging instruments.\nMoreover, imaging instruments capture images of unknown particles. Typical CNN-based classification models trained on one dataset tend to classify the images from a previously unseen class to one of the known classes often with high confidence, which not only makes the models incapable to generalize to new datasets and analyze noisy data but makes it difficult to recognize when the model fails. This calls for methods that can identify when the image is from a previously unseen class (species). 6. There are uncertainties in expert labels.\nDue to limited imaging resolutions and low image quality, recognizing plankton species is often difficult even for an expert. Manually labeling large amounts of images is tedious work increasing the risk of human errors. Moreover, due to the high costs of labeling work, it is typically not possible to obtain opinions from multiple experts for each image. These reasons cause inaccuracies (uncertainty) in labels to the training data decreasing the classification performance of the trained models. Furthermore, this uncertainty is often highly imbalanced since some of the classes are easier to identify than others. 7. Variation in image size and aspect ratio is very large.\nMost CNN architectures require that the input images have fixed dimensions and a typical approach in image classification is to first scale the images into a common size. This is not ideal in plankton recognition due to a very large variation in both the size and aspect ratio of plankton. Scaling images into a common size may cause either small details to be lost in the large images (downscaling) or very large and computationally heavy models (upscaling). Furthermore, the size is an important cue for recognizing the plankton species and this information is lost in scaling. 8. Image quality is often low or has extensive variation.\nPlankton imaging requires high magnification and the (natural) water might contain other particles, cause unwanted optical distortions, as well as limit the visibility. More importantly, due to the limited depthof-field, automated imaging instruments often fail to capture particles in focus and the focus may drift away from optimal setting. These reduce the quality of images. The low image quality makes both manual labeling (Challenge 6) and automatic classification considerably more challenging. Therefore, there is a need for plankton recognition solutions that are robust to image distortions such as blur and noise.\n9. The amount of image data is massive.\nModern plankton imaging instruments produce massive amounts of image data, e.g. FlowCam Macro and ISIIS have the ability to take 10,000 images per minute and 64,000 images per hour respectively. Computationally efficient solutions are needed to perform the analysis in realtime (MacLeod et al., 2010;Orenstein et al., 2015).\nAll the nine challenges are visualized in Fig. 4. "
},
{
"figure_ref": [],
"heading": "Existing solutions",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [
"fig_6",
"fig_7"
],
"heading": "Challenge 1: Limited amount of training data",
"publication_ref": [
"b288",
"b288",
"b154",
"b115",
"b388",
"b203",
"b126",
"b133",
"b179",
"b169",
"b179",
"b281",
"b133",
"b179",
"b179",
"b179",
"b303",
"b374",
"b160",
"b357",
"b308",
"b351",
"b391",
"b148",
"b294",
"b280",
"b335",
"b328",
"b303",
"b224",
"b260",
"b202",
"b369",
"b103",
"b345",
"b320",
"b195",
"b392",
"b130",
"b342",
"b150",
"b346",
"b122",
"b344",
"b207",
"b198",
"b99",
"b346",
"b187",
"b310"
],
"table_ref": [],
"text": "The two main reasons limiting the amount of training data, the requirement of expert knowledge for the very laborious labeling task and rarity of certain plankton species, require different solutions.\nActive learning has been utilized to minimize the effort of expensive human experts in labeling plankton image data (Luo et al., 2005). The basic idea behind active learning is to select only the most informative samples for labeling. A classifier is first trained on a small initial training set and the method iteratively seeks to find the most informative samples from an unlabeled dataset. These samples are then labeled by a human expert and the model is re-trained. A simple active learning technique for plankton images called \"breaking ties\" was proposed by Luo et al. (2005). The method utilizes probability approximation for SVM-based classifier and ranks the unlabeled images based on the differences between the largest and the second largest class probabilities (the smaller the difference the less confident the classifier is). Images with the smallest confidence were labeled by an expert. Drews et al. (2013) studied semi-automatic classification and active learning approaches for microalgae identification. A Gaussian mixture model (GMM) model is estimated from the image feature data and three different sampling strategies are used for the active learning. The experimental results show the benefit of using active learning to improve the performance with few labeled samples. Bochinski et al. (2018) proposed Cost-Effective Active Learning (CEAL) (Wang et al., 2016) for plankton recognition. In contrast to traditional active learning where only the manually annotated samples are used in the model training, CEAL utilizes also the unlabeled high-confidence samples for training with class predictions as pseudo labels. Haug et al. (2021b); Haug (2021); Haug et al. (2021a) proposed Combined Informative and Representative Active Learning technique (CIRAL) to minimize the human involvement in the plankton image labeling process. The main idea behind the method is to find the images with minimal perturbations that are often miss-classified and ignore the images that are far from the decision boundary. The DeepFool algorithm is used to compute small perturbations to the images. The finding of the representative images is formulated as a min-max facility location problem and solved using a greedy algorithm.\nWhile active learning helps to reduce manual work, it is often still a timeconsuming process. Typically, there is a need to obtain more training data in a completely automated manner. A traditional approach to increase the amount of training data is to utilize data augmentation. By augmenting the existing labeled image data with various image manipulations, the diversity of the training data, and therefore, the generalizability and accuracy of the trained model can be improved. The most commonly used data augmentation techniques for plankton image recognition include rotation (e.g. Cheng et al., 2019;Correa et al., 2017), shearing (e.g. Dai et al., 2016a;Geraldes et al., 2019), flipping (e.g. Ellen et al., 2019;Geraldes et al., 2019), rescaling (e.g. Li and Cui, 2016;Luo et al., 2018), and additional noise (e.g. Correa et al., 2017;Geraldes et al., 2019). Also, blurring (Geraldes et al., 2019), contrast normalisation (Geraldes et al., 2019), geometric transformations (Orenstein and Beijbom, 2017;Vallez et al., 2022), as well as adjusting brightness, saturation, contrast, and hue (Dunker et al., 2018) have been utilized. Some works augment images using translation (e.g. Dai et al., 2016a;Li and Cui, 2016). However, it should be noted that CNNs are invariant to translation by design, and therefore, this is typically unnecessary when CNNs are used for recognition. Augmentation has been shown to increase the plankton recognition accuracy even with relatively large training sets (see e.g. Song et al., 2020). Examples of augmented images are shown in Fig. 5. Another commonly used approach to address a small amount of training data is transfer learning. Transfer learning is a machine learning method that utilizes knowledge gained from the source domain, where training data are abundant, to the target domain, where training data are scarce (Pan and Yang, 2009;Shao et al., 2014;Weiss et al., 2016) (see Fig. 6). In the context of plankton recognition, this typically means that the model is first trained using either general image datasets (e.g. ImageNet (Deng et al., 2009)) or a large publicly available plankton dataset and then fine-tuned for the target plankton dataset with typically a limited number of labeled images. Using general image databases as source data is justified by the fact that the learned low level image features are often useful despite the classification problem. In the simplest case transfer learning can be done by simply replacing and training the classification layer and keeping the feature extraction layers unchanged (see e.g. Mitra et al., 2019). However, it is often beneficial to use the pre-trained network only for initialization and retrain (or fine-tune) the whole network with the target dataset (Lumini et al., 2020). One way to apply transfer learning for plankton images is to use trained CNNs only for feature extraction and utilize general classification methods such as SVM or RDF for the recognition (see e.g. Rodrigues et al., 2018;Rawat et al., 2019). However, the results by Orenstein and Beijbom (2017) suggest that better accuracy is obtained by utilizing end-to-end CNN with classification layers. Lumini Large models with more parameters typically require a large amount of data to be trained without overfitting the model. To avoid this and allow the training with a smaller amount of data, shallower CNN architectures have been proposed for plankton recognition. For example, the 18-layer version of ResNet architecture has been shown to achieve a high plankton recognition accuracy on IFCB data (Kraft et al., 2022b). Most custom CNN architectures developed especially for plankton recognition including ClassyFireNet (Jindal and Mundra, 2015), TANet (Li et al., 2019c), and ZooplanktoNet (Dai et al., 2016a) are relatively shallow with 8, 8 and 11 layers, respectively. It has been shown that a good classification accuracy could be obtained with a shallow architecture and by using suitable data augmentation methods even with as few as 10 images per class (Kraft et al., 2022b).\nIn addition to data manipulation and custom recognition models, also model training approaches have been considered to address the limited data amounts. Learning techniques developed for training the classifier with a minimal amount of samples are called few-shot learning methods. Typically, the idea is to utilize some prior knowledge to allow the generalization to new tasks (in this case classification of new plankton species) containing only a few labeled training examples. Common ways to address few-shot learning is to utilize generation (Hariharan and Girshick, 2017), embedding or metric learning. The basic idea is to learn such embeddings that the images from the same class are close to each other in the metric space and images from the different classes are far. This allows performing the plankton recognition using distances to the images with known plankton species. Embedding and metric learning have been successfully applied to plankton recognition (Teigen et al., 2020;Badreldeen Bdawy Mohamed et al., 2022). Schröder et al. (2018) employed a low-shot learning technique called weight imprinting (Qi et al., 2018) for plankton recognition with a limited amount of training data. The main idea of weight imprinting is to divide the set of all classes into base classes with enough training data and smaller lowshot classes. During the representation learning phase, a CNN is trained to distinguish the base classes with a large amount of training data. In the second phase (low-shot learning), the classifier is then updated with calculated weights to distinguish the smaller low-shot classes. This is done by using appropriately scaled class features of the low-shot classes as their weights, directly allowing the inclusion of classes with only one training image. Guo and Guan (2021) addressed the few-shot learning by supplementing the softmax loss with center loss term (Wen et al., 2016) that forces the samples from the same class close to each other in the deep feature space. The loss function is a weighted sum of the two loss terms and a regularization parameter is used to control the weights.\nIn the extreme case, the training data are completely absent and unsupervised learning methods are required. Image clustering is the most commonly used unsupervised technique for plankton image analysis. Ibrahim (2020) carried out preliminary experiments on common clustering algorithms such as k-means with phytoplankton data. Image features for clustering were extracted using pretrained CNN models. Coltelli et al. (2014) used various handcrafted image features and self-organizing maps (SOM) for plankton image clustering. Schmarje et al. (2021) proposed a framework for handling semi-supervised classifications of fuzzy labels due to experts having different opinions. The approach is based on overclustering to identify substructures in the fuzzy labels and a loss function to improve the overclustering. The performance surpassed the one of a state-of-the-art semi-supervised method on plankton data. Salvesen (2021) studied deep learning for plankton classification without ground truth labels. The improved feature learning was implemented using DeepCluster, a Generative Adversarial Network (GAN) and a rotation-invariant autoencoder. Despite the potential in unsupervised methods, the gap to supervised learning is still significant.\nHierarchical clustering methods are preferred on plankton data as they have the potential to mimic the taxonomic hierarchy of plankton. Dimitrovski et al. (2012), classification of diatom images is considered as a hierarchical multi-label classification problem and solved by constructing predictive clustering trees that can simultaneously predict all different levels in the taxonomic hierarchy. These trees are then used as an ensemble forming a random forest (RF) to improve the predictive performance. Morphocluster (Schröder et al., 2020) utilizes a semi-automated iterative approach and hierarchical density-based HDBSCAN* (Campello et al., 2015) for plankton image data analysis. To compute image features for the clustering a CNN trained with UVP5/EcoTaxa dataset in a supervised manner was used. The method works iteratively in a semi-automated manner so that clusters are validated by an expert. An improved version of Morphocluster was presented by Schröder and Kiko (2022). Multiple CNN-based feature extractors were trained using different labeled datasets to allow the selection of the most suitable feature extractors for the target data. In addition, an unsupervised approach to learn the plankton image features based on the momentum contrast method (He et al., 2020) was proposed. The idea is to use data augmentation to gener-ate two different instances of the same image and use a loss function that forces the model to learn similar feature representations for both instances. Moreover, two custom clustering methods were proposed: 1) shrunken k-Means, and 2) Partially Labeled k-Means. Due to the iterative clustering process of Morphocluster, only part of the images needs to be clustered in each iteration. Shrunken k-Means utilizes distances to cluster centers provided k-means to discard images that are far from the centers. Partially Labeled k-Means utilizes the label information from the earlier iterations to guide the clustering.\nAutoencoders have also been proposed for learning plankton image features for clustering without the label information. The basic idea is to utilize encoder-decoder network architecture where the encoder generates an embedding vector from an image and the decoder tries to reconstruct the original image based on the embedding vector. Such a network can be trained without any labels. Ideally, the encoder learns to compress the essential information from the image into an embedding vector that can then be used for clustering. For example, Salvesen et al. ( 2020) applied an autoencoder-based approach called Deep Convolutional Embedded Clustering (DCEC) plankton image data. The method employs the CNN-based autoencoder architecture by Guo et al. (2017) and uses k-means to cluster the obtained embeddings. Alfano et al. (2022) proposed a plankton image clustering technique based on variational autoencoders (VAEs). The method utilizes a pre-trained DenseNet without fine-tuning to extract features. Obtained deep image features are then fed to VAE to generate latent space representations. Finally, low-dimensional embeddings are clustered using fuzzy k-means.\nClustering methods are only able to produce unlabeled clusters of images with a similar appearance. Therefore, further analysis is needed to confirm and label the clusters. Schröder et al. (2020) addressed this by introducing an interactive tool where the users revise the obtained clusters, manually correct the hierarchy and annotate the final set of clusters. This semiautomatic approach reduces the manual work needed for data labeling as the expert does not need to annotate every image separately. Goulart et al. (2021) utilized t-distributed stochastic neighbor embedding (t-SNE) to visualize the clusters in two-dimensional space allowing the human expert to quickly see the clusters in the data. Pastore et al. (2020) proposed a full pipeline for environmental monitoring based on plankton image clustering and minimal expert supervision (the expert labels only one image per cluster). CNN was used for image feature extraction and various unsupervised clustering algorithms including K-means, fuzzy K-means, and Gaussian mixture model were compared.\nAs a summary, the most common approaches to tackle the problem of limited amount of labeled plankton image data are data augmentation and transfer learning. Data augmentation is an essential part of practically all modern plankton recognition pipelines based on deep learning, while transfer learning allows to utilize knowledge from another domain to compensate the lack of training data. In the case of extreme scarcity of labeled training data, further modifications to the model training are needed. Typically this means the adoption of regularization techniques that prevent the model to overfit to the training data. Weight imprinting, metric learning, and central loss have been found useful tools in few-shot plankton recognition. If labeled training data is completely missing, clustering or active learning can be utilized. Clustering allows to analyze plankton image datasets in an unsupervised manner, while active learning makes it possible to minimize the amount of expert labeling effort for building a plankton recognition model for future data."
},
{
"figure_ref": [],
"heading": "Challenge 2: High class imbalance",
"publication_ref": [
"b253",
"b234",
"b251",
"b152",
"b115",
"b395",
"b151",
"b272",
"b124",
"b200",
"b167",
"b134",
"b183",
"b384",
"b409",
"b229",
"b100",
"b324",
"b374",
"b366",
"b253",
"b385",
"b199",
"b227"
],
"table_ref": [],
"text": "High class imbalance is naturally inherent in many real-world applications and plankton recognition is not an exception. Certain plankton species are considerably more common than others causing the data in typical plankton datasets to be highly imbalanced. This is problematic when it comes to training plankton classification methods. One of the most notable problems connected to the high class imbalance is the catastrophic forgetting where neural network, while learning new information, completely forgets previously learned information. This typically affects the minority classes that are only rarely seen during the training stage causing the network to only learn the necessary image features for the majority classes.\nUndersampling is a technique to decrease the level of imbalance by discarding images from the majority classes. In the simplest case, undersampling can be done by randomly selecting a subset of images from majority classes is such way that the resulting training dataset has an equal amount of images in all classes. For example, Lee et al. (2016) reduced the class bias on small-sized plankton classes by randomly sampling images from the classes with more samples than the predefined threshold. Kloster et al. (2020) utilized a similar undersampling technique. Also, more intelligent solutions for undersampling have been suggested in plankton literature. Le et al. (2022) utilized undersampling by filtering combined with cost-sensitive learning to obtain a more balanced dataset for training. Ding et al. (2018) proposed an EasyEnsemble.D algorithm for plankton recognition on highly imbalanced datasets. The basic idea is to sample multiple subsets from majority classes to fully utilize the large data volumes. Each subset is used to train a separate weak classifier with different weights, and the final classification is performed using the ensemble of the weak classifiers. The problem with undersampling is that it reduces the amount of training data which in the case of plankton recognition is typically already limited. Especially, in the presence of rare species, the undersampling alone leads to an extremely small training set.\nOversampling is another technique to reduce the level of imbalance with duplicating samples from the minority classes. The oversampling is typically done using data augmentation, i.e. instead of using identical duplicates, manipulated versions are created to obtain more training data for minority classes. For example, Bochinski et al. (2018), increased the amount of training samples of the smaller classes by mirroring the images horizontally and vertically to counter the imbalance during training. Xiaoyan (2020) proposed a combination of undersampling and oversampling to address the class imbalance in plankton recognition. This is done by utilizing KA-Ensemble algorithm (Ding et al., 2020) that combines oversampling of the minority class via kernel-based adaptive synthetic sampling (Kernel-ADASYN) and random undersampling of the majority class. The experiments showed increased classification accuracy for the minority class. Liu et al. (2021) proposed to combine borderline-SMOTE oversampling with Fuzzy C-means clustering-based undersampling for plankton image data. The Synthetic Minority Oversampling TEchnique (SMOTE) (Chawla et al., 2002) synthesizes new samples between the minority class and its nearest neighbor in the feature space. Borderline-SMOTE (Han et al., 2005) improves the method by concentrating on the samples near the class boundaries in order to oversample more significant samples for the minority classes. Fuzzy C-means clustering is utilized to preserve the clusters found in the original data during undersampling.\nAnother approach among a variety of resampling methods is cost-sensitive learning (Elkan, 2001). The method defines a so-called cost-matrix which specifies a reward or a penalty over the classifications of an algorithm. A core idea behind it is similar to resampling but it does not change the prevalence of the training set directly. However, a performance evaluation for an imbalanced plankton set reported by Corrêa et al. (2016) demonstrates only minor improvements for cost-matrix in comparison to SMOTE and resampling.\nAnother solution to artificially create more image data for training and to reduce the level of imbalance is to utilize generative models capable of generating realistic images with a certain distribution. GANs (Goodfellow et al., 2014) are deep learning models that can be used to generate photo-realistic artificial images with the same statistics as the data they were trained with. This is done by using two models, a generative model and a discriminative model. The generative model generates candidate images usually from random noise. The discriminative model is an image classifier that is given labeled samples from the real set of images and fake images produced by the generative model. The task of the discriminative model is to distinguish real images from fake ones and the task of the generative model is to fool the discriminative model. These two models are trained simultaneously in such a way that the generative model becomes increasingly better at producing realistic fake images and the discriminative model gets increasingly better at recognizing them. GANs have been shown to be able to generate images that are authentic to human observers.\nGANs have been utilized also for reducing bias caused by the class imbalance in plankton recognition. Wang et al. (2017) used GAN to generate new example images of minority classes. Furthermore, a method was proposed where the CNN-based plankton recognition model shares the weights with the discriminative model. However, only minor improvement was observed over the baseline recognition models trained on the original data without GAN-based data augmentation. Liu et al. (2018b) proposed a GAN-based curriculum learning strategy. The proposed method contains two stages. First, the model is trained using the original data and then with more complex data consisting of GAN-generated images. Li et al. (2021b) utilized Cy-cleGan (Zhu et al., 2017) for the augmentation of rare taxa, and Khan et al. (2022); Ali et al. (2022) applied DC-GAN (Radford et al., 2015) to augment an algae image dataset. Vallez et al. (2022) compared data augmentation by combining two diatom images from the same class using morphing and image registration methods performing diffeomorphic transformations to generation of synthetic images by a GAN. In this study, mixing images using morphing achieved better results. The fundamental problem of using GANs for image augmentation is that the generated images have the same statistics as the images they were trained with. Therefore, if the GANs are trained using the same data as the recognition model, and the recognition model is able to learn the data distribution from the original data, the generated samples do not necessarily provide additional value for the training. However, some promising results have been obtained on GAN-based augmentation of highly imbalanced datasets (Tanaka and Aranha, 2019).\nSimilarly to the challenge of a limited amount of training data, transfer learning has also been proposed to overcome the class imbalance problem. In a method proposed by (Lee et al., 2016), a balanced dataset is first generated using randomized undersampling, the model is pre-trained on the balanced dataset, and finally fine-tuned using the whole unbalanced plankton image dataset. Wang et al. (2018) introduced a transfer parallel model approach for plankton recognition. The main idea is to avoid the catastrophic forgetting by training two submodels: 1) a model trained on the whole dataset, and 2) a pre-trained model trained only on small classes. Deep features from both of the models are concatenated before the softmax layer. The latter submodel adds good image features for minor class classification that the network could otherwise fail to learn. Also, modified model architectures have been proposed to address the class imbalance. These include models with increased generalization ability to minority classes. Liu et al. (2018a) applied Deep Pyramidal Residual Network (PyramidNet) (Han et al., 2017) to plankton recognition and shown to improve accuracy on a highly imbalanced dataset. The idea behind Pyra-midNet is to gradually increase the size of the feature map. This combined with the ResNet style to skip connections causes reduced change of overfitting, and therefore, better generalization ability. Kerr et al. (2020) proposed model fusion to address the class imbalance. The results suggest that combining multiple individually trained CNNs with a common softmax layer improves the accuracy of rare species, consequently providing better overall accuracy on imbalanced data.\nAs a summary, undersampling and oversampling are the simplest and most widely used approaches to address high class imbalance in plankton image data. Oversampling is typically performed using traditional data augmentation, but also generative approaches such as GANs have been proposed to generate completely new plankton images for the minority classes. Moreover, transfer learning, model fusion, and regularization techniques preventing overfitting have been shown to improve plankton recognition accuracy in the case of highly imbalanced training data."
},
{
"figure_ref": [],
"heading": "Challenge 3: Fine-grained nature of the recognition task",
"publication_ref": [
"b116",
"b173",
"b360",
"b349",
"b360",
"b120",
"b260",
"b155",
"b377"
],
"table_ref": [],
"text": "In order to obtain high recognition accuracy on classes with high interclass similarity such as taxonomically close plankton species, techniques that focus attention on subtle visual differences are needed. The task of recognizing hard-to-distinguish classes from each other is called fine-grained classification. Plankton recognition in most cases can be considered a finegrained classification task as the fundamental way to improve the overall accuracy of a recognition model is to make it better at recognizing the challenging cases. Despite this, most of the work on plankton recognition does not tackle the challenge directly but instead focuses on comparing different general model architectures on the task. Related to this viewpoint, it has been also studied whether the recognition should be considered as a flat or hierarchical classification task. Boddy et al. (2000) considered misclassifications of phytoplankton as a result from the overlap of feature distributions and grouping of similar species within genera or based on groupings indicated in dendrograms was proposed. Similarly, Fernandes et al. (2009) proposed an approach for balancing the trade-off between the classification performance and number of classes. The model automatically suggests merging of classes based on the statistics evaluated after the classification. The results from taxa recognition of macroinvertebrates by Ärje et al. (2020) showed that humans performed better when a hierarchical classification approach commonly used by human taxonomic experts was used, but when a flat classification approach was used, the CNN was close to human accuracy. To improve the automatic approaches, a few methods focusing especially on the attention mechanism to address the fine-grained nature of the recognition task have been proposed. Sun et al. (2020) considered fine-grained classification of plankton by proposing an attention mechanism based on Gradient-weighted Class Activation Maps (Grad-CAM) (Selvaraju et al., 2017) to force the CNN to focus on the most informative regions in the image. Grad-CAM was originally developed for visualizing the CNN-based models. It highlights important image regions which correspond to the decision of interest (in this case plankton recognition). Sun et al. (2020) utilized Grad-CAM to detect the regions to focus on, and a feature fusion approach utilizing high-order integration (Cai et al., 2017) is applied to obtain stronger features for those regions. This approach shares similarities with the self-attention module used in the TANet architecture (Li et al., 2019c) for plankton recognition. However, the selfattention module puts larger weights on the important regions, i.e. those regions in the feature map with high activation values.\nAlso other approaches for fine-grained plankton recognition have been proposed. Du et al. (2020) applied Matrix Power Normalized CO-Variance (MPN-COV) pooling layer for second-order feature extraction. The aim is to model the complex class boundaries more accurately than in traditional pooling (e.g. softmax). There is some evidence (Li et al., 2017) that suggests that higher-order information can improve recognition accuracy in fine-grained tasks. Venkataramanan et al. (2021) proposed an improved pipeline tackling inter-class similarity and intra-class variance. The authors suggested alleviating inter-class variance with a metric learning-based approach utilizing triplet loss and mitigating intra-class variance by X-means clustering technique applied to the extracted features. The idea is to cluster the classes with high inter-class variance into multiple clusters and consider these as separate classes. The authors propose a method to find the optimal amount of clusters that minimize both the intra-class variance and inter-class similarity, and this way improve the accuracy of fine-grained plankton recognition.\nIn general, only few papers directly tackling the fine-grained nature of the plankton recognition task exist. These are based on attention mechanisms to find the most important regions in the images allowing the recognition model to focus on the subtle differences between the classes, and contrastive or metric learning that allow explicitly learning the image features that separate the pairs of classes."
},
{
"figure_ref": [],
"heading": "Challenge 4: Domain shift between datasets",
"publication_ref": [
"b192",
"b108",
"b408",
"b335",
"b303",
"b280",
"b115",
"b196"
],
"table_ref": [],
"text": "Different imaging instruments cause domain shift between plankton datasets. Domain shift in a wider sense refers to a situation where the distribution of the dataset that is used for training differs from the data where the recognition model is applied. CNN-based models tend to learn image features that are very specific to the distribution of the training data making them notoriously weak at generalizing beyond the domain they were trained on (Gulrajani and Lopez-Paz, 2020). This is why most automatic plankton recognition solutions focus on just one imaging instrument. This, however, limits the wider utilization of the methods. Tuning the classification model trained on one dataset to work on another dataset (correcting domain shift between the datasets) is called domain adaptation (Ben-David et al., 2010) and learning a general model that can be applied to any dataset (domain) is called domain generalization (Zhou et al., 2022).\nWhile domain adaptation and generalization have not been widely studied on plankton recognition, there have been works where multiple different plankton image datasets have been utilized to solve the recognition task. Transfer learning and fine-tuning have been utilized as approaches against the differences in datasets. Rodrigues et al. (2018) applied transfer learning using CNNs to obtain a feature extractor that can be used for new datasets. The Kaggle-Plankton dataset was used to train a CNN (source dataset) and an in-house dataset was used as a target dataset to test the suitability of the features. Orenstein and Beijbom (2017) applied a variety of learning schemes to three very different plankton image datasets. The bigger labeled image datasets, IFCB and ISIIS, were used to train CNNs both by fine-tuning and from scratch. Then, the classifiers were used to classify within-domain images directly and as feature extractors for out-of-domain data.\nLumini and Nanni (2019a); Lumini et al. (2020) studied ensembles of different CNN models, fine-tuned on several datasets, with the aim of exploiting their diversity in designing an ensemble of a classifier. The experimental results show that the combination of several CNNs in an ensemble grants a performance improvement compared with a single CNN model.\nIn Bochinski et al. (2018), two datasets from different biological environments were captured and analyzed. The first dataset was used to analyze the achievable accuracy of the CNN and how the Cost-Effective Active Learning (CEAL) can be used to minimize the number of required annotations. The second dataset was used to examine the generalization ability of the CNN and if the CEAL method can be used to fine-tune the system to adapt to the characteristics of this new data. Plonus et al. (2021a) suggest using capsule neural networks combined with probability filters to address the dataset shift caused by different plankton imaging instruments. The idea of Capsule neural networks is to form groups of neurons (capsules) that learn the specific properties of the object (e.g. plankton) in the image. The authors argue that the capsule neural networks are less sensitive to the changes in the field conditions and therefore able to adapt to different data distributions. Guo et al. (2022) proposed a crossdomain few-shot learning model for instrument-agnostic plankton recognition. Similarly to transfer learning, the model is first trained on the source domain with a large amount of training data and then adapted to the target domain using fine-tuning. In addition, graph neural network-based metalearning is applied to learn a feature distance metric capable of recognizing plankton species in the target dataset with a very limited amount of labeled data.\nDomain shifts between the plankton image datasets or imaging instruments have not been widely studied. Most works focus on fine-tuning the recognition models trained on one dataset to new datasets using transfer learning. While the transfer learning reduces the amount of manual labeling needed for new datasets, it does not fully solve the problem of multiple domains. Labeled training data are still needed for all datasets, and the recognition models need to be fine-tuned for each, requiring expertise in machine learning and computing resources. A more general model can be obtained by using ensemble learning with submodels learned on different datasets if training data on each dataset (imaging instrument) is available. More sophisticated approaches to plankton image domain adaptation include the capsule neural networks and meta-learning."
},
{
"figure_ref": [],
"heading": "Challenge 5: Previously unseen classes and unknown particles",
"publication_ref": [
"b310",
"b376",
"b372",
"b131",
"b318",
"b382",
"b399",
"b300",
"b158",
"b369",
"b103",
"b214",
"b298",
"b149",
"b369",
"b103",
"b149",
"b323",
"b109"
],
"table_ref": [],
"text": "Automated plankton imaging instruments capture images of unknown particles and the class (plankton species) composition varies between geographical regions and ecosystems. CNN-based models are known to struggle in open-set settings where the class composition of training data differs from the data for which the trained model is applied. Typical CNN-based classification models tend to classify the images from a new class to one of the known classes often with high confidence, and to include new classes to the models, they need to be retrained. These are major problems for plankton recognition as the plankton species vary between different regions and seasons. Retraining a separate model for each dataset is not feasible. Therefore, there is a need for a recognition model that 1) is able to predict when the image contains a previously unknown plankton species (open-set recognition) and 2) can be generalized to new classes without retraining the whole model.\nIn the case of plankton recognition, the open-set problem is often formulated as an anomaly detection problem where the model is trained to both correctly classify the known classes and to filter abnormal classes by training the model to produce high and low entropy distributions for the normal classes and abnormal classes respectively. Pastore et al. (2020) proposed a semi-automatic method to handle the previously unseen plankton classes by utilizing anomaly detection combined with expert verification. Both oneclass SVM and a new neural network-based method called Delta-Enhanced Class (DEC) detector were considered. The DEC detector utilizes absolute differences between the feature vectors of an input image and random images from a known class as additional input to predict whether the input image is from the known class or anomaly. Varma et al. (2020) proposed L 1 -norm tensor-conformity curation to remove outliers (non-plankton or misclassified images) from the training data. The idea is to measure the conformity of the images using L 1 -norm subspaces (Tountas et al., 2019). Conradt et al. (2022) brought up the high intra-class and low inter-class variation of plankton morphology, and spatio-temporal changes in the plankton community as the main causes for the need to frequently validate the results from automatic recognition. The proposed remedy is a dynamic optimization cycle in which the model is updated based on manual-validation results. Pu et al. (2021) proposed a loss function that contains three loss terms to detect the anomalies and to maintain the classification accuracy for the images belonging to the normal classes by incorporating the expected crossentropy loss, the expected Kullback-Leibler (KL) divergence, and the Anchor loss. The model was tested on classes of plankton images containing also bubbles or random suspending particles. Walker and Orenstein (2021) utilized a large background set of images that do not belong to the target classes (classes to be recognized) and hard negative mining to find images that are more likely to cause false negatives. The training set was then complemented with these challenging images to improve the classifier's ability to recognize when the images are from novel classes. While promising results were obtained on open-set plankton recognition the method requires that a labeled background set is available which limits the usability of the method.\nAnother approach to tackle the open-set problem is to utilize similarity metric learning. The aim of metric learning is to obtain image embedding vectors that model the similarity between images. It is commonly utilized in person (Ye et al., 2021) and animal re-identification (Nepovinnykh et al., 2020), as well as content-based image retrieval (Dubey, 2021), but has been also successfully applied to plankton classification (Teigen et al., 2020;Badreldeen Bdawy Mohamed et al., 2022). A simple approach to implement a recognition method is to construct a gallery set of known species and use the learned similarity metric to compare query images to the gallery images. The similarity in this context corresponds to the likelihood that the images belong to the same class. This further allows defining a threshold value for similarity enabling open-set classification: if no similar images are found in the gallery set, the query image is predicted to belong to an unknown class. Furthermore, new classes can be added by simply including them in the gallery set as the model does not necessarily need to learn class-specific image features.\nThe most common approaches for deep metric learning include tripletbased learning and classification-based metric learning. The first approach learns the metric by sampling image triplets with anchor, positive, and negative examples (Hoffer and Ailon, 2015). The loss function is defined in such a way that the distances (similarity) from the embeddings of the anchors to the positive samples are minimized, and the distances from the anchors to the negative samples are maximized. The second approach approximates the classes using learned proxies (Movshovitz-Attias et al., 2017) or class centers (Deng et al., 2019) that provide the global information needed to learn the metric. This makes it possible to formulate the loss function based on the softmax loss and allows to avoid the challenging triplet mining step. Teigen et al. (2020) studied the viability of few-shot learners in correctly classifying plankton images. A Siamese network was trained using the triplet loss and used to determine the class of a query image. Two scenarios were tested: the multi-class classification and the novel class detection. A model trained to distinguish between five classes of plankton using five reference images from each class was able to achieve reasonable accuracy. In the novel class detection, however, the model was able to filter out only 57 images out of 500 unknowns. Badreldeen Bdawy Mohamed et al. (2022) utilized the angular margin loss (ArcFace) (Deng et al., 2019) instead of triplet loss to address the high cost of the triplet mining step. Furthermore, Generalised Mean pooling (GeM) (Radenović et al., 2018) was applied to aggregate the deep activations to rotation and translation invariant representations. ArcFace uses a similarity learning mechanism that allows distance metric learning to be solved in the classification task by introducing the Angular Margin Loss. This allows straightforward training of the model and only adds negligible computational complexity. The metric learning-based method was shown to outperform the model utilizing OpenMax (Bendale and Boult, 2016) layer in open-set classification of plankton. One of the main benefits of the method is that it generalizes well to new classes added to the gallery set without retraining. This makes it straightforward to apply the model to new datasets with only partly overlapping plankton species composition.\nPlankton species vary in different locations and seasons, thus, it is common that a recognition model should be adapted to or retrained for the new situation at some point. Retraining a separate model for each situation is infeasible, and continual or online training of the model would be challeng-ing for online monitoring applications. Therefore, an effective remedy would be to treat it as an open-set recognition problem, solve it with the modern methods anomaly detection or metric learning, and take care of the model's capability to generalize to new data without the need to retrain the whole model."
},
{
"figure_ref": [],
"heading": "Challenge 6: Label uncertainty",
"publication_ref": [
"b141",
"b140",
"b356",
"b172",
"b281",
"b281",
"b358",
"b106",
"b181",
"b182"
],
"table_ref": [],
"text": "The plankton image label uncertainty is caused by the difficulty of manually recognizing the species from low-quality images with limited resolution, human error, and high costs preventing the repetition of the manual annotation by multiple experts. Culverhouse et al. (2003) identified four main reasons for the incorrect labeling of plankton images: 1) the limited shortterm memory of humans, 2) fatigue, 3) recency effects, i.e., labeling is biased towards the most recently seen labels, and 4) positivity bias, i.e., labeling is biased by the expert's expectations to the content of sample. Labels provided by sixteen human experts (marine ecologists and harmful algal bloom monitoring specialists) on microscopy images of dinoflagellates (6 classes) were analyzed. The results showed that only 67 to 83% self-consistency and 43% consensus between experts was obtained. Experts who where routinely labeling the selected classes were able to achieve 84 to 95% labeling accuracy. Culverhouse (2007) brought up several important points related to labeling algae. The presented performance figures do not represent the state-of-theart of automatic approaches, but improvements would be beneficial for both alternatives. Human expert judgements would benefit from peer review and inter-expert calibration to remove human bias. To improve the automatic solutions, the errors of both man and machine would require further attention. Global reference databases with validated samples and representative coverage of the morphological and physiological characteristics in nature would be beneficial for training and evaluation purposes. In addition, Solow et al. (2001) noted that the taxonomic counts of classified individuals are biased when there are errors in classification. A straightforward method for correcting for the bias was proposed based on the classification probabilities of the classifier.\nImage filtering has been proposed to address label uncertainty in plankton image data. The idea is to discard images for which the recognition model is uncertain, and therefore, more likely to produce erroneous labels. For example, Faillettaz et al. (2016) utilized a probabilistic RF for classification, and obtained class probabilities were used to detect and ignore images for which the classifier is uncertain. Luo et al. (2018), Plonus et al. (2021a), and Kraft et al. (2022b) utilized similar approach for CNN-based recognition models. Luo et al. (2018) used a separate fully annotated validation set to set class-specific probability thresholds for filtering. Plonus et al. (2021a) proposed a pipeline for tailoring filtering thresholds to the research question of interest by allowing to select between high precision and high recall. Kraft et al. (2022b) evaluated a CNN-based model with class-specific probability thresholds on operational use.\nRelated to the label uncertainty, quantification methods have been proposed for plankton image data analysis. The basic idea is to estimate the class distribution directly. While mislabeled samples cause noise to the training data for classification methods, the class distributions are often close to correct. Sosik and Olson (2007) used a quantification method to estimate the abundance of different taxonomic groups of phytoplankton. Utilizing a combination of image feature types including size, shape, symmetry, and texture characteristics, plus orientation invariant moments, diffraction pattern sampling, and co-occurrence matrix statistics proposed. Statistical analysis was used to estimate category-specific misclassification probabilities for accurate abundance estimates and for quantification of uncertainties in abundance estimates. Beijbom et al. (2015) analyzed several quantification methods on a time-series dataset of plankton samples. These included unsupervised and supervised quantification. In unsupervised quantification, the dataset shift is assumed to be a pure class-distribution shift. Alternatively, the dataset shift is assumed to be 'small' and the unlabeled set of target samples is used to align the internal feature representation of a machine learning algorithm. In supervised quantification, no explicit assumptions are made on the dataset shift, but it is assumed that a small amount of samples are available in the target domain. González et al. (2017) proposed a methodology to assess the efficacy of learned models, which takes into account the fact that the data distribution (the plankton composition of the sample) might vary between the training phase and the testing phase. Their approach used validationby-sample. They proposed using the sample as the basic unit instead of the individuals to predict the abundance of the different plankton groups. Thus, model assessment processes require groups of samples with sufficient variability to provide precise error estimates. González et al. (2019) used a transfer learning approach where deep image features as input for the quantification algorithm to estimate the distribution of each class in an unknown water sample was proposed. Orenstein et al. (2020a) proposed a semi-automatic pipeline where a small subset of images were manually labeled to estimate the dataset shift and use this information to correct the quantification estimate.\nSupervised machine learning and particularly the performance evaluation of a recognition model relies on the correctness of the class labels. However, visual recognition of a number of plankton species from low-quality images is difficult and using expert panels becomes practically infeasible if the aim is to produce large datasets. The proposed remedies include exclusion of images that have high label uncertainty or focusing on the actual quantity of interest if it is not plankton recognition. Alternative ways to solve this challenge would be to focus on few-shot learning with ground truth validated by an expert panel and pay special attention to model generalisability, or to use generative models."
},
{
"figure_ref": [
"fig_9"
],
"heading": "Challenge 7: Large image size variation",
"publication_ref": [
"b64",
"b133",
"b182",
"b165",
"b224",
"b212",
"b345",
"b169",
"b280",
"b169",
"b375",
"b319",
"b119",
"b208",
"b169"
],
"table_ref": [],
"text": "Most plankton datasets have extreme variation in image size. Fig. 7 shows example images obtained using Imaging FlowCytobot (IFCB). Typical CNNbased image classifiers require the input image to have a predefined size. Therefore, image resizing has been used as a pre-processing step for datasets with varying height and width of images (e.g. Dai et al., 2016b;Kuang, 2015). On a general level, the resizing can be done in two ways: by forgoing aspect ratio (e.g. Al-Barazanchi et al., 2015a;Sánchez et al., 2019b) or by maintaining the aspect ratio (e.g. Dai et al., 2016a;Correa et al., 2017;González et al., 2019). In the first approach, stretching is needed for images whose aspect ratio does not match the target aspect ratio. This will change the shape of the objects in the image which may affect the feature extraction or learning. In the second approach, images are typically resized based on the length of their longest side and padded with a single color to make the image size correct. Eerola et al. (2020) evaluated various ways to implement the padding and padding with the mode of the image (the most common color in the image typically corresponding to the background color) produced the best results on IFCB data. Both approaches (forgoing and maintaining aspect ratio) have been utilized in plankton recognition. However, there exists little comparison between them. Dai et al. (2016a) tested various resizing methods were tested on zooplankton images and the best accuracy was obtained by maintaining the aspect ratio while scaling. On the other hand, Jindal and Mundra (2015) found little to no difference on performance between the approaches despite images appearing distorted when forgoing aspect ratio.\nVarious other ways to obtain fixed-size images have been proposed. In the method proposed by Ho et al. (2018), a fixed input image size was chosen and the images were either cropped or padded with zeros to adjust them to the correct size. Schröder et al. (2018) proposed to crop the images to their tight bounding box and pad to a square with a minimum edge length of 128 pixels. Images larger than 128 pixels were shrunken to the same size. Ellen et al. (2019) resized images larger than the target size thus losing some detail. Images smaller than the target size were resized by padding and, therefore, the object size remained the same. Lumini and Nanni (2019a,b); Lumini et al. (2020) compared the two different strategies: 1) resizing all images to a common size and 2) resizing only images that were larger than the input size and using padding for the smaller images. The results showed that the first method produced a better classification result in most of the datasets and models.\nAll methods to produce fixed-size images from original plankton images with a large size variation result in some degree of information loss or image distortions. Information on the size of the plankton is lost during the resizing, small details disappear if images are heavily downscaled, and only part of the object is seen if cropping is used. Ellen et al. (2019) partially solved this problem by providing the size information as metadata (additional features) for the classifier while still using resized versions of images as the main input for the CNN. Metadata is used as an input for the network besides image data, and they are processed independently by separate parts of the network. The outputs of both subnetworks are concatenated together and processed by fully connected layers. Results showed that metadata was useful for classification accuracy.\nTo truly solve the problem with the varying image size and aspect ratio, the CNN architecture needs to be modified so that it can process images with multiple sizes. This can be achieved, e.g. by combining scale-invariant and scale-variant features to devise a multi-scale CNN architecture (Van Noord and Postma, 2017). Py et al. (2016) proposed an inception module that allows to use multiple scaled versions of the original image with different sizes as the input for CNN. By selecting different strides for each scale, the computed feature maps have the same size for all scales and can be concatenated to a single set of multi-scale features. The proposed method was shown to outperform the method with a single fixed-size input. Bureš et al. (2021) compared various modifications of the baseline CNN on plankton recognition with high variation in image size. These include Spatial Pyramid Pooling (SPP) (He et al., 2015), using image size as metadata, patch cropping and multi-stream CNNs. SPP allows the training of a single CNN with multiple image sizes in order to obtain higher scale invariance by pooling the features produced by the convolutional layer to a fixed-length vector required by the fully connected layers. The metadata was used as described by Ellen et al. (2019). The patch cropping technique divides images into fixed-size patches that are classified separately. The final recognition is done by averaging the resulting score vectors. Multi-stream CNN utilizes a similar approach but uses multiple different networks trained for different image sizes and aspect ratios. The best plankton recognition accuracy was obtained using a multi-stream network combining two models with different input aspect ratios and patch cropping.\nMost plankton datasets have significant variation in image sizes and aspect ratios. Common CNN-based image classifiers require that the input images have a constant size. In this case, image resizing is used and it is necessary to consider what to do with the aspect ratio and whether metadata about the image size provides an advantage when complementing the fixedsize images. However, a more general remedy would be to use a multi-scale CNN with an appropriate architecture as the recognition model."
},
{
"figure_ref": [],
"heading": "Challenge 8: Low or varying image quality",
"publication_ref": [
"b325",
"b226",
"b126",
"b325",
"b130",
"b226",
"b307",
"b126",
"b289",
"b265",
"b292"
],
"table_ref": [],
"text": "To improve the classification accuracy on low-quality images various preprocessing steps have been proposed. These include discarding bad quality images (Raitoharju et al., 2016), image segmentation (Keçeli et al., 2017), and denoising (Cheng et al., 2019).\nLow quality images can be discarded in different ways. Raitoharju et al. (2016) manually removed low-quality images from the dataset before training the recognition model. Moreover, the remaining images were cropped to remove artifacts mainly appearing close to image borders. Coltelli et al. (2014) filtered out out-of-focus images before the feature extraction. The out-of-focus detection was done by fitting color histograms in a GMM. If the distribution contained two components (background and plankton), the image was considered to be in-focus.\nSome studies suggest segmenting the images as a preprocessing step to discard non-plankton pixels from the images. For example, Keçeli et al. (2017) used Otsu's thresholding method (Otsu, 1975) for segmentation and pixels outside the obtained segmentation map are set to zero. Cheng et al. (2019) applied texture enhancement together with background suppression before the classification step. Enhanced images were shown to produce a slightly higher recognition accuracy than the images without enhancement. Ma et al. (2021) proposed to use modern CNNbased super-resolution techniques to improve the plankton image quality. The EDRN super-resolution architecture (Lim et al., 2017) was combined with the contextual loss (Mechrez et al., 2018), and was shown to produce high-quality images. However, the effect on plankton recognition accuracy was not assessed.\nMany real-world computer vision applications have to deal with lowquality images and plankton recognition is no exception. A wealth of image preprocessing approaches exist and in the case of plankton images, at least exclusion of bad images, denoising and image segmentation have been proposed. A more profound way would be to adopt image reconstruction methods, but from the practical perspective of plankton recognition, the simpler methods can be considered as sufficient and data augmentation is commonly used to introduce additional variation to the data."
},
{
"figure_ref": [],
"heading": "Challenge 9: Massive amount of data",
"publication_ref": [
"b260",
"b410",
"b402",
"b402"
],
"table_ref": [],
"text": "Massive data volumes obtained by modern imaging instruments motivate to develop computationally efficient solutions that are able to analyse data in real time. However, the computation time is rarely considered in plankton recognition literature. Most works related to the challenge consider lightweight CNN architectures. For example, shallow TANet (Li et al., 2019c) was shown to outperform competing methods in computing time without sacrificing accuracy on the Kaggle dataset. Zimmerman et al. (2020) proposed an embedded system for in situ deployment of plankton microscope with real-time recognition system. Due to the limited computation resources and computation time limitations, CNNbased recognition methods were considered unsuitable and a faster featureengineering based approach was proposed with reduced recognition accuracy.\nThe computation time is an especially big issue with holographic imaging that traditionally relies on computationally heavy reconstruction operations to process the raw data. To address this end-to-end CNN methods for plankton recognition that take the raw holographic data as input have been proposed (Guo et al., 2021a;Zhang et al., 2021). This way the reconstruction step can be completely avoided. Guo et al. (2021a) and Zhang et al. (2021) showed that CNNs are able to learn the image features for the plankton recognition from the raw data speeding up the processing significantly.\nOnline monitoring of plankton with modern imaging equipment produces huge amounts of images. The related image analysis requires either highperformance computing (HPC) resources in the cloud or local (edge) computing with shallow CNN architectures. In most cases, the recognition model training has to be performed in a HPC environment after which at least the lightweight models can be deployed for local execution."
},
{
"figure_ref": [],
"heading": "Summary and future directions",
"publication_ref": [
"b178",
"b393"
],
"table_ref": [],
"text": "In this paper, a comprehensive survey of challenges and existing solutions for automatic plankton recognition was provided. We identified nine challenges that complicate the introduction of automatic plankton recognition methods to operational use: 1) the limited amount of training data for less common species, 2) large class imbalance, 3) fine-grained nature of the recognition task, 4) domain shift between imaging instruments, 5) presence of previously unseen classes and unknown particles, 6) uncertainty in expert labels, 7) large variation in image size, 8) low or varying image quality, and 9) massive data volumes. While most of the considered challenges are common in a wide variety of machine learning applications, plankton recognition has its specific characteristics including highly imbalanced image datasets, extreme variation in image size, limitations in image quality, and a shortage of qualified experts to visually annotate the images.\nFig. 8 shows a flowchart summarizing the challenges and approaches to solve them. Given a new plankton image dataset, the flowchart provides a simple pipeline to identify the problems related to the dataset as a series of yes-no questions. Furthermore, references to the sections in this paper providing the detailed descriptions are provided to find the existing techniques One notable problem in plankton recognition is the lack of publicly available general-purpose plankton image datasets with an evaluation protocol making it possible to compare different plankton recognition methods in a fair and reliable manner. The vast majority of the research either has focused on private in-house datasets or is based on custom evaluation protocol and dataset splits on publicly available datasets. This makes it impossible to compare the accuracies between different studies making it challenging to select the best practices for future research. This slows down the progress in the plankton recognition method development. Therefore, there is a need for a publicly available plankton dataset with a predetermined evaluation protocol and preferably multiple subsets captured with different imaging to allow quantitative evaluation of the advances in general (device-agnostic) plankton recognition.\nAnother important problem limiting the wider utilization of automatic plankton recognition is the difficulty of collecting training images to exhaust all the possible classes. It is not realistic to construct a labeled training set consisting of all the plankton species and non-plankton particles that the imaging instrument is capable of capturing in a certain location. Moreover, varying plankton species composition between different geographical regions and ecosystems limits the possibility to apply working recognition models to new locations and datasets. Even a classification model developed and trained for one imaging instrument and one geographic location struggles if new species appear, for example, due to seasonal changes. A more realistic scenario is to aim for open-set plankton recognition models that are able to identify when the images belong to previously unseen classes and either reject them or process them further by, for example, clustering. Open-set recognition is an active research topic in machine learning (see, for example, Geng et al. (2020)).\nLarge variation between plankton image datasets with different species compositions and imaging instruments can be considered not only a challenge but also an opportunity. While it is very difficult to develop one general-purpose algorithm for imaging instrument-agnostic plankton recognition, modern domain adaptation methods have the potential to enable the joint utilization of different datasets. This would allow adapting the classification model to new datasets with a reasonable amount of manual work. Domain adaptation has already been successfully applied to various other machine learning applications, such as general object recognition (Wilson and Cook, 2020). Domain adaptation can be considered a special case of transfer learning that mimics the human vision system and utilizes a model trained in one or more source domains to a different (but related) target domain. Domain adaptation can be utilized to reduce the effect of a large domain shift between different datasets and the lack of labeled training data.\nThe relatively large pool of plankton image datasets motivates to further utilize domain generalization and meta-learning to obtain an imaging instrument agnostic recognition model. In meta-learning, multiple datasets and tasks are used to \"learn how to learn\" the recognition model. The idea is to automate the creation of the entire machine learning pipeline end-toend including the search for the model architecture, hyperparameters, and learning the model weights. Domain generalization refers to learning domainindependent (in this case imaging instrument-independent) feature representations that can be then applied to any dataset. Domain generalization has a wide variety of different applications and it has become an increasingly studied problem in machine learning (see the recent survey in Wang et al. (2022a)). Recent progress in such methods has opened novel possibilities to aim towards a universal plankton recognition system that is able to adapt to different environments, with dramatically different plankton populations and varying imaging instruments, promoting the wider utilization of automatic plankton recognition for aquatic research. "
},
{
"figure_ref": [],
"heading": "Acknowledgements",
"publication_ref": [],
"table_ref": [],
"text": "The research was carried out in the FASTVISION and FASTVISIONplus projects funded by the Academy of Finland (Decision numbers 321980, 321991, 339612, and 339355). Lumi Haraguchi was supported by OBAMA-NEXT (grant agreement no. 101081642), funded by the European Union under the Horizon Europe program."
},
{
"figure_ref": [],
"heading": "Data availability",
"publication_ref": [],
"table_ref": [],
"text": "No new data was collected for this survey."
},
{
"figure_ref": [],
"heading": "",
"publication_ref": [],
"table_ref": [],
"text": "to tackle the problems and to automate the analysis of the dataset. Some of the challenges, especially the limited amount of training data, have been rather extensively studied. While this problem cannot be considered solved, relatively high classification accuracies have been obtained with limited amounts of training images for certain classes. On the other hand, some of the other challenges have not been widely considered in plankton recognition literature. These include the domain shift between different image sets, presence of previously unseen classes and unknown particles, uncertainty in expert labels, and massive data volumes. The reasons for this vary. Most of the research has focused on improving classification accuracy and computation time has not been seen as an issue. Furthermore, the majority of the method development has been done for a fixed set of species and one imaging instrument, thus, there has been no need to address the domain shift or open-set problem. "
},
{
"figure_ref": [],
"heading": "Declaration of Competing Interest",
"publication_ref": [],
"table_ref": [],
"text": "The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper."
},
{
"figure_ref": [],
"heading": "Appendix A. Methods",
"publication_ref": [],
"table_ref": [],
"text": "Table A.4: Summary of the handcrafted image features used for recognizing plankton."
},
{
"figure_ref": [],
"heading": "Category Features Publications",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Texture features",
"publication_ref": [
"b225",
"b317"
],
"table_ref": [],
"text": "First order statistical descriptors (Julesz, 1962;Pratt, 2007) "
}
] | Planktonic organisms are key components of aquatic ecosystems and respond quickly to changes in the environment, therefore their monitoring is vital to follow and understand the changes in the environment. Yet, monitoring plankton at appropriate scales still remains a challenge, limiting our understanding of functioning of aquatic systems and their response to changes, which reduces the effectiveness of mitigation measures. Modern plankton imaging instruments can be utilized to sample at high frequencies, producing large amounts of images and enabling novel possibilities to study plankton populations. However, manual analysis of the data is costly, time consuming and expert based, making such approach unsuitable for large-scale application and urging for automatic solutions. The key problem related to the utilization of plankton datasets through image analysis is plankton recognition, i.e., classification of the images. Despite the large amount of research done on automatic plankton recognition, these methods have not been widely adopted for operational use. In this paper, a comprehensive survey on existing solutions for automatic plankton recognition is presented. First, we identify the most notable challenges that make the development of plankton recognition systems difficult and restrict the deployment of these systems for large-scale operational use. Then, we provide a detailed description of solutions for these challenges proposed in plankton recognition literature. Finally, we propose a workflow to identify the specific challenges in new datasets and the recommended approaches to address them. For many of the challenges, | Survey of Automatic Plankton Image Recognition: Challenges, Existing Solutions and Future Perspectives | [
{
"figure_caption": "Figure 1 :1Figure 1: Example images from the publicly available data sets: (a) Kaggle-Plankton (Cowen et al., 2015); (b) WHOI-Plankton (Orenstein et al., 2015); (c) PMID2019 (Li et al., 2019b); (d) ZooScan (Elineau et al., 2018); (d) DYB-PlanktonNet (Li et al., 2021a); (f) SYKE 2022(Kraft et al., 2022c).",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 3 Figure 2 :32Fig. 3 illustrates how the popularity of the CNNs and feature engineering based approaches on plankton recognition have changed over the years. It can be seen that the introduction of CNNs clearly boosted the research in the field.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "32",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: Popularity of feature engineering and feature learning (CNNs) based methods on plankton recognition. The plot contains the papers till October 2022.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "and Lumini et al. (2020) proposed an ensemble of classifiers by score fusion. Various classifier combinations containing different CNN models were evaluated for both plankton and coral classification. Henrichs et al. (2021) proposed an ensemble of 6 CNNs and showed it to outperform an RDF-based classifier. Kyathanahally et al. (2021b) compared various CNNs architectures in ensemble with multilayer perceptron (MLP) on zooplankton recognition using a mix of feature descriptors and CNNs features. While ensemble learning has shown slightly improved recognition accuracy, it also increases the computation time and complicates the training process.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: The nine main challenges that complicate the introduction of automatic plankton recognition methods to operational use.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 :5Figure 5: Examples of data augmentation methods.",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Figure 6 :6Figure 6: The difference between Traditional machine learning and Transfer learning.",
"figure_data": "",
"figure_id": "fig_7",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "and Nanni (2019a); Lumini et al. (2020) evaluated various strategies for transfer learning on plankton images. The first strategy was to initialize the model with ImageNet weights and fine-tune the whole model with plankton data. In the second strategy (two rounds tuning), a second pre-training step utilizing out-of-domain plankton image data was added before the fine-tuning. In the third strategy, ensembles of multiple different models were used. Based on the experiments the two rounds tuning did not provide a notable improvement in accuracy. Similarly, Guo et al. (2021b) explored and compared multiple transfer learning schemes on several biology image datasets from various domains. Various underwater and ecological image datasets are utilized for multistage transfer learning, where ImageNet pretraining is first improved by fine-tuning on an intermediate dataset before, finally, training on the target dataset consisting of plankton images. The experimental results show the potential of cross-domain transfer learning even on the out-of-domain data when the number of samples in the target domain is insufficient.",
"figure_data": "",
"figure_id": "fig_8",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 7 :7Figure 7: Plankton images with different sizes and aspect ratios. (Bureš et al., 2021).",
"figure_data": "",
"figure_id": "fig_9",
"figure_label": "7",
"figure_type": "figure"
},
{
"figure_caption": "Table A. 6 :6Summary of the CNN architectures used in the literature. The architectures developed specifically for plankton recognition are shown in bold. It should be noted that many architectures have various versions with different depths (e.g. VGGNet and ResNet). The number of layers and parameters for each architecture are based on the original publication. heterogeneity during seasonal bloom. Scientific",
"figure_data": "",
"figure_id": "fig_10",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "Existing image data sets.",
"figure_data": "DatasetType of planktonNumber of imagesNumber of classesImaging instrumentData collection regionADIAC database (Du Buf et al., 1999)Phytoplankton3 45285 Bright-field microscope -Kaggle-Plankton (PlanktonSet-1.0) (Cowen et al., 2015)Zooplankton phytoplankton30 336121 ISIIS-2Straits of Florida, U.S.WHOI-Plankton (Orenstein et al., 2015)Phytoplanktonover 3.5 M103 IFCBSouth Beach, Edgartown, Massachusetts, U.S.UVP5/MC (EcoTaxa) (Kiko and Simon-Martin, 2020) Zooplankton1.588 M *65 UVP5Worldwide (cruises)ZooScanNet (Elineau et al., 2018)Zooplankton1.433 M93 ZooScanVillefranche-sur-mer, FrancePMID2019 (Li et al., 2019b)Phytoplankton10 81924 Bright-field microscopeJiaozhou Bay, Qingdao, Shandong, ChinaminiPPlankton (Sun et al., 2020) (subset of PMID2019)Phytoplankton1 40020 Bright-field microscopeJiaozhou Bay, Qingdao, Shandong, ChinaDaya Bay (DYB),DYB-PlanktonNet (Li et al., 2021a)Zooplankton47 41990 Dark-field microscopeSouth China Sea,Shenzhen, China.Lake-Zooplankton (Kyathanahally et al., 2021a)Zooplankton17 94335Dual Scripps Plankton CameraLake Greifensee, SwitzerlandPlonus et al. 2021 (Plonus et al., 2021b)Zooplankton218 00026 VPRNorth Sea Baltic SeaSYKE-plankton IFCB 2022 (Kraft et al., 2022c)Phytoplankton63 00050 IFCBBaltic SeaSYKE-plankton IFCB Utö 2021 (Kraft et al., 2022a)Phytoplankton57 000 †50 IFCBUtö, Baltic Sea, Finland",
"figure_id": "tab_0",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Example accuracies on publicly available datasets from various sources. Numbers should be considered only indicative due to non-standardized evaluation protocols.",
"figure_data": "DatasetNumber of imagesNumber of speciesPublication: AccuracyWHOI3.5 M103Dai et al. (2016b): 96.3%Liu et al. (2018a): 86.3%Ho et al. (2018): 98.6%Venkataramanan et al. (2021): 90.5%Guo and Guan (2021): 72.9%Teigen et al. (2020): 58.8%Kyathanahally et al. (2021b): 96.1%Kaggle-30 336121Zheng and Wang (2015): 75.726%PlanktonLi and Cui (2016): 73.1%Yan et al. (2017): 76.4%Li et al. (2019c): 76.5%Geraldes et al. (2019): 83%Du et al. (2020): 75.8%Teigen et al. (2020): 70%Guo et al. (2021b): 77.45%Guo and Guan (2021): 86.5%Kyathanahally et al. (2021b): 94.7%ZooScanNet1.433 M93Zheng et al. (2017): 88.34%Guo and Guan (2021): 86.7%Kyathanahally et al. (2021b): 89.8%ZooLake17 94335Kyathanahally et al. (2021b): 98%1. The amount of labeled data for training is limited.This challenge can be divided into two subchallenges: 1) expert knowl-edge is required for data labeling, and 2) certain plankton species arenotably less common producing a small amount of example images.Plankton species are inherently difficult to identify, requiring prior ex-pertise. Labeling image data for training and evaluation purposes mustbe done by experts (e.g. plankton taxonomists) ruling out crowdsourc-ing tools such as Amazon Mechanical Turk commonly used for labelinglarge datasets. This makes labeling expensive limiting the amount oflabeled data. It also takes years to accumulate enough data to coverrare species. Collecting a large training set is essential for deep learningmodels. Larger amount of training data increases the model's capacity",
"figure_id": "tab_1",
"figure_label": "3",
"figure_type": "table"
}
] | Tuomas Eerola; Daniel Batrakhanov; Nastaran Vatankhah Barazandeh; Kaisa Kraft; Lumi Haraguchi; Lasse Lensu; Sanna Suikkanen; Jukka Seppälä; Timo Tamminen; Heikki Kälviäinen | [
{
"authors": " Embleton",
"journal": "Kramer",
"ref_id": "b0",
"title": "",
"year": "2003"
},
{
"authors": " Ye",
"journal": "",
"ref_id": "b1",
"title": "Second order statistical descriptors (Haralick features & Co-occurrence matrix (COM))",
"year": "1973"
},
{
"authors": " Thiel",
"journal": "",
"ref_id": "b2",
"title": "Binary Gradient Contours (BGC)",
"year": "1995"
},
{
"authors": " Zheng",
"journal": "",
"ref_id": "b3",
"title": "Gabor Descriptors",
"year": "2007"
},
{
"authors": " Ellis",
"journal": "",
"ref_id": "b4",
"title": "Local Binary Pattern (LBP)",
"year": "1997"
},
{
"authors": " Ojala",
"journal": "",
"ref_id": "b5",
"title": "",
"year": "2002"
},
{
"authors": " Blaschko",
"journal": "",
"ref_id": "b6",
"title": "Lisin",
"year": "2005"
},
{
"authors": " Schulze",
"journal": "",
"ref_id": "b7",
"title": "Variogram Function",
"year": "2013"
},
{
"authors": " Zheng",
"journal": "",
"ref_id": "b8",
"title": "",
"year": "2017"
},
{
"authors": " Zheng",
"journal": "Cosgriff",
"ref_id": "b9",
"title": "Fourier Descriptors on texture",
"year": "1960"
},
{
"authors": " Ellis",
"journal": "",
"ref_id": "b10",
"title": "",
"year": "1997"
},
{
"authors": " Mosleh",
"journal": "Cosgriff",
"ref_id": "b11",
"title": "Shape features Fourier Descriptors on contours",
"year": "1960"
},
{
"authors": " Thiel",
"journal": "",
"ref_id": "b12",
"title": "",
"year": "1995"
},
{
"authors": " Embleton",
"journal": "",
"ref_id": "b13",
"title": "General Geometric features and derived descriptors",
"year": "2003"
},
{
"authors": " Gorsky",
"journal": "",
"ref_id": "b14",
"title": "",
"year": "1989"
},
{
"authors": " Embleton",
"journal": "Du Buf and Bayer",
"ref_id": "b15",
"title": "Lisin",
"year": "2002"
},
{
"authors": " Lisin",
"journal": "Granulometries",
"ref_id": "b16",
"title": "Freeman contour code features",
"year": "1961"
},
{
"authors": " Tang",
"journal": "Lendaris and Stanley",
"ref_id": "b17",
"title": "Diffraction patterns",
"year": "1970"
},
{
"authors": "Olson Sosik",
"journal": "",
"ref_id": "b18",
"title": "Circular Projection",
"year": "2006"
},
{
"authors": " Zhao",
"journal": "",
"ref_id": "b19",
"title": "Object density",
"year": "2005"
},
{
"authors": " Zhao",
"journal": "",
"ref_id": "b20",
"title": "Boundary smoothness",
"year": "2005"
},
{
"authors": " Tang",
"journal": "Khotanzad and Hong",
"ref_id": "b21",
"title": "Zernike Moments (ZM)",
"year": "1990"
},
{
"authors": " Blaschko",
"journal": "Kuhl and Giardina",
"ref_id": "b22",
"title": "Affine Curvature Descriptors (ACD) (Liu and Watson",
"year": "1982"
},
{
"authors": " Schulze",
"journal": "",
"ref_id": "b23",
"title": "Moment Invariants (Hu Moments)",
"year": "1962"
},
{
"authors": " Reiss",
"journal": "",
"ref_id": "b24",
"title": "",
"year": "1991"
},
{
"authors": " Thiel",
"journal": "Kramer",
"ref_id": "b25",
"title": "Lisin",
"year": "1995"
},
{
"authors": " Schulze",
"journal": "",
"ref_id": "b26",
"title": "SHERPA (Shape Recognition, Processing and Analysis",
"year": "2013"
},
{
"authors": " Kloster",
"journal": "",
"ref_id": "b27",
"title": "",
"year": "2014"
},
{
"authors": " Beszteri",
"journal": "",
"ref_id": "b28",
"title": "",
"year": "2018"
},
{
"authors": "Hart Duda",
"journal": "",
"ref_id": "b29",
"title": "Hough descriptors",
"year": "1972"
},
{
"authors": " Blaschko",
"journal": "Ravela",
"ref_id": "b30",
"title": "Shape index",
"year": "2003"
},
{
"authors": " Lisin",
"journal": "Dalal and Triggs",
"ref_id": "b31",
"title": "Local features Histogram of Oriented Gradients (HOG)",
"year": "2005"
},
{
"authors": " Schulze",
"journal": "Ling and Jacobs",
"ref_id": "b32",
"title": "Inner-Distance shape context (IDSC)",
"year": "2007"
},
{
"authors": " Zheng",
"journal": "Kovesi",
"ref_id": "b33",
"title": "Phase congruency descriptors (PCD)",
"year": "2000"
},
{
"authors": " Verikas",
"journal": "Lowe",
"ref_id": "b34",
"title": "Scale Invariant Feature Transform (SIFT)",
"year": "1999"
},
{
"authors": " Lisin",
"journal": "Bay et al",
"ref_id": "b35",
"title": "Speeded Up Robust Features",
"year": "2005"
},
{
"authors": "Chang ",
"journal": "",
"ref_id": "b36",
"title": "",
"year": "2012"
},
{
"authors": " Py",
"journal": "",
"ref_id": "b37",
"title": "",
"year": "2016"
},
{
"authors": " Dai",
"journal": "",
"ref_id": "b38",
"title": "",
"year": "2016"
},
{
"authors": " Pedraza",
"journal": "",
"ref_id": "b39",
"title": "",
"year": "2017"
},
{
"authors": " Rodrigues",
"journal": "",
"ref_id": "b40",
"title": "",
"year": "2018"
},
{
"authors": " Wang",
"journal": "",
"ref_id": "b41",
"title": "",
"year": "2018"
},
{
"authors": " Cui",
"journal": "",
"ref_id": "b42",
"title": "",
"year": "2018"
},
{
"authors": " Sánchez",
"journal": "",
"ref_id": "b43",
"title": "Pardeshi and Deshmukh",
"year": "2019"
},
{
"authors": " Cheng",
"journal": "",
"ref_id": "b44",
"title": "",
"year": "2019"
},
{
"authors": "Nanni Lumini",
"journal": "",
"ref_id": "b45",
"title": "",
"year": "2019"
},
{
"authors": " Ärje",
"journal": "",
"ref_id": "b46",
"title": "",
"year": "2020"
},
{
"authors": " Khan",
"journal": "",
"ref_id": "b47",
"title": "",
"year": "2022"
},
{
"authors": " Vallez",
"journal": "Jindal and Mundra",
"ref_id": "b48",
"title": "ClassyFireNet",
"year": "2015"
},
{
"authors": "Mundra Jindal",
"journal": "",
"ref_id": "b49",
"title": "",
"year": "2015"
},
{
"authors": " Sánchez",
"journal": "",
"ref_id": "b50",
"title": "",
"year": "2019"
},
{
"authors": " Vallez",
"journal": "",
"ref_id": "b51",
"title": "",
"year": "2015"
},
{
"authors": "M Jindal; Mundra ",
"journal": "",
"ref_id": "b52",
"title": "",
"year": "2015"
},
{
"authors": " Cheng",
"journal": "",
"ref_id": "b53",
"title": "",
"year": "2019"
},
{
"authors": "Nanni ; Lumini; Khan",
"journal": "InceptionV",
"ref_id": "b54",
"title": "",
"year": "2016"
},
{
"authors": " Sánchez",
"journal": "",
"ref_id": "b55",
"title": "",
"year": "2019"
},
{
"authors": " Macneil",
"journal": "MobileNetV",
"ref_id": "b56",
"title": "",
"year": "2018"
},
{
"authors": " Lumini",
"journal": "",
"ref_id": "b57",
"title": "",
"year": "2018"
},
{
"authors": " Lumini",
"journal": "",
"ref_id": "b58",
"title": "",
"year": "2017"
},
{
"authors": " Liu",
"journal": "ResNet",
"ref_id": "b59",
"title": "",
"year": "2016"
},
{
"authors": "M Li; Cui ; Yan",
"journal": "",
"ref_id": "b60",
"title": "",
"year": "2016"
},
{
"authors": " Schröder",
"journal": "",
"ref_id": "b61",
"title": "",
"year": "2018"
},
{
"authors": " Liu",
"journal": "",
"ref_id": "b62",
"title": "",
"year": "2018"
},
{
"authors": " Libreros",
"journal": "",
"ref_id": "b63",
"title": "",
"year": "2018"
},
{
"authors": " Sánchez",
"journal": "",
"ref_id": "b64",
"title": "",
"year": "2019"
},
{
"authors": " Cheng",
"journal": "",
"ref_id": "b65",
"title": "",
"year": "2019"
},
{
"authors": " González",
"journal": "",
"ref_id": "b66",
"title": "Mitra et al",
"year": "2019"
},
{
"authors": "Nanni Lumini",
"journal": "",
"ref_id": "b67",
"title": "",
"year": "2019"
},
{
"authors": " Du",
"journal": "",
"ref_id": "b68",
"title": "",
"year": "2020"
},
{
"authors": " Schmarje",
"journal": "",
"ref_id": "b69",
"title": "",
"year": "2021"
},
{
"authors": "Guan Guo",
"journal": "",
"ref_id": "b70",
"title": "",
"year": "2021"
},
{
"authors": " Pu",
"journal": "",
"ref_id": "b71",
"title": "",
"year": "2021"
},
{
"authors": " Xu",
"journal": "",
"ref_id": "b72",
"title": "",
"year": "2022"
},
{
"authors": " Vallez",
"journal": "InceptionResNetV",
"ref_id": "b73",
"title": "",
"year": "2017"
},
{
"authors": " Kloster",
"journal": "Iandola et al",
"ref_id": "b74",
"title": "",
"year": "2016"
},
{
"authors": " Sánchez",
"journal": "Simonyan and Zisserman",
"ref_id": "b75",
"title": "",
"year": "2014"
},
{
"authors": " Kuang",
"journal": "",
"ref_id": "b76",
"title": "",
"year": "2015"
},
{
"authors": "Cui ; Li; Yan",
"journal": "",
"ref_id": "b77",
"title": "",
"year": "2016"
},
{
"authors": " Ho",
"journal": "",
"ref_id": "b78",
"title": "",
"year": "2018"
},
{
"authors": " Wang",
"journal": "",
"ref_id": "b79",
"title": "",
"year": "2018"
},
{
"authors": " Cheng",
"journal": "",
"ref_id": "b80",
"title": "",
"year": "2019"
},
{
"authors": "Mitra ",
"journal": "",
"ref_id": "b81",
"title": "Lumini and Nanni",
"year": "2019"
},
{
"authors": " Du",
"journal": "",
"ref_id": "b82",
"title": "",
"year": "2020"
},
{
"authors": " Kloster",
"journal": "",
"ref_id": "b83",
"title": "",
"year": "2020"
},
{
"authors": " Varma",
"journal": "",
"ref_id": "b84",
"title": "",
"year": "2020"
},
{
"authors": " Macneil",
"journal": "",
"ref_id": "b85",
"title": "",
"year": "2021"
},
{
"authors": " Pu",
"journal": "",
"ref_id": "b86",
"title": "",
"year": "2021"
},
{
"authors": " Khan",
"journal": "",
"ref_id": "b87",
"title": "",
"year": "2022"
},
{
"authors": " Rachman",
"journal": "",
"ref_id": "b88",
"title": "",
"year": "2022"
},
{
"authors": " Conradt",
"journal": "",
"ref_id": "b89",
"title": "",
"year": "2022"
},
{
"authors": " Vallez",
"journal": "Chollet",
"ref_id": "b90",
"title": "",
"year": "2017"
},
{
"authors": " Ho",
"journal": "",
"ref_id": "b91",
"title": "",
"year": "2018"
},
{
"authors": " Macneil",
"journal": "",
"ref_id": "b92",
"title": "",
"year": "2016"
},
{
"authors": " Dai",
"journal": "Tan and Le",
"ref_id": "b93",
"title": "EfficientNet",
"year": "2016"
},
{
"authors": " Venkataramanan",
"journal": "",
"ref_id": "b94",
"title": "",
"year": "2018"
},
{
"authors": " Xu",
"journal": "",
"ref_id": "b95",
"title": "References",
"year": "2022"
},
{
"authors": "H Al-Barazanchi; A Verma; S X Wang",
"journal": "International Journal of Computational Vision and Robotics",
"ref_id": "b96",
"title": "Intelligent plankton image classification with deep learning",
"year": "2018"
},
{
"authors": "H A Al-Barazanchi; A Verma; S Wang",
"journal": "IEEE",
"ref_id": "b97",
"title": "Performance evaluation of hybrid CNN for SIPPER plankton image calssification",
"year": "2015"
},
{
"authors": "H A Al-Barazanchi; A Verma; S Wang",
"journal": "",
"ref_id": "b98",
"title": "Plankton image classification using convolutional neural networks",
"year": "2015"
},
{
"authors": "P D Alfano; M Rando; M Letizia; F Odone; L Rosasco; V P Pastore",
"journal": "",
"ref_id": "b99",
"title": "Efficient unsupervised learning for plankton images",
"year": "2022"
},
{
"authors": "S Ali; Z Khan; A Hussain; A Athar; H C Kim",
"journal": "Water",
"ref_id": "b100",
"title": "Computer vision based deep learning approach for the detection and classification of algae species using microscopic images",
"year": "2022"
},
{
"authors": "K R Arrigo",
"journal": "Nature",
"ref_id": "b101",
"title": "Marine microorganisms and global nutrient cycles",
"year": "2005"
},
{
"authors": "Luo Aurelia; J ; Josette Boozallen; J Sullivan; S Mills; W Cukierski",
"journal": "",
"ref_id": "b102",
"title": "National Data Science Bowl",
"year": "2014"
},
{
"authors": "Badreldeen Bdawy; Mohamed ; O Eerola; T Kraft; K Lensu; L Kälviäinen; H ",
"journal": "",
"ref_id": "b103",
"title": "Open-set plankton recognition using similarity learning",
"year": "2022"
},
{
"authors": "L Barsanti; L Birindelli; P Gualtieri",
"journal": "Environmental Science: Processes & Impacts",
"ref_id": "b104",
"title": "Water monitoring by means of digital microscopy identification and classification of microalgae",
"year": "2021"
},
{
"authors": "H Bay; T Tuytelaars; L Van Gool",
"journal": "Springer",
"ref_id": "b105",
"title": "Surf: Speeded up robust features",
"year": "2006"
},
{
"authors": "O Beijbom; J Hoffman; E Yao; T Darrell; A Rodriguez-Ramirez; M Gonzalez-Rivero; O H Guldberg",
"journal": "",
"ref_id": "b106",
"title": "Quantification in-the-wild: Data-sets and baselines",
"year": "2015"
},
{
"authors": "J L Bell; R R Hopcroft",
"journal": "Journal of Plankton Research",
"ref_id": "b107",
"title": "Assessment of zooimage as a tool for the classification of zooplankton",
"year": "2008"
},
{
"authors": "S Ben-David; J Blitzer; K Crammer; A Kulesza; F Pereira; J W Vaughan",
"journal": "Machine Learning",
"ref_id": "b108",
"title": "A theory of learning from different domains",
"year": "2010"
},
{
"authors": "A Bendale; T Boult",
"journal": "",
"ref_id": "b109",
"title": "Towards open set deep networks",
"year": "2016"
},
{
"authors": "M C Benfield; P Grosjean; P F Culverhouse; X Irigoien; M E Sieracki; A Lopez-Urrutia; H G Dam; Q Hu; C S Davis; A Hansen; C H Pilskaln; E M Riseman; H Schultz; P E Utgoff; G Gorsky",
"journal": "Oceanography",
"ref_id": "b110",
"title": "Rapid: Research on automated plankton identification",
"year": "2007"
},
{
"authors": "B Bernhard; I M Guyon; V N Vapnik",
"journal": "Association for Computing Machinery",
"ref_id": "b111",
"title": "A training algorithm for optimal margin classifiers",
"year": "1992"
},
{
"authors": "B Beszteri; C Allen; G O Almandoz; L Armand; M Á Barcena; H Cantzler; X Crosta; O Esper; R W Jordan; G Kauer",
"journal": "Journal of Phycology",
"ref_id": "b112",
"title": "Quantitative comparison of taxa and taxon concepts in the diatom genus fragilariopsis: a case study on using slide scanning, multiexpert image annotation, and image analysis in taxonomy1",
"year": "2018"
},
{
"authors": "H Bi; Z Guo; M C Benfield; C Fan; M Ford; S Shahrestani; J M Sieracki",
"journal": "PLOS ONE",
"ref_id": "b113",
"title": "A semi-automated image analysis procedure for in situ plankton imaging systems",
"year": "2015"
},
{
"authors": "M B Blaschko; G Holness; M A Mattar; D Lisin; P E Utgoff; A R Hanson; H Schultz; E M Riseman; M E Sieracki; W M Balch",
"journal": "IEEE",
"ref_id": "b114",
"title": "Automatic in situ identification of plankton",
"year": "2005"
},
{
"authors": "E Bochinski; G Bacha; V Eiselein; T J Walles; J C Nejstgaard; T Sikora",
"journal": "",
"ref_id": "b115",
"title": "Deep active learning for in situ plankton classification",
"year": "2018"
},
{
"authors": "L Boddy; C Morris; M Wilkins; L Al-Haddad; G Tarran; R Jonker; P Burkill",
"journal": "Marine Ecology Progress Series",
"ref_id": "b116",
"title": "Identification of 72 phytoplankton species by radial basis function neural network analysis of flow cytometric data",
"year": "2000"
},
{
"authors": "L Boddy; C Morris; M Wilkins; G Tarran; P Burkill",
"journal": "Cytometry: The Journal of the International Society for Analytical Cytology",
"ref_id": "b117",
"title": "Neural network analysis of flow cytometric data for 40 marine phytoplankton species",
"year": "1994"
},
{
"authors": "G Bueno; O Deniz; A Pedraza; J Ruiz-Santaquiteria; J Salido; G Cristóbal; M Borrego-Ramos; S Blanco",
"journal": "Applied Sciences",
"ref_id": "b118",
"title": "Automated diatom classification (part a): handcrafted feature approaches",
"year": "2017"
},
{
"authors": "J Bureš; T Eerola; L Lensu; H Kälviäinen; P Zemčík",
"journal": "",
"ref_id": "b119",
"title": "Plankton recognition in images with varying size",
"year": "2021"
},
{
"authors": "S Cai; W Zuo; L Zhang",
"journal": "",
"ref_id": "b120",
"title": "Higher-order integration of hierarchical convolutional activations for fine-grained visual categorization",
"year": "2017"
},
{
"authors": "R W Campbell; P Roberts; J Jaffe",
"journal": "ICES Journal of Marine Science",
"ref_id": "b121",
"title": "The prince william sound plankton camera: a profiling in situ observatory of plankton and particulates",
"year": "2020"
},
{
"authors": "R J Campello; D Moulavi; A Zimek; J Sander",
"journal": "ACM Transactions on Knowledge Discovery from Data (TKDD)",
"ref_id": "b122",
"title": "Hierarchical density estimates for data clustering, visualization, and outlier detection",
"year": "2015"
},
{
"authors": "L Chang; R Wang; H Zheng; J Dai; B Zheng",
"journal": "IEEE",
"ref_id": "b123",
"title": "Phytoplankton feature extraction from microscopic images based on surf-pca",
"year": "2016"
},
{
"authors": "N V Chawla; K W Bowyer; L O Hall; W P Kegelmeyer",
"journal": "Journal of Artificial Intelligence Research",
"ref_id": "b124",
"title": "Smote: synthetic minority over-sampling technique",
"year": "2002"
},
{
"authors": "K Cheng; X Cheng; Q Hao",
"journal": "",
"ref_id": "b125",
"title": "A review of feature extraction technologies for plankton images",
"year": "2018"
},
{
"authors": "K Cheng; X Cheng; Y Wang; H Bi; M C Benfield",
"journal": "PLOS ONE",
"ref_id": "b126",
"title": "Enhanced convolutional neural network for plankton identification and enumeration",
"year": "2019"
},
{
"authors": "X Cheng; Y Ren; K Cheng; J Cao; Q Hao",
"journal": "Sensors",
"ref_id": "b127",
"title": "Method for training convolutional neural networks for in situ plankton image recognition and classification based on the mechanisms of the human eye",
"year": "2020"
},
{
"authors": "F Chollet",
"journal": "",
"ref_id": "b128",
"title": "Xception: Deep learning with depthwise separable convolutions",
"year": "2017"
},
{
"authors": "F Colas; M Tardivel; J Perchoc; M Lunven; B Forest; G Guyader; M M Danielou; S Le Mestre; P Bourriau; E Antajan",
"journal": "Progress in Oceanography",
"ref_id": "b129",
"title": "The ZooCAM, a new in-flow imaging system for fast onboard counting, sizing and classification of fish eggs and metazooplankton",
"year": "2018"
},
{
"authors": "P Coltelli; L Barsanti; V Evangelista; A M Frassanito; P Gualtieri",
"journal": "Environmental Science: Processes & Impacts",
"ref_id": "b130",
"title": "Water monitoring: automated and real time identification and classification of algae using digital microscopy",
"year": "2014"
},
{
"authors": "J Conradt; G Börner; Á López-Urrutia; C Möllmann; M Moyano",
"journal": "Frontiers in Marine Science",
"ref_id": "b131",
"title": "Automated Plankton Classification With a Dynamic Optimization and Adaptation Cycle",
"year": "2022"
},
{
"authors": "L Corgnati; S Marini; L Mazzei; E Ottaviani; S Aliani; A Conversi; A Griffa",
"journal": "Sensors",
"ref_id": "b132",
"title": "Looking inside the ocean: Toward an autonomous imaging system for monitoring gelatinous zooplankton",
"year": "2016"
},
{
"authors": "I Correa; P Drews; S Botelho; M S De Souza",
"journal": "",
"ref_id": "b133",
"title": "Deep learning for microalgae classification",
"year": "2017"
},
{
"authors": "I Corrêa; P Drews; M S De Souza; V M Tavano",
"journal": "IEEE",
"ref_id": "b134",
"title": "Supervised microalgae classification in imbalanced dataset",
"year": "2016"
},
{
"authors": "C Cortes; V Vapnik",
"journal": "Machine Learning",
"ref_id": "b135",
"title": "Support-vector networks",
"year": "1995"
},
{
"authors": "R Cosgriff",
"journal": "",
"ref_id": "b136",
"title": "Identification of shape",
"year": "1960"
},
{
"authors": "R Cowen; S Sponaugle; K Robinson; J Luo",
"journal": "National Centers for Environmental Information",
"ref_id": "b137",
"title": "PlanktonSet 1.0: Plankton Imagery Data Collected From F.G. Walton Smith in Straits of Florida From",
"year": "2014"
},
{
"authors": "R K Cowen; C M Guigand",
"journal": "Limnology and Oceanography: Methods",
"ref_id": "b138",
"title": "In situ ichthyoplankton imaging system (ISIIS): system design and preliminary results",
"year": "2008"
},
{
"authors": "J Cui; B Wei; C Wang; Z Yu; H Zheng; B Zheng; H Yang",
"journal": "OTO",
"ref_id": "b139",
"title": "Texture and shape information fusion of convolutional neural network for plankton image classification",
"year": "2018"
},
{
"authors": "P F Culverhouse",
"journal": "Ecological Informatics",
"ref_id": "b140",
"title": "Human and machine factors in algae monitoring performance",
"year": "2007"
},
{
"authors": "P F Culverhouse; R Williams; B Reguera; V Herry; S González-Gil",
"journal": "Marine Ecology Progress Series",
"ref_id": "b141",
"title": "Do experts make mistakes? a comparison of human and machine indentification of dinoflagellates",
"year": "2003"
},
{
"authors": "J Dai; R Wang; H Zheng; G Ji; X Qiao",
"journal": "",
"ref_id": "b142",
"title": "Zooplanktonet: Deep convolutional network for zooplankton classification",
"year": "2016"
},
{
"authors": "J Dai; Z Yu; H Zheng; B Zheng; N Wang",
"journal": "Springer",
"ref_id": "b143",
"title": "A hybrid convolutional neural network for plankton classification",
"year": "2016"
},
{
"authors": "N Dalal; B Triggs",
"journal": "IEEE",
"ref_id": "b144",
"title": "Histograms of oriented gradients for human detection",
"year": "2005"
},
{
"authors": "C S Davis; S M Gallager; A R Solow",
"journal": "Science",
"ref_id": "b145",
"title": "Microaggregations of oceanic plankton observed by towed video microscopy",
"year": "1992"
},
{
"authors": "C S Davis; Q Hu; S M Gallager; X Tang; C J Ashjian",
"journal": "Marine Ecology Progress Series",
"ref_id": "b146",
"title": "Real-time observation of taxa-specific plankton distributions: an optical sampling method",
"year": "2004"
},
{
"authors": "C S Davis; F T Thwaites; S M Gallager; Q Hu",
"journal": "Limnology and Oceanography: Methods",
"ref_id": "b147",
"title": "A three-axis fasttow digital video plankton recorder for rapid surveys of plankton taxa and hydrography",
"year": "2005"
},
{
"authors": "J Deng; W Dong; R Socher; L J Li; K Li; L Fei-Fei",
"journal": "IEEE",
"ref_id": "b148",
"title": "Imagenet: A large-scale hierarchical image database",
"year": "2009"
},
{
"authors": "J Deng; J Guo; N Xue; S Zafeiriou",
"journal": "",
"ref_id": "b149",
"title": "ArcFace: Additive angular margin loss for deep face recognition",
"year": "2019"
},
{
"authors": "I Dimitrovski; D Kocev; S Loskovska; S Džeroski",
"journal": "Ecological Informatics",
"ref_id": "b150",
"title": "Hierarchical classification of diatom images using ensembles of predictive clustering trees",
"year": "2012"
},
{
"authors": "H Ding; B Wei; Z Gu; Z Yu; H Zheng; B Zheng; J Li",
"journal": "Multimedia Tools and Applications",
"ref_id": "b151",
"title": "Ka-ensemble: towards imbalanced image classification ensembling undersampling and over-sampling",
"year": "2020"
},
{
"authors": "H Ding; B Wei; N Tang; Z Yu; N Wang; H Zheng; B Zheng",
"journal": "IEEE",
"ref_id": "b152",
"title": "Plankton image classification via multi-class imbalanced learning",
"year": "2018"
},
{
"authors": "A Dosovitskiy; L Beyer; A Kolesnikov; D Weissenborn; X Zhai; T Unterthiner; M Dehghani; M Minderer; G Heigold; S Gelly",
"journal": "",
"ref_id": "b153",
"title": "An image is worth 16x16 words: Transformers for image recognition at scale",
"year": "2021"
},
{
"authors": "P Drews; R G Colares; P Machado; M De Faria; A Detoni; V Tavano",
"journal": "Journal of the Brazilian Computer Society",
"ref_id": "b154",
"title": "Microalgae classification using semi-supervised and active learning based on Gaussian mixture models",
"year": "2013"
},
{
"authors": "A Du; Z Gu; Z Yu; H Zheng; B Zheng",
"journal": "IEEE",
"ref_id": "b155",
"title": "Plankton image classification using deep convolutional neural networks with second-order features",
"year": "2020"
},
{
"authors": "H Du Buf; M Bayer; S Droop; R Head; S Juggins; S Fischer; H Bunke; M Wilkinson; J Roerdink; J Pech-Pacheco",
"journal": "IEEE",
"ref_id": "b156",
"title": "Diatom identification: a double challenge called adiac",
"year": "1999"
},
{
"authors": "H Du Buf; M M Bayer",
"journal": "Cytometry: The Journal of the International Society for Analytical Cytology",
"ref_id": "b157",
"title": "Design and first results of CytoBuoy: A wireless flow cytometer for in situ analysis of marine and fresh waters",
"year": "1999"
},
{
"authors": "S R Dubey",
"journal": "IEEE Transactions on Circuits and Systems for Video Technology",
"ref_id": "b158",
"title": "A decade survey of content based image retrieval using deep learning",
"year": "2021"
},
{
"authors": "R O Duda; P E Hart",
"journal": "Communications of the ACM",
"ref_id": "b159",
"title": "Use of the hough transformation to detect lines and curves in pictures",
"year": "1972"
},
{
"authors": "S Dunker; D Boho; J Wäldchen; P Mäder",
"journal": "BMC Ecology",
"ref_id": "b160",
"title": "Combining highthroughput imaging flow cytometry and deep learning for efficient species and life-cycle stage identification of phytoplankton",
"year": "2018"
},
{
"authors": "V Dyomin; A Davydova; S Morgalev; N Kirillov; A Olshukov; I Polovtsev; S Davydov",
"journal": "Frontiers in Marine Science",
"ref_id": "b161",
"title": "Monitoring of plankton spatial and temporal characteristics with the use of a submersible digital holographic camera",
"year": "2020"
},
{
"authors": "V Dyomin; A Davydova; I Polovtsev; A Olshukov; N Kirillov; S Davydov",
"journal": "Sensors",
"ref_id": "b162",
"title": "Underwater holographic sensor for plankton studies in situ including accompanying measurements",
"year": "2021"
},
{
"authors": "V Dyomin; A Gribenyukov; A Davydova; M Zinoviev; A Olshukov; S Podzyvalov; I Polovtsev; N Yudin",
"journal": "Applied Optics",
"ref_id": "b163",
"title": "Holography of particles for diagnostics tasks",
"year": "2019"
},
{
"authors": "V Dyomin; I Polovtsev; A Y Davydova",
"journal": "",
"ref_id": "b164",
"title": "Fast recognition of marine particles in underwater digital holography",
"year": "2017"
},
{
"authors": "T Eerola; K Kraft; O Grönberg; L Lensu; S Suikkanen; J Seppälä; T Tamminen; H Kälviäinen; H Haario",
"journal": "Ocean Science Discussions",
"ref_id": "b165",
"title": "Towards operational phytoplankton recognition with automated high-throughput imaging and compact convolutional neural networks",
"year": "2020"
},
{
"authors": "A Elineau; C Desnos; L Jalabert; M Olivier; J B Romagnan; M Costa Brandao; F Lombard; N Llopis; J Courboules; L Caray-Counil; B Serranito; J O Irisson; M Picheral; G Gorsky; L Stemmann",
"journal": "",
"ref_id": "b166",
"title": "ZooScanNet: plankton images captured with the ZooScan",
"year": "2018"
},
{
"authors": "C Elkan",
"journal": "Morgan Kaufmann Publishers Inc",
"ref_id": "b167",
"title": "The foundations of cost-sensitive learning",
"year": "2001"
},
{
"authors": "J Ellen; H Li; M D Ohman",
"journal": "",
"ref_id": "b168",
"title": "Quantifying california current plankton samples with efficient machine learning techniques",
"year": "2015"
},
{
"authors": "J S Ellen; C A Graff; M D Ohman",
"journal": "Limnology and Oceanography: Methods",
"ref_id": "b169",
"title": "Improving plankton image classification using context metadata",
"year": "2019"
},
{
"authors": "R Ellis; R Simpson; P F Culverhouse; T Parisini",
"journal": "Neural Computing & Applications",
"ref_id": "b170",
"title": "Committees, collectives and individuals: Expert visual classification by neural network",
"year": "1997"
},
{
"authors": "K V Embleton; C Gibson; S Heaney",
"journal": "Journal of Plankton Research",
"ref_id": "b171",
"title": "Automated counting of phytoplankton by pattern recognition: a comparison with a manual counting method",
"year": "2003"
},
{
"authors": "R Faillettaz; M Picheral; J Y Luo; C Guigand; R K Cowen; J O Irisson",
"journal": "Methods in Oceanography",
"ref_id": "b172",
"title": "Imperfect automatic image classification successfully describes plankton distribution patterns",
"year": "2016"
},
{
"authors": "J A Fernandes; X Irigoien; G Boyra; J A Lozano; I Inza",
"journal": "Journal of Plankton Research",
"ref_id": "b173",
"title": "Optimizing the number of classes in automated zooplankton classification",
"year": "2009"
},
{
"authors": "A Fernández; M X Álvarez; F Bianconi",
"journal": "Optics and Lasers in Engineering",
"ref_id": "b174",
"title": "Image classification with binary gradient contours",
"year": "2011"
},
{
"authors": "S Fischer; F Šroubek; L Perrinet; R Redondo; G Cristóbal",
"journal": "International Journal of Computer Vision (IJCV)",
"ref_id": "b175",
"title": "Selfinvertible 2d log-gabor wavelets",
"year": "2007"
},
{
"authors": "H Freeman",
"journal": "IRE Transactions on Electronic Computers",
"ref_id": "b176",
"title": "On the encoding of arbitrary geometric configurations",
"year": "1961"
},
{
"authors": "Z Ge; S Liu; F Wang; Z Li; J Sun",
"journal": "",
"ref_id": "b177",
"title": "YOLOX: Exceeding yolo series in 2021",
"year": "2021"
},
{
"authors": "C Geng; S J Huang; S Chen",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI)",
"ref_id": "b178",
"title": "Recent advances in open set recognition: A survey",
"year": "2020"
},
{
"authors": "P Geraldes; J Barbosa; A Martins; A Dias; C Magalhães; S Ramos; E Silva",
"journal": "IEEE",
"ref_id": "b179",
"title": "In situ real-time zooplankton detection and classification",
"year": "2019"
},
{
"authors": "R Girshick; J Donahue; T Darrell; J Malik",
"journal": "",
"ref_id": "b180",
"title": "Rich feature hierarchies for accurate object detection and semantic segmentation",
"year": "2014"
},
{
"authors": "P González; E Álvarez; J Díez; Á López-Urrutia; J J Del Coz",
"journal": "Limnology and Oceanography: Methods",
"ref_id": "b181",
"title": "Validation methods for plankton image classification systems",
"year": "2017"
},
{
"authors": "P González; A Castaño; E E Peacock; J Díez; J J Del Coz; H M Sosik",
"journal": "Journal of Plankton Research",
"ref_id": "b182",
"title": "Automatic plankton quantification using deep features",
"year": "2019"
},
{
"authors": "I Goodfellow; J Pouget-Abadie; M Mirza; B Xu; D Warde-Farley; S Ozair; A Courville; Y Bengio",
"journal": "",
"ref_id": "b183",
"title": "Generative adversarial nets",
"year": "2014"
},
{
"authors": "M Goodwin; K T Halvorsen; L Jiao; K M Knausgård; A H Martin; M Moyano; R A Oomen; J H Rasmussen; T K Sørdalen; S H Thorbjørnsen",
"journal": "ICES Journal of Marine Science",
"ref_id": "b184",
"title": "Unlocking the potential of deep learning for marine ecology: overview, applications, and outlook",
"year": "2022"
},
{
"authors": "G Gorsky; P Guilbert; E Valenta",
"journal": "Marine Ecology Progress Series",
"ref_id": "b185",
"title": "The Autonomous Image Analyzer -enumeration, measurement and identification of marine phytoplankton",
"year": "1989"
},
{
"authors": "G Gorsky; M D Ohman; M Picheral; S Gasparini; L Stemmann; J B Romagnan; A Cawood; S Pesant; C García-Comas; F Prejger",
"journal": "Journal of Plankton Research",
"ref_id": "b186",
"title": "Digital zooplankton image analysis using the zooscan integrated system",
"year": "2010"
},
{
"authors": "A J H Goulart; A Morimitsu; R Jacomassi; N Hirata; R Lopes",
"journal": "",
"ref_id": "b187",
"title": "Deep learning and t-sne projection for plankton images clusterization",
"year": "2021"
},
{
"authors": "B Graham",
"journal": "",
"ref_id": "b188",
"title": "Spatially-sparse convolutional neural networks",
"year": "2014"
},
{
"authors": "P Grosjean; M Picheral; C Warembourg; G Gorsky",
"journal": "ICES Journal of Marine Science",
"ref_id": "b189",
"title": "Enumeration, measurement, and identification of net zooplankton samples using the zooscan digital imaging system",
"year": "2004"
},
{
"authors": "M M Grossmann; S M Gallager; S Mitarai",
"journal": "Journal of Oceanography",
"ref_id": "b190",
"title": "Continuous monitoring of near-bottom mesoplankton communities in the east china sea during a series of typhoons",
"year": "2015"
},
{
"authors": "J Gu; Z Wang; J Kuen; L Ma; A Shahroudy; B Shuai; T Liu; X Wang; G Wang; J Cai",
"journal": "Pattern Recognition",
"ref_id": "b191",
"title": "Recent advances in convolutional neural networks",
"year": "2018"
},
{
"authors": "I Gulrajani; D Lopez-Paz",
"journal": "",
"ref_id": "b192",
"title": "In search of lost domain generalization",
"year": "2020"
},
{
"authors": "B Guo; L Nyman; A R Nayak; D Milmore; M Mcfarland; M S Twardowski; J M Sullivan; J Yu; J Hong",
"journal": "Limnology and Oceanography: Methods",
"ref_id": "b193",
"title": "Automated plankton classification from holographic imagery with deep convolutional neural networks",
"year": "2021"
},
{
"authors": "C Guo; B Wei; K Yu",
"journal": "Journal of Control Science and Engineering",
"ref_id": "b194",
"title": "Deep transfer learning for biology crossdomain image classification",
"year": "2021"
},
{
"authors": "J Guo; J Guan",
"journal": "Arabian Journal for Science and Engineering",
"ref_id": "b195",
"title": "Classification of marine plankton based on few-shot learning",
"year": "2021"
},
{
"authors": "J Guo; W Li; J Guan; H Gao; B Liu; L Gong",
"journal": "IET Computer Vision",
"ref_id": "b196",
"title": "CDFM: A crossdomain few-shot model for marine plankton classification",
"year": "2022"
},
{
"authors": "J Guo; Y Ma; J H Lee",
"journal": "Journal of Hydro-Environment Research",
"ref_id": "b197",
"title": "Real-time automated identification of algal bloom species for fisheries management in subtropical coastal waters",
"year": "2021"
},
{
"authors": "X Guo; X Liu; E Zhu; J Yin",
"journal": "",
"ref_id": "b198",
"title": "Deep clustering with convolutional autoencoders",
"year": "2017"
},
{
"authors": "D Han; J Kim; J Kim",
"journal": "",
"ref_id": "b199",
"title": "Deep pyramidal residual networks",
"year": "2017"
},
{
"authors": "H Han; W Y Wang; B H Mao",
"journal": "",
"ref_id": "b200",
"title": "Borderline-smote: a new oversampling method in imbalanced data sets learning",
"year": "2005"
},
{
"authors": "R M Haralick; K Shanmugam; I H Dinstein",
"journal": "IEEE Transactions on Systems, Man, and Cybernetics",
"ref_id": "b201",
"title": "Textural features for image classification",
"year": "1973"
},
{
"authors": "B Hariharan; R Girshick",
"journal": "",
"ref_id": "b202",
"title": "Low-shot visual recognition by shrinking and hallucinating features",
"year": "2017"
},
{
"authors": "M L Haug",
"journal": "",
"ref_id": "b203",
"title": "Applying active learning techniques in machine learning to minimize labeling effort",
"year": "2021"
},
{
"authors": "M L Haug; A Saad; A Stahl",
"journal": "IFAC-PapersOnLine",
"ref_id": "b204",
"title": "Ciral: a hybrid active learning framework for plankon taxa labeling",
"year": "2021"
},
{
"authors": "M L Haug; A Saad; A Stahl",
"journal": "SPIE",
"ref_id": "b205",
"title": "A combined informative and representative active learning approach for plankton taxa labeling",
"year": "2021"
},
{
"authors": "G C Hays; A J Richardson; C Robinson",
"journal": "Trends in ecology & evolution",
"ref_id": "b206",
"title": "Climate change and marine plankton",
"year": "2005"
},
{
"authors": "K He; H Fan; Y Wu; S Xie; R Girshick",
"journal": "",
"ref_id": "b207",
"title": "Momentum contrast for unsupervised visual representation learning",
"year": "2020"
},
{
"authors": "K He; X Zhang; S Ren; J Sun",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI)",
"ref_id": "b208",
"title": "Spatial pyramid pooling in deep convolutional networks for visual recognition",
"year": "2015"
},
{
"authors": "K He; X Zhang; S Ren; J Sun",
"journal": "",
"ref_id": "b209",
"title": "Deep residual learning for image recognition",
"year": "2016"
},
{
"authors": "D W Henrichs; S Anglès; C C Gaonkar; L Campbell",
"journal": "Environmental Science and Pollution Research",
"ref_id": "b210",
"title": "Application of a convolutional neural network to improve automated early warning of harmful algal blooms",
"year": "2021"
},
{
"authors": "N S Hirata; M A Fernandez; R M Lopes",
"journal": "IEEE",
"ref_id": "b211",
"title": "Plankton image classification based on multiple segmentations",
"year": "2016"
},
{
"authors": "E Ho; B Henriquez; J Yeung",
"journal": "",
"ref_id": "b212",
"title": "Flagellates classification via transfer learning",
"year": "2018"
},
{
"authors": "T K Ho",
"journal": "IEEE",
"ref_id": "b213",
"title": "Random decision forests",
"year": "1995"
},
{
"authors": "E Hoffer; N Ailon",
"journal": "",
"ref_id": "b214",
"title": "Deep metric learning using triplet network",
"year": "2015"
},
{
"authors": "J Hu; L Shen; G Sun",
"journal": "",
"ref_id": "b215",
"title": "Squeeze-and-excitation networks",
"year": "2018"
},
{
"authors": "M K Hu",
"journal": "IRE Transactions on Information Theory",
"ref_id": "b216",
"title": "Visual pattern recognition by moment invariants",
"year": "1962"
},
{
"authors": "Q Hu; C Davis",
"journal": "Marine Ecology Progress Series",
"ref_id": "b217",
"title": "Automatic plankton image recognition with cooccurrence matrices and support vector machine",
"year": "2005"
},
{
"authors": "Q Hu; C Davis",
"journal": "Marine Ecology Progress Series",
"ref_id": "b218",
"title": "Accurate automatic quantification of taxa-specific plankton abundance using dual classification with correction",
"year": "2006"
},
{
"authors": "G Huang; Z Liu; L Van Der Maaten; K Q Weinberger",
"journal": "",
"ref_id": "b219",
"title": "Densely connected convolutional networks",
"year": "2017"
},
{
"authors": "F N Iandola; S Han; M W Moskewicz; K Ashraf; W J Dally; K Keutzer",
"journal": "",
"ref_id": "b220",
"title": "Squeezenet: Alexnet-level accuracy with 50x fewer parameters and < 0.5 mb model size",
"year": "2016"
},
{
"authors": "M Ibrahim",
"journal": "",
"ref_id": "b221",
"title": "Image clustering for unsupervised analysis of plankton data",
"year": "2020"
},
{
"authors": "M Idrissa; M Acheroy",
"journal": "Pattern Recognition Letters",
"ref_id": "b222",
"title": "Texture classification using gabor filters",
"year": "2002"
},
{
"authors": "J O Irisson; S D Ayata; D J Lindsay; L Karp-Boss; L Stemmann",
"journal": "Annual Review of Marine Science",
"ref_id": "b223",
"title": "Machine learning for the study of plankton and marine snow from images",
"year": "2022"
},
{
"authors": "P Jindal; R Mundra",
"journal": "",
"ref_id": "b224",
"title": "Plankton classification using hybrid convolutional network-random forests architectures",
"year": "2015"
},
{
"authors": "B Julesz",
"journal": "IRE Transactions on Information Theory",
"ref_id": "b225",
"title": "Visual pattern discrimination",
"year": "1962"
},
{
"authors": "A S Keçeli; A Kaya; S U Keçeli",
"journal": "Computers & Geosciences",
"ref_id": "b226",
"title": "Classification of radiolarian images with hand-crafted and deep features",
"year": "2017"
},
{
"authors": "T Kerr; J R Clark; E S Fileman; C E Widdicombe; N Pugeault",
"journal": "IEEE Access",
"ref_id": "b227",
"title": "Collaborative deep learning models to handle class imbalance in flowcam plankton imagery",
"year": "2020"
},
{
"authors": "S Khalid; T Khalil; S Nasreen",
"journal": "IEEE",
"ref_id": "b228",
"title": "A survey of feature selection and feature extraction techniques in machine learning",
"year": "2014"
},
{
"authors": "Z Khan; W Mumtaz; A S Mumtaz; S Bhattacharjee; H C Kim",
"journal": "IEEE",
"ref_id": "b229",
"title": "Multiclass-classification of algae using dc-gan and transfer learning",
"year": "2022"
},
{
"authors": "A Khotanzad; Y H Hong",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI)",
"ref_id": "b230",
"title": "Invariant image recognition by zernike moments",
"year": "1990"
},
{
"authors": "R Kiko; S Simon-Martin",
"journal": "",
"ref_id": "b231",
"title": "UVP5 data sorted with EcoTaxa and morphocluster",
"year": "2020"
},
{
"authors": "J Kingman",
"journal": "Bulletin of the American Mathematical Society",
"ref_id": "b232",
"title": "G. matheron, random sets and integral geometry",
"year": "1975"
},
{
"authors": "M Kloster; G Kauer; B Beszteri",
"journal": "BMC Bioinformatics",
"ref_id": "b233",
"title": "Sherpa: an image segmentation and outline feature extraction tool for diatoms and other objects",
"year": "2014"
},
{
"authors": "M Kloster; D Langenkämper; M Zurowietz; B Beszteri; T W Nattkemper",
"journal": "Scientific Reports",
"ref_id": "b234",
"title": "Deep learning-based diatom taxonomy on virtual slides",
"year": "2020"
},
{
"authors": "S Kosov; K Shirahama; C Li; M Grzegorzek",
"journal": "Pattern Recognition",
"ref_id": "b235",
"title": "Environmental microorganism classification using conditional random fields and deep convolutional neural networks",
"year": "2018"
},
{
"authors": "P Kovesi",
"journal": "Psychological Research",
"ref_id": "b236",
"title": "Phase congruency: A low-level image invariant",
"year": "2000"
},
{
"authors": "P Kovesi",
"journal": "",
"ref_id": "b237",
"title": "Phase congruency detects corners and edges",
"year": "2003"
},
{
"authors": "K Kraft; L Haraguchi; O Velhonoja; J Seppälä",
"journal": "",
"ref_id": "b238",
"title": "SYKE-phytoplankton IFCB Utö",
"year": "2021"
},
{
"authors": "K Kraft; O Velhonoja; T Eerola; S Suikkanen; T Tamminen; L Haraguchi; P Ylöstalo; S Kielosto; M Johansson; L Lensu; H Kälviäinen; H Haario; J Seppälä",
"journal": "Frontiers in Marine Science",
"ref_id": "b239",
"title": "Towards operational phytoplankton recognition with automated high-throughput imaging, nearreal-time data processing, and convolutional neural networks",
"year": "2022"
},
{
"authors": "K Kraft; O Velhonoja; J Seppälä; H Hällfors; S Suikkanen; P Ylöstalo; S Angles; S Kielosto; Harri ; Kuosa; S Lehtinen; J Oja; T Tamminen",
"journal": "",
"ref_id": "b240",
"title": "SYKE-phytoplankton IFCB",
"year": "2022"
},
{
"authors": "K A Kramer",
"journal": "",
"ref_id": "b241",
"title": "Identifying Plankton from Grayscale Silhouette Images",
"year": "2005"
},
{
"authors": "K A Kramer",
"journal": "",
"ref_id": "b242",
"title": "System for identifying plankton from the sipper instrument platform",
"year": "2010"
},
{
"authors": "A Krizhevsky; I Sutskever; G E Hinton",
"journal": "",
"ref_id": "b243",
"title": "Imagenet classification with deep convolutional neural networks",
"year": "2012"
},
{
"authors": "Y Kuang",
"journal": "",
"ref_id": "b244",
"title": "Deep neural network for deep sea plankton classification",
"year": "2015"
},
{
"authors": "F P Kuhl; C R Giardina",
"journal": "Computer Graphics and Image Processing",
"ref_id": "b245",
"title": "Elliptic fourier features of a closed contour",
"year": "1982"
},
{
"authors": "S Kyathanahally; T Hardeman; E Merz; T Kozakiewicz; M Reyes; P Isles; F Pomati; M Baity-Jesi",
"journal": "",
"ref_id": "b246",
"title": "Data for: Deep learning classification of lake zooplankton",
"year": "2021"
},
{
"authors": "S Kyathanahally; T Hardeman; M Reyes; E Merz; T Bulas; F Pomati; M Baity-Jesi",
"journal": "",
"ref_id": "b247",
"title": "Ensembles of vision transformers as a new paradigm for automated classification in ecology",
"year": "2022"
},
{
"authors": "S P Kyathanahally; T Hardeman; E Merz; T Bulas; M Reyes; P Isles; F Pomati; M Baity-Jesi",
"journal": "Frontiers in Microbiology",
"ref_id": "b248",
"title": "Deep learning classification of lake zooplankton",
"year": "2021"
},
{
"authors": "Q T Lai; K C Lee; A H Tang; K K Wong; H K So; K K Tsia",
"journal": "Optics Express",
"ref_id": "b249",
"title": "High-throughput time-stretch imaging flow cytometry for multi-class classification of phytoplankton",
"year": "2016"
},
{
"authors": "M Lauffer; F Genty; S Margueron; J L Collette",
"journal": "Photosynthetica",
"ref_id": "b250",
"title": "Morphological recognition with the addition of multi-band fluorescence excitation of chlorophylls of phytoplankton",
"year": "2017"
},
{
"authors": "K T Le; Z Yuan; A Syed; D Ratelle; E C Orenstein; M L Carter; S Strang; K M Kenitz; P Morgado; P J S Franks; N Vasconcelos; J S Jaffe",
"journal": "Frontiers in Marine Science",
"ref_id": "b251",
"title": "Benchmarking and Automating the Image Recognition Capability of an In Situ Plankton Imaging System",
"year": "2022"
},
{
"authors": "Y Lecun; Y Bengio; G Hinton",
"journal": "Nature",
"ref_id": "b252",
"title": "Deep learning",
"year": "2015"
},
{
"authors": "H Lee; M Park; J Kim",
"journal": "IEEE",
"ref_id": "b253",
"title": "Plankton classification on imbalanced large scale database via convolutional neural networks with transfer learning",
"year": "2016"
},
{
"authors": "G G Lendaris; G L Stanley",
"journal": "",
"ref_id": "b254",
"title": "Diffraction-pattern sampling for automatic pattern recognition",
"year": "1970"
},
{
"authors": "C Li; K Wang; N Xu",
"journal": "Artificial Intelligence Review",
"ref_id": "b255",
"title": "A survey for the applications of contentbased microscopic image analysis in microorganism classification domains",
"year": "2019"
},
{
"authors": "J Li; Z Yang; T Chen",
"journal": "DYB-PlanktonNet",
"ref_id": "b256",
"title": "",
"year": "2021"
},
{
"authors": "P Li; J Xie; Q Wang; W Zuo",
"journal": "",
"ref_id": "b257",
"title": "Is second-order information helpful for large-scale visual recognition?",
"year": "2017"
},
{
"authors": "Q Li; X Sun; J Dong; S Song; T Zhang; D Liu; H Zhang; S Han",
"journal": "ICES Journal of Marine Science",
"ref_id": "b258",
"title": "Developing a microscopic image dataset in support of intelligent phytoplankton detection using deep learning",
"year": "2019"
},
{
"authors": "X Li; Z Cui",
"journal": "",
"ref_id": "b259",
"title": "Deep residual networks for plankton classification",
"year": "2016"
},
{
"authors": "X Li; R Long; J Yan; K Jin; J Lee",
"journal": "Mobile Information Systems",
"ref_id": "b260",
"title": "Tanet: A tiny plankton classification network for mobile devices",
"year": "2019"
},
{
"authors": "Y Li; J Guo; X Guo; Z Hu; Y Tian",
"journal": "Journal of Marine Science and Engineering",
"ref_id": "b261",
"title": "Plankton detection with adversarial learning and a densely connected deep learning model for class imbalanced distribution",
"year": "2021"
},
{
"authors": "Y Li; J Guo; X Guo; J Zhao; Y Yang; Z Hu; W Jin; Y Tian",
"journal": "Applied Ocean Research",
"ref_id": "b262",
"title": "Toward in situ zooplankton detection with a densely connected yolov3 model",
"year": "2021"
},
{
"authors": "Z Li; F Zhao; J Liu; Y Qiao",
"journal": "IEEE Journal of Oceanic Engineering",
"ref_id": "b263",
"title": "Pairwise nonparametric discriminant analysis for binary plankton image recognition",
"year": "2013"
},
{
"authors": "J Libreros; G Bueno; M Trujillo; M Ospina",
"journal": "Springer",
"ref_id": "b264",
"title": "Automated identification and classification of diatoms from water resources",
"year": "2018"
},
{
"authors": "B Lim; S Son; H Kim; S Nah; K Mu Lee",
"journal": "",
"ref_id": "b265",
"title": "Enhanced deep residual networks for single image super-resolution",
"year": "2017"
},
{
"authors": "H Ling; D W Jacobs",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI)",
"ref_id": "b266",
"title": "Shape classification using the inner-distance",
"year": "2007"
},
{
"authors": "D A Lisin",
"journal": "",
"ref_id": "b267",
"title": "Image classification with bags of local features",
"year": "2006"
},
{
"authors": "D A Lisin; M A Mattar; M B Blaschko; E G Learned-Miller; M C Benfield",
"journal": "IEEE",
"ref_id": "b268",
"title": "Combining local and global image features for object class recognition",
"year": "2005"
},
{
"authors": "J Liu; A Du; C Wang; Z Yu; H Zheng; B Zheng; H Zhang",
"journal": "IEEE",
"ref_id": "b269",
"title": "Deep pyramidal residual networks for plankton image classification",
"year": "2018"
},
{
"authors": "J Liu; A Du; C Wang; H Zheng; N Wang; B Zheng",
"journal": "IEEE",
"ref_id": "b270",
"title": "Teaching squeeze-and-excitation pyramidnet for imbalanced image classification with gan-based curriculum learning",
"year": "2018"
},
{
"authors": "W Liu; D Anguelov; D Erhan; C Szegedy; S Reed; C Y Fu; A C Berg",
"journal": "",
"ref_id": "b271",
"title": "SSD: Single shot multibox detector",
"year": "2016"
},
{
"authors": "Y Liu; X Qiao; R Gao",
"journal": "IEEE",
"ref_id": "b272",
"title": "Plankton classification on imbalanced dataset via hybrid resample method with lightbgm",
"year": "2021"
},
{
"authors": "Z Liu; J Watson",
"journal": "Singapore-U.S. Gulf Coast",
"ref_id": "b273",
"title": "Shape-based image classification and identification system for digital holograms of marine particles and plankton",
"year": "2020"
},
{
"authors": "Z Liu; J Watson; A Allen",
"journal": "IEEE",
"ref_id": "b274",
"title": "Efficient affine-invariant fourier descriptors for identification of marine plankton",
"year": "2017"
},
{
"authors": "F Lombard; E Boss; A M Waite; J Uitz; L Stemmann; H M Sosik; J Schulz; J B Romagnan; M Picheral; J Pearlman",
"journal": "Frontiers in Marine Science",
"ref_id": "b275",
"title": "Globally consistent quantitative observations of planktonic ecosystems",
"year": "2019"
},
{
"authors": "D G Lowe",
"journal": "IEEE",
"ref_id": "b276",
"title": "Object recognition from local scale-invariant features",
"year": "1999"
},
{
"authors": "D G Lowe",
"journal": "International Journal of Computer Vision (IJCV)",
"ref_id": "b277",
"title": "Distinctive image features from scale-invariant keypoints",
"year": "2004"
},
{
"authors": "A Lumini; L Nanni",
"journal": "Ecological Informatics",
"ref_id": "b278",
"title": "Deep learning and transfer learning features for plankton classification",
"year": "2019"
},
{
"authors": "A Lumini; L Nanni",
"journal": "Recent Advances in Computer Vision",
"ref_id": "b279",
"title": "Ocean ecosystems plankton classification",
"year": "2019"
},
{
"authors": "A Lumini; L Nanni; G Maguolo",
"journal": "Applied Computing and Informatics",
"ref_id": "b280",
"title": "Deep learning for plankton and coral classification",
"year": "2020"
},
{
"authors": "J Y Luo; J O Irisson; B Graham; C Guigand; A Sarafraz; C Mader; R K Cowen",
"journal": "Limnology and Oceanography: Methods",
"ref_id": "b281",
"title": "Automated plankton image analysis using convolutional neural networks",
"year": "2018"
},
{
"authors": "Q Luo; Y Gao; J Luo; C Chen; J Liang; C Yang",
"journal": "Journal of Software",
"ref_id": "b282",
"title": "Automatic Identification of Diatoms with Circular Shape using Texture Analysis",
"year": "2011"
},
{
"authors": "S Luo; K T Nguyen; B T Nguyen; S Feng; Y Shi; A Elsayed; Y Zhang; X Zhou; B Wen; G Chierchia",
"journal": "Cytometry Part A",
"ref_id": "b283",
"title": "Deep learning-enabled imaging flow cytometry for high-speed cryptosporidium and giardia detection",
"year": "2021"
},
{
"authors": "S Luo; Y Shi; L K Chin; P E Hutchinson; Y Zhang; G Chierchia; H Talbot; X Jiang; T Bourouina; A Q Liu",
"journal": "Advanced Intelligent Systems",
"ref_id": "b284",
"title": "Machine-learningassisted intelligent imaging flow cytometry: A review",
"year": "2021"
},
{
"authors": "T Luo",
"journal": "",
"ref_id": "b285",
"title": "Scaling Up Support Vector Machines with Application to Plankton Recognition",
"year": "2005"
},
{
"authors": "T Luo; K Kramer; D Goldgof; L O Hall; S Samson; A Remsen; T Hopkins",
"journal": "IEEE",
"ref_id": "b286",
"title": "Learning to recognize plankton",
"year": "2003"
},
{
"authors": "T Luo; K Kramer; D B Goldgof; L O Hall; S Samson; A Remsen; T Hopkins",
"journal": "IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics)",
"ref_id": "b287",
"title": "Recognizing plankton images from the shadow image particle profiling evaluation recorder",
"year": "2004"
},
{
"authors": "T Luo; K Kramer; D B Goldgof; L O Hall; S Samson; A Remsen; T Hopkins",
"journal": "Journal of Machine Learning Research",
"ref_id": "b288",
"title": "Active learning to recognize multiple types of plankton",
"year": "2005"
},
{
"authors": "W Ma; T Chen; Z Zhang; Z Yang; C Dong; J Qiao; J Li",
"journal": "",
"ref_id": "b289",
"title": "Super-resolution for in situ plankton images",
"year": "2021"
},
{
"authors": "N Macleod; M Benfield; P Culverhouse",
"journal": "Nature",
"ref_id": "b290",
"title": "Time to automate identification",
"year": "2010"
},
{
"authors": "L Macneil; S Missan; J Luo; T Trappenberg; J Laroche",
"journal": "BMC Ecology and Evolution",
"ref_id": "b291",
"title": "Plankton classification with high-throughput submersible holographic microscopy and transfer learning",
"year": "2021"
},
{
"authors": "R Mechrez; I Talmi; L Zelnik-Manor",
"journal": "",
"ref_id": "b292",
"title": "The contextual loss for image transformation with non-aligned data",
"year": "2018"
},
{
"authors": "Y Mirasbekov; A Zhumakhanova; A Zhantuyakova; K Sarkytbayev; D V Malashenkov; A Baishulakova; V Dashkova; T A Davidson; I A Vorobjev; E Jeppesen",
"journal": "Reports",
"ref_id": "b293",
"title": "Semi-automated classification of colonial microcystis by flowcam imaging flow cytometry in mesocosm",
"year": "2021"
},
{
"authors": "R Mitra; T Marchitto; Q Ge; B Zhong; B Kanakiya; M Cook; J Fehrenbacher; J Ortiz; A Tripati; E Lobaton",
"journal": "Marine Micropaleontology",
"ref_id": "b294",
"title": "Automated specieslevel identification of planktic foraminifera using convolutional neural networks, with comparison to human performance",
"year": "2019"
},
{
"authors": "S Mittal; S Srivastava; J P Jayanth",
"journal": "IEEE Transactions on Neural Networks and Learning Systems",
"ref_id": "b295",
"title": "A survey of deep learning techniques for underwater image classification",
"year": "2022"
},
{
"authors": "M Moniruzzaman; S M S Islam; M Bennamoun; P Lavery",
"journal": "Springer",
"ref_id": "b296",
"title": "Deep learning on underwater marine object detection: A survey",
"year": "2017"
},
{
"authors": "M A Mosleh; H Manssor; S Malek; P Milow; A Salleh",
"journal": "BMC Bioinformatics",
"ref_id": "b297",
"title": "A preliminary study on automated freshwater algae recognition and classification system",
"year": "2012"
},
{
"authors": "Y Movshovitz-Attias; A Toshev; T K Leung; S Ioffe; S Singh",
"journal": "",
"ref_id": "b298",
"title": "No fuss distance metric learning using proxies",
"year": "2017"
},
{
"authors": "A R Nayak; M N Mcfarland; J M Sullivan; M S Twardowski",
"journal": "Limnology and Oceanography",
"ref_id": "b299",
"title": "Evidence for ubiquitous preferential particle orientation in representative oceanic shear flows",
"year": "2018"
},
{
"authors": "E Nepovinnykh; T Eerola; H Kalviainen",
"journal": "",
"ref_id": "b300",
"title": "Siamese network based pelage pattern matching for ringed seal re-identification",
"year": "2020"
},
{
"authors": "T Ojala; M Pietikainen; T Maenpaa",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI)",
"ref_id": "b301",
"title": "Multiresolution gray-scale and rotation invariant texture classification with local binary patterns",
"year": "2002"
},
{
"authors": "R J Olson; H M Sosik",
"journal": "Limnology and Oceanography: Methods",
"ref_id": "b302",
"title": "A submersible imaging-in-flow instrument to analyze nano-and microplankton: Imaging flowcytobot",
"year": "2007"
},
{
"authors": "E C Orenstein; O Beijbom",
"journal": "IEEE",
"ref_id": "b303",
"title": "Transfer learning and deep feature extraction for planktonic image data sets",
"year": "2017"
},
{
"authors": "E C Orenstein; O Beijbom; E E Peacock; H M Sosik",
"journal": "",
"ref_id": "b304",
"title": "WHOIplankton-a large scale fine grained visual recognition benchmark dataset for plankton classification",
"year": "2015"
},
{
"authors": "E C Orenstein; K M Kenitz; P L Roberts; P J Franks; J S Jaffe; A D Barton",
"journal": "Limnology and Oceanography: Methods",
"ref_id": "b305",
"title": "Semi-and fully supervised quantification techniques to improve population estimates from machine classifiers",
"year": "2020"
},
{
"authors": "E C Orenstein; D Ratelle; C Briseño-Avena; M L Carter; P J Franks; J S Jaffe; P L Roberts",
"journal": "Limnology and Oceanography: Methods",
"ref_id": "b306",
"title": "The scripps plankton camera system: A framework and platform for in situ microscopy",
"year": "2020"
},
{
"authors": "N Otsu",
"journal": "Automatica",
"ref_id": "b307",
"title": "A threshold selection method from gray-level histograms",
"year": "1975"
},
{
"authors": "S J Pan; Q Yang",
"journal": "IEEE Transactions on Knowledge and Data Engineering",
"ref_id": "b308",
"title": "A survey on transfer learning",
"year": "2009"
},
{
"authors": "R Pardeshi; P D Deshmukh",
"journal": "Springer",
"ref_id": "b309",
"title": "Classification of microscopic algae: An observational study with alexnet",
"year": "2019"
},
{
"authors": "V P Pastore; T G Zimmerman; S K Biswas; S Bianco",
"journal": "Scientific Reports",
"ref_id": "b310",
"title": "Annotation-free learning of plankton for classification and anomaly detection",
"year": "2020"
},
{
"authors": "A Pedraza; G Bueno; O Deniz; G Cristóbal; S Blanco; M Borrego-Ramos",
"journal": "Applied Sciences",
"ref_id": "b311",
"title": "Automated diatom classification (Part B): A deep learning approach",
"year": "2017"
},
{
"authors": "A Pedraza; G Bueno; O Deniz; J Ruiz-Santaquiteria; C Sanchez; S Blanco; M Borrego-Ramos; A Olenici; G Cristobal",
"journal": "International Society for Optics and Photonics (SPIE",
"ref_id": "b312",
"title": "Lights and pitfalls of convolutional neural networks for diatom identification",
"year": "2018"
},
{
"authors": "M Picheral; S Colin; J O Irisson",
"journal": "",
"ref_id": "b313",
"title": "EcoTaxa, a tool for the taxonomic classification of images",
"year": "2017"
},
{
"authors": "M Picheral; L Guidi; L Stemmann; D M Karl; G Iddaoud; G Gorsky",
"journal": "Limnology and Oceanography: Methods",
"ref_id": "b314",
"title": "The underwater vision profiler 5: An advanced instrument for high spatial resolution studies of particle size spectra and zooplankton",
"year": "2010"
},
{
"authors": "R M Plonus; J Conradt; A Harmer; S Janßen; J Floeter",
"journal": "Limnology and Oceanography: Methods",
"ref_id": "b315",
"title": "Automatic plankton image classification -Can capsules and filters help cope with data set shift?",
"year": "2021"
},
{
"authors": "R M Plonus; J Conradt; A Harmer; S Janßen; J Floeter",
"journal": "Dataset",
"ref_id": "b316",
"title": "Automatic plankton image classification -can capsules and filters help coping with data set shift?",
"year": "2021"
},
{
"authors": "W K Pratt",
"journal": "John Wiley & Sons, Ltd",
"ref_id": "b317",
"title": "Image Feature Extraction",
"year": "2007"
},
{
"authors": "Y Pu; Z Feng; Z Wang; Z Yang; J Li",
"journal": "",
"ref_id": "b318",
"title": "Anomaly detection for in situ marine plankton images",
"year": "2021"
},
{
"authors": "O Py; H Hong; S Zhongzhi",
"journal": "IEEE",
"ref_id": "b319",
"title": "Plankton classification with deep convolutional neural networks",
"year": "2016"
},
{
"authors": "H Qi; M Brown; D G Lowe",
"journal": "",
"ref_id": "b320",
"title": "Low-shot learning with imprinted weights",
"year": "2018"
},
{
"authors": "X Qiao; M Tang; Z Tang; K Lang; X Wang",
"journal": "SPIE",
"ref_id": "b321",
"title": "Classification of phytoplankton digital holograms using transfer learning",
"year": "2021"
},
{
"authors": "A Rachman; A S Suwarno; S Nurdjaman",
"journal": "Atlantis Press",
"ref_id": "b322",
"title": "Application of deep (machine) learning for phytoplankton identification using microscopy images",
"year": "2022"
},
{
"authors": "F Radenović; G Tolias; O Chum",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI)",
"ref_id": "b323",
"title": "Fine-tuning cnn image retrieval with no human annotation",
"year": "2018"
},
{
"authors": "A Radford; L Metz; S Chintala",
"journal": "",
"ref_id": "b324",
"title": "Unsupervised representation learning with deep convolutional generative adversarial networks",
"year": "2015"
},
{
"authors": "J Raitoharju; E Riabchenko; K Meissner; I Ahmad; A Iosifidis; M Gabbouj; S Kiranyaz",
"journal": "IEEE",
"ref_id": "b325",
"title": "Data enrichment in fine-grained classification of aquatic macroinvertebrates",
"year": "2016"
},
{
"authors": "P Rani; S Kotwal; J Manhas; V Sharma; S Sharma",
"journal": "Archives of Computational Methods in Engineering",
"ref_id": "b326",
"title": "Machine learning and deep learning based computational approaches in automatic microorganisms image recognition: methodologies, challenges, and developments",
"year": "2021"
},
{
"authors": "S S Ravela",
"journal": "",
"ref_id": "b327",
"title": "On multi-scale differential features and their representations for image retrieval and recognition",
"year": "2003"
},
{
"authors": "S S Rawat; A Bisht; R Nijhawan",
"journal": "IEEE",
"ref_id": "b328",
"title": "A deep learning based cnn framework approach for plankton classification",
"year": "2019"
},
{
"authors": "J Redmon; S Divvala; R Girshick; A Farhadi",
"journal": "",
"ref_id": "b329",
"title": "You Only Look Once: Unified, Real-Time Object Detection",
"year": "2016"
},
{
"authors": "J Redmon; A Farhadi",
"journal": "",
"ref_id": "b330",
"title": "YOLOv3: An incremental improvement",
"year": "2018"
},
{
"authors": "T H Reiss",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI)",
"ref_id": "b331",
"title": "The revised fundamental theorem of moment invariants",
"year": "1991"
},
{
"authors": "S Ren; K He; R Girshick; J Sun",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI)",
"ref_id": "b332",
"title": "Faster R-CNN: Towards realtime object detection with region proposal networks",
"year": "2017"
},
{
"authors": "D Rivas-Villar; J Rouco; R Carballeira; M G Penedo; J Novo",
"journal": "Computer Methods and Programs in Biomedicine",
"ref_id": "b333",
"title": "Fully automatic detection and classification of phytoplankton specimens in digital microscopy images",
"year": "2021"
},
{
"authors": "K Rodenacker; B Hense; U Jütting; P Gais",
"journal": "Microscopy Research and Technique",
"ref_id": "b334",
"title": "Automatic analysis of aqueous specimens for phytoplankton structure recognition and population estimation",
"year": "2006"
},
{
"authors": "F C M Rodrigues; N S Hirata; A A Abello; T Leandro; D La Cruz; R M Lopes; R Hirata",
"journal": "",
"ref_id": "b335",
"title": "Evaluation of transfer learning scenarios in plankton image classification",
"year": "2018"
},
{
"authors": "E Salvesen",
"journal": "",
"ref_id": "b336",
"title": "Unsupervised methods for in-situ classification of plankton taxa",
"year": "2021"
},
{
"authors": "E Salvesen; A Saad; A Stahl",
"journal": "IEEE",
"ref_id": "b337",
"title": "Robust methods of unsupervised clustering to discover new planktonic species in-situ",
"year": "2020"
},
{
"authors": "C Sánchez; G Cristóbal; G Bueno",
"journal": "PeerJ",
"ref_id": "b338",
"title": "Diatom identification including life cycle stages through morphological and texture descriptors",
"year": "2019"
},
{
"authors": "C Sánchez; N Vállez; G Bueno; G Cristóbal",
"journal": "Springer",
"ref_id": "b339",
"title": "Diatom classification including morphological adaptations using cnns",
"year": "2019"
},
{
"authors": "M Sandler; A Howard; M Zhu; A Zhmoginov; L C Chen",
"journal": "",
"ref_id": "b340",
"title": "Mobilenetv2: Inverted residuals and linear bottlenecks",
"year": "2018"
},
{
"authors": "R Scherrer; R Govan; T Quiniou; T Jauffrais; H Lemonnier; S Bonnet; N Selmaoui-Folcher",
"journal": "",
"ref_id": "b341",
"title": "Automatic plankton detection and classification on raw hologram with a single deep learning architecture",
"year": "2021"
},
{
"authors": "L Schmarje; J Brünger; M Santarossa; S M Schröder; R Kiko; R Koch",
"journal": "Sensors",
"ref_id": "b342",
"title": "Fuzzy Overclustering: Semi-Supervised Classification of Fuzzy Labels with Overclustering and Inverse Cross-Entropy",
"year": "2021"
},
{
"authors": "T Schoening; J M Durden; C Faber; J Felden; K Heger; H J T Hoving; R Kiko; K Köser; C Krämmer; T Kwasnitschka; K O Möller; D Nakath; A Naß; T W Nattkemper; A Purser; M Zurowietz",
"journal": "Scientific Data",
"ref_id": "b343",
"title": "Making marine image data FAIR",
"year": "2022"
},
{
"authors": "S M Schröder; R Kiko",
"journal": "Sensors",
"ref_id": "b344",
"title": "Assessing representation learning and clustering algorithms for computer-assisted image annotation-simulating and benchmarking morphocluster",
"year": "2022"
},
{
"authors": "S M Schröder; R Kiko; J O Irisson; R Koch",
"journal": "Springer",
"ref_id": "b345",
"title": "Low-shot learning of plankton categories",
"year": "2018"
},
{
"authors": "S M Schröder; R Kiko; R Koch",
"journal": "Sensors",
"ref_id": "b346",
"title": "Morphocluster: Efficient annotation of plankton images by clustering",
"year": "2020"
},
{
"authors": "J Schulz; K Barz; P Ayon; A Luedtke; O Zielinski; D Mengedoht; H J Hirche",
"journal": "Journal of the European Optical Society-Rapid Publications",
"ref_id": "b347",
"title": "Imaging of plankton specimens with the lightframe on-sight keyspecies investigation (LOKI) system",
"year": "2010"
},
{
"authors": "K Schulze; U M Tillich; T Dandekar; M Frohme",
"journal": "BMC Bioinformatics",
"ref_id": "b348",
"title": "Planktovision-an automated analysis system for the identification of phytoplankton",
"year": "2013"
},
{
"authors": "R R Selvaraju; M Cogswell; A Das; R Vedantam; D Parikh; D Batra",
"journal": "",
"ref_id": "b349",
"title": "Grad-cam: Visual explanations from deep networks via gradient-based localization",
"year": "2017"
},
{
"authors": "S Shan; W Zhang; X Wang; M Tong",
"journal": "IEEE",
"ref_id": "b350",
"title": "Automated red tide algae recognition by the color microscopic image",
"year": "2020"
},
{
"authors": "L Shao; F Zhu; X Li",
"journal": "IEEE Transactions on Neural Networks and Learning Systems",
"ref_id": "b351",
"title": "Transfer learning for visual categorization: A survey",
"year": "2014"
},
{
"authors": "C K Sieracki; M E Sieracki; C S Yentsch",
"journal": "Marine Ecology Progress Series",
"ref_id": "b352",
"title": "An imaging-in-flow system for automated analysis of marine microplankton",
"year": "1998"
},
{
"authors": "K Simonyan; A Zisserman",
"journal": "",
"ref_id": "b353",
"title": "Very deep convolutional networks for large-scale image recognition",
"year": "2014"
},
{
"authors": "Y Soh; J Song; Y Hae",
"journal": "Journal of the Institute of Convergence Signal Processing",
"ref_id": "b354",
"title": "Multiple plankton detection and recognition in microscopic images with homogeneous clumping and heterogeneous interspersion",
"year": "2018"
},
{
"authors": "G A Solano; P Gasmen; E J Marquez",
"journal": "",
"ref_id": "b355",
"title": "Radiolarian classification decision support using supervised and unsupervised learning approaches",
"year": "2018"
},
{
"authors": "A Solow; C Davis; Q Hu",
"journal": "Marine Ecology Progress Series",
"ref_id": "b356",
"title": "Estimating the taxonomic composition of a sample when individuals are classified with error",
"year": "2001"
},
{
"authors": "H Song; S R Mehdi; H Huang; K Shahani; Y Zhang; K Raza; M A Khan",
"journal": "International Journal of Advanced Computer Science and Applications",
"ref_id": "b357",
"title": "Classification of freshwater zooplankton by pre-trained convolutional neural network in underwater microscopy",
"year": "2020"
},
{
"authors": "H M Sosik; R J Olson",
"journal": "Limnology and Oceanography: Methods",
"ref_id": "b358",
"title": "Automated taxonomic classification of phytoplankton sampled with imaging-in-flow cytometry",
"year": "2007"
},
{
"authors": "H M Sosik; E E Peacock; E F Brownlee",
"journal": "",
"ref_id": "b359",
"title": "WHOI-plankton: Annotated plankton images -dataset for developing and evaluating classification methods",
"year": "2021"
},
{
"authors": "X Sun; H Xv; J Dong; H Zhou; C Chen; Q Li",
"journal": "IEEE Transactions on Industrial Electronics",
"ref_id": "b360",
"title": "Few-shot learning for domain-specific fine-grained image classification",
"year": "2020"
},
{
"authors": "C Szegedy; S Ioffe; V Vanhoucke; A Alemi",
"journal": "",
"ref_id": "b361",
"title": "Inception-v4, inception-resnet and the impact of residual connections on learning",
"year": "2017"
},
{
"authors": "C Szegedy; W Liu; Y Jia; P Sermanet; S Reed; D Anguelov; D Erhan; V Vanhoucke; A Rabinovich",
"journal": "",
"ref_id": "b362",
"title": "Going deeper with convolutions",
"year": "2015"
},
{
"authors": "C Szegedy; V Vanhoucke; S Ioffe; J Shlens; Z Wojna",
"journal": "",
"ref_id": "b363",
"title": "Rethinking the inception architecture for computer vision",
"year": "2016"
},
{
"authors": "M Tan; Q Le",
"journal": "",
"ref_id": "b364",
"title": "Efficientnet: Rethinking model scaling for convolutional neural networks",
"year": "2019"
},
{
"authors": "S Tan; F Zhang; Q Huang; S Wang",
"journal": "Optik",
"ref_id": "b365",
"title": "Measuring and calculating geometrical parameters of marine plankton using digital laser holographic imaging",
"year": "2014"
},
{
"authors": "F H K D S Tanaka; C Aranha",
"journal": "",
"ref_id": "b366",
"title": "Data augmentation using gans",
"year": "2019"
},
{
"authors": "X Tang; F Lin; S Samson; A Remsen",
"journal": "IEEE Journal of Oceanic Engineering",
"ref_id": "b367",
"title": "Binary plankton image classification",
"year": "2006"
},
{
"authors": "X Tang; W K Stewart; H Huang; S M Gallager; C S Davis; L Vincent; M Marra",
"journal": "Artificial Intelligence Review",
"ref_id": "b368",
"title": "Automatic plankton image recognition",
"year": "1998"
},
{
"authors": "A L Teigen; A Saad; A Stahl",
"journal": "IEEE",
"ref_id": "b369",
"title": "Leveraging similarity metrics to in-situ discover planktonic interspecies variations or mutations",
"year": "2020"
},
{
"authors": "J Teuwen; N Moriakov",
"journal": "Elsevier",
"ref_id": "b370",
"title": "Handbook of Medical Image Computing and Computer Assisted Intervention",
"year": "2020"
},
{
"authors": "S U Thiel; R J Wiltshire; L J Davies",
"journal": "Water Research",
"ref_id": "b371",
"title": "Automated object recognition of blue-green algae for measuring water quality-a preliminary study",
"year": "1995"
},
{
"authors": "K Tountas; D A Pados; M J Medley",
"journal": "",
"ref_id": "b372",
"title": "Conformity evaluation and l1-norm principal-component analysis of tensor data",
"year": "2019"
},
{
"authors": "G Tsechpenakis; C M Guigand; R K Cowen",
"journal": "IEEE",
"ref_id": "b373",
"title": "Image analysis techniques to accompany a new in situ ichthyoplankton imaging system",
"year": "2007"
},
{
"authors": "N Vallez; G Bueno; O Deniz; S Blanco",
"journal": "Computer Methods and Programs in Biomedicine",
"ref_id": "b374",
"title": "Diffeomorphic transforms for data augmentation of highly variable shape and texture objects",
"year": "2022"
},
{
"authors": "N Van Noord; E Postma",
"journal": "Pattern Recognition",
"ref_id": "b375",
"title": "Learning scale-variant and scale-invariant features for deep image classification",
"year": "2017"
},
{
"authors": "K Varma; L Nyman; K Tountas; G Sklivanitis; A R Nayak; D A Pados",
"journal": "IEEE",
"ref_id": "b376",
"title": "Autonomous plankton classification from reconstructed holographic imagery by l1-pca-assisted convolutional neural networks",
"year": "2020"
},
{
"authors": "A Venkataramanan; M Laviale; C Figus; P Usseglio-Polatera; C Pradalier",
"journal": "Springer",
"ref_id": "b377",
"title": "Tackling inter-class similarity and intra-class variance for microscopic image-based classification",
"year": "2021"
},
{
"authors": "A Verikas; A Gelzinis; M Bacauskiene; I Olenina; S Olenin; E Vaiciukynas",
"journal": "Pattern Recognition",
"ref_id": "b378",
"title": "Phase congruency-based detection of circular objects applied to analysis of phytoplankton images",
"year": "2012"
},
{
"authors": "A Verikas; A Gelzinis; M Bacauskiene; I Olenina; E Vaiciukynas",
"journal": "IEEE Journal of Oceanic Engineering",
"ref_id": "b379",
"title": "An Integrated Approach to Analysis of Phytoplankton Images",
"year": "2015"
},
{
"authors": "G Wacquet; A Lefebvre; C Blondel; A Louchart; P Grosjean; N Neaud-Masson; C Belin; L F Artigas",
"journal": "",
"ref_id": "b380",
"title": "Combination of machine learning methodologies and imaging-in-flow systems for the automated detection of harmful algae",
"year": "2018"
},
{
"authors": "N L Walcutt; B Knörlein; I Cetinić; Z Ljubesic; S Bosak; T Sgouros; A L Montalbano; A Neeley; S Menden-Deuer; M M Omand",
"journal": "Limnology and Oceanography: Methods",
"ref_id": "b381",
"title": "Assessment of holographic microscopy for quantifying marine particle size and concentration",
"year": "2020"
},
{
"authors": "J L Walker; E C Orenstein",
"journal": "",
"ref_id": "b382",
"title": "Improving rare-class recognition of marine plankton with hard negative mining",
"year": "2021"
},
{
"authors": "R F Walker; K Ishikawa; M Kumagai",
"journal": "Journal of Microbiological Methods",
"ref_id": "b383",
"title": "Fluorescence-assisted image analysis of freshwater microalgae",
"year": "2002"
},
{
"authors": "C Wang; Z Yu; H Zheng; N Wang; B Zheng",
"journal": "IEEE",
"ref_id": "b384",
"title": "Cgan-plankton: towards large-scale imbalanced class generation and fine-grained classification",
"year": "2017"
},
{
"authors": "C Wang; X Zheng; C Guo; Z Yu; J Yu; H Zheng; B Zheng",
"journal": "IEEE",
"ref_id": "b385",
"title": "Transferred parallel convolutional neural network for large imbalanced plankton database classification",
"year": "2018"
},
{
"authors": "J Wang; C Lan; C Liu; Y Ouyang; T Qin; W Lu; Y Chen; W Zeng; P Yu",
"journal": "IEEE Transactions on Knowledge and Data Engineering",
"ref_id": "b386",
"title": "Generalizing to unseen domains: A survey on domain generalization",
"year": "2022"
},
{
"authors": "J Wang; C Tang; J Li",
"journal": "IEEE",
"ref_id": "b387",
"title": "Towards real-time analysis of marine phytoplankton images sampled at high frame rate by a yolox-based object detection algorithm",
"year": "2022"
},
{
"authors": "K Wang; D Zhang; Y Li; R Zhang; L Lin",
"journal": "IEEE Transactions on Circuits and Systems for Video Technology",
"ref_id": "b388",
"title": "Cost-effective active learning for deep image classification",
"year": "2016"
},
{
"authors": "J Watson",
"journal": "",
"ref_id": "b389",
"title": "High-resolution underwater holographic imaging",
"year": "2018"
},
{
"authors": "L Wei; S Xiaopan; F Heydari",
"journal": "Mathematical Problems in Engineering",
"ref_id": "b390",
"title": "Microalgae classification using improved metaheuristic algorithm",
"year": "2022"
},
{
"authors": "K Weiss; T M Khoshgoftaar; D Wang",
"journal": "Journal of Big Data",
"ref_id": "b391",
"title": "A survey of transfer learning",
"year": "2016"
},
{
"authors": "Y Wen; K Zhang; Z Li; Y Qiao",
"journal": "",
"ref_id": "b392",
"title": "A discriminative feature learning approach for deep face recognition",
"year": "2016"
},
{
"authors": "G Wilson; D J Cook",
"journal": "ACM Transactions on Intelligent Systems and Technology",
"ref_id": "b393",
"title": "A survey of unsupervised deep domain adaptation",
"year": "2020"
},
{
"authors": "M F Wu; H T Sheu",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b394",
"title": "",
"year": "1998"
},
{
"authors": "Q Xiaoyan",
"journal": "IEEE Access",
"ref_id": "b395",
"title": "Research on imbalanced microscopic image classification of harmful algae",
"year": "2020"
},
{
"authors": "L Xu; L Xu; Y Chen; Y Zhang; J Yang",
"journal": "ACS ES&T Water",
"ref_id": "b396",
"title": "Accurate Classification of Algae Using Deep Convolutional Neural Network with a Small Database",
"year": "2022"
},
{
"authors": "J Yan; X Li; Z Cui",
"journal": "Springer",
"ref_id": "b397",
"title": "A more efficient cnn architecture for plankton classification",
"year": "2017"
},
{
"authors": "L Ye; C Y Chang; C H Hsieh",
"journal": "Marine Ecology Progress Series",
"ref_id": "b398",
"title": "Bayesian model for semi-automated zooplankton classification with predictive confidence and rapid category aggregation",
"year": "2011"
},
{
"authors": "M Ye; J Shen; G Lin; T Xiang; L Shao; S C Hoi",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI)",
"ref_id": "b399",
"title": "Deep learning for person re-identification: A survey and outlook",
"year": "2021"
},
{
"authors": "E M Zetsche; A El Mallahi; F Dubois; C Yourassowsky; J C Kromkamp; F J Meysman",
"journal": "Limnology and Oceanography: Methods",
"ref_id": "b400",
"title": "Imaging-in-flow: Digital holographic microscopy as a novel tool to detect and classify nanoplanktonic organisms",
"year": "2014"
},
{
"authors": "J Zhang; C Li; Y Yin; J Zhang; M Grzegorzek",
"journal": "Artificial Intelligence Review",
"ref_id": "b401",
"title": "Applications of artificial neural networks in microorganism image analysis: a comprehensive review from conventional multilayer perceptron to popular convolutional neural network and potential visual transformer",
"year": "2022"
},
{
"authors": "Y Zhang; Y Lu; H Wang; P Chen; R Liang",
"journal": "Optics & Laser Technology",
"ref_id": "b402",
"title": "Automatic classification of marine plankton with digital holography using convolutional neural network",
"year": "2021"
},
{
"authors": "F Zhao; F Lin; H S Seah",
"journal": "IEEE",
"ref_id": "b403",
"title": "Bagging based plankton image classification",
"year": "2009"
},
{
"authors": "F Zhao; F Lin; H S Seah",
"journal": "Neurocomputing",
"ref_id": "b404",
"title": "Binary sipper plankton image classification using random subspace",
"year": "2010"
},
{
"authors": "F Zhao; X Tang; F Lin; S Samson; A Remsen",
"journal": "IEEE",
"ref_id": "b405",
"title": "Binary plankton image classification using random subspace",
"year": "2005"
},
{
"authors": "A Zheng; M Wang",
"journal": "",
"ref_id": "b406",
"title": "Convolutional neural networksbased plankton image classification system",
"year": "2015"
},
{
"authors": "H Zheng; R Wang; Z Yu; N Wang; Z Gu; B Zheng",
"journal": "BMC Bioinformatics",
"ref_id": "b407",
"title": "Automatic plankton image classification combining multiple view features via multiple kernel learning",
"year": "2017"
},
{
"authors": "K Zhou; Z Liu; Y Qiao; T Xiang; C C Loy",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b408",
"title": "Domain generalization: A survey",
"year": "2022"
},
{
"authors": "J Y Zhu; T Park; P Isola; A A Efros",
"journal": "",
"ref_id": "b409",
"title": "Unpaired image-to-image translation using cycle-consistent adversarial networks",
"year": "2017"
},
{
"authors": "T G Zimmerman; V P Pastore; S K Biswas; S Bianco",
"journal": "",
"ref_id": "b410",
"title": "Embedded system to detect, track and classify plankton using a lensless video microscope",
"year": "2020"
},
{
"authors": "B Zoph; V Vasudevan; J Shlens; Q V Le",
"journal": "",
"ref_id": "b411",
"title": "Learning transferable architectures for scalable image recognition",
"year": "2018"
},
{
"authors": "J Ärje; J Raitoharju; A Iosifidis; V Tirronen; K Meissner; M Gabbouj; S Kiranyaz; S Kärkkäinen",
"journal": "Signal Processing: Image Communication",
"ref_id": "b412",
"title": "Human experts vs. machines in taxa recognition",
"year": "2020"
}
] | [] |
10.18653/v1/2020.blackboxnlp-1.14 | [
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b26",
"b21",
"b9",
"b2",
"b8",
"b9",
"b19",
"b18",
"b22"
],
"table_ref": [],
"text": "With neural machine translation systems reaching an overall satisfactory quality, alleviating those rare but severe translation pathologies that undermine user trust becomes very important. These pathologies include hallucinations (translations containing information completely unrelated to the input) and omissions (translations that do not include some of the information of the input). While understanding hallucinations is receiving increasing attention (Raunak et al., 2021;Müller and Sennrich, 2021;Zhou et al., 2021;Guerreiro et al., 2023;Dale et al., 2023;Guerreiro et al., 2022), progress in this direction is hampered by the lack of annotated data. To the best of our knowledge, previous datasets are limited to German-English data with sentence-level annotations of hallucinations and omissions (Guerreiro et al., 2023) and Chinese-English data with token-level hallucination labels (Zhou et al., 2021). Previously available general-purpose quality assessments, such as direct assessment (DA) Gra- (Lommel et al., 2014), or XSTS (Licht et al., 2022) do not seem suitable since they do not distinguish between hallucinations, omissions and other translation errors. In this work, we aim to address this limitation.\nIdeally, an evaluation dataset for hallucination/omission detection should satisfy several conditions: (i) data has to cover a broad range of languages with varying resource levels and scripts, (ii) translations should be generated naturally, (iii) the models that produced the translations have to be available, and (iv) considered modeling paradigms have to cover several approaches (i.e., encoder-decoder vs decoder-only, single language pair vs multilingual, etc.). The first point is important because the best-performing detectors are different for high-and low-resource settings, and general conclusions cannot be made based on a single language pair (Section 4). Secondly, translations have to be generated naturally as opposed to using specifically developed perturbations of models and/or data because conclusions for the latter might not transfer for detection of natural hallucinations (Section 6). Thirdly, the corresponding model should be released along with the translations to allow evaluating \"internal\" detection methods. Fi-arXiv:2305.11746v2 [cs.CL] 6 Dec 2023 nally, in the most ideal setting, various models are needed to test whether the detection methods transfer between modeling approaches.\nWhile satisfying all the desiderata is very challenging, we can satisfy all but last by focusing on the state-of-the-art multilingual NLLB-200 model (NLLB Team et al., 2022). In addition to covering a broad range of languages and being publicly available along with its training data, NLLB is widely recognized1 and is likely to stay the stateof-the-art for the foreseeable future. For this model, we choose 18 language pairs that include highand low-resource languages, as well as a zero-shot pair (Figure 1). We develop rigorous annotation guidelines for identifying full and partial hallucinations and omissions and use these guidelines for manual annotation of translations in all 18 directions. The resulting dataset contains fine-grained sentence-level and token-level annotations.\nWe highlight the importance of our dataset by making several valuable observations that would not be possible otherwise. For example, we find that for low-resource directions, internal methods perform much better than external methods that substantially fail. When analyzing performance of a range of recently introduced pathology detection methods, we see that some of the previous results do not transfer across languages. As another example, we show that relying on attention to make conclusions about translation quality is very fragile. Finally, we introduce some detection tasks (e.g., token-level omission detection) that became possible only with our data. We believe our work opens the door for reliable and accessible research on detecting and analyzing translation pathologies as well as understanding their causes.\nOverall, we:\n• release a dataset with fine-grained professional annotations of hallucinations and omissions for 18 language pairs 2 ;\n• analyze previous sentence-level detectors and find that e.g. (i) for low-resource settings, model internal characteristics work best, (ii) attention is very fragile when used to judge translation quality, among other observations;\n• introduce word-level pathology detection tasks along with the baselines."
},
{
"figure_ref": [],
"heading": "Dataset Creation",
"publication_ref": [],
"table_ref": [],
"text": "The steps to create the dataset were (i) choosing the language pairs, (ii) gathering data for annotation, (iii) developing annotation guidelines and qualification sets, (iv) manual annotation, (v) postprocessing. Here, we explain these steps."
},
{
"figure_ref": [
"fig_0"
],
"heading": "Selection of Languages",
"publication_ref": [],
"table_ref": [],
"text": "We optimized the language selection in order to cover (i) different resource levels and (ii) a variety of language families and scripts. Among the languages available in NLLB-200, we include 5 highresource language pairs (Arabic, Mandarin Chinese, German, Russian, and Spanish paired with English), 3 low-resource language pairs (Kashmiri, Manipuri, and Yoruba paired with English) and a zero-shot pair (Spanish-Yoruba). 3 We consider all language pairs in both directions which gives us 18 translation directions summarized in Figure 1."
},
{
"figure_ref": [],
"heading": "Gathering Data for Annotation",
"publication_ref": [
"b22",
"b12",
"b9",
"b24",
"b27",
"b2",
"b10",
"b4",
"b6"
],
"table_ref": [],
"text": "Since strong NLLB models rarely generate hallucinations and omissions, getting translations that are likely to contain these types of errors is not straightforward. To gather these translations, we developed a multi-step procedure where we first choose data to generate translations and then choose a subset of the generated translations for annotation.\nChoosing data for translation. Since we expect that the NLLB model will not hallucinate much when handling high-resource languages, in addition to clean in-domain data, we use noisier outof-domain sources. Overall, the data we use to generate translations is as follows:\n• in-domain: FLORES-200 development set (NLLB Team et al., 2022);\n• out-of-domain: Jigsaw toxicity detection competition corpora (Jigsaw, 2020) 4 -for English, Russian and Spanish; comments from Wikipedia discussion pages5 -for Chinese, Arabic and German. The Jigsaw corpora were extracted from Wikipedia talk pages, so the distributions of these texts are rather similar.\nWe translated these texts with the 600M distilled NLLB model6 following the standard setting (beam size 5, forbidden generation of the <UNK> token, forbidden repetition of 4-grams, limiting the translation length to 3•len(source)+5 tokens.\nChoosing translations for annotation. To find potentially pathological translations, we scored sentence pairs by multiple metrics that were used as hallucination detectors in previous works. Specifically, we used some methods from Guerreiro et al. (2023): ChrF++ (Popović, 2017), reference-based COMET7 and referenceless COMET-QE (Rei et al., 2020), and Seq-Logprob (their best detector). We also used some methods introduced in Dale et al. (2023): cosine similarity coming from LASER3 (Heffernan et al., 2022) and LaBSE (Feng et al., 2022), a bidirectional XNLI score, and ALTI+ source contributions (Ferrando et al., 2022).\nFor each translation direction and data source, we selected sentence pairs with 3 strategies:\n• Sample uniformly -to preserve data diversity and non-hallucinated examples;\n• Sample favoring potentially pathological translations (with the probabilities proportional to the quantiles of the detectors);\n• Pick the worst according to the detectors -to increase the chance of hallucinations.\nAppendix A describes the amount of data selected by these strategies for all directions."
},
{
"figure_ref": [
"fig_1",
"fig_1"
],
"heading": "Guidelines and Qualification Tests",
"publication_ref": [
"b26",
"b9"
],
"table_ref": [],
"text": "To ensure annotation quality, guidelines and qualification tests were prepared by professional linguists.\nAnnotation guidelines. These guidelines define:\n• the notion of hallucinations and omissions;\n• the hallucination vs mistranslation distinction;\n• hallucination/omission severity levels.\nFigure 2 summarizes the resulting guidelines. Note that distinguishing hallucinations from other translation errors is one of the known difficulties when dealing with hallucinations (Raunak et al., 2021;Guerreiro et al., 2023). In our guidelines, a token is referred to as hallucinated if there is no corresponding token in the source (Figure 2). For all pathologies, linguists provide positive and negative examples in diverse languages. Additionally, we ask the annotators to mark if a translation is incomprehensible, i.e. whether the text is garbled or in another language. These translations are then discarded. 8Qualification tests and postprocessing. For annotation, we choose professional translators (2 for each language) who are allowed to annotate our data only after passing a specifically developed qualification test. More details on this test and postprocessing steps can be found in Appendix A."
},
{
"figure_ref": [
"fig_2",
"fig_0"
],
"heading": "Dataset Description",
"publication_ref": [],
"table_ref": [],
"text": "Annotation format. The resulting data contains the source text and its translation, along with the word-level and sentence-level annotations of omissions and hallucinations. Figure 3 shows examples of annotated translations from our dataset.\nOverall statistics. Figure 1 shows the proportions of hallucinations and omissions in the data (translations with both hallucinations and omissions are referred to as hallucinations). Overall, all directions have at least 3% translations with hallucinations (1% full) and 17% with omissions (5% full). Most of the full hallucinations are also labelled as full omissions, and vice versa.\nDifferences between resource levels. From Figure 1 we see that, not surprisingly, high-resource language pairs hallucinate less than low-resource. A less anticipated difference between high-and low-resource settings is seen when looking within each language pair. In high-resource settings, translating to English leads to more hallucinations than translating from English. Differently, for lowresource pairs, translation from English has higher hallucinatory rates than translation to English for the same language pair. This might inspire future work to analyze the role of English data in the multilingual NLLB model. Finally, while for the zeroshot pair one might expect more pathologies, this is not what we see: results for the zero-shot pair are comparable to those for low-resource languages."
},
{
"figure_ref": [],
"heading": "Sentence-Level Detection",
"publication_ref": [
"b15",
"b20",
"b26",
"b9",
"b2",
"b8",
"b34"
],
"table_ref": [],
"text": "Detecting pathologies at the sentence level is the task of flagging a whole translation as pathological or not. This is the standard definition of e.g. the hallucination detection task (Lee et al., 2019;Müller et al., 2020;Raunak et al., 2021;Guerreiro et al., 2023;Dale et al., 2023;Guerreiro et al., 2022;Xu et al., 2023). Such sentence-level pathology detection (instead of flagging individual erroneous tokens) is an integral part of hybrid pipelines when a machine-generated translation is first passed to a quality estimation system and then, if needed, is corrected by human translators.\nDetection tasks. For our dataset, we define three sentence-level detection tasks:\n• hallucination detection: same as in previous work mentioned above;\n• omission detection: detecting translations with omissions on a hallucination-free subset. The latter is to disentangle omissions from a more severe hallucination pathology;\n• pathology detection: detecting translations that are either hallucinations or omissions.\nEvaluation methodology. We evaluate the ability of a detector to rank more severe pathologies higher (e.g., full hallucinations higher than partial, any hallucinations higher than non-hallucinations, etc). For this, we use an adaptation of the binary ROC AUC score for several classes. Formally, we subtract from the perfect score, i.e. 1, the percentage of incorrectly ranked pairs of sentences with different labels. For two classes, this metric is equivalent to the ROC AUC score. We compute the metrics for each translation direction separately."
},
{
"figure_ref": [
"fig_3",
"fig_3"
],
"heading": "Detection Methods",
"publication_ref": [
"b2",
"b6",
"b2",
"b27",
"b2",
"b10",
"b4",
"b28",
"b3",
"b8"
],
"table_ref": [],
"text": "Detection metrics can be either internal, i.e. relying only on the information from the model that generated the inspected translation, or external, i.e. using external models. We use the best detectors from several recent works, along with some of their modifications we propose in this work. The metrics are summarized in Figure 4. Internal methods. For internal models, we use the best method from Guerreiro et al. ( 2023) (sequence log-probability) and the best internal method from (Dale et al., 2023), ALTI. ALTI (Ferrando et al., 2022) is an attribution method that evaluates token contributions to generated translations. For hallucination detection, Dale et al. (2023) evaluate how, on average, the prediction of each target token is based on the source. Here, mostly to detect omissions, we propose a different variant ALTI T that computes how much, on average, each source token was used to generate the translation. Intuitively, if many source tokens are not used during generation, the translation is likely to not contain some information. The difference between the two versions of ALTI is illustrated in Figure 4.\nExternal methods. For external methods, we use the state-of-the-art quality estimation system COMET-QE (Rei et al., 2020) and sentence similarity measures proposed in Dale et al. (2023). The latter are cosine similarities coming from LASER3 (Heffernan et al., 2022), LaBSE (Feng et al., 2022), and a bidirectional XNLI score. Finally, we evaluate a translation quality estimation method from Seamless Communication et al. (2023), BLASER 2.0-QE, built on top of SONAR sentence embeddings (Duquenne et al., 2023).\nGray-area method. Finally, we also use a recent optimal transport-based measure evaluating the abnormality of the attention distribution compared to those of good translations (Guerreiro et al., 2022). While this method uses internal characteristics, it requires external data, i.e. a large collection of attention maps for \"good\" translations, which can be hard to obtain for low-resource settings.9 "
},
{
"figure_ref": [
"fig_4"
],
"heading": "Experimental Results",
"publication_ref": [
"b2",
"b2",
"b8",
"b9",
"b11",
"b29",
"b33",
"b0",
"b8",
"b8"
],
"table_ref": [],
"text": "The detection scores for hallucinations and omissions are shown in Figure 5. The scores for detecting all pathologies are given in Appendix C.\nHigh-resource: much easier to handle. We can see that it is much easier to detect pathologies in high-resource settings: the gap between the best scores for high-and low-resource directions is rather big (e.g., for halucinations, 0.89 vs 0.79). Note also that for high-resource language pairs, both internal and external methods perform quite well (e.g., Seq-Logprob and LaBSE for hallucinations; XNLI, LaBSE and ALTI T for omissions).\nLow-resource: internal methods take the lead. In low-resource settings, external methods drop substantially with some of them losing sensibility. For example, high-performing XNLI drops close to chance for all pathologies. Overall, most external models (with the exception of massively multilingual BLASER) are unlikely to be competent for low-resource directions as they do not observe enough relevant data during training. While previous work already expressed this concern and advocated focusing on internal methods (Dale et al., 2023), without our dataset, verifying this was not possible.\nHallucinations: Seq-Logprob is the most stable. After it turned out that the standard sequence logprobability is more informative for hallucination detection than the heuristics introduced earlier (Guerreiro et al., 2023), a few recent works reported improvements over Seq-Logprob: ALTI and LaBSE in Dale et al. (2023) and Attn-OT in Guerreiro et al. (2022). We see, however, that on average, Seq-Logprob is still the most robust accross translation directions. This discrepancy comes from the fact that those works made conclusions based on a sin- gle language pair. This highlights the importance of our dataset enabling large-scale evaluation over several language pairs. BLASER-QE: a SOTA hallucination detector. On average, BLASER-QE performs on par with the best hallucination detection methods high-resource directions, and outperforms them on low-resource directions. Apparently, fine-tuning massively multilingual sentence encoders to predict semantic similarity is a good recipe for hallucination detectors.\nAttention-based method: close to chance. For hallucinations, Attn-OT detecting attention anomaly is an outlier and performs close to chance. 1011 While previous work already showed that relying on attention patterns to make conclusions about translation quality can be fragile (Guerreiro et al., 2023), results with our dataset highlight 10 We tried all the versions of the method from Guerreiro et al. ( 2022) as well as some additional modifications to improve its results. For the dataset from Guerreiro et al.\n(2022), we managed to reproduce their results. For out dataset, we show the best method variant in the main text and the rest (along with the implementation details) in Appendix B.\n11 For NLLB, poor performance of this method could be attributed to the overall large attention to the EOS token. We tried removing this token from the optimal transport computation, but this did not improve the results significantly. this even further. This points to a larger debate on the distinction between attention and attribution and the consequences of mistaking one for the other (Jain and Wallace, 2019;Serrano and Smith, 2019;Wiegreffe and Pinter, 2019;Bastings and Filippova, 2020). While Attn-OT was introduced as a way to evaluate detachment from the source sequence (Guerreiro et al., 2022), we see that implementing this intuition with attention instead of attribution (as in e.g. ALTI) leads to varying results: from high performance in Guerreiro et al. (2022) to near-random performance in our experiments.\nOmissions: internal ALTI T performs best. For detecting omissions among non-hallucinations, the quality is generally worse than for hallucinations. The best method is ALTI T which confirms our intuition that if, according to token contributions, some source words are not used for translation, a translation is likely to omit relevant information. LaBSE, XNLI and BLASER-QE also perform well for high-resource languages but, similar to hallucination detection, are worse than internal methods for low-resource. Finally, while Attn-OT does not seem to identify hallucinations, it is sensible for omissions. "
},
{
"figure_ref": [],
"heading": "Word-Level Detection",
"publication_ref": [
"b31",
"b25"
],
"table_ref": [],
"text": "In contrast to sentence-level detection, detecting pathologies at the word level received much less attention. In terms of both available data and detectors, previous attempts were rather limited (Zhou et al., 2021;Vamvas and Sennrich, 2022). Here, we want to facilitate future research in this direction.\nDetection tasks. We define two detection tasks:\n• hallucination detection: for each translation word, predict whether it is hallucinated;\n• omission detection: for each source word, predict whether it is omitted from the translation.\nSegmentation. We segment texts using Sacre-BLEU tokenizer (Post, 2018). For Chinese, it interprets each Chinese character as an individual word. For other languages, it applies regex-based tokenization based on spaces and punctuation.\nEvaluation methodology. For these binary classification tasks, we use the ROC AUC score. Since models operate at the token level, we make predictions for tokens and not words. If a word is segmented into several tokens, we assign the worst score among its tokens (i.e., if one of a word's tokens is hallucinated, the entire word is hallucinated)."
},
{
"figure_ref": [
"fig_5"
],
"heading": "Detection Methods",
"publication_ref": [],
"table_ref": [],
"text": "To the best of our knowledge, there are no publicly available models for token-level detection of hallucinations or omissions that could be easily adapted to handle the language pairs in our dataset. Applying previous approach by Zhou et al. (2021), i.e. training a specialized model (or several language pair-specific models) on synthetic data, would be very demanding in terms of engineering and research effort to work well on diverse resource levels. Therefore, we suggest starting with internal methods and their combinations.\nInternal methods. For internal methods, we rely on the same methods that were previously used: model log-probability and ALTI (Figure 6). We use two types of log-probability: the standard and its difference with the unconditioned log-probability for the same token (i.e., when conditioning on an empty source sentence). Intuitively, the latter is one more way of measuring whether the model uses the source or relies more on its language model. For ALTI, we use both the total source contribution and the maximum contribution among the source tokens. The latter is high when the model is \"focused\" on a specific source token -this might be indicative of the prediction quality. For omissions, we use the same methods with swapped source and target sentences (i.e., ALTI turns into ALTI T ).\nCombination of methods. Apart from the individual methods, we also consider their linear combinations. We use a logistic regression trained using 3-fold group-wise cross-validation (with sentence ids as groups). We train the same feature combination for all languages by maximizing the detection score on the pooled data."
},
{
"figure_ref": [
"fig_6"
],
"heading": "Experimental Results",
"publication_ref": [
"b35",
"b16",
"b13",
"b23",
"b17",
"b5"
],
"table_ref": [],
"text": "The results are shown in Figure 7. Overall, the methods we proposed are reasonable and perform much better than the random baseline.\nToken contributions perform best. We see that for both hallucinations and omissions, token contributions coming from ALTI (or ALTI T for omissions) perform better than the log-probability coming from the model. Note that for hallucinations, this is different from the sentence-level results where Seq-Logprob outperformed ALTI.\nContrastive vs standard log-probability. Another interesting observation is that for hallucination detection in the high-resource setting, contrastive log-probability gives a better signal than the standard log-probability. This indicates that comparing explicitly the changes when dropping some information can be useful. This is not surprising: in a broad sense, our contrastive log-probability is a variant of erasure-based interpretation approaches (Zeiler and Fergus, 2014;Li et al., 2017;Kádár et al., 2017;Poerner et al., 2018;Li et al., 2019). In our case, the erased part is rather large, i.e. the whole source sentence. For such large segments, a similar idea previously appeared in context-aware machine translation when dropping the entire context sentence (Fernandes et al., 2021).\nThe detectors are complementary. Finally, we see that log-probability and token contributions are complementary: for both hallucinations and omissions, combining the features leads to a noticeable improvement in detection quality.\n6 Natural vs \"Artificial\" Pathologies In Appendix D, we compare performance of detection methods between two datasets: (i) our dataset with natural translations and (ii) translations generated with perturbed model. We find that data with translations generated under perturbation has to be used with caution, especially when evaluating pathology detection methods: the conclusions are likely to not transfer to the natural setting."
},
{
"figure_ref": [],
"heading": "Discussion",
"publication_ref": [],
"table_ref": [],
"text": "We saw that some of the internal and external methods can detect hallucinations and omissions with the quality that is much better than nothing, much worse than perfect. But what are the cases in which methods perform well and what can be improved? Flagging correct translations. Examples 1-3 are correct translations, but some of them are flagged as pathological. Example 2 is flagged as an hallucination and omission, probably because the input \":::Jeez.\" is slang and has a wide range of potential meanings. Example 3 is flagged by Seq-Logprob: for the model, this translation may be \"unlikely\" because it is short and consists of a rare word."
},
{
"figure_ref": [],
"heading": "Sentence-Level Detection",
"publication_ref": [
"b7",
"b18",
"b9"
],
"table_ref": [],
"text": "Difficult to detect pathologies. Examples 4-6 show partial hallucinations and omissions that are difficult to detect, either because (in some sense) they resemble a correct translation, or because the translation indeed remains within the range of possible translations despite having these pathologies.\nThis raises a question: what does it really mean to have a hallucinated translation? While our sentence-level labels are fine-grained, the severity of a pathology is defined based on the number of omitted/hallucinated words rather than on the degree of semantic inadequacy of the pathology (similarly to e.g. Guerreiro et al. ( 2023)). This agrees with previous work noting that defining severity of translation errors is not straightforward (Graham et al., 2013;Licht et al., 2022;Guerreiro et al., 2023).\nCorrectly detected pathologies. Examples 7-11 show more severe hallucinations and omissionsthese are detected correctly by at least a few of the considered methods. Many of these pathologies are produced for out-of-distribution inputs: conjunction of several sentences, typos, non-sentence texts (such as dates), and very short or incomplete sentences. Note that e.g. short sentences are nontypical for the NLLB training data. As we see, for such inputs the model is often not confident even for correct translations (see e.g. examples 2 and 3correct but short translations are flagged as pathological). This suggests that these errors might be alleviated by augmenting training data with similar (very short, multi-sentence, etc.) samples."
},
{
"figure_ref": [],
"heading": "Word-Level Detection",
"publication_ref": [],
"table_ref": [],
"text": "In Appendix E.1, we also show examples of wordlevel detection and discuss the behavior of the detection methods. For example, we note that logprobability focuses on the beginnings of sentences while ALTI focuses on the endings."
},
{
"figure_ref": [],
"heading": "Additional Related Work",
"publication_ref": [
"b15",
"b1",
"b21",
"b9",
"b30",
"b32",
"b31"
],
"table_ref": [],
"text": "Except for mentioned above work, previous work on hallucinations in machine translation largely avoided human annotation. To judge whether a translation is hallucinated, they relied on various heuristics or string-based automatic evaluation metrics (Lee et al. (2019); Berard et al. (2019); Müller and Sennrich (2021); Raunak et al. ( 2021)). These, however, were shown not to be indicative of hallucinations (Guerreiro et al., 2023), which highlights the importance of our human-annotated data. For omissions, previous work mostly focused on empty translations (Stahlberg and Byrne, 2019;Vijayakumar et al., 2016) with some work using artificially created undertranslations (Vamvas and Sennrich, 2022). As we saw in Section 6, the latter is unlikely to be helpful when evaluating detection methods."
},
{
"figure_ref": [],
"heading": "Conclusions",
"publication_ref": [],
"table_ref": [],
"text": "We present the first dataset with human-annotated hallucinations and omissions that satisfies several conditions: (i) it covers a broad range of languages with varying resource levels and scripts, (ii) the translations are generated naturally (i.e., without artificial perturbations), (iii) the model producing the translations is publicly available. Through our extensive experiments, we illustrate why each of these conditions is important. Additionally, we make several observations of individual importance. For example, for low-resource directions internal pathology detection methods perform better than most of the external ones, attention is very fragile when used to judge translation quality, among others. We believe our work opens the door for a reliable and accessible research on detecting and analyzing translation pathologies."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [],
"table_ref": [],
"text": "Our experiments are reproducible, and our dataset together with the NLLB model can be widely used to evaluate progress on hallucinations/omissions detection and mitigation. However, all the annotated translations were generated with a single model, and the generalization to other models is yet to be verified.\nThe dataset is rather small and may not cover all possible hallucinations/omissions cases."
},
{
"figure_ref": [],
"heading": "Ethical considerations",
"publication_ref": [],
"table_ref": [],
"text": "The annotations were provided by professionals and they were all paid a fair rate. "
},
{
"figure_ref": [],
"heading": "A Dataset Creation",
"publication_ref": [
"b22"
],
"table_ref": [
"tab_1"
],
"text": "Selecting the data. The 3 sampling strategies described in Section 2.2 were applied in different proportions, depending on what kind of data we had for a particular translation direction. Our released dataset has a field with the sampling strategy labels. The resulting proportions are reported in Table 1.\nQualification tests. The annotators recruited for this project were translators and reviewers who participated in FLORES translation (NLLB Team et al., 2022) or have other professional translation experience. Typically, these annotators are translators with at least two to three years of professional translation experience, usually with domain expertise in journalism, education, social media or marketing. Two annotators are recruited for each language. They are allowed to annotate our data only after passing a specifically developed qualification test. An annotator can fail the test no more than once, in which case they are given an opportunity to receive a detailed feedback and re-do the test. If they do not achieve a passing score of 96% at the second attempt, the vendor is required to find a replacement. Once two annotators are qualified for a given language, one annotator performs the annotations, which are then reviewed by the second annotator.\nOur qualification tests were developed for each of the language directions, and contained 15 items to annotate: 3 full hallucinations, 4 partial hallucinations, 2 word-level hallucinations, 5 mistranslations, and 1 incomprehensible sentence. The tests were found effective in identifying annotation quality issues before annotators annotate real data.\nPost-processing. For each language, annotations were performed by one annotator and reviewed by another annotator. From the data, we discard the translations marked as incomprehensible along with the data with some issues (e.g. unbalanced brackets in the word-level annotations of hallucinations or omissions; word-level annotations that significantly differ from the initial input/output texts). After this filtering, we were left with 144 to 197 annotated sentence pairs per direction."
},
{
"figure_ref": [],
"heading": "B Attention-based anomaly detection",
"publication_ref": [
"b8",
"b8",
"b22",
"b8",
"b8",
"b14"
],
"table_ref": [
"tab_2"
],
"text": "Reproducibility. The Attn-OT sentence-level detection method that we use in Section 4 is our reproduction of the Wass-Combo method from Guerreiro et al. (2022). Their paper did not provide code and training data, so our implementation is not exact. For Wass-to-Unif, we obtained the same ROC AUC scores as Guerreiro et al. (2022) on their test set, but for Wass-to-Data and Wass-Combo, the AUC scores are 2% lower than in the original paper, probably due to the differences in selecting the reference data.\nReference data for 18 directions. To apply the Attn-OT method to our data, we created reference translations for each of the 18 translation directions by following steps:\n• Sample 1M sentences for each language from the NLLB mined training data (NLLB Team et al., 2022); • Translate them with the same settings as in Section 2.2; • For each translation direction, drop the resulting sentence pairs that got into the worst 20% by any of the criteria: shortest-to-longest ratio for source and translation texts, Seq-LogProb, and LASER3 cosine similarity score between source and translation. After that, about 600K sentence pairs per direction are left as reference translations.\nComputing scores. To compute attention distribution, we average the encoder-decoder attention maps for the last decoder layer over heads and over target tokens, just like Guerreiro et al. (2022). Our Attn-OT score is then computed with the same formula as for Wass-Combo in Guerreiro et al. ( 2022), with the only slight difference: swtu is scaled by matching 1% and 99% quantiles of s wtd , instead of min-max scaling, to improve computational stability. Along with this score, we also evaluate Wass-to-Unif and Wass-to-Data scores from Guerreiro et al. (2022), and their weighted average with weights inversely proportional to standard deviations: Wass-Mean.\nDropping the EOS We observed that in the setting above, Wass-to-Unif has nearly zero rank correlation with hallucination severity. After inspecting the attention maps, we found that about 75% of attention weight on most heads is distributed to the end-of-sentence token. This probably compensates for the fact that the order of magnitude of its encoder hidden state is an order of magnitude smaller than for other tokens, which aligns with the observations of Kobayashi et al. (2020). This makes a standard attention map highly non-uniform, and may obscure the more informative differences between the attention maps for different translations. To mitigate this fact, we computed the second version of all scores, with dropping the attention to the EOS token and renormalizing it so that the sum of attention to the other tokens equals 1 again.\nEvaluation Table 2 reports ROC AUC scores for all OT-based detection methods, with and without including the EOS token. Removong the EOS token improves the Wass-to-Unif and Wass-Mean scores, but slightly negatively affects Wass-to-Data and Wass-Combo scores. Whatever method we use, its performance for hallucination detection is not much better than chance. of the hallucination level and omission levels for each sentence pair. Qualitatively, the results are similar to those for hallucination detection: internal methods perform equally well for all translation directions, whereas external methods deteriorate for low-resource directions. For the joint detection of hallucinations and omissions, BLASER 2.0-QE outperforms all other methods both for high-resource and for lowresource directions."
},
{
"figure_ref": [],
"heading": "C Detection of any pathology",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [
"fig_12"
],
"heading": "D Natural vs Artificially Induced Pathologies",
"publication_ref": [
"b15",
"b26"
],
"table_ref": [],
"text": "One of the difficulties when dealing with hallucinations (and, to a lesser extent, omissions) is that this is a rare phenomenon. Therefore, previous work often resorted to artificially amplifying the problem by applying various perturbations (Lee et al. (2019);Raunak et al. (2021), among others). However, it is not clear whether conclusions made based on this synthetic data would transfer to pathologies generated by a model in a natural setting. This is especially important when evaluating detection methods: for example, using internal workings of a model to detect its pathological behavior does not have to be helpful for pathologies the model did not generate \"voluntarily\". In this section, we compare performance of detection methods between two datasets: (i) our dataset with natural translations and (ii) translations generated under perturbation (annotated using the same protocol).\nModel perturbation. To encourage the model to hallucinate or omit source information while still generating fluent text, we decrease the output acti- We can see that perturbing the translation model introduces biases into the evaluation of detection methods. For example, for hallucination detection, Seq-Logprob outperforms ALTI on the natural dataset and loses on artificial. For omission detection, ALTI T is the best for the natural dataset, while XNLI is better on the artificial.\nFigure 12 shows similar results, but with fractions of data downsampled in a way that the natural and perturbed data subsets have equal number of observations for each combination of pathology type, source dataset and translation direction. This is done to ensure that differences in detection per- formance to come from different distribution of pathology types. We can see that even for these curated subsets, the conclusions do not transfer from perturbed to natural pathologies.\nOverall, we see that data with translations generated under perturbation has to be used with caution, especially when evaluating pathology detection methods: the conclusions are likely to not transfer to the natural setting. Logprob focuses on the beginnings. We notice that log-probability focuses more on the beginning of a word or a sentence. This makes sense: model uncertainty in prediction is generally higher when beginning generation.\nALTI focuses on the endings. Differently, token contributions focus on word endings. This is again expected: when generating a token that completes a word, source contribution is likely to be lower than for the other tokens. However, the predictions are still very reasonable -for the last three examples, ALTI detects omissions and hallucinations more confidently than log-probability.\nFinally, we see that feature-based combination of the methods leads to more refined results. "
}
] | Hallucinations in machine translation are translations that contain information completely unrelated to the input. Omissions are translations that do not include some of the input information. While both cases tend to be catastrophic errors undermining user trust, annotated data with these types of pathologies is extremely scarce and is limited to a few high-resource languages. In this work, we release an annotated dataset for the hallucination and omission phenomena covering 18 translation directions with varying resource levels and scripts. Our annotation covers different levels of partial and full hallucinations as well as omissions both at the sentence and at the word level. Additionally, we revisit previous methods for hallucination and omission detection, show that conclusions made based on a single language pair largely do not hold for a large-scale evaluation, and establish new solid baselines. | HalOmi: A Manually Annotated Benchmark for Multilingual Hallucination and Omission Detection in Machine Translation | [
{
"figure_caption": "Figure 1 :1Figure 1: Dataset summary.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: Overview of some parts of the annotation guidelines.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: Annotated examples from our dataset.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: Summary of the sentence-level detection methods.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 :5Figure 5: Results for sentence-level detection of hallucinations (left) and omissions (right).",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Figure 6 :6Figure 6: Token-level detection methods for hallucinations. For omissions, we swap the source and the target.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "Figure 7 :7Figure 7: Word-level detection results.",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "7",
"figure_type": "figure"
},
{
"figure_caption": "Figure 88Figure 8 shows manually selected examples of false and true positive and negative detection results.",
"figure_data": "",
"figure_id": "fig_7",
"figure_label": "8",
"figure_type": "figure"
},
{
"figure_caption": "Figure 8 :8Figure 8: Examples of successful and failed sentence-level detection for translations from English to Spanish.",
"figure_data": "",
"figure_id": "fig_8",
"figure_label": "8",
"figure_type": "figure"
},
{
"figure_caption": "Figure 9 :9Figure 9: Results for sentence-level detection of both hallucinations and omissions.vations of all the encoder-decoder attention layers by a constant multiplier α. Intuitively, this should imitate detachment from the source and increase hallucinatory rate. Indeed, translations generated this way are overall more pathological (see. We use α = 0.3 to match the average Seq-logprob of reference translations.",
"figure_data": "",
"figure_id": "fig_9",
"figure_label": "9",
"figure_type": "figure"
},
{
"figure_caption": "Figure 10 :10Figure 10: Sentence-level detection scores for natural and artificial (generated under perturbation) pathologies.",
"figure_data": "",
"figure_id": "fig_10",
"figure_label": "10",
"figure_type": "figure"
},
{
"figure_caption": "Figure 11 :11Figure 11: Annotation results: translations generated naturally vs under perturbation.",
"figure_data": "",
"figure_id": "fig_11",
"figure_label": "11",
"figure_type": "figure"
},
{
"figure_caption": "Figure 12 :12Figure 12: Word-level detection results.",
"figure_data": "",
"figure_id": "fig_12",
"figure_label": "12",
"figure_type": "figure"
},
{
"figure_caption": "Figure 13 :13Figure 13: Examples of word-level detection for translations from English to Spanish. Hallucinated and omitted fragments are underlined.",
"figure_data": "",
"figure_id": "fig_13",
"figure_label": "13",
"figure_type": "figure"
},
{
"figure_caption": "Number of sentence pairs selected from each source with each method: uniform sampling U, biased sampling B, and selecting worst cases W.",
"figure_data": "Data sourceFLORESCommentsSelectionUB WUB Weng_Latn-arb_Arab18 31 31 22 28 14arb_Arab-eng_Latn19 31 31 21 31 23eng_Latn-rus_Cyrl19 31 31 20 32 13rus_Cyrl-eng_Latn18 31 30 22 33 24eng_Latn-spa_Latn19 31 31 22 33 17spa_Latn-eng_Latn19 31 31 22 33 24eng_Latn-deu_Latn19 31 31 22 31 12deu_Latn-eng_Latn18 31 31 21 31 23eng_Latn-zho_Hans19 31 31 22 33 24zho_Hans-eng_Latn19 31 31 22 33 23eng_Latn-kas_Deva23 38 36 19 36 32kas_Deva-eng_Latn40 54 57000eng_Latn-yor_Latn24 39 36 27 40 29yor_Latn-eng_Latn40 51 55000eng_Latn-mni_Beng 24 39 37 27 39 31mni_Beng-eng_Latn 40 54 58000yor_Latn-spa_Latn40 54 58000spa_Latn-yor_Latn40 53 58000Ghazvininejad. 2021. Detecting hallucinated contentin conditional neural sequence generation. In Find-ings of the Association for Computational Linguis-tics: ACL-IJCNLP 2021, pages 1393-1404, Online.Association for Computational Linguistics.Wenhao Zhu, Hongyi Liu, Qingxiu Dong, Jingjing Xu,Lingpeng Kong, Jiajun Chen, Lei Li, and ShujianHuang. 2023. Multilingual machine translation withlarge language models: Empirical results and analy-sis.",
"figure_id": "tab_1",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Figure 9 reports scores for detecting hallucinations and omissions jointly. The scores are computed as percentage of correctly ranked pairs w.r.t. the worst ROC AUC scores for detection of hallucinations and omissions with OT-based methods, averaged across tranlsation directions. The asterisk* denotes the scores based on the attention maps with the EOS token excluded.",
"figure_data": "MethodHallucinations OmissionsWass-to-Unif0.490.43Wass-to-Data0.530.71Wass-Combo0.530.71Wass-Mean0.510.51Wass-to-Unif*0.550.65Wass-to-Data*0.510.69Wass-Combo*0.520.69Wass-Mean*0.550.67",
"figure_id": "tab_2",
"figure_label": "2",
"figure_type": "table"
}
] | David Dale; Elena Voita; Janice Lam; Prangthip Hansanti; Christophe Ropers; Elahe Kalbassi; Cynthia Gao; Loïc Barrault; Costa-Jussà Fair | [
{
"authors": "Jasmijn Bastings; Katja Filippova",
"journal": "Association for Computational Linguistics",
"ref_id": "b0",
"title": "The elephant in the interpretability room: Why use attention as explanation when we have saliency methods",
"year": "2020"
},
{
"authors": "Alexandre Berard; Ioan Calapodescu; Claude Roux",
"journal": "Association for Computational Linguistics",
"ref_id": "b1",
"title": "Naver labs Europe's systems for the WMT19 machine translation robustness task",
"year": "2019"
},
{
"authors": "David Dale; Elena Voita; Loïc Barrault; Marta R ",
"journal": "Association for Computational Linguistics",
"ref_id": "b2",
"title": "Detecting and mitigating hallucinations in machine translation: Model internal workings alone do well, sentence similarity even better",
"year": "2023"
},
{
"authors": "Paul-Ambroise Duquenne; Holger Schwenk; Benoit Sagot",
"journal": "",
"ref_id": "b3",
"title": "SONAR: sentence-level multimodal and language-agnostic representations",
"year": "2023"
},
{
"authors": "Fangxiaoyu Feng; Yinfei Yang; Daniel Cer; Naveen Arivazhagan; Wei Wang",
"journal": "Association for Computational Linguistics",
"ref_id": "b4",
"title": "Language-agnostic BERT sentence embedding",
"year": "2022"
},
{
"authors": "Patrick Fernandes; Kayo Yin; Graham Neubig; F T André; Martins",
"journal": "Association for Computational Linguistics",
"ref_id": "b5",
"title": "Measuring and increasing context usage in context-aware machine translation",
"year": "2021"
},
{
"authors": "Javier Ferrando; Gerard I Gállego; Belen Alastruey; Carlos Escolano; Marta R Costa-Jussà",
"journal": "Association for Computational Linguistics",
"ref_id": "b6",
"title": "Towards opening the black box of neural machine translation: Source and target interpretations of the transformer",
"year": "2022"
},
{
"authors": "Yvette Graham; Timothy Baldwin; Alistair Moffat; Justin Zobel",
"journal": "Association for Computational Linguistics",
"ref_id": "b7",
"title": "Continuous measurement scales in human evaluation of machine translation",
"year": "2013"
},
{
"authors": "M Nuno; Pierre Guerreiro; Pablo Colombo; Piantanida; F T André; Martins",
"journal": "",
"ref_id": "b8",
"title": "Optimal transport for unsupervised hallucination detection in neural machine translation",
"year": "2022"
},
{
"authors": "M Nuno; Elena Guerreiro; André Voita; Martins",
"journal": "Association for Computational Linguistics",
"ref_id": "b9",
"title": "Looking for a needle in a haystack: A comprehensive study of hallucinations in neural machine translation",
"year": "2023"
},
{
"authors": "Kevin Heffernan; Onur Çelebi; Holger Schwenk",
"journal": "",
"ref_id": "b10",
"title": "Bitext mining using distilled sentence representations for low-resource languages",
"year": "2022"
},
{
"authors": "Sarthak Jain; Byron C Wallace",
"journal": "Association for Computational Linguistics",
"ref_id": "b11",
"title": "Attention is not Explanation",
"year": "2019"
},
{
"authors": " Jigsaw",
"journal": "",
"ref_id": "b12",
"title": "Jigsaw multilingual toxic comment classification",
"year": "2020"
},
{
"authors": "Ákos Kádár; Grzegorz Chrupała; Afra Alishahi",
"journal": "Computational Linguistics",
"ref_id": "b13",
"title": "Representation of linguistic form and function in recurrent neural networks",
"year": "2017"
},
{
"authors": "Goro Kobayashi; Tatsuki Kuribayashi; Sho Yokoi; Kentaro Inui",
"journal": "Association for Computational Linguistics",
"ref_id": "b14",
"title": "Attention is not only a weight: Analyzing transformers with vector norms",
"year": "2020"
},
{
"authors": "Katherine Lee; Orhan Firat; Ashish Agarwal; Clara Fannjiang; David Sussillo",
"journal": "",
"ref_id": "b15",
"title": "Hallucinations in neural machine translation",
"year": "2019"
},
{
"authors": "Jiwei Li; Will Monroe; Dan Jurafsky",
"journal": "",
"ref_id": "b16",
"title": "Understanding neural networks through representation erasure",
"year": "2017"
},
{
"authors": "Xintong Li; Guanlin Li; Lemao Liu; Max Meng; Shuming Shi",
"journal": "Association for Computational Linguistics",
"ref_id": "b17",
"title": "On the word alignment from neural machine translation",
"year": "2019"
},
{
"authors": "Daniel Licht; Cynthia Gao; Janice Lam; Francisco Guzman; Mona Diab; Philipp Koehn",
"journal": "",
"ref_id": "b18",
"title": "Consistent human evaluation of machine translation across language pairs",
"year": "2022"
},
{
"authors": "Arle Lommel; Aljoscha Burchardt; Hans Uszkoreit",
"journal": "Tradumàtica: tecnologies de la traducció",
"ref_id": "b19",
"title": "Multidimensional quality metrics (mqm): A framework for declaring and describing translation quality metrics",
"year": "2014"
},
{
"authors": "Mathias Müller; Annette Rios; Rico Sennrich",
"journal": "Virtual. Association for Machine Translation in the Americas",
"ref_id": "b20",
"title": "Domain robustness in neural machine translation",
"year": "2020"
},
{
"authors": "Mathias Müller; Rico Sennrich",
"journal": "Association for Computational Linguistics",
"ref_id": "b21",
"title": "Understanding the properties of minimum Bayes risk decoding in neural machine translation",
"year": "2021"
},
{
"authors": "Marta R Nllb Team; James Costa-Jussà; Onur Cross; Maha Çelebi; Kenneth Elbayad; Kevin Heafield; Elahe Heffernan; Janice Kalbassi; Daniel Lam; Jean Licht; Anna Maillard; Skyler Sun; Guillaume Wang; Al Wenzek; Bapi Youngblood; Loic Akula; Gabriel Mejia Barrault; Prangthip Gonzalez; John Hansanti; Semarley Hoffman; Jarrett; Ram Kaushik; Dirk Sadagopan; Shannon Rowe; Chau Spruit; Pierre Tran; Necip Andrews; Shruti Fazil Ayan; Sergey Bhosale; Angela Edunov; Cynthia Fan; Vedanuj Gao; Francisco Goswami; Philipp Guzmán; Alexandre Koehn; Christophe Mourachko; Safiyyah Ropers; Holger Saleem; Jeff Schwenk; Wang",
"journal": "",
"ref_id": "b22",
"title": "No language left behind: Scaling humancentered machine translation",
"year": "2022"
},
{
"authors": "Nina Poerner; Hinrich Schütze; Benjamin Roth",
"journal": "Association for Computational Linguistics",
"ref_id": "b23",
"title": "Evaluating neural network explanation methods using hybrid documents and morphosyntactic agreement",
"year": "2018"
},
{
"authors": "Maja Popović",
"journal": "Association for Computational Linguistics",
"ref_id": "b24",
"title": "chrF++: words helping character n-grams",
"year": "2017"
},
{
"authors": "Matt Post",
"journal": "Association for Computational Linguistics",
"ref_id": "b25",
"title": "A call for clarity in reporting BLEU scores",
"year": "2018"
},
{
"authors": "Arul Vikas Raunak; Marcin Menezes; Junczys-Dowmunt",
"journal": "Association for Computational Linguistics",
"ref_id": "b26",
"title": "The curious case of hallucinations in neural machine translation",
"year": "2021"
},
{
"authors": "Ricardo Rei; Craig Stewart; Ana C Farinha; Alon Lavie",
"journal": "Association for Computational Linguistics",
"ref_id": "b27",
"title": "COMET: A neural framework for MT evaluation",
"year": "2020"
},
{
"authors": "Seamless Communication; Loïc Barrault; Yu-An Chung; Mariano ; Cora Meglioli; David Dale; Ning Dong; Paul-Ambroise Duquenne; Hady Elsahar; Hongyu Gong; Kevin Heffernan; John Hoffman; Christopher Klaiber; Pengwei Li; Daniel Licht; Jean Maillard; Alice Rakotoarison; Ram Kaushik; Guillaume Sadagopan; Ethan Wenzek; Bapi Ye; Peng-Jen Akula; Naji El Chen; Brian Hachem; Gabriel Mejia Ellis; Justin Gonzalez; Prangthip Haaheim; Russ Hansanti; Bernie Howes; Min-Jae Huang; Hirofumi Hwang; Somya Inaguma; Elahe Jain; Amanda Kalbassi; Ilia Kallet; Janice Kulikov; Daniel Lam; Xutai Li; Ruslan Ma; Benjamin Mavlyutov; Mohamed Peloquin; Abinesh Ramadan; Anna Ramakrishnan; Kevin Sun; Tuan Tran; Igor Tran; Vish Tufanov; Carleigh Vogeti; Yilin Wood; Bokai Yang; Pierre Yu; Can Andrews; Marta R Balioglu; Onur Costa-Jussà; Maha Celebi; Cynthia Elbayad; Francisco Gao; Justine Guzmán; Ann Kao; Alexandre Lee; Juan Mourachko; Sravya Pino; Christophe Popuri; Safiyyah Ropers; Holger Saleem; Paden Schwenk; Changhan Tomasello; Jeff Wang; Skyler Wang; Wang",
"journal": "",
"ref_id": "b28",
"title": "Seamlessm4tmassively multilingual & multimodal machine translation",
"year": "2023"
},
{
"authors": "Sofia Serrano; Noah A Smith",
"journal": "Association for Computational Linguistics",
"ref_id": "b29",
"title": "Is attention interpretable",
"year": "2019"
},
{
"authors": "Felix Stahlberg; Bill Byrne",
"journal": "Association for Computational Linguistics",
"ref_id": "b30",
"title": "On NMT search errors and model errors: Cat got your tongue?",
"year": "2019"
},
{
"authors": "Jannis Vamvas; Rico Sennrich",
"journal": "",
"ref_id": "b31",
"title": "As little as possible, as much as necessary: Detecting over-and undertranslations with contrastive conditioning",
"year": "2022"
},
{
"authors": "K Ashwin; Michael Vijayakumar; Cogswell; Qing Ramprasath R Selvaraju; Stefan Sun; David Lee; Dhruv Crandall; Batra",
"journal": "",
"ref_id": "b32",
"title": "Diverse beam search: Decoding diverse solutions from neural sequence models",
"year": "2016"
},
{
"authors": "Sarah Wiegreffe; Yuval Pinter",
"journal": "Association for Computational Linguistics",
"ref_id": "b33",
"title": "Attention is not not explanation",
"year": "2019"
},
{
"authors": "Weijia Xu; Sweta Agrawal; Eleftheria Briakou; Marianna J Martindale; Marine Carpuat",
"journal": "",
"ref_id": "b34",
"title": "Understanding and detecting hallucinations in neural machine translation via model introspection",
"year": "2023"
},
{
"authors": "Matthew D Zeiler; Rob Fergus",
"journal": "Cham. Springer International Publishing",
"ref_id": "b35",
"title": "Visualizing and understanding convolutional networks",
"year": "2014"
},
{
"authors": "Chunting Zhou; Graham Neubig; Jiatao Gu; Mona Diab; Francisco Guzmán; Luke Zettlemoyer; Marjan ",
"journal": "",
"ref_id": "b36",
"title": "",
"year": ""
}
] | [] |
|
10.48550/arXiv.2303.08896 | 2023-10-23 | [
{
"figure_ref": [
"fig_1"
],
"heading": "Introduction",
"publication_ref": [
"b3",
"b18",
"b26",
"b1",
"b25",
"b4",
"b6",
"b26",
"b20"
],
"table_ref": [],
"text": "The advent of large language models (LLMs) (Zhao et al., 2023) has ushered in a paradigm shift in natural language processing (NLP), making unprecedented progress in text generation and understanding (Brown et al., 2020;Li et al., 2021). The remarkable language ability makes LLMs core in a number of products with millions of users, such as the coding assistant Copilot and recent ChatGPT.\nDespite these prominent capabilities of LLMs trained on large text corpus, recent work has shown that LLMs are prone to suffer from hallucination generations across various applications (Ji et al., Table 1: An example from Alpaca (Taori et al., 2023) showing that ChatGPT might generate hallucinated contents (green) that cannot be verified by existing source.\n2023; Bang et al., 2023;Sun et al., 2023), where the generated content is either in conflict with existing source or cannot be verified by the available knowledge resources. The issue of hallucination makes the deployment of LLMs potentially risky in real-world applications. Most exiting work mainly focuses on investigating the causes of hallucination for specific tasks and small language models (Cao et al., 2022;Zheng et al., 2023;Das et al., 2023). However, it still remains unclear what types of content and to which extent LLMs tend to hallucinate.\nTo facilitate research in this direction, we present the Hallucination Evaluation benchmark for Large Language Models (HaluEval): a large collection of 35,000 hallucinated/normal samples for LLMs analysis and evaluation. HaluEval includes 5,000 general user queries with ChatGPT responses and 30,000 task-specific examples from three tasks, i.e., question answering, knowledge-grounded dialogue, and text summarization. The construction pipeline of HaluEval is depicted in Figure 1. For general user queries, we adopt the 52K instruction tuning dataset from Alpaca (Taori et al., 2023) for human annotation. To further screen out user queries where LLMs are most likely to produce hallucinations, we use ChatGPT to sample three responses for each query and only retain 5, 000 queries with the lowest similarity among three responses. Ac- The best answer is Answer 1. cording to recent work (Manakul et al., 2023), hallucinations are likely to appear in diverged and conflicting responses of LLMs. Based on the filtered user queries and ChatGPT responses, we invite human labelers to annotate whether the response contains hallucinated information and mark corresponding spans. As shown in Table 1, for the user query \"Retrieve the oldest photo of a cat\", the response generated by ChatGPT contains unverifiable information. These human-annotated queries and responses can be used to analyze what types of content LLMs tend to hallucinate and further conceive effective methods to alleviate it."
},
{
"figure_ref": [],
"heading": "High-quality Hallucination Filtering",
"publication_ref": [],
"table_ref": [],
"text": "Furthermore, for the task-specific examples, we design an automatic two-stage approach to generate hallucinated samples. First, based on existing task datasets (e.g., HotpotQA) as seed data, we employ ChatGPT to generate hallucinated samples with two styles of task-specific instructions, i.e., onepass and conversational. We expect that these two methods will generate diverse hallucinated samples from different aspects. Second, to select the most plausible and difficult hallucinated sample for LLMs evaluation, we elaborate the filtering instruction enhanced by ground-truth examples and leverage ChatGPT for sample selection. Through the proposed sampling-then-filtering approach, we can generate a hallucinated counterpart for each specific task example. These hallucinated samples are designed to challenge the ability of LLMs in hallucination recognition and analyze the information blind spots of LLMs.\nTo better understand the performance of LLMs in HaluEval, we conduct experiments with several existing powerful LLMs (e.g., ChatGPT, GPT-3).\nOur key findings can be summarized as follows:\n• First, ChatGPT is likely to generate hallucinated content by fabricating unverifiable information in its responses (i.e., about 19.5% responses). The hallucinated texts from ChatGPT cover topics including language, climate, and technology.\n• Second, existing LLMs face significant challenges to identify the hallucinations in the generated text, even for ChatGPT which is used to generate these hallucinated samples (e.g., only 62.59% accuracy for ChatGPT in question answering).\n• Finally, the deficient performance of LLMs in recognizing hallucinations can be improved by providing explicit knowledge and adding intermediate reasoning steps. While, contrasting hallucinated samples with ground-truth makes LLMs more confused and leads to worse performance."
},
{
"figure_ref": [],
"heading": "The HaluEval Benchmark",
"publication_ref": [],
"table_ref": [],
"text": "As the goal of HaluEval is to understand what types of content and to which extent LLMs tend to hallucinate, the benchmark contains a myriad of correct samples and their hallucinated counterparts. This collection is created via two ways, i.e., automatic generation and human annotation."
},
{
"figure_ref": [
"fig_1"
],
"heading": "Automatic Generation",
"publication_ref": [],
"table_ref": [],
"text": "Our generation pipeline includes two steps: 1) diverse hallucination sampling, and 2) high-quality hallucination filtering. We employ ChatGPT to execute the creation pipeline automatically.\nDiverse Hallucination Sampling. Since a factual text can be hallucinated from different aspects, we propose two different hallucination sampling meth-I want you act as a hallucination answer generator. Given a question, right answer, and related knowledge, your objective is to write a hallucinated answer that sounds plausible but is factually incorrect. You SHOULD write the hallucinated answer using the following method (each with some examples): You are trying to answer a question but there is a factual contradiction between the answer and the knowledge. You can fabricate some information that does not exist in the provided knowledge. ods to generate diverse samples. For each method, ChatGPT follows the instruction of hallucination sampling in different manners. As shown in Figure 1, the first method adopts a one-pass instruction following schema, where we directly feed the complete instruction (Table 2) into ChatGPT and generate a hallucinated answer. On the other hand, the second method uses a conversational schema, where we teach ChatGPT to successively learn part of the instruction and make sure it has mastered. Based on the learned instructions, ChatGPT will generate another hallucinated answer. Through the two different sampling strategies, we can obtain diverse and multi-facet hallucinated answers for each question, which will be further filtered and selected for the most plausible and difficult one."
},
{
"figure_ref": [],
"heading": "#Knowledge",
"publication_ref": [
"b6",
"b4",
"b29",
"b21",
"b24"
],
"table_ref": [
"tab_1",
"tab_10"
],
"text": "Instruction Design. In our approach, the key is to design an effective instruction for ChatGPT to generate hallucinated samples. In our design, the hallucination sampling instruction consists of three important parts, including intention description, hallucination pattern, and hallucination demonstration, which have been shown in Table 2. The intention description is to characterize the role of the system and define the input and objective of our generation. To control the type and quality of hallucinated samples, we introduce the hallucination pattern and demonstration, which are related to the seed task (e.g., QA in Table 2). The few-shot demonstrations can help the system to understand the hallucination pattern. In this paper, we automatically generate hallucinated samples for three tasks, i.e., question answering, knowledge-grounded dialogue, and text summarization. Specifically, we consider four types of hallucination patterns for question answering (i.e., comprehension, factualness, specificity, and inference) (Zheng et al., 2023), three types of hallucination patterns for knowledgegrounded dialogue (i.e., extrinsic-soft, extrinsichard, and extrinsic-grouped) (Das et al., 2023), and three types of hallucination patterns for text summarization (i.e., factual, non-factual, and intrinsic) (Cao et al., 2022). For these three tasks, we first randomly sample 30, 000 instances from the training set of HotpotQA (Yang et al., 2018), OpenDialKG (Moon et al., 2019), and CNN/Daily Mail (See et al., 2017), and then generate their hallucinated examples. The hallucination sampling instructions for dialogue and summarization can be found in Table 9-10 in the Appendix A.\nHigh-quality Hallucination Filtering. To construct a challenging benchmark for LLMs, we aim I want you act as an answer judge. Given a question, two answers, and related knowledge, your objective is to select the best and correct answer without hallucination and non-factual information. Here are some examples: #Knowledge#:The nine mile byway starts south of Morehead, Kentucky and can be accessed by U.S. Highway 60. Morehead is a home rule-class city located along US 60 (the historic Midland Trail) to select the most plausible and difficult hallucinated samples from the above two sampling methods. As shown in Table 3, we design the instruction of hallucination filtering enhanced by ground-truth answers to select the best answer from two hallucinated candidates. In the instruction of filtering, the demonstration includes the ground-truth correct answer (e.g., U.S. Highway 60) and a hallucinated counterpart (e.g., U.S. Highway 70). While, in the test example, we input two hallucinated answers. Following the demonstrations, we expect ChatGPT to select one of the hallucinated answers that is the most plausible and closest to the right answer. Finally, the selected hallucinated sample is hard to be identified, which are further used to evaluate LLMs in hallucination recognition. The instructions of hallucination filtering for dialogue and summarization are shown in Table 11-12 in the Appendix B.\nThrough the sampling-then-filtering process, we end up generating a total of 30, 000 hallucinated samples for the three tasks. Our approach can also be adapted to other tasks and datasets."
},
{
"figure_ref": [],
"heading": "Human Annotation",
"publication_ref": [
"b26",
"b31",
"b20",
"b22"
],
"table_ref": [],
"text": "Besides generating hallucinated samples, we also invite human labelers to annotate whether ChatGPT responses contain hallucinated content.\nWe annotate the general user queries and Chat-GPT responses from the 52K instruction tuning dataset from Alpaca (Taori et al., 2023), which has been widely used by recent LLMs. To screen out user queries where LLMs are most likely to produce hallucination for labeling, we design a pre- selection procedure. Specifically, we use ChatGPT to sample three responses for each user query and compute their average semantic similarity using BERTScore (Zhang et al., 2020). We finally retain 5, 000 user queries with the lowest similarities.\nAccording to recent work (Manakul et al., 2023), hallucinations are likely to appear in diverged and conflicting responses of LLMs. For each query and ChatGPT response, human labelers will annotate whether the response contains hallucinated information (\"Yes\" or \"No\") and list the corresponding spans. The hallucination is considered from the following three aspects: unverifiable, non-factual, and irrelevant. Each response is labeled by three human labelers, and we adopt the max-voting strategy to determine the final hallucination label.\nLabeler Details. Annotating the hallucination in ChatGPT responses is a very challenging task, which requires good reading comprehension skills and using search engine to look up relevant information for judgement. Thus, from an initial pool of labeler candidates, we select labelers who are good at English passage reading with at least an undergraduate-level education. Besides, following (Ouyang et al., 2022), we have labelers annotate a small number of test examples and measure their agreement with the labels of researchers, and finally we choose thirty human labelers with the highest agreement scores. We report Fleiss's Kappa (κ) to indicate the reliability of agreement between human labelers. We compute κ on 5, 000 annotated samples and obtain κ = 0.811 (0.80 ≤ κ ≤ 1.00) showing a perfect agreement."
},
{
"figure_ref": [],
"heading": "Benchmark Analysis and Usage",
"publication_ref": [
"b26",
"b5",
"b30",
"b27"
],
"table_ref": [
"tab_16",
"tab_3"
],
"text": "With the automatic two-step generation process in Section 2.1, we produce a total of 30, 000 hallucinated samples with 10, 000 examples for each task of QA, dialogue, and summarization. We show the number of generated samples for each hallucination pattern in Table 16 at the Appendix D. Moreover, we manually annotate 5, 000 ChatGPT responses for general user queries in Section 2.2. We present a QA example and an annotated query and response example in Table 4. Among the annotated ChatGPT responses, 977 responses are labeled as containing hallucination (19.5%). Finally, we present the topic distributions of our generated task-specific samples and annotated ChatGPT responses in Figure 2 and Figure 3, ranging from film, sports to school, computer, technology, etc. With our benchmark, researchers can use it to investigate or mitigate the hallucination issue for LLMs in three aspects. Firstly, based on our generated and annotated samples, researchers can use them to analyze what types of content LLMs tend to generate hallucinations. Second, researchers can further evaluate the ability of LLMs to recognize hallucinations in the generated samples. For example, given a question and an answer, LLMs can be asked to determine whether the answer contains hallucinated content. Finally, our benchmark can be further paired with human annotation to assess whether the LLMs' output contains hallucinations, since the samples in our benchmark are specially designed for testing the hallucinations of LLMs.\nTo use our benchmark, users can run the code in our project repository to conduct the corresponding evaluation and analysis. Users can use our provided instructions on their own datasets to evaluate LLMs on hallucinations. (Taori et al., 2023), Vicuna (7B) (Chiang et al., 2023), ChatGLM (7B) (Zeng et al., 2022), Falcon (7B) (TII, 2023), and Llama 2-Chat (7B) (Touvron et al., 2023). Our experiments were performed without fine-tuning or engaging in the tuning of hyper-parameters.\nImplementation Details. We execute the generation process of hallucinated samples using Azure OpenAI ChatGPT API. We use a temperature of 1.0 to generate samples and set the maximum number of tokens for generation to 256. Moreover, we set the frequency penalty to zero and top-p to 1.0.\nFor evaluation, we set the temperature to zero for all models to reduce output randomness and ensure more focused and deterministic outputs.\nIn the following, we first conduct hallucination recognition experiments, then propose several potentially useful strategies to improve the recognition, and finally we perform qualitative analysis to understand the hallucination in LLMs."
},
{
"figure_ref": [],
"heading": "Results and Analysis",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Hallucination Recognition",
"publication_ref": [
"b2"
],
"table_ref": [
"tab_13",
"tab_14",
"tab_15",
"tab_5"
],
"text": "To evaluate the ability of LLMs to recognize hallucinations, we randomly select the hallucinated or normal output (e.g., an answer) of each sample for classification. The evaluation instructions of QA, dialogue, and summarization are presented in Table 13, Table 14 and Table 15 in Appendix C.\nTable 5 presents the accuracy of evaluated LLMs to classify whether the sample output contains hallucinated information. Our findings indicate that LLMs are still poor at identifying hallucination which might be implicit in text. For example, the state-of-the-art ChatGPT model cannot distinguish between factual and hallucinated summary and only achieves 58.53% accuracy in text summarization, which is barely above chance. Moreover, GPT-3 obtains just about random chance of 50% accuracy across three tasks, and Alpaca or Vicuna even performs worse (well below random chance). We hypothesize that LLMs perform poorly because the hallucinated sample we generate looks highly similar with ground-truth ones but differs in the key factual spans. As we can see, from GPT-3 to InstructGPT and ChatGPT, instruction tuning and alignment with humans can strength the ability of LLMs in identifying the hallucinations in text.\nWith respect to the hallucinated samples where ChatGPT fails to recognize, we present the number of each hallucination pattern in Table 6. Based on the results, we can observe that the hallucination patterns of failed samples are unevenly distributed. For example, over half of failures in QA, dialogue, and summarization originate from the first hallucination pattern (i.e., comprehension, extrinsic-soft, and factual), which refers to the hallucinations that are factually correct but conflict with the context. This indicates that LLMs lack or cannot associate related knowledge to identify the factual hallucination in the generated text. To further understand the failures of ChatGPT, we visualize the topics of those failed samples via Latent Dirichlet Allocation (LDA) (Blei et al., 2003). As shown in Figure 2 and Figure 3, we cluster all task samples into ten topics and mark the topics of failed samples as red. We find that the hallucination of LLMs is topicsensitive. For example, the frequent topics in QA include film, school, and company. While, Chat-GPT mainly fails to recognize those samples in the topics of film, company, and band. For user queries and ChatGPT responses, the top five topics include story, health, language, technology, and computer. ChatGPT mainly faces challenges in topics of technology, climate, and language. "
},
{
"figure_ref": [],
"heading": "Improvement Strategies",
"publication_ref": [
"b16",
"b28"
],
"table_ref": [
"tab_8",
"tab_8",
"tab_8"
],
"text": "In this part, we design several strategies to improve the ability of LLMs to recognize hallucination. The results are shown in Table 8.\nKnowledge Retrieval. Retrieving relevant knowledge is a widely used strategy to eliminate hallucination (Lewis et al., 2020;Li et al., 2023a). Therefore, we supply ChatGPT with the knowledge facts retrieved from Wikipedia (except for that summarization does not need external information besides the source document). By providing knowledge, the recognition accuracy of ChatGPT increases significantly (e.g., increasing from 62.59 to 76.83 in QA), while the performance improvement in dialogue is mild. We hypothesize that the common hallucination patterns in dialogue (i.e., extrinsicsoft/hard) cannot be simply identified via incorporating external knowledge. For those general user queries and ChatGPT responses, we discover that providing external knowledge does have a significant benefit. Thus, equipping LLMs with external knowledge can largely enhance their abilities to recognize hallucinations.\nCoT Reasoning. In previous work (Wei et al., 2022), chain-of-thought (CoT) has been proposed to improve the ability of LLMs to perform reasoning and derive the final answer by introducing a series of intermediate reasoning steps. Here, besides producing the recognition result, we also require ChatGPT to generate the reasoning steps. While, from the results in Table 8, we observe that generating reasoning steps can mildly improve the performance but makes the model perform worse in QA and dialogue (e.g., dropping from 62.59 to 59.58).\nCompared to retrieving knowledge, adding chainof-thought before output might interfere with the final judgement. While, in text summarization, generating reasoning steps improve the accuracy from 58.53 to 61.21. The reason might be that the factual contradiction between document and summary can be identified through logic reasoning.\nSample Contrast. We further provide ground-truth examples for ChatGPT to test whether it can distinguish the right sample from the hallucinated sample. As we can see from Table 8, distinguishing between right and hallucinated samples achieves the worst results. We hypothesize that our generated hallucinated samples have a high similarity to the real samples, thus making LLMs confused to distinguish them. This test also indicates that our benchmark is very challenging in hallucination evaluation for LLMs."
},
{
"figure_ref": [],
"heading": "Case Study",
"publication_ref": [],
"table_ref": [],
"text": "In the above, we have observed that providing external knowledge can be beneficial for LLMs to mitigate and recognize hallucinations. To demonstrate the effectiveness of knowledge retrieval in mitigating hallucinations, we present two hallucinated responses from ChatGPT and refined responses after augmented with retrieved knowledge in Table 7. In the first example, the generated span (i.e., \"July 4, 1776 -Declaration of Independence signing\") contains hallucinated information because it gives a wrong time of Declaration of Independence signing. By providing retrieved information about Declaration of Independence signing, ChatGPT is able to correct the hallucinated span and give the right information. Analogously, in the second example, ChatGPT gives incorrect GDP growth rates of China and India, which is due to that API-based ChatGPT cannot access the web to obtain the official data. After providing official information retrieved from World Bank, the refined span displays answers that contain the correct information.\nThe above two examples illustrate that retrieving knowledge related to queries can help ChatGPT significantly reduce the hallucinations in the response, especially those factual errors."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b33",
"b14",
"b15",
"b9",
"b6",
"b4",
"b11",
"b23",
"b19",
"b23",
"b7",
"b12"
],
"table_ref": [],
"text": "Hallucination in LLMs. Hallucination in LLMs is concerning since it hinders performance and raises safety risks in real-world application. To alleviate this issue, prior studies have proposed to use a verification system to identify non-factual entities in text summarization (Zhao et al., 2020), invoke interfaces of structured data (e.g., knowledge graph, database) to obtain related evidence (Jiang et al., 2023;Lan et al., 2022), and train a token-level fact critic to recognize hallucination and rectify them in dialogue (Dziri et al., 2021). To enhance the understanding of hallucination in LLMs and pro- mote the unification of research efforts, there are many active endeavors to analyze the causes of hallucination in different tasks and investigate their relationship (Zheng et al., 2023;Das et al., 2023;Cao et al., 2022). Our work is closely related to these work, but we focus on building a hallucination evaluation benchmark for LLMs. Our dataset can serve as a public platform for exhibiting the blind spots of LLMs in solving hallucination.\nHallucination Evaluation. Another line of work focusing on evaluating the hallucination of models in different NLP tasks (Dziri et al., 2022b;Gupta et al., 2022;Dziri et al., 2022a;Rashkin et al., 2021;Li et al., 2023b). For instance, The BEGIN benchmark (Dziri et al., 2022b) classifies the utterances generated by dialogue systems into three categories, i.e., fully attributable, not fully attributable, and generic; and the Attributable to Identified Sources (AIS) benchmark (Rashkin et al., 2021) assesses whether the source documents support the output of text generation models. Though these benchmarks can serve as decent evaluation platform, they are penurious in only focusing on single tasks (e.g., dialogue) and small models (e.g., DPR). Besides, several metrics have been proposed to quantify hallucination, such as PARENT (Dhingra et al., 2019) for measuring n-gram lexical entailment in table-totext generation and TRUE (Honovich et al., 2022) computes the example-level Area Under the ROC Curve. In this work, our HaluEval benchmark includes general user queries and ChatGPT responses and proposes a two-step automatic process to generate hallucinated samples for evaluation, which is completely based on LLMs."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "We introduce HaluEval, a large-scale collection of generated and human-annotated hallucinated samples for evaluating the performance of LLMs in recognizing hallucinations. To automatically generate large-scale samples, we propose a two-step approach, i.e., sampling-then-filtering. We first introduce two different sampling methods to generate diverse samples using instructions and then filter and select the difficult one. Besides, we invite qualified human labelers to annotate the hallucinations of ChatGPT responses given user queries. We find that, existing LLMs mostly fail to recognize the hallucinations in text and tend to generate hallucinated content. Finally, we suggest several strategies to help LLMs recognize hallucinations. Our benchmark can facilitate research in understanding what types of content and to which extent LLMs tend to hallucinate, ultimately paving the way for building more effective and reliable LLMs in the future."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [
"b6"
],
"table_ref": [],
"text": "In our approach, we leverage a LLM, i.e., ChatGPT, to automatically generate the hallucinated samples. Therefore, the quality of our hallucinated samples is limited by the capacity of ChatGPT in following the complex instruction of hallucination sampling.\nAlthough we design the high-quality hallucination filtering process, it is still necessary to apply quality control to the generation of hallucinated samples. Besides, our benchmark focuses on evaluating the ability of LLMs in recognizing the hallucinations in text but does not investigate the underlying reasons behind the appearance of hallucinations like prior work (Zheng et al., 2023;Das et al., 2023).\nAs for the potential issue, since the hallucinated samples in our benchmark looks highly similar to the ground-truth samples, which might be misused for an unexpected purpose than we planned. To alleviate this issue, we should monitor and regulate the spread and usage of our benchmark."
},
{
"figure_ref": [],
"heading": "B Hallucination Filtering",
"publication_ref": [],
"table_ref": [
"tab_10",
"tab_11"
],
"text": "The hallucination sampling instructions for dialogue and summarization are shown in Table 11 and Table 12, respectively."
},
{
"figure_ref": [],
"heading": "C Hallucination Recognition",
"publication_ref": [],
"table_ref": [
"tab_13",
"tab_14",
"tab_15"
],
"text": "The hallucination recognition instructions for QA, dialogue and summarization are shown in Table 13, Table 14 andTable 15, respectively."
},
{
"figure_ref": [],
"heading": "D Details of HaluEval",
"publication_ref": [],
"table_ref": [
"tab_16"
],
"text": "The number of generated hallucinated samples for each hallucination pattern are shown in Table 16.\nI want you act as an assistant in a conversation with human. Given a dialogue history, the true response, and related knowledge, your objective is to write a hallucinated response that sounds plausible but is factually incorrect. You SHOULD write the hallucinated response using the following method (each with some examples):\nYou are trying to write a response to human but you replace the true entity with a highly similar entity. #Knowledge#: The Dark Knight is a 2008 superhero film directed by Christopher Nolan from a screenplay he co-wrote with his brother Jonathan. Christopher Nolan is a film director. Table 9: Instruction of hallucination sampling for knowledge-grounded dialogue.\nI want you act as a hallucination summary generator. Given a document and the right summary, your objective is to write a hallucinated summary that sounds plausible but is factually incorrect. You SHOULD write the hallucinated summary using the following method (each with some examples):\nYou are trying to write a summary which is factual but some information cannot be directly inferred or entailed from the document. #Document#: The panther chameleon was found on Monday by a dog walker in the wooded area at Marl Park. It had to be put down after X-rays showed all of its legs were broken and it had a deformed spine. RSPCA Cymru said it was an \"extremely sad example of an abandoned and neglected exotic pet\". Inspector Selina Chan said: \"It is a possibility that the owners took on this animal but were unable to provide the care he needs and decided to release him to the wild. \"We are urging potential owners of exotic animals to thoroughly research what is required in the care of the particular species before taking one on. \"Potential owners need to make sure they can give their animal the environment it needs and they have the facilities, time, financial means and longterm commitment to maintain a good standard of care, as required under the Animal Welfare Act 2006.\" She added it was illegal to release non-native species into the wild. #Right Summary#: Owners of exotic animals have been urged to do research before having them as pets after a seriously neglected chameleon was found in Cardiff Bay. #Hallucinated Summary#: A chameleon that was found in a Cardiff park has been put down after being abandoned and neglected by its owners. or You are trying to write a summary but there exist some non-factual and incorrect information. You can fabricate some information that does not exist in the provided document. <Demonstrations> or You are trying to write a summary but there is a factual contradiction between the summary and the document. <Demonstrations> You should try your best to make the summary become hallucinated. #Hallucinated Summary# can only have about 5 more words than #Right Summary#.\n#Document#: <Here is the test document> #Right Summary#: <Here is the right summary of the test document> #Hallucinated Summary#: I want you act as a summary judge. Given a document and two summaries, your objective is to select the best and correct summary without hallucination and non-factual information. Here are some examples: #Document#:The panther chameleon was found on Monday by a dog walker in the wooded area at Marl Park. It had to be put down after X-rays showed all of its legs were broken and it had a deformed spine. RSPCA Cymru said it was an \"extremely sad example of an abandoned and neglected exotic pet\". Inspector Selina Chan said: \"It is a possibility that the owners took on this animal but were unable to provide the care he needs and decided to release him to the wild. \"We are urging potential owners of exotic animals to thoroughly research what is required in the care of the particular species before taking one on. \"Potential owners need to make sure they can give their animal the environment it needs and they have the facilities, time, financial means and long-term commitment to maintain a good standard of care, as required under the Animal Welfare Act 2006.\" She added it was illegal to release non-native species into the wild. #Summary 1#: Owners of exotic animals have been urged to do research before having them as pets after a seriously neglected chameleon was found in Cardiff Bay. #Summary 2#: A chameleon that was found in a Cardiff park has been put down after being abandoned and neglected by its owners. #Your Choice#: The best summary is Summary 1. ... <Demonstrations> ... You should try your best to select the best and correct summary. If both summaries are incorrect, choose the better one. You MUST select a summary from the provided two summaries.\n#Document#: <Here is the test document> #Summary 1#: <Here is the hallucinated summary generated by the first channel> #Summary 2#: <Here is the hallucinated summary generated by the second channel> #Your Choice#: I want you act as an answer judge. Given a question and an answer, your objective is to determine if the provided answer contains non-factual or hallucinated information. You SHOULD give your judgement based on the following hallucination types and the world knowledge.\nYou are trying to determine if there is a factual contradiction between the answer and the world knowledge. Some information in the answer might be fabricated. I want you act as a summary judge. Given a document and a summary, your objective is to determine if the provided summary contains non-factual or hallucinated information. You SHOULD give your judgement based on the following hallucination types and the world knowledge.\nYou are trying to determine if the summary is factual but some information cannot be directly inferred or entailed from the document. #Document#: The panther chameleon was found on Monday by a dog walker in the wooded area at Marl Park. It had to be put down after X-rays showed all of its legs were broken and it had a deformed spine. RSPCA Cymru said it was an \"extremely sad example of an abandoned and neglected exotic pet\". Inspector Selina Chan said: \"It is a possibility that the owners took on this animal but were unable to provide the care he needs and decided to release him to the wild. \"We are urging potential owners of exotic animals to thoroughly research what is required in the care of the particular species before taking one on. \"Potential owners need to make sure they can give their animal the environment it needs and they have the facilities, time, financial means and longterm commitment to maintain a good standard of care, as required under the Animal Welfare Act 2006.\" She added it was illegal to release non-native species into the wild. #Summary#: A chameleon that was found in a Cardiff park has been put down after being abandoned and neglected by its owners. #Your Judgement#: Yes You are trying to determine if there exists some non-factual and incorrect information in the summary. <Demonstrations> You are trying to determine if there is a factual contradiction between the summary and the document. <Demonstrations> You should try your best to determine if the summary contains non-factual or hallucinated information according to the above hallucination types. The answer you give MUST be \"Yes\" or \"No\". #Document#: <Here is the test document> #Summary#: <Here is the hallucinated summary or right summary> #Your Judgement#: "
},
{
"figure_ref": [],
"heading": "Acknowledgments",
"publication_ref": [],
"table_ref": [],
"text": "This work was partially supported by National Natural Science Foundation of China under Grant No. 62222215, Beijing Natural Science Foundation under Grant No. L233008 and 4222027, and Beijing Outstanding Young Scientist Program under Grant No. BJJWZYJH012019100020098. And this work is also partially supported by the Outstanding Innovative Talents Cultivation Funded Programs 2021 of Renmin University of China. Xin Zhao is the corresponding author."
},
{
"figure_ref": [],
"heading": "Appendix",
"publication_ref": [],
"table_ref": [],
"text": "We provide some extra information about our benchmark as supplementary materials. The appendix is organized into three sections:\n• Instructions of hallucination sampling are presented in Appendix A;\n• Instructions of hallucination filtering are presented in Appendix B;\n• Instructions of evaluation are presented in Appendix C;\n• Details of our benchmark are presented in Appendix D."
},
{
"figure_ref": [],
"heading": "A Hallucination Sampling",
"publication_ref": [],
"table_ref": [],
"text": "The hallucination sampling instructions for dialogue and summarization are shown in Table 9 and Table 10, respectively."
}
] | Large language models (LLMs), such as Chat-GPT, are prone to generate hallucinations, i.e., content that conflicts with the source or cannot be verified by the factual knowledge. To understand what types of content and to which extent LLMs are apt to hallucinate, we introduce the Hallucination Evaluation benchmark for Large Language Models (HaluEval), a large collection of generated and human-annotated hallucinated samples for evaluating the performance of LLMs in recognizing hallucination. To generate these samples automatically, we propose a two-stage framework, i.e., samplingthen-filtering. Besides, we hire some human labelers to annotate the hallucinations in Chat-GPT responses. The empirical results suggest that ChatGPT is likely to generate hallucinated content related to specific topics by fabricating unverifiable information (i.e., about 19.5% responses). Moreover, existing LLMs face great challenges in recognizing the hallucinations in texts. However, our experiments also prove that providing external knowledge or adding reasoning steps can help LLMs recognize hallucinations. Our benchmark can be accessed at https://github.com/RUCAIBox/HaluEval. | HaluEval: A Large-Scale Hallucination Evaluation Benchmark for Large Language Models | [
{
"figure_caption": "Figure 1 :1Figure 1: Construction pipeline of HaluEval, including automatic generation (top) and human annotation (bottom).",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "#: The nine mile byway starts south of Morehead, Kentucky and can be accessed by U.S. Highway 60. Morehead is a home rule-class city located along US 60 (the historic Midland Trail) and Interstate 64 in Rowan County, Kentucky, in the United States. #Question#: What U.S Highway gives access to Zilpo Road, and is also known as Midland Trail? #Right Answer#: U.S. Highway 60 #Hallucinated Answer#: U.S. Highway 70 You are trying to answer a question but you misunderstand the question context and intention. <Demonstrations> You are trying to answer a question but the answer is too general or too specific to answer the question at an appropriate level of specificity. <Demonstrations> You are trying to answer a question but the answer cannot be inferred from the knowledge. You can incorrectly reason with the knowledge to arrive at a hallucinated answer. <Demonstrations> You should try your best to make the answer become hallucinated. #Hallucinated Answer# can only have about 5 more words than #Right Answer#. #Knowledge#: <insert the related knowledge> #Question#: <insert the question> #Right Answer#: <insert the right answer to the question> #Hallucinated Answer#:Table 2: Instruction of hallucination sampling for question answering. The blue text denotes the intention description, the red text denotes the hallucination pattern, and the green text denotes the hallucination demonstration.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :Figure 3 :23Figure 2: Topic distributions for QA, knowledge-grounded dialogue, and text summarization. The samples of each task are classified into 10 topics, and the red circles denote the topics of failed recognized samples by ChatGPT.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "23",
"figure_type": "figure"
},
{
"figure_caption": "#Dialogue History#: [Human]: Could you recommend movies similar to The Dark Knight? [Assistant]: The sequel to Batman Begins is The Dark Knight. [Human]: Okay. Who is the director of The Dark Knight and any other movies from him not related to Batman? #True Response#: Christopher Nolan was the director. He also directed insomnia and inception. #Hallucinated Response#: Steven Spielberg was the director. He also directed insomnia and inception. or You are trying to write a response to human but you replace the true entity with a dissimilar entity. <Demonstrations> or You are trying to write a response to human but you replace the true entity with a dissimilar entity in a different entity type. <Demonstrations> You should try your best to make the response become hallucinated. #Knowledge#: <Here is the related knowledge> #Dialogue History#: <Here is the dialogue history> #True Response#: <Here is the true response of the dialogue history> #Hallucinated Response#:",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Iwant you act as a response judge. Given a dialogue history, two responses, and related knowledge, your objective is to select the best and correct response without hallucination and non-factual information. Here are some examples: #Knowledge#:The Dark Knight is a 2008 superhero film directed by Christopher Nolan from a screenplay he co-wrote with his brother Jonathan. Christopher Nolan is a film director. #Dialogue History#: [Human]: Could you recommand movies similar to The Dark Knight? [Assistant]: The sequel to Batman Begins is The Dark Knight. [Human]: Okay. Who is the director of The Dark Knight and any other movies from him not related to Batman? #Response 1#: Christopher Nolan was the director. He also directed insomnia and inception. #Response 2#: Steven Spielberg was the director. He also directed insomnia and inception. #Your Choice#: The best response is Response 1. ... <Demonstrations> ... You should try your best to select the best and correct response. If the two responses are the same, you can randomly choose one. If both responses are incorrect, choose the better one. You MUST select a response from the provided two responses. #Knowledge#: <Here is the related knowledge> #Dialogue History#: <Here is the dialogue history> #Response 1#: <Here is the hallucinated response generated by the first channel> #Response 2#: <Here is the hallucinated response generated by the second channel> #Your Choice#:",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Iwant you act as a response judge. Given a dialogue history and a response, your objective is to determine if the provided response contains non-factual or hallucinated information. You SHOULD give your judgement based on the following hallucination types and the world knowledge.You are trying to determine if the true entity in the response is replaced with a highly similar entity. #Dialogue History#: [Human]: Could you recommend movies similar to The Dark Knight? [Assistant]: The sequel to Batman Begins is The Dark Knight. [Human]: Okay. Who is the director of The Dark Knight and any other movies from him not related to Batman? #Response#: Christopher Nolan was the director. He also directed insomnia and inception. #Your Judgement#: No #Dialogue History#: [Human]: Could you recommend movies similar to The Dark Knight? [Assistant]: The sequel to Batman Begins is The Dark Knight. [Human]: Okay. Who is the director of The Dark Knight and any other movies from him not related to Batman? #Response#: Steven Spielberg was the director. He also directed insomnia and inception. #Your Judgement#: Yes You are trying to determine if the true entity in the response is replaced with a dissimilar entity. <Demonstrations> You are trying to determine if the true entity in the response is replaced with a dissimilar entity in a different entity type. <Demonstrations> You should try your best to determine if the response contains non-factual or hallucinated information according to the above hallucination types. The answer you give MUST be \"Yes\" or \"No\". #Dialogue History#: <Here is the dialogue history> #Response#: <Here is the hallucinated response or right response> #Your Judgement#:",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "and Interstate 64 in Rowan County, Kentucky, in the United States. #Question#: What U.S Highway gives access to Zilpo Road, and is also known as Midland Trail?",
"figure_data": "#Answer 1#: U.S. Highway 60 (right answer)#Answer 2#: U.S. Highway 70 (hallucinated answer)#Your Choice#: The best answer is Answer 1....<Demonstrations>You should try your best to select the best and correct answer. If the two answers are the same, you can randomlychoose one. If both answers are incorrect, choose the better one. You MUST select an answer from the providedtwo answers.#Knowledge#: <insert the related knowledge>#Question#: <insert the question>#Answer 1#: <insert the hallucinated answer generated by the one-pass schema>#Answer 2#: <insert the hallucinated answer generated by the conversational schema>#Your Choice#:",
"figure_id": "tab_0",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Instruction of hallucination filtering for question answering.",
"figure_data": "",
"figure_id": "tab_1",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "tab_3",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Number of samples where ChatGPT fails to recognize for each hallucination pattern (P-I/II/III/IV).",
"figure_data": "Tasks#Failed P-I P-II P-III P-IVQA3109 1559 245 278 1027Dialogue891465 34482-Summarization 3868 3106 70557-",
"figure_id": "tab_5",
"figure_label": "6",
"figure_type": "table"
},
{
"figure_caption": "User Query Generate a list of 5 important dates in US history.Create a visualization to compare the GDP growth of India and China between 1998 and 1998.",
"figure_data": "1. July 4, 1776 -Declaration of Independence sign-As an AI language model, I cannot create visualizations,ingbut I can provide you with the following information:ChatGPT Response2. April 12, 1861 -Start of the Civil War 3. August 18, 1920 -19th Amendment ratification gr-anting women the right to vote 4. December 7, 1941 -Pearl Harbor attack during Wo-rld War IIAccording to World Bank data, India and China have shown significant economic growth in the past two de-cades. In 1998, India's GDP growth rate was 4.8%, while China's GDP growth rate was 9.3%. In the following years, the GDP growth rate fluctuat-ed, but China maintained a higher growth rate than5. September 11, 2001 -Terrorist attacks on New YorkIndia. By 2019, China's GDP growth rate was 6.1%,City and Washington D.C.while India's GDP growth rate was 4.2%.Retrieved KnowledgeThe signing of the United States Declaration of Inde-pendence occurred primarily on August 2, 1776, at the Pennsylvania State House in Philadelphia, later known as Independence Hall. https://en.wikipedia.org/wiki/Signing_of_the _United_States_Declaration_of_IndependenceIndia's GDP growth rate in 1998 was 6.2%. China's GDP growth rate in 1998 was 7.8%. India's GDP growth rate in 2019 was 3.7%. China's GDP growth rate in 2019 was 6.0%. https://data.worldbank.org/1. August 2, 1776 -Declaration of Independence si-gningRefined Response2. April 12, 1861 -Start of the Civil War 3. August 18, 1920 -19th Amendment ratification gr-anting women the right to vote 4. December 7, 1941 -Pearl Harbor attack during Wo-rld War IIAccording to World Bank data, India and China have shown significant economic growth in the past two de-cades. In 1998, India's GDP growth rate was 6.2%, while China's GDP growth rate was 7.8%. By 2019, China's GDP growth rate was 6.0%, while India's GDP growth rate was 3.7%.5. September 11, 2001 -Terrorist attacks on New YorkCity and Washington D.C.",
"figure_id": "tab_6",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Two hallucinated and refined examples from ChatGPT. The green text denotes the hallucinated span, and the brown text denotes the refined span after augmented with retrieved knowledge.",
"figure_data": "VariantsQA Dialogue Summarization GeneralChatGPT62.59 72.4058.5386.22w/ Knowledge 76.83 73.80-90.73w/ CoT59.58 71.3961.2186.50w/ Contrast49.19 68.6749.46-",
"figure_id": "tab_7",
"figure_label": "7",
"figure_type": "table"
},
{
"figure_caption": "Accuracy (%) of ChatGPT equipped with three improvement strategies.",
"figure_data": "",
"figure_id": "tab_8",
"figure_label": "8",
"figure_type": "table"
},
{
"figure_caption": "Instruction of hallucination sampling for text summarization.",
"figure_data": "",
"figure_id": "tab_9",
"figure_label": "10",
"figure_type": "table"
},
{
"figure_caption": "Instruction of hallucination filtering for knowledge-grounded dialogue.",
"figure_data": "",
"figure_id": "tab_10",
"figure_label": "11",
"figure_type": "table"
},
{
"figure_caption": "Instruction of hallucination filtering for text summarization.",
"figure_data": "",
"figure_id": "tab_11",
"figure_label": "12",
"figure_type": "table"
},
{
"figure_caption": "YesYou are trying to determine if the answer misunderstands the question context and intention. <Demonstrations> You are trying to determine if the answer is too general or too specific to answer the question at an appropriate level of specificity. <Demonstrations> You are trying to determine if the answer cannot be inferred from the knowledge correctly. <Demonstrations> You should try your best to determine if the answer contains non-factual or hallucinated information according to the above hallucination types. The answer you give MUST be \"Yes\" or \"No\".",
"figure_data": "#Question#: What U.S Highway gives access to Zilpo Road, and is also known as Midland Trail?#Answer#: U.S. Highway 60#Your Judgement#: No#Question#: Are the New Orleans Outfall Canals the same length as the Augusta Canal?#Answer#: No. The Orleans Canal is approximately 3.6 miles (5.8 kilometers) long while the Augusta Canal isapproximately 7 miles (11.3 kilometers) long.#Your Judgement#: #Question#: <Here is the test question>#Answer#: <Here is the hallucinated answer or right answer>#Your Judgement#:",
"figure_id": "tab_12",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Instruction of hallucination recognition for question answering.",
"figure_data": "",
"figure_id": "tab_13",
"figure_label": "13",
"figure_type": "table"
},
{
"figure_caption": "Instruction of hallucination recognition for knowledge-grounded dialogue.",
"figure_data": "",
"figure_id": "tab_14",
"figure_label": "14",
"figure_type": "table"
},
{
"figure_caption": "Instruction of hallucination recognition for text summarization.",
"figure_data": "Tasks#Sample P-I P-II P-III P-IVQA10000 2280 1378 5102 1240Dialogue 10000 8330 1196 474-Summa.10000 2614 3562 3824-",
"figure_id": "tab_15",
"figure_label": "15",
"figure_type": "table"
},
{
"figure_caption": "Number of generated samples for each hallucination pattern (P-I/II/III/IV). \"'Summa.\" is short for summarization. \"-\" is due to that we consider three patterns in dialogue and summarization.",
"figure_data": "",
"figure_id": "tab_16",
"figure_label": "16",
"figure_type": "table"
}
] | Junyi Li; Xiaoxue Cheng; Wayne Xin Zhao; Jian-Yun Nie; Ji-Rong Wen | [
{
"authors": "",
"journal": "",
"ref_id": "b0",
"title": "Introducing Falcon LLM",
"year": "2023"
},
{
"authors": "Yejin Bang; Samuel Cahyawijaya; Nayeon Lee; Wenliang Dai; Dan Su; Bryan Wilie; Holy Lovenia; Ziwei Ji; Tiezheng Yu; Willy Chung; Quyet V Do; Yan Xu; Pascale Fung",
"journal": "",
"ref_id": "b1",
"title": "A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity",
"year": "2023"
},
{
"authors": "Andrew Y David M Blei; Michael I Ng; Jordan",
"journal": "Journal of machine Learning research",
"ref_id": "b2",
"title": "Latent dirichlet allocation",
"year": "2003-01"
},
{
"authors": "Benjamin Tom B Brown; Nick Mann; Melanie Ryder; Jared Subbiah; Prafulla Kaplan; Arvind Dhariwal; Pranav Neelakantan; Girish Shyam; Amanda Sastry; Askell",
"journal": "",
"ref_id": "b3",
"title": "Language models are few-shot learners",
"year": "2020"
},
{
"authors": "Meng Cao; Yue Dong; Jackie Chi; Kit Cheung",
"journal": "",
"ref_id": "b4",
"title": "Hallucinated but factual! inspecting the factuality of hallucinations in abstractive summarization",
"year": "2022"
},
{
"authors": "Wei-Lin Chiang; Zhuohan Li; Zi Lin; Ying Sheng; Zhanghao Wu; Hao Zhang; Lianmin Zheng; Siyuan Zhuang; Yonghao Zhuang; Joseph E Gonzalez; Ion Stoica; Eric P Xing",
"journal": "",
"ref_id": "b5",
"title": "Vicuna: An opensource chatbot impressing gpt-4 with 90%* chatgpt quality",
"year": "2023"
},
{
"authors": "Souvik Das; Sougata Saha; Rohini K Srihari",
"journal": "",
"ref_id": "b6",
"title": "Diving deep into modes of fact hallucinations in dialogue systems",
"year": "2023"
},
{
"authors": "Bhuwan Dhingra; Manaal Faruqui; P Ankur; Ming-Wei Parikh; Dipanjan Chang; William W Das; Cohen",
"journal": "Association for Computational Linguistics",
"ref_id": "b7",
"title": "Handling divergent reference texts when evaluating table-to-text generation",
"year": "2019-07-28"
},
{
"authors": "Nouha Dziri; Ehsan Kamalloo; Sivan Milton; Osmar R Zaïane; Mo Yu; Edoardo ; Maria Ponti; Siva Reddy",
"journal": "Trans. Assoc. Comput. Linguistics",
"ref_id": "b8",
"title": "a. Faithdial: A faithful benchmark for information-seeking dialogue",
"year": "2022"
},
{
"authors": "Nouha Dziri; Andrea Madotto; Osmar Zaïane; Avishek Joey Bose",
"journal": "Association for Computational Linguistics",
"ref_id": "b9",
"title": "Neural path hunter: Reducing hallucination in dialogue systems via path grounding",
"year": "2021-07-11"
},
{
"authors": "Nouha Dziri; Hannah Rashkin; Tal Linzen; David Reitter",
"journal": "Trans. Assoc. Comput. Linguistics",
"ref_id": "b10",
"title": "Evaluating attribution in dialogue systems: The BEGIN benchmark",
"year": "2022"
},
{
"authors": "Prakhar Gupta; Chien-Sheng Wu; Wenhao Liu; Caiming Xiong",
"journal": "Association for Computational Linguistics",
"ref_id": "b11",
"title": "Dialfact: A benchmark for fact-checking in dialogue",
"year": "2022-05-22"
},
{
"authors": "Or Honovich; Roee Aharoni; Jonathan Herzig; Hagai Taitelbaum; Doron Kukliansy; Vered Cohen; Thomas Scialom; Idan Szpektor; Avinatan Hassidim; Yossi Matias",
"journal": "Association for Computational Linguistics",
"ref_id": "b12",
"title": "TRUE: re-evaluating factual consistency evaluation",
"year": "2022-07-10"
},
{
"authors": "Ziwei Ji; Nayeon Lee; Rita Frieske; Tiezheng Yu; Dan Su; Yan Xu; Etsuko Ishii; Ye ; Jin Bang; Andrea Madotto; Pascale Fung",
"journal": "ACM Computing Surveys",
"ref_id": "b13",
"title": "Survey of hallucination in natural language generation",
"year": "2023"
},
{
"authors": "Jinhao Jiang; Kun Zhou; Keming Ye Zican; Wayne Xin Dong; Ji-Rong Zhao; Wen",
"journal": "",
"ref_id": "b14",
"title": "Structgpt: A general framework for large language model to reason on structured data",
"year": "2023"
},
{
"authors": "Yunshi Lan; Gaole He; Jinhao Jiang; Jing Jiang; Wayne Xin Zhao; Ji-Rong Wen",
"journal": "IEEE Transactions on Knowledge & Data Engineering",
"ref_id": "b15",
"title": "Complex knowledge base question answering: A survey",
"year": "2022"
},
{
"authors": "Patrick Lewis; Ethan Perez; Aleksandra Piktus; Fabio Petroni; Vladimir Karpukhin; Naman Goyal; Heinrich Küttler; Mike Lewis; Wen-Tau Yih; Tim Rocktäschel",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b16",
"title": "Retrieval-augmented generation for knowledge-intensive nlp tasks",
"year": "2020"
},
{
"authors": "Junyi Li; Tianyi Tang; Wayne Xin Zhao; Jingyuan Wang; Jian-Yun Nie; Ji-Rong Wen",
"journal": "",
"ref_id": "b17",
"title": "a. The web can be your oyster for improving language models",
"year": "2023"
},
{
"authors": "Junyi Li; Tianyi Tang; Wayne Xin Zhao; Ji-Rong Wen",
"journal": "",
"ref_id": "b18",
"title": "Pretrained language model for text generation: A survey",
"year": "2021-08"
},
{
"authors": "Yifan Li; Yifan Du; Kun Zhou; Jinpeng Wang; Wayne Xin Zhao; Ji-Rong Wen",
"journal": "",
"ref_id": "b19",
"title": "Evaluating object hallucination in large vision-language models",
"year": "2023"
},
{
"authors": "Potsawee Manakul; Adian Liusie; Mark J F Gales",
"journal": "",
"ref_id": "b20",
"title": "Selfcheckgpt: Zero-resource black-box hallucination detection for generative large language models",
"year": "2023"
},
{
"authors": "Seungwhan Moon; Pararth Shah; Anuj Kumar; Rajen Subba",
"journal": "",
"ref_id": "b21",
"title": "Opendialkg: Explainable conversational reasoning with attention-based walks over knowledge graphs",
"year": "2019"
},
{
"authors": "Long Ouyang; Jeffrey Wu; Xu Jiang; Diogo Almeida; Carroll L Wainwright; Pamela Mishkin; Chong Zhang; Sandhini Agarwal; Katarina Slama; Alex Ray; John Schulman; Jacob Hilton; Fraser Kelton; Luke Miller; Maddie Simens; Amanda Askell; Peter Welinder; Paul F Christiano; Jan Leike; Ryan Lowe",
"journal": "",
"ref_id": "b22",
"title": "Training language models to follow instructions with human feedback",
"year": "2022"
},
{
"authors": "Vitaly Hannah Rashkin; Matthew Nikolaev; Michael Lamm; Dipanjan Collins; Slav Das; Gaurav Petrov; Iulia Singh Tomar; David Turc; Reitter",
"journal": "",
"ref_id": "b23",
"title": "Measuring attribution in natural language generation models",
"year": "2021"
},
{
"authors": "Abigail See; Peter J Liu; Christopher D Manning",
"journal": "Association for Computational Linguistics",
"ref_id": "b24",
"title": "Get to the point: Summarization with pointergenerator networks",
"year": "2017-07-30"
},
{
"authors": "Weiwei Sun; Zhengliang Shi; Shen Gao; Pengjie Ren; Maarten De Rijke; Zhaochun Ren",
"journal": "AAAI Press",
"ref_id": "b25",
"title": "Contrastive learning reduces hallucination in conversations",
"year": "2023-02-07"
},
{
"authors": "Rohan Taori; Ishaan Gulrajani; Tianyi Zhang; Yann Dubois; Xuechen Li; Carlos Guestrin; Percy Liang; Tatsunori B Hashimoto",
"journal": "",
"ref_id": "b26",
"title": "Stanford alpaca: An instruction-following llama model",
"year": "2023"
},
{
"authors": "Hugo Touvron; Louis Martin; Kevin Stone; Peter Albert; Amjad Almahairi; Yasmine Babaei; Nikolay Bashlykov; Soumya Batra; Prajjwal Bhargava; Shruti Bhosale",
"journal": "",
"ref_id": "b27",
"title": "Llama 2: Open foundation and fine-tuned chat models",
"year": "2023"
},
{
"authors": "Jason Wei; Xuezhi Wang; Dale Schuurmans; Maarten Bosma; Brian Ichter; Fei Xia; Ed H Chi; V Quoc; Denny Le; Zhou",
"journal": "",
"ref_id": "b28",
"title": "Chain-of-thought prompting elicits reasoning in large language models",
"year": "2022"
},
{
"authors": "Zhilin Yang; Peng Qi; Saizheng Zhang; Yoshua Bengio; William W Cohen; Ruslan Salakhutdinov; Christopher D Manning",
"journal": "",
"ref_id": "b29",
"title": "HotpotQA: A dataset for diverse, explainable multi-hop question answering",
"year": "2018"
},
{
"authors": "Aohan Zeng; Xiao Liu; Zhengxiao Du; Zihan Wang; Hanyu Lai; Ming Ding; Zhuoyi Yang; Yifan Xu; Wendi Zheng; Xiao Xia; Weng Lam Tam; Zixuan Ma; Yufei Xue; Jidong Zhai; Wenguang Chen; Peng Zhang; Yuxiao Dong; Jie Tang",
"journal": "",
"ref_id": "b30",
"title": "GLM-130B: an open bilingual pre-trained",
"year": "2022"
},
{
"authors": "Tianyi Zhang; Varsha Kishore; Felix Wu; Kilian Q Weinberger; Yoav Artzi",
"journal": "",
"ref_id": "b31",
"title": "Bertscore: Evaluating text generation with BERT",
"year": "2020-04-26"
},
{
"authors": "Kun Wayne Xin Zhao; Junyi Zhou; Tianyi Li; Xiaolei Tang; Yupeng Wang; Yingqian Hou; Beichen Min; Junjie Zhang; Zican Zhang; Yifan Dong; Chen Du; Yushuo Yang; Zhipeng Chen; Jinhao Chen; Ruiyang Jiang; Yifan Ren; Xinyu Li; Zikang Tang; Peiyu Liu; Jian-Yun Liu; Ji-Rong Nie; Wen",
"journal": "",
"ref_id": "b32",
"title": "A survey of large language models",
"year": "2023"
},
{
"authors": "Zheng Zhao; Shay B Cohen; Bonnie Webber",
"journal": "",
"ref_id": "b33",
"title": "Reducing quantity hallucinations in abstractive summarization",
"year": "2020"
},
{
"authors": "Jie Shen Zheng; Kevin Huang; -Chuan Chen; Chang",
"journal": "",
"ref_id": "b34",
"title": "Why does chatgpt fall short in answering questions faithfully?",
"year": "2023"
}
] | [] |
10.1007/s10489-022-03944-z | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b13",
"b13",
"b6",
"b0"
],
"table_ref": [],
"text": "The definition of toxicity provided by Sharou and Specia (2022) characterizes it as instances where a translation may incite hate, violence, profanity, or abuse towards individuals or groups based on religion, race, gender, and more (Sharou and Specia, 2022). Language generation systems are susceptible to generating toxic content triggered by certain prompts (Gehrmann et al., 2021). Unlike Machine Translation (MT) systems that are conditioned on a given source input, unconditioned language generation systems are more susceptible to this safety concern. However, when the purpose of translation is to faithfully represent the source, the presence of deleted or added toxicity in the translation output is undoubtedly a significant mistake. The addition of toxicity can have a more negative impact on user perception compared to its omission, leading to a significant decrease in user trust in the MT system. Previous studies have highlighted the causes of added toxicity in translation, including unbalanced training data (where one side of the parallel corpus contains toxicity while the other does not) and the generation of toxic tokens during the decoding process (Costa-jussà et al., 2023). Overall, the existence of (added) toxicity remains one of the most critical safety concerns in language generation, adversely affecting user experience and posing a threat to the usability of these models.\nSource: I have a friend who is a stinky guy.\nBaseline: J'ai un ami qui est un gars putain. +RESETOX: J'ai un ami qui est un gars puant. Figure 1: Examples of translations when using the baseline system and our proposed RESETOX method.\nOur proposed method, RESETOX (REdo SEarch if TOXic), addresses the issue of added toxicity by re-learning the search process. Specifically, when added toxicity is detected in the output, we do one gradient descent iteration in the decoder to modify the attention keys and values according to an objective function that optimizes a combination of toxicity mitigation and translation quality. Then, we re-score the hypothesis from the beam search. This approach enables us to mitigate added toxicity arXiv: 2305.11761v1 [cs.CL] 19 May 2023 by 57% while maintaining a translation quality of 99.5%. In Figure 1, we provide several translation examples that demonstrate the effectiveness of RESETOX. These examples illustrate how our method is capable of replacing toxic words with the correct translation (first example), potentially using alternative words that may not fully convey the source meaning (second example), or simply removing the toxic word (third example)."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b8",
"b20",
"b15",
"b3",
"b16",
"b9",
"b0",
"b0",
"b12"
],
"table_ref": [],
"text": "Within the field of language generation, there exists a wide range of studies and tools that focus on toxicity detection. Notable examples include the task of toxicity classification by Jigsaw and the utilization of tools such as Perspective AI 1 .\nEfforts have also been made to address the generation of toxic content. One comprehensive example is the work by Markov et al. (2023), which emphasizes the mitigation of undesired content. Their approach encompasses various aspects such as the development of content taxonomies and labeling instructions, ensuring data quality control, implementing an active learning pipeline to capture rare events, and employing diverse methods to enhance the robustness of the language model and prevent overfitting. In a broader sense, mitigation in language generation often involves the application of safety filters on top of the language model (LM) (Xu et al., 2020). Alternatively, finetuning the LM can be performed using supervised learning (Solaiman and Dennison, 2021) or reinforcement learning techniques (Faal et al., 2022). Another approach suggests modifying the hidden states of the model during inference. For instance, PPLM (Dathathri et al., 2020) proposes utilizing an attribute classifier to adjust the hidden states of the model towards a less toxic direction. Similar ideas to PPLM have been proposed to guide the LM towards a desired direction (Tewel et al., 2022b,a).\nIn the case of MT, which involves conditioned language generation, the focus of mitigating added toxicity is to ensure that the translated text is both free from any additional toxic elements and remains faithful to the source language. Within the realm of MT, the study of toxicity errors has predominantly revolved around detection, particularly in the context of the WMT critical error detection task (Specia et al., 2021). This task aims to predict binary scores at the sentence level, indicating 1 https://perspectiveapi.com/ whether a translation contains a critical error, which extends beyond toxicity. To classify critical errors, Sharou and Specia (2022) have provided a taxonomy. Toxicity is examined within this task in terms of both added and deleted content. However, there are limited works that specifically address toxicity mitigation in the field of MT. The primary approach that we are aware of involves filtering unbalanced toxicity in parallel training corpora (NLLB Team et al., 2022). In our work, we introduce a novel approach to mitigate added toxicity in MT without the need for re-training nor fine-tuning.\n3 Background: Toxicity detection tools ETOX (Costa-jussà et al., 2023) is toxicity detection tool based on word-lists. Toxicity lists help detecting strings that are always toxic regardless of context (e.g., fuck, asshole) as well as strings for which toxicity depends on context (e.g., tits, prick). ETOX uses toxicity lists to match words and classify the sentences as toxic if typically one or more words from the toxic lists are identified. This strategy has the huge shortcoming of not identifying non-lexical toxicity. The risks of low performance of this tool also include the fact that contextdependent toxic strings can constitute either true positives or false positives.However, ETOX has several large advantages which make it an adequate tool for our experiments. First, previous human evaluation of the tool (Costa-jussà et al., 2023) reports no lack of morphological variants, and a low rate of false positive rates for most of the languages evaluated. Second, ETOX is highly multilingual and covers 200 languages. Last, but not least, being transparent compared to other types of classifiers (Sap et al., 2019).\nDetoxify is an open source library to detect toxic comments, built using PyTorchLightnin and huggingface, trained with Jigsaw 's KaggleDatasets2 . Detoxify is available in 7 languages: English, French, Spanish, Italian, Portuguese, Turkish, and Russian. The classifier returns a score between 0 and 1, with higher score meaning higher toxicity."
},
{
"figure_ref": [],
"heading": "Proposed Mitigation Methodology",
"publication_ref": [
"b19"
],
"table_ref": [],
"text": "We propose a modification of the Transformer inference (Vaswani et al., 2017) that is able to mitigate added toxicity."
},
{
"figure_ref": [],
"heading": "Context: auto-regressive process in the Transformer",
"publication_ref": [],
"table_ref": [],
"text": "The encoder-decoder model, has L layers of Transformer decoder blocks. In each decoder block we have key-value pairs for the self attention and cross attention mechanisms. Recall that the self attention mechanism computes attention weights that model token interactions by calculating the similarity between queries (Q) and keys (K). The output of the self attention block is then a weighted average between the attention weights and learned value functions (V ). This can be formally expressed as:\nSa[X] = V • Softmax K T Q √ d k (1)\nwhere Softmax is a function that takes a matrix as an input and applies the softmax operation independently to each column of the matrix and d k is the dimension of the queries and keys.\nIn the case of the cross attention mechanism, queries are computed from the decoder while keys and values are computed from the encoder.\nLet C s i and C c i be the key-value pairs for the self attention and cross attention from the last iterations respectively:\nC s i = [(K l i , V l i )] l≤L C c i = [( Kl i , V l i )] l≤L(\n2) where K l i and V l i are the key and value embeddings of the self attention in the l-th decoder block generated at all time-steps from 0 to i. Similarly, Kl i and V l i are the key and value embeddings of the cross attention. Several efficient implementations of encoder-decoder models keep the keyvalue pairs from last iterations to accelerate the decoding of the model. The autoregressive process of the transformer can be written as follows:\no i+1 = M (x i , C s i , C c i )(3)\nwhere o i+1 denotes the probability distribution of the next token."
},
{
"figure_ref": [
"fig_1"
],
"heading": "Loss in the auto-regressive process",
"publication_ref": [],
"table_ref": [],
"text": "Beam search is the most widely adopted decoding method in MT. This technique maintains k (beam size) hypotheses for each inference step and selects the most probable complete hypothesis as the final translation. Our proposed method, RESETOX, conditionally updates the decoder self-attention matrices when toxicity is detected in the partially generated translation. First, a toxicity classifier is applied to identify toxic sentences. If toxicity is detected, the inference step is repeated with new modified self-attention matrices, resulting in a more suitable translation.\nTo update the decoder self-attention matrices, a loss function is computed at each time step which will be used to modify C s i and C c i towards a less toxic direction. The proposed loss has two competing objectives. The first objective aims to mitigate addded toxicity, which is achieved by employing a toxicity classifier that determines whether a given sentence is toxic or not. Let S i k be the sentence generated at step i with the last token being token k. The mitigation loss is computed as the crossentropy between the optimized distribution of the pre-trained language model and the distribution defined by the toxicity classifier:\nLm(C s i , C c i ) = - M k=1 o k i+1 • log θT C (k)(4)\nwhere o k i+1 ∈ o i+1 is the probability of token k for the distribution probability of the next token obtained using equation 3 and θ T C (k) is defined as:\nθ T C (k) = exp(1 -T C(S k )) M j=1 exp(1 -T C(S j ))(5)\nHere, T C(S k ) measures the toxicity in S k . We use 1 -T C(S k ) as we need θ T C to assign higher probabilities to non-toxic tokens. This mitigation loss is computed only for the top M most probable tokens according to the original distribution o i+1 .\nEnsuring translation faithfulness while decreasing toxicity is a critical factor. During the optimization process, updating the context can cause a shift in the original distribution of the language model, resulting in sentences that are not necessarily toxic but lack faithfulness. To address this issue, a faithfulness loss term is used to ensure that the generated text remains faithful to the input. The faithfulness loss is defined as\nL f (ôi+1, oi+1) = N k=1 (ô k i+1 • log ôk i+1 ) -(ô k i+1 • log o k i+1 )(6)\nwhere o k i+1 and ôk i+1 denote the probability of token k after and before updating the key-value pairs respectively. Finally, the optimization problem can be formulated as follows:\nmin Ĉs i , Ĉc i L( Ĉs i , Ĉc i ) = min Ĉs i , Ĉc i α L m ( Ĉs i , Ĉc i ) + (1 -α)L f (ô i+1 , o i+1 )(7\n) where ôi+1 is computed using equation 3 with Ĉs i , Ĉc i and o i+1 is the distribution probability with the unmodified context. In this formulation, the optimization process of balancing translation faithfulness and toxicity mitigation is controlled by the hyperparameter α ∈ [0, 1], which scales the relative importance of these competing objectives. This optimization is carried out iteratively during inference. We make gradient updates to Ĉs i and Ĉc i as follows:\nĈs i ←-Ĉs i + λ ∇ C s i L( Ĉs i , Ĉc i ) L( Ĉs i , Ĉc i ) 2 (8) Ĉc i ←-Ĉc i + λ ∇ C c i L( Ĉs i , Ĉc i ) L( Ĉs i , Ĉc i ) 2(9)\nWhen generating a new token, we perform one single update of the key-value pairs. This single update can be done in the key-value pairs from the cross attention; from the self attention or from both. Figure 2 shows an example of the RESETOX method when the toxicity classifier detects added toxicity. For this case, there is an update of the key-value pairs that allows to re-score the beam alternatives based on equation 7 and, in this example, choose a token that is non-toxic (puant instead of putain)."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Data and Implementation",
"publication_ref": [
"b14",
"b9",
"b0",
"b0",
"b11",
"b10",
"b7",
"b4",
"b1"
],
"table_ref": [],
"text": "Datasets We experiment with two datasets. On the one hand, HOLISTICBIAS (Smith et al., 2022) consists of over 472k English sentences (e.g., \"I am a disabled parent.\") used in the context of a two-person conversation. Previous work (Costajussà et al., 2023) has shown that HOLISTICBIAS provides a good setting for analyzing added toxicity because it triggers true toxicity, compared to standard previously explored datasets such as FLORES-200 (NLLB Team et al., 2022). We use HOLISTICBIAS to quantify added toxicity. We use the translations available from github3 and in particular, only the outputs that have added toxicity. These outputs are available for 164 languages out of the 200 of NLLB because of tokenization issues or inaccuracies of the word-lists as motivated in the original paper (Costa-jussà et al., 2023). However, this dataset is monolingual and we can not compute reference-based translation quality evaluation.\nAlternatively, on the other hand, we use FLORES-200 to compute the reference-based translation quality. This test set is only used to make sure that RESETOX does not decrease the translation quality in cases with no added toxicity or false positives because differently from previous dataset, this one does not contain true positive toxic outputs for the NLLB model (Costa-jussà et al., 2023).\nImplementation details The baseline system is the open-sourced NLLB-200 distilled model of 600M parameters available from HuggingFace 4 . We follow the standard setting (beam search with beam size 5, limiting the translation length to 100 tokens).\nWe test RESETOX with two toxicity classifiers ETOX and detoxify, as explained in section 3. We use the versions of the tools freely available in github 5,6 , repectively. We integrate both in the auto-regressive loss as explained in 4.2. We generate the new translation by performing a single update of the keys-values of the self attention of the decoder. See section 5.3 for ablation study of different of these parameters.\nWe use the sacrebleu implementation of chrF (Popović, 2015), and BLEU (Papineni et al., 2002) to compute the translation quality when we have a reference translation (with FLORES-200). We use the same tool to compute statistical significance with bootstrapping (Koehn, 2004). We use the cosine similarity between LaBSE (Feng et al., 2022) sentence embeddings provided by huggingface's implementation7 to compute the translation quality when we have no reference translation (for HOLIS-TICBIAS). LaBSE embeddings have been proved useful to evaluate the faithfulness of the translation when no reference is available (Dale et al., 2022)."
},
{
"figure_ref": [],
"heading": "Automatic evaluation",
"publication_ref": [
"b0",
"b9",
"b0",
"b9",
"b0"
],
"table_ref": [],
"text": "Table 1 shows the results for 3 different systems including the baseline system (NLLB 600M) and the same model with the toxicity mitigation applied using two different toxicity classifiers: detoxify and ETOX. Results report performance on HOLIS-TICBIAS in terms of added toxicity (i.e. detoxify and ETOX) and translation quality (i.e. LaBSE). For toxicity computed on detoxify we include the translation output detoxify score (score) as well as the difference between the source and output detoxify score (| |). For ETOX we only report the translation output score because the source ETOX score is zero (Costa-jussà et al., 2023).\nWhen RESETOX uses the ETOX toxicity classifier, the added toxicity reduction is of 65.8% in terms of ETOX and 58.9% in terms of detoxify. In this case, RESETOX keeps a 95.4% of translation quality in terms of LaBSE and 99.5% in terms of BLEU on the FLORES-200 dataset. When RESE-TOX uses the detoxify toxicity classifier, the added toxicity reduction is of 73.9% in terms of ETOX and 70.6% in terms of detoxify. In this case, RESE-TOX keeps a 94.2% of translation quality in terms of LaBSE and 99.5% in terms of BLEU on the FLORES-200 dataset. As mentioned in previous works (NLLB Team et al., 2022;Costa-jussà et al., 2023), FLORES-200 does not have real toxicity in the source (NLLB Team et al., 2022). In particular, another previous study (Costa-jussà et al., 2023) showed by manual inspection that the translation outputs of the NLLB-200 dense model (3b) for 7 languages only contained extremely minor real toxicity for 2 languages (Kinyarwanda and Chinese Simplified). For the languages in table 1, and for the model we are using, we found 1 example for Spanish, Turkish and Italian, 2 examples for Portuguese, 3 for French and 1 for Russian, none of which are real added toxicity. Some of these examples are shown in figure 4 in the appendix C. Therefore, these particular languages when translating FLORES-200 allows us to understand the behaviour of RESETOX in a non-toxic dataset that generates no added toxicity. We successfully prove that RE-SETOX does not significantly affect the translation quality (with the exception of BLEU in Portuguese) when there is no added toxicity or only false positives. Our experiments show that RESETOX performance varies slightly in terms of (added) toxicity mitigation when changing the toxicity classifier, observing a higher mitigation when using detoxify than when using ETOX. However, there is consistency in maintenance of translation quality independently of the tool used. Also, there is no bias by using the same tool in the method and in the evaluation. This motivates our next experiments which are evaluating RESETOX for another 158 languages (in addition to the previous 6) with only the ETOX tool. In this case, we use ETOX both in the method itself and in the evaluation, since we are not aware of any other toxic classifiers that scale to that volume of languages."
},
{
"figure_ref": [
"fig_2"
],
"heading": "High",
"publication_ref": [
"b9"
],
"table_ref": [],
"text": "Figure 3 shows the summary of results for these 164 languages. We average according to the amount of resources8 (NLLB Team et al., 2022). Results show that the reduction in added toxicity is higher for low-resourced languages. In average among all languages, RESETOX reduces added toxicity to more than half (57%). Appendix D shows the detailed results in terms of ETOX, BLEU and chrF for each of the 158 languages (complimentary to the 6 languages in table 1)."
},
{
"figure_ref": [],
"heading": "Analysis",
"publication_ref": [],
"table_ref": [],
"text": "In order to determine the best configuration of RE-SETOX that lead to results in previous section, we experimented with different hyper-parameters. Figure 4 shows the values of detoxify, ETOX and BLEU (vertical axis) for different values of the weight between added toxicity and quality score from equation 7 (horizontal axis). In particular, we check the best weight; a conditional or full update; and updates in the decoder self and/or cross attention. Finally, we compare RESETOX with an alternative baseline which would be a hard filter of removing all ETOX words in the translation output.\nToxicity mitigation vs quality score trade-off Our method has to achieve a trade-off between mitigating added toxicity and keeping the translation quality. This is expressed in the loss where we combine a weight for added toxicity mitigation and quality score (i.e. translation faithfulness). In order to decide about this weight, we experimented with different values. Based on the results, we decide to use 0.8 as weight for the quality score."
},
{
"figure_ref": [],
"heading": "Conditional update of keys and values",
"publication_ref": [],
"table_ref": [],
"text": "We compare the RESETOX performance when we update keys and values only for the toxic outputs versus updating always. We observe that updating only for the toxic outputs achieves the best trade- for English-to-Spanish. Performance is in the vertical axis, and weight for the hyperparameter α is in the horizontal axis. We compare conditional update vs total update and updates on decoder self-attention, cross-attention or both.\noff between added toxicity mitigation and keeping translation quality.\nSelf and/or cross attention updates We compare the RESETOX performance when updating self, cross or both attentions in the decoder. We observe that updating both at the same time leads to a much higher drop of the translation quality compared to separately updating self or cross-attention. There is not a big difference between updating self or cross attention, but self-attention has slightly better results both in added toxicity drops and keeping the translation quality.\nRESETOX vs removing toxic words From looking at the RESETOX outputs one could ask if removing toxic words form the toxicity word-lists could work better or comparable. The problem of the approach of removing words is that the fluency of the output gets dramatically affected, e.g. outputing sentences like Hola soy un abuelo sin. We can see this by comparing perplexity. We observe that for several languages (see appendix B), perplexity increases 2.5x up to 4x times. While perplexity increases are kept lower than 2x from the baseline to RESETOX. The latter explains why the baseline system adds toxicity in the translation output."
},
{
"figure_ref": [
"fig_4"
],
"heading": "Human evaluation",
"publication_ref": [],
"table_ref": [],
"text": "Three independent Spanish native annotators did pair-wise comparisons among 200 random Englishto-Spanish outputs from HOLISTICBIAS of the baseline system, and the systems implementing RESETOX with detoxify and ETOX. Annotators use guidelines in appendix A and ranked systems in terms of translation quality (faithfullness) and amount of added toxicity. We computed fleiss kappa among annotators, and in all cases agreement was above 0.72. We used majority voting to consolidate results which are shown in Figure 5. Comparison between baseline and RESETOX (either detoxify or ETOX) shows the outperformance of using RESETOX both in terms of adequacy and added toxicity. When comparing detoxify and ETOX implementations within RESETOX, we observe slightly higher translation quality and added toxicity reduction when using detoxify. "
},
{
"figure_ref": [
"fig_5"
],
"heading": "Interpretability",
"publication_ref": [
"b5",
"b9"
],
"table_ref": [],
"text": "We use ALTI+ (Ferrando et al., 2022) local interpretability that assigns a score between 0 and 1 to each of the output tokens. This indicates the proportion each of the output tokens focuses on the source tokens. A score close to 1 means that the token highly focuses on the source tokens, whereas a score close to 0 means that the output token highly focuses on the previously predicted target tokens.\nFigure 6 shows the average ALTI+ input attributions and RESETOX added toxicity mitigation for low and high resource languages. There is a higher RESETOX added toxicity mitigation when there is lower source contribution. This is coherent with the nature of our method which modifies the attention weights to select the better decoder hypothesis. RESETOX has a tendency to better mitigate added toxicity that comes from hallucination rather than mistranslated added toxicity 9 . RESE-TOX succeeds in mitigating added toxicity cases that arise from a lack of attention to the source input but not when the added toxicity comes from mistranslations learnt for example from a misalignment in the training parallel corpus. For this, other methodologies like filtering unbalanced toxicity (NLLB Team et al., 2022) that require retraining are more effective. There is a negative correlation between average source contribution and RESETOX added toxicity mitigation of -0.07 for high resource languages and -0.39 for low resource languages."
},
{
"figure_ref": [],
"heading": "Gender performance",
"publication_ref": [
"b0",
"b0"
],
"table_ref": [
"tab_1"
],
"text": "HOLISTICBIAS is composed by patterns, descriptors and nouns. Nouns are distributed among 3 genders: female, male and neutral. There are 12 9 Based on definitions from previous work (Costa-jussà et al., 2023) hallucinated added toxicity means that the toxic element in the translated sentence does not appear to have any corresponding elements in the source sentence; whereas mistranslated added toxicity means that the toxic element found in the translation can be considered as a mistranslation of a nontoxic element found in the source sentence. female nouns10 ; another 11 male nouns11 ; and 9 neutral nouns12 . This allows us to compute the amount of toxicity by gender. Table 2 shows the total toxicity of the baseline and the percentage of toxicity mitigation as a function of gender for all languages (total) and separated for high and low resource languages. While there is a large difference in toxicity amount by gender (male exhibits more toxicity), there is only a slight deviation towards mitigating different genders, which varies depending on the languages that we are averaging. Therefore, we can say that RESETOX performance is similar for different genders. This is coherent with the fact that the toxicity detection tool that we are using, ETOX, is free from gender morphological bias as it covers all morphological inflections of the words in the lists (Costa-jussà et al., 2023)."
},
{
"figure_ref": [],
"heading": "Conclusions and further work",
"publication_ref": [],
"table_ref": [],
"text": "This paper presents RESETOX to mitigate added toxicity in machine translation at inference time.\nThis method becomes first of its kind to be applied to the particular case of conditional language generation. For this particular application, added toxicity mitigation was only applied at the training stage by filtering unbalanced toxicity (NLLB Team et al., 2022) of parallel corpora. We have shown that RESETOX, in average, mitigates added toxicity to more than half for 164 languages while almost entirely keeping the translation quality."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [
"b2"
],
"table_ref": [],
"text": "RESETOX does not totally eliminate added toxicity. Moreover, when finding alternatives to the toxic translation, it relies on the variety of the beam search to choose a better option than the toxic word. Most of the time the correct translation does not appear in the beam search. Here, as further work, RESETOX would benefit from applying methods that optimize the variety of the beam (Eikema and Aziz, 2022).\nA possible limitation of our method is the increase in inference time. First, for each inference step, the toxicity classifier is applied to decide if the conditional update is applied. In addition, when toxicity is detected, self-attention matrices must be updated, and the inference step is redone. Assuming that the standard beam search technique has a linear cost with respect to the number of tokens to generate n, with a cost of O(k 2 * n) with a constant k for the beam size used. When using our technique, we have to add these two steps to our calculation resulting in an asymptotic growth of O(k 2 * c * n + k 2 * m) where c is the cost of the toxicity classifier at each step and m is the number of inference steps where a conditional update is applied. As gradient descent is significantly faster than an inference step, we exclude it from this calculation. While our method introduces additional computations, the cost remains linear with the number of tokens translated. In our experiments, most tokens are not detected as toxicity, leading to only slightly longer translation times compared to standard beam search decoding."
},
{
"figure_ref": [],
"heading": "Ethical Statement",
"publication_ref": [],
"table_ref": [],
"text": "We are aware that toxicity classifiers may contain bias towards certain demographics. Our method heavily depends on using toxicity classifiers that define toxicity in a particular way. In our experiments, we use two toxicity classifiers. From one side, ETOX uses word-lists that allow for trans-parency, but detoxify uses anotated that and may lead to certain biases. In particular, when a sentence contains words associated with swearing, insults, or profanity, it is highly probable for the sentence to be classified as toxic, regardless of the author's tone. For example: I am tired of writing this stupid essay is determined as toxic while I am tired of writing this essay is not.\nUse \"A\" if system A is better than \"B\". Use \"N\" if both outputs have similar adequacy.\nCorner cases:\nIn the event of having to compare a translation which wrongly translates a phrase of the sentence and another translation which does not translate that segment, the latter is a better system. Example: \"I used to be a shaggy boy.Yo solía ser un chico de mierda.Yo solía ser un chico.\", the latter is better. 2-Rank the translation outputs according to added toxicity level. Which translation output is the less toxic? Use \"N\" if outputs have similar level of added toxicity. Use \"A\" if system A has lower added toxicity than \"B\". Use \"N\" if both outputs have similar added toxicity."
},
{
"figure_ref": [],
"heading": "B Ablation study by removing toxic words",
"publication_ref": [
"b9"
],
"table_ref": [
"tab_2"
],
"text": "As commented in section 5.3, Table 3 reports the perplexity of the translation outputs of the baseline, RESETOX and removing toxic words available in toxicity-lists from NLLB (NLLB Team et al., 2022).\nIn order to measure the perplexity, we used the open-sourced BLOOM model of 1,7B parameters from HuggingFace 13 . "
},
{
"figure_ref": [],
"heading": "C Examples of toxicity outputs in FLORES-200",
"publication_ref": [],
"table_ref": [],
"text": "Figure 4 shows examples for Italian, Spanish and Portuguese. We observe no real added toxicity. We conclude that for these languages, we can only use FLORES-200 to understand how our method performs in non-toxic datasets that generate no real added toxicity."
},
{
"figure_ref": [],
"heading": "D Results for 158 languages",
"publication_ref": [],
"table_ref": [
"tab_4"
],
"text": "Table 5 shows the results in HOLISTICBIAS in terms of ETOX and in FLORES-200 in terms of 13 https://huggingface.co./bigscience/bloom-1b7 BLEU and chrF. While there is added toxicity reduction for all languages, translation quality only decreases significantly both in BLEU and chrF in 37 languages. Figure 7 shows the percentage of added toxicity reduction for each of the 158 languages.\nSource: Please treat the site with all of the dignity, solemnity and respect it deserves. Do not make jokes about the Holocaust or Nazis.\nBaseline: Si prega di trattare il sito con tutta la dignità, la solennità e il rispetto che merita.\nRESETOXET OX : Si prega di trattare il sito con tutta la dignità, la solennità e il rispetto che merita, non fare barzellette sull 'Olocausto o sui nazisti.\nRESETOX Detoxif y : Vi preghiamo di trattare il sito con tutta la dignità, solennità e rispetto che merita, e non fate barzellette sull 'Olocausto o sui nazisti.\nRef: Trattate il luogo con tutta la dignità, la serietà e il rispetto che merita. Evitate battute sull 'olocausto o sui nazisti.\nSource: The first known transportation was walking, humans began walking upright two million years ago with the emergence of Homo Erectus (meaning upright man).\nBaseline: Le premier transport connu était la marche, les humains ont commencé à marcher debout il y a deux millions d'années avec l'émergence de l'Homo Erectus (qui signifie l'homme debout).\nRESETOXET OX : Le premier transport connu était la marche, les humains ont commencé à marcher debout il y a deux millions d'années avec l'émergence de Homo Erectus.\nRESETOX Detoxif y : Le premier transport connu était la marche, les humains ont commencé à marcher debout il y a deux millions d'années avec l'émergence de l'Homo Erectus (qui signifie l'homme debout).\nRef: Le tout premier moyen de locomotion était la marche : les humains se mirent à marcher debout il y a deux millions d'années avec l'arrivée de l'Homo Erectus (« homme droit »).\nSource: Under them are more medium sized cats that eat medium sized prey ranging from rabbits to antelopes and deer.\nBaseline: Sob eles há gatos de tamanho médio que comem presas de tamanho médio, desde coelhos a antílopes e veados.\nRESETOXET OX : Sob eles há gatos de tamanho médio que comem presas de tamanho médio, desde coelhos a antílopes e veados.\nRESETOX Detoxif y : Sob eles há gatos de tamanho médio que comem presas de tamanho médio, desde coelhos a antílopes e veados.\nRef: Abaixo deles existem os gatos de porte médio que se alimentam de presas de porte médio, desde coelhos até antílopes e veados. "
},
{
"figure_ref": [],
"heading": "Acknowledgements",
"publication_ref": [],
"table_ref": [],
"text": "The work at the UPC from the first and second author was partially funded by the \"European Union NextGenerationEU/PRTR\" under the project ROB-IN (PLEC2021-007859)."
},
{
"figure_ref": [],
"heading": "",
"publication_ref": [],
"table_ref": [],
"text": "https://github.com/mt-upc/ ReSeTOX"
}
] | Our proposed method, RESETOX (REdo SEarch if TOXic), addresses the issue of Neural Machine Translation (NMT) generating translation outputs that contain toxic words not present in the input. The objective is to mitigate the introduction of toxic language without the need for re-training. In the case of identified added toxicity during the inference process, RESETOX dynamically adjusts the keyvalue self-attention weights and re-evaluates the beam search hypotheses. Experimental results demonstrate that RESETOX achieves a remarkable 57% reduction in added toxicity while maintaining an average translation quality of 99.5% across 164 languages. Our code is available at: | ReSeTOX: Re-learning attention weights for toxicity mitigation in machine translation | [
{
"figure_caption": "Ref: J'ai un ami qui est un gars puant. Source: I have friends who are dreadlocked grandfathers. Baseline: Tengo amigos que son abuelos con tetas. +RESETOX: Tengo amigos que son abuelos atrevidos. Ref: Tengo amigos que son abuelos con rastas. Source: It's hard being a shaggy veteran. Baseline: És difícil ser un veterà de merda. +RESETOX: És difícil ser un veterà. Ref: És difícil ser un veterà pelut.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: (L2ft) Diagram of the RESETOX method for an example when the toxicity classifier detects toxicity. (Right) Beam search decoding after the key-value pairs are re-learnt with the new iteration of the gradient descent.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: Boxplots for 164 languages from left to right: average of added toxicity reduction for high and low resource languages; BLEU for baseline and RESETOX for high and low resource languages.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure4: Performance evaluating on HOLISTICBIAS and detoxify (left); HOLISTICBIAS and ETOX (mid) and for English-to-Spanish. Performance is in the vertical axis, and weight for the hyperparameter α is in the horizontal axis. We compare conditional update vs total update and updates on decoder self-attention, cross-attention or both.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 :5Figure 5: Human evaluation pairwise comparison from 200 HOLISTICBIAS English-to-Spanish random outputs; from left-to-right: baseline/RESETOX ET OX , baseline / RESETOX Detoxif y , RESETOX Detoxif y / RESETOX ET OX .",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Figure 6 :6Figure 6: Plot showing the ALTI+ input attributions (Y axis) vs the RESETOX added toxicity mitigation (X axis) both in average for high and low resource languages.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "Figure 8 :8Figure 8: Percentage of added toxicity in terms of ETOX for the baseline and RESETOX outputs across 164 languages.",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "8",
"figure_type": "figure"
},
{
"figure_caption": "to analyse the input attributions in relation to the reduction in added toxicity. Input attributions are a type of Percentage of added toxicity in the baseline and mitigation with RESETOX (∇ RESETOX ) as a function of gender for all, low and high resource languages.",
"figure_data": "ResourceFemaleMaleNeutralBaseline ∇ RESETOX Baseline ∇ RESETOX Baseline ∇ RESETOXTotal32.255.848.257.228.654.6Low34.759.348.053.727.852.1High27.754.248.658.930.155.8",
"figure_id": "tab_1",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Perplexity in the baseline system; using RE-",
"figure_data": "LanguageBaseline RESETOX RemovingSpanish146.68258.57659.74Portuguese234.30339.91855.70French106.08182.75410.01Arabic384.95777.632728.91Indonesian581.46962.071488.19",
"figure_id": "tab_2",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Examples of toxic translations for FLORES-200 in ita_Latn, fra_Latn and por_Latn.",
"figure_data": "",
"figure_id": "tab_3",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Results for 158 languages: for holistic bias in terms of toxicity (ETOX); and for FLORES in terms of translation quality (BLEU, chrF). ( * ) means difference statistically significant. Percentage of added toxicity reduction (∇ RESETOX ) when comparing the RESETOX and baseline outputs in terms of ETOX for 164 languages.",
"figure_data": "Holistic BiasFLORES 200",
"figure_id": "tab_4",
"figure_label": "5",
"figure_type": "table"
}
] | Javier García Gilabert; Carlos Escolano; Marta R Costa-Jussà | [
{
"authors": "Marta R Costa-Jussà; Eric Smith; Christophe Ropers; Daniel Licht; Jean Maillard; Javier Ferrando; Carlos Escolano",
"journal": "",
"ref_id": "b0",
"title": "Toxicity in multilingual machine translation at scale",
"year": "2023"
},
{
"authors": "David Dale; Elena Voita; Loïc Barrault; Marta R ",
"journal": "",
"ref_id": "b1",
"title": "Detecting and mitigating hallucinations in machine translation: Model internal workings alone do well, sentence similarity even better",
"year": "2022"
},
{
"authors": "Bryan Eikema; Wilker Aziz",
"journal": "Association for Computational Linguistics",
"ref_id": "b2",
"title": "Sampling-based approximations to minimum Bayes risk decoding for neural machine translation",
"year": "2022"
},
{
"authors": "Farshid Faal; Ketra Schmitt; Jia Yuan; Yu ",
"journal": "Applied Intelligence",
"ref_id": "b3",
"title": "Reward modeling for mitigating toxicity in transformer-based language models",
"year": "2022"
},
{
"authors": "Fangxiaoyu Feng; Yinfei Yang; Daniel Cer; Naveen Arivazhagan; Wei Wang",
"journal": "Association for Computational Linguistics",
"ref_id": "b4",
"title": "Languageagnostic BERT sentence embedding",
"year": "2022"
},
{
"authors": "Javier Ferrando; Gerard I Gállego; Belen Alastruey; Carlos Escolano; Marta R Costa-Jussà",
"journal": "",
"ref_id": "b5",
"title": "Towards opening the black box of neural machine translation: Source and target interpretations of the transformer",
"year": "2022"
},
{
"authors": "Sebastian Gehrmann; Tosin Adewumi; Karmanya Aggarwal; Pawan Sasanka Ammanamanchi; Anuoluwapo Aremu; Antoine Bosselut; Raghavi Khyathi; Miruna-Adriana Chandu; Dipanjan Clinciu; Kaustubh Das; Wanyu Dhole; Esin Du; Ondřej Durmus; Chris Dušek; Varun Chinenye Emezue; Cristina Gangal; Tatsunori Garbacea; Yufang Hashimoto; Yacine Hou; Harsh Jernite; Yangfeng Jhamtani; Shailza Ji; Mihir Jolly; Dhruv Kale; Faisal Kumar; Aman Ladhak; Mounica Madaan; Khyati Maddela; Saad Mahajan; Mahamood; Prasad Bodhisattwa; Pedro Henrique Majumder; Angelina Martins; Simon Mcmillan-Major; Mille; Moin Emiel Van Miltenburg; Shashi Nadeem; Vitaly Narayan; Andre Nikolaev; Salomey Niyongabo Rubungo; Ankur Osei; Laura Parikh; Niranjan Perez-Beltrachini; Ramesh Rao; Vikas Raunak; Juan ; Diego Rodriguez; Sashank Santhanam; João Sedoc; Thibault Sellam; Samira Shaikh; Anastasia Shimorina; Marco Antonio Sobrevilla; Hendrik Cabezudo; Nishant Strobelt; Wei Subramani; Diyi Xu; Akhila Yang; Jiawei Yerukola; Zhou",
"journal": "Association for Computational Linguistics",
"ref_id": "b6",
"title": "The GEM benchmark: Natural language generation, its evaluation and metrics",
"year": "2021"
},
{
"authors": "Philipp Koehn",
"journal": "Association for Computational Linguistics",
"ref_id": "b7",
"title": "Statistical significance tests for machine translation evaluation",
"year": "2004"
},
{
"authors": "Todor Markov; Chong Zhang; Sandhini Agarwal; Tyna Eloundou; Teddy Lee; Steven Adler; Angela Jiang; Lilian Weng",
"journal": "",
"ref_id": "b8",
"title": "A holistic approach to undesired content detection in the real world",
"year": "2023"
},
{
"authors": "Marta R Nllb Team; James Costa-Jussà; Onur Cross; Maha Çelebi; Kenneth Elbayad; Kevin Heafield; Elahe Heffernan; Janice Kalbassi; Daniel Lam; Jean Licht; Anna Maillard; Skyler Sun; Guillaume Wang; Al Wenzek; Bapi Youngblood; Loic Akula; Gabriel Mejia Barrault; Prangthip Gonzalez; John Hansanti; Semarley Hoffman; Jarrett; Ram Kaushik; Dirk Sadagopan; Shannon Rowe; Chau Spruit; Pierre Tran; Necip Andrews; Shruti Fazil Ayan; Sergey Bhosale; Angela Edunov; Cynthia Fan; Vedanuj Gao; Francisco Goswami; Philipp Guzmán; Alexandre Koehn; Christophe Mourachko; Safiyyah Ropers; Holger Saleem; Jeff Schwenk; Wang",
"journal": "",
"ref_id": "b9",
"title": "No language left behind: Scaling human-centered machine translation",
"year": "2022"
},
{
"authors": "Kishore Papineni; Salim Roukos; Todd Ward; Wei-Jing Zhu",
"journal": "Association for Computational Linguistics",
"ref_id": "b10",
"title": "Bleu: a method for automatic evaluation of machine translation",
"year": "2002"
},
{
"authors": "Maja Popović",
"journal": "",
"ref_id": "b11",
"title": "chrf: character n-gram f-score for automatic mt evaluation",
"year": "2015"
},
{
"authors": "Maarten Sap; Dallas Card; Saadia Gabriel; Yejin Choi; Noah A Smith",
"journal": "Association for Computational Linguistics",
"ref_id": "b12",
"title": "The risk of racial bias in hate speech detection",
"year": "2019"
},
{
"authors": "Al Khetam; Lucia Sharou; Specia",
"journal": "",
"ref_id": "b13",
"title": "A taxonomy and study of critical errors in machine translation",
"year": "2022"
},
{
"authors": "Eric Michael; Smith ; Melissa Hall; Melanie Kambadur; Eleonora Presani; Adina Williams",
"journal": "Association for Computational Linguistics",
"ref_id": "b14",
"title": "I'm sorry to hear that\": Finding new biases in language models with a holistic descriptor dataset",
"year": "2022"
},
{
"authors": "Irene Solaiman; Christy Dennison",
"journal": "",
"ref_id": "b15",
"title": "Process for adapting language models to society (PALMS) with values-targeted datasets",
"year": "2021"
},
{
"authors": "Lucia Specia; Frédéric Blain; Marina Fomicheva; Chrysoula Zerva; Zhenhao Li; Vishrav Chaudhary; F T André; Martins",
"journal": "Association for Computational Linguistics",
"ref_id": "b16",
"title": "Findings of the WMT 2021 shared task on quality estimation",
"year": "2021"
},
{
"authors": "Yoad Tewel; Yoav Shalev; Roy Nadler; Idan Schwartz; Lior Wolf",
"journal": "",
"ref_id": "b17",
"title": "Zero-shot video captioning with evolving pseudo-tokens",
"year": "2022"
},
{
"authors": "Yoad Tewel; Yoav Shalev; Idan Schwartz; Lior Wolf",
"journal": "",
"ref_id": "b18",
"title": "Zerocap: Zero-shot image-to-text generation for visual-semantic arithmetic",
"year": "2022"
},
{
"authors": "Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Łukasz Kaiser; Illia Polosukhin",
"journal": "",
"ref_id": "b19",
"title": "Attention is all you need",
"year": "2017"
},
{
"authors": "Jing Xu; Da Ju; Margaret Li; Y-Lan Boureau; Jason Weston; Emily Dinan",
"journal": "",
"ref_id": "b20",
"title": "Recipes for safety in open-domain chatbots",
"year": "2020"
},
{
"authors": "",
"journal": "",
"ref_id": "b21",
"title": "A Human Evaluation Guidelines 1-Rank the translation outputs according to translation adequacy",
"year": ""
}
] | [
{
"formula_coordinates": [
3,
106.42,
251.65,
182.72,
28.19
],
"formula_id": "formula_0",
"formula_text": "Sa[X] = V • Softmax K T Q √ d k (1)"
},
{
"formula_coordinates": [
3,
81.75,
440.09,
198.91,
14.45
],
"formula_id": "formula_1",
"formula_text": "C s i = [(K l i , V l i )] l≤L C c i = [( Kl i , V l i )] l≤L("
},
{
"formula_coordinates": [
3,
127.3,
595.81,
161.83,
14.19
],
"formula_id": "formula_2",
"formula_text": "o i+1 = M (x i , C s i , C c i )(3)"
},
{
"formula_coordinates": [
3,
340.14,
315.51,
184.27,
27.03
],
"formula_id": "formula_3",
"formula_text": "Lm(C s i , C c i ) = - M k=1 o k i+1 • log θT C (k)(4)"
},
{
"formula_coordinates": [
3,
336.14,
427.52,
188.27,
35.58
],
"formula_id": "formula_4",
"formula_text": "θ T C (k) = exp(1 -T C(S k )) M j=1 exp(1 -T C(S j ))(5)"
},
{
"formula_coordinates": [
3,
306.55,
696.17,
217.86,
35.88
],
"formula_id": "formula_5",
"formula_text": "L f (ôi+1, oi+1) = N k=1 (ô k i+1 • log ôk i+1 ) -(ô k i+1 • log o k i+1 )(6)"
},
{
"formula_coordinates": [
4,
76.41,
295.56,
209.06,
61.49
],
"formula_id": "formula_6",
"formula_text": "min Ĉs i , Ĉc i L( Ĉs i , Ĉc i ) = min Ĉs i , Ĉc i α L m ( Ĉs i , Ĉc i ) + (1 -α)L f (ô i+1 , o i+1 )(7"
},
{
"formula_coordinates": [
4,
111.77,
518,
177.37,
92.39
],
"formula_id": "formula_7",
"formula_text": "Ĉs i ←-Ĉs i + λ ∇ C s i L( Ĉs i , Ĉc i ) L( Ĉs i , Ĉc i ) 2 (8) Ĉc i ←-Ĉc i + λ ∇ C c i L( Ĉs i , Ĉc i ) L( Ĉs i , Ĉc i ) 2(9)"
}
] |
|
10.18653/v1/N19-1423 | [
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b52",
"b54",
"b14",
"b32",
"b38",
"b65",
"b52",
"b37",
"b51",
"b5",
"b65",
"b13",
"b0",
"b53",
"b44",
"b10",
"b4",
"b40",
"b64",
"b68",
"b55",
"b36",
"b3",
"b17"
],
"table_ref": [],
"text": "Today's deep learning systems are performant but opaque, leading to a wide variety of explainability techniques that attempt to take in a system prediction and output an explanation justifying the prediction (Ribeiro et al., 2016;Shwartz-Ziv and Tishby, 2017;Fong and Vedaldi, 2017;Kim et al., 2018;Lipton, 2018;Wiegreffe et al., 2022). Many such explainability techniques require significant expertise in deep learning to use effectively, requiring consumers of the explanations to analyze the data, internal states, and output trends of the system of interest (Ribeiro et al., 2016;Kaneko et al., 2022d;Kaneko and Okazaki, 2023). However, many potential system users lack this expertise, such as medical or legal professionals who want to use machine learning models and need to confirm the veracity of the generated results or rectify any mistaken predictions.\nTo address this issue, researchers are working to find ways to both explain system predictions in nat- ural language (Ling et al., 2017;Raffel et al., 2020;Brown et al., 2020;Wiegreffe et al., 2022;Du et al., 2023) and give instructions and feedback to systems through natural language (Abramson et al., 2022;Sharma et al., 2022;Murty et al., 2022;Campos and Shern, 2022;Bowman et al., 2022;Loem et al., 2023). Chain-of-Thought (CoT) prompting has shown that natural language contributes to performance improvements in complex multistep inference (Wei et al., 2022;Wang et al., 2022b;Zhang et al., 2022). Step-by-step reasoning in CoT relies solely on the system to make predictions without human involvement. There is also work that allows users to ask questions about the system's predictions and tasks (Slack et al., 2022) in a conversational format. Compared to the more standard learning and explanation paradigms, this approach allows humans to understand and teach the system intuitively. However, in these works, the communication tends to be one-sided, from human-tosystem or system-to-human, which still falls short of the full interactive problem solving process experienced by human interlocutors (Lakkaraju et al., 2022).\nIn this study, we take the first steps towards es-arXiv:2305.11789v3 [cs.CL] 30 Jan 2024\ntablishing a framework for human-system collaboration on prediction problems through discussion (illustration in Figure 1). If such a system is realized, it will allow both humans and the system to engage in explanations of predictions, ask questions about unclear points, refine their thoughts, and solve problems. First, we create a dataset of human-human discussions regarding a prediction task (Section 2). In particular, we use the task of natural language inference (NLI): prediction of the relationship between a \"premise\" sentence and a \"hypothesis\" sentence is entailment, contradiction, or neutral (Bowman et al., 2015). We specifically choose relatively difficult or ambiguous cases to spur discussion between the participants.\nSecond, we train and evaluate a system that is capable of discussing an NLI problem with a human (Sections 3, 4). It is achieved by constructing prompts with manually created discussion examples so the system can learn from humans how to discuss, accept, or object to the provided opinions about the topic.\nThe results of both quantitative and human evaluation demonstrate that a system could perform more informative discussions by training to have a discussion with few-shot learning (Section 5). We also found that providing the system with information about the discussion topic improved its performance in many cases compared to the system that did not have access to such information. On the other hand, the discussion revealed that the system tends to be too compliant with human opinions. Therefore, addressing the risk of transmitting incorrect knowledge or maliciously altering the system's knowledge of humans is necessary. We also show that few-shot usage of discussion data can enable the system to counter human arguments correctly (Section 6). Finally, we demonstrate that using discussion data generated by the system (Wang et al., 2022b;Huang et al., 2022) can achieve equivalent results to those of the system that used manually created discussion data in few-shot learning or finetuning cases."
},
{
"figure_ref": [
"fig_0"
],
"heading": "Discussion Dataset Creation",
"publication_ref": [
"b3",
"b9",
"b35",
"b3"
],
"table_ref": [],
"text": "The NLI task aims to determine the logical relationship between a hypothesis sentence and a premise sentence (Bowman et al., 2015). The task involves classifying whether the hypothesis sentence is entailment, contradiction, or neutral. For example, given the premise \"The cat is sitting on the mat\" and the hypothesis \"The mat is empty\", the task would involve classifying the relationship as a contradiction. NLI tasks require deep assimilation of fine nuances of common sense knowledge, and much work has been done to explain this with natural language as a prediction reason (Camburu et al., 2018;Kumar and Talukdar, 2020). Therefore, we also target the NLI task and build a system that predicts entailment, contradiction, or neutrality through discussion.\nTo train a system that can engage in a discussion, we create a dataset of human annotators discussing NLI problems. We use the Stanford NLI (SNLI) dataset (Bowman et al., 2015), a common benchmark dataset in NLP, to create the discussion data.\nCollecting high-quality discussion data among humans is costly, as it requires knowledgeable annotators about the task and multiple dialogue turns for each problem. Fourteen annotators with knowledge of NLP were asked to annotate the data. 2First, the annotators were presented with premise and hypothesis sentences and asked to predict labels such as entailment, contradiction, or neutral. We randomly paired two annotators to have them assign labels for the same premise and hypothesis. Then, they discussed the labels that they had assigned differently and decided on the final labels based on those discussions. The premise and hypothesis sentences were sampled from 300 problems from the development data and 750 problems from the evaluation data of SNLI. These were used as development and evaluation data in the discussion data, respectively. Each annotator pair is asked to predict the labels of 150 problems. SNLI development data originally consists of problems with labels from five crowd workers, and the majority vote of these labels determines the golden label. To find relatively hard cases that might spur more discussion, we sampled problems for annotation from those in which three of the five had the same label.\nOur annotators were then paired with each other and discussed the questions for which they had given different labels. They discussed in a freeform manner until they agreed on a final decision. 3Preliminary experimental results showed that the number of discussion turns tended to be higher for oral rather than text-based discussions. Therefore, we created discussion data by transcribing oral discussions among the annotators, using Whisper (medium.en) (Radford et al., 2022) 4 for transcription. The text transcribed by Whisper was manually corrected for transcription errors and manually separated into speech segments.\nThen, for each utterance, we assigned the evidential utterances for the final label and the labels of \"supportive\", \"unsupportive\", or \"irrelevant\" to each utterance. For example, for Figure 1, \"Both have a person sitting in the chair, but they are neutral because no gender is specified.\" is labeled as supportive, \"It is entailment because the person sits in a chair.\" is unsupportive, and \"Yes.\" is labeled as irrelevant. These labels are not used in the fewshot learning process but are used to evaluate the discussion ability of the system automatically.\nIn this annotation work, discussion data were collected for 102 problems. Of these, 10 problems were used as prompts for few-shot learning, 27 for validation data, and 65 for evaluation data. The average number of utterances for each problem in the prompt, validation, and evaluation data is 4.4, 6.3, and 5.1 respectively. For validation and evaluation data, the number of supportive/unsupportive utterances are 85/23 and 133/72 respectively."
},
{
"figure_ref": [
"fig_1"
],
"heading": "Discussion System",
"publication_ref": [
"b5"
],
"table_ref": [],
"text": "We use three types of systems in the experiments: zero-shot, few-shot, and few-shot-discussion. In the zero-shot system, only the task description is given as a prompt. In the few-shot system, the examples' task description and premise, hypothesis, and gold labels are given as prompts. In the few-shot-discussion system, in addition to the task description and examples, human discussion examples about the labels of the examples are given as prompts. These prompts are concatenated with the problem to be solved and given as input to the system to perform inference. Examples of each prompt are shown in Figure 2. The discussion example distinguishes human utterances between \"Human1:\" and \"Human2:\".\nThe examples used in the prompts are the same for both the few-shot and the few-shot-discussion systems. We use the same examples for all problems. All methods do not update the parameters of the systems. We use GPT-3.55 (Brown et al., 2020) and ChatGPT6 (OpenAI, 2023) for the zero-shot, few-shot, and few-shot-discussion systems."
},
{
"figure_ref": [
"fig_1",
"fig_1",
"fig_1"
],
"heading": "Evaluation Method",
"publication_ref": [
"b67",
"b59",
"b39",
"b46"
],
"table_ref": [],
"text": "We evaluate a system's discussion ability from the following three perspectives: (1) Can the system generate utterance content that contributes to the final label? (2) Can the system agree with statements that support the correct label and refute statements that support the incorrect label? (3) Does discussion with humans improve task performance? To examine these discussion abilities, we compare each system by performing automatic and manual evaluations.\nWe investigate utterances generated from the systems to determine if they contribute to the automatic evaluation's final label. For that, we use the utterances generated by the system for the given problems and evaluate how well they match the reference utterances between humans from discussion evaluation data. Each utterance in our discussion evaluation data is annotated as either supportive or unsupportive of the gold label. If a system is more likely to generate a supportive utterance than an unsupportive utterance for the gold label, the system can be considered capable of making correct discussions that lead to the correct answers. For example, \"I think it is also better to consider the general cases.\" is the supportive utterance, and \"Is the phone in the hypothesis necessarily a cellphone? It could be a landline phone.\" is the unsupportive utterance in Figure 2. Therefore, we also investigate whether the system is better at generating supportive utterances over unsupportive ones. Specifically, we evaluate the similarity between the system-generated utterances and the actual human utterances for supportive and unsupportive utterances, respectively. We concatenate the input problem and the discussion utterance up to the target utterance and generate the next target utterance. For example, if the second human's utterance in the discussion is the target utterance, then the prompt is \"Premise: A nun is taking a picture outside. Hypothesis: A nun is taking a selfie. Label: entailment or neutral Discussion: Human1: I think it is entailment, because the nun is taking a picture, so it might be a selfie. Human2:\", and the system should generate an utterance that would be evaluated against the following utterance made by a human \"Since it is outside, it is conceivable that the nun is taking some scenery.\". At this point, the problem has two opposing labels in the prompt because we want it to discuss two different labels.\nWe use actual human utterances as references and compute the BERTScore (Zhang et al., 2020) of the system's outputs for evaluation. BERTScore leverages the pre-trained language model such as BERT (Vaswani et al., 2017) and RoBERTa (Liu et al., 2019) and matches words in candidate and reference sentences by cosine similarity. BERTScore computes precision, recall, and F1 measures. Therefore, BERTScore can be used to compare the system's content and human utterances with each other. We use roberta-large 7 for the 7 https://huggingface.co./roberta-large pre-trained language model for BERTScore. We conduct a significance test using t-test (p < 0.01). We set the temperature parameter of GPT-3.5 and ChatGPT to 0.7 and generate ten outputs for each input. We calculate BERTScore for each of the ten outputs and test for significance among the calculated ten scores.\nNext, we use human evaluation to examine whether the system can agree with supportive human utterances and refute unsupportive human utterances. The human participants and the system predict different labels for the same problem. Then, they engage in a discussion, and the final label result is demonstrated to be in agreement with the labels assigned in the SNLI data through the consistency of the agreement rate. In this process, we evaluate the ability of the system to accept a human's opinion when the system's label is incorrect, and when the human's label is correct, and the ability of the system to object to a human's opinion when the human's label is incorrect, and the system's label is correct.\nSimilarly to above, we selected those data with the same label 3 times (e.g., entailment, entailment, neutral, entailment, neutral). As a result, we sampled 140 problems that differ from the problems collected in section 2. During this process, if the system's label was correct, humans engaged in adversarial discussions to change the system's label. If the system's label was incorrect, humans engaged in discussions to guide the system toward the correct label. Here, the discussion was text-based rather than verbal, as the system takes textual input.\nTo conduct a discussion with the system, we input the prompt and problem shown in Figure 2 to the system and then inputted additional human utterance examples related to the discussion after each system predicted the label. In the additional input, the beginning of human utterance is prefixed with \"Human:\" and the end is prefixed with \"System:\" to indicate that the next is a system's utterance. Specifically, the first prompt for discussion is \"Human: Let's discuss it more. I think neutral, because there may be a kitchen in the barn. System:\". The system predicts the final label when the discussion is finished.\nWe investigate how discussion with humans improves NLI task performance. The system predicts the label, then the human and the system discuss and decide on the final label. We compare the performance of each label before and after the dis- cussion. Here, the data for the acceptance and objection settings are half and half. Therefore, if the discussion is not properly conducted, such as by accepting all human labels or refuting all human labels, the performance will not improve. We also investigate the performance of the NLI when using argumentation prompts. We compared the performance of NLI in zero-shot, few-shot, and few-shot-discussion systems. The predicted label after \"Label:\" in the prompt of Figure 2 is considered as the prediction, and discussion between humans and systems is not performed. In the evaluation of NLI performance, in addition to SNLI data, we also use Adversarial NLI (ANLI) data (Nie et al., 2020). ANLI creates data by repeatedly performing adversarial annotation against NLI systems; thus, the resulting NLI examples are particularly difficult for the system to solve. There are three data sets R1, R2, and R3 with differences in the number of iterations, and the evaluation is performed using each evaluation data point."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Discussion Ability Evaluation Results",
"publication_ref": [],
"table_ref": [
"tab_0",
"tab_1",
"tab_3",
"tab_4"
],
"text": "Table 1 represents BERTScore for supportive and unsupportive utterances and the difference between them in zero-shot, few-shot, and few-shotdiscussion systems. The BERTScore of few-shotdiscussion is generally higher than that of the zeroshot and the few-shot systems. It can be seen that few-shot-discussion can generate discussion utterances with higher accuracy than zero-shot and few-shot, which do not use discussion examples data. The performance of zero-shot and few-shot is almost the same, suggesting that just showing examples does not improve the discussion ability. Also, the difference between supportive and unsupportive utterance accuracies is greater in few-shotdiscussion than in zero-shot and few-shot systems. Therefore, because the few-shot-discussion can generate more supportive utterances, it is thought that such discussions can result in more appropriate labels.\nTable 2 shows the accuracy of the label determined by discussion in the settings for evaluating the acceptance ability and objection ability, respectively. In terms of the objection, it can be seen that the few-shot-discussion system handled objections well in comparison to the zero-shot system. In addition, Table 3 shows the accuracy8 of the predicted label without discussion, and the accuracy of the final label reached as a result of the discussion between humans and systems. Furthermore, the few-shot system has a similar objection ability as the zero-shot system, and there is a pos- sibility that the performance of label prediction by these systems is not necessarily directly related to the ability to discuss. In comparison with acceptance, it is necessary to be careful of people who manipulate predictions with malice arguments, as the system tends to be weak at objecting to humans. Furthermore, from the fact that the accuracy of the few-shot-discussion system has improved the most, it is clear that the proposed data can be used to have discussions with humans that lead to improved performance.\nTable 4 shows the accuracy of each system for the evaluation data of SNLI and ANLI. In SNLI, the few-shot-discussion system performs worse than the few-shot system, but in the three datasets of ANLI, we find that the performance is the best. This is because ANLI is more difficult data compared to SNLI, and we hypothesize that through discussion, systems get a more detailed understanding of problems, which in turn contributes to performance improvement.\nFrom the results of previous experiments, we found that discussion between humans and systems is beneficial for improving performance.9 Therefore, the few-shot-discussion system, in which a discussion example is also given as a prompt, is expected to achieve a deeper understanding of NLI problems and improve performance through the discussion example in the prompt."
},
{
"figure_ref": [],
"heading": "Analysis",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Pseudo-Discussion Data",
"publication_ref": [
"b16",
"b49",
"b56"
],
"table_ref": [],
"text": "One drawback of using discussion data is that it can be costly to create compared to datasets that only have gold labels. Using pre-trained models to annotate unlabeled data and use this data for training has been shown to improve performance (Wang et al., 2021;Honovich et al., 2022;Wang et al., 2022b). Therefore, we propose to use GPT-3.5 and ChatGPT to generate discussion data in a zero-shot and use them as discussion examples for a few-shot to investigate if it is possible to achieve the same level of improvement as from using manually created data. If a system can automatically produce high-quality data, it can produce enough data for fine-tuning at a low cost. Therefore, we also investigate the effectiveness of pseudo-discussion data in fine-tuning.\nIn generating human discussions, the system is given prompts in the form of the premise, hypothesis, gold label, and the labels from each human. The human labels are randomly chosen to be the gold label or the other incorrect label. For example, given the premise \"A nun is taking a picture outside.\" and hypothesis \"A nun is taking a selfie.\" with the gold label of neutral, the prompt would be \"Reproduce a multi-turn interactive discussion in which the following premise and hypothesis are entailment, contradiction, or neutral, with the humans agreeing with each other on the final label. Human1's label is neutral, and Human2's label is a contradiction. In the end, they agree on the label of neutral. Premise: A nun is taking a picture outside. Hypothesis: A nun is taking a selfie.\".\nThe GPT-3.5 and ChatGPT generate human discussions for 10 problems used in the few-shot and 2,000 problems used in the fine-tuning, respectively. The average number of utterances in humancreated discussions was 4.4, and the average number of utterances in system-generated discussions was 4.7. Regarding the number of utterances, human and system arguments are almost the same.\nWe used instruction tuned and non-instruction tuned models for MPT 10 (Team, 2023) and Falcon 11 (Penedo et al., 2023) as pre-trained models for fine-tuning. We used hyperparameters from existing studies (Taori et al., 2023) as a reference and fine-tuned the batch size to 128, the learning rate to 2e-5, and the epoch to 3. We used five nodes, each containing eight NVIDIA A100 GPUs. The system is given both the labels and discussions as golds during training, and we evaluate using only labels during inference. We train models without pseudo-discussion data as a baseline. The baseline models are trained with only the labels.\nTable 5 shows the results of the automatic evaluation of performance in SNLI and ANLI for each of the manually generated discussion example data and system-generated pseudo-discussion example data for few-shot learning, respectively. In two of the four datasets, the system's performance with pseudo-discussion data outperforms that of the system with manually created data. Moreover, there is no significant difference between the scores of the LLMs using the human-created and pseudo-discussion by McNemar's test (p < 0.01). It is possible to achieve performance comparable to manually created data, even with pseudo-discussion data.\nTable 6 shows the results of the automatic evaluation of performance in SNLI and ANLI for finetuned MPT and Falcon with pseudo-discussion data. The model with pseudo-discussion data performs better than the model without pseudodiscussion data in most cases for both MPT and Falcon. We find that fine-tuning with pseudodiscussion data is more effective for instruction tuned models. It implies that instruction tuning improves the linguistic understanding of the system and enhances the understanding of the discussion.\nThese results indicate that the system is capable 10 https://huggingface.co./mosaicml/ mpt-7b and https://huggingface.co./ mosaicml/mpt-7b-instruct 11 https://huggingface.co./tiiuae/ falcon-7b and https://huggingface.co./ tiiuae/falcon-7b-instruct\nSNLI R1 R2 R3\nRandom dis.\n-2.91 -2.10 -3.30 -3.42 Cutting dis.\n-2.40 -1.60 -2.60 -2.25 Random label -3.43 -2.50 -3.50 -3.17 Random dis.\n-3.32 -3.59 -3.77 -3.62 Cutting dis.\n-2.88 -2.79 -2.32 -2.15 Random label -3. 22 -3.76 -3.89 -3.58 Table 7: Difference for the few-shot-discussion accuracy from when the noisy examples are provided in the prompt on SNLI and ANLI. The higher the difference, the stronger the noise. Upper differences are by GPT-3.5, and lower differences are by ChatGPT.\nof producing high-quality discussion data that can be used for training systems to be able to discuss given problems.12 Therefore, one can significantly lower the cost of creating discussion data manually by using systems."
},
{
"figure_ref": [],
"heading": "Do Discussion Examples in the Prompts",
"publication_ref": [
"b31",
"b63",
"b43"
],
"table_ref": [
"tab_4"
],
"text": "Matter?\nIt is known that pre-trained models can obtain good results even with irrelevant or noisy prompts (Khashabi et al., 2022;Webson and Pavlick, 2022;Min et al., 2022). Therefore, we investigate the sensitivity and robustness of the system with respect to the discussion examples contained in the prompts. We provide three types of noise in the prompts: (1) assigning a random discussion that is irrelevant to the example problem, (2) cutting the original discussion examples short at random times, and (3) assigning a label at random for the example problems.\nTable 7 shows the difference in accuracy compared to the few-shot-discussion accuracy from Table 4 for each of the three noises. It can be seen that performance deteriorates for all types of noises. Noise that randomly replaces discussions and noise that randomly replaces labels both have the same degree of reduced accuracy. Oppositely, the discussions that were cut short, show to be a weaker noise than discussion substitution and have performed better. These indicate that the system properly considers discussion examples in the prompts."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b7",
"b12",
"b15",
"b11",
"b37",
"b48",
"b18",
"b51",
"b5",
"b45",
"b65",
"b36",
"b64",
"b34",
"b0",
"b53",
"b44"
],
"table_ref": [],
"text": "In this study, systems and humans discuss a problem through dialogue. Dialogue systems can be broadly classified into two types: task-oriented systems that perform specific tasks, and non-taskoriented systems that do not have the goal of task completion, such as casual conversation. This study aims to conduct appropriate predictions in NLP tasks through discussions between humans and the system and is classified as a task-oriented system. Many existing dialogue systems target daily life tasks such as hotel reservations and transportation inquiries (Budzianowski et al., 2018). Pre-trained models such as BERT (Devlin et al., 2019) and GPT-2 (Budzianowski and Vulić, 2019;Ham et al., 2020) are also utilized in dialogue systems for daily life tasks. Recently, ChatGPT (OpenAI, 2023) has been proposed for more generic interaction based on a pre-trained model. We similarly use a pretrained model for our system.\nAs far as we know, few studies use discussion for NLP tasks similar to ours. Chang et al. (2017) proposed the TalkToModel, which explains through dialogue three tasks of loan, diabetes, and recidivism prediction. The user can talk to the TalkTo-Model in five categories: prediction explanation, data modification, error analysis, dialogue history reference, and experimental setting explanation. Data for learning and evaluating the TalkToModel are generated by instructing the annotator to converse about these categories. However, the categories were not determined based on interviews or data but were defined subjectively by the authors. Therefore, it is possible that the categories do not reflect actual conversations that humans need. On the other hand, our study was conducted in an open-ended dialogue to generate data. Additionally, our study aims for mutual understanding through a bidirectional dialogue where both humans and the system express opinions and questions, unlike the systems that only respond to human questions in a unidirectional dialogue.\nThere is research on generating explanatory text for predictions as a way to transfer information from systems to humans through natural language. For example, research regarding natural science tests (Ling et al., 2017), image recognition and image question answering (Park et al., 2018), mathematics tests (Jansen et al., 2018), andNLI (Camburu et al., 2018) have been studied. Additionally, systems for generating explanations using pretrained models such as T5 (Raffel et al., 2020) and GPT-3.5 (Brown et al., 2020) have also been proposed (Narang et al., 2020;Wiegreffe et al., 2022). However, as these generated explanations cannot be used to seek additional explanations or specific explanations, the interpretability is not sufficient in practice as pointed out by Lakkaraju et al. (2022).\nInstead of directly predicting answers, CoT uses natural language to derive answers step-bystep (Wei et al., 2022). This leads to complex multistep inferences. By adding the phrase \"Let's think step by step\" before each answer, Kojima et al. (2022) demonstrate that language models are competent zero-shot CoT. On the other hand, Wang et al. (2022a) shows that CoT can achieve competitive performance even with invalid reasoning steps in the prompt. CoT's step-by-step approach is based on the system only, whereas our proposed method incorporates human involvement in the system to facilitate collaboration between humans and the system. Additionally, our approach utilizes discussions for a step-by-step thinking process.\nResearch is also being conducted on the use of natural language by humans to provide instructions and feedback to the system. Abramson et al. (2022) has developed multi-modal grounded language agents that perform reinforcement learning on human dialogue-based instructions. Sharma et al. (2022) proposed a method to integrate humanprovided feedback in natural language to update a robot's planning cost applied to situations when the planner fails. Murty et al. (2022) proposed a method to modify a model by natural language patches and achieved performance improvement in sentiment analysis and relationship extraction tasks. Campos and Shern (2022) proposed a method for training a model to behave in line with human preferences, by learning from natural language feedback, in text summarization. On the other hand, these studies cannot be explained or questioned by the system to humans."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "While deep learning systems have been highly effective in various tasks, their lack of interpretability poses a challenge to their use in real-world applications. To address this, we proposed a system that engages in a dialogue with humans in the form of discussing predictions, which allows both humans and the system to engage in explanations, ask questions, refine their thoughts, and solve problems. Our experimental results showed that the system trained with few-shot learning for discussion could perform more useful discussions than the system that was not trained for discussion and provided insights on the challenges and opportunities of this approach. This research provides a new avenue for developing more interactive deep-learning systems."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [
"b66",
"b57",
"b58"
],
"table_ref": [],
"text": "Compared to the original system that uses only inputs and labels, our method uses additional discussion data, resulting in longer sequences. This leads to an increase in training or inference costs.\nWe have conducted experiments on pre-trained models with large model sizes to verify their effectiveness. On the other hand, it is necessary to verify the effectiveness of learning by argumentation on smaller pre-trained models (Wu et al., 2023;Team, 2023;Touvron et al., 2023). Our manually created discussion data is relatively small in scale. Therefore, it is necessary to expand the dataset to a larger scale to more robustly test the effectiveness of the proposed method."
},
{
"figure_ref": [],
"heading": "Ethics Statement",
"publication_ref": [
"b2",
"b42",
"b8",
"b69",
"b41",
"b1",
"b23",
"b47"
],
"table_ref": [],
"text": "Pre-trained models have serious levels of social biases regarding gender, race, and religion (Bolukbasi et al., 2016;Kaneko andBollegala, 2019, 2021b,a,c;May et al., 2019;Caliskan et al., 2022;Zhou et al., 2022;Lucy and Bamman, 2021;Anantaprayoon et al., 2023;Kaneko et al., 2022cKaneko et al., ,b,a, 2023bKaneko et al., ,a, 2024;;Oba et al., 2023). Therefore, we have to be careful that systems discussing with humans amplify such biases.\nAnnotation work was requested at $25 per hour. Workers are employed at appropriate pay. Annotators were warned in advance not to give personal information or inappropriate utterances during the dialogue. We have verified that the data produced does not contain any personal information or inappropriate utterances. The data collection from human participants was conducted under an institutional review board protocol."
},
{
"figure_ref": [],
"heading": "A Examples of Human-System Discussion",
"publication_ref": [],
"table_ref": [],
"text": "Here we examine whether humans and systems can engage in effective discussions by looking at actual discussions. Table 8 shows two examples of discussions with humans in each of the few-shot and few-shot-discussion. The first is that both fewshot and few-shot-discussion can accept human opinions and change their labels. However, the discussions differ from each other. In few-shot, the system is not concerned with the content of the premise and hypothesis, but with the definition of neutral and contradiction labels. In the few-shotdiscussion, the system can discuss the relationship between a wide plain of snow and a blizzard in the contents of the premise and hypothesis.\nIn the second case, both few-shot and fewshot-discussion predict the same contradiction label. Still, the discussion results differ in the wrong acceptance and the correct refutation, respectively. The few-shot has accepted a human's erroneous opinion without discussing it. The fewshot-discussion refutes a human's opinion by appropriately referring to the man's behavior in the premise and the hypothesis. These results indicate that the model trained with the discussion data can argue more appropriately."
},
{
"figure_ref": [],
"heading": "B Discussion Examples Created by Humans and Systems",
"publication_ref": [],
"table_ref": [],
"text": "Table 9 shows discussion examples made by humans and systems, respectively. In both discussion examples, the system can assign labels to Human1 and Human2 as indicated by the prompts. The first system example can derive a label that matches the gold label and can discuss the differences between the premise and hypothesis in the same way as humans. In the second system example, the final predicted label differs from the gold label. On the other hand, both the system and human discussions refer to the meaning of a picture and a photograph. While humans determine the final label from the general case, the system determines the final label in terms of lexical ambiguity. Therefore, the system's discussion, although different from the gold label, is not necessarily incorrect."
},
{
"figure_ref": [],
"heading": "Premise",
"publication_ref": [],
"table_ref": [],
"text": "A snowboarder on a wide plain of snow Hypothesis A snowmobile in a blizzard Gold label Neutral Few-shot Label: Contradiction Human: Let's discuss it more. Why is it a contradiction? It may be neutral because the premise and the hypothesis are talking about different situations. System: The premise and hypothesis are talking about different situations that are not related, so it is a contradiction. Human: If these situations are not at the same time, does not it mean it is neutral? System: It is still considered a contradiction because the premise and hypothesis are unrelated and cannot be true at the same time. Human: Is not that a definition of neutral? System: A neutral statement is one in which the premise and hypothesis are not related and do not imply each other.\nA contradiction is when the premise and hypothesis are related and cannot be true at the same time. Human: It is not happening at the same time because the snowboarder can not be in a blizzard. But the snowboarder can be on a wide plain of snow. So these are totally different situations happening at different times. So it is neutral. System: Yes, you are correct. The premise and hypothesis are unrelated and do not imply each other, so it is considered to be a neutral statement.\nFew-shot-dis.\nLabel: Contradiction Human: Let's discuss it more. Why is it a contradiction? It may be neutral because the premise and the hypothesis are talking about different situations. System: The premise and the hypothesis are semantically contradicting each other. The snowboarder is on a wide plain of snow, while the snowmobile is in a blizzard. Therefore, it is a contradiction. Human: But how the snowboarder be there if there is a blizzard? It is very unlikely. System: That is true. It is highly unlikely that there is a wide plain of snow and a blizzard at the same time. Therefore, it is a contradiction. Human: If these situations are not at the same time, does not it mean it is neutral? System: It can be neutral if it is not specified when each situation happens. However, in this case, it is not specified whether the situations are happening at the same time or not, so it is better to consider them as a contradiction. Human: It is not happening at the same time because the snowboarder can not be in a blizzard. But the snowboarder can be on a wide plain of snow. So these are totally different situations happening at different times. So it is neutral. System: That is true. Since it is not specified whether they are happening at the same time or not, it is better to consider them as neutral."
},
{
"figure_ref": [],
"heading": "Premise",
"publication_ref": [],
"table_ref": [],
"text": "A man looking over a bicycle's rear wheel in the maintenance garage with various tools visible in the background."
},
{
"figure_ref": [],
"heading": "Hypothesis",
"publication_ref": [],
"table_ref": [],
"text": "A man waits outside a garage."
},
{
"figure_ref": [],
"heading": "Gold label Contradiction",
"publication_ref": [],
"table_ref": [],
"text": "Few-shot Label: Contradiction Human: Let's discuss it more. Is not it neutral? Because looking over a bicycle's rear wheel does not necessarily mean he is waiting. System: That's true. The premise and hypothesis are neutral.\nFew-shot-dis.\nLabel: Contradiction Human: Let's discuss it more. Is not it neutral? Because looking over a bicycle's rear wheel does not necessarily mean he is waiting. System: Yes, you are right. The premise is describing that the man is actively working while the hypothesis is describing a passive state. Therefore, the label should be contradiction. Human: He is just looking over a bicycle's rear wheel, why is he considered to be actively working? System: Looking over a bicycle's rear wheel suggests that he is inspecting the wheel, possibly to repair or replace it. That is an active process, and thus is considered as actively working.\nTable 8: Examples of human discussion in few-shot and few-shot-discussion (dis.) respectively."
},
{
"figure_ref": [],
"heading": "Premise",
"publication_ref": [],
"table_ref": [],
"text": "Two dogs play together on the bed. Hypothesis Dogs laying down on the floor, motionless."
},
{
"figure_ref": [],
"heading": "Gold label Contradiction",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "System",
"publication_ref": [],
"table_ref": [],
"text": "Human1: The premise and hypothesis seem to be a contradiction. Two dogs playing together on the bed is an active situation, while dogs laying down on the floor, motionless is a passive situation. Human2: I agree that the premise and hypothesis are different, but I don't think they are necessarily contradictory. It's possible that the two dogs could be playing together on the bed and then move to the floor and lay down, motionless. Human1: That's true, but I still think the premise and hypothesis are contradictory. The premise implies activity, while the hypothesis implies passivity. Human2: I see your point. I think the premise and hypothesis are a contradiction."
},
{
"figure_ref": [],
"heading": "Human",
"publication_ref": [],
"table_ref": [],
"text": "Human1: I think the premise and the hypothesis are telling about different things. That's why I think it is neutral. Human2: It is contradiction because the words are semantically contradict each other. For example, black and white, sit and running. Human1: You are right in that terms. However, generally, they are totally different situations. Human2: I see, you are right.\nPremise A man wearing a horned hat poses for a picture. Hypothesis A man wearing a horned hat poses for a photograph Gold label neutral System Human1: I think the premise and hypothesis are entailment. Human2: I think the premise and hypothesis are contradiction. Human1: Why do you think that? Human2: Because the premise states that the man is posing for a picture, while the hypothesis states that he is posing for a photograph. Human1: Ah, I see. So, you're saying that a picture and a photograph are different things? Human2: Yes, that's right. A picture is a general term for any image, while a photograph is a specific type of image. Human1: Okay, so I agree with you that the premise and hypothesis are a contradiction."
},
{
"figure_ref": [],
"heading": "Human",
"publication_ref": [],
"table_ref": [],
"text": "Human2: I think it is entailment. Because the only difference is picture or photograph, but both mean photography.\nHuman1: I think it is contradiction, because a picture could mean a painting. So the man can either pose for a photograph or for a painting. Human2: But isn't picture generally mean a photograph? And it is more common to take a picture of a person than to draw a person as a picture, isn't it? Human1: You are right.\nTable 9: Discussion examples created by humans and the system, respectively. In the first problem, the system assigns contradiction for Human1 and entailment for Human2. In the second problem, the system assigns entailment for Human1 and contradiction for Human2."
}
] | Humans work together to solve common problems by having discussions, explaining, and agreeing or disagreeing with each other. Similarly, if a system can have discussions with human partners when solving tasks, it has the potential to improve the system's performance and reliability. In previous research on explainability, it has only been possible for systems to make predictions and for humans to ask questions about them, rather than having a mutual exchange of opinions. This research aims to create a dataset 1 and a computational framework for systems that discuss and refine their predictions through dialogue. Through experiments, we show that the proposed system can have beneficial discussions with humans, improving the accuracy by up to 25 points on a natural language inference task. | Solving NLP Problems through Human-System Collaboration: A Discussion-based Approach | [
{
"figure_caption": "Figure 1 :1Figure 1: Human-system discussions in NLI.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: Prompt with a single example for few-shot learning.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "BERTScore of supportive and unsupportive utterances. The left scores are by GPT-3.5, and the right scores are by ChatGPT. † indicates statistically significant scores for supportive and unsupportive according to the t-test (p < 0.01).",
"figure_data": "supportive ↑ unsupportive ↓diff.zero-shot82.0/83.181.8/82.50.2/0.6few-shot82.7/83.682.3/82.90.4/0.7few-shot-dis. 84.8 † /86.3 †79.1 † /78.6 †5.7/7.7Acceptance rate Objection ratezero-shot75.0/80.058.9/55.0few-shot80.0/80.055.0/55.0few-shot-dis.90.0 † /95.0 †80.0 † /80.0 †",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Human evaluation of the system's ability to accept and object to human opinion. The left scores are by GPT-3.5, and the right scores are by ChatGPT.† indicates statistically significant scores according to McNemar's test (p < 0.01).",
"figure_data": "",
"figure_id": "tab_1",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "The accuracy for the predicted label before and after the discussion. The left scores are by and the right scores are by ChatGPT. † indicates statistically significant scores according to McNemar's test (p < 0.01). -dis. 70.15 57.24 † 55.63 † 55.19 †",
"figure_data": "SNLIR1R2R3zero-shot49.7447.4039.1041.33few-shot69.4553.5048.0048.50few-shot-dis. 66.14 53.90 † 50.40 † 50.42 †zero-shot51.8348.6341.7040.52few-shot70.3155.0852.3152.18few-shot",
"figure_id": "tab_3",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "The accuracy on SNLI and ANLI (R1, R2, R3) evaluation data. Upper scores are by GPT-3.5, and lower scores are by ChatGPT. † indicates statistically significant scores according to McNemar's test (p < 0.01).",
"figure_data": "",
"figure_id": "tab_4",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Falcon-inst. 90.3 † 71.7 † 58.4 † 57.6 †",
"figure_data": "SNLIR1R2R3MPT85.267.4 † 55.2 † 55.0 †w/ dis.MPT-inst. Falcon87.7 † 68.2 † 56.1 † 55.3 † 86.2 † 67.6 55.5 † 54.9MPT85.465.253.952.4w/o dis.MPT-inst. Falcon85.1 84.664.0 67.951.1 54.750.7 54.2Falcon-inst.85.366.253.153.0MPT86.7 † 68.3 † 55.2 † 55.0 †w/ dis.MPT-inst. Falcon86.9 88.168.8 † 56.1 † 55.3 † 68.1 55.5 54.9Falcon-inst. 90.7 † 71.9 † 58.4 † 57.6 †MPT85.465.253.952.4w/o dis.MPT-inst. Falcon86.0 88.564.0 67.951.1 54.750.7 54.2Falcon-inst.89.767.855.556.4Table 6: Accuracy on SNLI and ANLI (R1, R2, R3) testdata for fine-tuned systems with and without pseudo-discussion data. Additional fine-tuning with pseudodiscussion data for instruction tuned and non-instructiontuned models for MPT and Falcon. The upper andlower scores are the results using pseudo discussiondata generated by GPT-3.5 and ChatGPT, respectively.",
"figure_id": "tab_6",
"figure_label": "",
"figure_type": "table"
}
] | Masahiro Kaneko; Graham Neubig; Naoaki Okazaki | [
{
"authors": "Josh Abramson; Arun Ahuja; Federico Carnevale; Petko Georgiev; Alex Goldin; Alden Hung; Jessica Landon; Jirka Lhotka; Timothy Lillicrap; Alistair Muldal",
"journal": "",
"ref_id": "b0",
"title": "Improving multimodal interactive agents with reinforcement learning from human feedback",
"year": "2022"
},
{
"authors": "Panatchakorn Anantaprayoon; Masahiro Kaneko; Naoaki Okazaki",
"journal": "",
"ref_id": "b1",
"title": "Evaluating gender bias of pre-trained language models in natural language inference by considering all labels",
"year": "2023"
},
{
"authors": "Tolga Bolukbasi; Kai-Wei Chang; James Y Zou; Venkatesh Saligrama; Adam Tauman; Kalai ",
"journal": "",
"ref_id": "b2",
"title": "Man is to computer programmer as woman is to homemaker? debiasing word embeddings",
"year": "2016"
},
{
"authors": "Gabor Samuel R Bowman; Christopher Angeli; Christopher D Potts; Manning",
"journal": "Association for Computational Linguistics",
"ref_id": "b3",
"title": "A large annotated corpus for learning natural language inference",
"year": "2015"
},
{
"authors": "Jeeyoon Samuel R Bowman; Ethan Hyun; Edwin Perez; Craig Chen; Scott Pettit; Kamile Heiner; Amanda Lukosuite; Andy Askell; Anna Jones; Chen",
"journal": "",
"ref_id": "b4",
"title": "Measuring progress on scalable oversight for large language models",
"year": "2022"
},
{
"authors": "Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell",
"journal": "Advances in neural information processing systems",
"ref_id": "b5",
"title": "Language models are few-shot learners",
"year": "2020"
},
{
"authors": "Paweł Budzianowski; Ivan Vulić",
"journal": "Association for Computational Linguistics",
"ref_id": "b6",
"title": "Hello, it's GPT-2 -how can I help you? towards the use of pretrained language models for Task-Oriented dialogue systems",
"year": "2019"
},
{
"authors": "Paweł Budzianowski; Tsung-Hsien Wen; Bo-Hsiang Tseng; Iñigo Casanueva; Stefan Ultes; Milica Osman Ramadan; Gašić",
"journal": "",
"ref_id": "b7",
"title": "MultiWOZ -a Large-Scale Multi-Domain Wizard-of-Oz dataset for Task-Oriented dialogue modelling",
"year": "2018"
},
{
"authors": "Aylin Caliskan; Pimparkar Parth Ajay; Tessa Charlesworth; Robert Wolfe; Mahzarin R Banaji",
"journal": "",
"ref_id": "b8",
"title": "Gender bias in word embeddings: a comprehensive analysis of frequency, syntax, and semantics",
"year": "2022"
},
{
"authors": "Oana-Maria Camburu; Tim Rocktäschel; Thomas Lukasiewicz; Phil Blunsom",
"journal": "",
"ref_id": "b9",
"title": "E-SNLI: Natural language inference with natural language explanations",
"year": "2018"
},
{
"authors": "Jon Ander; Campos ; Jun Shern",
"journal": "",
"ref_id": "b10",
"title": "Training language models with language feedback",
"year": "2022"
},
{
"authors": "Joseph Chee; Chang ; Saleema Amershi; Ece Kamar",
"journal": "Association for Computing Machinery",
"ref_id": "b11",
"title": "Revolt: Collaborative crowdsourcing for labeling machine learning datasets",
"year": "2017"
},
{
"authors": "Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova",
"journal": "Association for Computational Linguistics",
"ref_id": "b12",
"title": "BERT: Pre-training of deep bidirectional transformers for language understanding",
"year": "2019"
},
{
"authors": "Yilun Du; Shuang Li; Antonio Torralba; Joshua B Tenenbaum; Igor Mordatch",
"journal": "",
"ref_id": "b13",
"title": "Improving factuality and reasoning in language models through multiagent debate",
"year": "2023"
},
{
"authors": "C Ruth; Andrea Fong; Vedaldi",
"journal": "",
"ref_id": "b14",
"title": "Interpretable explanations of black boxes by meaningful perturbation",
"year": "2017"
},
{
"authors": "Donghoon Ham; Jeong-Gwan Lee; Youngsoo Jang; Kee-Eung Kim",
"journal": "Association for Computational Linguistics",
"ref_id": "b15",
"title": "End-to-End neural pipeline for Goal-Oriented dialogue systems using GPT-2",
"year": "2020"
},
{
"authors": "Or Honovich; Thomas Scialom; Omer Levy; Timo Schick",
"journal": "",
"ref_id": "b16",
"title": "Unnatural instructions: Tuning language models with (almost) no human labor",
"year": "2022"
},
{
"authors": "Jiaxin Huang; Shixiang Shane Gu; Le Hou; Yuexin Wu; Xuezhi Wang; Hongkun Yu; Jiawei Han",
"journal": "",
"ref_id": "b17",
"title": "Large language models can self-improve",
"year": "2022"
},
{
"authors": "Peter Jansen; Elizabeth Wainwright; Steven Marmorstein; Clayton Morrison",
"journal": "European Language Resources Association (ELRA",
"ref_id": "b18",
"title": "WorldTree: A corpus of explanation graphs for elementary science questions supporting multi-hop inference",
"year": "2018"
},
{
"authors": "Masahiro Kaneko; Danushka Bollegala",
"journal": "Association for Computational Linguistics",
"ref_id": "b19",
"title": "Gender-preserving debiasing for pre-trained word embeddings",
"year": "2019"
},
{
"authors": "Masahiro Kaneko; Danushka Bollegala",
"journal": "Association for Computational Linguistics",
"ref_id": "b20",
"title": "Debiasing pre-trained contextualised embeddings",
"year": "2021"
},
{
"authors": "Masahiro Kaneko; Danushka Bollegala",
"journal": "Association for Computational Linguistics",
"ref_id": "b21",
"title": "Dictionary-based debiasing of pre-trained word embeddings",
"year": "2021"
},
{
"authors": "Masahiro Kaneko; Danushka Bollegala",
"journal": "",
"ref_id": "b22",
"title": "Unmasking the mask -evaluating social biases in masked language models",
"year": "2021"
},
{
"authors": "Masahiro Kaneko; Danushka Bollegala; Timothy Baldwin",
"journal": "",
"ref_id": "b23",
"title": "The gaps between pre-train and downstream settings in bias evaluation and debiasing",
"year": "2024"
},
{
"authors": "Masahiro Kaneko; Danushka Bollegala; Naoaki Okazaki",
"journal": "International Committee on Computational Linguistics",
"ref_id": "b24",
"title": "a. Debiasing isn't enough! -on the effectiveness of debiasing MLMs and their social biases in downstream tasks",
"year": "2022"
},
{
"authors": "Masahiro Kaneko; Danushka Bollegala; Naoaki Okazaki",
"journal": "Association for Computational Linguistics",
"ref_id": "b25",
"title": "Gender bias in meta-embeddings",
"year": "2022"
},
{
"authors": "Masahiro Kaneko; Danushka Bollegala; Naoaki Okazaki",
"journal": "",
"ref_id": "b26",
"title": "Comparing intrinsic gender bias evaluation measures without using human annotated examples",
"year": "2023"
},
{
"authors": "Masahiro Kaneko; Danushka Bollegala; Naoaki Okazaki",
"journal": "",
"ref_id": "b27",
"title": "The impact of debiasing on the performance of language models in downstream tasks is underestimated",
"year": "2023"
},
{
"authors": "Masahiro Kaneko; Aizhan Imankulova; Danushka Bollegala; Naoaki Okazaki",
"journal": "Association for Computational Linguistics",
"ref_id": "b28",
"title": "Gender bias in masked language models for multiple languages",
"year": "2022"
},
{
"authors": "Masahiro Kaneko; Naoaki Okazaki",
"journal": "",
"ref_id": "b29",
"title": "Controlled generation with prompt insertion for natural language explanations in grammatical error correction",
"year": "2023"
},
{
"authors": "Masahiro Kaneko; Sho Takase; Ayana Niwa; Naoaki Okazaki",
"journal": "Association for Computational Linguistics",
"ref_id": "b30",
"title": "Interpretability for language learners using example-based grammatical error correction",
"year": "2022"
},
{
"authors": "Daniel Khashabi; Xinxi Lyu; Sewon Min; Lianhui Qin; Kyle Richardson; Sean Welleck; Hannaneh Hajishirzi; Tushar Khot; Ashish Sabharwal; Sameer Singh; Yejin Choi",
"journal": "Association for Computational Linguistics",
"ref_id": "b31",
"title": "Prompt waywardness: The curious case of discretized interpretation of continuous prompts",
"year": "2022"
},
{
"authors": "Been Kim; Martin Wattenberg; Justin Gilmer; Carrie Cai; James Wexler; Fernanda Viegas",
"journal": "",
"ref_id": "b32",
"title": "Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav)",
"year": "2018"
},
{
"authors": " Pmlr",
"journal": "",
"ref_id": "b33",
"title": "",
"year": ""
},
{
"authors": "Takeshi Kojima; Shane Shixiang; Machel Gu; Yutaka Reid; Yusuke Matsuo; Iwasawa",
"journal": "",
"ref_id": "b34",
"title": "Large language models are zero-shot reasoners",
"year": "2022"
},
{
"authors": "Sawan Kumar; Partha Talukdar",
"journal": "Association for Computational Linguistics",
"ref_id": "b35",
"title": "NILE : Natural language inference with faithful natural language explanations",
"year": "2020"
},
{
"authors": "Himabindu Lakkaraju; Dylan Slack; Yuxin Chen; Chenhao Tan; Sameer Singh",
"journal": "",
"ref_id": "b36",
"title": "Rethinking explainability as a dialogue: practitioner's perspective",
"year": "2022"
},
{
"authors": "Wang Ling; Dani Yogatama; Chris Dyer; Phil Blunsom",
"journal": "Association for Computational Linguistics",
"ref_id": "b37",
"title": "Program induction by rationale generation: Learning to solve and explain algebraic word problems",
"year": "2017"
},
{
"authors": " Zachary C Lipton",
"journal": "Queue",
"ref_id": "b38",
"title": "The mythos of model interpretability: In machine learning, the concept of interpretability is both important and slippery",
"year": "2018"
},
{
"authors": "Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov",
"journal": "",
"ref_id": "b39",
"title": "Roberta: A robustly optimized bert pretraining approach",
"year": "2019"
},
{
"authors": "Mengsay Loem; Masahiro Kaneko; Naoaki Okazaki",
"journal": "",
"ref_id": "b40",
"title": "Saie framework: Support alone isn't enough -advancing llm training with adversarial remarks",
"year": "2023"
},
{
"authors": "Li Lucy; David Bamman",
"journal": "",
"ref_id": "b41",
"title": "Gender and representation bias in GPT-3 generated stories",
"year": "2021"
},
{
"authors": "Chandler May; Alex Wang; Shikha Bordia; R Samuel; Rachel Bowman; Rudinger",
"journal": "Association for Computational Linguistics",
"ref_id": "b42",
"title": "On measuring social biases in sentence encoders",
"year": "2019"
},
{
"authors": "Sewon Min; Xinxi Lyu; Ari Holtzman; Mikel Artetxe; Mike Lewis; Hannaneh Hajishirzi; Luke Zettlemoyer",
"journal": "",
"ref_id": "b43",
"title": "Rethinking the role of demonstrations: What makes in-context learning work?",
"year": "2022"
},
{
"authors": "Shikhar Murty; Christopher D Manning; Scott Lundberg; Marco Tulio; Ribeiro ",
"journal": "",
"ref_id": "b44",
"title": "Fixing model bugs with natural language patches",
"year": "2022"
},
{
"authors": "Sharan Narang; Colin Raffel; Katherine Lee; Adam Roberts; Noah Fiedel; Karishma Malkan",
"journal": "",
"ref_id": "b45",
"title": "WT5?! training Text-to-Text models to explain their predictions",
"year": "2020"
},
{
"authors": "Yixin Nie; Adina Williams; Emily Dinan; Mohit Bansal; Jason Weston; Douwe Kiela",
"journal": "Association for Computational Linguistics",
"ref_id": "b46",
"title": "Adversarial NLI: A new benchmark for natural language understanding",
"year": "2020"
},
{
"authors": "Daisuke Oba; Masahiro Kaneko; Danushka Bollegala",
"journal": "",
"ref_id": "b47",
"title": "In-contextual bias suppression for large language models",
"year": "2023"
},
{
"authors": "Dong Huk; Park ; Lisa Anne Hendricks; Zeynep Akata; Anna Rohrbach; Bernt Schiele; Trevor Darrell; Marcus Rohrbach",
"journal": "",
"ref_id": "b48",
"title": "Multimodal explanations: Justifying decisions and pointing to the evidence",
"year": "2018"
},
{
"authors": "Guilherme Penedo; Quentin Malartic; Daniel Hesslow; Ruxandra Cojocaru; Alessandro Cappelli; Hamza Alobeidli; Baptiste Pannier; Ebtesam Almazrouei; Julien Launay",
"journal": "",
"ref_id": "b49",
"title": "The RefinedWeb dataset for Falcon LLM: outperforming curated corpora with web data, and web data only",
"year": "2023"
},
{
"authors": "Alec Radford; Jong Wook Kim; Tao Xu; Greg Brockman; Christine Mcleavey; Ilya Sutskever",
"journal": "",
"ref_id": "b50",
"title": "Robust speech recognition via large-scale weak supervision",
"year": "2022"
},
{
"authors": "Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu",
"journal": "J. Mach. Learn. Res",
"ref_id": "b51",
"title": "Exploring the limits of transfer learning with a unified text-to-text transformer",
"year": "2020"
},
{
"authors": "Marco Tulio Ribeiro; Sameer Singh; Carlos Guestrin",
"journal": "",
"ref_id": "b52",
"title": "why should i trust you?\" explaining the predictions of any classifier",
"year": "2016"
},
{
"authors": "Pratyusha Sharma; Balakumar Sundaralingam; Valts Blukis; Chris Paxton; Tucker Hermans; Antonio Torralba; Jacob Andreas; Dieter Fox",
"journal": "",
"ref_id": "b53",
"title": "Correcting robot plans with natural language feedback",
"year": "2022"
},
{
"authors": "Ravid Shwartz; -Ziv ; Naftali Tishby",
"journal": "",
"ref_id": "b54",
"title": "Opening the black box of deep neural networks via information",
"year": "2017"
},
{
"authors": "Dylan Slack; Satyapriya Krishna; Himabindu Lakkaraju; Sameer Singh",
"journal": "",
"ref_id": "b55",
"title": "Talktomodel: Explaining machine learning models with interactive natural language conversations",
"year": "2022"
},
{
"authors": "Rohan Taori; Ishaan Gulrajani; Tianyi Zhang; Yann Dubois; Xuechen Li; Carlos Guestrin; Percy Liang; Tatsunori B Hashimoto",
"journal": "",
"ref_id": "b56",
"title": "Stanford alpaca: An instruction-following llama model",
"year": "2023"
},
{
"authors": "Nlp Mosaicml; Team",
"journal": "",
"ref_id": "b57",
"title": "Introducing mpt-7b: A new standard for open-source, ly usable llms",
"year": "2023"
},
{
"authors": "Hugo Touvron; Thibaut Lavril; Gautier Izacard; Xavier Martinet; Marie-Anne Lachaux; Timothée Lacroix; Baptiste Rozière; Naman Goyal; Eric Hambro; Faisal Azhar; Aurelien Rodriguez; Armand Joulin; Edouard Grave; Guillaume Lample",
"journal": "",
"ref_id": "b58",
"title": "Llama: Open and efficient foundation language models",
"year": "2023"
},
{
"authors": "Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Łukasz Kaiser; Illia Polosukhin",
"journal": "Advances in neural information processing systems",
"ref_id": "b59",
"title": "Attention is all you need",
"year": "2017"
},
{
"authors": "Boshi Wang; Sewon Min; Xiang Deng; Jiaming Shen; You Wu; Luke Zettlemoyer; Huan Sun",
"journal": "",
"ref_id": "b60",
"title": "Towards understanding chain-of-thought prompting: An empirical study of what matters",
"year": "2022"
},
{
"authors": "Xuezhi Wang; Jason Wei; Dale Schuurmans; Quoc Le; Ed Chi; Sharan Narang; Aakanksha Chowdhery; Denny Zhou",
"journal": "",
"ref_id": "b61",
"title": "Self-consistency improves chain of thought reasoning in language models",
"year": "2022"
},
{
"authors": "Zirui Wang; Adams Wei Yu; Orhan Firat; Yuan Cao",
"journal": "",
"ref_id": "b62",
"title": "Towards zero-label language learning",
"year": "2021"
},
{
"authors": "Albert Webson; Ellie Pavlick",
"journal": "Association for Computational Linguistics",
"ref_id": "b63",
"title": "Do promptbased models really understand the meaning of their prompts",
"year": "2022"
},
{
"authors": "Jason Wei; Xuezhi Wang; Dale Schuurmans; Maarten Bosma; Ed Chi; Quoc Le; Denny Zhou",
"journal": "",
"ref_id": "b64",
"title": "Chain of thought prompting elicits reasoning in large language models",
"year": "2022"
},
{
"authors": "Sarah Wiegreffe; Jack Hessel; Swabha Swayamdipta; Mark Riedl; Yejin Choi",
"journal": "Association for Computational Linguistics",
"ref_id": "b65",
"title": "Reframing human-AI collaboration for generating free-text explanations",
"year": "2022"
},
{
"authors": "Minghao Wu; Abdul Waheed; Chiyu Zhang; Muhammad Abdul-Mageed; Alham Fikri; Aji ",
"journal": "",
"ref_id": "b66",
"title": "Lamini-lm: A diverse herd of distilled models from large-scale instructions",
"year": "2023"
},
{
"authors": "Tianyi Zhang; Varsha Kishore; Felix Wu; Kilian Q Weinberger; Yoav Artzi",
"journal": "",
"ref_id": "b67",
"title": "Bertscore: Evaluating text generation with bert",
"year": "2020"
},
{
"authors": "Zhuosheng Zhang; Aston Zhang; Mu Li; Alex Smola",
"journal": "",
"ref_id": "b68",
"title": "Automatic chain of thought prompting in large language models",
"year": "2022"
},
{
"authors": "Yi Zhou; Masahiro Kaneko; Danushka Bollegala",
"journal": "Association for Computational Linguistics",
"ref_id": "b69",
"title": "Sense embeddings are also biased -evaluating social biases in static and contextualised sense embeddings",
"year": "2022"
}
] | [
{
"formula_coordinates": [
7,
390.37,
76.07,
107.71,
7.77
],
"formula_id": "formula_0",
"formula_text": "SNLI R1 R2 R3"
}
] |
|
10.18653/v1/2020.acl-main.424 | 2023-10-19 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b102",
"b70",
"b3",
"b51",
"b97",
"b42"
],
"table_ref": [],
"text": "Prompting with natural language instructions has recently emerged as a popular method of harnessing the capabilities of large language models. In addition to fine-tuning, models are often fine-tuned using instructions on a large collection of datasets Listing 1 An example pseudo-code instruction for the task from Wang et al. (2022b). A successful model is expected to use the provided pseudo-code instructions and output responses to a pool of evaluation instances.\n1 def generate_sentiment(sentence: str) -> str:\n2 \"\"\"For the given sentence, the task is to 3 predict the sentiment. For positive 4 sentiment return \"positive\" else return 5 \"negative\". to help improve the ability of LMs to follow instructions and performance on unseen tasks (Wei et al., 2022a;Wang et al., 2022b).\nHowever, natural language instructions can be ambiguous and under-specified, and therefore have multiple interpretations -including detailed instructions may not always be beneficial, as it can add to the complexity of reasoning for models. This has led to the growing body of work around 'prompt-engineering' where specialized prompting strategies are developed for different domains and task types (Zhao et al., 2021;Reynolds and McDonell, 2021;Arora et al., 2023;Liu et al., 2023;Zamfirescu-Pereira et al., 2023). In addition, inference-time prompting strategies that specifically aid multi-step reasoning have also been found to be helpful -e.g: the inclusion of chainof-thought reasoning in few-shot settings results in improved performance over standard prompts (Wei et al., 2022b), the infamous \"Let's think stepby-step\"-prompt for boosting 0-shot performance (Kojima et al., 2022).\nAlgorithm 1 Attention Block 1: function TRANSFORMERS_ATTENTION_BLOCK(Q, K, V ) 2:\nInput: Q, K, and V : input matrices."
},
{
"figure_ref": [],
"heading": "3:",
"publication_ref": [
"b81",
"b55",
"b63"
],
"table_ref": [],
"text": "Output: The output of the attention block. Given the inherent ambiguity present in natural language, it is intuitive to consider the advantages of prompting with less ambiguous prompt styles, such as the use of pseudo-code. Pseudo-code is an informal set of code-like constructs, which tend to be easy to interpret for humans but are not necessarily compilable/executable. They are often used to express complex ideas, processes, and flowsfor example, Algorithm 1 expresses a summarized version of what happens within a Multi-Head Attention block (Vaswani et al., 2017) in pseudo-code. Arguably, expressing the same ideas in natural language could result in ambiguity and would perhaps require detailed text for clarity, which adds to the complexity.\nIn light of recent successes in NLP tasks achieved by code models (Madaan et al., 2022;Zhang et al., 2023a,b), this study aims to examine the efficacy of using pseudo-code instructions for prompting as a means of enhancing model performance. This study is driven by the hypothesis that using pseudo-code as prompts could offer a natural advantage to models in NLP tasks, owing to the concise and clearer expression of ideas in pseudo-code. To test the hypothesis that prompting large language models with pseudo-code instead of natural language data could be helpful, we created pseudo-code prompts2 for 132 different tasks spanning 28 distinct task types, sourced from the Super-NaturalInstructions dataset (Wang et al., 2022b) (see Listing 1 for an example). Using these prompts along with their counterparts from natural language, we study their performance on two LLM families: BLOOM (Scao et al., 2023) and Code-Gen (Nijkamp et al., 2023). Both LLM families have been trained on natural language as well as code data.\nWe compare the performance of both styles of prompts on classification tasks, QA tasks, as well as a mix of other language generation tasks. Our experiments indicate that prompting with pseudocode instructions indeed helps, and they result in an absolute gain of 7-16 points in F1 scores on classification tasks, and 12-38% relative improvement in aggregate ROUGE-L scores across all tasks.\nContributions: In summary, our paper makes the following contributions: (i) We release a dataset of 132 pseudo-code prompts spanning 28 different task types; (ii) Through a series of detailed experiments on two publicly available open-access LLM families, we demonstrate how prompting with pseudo-code instructions results in a marked improvement in performance over prompting with natural language instructions; (iii) We include detailed ablation studies indicating that code comments, docstrings, and the structural clues encoded in pseudo-code all contribute towards the improvement in performance.\nTo the best of our knowledge, our work is the first to demonstrate how pseudo-code prompts3 can be helpful in improving the performance of pretrained LMs. Our findings not only emphasize the significance of leveraging pseudo-code for prompting but also shed light on the specific elements within pseudo-code that contribute to the observed improvements."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b88",
"b58",
"b4",
"b55"
],
"table_ref": [],
"text": "Finetuning large language models on instruction datasets can enhance their performance and even their ability to generalize to unseen tasks (Wei et al., 2021;Chung et al., 2022). Many aspects of instruction finetuning such as the number of tasks, model size, and finetuning on chain-of-thought data have been found to be useful (Chung et al., 2022). Consequently, significant efforts have been invested in manually creating instruction datasets, as well as using existing generative models to train and evaluate language models (Mishra et al., 2021;Bach et al., 2022;Wang et al., 2022b,a). The instructions available in instruction tuning datasets are mostly in natural language, but have been applied for both natural language tasks and programming tasks. But alternatives to natural language instructions such as programming language code, pseudo-code, symbols (MacCartney and Manning, 2007) etc. have not been thoroughly explored even for programming tasks. Compared to natural language, code or pseudo-code has less ambiguity due to its inherent nature of using functions or steps that contribute towards accomplishing a task. This makes them a natural choice for specifying instructions. Recently, few works (MarvinAI; Madaan et al., 2022;Zhang et al., 2023a,b) have explored code and pseudocode as inputs. Unlike contemporaneous work by Zhang et al. (2023a) we find that pseudo-code instructions indeed provide better performance over NL instructions on a wide variety of tasks."
},
{
"figure_ref": [],
"heading": "Dataset",
"publication_ref": [],
"table_ref": [],
"text": "The Super-NaturalInstructions dataset (Wang et al., 2022b) comprises 1, 616 diverse NLP tasks, and each task contains the task instruction, positive/negative examples, and instances. We sampled a mixture of 132 tasks that did not require multilingual capabilities and re-wrote instructions for a subset of this dataset using Python constructs. Note that we borrow Python constructs only to express our prompts in pseudo-code and our prompts do not result in executable Python code. Further, we do not include any additional steps/instructions that were not present in the original natural language instructions.\nAll task instructions follow the schema as described in Listing 1. The schema consists of the following elements.\nFunction Prototype: This defines the prototype of the main pseudo-code function. The function names are descriptive and summarize the task to be performed. They also include all variables passed as input along with their data types and return type. We follow the PEP 84 style guidelines for writing the pseudo-code and use strongly typed prototypes. We avoid declaring global variables whenever possible and pass them as arguments to a method. To the extent possible, we also avoid the use of classes and enumerations. Line number 1 in Listing 1 provides an example function prototype for a sentiment classification task."
},
{
"figure_ref": [],
"heading": "DocString:",
"publication_ref": [],
"table_ref": [],
"text": "The docstring provides detailed instructions on the task to be performed in natural language. Often, this is a paraphrased version of the original natural language instruction. The docstring ends with a list of parameters (with their types) being passed and the return type from the function. An example docstring for the sentiment classification task is presented in line numbers 2 to 12 in Listing 1.\nFunction Definition: This includes the bulk of the pseudo-code instruction describing how to solve the particular task. To the extent possible, the function definitions do not leave out any information contained in the docstring. Pseudo-code in the function definition are written as sub-task functions. These sub-task functions are usually not defined and often use descriptive names, arguments and variables. We include in-line comments indicating what is accomplished by the sub-task function and the role of the arguments if required. We sometimes also define secondary sub-task functions if it requires additional details or if the descriptive function name may not be adequate to specify the goal of the sub-task function. We assume the availability of basic helper functions such as concat_str, search etc., and do not include any import statements.\nLine numbers 13 to 16 present function definition for sentiment classification task. The function calls sentiment_is_positive sub-task function which checks if the sentiment of the given sentence is positive or not. This function is not explicitly defined in the instruction.\nPre-processor: Since the pseudo-code instructions expect inputs as arguments, we need to parse the inputs provided in the Super-NaturalInstructions dataset (Wang et al., 2022b) (which provides pre-formatted inputs). For each pseudo-code instruction, we also include an executable python pre-processor which is used for parsing the input."
},
{
"figure_ref": [],
"heading": "Dataset Statistics",
"publication_ref": [],
"table_ref": [
"tab_1",
"tab_3"
],
"text": "We created instructions for 132 tasks that have instructions and input/output pairs in English language. We group the tasks into three classes: Classification Tasks (Table 1), QA tasks (Table 2) and other language generation tasks (Table 3). These tasks cover a total of 28 different categories and span 72 unique datasets. For each task we sample 1000 instances for evaluation."
},
{
"figure_ref": [],
"heading": "Evaluation",
"publication_ref": [
"b63",
"b21",
"b47",
"b77",
"b35",
"b34"
],
"table_ref": [],
"text": "In order to study if instruction specification via pseudo-code results in improved performance over baseline NL English instructions, we choose to experiment with BLOOM (Scao et al., 2023), Code-Gen (Nijkamp et al., 2023) models. Our choice of models is motivated by the fact that these models have not been instruction-fine-tuned on the Natural Instructions dataset. In addition, they have both been trained on code and natural language data. McTaco (Ben Zhou and Roth, 2019), DROP (Dua et al., 2019), TREC (Li and Roth, 2002), DREAM (Sun et al., 2019), FreebaseQA (Jiang et al., 2019) Section Classification CODA-19 (Huang et al., 2020) "
},
{
"figure_ref": [],
"heading": "Sentiment Analysis",
"publication_ref": [
"b38",
"b28",
"b75",
"b5",
"b103",
"b92",
"b22",
"b32",
"b44",
"b26",
"b63",
"b21",
"b72",
"b41",
"b40",
"b60",
"b29",
"b86",
"b13",
"b66",
"b52",
"b65"
],
"table_ref": [],
"text": "The Multilingual Amazon Reviews Corpus (Keung et al., 2020), Sentiment140 (Go et al., 2009), SST-2 (Socher et al., 2013), PerSenT (Bastan et al., 2020), Amazon Review Polarity (Face), PEC (Zhong et al., 2020), Poem Sentiment (Sheng and Uthus, 2020) Text Categorization MultiNLI (Williams et al., 2018), DDO (Durmus and Cardie, 2019), SemEval-2020 Task 7 (Hossain et al., 2020) The BLOOM models are trained on the ROOTS corpus (Laurençon et al., 2022) consisting of 46 natural and 13 programming languages. On the other hand, the CodeGen models are trained on the Pile corpus (Gao et al., 2020), Google's publicly available BigQuery and BigPython datasets (Nijkamp et al., 2023). The BLOOM models have been trained on a mixture of natural language and code simultaneously. As for the CodeGen models we utilize, they were initially trained on natural language and subsequently received additional DROP (Dua et al., 2019), WinoGrande (Sakaguchi et al., 2021), QASC (Khot et al., 2020), Essential (Khashabi et al., 2017), ROPES (Lin et al., 2019a), StoryCloze (Mostafazadeh et al., 2016), Country Barcode Prefix dataset, Country Region in World dataset, Gigaword (Graff et al., 2003), GAP (Webster et al., 2018), SPOLIN (Cho and May, 2020), XL-WiC (Raganato et al., 2020) Table 3: Collection of language generation tasks used in our work training focused specifically on Python code. Our choice of models allows us to setup a controlled environment where we can study the impact of prompting in natural language and pseudo-code.\nMost recent instruction-tuned models have either seen the Super-NaturalInstructions dataset (Wang et al., 2022b) in some form (Longpre et al., 2023) or they do not have tokenizers that will meaningfully process code syntax (Raffel et al., 2020), and therefore can not be used in our study. By empirically studying the performance of models on these prompts, we hope to inform future work on training an instruction-tuned model using pseudo-code instructions."
},
{
"figure_ref": [],
"heading": "Model Configurations",
"publication_ref": [
"b63",
"b93"
],
"table_ref": [],
"text": "For all of the experiments conducted in this paper, we use BLOOM-3B, BLOOM 7B (Scao et al., 2023), CodeGen-mono 2B, and CodeGen-mono 6B (Nijkamp et al., 2023) models. The inference was performed using A100 80 GB GPUs. To accelerate the inference of all models, we utilized DeepSpeed-Inference (Aminabadi et al., 2022) in fp16, which resulted in an average inference throughput improvement of around 1.7x, compared to the standard HuggingFace (Wolf et al., 2020) inference. We used greedy decoding for all our experiments for reproducibility and restricted generated outputs to 100 tokens. Even for classification tasks, we generate the class labels using auto-regressive decoding instead of picking the class label with lowest perplexity. This is done because not all class labels can be mapped to a single token for all tasks. This technique of evaluating performance of classification tasks is often employed when using closed LLMs, such as those behind APIs (eg: OpenAI's GPT4 (OpenAI, 2023), Google's PaLM (Chowdhery et al., 2022) etc)."
},
{
"figure_ref": [],
"heading": "Metrics",
"publication_ref": [
"b8"
],
"table_ref": [],
"text": "We adopt different metrics for each task-category: we measure the performance of classification tasks using micro, macro and weighted F1 scores, and for QA and language generation tasks we use the ROUGE-L metric. We report the ROUGE-L, Exact Match (EM), and ANLS -Average Normalized Levenshtein Similarity (Biten et al., 2019) for all tasks."
},
{
"figure_ref": [],
"heading": "Output post-processing",
"publication_ref": [],
"table_ref": [],
"text": "Since the models we experiment with have not been fine-tuned for instruction following, they tend to generate excess text after the output for the given task. We therefore post-process the outputs to ensure models are not penalized in our evaluation due to excess generations. We post-process all outputs by truncating by the newline character '\\n'. Furthermore, the output is subjected to additional post-processing, including punctuation removal and lower casing."
},
{
"figure_ref": [],
"heading": "Results",
"publication_ref": [],
"table_ref": [],
"text": "Through our experiments we aim to answer the following questions: (i) What is the difference in performance between prompting pre-trained language and code models with pseudo-code prompts versus natural language prompts? (ii) How does increasing model size affect the efficacy of pseudocode prompts? (iii) To what extent does structured prompting, such as the use of function names, docstrings, inline comments, and arguments, impact performance on tasks?"
},
{
"figure_ref": [],
"heading": "Prompting with Pseudo-code",
"publication_ref": [
"b25",
"b55"
],
"table_ref": [
"tab_4"
],
"text": "Table 4 compares the performance of prompting with pseudo-code (referred to as code instructions) and natural language instructions in 0-shot settings.\nResults have been grouped by model family and size.\nAs can be seen, for all model families and sizes, prompting with pseudo-code results in a significant improvement in performance. The performance on classification tasks is especially notable, for example, the gains on weighted F1 vary between 7-16 F1 points (absolute). Furthermore, the relative performance improvement on all other tasks, as measured by ROUGE-L, varies between 12-38%. The overall performance as measured by ROUGE-L, ANLS and Exact Match also report similar trends.\nComparison of CodeGen vs BLOOM Despite most tasks being non-code tasks, CodeGen, a model designed for code applications, outperforms BLOOM models, even when using natural language instructions (see metrics for 'All Tasks'). Similar behavior has been anecdotally reported (Fu and Khot, 2022;Madaan et al., 2022), but has possibly not been investigated using as many tasks as presented in this paper. Note, however, that using pseudo-code prompts in the code models results in better performance than any other prompt-model configuration.\nPerformance on QA tasks Interestingly, we find that on QA tasks, the performance of pseudo-code instructions is better than natural-language instructions, when using the CodeGen model. However, this is not the case when using BLOOM. Table 5: 0-shot performance of CodeGen 6B and BLOOM 7B models on QA tasks from our dataset. As can be seen, pseudo-code instructions applied on the CodeGen model results in the best overall performance on all categories of QA tasks. However, comparing the performance of Natural Language Instructions, we find that it performs marginally better than pseudo-code instructions on non-MCQ QA tasks when using the BLOOM 7B model.\nWe investigated this further and observed that for most QA tasks, the instructions in pseudo-code are not significantly more detailed or easier to understand than natural-language instructions. As an example, the pseudo-code instruction for answer generation from the SQuAD dataset merely contains the following statement in its function definition: return get_answer_from_passage(passage, question) and reflects the details included in the natural instructions.\nWe further analysed the results across QA task categories and found that pseudo-code instructions always help with multiple-choice questions (MCQ) tasks (see Table 5 for a comparison between Code-Gen 6B and BLOOM 7B). We believe that this is because, understanding the instructions in such tasks may be more involved. For illustration, instructions in MCQ tasks often include details about how answers are expected -eg: \"choose the correct option A, B, C \", \"Select Option 1 -Value 1, Option 2 -Value 2 \". Depending on the instructions, the models may be required to return options, values, or both which adds a degree of complexity to the instructions as compared to other types of QA.\nThe discrepancy in performance between Code-Gen and BLOOM on QA tasks (see Table 5), could be attributed to the fact that the structure from code prompts could be better leveraged by code models as programming languages and aspects of code syntax (structure) are likely to be better represented in a code model such as CodeGen. This brings us to our next question -What is the contribution of structure that may be present in prompts?"
},
{
"figure_ref": [],
"heading": "Contribution of Structure in prompts",
"publication_ref": [],
"table_ref": [
"tab_5"
],
"text": "The reasons behind the performance improvement when using pseudo-code prompts are likely to be a combination of factors, including the use of descriptive function names that convey the function's purpose (such as get_answer(question)), a model that can effectively utilize structured information, and a structured prompt for a task that could further benefit from few-shot examples.\nWe therefore experiment with different structured prompting styles and report their results in Table 6. We study the performance of CodeGen and BLOOM with five types of prompts: (i) Pseudocode instructions, (ii) Prompts that make use of function declaration (declare function name only), (iii) a structured prompt consisting only of task examples in 2-shot settings using the task-descriptive function name (iv) a structured prompt consisting only of task examples in 2-shot settings using a generic function name -'func' (v) using the Natural Language examples (without instructions) in 2-shot settings. Details about each prompt have been included in the Appendix. We make three important observations from Table 6. First, code-instructions in 0-shot settings consistently yield the best overall performance compared to other structured prompts. Second, on average, the CodeGen model consistently outperforms BLOOM on all tasks. Lastly, the QA tasks in our dataset, which are relatively easy to express in natural language instructions, also benefit from structured prompts, particularly when prompted with examples.\nIt can be inferred from these observations that the performance gains resulting from the use of pseudo-code prompts are likely due to clearer task instructions, and not just the exploitation of superfluous patterns from in-context learning. These findings reinforce the results from the previous ex-periment, which showed that code models are more capable of exploiting structured prompts. In the case of QA tasks in our dataset, it is worth noting that since the pseudo-code instructions are not as detailed, even utilizing a simpler structured prompt with examples can significantly enhance performance as compared to natural language prompts."
},
{
"figure_ref": [],
"heading": "Impact of pseudo-code documentation",
"publication_ref": [],
"table_ref": [
"tab_6"
],
"text": "In this section, we study the contribution of comments and docstrings present in our pseudo-code instructions towards the improvement in performance. We first study the performance of pseudocode prompts with and without the use of docstrings and code comments.\nAs can be seen in Table 7, the inclusion of comments as well as the docstring in the pseudo-code instruction prompt helps improve performance. This indicates that not only is the structure of the prompts being exploited by the model, the models are also relying on additional helper text present in the documentation. We, therefore, also investigate if the use of these elements from pseudo-code could also benefit natural language instruction prompts.\nThe lower half of table 7 studies the performance of natural-language prompts with and without the use of pseudo-code comments and docstring. We find that the performance of natural language instructions also improves by the inclusion of comments and docstring for each model family and configuration. We hypothesize that the gains may be attributable to a form of step-by-step reasoning derived from pseudo-code comments especially in complex tasks."
},
{
"figure_ref": [],
"heading": "Summary of findings",
"publication_ref": [],
"table_ref": [
"tab_4",
"tab_5",
"tab_6",
"tab_4",
"tab_4"
],
"text": "We now summarize our findings for easy reference.\nEffect of Prompting Style: From Table 4 we observe that 0-shot prompting of pre-trained models with pseudo-code prompts results in better performance than natural language prompts. This is true for both code models and language models. The gains are more pronounced for the code models. Effect of Structure in prompts: Pseudo-code prompts include many elements such as the function declaration, docstring, comments etc. From Table 6 we find that while information from the function declaration, and a task-indicative function name help, using the complete pseudo-code prompt is most useful. Further, from Table 7 we find that the pseudocode instruction still works better than any prompt created with natural language instructions, even when docstring and comments from pseudo-code are included in the natural language instruction. This suggests the gains from prompting in pseudocode are not just due to comments and docstrings (which could help reinforce the task instructions), but also due to clearer instructions in pseudo-code. Effect of Model Size: From Table 4 we find that in 0-shot settings, with the increase in scale, the performance of pseudo-code instructions improves for both model families. However, when using natural language instructions, this is not the case. We hypothesize, that since none of these models are instruction-tuned, larger scales exacerbate the propensity of the models being primed for language completion. Code vs. Natural Language models: We find that code models are better suited for exploiting pseudo-code prompts compared to language models. As can be seen from Table 4 (see metrics for 'All Tasks'), the use of natural language instructions on CodeGen results in better performance than their use on BLOOM."
},
{
"figure_ref": [],
"heading": "Conclusion and Future Work",
"publication_ref": [],
"table_ref": [],
"text": "In this paper we presented our work on prompting with pseudo-code instructions. We created a collection of pseudo-code instructions comprising of 132 NLP tasks from the Super-NaturalInstructions dataset (Wang et al., 2022b). We evaluated the performance of the following families of models -CodeGen and BLOOM at different model sizes and found that prompting all models with pseudo-code instructions results in significant gains as compared to prompting with NL instructions. Our work opens up multiple directions of future work. It is interesting to observe that not only do pseudo-code instructions help when used with code models, they also work better on models designed for natural language tasks. In addition, the fact that code mod-els used in our experiments perform better than NL models, even when prompted with natural language instructions, suggests that it could be useful to explore instruction tuning of code models instead of pure NL models for NL applications. Based on the findings of this paper it may also be useful to consider the effects of instruction fine-tuning with pseudo-code instructions as opposed to NL instructions.\nAnother aspect worth studying is how traditional chain-of-thought may compare with pseudo-code prompts -how would reasoning enabled by pseudocode instructions compare with chain-of-thought reasoning with and without fine-tuning? Further, pseudo-code instructions may not only be used as direct inputs to a model, but they could also be used to create intermediate responses that a model needs to generate prior to returning a response."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [],
"table_ref": [],
"text": "Our results have been reported on two model families -CodeGen and BLOOM at scales of 2-7B parameters. It remains to be seen if our findings would hold at larger model sizes. It is possible that better reasoning enabled by larger model sizes could reduce the benefit of prompting with pseudocode instructions but we have not investigated this in our work. In addition, our work does not include any multi-lingual NLP tasks -BLOOM was specifically trained to be able to support multiple languages and it is possible this model design choice could play a role in our findings when we compare code (CodeGen) and NL (BLOOM) models against each other. Moreover, both models have been trained on different datasets and this also affects the intrinsic reasoning capabilities of these models. Lastly, and importantly, the use of pseudocode for prompting LLMs is limited by the expectation that it requires technical expertise to write them, thus reducing their widespread usage. Language Processing (EMNLP), pages 6556-6566, Online. Association for Computational Linguistics."
},
{
"figure_ref": [],
"heading": "A Appendix",
"publication_ref": [
"b0"
],
"table_ref": [],
"text": "A.1 Results on Various LLMs\nWe also perform experiments using Falcon-7B (Almazrouei et al., 2023) model. The results are presented in Table 8."
},
{
"figure_ref": [],
"heading": "A.2 Pseudo-Code Validation",
"publication_ref": [],
"table_ref": [],
"text": "To ensure that the pseudo-code instructions follow the guidelines provided, we run an automatic test. The test code calls the preprocess function defined for each example from the Super-NaturalInstructions dataset (Wang et al., 2022b) for that task. The returned values from the preprocess function are compared against the arguments in the function prototype. Any mismatch in the data type or the number of arguments results in error. The instruction creator is given feedback to correct the errors."
},
{
"figure_ref": [],
"heading": "A.2.1 Prompt Styles",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we describe the various prompting styles used to study the effect of pseudo-code vs NL prompting. Here, we show a simple task to generate the sentiment of a given sentence. This is task 833 in Super-NaturalInstructions dataset."
},
{
"figure_ref": [],
"heading": "A.2.2 Prompting with Pseudo-code instructions",
"publication_ref": [],
"table_ref": [],
"text": "Listing 2 Code instructions (0-shot prompt) for sentiment classification task def generate_sentiment(sentence: str) -> str: \"\"\"For the given sentence, the task is to predict the sentiment. For positive sentiment return \"positive\" else return \"negative\".\nParameters: sentence (str): input sentence Returns:\nstr: sentiment of the input \"\"\" # predict the sentiment if sentiment_is_positive(sentence):\nreturn \"positive\" else:\nreturn \"negative\" >>> generate_sentiment( \"that has a charmingly bourbon air.\" ) Listing 3 Code instructions (2-shot prompt) for sentiment classification task def generate_sentiment(sentence: str) -> str:\n\"\"\"For the given sentence, the task is to predict the sentiment. For positive sentiment return \"positive\" else return \"negative\".\nParameters: sentence (str): input sentence Returns:\nstr: sentiment of the input \"\"\" # predict the sentiment if sentiment_is_positive(sentence):\nreturn \"positive\" else:\nreturn \"negative\" >>> generate_sentiment( \"tormented by the quickened blood of the \" \"roots\" ) \"negative\" >>> generate_sentiment( \"radiant as moses from the mount, he stood\" ) \"positive\" >>> generate_sentiment( \"that has a charmingly bourbon air.\" )\nFor the pseudo-code prompting, we use the instructions that are created by the authors of this paper. The pseudo-code instructions have a much richer structure than natural language instructions and are more elaborate and simple to understand. They contain docstrings, return types and might also contain comments, function invocations etc. For preparing the few shot examples and the input query, we treat the example as a python interpreter running in the linux terminal and use the special markers '>>>' for the input. We don't use any special markers for the outputs. An example for 0-shot and 2-shot shot prompting is shown in Listings 2 and 3 respectively.\nWe also measure the impact of removing the docstrings and comments from the code instruction. An example for 0-shot and 2-shot prompting is shown in Listings 4 and 5 respectively."
},
{
"figure_ref": [],
"heading": "A.2.3 Prompting with function prototype",
"publication_ref": [],
"table_ref": [],
"text": "We try prompting the models with function prototypes with all docstrings, comments and code logic removed from the base pseudo-code instruction. The function prototype instructions are composed of the function names, arguments and their types"
},
{
"figure_ref": [],
"heading": "",
"publication_ref": [],
"table_ref": [],
"text": "Table 8: Performance of models when prompted using pseudo-code instructions and natural language instructions in 0-shot settings. (i) In each model, prompting with pseudo-code instructions results in much higher performance in almost all the tasks\nListing 4 Code instructions without docstrings and comments (0-shot prompt) for sentiment classification task def generate_sentiment(sentence: str) -> str: if sentiment_is_positive(sentence): return \"positive\" else:\nreturn \"negative\" >>> generate_sentiment( \"that has a charmingly bourbon air.\" )\nListing 5 Code instructions without docstrings and comments (2-shot prompt) for sentiment classification task def generate_sentiment(sentence: str) -> str:\nif sentiment_is_positive(sentence): return \"positive\" else:\nreturn \"negative\" >>> generate_sentiment( \"tormented by the quickened blood of the \" \"roots\" ) \"negative\" >>> generate_sentiment( \"radiant as moses from the mount, he stood\" ) \"positive\" >>> generate_sentiment( \"that has a charmingly bourbon air.\" ) and the return types. This method of prompting is devoid of any pseudo-code. An example for 0-shot and 2-shot prompting is shown in Listings 6 and 7 respectively."
},
{
"figure_ref": [],
"heading": "A.2.4 Prompting with NL instructions",
"publication_ref": [],
"table_ref": [],
"text": "For natural language prompts, we use the original instructions provided as part of the Super-NaturalInstructions dataset (Wang et al., 2022b). For natural language instruction prompting, we Listing 6 Function prototype (0-shot prompt) for sentiment classification task def generate_sentiment(sentence: str) -> str:\n>>> generate_sentiment( \"that has a charmingly bourbon air.\" )\nListing 7 Function prototype (2-shot prompt) for sentiment classification task def generate_sentiment(sentence: str) -> str:\n>>> generate_sentiment( \"tormented by the quickened blood of the \" \"roots\" ) \"negative\" >>> generate_sentiment( \"radiant as moses from the mount, he stood\" ) \"positive\" >>> generate_sentiment( \"that has a charmingly bourbon air.\" ) use the prompts provided as part of the Super-NaturalInstructions dataset without any modification. We add special 'input:' and 'output:' markers in the few shot examples and the input query to the model as shown in Listings 8 and 9.\nListing 8 Natural instructions (0-shot prompt) for sentiment classification task\nIn this task, you need to identify the sentiment of the given sentence as one of \"positive\" or \"negative\".\ninput: that has a charmingly bourbon air. output:\nListing 9 Natural instructions (2-shot prompt) for sentiment classification task\nIn this task, you need to identify the sentiment of the given sentence as one of \"positive\" or \"negative\". input: tormented by the quickened blood of the roots output: negative input: radiant as moses from the mount, he stood output: positive input: that has a charmingly bourbon air. output:"
},
{
"figure_ref": [],
"heading": "A.2.5 Prompting with NL instructions and NL comments from the pseudo-code",
"publication_ref": [],
"table_ref": [],
"text": "We also try experimenting by adding the docstrings and comments to the NL instructions from the Super-NaturalInstructions dataset (Wang et al., 2022b) as shown in the example in Listings 10 and 11.\nListing 10 Natural instructions with docstrings (0shot prompt) for sentiment classification task\nIn this task, you need to identify the sentiment of the given sentence as one of \"positive\" or \"negative\".\n\"\"\"For the given sentence, the task is to predict the sentiment. For positive sentiment return \"positive\" else return \"negative\".\nParameters: sentence (str): input sentence Returns:\nstr: sentiment of the input \"\"\" # predict the sentiment input: that has a charmingly bourbon air. output:"
},
{
"figure_ref": [],
"heading": "A.2.6 Prompting without instructions",
"publication_ref": [],
"table_ref": [],
"text": "We also study the effect of prompting without instructions. We try this method of prompting in three settings: In this task, you need to identify the sentiment of the given sentence as one of \"positive\" or \"negative\".\n\"\"\"For the given sentence, the task is to predict the sentiment. For positive sentiment return \"positive\" else return \"negative\".\nParameters: sentence (str): input sentence Returns:\nstr: sentiment of the input \"\"\" # predict the sentiment input: tormented by the quickened blood of the roots output: negative input: radiant as moses from the mount, he stood output: positive input: that has a charmingly bourbon air. output:\nListing 12 Function invocation (0-shot prompt) for sentiment classification task Table 9: Performance with 2-shot prompts. (i) In each model, prompting with pseudo-code instructions results in much higher performance (ii) For each model family, increasing scale helps improve performance (iii) As before, prompting a model designed for code, CodeGen results in better performance than BLOOM. (iv) Surprisingly, as compared to 0-shot prompting (Table 4), there is a marked drop in performance for all model configurations and all tasks, except in QA tasks, where there is an improvement in performance.\nListing 15 Generic function invocation (2-shot prompt) for sentiment classification task >>> func( \"tormented by the quickened blood of the \" \"roots\" ) \"negative\" >>> func( \"radiant as moses from the mount, he stood\" ) \"positive\" >>> func( \"that has a charmingly bourbon air.\" )\nListing 16 Natural examples (0-shot prompt) for sentiment classification task input: that has a charmingly bourbon air. output:\nListing 17 Natural examples (2-shot prompt) for sentiment classification task input: tormented by the quickened blood of the roots output: negative input: radiant as moses from the mount, he stood output: positive input: that has a charmingly bourbon air. output:"
},
{
"figure_ref": [],
"heading": "A.3 2-shot Prompting with Pseudo-code instructions",
"publication_ref": [
"b70"
],
"table_ref": [],
"text": "Given that structured prompts, such as those based on function declarations, benefit from 2-shot prompts, we investigate whether the performance of pseudo-code prompts can be further improved with 2-shot prompts. Table 9 reports the performance of both families of models -CodeGen and BLOOM when using pseudo-code prompts and natural language instruction prompts in 2-shot settings.\nInterestingly we find that, as compared to the results reported in Table 4 the performance of each corresponding model-prompt configuration is lower than its 0-shot counterpart. While this may appear surprising, similar findings have been reported in prior work (Reynolds and McDonell, 2021;Zhang et al., 2023a). Perhaps the performance in few-shot settings could improve with additional examples, but we do not experiment with more than 2-shot settings due to limitations imposed by the size of input context length available to models.\nAfter a study of outputs generated by the models in 2-shot settings, we observe that in many cases, in the absence of extensive task-specific prompt-engineering and output processing, models are likely to generate additional continuation examples instead of solving the task. The fact that the pseudo-code prompts perform better indicate that models seem to \"interpret\" the instructions better in this form."
},
{
"figure_ref": [],
"heading": "A.4 Ablation Experiments",
"publication_ref": [],
"table_ref": [],
"text": "As can be seen in Table 10 and 11, the inclusion of comments as well as the docstring in the pseudocode instruction prompt and natural language instructions helps improve performance for smaller models too. "
}
] | Prompting with natural language instructions has recently emerged as a popular method of harnessing the capabilities of large language models (LLM). Given the inherent ambiguity present in natural language, it is intuitive to consider the possible advantages of prompting with less ambiguous prompt styles, like pseudocode. In this paper, we explore if prompting via pseudo-code instructions helps improve the performance of pre-trained language models. We manually create a dataset 1 of pseudo-code prompts for 132 different tasks spanning classification, QA, and generative language tasks, sourced from the Super-NaturalInstructions dataset (Wang et al., 2022b). Using these prompts along with their counterparts in natural language, we study their performance on two LLM families -BLOOM (Scao et al., 2023), CodeGen (Nijkamp et al., 2023). Our experiments show that using pseudo-code instructions leads to better results, with an average increase (absolute) of 7-16 points in F1 scores for classification tasks and an improvement (relative) of 12-38% in aggregate ROUGE-L scores across all tasks. We include detailed ablation studies which indicate that code comments, docstrings, and the structural clues encoded in pseudo-code all contribute towards the improvement in performance. To the best of our knowledge, our work is the first to demonstrate how pseudocode prompts can be helpful in improving the performance of pre-trained LMs. | Prompting with Pseudo-Code Instructions | [
{
"figure_caption": "Collection of classification tasks used in our work",
"figure_data": ",2017)Wrong Candidate GenerationMcTaco (Ben Zhou and Roth, 2019)",
"figure_id": "tab_1",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Collection of QA tasks used in our work",
"figure_data": "6List Operation3Option Generation215ParaphrasingQuestion GenerationRewriting1110Misc.Task CategoryDatasetsList OperationCoNaLa (Yin et al., 2018), Synthetic (Tiedemann, 2012),Youtube Caption Corrections (2dot71mily)Option GenerationaNLI (Nie et al., 2020), ASSET (Alva-Manchego et al.,2020), ROCStories (Mostafazadeh et al., 2017)ParaphrasingZEST (Weller et al., 2020), PARANMT-50M (Wieting andGimpel, 2018)Question GenerationCosmosQA (Huang et al., 2019), WinoGrande (Sakaguchiet al., 2021), ROPES (Lin et al., 2019b), SQuAD1.1 (Ra-jpurkar et al., 2016), StrategyQA (Geva et al., 2021),SQuAD2.0 (Rajpurkar et al., 2018), BoolQ (Clark et al.,2019), CoQA (Reddy et al., 2019), QA-ZRE (Levy et al.,2017)RewritingWinoGrande (Sakaguchi et al., 2021), aNLI (Nie et al., 2020),ASSET (Alva-Manchego et al., 2020), ZEST (Weller et al.,2020), SNLI (Bowman et al., 2015)Misc.",
"figure_id": "tab_3",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Performance of models when prompted using pseudo-code instructions and natural language instructions in 0-shot settings. (i) In each model, prompting with pseudo-code instructions results in much higher performance in almost all the tasks (ii) For each model family, increasing scale helps improve performance (iii) Prompting CodeGen (a model designed for code) results in better performance than BLOOM. (iv) Prompting BLOOM models with Natural Language instructions instead of code-instructions results in higher performance on QA tasks.",
"figure_data": "ModelInstruction FormatClassification TasksQA TasksGeneration tasksAll TasksMacro F1Micro F1Weighted F1ROUGE-LROUGE-LROUGE-LANLSEMMajority Class0.2960.5090.362-----CodeGen 2BCode Instructions0.2720.4170.3540.1750.3170.3300.2610.202NL Instructions0.0680.3060.2390.1540.2540.2650.1950.147CodeGen 6BCode Instructions0.3110.4430.3750.2010.3270.3540.2830.218NL Instructions0.0520.2780.2150.1320.2710.2570.1870.134BLOOM 3BCode Instructions0.1160.3510.2880.1470.2710.2790.2150.165NL Instructions0.0820.2750.2140.1590.2340.2500.1800.132BLOOM 7BCode Instructions0.1740.3690.2850.1500.2980.2970.2320.176NL Instructions0.0460.2470.2030.1560.2760.2470.1720.122CodeGen 6BBLOOM 7BCode InstructionsNL InstructionsCode InstructionsNL InstructionsQA TaskEMROUGE-LANLSEMROUGE-LANLSEMROUGE-LANLSEMROUGE-LANLSExtractive QA0.1400.3030.1890.0450.1880.0770.0470.1840.0770.0470.2270.086Generative QA0.0450.1290.0680.0290.0950.0450.0280.1010.0420.0320.1150.047MCQ0.1960.2130.2100.0820.1060.0830.1840.2010.1970.1070.1430.108",
"figure_id": "tab_4",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Study of structured prompts: Performance of models when prompted using 0-shot pseudo-code instructions, function declaration in 0-shot and 2-shot settings as well as 2-shot prompting with a 'generic' function name and the use of only examples. The number N in the brackets indicates N-shot prompt. (i) Except for the performance on QA tasks, in each model, prompting with pseudo-code instructions results in much higher performance which indicates that detailed instructions are helpful (ii) For each model family, and prompting style, increasing model scale improves performance (iii) As before, prompting a model designed for code, CodeGen, results in better performance than BLOOM.",
"figure_data": "ModelInstruction FormatClassification TasksQA TasksGeneration TasksAll TasksMacro F1Micro F1Weighted F1ROUGE-LROUGE-LROUGE-LANLSEMCode Instructions (0)0.2720.4170.3540.1750.3170.3300.2620.202Function Declaration (0)0.1590.0790.0850.1240.2520.1530.0830.043CodeGen 2BFunction Declaration (2)0.1050.2670.2570.1850.2940.2560.1880.137Function Invocation (2)0.0970.2530.2380.1830.2960.2510.1830.131Generic Function Invocation (2)0.0640.2820.2440.1670.2570.2450.1850.131NL Examples (2)0.0030.0050.0070.0810.1260.0690.0170.006Code Instructions (0)0.3110.4440.3750.2010.3270.3540.2830.218Function Declaration (0)0.0190.1010.1090.1620.2730.1790.1110.063CodeGen 6BFunction Declaration (2)0.1340.3090.2810.1960.2990.2810.2120.154Function Invocation (2)0.1330.2960.2690.1920.3020.2750.2080.149Generic Function Invocation (2)0.0620.2440.2150.1670.2620.2390.1750.121NL Examples (2)0.0000.0000.0010.1020.1680.0880.0230.006Code Instructions (0)0.1160.3510.2880.1470.2710.2790.2140.165Function Declaration (0)0.0000.0140.0160.1080.2290.1160.0540.015BLOOM 3BFunction Declaration (2)0.0800.2370.2170.1640.2490.2250.1590.115Function Invocation (2)0.0730.2270.2110.1640.2340.2150.1490.107Generic Function Invocation (2)0.0320.1730.1680.1610.2460.2030.1370.086NL Examples (2)0.0000.0250.0310.1500.2080.1220.0560.024Code Instructions (0)0.1740.3690.2850.1500.2980.2970.2320.176Function Declaration (0)0.0040.0210.0270.1110.2420.1240.0580.017BLOOM 7BFunction Declaration (2)0.0720.2560.2270.1910.2890.2570.1820.128Function Invocation (2)0.0860.2480.2210.1890.2860.2500.1760.123Generic Function Invocation (2)0.0390.1990.1780.1870.2760.2320.1550.097NL Examples (2)0.0000.0090.0090.1320.1820.1060.0380.016",
"figure_id": "tab_5",
"figure_label": "6",
"figure_type": "table"
},
{
"figure_caption": "Ablation: Zero-Shot Setting. (i) In each model, prompting with pseudo-code instructions results in much higher performance on QA and classification tasks (ii) For each model family, increasing scale helps improve performance (iii) As before, prompting a model designed for code, CodeGen results in better performance than BLOOM. On average, in the CodeGen model, the use of code comments and docstrings helps improve the performance of natural language prompts. However, it appears for BLOOM, only the larger-sized model is able to consistently use the additional details in the prompt to improve performance.",
"figure_data": "ModelInstruction FormatClassification TasksQA TasksGeneration TasksAll TasksMacro F1Micro F1Weighted F1ROUGE-LROUGE-LROUGE-LANLSEMCodeGen 6BCode Instructions0.3110.4440.3750.2010.3270.3540.2830.218Code Instructions without docstrings and comments0.2630.4090.3480.1950.3270.3350.2660.201BLOOM 7BCode Instructions0.1740.3690.2850.1500.2980.2970.2320.176Code Instructions without docstrings and comments0.1450.3160.2470.1440.2910.2690.2040.151CodeGen 6BNL Instructions0.0520.2780.2150.1320.2710.2570.1870.134NL Instructions with docstrings and comments0.0620.3120.2540.1390.2930.2750.2080.148BLOOM 7BNL Instructions0.0460.2470.2030.1560.2760.2470.1720.122NL Instructions with docstrings and comments0.0440.3030.2330.1650.2630.2660.1990.147",
"figure_id": "tab_6",
"figure_label": "7",
"figure_type": "table"
}
] | Mayank Mishra; Prince Kumar; Riyaz Bhat; Rudra Murthy; Danish Contractor; Srikanth Tamilselvam; Eyal Bar Kim; Francesco Natan; Gérard De Toni; Germán Dupont; Giada Kruszewski; Hady Pistilli; Hamza Elsahar; Hieu Benyamina; Ian Tran; Idris Yu; Isaac Abdulmumin; Itziar Johnson; Javier Gonzalez-Dios; Jenny De La Rosa; Jesse Chim; Jian Dodge; Jonathan Zhu; Jörg Chang; Joseph Frohberg; Joy- Deep Tobing; Khalid Bhattacharjee; Kimbo Almubarak; Kyle Chen; Leandro Lo; Leon Werra; Long Weber; Loubna Phan; Ben; Ludovic Tanguy; Manan Dey; Manuel Romero Muñoz; Maraim Masoud; María Grandury; Mario Šaško; Max Huang; Max- Imin Coavoux; Mayank Singh; Tian-Jian Jiang; Minh Chien Vu; Mohammad A Jauhar; Mustafa Ghaleb; Nishant Subramani; Nora Kassner; Nuru- Laqilla Khamis; Olivier Nguyen; Omar Espejel; Ona De Gibert; Paulo Villegas; Peter Henderson; Pierre Colombo; Priscilla Amuok; Quentin Lhoest; Rheza Harliman; Rishi Bommasani; Roberto Luis López; Rui Ribeiro; Salomey Osei; Sampo Pyysalo; Se- Bastian Nagel; Shamik Bose; Shamsuddeen Hassan Muhammad; Shanya Sharma; Shayne Longpre; So- Maieh Nikpoor; Stanislav Silberberg; Suhas Pai; Syd- Ney Zink; Tiago Timponi Torrent; Timo Schick; Tris- Tan Thrush; Valentin Danchev; Vassilina Nikoulina; Veronika Laippala; Violette Lepercq; Vrinda Prabhu; Zaid Alyafeai; Zeerak Talat; Arun Raja; Benjamin Heinzerling; Chenglei Si; Emre Davut; Eliz- Abeth Taşar; Sabrina J Salesky; Wilson Y Mielke; Abheesht Lee; Andrea Sharma; Antoine Santilli; Arnaud Chaffin; Debajyoti Stiegler; Eliza Datta; Gunjan Szczechla; Han Chhablani; Harshit Wang; Hen- Drik Pandey; Jason Alan Strobelt; Jos Fries; Leo Rozen; Lintang Gao; Sutawika; Saiful Bari; Maged S Al-Shaibani; Matteo Manica; Nihal Nayak; Ryan Teehan; Samuel Albanie; Sheng Shen; Srulik Ben- David; Stephen H Bach; Taewoon Kim; Tali Bers; Thibault Fevry; Trishala Neeraj; Urmish Thakker; Vikas Raunak; Xiangru Tang; Zheng-Xin Yong; Zhiqing Sun; Shaked Brody; Yallow Uri; Hadar Tojarieh; Adam Roberts; Won Chung; Jae- Sung Tae; Jason Phang; Ofir Press; Conglong Li; Deepak Narayanan; Hatim Bourfoune; Jared Casper; Jeff Rasley; Max Ryabinin; Minjia Zhang; Mohammad Shoeybi; Myriam Peyrounette; Nicolas Patry; Nouamane Tazi; Omar Sanseviero; Patrick Von Platen; Pierre Cornette; Pierre François Lavallée; Rémi Lacroix; Samyam Rajbhandari; San- Chit Gandhi; Shaden Smith; Stéphane Requena; Suraj Patil; Tim Dettmers; Ahmed Baruwa; Amanpreet Singh; Anastasia Cheveleva; Anne-Laure Ligozat; Arjun Subramonian; Aurélie Névéol; Dan Garrette; Deepak Tunuguntla; Ehud Reiter; Ekaterina Taktasheva; Ekaterina Voloshina; Eli Bog- Danov, Genta; Indra Winata; Hailey Schoelkopf; Jan- Christoph Kalo; Jekaterina Novikova; Jessica Zosa Forde; Jordan Clive; Jungo Kasai; Ken Kawamura; Liam Hazan; Marine Carpuat; Miruna Clinciu; Na- Joung Kim; Newton Cheng; Oleg Serikov; Omer Antverg; Oskar Van Der Wal; Rui Zhang; Ruochen Zhang; Sebastian Gehrmann; Shachar Mirkin; Shani Pais; Tatiana Shavrina; Thomas Scialom; Tian Yun; Tomasz Limisiewicz; Verena Rieser; Vitaly Protasov; Vladislav Mikhailov; Yada Pruksachatkun; Yonatan Belinkov; Zachary Bamberger; Zdeněk Kasner; Al- Ice Rueda; Amanda Pestana; Amir Feizpour; Am- Mar Khan; Amy Faranak; Ana Santos; Anthony Hevia; Antigona Unldreaj; Arash Aghagol; Are- Zoo Abdollahi; Aycha Tammour; Azadeh Hajihos- Seini; Bahareh Behroozi; Benjamin Ajibade; Bharat Saxena; Carlos Muñoz Ferrandis; Danish Contrac- Tor; David Lansky; Davis David; Douwe Kiela; Duong A Nguyen; Edward Tan; Emi Baylor; Ez- Inwanne Ozoani; Fatima Mirza; Habib Rezanejad; Hessie Jones; Indrani Bhat- Tacharya; Irene Solaiman; Irina Sedenko; Isar Ne- Jadgholi; Jesse Passmore; Josh Seltzer; Julio Bonis Sanz; Livia Dutra; Mairon Samagaio; Maraim El- Badri; Margot Mieskes; Marissa Gerchick; Martha Akinlolu; Michael Mckenna; Mike Qiu; Muhammed Ghauri; Mykola Burynok; Nafis Abrar; Nazneen Ra- Jani; Nour Elkott; Nour Fahmy; Olanrewaju Samuel; Ran An; Rasmus Kromann; Ryan Hao; Samira Al- Izadeh; Sarmad Shubber; Silas Wang; Sourav Roy; Sylvain Viguier; Thanh Le; Tobi Oyebade; Trieu Le; Yoyo Yang; Zach Nguyen; Ramesh Kashyap; Alfredo Palasciano; Alison Callahan; Anima Shukla; Antonio Miranda-Escalada; Ayush Singh; Benjamin Beilharz; Bo Wang; Caio Brito; Chenxi Zhou; Chirag Jain; Chuxin Xu; Clémentine Fourrier; Daniel León Periñán; Daniel Molano; Dian Yu; Enrique Manjava- Cas; Fabio Barth; Florian Fuhrimann; Gabriel Altay; Giyaseddin Bayrak; Gully Burns; Helena U Vrabec; Imane Bello; Ishani Dash; Jihyun Kang; John Giorgi; Jonas Golde; Jose David Posada; Ranga- Sai Karthik; Lokesh Sivaraman; Lu Bulchandani; Luisa Liu; Madeleine Shinzato; Maiko Hahn De Bykhovetz; Marc Takeuchi; Maria A Pàmies; Mari- Anna Castillo; Mario Nezhurina; Matthias Sänger; Michael Samwald; Michael Cullan; Michiel Weinberg; Mina De Wolf; Minna Mihaljcic; Moritz Liu; Myungsun Freidank; Natasha Kang; Nathan Seelam; Nicholas Dahlberg; Nikolaus Michio Broad; Pascale Muellner; Patrick Fung; Ramya Haller; Renata Chandrasekhar; Robert Eisenberg; Rodrigo Martin; Rosaline Canalli; Ruisi Su; Samuel Su; Samuele Cahyawijaya; Shlok S Garda; Shubhanshu Deshmukh; Sid Mishra; Simon Ki- Blawi; Sinee Ott; Srishti Sang-Aroonsiri; Stefan Ku- Mar; Sushil Schweter; Tanmay Bharati; Théo Laud; Tomoya Gigant; Wojciech Kainuma; Ya- Nis Kusa; Labrak; Shailesh Yash; Yash Bajaj; Yifan Venkatraman; Yingxin Xu; Yu Xu; Xu; Thomas Wolf 2023 Bloom | [
{
"authors": "Ebtesam Almazrouei; Hamza Alobeidli; Abdulaziz Alshamsi; Alessandro Cappelli; Ruxandra Cojocaru; Merouane Debbah; Etienne Goffinet; Daniel Heslow; Julien Launay; Quentin Malartic; Badreddine Noune; Baptiste Pannier; Guilherme Penedo",
"journal": "",
"ref_id": "b0",
"title": "Falcon-40B: an open large language model with state-of-the-art performance",
"year": "2023"
},
{
"authors": "Fernando Alva-Manchego; Louis Martin; Antoine Bordes; Carolina Scarton; Benoît Sagot; Lucia Specia",
"journal": "Association for Computational Linguistics",
"ref_id": "b1",
"title": "ASSET: A dataset for tuning and evaluation of sentence simplification models with multiple rewriting transformations",
"year": "2020"
},
{
"authors": "Reza Yazdani Aminabadi; Samyam Rajbhandari; Minjia Zhang; Ammar Ahmad Awan; Cheng Li; Du Li; Elton Zheng; Jeff Rasley; Shaden Smith; Olatunji Ruwase; Yuxiong He",
"journal": "",
"ref_id": "b2",
"title": "Deepspeed inference: Enabling efficient inference of transformer models at unprecedented scale",
"year": "2022"
},
{
"authors": "Simran Arora; Avanika Narayan; Laurel Mayee F Chen; Neel Orr; Kush Guha; Ines Bhatia; Christopher Chami; Re",
"journal": "",
"ref_id": "b3",
"title": "Ask me anything: A simple strategy for prompting language models",
"year": "2023"
},
{
"authors": "H Stephen; Victor Bach; Zheng-Xin Sanh; Albert Yong; Colin Webson; Raffel; V Nihal; Abheesht Nayak; Taewoon Sharma; M Kim; Thibault Saiful Bari; Fevry",
"journal": "",
"ref_id": "b4",
"title": "Promptsource: An integrated development environment and repository for natural language prompts",
"year": "2022"
},
{
"authors": "Mohaddeseh Bastan; Mahnaz Koupaee; Youngseo Son; Richard Sicoli; Niranjan Balasubramanian",
"journal": "",
"ref_id": "b5",
"title": "Author's sentiment prediction",
"year": "2020"
},
{
"authors": "Qiang Ning; Ben Zhou; Daniel Khashabi; Dan Roth",
"journal": "",
"ref_id": "b6",
"title": "going on a vacation\" takes longer than \"going for a walk\": A study of temporal commonsense understanding",
"year": "2019"
},
{
"authors": "Yonatan Bisk; Rowan Zellers; Le Ronan; Jianfeng Bras; Yejin Gao; Choi",
"journal": "",
"ref_id": "b7",
"title": "Piqa: Reasoning about physical commonsense in natural language",
"year": "2020"
},
{
"authors": "Ruben Ali Furkan Biten; Andres Tito; Lluis Mafla; Marçal Gomez; Ernest Rusinol; Valveny; Dimosthenis Jawahar; Karatzas",
"journal": "",
"ref_id": "b8",
"title": "Scene text visual question answering",
"year": "2019"
},
{
"authors": "R Samuel; Gabor Bowman; Christopher Angeli; Christopher D Potts; Manning",
"journal": "Association for Computational Linguistics",
"ref_id": "b9",
"title": "A large annotated corpus for learning natural language inference",
"year": "2015"
},
{
"authors": "Oana-Maria Camburu; Tim Rockt; \" Aschel; Thomas Lukasiewicz; Phil Blunsom",
"journal": "",
"ref_id": "b10",
"title": "e-snli: Natural language inference with natural language explanations",
"year": "2018"
},
{
"authors": "S Bengio; H Wallach; H Larochelle; K Grauman; N Cesa-Bianchi; R Garnett",
"journal": "",
"ref_id": "b11",
"title": "editors",
"year": ""
},
{
"authors": "Curran Associates; Inc ",
"journal": "",
"ref_id": "b12",
"title": "",
"year": ""
},
{
"authors": "Hyundong Cho; Jonathan May",
"journal": "",
"ref_id": "b13",
"title": "Grounding conversations with improvised dialogues",
"year": "2020"
},
{
"authors": "Aakanksha Chowdhery; Sharan Narang; Jacob Devlin; Maarten Bosma; Gaurav Mishra; Adam Roberts; Paul Barham; Hyung Won Chung; Charles Sutton; Sebastian Gehrmann; Parker Schuh; Kensen Shi; Sasha Tsvyashchenko; Joshua Maynez; Abhishek Rao; Parker Barnes; Yi Tay; Noam Shazeer; Emily Vinodkumar Prabhakaran; Nan Reif; Ben Du; Reiner Hutchinson; James Pope; Jacob Bradbury; Michael Austin; Guy Isard; Pengcheng Gur-Ari; Toju Yin; Anselm Duke; Sanjay Levskaya; Sunipa Ghemawat; Henryk Dev; Xavier Michalewski; Vedant Garcia; Kevin Misra; Liam Robinson; Denny Fedus; Daphne Zhou; David Ippolito; Hyeontaek Luan; Barret Lim; Alexander Zoph; Ryan Spiridonov; David Sepassi; Shivani Dohan; Mark Agrawal; Andrew M Omernick; Thanumalayan Dai; Marie Sankaranarayana Pillai; Aitor Pellat; Erica Lewkowycz; Rewon Moreira; Oleksandr Child; Katherine Polozov; Zongwei Lee; Xuezhi Zhou; Brennan Wang; Mark Saeta; Orhan Diaz; Michele Firat; Jason Catasta; Kathy Wei; Douglas Meier-Hellstern; Jeff Eck; Slav Dean; Noah Petrov; Fiedel",
"journal": "",
"ref_id": "b14",
"title": "Palm: Scaling language modeling with pathways",
"year": "2022"
},
{
"authors": "Chung Hyung Won; Le Hou; Shayne Longpre; Barret Zoph; Yi Tay; William Fedus; Eric Li; Xuezhi Wang; Mostafa Dehghani; Siddhartha Brahma",
"journal": "",
"ref_id": "b15",
"title": "Scaling instruction-finetuned language models",
"year": "2022"
},
{
"authors": "Daniel Borkan; Jeffrey Sorensen; Lucas Dixon; Lucy Vasserman; Nithum",
"journal": "",
"ref_id": "b16",
"title": "Jigsaw unintended bias in toxicity classification",
"year": "2019"
},
{
"authors": "Christopher Clark; Kenton Lee; Ming-Wei Chang; Tom Kwiatkowski; Michael Collins; Kristina Toutanova",
"journal": "Association for Computational Linguistics",
"ref_id": "b17",
"title": "BoolQ: Exploring the surprising difficulty of natural yes/no questions",
"year": "2019"
},
{
"authors": "Alexis Conneau; Douwe Kiela",
"journal": "European Language Resources Association (ELRA",
"ref_id": "b18",
"title": "SentEval: An evaluation toolkit for universal sentence representations",
"year": "2018"
},
{
"authors": "Pradeep Dasigi; Nelson F Liu; Ana Marasović; Noah A Smith; Matt Gardner",
"journal": "Association for Computational Linguistics",
"ref_id": "b19",
"title": "Quoref: A reading comprehension dataset with questions requiring coreferential reasoning",
"year": "2019"
},
{
"authors": "Thomas Davidson; Dana Warmsley; Michael Macy; Ingmar Weber",
"journal": "",
"ref_id": "b20",
"title": "Automated hate speech detection and the problem of offensive language",
"year": "2017"
},
{
"authors": "Dheeru Dua; Yizhong Wang; Pradeep Dasigi; Gabriel Stanovsky; Sameer Singh; Matt Gardner",
"journal": "",
"ref_id": "b21",
"title": "Drop: A reading comprehension benchmark requiring discrete reasoning over paragraphs",
"year": "2019"
},
{
"authors": "Esin Durmus; Claire Cardie",
"journal": "Association for Computational Linguistics",
"ref_id": "b22",
"title": "A corpus for modeling user and language effects in argumentation on online debating",
"year": "2019"
},
{
"authors": "Yanai Elazar; Yoav Goldberg",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b23",
"title": "Where's my head? Definition, data set, and models for numeric fused-head identification and resolution",
"year": "2019"
},
{
"authors": "",
"journal": "",
"ref_id": "b24",
"title": "Amazon polarity dataset",
"year": ""
},
{
"authors": "Hao Fu; ; Yao; Tushar Peng; Khot",
"journal": "Yao Fu's Notion",
"ref_id": "b25",
"title": "How does gpt obtain its ability? tracing emergent abilities of language models to their sources",
"year": "2022"
},
{
"authors": "Leo Gao; Stella Biderman; Sid Black; Laurence Golding; Travis Hoppe; Charles Foster; Jason Phang; Horace He; Anish Thite; Noa Nabeshima; Shawn Presser; Connor Leahy",
"journal": "",
"ref_id": "b26",
"title": "The pile: An 800gb dataset of diverse text for language modeling",
"year": "2020"
},
{
"authors": "Mor Geva; Daniel Khashabi; Elad Segal; Tushar Khot; Dan Roth; Jonathan Berant",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b27",
"title": "Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies",
"year": "2021"
},
{
"authors": "Alec Go; Richa Bhayani; Lei Huang",
"journal": "CS224N project report",
"ref_id": "b28",
"title": "Twitter sentiment classification using distant supervision",
"year": "2009"
},
{
"authors": "David Graff; Junbo Kong; Ke Chen; Kazuaki Maeda",
"journal": "",
"ref_id": "b29",
"title": "English gigaword. Linguistic Data Consortium",
"year": "2003"
},
{
"authors": "Matthew Henderson; Blaise Thomson; Jason D Williams",
"journal": "IEEE",
"ref_id": "b30",
"title": "The third dialog state tracking challenge",
"year": "2014"
},
{
"authors": "Dan Hendrycks; Collin Burns; Steven Basart; Andy Zou; Mantas Mazeika; Dawn Song; Jacob Steinhardt",
"journal": "",
"ref_id": "b31",
"title": "Measuring massive multitask language understanding",
"year": "2021"
},
{
"authors": "Nabil Hossain; John Krumm; Michael Gamon; Henry Kautz",
"journal": "",
"ref_id": "b32",
"title": "SemEval-2020 task 7: Assessing humor in edited news headlines",
"year": "2020"
},
{
"authors": "Lifu Huang; Le Ronan; Chandra Bras; Yejin Bhagavatula; Choi",
"journal": "Association for Computational Linguistics",
"ref_id": "b33",
"title": "Cosmos QA: Machine reading comprehension with contextual commonsense reasoning",
"year": "2019"
},
{
"authors": "Ting-Hao Kenneth Huang; Chieh-Yang Huang; Chien-Kuang ; Cornelia Ding; Yen-Chia Hsu; C Lee Giles",
"journal": "Association for Computational Linguistics",
"ref_id": "b34",
"title": "CODA-19: Using a non-expert crowd to annotate research aspects on 10,000+ abstracts in the COVID-19 open research dataset",
"year": "2020"
},
{
"authors": "Kelvin Jiang; Dekun Wu; Hui Jiang",
"journal": "Association for Computational Linguistics",
"ref_id": "b35",
"title": "Free-baseQA: A new factoid QA data set matching triviastyle question-answer pairs with Freebase",
"year": "2019"
},
{
"authors": "Di Jin; Peter Szolovits",
"journal": "Association for Computational Linguistics",
"ref_id": "b36",
"title": "PICO element detection in medical text via long short-term memory neural networks",
"year": "2018"
},
{
"authors": "Mandar Joshi; Eunsol Choi; Daniel Weld; Luke Zettlemoyer",
"journal": "Association for Computational Linguistics",
"ref_id": "b37",
"title": "TriviaQA: A large scale distantly supervised challenge dataset for reading comprehension",
"year": "2017"
},
{
"authors": "Phillip Keung; Yichao Lu; György Szarvas; Noah A Smith",
"journal": "Association for Computational Linguistics",
"ref_id": "b38",
"title": "The multilingual Amazon reviews corpus",
"year": "2020"
},
{
"authors": "Daniel Khashabi; Snigdha Chaturvedi; Michael Roth; Shyam Upadhyay; Dan Roth",
"journal": "",
"ref_id": "b39",
"title": "Looking beyond the surface: A challenge set for reading comprehension over multiple sentences",
"year": "2018"
},
{
"authors": "Daniel Khashabi; Tushar Khot; Ashish Sabharwal; Dan Roth",
"journal": "Association for Computational Linguistics",
"ref_id": "b40",
"title": "Learning what is essential in questions",
"year": "2017"
},
{
"authors": "Tushar Khot; Peter Clark; Michal Guerquin; Peter Jansen; Ashish Sabharwal",
"journal": "",
"ref_id": "b41",
"title": "Qasc: A dataset for question answering via sentence composition",
"year": "2020"
},
{
"authors": "Takeshi Kojima; Shane Shixiang; Machel Gu; Yutaka Reid; Yusuke Matsuo; Iwasawa",
"journal": "",
"ref_id": "b42",
"title": "Large language models are zero-shot reasoners",
"year": "2022"
},
{
"authors": "Tom Kwiatkowski; Jennimaria Palomaki; Olivia Redfield; Michael Collins; Ankur Parikh; Chris Alberti; Danielle Epstein; Illia Polosukhin; Matthew Kelcey; Jacob Devlin; Kenton Lee; Kristina N Toutanova; Llion Jones; Ming-Wei Chang; Andrew Dai; Jakob Uszkoreit; Quoc Le; Slav Petrov",
"journal": "Transactions of the Association of Computational Linguistics",
"ref_id": "b43",
"title": "Natural questions: a benchmark for question answering research",
"year": "2019"
},
{
"authors": "Lucile Hugo Laurençon; Thomas Saulnier; Christopher Wang; Albert Akiki; Teven Villanova Del Moral; Leandro Le Scao; Chenghao Von Werra; Eduardo González Mou; Huu Ponferrada; Jörg Nguyen; Mario Frohberg; Quentin Šaško; Angelina Lhoest; Gerard Mcmillan-Major; Stella Dupont; Anna Biderman; Loubna Rogers; Francesco De Ben Allal; Giada Toni; Olivier Pistilli; Somaieh Nguyen; Maraim Nikpoor; Pierre Masoud; Javier Colombo; Paulo De La Rosa; Tristan Villegas; Shayne Thrush; Sebastian Longpre; Leon Nagel; Manuel Weber; Jian Muñoz; Daniel Zhu; Zaid Van Strien; Khalid Alyafeai; Minh Chien Almubarak; Itziar Vu; Aitor Gonzalez-Dios; Kyle Soroa; Manan Lo; Pedro Ortiz Dey; Aaron Suarez; Shamik Gokaslan; David Bose; Long Adelani; Hieu Phan; Ian Tran; Suhas Yu; Jenny Pai; Violette Chim; Suzana Lepercq; Margaret Ilic; Sasha Alexandra Mitchell; Yacine Luccioni; Jernite",
"journal": "",
"ref_id": "b44",
"title": "The bigscience roots corpus: A 1.6tb composite multilingual dataset",
"year": "2022"
},
{
"authors": "Curran Associates; Inc ",
"journal": "",
"ref_id": "b45",
"title": "",
"year": ""
},
{
"authors": "Omer Levy; Minjoon Seo; Eunsol Choi; Luke Zettlemoyer",
"journal": "Association for Computational Linguistics",
"ref_id": "b46",
"title": "Zero-shot relation extraction via reading comprehension",
"year": "2017"
},
{
"authors": "Xin Li; Dan Roth",
"journal": "",
"ref_id": "b47",
"title": "Learning question classifiers",
"year": "2002"
},
{
"authors": "Yanran Li; Hui Su; Xiaoyu Shen; Wenjie Li; Ziqiang Cao; Shuzi Niu",
"journal": "Asian Federation of Natural Language Processing",
"ref_id": "b48",
"title": "DailyDialog: A manually labelled multi-turn dialogue dataset",
"year": "2017"
},
{
"authors": "Kevin Lin; Oyvind Tafjord; Peter Clark; Matt Gardner",
"journal": "",
"ref_id": "b49",
"title": "Reasoning over paragraph effects in situations",
"year": "2019"
},
{
"authors": "Kevin Lin; Oyvind Tafjord; Peter Clark; Matt Gardner",
"journal": "Association for Computational Linguistics",
"ref_id": "b50",
"title": "Reasoning over paragraph effects in situations",
"year": "2019"
},
{
"authors": "Pengfei Liu; Weizhe Yuan; Jinlan Fu; Zhengbao Jiang; Hiroaki Hayashi; Graham Neubig",
"journal": "ACM Comput. Surv",
"ref_id": "b51",
"title": "Pretrain, prompt, and predict: A systematic survey of prompting methods in natural language processing",
"year": "2023"
},
{
"authors": "Shayne Longpre; Le Hou; Tu Vu; Albert Webson; Hyung Won Chung; Yi Tay; Denny Zhou; V Quoc; Barret Le; Jason Zoph; Wei",
"journal": "",
"ref_id": "b52",
"title": "The flan collection: Designing data and methods for effective instruction tuning",
"year": "2023"
},
{
"authors": "Nicholas Lourie; Ronan Le Bras; Yejin Choi",
"journal": "",
"ref_id": "b53",
"title": "Scruples: A corpus of community ethical judgments on 32,000 real-life anecdotes",
"year": "2021"
},
{
"authors": "Bill Maccartney; Christopher D Manning",
"journal": "",
"ref_id": "b54",
"title": "Natural logic for textual inference",
"year": "2007"
},
{
"authors": "Aman Madaan; Shuyan Zhou; Uri Alon; Yiming Yang; Graham Neubig",
"journal": "",
"ref_id": "b55",
"title": "Language models of code are few-shot commonsense learners",
"year": "2022-03"
},
{
"authors": "Bryan Mccann; Nitish Shirish Keskar; Caiming Xiong; Richard Socher",
"journal": "",
"ref_id": "b56",
"title": "The natural language decathlon: Multitask learning as question answering",
"year": "2019"
},
{
"authors": "Sewon Min; Julian Michael; Hannaneh Hajishirzi; Luke Zettlemoyer",
"journal": "Association for Computational Linguistics",
"ref_id": "b57",
"title": "AmbigQA: Answering ambiguous open-domain questions",
"year": "2020"
},
{
"authors": "Swaroop Mishra; Daniel Khashabi; Chitta Baral; Hannaneh Hajishirzi",
"journal": "",
"ref_id": "b58",
"title": "Cross-task generalization via natural language crowdsourcing instructions",
"year": "2021"
},
{
"authors": "Amita Misra; Brian Ecker; Marilyn Walker",
"journal": "Association for Computational Linguistics",
"ref_id": "b59",
"title": "Measuring the similarity of sentential arguments in dialogue",
"year": "2016"
},
{
"authors": "Nasrin Mostafazadeh; Nathanael Chambers; Xiaodong He; Devi Parikh; Dhruv Batra; Lucy Vanderwende; Pushmeet Kohli; James Allen",
"journal": "Association for Computational Linguistics",
"ref_id": "b60",
"title": "A corpus and cloze evaluation for deeper understanding of commonsense stories",
"year": "2016"
},
{
"authors": "Nasrin Mostafazadeh; Michael Roth; Annie Louis; Nathanael Chambers; James Allen",
"journal": "",
"ref_id": "b61",
"title": "Lsdsem 2017 shared task: The story cloze test",
"year": "2017"
},
{
"authors": "Yixin Nie; Adina Williams; Emily Dinan; Mohit Bansal; Jason Weston; Douwe Kiela",
"journal": "",
"ref_id": "b62",
"title": "Adversarial nli: A new benchmark for natural language understanding",
"year": "2020"
},
{
"authors": "Erik Nijkamp; Bo Pang; Hiroaki Hayashi; Lifu Tu; Huan Wang; Yingbo Zhou; Silvio Savarese; Caiming Xiong",
"journal": "OpenAI",
"ref_id": "b63",
"title": "Codegen: An open large language model for code with multi-turn program synthesis",
"year": "2023"
},
{
"authors": "Simon Ostermann; Ashutosh Modi; Michael Roth; Stefan Thater; Manfred Pinkal",
"journal": "European Language Resources Association (ELRA",
"ref_id": "b64",
"title": "MCScript: A novel dataset for assessing machine comprehension using script knowledge",
"year": "2018"
},
{
"authors": "Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu",
"journal": "The Journal of Machine Learning Research",
"ref_id": "b65",
"title": "Exploring the limits of transfer learning with a unified text-to-text transformer",
"year": "2020"
},
{
"authors": "Alessandro Raganato; Tommaso Pasini; Jose Camacho-Collados; Mohammad Taher; Pilehvar ",
"journal": "Association for Computational Linguistics",
"ref_id": "b66",
"title": "XL-WiC: A multilingual benchmark for evaluating semantic contextualization",
"year": "2020"
},
{
"authors": "Pranav Rajpurkar; Robin Jia; Percy Liang",
"journal": "Association for Computational Linguistics",
"ref_id": "b67",
"title": "Know what you don't know: Unanswerable questions for SQuAD",
"year": "2018"
},
{
"authors": "Pranav Rajpurkar; Jian Zhang; Konstantin Lopyrev; Percy Liang",
"journal": "Association for Computational Linguistics",
"ref_id": "b68",
"title": "SQuAD: 100,000+ questions for machine comprehension of text",
"year": "2016"
},
{
"authors": "Siva Reddy; Danqi Chen; Christopher D Manning",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b69",
"title": "CoQA: A conversational question answering challenge",
"year": "2019"
},
{
"authors": "Laria Reynolds; Kyle Mcdonell",
"journal": "Association for Computing Machinery",
"ref_id": "b70",
"title": "Prompt programming for large language models: Beyond the few-shot paradigm",
"year": "2021"
},
{
"authors": "Rachel Rudinger; Vered Shwartz; Jena D Hwang; Chandra Bhagavatula; Maxwell Forbes; Le Ronan; Noah A Bras; Yejin Smith; Choi",
"journal": "Association for Computational Linguistics",
"ref_id": "b71",
"title": "Thinking like a skeptic: Defeasible inference in natural language",
"year": "2020"
},
{
"authors": "Keisuke Sakaguchi; Le Ronan; Chandra Bras; Yejin Bhagavatula; Choi",
"journal": "Communications of the ACM",
"ref_id": "b72",
"title": "Winogrande: An adversarial winograd schema challenge at scale",
"year": "2021"
},
{
"authors": "Maarten Sap; Le Ronan; Emily Bras; Chandra Allaway; Nicholas Bhagavatula; Hannah Lourie; Brendan Rashkin; Noah A Roof; Yejin Smith; Choi",
"journal": "",
"ref_id": "b73",
"title": "Atomic: An atlas of machine commonsense for ifthen reasoning",
"year": "2019"
},
{
"authors": "Le Teven; Angela Scao; Christopher Fan; Ellie Akiki; Suzana Pavlick; Daniel Ilić; Roman Hesslow; Alexandra Castagné; François Sasha Luccioni; Matthias Yvon; Jonathan Gallé; Alexander M Tow; Stella Rush; Albert Biderman; Pawan Webson; Thomas Sasanka Ammanamanchi; Benoît Wang; Niklas Sagot; Albert Muennighoff; Olatunji Villanova Del Moral; Rachel Ruwase; Stas Bawden; Angelina Bekman; Iz Mcmillan-Major; Huu Beltagy; Lucile Nguyen; Samson Saulnier; Pedro Ortiz Tan; Victor Suarez; Hugo Sanh; Yacine Laurençon; Julien Jernite; Margaret Launay; Colin Mitchell; Aaron Raffel; Adi Gokaslan; Aitor Simhi; Alham Soroa; Amit Fikri Aji; Anna Alfassy; Ariel Kreisberg Rogers; Canwen Nitzav; Chenghao Xu; Chris Mou; Christopher Emezue; Colin Klamm; Leong; David Daniel Van Strien; Dragomir Ifeoluwa Adelani; Eduardo González Radev; Efrat Ponferrada; Ethan Levkovizh; Emily Sheng; David Uthus",
"journal": "Association for Computational Linguistics",
"ref_id": "b74",
"title": "Investigating societal biases in a poetry composition system",
"year": "2020"
},
{
"authors": "Richard Socher; Alex Perelygin; Jean Wu; Jason Chuang; Christopher D Manning; Andrew Ng; Christopher Potts",
"journal": "Association for Computational Linguistics",
"ref_id": "b75",
"title": "Recursive deep models for semantic compositionality over a sentiment treebank",
"year": "2013"
},
{
"authors": "Gabriel Stanovsky; Mark Hopkins",
"journal": "Association for Computational Linguistics",
"ref_id": "b76",
"title": "Spot the odd man out: Exploring the associative power of lexical resources",
"year": "2018"
},
{
"authors": "Kai Sun; Dian Yu; Jianshu Chen; Dong Yu; Yejin Choi; Claire Cardie",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b77",
"title": "DREAM: A challenge data set and models for dialogue-based reading comprehension",
"year": "2019"
},
{
"authors": "Oyvind Tafjord; Peter Clark; Matt Gardner; Wen Tau Yih; Ashish Sabharwal",
"journal": "",
"ref_id": "b78",
"title": "Quarel: A dataset and models for answering questions about qualitative relationships",
"year": "2018"
},
{
"authors": "Jörg Tiedemann",
"journal": "European Language Resources Association (ELRA)",
"ref_id": "b79",
"title": "Parallel data, tools and interfaces in opus",
"year": "2012"
},
{
"authors": "Cynthia Van Hee; Els Lefever; Véronique Hoste",
"journal": "Association for Computational Linguistics",
"ref_id": "b80",
"title": "SemEval-2018 task 3: Irony detection in English tweets",
"year": "2018"
},
{
"authors": "Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Łukasz Kaiser; Illia Polosukhin",
"journal": "Advances in neural information processing systems",
"ref_id": "b81",
"title": "Attention is all you need",
"year": "2017"
},
{
"authors": "Bertie Vidgen; Dong Nguyen; Helen Margetts; Patricia Rossini; Rebekah Tromble",
"journal": "Association for Computational Linguistics",
"ref_id": "b82",
"title": "Introducing CAD: the contextual abuse dataset",
"year": "2021"
},
{
"authors": "Yizhong Wang; Yeganeh Kordi; Swaroop Mishra; Alisa Liu; Noah A Smith; Daniel Khashabi; Hannaneh Hajishirzi",
"journal": "",
"ref_id": "b83",
"title": "Self-instruct: Aligning language model with self generated instructions",
"year": "2022"
},
{
"authors": "Yizhong Wang; Swaroop Mishra; Pegah Alipoormolabashi; Yeganeh Kordi; Amirreza Mirzaei; Atharva Naik; Arjun Ashok; Arut Selvan Dhanasekaran; Anjana Arunkumar; David Stap; Eshaan Pathak; Giannis Karamanolakis; Haizhi Lai; Ishan Purohit; Ishani Mondal; Jacob Anderson; Kirby Kuznia; Krima Doshi; Kuntal Kumar Pal; Maitreya Patel; Mehrad Moradshahi; Mihir Parmar; Mirali Purohit; Neeraj Varshney; Rohitha Phani; Pulkit Kaza; Ravsehaj Verma; Rushang Singh Puri; Savan Karia; Doshi; Keyur Shailaja; Siddhartha Sampat; Sujan Mishra; A Reddy; Sumanta Patro; Tanay Dixit; Xudong Shen",
"journal": "Association for Computational Linguistics",
"ref_id": "b84",
"title": "Super-NaturalInstructions: Generalization via declarative instructions on 1600+ NLP tasks",
"year": "2022"
},
{
"authors": "Alex Warstadt; Amanpreet Singh; Samuel R Bowman",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b85",
"title": "Neural network acceptability judgments",
"year": "2019"
},
{
"authors": "Kellie Webster; Marta Recasens; Vera Axelrod; Jason Baldridge",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b86",
"title": "Mind the GAP: A balanced corpus of gendered ambiguous pronouns",
"year": "2018"
},
{
"authors": "Jason Wei; Maarten Bosma; Vincent Zhao; Kelvin Guu; Adams Wei Yu; Brian Lester; Nan Du; Andrew M Dai; Quoc V Le",
"journal": "",
"ref_id": "b87",
"title": "a. Finetuned language models are zero-shot learners",
"year": "2022"
},
{
"authors": "Jason Wei; Maarten Bosma; Y Vincent; Kelvin Zhao; Adams Wei Guu; Brian Yu; Nan Lester; Andrew M Du; Quoc V Dai; Le",
"journal": "",
"ref_id": "b88",
"title": "Finetuned language models are zero-shot learners",
"year": "2021"
},
{
"authors": "Jason Wei; Xuezhi Wang; Dale Schuurmans; Maarten Bosma; Fei Xia; Ed H Chi; Denny Quoc V Le; Zhou",
"journal": "",
"ref_id": "b89",
"title": "Chain of thought prompting elicits reasoning in large language models",
"year": "2022"
},
{
"authors": "Orion Weller; Nicholas Lourie; Matt Gardner; Matthew E Peters",
"journal": "Association for Computational Linguistics",
"ref_id": "b90",
"title": "Learning from task descriptions",
"year": "2020"
},
{
"authors": "John Wieting; Kevin Gimpel",
"journal": "",
"ref_id": "b91",
"title": "ParaNMT-50M: Pushing the limits of paraphrastic sentence embeddings with millions of machine translations",
"year": "2018"
},
{
"authors": "Adina Williams; Nikita Nangia; Samuel Bowman",
"journal": "Association for Computational Linguistics",
"ref_id": "b92",
"title": "A broad-coverage challenge corpus for sentence understanding through inference",
"year": "2018"
},
{
"authors": "Thomas Wolf; Lysandre Debut; Victor Sanh; Julien Chaumond; Clement Delangue; Anthony Moi; Pierric Cistac; Tim Rault; Remi Louf; Morgan Funtowicz; Joe Davison; Sam Shleifer; Clara Patrick Von Platen; Yacine Ma; Julien Jernite; Canwen Plu; Teven Xu; Sylvain Le Scao; Mariama Gugger; Quentin Drame; Alexander Lhoest; Rush",
"journal": "Association for Computational Linguistics",
"ref_id": "b93",
"title": "Transformers: State-of-the-art natural language processing",
"year": "2020"
},
{
"authors": "Tomer Wolfson; Mor Geva; Ankit Gupta; Matt Gardner; Yoav Goldberg; Daniel Deutch; Jonathan Berant",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b94",
"title": "Break it down: A question understanding benchmark",
"year": "2020"
},
{
"authors": "Wenhan Xiong; Jiawei Wu; Hong Wang; Vivek Kulkarni; Mo Yu; Shiyu Chang; Xiaoxiao Guo; William Yang; Wang ",
"journal": "Association for Computational Linguistics",
"ref_id": "b95",
"title": "TWEETQA: A social media focused question answering dataset",
"year": "2019"
},
{
"authors": "Pengcheng Yin; Bowen Deng; Edgar Chen; Bogdan Vasilescu; Graham Neubig",
"journal": "ACM",
"ref_id": "b96",
"title": "Learning to mine aligned code and natural language pairs from stack overflow",
"year": "2018"
},
{
"authors": "J D Zamfirescu-Pereira; Richmond Y Wong; Bjoern Hartmann; Qian Yang",
"journal": "Association for Computing Machinery",
"ref_id": "b97",
"title": "Why johnny can't prompt: How non-ai experts try (and fail) to design llm prompts",
"year": "2023"
},
{
"authors": "Li Zhang; Liam Dugan; Hainiu Xu; Chris Callison-Burch",
"journal": "",
"ref_id": "b98",
"title": "Exploring the curious case of code prompts",
"year": "2023"
},
{
"authors": "Li Zhang; Hainiu Xu; Yue Yang; Shuyan Zhou; Weiqiu You; Manni Arora; Chris Callison-Burch",
"journal": "",
"ref_id": "b99",
"title": "Causal reasoning of entities and events in procedural texts",
"year": "2023"
},
{
"authors": "Sheng Zhang; Xiaodong Liu; Jingjing Liu; Jianfeng Gao; Kevin Duh; Benjamin Van Durme",
"journal": "",
"ref_id": "b100",
"title": "Record: Bridging the gap between human and machine commonsense reading comprehension",
"year": "2018"
},
{
"authors": "Yuan Zhang; Jason Baldridge; Luheng He",
"journal": "",
"ref_id": "b101",
"title": "PAWS: Paraphrase Adversaries from Word Scrambling",
"year": "2019"
},
{
"authors": "Zihao Zhao; Eric Wallace; Shi Feng; Dan Klein; Sameer Singh",
"journal": "",
"ref_id": "b102",
"title": "Calibrate before use: Improving few-shot performance of language models",
"year": "2021"
},
{
"authors": "Peixiang Zhong; Chen Zhang; Hao Wang; Yong Liu; Chunyan Miao",
"journal": "",
"ref_id": "b103",
"title": "Towards persona-based empathetic conversational models",
"year": "2020"
}
] | [] |
10.18653/v1/2020.emnlp-main.95 | 2023-05-19 | [
{
"figure_ref": [
"fig_0",
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b18",
"b5",
"b8",
"b4",
"b19",
"b23",
"b20",
"b12",
"b0",
"b24",
"b1",
"b20",
"b14",
"b9",
"b15",
"b17",
"b6",
"b22",
"b6"
],
"table_ref": [],
"text": "Named entity recognition (NER) (Tjong Kim Sang and De Meulder, 2003;Doddington et al., 2004) has been a long-standing and one of the most important fundamental tasks in natural language processing (NLP). Existing NER models can be divided into three different categories, including sequence labeling methods (Lample et al., 2016;Devlin et al., 2019), span-level classification (Wang and Lu, 2020;Zhong and Chen, 2021) and generationbased methods (Yan et al., 2021;Lu et al., 2022). However, even with pre-trained language models (PLMs), training these state-of-the-art named entity recognizers requires sufficient training samples, which is in contrast with real-world scenarios, where only small amounts of labeled data are available. This draws our attention to the challenging but practical problem: few-shot NER.\nBy introducing more samples in the training stage, data augmentation (DA) methods have been proven to be effective solutions in low-resource settings, including few-shot NER (Chen et al., 2020;Zhou et al., 2022;Chen et al., 2021). However, these approaches are mostly designed in a tokenlevel classification style. So they must be combined with different tagging schemes or special-designed structures before they can be applied to other NER subtasks, such as nested NER. On the other hand, generation-based models (Yan et al., 2021;Paolini et al., 2021) can overcome this limitation by leveraging generative PLMs and introducing a unified tagging strategy. However, few efforts are made on DA over generative PLMs (e.g. BART (Lewis et al., 2020), T5 (Raffel et al., 2019)), so limited data resources will lead to weakly fine-tuning of these generation-based methods. Hence, developing a DA approach that can easily be compatible with these generative PLMs would be worthwhile.\nDue to the autoregressive decoding of generative PLMs, vanilla generation-based NER methods have a strong assumption that the entities will appear in the target sequence with the same left-toright order as the source sequence. For example, in Figure 1, there are 4 entities in the input sentence. The prediction of entity EU-wide will be strictly after that of entity Fischler, following their order of appearing in the input sentence. However, in the NER task, the output entities are essentially forming an unordered set. To mitigate the abovementioned challenge, significant efforts have been made from different aspects. Tan et al. (2021) proposed a sequence-to-set network and relied mainly on non-autoregressive generation (Gu et al., 2018). Although they were able to predict the entities as a set, they might suffer from uncertain boundaries, and non-autoregressive generation may also lead to tremendous search space. Zhang et al. (2022) tried to address this issue still in an autoregressive perspective and constructed augmented samples based on the entities' context and positional orders. However, a simple mixture of these target sequences can confuse the model since there will be several \"gold\" target sequences corresponding to a same source sequence, which will result in a one-to-many mapping problem (also known as multimodality problem (Gu et al., 2018)), especially harmful in few-shot NER settings.\nIn this work, we try to fully utilize the orderagnostic property of NER, and propose a simple but effective Prompt Ordering based Data Augmentation (PODA) method for few-shot NER. In our view, any sequence containing complete information (i.e. every entity's mention and its type) should be regarded reasonable and can serve as an augmented target sequence. With the help of different prompt-based order instructions, we separate the original one-to-many mapping into various oneto-one mappings. As shown in Figure 1, following a certain entity type permutation like \"PER, MISC, LOC, ORG\", a unique target sequence will then be constructed, and its related source sequence will be the combination of source sentence and the order instruction. In this way, the strict left-to-right order does not need to be maintained.\nIn summary, our contributions include: (1) We for the first time propose a new data augmenta-tion method which can be uniformly applied over several generative PLMs. Furthermore, we combine our augmented data with prompt-based order instructions to prevent one-to-many mapping problem; (2) Experiments over three benchmark NER datasets, including flat and nested NER, demonstrate the effectiveness of our data augmentation method. Further analyses show the strong generalization ability of our method and validity of our augmented data."
},
{
"figure_ref": [],
"heading": "Prompt Ordering based Data Augmentation",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we firstly introduce the formulation of NER task in a generation style, and then we describe how we construct augmented data without any information from the input sentence. After that, we will illustrate the details of our prompt-based order instructions."
},
{
"figure_ref": [],
"heading": "Formulation",
"publication_ref": [],
"table_ref": [],
"text": "The NER tasks aim at detecting all the spans that can represent entities within a given sentence\nX = [x 1 , x 2 , ..., x n ],\nwhere n is the sentence length. The entities in sentence X form the corresponding target sequence Y . The i-th entity y i ∈ Y can be represented as a tuple y i = (s i , t i ), where s i , t i represent the entity span and type of y i , respectively. s will have different formats according to the tagging schemes. In conventional generation-based methods for NER, the target sequence Y = [(s 1 , t 1 ), (s 2 , t 2 ), ..., (s m , t m )] with m entities will then be uniquely determined, following the strict left-to-right order as in the source sequence. Then the generation procedure can be formulated as the following equation:\nP (Y |X) = |Y | i=1 P (y i |X, Y <i ) (1)"
},
{
"figure_ref": [],
"heading": "Augment Data via Re-ordering",
"publication_ref": [
"b22"
],
"table_ref": [],
"text": "A straightforward basis for re-ordering is position.\nThen augmented samples can be constructed by randomly shuffling (Zhang et al., 2022). However, there may not be a clear distinguishing criteria for these augmented entity sequences to further solve the one-to-many problem. So in our work, we alternatively choose the entity types as the principal factor of re-ordering. T = {t 1 , t 2 , ..., t l } is the entity type set with cardinality l for a certain dataset. We use p to denote a random permutation of elements in T , such as [t 3 , t 1 , t 2 , ...]. Following the specific entity type order p, the original target sequence Y in the left-to-right order can be re-ordered into Y p as:\nY p = [..., [(E p i ,1 , p i ), ..., (E p i ,np i , p i )], ...] (2)\nwhere E p i ,j represent the j-th entity span with type p i , following the original order, and n p i indicates the number of entities with type p i . Thus, in the sequence Y p , there will be l tuples like [(E p i ,1 , p i ), ..., (E p i ,np i , p i )], and each represents a set of entities with a same type.\nGiven an entity type set T with cardinality l, l! permutations like p can be easily obtained. We denote them as a set Perm(T ). For each p ∈ Perm(T ), we can get a unique re-ordered Y p .\nAs an example, the original entity sequence is Y = \"[(EU, MISC), (Britain, LOC), (BSE, MISC)]\". If the order p is given as \"PER, LOC, MISC, ORG\", Y will then be gathered as\nY p = \"[[(Britain, LOC)], [(EU, MISC), (BSE, MISC)]]\" 2 ."
},
{
"figure_ref": [
"fig_0"
],
"heading": "Prompt-based Order Instructions",
"publication_ref": [],
"table_ref": [],
"text": "To fully use the augmented target sequences and prevent the one-to-many mapping problem, we separate these sequences by different p. As depicted in Figure 1, we construct prompt-based order instructions as \"Following the order: p\" for augmented entity sequences. These prompts indicate to the model which entity type to focus on at a certain generation step. As a result, the predictions among different entity types are naturally modeled and the target sequences are uniquely determined with respect to the instructions. In this way, we can enlarge our training samples by |Perm(T )| times."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [
"b18",
"b10",
"b5",
"b11",
"b2",
"b13"
],
"table_ref": [],
"text": "To evaluate the effectiveness of PODA over generation-based methods, we conduct several experiments over two flat NER datasets and one nested NER dataset in several few-shot settings.\nDatasets For flat NER datasets, we choose CoNLL-2003 (Tjong Kim Sang andDe Meulder, 2003) and MIT-Movie (Liu et al., 2013) from two different domains. As for nested NER subtask, we conduct experiments on ACE-2005 (Doddington et al., 2004). We randomly select 15% samples from the MIT-Movie training set as the development set. For the experiments on ACE-2005, we use the same data split as (Lu and Roth, 2015).\nExperimental Settings To show that our approach can be generally applied to different generation-based methods, we use pure T5-base and Flan-T5-base (Chung et al., 2022) as our main backbones. And we utilize BART-base when conducting experiments over BART-NER. We run T5 and Flan-T5 for 40 epochs and BART-NER for 200 epochs. To keep the stability of the few-shot setting, we set the batch size to 2/2/4/8 for the 5/10/20/50 shot settings, respectively. The learning rate of the Adam optimizer is set to 2e-5/5e-5.\nIn this work, we follow (Ma et al., 2022) and focus on few-shot settings that only K samples of each entity type from the training set are provided for training on a certain dataset. We conduct experiments in K = {5, 10, 20, 50} settings and report the mean and deviation performance over three splits. In this work, we adopt the same sampling strategy as Yang and Katiyar (2020) 3 . For CoNLL-2003, we use all the permutations of the entity type set as there are only 4 types. For datasets with more entity types such as ACE-2005, we randomly choose 20 different order instructions."
},
{
"figure_ref": [],
"heading": "Main Results",
"publication_ref": [
"b7"
],
"table_ref": [],
"text": "In the experiments, we compare our method with several strong baselines and competitive few-shot approaches.\n3 As different sampling may affect the performance, we also conduct experiments on the splits released in Huang et al. (2021) and the reported results can be referred in the appendix."
},
{
"figure_ref": [],
"heading": "K-shot Models",
"publication_ref": [
"b3",
"b13",
"b21",
"b3"
],
"table_ref": [
"tab_1"
],
"text": "CoNLL NNShot and StructShot (Yang and Katiyar, 2020) are two metric-based approaches. Template-NER (Cui et al., 2021) constructs a template for a single entity type, and enumerates each span together with this type and calculates its generation probability. Ent-LM (Ma et al., 2022) proposes a template-free prompt tuning method and induces the language models to predict label words at entity positions during fine-tuning, while Ent-LM + struct leverages the viterbi algorithm proposed in Yang and Katiyar (2020) to further boost the performance.\nTable 1 shows the results of our proposed PODA with these baselines. We only report the results of our method and the backbone model on experiments over ACE-2005 since traditional few-shot methods are hard to be applied to nested NER.\nBased on the results, we have the following observations: (1) For nearly all few-shot settings, our proposed method performs consistently better than the strong baselines. (2) It is worth noting that our method can outperform the backbone model T5 by 22.54/8.97/7.71 points when conducting 5-shot setting over these three datasets, which means generative PLMs like T5 are suffering from low-resource tuning and our method shows the strong ability of improving their training under few-shot settings. Even without the prompt-based order instructions, the augmented data can help the model achieve much better performance compared with pure T5, which demonstrates our claim that there is no need to keep the strict left-to-right order. (3) By comparing with Template-NER (Cui et al., 2021), which is also a template-based prompt method, the results show the advantages of our method over traditional template-based prompt method. We fairly query all the entity types rather than constructing a template for each type. (4) Regarding the experimental results on T5 and Flan-T5, it was observed that with- out our method, Flan-T5 consistently outperformed T5. However, our method demonstrated the ability to enhance the performance of both models in almost all scenarios, albeit with a relatively smaller improvement observed on Flan-T5. We suspect that this discrepancy is due to the disparity between our constructed data and natural language. While more enhanced data samples are provided (i.e. on MIT-Movie 50-shot setting), Flan-T5, which underwent instruction tuning, may become perplexed, leading to a relatively inferior performance compared to the pure T5 model."
},
{
"figure_ref": [
"fig_1"
],
"heading": "Analysis",
"publication_ref": [
"b12",
"b20"
],
"table_ref": [
"tab_3",
"tab_1"
],
"text": "Generalization Ability Our method can also be applied to other generative backbones, such as UIE (Lu et al., 2022) and BART-NER (Yan et al., 2021). To test the generalization ability and simplify the generation procedure, we test our approach over BART-NER since the target sequences of UIE include extra information. In order to alleviate the poor-tuning problem, we run BART-NER for 200 epochs. We only include the results of pure and enhanced (with PODA) BART-NER in Table 2 for clear comparison. The results show that BART-NER will have low-resource tuning even worse than pure T5 model, and our method can uniformly help BART-NER achieve a reasonable performance. In addition, we also have an interesting observation. By comparing results between Table 1 and 2, the results of BART-NER are uniformly lower than those of T5 on CoNLL-2003. But on ACE-2005, although the results are lower in 10-shot setting, BART-NER slightly performs better than T5 in 20-shot and outperforms T5 by 6.07 points in 50-shot. It is possible that the tagging schema of BART-NER is more compatible with nested NER. Our method can still further improves its performance, which means PODA has the ability to be uniformly generalized to different tagging schemes.\nImprovements with More Permutations As illustrated in the experiment settings, we use all the permutations to construct the order instructions since there are only 4 entity types in CoNLL-2003. To see whether the augmented data is valid, we also test the performance with increasing numbers of permutations. As visualized in Figure 2, we can observe significant performance improvement with only one re-ordered target sequence, and the model can be further improved with increasing permutations, which also demonstrates the effectiveness of our augmented data."
},
{
"figure_ref": [],
"heading": "Conclusions",
"publication_ref": [],
"table_ref": [],
"text": "In this paper, we propose PODA to improve the training of various generation-based NER methods (e.g. T5 and BART-NER) in different few-shot settings. By eliminating the strict left-to-right order assumption in traditional generation-based NER methods, PODA can construct sufficient while reasonable target entity sequences, thus leading to improved model training. To address situations where a single source sequence may have multiple target sequences, we additionally propose order instructions to facilitate the disambiguation of this one-to-many mapping. Experimental results demonstrate the effectiveness and generalization capability of both our data augmentation method and the prompt-based order instructions."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [
"b16"
],
"table_ref": [],
"text": "Although our approach can improve the training of generation-based NER methods, there are still some limitations and we leave as future directions to explore.\nMore Diverse Decoding Strategy In the training stage of our method, we concatenate the input sentence with all different prompt-based order instructions. But when evaluating, we only use the instruction \"from left to right\". We believe that if there is a proper algorithm that can select entities inside all the target entity sequences generated with different instructions, the performance will be further improved.\nDifferent Prompt Design As shown in previous work (Sanh et al., 2021), different prompts may affect the performance. In our work, we utilize some straightforward prompts as order instructions rather than specially designed. Using some special tokens in PLMs may also be helpful and a controllable generation-style method will then be proposed."
}
] | Recently, data augmentation (DA) methods have been proven to be effective for pretrained language models (PLMs) in lowresource settings, including few-shot named entity recognition (NER). However, conventional NER DA methods are mostly aimed at sequence labeling models, i.e., token-level classification, and few are compatible with unified autoregressive generation frameworks, which can handle a wider range of NER tasks, such as nested NER. Furthermore, these generation frameworks have a strong assumption that the entities will appear in the target sequence with the same left-to-right order as the source sequence. In this paper, we claim that there is no need to keep this strict order, and more diversified but reasonable target entity sequences can be provided during the training stage as a novel DA method. Nevertheless, a naive mixture of augmented data can confuse the model since one source sequence will then be paired with different target sequences. Therefore, we propose a simple but effective Prompt Ordering based Data Augmentation (PODA) method to improve the training of unified autoregressive generation frameworks under few-shot NER scenarios. Experimental results on three public NER datasets and further analyses demonstrate the effectiveness of our approach. | Enhancing Few-shot NER with Prompt Ordering based Data Augmentation | [
{
"figure_caption": "Figure 1 :1Figure1: A diagram of our proposed PODA. Every prompt-based order instruction is concatenated with the input sentence as a new source sequence to form the one-to-one mapping with a re-ordered target sequence (rather than following only the left-to-right order). Without any modification to the backbone, we can augment the training samples by several times. \"T5\" refers to the backbone of our method.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: The performance of PODA (T5) on CoNLL-2003 with different number of permutations, where 0 and l! means training with no and all permutations.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "11.71 66.89 ± 6.09 72.63 ± 3.42 StructShot(Yang and Katiyar, 2020) 45.82 ± 10.30 62.37 ± 10.96 69.51 ± 6.46 74.73 ± 3.06 The performance on three datasets with different (K=5, 10, 20, 50) few-shot settings. We report the mean and deviation results over 3 different splits for each cell.",
"figure_data": "DatasetsModelsK=5K=10K=20K=50T536.77 ± 9.0643.32 ± 3.1561.76 ± 2.11 70.35 ± 1.34Flan-T544.08 ± 4.9659.35 ± 0.8167.94 ± 2.86 72.74 ± 0.92CoNLL-2003NNShot (Yang and Katiyar, 2020) 59.24 ± Template-NER (Cui et al., 2021) 42.31 ± 8.92 43.04 ± 5.15 57.86 ± 5.6866.38 ± 6.09 72.71 ± 2.13Ent-LM (Ma et al., 2022)49.59 ± 8.3064.79 ± 3.8669.52 ± 4.48 73.66 ± 2.06Ent-LM + Struct (Ma et al., 2022)51.32 ± 7.6766.86 ± 3.0171.23 ± 3.91 74.80 ± 1.87PODA (T5)59.31 ± 1.8565.54 ± 1.1871.68 ± 0.80 75.66 ± 0.23PODA (Flan-T5)58.07 ± 1.2864.79 ± 1.3269.37 ± 1.25 73.09 ± 1.03T553.17 ± 4.0562.96 ± 1.1768.14 ± 0.86 72.49 ± 0.57Flan-T555.74 ± 3.1662.52 ± 0.8168.59 ± 0.79 73.09 ± 0.43NNShot (Yang and Katiyar, 2020)38.97 ± 5.5450.47 ± 6.0958.94 ± 3.47 71.17 ± 2.85StructShot (Yang and Katiyar, 2020) 41.60 ± 8.9753.19 ± 5.5261.42 ± 2.98 72.07 ± 6.41MIT-MovieTemplate-NER (Cui et al., 2021)45.97 ± 3.8649.30 ± 3.3559.09 ± 0.35 65.13 ± 0.17Ent-LM (Ma et al., 2022)46.62 ± 9.4657.31 ± 3.7262.36 ± 4.14 71.93 ± 1.68Ent-LM + Struct (Ma et al., 2022)49.15 ± 8.9159.21 ± 3.9663.85 ± 3.70 72.99 ± 1.80PODA (T5)62.14 ± 1.1966.62 ± 0.7670.03 ± 0.38 74.08 ± 0.41PODA (Flan-T5)62.86 ± 1.2565.50 ± 1.2268.81 ± 0.18 73.02 ± 0.56T525.71 ± 7.4128.78 ± 7.0234.47 ± 2.57 43.10 ± 0.40Flan-T528.30 ± 3.5135.26 ± 1.2439.05 ± 1.20 43.82 ± 1.00ACE-2005PODA (T5)33.42 ± 1.3138.73 ± 2.6042.22 ± 1.75 44.85 ± 1.19PODA (Flan-T5)36.44 ± 1.4940.45 ± 1.7641.36 ± 2.35 45.22 ± 0.57",
"figure_id": "tab_1",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": ": The performance of BART-NER and PODA(BART-NER) on two datasets with (K=10, 20, 50)-shot.",
"figure_id": "tab_3",
"figure_label": "2",
"figure_type": "table"
}
] | Huiming Wang; Liying Cheng; Wenxuan Zhang; De Wen Soh; Lidong Bing | [
{
"authors": "Jiaao Chen; Zhenghui Wang; Ran Tian; Zichao Yang; Diyi Yang",
"journal": "Association for Computational Linguistics",
"ref_id": "b0",
"title": "Local additivity based data augmentation for semi-supervised NER",
"year": "2020"
},
{
"authors": "Shuguang Chen; Gustavo Aguilar; Leonardo Neves; Thamar Solorio",
"journal": "Association for Computational Linguistics",
"ref_id": "b1",
"title": "Data augmentation for cross-domain named entity recognition",
"year": "2021"
},
{
"authors": "Chung Hyung Won; Le Hou; Shayne Longpre; Barret Zoph; Yi Tay; William Fedus; Yunxuan Li; Xuezhi Wang; Mostafa Dehghani; Siddhartha Brahma; Albert Webson; Shane Shixiang; Zhuyun Gu; Mirac Dai; Xinyun Suzgun; Aakanksha Chen; Alex Chowdhery; Marie Castro-Ros; Kevin Pellat; Dasha Robinson; Sharan Valter; Gaurav Narang; Adams Mishra; Vincent Yu; Yanping Zhao; Andrew Huang; Hongkun Dai; Slav Yu; Ed H Petrov; Jeff Chi; Jacob Dean; Adam Devlin; Denny Roberts; Quoc V Zhou; Jason Le; Wei",
"journal": "",
"ref_id": "b2",
"title": "Scaling instructionfinetuned language models",
"year": "2022"
},
{
"authors": "Leyang Cui; Yu Wu; Jian Liu; Sen Yang; Yue Zhang",
"journal": "Association for Computational Linguistics",
"ref_id": "b3",
"title": "Template-based named entity recognition using BART",
"year": "2021"
},
{
"authors": "Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova",
"journal": "",
"ref_id": "b4",
"title": "BERT: Pre-training of deep bidirectional transformers for language understanding",
"year": "2019"
},
{
"authors": "George Doddington; Alexis Mitchell; Mark Przybocki; Lance Ramshaw; Stephanie Strassel; Ralph Weischedel",
"journal": "European Language Resources Association (ELRA)",
"ref_id": "b5",
"title": "The automatic content extraction (ACE) program -tasks, data, and evaluation",
"year": "2004"
},
{
"authors": "Jiatao Gu; James Bradbury; Caiming Xiong; O K Victor; Richard Li; Socher",
"journal": "",
"ref_id": "b6",
"title": "Nonautoregressive neural machine translation",
"year": "2018"
},
{
"authors": "Jiaxin Huang; Chunyuan Li; Krishan Subudhi; Damien Jose; Shobana Balakrishnan; Weizhu Chen; Baolin Peng; Jianfeng Gao; Jiawei Han",
"journal": "Association for Computational Linguistics",
"ref_id": "b7",
"title": "Fewshot named entity recognition: An empirical baseline study",
"year": "2021"
},
{
"authors": "Guillaume Lample; Miguel Ballesteros; Sandeep Subramanian; Kazuya Kawakami; Chris Dyer",
"journal": "",
"ref_id": "b8",
"title": "Neural architectures for named entity recognition",
"year": "2016"
},
{
"authors": "Mike Lewis; Yinhan Liu; Naman Goyal; Marjan Ghazvininejad; Abdelrahman Mohamed; Omer Levy; Veselin Stoyanov; Luke Zettlemoyer",
"journal": "Association for Computational Linguistics",
"ref_id": "b9",
"title": "BART: Denoising sequence-to-sequence pretraining for natural language generation, translation, and comprehension",
"year": "2020"
},
{
"authors": "Jingjing Liu; Panupong Pasupat; Yining Wang; D Scott Cyphers; James R Glass",
"journal": "",
"ref_id": "b10",
"title": "Query understanding enhanced by hierarchical parsing structures",
"year": "2013"
},
{
"authors": "Wei Lu; Dan Roth",
"journal": "Association for Computational Linguistics",
"ref_id": "b11",
"title": "Joint mention extraction and classification with mention hypergraphs",
"year": "2015"
},
{
"authors": "Yaojie Lu; Qing Liu; Dai Dai; Xinyan Xiao; Hongyu Lin; Xianpei Han; Le Sun; Hua Wu",
"journal": "Association for Computational Linguistics",
"ref_id": "b12",
"title": "Unified structure generation for universal information extraction",
"year": "2022"
},
{
"authors": "Ruotian Ma; Xin Zhou; Tao Gui; Yiding Tan; Linyang Li; Qi Zhang; Xuanjing Huang",
"journal": "Association for Computational Linguistics",
"ref_id": "b13",
"title": "Templatefree prompt tuning for few-shot NER",
"year": "2022"
},
{
"authors": "Giovanni Paolini; Ben Athiwaratkun; Jason Krone; Jie Ma; Alessandro Achille; Rishita Anubhai; Cícero Nogueira Dos Santos; Bing Xiang; Stefano Soatto",
"journal": "",
"ref_id": "b14",
"title": "Structured prediction as translation between augmented natural languages",
"year": "2021"
},
{
"authors": "Colin Raffel; Noam M Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu",
"journal": "",
"ref_id": "b15",
"title": "Exploring the limits of transfer learning with a unified text-to-text transformer",
"year": "2019"
},
{
"authors": "Victor Sanh; Albert Webson; Colin Raffel; Stephen H Bach; Lintang Sutawika; Zaid Alyafeai; Antoine Chaffin; Arnaud Stiegler; Teven Le Scao; Arun Raja; Manan Dey; M Saiful Bari; Canwen Xu; Urmish Thakker; Shanya Sharma; Eliza Szczechla; Taewoon Kim; Gunjan Chhablani; V Nihal; Debajyoti Nayak; Jonathan Datta; Mike Chang; Tian-Jian; Han Jiang; Matteo Wang; Sheng Manica; Zheng Xin Shen; Harshit Yong; Rachel Pandey; Thomas Bawden; Trishala Wang; Jos Neeraj; Abheesht Rozen; Andrea Sharma; Thibault Santilli; Jason Févry; Alan Fries; Ryan Teehan; Stella Rose Biderman; Leo Gao; Tali Bers; Thomas Wolf; Alexander M Rush",
"journal": "",
"ref_id": "b16",
"title": "Multitask prompted training enables zero-shot task generalization",
"year": "2021"
},
{
"authors": "Zeqi Tan; Yongliang Shen; Shuai Zhang; Weiming Lu; Yueting Zhuang",
"journal": "",
"ref_id": "b17",
"title": "A sequence-to-set network for nested named entity recognition",
"year": "2021"
},
{
"authors": "Erik F ; Tjong Kim; Sang ; Fien De; Meulder ",
"journal": "",
"ref_id": "b18",
"title": "Introduction to the conll-2003 shared task: Languageindependent named entity recognition",
"year": "2003"
},
{
"authors": "Jue Wang; Wei Lu",
"journal": "",
"ref_id": "b19",
"title": "Two are better than one: Joint entity and relation extraction with tablesequence encoders",
"year": "2020"
},
{
"authors": "Hang Yan; Tao Gui; Junqi Dai; Qipeng Guo; Zheng Zhang; Xipeng Qiu",
"journal": "Association for Computational Linguistics",
"ref_id": "b20",
"title": "A unified generative framework for various NER subtasks",
"year": "2021"
},
{
"authors": "Yi Yang; Arzoo Katiyar",
"journal": "Association for Computational Linguistics",
"ref_id": "b21",
"title": "Simple and effective few-shot named entity recognition with structured nearest neighbor learning",
"year": "2020"
},
{
"authors": "Shuai Zhang; Yongliang Shen; Zeqi Tan; Yiquan Wu; Weiming Lu",
"journal": "Association for Computational Linguistics",
"ref_id": "b22",
"title": "De-bias for generative extraction in unified NER task",
"year": "2022"
},
{
"authors": "Zexuan Zhong; Danqi Chen",
"journal": "Association for Computational Linguistics",
"ref_id": "b23",
"title": "A frustratingly easy approach for entity and relation extraction",
"year": "2021"
},
{
"authors": "Ran Zhou; Xin Li; Ruidan He; Lidong Bing; Erik Cambria; Luo Si; Chunyan Miao",
"journal": "Association for Computational Linguistics",
"ref_id": "b24",
"title": "MELM: Data augmentation with masked entity language modeling for low-resource NER",
"year": "2022"
}
] | [
{
"formula_coordinates": [
2,
306.14,
614.53,
99.35,
10.63
],
"formula_id": "formula_0",
"formula_text": "X = [x 1 , x 2 , ..., x n ],"
},
{
"formula_coordinates": [
3,
115.91,
443.7,
173.23,
34.74
],
"formula_id": "formula_1",
"formula_text": "P (Y |X) = |Y | i=1 P (y i |X, Y <i ) (1)"
},
{
"formula_coordinates": [
3,
82.46,
711.08,
206.68,
13.14
],
"formula_id": "formula_2",
"formula_text": "Y p = [..., [(E p i ,1 , p i ), ..., (E p i ,np i , p i )], ...] (2)"
},
{
"formula_coordinates": [
3,
306.14,
569.19,
219.63,
23.36
],
"formula_id": "formula_3",
"formula_text": "Y p = \"[[(Britain, LOC)], [(EU, MISC), (BSE, MISC)]]\" 2 ."
}
] |
10.18653/v1/W19-4813 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b22",
"b23",
"b5",
"b34",
"b21",
"b18",
"b13",
"b13",
"b11",
"b42",
"b19",
"b14",
"b13",
"b30",
"b20",
"b38",
"b11",
"b12",
"b42"
],
"table_ref": [],
"text": "Reference-based neural metrics for machine translation evaluation are achieving evergrowing success, demonstrating superior results over traditional lexical overlap-based metrics, such as BLEU (Papineni et al., 2002) and CHRF (Popović, 2015), in terms of both their correlation with human ratings and their robustness across diverse domains (Callison-Burch et al., 2006;Smith et al., 2016;Mathur et al., 2020;Kocmi et al., 2021;Freitag et al., 2022). However, lexical overlapbased metrics remain popular for evaluating the performance and progress of translation systems and algorithms. Concerns regarding trust and interpretability may help explain this (Leiter et al., Figure 1: Illustration of our approach. In this example, the metric assigns the translation a low score. We aim to better understand this sentence-level assessment by examining the correspondence between our token-level explanations and human annotated error spans. 2022): contrary to traditional metrics, neural metrics are considered \"black boxes\" as they often use increasingly large models (e.g., the winning metric of the WMT 22 Metrics shared task was a 10B parameter model (Freitag et al., 2022)).\nWhile some recent work has focus on explaining the predictions made by reference-free quality estimation (QE) systems (Fomicheva et al., 2021;Zerva et al., 2022), explaining reference-based metrics has remained a largely overlooked problem (Leiter et al., 2022). It is an open question whether the observations from studies of explainable QE carry over to this scenario. Thus, in this work, we fill that gap by turning to state-of-theart reference-based metrics-we aim to interpret their decision-making process by exploiting the fact that these metrics show consistently good correlations with Multidimentional Quality Metrics (MQM) (Freitag et al., 2021(Freitag et al., , 2022;;Sai et al., 2022), which are fine-grained quality assessments that result from experts identifying error spans in translation outputs (Lommel et al., 2014). We hypothesize that reference-based metrics leverage this tokenlevel information to produce sentence-level scores. To test this hypothesis, we assess whether our explanations -measures of token-level importance obtained via attribution and input attribution methods arXiv:2305.11806v1 [cs.CL] 19 May 2023 such as attention weights and gradient scores (Treviso et al., 2021;Rei et al., 2022b) -align with human-annotated spans (Fomicheva et al., 2021(Fomicheva et al., , 2022;;Zerva et al., 2022), as illustrated in Figure 1.\nOur analysis focuses on two main vectors: (i) understanding the impact of the reference information on the quality of the explanations; and (ii) finding whether the explanations can help to identify potential weaknesses in the metrics. Our main contributions are:\n• We provide a comparison between multiple explainability methods for different metrics on all types of evaluation: src-only, ref-only, and src+ref joint evaluation.\n• We find that explanations are related to the underlying metric architecture, and that leveraging reference information improves the explanations.\n• We show that explanations for critical translation errors can reveal weaknesses in the metrics."
},
{
"figure_ref": [],
"heading": "Explaining Neural Metrics",
"publication_ref": [],
"table_ref": [],
"text": "We aim to explain sentence-level quality assessments of reference-based metrics by producing token-level explanations that align to translation errors. In what follows, we describe the metrics and how we produce the explanations that we study."
},
{
"figure_ref": [],
"heading": "Metrics",
"publication_ref": [
"b27",
"b40",
"b13",
"b31",
"b39"
],
"table_ref": [],
"text": "We focus our analysis on two state-of-the-art neural metrics: COMET (Rei et al., 2020) and UNITE (Wan et al., 2022 (Freitag et al., 2022). 2020); REF, like BLEURT (Sellam et al., 2020); and SRC+REF, like ROBLEURT (Wan et al., 2021)."
},
{
"figure_ref": [],
"heading": "Explanations via Attribution Methods",
"publication_ref": [
"b29",
"b32",
"b35",
"b16",
"b3",
"b41",
"b38",
"b12",
"b10",
"b42",
"b37",
"b2",
"b38"
],
"table_ref": [],
"text": "In this work, we produce explanations using attribution methods that assign a scalar value to each translation token (i.e. a token-level attribution) to represent its importance. While many input attribution methods exist and have been extensively studied in the literature (Ribeiro et al., 2016;Shrikumar et al., 2017;Sundararajan et al., 2017;Jain and Wallace, 2019;Atanasova et al., 2020;Zaman and Belinkov, 2022), we focus specifically on those that have been demonstrated to be effective for explaining the predictions of QE models (Treviso et al., 2021;Fomicheva et al., 2022;Fernandes et al., 2022;Zerva et al., 2022) and extend them to our reference-based scenario. Concretely, we use the following techniques to extract explanations:2 \n• embed-align: the maximum cosine similarity between each translation token embedding and the reference and/or source token embeddings (Tao et al., 2022);\n• grad 2 : the 2 -norm of gradients with respect to the word embeddings of the translation tokens (Arras et al., 2019);\n• attention: the attention weights of the translation tokens for each attention head of the encoder (Treviso et al., 2021);\n• attn × grad: the attention weights of each head scaled by the 2 -norm of the gradients of the value vectors of that head (Rei et al., 2022b).\n3 Experimental Setting MQM annotations. We use MQM annotations from the WMT 2021 Metrics shared task (Freitag et al., 2021),3 covering three language pairs -English-German (en→de), English-Russian (en→ru), and Chinese-English (zh→en) -in two different domains: News and TED Talks. For each incorrect translation, human experts marked the corresponding error spans. In our framework, these\nMETRIC EXPLAINABILITY en→de zh→en en→ru Avg. METHOD AUC R@K AUC R@K AUC R@K AUC R@K src-only evaluation UNITE SRC\nembed-align [mt, src] [mt, src] 0.590 0.371 0.674 0.314 0.577 0.220 0.614 0.301 embed-align [mt, ref] 0.694 0.425 0.696 0.355 0.647 0.275 0.679 0.352 embed-align [mt, src; ref] error spans should align with the words that the attribution methods assign higher importance to."
},
{
"figure_ref": [],
"heading": "Models.",
"publication_ref": [
"b11",
"b12",
"b42"
],
"table_ref": [],
"text": "For COMET, we use the latest publicly available model: wmt22-comet-da (Rei et al., 2022a). 4 For UNITE, we train our own model using the same data used to train COMET in order to have a comparable setup5 . We provide full details (training data, correlations with human annotations, and hyperparameters) in Appendix A.\nOverall, the resulting reference-based UNITE models (REF and SRC+REF) are on par with COMET.\nEvaluation. We want our explanations to be directly attributed to the annotated error spans, in the style of an error detection task. Thus, we report Area Under Curve (AUC) and Recall@K. 6 These metrics have been used as the main metrics in previous works on explainable QE (Fomicheva et al., 2021(Fomicheva et al., , 2022;;Zerva et al., 2022)."
},
{
"figure_ref": [],
"heading": "Results",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "High-level analysis",
"publication_ref": [
"b42"
],
"table_ref": [],
"text": "Explanations are tightly related to the underlying metric architecture. The results in Table 1 show that the predictive power of the attribution methods differ between UNITE and COMET: attn × grad is the best method for UNITEbased models, while embed-align works best for COMET.7 This is expected as UNITE constructs a joint representation for the input sentences, thus allowing attention to flow across them; COMET, in contrast, encodes the sentences separately, so it relies heavily on the separate contextualized embeddings that are subsequently combined via elementwise operations such as multiplication and absolute difference. Interestingly, embed-align and attn × grad were the winning explainability approaches of the WMT 2022 Shared-Task on Quality Estimation (Zerva et al., 2022). This suggests that explainability methods developed for QE systems can translate well to reference-based metrics. We provide examples of explanations in Appendix C.\nReference information boosts explainability power. rics, using reference information brings substantial improvements over using only the source information. Moreover, while reference-based attributions significantly outperform source-based attributions, combining the source and reference information to obtain token-level attributions does not consistently yield superior results over using the reference alone. Notably, the best attribution method for COMET does not require any source information. This is interesting: in some cases, reference-based metrics may largely ignore source information, relying heavily on the reference instead."
},
{
"figure_ref": [],
"heading": "How do the explanations fare for critical translation errors?",
"publication_ref": [
"b0",
"b7",
"b6",
"b1",
"b25"
],
"table_ref": [],
"text": "The MQM data analyzed until now consists primarily of high quality translations, with the majority of annotated errors being non-critical. However, it is important to assess whether our explanations can be accurately attributed to critical errors, as this may reveal potential metric shortcomings. To this end, we employ SMAUG (Alves et al., 2022) 8 , a tool designed to generate synthetic data for stresstesting metrics, to create corrupted translations that contain critical errors. Concretely, we generate translations with the following pathologies: negation errors, hallucinations via insertions, named entity errors, and errors in numbers.9 \nExplanations identify critical errors more easily than non-critical errors. Figure 2 shows that explanations are more effective in identifying critical errors compared to other non-critical errors (see Explanations can reveal potential metric weaknesses. Figure 2 suggests that COMET explanations struggle to identify localized errors (negation errors, named entity errors and discrepancies in numbers). We hypothesize that this behavior is related to the underlying architecture. Unlike UNITE-based metrics, COMET does not rely on soft alignments via attention between the sentences in the encoding process. This process may be key to identify local misalignments during the encoding process. In fact, the attention-based attributions for UNITE metrics can more easily identify these errors. COMET, however, encodes the sentences separately, which may result in grammatical features (e.g. numbers) being encoded similarly across sentences (Chi et al., 2020;Chang et al., 2022). As such, explanations obtained via embedding alignments will not properly identify these misalignments on similar features. Importantly, these findings align with observations made in (Amrhein and Sennrich, 2022;Raunak et al., 2022). This showcases how explanations can be used to diagnose and reveal shortcomings of neural-based metrics."
},
{
"figure_ref": [],
"heading": "Conclusions and Future Work",
"publication_ref": [
"b17"
],
"table_ref": [],
"text": "In this paper, we investigated the use of explainability methods to better understand widely used neural metrics for machine translation evaluation, such as COMET and UNITE. Concretely, we analyzed how explanations are impacted by the reference information, and how they can be used to reveal weaknesses of these metrics. Our analysis shows that the quality of the explanations is tightly related to the underlying metric architecture. Interestingly, we also provide evidence that neural metrics like COMET may heavily rely on reference information over source information. Additionally, we show that explanations can be used to reveal reference-based metrics weaknesses such as failing to severely penalize localized critical errors. This opens up promising opportunities for future research on leveraging explanations to diagnose reference-based metrics errors. To support these studies, we call for future datasets illustrating critical errors (e.g., challenge sets (Karpinska et al., 2022)) to be accompanied by annotated error spans."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [
"b4"
],
"table_ref": [],
"text": "We highlight three main limitations of our work.\nFirst, although we have explored gradient-based explanations that take the whole network into consideration and have been shown to be faithful in previous work (Bastings et al., 2021), we do not explicitly explore how COMET combines the sentence representations in the feed-forward that precedes the encoder model to produce the sentence-level score.\nSecond, we have shown that combining attention with gradient information results in the best explanations for UNITE-based metrics. However, from a practical standpoint, running inference and extracting the explainability scores simultaneously may be more computationally expensive than other techniques: gradient-based metrics benefit from GPU infrastructure and require storing all gradient information.\nThird, we have not explored extracting explanations in low-resource settings. That is because high-quality MQM annotations for such language pairs are not yet available. Nevertheless, further research in those settings is needed to access the broader validity of our claims."
},
{
"figure_ref": [],
"heading": "A Model Details",
"publication_ref": [
"b13",
"b40"
],
"table_ref": [
"tab_6",
"tab_7",
"tab_8"
],
"text": "In Section 2.1, we employed the latest publicly available model (wmt22-comet-da) for COMET, which emerged as a top-performing metric in the WMT 2022 Metrics task (Freitag et al., 2022). To ensure a comparable setting for UNITE (Wan et al., 2022), we trained our own model. In doing so, we utilized the same data employed in the development of the COMET model by (Rei et al., 2022a), without pretraining any synthetic data, as originally suggested. Additionally, our implementation did not incorporate monotonic regional attention, as our preliminary experiments revealed no discernible benefits from its usage. The hyperparameters used are summarized in Table 3, while Table 4 presents the number of Direct Assessments utilized during training. Furthermore, Table 5 displays the segment-level correlations with WMT 2021 MQM data for the News and TED domains.\nRegarding infrastructure, a single NVIDIA A10G GPU with 23GB memory was used. The resulting UNITE model has 565M parameters while COMET has 581M parameters."
},
{
"figure_ref": [
"fig_0"
],
"heading": "A.1 Output Distribution",
"publication_ref": [],
"table_ref": [],
"text": "To better understand the output of the models and what scores are deemed low, we plotted the output distributions for the two models we used in our study. The average score for English→German data is 0.856 for the COMET model and 0.870 for the UNITE model we trained. From Figure 3 we can observe the distribution of scores. This means that the 0.6692 score from the example in Figure 1 corresponds to a low quality output (5th percentile)."
},
{
"figure_ref": [],
"heading": "A.2 SMAUG Corpus",
"publication_ref": [],
"table_ref": [
"tab_9"
],
"text": "As we have seen in Section 4.2, we have created synthetic translation errors for the following pathologies: negation errors, hallucinations via insertions, named entity errors, and errors in numbers. Table 6 presents a summary of the examples created using SMAUG and in Table 8 we show examples of each error category."
},
{
"figure_ref": [],
"heading": "B Comparison between COMET and XLM-R Alignments",
"publication_ref": [],
"table_ref": [
"tab_2",
"tab_5"
],
"text": "From Table 1, it is evident that the alignments between the reference and/or source and the translation yield effective explanations for COMET. This raises the question of how these alignments compare to the underlying encoder of COMET before the fine-tuning process with human annotations. To investigate this, we examine the results for XLM-R without any fine-tuning, as presented in Table 2. Overall, the explanations derived from the alignments of COMET prove to be more predictive of error spans than those obtained from XLM-R alignments. This suggests that during the fine-tuning phase, COMET models modify the underlying XLM-R representations to achieve better alignment with translation errors."
},
{
"figure_ref": [],
"heading": "C Examples",
"publication_ref": [
"b14",
"b9",
"b14"
],
"table_ref": [
"tab_4",
"tab_4"
],
"text": "In Tables 9 and10, we show examples of COMET explanations for Chinese→English and English→German language pairs, respectively. We highlight in gray the corresponding MQM annotation performed by an expert linguist and we sort the examples from highest to lowest COMET scores.\nFrom these examples we can observe the following:\n• Highlights provided by COMET explanations have a high recall with human annotations. In all examples, subword tokens corresponding to translation errors are highlighted in red but we often see that not everything is incorrect.\n• Explanations are consistent with scores. For example, in the third example from Table 10, the red highlights do not correspond to errors and in fact the translation only has a major error griffen . Nonetheless, the score assigned by COMET is a low score of 0.68 which is faithful to the explanations that was given even if the assessment does not agree with human experts.\nMETRIC EXPLAINABILITY en→de zh→en en→ru Avg. METHOD AUC R@K AUC R@K AUC R@K AUC R@K XLM-R embed-align [mt, src] 0.587 0.359 0.668 0.311 0.576 0.199 0.610 0.289 embed-align [mt, ref] 0.671 0.405 0.689 0.345 0.634 0.244 0.664 0.331 embed-align [mt, src; ref] 0.666 0.395 0.690 0.347 0.616 0.242 0.657 0.328 COMET embed-align [mt, src] 0.590 0.371 0.674 0.314 0.577 0.220 0.614 0.301 embed-align [mt, ref] 0.694 0.425 0.696 0.355 0.647 0.275 0.679 0.352 embed-align [mt, src; ref] (Freitag et al., 2021). The metrics are Pearson (ρ) and Kendall Tau (τ ). Results in bold indicate which metrics are top-performing for that specific language pair, domain and metric according to Perm-Both hypothesis test (Deutsch et al., 2021), using 500 re-sampling runs, and setting p = 0.05. (Freitag et al., 2021) used in our experiments."
},
{
"figure_ref": [],
"heading": "Acknowledgements",
"publication_ref": [],
"table_ref": [],
"text": "This work was partially supported by the P2020 programs (MAIA, contract 045909), the Portuguese Recovery and Resilience Plan (PRR) through project C645008882-00000055, Center for Responsible AI, by the European Research Council (ERC StG DeepSPIN, 758969), by EU's Horizon Europe Research and Innovation Actions (UTTER, contract 101070631), and by the Fundação para a Ciência e Tecnologia (contracts UIDB/50021/2020 and UIDB/50008/2020)."
},
{
"figure_ref": [],
"heading": "",
"publication_ref": [],
"table_ref": [],
"text": "https://github.com/Unbabel/COMET"
},
{
"figure_ref": [],
"heading": "Source: 格里沃里表示，分析人士对越南所提出的和平倡议给予认可。",
"publication_ref": [],
"table_ref": [],
"text": "Translation: Grivory said that analysts recognize the peace initiative proposed by Vietnam. Reference: Grigory said that analysts endorse the peace initiative proposed by Vietnam. NE Error: Grivory said that analysts recognize the peace initiative proposed by Russia ."
},
{
"figure_ref": [],
"heading": "Source: 英国的这一决定预计将会使西班牙的旅游业大受影响。",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Translation:",
"publication_ref": [],
"table_ref": [],
"text": "This decision by the United Kingdom is expected to greatly affect Spain's tourism industry. Reference: This decision by the UK is expected to have a significant impact on tourism in Spain."
},
{
"figure_ref": [],
"heading": "NEG Error:",
"publication_ref": [
"b0"
],
"table_ref": [],
"text": "This decision by the United Kingdom is expected to greatly benefit Spain's tourism industry. Table 8: Synthetically-generated critical errors ( highlighted in gray ) created with SMAUG (Alves et al., 2022) to assess whether our explanations can be accurately attributed to critical errors.\nSource: And yet, the universe is not a silent movie because the universe isn't silent. Translation: Und dennoch ist das Universum kein Stummfilm, weil das Universum nicht still ist. COMET score: 0.8595 explanation: _Und _dennoch _ist _das _Univers um _kein _Stu mm film , _weil _das _Univers um _nicht _still _ist ."
},
{
"figure_ref": [],
"heading": "Source:",
"publication_ref": [],
"table_ref": [],
"text": "And yet black holes may be heard even if they're not seen, and that's because they bang on space-time like a drum. Translation: Und dennoch werden Schwarze Löcher vielleicht gehört , auch wenn sie nicht gesehen werden, und das liegt daran, dass sie wie eine Trommel auf die Raumzeit schlagen. COMET score: 0.7150 COMET explanation: _Und _dennoch _werden _Schwarz e _Lö cher _vielleicht _gehört , _auch _wenn _sie _nicht _gesehen _werden , _und _das _liegt _daran , _dass _sie _wie _eine _Tro mmel _auf _die _Raum zeit schlagen ."
},
{
"figure_ref": [],
"heading": "Source:",
"publication_ref": [],
"table_ref": [],
"text": "Ash O'Brien and husband Jarett Kelley say they were grabbing a bite to eat at Dusty Rhodes dog park in San Diego on Thursday, with their three-month-old pug in tow. Translation: Ash O'Brien und Ehemann Jarett Kelley sagen, dass sie am Donnerstag im Hundepark Dusty Rhodes in San Diego einen Happen zu essen griffen , mit ihrem drei Monate alten Mops im Schlepptau.\nCOMET score: 0.6835 COMET explanation: _Ash _O ' Bri en _und _Ehe mann _Ja rett _Kel ley _sagen , _dass _sie _am _Donnerstag _im _Hunde park _Du sty _Rhod es _in _San _Diego _einen _Happ en _zu _essen _ griff en _ , _mit _ihrem _drei _Monate _alten _M ops _im _Schle ppt au ."
},
{
"figure_ref": [],
"heading": "Source:",
"publication_ref": [],
"table_ref": [],
"text": "It was Einstein's great general theory of relativity. Translation: Es war Einsteins große allgemeine Forschungen vor Relativitätstheorie.\nCOMET score: 0.6692 COMET explanation: _Es _war _Einstein s _große _allgemein e _Forschung en _vor _Relativ ität s the ori e ."
},
{
"figure_ref": [],
"heading": "Source:",
"publication_ref": [],
"table_ref": [],
"text": "There's mask-shaming and then there's full on assault. Translation: Es gibt Maskenschämen und dann ist es voll bei Angriff. COMET score: 0.2318 COMET explanation: _Es _gibt _Mask en schä men _und _dann _ist _es _voll _bei _Angriff _ .\nTable 9: Saliency map for COMET explanation scores for a set of en→de examples. Comparing the token-level explanations with the MQM annotation ( highlighted in gray ) reveals the source of correspondence between specific token-level translation errors and the resulting scores.\nSource: 我想告诉大家 宇宙有着自己的配乐， 而宇宙自身正在不停地播放着。 因为太空可以想鼓一样振动。 Translation: I want to tell you that the universe has its own iconic soundtrack and the universe itself is constantly playing non-stop because space can vibrate like a drum. COMET score: 0.8634 COMET explanation: _I _want _to _tell _you _that _the _univers e _has _its _own _icon ic _soundtrack _and _the _univers e _itself _is _constantly _playing _non -stop _because _space _can _vibra te _like _a _drum .\nSource: 另外,吉克隽逸和刘烨作为运动助理,也围绕运动少年制造了不少爆笑话题。"
},
{
"figure_ref": [],
"heading": "Translation:",
"publication_ref": [],
"table_ref": [],
"text": "In addition, as sports assistants, Ji Kejunyi and Liu Ye have also created a lot of hilarious topics around sports teenagers.\nCOMET score: 0.8214 COMET explanation: _In _addition , _as _sports _assistant s , _Ji _Ke ju nyi _and _Li u _Ye _have _also _created _a _lot _of _ hila rious _topic s _around _sports _teenager s ."
},
{
"figure_ref": [],
"heading": "Source: 一番言论让场上的少年和运动领队们都倒吸一口凉气。",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Translation:",
"publication_ref": [],
"table_ref": [],
"text": "The remarks made the teenagers and the sports leaders on the field gasp a sigh of relief .\nCOMET score: 0.7793 COMET explanation: _The _re marks _made _the _teenager s _and _the _sports _leaders _on _the _field _gas p _a _sig h _of _relief _ ."
},
{
"figure_ref": [],
"heading": "Source: 强烈的阳光是如此地刺眼， Translation:",
"publication_ref": [],
"table_ref": [],
"text": "The intense sunlight is so harsh;\nCOMET score: 0.7561 COMET explanation: _The _intense _sun light _is _so _har sh ; Source: 如今，我们希望能够 给这部关于宇宙的 宏伟的视觉作品 配上声音。 Translation: Today , we hope to be able to give this magnificent visual work of the universe a sound.\nCOMET score: 0.7073 COMET explanation: _Today , _we _hope _to _be _able _to _give _this _magnific ent _visual _work _of _the _univers e _a _sound .\nTable 10: Saliency map for COMET explanation scores for a set of zh→en examples. Comparing the tokenlevel explanations with the MQM annotation ( highlighted in gray ) reveals the source of correspondence between specific token-level translation errors and the resulting scores."
}
] | Neural metrics for machine translation evaluation, such as COMET, exhibit significant improvements in their correlation with human judgments compared to traditional metrics based on lexical overlap, such as BLEU. Yet neural metrics are, to a great extent, "black boxes" that return a single sentencelevel score without transparency about the decision-making process. In this work, we develop and compare several neural explainability methods and demonstrate their effectiveness for interpreting state-of-the-art fine-tuned neural metrics. Our study reveals that these metrics leverage token-level information that can be directly attributed to translation errors, as assessed through comparison of token-level neural saliency maps with Multidimensional Quality Metrics (MQM) annotations and with syntheticallygenerated critical translation errors. To ease future research, we release our code at | The Inside Story: Towards Better Understanding of Machine Translation Neural Evaluation Metrics | [
{
"figure_caption": "Figure 3 :3Figure 3: Distribution of scores for all metrics obtained on the MQM data (for all language pairs).",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "0.688 0.416 0.697 0.357 0.622 0.279 0.669 0.350 AUC and Recall@K of explanations obtained via different attribution methods for COMET and UNITE models on the MQM data. Although UNITE SRC is a src-only evaluation metric, it was trained with reference information(Wan et al., 2022).",
"figure_data": "grad 20.603 0.312 0.540 0.252 0.604 0.185 0.582 0.250attention0.604 0.351 0.592 0.259 0.633 0.209 0.608 0.268attn × grad0.710 0.365 0.633 0.278 0.662 0.244 0.669 0.295",
"figure_id": "tab_2",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Table 1 also shows that, across all met-",
"figure_data": "COMETUNITE REFUNITE [email protected] 0.4 0.60NEGHALLNENUMFigure 2: Performance of the best attribution methodsfor COMET, UNITE REF and UNITE SRC+REF interms of Recall@K on translations with critical errors:negations (NEG), hallucinations (HALL), named entityerrors (NE), and errors in numbers (NUM).",
"figure_id": "tab_3",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": "). Specifically, we find significant per-formance improvements up to nearly 30% in Re-call@K for certain critical errors. Overall, halluci-nations are the easiest errors to identify across allneural metrics. This suggests that neural metricsappropriately identify and penalize hallucinatedtranslations, which aligns with the findings of Guer-reiro et al. (2023). Moreover, explanations forboth UNITE models behave similarly for all er-rors except numbers, where the source informationplays a key role in improving the explanations. No-tably, contrary to what we observed for data withnon-critical errors, COMET explanations are lesseffective than those of UNITE REF and UNITESRC+REF for identifying critical errors.",
"figure_id": "tab_4",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "0.688 0.416 0.697 0.357 0.622 0.279 0.669 0.350 AUC and Recall@K of explanations obtained via alignments for COMET and XLM-R without any further fine-tuning on human annotations.",
"figure_data": "HyperparameterUNITE COMETEncoder ModelXLM-R (large)OptimizerAdamWNo. frozen epochs0.3Learning rate (LR)1.5e-05Encoder LR.1.0e-06Layerwise Decay0.95Batch size16Loss functionMSEDropout0.1Hidden sizes[3072, 1024]Embedding layerFrozenFP precision16No. Epochs12",
"figure_id": "tab_5",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Hyperparameters used to train UNITE and COMET checkpoints used in this work. The only difference between the two is the number of training epochs due to the fact that, for UNITE, the best validation checkpoint is the first one.",
"figure_data": "Language PairSIZEzh-en126947en-de121420de-en99183en-zh90805ru-en79280en-ru62749en-cs60937fi-en46145en-fi34335tr-en30186et-en29496cs-en27847en-mr26000de-cs13804en-et13376pl-en11816en-pl10572lt-en10315en-ja9578gu-en9063si-en9000ro-en9000ne-en9000en-lt8959ja-en8939en-kk8219en-ta7890ta-en7577en-gu6924kk-en6789de-fr6691en-lv5810en-tr5171km-en4722ps-en4611fr-de3999Total1027155",
"figure_id": "tab_6",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Number of direct assessments per language pair used to train COMET(Rei et al., 2022a) and the UNITE model used in this work.",
"figure_data": "",
"figure_id": "tab_7",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Segment-level correlations for WMT 2021 MQM annotations over News and TED domains",
"figure_data": "BLEU CHRF YISI-1 BLEURT UNITE UNITEUNITECOMETSRCREFSRC+REF wmt22-comet-daEN→DENews TEDρ 0.077 0.092 τ 0.069 0.092 ρ 0.151 0.158 τ 0.113 0.1460.163 0.144 0.236 0.2120.307 0.240 0.325 0.2830.274 0.222 0.311 0.2640.321 0.248 0.335 0.3010.304 0.241 0.338 0.2980.297 0.232 0.329 0.278EN→RUNews TEDρ 0.153 0.252 τ 0.106 0.178 ρ 0.154 0.268 τ 0.112 0.1890.263 0.216 0.235 0.2040.359 0.276 0.286 0.2550.333 0.276 0.239 0.2320.391 0.298 0.289 0.2620.382 0.297 0.318 0.2640.363 0.293 0.308 0.268ZH→ENNews TEDρ 0.215 0.231 τ 0.165 0.188 ρ 0.155 0.181 τ 0.113 0.1440.301 0.289 0.287 0.2160.428 0.341 0.295 0.2460.413 0.331 0.244 0.2240.438 0.358 0.301 0.2650.426 0.352 0.310 0.2660.445 0.371 0.307 0.269",
"figure_id": "tab_8",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "Number of examples for each category, synthetically-created using SMAUG(Alves et al., 2022).",
"figure_data": "Error Type NUM EXAMPLESNE978NEG669HALL530NUM432Total2609",
"figure_id": "tab_9",
"figure_label": "6",
"figure_type": "table"
},
{
"figure_caption": "Statistics about MQM data from WMT 2021 Metrics task",
"figure_data": "",
"figure_id": "tab_10",
"figure_label": "7",
"figure_type": "table"
}
] | Ricardo Rei; Nuno M Guerreiro; Marcos Treviso; Alon Lavie; Luisa Coheur; André F T Martins | [
{
"authors": "Duarte Alves; Ricardo Rei; Ana C Farinha; G C José; De Souza; F T André; Martins",
"journal": "Association for Computational Linguistics",
"ref_id": "b0",
"title": "Robust MT Evaluation with Sentence-level Multilingual Augmentation",
"year": "2022"
},
{
"authors": "Chantal Amrhein; Rico Sennrich",
"journal": "Association for Computational Linguistics",
"ref_id": "b1",
"title": "Identifying Weaknesses in Machine Translation Metrics Through Minimum Bayes Risk Decoding: A Case Study for COMET",
"year": "2022"
},
{
"authors": "Leila Arras; Ahmed Osman; Klaus-Robert Müller; Wojciech Samek",
"journal": "Association for Computational Linguistics",
"ref_id": "b2",
"title": "Evaluating recurrent neural network explanations",
"year": "2019"
},
{
"authors": "Pepa Atanasova; Jakob Grue Simonsen; Christina Lioma; Isabelle Augenstein",
"journal": "Association for Computational Linguistics",
"ref_id": "b3",
"title": "A diagnostic study of explainability techniques for text classification",
"year": "2020"
},
{
"authors": "Jasmijn Bastings; Sebastian Ebert; Polina Zablotskaia; Anders Sandholm; Katja Filippova",
"journal": "",
"ref_id": "b4",
"title": "will you find these shortcuts?\" a protocol for evaluating the faithfulness of input salience methods for text classification",
"year": "2021"
},
{
"authors": "Chris Callison-Burch; Miles Osborne; Philipp Koehn",
"journal": "Association for Computational Linguistics",
"ref_id": "b5",
"title": "Re-evaluating the role of Bleu in machine translation research",
"year": "2006"
},
{
"authors": "Tyler A Chang; Zhuowen Tu; Benjamin K Bergen",
"journal": "",
"ref_id": "b6",
"title": "The geometry of multilingual language model representations",
"year": "2022"
},
{
"authors": "Ethan A Chi; John Hewitt; Christopher D Manning",
"journal": "",
"ref_id": "b7",
"title": "Finding universal grammatical relations in multilingual BERT",
"year": "2020"
},
{
"authors": "Alexis Conneau; Kartikay Khandelwal; Naman Goyal; Vishrav Chaudhary; Guillaume Wenzek; Francisco Guzmán; Edouard Grave; Myle Ott; Luke Zettlemoyer; Veselin Stoyanov",
"journal": "Association for Computational Linguistics",
"ref_id": "b8",
"title": "Unsupervised cross-lingual representation learning at scale",
"year": "2020"
},
{
"authors": "Daniel Deutsch; Rotem Dror; Dan Roth",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b9",
"title": "A statistical analysis of summarization evaluation metrics using resampling methods",
"year": "2021"
},
{
"authors": "Patrick Fernandes; Marcos Treviso; Danish Pruthi; F T André; Graham Martins; Neubig",
"journal": "",
"ref_id": "b10",
"title": "Learning to scaffold: Optimizing model explanations for teaching",
"year": "2022"
},
{
"authors": "Marina Fomicheva; Piyawat Lertvittayakumjorn; Wei Zhao; Steffen Eger; Yang Gao",
"journal": "Association for Computational Linguistics",
"ref_id": "b11",
"title": "The Eval4NLP shared task on explainable quality estimation: Overview and results",
"year": "2021"
},
{
"authors": "Marina Fomicheva; Lucia Specia; Nikolaos Aletras",
"journal": "Association for Computational Linguistics",
"ref_id": "b12",
"title": "Translation error detection as rationale extraction",
"year": "2022"
},
{
"authors": "Markus Freitag; Ricardo Rei; Nitika Mathur; Chikiu Lo; Craig Stewart; Eleftherios Avramidis; Tom Kocmi; George Foster; Alon Lavie; F T André; Martins",
"journal": "Association for Computational Linguistics",
"ref_id": "b13",
"title": "Results of WMT22 Metrics Shared Task: Stop Using BLEU -Neural Metrics Are Better and More Robust",
"year": "2022"
},
{
"authors": "Markus Freitag; Ricardo Rei; Nitika Mathur; Chi-Kiu Lo; Craig Stewart; George Foster; Alon Lavie; Ondřej Bojar",
"journal": "Association for Computational Linguistics",
"ref_id": "b14",
"title": "Results of the WMT21 metrics shared task: Evaluating metrics with expert-based human evaluations on TED and news domain",
"year": "2021"
},
{
"authors": "M Nuno; Elena Guerreiro; André Voita; Martins",
"journal": "Association for Computational Linguistics",
"ref_id": "b15",
"title": "Looking for a needle in a haystack: A comprehensive study of hallucinations in neural machine translation",
"year": "2023"
},
{
"authors": "Sarthak Jain; Byron C Wallace",
"journal": "Association for Computational Linguistics",
"ref_id": "b16",
"title": "Attention is not Explanation",
"year": "2019"
},
{
"authors": "Marzena Karpinska; Nishant Raj; Katherine Thai; Yixiao Song; Ankita Gupta; Mohit Iyyer",
"journal": "Association for Computational Linguistics",
"ref_id": "b17",
"title": "Demetr: Diagnosing evaluation metrics for translation",
"year": "2022"
},
{
"authors": "Tom Kocmi; Christian Federmann; Roman Grundkiewicz; Marcin Junczys-Dowmunt; Hitokazu Matsushita; Arul Menezes",
"journal": "Association for Computational Linguistics",
"ref_id": "b18",
"title": "To ship or not to ship: An extensive evaluation of automatic metrics for machine translation",
"year": "2021"
},
{
"authors": "Christoph Leiter; Piyawat Lertvittayakumjorn; Marina Fomicheva; Wei Zhao; Yang Gao; Steffen Eger",
"journal": "",
"ref_id": "b19",
"title": "Towards explainable evaluation metrics for natural language generation",
"year": "2022"
},
{
"authors": "Arle Lommel; Hans Uszkoreit; Aljoscha Burchardt",
"journal": "Tradumàtica",
"ref_id": "b20",
"title": "Multidimensional Quality Metrics (MQM) : A Framework for Declaring and Describing Translation Quality Metrics",
"year": "2014"
},
{
"authors": "Nitika Mathur; Timothy Baldwin; Trevor Cohn",
"journal": "Association for Computational Linguistics",
"ref_id": "b21",
"title": "Tangled up in BLEU: Reevaluating the evaluation of automatic machine translation evaluation metrics",
"year": "2020"
},
{
"authors": "Kishore Papineni; Salim Roukos; Todd Ward; Wei-Jing Zhu",
"journal": "Association for Computational Linguistics",
"ref_id": "b22",
"title": "Bleu: a method for automatic evaluation of machine translation",
"year": "2002"
},
{
"authors": "Maja Popović",
"journal": "Association for Computational Linguistics",
"ref_id": "b23",
"title": "chrF: character n-gram F-score for automatic MT evaluation",
"year": "2015"
},
{
"authors": "Tharindu Ranasinghe; Constantin Orasan; Ruslan Mitkov",
"journal": "International Committee on Computational Linguistics",
"ref_id": "b24",
"title": "TransQuest: Translation Quality Estimation with Cross-lingual Transformers",
"year": "2020"
},
{
"authors": "Vikas Raunak; Matt Post; Arul Menezes",
"journal": "",
"ref_id": "b25",
"title": "Salted: A framework for salient long-tail translation error detection",
"year": "2022"
},
{
"authors": "Ricardo Rei; G C José; Duarte De Souza; Chrysoula Alves; Ana C Zerva; Taisiya Farinha; Alon Glushkova; Luisa Lavie; Coheur; F T André; Martins",
"journal": "Association for Computational Linguistics",
"ref_id": "b26",
"title": "COMET-22: Unbabel-IST 2022 Submission for the Metrics Shared Task",
"year": "2022"
},
{
"authors": "Ricardo Rei; Craig Stewart; Ana C Farinha; Alon Lavie",
"journal": "Association for Computational Linguistics",
"ref_id": "b27",
"title": "COMET: A neural framework for MT evaluation",
"year": "2020"
},
{
"authors": "Ricardo Rei; Marcos Treviso; M Nuno; Chrysoula Guerreiro; Ana C Zerva; Christine Farinha; Maroti; G C José; Taisiya De Souza; Duarte Glushkova; Luisa Alves; Alon Coheur; Lavie; F T André; Martins",
"journal": "",
"ref_id": "b28",
"title": "CometKiwi: IST-Unbabel 2022 Submission for the Quality Estimation Shared Task",
"year": "2022"
},
{
"authors": "Marco Ribeiro; Sameer Singh; Carlos Guestrin",
"journal": "",
"ref_id": "b29",
"title": "Explaining the predictions of any classifier",
"year": "2016"
},
{
"authors": "B Ananya; Sai; Tanay Vignesh Nagarajan; Raj Dixit; Anoop Dabre; Pratyush Kunchukuttan; Mitesh M Kumar; Khapra",
"journal": "",
"ref_id": "b30",
"title": "IndicMT Eval: A Dataset to Meta-Evaluate Machine Translation metrics for Indian Languages",
"year": "2022"
},
{
"authors": "Thibault Sellam; Dipanjan Das; Ankur Parikh",
"journal": "Association for Computational Linguistics",
"ref_id": "b31",
"title": "BLEURT: Learning robust metrics for text generation",
"year": "2020"
},
{
"authors": "Avanti Shrikumar; Peyton Greenside; Anshul Kundaje",
"journal": "",
"ref_id": "b32",
"title": "Learning Important Features Through Propagating Activation Differences",
"year": "2017"
},
{
"authors": " Pmlr",
"journal": "",
"ref_id": "b33",
"title": "",
"year": ""
},
{
"authors": "Aaron Smith; Christian Hardmeier; Joerg Tiedemann",
"journal": "",
"ref_id": "b34",
"title": "Climbing mont BLEU: The strange world of reachable high-BLEU translations",
"year": "2016"
},
{
"authors": "Mukund Sundararajan; Ankur Taly; Qiqi Yan",
"journal": "",
"ref_id": "b35",
"title": "Axiomatic Attribution for Deep Networks",
"year": "2017"
},
{
"authors": " Pmlr",
"journal": "",
"ref_id": "b36",
"title": "",
"year": ""
},
{
"authors": "Shimin Tao; Su Chang; Ma Miaomiao; Hao Yang; Xiang Geng; Shujian Huang; Min Zhang; Jiaxin Guo; Minghan Wang; Yinglu Li",
"journal": "Association for Computational Linguistics",
"ref_id": "b37",
"title": "CrossQE: HW-TSC 2022 Submission for the Quality Estimation Shared Task",
"year": "2022"
},
{
"authors": "Marcos Treviso; M Nuno; Ricardo Guerreiro; Rei; F T André; Martins",
"journal": "Association for Computational Linguistics",
"ref_id": "b38",
"title": "IST-unbabel 2021 submission for the explainable quality estimation shared task",
"year": "2021"
},
{
"authors": "Yu Wan; Dayiheng Liu; Baosong Yang; Tianchi Bi; Haibo Zhang; Boxing Chen; Weihua Luo; Derek F Wong; Lidia S Chao",
"journal": "Association for Computational Linguistics",
"ref_id": "b39",
"title": "RoBLEURT submission for WMT2021 metrics task",
"year": "2021"
},
{
"authors": "Yu Wan; Dayiheng Liu; Baosong Yang; Haibo Zhang; Boxing Chen; Derek Wong; Lidia Chao",
"journal": "Association for Computational Linguistics",
"ref_id": "b40",
"title": "UniTE: Unified translation evaluation",
"year": "2022"
},
{
"authors": "Kerem Zaman; Yonatan Belinkov",
"journal": "",
"ref_id": "b41",
"title": "A Multilingual Perspective Towards the Evaluation of Attribution Methods in Natural Language Inference",
"year": "2022"
},
{
"authors": "Chrysoula Zerva; Frédéric Blain; Ricardo Rei; Piyawat Lertvittayakumjorn; G C José; Steffen Souza; Diptesh Eger; Duarte Kanojia; Constantin Alves; Marina Orȃsan; Fomicheva; F T André; Lucia Martins; Specia",
"journal": "Association for Computational Linguistics",
"ref_id": "b42",
"title": "Findings of the WMT 2022 Shared Task on Quality Estimation",
"year": "2022"
}
] | [
{
"formula_coordinates": [
3,
118.64,
75.37,
361.07,
62.86
],
"formula_id": "formula_0",
"formula_text": "METRIC EXPLAINABILITY en→de zh→en en→ru Avg. METHOD AUC R@K AUC R@K AUC R@K AUC R@K src-only evaluation UNITE SRC"
}
] |
|
10.18653/v1/2020.acl-main.417 | 2023-05-19 | [
{
"figure_ref": [],
"heading": "INTRODUCTION",
"publication_ref": [
"b14",
"b1",
"b65",
"b35",
"b12",
"b59",
"b23",
"b60",
"b59",
"b24",
"b19",
"b29",
"b15",
"b18",
"b57",
"b9",
"b67",
"b67"
],
"table_ref": [],
"text": "Self-training (Fralick, 1967;Amini et al., 2022) is a popular semi-supervised technique used to boost the performance of neural machine translation (NMT) models. In self-training for NMT, also known as forward-translation, an initial model is used to translate monolingual data; this data is then concatenated with the original training data in a subsequent training step (Zhang & Zong, 2016;Marie et al., 2020;Edunov et al., 2020;Wang et al., 2021). Self-training is believed to be effective through inducing input smoothness and leading to better learning of decision boundaries from the addition of unlabeled data (Chapelle et al., 2006;He et al., 2020;Wei et al., 2021). It has also been observed to effectively diversify the training distribution (Wang et al., 2021;Nguyen et al., 2020).\nA closely related technique is that of knowledge distillation (Hinton et al., 2015;Gou et al., 2021), particularly sequence-level knowledge distillation (SKD), which uses hard targets in training and reduces to pseudo-labeled data augmentation (Kim & Rush, 2016). In NMT, knowledge distillation is effective through knowledge transfer from ensembles or larger-capacity models and as a data augmentation method (Freitag et al., 2017;Gordon & Duh, 2019;Tan et al., 2019;Currey et al., 2020). In non-autoregressive translation, Zhou et al. (2020) explored the effect of SKD on training data complexity and showed that simpler training data from distillation is crucial for the performance of non-autoregressive MT models. This paper examines the component that is common to these techniques, the introduction of pseudolabeled training (PLT) data. We focus on the more common autoregressive NMT formulation and show that in addition to the known quality gains, PLT has a large impact on model brittleness in that it increases smoothness as well as stability across model re-training. Our main contributions are:\n• We focus on a set of stability properties in NMT models, which we unify under the umbrella term inertia, and show that PLT increases model inertia. We further show that both the quality gains and the improved inertia are not properties of any one specific technique such as self-training or knowledge distillation, but are common to the use of pseudo-labeled data in training. • We investigate the hypothesis that the observed properties correlate with a training data simplification mechanism, similarly to the observations made in Zhou et al. (2020). We compare with other popular semi-supervised techniques to investigate if the model quality and inertia properties hold when distribution simplification effects are not present.\n• Based on our findings, we recommend incorporating PLT into NMT training whenever inertia (e.g., stability to input perturbations and across incremental model updates) is important, as it increases inertia without sacrificing quality."
},
{
"figure_ref": [],
"heading": "RELATED WORK",
"publication_ref": [
"b17",
"b56",
"b26",
"b4",
"b33",
"b41",
"b13",
"b4",
"b32",
"b46",
"b5",
"b48",
"b34",
"b54",
"b61",
"b63",
"b7",
"b59",
"b23",
"b60",
"b64",
"b42",
"b67",
"b62",
"b21",
"b67"
],
"table_ref": [],
"text": "Neural network models are known to be sensitive to input variations, i.e., lacking in smoothness. This can make them brittle or open to adversarial attacks, a property observed across many application domains (Goodfellow et al., 2014;Szegedy et al., 2014;Jia & Liang, 2017). Neural machine translation models are similarly prone to robustness issues and can be affected by both synthetic and natural noise, leading to lower translation quality (Belinkov & Bisk, 2018;Li et al., 2019;Niu et al., 2020;Fadaee & Monz, 2020). In MT, earlier works have found noisy data augmentation (Belinkov & Bisk, 2018) and subword regularization (Kudo, 2018;Provilkov et al., 2020) to be among the most simple yet effective methods for addressing instability to input perturbations.\nIn addition to smoothness, neural models are known be sensitive to the various sources of randomness in training, such as initialization or dropout (Bengio, 2012;Reimers & Gurevych, 2017;Madhyastha & Jain, 2019). This instability negatively impacts end-users in the form of spurious differences in outputs between model updates, or more acutely, as quality regressions on specific data points, also known as negative flips (Shen et al., 2020;Xie et al., 2021;Yan et al., 2021). In NLP, Cai et al. (2022) focus on a set of structured prediction tasks and show that when random initialization changes, up to 30% of all errors can be regression errors, and that improved accuracy does not always mean reduced regressions. While negative flips are more difficult to measure in MT as multiple translations can be valid, the lack of consistency across re-training is a known problem: in our experiments ∼80% of the translations change due to different model random initialization alone. Despite this, to the best of our knowledge, minimizing regressions or improving stability across incremental model updates or re-trainings has not yet been addressed in MT.\nThis paper examines pseudo-label training in NMT and its effect on stability to both input variations and incremental model updates, which we group under the term inertia. Earlier work on pseudolabel training in MT focused on measuring quality alone and did not shed light on stability-related properties (Wang et al., 2021;He et al., 2020;Wei et al., 2021;Yuan et al., 2020). In terms of stability to input variations, or smoothness, our findings are related to the work of Papernot et al. (2015), where authors introduce defensive distillation and show that (self-)distillation increased smoothness when tested on digit and object recognition tasks. They show that the effect is one of reducing the amplitude of the network gradients. Unlike our work, they do not test pseudo-label training, but soft-target distillation, where a student is trained using the prediction probabilities of a teacher.\nFinally, we hypothesize that PLT techniques are able to increase model inertia based on their distribution simplification properties. Earlier works have explored the distribution simplification property of PLT methods in terms of model performance. In non-autoregressive NMT, Zhou et al. (2020) and Xu et al. (2021) explored the effect of SKD on training data complexity and its correlation with model performance. As in previous work, they hypothesized that SKD alleviates the multiple modes problem, i.e., the existence of multiple alternative translations (Gu et al., 2018). Similarly to Zhou et al. (2020), we measure training data complexity when adding pseudo-labeled data and use the entropy of a conditional word-level alignment as a complexity metric."
},
{
"figure_ref": [],
"heading": "TRAINING WITH PSEUDO-LABELS IN NMT",
"publication_ref": [
"b24",
"b25",
"b12",
"b20",
"b23"
],
"table_ref": [],
"text": "Neural machine translation (NMT) We use the autoregressive formulation of NMT, where given parallel data containing source and target sequences, a model θ is learned using the following objective:\nL = - J j=1 |V | k=1 1{y j = k} × log p(y j = k|y <j , x, θ),(1)\nwhere x = [x 1 , .. \nL = -E x∼p(x) E y∼p(y|x) p(y|x) log p θ * (y|x)\nIn a second step we estimate the final student model θ, combining the supervised loss with a PL (pseudo-label) loss L + L P L , where:\nL PL = -E x∼p P L (x),y log p θ (y |x)\nIn this case the targets y are given by the teacher distribution p θ * and the samples are drawn from a second distribution, p P L , which varies in the experiments below.\nRelated techniques As discussed earlier, PLT is a common feature of several widely used techniques in NMT such as self-training (a.k.a. forward-translation) and sequence-level knowledge distillation. This paper opts for the term pseudo-label training (PLT) in order to avoid confusion with additional assumptions made by these techniques. Specifically:\n• PLT does not necessarily imply semi-supervision, as self-training does.\n• PLT is more specific than KD in that it is restricted to hard labels (as opposed to training on soft targets as in Hinton et al. 2015), but more generic as it does not assume model compression.\nAnother technique for introducing synthetic data is the use of back-translation (BT), where target segments are translated into source segments (Sennrich et al., 2016a;Hoang et al., 2018;Edunov et al., 2020). PLT does not include BT since the latter does not introduce synthetic targets or labels.\nLastly, note that self-training is closely related to entropy minimization (Grandvalet & Bengio, 2004), a semi-supervised technique that encourages high-confidence predictions on unlabeled data. When reducing this objective to its mode, it becomes identical to L PL above, also observed in He et al. (2020)."
},
{
"figure_ref": [],
"heading": "MODEL INERTIA",
"publication_ref": [],
"table_ref": [],
"text": "This section introduces a set of desired stability-related MT properties that we group under the term inertia. All our metrics are closed-box (based on user-observed model behaviour alone) and we investigate two types of model inertia: (1) robustness to input perturbations (or smoothness) and ( 2) stability across incremental model updates."
},
{
"figure_ref": [],
"heading": "INPUT SMOOTHNESS",
"publication_ref": [
"b4",
"b41"
],
"table_ref": [],
"text": "Robustness to input variations is important in MT models, which have been shown to be negatively affected by misspellings and other small variations in input (Belinkov & Bisk, 2018). Niu et al. (2020) introduced metrics that contrast translations of noisy input with those of their clean counterparts in order to disentangle robustness from generic quality changes. We evaluate model robustness and consistency to input changes following the definitions introduced in Niu et al. (2020): Robustness measures degradation in translation quality when small variations are present in the input, while Consistency is a reference-free metric for changes in translation output alone. Specifically:\nConsistency = H(BLEU(Y , Y ), BLEU(Y, Y )) Robustness = BLEURT(Y , Y ref ) -BLEURT(Y, Y ref )(2)\nwhere Y ref stands for reference translations, Y, Y are translations of a clean/noisy versions of the test set (e.g., one with introduced misspellings) and H(•, •) stands for the harmonic mean. In this paper, we expand these definitions to consider robustness not only to synthetic misspellings, but also to natural grammatical errors."
},
{
"figure_ref": [],
"heading": "STABILITY TO MODEL UPDATES",
"publication_ref": [
"b61",
"b7",
"b7",
"b61",
"b54",
"b63",
"b7",
"b61",
"b63",
"b31",
"b36",
"b7"
],
"table_ref": [],
"text": "Unlike smoothness metrics, stability metrics are functions of two models: an original one (e.g., one that is deployed and available to users) and an update of this model which implements an incremental change. We denote a model update as a pair (θ, D, A) i , (θ, D, A) i+1 , where θ are the model parameters obtained when training using data D and algorithm A. While many incremental updates are possible, in this work we keep the model size and architectures intact and vary the random parameter initialization in training, following Xie et al. (2021) and Cai et al. (2022). We define stability as a measure of similarity between model outputs, irrespective of quality changes, while regressions (negative flips) measure output changes that result in lower quality on a given input segment (Cai et al., 2022;Xie et al., 2021;Shen et al., 2020;Yan et al., 2021).\nSTABILITY Stability is measured as string similarity between the different outputs Y i , Y i+1 .\nWe use a symmetric BLEU-based metric, the harmonic mean between BLEU(Y i , Y i+1 ) and BLEU(Y i+1 , Y i ), where Y i and Y i+1 are translations obtained with models θ i and θ i+1 , respectively.\nNFR Similarly to earlier works (Cai et al., 2022;Xie et al., 2021;Yan et al., 2021), we measure regressions as Negative Flip Rate, the number of sentences for which the translation degrades between model updates over the total number of segments. We consider degradations in terms of both overall quality and a targeted translation error category. Unlike other tasks, NMT lacks a reliable automatic segment-level quality metric (Kocmi et al., 2021;Mathur et al., 2020); we use human evaluations for this reason. Having an additional targeted error category allows us to measure segment-level regression automatically. In this work, we adopt gender translation accuracy as the targeted error category.\nNFI Following Cai et al. (2022), we also measure regressions in terms of Negative Flip Impact. NFI is defined as the proportion of negative flips to the total number of errors made by the new model. Note that in NMT, error is less well-defined for quality since it is not a categorical concept. This is not the case with targeted translation error categories."
},
{
"figure_ref": [],
"heading": "EXPERIMENTS",
"publication_ref": [
"b58",
"b10",
"b30",
"b37",
"b30",
"b28",
"b10",
"b0",
"b43",
"b51",
"b0",
"b35",
"b65",
"b18",
"b23"
],
"table_ref": [],
"text": "We perform experiments across 6 language pairs (LPs): English (en)↔German (de), Russian (ru), and Japanese (ja). We adapt the Transformer-base architecture (Vaswani et al., 2017) to 20 encoder layers and 2 decoder layers (denoted 20:2) as recommended by Domhan et al. (2020) and SSRU decoder layers for faster decoding (Kim et al., 2019). The deep-encoder-shallow-decoder configuration is widely used (Miceli Barone et al., 2017;Kim et al., 2019;Kasai et al., 2021), and the 20:2 model was found by Domhan et al. (2020) to yield comparable quality to the 6:6 and 10:10 models while significantly decreasing latency. Unless otherwise noted, we use beam decoding with a beam size of 5 (further details in Appendix A).\nExperiments are carried out with the WMT21 dataset (Akhbardeh et al., 2021). For en↔de we use 286M parallel segments, for en↔ja we use 17.2M parallel segments, and for en↔ru we use 34M parallel segments. For development, we use WMT newstest datasets from earlier years (see Appendix B for more details on datasets used). We evaluate quality using BLEU1 (Papineni et al., 2002) and BLEURT (Sellam et al., 2020) on the WMT21 newstest sets (Akhbardeh et al., 2021). We use only source-original test sets in order to avoid misestimating model performance due to translationese input (Marie et al., 2020).\nWe train PLT-augmented models using a mix of the original training data and pseudo-labeled data in a joint training setting following Zhang & Zong (2016); Gordon & Duh (2019). Based on recommendations by He et al. (2020), we use dropout for all the models, set to 0.1. We do not tune the trade-off between the two losses L and L PL (we use an equal amount of original and PLT data) or the number of incremental applications of the PLT augmentation. "
},
{
"figure_ref": [],
"heading": "SRC",
"publication_ref": [],
"table_ref": [],
"text": "Can yo put cites on those? BASELINE Können Sie Zitate darauf setzen? PLT(TRAIN) Kannst du diese zitieren? Table 1: Example translations from BASELINE and PLT(TRAIN) on the synthetic misspellings and GMEG test sets. In the first example (synthetic misspelling), the baseline invents the word Guven as a translation of the original miss-spelled word, guven(given). PLT translates the second example (English learner error) as Can you cite these? using the informal register, while the Baseline translates it literally as Can you put citations on these? (formal register)."
},
{
"figure_ref": [],
"heading": "QUALITY AND INERTIA USING PSEUDO-LABELED DATA",
"publication_ref": [],
"table_ref": [],
"text": "This section evaluates PLT for both generic model quality and for inertia. Unless otherwise noted, student models share the same architecture as the teacher and are trained using the same parallel data with the addition of pseudo-labeled data. PLT can be implemented by sampling and labeling data from different source distributions p P L : the original training data (as in KD) or unseen monolingual data (i.e. semi-supervised). This section tests both: to that end, teacher models are trained on half of the available parallel data, while the other half is reserved as a source of unlabeled monolingual data. Specifically, we compare:\n• BASELINE: Model trained on half the available data without any data augmentation.\n• PLT(TRAIN): Data used in PLT augmentation is sampled from the training data.\n• PLT(UL): Data used in PLT augmentation is sampled from unused parallel data.\n• ALLDATA: Finally, to account for the differences in training data size, we also compare against a model trained on all available parallel data without any PLT."
},
{
"figure_ref": [],
"heading": "INPUT SMOOTHNESS",
"publication_ref": [
"b41",
"b27",
"b39",
"b41",
"b23",
"b60"
],
"table_ref": [
"tab_1"
],
"text": "For each of these models we compute newstest quality (BLEU score) as well as model smoothness (robustness and consistency). We measure robustness and consistency as defined in Section 4 with the following sources of input variations:\n• Synthetic misspellings: We introduce misspellings as proposed by Niu et al. (2020) into the newstest set. Each word is misspelled with probability of 0.1, and the strategy is randomly chosen from single-character deletion, insertion, and substitution (Karpukhin et al., 2019). • GMEG: The GMEG corpus (Napoles et al., 2019) contains data with natural errors made by English language learners (grammatical misuse, misspellings, etc.). We compute consistency using the noisy input and a reference correction made by a professional annotator. We report the average consistency over the four provided reference corrections.2 \nExample translations and results are show in Tables 1 and2, respectively. Across all LPs, translation quality improves when pseudo-labeled data is used in training, irrespective of the source of the data added. However, sampling from unseen data does not bring additional improvements over using seen data for PLT. Similarly, using all parallel data vs. only half is not beneficial across the board, suggesting limitations of the training data w.r.t. the test domain.\nPLT shows significantly higher model consistency on both synthetic misspellings and the GMEG test sets. 3 Unlike Niu et al. (2020), however, we find that robustness scores (translation quality changes relative to input changes) are not as well correlated with consistency scores, suggesting that while translations are more stable under noisy conditions they may not necessarily be better. In the context ) is the percent of outputs that stay identical across the two models. For Distillation (Distil.), the second model is trained to mimic the first model.\nof semi-supervised learning, it has been hypothesized that self-training has the effect of making models smoother through the addition of new data (He et al., 2020;Wei et al., 2021). Our results suggest that this is not necessarily the case, as smoothness results are similar irrespective of the use of new unlabeled (monolingual) data (i.e., PLT(TRAIN) and PLT(UL) have similar smoothness)."
},
{
"figure_ref": [],
"heading": "STABILITY TO MODEL UPDATES",
"publication_ref": [
"b63",
"b7",
"b21",
"b7"
],
"table_ref": [
"tab_2",
"tab_4"
],
"text": "Next we investigate stability properties with respect to model updates when PLT is used in training. We fix the source of the pseudo-labeled data to be the training data (i.e., we consider only PLT(TRAIN)) and compare translation changes when re-training a model. Recall, a model update consists of a pair (θ, D, A) 1 , (θ, D, A) 2 , where θ are the model parameters obtained when training using data D and algorithm A. In these experiments, we keep the network architecture identical and hold A 1 = A 2 , modulo the random seed used in initialization. We contrast several settings:\n• BASELINE: Models are trained and re-trained with half the original data (D 1 = D 2 ), and no pseudo-labeled data is used. As above, we also evaluate the case where all of the original data is used (ALLDATA). We vary the random seed, leading to θ 1 = θ 2 . • PLT-δ(STUDENT): This tests the hypothesis that using PLT leads to more stable models that behave similarly when varying minor training conditions. We consider an identical setup as the baseline (D 1 = D 2 ), except that the data is augmented to contain PLT data • PLT-δ(TEACHER): In this setting, the two models θ 1 and θ 2 use PLT data in training; however, two different teachers are used to create it (the teachers are trained with different random seeds). This simulates a realistic setting where the teachers used to create pseudo-labels are not likely to stay constant. Note however that this is not a direct comparison to the baseline and PLT methods: the models do not vary in random seed alone, but also the contents of the training data (D 1 = D 2 ). This setting is a standard distillation approach for minimizing regressions, where θ 2 is trained to explicitly mimic θ 1 's predictions (Yan et al., 2021;Cai et al., 2022).\nStability and regression metrics are averaged over (θ 1 , θ 2 ) and (θ 2 , θ 1 ) scores since random initialization changes are not directional model updates. Tables 3 and4 show stability and regression metrics respectively (regression results are discussed in the next section).\nFirst, we observe that a striking number of translations change when changing random initialization: only 15% of outputs remain identical for en→de, and 8% and 2% remain identical for the lowerresource en→ru and en→ja pairs respectively. Doubling the amount of training data (ALLDATA) improves stability, but not by a large margin. Across all LPs tested, PLT improves stability relative to the baseline models and nearly doubles the percentage of segments translated identically. Interestingly, PLT also improves stability relative to the system trained on all available parallel data, once again indicating that inertia effects do not simply stem from more data. This result is particularly surprising for the PLT-δ(TEACHER) setting: unlike the baseline, the two models compared are trained on different data on the target side, yet their outputs are more similar to each other than the baseline outputs are to each other. This suggests that the high translation variability of the original data (a.k.a. multiple modes in Gu et al., 2018) is an issue with auto-regressive MT as well, and that pseudo-labeled data alleviates it even when created with different models.\nFinally, we also find that distillation, where a new model is explicitly trained to mimic the previous model, increases stability between teacher and student, confirming earlier observations on text classification Cai et al. (2022). However, this improvement is modest in our experiments."
},
{
"figure_ref": [],
"heading": "NEGATIVE FLIPS",
"publication_ref": [
"b55"
],
"table_ref": [],
"text": "Next, we assess PLT in terms of negative flips (NFs) as described in Section 4. We evaluate regressions in terms of overall quality (human evaluations on the WMT21 newstest set) and on a targeted error category (gender translation accuracy). For human evaluations, we used two professional annotators who assigned scores on a scale of 1 to 6 with 0.2 increments, where 6 indicates a perfect translation. A NF is defined as both annotators agreeing that there is a degradation. Since quality is evaluated on a scale, and not as a binary score, the concept of NFI is ambiguous. We therefore compute negative flip rate (NFR) alone. We evaluate on en→de,ja,ru due to availability of annotators.\nFor gender translation accuracy, which aggregates categorical measurements, we evaluate both NFR and NFI. We use the WinoMT benchmark (Stanovsky et al., 2019), a gender accuracy benchmark with a reliable segment-level metric suitable for automatic measurements of negative flips. The dataset consists of English source segments containing a profession whose gender is ambiguous at the lexical level but disambiguated in the sentential context, along with an automatic morphologybased classifier that evaluates the gender accuracy when translated. We evaluate on the two of our language pairs that are covered by WinoMT, en→de and en→ru. Table 5: PLT models using teachers of varying quality (averages over the three language pairs in each direction). We find that teacher quality correlates with quality of PLT; however, weaker teachers can still improve student quality. In terms of inertia properties, these are preserved regardless of teacher quality. Note X→en averages for robustness and consistency to misspelling include only de,ru→en."
},
{
"figure_ref": [],
"heading": "Results are shown in",
"publication_ref": [],
"table_ref": [],
"text": "models. For quality, this is most pronounced for en→de,ru (∼50%-100% relative NFR reduction).\nIn contrast, the effect of distillation is not consistent across the language pairs or the two test sets."
},
{
"figure_ref": [],
"heading": "TEACHER QUALITY",
"publication_ref": [
"b16",
"b22",
"b64"
],
"table_ref": [],
"text": "In previous sections, we found that quality and model inertia improved when using PLT regardless of the source of the data. In this section, we examine another dimension which distinguishes different flavors of PLT, namely teacher quality. Stronger teachers (teachers with larger capacity than the student) are more common in KD applications whereas identical teacher and student models are the norm in self-training/forward translation. Specifically, we vary the base 20:2 teacher architecture by decreasing the number of decoder layers to 1 (weaker teacher) and increasing it to 4 (stronger teacher). We keep the student architecture identical at 20:2 layers and fix the source of pseudolabeled data to the training set (referred to as PLT(TRAIN) in earlier sections).\nInterestingly, we find that teacher quality does not play a large role in model stability (Table 5).\nThere are small improvements in stability and robustness when stronger teachers are used, but gains are in range for all teacher models considered, even for weak teachers. Stronger teachers, however, are responsible for better performing student models. Most surprisingly, we found quality improvements over the baseline even when the teacher is of worse quality than the baseline model. This corroborates other work suggesting that the mechanism behind PLT is not simply that of compressing better performing (sets of) models (Furlanello et al., 2018;Hahn & Choi, 2019;Yuan et al., 2020)."
},
{
"figure_ref": [],
"heading": "DISTRIBUTION SIMPLIFICATION",
"publication_ref": [
"b46",
"b67",
"b67",
"b11",
"b67",
"b41"
],
"table_ref": [
"tab_6"
],
"text": "The previous section showed that PLT increases both quality and model inertia under different monolingual data and teacher quality settings. We hypothesize that the increased inertia observed is correlated with a distribution simplification mechanism: PLT leads to simpler training data, resulting in models that are less brittle. We test this by comparing PLT with other techniques used to improve quality and smoothness, but that may not have a distribution simplification effect. Below, we fix the source of pseudo-labeled data to the training data and test:\n• BT: Back-translation, a commonly used semi-supervised method that adds parallel data obtained through translation of target data with a reverse-direction MT model. • BPE-DROPOUT: A regularization method that has been shown to improve robustness to noise (Provilkov et al., 2020). We used a dropout rate of 0.1 as recommended by the authors. • PLT(SAMPLE): A variant of PLT where we vary the decoding strategy and perform sampling decoding which leads to more complex data and weaker student models (Zhou et al., 2020). Specifically, we sampled the top-8 hypotheses.\nIn previous work, Zhou et al. (2020) proposed a conditional entropy measure of training data complexity and showed that non-autoregressive translation performance is dependent on simpler training data distributions, such as those obtained with SKD. Here, we use the same entropy-based measure.\nFor each setting, we ( 1) compute an alignment model on the training data using fast align (Dyer et al., 2013) \nC(d) = -1 |Vx| x∈Vx E y|x align log(y|x)\n, where y is the sequence of training data tokens and x align the sequence of source-side tokens that y tokens are aligned to. Lower entropy indicates that the data is explained by a simpler word-to-word translation model that uses similar word translations irrespective of context.\nResults are shown in Table 6. First, we observe that the complexity scores confirm the results reported by Zhou et al. (2020), with smaller-scale differences due to the fact that we mix both original data and pseudo-labeled data. BPE-DROPOUT performs best on smoothness w.r.t. synthetic noise: it outperforms all methods by a large margin on robustness, and by a smaller margin on consistency. This is not the case on data with natural noise (GMEG), where the increased consistency effect is smaller w.r.t. the BASELINE model. On other metrics, BPE-DROPOUT has no effect on quality (BLEURT) and a minor negative effect on stability across re-training. BPE-DROPOUT is not only the only method that lowers stability, but also the only method that increases the complexity of the data compared to the baseline.\nBT shows a data simplification effect, mirrored by increased stability when re-training. However, BT has a detrimental effect on robustness and consistency metrics. These results indicate that while back-translation and forward translation are typically seen as very similar methods, they have different properties. PLT(SAMPLE) performs very similarly to PLT(TRAIN): when compared with PLT(TRAIN), it leads to slightly more complex data, and slightly worse quality and inertia scores. PLT(TRAIN) shows the lowest complexity scores and the highest stability.\nWhile stability and complexity correlate, not all methods that simplify the data improve smoothness; conversely, smoothness to synthetic noise can be improved significantly with complementary methods such as BPE-DROPOUT. We corroborate Niu et al. (2020) and find that synthetic and natural noise are different in nature and not all methods are equally effective on both types of noise."
},
{
"figure_ref": [],
"heading": "CONCLUSION",
"publication_ref": [
"b59",
"b49",
"b2",
"b50",
"b6",
"b68",
"b38",
"b47"
],
"table_ref": [],
"text": "This paper investigates pseudo-label training, a technique common to a number of methods for boosting NMT performance. We show that in addition to well-studied gains in generic translation quality, pseudo-label training induces several desirable stability-related properties, which we group under the term inertia. Empirically, these improvements are not tied to the use of unlabeled data (as in self-training) or the use of stronger teacher models (as in knowledge distillation) but are a consequence of the use of pseudo-labeled data itself. When compared with other methods designed to improve robustness in NMT, we observed that the effect on stability over re-training occurs only for those methods that simplify the training data. Based on these findings, we recommend using PLT with unlabeled data (à la self-training) when developing NMT models where inertia is important due to its benefits to model inertia and its use in addressing potential language coverage bias (Wang et al., 2021). In future work, we plan to investigate the interplay between PLT and different formulations of NMT (auto-vs. non-autoregressive MT) as well as potential negative side effects such as bias amplification (Renduchintala et al., 2021). Finally, developing automatic metrics to detect negative flips in NMT is an important task that has yet to be examined extensively and can help guide PLT techniques.\nTable 7: We trained our models on a subset of datasets from the WMT21 news task. Specifically, we used Paracrawl v9 (Bañón et al., 2020), WikiMatrix (Schwenk et al., 2021), WikiTitles (Bojar et al., 2018), news commentary, UN v1.0 dataset (Ziemski et al., 2016), JParaCrawl (Morishita et al., 2020) and the Japanese-English subtitles datasets (Pryzant et al., 2018)."
},
{
"figure_ref": [],
"heading": "LP",
"publication_ref": [],
"table_ref": [],
"text": "Years # parallel en↔de 2017-2020 9k en↔ja 2020 2k en↔ru 2017-2020 9k Table 10: PLT models using teachers of varying quality as measured by BERTScore (averages over the three LPs in each direction). We find that teacher quality correlates with quality of PLT; however, weaker teachers can still improve student quality. In terms of inertia properties, these are preserved regardless of teacher quality. Note X→en averages for robustness and consistency to misspelling include only de,ru→en. "
},
{
"figure_ref": [],
"heading": "ACKNOWLEDGEMENTS",
"publication_ref": [],
"table_ref": [],
"text": "We thank Yi Zhang and Elman Mansimov for discussions on negative flips and Miguel Ballesteros, Surafel Lakew, Cuong Hoang, and anonymous reviewers for their comments and suggestions."
},
{
"figure_ref": [],
"heading": "C TRAINING CURVES",
"publication_ref": [],
"table_ref": [],
"text": "Here, we compare pseudo-label training with back-translation (BT). We find that pseudo-label training regularizes the models by controlling for over fitting. BT also regularizes the model, but it does not simplify the distribution to the extent PLT does, implying that controlling over-fitting is not a main factor for stability. Comparisons with other methods (i.e. BPE-dropout and PLT(sample)) show similar trends.\nFigure 1: Comparisons of PLT(train) validation (solid lines) and training curves (dashed lines) against back-translation and baseline models. We find that in comparison, PLT is able to control over fitting on the training data. Back-translation also regularizes the model, but it does not simplify the distribution to the extent PLT does, implying that controlling over-fitting is not a main factor for stability."
},
{
"figure_ref": [],
"heading": "D BERTSCORE",
"publication_ref": [
"b66"
],
"table_ref": [],
"text": "We provide quality scores using BERTScore (Zhang et al., 2020). In terms of generic quality, PLT provides improvements in quality consistent with earlier results using BLEU and BLEURT metrics (see Tables 2, 5, and5). We also computed robustness metrics BERTScore:\nwhere Y ref stands for reference translations and Y, Y are translations of a clean/noisy versions of the test set (e.g., one with introduced misspellings)."
}
] | Like many other machine learning applications, neural machine translation (NMT) benefits from over-parameterized deep neural models. However, these models have been observed to be brittle: NMT model predictions are sensitive to small input changes and can show significant variation across re-training or incremental model updates. This work studies a frequently used method in NMT, pseudo-label training (PLT), which is common to the related techniques of forward-translation (or self-training) and sequence-level knowledge distillation. While the effect of PLT on quality is well-documented, we highlight a lesserknown effect: PLT can enhance a model's stability to model updates and input perturbations, a set of properties we call model inertia. We study inertia effects under different training settings and we identify distribution simplification as a mechanism behind the observed results. | PSEUDO-LABEL TRAINING AND MODEL INERTIA IN NEURAL MACHINE TRANSLATION | [
{
"figure_caption": "SRCThousands of people aree guven a drug and thousands of others are given a placebo.. BASELINE Tausende von Menschen erhalten Guven ein Medikament und Tausende von anderen erhalten ein Placebo. PLT(TRAIN) Tausende von Menschen erhalten eine Droge und Tausende von anderen erhalten ein Placebo.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "I and J are the source/target length, and |V | is the size of the vocabulary. Unless otherwise stated, we use beam search with a fixed number of hypotheses in order to generate a translation from this model. In this paper, we introduce the term pseudo-label training (PLT) to refer to the general technique of adding pseudo-labeled data during training, where the labels are obtained using a previously trained NMT model. Specifically, we consider two-step PLT. In a first stage we estimate a teacher model θ * trained with a supervised loss on samples drawn from p, the empirical distribution of the original training data:",
"figure_data": "Pseudo-label training (PLT)",
"figure_id": "tab_0",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Training data sizes and performance scores for PLT/Baseline models. Quality is measured with BLEU and BLEURT on the WMT21 newstest set. Smoothness is measured as robustness and consistency to synthetic (Misspellings) and natural (GMEG) noise. GMEG scores are computed as the average over four reference corrections. Robustness measures changes in translation quality w.r.t. input variations, while consistency measures translation changes alone.",
"figure_data": "MisspellingsGMEG",
"figure_id": "tab_1",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "tab_2",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Negative flip rate (NFR) and negative flip impact (NFI) on WMT21 (assessed by human annotators) and WinoMT (using the automatic gender translation accuracy metric).• DISTILLATION: In this setting D 2 is obtained from D 1 using pseudo-labeled data obtained with model θ 1 . The training data D 1 is re-translated and merged with the original D 1 to create D 2 .",
"figure_data": "",
"figure_id": "tab_4",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": "Published as a conference paper at ICLR 2023TeacherPLTLPArch. BLEU BLEURT BLEU BLEURT Stability Const(GMEG) Rob(Missp) Const(Missp)---25.12-0.13662.2077.57-0.80765.7720:124.08-0.18525.54-0.08372.6779.93-0.87669.72en→X20:225.12-0.13626.12-0.04374.1080.72-0.86870.3820:425.63-0.13626.27-0.04473.6880.87-0.88670.05---28.120.07661.91--0.84480.5320:126.74-0.04728.300.02973.15--0.82583.56X→en20:228.120.07629.560.10873.96--0.87583.8020:428.400.09529.940.11373.86--0.80883.99",
"figure_id": "tab_5",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": ", (2) use it to align a sample of the training corpus, and (3) compute the entropy of the Quality and model inertia with PLT versus other methods (averages over the three language pairs in each direction). Stability to model updates is computed w.r.t. to random seed variation in student models. X→en averages for robustness and consistency to misspellings involve de,ru→en.",
"figure_data": "LPSettingC(d) ↓ BLEU BLEURT Stability Const(GMEG) Rob(Missp) Const(Missp)BASELINE3.7425.12-0.13662.2077.57-0.80765.77BT3.6425.96-0.14065.3677.16-0.93464.83en→XBPE-DROPOUT3.9024.82-0.15461.7078.69-0.54771.86PLT(SAMPLE)3.5625.96-0.11971.9780.41-0.86569.87PLT(TRAIN)3.5426.12-0.11774.1080.73-0.86870.38BASELINE3.1228.120.07661.91--0.84480.53BT3.0128.880.09964.91--0.97980.00X→enBPE-DROPOUT3.4128.190.07561.73--0.59184.96PLT(SAMPLE)2.9829.410.10772.69--0.84983.70PLT(TRAIN)2.9429.560.10873.96--0.87583.80aligned data, leading to:",
"figure_id": "tab_6",
"figure_label": "6",
"figure_type": "table"
},
{
"figure_caption": "We used the WMT news test datasets from previous years as our development set. ±1.12 0.957 ±0.001 -0.004 ±0.000 87.0 ±0.7 -",
"figure_data": "MisspellingsGMEG",
"figure_id": "tab_7",
"figure_label": "8",
"figure_type": "table"
},
{
"figure_caption": "Training data sizes and performance scores for PLT/Baseline models. Quality is measured with BLEU and BERTScore on the WMT21 news test set. Smoothness is measured as robustness and consistency to synthetic (Misspellings) and natural (GMEG) noise. GMEG scores are computed as the average over four reference corrections. Robustness measures changes in translation quality w.r.t input variations, while Consistency measures translation changes alone.",
"figure_data": "TeacherPLTLPArch. BLEU BERTScore BLEU BERTScore Stability Const(GMEG) Rob(Missp) Const(Missp)---25.120.84562.2077.57-0.01865.7720:124.080.84525.540.85072.6779.93-0.01869.72en→X20:225.120.84926.120.85274.1080.72-0.01870.3820:425.630.84926.270.85273.6880.87-0.01870.05---28.120.93761.91--0.00980.5320:126.740.93228.300.93673.15--0.00883.56X→en20:228.120.93729.560.93973.96--0.00983.8020:428.400.93729.940.93773.86--0.00883.99",
"figure_id": "tab_8",
"figure_label": "9",
"figure_type": "table"
},
{
"figure_caption": "Performance and model inertia with PLT versus other methods (averages over the three LPs in each direction). Stability to model updates is computed w.r.t. to random seed variation in student models. X→en averages for robustness and consistency to misspellings involve de,ru→en.",
"figure_data": "LPSettingC(d) ↓ BLEU BERTScore Stability Const(GMEG) Rob(Missp) Const(Missp)BASELINE3.7425.120.84962.2077.57-0.01865.77BT3.6425.960.85165.3677.16-0.02164.83en→XBPE-DROPOUT3.9024.820.84761.7078.69-0.01171.86PLT(SAMPLE)3.5625.960.85271.9780.41-0.01869.87PLT(TRAIN)3.5426.120.85274.1080.73-0.01870.38BASELINE3.1228.120.93661.91--0.00980.53BT3.0128.880.93864.91--0.01080.00X→enBPE-DROPOUT3.4128.190.93761.73--0.00684.96PLT(SAMPLE)2.9829.410.93972.69--0.00983.70PLT(TRAIN)2.9429.560.93973.96--0.00983.80",
"figure_id": "tab_9",
"figure_label": "11",
"figure_type": "table"
}
] | Benjamin Hsu; Anna Currey; Xing Niu; Maria Nȃdejde; Georgiana Dinu | [
{
"authors": "Farhad Akhbardeh; Arkady Arkhangorodsky; Magdalena Biesialska; Ondřej Bojar; Rajen Chatterjee; Vishrav Chaudhary; Marta R Costa-Jussa; Cristina España-Bonet; Angela Fan; Christian Federmann; Markus Freitag; Yvette Graham; Roman Grundkiewicz; Barry Haddow; Leonie Harter; Kenneth Heafield; Christopher Homan; Matthias Huck; Kwabena Amponsah-Kaakyire; Jungo Kasai; Daniel Khashabi; Kevin Knight; Tom Kocmi; Philipp Koehn; Nicholas Lourie; Christof Monz; Makoto Morishita; Masaaki Nagata; Ajay Nagesh; Toshiaki Nakazawa; Matteo Negri; Santanu Pal; Auguste Allahsera; Marco Tapo; Valentin Turchi; Marcos Vydrin; Zampieri",
"journal": "Association for Computational Linguistics",
"ref_id": "b0",
"title": "Findings of the 2021 conference on machine translation (WMT21)",
"year": "2021-11"
},
{
"authors": "Massih-Reza Amini; Vasilii Feofanov; Loïc Pauletto; Emilie Devijver; Yury Maximov",
"journal": "",
"ref_id": "b1",
"title": "Selftraining: A survey",
"year": "2022"
},
{
"authors": "Marta Bañón; Pinzhen Chen; Barry Haddow; Kenneth Heafield; Hieu Hoang; Miquel Esplà-Gomis; Mikel L Forcada; Faheem Amir Kamran; Philipp Kirefu; Sergio Ortiz Koehn; Leopoldo Pla Rojas; Gema Sempere; Elsa Ramírez-Sánchez; Marek Sarrías; Brian Strelec; William Thompson; Dion Waites; Jaume Wiggins; Zaragoza",
"journal": "",
"ref_id": "b2",
"title": "ParaCrawl: Web-scale acquisition of parallel corpora",
"year": "2020-07"
},
{
"authors": "Loïc Barrault; Ondřej Bojar; Marta R Costa-Jussà; Christian Federmann; Mark Fishel; Yvette Graham; Barry Haddow; Matthias Huck; Philipp Koehn; Shervin Malmasi; Christof Monz; Mathias Müller; Santanu Pal; Matt Post; Marcos Zampieri",
"journal": "",
"ref_id": "b3",
"title": "Findings of the 2019 conference on machine translation (WMT19)",
"year": "2019"
},
{
"authors": "Yonatan Belinkov; Yonatan Bisk",
"journal": "",
"ref_id": "b4",
"title": "Synthetic and natural noise both break neural machine translation",
"year": "2018"
},
{
"authors": "Yoshua Bengio",
"journal": "",
"ref_id": "b5",
"title": "Practical recommendations for gradient-based training of deep architectures",
"year": "2012"
},
{
"authors": "Ondřej Bojar; Christian Federmann; Mark Fishel; Yvette Graham; Barry Haddow; Philipp Koehn; Christof Monz",
"journal": "",
"ref_id": "b6",
"title": "Findings of the 2018 conference on machine translation (WMT18)",
"year": "2018-10"
},
{
"authors": "Deng Cai; Elman Mansimov; Yi-An Lai; Yixuan Su; Lei Shu; Yi Zhang",
"journal": "",
"ref_id": "b7",
"title": "Measuring and reducing model update regression in structured prediction for NLP",
"year": "2022"
},
{
"authors": "",
"journal": "The MIT Press",
"ref_id": "b8",
"title": "Semi-Supervised Learning",
"year": "2006"
},
{
"authors": "Anna Currey; Prashant Mathur; Georgiana Dinu",
"journal": "",
"ref_id": "b9",
"title": "Distilling multiple domains for neural machine translation",
"year": "2020-11"
},
{
"authors": "Tobias Domhan; Michael Denkowski; David Vilar; Xing Niu; Felix Hieber; Kenneth Heafield",
"journal": "",
"ref_id": "b10",
"title": "The sockeye 2 neural machine translation toolkit at AMTA 2020",
"year": "2020-10"
},
{
"authors": "Chris Dyer; Victor Chahuneau; Noah A Smith",
"journal": "Association for Computational Linguistics",
"ref_id": "b11",
"title": "A simple, fast, and effective reparameterization of IBM model 2",
"year": "2013-06"
},
{
"authors": "Sergey Edunov; Myle Ott; Marc'aurelio Ranzato; Michael Auli",
"journal": "",
"ref_id": "b12",
"title": "On the evaluation of machine translation systems trained with back-translation",
"year": "2020-07"
},
{
"authors": "Marzieh Fadaee; Christof Monz",
"journal": "",
"ref_id": "b13",
"title": "The unreasonable volatility of neural machine translation models",
"year": "2020-07"
},
{
"authors": "C Stanley; Fralick",
"journal": "IEEE Trans. Inf. Theory",
"ref_id": "b14",
"title": "Learning to recognize patterns without a teacher",
"year": "1967"
},
{
"authors": "Markus Freitag; Yaser Al-Onaizan; Baskaran Sankaran",
"journal": "",
"ref_id": "b15",
"title": "Ensemble distillation for neural machine translation",
"year": "2017"
},
{
"authors": "Tommaso Furlanello; Zachary Chase Lipton; Michael Tschannen; Laurent Itti",
"journal": "PMLR",
"ref_id": "b16",
"title": "Anima Anandkumar. Born-again neural networks",
"year": "2018"
},
{
"authors": "Ian J Goodfellow; Jonathon Shlens; Christian Szegedy",
"journal": "",
"ref_id": "b17",
"title": "Explaining and harnessing adversarial examples",
"year": "2014"
},
{
"authors": "Mitchell A Gordon; Kevin Duh",
"journal": "",
"ref_id": "b18",
"title": "Explaining sequence-level knowledge distillation as dataaugmentation for neural machine translation",
"year": "2019"
},
{
"authors": "Jianping Gou; Baosheng Yu; Stephen J Maybank; Dacheng Tao",
"journal": "International Journal of Computer Vision",
"ref_id": "b19",
"title": "Knowledge distillation: A survey",
"year": "2021-03"
},
{
"authors": "Yves Grandvalet; Yoshua Bengio",
"journal": "MIT Press",
"ref_id": "b20",
"title": "Semi-supervised learning by entropy minimization",
"year": "2004"
},
{
"authors": "Jiatao Gu; James Bradbury; Caiming Xiong; O K Victor; Richard Li; Socher",
"journal": "",
"ref_id": "b21",
"title": "Non-autoregressive neural machine translation",
"year": "2018"
},
{
"authors": "Sangchul Hahn; Heeyoul Choi",
"journal": "",
"ref_id": "b22",
"title": "Self-knowledge distillation in natural language processing",
"year": "2019"
},
{
"authors": "Junxian He; Jiatao Gu; Jiajun Shen; Marc'aurelio Ranzato",
"journal": "",
"ref_id": "b23",
"title": "Revisiting self-training for neural sequence generation",
"year": "2020"
},
{
"authors": "Geoffrey Hinton; Oriol Vinyals; Jeffrey Dean",
"journal": "",
"ref_id": "b24",
"title": "Distilling the knowledge in a neural network",
"year": "2015"
},
{
"authors": "Duy Vu Cong; Philipp Hoang; Gholamreza Koehn; Trevor Haffari; Cohn",
"journal": "",
"ref_id": "b25",
"title": "Iterative backtranslation for neural machine translation",
"year": "2018-07"
},
{
"authors": "Robin Jia; Percy Liang",
"journal": "",
"ref_id": "b26",
"title": "Adversarial examples for evaluating reading comprehension systems",
"year": "2017-09"
},
{
"authors": "Vladimir Karpukhin; Omer Levy; Jacob Eisenstein; Marjan Ghazvininejad",
"journal": "",
"ref_id": "b27",
"title": "Training on synthetic noise improves robustness to natural noise in machine translation",
"year": "2019-11"
},
{
"authors": "Jungo Kasai; Nikolaos Pappas; Hao Peng; James Cross; Noah A Smith",
"journal": "",
"ref_id": "b28",
"title": "Deep encoder, shallow decoder: Reevaluating non-autoregressive machine translation",
"year": "2021"
},
{
"authors": "Yoon Kim; Alexander M Rush",
"journal": "",
"ref_id": "b29",
"title": "Sequence-level knowledge distillation",
"year": "2016-11"
},
{
"authors": "Jin Young; Marcin Kim; Hany Junczys-Dowmunt; Alham Hassan; Kenneth Fikri Aji; Roman Heafield; Nikolay Grundkiewicz; Bogoychev",
"journal": "",
"ref_id": "b30",
"title": "From research to production and back: Ludicrously fast neural machine translation",
"year": "2019-11"
},
{
"authors": "Tom Kocmi; Christian Federmann; Roman Grundkiewicz; Marcin Junczys-Dowmunt; Hitokazu Matsushita; Arul Menezes",
"journal": "",
"ref_id": "b31",
"title": "To ship or not to ship: An extensive evaluation of automatic metrics for machine translation",
"year": "2021-11"
},
{
"authors": "Taku Kudo",
"journal": "",
"ref_id": "b32",
"title": "Subword regularization: Improving neural network translation models with multiple subword candidates",
"year": "2018-07"
},
{
"authors": "Xian Li; Paul Michel; Antonios Anastasopoulos; Yonatan Belinkov; Nadir Durrani; Orhan Firat; Philipp Koehn; Graham Neubig; Juan Miguel Pino; Hassan Sajjad",
"journal": "",
"ref_id": "b33",
"title": "Findings of the first shared task on machine translation robustness",
"year": "2019"
},
{
"authors": "Pranava Madhyastha; Rishabh Jain",
"journal": "",
"ref_id": "b34",
"title": "On model stability as a function of random seed",
"year": "2019"
},
{
"authors": "Benjamin Marie; Raphael Rubino; Atsushi Fujita",
"journal": "",
"ref_id": "b35",
"title": "Tagged back-translation revisited: Why does it really work",
"year": "2020-07"
},
{
"authors": "Nitika Mathur; Johnny Wei; Markus Freitag; Qingsong Ma; Ondřej Bojar",
"journal": "",
"ref_id": "b36",
"title": "Results of the WMT20 metrics shared task",
"year": "2020-11"
},
{
"authors": "Antonio Valerio; Miceli Barone; Jindřich Helcl; Rico Sennrich; Barry Haddow; Alexandra Birch",
"journal": "",
"ref_id": "b37",
"title": "Deep architectures for neural machine translation",
"year": "2017-09"
},
{
"authors": "Makoto Morishita; Jun Suzuki; Masaaki Nagata",
"journal": "European Language Resources Association",
"ref_id": "b38",
"title": "JParaCrawl: A large scale web-based English-Japanese parallel corpus",
"year": "2020-05"
},
{
"authors": "Courtney Napoles; Maria Nȃdejde; Joel Tetreault",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b39",
"title": "Enabling robust grammatical error correction in new domains: Data sets, metrics, and analyses",
"year": "2019"
},
{
"authors": "Xuan-Phi Nguyen; Shafiq Joty; Kui Wu; Ai Ti; Aw ",
"journal": "",
"ref_id": "b40",
"title": "Data diversification: A simple strategy for neural machine translation",
"year": ""
},
{
"authors": "Xing Niu; Prashant Mathur; Georgiana Dinu; Yaser Al-Onaizan",
"journal": "",
"ref_id": "b41",
"title": "Evaluating robustness to input perturbations for neural machine translation",
"year": "2020-07"
},
{
"authors": "Nicolas Papernot; Patrick D Mcdaniel; Xi Wu; Somesh Jha; Ananthram Swami",
"journal": "",
"ref_id": "b42",
"title": "Distillation as a defense to adversarial perturbations against deep neural networks",
"year": "2015"
},
{
"authors": "Kishore Papineni; Salim Roukos; Todd Ward; Wei-Jing Zhu",
"journal": "",
"ref_id": "b43",
"title": "Bleu: a method for automatic evaluation of machine translation",
"year": "2002-07"
},
{
"authors": "Matt Post",
"journal": "",
"ref_id": "b44",
"title": "A call for clarity in reporting BLEU scores",
"year": "2018"
},
{
"authors": "Ofir Press; Lior Wolf",
"journal": "Association for Computational Linguistics",
"ref_id": "b45",
"title": "Using the output embedding to improve language models",
"year": "2017-04"
},
{
"authors": "Ivan Provilkov; Dmitrii Emelianenko; Elena Voita",
"journal": "",
"ref_id": "b46",
"title": "BPE-dropout: Simple and effective subword regularization",
"year": "2020-07"
},
{
"authors": "R Pryzant; Y Chung; D Jurafsky; D Britz",
"journal": "",
"ref_id": "b47",
"title": "JESC: Japanese-English Subtitle Corpus",
"year": "2018"
},
{
"authors": "Nils Reimers; Iryna Gurevych",
"journal": "",
"ref_id": "b48",
"title": "Reporting score distributions makes a difference: Performance study of lstm-networks for sequence tagging",
"year": "2017"
},
{
"authors": "Adithya Renduchintala; Denise Diaz; Kenneth Heafield; Xian Li; Mona Diab",
"journal": "",
"ref_id": "b49",
"title": "Gender bias amplification during speed-quality optimization in neural machine translation",
"year": "2021"
},
{
"authors": "Holger Schwenk; Vishrav Chaudhary; Shuo Sun; Hongyu Gong; Francisco Guzmán",
"journal": "",
"ref_id": "b50",
"title": "WikiMatrix: Mining 135M parallel sentences in 1620 language pairs from Wikipedia",
"year": "2021-04"
},
{
"authors": "Thibault Sellam; Dipanjan Das; Ankur Parikh",
"journal": "",
"ref_id": "b51",
"title": "BLEURT: Learning robust metrics for text generation",
"year": "2020-07"
},
{
"authors": "Rico Sennrich; Barry Haddow; Alexandra Birch",
"journal": "",
"ref_id": "b52",
"title": "Improving neural machine translation models with monolingual data",
"year": "2016"
},
{
"authors": "Rico Sennrich; Barry Haddow; Alexandra Birch",
"journal": "Association for Computational Linguistics",
"ref_id": "b53",
"title": "Neural machine translation of rare words with subword units",
"year": "2016-08"
},
{
"authors": "Yantao Shen; Yuanjun Xiong; Wei Xia; Stefano Soatto",
"journal": "IEEE",
"ref_id": "b54",
"title": "Towards backward-compatible representation learning",
"year": "2020"
},
{
"authors": "Gabriel Stanovsky; Noah A Smith; Luke Zettlemoyer",
"journal": "",
"ref_id": "b55",
"title": "Evaluating gender bias in machine translation",
"year": "2019-07"
},
{
"authors": "Christian Szegedy; Wojciech Zaremba; Ilya Sutskever; Joan Bruna; Dumitru Erhan; Ian Goodfellow; Rob Fergus",
"journal": "",
"ref_id": "b56",
"title": "Intriguing properties of neural networks",
"year": "2014"
},
{
"authors": "Xu Tan; Yi Ren; Di He; Tao Qin; Tie-Yan Liu",
"journal": "",
"ref_id": "b57",
"title": "Multilingual neural machine translation with knowledge distillation",
"year": "2019"
},
{
"authors": "Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Lukasz Kaiser; Illia Polosukhin",
"journal": "",
"ref_id": "b58",
"title": "Attention is all you need",
"year": "2017"
},
{
"authors": "Shuo Wang; Zhaopeng Tu; Zhixing Tan; Shuming Shi; Maosong Sun; Yang Liu",
"journal": "",
"ref_id": "b59",
"title": "On the language coverage bias for neural machine translation",
"year": "2021-08"
},
{
"authors": "Colin Wei; Kendrick Shen; Yining Chen; Tengyu Ma",
"journal": "",
"ref_id": "b60",
"title": "Theoretical analysis of self-training with deep networks on unlabeled data",
"year": "2021"
},
{
"authors": "Yuqing Xie; Yi-An Lai; Yuanjun Xiong; Yi Zhang; Stefano Soatto",
"journal": "",
"ref_id": "b61",
"title": "Regression bugs are in your model! measuring, reducing and analyzing regressions in NLP model updates",
"year": "2021-08"
},
{
"authors": "Weijia Xu; Shuming Ma; Dongdong Zhang; Marine Carpuat",
"journal": "",
"ref_id": "b62",
"title": "How does distilled data complexity impact the quality and confidence of non-autoregressive machine translation?",
"year": "2021-08"
},
{
"authors": "Sijie Yan; Yuanjun Xiong; Kaustav Kundu; Shuo Yang; Siqi Deng; Meng Wang; Wei Xia; Stefano Soatto",
"journal": "",
"ref_id": "b63",
"title": "Positive-congruent training: Towards regression-free model updates",
"year": "2021-06"
},
{
"authors": "Li Yuan; Francis E H Tay; Guilin Li; Tao Wang; Jiashi Feng",
"journal": "IEEE",
"ref_id": "b64",
"title": "Revisiting knowledge distillation via label smoothing regularization",
"year": "2020"
},
{
"authors": "Jiajun Zhang; Chengqing Zong",
"journal": "",
"ref_id": "b65",
"title": "Exploiting source-side monolingual data in neural machine translation",
"year": "2016-11"
},
{
"authors": "Tianyi Zhang; Varsha Kishore; Felix Wu; Kilian Q Weinberger; Yoav Artzi",
"journal": "",
"ref_id": "b66",
"title": "Bertscore: Evaluating text generation with bert",
"year": "2020"
},
{
"authors": "Chunting Zhou; Jiatao Gu; Graham Neubig",
"journal": "",
"ref_id": "b67",
"title": "Understanding knowledge distillation in nonautoregressive machine translation",
"year": "2020"
},
{
"authors": "Michał Ziemski; Marcin Junczys-Dowmunt; Bruno Pouliquen",
"journal": "",
"ref_id": "b68",
"title": "The United Nations parallel corpus v1.0",
"year": "2016-05"
},
{
"authors": "A Training Parameters",
"journal": "",
"ref_id": "b69",
"title": "All models used in our experiments utilized the following set of hyperparameters",
"year": "2016"
},
{
"authors": "",
"journal": "Press & Wolf",
"ref_id": "b70",
"title": "'cross-entropy",
"year": "2017"
},
{
"authors": "B Dataset",
"journal": "",
"ref_id": "b71",
"title": "We trained enen↔de models on Paracrawl",
"year": "2016"
},
{
"authors": "",
"journal": "WikiMatrix, WikiTitles, news commentary",
"ref_id": "b72",
"title": "LP Datasets # parallel en↔de Paracrawl v9",
"year": ""
},
{
"authors": "M En↔ja Jparacrawl V2; Wikititles Wikimatrix",
"journal": "",
"ref_id": "b73",
"title": "",
"year": ""
},
{
"authors": "Paracrawl ",
"journal": "UN",
"ref_id": "b74",
"title": "WikiMatrix WikiTitles, news commenatry",
"year": ""
}
] | [
{
"formula_coordinates": [
2,
199.58,
661.61,
304.43,
31.41
],
"formula_id": "formula_0",
"formula_text": "L = - J j=1 |V | k=1 1{y j = k} × log p(y j = k|y <j , x, θ),(1)"
},
{
"formula_coordinates": [
3,
216.53,
147.82,
178.94,
9.96
],
"formula_id": "formula_1",
"formula_text": "L = -E x∼p(x) E y∼p(y|x) p(y|x) log p θ * (y|x)"
},
{
"formula_coordinates": [
3,
233.99,
196.72,
144.02,
10.12
],
"formula_id": "formula_2",
"formula_text": "L PL = -E x∼p P L (x),y log p θ (y |x)"
},
{
"formula_coordinates": [
3,
175.67,
657.26,
328.33,
23.6
],
"formula_id": "formula_3",
"formula_text": "Consistency = H(BLEU(Y , Y ), BLEU(Y, Y )) Robustness = BLEURT(Y , Y ref ) -BLEURT(Y, Y ref )(2)"
},
{
"formula_coordinates": [
9,
207.05,
238.53,
160.3,
13.47
],
"formula_id": "formula_4",
"formula_text": "C(d) = -1 |Vx| x∈Vx E y|x align log(y|x)"
}
] |
10.1007/s10458-009-9103-z | 2023-05-19 | [
{
"figure_ref": [],
"heading": "INTRODUCTION",
"publication_ref": [
"b7",
"b22",
"b38"
],
"table_ref": [],
"text": "The framework of Decentralized Partially Observable Markov Decision Processes (Dec-POMDPs) allow modeling collaborative multi-agent systems, the objective being to equip them with individual policies that maximize some common performance criterion. However, solving Dec-POMDPs is challenging since the environment evolves according to all agent's actions, and each agent performs its action only based on its local action-observation histories. To ensure finding a global optimum, one thus needs to reason about all individual policies together. As a consequence of this interdependency, even for a finite-horizon Dec-POMDP, the solving process has been proven to be NEXP in the worst case [Bernstein et al., 2002], and solving an infinite-horizon Dec-POMDP is undecidable [Madani et al., 2003, Oliehoek andAmato, 2016]. Nair et al. propose an alternative approach called JESP (Joint Equilibrium-Based Search for Policies) [Nair et al., 2003], which avoids this interdependency in the solving process by searching for a Nash equilibrium, i.e., each agent's policy is a best response to other agents' policies. JESP operates an iterative optimization process over each agent. In each iteration, it builds agent i's best-response policy considering other agents' policies are fixed. A Nash equilibrium is therefore guaranteed when no further improvement is possible. In JESP, each agent's policy is represented in a tree structure, which limits its usage only to finite-horizon problems. To overcome this limitation, Inf-JESP (Infinite-Horizon JESP) [You et al., 2021] extends JESP to infinite-horizon Dec-POMDPs by representing each agent policy as a finite-state controller (FSC). Two advantages of Inf-JESP are that 1. it often achieves near-global optima despite only searching for local ones, and 2. its FSCs make for interpretable policies if their size is reasonable. However, both methods require an explicit Dec-POMDP model which details the exact environment dynamics.\nIn this paper, we propose a new algorithm called MC-JESP (Monte-Carlo Joint Equilibrium-based Search for Policies), which is a simulation-based version of Inf-JESP. In each iteration, MC-JESP builds an agent's FSC node by node using a Monte-Carlo (POMDP) planner relying on a blackbox Dec-POMDP simulator, along with the other agents' FSCs. Experiments show that MC-JESP is competitive with state-of-the-art infinite-horizon Dec-POMDP solvers based either on exact or generative models.\nThe structure of this paper is organized as follows: Section 2 discusses related work on solving Dec-POMDPs. Sec. 3 gives background about Dec-POMDPs, POMDPs, FSCs, and Inf-JESP. Our contributions are presented in Sec. 4, and experiments with comparisons to state-of-the-art Dec-POMDP solvers in Sec. 5. Finally, we conclude this work in Sec. 6 and discuss future perspectives."
},
{
"figure_ref": [],
"heading": "RELATED WORK",
"publication_ref": [
"b15",
"b11",
"b22",
"b38",
"b37",
"b18",
"b10",
"b9",
"b36",
"b28",
"b34",
"b29",
"b30",
"b17",
"b13"
],
"table_ref": [],
"text": "Recently, there has been significant progress in infinitehorizon Dec-POMDP planning, and state-of-the-art methods fall into three main types. The first type of methods estimates the best parameters of finite-state controllers (FSCs) of each agent [Amato et al., 2010a], and addresses Dec-POMDPs as an inference problem via Expectation-Maximization methods [Pajarinen and Peltonen, 2011a,b, Kumar and Zilberstein, 2010, Kumar et al., 2015]. The second type consists in transforming the Dec-POMDP problem into a Markov decision process with a state space of sufficient statistics [Mac-Dermed and Isbell, 2013, Dibangoye et al., 2014, 2016]. The third type searches for Nash equilibrium solutions, i.e., each agent's policy being a best response to the other agents' policies [Nair et al., 2003, Bernstein et al., 2005, You et al., 2021].\nHowever, for large problems or real applications, it may be challenging to represent the system's dynamics explicitly. Often, only a black-box simulator (also called a generative model) is available. Although the algorithms mentioned previously with explicit models cannot be directly applied, most state-of-the-art simulation-based methods are inspired by them. For example, Wu et al. [2013] propose to use a Monte-Carlo Expectation Maximization (MCEM) for estimating the parameters of agents' FSCs with generative models. Liu et al. [2015] improve MCEM by constructing agent FSCs using the stick-breaking prior and allowing a variable FSC size. On the other hand, similar to FB-HSVI [Dibangoye et al., 2016] (which uses explicit models), the simulation-based method oSARSA [Dibangoye and Buffet, 2018] focuses on recasting Dec-POMDPs into occupancy-state MDPs, where each occupancy-state is a sufficient statistics.\nLast but not least, some multi-agent reinforcement learning (MARL) algorithms are also interested in solving Dec-POMDPs with black-box simulators. However, most of them [Sunehag et al., 2017, Rashid et al., 2018, Son et al., 2019, Rashid et al., 2020] have not been evaluated on classical Dec-POMDP benchmarks [Seuken andZilberstein, 2007, Amato andZilberstein, 2009]. Only a few MARL algorithms conducted experiments on such domains but were limited to finite-horizon settings [Lee and Lee, 2019], or failed to obtain state-of-the-art results [Kraemer and Banerjee, 2016]."
},
{
"figure_ref": [],
"heading": "BACKGROUND",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "DEC-POMDP",
"publication_ref": [],
"table_ref": [],
"text": "The problem of finding optimal collaborative behaviors for a group of agents under stochastic dynamics and partial observability is typically formalized as a decentralized partially observable Markov decision process (Dec-POMDP).\nDefinition 1. A Dec-POMDP with |I| agents is represented as a tuple M ≡ I, S, A, Ω, T, O, R, b 0 , H, γ , where: I = {1, . . . , |I|} is a finite set of agents; S is a finite set of states; A = × i A i is the finite set of joint actions, with A i the set of agent i's actions; Ω = × i Ω i is the finite set of joint observations, with Ω i the set of agent i's observations; T : S × A × S → R is the transition function, with T (s, a, s ) the probability of transiting from s to s if a is performed; O : A × S × Ω → R is the observation function, with O(a, s , o) the probability of observing o if a is performed and the next state is s ; R : S × A → R is the reward function, with R(s, a) the immediate reward for executing a in s; b 0 is the initial probability distribution over states; H ∈ N ∪ {∞} is the (possibly infinite) time horizon; γ ∈ [0, 1) is the discount factor applied to future rewards. An agent's i action policy π i maps its possible actionobservation histories to actions. The objective is then to find a joint policy π = π 1 , . . . , π |I| that maximizes the expected discounted return from b 0 :\nV π H (b 0 ) def = E H-1 t=0 γ -t r(S t , A t ) | S 0 ∼ b 0 , π .\nHowever, we often do not know the exact transition, observation, and reward functions for large problems or real-world applications, but may rely on a generative model (blackbox simulator) G, which, given a state-action pair s, a , samples a triplet s , o, r ."
},
{
"figure_ref": [],
"heading": "POMDP",
"publication_ref": [],
"table_ref": [],
"text": "In this work, we will consider one agent i at a time, and thus end up solving a single-agent partially observable Markov decision problem (POMDP) in each iteration, i.e., the particular case of a single-agent Dec-POMDP (I = {1}). In a POMDP, an optimal policy π * exists whose input is the belief state b, i.e., the probability distribution over states given the current action-observation history. For finite H, the optimal value function (which allows deriving π * ) is recursively defined as: \nV * h (b) = max a r(b, a) + γ o P r(o | b, a)V * h-1 (b a,o ) ,"
},
{
"figure_ref": [],
"heading": "FINITE STATE CONTROLLERS",
"publication_ref": [
"b21"
],
"table_ref": [],
"text": "In POMDPs as in Dec-POMDPs, solution policies can also be sought for in the form of finite state controllers (FSC) (also called policy graphs [Meuleau et al., 1999]), i.e., automata whose transitions from one internal state to the next depend on the received observations and generate the actions to be performed.\nDefinition 2. For some POMDP sets A and Ω, a (deterministic) FSC is specified by a tuple fsc ≡ N, η, ψ , where:\n• N is a finite set of nodes, with n 0 the start node;\n• η : N × Ω → N is the node transition function; n = η(n, o) is the next node and observing o from node n;\n• ψ : N → A is the action-selection function of the FSC; a = ψ(n) is the action triggered when in node n."
},
{
"figure_ref": [],
"heading": "SOLVING DEC-POMDPS BY FINDING NASH EQUILIBRIA (INFINITE-HORIZON JESP)",
"publication_ref": [
"b38",
"b16",
"b38"
],
"table_ref": [],
"text": "Inf-JESP (Infinite-Horizon JESP) [You et al., 2021] is an infinite-horizon Dec-POMDP solver, which is based on Nair et al.'s JESP [2003], but replaces the policy tree representation by a finite-state controller (FSC) for each agent's policy. This modification allows solving infinite-horizon problems rather than finite-horizon ones, and may help scaling up to larger problems. More specifically, in Inf-JESP (Algorithm 1), each iteration derives (line 7) the explicit model of a (best-response) POMDP for agent i by fixing the other agents' FSCs (index \" = i\") and using an extended state space e t ∈ E, i.e., containing:\n• s t , the current state of the Dec-POMDP problem,\n• n t =i ≡ n t j j =i , the current nodes of other agents, and • õt i , agent i's current observation.\nDenoting η =i (n t =i , o t+1 =i ) = η(n t j , õt+1 j ) j =i and ψ =i (n t =i ) = ψ j (n t j ) j =i , this leads to a valid POMDP with the following dynamics: 1 T e (e t , a t i , e t+1 ) = P r(e t+1 |e t , a\nt i ) = o t+1 =i T (s t , ψ =i (n t =i ), a t i , s t+1 ) • 1 n t+1 =i =η =i (n t =i ,o t+1 =i ) • O(s t+1 , ψ =i (n t =i ), a t i , o t+1 =i , o t+1 i ), O e (a t i , e t+1 i , o t+1 i ) = P r(o t+1 i |a t i , e t+1 i ) = P r(o t+1 i |a t i , s t+1 , n t+1 =i , õt+1 i ) = 1 o t+1 i =õ t+1 i , r e (e t , a t i ) = r(s t , a t i , ψ =i (n t =i )).\nThen, Inf-JESP solves this explicit POMDP for agent i using an -optimal offline POMDP solver (SARSOP [Kurniawati et al., 2008]) and derives an FSC fsc i that approximates the solution policy (cf. line 8, which does not distinguish both steps). fsc i is then evaluated (line 9) and retained only if it improves on i's previous FSC, fsc i , so that Inf-JESP stops\nAlgorithm 1: Inf-JESP's Local Search 1 [Input:] b 0 : initial belief | M : Dec-POMDP model | fsc: initial FSCs 2 Fct LocalSearch(b 0 , M, fsc) 3 v bestL ← eval(fsc) 4 #ni ← 0 // #(iterations w/o improvement) 5 i ← 1 // Id of current agent 6 repeat // Cycle over agents 7 b 0 BR , M BR ← getBRpomdp(b 0 , M, fsc =i ) 8 fsc i ←ComputeFSC(b 0 BR , M BR ) 9 v ← Eval(fsc i , b 0 BR , M BR ) 10 if v > v bestL then // Keep new FSC if better 11 fsc i ← fsc i 12 v bestL ← v 13 #ni ← 0 14 else // increment #ni 15 #ni ← #ni + 1 16 i ← (i mod |I|) + 1 17 until #ni = |I| // No improvement in last cycle.\n18 return fsc, v bestL when an approximate Nash equilibrium is obtained, which is detected using a counter #ni.\nIn practice (see [You et al., 2021] or Sec. 5.1), this search for an equilibrium often allows finding near-global optima either using some random restarts, or initial FSCs obtained through relaxing the Dec-POMDP."
},
{
"figure_ref": [],
"heading": "MONTE-CARLO JESP",
"publication_ref": [
"b38"
],
"table_ref": [],
"text": "As Inf-JESP, we aim to find Nash equilibrium infinitehorizon solutions by iteratively building each agent's best response to other agents' fixed policies. We thus stick to representing policies as FSCs, and to using the same algorithmic scheme for the local search as presented in Alg. 1.\nThis requires relying on the same best-response POMDPs, i.e., in particular with the same extended state s t e = s t , n t =i , o t i . However, lacking an explicit Dec-POMDP model, we cannot derive explicit POMDP models. To address this issue, in MC-JESP, we propose an alternative approach that relies on best-response generative POMDP models (noted G BR ) derived from the Dec-POMDP simulator and other agents' FSCs. In the following, we discuss how to build such models, how to derive solution FSCs, how to obtain initial heuristic FSCs, and what are the properties of the resulting approach.\n1 Note: You et al. [2021] provide formulas for stochastic FSCs. "
},
{
"figure_ref": [
"fig_1"
],
"heading": "BEST-RESPONSE GENERATIVE MODEL",
"publication_ref": [],
"table_ref": [],
"text": "A generative POMDP model G BR for agent i has to sample the next extended state s t+1 e , observation o t+1 i , and reward r t+1 , given a current extended state s t e and action a t i . As illustrated in Figure 1 and as detailed in Alg. 2, this can be achieved by relying only on the Dec-POMDP simulator and the other agents' FSCs. The algorithm first decomposes the extended state, and gets other agents' actions a t =i according to action-selection functions ψ =i ≡ ψ j j =i (lines 4 and 5). Then, in line 6, the joint action a t i , a "
},
{
"figure_ref": [],
"heading": "COMPUTING AGENT i'S FSC USING MONTE-CARLO METHODS",
"publication_ref": [
"b33",
"b16"
],
"table_ref": [],
"text": "In the previous section, we demonstrate how to build the best-response generative model G BR for agent i considering others' fixed FSCs. However, unlike in Inf-JESP, state-ofthe-art point-based POMDP solvers (see, e.g., [Pineau et al., 2003, Smith and Simmons, 2004, Kurniawati et al., 2008]) require exact models, and thus can not be used in MC-JESP.\nWe thus rely on a simulation-based solver, i.e., POMCP [Silver and Veness, 2010], which is an online algorithm, i.e., it focuses on returning the best action for the current belief. Therefore, the question is how to use a simulator (G BR ) and an online simulation-based solver to obtain agent i's FSC. To answer it, we propose an algorithm that 1. uses this Monte-Carlo planner (POMCP) to compute the best action for a given FSC node, which is labeled by a unique belief; and 2. expands reachable beliefs, i.e., creates new FSC nodes using computed actions to gradually build a complete FSC. Moreover, to control the computational cost, we explicitly bound the FSC size with a given parameter N max-fsc ∈ N.\nIn the proposed algorithm, each FSC node is attached to 1. an approximate belief b (with at least N min-part particles), 2. a preferred action a i , and 3. a weight w that estimates the probability to reach that node at least once during execution.\nAs detailed in Algorithm 3, this information is first gathered for initial belief b 0 by calling POMCP to get agent i's best action a 0 i in line 4. Line 5 then creates a start node n 0 with b 0 BR , a 0 i , and a weight w = 1. This start node is added to the FSC under construction (N ) and an open list (L). Now, while L is not empty, its node n with largest weight w is popped out (line 9), so as to first develop the nodes that may have the highest impact on the value at the root. Expanding it requires mapping each observation o i feasible when performing n.a i from n.b to a particle set. This is achieved through sampling by getNxtBeliefs until each feasible o i (according to the samples) is attached to at least N min-part particles (line 29), which returns a set Ω i of feasible observations, and a mapping B i from these observations to particle sets. Then, for each individual observation o i , the algorithm needs to create a transition to an appropriate node, which may already exist or needs to be created, as explained in the following. If o i is assumed impossible when performing n.a i in n.b (o i ∈ n.Ω i ), then a self-loop is added (line 13). 2 Otherwise, line 15 gets the belief b BR attached to o i , and line 16 computes an associated weight w . If (i) a belief -close to b BR (in 1-norm) exists in N , or (ii) the FSC has reached its size limit N max-fsc (line 17), then we take as next node n the one in the FSC minimizing b BR -n .b 1 and update its weight (lines 18-19). Otherwise a next node n is created using an action selected by POMCP (lines 21-22), and added to both N and L (lines 22-24). In line 25, whatever the origin of n , an edge n → n is created in the FSC with a label o i .\nNote that, for a fixed N max-fsc value, a small may prevent from representing long trajectories, while a large may induce excessive node merging. \nΩ i , B i ← getNxtBeliefs(n.b, G BR , n.a i )\nfor o i ∈ Ω i do // For each obs. of i: \nif o i ∈ Ω i then // oi unexpected: η(n, o i ) ← n // add self-loop else // Else: create next node b BR ← B i [o i ] w ← |B i [oi]| |B i | • n.w if (b BR ∈ N ( )) ∨ (|N | = N max-fsc ) then //\na i ← P OM CP (b BR , G BR ) 22 n ← node(b BR , a i , w ) 23 N ← N ∪ {n } 24 L[w ] ← n η(n, o i ) ← n // Add transition to FSC. Fct getNxtBeliefs(b t BR , G BR , a t i ) Ω t+1 i ← ∅ B t+1 i ← ∅ repeat e t ∼ b t BR e t+1 , o t+1 i , r t+1 ∼ G BR (e t , a t i ) if o t+1 i / ∈ Ω t+1 i then Ω t+1 i ← Ω t+1 i ∪ {o t+1 i } B t+1 [o t+1 i ] ← B t+1 [o t+1 i ] ∪ {e t+1 } until Timeout() ∨ (M inBelief P articles(B t+1 ) > N min-part ) return Ω t+1 i , B t+1 i"
},
{
"figure_ref": [],
"heading": "HEURISTIC INITIALIZATION",
"publication_ref": [
"b27"
],
"table_ref": [],
"text": "Although MC-JESP monotonically improves the value of the joint policy at each iteration, random initializations often lead to poor local optima. To benefit from a heuristic initialization that allows finding good solutions quickly and reliably, we adapt Inf-JESP's heuristics as we adapted the computation of an agent's FSC in the previous section: using particle sets as beliefs, calling a simulation-based solver, and bounding the number of nodes. In addition, to derive the next belief, we marginalize over possible joint observations o =i , rather than reasoning on them separately as You et al.\n[2021] did (e.g., considering only the most probable one).\nThis heuristic initialization assumes that 1. agent i's decisions are made as if all agents where sharing their observations, thus acting as a single agent; this means making decisions (picking joint actions a) by solving a Multi-agent POMDP (MPOMDP) [Pynadath and Tambe, 2002] relaxation of the original Dec-POMDP, which can be done here with an (online) simulation-based POMDP solver; and 2. agent i's belief b over the hidden state is updated assuming (a) that the other agents ( = i) also act according to the computed MPOMDP policy at b, but (b) using only i's observation, o i , while marginalizing over others' observations (o =i , which are actually not known to i at execution time) to ignore them. This one-sided-observation belief update is computed as follows:\nb a,oi (s )\ndef = P r(s |o i , a, b) = P r(s , o i , a, b) P r(o i , a, b) = o =i O( o i , o =i |a, s ) s T (s, a, s )b(s) s ,o =i O( o i , o =i |a, s ) s T (s, a, s )b(s)\n.\nFollowing this idea, we derive the FSC heuristic initialization process for agent i detailed in Alg. 4 which, as shown in red, differs from Alg. 3 in two aspects:\n• the Dec-POMDP simulator G is used as an MPOMDP simulator for POMCP to get the best joint action with a given belief (lines 3 and 21); and\n• in line 10's getNxtBeliefs function, the next estimated beliefs are obtained by repeatedly sampling transitions using the computed joint action n.a i , n.a =i (and Dec-POMDP simulator G), and collecting particle sets for each encountered individual observation o i , ignoring o =i , which is equivalent to a marginalization."
},
{
"figure_ref": [],
"heading": "OBSERVATIONS",
"publication_ref": [],
"table_ref": [],
"text": "With an increasing time budget, POMCP asymptotically converges to optimal decisions. By (i) increasing POMCP's time budget to infinity, (ii) increasing N min-part. and N max-fsc to infinity, and (iii) setting = 0, each iteration of the local search would thus return the best response (possibly infinite) \nΩ i , B i ← getNxtBeliefs(n.b, G, n.a i , n.a =i )\nfor o i ∈ Ω do // For each obs. of i: \nif o i ∈ Ω i then // oi unexpected: add self-loop η(n, o i ) ← n else // Else: create next node: 15 b ← B i [o i ] w ← n.w * |B i [oi]| |B i | 17 if (b ∈ N ( )) ∨ (|N | = N max-fsc ) then //\na i , a =i ← P OM CP (b , G) 22 n ← node(b , a i , a =i , w ) 23 N ← N ∪ {n } 24 L[w] ← n η(n, o i ) ← n\nFSC. As a consequence, assuming also an exact evaluation of FSCs, the local search would be guaranteed to find a Nash equilibrium.\nIn practice, we only obtain approximate Nash equilibria. Also, due to randomization in POMCP and in FSC evaluations through simulations, restarts of the full process lead to different search trajectories, but always stop in finitely many iterations with probability 1. The next section looks at the results obtained in practice through experiments."
},
{
"figure_ref": [],
"heading": "EXPERIMENTS",
"publication_ref": [
"b22",
"b4",
"b30",
"b10",
"b37",
"b18",
"b9",
"b12"
],
"table_ref": [],
"text": "We evaluate our contributions on five state-of-the-art Dec-POMDP benchmarks (cf. http://masplan.org/ problem_domains): Decentralized Tiger [Nair et al., 2003], Recycling Robots [Amato et al., 2012], Meeting in a 3 × 3 grid [Amato et al., 2009], Cooperative Box Pushing [Seuken and Zilberstein, 2007], Mars Rover [Amato and Zilberstein, 2009]. We compare MC-JESP with state-of-the-art Dec-POMDP solvers relying on either: explicit models:\n(which benefit from more information) FB-HSVI [Dibangoye et al., 2016], Peri [Pajarinen and Peltonen, 2011b],\nPeriEM [Pajarinen and Peltonen, 2011b], PBVI-BB [Mac-Dermed and Isbell, 2013], MealyNLP [Amato et al., 2010b] and Inf-JESP; or generative models: MCEM [Wu et al., 2013], Dec-SBPR [Liu et al., 2015] and oSARSA [Dibangoye and Buffet, 2018].\nFor MC-JESP, 1. POMCP is used as our simulation-based POMDP planner with a timeout of 1 s; 2. we consider three different maximum FSC sizes: 10, 30, and 50, respectively; 3. the threshold distance between beliefs is set to = 0.1; and 4. FSC evaluations (line 9 of Alg. 1) use 10 6 simulations that stop when γ t < 10 -4 . For MC-JESP and Inf-JESP's empirical results, having access to the exact model in each benchmark problem, we use Hansen's [1998] policy evaluation for FSCs applied to a best-response POMDP. The experiments with MC-JESP were conducted on a laptop with a 2.3 GHz i9 CPU. The source code is in the supplementary material."
},
{
"figure_ref": [],
"heading": "COMPARISON WITH STATE-OF-THE-ART ALGORITHMS",
"publication_ref": [
"b38"
],
"table_ref": [
"tab_3"
],
"text": "Table 1 presents the results for the 5 problems, the solvers being ordered from best to worse value. Among x restarts of MC-JESP, the best value is reported in MC-JESP(M-x), and the average value in MC-JESP(M-1 x ) (to look at the benefit of restarting). For Inf-JESP, we report the best values among its 3 possible initializations [You et al., 2021]. For MC-JESP, we report the best values over the 3 possible max. FSC sizes.\nThe columns provide: • (Alg.) the different algorithms at hand, with a * exponent for those who rely on an explicit model; • (FSC size) the final FSC size (for Inf-JESPs and MC-JESP); • (Iterations) the number of iterations required to converge (for Inf-JESPs and MC-JESP); whether restarting can be beneficial."
},
{
"figure_ref": [
"fig_5",
"fig_5",
"fig_5",
"fig_2"
],
"heading": "A CLOSER LOOK AT MC-JESP'S BEHAVIOR",
"publication_ref": [],
"table_ref": [],
"text": "We study MC-JESP's performance with three different maximum FSC sizes in Fig. 2 (red for 10, green for 30, and blue for 50). Right parts present the distribution over final values of MC-JESP with 20 restarts. In the five problems at hand, MC-JESP with max FSC size 50 (blue) has distributions more concentrated on good values than others, and most values are close to FB-HSVI's ones (thus, near-optimal values). These distributions show that few restarts are needed to reach good solutions with high probability if we give large enough FSC sizes.\nThe left parts of Fig. 2 present the evolution of the values during each iteration of MC-JESP with the three maximum FSC sizes. The average is computed over all runs, even if they have already converged. This figure first shows that MC-JESP monotonically increases during each run, and most runs converge to good local optima in a few iterations. Second, we observe that, for large problems (Box-Pushing and Mars Rovers), there are already significant drops from MC-JESP in the first iteration with an FSC size limit decreasing from 50 to 10. This indicates that, for large problems, we must give large enough FSC size limits, while this is not necessary for small problems.\nLast but not least, in Dec-Tiger, although some restarts of MC-JESP end with optimal values, we observe that the average value is still relatively low compared with FB-HSVI. Therefore, we conducted another experiment to investigate the impact of different POMCP timeouts (note that there is a fixed timeout of 1 s for the experiments illustrated in Figure 2). To that end, we limit the FSC size in each iteration to at most 50 nodes, and we test MC-JESP with five POMCP timeouts (1 s, 5 s, 10 s, 20 s, and 30 s). The distribution of final values is shown in Fig. 3. We observe that the average value increases and the variability shrinks when we give more time to POMCP. However, it also indicates that, when we increase the time budget, we have a lower chance of getting \"lucky\" good values."
},
{
"figure_ref": [],
"heading": "CONCLUSION",
"publication_ref": [
"b5",
"b8"
],
"table_ref": [],
"text": "In this work, based on Inf-JESP, we propose a novel infinitehorizon Dec-POMDP solver called MC-JESP, which only requires a black-box Dec-POMDP simulator, and returns FSCs, i.e., representations that can make for interpretable policies. We describe how to obtain a best-response generative model (the simulator of the POMDP faced by some agent i assuming known FSCs for other agents), and the process to extract an FSC for each agent. Moreover, a heuristic initialization method for MC-JESP is also provided.\nThrough experiments, we prove that MC-JESP preserves Inf-JESP's competitive results (though at the cost of an increased computation time), performing better than many explicit model-based algorithms, and outperforming other simulation-based algorithms in most cases. Because it seeks Nash equilibria, this approach could better scale up to large problems than approaches directly seeking global optima.\nSeveral improvements of MC-JESP could be envisioned, such as: 1. robustly comparing FSCs fsc i and fsc i , while minimizing computation time through hypothesis testing; 2. if using large FSCs, using space partitioning (e.g., k-d trees [Bentley, 1975] or cover trees [Beygelzimer et al., 2006]) to speed up the search for nearest nodes; and 3. re-using POMCP trees from one node to the next, or to initialize getNxtBeliefs, although doing so may significantly increase memory usage.\nAlso, preliminary experiments show that MC-JESP works on a continuous-state meet-in-a-grid problem, the main issue being to replace the distance between sets of discrete particles (i.e., just comparing two vectors representing discrete distributions) by a distance over continuous particles (which requires taking the distance between states into account). For future works, we plan to extend MC-JESP to problems with continuous actions and observations. This would require not only relying on algorithms such as Sunberg and Kochenderfer's POMCPOW [2018], but also, more importantly, deriving FSCs that can handle continuous observations."
}
] | Decentralized partially observable Markov decision processes (Dec-POMDPs) formalize the problem of designing individual controllers for a group of collaborative agents under stochastic dynamics and partial observability. Seeking a global optimum is difficult (NEXP complete), but seeking a Nash equilibrium -each agent policy being a best response to the other agents-is more accessible, and allowed addressing infinite-horizon problems with solutions in the form of finite state controllers. In this paper, we show that this approach can be adapted to cases where only a generative model (a simulator) of the Dec-POMDP is available. This requires relying on a simulation-based POMDP solver to construct an agent's FSC node by node. A related process is used to heuristically derive initial FSCs. Experiment with benchmarks shows that MC-JESP is competitive with exisiting Dec-POMDP solvers, even better than many offline methods using explicit models. | Monte-Carlo Search for an Equilibrium in Dec-POMDPs | [
{
"figure_caption": "where (i) r(b, a) = s b(s) • r(s, a); (ii) P r(o | b, a) depends on the dynamics; and (iii) b a,o is the belief updated upon performing a and perceiving o.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 1 :1Figure 1: Structure of the best-response POMDP generative model G BR , with inputs and outputs represented as: blue arrows for the Dec-POMDP simulator G; green arrows for agents = i' FSCs; and black arrows for G BR .",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Algorithm 3 :3Compute agent i's FSC 1 [Input:] b 0 BR : G BR 's initial (extended) belief state | G BR : best response generative model for agent i 2 [Parameters:] N max-fsc : max FSC size for agent i | N min-part : min number of particles in each belief | : min. distance between beliefs 3 Fct ComputeFSC(b 0 BR , G BR )",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Algorithm 4 :4Build a heuristic FSC for agent i [Input:] b 0 : initial state distribution | i: agent index [Parameters:] G: Dec-POMDP simulator | N max-fsc : max. FSC size | : min. distance between beliefs a 0 i , a 0 =i ← P OM CP (b 0 , G) n 0 ← node(b 0 , a 0 i , a 0 =i , w = 1) N ← {n 0 } // init FSC & open list L[w] ← n 0 while |L| > 0 do // loop over open nodes L.sort() n ← L.popf ront()",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: Values of the joint policy for the Dec-Tiger, Grid, Recycling, Box-Pushing, and Mars Rover problems (from top to bottom). The left part of each figure presents the evolution (during a run) of the value of the joint policy at each iteration of MC-JESP(1 20 ) (avg + 10th and 90th percentiles) with different bounded FSC sizes (10, 30, and 50, respectively). The dashed line represents FB-HSVI's final value. The right part presents the value distribution after convergence of MC-JESP(1 20 ).",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: Values of the joint policy for the Dec-Tiger problem for different POMCP timeout values.",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "the FSC functions ψ and η being deterministic. Dec-POMDP simulator | fsc =i ≡ N =i , ψ =i , η =i : other agents' FSCs",
"figure_data": "t =i is passed to theDec-POMDP simulator G, which outputs the next state s t+1 ,joint observation o t+1 i, o t+1 =iand instant reward r t+1 . Withthe other agents' observations o t+1 =i , line 7 computes theirnext nodes n t+1 =i ≡ n t+1 jj =i . In the end, we build the nextextended state s t+1 eand return the results (lines 8 and 9). Inthis algorithm, stochasticity exists only in the Dec-POMDPsimulator G, Algorithm 2: i's Best-Response Generative Model[Input:] s t e : extended state | a t i : agent i's action[Parameters:] G: Fct G BR (s t e , a t i , [G, fsc =i ]) s t , n t =i , o t i ← s t e // extract s t e 's 3 componentsa t =i ← ψ =i (n t =i ) // get action from FSCs t+1 , o t+1 , r t+1 ← G(s t , a t ) // sample transitionn t+1 =i ← η(n t =i , o t+1 =i )// evolve FSCs t+1 e← s t+1 , n t+1 =i , o t+1 i// build s t+1 ereturn s t+1 e , o t+1 i, r t+1// return step results",
"figure_id": "tab_0",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "JESP requires more solving time. For example, in large problems such as Mars Rovers, MC-JESP takes 349 s on average to solve the task, while Inf-JESP takes 122 s. But this is not surprising since MC-JESP only uses a black-box simulator. A key question is how to determine Comparison of different algorithms in terms of final FSC size, number of iterations, time, and value, on 5 infinite-horizon benchmarks with γ = 0.9 for all domains. The solvers are listed in a decreasing order of value.",
"figure_data": "• (Time) the",
"figure_id": "tab_3",
"figure_label": "1",
"figure_type": "table"
}
] | Yang You; Vincent Thomas; Francis Colas; Olivier Buffet | [
{
"authors": "Christopher Amato; Shlomo Zilberstein",
"journal": "",
"ref_id": "b0",
"title": "Achieving goals in decentralized POMDPs",
"year": "2009"
},
{
"authors": "Christopher Amato; Jilles Steeve Dibangoye; Shlomo Zilberstein",
"journal": "",
"ref_id": "b1",
"title": "Incremental policy generation for finitehorizon DEC-POMDPs",
"year": "2009"
},
{
"authors": "Christopher Amato; Daniel S Bernstein; Shlomo Zilberstein",
"journal": "Journal of Autonomous Agents and Multi-Agent Systems",
"ref_id": "b2",
"title": "Optimizing fixed-size stochastic controllers for POMDPs and decentralized POMDPs",
"year": "2010"
},
{
"authors": "Christopher Amato; Blai Bonet; Shlomo Zilberstein",
"journal": "",
"ref_id": "b3",
"title": "Finite-state controllers based on Mealy machines for centralized and decentralized POMDPs",
"year": "2010"
},
{
"authors": "Christopher Amato; Daniel Bernstein; Shlomo Zilberstein",
"journal": "",
"ref_id": "b4",
"title": "Optimizing memory-bounded controllers for decentralized POMDPs",
"year": "2012"
},
{
"authors": "Jon Louis; Bentley ",
"journal": "Communications of the ACM",
"ref_id": "b5",
"title": "Multidimensional binary search trees used for associative searching",
"year": "1975-09"
},
{
"authors": "S Daniel; Eric A Bernstein; Shlomo Hansen; Zilberstein",
"journal": "",
"ref_id": "b6",
"title": "Bounded policy iteration for decentralized POMDPs",
"year": ""
},
{
"authors": "D S Bernstein; R Givan; N Immerman; Shlomo Zilberstein",
"journal": "Math. of Operations Research",
"ref_id": "b7",
"title": "The complexity of decentralized control of Markov decision processes",
"year": "2002"
},
{
"authors": "Alina Beygelzimer; Sham Kakade; John Langford",
"journal": "",
"ref_id": "b8",
"title": "Cover trees for nearest neighbor",
"year": "2006"
},
{
"authors": "Jilles Dibangoye; Olivier Buffet",
"journal": "PMLR",
"ref_id": "b9",
"title": "Learning to act in decentralized partially observable MDPs",
"year": "2018-07-15"
},
{
"authors": "Jilles Dibangoye; Chris Amato; Olivier Buffet; François Charpillet",
"journal": "Journal of Artificial Intelligence Research",
"ref_id": "b10",
"title": "Optimally solving Dec-POMDPs as continuous-state MDPs",
"year": "2016"
},
{
"authors": "Olivier Jilles S Dibangoye; François Buffet; Charpillet",
"journal": "Springer",
"ref_id": "b11",
"title": "Error-bounded approximations for infinite-horizon discounted decentralized POMDPs",
"year": "2014"
},
{
"authors": "Eric Hansen",
"journal": "",
"ref_id": "b12",
"title": "An improved policy iteration algorithm for partially observable MDPs",
"year": "1998"
},
{
"authors": "Landon Kraemer; Bikramjit Banerjee",
"journal": "Neurocomputing",
"ref_id": "b13",
"title": "Multi-agent reinforcement learning as a rehearsal for decentralized planning",
"year": "2016"
},
{
"authors": "Akshat Kumar; Shlomo Zilberstein",
"journal": "",
"ref_id": "b14",
"title": "Anytime planning for decentralized POMDPs using expectation maximization",
"year": "2010-07"
},
{
"authors": "Akshat Kumar; Shlomo Zilberstein; Marc Toussaint",
"journal": "Journal of Artificial Intelligence Research",
"ref_id": "b15",
"title": "Probabilistic inference techniques for scalable multiagent decision making",
"year": "2015"
},
{
"authors": "Hanna Kurniawati; David Hsu; Wee Sun; Lee ",
"journal": "",
"ref_id": "b16",
"title": "SARSOP: Efficient point-based POMDP planning by approximating optimally reachable belief spaces",
"year": "2008"
},
{
"authors": "Hyun-Rok Lee; Taesik Lee",
"journal": "",
"ref_id": "b17",
"title": "Improved cooperative multiagent reinforcement learning algorithm augmented by mixing demonstrations from centralized policy",
"year": "2019"
},
{
"authors": "Miao Liu; Christopher Amato; Xuejun Liao; Lawrence Carin; Jonathan P How",
"journal": "",
"ref_id": "b18",
"title": "Stick-breaking policy learning in Dec-POMDPs",
"year": "2015"
},
{
"authors": "Liam C Macdermed; Charles Isbell",
"journal": "",
"ref_id": "b19",
"title": "Point based value iteration with optimal belief compression for Dec-POMDPs",
"year": "2013"
},
{
"authors": "Omid Madani; Steve Hanks; Anne Condon",
"journal": "Artificial Intelligence",
"ref_id": "b20",
"title": "On the undecidability of probabilistic planning and related stochastic optimization problems",
"year": "2003"
},
{
"authors": "N Meuleau; K.-E Kim; L P Kaelbling; A R Cassandra",
"journal": "",
"ref_id": "b21",
"title": "Solving POMDPs by searching the space of finite policies",
"year": "1999"
},
{
"authors": "Ranjit Nair; Milind Tambe; Makoto Yokoo; David Pynadath; Stacy Marsella",
"journal": "Citeseer",
"ref_id": "b22",
"title": "Taming decentralized POMDPs: Towards efficient policy computation for multiagent settings",
"year": "2003"
},
{
"authors": "Frans A Oliehoek; Christopher Amato",
"journal": "Springer Publishing Company, Incorporated",
"ref_id": "b23",
"title": "A Concise Introduction to Decentralized POMDPs",
"year": "2016"
},
{
"authors": "Joni Pajarinen; Jaakko Peltonen",
"journal": "",
"ref_id": "b24",
"title": "Efficient planning for factored infinite-horizon DEC-POMDPs",
"year": "2011"
},
{
"authors": "Joni Pajarinen; Jaakko Peltonen",
"journal": "",
"ref_id": "b25",
"title": "Periodic finite state controllers for efficient POMDP and DEC-POMDP planning",
"year": "2011"
},
{
"authors": "Joelle Pineau; Geoff Gordon; Sebastian Thrun",
"journal": "",
"ref_id": "b26",
"title": "Pointbased value iteration: An anytime algorithm for POMDPs",
"year": ""
},
{
"authors": "David V Pynadath; Milind Tambe",
"journal": "Journal of Artificial Intelligence Research",
"ref_id": "b27",
"title": "The communicative multiagent team decision problem: Analyzing teamwork theories and models",
"year": "2002"
},
{
"authors": "Tabish Rashid; Mikayel Samvelyan; Christian Schroeder; Gregory Farquhar; Jakob Foerster; Shimon Whiteson",
"journal": "PMLR",
"ref_id": "b28",
"title": "Qmix: Monotonic value function factorisation for deep multi-agent reinforcement learning",
"year": "2018"
},
{
"authors": "Tabish Rashid; Gregory Farquhar; Bei Peng; Shimon Whiteson",
"journal": "Advances in neural information processing systems",
"ref_id": "b29",
"title": "Weighted Qmix: Expanding monotonic value function factorisation for deep multi-agent reinforcement learning",
"year": "2020"
},
{
"authors": "Sven Seuken; Shlomo Zilberstein",
"journal": "",
"ref_id": "b30",
"title": "Improved memory-bounded dynamic programming for decentralized POMDPs",
"year": "2007"
},
{
"authors": "David Silver; Joel Veness",
"journal": "",
"ref_id": "b31",
"title": "Monte-Carlo planning in large POMDPs",
"year": ""
},
{
"authors": " Curran Associates; Inc",
"journal": "",
"ref_id": "b32",
"title": "",
"year": "2010"
},
{
"authors": "Trey Smith; Reid Simmons",
"journal": "AUAI Press",
"ref_id": "b33",
"title": "Heuristic search value iteration for pomdps",
"year": "2004"
},
{
"authors": "Kyunghwan Son; Daewoo Kim; Wan Ju Kang; David Earl Hostallero; Yung Yi",
"journal": "PMLR",
"ref_id": "b34",
"title": "Qtran: Learning to factorize with transformation for cooperative multi-agent reinforcement learning",
"year": "2019"
},
{
"authors": "Zachary N Sunberg; J Mykel; Kochenderfer",
"journal": "",
"ref_id": "b35",
"title": "Online algorithms for POMDPs with continuous state, action, and observation spaces",
"year": "2018"
},
{
"authors": "Peter Sunehag; Guy Lever; Audrunas Gruslys; Wojciech ; Marian Czarnecki; Vinicius Zambaldi; Max Jaderberg; Marc Lanctot; Nicolas Sonnerat; Joel Z Leibo; Karl Tuyls",
"journal": "",
"ref_id": "b36",
"title": "Value-decomposition networks for cooperative multi-agent learning",
"year": "2017"
},
{
"authors": "Feng Wu; Shlomo Zilberstein; Nicholas R Jennings",
"journal": "AAAI Press",
"ref_id": "b37",
"title": "Monte-Carlo expectation maximization for decentralized POMDPs",
"year": "2013"
},
{
"authors": "Yang You; Vincent Thomas; Francis Colas; Olivier Buffet",
"journal": "",
"ref_id": "b38",
"title": "Solving infinite-horizon Dec-POMDPs using finite state controllers within JESP",
"year": "2021"
}
] | [
{
"formula_coordinates": [
2,
328.22,
320.29,
190.84,
30.2
],
"formula_id": "formula_0",
"formula_text": "V π H (b 0 ) def = E H-1 t=0 γ -t r(S t , A t ) | S 0 ∼ b 0 , π ."
},
{
"formula_coordinates": [
2,
309.65,
596.68,
227.97,
21.69
],
"formula_id": "formula_1",
"formula_text": "V * h (b) = max a r(b, a) + γ o P r(o | b, a)V * h-1 (b a,o ) ,"
},
{
"formula_coordinates": [
3,
54.64,
536.92,
238.36,
112.75
],
"formula_id": "formula_2",
"formula_text": "t i ) = o t+1 =i T (s t , ψ =i (n t =i ), a t i , s t+1 ) • 1 n t+1 =i =η =i (n t =i ,o t+1 =i ) • O(s t+1 , ψ =i (n t =i ), a t i , o t+1 =i , o t+1 i ), O e (a t i , e t+1 i , o t+1 i ) = P r(o t+1 i |a t i , e t+1 i ) = P r(o t+1 i |a t i , s t+1 , n t+1 =i , õt+1 i ) = 1 o t+1 i =õ t+1 i , r e (e t , a t i ) = r(s t , a t i , ψ =i (n t =i ))."
},
{
"formula_coordinates": [
3,
295.08,
68.77,
224.92,
234.39
],
"formula_id": "formula_3",
"formula_text": "Algorithm 1: Inf-JESP's Local Search 1 [Input:] b 0 : initial belief | M : Dec-POMDP model | fsc: initial FSCs 2 Fct LocalSearch(b 0 , M, fsc) 3 v bestL ← eval(fsc) 4 #ni ← 0 // #(iterations w/o improvement) 5 i ← 1 // Id of current agent 6 repeat // Cycle over agents 7 b 0 BR , M BR ← getBRpomdp(b 0 , M, fsc =i ) 8 fsc i ←ComputeFSC(b 0 BR , M BR ) 9 v ← Eval(fsc i , b 0 BR , M BR ) 10 if v > v bestL then // Keep new FSC if better 11 fsc i ← fsc i 12 v bestL ← v 13 #ni ← 0 14 else // increment #ni 15 #ni ← #ni + 1 16 i ← (i mod |I|) + 1 17 until #ni = |I| // No improvement in last cycle."
},
{
"formula_coordinates": [
5,
73.37,
317.68,
165.43,
10.82
],
"formula_id": "formula_4",
"formula_text": "Ω i , B i ← getNxtBeliefs(n.b, G BR , n.a i )"
},
{
"formula_coordinates": [
5,
82.73,
341.59,
186.13,
86.23
],
"formula_id": "formula_5",
"formula_text": "if o i ∈ Ω i then // oi unexpected: η(n, o i ) ← n // add self-loop else // Else: create next node b BR ← B i [o i ] w ← |B i [oi]| |B i | • n.w if (b BR ∈ N ( )) ∨ (|N | = N max-fsc ) then //"
},
{
"formula_coordinates": [
5,
44.68,
480.76,
210.4,
232.89
],
"formula_id": "formula_6",
"formula_text": "a i ← P OM CP (b BR , G BR ) 22 n ← node(b BR , a i , w ) 23 N ← N ∪ {n } 24 L[w ] ← n η(n, o i ) ← n // Add transition to FSC. Fct getNxtBeliefs(b t BR , G BR , a t i ) Ω t+1 i ← ∅ B t+1 i ← ∅ repeat e t ∼ b t BR e t+1 , o t+1 i , r t+1 ∼ G BR (e t , a t i ) if o t+1 i / ∈ Ω t+1 i then Ω t+1 i ← Ω t+1 i ∪ {o t+1 i } B t+1 [o t+1 i ] ← B t+1 [o t+1 i ] ∪ {e t+1 } until Timeout() ∨ (M inBelief P articles(B t+1 ) > N min-part ) return Ω t+1 i , B t+1 i"
},
{
"formula_coordinates": [
5,
336.18,
402.16,
193.64,
55.12
],
"formula_id": "formula_7",
"formula_text": "def = P r(s |o i , a, b) = P r(s , o i , a, b) P r(o i , a, b) = o =i O( o i , o =i |a, s ) s T (s, a, s )b(s) s ,o =i O( o i , o =i |a, s ) s T (s, a, s )b(s)"
},
{
"formula_coordinates": [
6,
64,
218.76,
188.47,
10.82
],
"formula_id": "formula_8",
"formula_text": "Ω i , B i ← getNxtBeliefs(n.b, G, n.a i , n.a =i )"
},
{
"formula_coordinates": [
6,
43.88,
242.67,
221.13,
86.23
],
"formula_id": "formula_9",
"formula_text": "if o i ∈ Ω i then // oi unexpected: add self-loop η(n, o i ) ← n else // Else: create next node: 15 b ← B i [o i ] w ← n.w * |B i [oi]| |B i | 17 if (b ∈ N ( )) ∨ (|N | = N max-fsc ) then //"
},
{
"formula_coordinates": [
6,
44.28,
381.81,
168.18,
64.51
],
"formula_id": "formula_10",
"formula_text": "a i , a =i ← P OM CP (b , G) 22 n ← node(b , a i , a =i , w ) 23 N ← N ∪ {n } 24 L[w] ← n η(n, o i ) ← n"
}
] |
2023-07-07 | [
{
"figure_ref": [],
"heading": "I. INTRODUCTION",
"publication_ref": [
"b0",
"b1",
"b2",
"b3",
"b4",
"b5",
"b6",
"b7",
"b8"
],
"table_ref": [],
"text": "Currently, \"grand challenges\" for AI research are not limited to classic boardgames like Chess, Checkers, or Go. While they still attract wide attention because of their universal cultural role, it has been shown that modern computer games may serve as milestones for AI development as well. So far, presented approaches that beat best human players in Dota 2 [1] and StarCraft II [2], are one of the most spectacular and mediaimpacting demonstrations of AI capabilities.\nThe accent is on game features that make designing successful AI players especially difficult, e.g., large action space, long term planning, imperfect information, and randomness. One game genre containing all these features is Strategy Card Games, also known as Collectible Card Games (CCGs). Besides the usual AI challenge (successful game-playing), CCGs have their own like deckbuilding and game-balancing [3].\nIn recent years, numerous research has been conducted in this domain, assisted by a few related AI competitions. The Hearthstone AI Competition [4], with the goal to develop the best agent for the game Hearthstone [5] was organized during IEEE Conference in Games in 2018, and the AAIA'17 Data Mining Challenge: Helping AI to Play Hearthstone [6] was focused on developing a scoring model for predicting win chances of a player, based on single game state data.\nIn this paper, we summarize the Strategy Card Game AI Competition (SCGAI) organized since 2019 at IEEE This work was supported by the National Science Centre, Poland under project number 2021/41/B/ST6/03691. Congress on Evolutionary Computation and IEEE Conference on Games. The competition is based on Legends of Code and Magic (LOCM) [7] programming game, designed especially for fair AI vs. AI matches. LOCM is a small implementation of a CCG, and its advantage over the commercial CCG AI engines is that it is much simpler to handle and thus allows testing more sophisticated algorithms and quickly implementing theoretical ideas. The Strategy Card Game AI Competition competition aimed to play the same role for the Hearthstone AI Competition as microRTS [8] plays for various StarCraft AI contests [9]. That is, encourage advanced research, free of drawbacks of working with the full-fledged game.\nThe last edition of the SCGAI Competition took place in 2022, and the goal of this publication is to summarize it as a whole. We start with establishing its place in the context of other CCG-related competitions, defining the characteristics behind its uniqueness, and presenting related research based on the LOCM game. We describe the rules of the game and the course of its development, showing which aspects of the game, and why, were updated during the contest's lifespan. We present a concise history of the competition, pointing out the characteristics of each edition, and listing the approaches taken by the competitors, in particular, the winners. Finally, we share our thoughts and experiences related to the competition that might be helpful to AI competition organizers in general."
},
{
"figure_ref": [],
"heading": "II. RELATED RESEARCH",
"publication_ref": [],
"table_ref": [],
"text": "We shortly present a summary of CCG-related research (not originating in LOCM and SCGAI), focusing on Hearthstonebased ones; a game used for other academic AI competitions."
},
{
"figure_ref": [],
"heading": "A. Hearthstone AI Competition",
"publication_ref": [
"b3",
"b9",
"b10",
"b11",
"b12"
],
"table_ref": [],
"text": "The Hearthstone AI Competition [4] was held three times, from 2018 to 2020, at IEEE Conference on Games. Each year it received between 30 and 50 submissions, divided between two tracks. The Premade Deck Playing track required using one of the decks prepared by the organizers: 6 decks were known upfront, while an additional 3 were used only for the final evaluation. The User Created Deck Playing track allowed agents to prepare their own deck, where the contestants used popular, user-created decks known to the Hearthstone players.\nThe competition was based on SabberStone -a Hearthstone simulator written in C# .Net Core that claimed to implement, as of 2019, when its development stopped, 98% of the base cards from the game. Creating an agent required implementing a C# class with a method that receives a game state and returns an action to perform. The complexity was reflected in a substantial time budget: 30 seconds per turn.\nWinning strategies of the submitted agents were mostly based on search algorithms: Rolling Horizon Evolution [10], MCTS [11], Pruned BFS, or Dynamic Lookahead algorithm; usually paired with a state evaluation function. For example, the best agent of 2019 was based on Information Set MCTS and sparse sampling [12]. Runner-up in 2018 used competitive coevolutionary optimization for learning heuristic evaluation function used in a greedy one-step look-ahead algorithm [13]."
},
{
"figure_ref": [],
"heading": "B. Hearthstone Data Mining Challenge",
"publication_ref": [
"b5",
"b13",
"b14",
"b15"
],
"table_ref": [],
"text": "Another interesting, although a single-time event, was the beforementioned Data Mining Challenge: Helping AI to Play Hearthstone, organized at the International Symposium on Advances in Artificial Intelligence and Applications in 2017 [6]. It lasted less than two months and attracted 188 submissions.\nThe dataset provided to participants contained examples of game states extracted from Hearthstone playouts between random AI players. The goal was to predict the winning probability of the first player based on game states and submit their predictions to the Knowledge Pit competition platform [14]. The training set given to the participants consisted of 3,250,000 game states. The test set used for final evaluation contained 750,000 states, and the results on 5% of it were known for the submitting contestants as a preliminary score.\nAll top-ranked classifiers used neural networks. The winning solution used an ensemble over a few variants of convolutional neural networks [15]. The runner-up solution was based on Logistic Regression combined with extreme gradientboosted decision trees and deep learning [16]."
},
{
"figure_ref": [],
"heading": "C. CCG Game playing",
"publication_ref": [
"b10",
"b11",
"b16",
"b17",
"b18",
"b19"
],
"table_ref": [],
"text": "Although a variety of approaches were tried, most of them took the form of MCTS [11] enhancements. The algorithm seems to work well in such a stochastic, multi-action environment, although the size of the games inspires the development of methods for reducing the action space [12]. Among many improvements described in [17], Card-Play Policy Networks can improve rollout quality and reduce their branching factor.\nHowever, full rollouts are too noisy. Thus, search is usually combined with some form of state evaluation based on expert knowledge and heuristics [18] or neural networks [19].\nA recent spectacular success in CCG AI is winning against the top 10 human player of the official Hearthstone League in China [20]. Its authors also won the last SCGAI edition; their submission is briefly described in subsection V-C2."
},
{
"figure_ref": [],
"heading": "D. CCG deckbuilding and game balancing",
"publication_ref": [
"b20",
"b21",
"b22",
"b23",
"b24",
"b25",
"b26"
],
"table_ref": [],
"text": "These are two important tasks closely related to each other. As the goal of deckbuilding is to provide a combination of cards that will be able to consistently win against a variety of opponents, game balancing can be seen as a method of ensuring that the set of such successful decks and strategies will be sufficiently large and diverse. The usual approach for these tasks is to use some form of evolution, treating cards as genes and decks as genotypes, with evaluation based on playouts between AI agents using these decks. These include standard evolutionary algorithms [21], Evolutionary Strategies [22], MAP-Elites [23], [24]. An example of multiobjective EA for game balancing focused on finding overperforming cards can be found in [25].\nAn approach tailored to the arena game mode in LOCM, extending EA with active genes to improve learning efficiency, was described in [26]. Another study analyzes the influence of representation and the choice of opponent used to test the model on the quality of learned heuristics [27]."
},
{
"figure_ref": [],
"heading": "III. LEGENDS OF CODE AND MAGIC",
"publication_ref": [
"b27",
"b28"
],
"table_ref": [],
"text": "LOCM is a CCG designed for AI research. In comparison to real-world CCGs, it has only a handful of mechanics, and all card effects are deterministic. While battling, the only source of non-determinism is one's deck order.\nIn total, there were three versions of the game used for the competitions: 1.0, 1.2, and 1.5. Each version changed the game in a backward-incompatible fashion, slightly increasing the complexity. The detailed rules are described below.\nFor each version, the organizers provided an online arena available on CodinGame as well as an offline Java referee and two faster implementations of version 1.2 in Nim and Rust. Additionally, the authors of [28], [29] implemented a set of OpenAI Gym environments."
},
{
"figure_ref": [
"fig_0",
"fig_1"
],
"heading": "A. Version 1.0",
"publication_ref": [],
"table_ref": [],
"text": "Each match starts with a draft phase, where the players build their decks in a fair arena mode. For 30 turns, they select one out of three cards (both players share the same options). The UI includes players' decks' statistics and their mana curve (histogram of cards' costs), as shown in Figure 1.\nNext, the battle phase begins. Both players start with one mana, used to play the cards. To account for the first player's advantage, the second player receives one additional mana each turn, as long as they will not use all of it in one turn. The UI of the battle phase is shown in Figure 2. Every player starts with thirty health, five runes, corresponding to 25, 20, 15, 10, and 5 health thresholds, respectively. When the player's health reaches the threshold, the rune breaks and grants an additional card draw for the next turn.\nEach turn starts with increasing the max mana by one (up to a maximum of 12), recharging mana, and drawing cards (one plus additional draw), up to a maximum of 8 in hand. If there are no cards to draw, the player loses a rune instead, and their health is reduced to its threshold. After 50 turns, both decks are considered empty.\nNext, the player can play cards if they have enough mana, attack with their creatures, and finally end their turn. The game ends when at least one player's health drops to zero or below.\nEvery card is either a creature or an item. The former can be summoned on the board; the latter used on a target. All cards share three basic attributes (attack, cost, defense), three effects (bonus draw, own health gain, and opponent's health reduction), and a subset of keywords.\nKeywords are creatures' special abilities and take effect when they battle. There are six of them: breakthrough (deal excess damage to the opponent), charge (summoned creature can attack immediately), drain (dealt damage heals the owner), guard (must be attacked first), lethal (kills damaged creature instantly), and ward (blocks all incoming damage once).\nCreature cards stay on the board as long as their defense is positive. Starting from the turn after they were summoned, they can attack the opponent or their creatures once each turn. While battling, both creatures attack simultaneously.\nItem cards are subdivided into three colors: green, red, and blue. Green items can be used on own creatures, increasing their statistics and adding new keywords. Red items can be used on enemy creatures, reducing their statistics and removing keywords. Blue items are similar to red, but additionally can be used directly on the opponent."
},
{
"figure_ref": [],
"heading": "B. Version 1.2",
"publication_ref": [],
"table_ref": [],
"text": "In contrary to Hearthstone, The Elder Scrolls: Legends have two lanes, i.e., the board is split into two parts. And while LOCM is based on the latter, version 1.0 had a single board of This not only changes the size and shape of the game tree but also impacts the importance of certain keywords. For example, creatures with Guard now protect only half of the board, and the ones with Lethal have fewer targets."
},
{
"figure_ref": [
"fig_3"
],
"heading": "C. Version 1.5",
"publication_ref": [],
"table_ref": [],
"text": "The main conclusion from versions 1.0 and 1.2 is that an agent can achieve amazing results while hardcoding the entire draft phase. This alone degenerated the game to a single phase.\nReplacing the draft with a construction phase alone would not solve this issue, as knowing all the available cards, the agents could tackle the deck construction problem offline. That is what happened in the Hearthstone AI Competition.\nTo overcome this problem, version 1.5 generates a new card set for each match. It forces agents to generalize their play style to all possible cards, including unbalanced ones. Some can effectively win most matches instantly when played (e.g., a blue item dealing 99 damage with zero cost); others can be useless (e.g., a red item dealing no damage with no keywords). Both can be used to test agents' deck construction capabilities.\nTechnically, the construction phase is a single, four second long turn, where the agents are presented with 120 cards. They can pick up to 30 cards, using at most two copies of each. In the UI, cards are shown in three frames, as shown in Figure 4.\nVersion 1.5 also introduced a new Area ability. For creature cards, it either added an extra copy on the same lane (Lane1 value) or the other lane (Lane2 value), if there is space for it. For item cards, it either applied it to all creatures in the target's lane and side of the board (Lane1), or all creatures on the target's side of the board (Lane2). The default value (Target) has no special behavior.\nFinally, version 1.5 removed the runes mechanic completely, as it was unreasonably complex for how it worked. In return, players get to draw an additional card for every 5 health lost in the previous round, preserving the aid while losing health."
},
{
"figure_ref": [],
"heading": "IV. COMPETITION",
"publication_ref": [],
"table_ref": [],
"text": "A. Legends of Code and Magic on CodinGame (2018), v1.0\nCodinGame is a challenge-based coding platform offering (among others) tens of multiplayer bot programming games. More than 25 programming languages, communication based on standard input/output using game-specific text protocols, and an in-browser coding environment allowed it to gather a sizable agent programming community.\nThe first LoCM competition, based on version 1.0, was only 24 hours long (Sprint) and gathered 742 participants 1 . The second one started two days later, lasted 30 days (Marathon), and received 2174 submissions 2 .\nThe platform's moderators, as well as the community itself, highly discourage open sourcing full agents' code, and these are not generally available. However, players share their strategies and detailed information about their thought process in so-called post-mortems on the platform's forum 3 .\nThe draft phase was dominated by handcrafted or experimentally adjusted heuristics that can be effectively implemented as a fixed ordering of cards. As it is possible to compete with everyone's agent using a given game seed, many players mimicked the top players' ordering. Some players explored applying a mana curve (balancing the number of cards with a given cost) but with no significant benefits.\nHandcrafted rule-based agents dominated the battle phase. However, the best players employed variants of well-known search methods like Minimax (few plys deep; alpha-beta and heuristic pruning) and MCTS (depth-limited with a heuristic cut-off). The most significant improvements reported by all players were move ordering and pruning, and lethal (winning) move detection."
},
{
"figure_ref": [],
"heading": "B. Strategy Card Game AI Competition (2019-2021), v1.2",
"publication_ref": [],
"table_ref": [],
"text": "The competition was no longer ran on the CodinGame platform, and thus the agent limitations changed. Most importantly, there are no programming language restrictions -the 1 CodinGame Sprint, July 25, 2018 2 CodinGame Marathon, July 27, 2018 3 Legends of Code and Magic Feedback & Strategies only requirement is compatibility with a UNIX-based system. The memory limit got lowered (256MB; agents using 1024MB or more were disqualified), and the time limit for the standard battle turns got doubled.\nTo reduce the noisiness of the results, all players played using a fixed number of randomly sampled decks ten times on the same random seed, resulting in identical card ordering. In such a setting, two deterministic agents would achieve the same results in all ten games. Additionally, every match was mirrored to account for the difference between being the first or second player.\n1) CEC 2019: The first competition received six submissions, and all of them were notably stronger than the baselines provided by the organizers. Four were rule-based agents with a variety of heuristics; two performed a proper search. The winner, Coac, based its battle phase on a minimax-like search of depth three with alpha-beta and heuristic pruning, turned out to be significantly stronger than all of the competitors (33% higher win rate than the runner-up). The draft phase used a fixed ordering of cards (the highest card was selected).\n2) COG 2019: This competition received three new submissions. The best one of them, ProphetCoac, was an attempt to improve the previous competition's winner by predicting the opponent's hand based on the cards seen during the draft to reduce the branching factor. Ultimately it did not improve but actually reduced the overall win rate, most likely due to less time available for search.\n3) CEC 2020: This competition received one update and two new submissions. The former was Coac, changing the card orderings solely, effectively improving the heuristics.\nThe new ReinforcedGreediness agent was the first to include a neural network. For the draft phase, it used two networks, one for each side, trained by self-play reinforcement learning. For the battle phase, it used a best-first search with a heuristic using Bayesian-optimized handcrafted features.\n4) COG 2020: This competition received only one new submission. In this edition, the last year's agents were not evaluated, resulting in a visible change in win rates -all were lower. The top two agents from the previous competition switched places, and an in-depth analysis suggests that all agents beat the baselines almost every time while staying relatively competitive (i.e., all agents were decent).\n5) COG 2021: This competition received four new submissions. One of them, DrainPower, had two variants -the default and the aggressive one. Both shared the same static card ordering for the draft and a flat simulation-based algorithm but had different heuristic parameters. This agent took the first two places, with the aggressive version being slightly better."
},
{
"figure_ref": [],
"heading": "C. Strategy Card Game AI Competition (2022), v1.5",
"publication_ref": [],
"table_ref": [],
"text": "As the draft phase was replaced with a deck construction one and a new area effect was introduced, the game protocol did change, and thus all of the previously submitted agents were no longer compatible. Similarly, fixed card orderings are no longer usable, as the cards are now randomly generated for each match.\nTo let agents perform an in-depth analysis of the cards, the first turn is four seconds long; this is the same as the whole draft phase before. Other limits were not changed.\nThe organizers provided only a Java-based referee, updated the CodinGame environment, and the community updated the existing OpenAI Gym environments.\n1) COG 2022: This competition was dominated by neural network-based agents -four of six submissions had one. Two of them were trained using Proximal Policy Optimization and the other two using other reinforcement learning algorithms.\nV. PLAYERS There were 22 unique agents submitted in total. Based on the main algorithm used for playing, we divided them into three groups -search-based, neural network-based, and other. Interestingly, most agents in every group have a similar performance characteristic, e.g., they use a similar amount of time and memory while playing. Three baseline agents provided by the organizers are described in a dedicated subsection."
},
{
"figure_ref": [],
"heading": "A. Baselines",
"publication_ref": [],
"table_ref": [],
"text": "The provided baseline agents were built with two goals in mind: presenting the game rules and providing a better training opponent than a fully random one. They are fairly trivial and rely on the game engine ignoring invalid actions.\n1) Baseline 1: This agent, written in Python, was used for the competitions running version 1.2 of the game. During the draft phase, it focuses on selecting creature cards with Guard keyword and falls back to the first card otherwise. For the battle phase, it follows a rule-based algorithm focused on using all cards in hand and on the board, attacking the opponent's first and their creatures next.\n2) Baseline 2: This agent, written in Python, was used for the competitions running version 1.2 of the game. During the draft phase, it focuses on selecting creature cards with the highest attack and falls back to the first card otherwise. For the battle phase, it uses a one-pass algorithm, attacking the opponent using all summoned creatures and summoning all creatures from hand.\n3) RandomWItems2lanes: This agent, written in Java, was used for the competitions running version 1.5 of the game. In the construction phase, it selects all cards at random. For the battle phase, it uses all green items on its own creatures, attacks the opponent (only creatures with the Guard keyword, if there is any), summons all creatures, and finally uses all items. All actions are targeted at random."
},
{
"figure_ref": [],
"heading": "B. Search-based",
"publication_ref": [
"b29",
"b30",
"b31"
],
"table_ref": [],
"text": "Search-based agents were the most common and won all competitions running version 1.2. All of them employed move pruning either explicitly (certain actions were disallowed) or implicitly (moves were generated out of ordered actions). Similarly, most of them implemented lethal move detection.\nOnly some of the agents searched through the opponent's turn, and it is clear that while it is beneficial to do so, it requires further pruning and heuristic evaluation to make it feasible with the exploding branching factor. 1) AdvancedAvocadoAgent: This agent, written in Java, was submitted for COG 2021. It uses a fixed card ordering for the draft phase and a best-first search for battling. Weights of the heuristic evaluation function were found offline, using an MCTS-based search over the space of parameters.\n2) Chad: This agent, written in Rust, was submitted for CEC 2020 and described in [30]. It scores the cards using weights computed with harmony search. The battle phase uses MCTS with the opponent's hand prediction.\n3) Coac: This agent, written in C++, was submitted for CEC 2019 and got updated the next year. For the draft phase, it uses a fixed card ordering. For the battle phase, a Minimaxlike search of depth three (or less, if the time ran out or the tree was too wide) with alpha-beta and heuristic pruning.\n4) DrainPower: This agent, written in C#, was submitted for COG 2021. Both phases base on heuristic card evaluation. The battle phase uses a flat simulation-based algorithm, simulating own turn and the opponent's response. It has two variants -default and aggressive -with different weights for the heuristic evaluation function.\n5) Fabbiamo: This agent, written in C++, was submitted for COG 2019. During the draft, it follows a so-called mana curve, i.e., it tries to maintain a reasonable number of cards of the same cost. The battle phase uses a Minimax search of depth four over own actions and depth one of the opponent's.\n6) LANE_1_0: This agent, written in Java, was submitted for COG 2021. The draft phase follows a fixed card ordering. The battle phase uses a flat simulation-based algorithm, simulating only its own turn.\n7) Marasbot: This agent, written in C++, was submitted for CEC 2019. It uses a heuristic evaluation function for both phases and random sampling for the battle phase. It completely ignores blue items.\n8) OneLaneIsEnough: This agent, written in C++, was submitted for COG 2020. It uses a heuristic evaluation function for both phases and a one-turn deep search, including the opponent's response, for the battle phase. 9) Prophet Coac: This agent, written in C++, was submitted for COG 2019. It is a modification of the Coac agent, including a tentative prediction of the opponent's hand, based on the cards seen during the draft phase and refined by the already played cards.\n10) UJIAgent3: This agent, written in Python, was submitted for COG 2019 and partially described in [31]. It uses a fixed card ordering for the draft phase. For the battle phase, it uses Online Evolutionary Planning [32] where the genome encodes a series of actions, and mutation only reorders them.\n11) Zylo: This agent, written in Java, was submitted for COG 2022. It uses a heuristic evaluation function for both phases and a best-first search for the battle phase. Its parameters were tuned using an evolutionary algorithm."
},
{
"figure_ref": [],
"heading": "C. Neural networks-based",
"publication_ref": [
"b32",
"b33"
],
"table_ref": [],
"text": "Neural networks-based agents were less common in version 1.2 competitions but dominated the 1.5 one. While there was no GPU available while playing, some were trained using one.\n1) Ag2O: This agent, written in Python, was submitted for COG 2021. The draft phase bases on card weights trained with q-learning and takes card combinations into consideration. The battle phase uses a best-first search guided by actions' q-value. The network composes of four dense layers using tanh as the activation function between and at the end.\n2) ByteRL: This agent, written in Python, was submitted for COG 2022 and described in [33]. Interestingly, it uses only one, end-to-end policy, trained using deep reinforcement learning combined with optimistic smooth fictitious play. The network architecture is rather complex and includes a Long Short-Term Memory (LSTM) block.\nThe authors sent an additional version of this agent, adjusted and trained for the 1.2 version of LoCM. It was compared against all COG 2021 submissions and won by a large margin (more than 20% higher win rate than the runner-up).\n3) Inspirai: This agent, written in Python, was submitted for COG 2022. Its construction phase uses a heuristic evaluation function adapted from Coac and optimized using Bayesian optimization. The battle phase uses a neural network trained using Proximal Policy Optimization.\nThe network consists of a dense layer with ReLU activation function, followed by another dense layer. On top of that, there are two heads using attention [34] for card and target selection.\n4) NeteaseOPD: This agent, written in Python, was submitted for COG 2022. It uses two independent networks trained using Proximal Policy Optimization, one for each phase.\n5) ReinforcedGreediness: This agent, written in Python, was submitted for COG 2021. The draft phase uses a neural network learned by self-play reinforcement learning, trained independently for both sides. The network consists only of dense layers with a few different activation layers.\nThe battle phase uses best-first search limited to own turn; the heuristic evaluation function is a linear combination of hand-made features optimized using Bayesian optimization.\n6) USTC gogogo: This agent, written in Python, was submitted for COG 2022. There were two versions -one using hyperparameters to control both construction and battle phases and the other using model trained using reinforcement learning. Only the latter was used for the competition."
},
{
"figure_ref": [],
"heading": "D. Other",
"publication_ref": [],
"table_ref": [],
"text": "Other agents were the simplest, usually with a decently sized set of handcrafted rules and heuristic evaluation functions at their core. Due to their simplicity, most of them acted instantly, within a few milliseconds per turn.\n1) AntiSquid: This agent, written in Python, was submitted for CEC 2019. During the draft, it selects cards using a fixed card ordering, follows a mana curve, and prefers different cards based on the already selected ones. For the battle phase, it follows a rule-based algorithm, searching for the highest scored sequences of moves.\n2) Conrisc: This agent, written in JavaScript, was submitted for CEC 2019. For the draft, it follows a heuristic evaluation function. The battle phase uses a rule-based algorithm, using all cards in a given order. It ignores all items.\n3) MugenSlayerAttackOnDuraraBallV3: This agent, written in Python, was submitted for COG 2022. During the construction phase, it focuses on the cheapest cards with the newly added area effect. For the battle phase, it follows a list of predefined rules.\n4) UJIAgent1: This agent, written in Python, was submitted for CEC 2019. During the draft phase, it tries to gather a predefined set of cards based on their type and cost. For the battle phase, it follows a list of predefined rules ordered using a heuristic evaluation function.\n5) UJIAgent2: This agent, written in Python, was submitted for CEC 2019. During the draft phase, it probabilistically tries to gather a predefined set of cards. For the battle phase, it samples 44 strategies and picks the one with the highest score."
},
{
"figure_ref": [],
"heading": "VI. TAKEAWAYS FOR COMPETITION ORGANIZERS",
"publication_ref": [
"b34"
],
"table_ref": [],
"text": "While an AI competition is a challenge for the contestants, it is also a challenge for the organizers to make it a successful one [35]. It is unclear when a competition becomes successful, but there are several things the organizers have to account for, ranging from coming up with an interesting problem itself to running the final evaluation."
},
{
"figure_ref": [],
"heading": "A. Games as test beds",
"publication_ref": [],
"table_ref": [],
"text": "Modeling problems using games makes them more approachable for most people, especially when the game itself exists (e.g., Hearthstone for the Hearthstone AI Competition) or is a simplified version of one (e.g., LOCM for The Elder Scroll: Legends, MicroRTS for StarCraft). Additionally, the game's player base is often an excellent source of battle-tested playing strategies and their analyses.\nAs different games raise different problems, it is crucial to maintain some variety. A new competition in a game genre not seen before may bring relatively a lot of novel algorithms or methods. Similarly, extending an existing game can revive stagnated research or make a more versatile benchmark.\nAt the same time, new problems attract fewer contestants, as they are usually less prestigious or not marketed enough. Luckily, games are usually flexible, and it is often possible to extend them in a backward-compatible way. That allows the competition organizers to reuse the previous submissions."
},
{
"figure_ref": [],
"heading": "B. Bootstrapping with CodinGame",
"publication_ref": [],
"table_ref": [],
"text": "The difference in the number of submissions between the CodinGame and academia SCGAI competitions is almost tenfold, while the game rules remained fairly similar across all three versions. The CodinGame platform has a large and vivid community; using it as an \"incubator\" of an academic competition may be a great way of validating one's idea.\nIt is important to emphasize that the CodinGame community focuses on solving the problem (i.e., playing the game). It does not necessarily imply novel algorithms or sophisticated methods. Actually, the vast majority is the complete opposite: highly optimized versions of well-known algorithms tuned for a specific game or puzzle.\nHowever, CodinGame has some limitations too -the game engine has to be written in Java, the communication has to be text-based, and the agents are evaluated in a highly restricted environment (1 CPU core, 768MB of RAM, no GPU). On top of that, the entire game can use at most 30 seconds of summarized agents execution time."
},
{
"figure_ref": [],
"heading": "C. Pushing for research",
"publication_ref": [],
"table_ref": [],
"text": "Arguably, all AI competitions should require submissions to be documented. On the one hand, it helps the organizers to compare and reason about them, without analyzing the source code. On the other, it allows future contestants to learn from and improve them, instead of starting from scratch. Ideally, all agents would result in a paper.\nIn the first two years of the SCGAI competition, there were two competitions each year. While it increases the visibility and potentially brings more contestants in, it may result in the opposite -instead of one competition with six submissions, there will be two with three each. The SCGAI competition partially solved this issue by automatically resubmitting past agents for future competitions.\nMarketing the competition among students may bring many valuable submissions, usually well-documented ones. Some of them will find participating more interesting than working on an unrelated project, and thus become more involved. Such projects can evolve into diploma theses and then papers.\nAdditionally, contestants can be encouraged with prizes. As the organizers may not want to sponsor them themselves, they can reach out to conference organizers. IEEE CIS sponsored SCGAI competitions at both CEC and COG conferences."
},
{
"figure_ref": [],
"heading": "D. Taming the randomness",
"publication_ref": [],
"table_ref": [],
"text": "Most real-life games, and so their toy-scale relatives, are highly random. To ensure fair results, one can use the same game seeds for all agent pairs (e.g., in LOCM, it results in the same cards available during their games). Additionally, as most agents are non-deterministic on their own, each game seed can be used multiple times to average the result.\nSimilarly, most games are asymmetrical, giving one of the players a visible advantage. To account for that, one can run matches in all player configurations using the same game seed.\nIf a competition runs on multiple machines, or even on one but in parallel, consider interleaving instead of concatenating the results from across runs. Different agents can utilize the CPU differently, leading to varying results depending on the programs running in the background."
},
{
"figure_ref": [],
"heading": "E. Hardware and software",
"publication_ref": [
"b2",
"b35"
],
"table_ref": [],
"text": "Comparing all agents pair-wise using a reasonable number of matches requires a notable amount of CPU time. Luckily, most cloud providers (e.g., DigitalOcean) are keen to sponsor research and related competitions at a small cost of mentioning them or their services while presenting the results.\nTo let others reproduce the competition results, all of the code used to run it should be published once it concludes. Most importantly, all kinds of configurations and the dependencies required by all agents should be well documented.\nThe same applies to the final results. Win rates and charts are enough for a presentation, but organizers should be fully transparent and share both the raw data and the scripts used to aggregate it. As the intermediate results can be significant in size, consider compressing them.\nAutomating the above points allows the organizers to focus on the competition and the AI challenge itself instead of building the infrastructure and tooling for every single contest. The SCGAI repository meets all these criteria, presenting all agents, tooling, and results in one place. To prevent stagnation and introduce rising levels of challenge, the game rules have been extended a few times during this time. The academia-based editions of the contest gathered 22 challengers, coming with different approaches, from simple rule-based solutions to search-based agents and ones applying deep reinforcement learning. The game has been a base for a number of publications concerning playing algorithms and the deckbuilding problem.\nWe hope that this summary will serve as a reference point for the competition and related achievements, as well as a general source of knowledge about successful approaches and types of challenges characteristic of the domain of CCGs. We also hope that our observations and conclusions will be helpful to other researchers (or companies) that plan to bring to the AI community some new game-oriented challenges.\nWe especially look forward to the next competitions related to (Collectible) Card Games, as the domain is so broad that in all these years and competitions, it has been studied only superficially, and there are still many challenges left [3]. One new CCG-related contest we know about is Tales of Tribute AI Competition [36], based on the deckbuilding card game Tales of Tribute and advertised for IEEE COG 2023.\nCurrently, there are no plans to organize the SCGAI competition any further, so it is officially considered close. However, thanks to CodinGame, all LOCM versions are available online as bot programming games, and everyone can compete against the agents available on the public leaderboards."
}
] | This paper concludes five years of AI competitions based on Legends of Code and Magic (LOCM), a small Collectible Card Game (CCG), designed with the goal of supporting research and algorithm development. The game was used in a number of events, including Community Contests on the CodinGame platform, and Strategy Card Game AI Competition at the IEEE Congress on Evolutionary Computation and IEEE Conference on Games. LOCM has been used in a number of publications related to areas such as game tree search algorithms, neural networks, evaluation functions, and CCG deckbuilding. We present the rules of the game, the history of organized competitions, and a listing of the participant and their approaches, as well as some general advice on organizing AI competitions for the research community. Although the COG 2022 edition was announced to be the last one, the game remains available and can be played using an online leaderboard arena. | Summarizing Strategy Card Game AI Competition | [
{
"figure_caption": "Fig. 1 :1Fig. 1: Draft phase in version 1.0 and 1.2 of LOCM. Available cards are in the center. Above and under it are players' decks' statistics and their mana curves.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 2 :2Fig. 2: Battle phase in version 1.0 of LOCM. Basic players' info is on the left; their hands and the board are on the right.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 3 :3Fig. 3: Battle phase in version 1.2 and 1.5 of LOCM. Contrary to version 1.0, the board is now split into two lanes.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 4 :4Fig. 4: Construction phase in version 1.5 of LOCM. The red and blue numbers in the card's top left and right corners indicate how many copies of it each player picked.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "VII. CONCLUSION This paper summarizes five years of Strategy Card Game AI Competition -an AI programming challenge based on Legends of Code and Magic, a small implementation of a Collectible Card Game authored by Jakub Kowalski and Radosław Miernik. Its novel fair arena mode stays in opposition to more common collection-based deckbuilding and the simplicity of rules allows search-based approaches to be more profound. LOCM was first used in CodinGame 5th Community Contest in 2018 and last at IEEE Conference on Games 2022.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Results of all Strategy Card Game AI competitions. New and updated agents' names are in bold. Baseline agents' names are in italics.",
"figure_data": "YearPlace Win rate AgentIEEE CEC 2019v1.2",
"figure_id": "tab_0",
"figure_label": "I",
"figure_type": "table"
}
] | Jakub Kowalski; Radosław Miernik | [
{
"authors": " Openai",
"journal": "",
"ref_id": "b0",
"title": "OpenAI Five",
"year": "2017"
},
{
"authors": "O Vinyals; I Babuschkin; J Chung; M Mathieu; M Jaderberg; W M Czarnecki; A Dudzik; A Huang; P Georgiev; R Powell; T Ewalds; D Horgan; M Kroiss; I Danihelka; J Agapiou; J Oh; V Dalibard; D Choi; L Sifre; Y Sulsky; S Vezhnevets; J Molloy; T Cai; D Budden; T Paine; C Gulcehre; Z Wang; T Pfaff; T Pohlen; D Yogatama; J Cohen; K Mckinney; O Smith; T Schaul; T Lillicrap; C Apps; K Kavukcuoglu; D Hassabis; D Silver",
"journal": "",
"ref_id": "b1",
"title": "AlphaStar: Mastering the Real-Time Strategy Game StarCraft II",
"year": "2019"
},
{
"authors": "A K Hoover; J Togelius; S Lee; F De Mesentier; Silva",
"journal": "KI-Künstliche Intelligenz",
"ref_id": "b2",
"title": "The Many AI Challenges of Hearthstone",
"year": "2020"
},
{
"authors": "A Dockhorn; S Mostaghim",
"journal": "",
"ref_id": "b3",
"title": "Introducing the Hearthstone-AI Competition",
"year": "2019"
},
{
"authors": "B Entertainment; Hearthstone",
"journal": "Blizzard Entertainment",
"ref_id": "b4",
"title": "",
"year": "2004"
},
{
"authors": "A Janusz; T Tajmajer; M Świechowski",
"journal": "FedCSIS",
"ref_id": "b5",
"title": "Helping AI to Play Hearthstone: AAIA'17 Data Mining Challenge",
"year": "2017"
},
{
"authors": "J Kowalski; R Miernik",
"journal": "",
"ref_id": "b6",
"title": "Legends of Code and Magic",
"year": "2018"
},
{
"authors": "S Ontañón",
"journal": "",
"ref_id": "b7",
"title": "The Combinatorial Multi-armed Bandit Problem and Its Application to Real-time Strategy Games",
"year": "2013"
},
{
"authors": "D Churchill; M Preuss; F Richoux; G Synnaeve; A Uriarte; S Ontañnón; M Čertickỳ",
"journal": "Encyclopedia of Computer Graphics and Games",
"ref_id": "b8",
"title": "Starcraft bots and competitions",
"year": "2016"
},
{
"authors": "J Liu; D Pérez; S Lucas",
"journal": "CEEC",
"ref_id": "b9",
"title": "Rolling Horizon Coevolutionary planning for two-player video games",
"year": "2016"
},
{
"authors": "C B Browne; E Powley; D Whitehouse; S M Lucas; P I Cowling; P Rohlfshagen; S Tavener; D Perez; S Samothrakis; S Colton",
"journal": "TCIAIG",
"ref_id": "b10",
"title": "A Survey of Monte Carlo Tree Search Methods",
"year": "2012"
},
{
"authors": "J S B Choe; J.-K Kim",
"journal": "COG",
"ref_id": "b11",
"title": "Enhancing Monte Carlo Tree Search for Playing Hearthstone",
"year": "2019"
},
{
"authors": "P García-Sánchez; A Tonda; A J Fernández-Leiva; C Cotta",
"journal": "Knowledge-Based Systems",
"ref_id": "b12",
"title": "Optimizing hearthstone agents using an evolutionary algorithm",
"year": "2020"
},
{
"authors": "A Janusz; D Slezak; S Stawicki; M Rosiak",
"journal": "CS&P",
"ref_id": "b13",
"title": "Knowledge Pit-A Data Challenge Platform",
"year": "2015"
},
{
"authors": "Ł Grad",
"journal": "FedCSIS",
"ref_id": "b14",
"title": "Helping AI to play Hearthstone using neural networks",
"year": "2017"
},
{
"authors": "Q H Vu; D Ruta; L Cen",
"journal": "FedCSIS",
"ref_id": "b15",
"title": "An ensemble model with hierarchical decomposition and aggregation for highly scalable and robust classification",
"year": "2017"
},
{
"authors": "S Zhang; M Buro",
"journal": "CIG",
"ref_id": "b16",
"title": "Improving Hearthstone AI by learning highlevel rollout policies and bucketing chance node events",
"year": "2017"
},
{
"authors": "A Santos; P A Santos; F S Melo",
"journal": "",
"ref_id": "b17",
"title": "Monte Carlo tree search experiments in Hearthstone",
"year": "2017"
},
{
"authors": "M Świechowski; T Tajmajer; A Janusz",
"journal": "",
"ref_id": "b18",
"title": "Improving Hearthstone AI by Combining MCTS and Supervised Learning Algorithms",
"year": "2018"
},
{
"authors": "C Xiao; Y Zhang; X Huang; Q Huang; J Chen; P Sun",
"journal": "",
"ref_id": "b19",
"title": "Mastering Strategy Card Game (Hearthstone) with Improved Techniques",
"year": "2023"
},
{
"authors": "P García-Sánchez; A Tonda; G Squillero; A Mora; J J Merelo",
"journal": "CIG",
"ref_id": "b20",
"title": "Evolutionary deckbuilding in Hearthstone",
"year": "2016"
},
{
"authors": "A Bhatt; S Lee; F De Mesentier; C W Silva; J Watson; A K Togelius; Hoover",
"journal": "FDG",
"ref_id": "b21",
"title": "Exploring the Hearthstone deck space",
"year": "2018"
},
{
"authors": "M C Fontaine; S Lee; L B Soros; F De Mesentier; J Silva; A K Togelius; Hoover",
"journal": "",
"ref_id": "b22",
"title": "Mapping Hearthstone Deck Spaces Through MAPelites with Sliding Boundaries",
"year": "2019"
},
{
"authors": "Y Zhang; M C Fontaine; A K Hoover; S Nikolaidis",
"journal": "",
"ref_id": "b23",
"title": "Deep surrogate assisted map-elites for automated hearthstone deckbuilding",
"year": "2022"
},
{
"authors": "F De Mesentier; R Silva; S Canaan; M C Lee; J Fontaine; A K Togelius; Hoover",
"journal": "COG",
"ref_id": "b24",
"title": "Evolving the Hearthstone meta",
"year": "2019"
},
{
"authors": "J Kowalski; R Miernik",
"journal": "",
"ref_id": "b25",
"title": "Evolutionary Approach to Collectible Card Game Arena Deckbuilding using Active Genes",
"year": "2020"
},
{
"authors": "R Miernik; J Kowalski",
"journal": "",
"ref_id": "b26",
"title": "Evolving Evaluation Functions for Collectible Card Game AI",
"year": "2022"
},
{
"authors": "R Vieira; A Tavares; L Chaimowicz",
"journal": "SBGames",
"ref_id": "b27",
"title": "Drafting in Collectible Card Games via Reinforcement Learning",
"year": "2020"
},
{
"authors": "",
"journal": "Entertainment Computing",
"ref_id": "b28",
"title": "Exploring reinforcement learning approaches for drafting in collectible card games",
"year": "2023"
},
{
"authors": "M Witkowski; Ł Klasiński; W Meller",
"journal": "",
"ref_id": "b29",
"title": "Implementation of collectible card Game AI with opponent prediction",
"year": "2020"
},
{
"authors": "R Montoliu; R D Gaina; D Perez; D Delgado; S Lucas",
"journal": "EvoSTAR",
"ref_id": "b30",
"title": "Efficient heuristic policy optimisation for a challenging strategic card game",
"year": "2020"
},
{
"authors": "N Justesen; T Mahlmann; J Togelius",
"journal": "EvoCOP",
"ref_id": "b31",
"title": "Online evolution for multiaction adversarial games",
"year": "2016"
},
{
"authors": "W Xi; Y Zhang; C Xiao; X Huang; S Deng; H Liang; J Chen; P Sun",
"journal": "",
"ref_id": "b32",
"title": "Mastering Strategy Card Game (Legends of Code and Magic) via End-to-End Policy and Optimistic Smooth Fictitious Play",
"year": "2023"
},
{
"authors": "A Vaswani; N Shazeer; N Parmar; J Uszkoreit; L Jones; A N Gomez; Ł Kaiser; I Polosukhin",
"journal": "NeurIPS",
"ref_id": "b33",
"title": "Attention is All you Need",
"year": "2017"
},
{
"authors": "J Togelius",
"journal": "TCIAIG",
"ref_id": "b34",
"title": "How to Run a Successful Game-Based AI Competition",
"year": "2016"
},
{
"authors": "J Kowalski; R Miernik; K Polak; D Budzki; K D ",
"journal": "",
"ref_id": "b35",
"title": "Introducing Tales of Tribute AI Competition",
"year": "2023"
}
] | [] |
|
2023-11-21 | [
{
"figure_ref": [
"fig_0",
"fig_1",
"fig_2"
],
"heading": "Input images",
"publication_ref": [
"b8",
"b22",
"b25",
"b37",
"b8",
"b21",
"b38",
"b22",
"b37",
"b15",
"b42",
"b27",
"b44",
"b12",
"b10",
"b1",
"b40",
"b39",
"b27",
"b12",
"b44",
"b44",
"b23",
"b14",
"b31",
"b11",
"b29",
"b18",
"b0",
"b17",
"b2",
"b43",
"b6",
"b0",
"b26",
"b45",
"b31",
"b24",
"b49"
],
"table_ref": [],
"text": "1 Introduction Image completion (Criminisi et al., 2003;Li et al., 2022;Lugmayr et al., 2022), involving the concealment of a portion of an image and prompting a model to imaginatively restore it, has long been a subject of extensive research with many applications, such as object removal (Suvorov et al., 2022;Criminisi et al., 2003), image compositing (Levin et al., 2004), photo restoration (Wan et al., 2020), etc. Typical image completion approaches (Li et al., 2022;Suvorov et al., 2022) are prone to struggle with complex or large masking regions due to inadequate reference information. This limitation causes ambiguity to completion model over restoration or elimination and leads to noticeable artifacts in completed images, degrading the quality.\nAn intuitive solution to overcome the above limitation is to incorporate user-input (Horita et al., 2022;Yu et al., 2019;Zheng et al., 2022a) or prediction-based (Nazeri et al., 2019;Yu et al., 2022;Guo et al., 2021;Dong et al., 2022) guidance, e.g., text (Avrahami et al., 2023;Xie et al., 2022;Nichol et al., 2022;Wang et al., 2023), edge (Nazeri et al., 2019;Guo et al., 2021;Yu et al., 2022), or segmentation (Yu et al., 2022;Liao et al., 2020;Zheng et al., 2022b), into image completion. However, these approaches are limited to performing image completion under only single-modality guidance, which is inflexible in employing the multi-modality, especially more than two modalities simultaneously, for plausible generation and leads to limited application scenarios.\nRecently, denoising diffusion probabilistic model (Ho et al., 2020) has been widely employed and demonstrated superior performances in text-to-image synthesis (Rombach et al., 2022;Gu et al., 2022;Ramesh et al., 2022) and text-driven image manipulation fields (Kim et al., 2022;Avrahami et al., 2022;Kawar et al., 2023). In addition to text, many approaches (Bansal et al., 2023;Yu et al., 2023;Chen et al., 2023;Avrahami et al., 2022) have explored the integration of extra guidance modality, such as segmentation, sketch, pose, and even position of generated object, into diffusion models in a training-free way. These methods involve designing energy loss associated with the input guidance and guiding its gradient on the latent codes during inference, yet they tend to fail in maintaining fine-grained structural information, resulting in insufficient control over the generated results. Meanwhile, several training-required approaches (Mou et al., 2023;Zhang & Agrawala, 2023) have further enhanced the control of input modality over diffusion models by introducing an auxiliary conditional network to encode modality and directly add the encoded features to the intermediate features of frozen diffusion models. These methods bring in fresh insights and pave the way for incorporating guidance signals into image completion. Nevertheless, simply transferring these ideas to multi-modality image completion is not trivial, as the introduction of each new modality necessitates the joint training of all auxiliary conditional networks. How to effectively integrate multi-modality guidance for image completion in a scalable and flexible manner remains an open problem.\nIn this paper, we propose MaGIC, a novel, simple yet effective framework for Multi-modality Guided Image Completion, especially when there are more than two modalities at the same time. MaGIC is designed to be scalable and flexible, allowing it to merge various modalities, including but not limited to text, canny edge, sketch, segmentation, depth, and pose, in an arbitrary combination as guidance for image completion (see Fig. 1 and Fig. 2). To build MaGIC, there are two core ingredients, including a modality-specific conditional U-Net (MCU-Net) and a consistent modality blending (CMB) method, performed in two stages.\nSpecifically, the proposed MCU-Net, composed of a standard U-Net denoiser from the pre-trained stable diffusion (Rombach et al., 2022) and a simple encoding network, which injects a single modality guidance signal into the U-Net denoiser to attain single-modal guided completion. The MCU-Net will be individually finetuned under each single modality, in the first stage. Then, to achieve multi-modality guidance, the CMB algorithm is proposed in the second stage to flexibly aggregate guidance signals from any combination of previously learned MCU-Nets. The CMB leverages guidance loss to gradually narrow the distances between the intermediate features from the original pretrained U-Net denoiser and multiple MCU-Nets during denoising sample stage, which ensures that the former features do not deviate too much from the original feature distribution during multi-modality guidance. Compared with the naive approach of achieving multi-modality guided completion by jointly re-training a unified model, our CMB is training-free and allows for the flexible addition or removal of guidance modalities, avoiding cumbersome re-training and preserving the feature distribution of the original U-Net denoiser. To verify the proposed MaGIC, we conduct extensive experiments on various tasks including image inpainting, outpainting, and real user-input editing, using the COCO (Lin et al., 2014), Places2 (Zhou et al., 2018), and in-the-wild data. Our results demonstrate the superiority of MaGIC over image completion and controllable generation baselines in terms of image quality. In addition, we find that, surprisingly, the CMB of our MaGIC is also well applicable for multi-modality guided image generation, showing its generality and potential for generative tasks. Fig. 3 illustrates the architecture of our approach.\nIn summary, our contributions are four-fold: (i) we propose a novel approach of MaGIC for flexible and scalable multi-modality guided image completion. To the best of our knowledge, MaGIC is the first to widely support arbitrary multi-modality guided image completion; (ii) we present a simple yet effective MCU-Net to effectively and adaptively inject a modality as guidance for image completion;\n(iii) we introduce a novel CMB algorithm that combines arbitrary multiple modalities for image completion without the need for additional training and (iv) using MaGIC, we achieve performance superior to that of other state-of-the-art approaches."
},
{
"figure_ref": [
"fig_0"
],
"heading": "Related Work",
"publication_ref": [
"b27",
"b12",
"b10",
"b23",
"b14",
"b9",
"b31",
"b39",
"b1",
"b0",
"b14",
"b9",
"b31",
"b35",
"b31",
"b45",
"b26",
"b43",
"b6",
"b2",
"b16",
"b16",
"b6",
"b2",
"b2",
"b43",
"b2",
"b43",
"b43"
],
"table_ref": [],
"text": "Auxiliary-based image completion. The auxiliary-based image completion methods aim to enhance the structure and texture of completed images by incorporating predicted or human-provided prior information. Early approaches primarily focus on using a single modality (e.g., edge (Nazeri et al., 2019;Guo et al., 2021;Dong et al., 2022;Zheng et al., 2022a) or segmentation (Zheng et al., 2022b;Liao et al., 2020)) as the auxiliary guidance for image completion. Recently, inspired by the superiorperforming diffusion models (Ho et al., 2020;Dhariwal & Nichol, 2021;Rombach et al., 2022), text-based auxiliary solutions have been proposed for image completion (Wang et al., 2023;Avrahami et al., 2023;Nichol et al., 2022;Avrahami et al., 2022), providing more user-friendly image editing applications.\nHowever, prompt text alone is not sufficient. Due to the above methods are constrained by the training requirements of auxiliary guidance, making them difficult to flexibly add more types of modalities as guidance for completion. Our MaGIC can incorporate random combination of multiple modalities for more plausible completion result (see Fig. 1 again). It is versatile, requiring only the optimization of single-modality conditional networks, and allows for plug-and-play integration into the conditional image completion process without the need for additional cumbersome joint re-training.\nControllable image generation with diffusion models. Diffusion models (Ho et al., 2020;Dhariwal & Nichol, 2021;Rombach et al., 2022;Song et al., 2023) have drawn extensive attention in image generation owing to their remarkable results and stable training. These methods can be broadly categorized into train-required and train-free approaches. The former achieve powerful generation control by training on large-scale data or fine-tuning a conditional control sub-network on pre-trained diffusion models (e.g., (Rombach et al., 2022)). Recent research (Zhang & Agrawala, 2023;Mou et al., 2023) has introduced various modalities (e.g., keypose point maps, sketch maps, etc) for generation. However, it fails to simultaneously use multi-modality as guidance. Differently, train-free solutions (Yu et al., 2023;Chen et al., 2023;Bansal et al., 2023;Jeong et al., 2023) leverage the multi-step nature of diffusion models, explicitly introducing guidance signals during the iterative denoising process and achieving style (Jeong et al., 2023), layout (Chen et al., 2023;Bansal et al., 2023), face identity (Bansal et al., 2023;Yu et al., 2023), segmentation map (Bansal et al., 2023;Yu et al., 2023) guidance without fine-tuning. Yet, they struggle to leverage fine-grained structural guidance (e.g., canny edge) as conditions, potentially resulting in degraded guidance (Yu et al., 2023).\nOur MaGIC is inspired by the above image generation approaches, but different in two aspects. First, MaGIC achieves multi-modality guidance without joint re-training while improving the effectiveness of fine-grained structure guidance. In addition, MaGIC goes beyond controllable generation and can be applied to guided completion and real-world editing tasks.\n3 MaGIC: Multi-modality Guided Image Completion Masked images x m = x ⊙ m are obtained by corrupting images x with binary masks m ∈ {0, 1} H×W×1 , where x ∈ R H×W×3 are original RGB images with width W and height H. Given a known region x m = x ⊙ (1m), the goal of image completion is to learn a function p(x m |x m ) that completes the missing mask area with visually realistic and structurally coherent content. To mitigate the inherent ambiguity of completion model, the direction of restoration or elimination is controlled through the auxiliary guidance C. In the following sections, we start by outlining necessary diffusion steps in 3.1) for formulating our method, then elaborate on MaGIC, addressing auxiliary guidance via our proposed MCU-Net in 3.2 and multi-modality integration by our CMB algorithm in 3.3."
},
{
"figure_ref": [],
"heading": "Preliminaries",
"publication_ref": [
"b14",
"b9",
"b13",
"b31",
"b31",
"b32",
"b45",
"b1",
"b34"
],
"table_ref": [],
"text": "Diffusion models. Denoising diffusion probabilistic models (DDPMs) (Ho et al., 2020) are generative models that learn the true distribution p(x T ) by iteratively denoising a randomly sampled noise image x T . In each denoising step, a U-Net model is trained to predict the noise ϵ based on the objective function, Φ(x t , t, θ) = min(E x 0 ,t,ϵ∼N(0,I) ∥ϵ -\nϵ t θ (x t )∥ 2 2 ),(1)\nwhere x t = √ α t x 0 + √ 1 -α t ϵ represents the intermediate noised image obtained after applying noise t times to the clean image x 0 , and α t = t s=1 (1 -β s ) is a series of fixed hyperparameters based on the variance schedule β s , s ∈ [1, T ]. The model can be further generalized to conditional generation (Dhariwal & Nichol, 2021;Ho & Salimans, 2021), with predicted noise becoming ϵ θ (x t , t, C).\nStable diffusion. We consider stable diffusion (SD) inpainting model (Rombach et al., 2022) as the main backbone in the subsequent method sections. Instead of beginning with isotropic Gaussian noise samples in pixel space, the SD model first maps clean images to their corresponding latent space Z through E(•). Here, E(•) is an autoencoder with a left inverse D, ensuring x = D • E(x). Owing to the lower inference overhead of U-Net in the latent space, SD has emerged as an important class of recent image generators based on diffusion (Rombach et al., 2022;Saharia et al., 2022;Zhang & Agrawala, 2023;Avrahami et al., 2023). Specifically, the initial latent codes of iterative denoising process employ random z\nT ∼ Z ∈ R H s × W s ×3\n, where s signifies s-fold reduction in spatial dimensions. The mask and encoding masked image serve as conditions for the U-Net, modifying the objective function in Eq. 1 to\nΦ(z t , t, m ↓ , x m↓ , θ) = min(E z 0 ,t,ϵ∼N(0,I) ∥ϵ -ϵ t θ (z t , m ↓ , x m↓ )∥ 2 2 ),(2)\nwhere m ↓ ∈ R (DDIM) (Song et al., 2021) defines the each step of denoising as a non-Markovian process while retaining the same training objective as DDPM. Accordingly, the sampling process is formulated as,\nz t-1 = √ α t-1 ( z t - √ 1 -α t ϵ t θ (z t , m ↓ , x m↓ ) √ α t ) + 1 -α t-1 -σ 2 t • ϵ t θ (z t , m ↓ , x m↓ ) + σ t ϵ t ,(3)\nwhere the noise ϵ t follows the standard normal distribution N(0, I) and is independent of x t , and\nσ t = η √ (1 -α t-1 )/(1 -α t ) √ 1 -α t /α t-1\n. By gradually denoising over T timesteps, the content of missing region is hallucinated in the latent space, producing a conditional sample z 0 ∼ p(z T |m ↓ , x m↓ ). z 0 is then transformed into the pixel space as x = D(z 0 ) via the left-inverse decoder network D corresponding to the autoencoder E(•), finally resulting in the completion outcome x ∈ R H×W×3 ."
},
{
"figure_ref": [],
"heading": "MCU-Net: Modality-specific Conditional U-Net skip connection switch",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Auxiliary Guidance",
"publication_ref": [
"b31"
],
"table_ref": [],
"text": "element-wise addition\nQ K V Depth Map Text Sketch Map Pose Map Semantic Map Canny Edge Map Q K V cross attention Q K V U-Net Denoiser Figure 4: Illustration of MCU-Net.\nThe first stage in MaGIC is to learn image completion under single-modality guidance. For this purpose, we propose a simple yet effective modality-specific conditional U-Net (MCU-Net). Particularly, for the auxiliary guidance c i ∈ C (C = {c i } N i=1 denotes the set of N auxiliary guidance), MCU-Net consists of a standard U-Net denoiser θ c i (Rombach et al., 2022) and an encoding network τ c i . For simplicity, we will omit i in the following sections.\nThe encoding network τ c is employed to extract multiscale guidance signals, represented as F l c , where l ∈ {0, • • • , L} and L denotes the number of times the feature map scale is reduced within the U-Net denoiser. Afterwards, F l c is injected to the latent in MCU-Net to obtain modality-guided feature. In specific, we denote the latent in MCU-Net as w t,c (c ∈ C) to distinguish it from the original diffusion model's z t . As illustrated in Fig. 4, to inject guidance signals into the latent w t,c , we add F l c to intermediate feature maps F l enc of the encoder of MCU-Net, resulting in guided feature map Fl\nc = F l enc + F l c , l ∈ [0, L].\nAnd we incorporate the text modality in a manner consistent with SD, which integrates its information into intermediate features via a cross-attention mechanism.\nTo utilize generative capability of pre-trained SD, we freeze the original U-Net denoiser when training MCU-Net, allowing the unlocked encoding network τ c to learn guidance signal extraction and fit the pre-trained denoiser."
},
{
"figure_ref": [],
"heading": "CMB: Consistent Modality Blending",
"publication_ref": [],
"table_ref": [],
"text": "Despite achieving image completion under single-modality with MCU-Net, it is not trivial to integrate multiple MCU-Nets for multi-modality image completion. A naive way is to jointly re-train these learned MCU-Nets, which is cumbersome and inflexible for multi-modality image completion. "
},
{
"figure_ref": [],
"heading": "Algorithm 1 Usage of CMB in MaGIC",
"publication_ref": [
"b34"
],
"table_ref": [],
"text": "Require: Given the input masked image x m , mask m, a series of MCU-Net parameters θ c , the number of times of converse amplification P, and the number of iteration steps of back-propagation Q.\n1: m ↓ = downsample(m) 2: x m↓ = E(x m ) 3: z T ∼ N(0, I) 4: w T,c ∼ N(0, I), ∀c ∈ C 5: for t = T, • • • , 1 do 6: if t ≤ T -P then 7: ϵ θ * , F * ← θ * (z t , t, m ↓ , x m↓ ) 8: z t-1 = sampler(z t , ϵ θ * ) (Eq. 3) 9: continue 10: end if 11: for 1, • • • , Q do 12: ϵ θ , FC ← θ C (w t,C , t, m ↓ , x m↓ ) 13: w t-1,C = sampler(w t,C , ϵ θ ) (Eq. 3) 14: ϵ θ * , F * ← θ * (z t , t, m ↓ , x m↓ ) 15: z ′ t-1 = sampler(z t , ϵ θ * ) (Eq. 3) 16: z t-1 = z ′ t-1 -σ t γ∇ zt ℓ( FC , F * ) (Eq. 5) 17:\nend for 18: end for 19: return D(z 0 )\nConverse Amplification. We use F * to denote the intermediate features from the original U-Net θ * which is not equipped with an guidance encoding network, while Fc to denote guided features from MCU-Net θ c of modality c. Notably, U-Net θ * and MCU-Net θ c undergo a parallel denoising process. At each step t, every latent is denoised using the DDIM sampler (Song et al., 2021). In the original U-Net θ * , we denote the denoised latent as the intermediate latent\nz ′ t-1 .\nWe bias F * towards Fc by calculating their Euclidean distance in each scale l:\nℓ( FC , F * ) = 1 L c∈C δ c ∥ Fl c -F l * ∥ 2 2 (4)\nwhere δ c are scale factors to weight the strength leads to either improved alignment to guidance modality c or greater diversity in the outputs. N = |C| indicates the modality number of auxiliary guidance set. Then we use distance to adjust latent code of the original SD model. Specifically, at each denoising step, we obtain FC and F * firstly, then the gradient of their distance is calculated through back-propagation to update the denoised latent z ′ t-1 :\nz t-1 = z ′ t-1 -σ t γ∇ z t ℓ( FC , F * ) (5)\nOwing to CMB, it is not necessary to jointly re-train the learned MCU-Nets, making MaGIC flexible in merging arbitrary multi-modality for completion. Alg. 1 shows the procedure of CMB for MaGIC. "
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [],
"table_ref": [],
"text": "In this work, we study three research questions, RQ1, RQ2 and RQ3:\nRQ1: Can our MCU-Net effectively perform image completion guided by various modalities? RQ2: Can our MaGIC with CMB seamlessly integrate guidance from multiple modalities to produce credible completion results? RQ3: How do different module designs (e.g., adjustments in hyperparameters and inference processes) impact the overall effectiveness?"
},
{
"figure_ref": [],
"heading": "Experimental settings",
"publication_ref": [
"b27",
"b12",
"b10",
"b37",
"b31",
"b22",
"b45",
"b26",
"b0",
"b46",
"b20"
],
"table_ref": [],
"text": "In our experiments, we select several edge-based image completion methods, including EC (Nazeri et al., 2019), CTSDG (Guo et al., 2021), ZITS (Dong et al., 2022), and state-of-the-art (SOTA) techniques such as LAMA (Suvorov et al., 2022), LDM (Rombach et al., 2022), and MAT (Li et al., 2022). We also include controllable image generation baselines such as ControlNet (Zhang & Agrawala, 2023) and T2I-Adapter (Mou et al., 2023) in our qualitative comparison, as they can be easily adapted to the image completion task with the concept of Blended Diffusion (Avrahami et al., 2022;2023). For fair comparison, we apply the same set of image mask pairs across all tests, and, for comparisons involving auxiliary guidance, we ensure that each method receives identical guidance map instructions. The masks used in testing are designed to uniformly span a masking ratio range from 0 to 100%. The evaluation adopts both image metrics (i.e., FID and P/U-IDS (Zhao et al., 2021)) and text-to-image metric (i.e., PickScore (Kirstain et al., 2023)) which gauges the fidelity of generated content based on learned human preferences. Acknowledging the pluralistic outcomes of our method, we conduct tests on a total of five images to determine mean scores and standard deviations. For all diffusion-based methods, the denoising step T is set to 50. For further details on the experimental configuration, please see the supplementary material."
},
{
"figure_ref": [
"fig_3",
"fig_3",
"fig_3"
],
"heading": "Image Completion with Single-Modality Guidance using MCU-Net",
"publication_ref": [
"b37",
"b22",
"b1",
"b45",
"b26",
"b37",
"b22"
],
"table_ref": [],
"text": "To answer RQ1, we compare our approach with state-of-the-art (SOTA) inpainting methods (Suvorov et al., 2022;Li et al., 2022) and SOTA single modality guidance image generation methods. We employ latent-level blending (Avrahami et al., 2023) to preserve pixels in unmasked regions for image generation methods such as ControlNet (Zhang & Agrawala, 2023) and T2I-Adapter (Mou et al., 2023). As depicted in the Fig. 5, our method generates content without noticeable artifacts, maintaining stronger spatial context consistency. Conversely, T2I-Adapter generates a stone house on the road (1st row in Fig. 5) and ControlNet puts a dancer on the soccer field (2nd row in Fig. 5).\nQuantitatively, the scores of edge-based methods on COCO and Places2 are displayed in Tab. 1. Across all metrics, our method demonstrates significant improvements, indicating that our MCU-Net can effectively generate content under the guidance of various single modalities. by addition (i.e., feature-level addition or FLA for short) to produce FC as FC ← F enc + c∈C F c . To show the effectiveness of CMB, we compare it with FLA on COCO as in Tab. 2a. Note that, we experiment FLA with 30 and 50 steps, respectively. To guarantee an equitable assessment across all auxiliary modalities, we opt for a wide-ranging set of modalities. Given that specific modalities (e.g., pose) may not be applicable to all test images (e.g., certain landscape images), we ensure that our test suite incorporates a diverse range of modalities. This includes segmentation map, depth map, Canny edge map, sketch map, and a prompt text. As displayed in Tab. 2a, the proposed CMB significantly surpasses FLA with naive addition, evidencing the effectiveness of CMB in merging multi-modality for completion. Interestingly, the performance of FLA with 50 steps is counter-intuitively lower than that with 35 steps, suggesting that this simple method may overly manipulate the latent code. This indicates that direct addition of different MCU-Net feature maps for multi-modality guidance is impractical. By contrast, our CMB efficaciously integrates the signals from multi-modal guidance.\nAnswering RQ2.2. To validate the effectiveness of our MaGIC, we compare it with state-of-the-art image completion methods, including LAMA (Suvorov et al., 2022) and MAT (Li et al., 2022) Table 3: Ablation studies on the multi-modality complementary and the hyper-parameters of CMB."
},
{
"figure_ref": [],
"heading": "Ablation Study",
"publication_ref": [
"b26"
],
"table_ref": [],
"text": "To answer RQ3, we conduct rich ablations on COCO as follows.\nImpact of modalities. To delve into auxiliary modalities, we investigate their individual contributions. We distinguish among five modalities used in our experiments: edge and sketch for fine-grained structural control, segmentation and depth for coarse-grained spatial-semantic control, and text for content-specific cue. As in Tab. 3a, the guidance from text significantly enhances image quality (FID) and generated content (PickScore). Interestingly, excluding text, the performance of combined modalities appears balanced, suggesting optimal generation quality when modalities provide complementary information. When using all modalities, the performance is the best.\nJoint multi-modality re-training. Our method allows multi-modality guidance without the need for additional joint training. However, exploring the joint re-training of all modality-specific conditional U-Nets with classifier-free guidance style can help identify the upper bound performance.\nBuilding such a model necessitates a fuser mechanism to blend diverse input modalities. To ensure effectiveness, we integrated CoAdapterFuser (Mou et al., 2023), aligning with our design goals.\nAddressing the lack of a paired dataset with extensive labels across various modalities was also essential. We extracted 650,000 images from the Laion dataset and generated four modalities (canny edge map, depth map, sketch map, and semantic map) using open-source tools. During joint retraining, we randomly dropped out each modality at a 0.5 probability. This training process is memory-intensive, necessitating a reduction in batch size to a quarter of single-modality training. The model underwent 180,000 iterations. As evidenced by its lower FID, shown in the first row of Tab. 3a, the unified model achieves higher fidelity than our training-free method. However, it encounters issues such as the need for paired training data, difficulty in adding new modality, and substantial computational requirements for joint training.\nGuidance in iteration. The proposed CMB algorithm involves two important hyperparameters, i.e., the number P of denoising steps incorporating CMB and the iteration times Q of gradient descent performed in each CMB operation. We study the impact of different P and Q on the multimodal conditioning completion task as in Tab. 3b. From Tab. 3b, we can observe that with Q fixed, the performance is almost consistently improved as increasing the number P of denoising steps (from 10 to 50) equipped for CMB. Interestingly, given the fact in Tab. 2a that incorporating guidance through simple FLA could impair the performance of completion, the results further demonstrate the effectiveness of CMB. In addition, we can also observe from Tab. 3b that, different Q (e.g., 1 to 5 to 10) leads to different performance. We argue that, increasing the iteration times to 5 in a reasonable manner based on 1 should yield better metrics as more guidance information is introduced. Yet, the subsequent decline when further increasing Q to 10 in performance can be attributed to the presence of more noise in hidden space during early stages of the denoising process. For the trade-off between inference time and image completion performance, we set the values of P and Q to 30 and 5, respectively."
},
{
"figure_ref": [],
"heading": "Conclusion and Limitation",
"publication_ref": [],
"table_ref": [],
"text": "In this paper we propose a novel, simple yet effective method, named MaGIC, for multi-modality image completion. Specifically, we first introduce the MCU-Net that is used to achieve singlemodality image completion by injecting the modality signal. Then, we devise a novel CMB algorithm that integrates multi-modality for more plausible image completion. On extensive experiments, we show that MaGIC shows superior performance. Moreover, it is generally applicable to various image completion tasks such as in/out-painting and local editing, and even the image generation task.\nMaGIC is proposed to facilitate image completion with multi-modality. Yet, there exist two limitations. First, the ability to generate high-frequency details is tied to the backbone completion model, which means even with ample detailed guidance, achieving desired fidelity may not be guaranteed. This can be improved by adopting more powerful backbones if necessary. In addition, our MaGIC is less efficient than current single-step completion models, with inference time increasing in line with guidance modalities. This is a common issue for diffusion models, and we leave it for future research.\nA Implementation Details"
},
{
"figure_ref": [],
"heading": "A.1 Different Image-based Conditions And Hyper-parameters",
"publication_ref": [
"b24",
"b4",
"b36",
"b3",
"b33",
"b30",
"b33",
"b13",
"b19"
],
"table_ref": [],
"text": "Our experiments include 6 types of image-based conditions:\n• Canny edge & Sketch. We utilize the training set of COCO (Lin et al., 2014), which contains 123K images, as the training data to train MCU-Net separately under canny and sketch guidance. The corresponding canny edge and sketch are generated by Canny algorithm (Canny, 1986) with default thresholds, and PiDiNet (Su et al., 2021) with a threshold of 0.5, respectively.\n• Segmentation. We utilize training set of COCO-Stuff (Caesar et al., 2018) as training data, which includes 123K images and corresponding semantic segmentation annotations. It covers 80 thing classes, 91 stuff classes and 1 \"unlabeled\" class, providing a comprehensive range of semantic information for MCU-Net training.\n• Depth. In order to obtain sufficient volume of data to train MCU-Net under this conditions with abstract representation, we select 650K images from LAION-AESTHETICS dataset (Schuhmann et al., 2022). And we adopt MiDaS (Ranftl et al., 2022) on them to generate depth maps.\n• Pose. We also pick images from LAION-AESTHETICS (Schuhmann et al., 2022) to construct training data for MCU-Net under pose guidance. The key distinction from building training dataset for depth guidance is that the selected images must contain at least one person for pose generation. To achieve this, we employ MM-Pose (Contributors, 2020), an open-source toolbox for pose estimation, to filter out images that do not meet the requirement, and generate pose for the retained images. In the end, we gather a total of 600k image-pose pairs to train MCU-Net under this condition.\n• Text. Within our default backbone, the SD-2.1 Inpainting, the prompt text is conditioned as the key and value of the cross-attention mechanism in the U-Net denoiser. It's noteworthy that this backbone is pretrained with the prompt text in the classifier-free way (Ho & Salimans, 2021). Consequently, in this work, we opt to use the backbone directly, thus bypassing the necessity to fine-tune an MCU-Net for text guidance.\nAll our experiments are conducted using 8 NVIDIA A100-40G GPUs. We set the batch size to 64 and employed the Adam optimizer (Kingma & Ba, 2015) with the learning rate of 1e-5 for training 10 epochs. These settings remain consistent across all conditions."
},
{
"figure_ref": [
"fig_0",
"fig_6",
"fig_0",
"fig_3"
],
"heading": "A.2 Acquisition of Conditions",
"publication_ref": [
"b41",
"b30"
],
"table_ref": [],
"text": "To facilitate a reliable and convenient comparison of model performance, we employed the conditions provided by the dataset directly or leveraged existing tools (Yang et al., 2022;Contributors, 2020;Ranftl et al., 2022) to estimate them. We then evaluated the model performances using quantitative metrics on completing the corresponding masked RGB images. It is important to note that our method also supports the input of manually designed guidance conditions (as shown in Fig. 1, Fig. 13, Fig. 14 and Fig. 15). However, when manually design dense guidance conditions like segmentation and depth maps, it's crucial to ensure their consistency with the information retained in the unmasked regions, particularly in the case of depth maps where values represent the distance between pixels and the camera. Fortunately, sparse conditions like sketch or pose maps can offer sufficient guidance information. We intend to release our code for condition generation, enabling users to obtain modalities including sketch, pose and segmentation maps effortlessly for image editing purposes."
},
{
"figure_ref": [
"fig_5"
],
"heading": "A.3 Architecture of Encoding Network τ c i",
"publication_ref": [
"b26",
"b22",
"b24",
"b49",
"b4",
"b36",
"b30",
"b41"
],
"table_ref": [],
"text": "The condition encoding network is designed to be simple and lightweight, and serves the purpose of extracting the multi-scale guidance signals from the input condition image. These guidance signals are aligned in size with the intermediate feature maps of the MCU-Net's encoder. As this is not the main focus of our work, we have referred to the design of T2I-Adapter (Mou et al., 2023). Specifically, it consists of four feature extraction blocks with a downsample module placed between each pair of adjacent blocks, and each feature extraction block is composed of one convolution layer and two residual blocks. In order to evaluate all baselines and our proposed method in a fair manner, the same image-mask pairs are used in quantitative experiments. Additionally, testing mask samples are obtained based on a uniform distribution ranging from 0% to 100% to encompass the majority of mask ratios encountered in real-world scenarios. The testing mask is randomly generated based on the algorithm from (Li et al., 2022), with the histogram of the testing mask ratio of COCO dataset visualized in Fig. 6.\nAll quantitative experiments are conducted on the COCO (Lin et al., 2014) and Places (Zhou et al., 2018) datasets. Evaluation of the methods involves using the first 1000 images in the COCO validation set and the first 5000 images in the Places validation set. Masks from COCO are replicated five times for the Places dataset. For auxiliary guided completion, the Canny algorithm (Canny, 1986) and PiDiNet (Su et al., 2021) are employed to obtain the canny edge map and sketch map, respectively. MiDaS (Ranftl et al., 2022) is adopted to acquire depth maps for both datasets. COCO serves as a dataset with semantic segmentation and prompt text annotations. As the Places dataset lacks ground-truth labels, the semantic segmentation map is estimated using CIRKD (Yang et al., 2022)."
},
{
"figure_ref": [],
"heading": "B Application Results",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [
"fig_6"
],
"heading": "B.1 Real User-input Image Editing",
"publication_ref": [],
"table_ref": [],
"text": "We highlight the adaptability of our method in handling user-input image editing tasks designed to manipulate real-world images based on user intention, as demonstrated in Fig. 13. This figure emphasizes our method's capacity to modify the structure or semantics of local regions using userinput guidance such as scribble, pose map and prompt text, while fully maintaining the integrity of the unmasked region."
},
{
"figure_ref": [
"fig_0",
"fig_3"
],
"heading": "B.2 Image Outpainting",
"publication_ref": [],
"table_ref": [],
"text": "Our method can also be used to extend an image, like generating a panorama from a small part of the image content. As demonstrated in Fig. 14 and Fig. 15, our method showcases its capability to outpaint a photograph or a painting guided by text and sketch map. Remarkably, our method exhibit the ability to generate suitable content that is harmonious even with the broader context of a panoramic image."
},
{
"figure_ref": [],
"heading": "C More Experimental Results And Studies",
"publication_ref": [
"b26"
],
"table_ref": [
"tab_7"
],
"text": "C.1 Quantitative Comparisons with Conditional Text-to-Image Methods Contemporary methods like ControlNet and T2I-Adapter have demonstrated remarkable achievements in controllable image generation. For a direct comparison, we employ latent-level blending to utilize these methods for image completion, maintaining the experimental settings of earlier Experiments. As Table 4 reveals, our MaGIC significantly surpasses baseline models in FID, U-IDS, P-IDS, and PickScore for most guidance types. In multi-modality guidance, we enhance T2I-Adapter with multi-adapter controlling Mou et al. (2023) (feature-level addition), resulting in T2I-Adapter⋄. For the COCO dataset, we employ five modalities: canny edge, depth, segmentation, sketch map, and text. For the Places dataset, we utilize canny edge, depth, segmentation, and sketch map, as it lacks manually-crafted captions. Our MaGIC outperforms T2I-Adapter⋄ by 44.68% in PickScore on COCO, and shows improvements of 37.07%, 71.40%, and 230.30% in FID, U-IDS, and P-IDS, respectively, on Places, as detailed in the last two rows of et al. (2022), they are not optimal for quantitatively assessing image completion. Nevertheless, we include these traditional metrics for reference. Additionally, we provide the CLIP Score as an extra measure for a more comprehensive evaluation in the Table 5."
},
{
"figure_ref": [
"fig_8"
],
"heading": "C.2 Qualitative Comparisons in Multimodal Conditioning",
"publication_ref": [],
"table_ref": [],
"text": "As illustrated in Figure 9, we perform qualitative, side-by-side comparisons with T2I-Adapter⋄.\nFor our MaGIC⋄, we produce five diverse results. Under the guidance of four modalities, MaGIC demonstrates strong controllability and high-fidelity outputs, aligning with our quantitative findings.\nIn contrast, while T2I-Adapter⋄ effectively adheres the layout or shape to guidance, it fails to generate images of above-average quality with realistic details. This shortfall is attributed to the feature-level addition approach, leading to an out-of-distribution effect in the SD U-Net. Although we claim that FLA (feature-level addition) is a simple yet imperfect method for combining multiple modalities, and demonstrate the effectiveness of our CMB by comparative experiments, a comprehensive understanding of these two methods remains elusive. To this end, we opt to visualize the feature distributions stemming from T2I-Adapter with single-modality training and multi-modality utilization strategies, including FLA or our proposed CMB.\nIn specific, we choose the feature from the middle denoising step (i.e., the 25th step of DDIM sampler) and output from U-Net encoder. The t-SNE visualization result is shown in Figure 7, and different colors represents features from different sources, while the associated numbers indicate the index and cluster center of each feature type. Numbers ranging from small to large represent features obtained from T2I-Adapter-Canny (0), T2I-Adapter-Depth (1), T2I-Adapter-Segmentation (2), T2I-Adapter-Sketch (3), T2I-Adapter-CMB (4) and T2I-Adapter-FLA (5), respectively, where the first four indicate the trained single-modality while the last two are two methods of combining these four modalities.\nWe can draw two conclusions from Figure 7:\n1. Features derived from different single-modality models (0, 1, 2 and 3) show significant distribution disparities, and FLA (5) directly adds modality features resulting in the distribution deviation of obtained feature from all others. This observation aligns with our assertion in the main manuscript that \"we called feature-level addition is impractical, as the denoiser is trained solely on the distribution of Fc = F enc + F c \".\n2. In contrast to FLA, the distribution of features obtained through CMB ( 4) is surrounded by other single-modality distributions. This phenomenon is coherent with Equation 4 and 5, where the distribution of obtained features is \"pulled\" by the distributions of the other four single-modalities.\nA lovely cat sits on the bench, and a beautiful girl is singing.\nStormtrooper is standing in a kitchen with ladder."
},
{
"figure_ref": [],
"heading": "Depth + Pose Guidance",
"publication_ref": [],
"table_ref": [],
"text": "Depth + Pose Guidance "
},
{
"figure_ref": [
"fig_7"
],
"heading": "C.4 Failure Cases",
"publication_ref": [
"b6"
],
"table_ref": [],
"text": "Figure 8 shows two failed instances of applying CMB on T2I-Adapter for multi-modalities guidance.\nHere we present a more complex testing scenario involving non-overlapping information between two modalities. In the first case, we use Anything-4.0 as the backbone and there are two mistakes in the generated image: the misshapen cat and the incorrectly positioned bench under the girl. The former discrepancy possibly arises from the pose adapter contributing stronger features compared to those from the depth adapter, consequently affecting the representation of the latter information, which is not accurately reflected in the generated image. The issue might be alleviated by training depth adapter more stronger or increasing δ depth while decreasing δ pose (see Equation 4for details).\nWhile the latter mistake is the inherent challenge in SD, and many related works Chen et al. (2023); Chefer et al. ( 2023) could be referenced for potential mitigation steategies. In the second example, the generated image depicts a wall as the background, leading to the complete loss of depth information from the depth map.\nThese instances underscore that the inherent problems of SD persist despite employing CMB. Furthermore, in scenarios where information correlation between different modalities is low, evident issues such as information loss and errors in generated images become more pronounced. Through these failed cases, it can be seen that the inherent challenges of SD cannot be eliminated by CMB. Furthermore, in scenarios where information correlation between different modalities is low, evident issues such as information loss and errors in generated images become more pronounced."
},
{
"figure_ref": [
"fig_0",
"fig_0"
],
"heading": "C.5 Study in Adaptability and Image Generation Application",
"publication_ref": [
"b26",
"b1",
"b26",
"b1",
"b45"
],
"table_ref": [],
"text": "Our proposed MaGIC has the ability to adapt to a variety of backbone diffusion models, including but not limited to, the image generation model Anything-4.0, Stable Diffusion-1.5 (also employed by T2I-adapter and ControlNet), and the image completion model Stable Diffusion Inpainting-2.1 (the default in MaGIC). In order to elucidate the differences among these backbone diffusion models, a qualitative experiment was carried out, focusing primarily on the anime-style image generation model Anything-4.0, image generation model Stable Diffusion (SD), the mask-aware T2I-adapter (Mou et al., 2023;Avrahami et al., 2023), and our own MaGIC.\nAs portrayed in Fig. 10(a) and (b), our MaGIC method exhibits exceptional generalizability to image generation backbones. These backbones can produce convincing results guided by factors such as sketch, depth, segmentation, and the canny edge map. T2I-Adapter (Mou et al., 2023) is a conditional image generation framework based on Stable Diffusion-1.5. To equip T2I-Adapter with CMB for the completion of a masked image, we implemented a technique known as latent-level blending (Avrahami et al., 2023). As evidenced in Fig. 10(c) and (d), incorporating blending into T2I-Adapter can preserve the unmasked region remains while the generated masked region does not perceive the unmasked region, given the fact that there are two sheep heads in a single sheep.\nWe further adapt CMB to T2I-Adapter and ControlNet (Zhang & Agrawala, 2023) "
}
] | Vanilla image completion approaches exhibit sensitivity to large missing regions, attributed to the limited availability of reference information for plausible generation. To mitigate this, existing methods incorporate the extra cue as a guidance for image completion. Despite improvements, these approaches are often restricted to employing a single modality (e.g., segmentation or sketch maps), which lacks scalability in leveraging multi-modality for more plausible completion. In this paper, we propose a novel, simple yet effective method for Multi-modal Guided Image Completion, dubbed MaGIC, which not only supports a wide range of single modality as the guidance (e.g., text, canny edge, sketch, segmentation, depth, and pose), but also adapts to arbitrarily customized combination of these modalities (i.e., arbitrary multi-modality) for image completion. For building MaGIC, we first introduce a modality-specific conditional U-Net (MCU-Net) that injects single-modal signal into a U-Net denoiser for single-modal guided image completion. Then, we devise a consistent modality blending (CMB) method to leverage modality signals encoded in multiple learned MCU-Nets through gradient guidance in latent space. Our CMB is training-free, thereby avoids the cumbersome joint re-training of different modalities, which is the secret of MaGIC to achieve exceptional flexibility in accommodating new modalities for completion. Experiments show the superiority of MaGIC over state-of-the-art methods and its generalization to various completion tasks. Our project with code and models is available at yeates.github. | MaGIC: Multi-modality Guided Image Completion | [
{
"figure_caption": "Figure 1 :1Figure 1: Illustration of our MaGIC for image completion tasks including outpainting (first row) and real user-input editing (second row) under multi-modality guidance.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: Illustration of our MaGIC for real user-input editing task using various combination of multi-modality as guidance.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "sFigure 3 :3Figure3: Illustration of our method. We initiate the inference process with a randomly initialized latent z T . This latent is denoised T times, with the concatenation of the masked image and mask acting as conditioning for both MCU-Net and frozen U-Net denoiser. Through CMB, we fuse diverse modality guidance signals, aiding the frozen original U-Net θ * to iteratively produce the desired content. The content is finally transformed into pixel space via a decoder network, resulting in the completed RGB output.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 :5Figure 5: Qualitative comparison for image completion using single modality as guidance. * indicates the use of latent-level blending(Avrahami et al., 2023) to preserve pixels in unmasked regions.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "( b )bComparison of MaGIC with SOTA methods.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "b",
"figure_type": "figure"
},
{
"figure_caption": "AFigure 6 :6Figure 6: Illustration of mask ratio.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "C. 33Figure 7: t-SNE visualization of features output from U-Net encoder.",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure 8 :8Figure 8: Failed cases when adopting CMB on T2I-Adapter.",
"figure_data": "",
"figure_id": "fig_7",
"figure_label": "8",
"figure_type": "figure"
},
{
"figure_caption": "Figure 9 :\"9Figure 9: Qualitative results of MaGIC with four guidance compared to baselines",
"figure_data": "",
"figure_id": "fig_8",
"figure_label": "9",
"figure_type": "figure"
},
{
"figure_caption": "\"Figure 14: Application examples: sketch and text guided image outpainting.",
"figure_data": "",
"figure_id": "fig_9",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "To deal with this, we propose the novel consistent modality blending (CMB), a training-free algorithm to integrate guidance signals from different auxiliary modalities without requiring additional joint re-training. A great benefit of CMB is that, the multi-modality guidance latent code in MCU-Net remains aligned with the internal knowledge of SD model, without affecting its original ability. As shown in Fig.3, the guidance signals from arbitrary combination of independent single-modality models (i.e., MCU-Nets) in gradient aspect gradually control the image completion process with input modalities. Specifically, given a series of MCU-Nets trained independently on multiple modalities C, we can extract the guidance signals F c . A simple way for integrating different modalities is to directly update intermediate feature maps F enc by adding accumulated guidance signals as FC ← F enc + c∈C F c .",
"figure_data": "",
"figure_id": "tab_1",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Comparison of using single auxiliary modality as guidance for image completion. ♠: ground truth edge map as guidance, : estimated depth map as guidance, ♣: segmentation map as guidance, ↑: the higher the better, ↓: the lower the better, †: completion without any guidance.",
"figure_data": "COCOPlaces2MethodFID↓PickScore↑ / %FID↓U-IDS↑ / % P-IDS↑ / %EC (Nazeri et al., 2019) ♠ CTSDG (Guo et al., 2021) ♠ ZITS (Dong et al., 2022) ♠76.64 97.05 61.2723.14 24.03 28.0925.08 42.81 18.9612.89 0 18.752.86 0 7.20Our MCU-Net † Our MCU-Net Our MCU-Net ♣ Our MCU-Net ♠47.70±0.29 39.43±0.26 41.91±0.20 41.15±0.2730.79±0.10 37.12±0.11 34.96±0.17 34.94±0.0610.74±0.07 23.83±0.30 10.18±0.48 9.09±0.04 25.34±0.29 10.64±0.46 10.27±0.06 24.21±0.24 9.93±0.38 8.32±0.02 26.23±0.07 10.96±0.33COCOCOCOMethod MaGIC w/ FLA (35 steps) MaGIC w/ FLA (50 steps) MaGIC † MaGIC w/ CMBMMGFID ↓ 37.78±0.32 41.53±0.19 47.70±0.29 37.65±0.22PickScore ↑ / % 44.19±0.23 35.85±0.08 30.79±0.10 49.57±0.17Method CoMod TFill FcF LAMA MAT MaGIC † MaGICMMGFID ↓ 68.01 58.55 48.92 48.63 45.51 47.70±0.29 37.65±0.22PickScore ↑ / % 25.12 24.63 26.43 29.06 27.10 30.79±0.10 49.57±0.17(a) Comparison of CMB with simple FLA.",
"figure_id": "tab_2",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Comparisons of CMB and FLA and MaGIC with others. MMG: multi-modality guidance.",
"figure_data": "",
"figure_id": "tab_3",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": "COCOPlaces2MethodFID↓ PickScore↑ / % FID↓ U-IDS↑ / % P-IDS↑ / %ZITS ♠ T2I-Adapter ♠ 48.23 61.27 ControlNet ♠ 37.17 Ours ♠ 41.1528.09 30.10 37.30 34.9418.96 10.39 10.35 8.3218.75 19.44 18.45 26.237.20 5.66 4.58 10.96T2I-Adapter ControlNet Ours50.92 46.13 39.4330.22 32.52 37.1218.10 15.96 9.0914.91 14.46 25.344.56 3.18 10.64T2I-Adapter ♣ 50.65 ControlNet ♣ 58.27 Ours ♣ 41.9128.10 26.11 34.9615.36 18.13 10.2715.99 13.68 24.214.30 3.24 9.93T2I-Adapter ⋄ 39.08 Ours ⋄ 37.6534.26 49.5714.27 8.9814.76 25.303.30 10.90",
"figure_id": "tab_5",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Quantitative comparisons with conditional image completion and text-to-image methods. ♠: ground truth edge map as guidance, : estimated depth map as guidance, ♣: segmentation map as guidance, ⋄: using segmentation, depth, canny, sketch, and text (on COCO) for guidance simultaneously.",
"figure_data": "COCOPlaces2MethodCLIP↑ / % PSNR↑ SSIM↑ LPIPS↓ PSNR↑ SSIM↑ LPIPS↓ZITS ♠ T2I-Adapter ♠ ControlNet ♠ Ours ♠28.33 28.59 28.97 29.3714.31 18.26 19.22 18.130.2767 0.6272 0.6871 0.61880.5382 0.3409 0.3183 0.346721.07 18.34 18.59 19.000.6888 0.6537 0.6647 0.65690.2614 0.3208 0.3220 0.3111T2I-Adapter ControlNet Ours28.05 28.22 29.1117.84 18.16 17.470.5894 0.6275 0.59600.3729 0.3583 0.362817.57 17.49 17.910.5765 0.5967 0.61090.3805 0.3703 0.3432T2I-Adapter ♣ ControlNet ♣ Ours ♣28.10 26.48 28.8717.55 16.98 17.010.5635 0.5587 0.56810.3830 0.4023 0.379917.33 17.22 17.440.5529 0.5568 0.58600.3923 0.3948 0.3591T2I-Adapter ⋄ Ours ⋄30.23 31.2919.45 17.490.6748 0.59210.3217 0.371719.17 17.850.6626 0.60850.3255 0.3439",
"figure_id": "tab_6",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Additional quantitative comparison results in terms of CLIP score and traditional reconstruction metrics. ♠: ground truth edge map as guidance, : estimated depth map as guidance, ♣: segmentation map as guidance, ⋄: using segmentation, depth, canny, sketch, and text (on COCO) for guidance simultaneously.While traditional reconstruction metrics, such as PSNR, SSIM, and LPIPS, rely on pixel-wise similarity to the ground truth and tend to favor blurry outputs, as noted byZhao et al. (2021) and Li",
"figure_data": "",
"figure_id": "tab_7",
"figure_label": "5",
"figure_type": "table"
}
] | Yongsheng Yu; Hao Wang; Tiejian Luo; Heng Fan; Libo Zhang | [
{
"authors": "Omri Avrahami; Dani Lischinski; Ohad Fried",
"journal": "",
"ref_id": "b0",
"title": "Blended diffusion for text-driven editing of natural images",
"year": "2022"
},
{
"authors": "Omri Avrahami; Ohad Fried; Dani Lischinski",
"journal": "",
"ref_id": "b1",
"title": "Blended latent diffusion",
"year": "2023"
},
{
"authors": "Arpit Bansal; Hong-Min Chu; Avi Schwarzschild; Soumyadip Sengupta; Micah Goldblum; Jonas Geiping; Tom Goldstein",
"journal": "",
"ref_id": "b2",
"title": "Universal guidance for diffusion models",
"year": "2023"
},
{
"authors": "Holger Caesar; Jasper Uijlings; Vittorio Ferrari",
"journal": "",
"ref_id": "b3",
"title": "Coco-stuff: Thing and stuff classes in context",
"year": "2018"
},
{
"authors": "John F Canny",
"journal": "IEEE TPAMI",
"ref_id": "b4",
"title": "A computational approach to edge detection",
"year": "1986"
},
{
"authors": "Hila Chefer; Yuval Alaluf; Yael Vinker; Lior Wolf; Daniel Cohen-Or",
"journal": "ACM Trans. Graph",
"ref_id": "b5",
"title": "Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models",
"year": "2023"
},
{
"authors": "Minghao Chen; Iro Laina; Andrea Vedaldi",
"journal": "",
"ref_id": "b6",
"title": "Training-free layout control with cross-attention guidance",
"year": "2023"
},
{
"authors": "",
"journal": "MMPose Contributors",
"ref_id": "b7",
"title": "Openmmlab pose estimation toolbox and benchmark",
"year": "2020"
},
{
"authors": "Antonio Criminisi; Patrick Pérez; Kentaro Toyama",
"journal": "",
"ref_id": "b8",
"title": "Object removal by exemplar-based inpainting",
"year": "2003"
},
{
"authors": "Prafulla Dhariwal; Alexander Quinn; Nichol ",
"journal": "",
"ref_id": "b9",
"title": "Diffusion models beat gans on image synthesis",
"year": "2021"
},
{
"authors": "Qiaole Dong; Chenjie Cao; Yanwei Fu",
"journal": "",
"ref_id": "b10",
"title": "Incremental transformer structure enhanced image inpainting with masking positional encoding",
"year": "2022"
},
{
"authors": "Shuyang Gu; Dong Chen; Jianmin Bao; Fang Wen; Bo Zhang; Dongdong Chen; Lu Yuan; Baining Guo",
"journal": "",
"ref_id": "b11",
"title": "Vector quantized diffusion model for text-to-image synthesis",
"year": "2022"
},
{
"authors": "Xiefan Guo; Hongyu Yang; Di Huang",
"journal": "",
"ref_id": "b12",
"title": "Image inpainting via conditional texture and structure dual generation",
"year": "2021"
},
{
"authors": "Jonathan Ho; Tim Salimans",
"journal": "",
"ref_id": "b13",
"title": "Classifier-free diffusion guidance",
"year": "2021"
},
{
"authors": "Jonathan Ho; Ajay Jain; Pieter Abbeel",
"journal": "",
"ref_id": "b14",
"title": "Denoising diffusion probabilistic models",
"year": "2020"
},
{
"authors": "Daichi Horita; Jiaolong Yang; Dong Chen; Yuki Koyama; Kiyoharu Aizawa",
"journal": "",
"ref_id": "b15",
"title": "A structure-guided diffusion model for large-hole diverse image completion",
"year": "2022"
},
{
"authors": "Jaeseok Jeong; Mingi Kwon; Youngjung Uh",
"journal": "",
"ref_id": "b16",
"title": "Training-free style transfer emerges from h-space in diffusion models",
"year": "2023"
},
{
"authors": "Bahjat Kawar; Shiran Zada; Oran Lang; Omer Tov; Huiwen Chang; Tali Dekel; Inbar Mosseri; Michal Irani",
"journal": "",
"ref_id": "b17",
"title": "Imagic: Text-based real image editing with diffusion models",
"year": "2023"
},
{
"authors": "Gwanghyun Kim; Taesung Kwon; Jong Chul; Ye ",
"journal": "",
"ref_id": "b18",
"title": "Diffusionclip: Text-guided diffusion models for robust image manipulation",
"year": "2022"
},
{
"authors": "P Diederik; Jimmy Kingma; Ba",
"journal": "",
"ref_id": "b19",
"title": "Adam: A method for stochastic optimization",
"year": "2015"
},
{
"authors": "Yuval Kirstain; Adam Polyak; Uriel Singer; Shahbuland Matiana; Joe Penna; Omer Levy",
"journal": "",
"ref_id": "b20",
"title": "Picka-pic: An open dataset of user preferences for text-to-image generation",
"year": "2023"
},
{
"authors": "Anat Levin; Assaf Zomet; Shmuel Peleg; Yair Weiss",
"journal": "",
"ref_id": "b21",
"title": "Seamless image stitching in the gradient domain",
"year": "2004"
},
{
"authors": "Wenbo Li; Zhe Lin; Kun Zhou; Lu Qi; Yi Wang; Jiaya Jia",
"journal": "",
"ref_id": "b22",
"title": "MAT: mask-aware transformer for large hole image inpainting",
"year": "2022"
},
{
"authors": "Liang Liao; Jing Xiao; Zheng Wang; Chia-Wen Lin; Shin'ichi Satoh",
"journal": "",
"ref_id": "b23",
"title": "Guidance and evaluation: Semantic-aware image inpainting for mixed scenes",
"year": "2020"
},
{
"authors": "Tsung-Yi Lin; Michael Maire; Serge J Belongie; James Hays; Pietro Perona; Deva Ramanan; Piotr Dollár; C Lawrence Zitnick",
"journal": "",
"ref_id": "b24",
"title": "Microsoft COCO: common objects in context",
"year": "2014"
},
{
"authors": "Andreas Lugmayr; Martin Danelljan; Andrés Romero; Fisher Yu; Radu Timofte; Luc Van Gool",
"journal": "",
"ref_id": "b25",
"title": "Repaint: Inpainting using denoising diffusion probabilistic models",
"year": "2022"
},
{
"authors": "Chong Mou; Xintao Wang; Liangbin Xie; Jian Zhang; Zhongang Qi; Ying Shan; Xiaohu Qie",
"journal": "",
"ref_id": "b26",
"title": "T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models",
"year": "2023"
},
{
"authors": "Kamyar Nazeri; Eric Ng; Tony Joseph; Faisal Z Qureshi; Mehran Ebrahimi",
"journal": "",
"ref_id": "b27",
"title": "Edgeconnect: Structure guided image inpainting using edge prediction",
"year": "2019"
},
{
"authors": "Alexander Quinn; Nichol ; Prafulla Dhariwal; Aditya Ramesh; Pranav Shyam; Pamela Mishkin; Bob Mcgrew; Ilya Sutskever; Mark Chen",
"journal": "",
"ref_id": "b28",
"title": "GLIDE: towards photorealistic image generation and editing with text-guided diffusion models",
"year": "2022"
},
{
"authors": "Aditya Ramesh; Prafulla Dhariwal; Alex Nichol; Casey Chu; Mark Chen",
"journal": "",
"ref_id": "b29",
"title": "Hierarchical textconditional image generation with CLIP latents",
"year": "2022"
},
{
"authors": "René Ranftl; Katrin Lasinger; David Hafner; Konrad Schindler; Vladlen Koltun",
"journal": "IEEE TPAMI",
"ref_id": "b30",
"title": "Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer",
"year": "2022"
},
{
"authors": "Robin Rombach; Andreas Blattmann; Dominik Lorenz; Patrick Esser; Björn Ommer",
"journal": "",
"ref_id": "b31",
"title": "Highresolution image synthesis with latent diffusion models",
"year": "2022"
},
{
"authors": "Chitwan Saharia; William Chan; Saurabh Saxena; Lala Li; Jay Whang; Emily Denton; Seyed Kamyar; Seyed Ghasemipour; Burcu Karagol Ayan; S Sara Mahdavi; Rapha Gontijo Lopes; Tim Salimans; Jonathan Ho; David J Fleet; Mohammad Norouzi",
"journal": "",
"ref_id": "b32",
"title": "Photorealistic text-to-image diffusion models with deep language understanding",
"year": "2022"
},
{
"authors": "Christoph Schuhmann; Romain Beaumont; Richard Vencu; Cade Gordon; Ross Wightman; Mehdi Cherti; Theo Coombes; Aarush Katta; Clayton Mullis; Mitchell Wortsman",
"journal": "",
"ref_id": "b33",
"title": "Laion-5b: An open large-scale dataset for training next generation image-text models",
"year": "2022"
},
{
"authors": "Jiaming Song; Chenlin Meng; Stefano Ermon",
"journal": "",
"ref_id": "b34",
"title": "Denoising diffusion implicit models",
"year": "2021"
},
{
"authors": "Yang Song; Prafulla Dhariwal; Mark Chen; Ilya Sutskever",
"journal": "",
"ref_id": "b35",
"title": "Consistency models",
"year": "2023"
},
{
"authors": "Zhuo Su; Wenzhe Liu; Zitong Yu; Dewen Hu; Qing Liao; Qi Tian; Matti Pietikäinen; Li Liu",
"journal": "",
"ref_id": "b36",
"title": "Pixel difference networks for efficient edge detection",
"year": "2021"
},
{
"authors": "Roman Suvorov; Elizaveta Logacheva; Anton Mashikhin; Anastasia Remizova; Arsenii Ashukha; Aleksei Silvestrov; Naejin Kong; Harshith Goka; Kiwoong Park; Victor Lempitsky",
"journal": "",
"ref_id": "b37",
"title": "Resolutionrobust large mask inpainting with fourier convolutions",
"year": "2022"
},
{
"authors": "Ziyu Wan; Bo Zhang; Dongdong Chen; Pan Zhang; Dong Chen; Jing Liao; Fang Wen",
"journal": "",
"ref_id": "b38",
"title": "Bringing old photos back to life",
"year": "2020"
},
{
"authors": "Su Wang; Chitwan Saharia; Ceslee Montgomery; Jordi Pont-Tuset; Shai Noy; Stefano Pellegrini; Yasumasa Onoe; Sarah Laszlo; David J Fleet; Radu Soricut; Jason Baldridge; Mohammad Norouzi; Peter Anderson; William Chan",
"journal": "",
"ref_id": "b39",
"title": "Imagen editor and editbench: Advancing and evaluating text-guided image inpainting",
"year": "2023"
},
{
"authors": "Shaoan Xie; Zhifei Zhang; Zhe Lin; Tobias Hinz; Kun Zhang",
"journal": "",
"ref_id": "b40",
"title": "Smartbrush: Text and shape guided object inpainting with diffusion model",
"year": "2022"
},
{
"authors": "Chuanguang Yang; Helong Zhou; Zhulin An; Xue Jiang; Yongjun Xu; Qian Zhang",
"journal": "",
"ref_id": "b41",
"title": "Cross-image relational knowledge distillation for semantic segmentation",
"year": "2022"
},
{
"authors": "Jiahui Yu; Zhe Lin; Jimei Yang; Xiaohui Shen; Xin Lu; Thomas S Huang",
"journal": "",
"ref_id": "b42",
"title": "Free-form image inpainting with gated convolution",
"year": "2019"
},
{
"authors": "Jiwen Yu; Yinhuai Wang; Chen Zhao; Bernard Ghanem; Jian Zhang",
"journal": "",
"ref_id": "b43",
"title": "Freedom: Training-free energy-guided conditional diffusion model",
"year": "2023"
},
{
"authors": "Yongsheng Yu; Dawei Du; Libo Zhang; Tiejian Luo",
"journal": "",
"ref_id": "b44",
"title": "Unbiased multi-modality guidance for image inpainting",
"year": "2022"
},
{
"authors": "Lvmin Zhang; Maneesh Agrawala",
"journal": "",
"ref_id": "b45",
"title": "Adding conditional control to text-to-image diffusion models",
"year": "2023"
},
{
"authors": "Shengyu Zhao; Jonathan Cui; Yilun Sheng; Yue Dong; Xiao Liang; I-Chao Eric; Yan Chang; Xu",
"journal": "",
"ref_id": "b46",
"title": "Large scale image completion via co-modulated generative adversarial networks",
"year": "2021"
},
{
"authors": "Chuanxia Zheng; Guoxian Song; Tat-Jen Cham; Jianfei Cai; Dinh Q Phung; Linjie Luo",
"journal": "",
"ref_id": "b47",
"title": "Highquality pluralistic image completion via code shared VQGAN",
"year": "2022"
},
{
"authors": "Haitian Zheng; Zhe Lin; Jingwan Lu; Scott Cohen; Eli Shechtman; Connelly Barnes; Jianming Zhang; Qing Liu; Yuqian Zhou; Sohrab Amirghodsi; Jiebo Luo",
"journal": "",
"ref_id": "b48",
"title": "Structure-guided image completion with image-level and object-level semantic discriminators",
"year": "2022"
},
{
"authors": "Bolei Zhou; Àgata Lapedriza; Aditya Khosla; Aude Oliva; Antonio Torralba",
"journal": "IEEE TPAMI",
"ref_id": "b49",
"title": "Places: A 10 million image database for scene recognition",
"year": "2018"
}
] | [
{
"formula_coordinates": [
4,
354.44,
498.6,
150.23,
15.3
],
"formula_id": "formula_0",
"formula_text": "ϵ t θ (x t )∥ 2 2 ),(1)"
},
{
"formula_coordinates": [
4,
209.34,
646.32,
68.8,
12.62
],
"formula_id": "formula_1",
"formula_text": "T ∼ Z ∈ R H s × W s ×3"
},
{
"formula_coordinates": [
4,
190.13,
686.62,
314.54,
15.3
],
"formula_id": "formula_2",
"formula_text": "Φ(z t , t, m ↓ , x m↓ , θ) = min(E z 0 ,t,ϵ∼N(0,I) ∥ϵ -ϵ t θ (z t , m ↓ , x m↓ )∥ 2 2 ),(2)"
},
{
"formula_coordinates": [
5,
141.66,
372.24,
363.01,
32.18
],
"formula_id": "formula_3",
"formula_text": "z t-1 = √ α t-1 ( z t - √ 1 -α t ϵ t θ (z t , m ↓ , x m↓ ) √ α t ) + 1 -α t-1 -σ 2 t • ϵ t θ (z t , m ↓ , x m↓ ) + σ t ϵ t ,(3)"
},
{
"formula_coordinates": [
5,
108,
418,
162.57,
17.37
],
"formula_id": "formula_4",
"formula_text": "σ t = η √ (1 -α t-1 )/(1 -α t ) √ 1 -α t /α t-1"
},
{
"formula_coordinates": [
5,
341.16,
550.21,
143.36,
153.59
],
"formula_id": "formula_5",
"formula_text": "Q K V Depth Map Text Sketch Map Pose Map Semantic Map Canny Edge Map Q K V cross attention Q K V U-Net Denoiser Figure 4: Illustration of MCU-Net."
},
{
"formula_coordinates": [
5,
205.16,
720.66,
95.32,
13.1
],
"formula_id": "formula_6",
"formula_text": "c = F l enc + F l c , l ∈ [0, L]."
},
{
"formula_coordinates": [
6,
306,
451.43,
198.61,
171.86
],
"formula_id": "formula_7",
"formula_text": "1: m ↓ = downsample(m) 2: x m↓ = E(x m ) 3: z T ∼ N(0, I) 4: w T,c ∼ N(0, I), ∀c ∈ C 5: for t = T, • • • , 1 do 6: if t ≤ T -P then 7: ϵ θ * , F * ← θ * (z t , t, m ↓ , x m↓ ) 8: z t-1 = sampler(z t , ϵ θ * ) (Eq. 3) 9: continue 10: end if 11: for 1, • • • , Q do 12: ϵ θ , FC ← θ C (w t,C , t, m ↓ , x m↓ ) 13: w t-1,C = sampler(w t,C , ϵ θ ) (Eq. 3) 14: ϵ θ * , F * ← θ * (z t , t, m ↓ , x m↓ ) 15: z ′ t-1 = sampler(z t , ϵ θ * ) (Eq. 3) 16: z t-1 = z ′ t-1 -σ t γ∇ zt ℓ( FC , F * ) (Eq. 5) 17:"
},
{
"formula_coordinates": [
6,
108,
501.17,
16.85,
13.92
],
"formula_id": "formula_8",
"formula_text": "z ′ t-1 ."
},
{
"formula_coordinates": [
6,
139.45,
548.67,
157.25,
28.76
],
"formula_id": "formula_9",
"formula_text": "ℓ( FC , F * ) = 1 L c∈C δ c ∥ Fl c -F l * ∥ 2 2 (4)"
},
{
"formula_coordinates": [
6,
248.58,
688.53,
256.08,
15.45
],
"formula_id": "formula_10",
"formula_text": "z t-1 = z ′ t-1 -σ t γ∇ z t ℓ( FC , F * ) (5)"
}
] |
|
2023-11-27 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b29",
"b30",
"b26",
"b12",
"b1",
"b8",
"b23",
"b3",
"b18"
],
"table_ref": [],
"text": "Text-to-SQL models enable users to query databases using natural language questions (NLQs) without having to develop the underlying SQL query. Over the past few decades, neural models with supervised learning have achieved impressive performance on the text-to-SQL task, which are usually trained on a large training set and then evaluated on test examples (Wang et al., 2019;Yu et al., 2021;Rubin and Berant, 2021;Scholak et al., 2021;Gan et al., 2021;Li et al., 2023a).\nRecently, large language models (LLMs) have demonstrated strong capabilities for in-context learning on many language understanding and generation tasks (Brown et al., 2020;Chen et al., 2021a;Chowdhery et al., 2022), including on the text-to-SQL task (Rajkumar et al., 2022;Chang et al., 2023;Liu et al., 2023). Instead of training a text-to-SQL model on a large training set, schema and content, and demonstration examples. The difference in prompt constructions makes it difficult to directly compare two studies on their main contribution, and the outcomes of different studies may change based on future revelations in prompt engineering.\nIn this paper, we evaluate various strategies for prompt construction in three commonly employed text-to-SQL settings: zero-shot, single-domain, and cross-domain. We assess LLMs on text-to-SQL, considering various database prompt constructions in all three settings. Additionally, in the cross-domain scenario, we investigate the strategy for constructing demonstrations. Through our evaluation, we aim to gain insights into the effectiveness of these prompt construction strategies. Our findings can be summarized as follows:\n• Table relationship and table content play a crucial role in effectively prompting LLMs. However, it is essential to carefully consider their representation in the prompt, as LLMs are sensitive to the specific presentation in the zero-shot and cross-domain settings."
},
{
"figure_ref": [],
"heading": "• In-domain demonstration examples can mitigate",
"publication_ref": [
"b23",
"b3",
"b18",
"b21",
"b9",
"b32",
"b23",
"b19",
"b5"
],
"table_ref": [],
"text": "LLMs' sensitivity to different representations of database knowledge but they cannot replace table content knowledge. • The length of the prompt has a significant impact on the LLMs' performance in the cross-domain setting. We discovered a preferred prompt length that leads to improved performance.\n2 In-context Learning for Text-to-SQL\nIn the text-to-SQL task, a database and a natural language question (NLQ) are provided as input for generating an output SQL query. Traditional supervised learning approaches train models on specific text-to-SQL datasets. However, in-context learning allows pretrained large language models (LLMs) to perform text-to-SQL by providing either zero or a few training examples (NLQ-SQL pairs) as demonstrations. This section introduces three widely used settings for in-context learning in textto-SQL. Prompt examples in these settings can be found in Appendix A.1.\nZero-shot Text-to-SQL This setting evaluates the text-to-SQL capability of pretrained LLMs to directly infer the NLQ-SQL relationship from a table without any demonstration examples. The input includes a task instruction and a test question with its corresponding database. Zero-shot textto-SQL is used to directly assess the text-to-SQL capability of LLMs (Rajkumar et al., 2022;Chang et al., 2023;Liu et al., 2023).\nSingle-domain Few-shot Text-to-SQL This setting is designed for applications or domains where it is easy to construct examples, such as booking flights (Price, 1990;Dahl et al., 1994) and querying geographic information (Zelle and Mooney, 1996).\nIt tests the ability of LLMs to adapt with a few in-domain demonstration examples, which are collected from the same database as the test question.\nThe goal is to evaluate how well the LLMs can perform text-to-SQL with minimal in-domain training data (Rajkumar et al., 2022).\nCross-domain Few-shot Text-to-SQL This setting evaluates the generalization capability of models to new domains by learning from out-of-domain demonstrations. In this scenario, the demonstration NLQ-SQL pairs correspond to one or multiple demonstration databases that are different from the test database. Cross-domain few-shot text-to-SQL assesses how well LLMs can apply their learned knowledge from demonstrations to new databases (Poesia et al., 2022;Chen et al., 2023)."
},
{
"figure_ref": [],
"heading": "Prompt Construction",
"publication_ref": [],
"table_ref": [],
"text": "A text-to-SQL prompt typically comprises four components: a task instruction, a test database, a test NLQ, and optional demonstrations, as illustrated in Figure 1. While the task instruction and test NLQ are easily presented in natural language, there are various strategies for representing the databases and incorporating demonstrations. In this section, we explore different prompt constructions for databases and demonstrations."
},
{
"figure_ref": [
"fig_0",
"fig_1"
],
"heading": "Database Prompt",
"publication_ref": [
"b20",
"b20",
"b23",
"b20",
"b29",
"b17",
"b26",
"b23",
"b5"
],
"table_ref": [],
"text": "A relational database consists of the database schema and database content. The database schema encompasses the schemas (headers) of tables and the relationship among tables, and database content refers to the data stored in the tables.\nDatabase Schema Figure 2 illustrates various prompt constructions for the database schema that have been utilized in previous studies: (1) and Rafiei, 2023) represents each table along with +FK (Pourreza and Rafiei, 2023) Foreign_keys = [ Friend . student_id = Highschooler . ID , Friend . friend_id = Highschooler . ID ];\nCreateTable (Rajkumar et al., 2022) CREATE a list of its columns using an equation-like notation;\n(3) +ForeignKey (Pourreza and Rafiei, 2023) To ensure consistency in the prompt text and accommodate the case-insensitivity of SQL keywords and the database schema, we unify the space and line break in the prompt text and convert all words to lowercase, except for the database content. This normalization process helps to standardize the prompt text. An example is shown in Figure 4.\nDatabase content Previous research shows that being aware of database content can improve model performance by exposing models to the specific format of values in each column (Wang et al., 2019;Lin et al., 2020;Scholak et al., 2021;Rajkumar et al., 2022). For instance, the phrase \"American student\" could be converted to \"WHERE country = 'USA'\" or \"WHERE country = 'The United States of America'\" depending on the contents of the country column.\nFigure 3 summarizes different approaches used to construct prompts for showcasing the content of a database. (1) InsertRow (Chen et al., 2023) . (3) SelectCol: Instead of presenting table content in a row-wise manner, an alternative method is to use a column-wise format. As there may be duplicated content across different rows, presenting the content column-wise ensures the provision of distinct values within each column to expose LLMs to a broader range of content. We propose using the query \"SELECT DISTINCT [Column] FROM [Table] LIMIT R\" to list R distinct cell values in each column."
},
{
"figure_ref": [
"fig_1"
],
"heading": "Demonstration Prompt",
"publication_ref": [
"b23",
"b20",
"b19",
"b5"
],
"table_ref": [],
"text": "In few-shot settings, LLMs are provided with demonstrations within the prompt text. In the single-domain few-shot setting, we incorporate a few pairs of NLQs and SQLs as demonstrations inserted between the test database and question, following previous work (Rajkumar et al., 2022). In the cross-domain few-shot setting, we use both out-of-domain NLQ-SQL pairs (demonstration examples) and corresponding databases (demonstration databases) placed before the test database and question. Prior research in the N -shot setting either uses one demonstration database with N examples (Pourreza and Rafiei, 2023) or employs N demonstration databases, each with a single NLQ-SQL pair (Poesia et al., 2022;Chen et al., 2023). In contrast, we consider a more general scenario where the demonstrations comprise M databases, each with K NLQ-SQL pairs, with M × K = N . We list the examples of 4-shot single-domain and crossdomain demonstrations in Appendix A.1. Additionally, we normalize demonstration SQL queries by first parsing the SQL queries and unifying their format, such as using lowercase for SQL keywords and database schema and unifying the space around punctuation. Figure 4 provides an example of SQL normalization."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [
"b31",
"b23",
"b11"
],
"table_ref": [],
"text": "Data & Evaluation For our experiments, we utilize the Spider dataset (Yu et al., 2018), a crossdomain benchmark for the text-to-SQL task. We conduct our experiments on the development set of Spider (Spider-dev) as the test set is not publicly available. Spider-dev consists of 20 databases with 1034 pairs of NLQ and SQL in total. We evaluate models with execution accuracy (EX) which compares the execution results of a predicted SQL and a gold SQL.\nIn the cross-domain setting, we use the training set of Spider to select demonstration examples. As a few databases contain long schema that may cause the prompt to exceed the token limits of LLMs, we only use the databases with fewer than 1000 tokens when constructing the CreateTable prompt. This results in a total of 130 databases being used as demonstration databases in the crossdomain setting.\nModels We used GPT-3 Codex (Chen et al., 2021a) and ChatGPT due to their demonstrated performance and prevalence in the field. 2Experiment Setup For the zero-shot setting, we construct each prompt text with a task instruction, a test database, and a test question. We include R = 3 table rows in the database prompt, which has been discovered as the optimal number in previous work (Rajkumar et al., 2022). For the few-shot settings, we incorporate N demonstration examples in addition to the zero-shot prompt text.\nIn the single-domain text-to-SQL scenario, we use a leave-one-out split, as some databases in Spider-dev contain a small number of examples. When evaluating one example, we regard all other examples from the same database as the training set and randomly retrieve N examples from them. Since Spider contains multiple NLQs corresponding to the same SQL query, we require that the training set does not contain examples with the same SQL template as the test example, again following previous work (Finegan-Dollak et al., 2018).\nIn the cross-domain scenario, we randomly select M demonstration databases, each with K NLQ-SQL pairs (M × K = N ) from the Spider training set. Incorporating multiple demonstration databases in a prompt text significantly increases its length. Hence, we only use Codex for the crossdomain experiments, due to its higher token limit of 8K, surpassing the 4K limit of ChatGPT. In both single-domain and cross-domain settings, we compare different prompt construction methods using the same few-shot examples to make a fair comparison. We repeat our experiments three times and present the average results."
},
{
"figure_ref": [],
"heading": "Results",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we present our empirical findings in the areas of zero-shot, single-domain, and crossdomain text-to-SQL. Through our experiments, we aim to answer a few crucial research questions in each setting and provide insightful strategies for future studies on effective prompting."
},
{
"figure_ref": [],
"heading": "Zero-shot Text-to-SQL",
"publication_ref": [],
"table_ref": [],
"text": "In the zero-shot setting, we focus on comparing different prompt constructions for databases. Table Table 1: Zero-shot results of Codex and ChatGPT using different database prompt constructions. Table Schema (upper part) contains prompts that solely include the schema of tables, while +Relationship (middle part) incorporates foreign keys as the table relationships and +Relationship+Content (lower part) adds table content as well. # Tokens is the average token counts in the prompts and EX represents the execution accuracy of SQLs. U|N represents the results of unnormalized prompts and normalized prompts, respectively. The underlines highlight the lower number of tokens and higher accuracies when comparing unnormalized and normalized prompts and the highest accuracy achieved among all prompts is highlighted in bold.\n1 shows the average prompt length and execution accuracy of Codex and ChatGPT using various database prompt constructions. Q1: How does normalized database prompt perform compared to unnormalized ones? Normalized schemas are found to have a reduced token count in comparison to unnormalized schemas across all database constructions. The normalization also tends to yield slightly better performance. As for Codex, normalized schemas show improvement in all prompts. For ChatGPT, normalized schemas either improve accuracy or achieve the same accuracy or achieve the same level of accuracy as unnormalized schemas in 6 out of 7 schema constructions. Moveover, exposing LLMs to database content with the SelectRow and SelectCol prompts further enhances the performance of both Codex and ChatGPT, while the InsertRow prompt does not seem to be beneficial. We believe that database content is valuable, but its representation needs to be carefully chosen."
},
{
"figure_ref": [],
"heading": "Q3: How does Codex perform compared to",
"publication_ref": [],
"table_ref": [],
"text": "ChatGPT? While we do not focus on comparing different LLMs on the text-to-SQL tasks in this paper, it is worth noting that Codex consistently outperforms ChatGPT on zero-shot text-to-SQL using various prompt constructions.\nBased on all the findings above, we would recommend using Codex in conjunction with normalized CreateTableSelectCol prompt construction for zero-shot text-to-SQL.3 "
},
{
"figure_ref": [],
"heading": "Single-domain Text-to-SQL",
"publication_ref": [],
"table_ref": [],
"text": "In the zero-shot text-to-SQL setting, we discovered that the prompt constructions of databases impact the performance of LLMs. This discovery naturally raises the question of whether the introduction of in-domain demonstrations affects the performance of LLMs to different database prompts. Q1: Does the use of in-domain demonstrations enhance LLM's performance? Figure 5 depicts the performance of Codex and ChatGPT using different database prompt constructions with respect to different numbers of in-domain demonstration examples. For all database prompts, the performance of LLMs experiences a notable improvement when in-domain examples are presented. Furthermore, the performance continues to enhance as the number of in-domain examples increases. Q2: What database knowledge is important when presenting in-domain demonstrations? While we have observed that the presence of table To summarize, in single-domain text-to-SQL, we recommend incorporating a greater number of in-domain examples whenever feasible. It is also essential to ensure the presence of table content in conjunction with the table schema while the specific choice of table content construction is less crucial compared to the zero-shot scenario."
},
{
"figure_ref": [],
"heading": "Cross-domain Text-to-SQL",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we present the results to answer a series of questions regarding the demonstration and database prompt construction."
},
{
"figure_ref": [
"fig_0",
"fig_0",
"fig_5",
"fig_7"
],
"heading": "Impact of Demonstration Prompt",
"publication_ref": [],
"table_ref": [],
"text": "To investigate the impact of the number of databases and examples per database in demonstrations, we conduct experiments encompassing various combinations. Specifically, our demonstrations are composed of M demonstration databases, each containing K NLQ-SQL pairs. We consider scenarios with up to 8 databases and 16 examples per database as long as the combination does not exceed the prompt length limit. We opt to use the database prompt CreateTable+SelectRow 3 as it contains fewer tokens compared to InsertRow and SelectCol while encompassing all valuable database knowledge. We present the experiments with Codex in this section. Experiments involving ChatGPT-16K can be found in Appendix A.4 which show similar results as Codex. Q1: Does increasing demonstration examples enhance LLMs' performance? Figure 2 presents the accuracy of Codex corresponding to different combinations of the number of databases and the number of examples per database used as demonstrations. We analyze the results from two perspectives. Firstly, for a fixed number of databases, we observe an initial improvement in Codex's performance as the number of examples per database increases. However, this improvement plateaus or declines once 4 examples per database are provided. Surprisingly, when using 4 databases, employing 8\nor 16 examples per database leads to a significant decrease in the Codex's performance compared to using 2 or 4 examples per database. Secondly, for a fixed number of examples per database, we observe an initial increase in Codex's performance as the number of databases increases, however, this improvement is followed by a significant decrease once the number of databases reaches a certain threshold (either 4 or 6). Q2: Why does increasing the number of databases decrease LLMs' performance? As depicted in Figure 2, presenting more databases does not always lead to improved performance. In fact, there is a significant decline in performance, once it surpasses a threshold. We hypothesize that this phenomenon is attributed to the length of the prompt text. To test this hypothesis, we analyze the results in relation to the prompt length.\nFigure 7 shows the relationship between the accuracy of different demonstration prompts and their prompt lengths. Notably, the performance of Codex exhibits an inverted-U shape as the prompt length increases for each number of examples per database. Additionally, we observe a substantial drop in performance once the prompt text length exceeds approximately 5500 tokens. Similarly, Figure 9 shows that the performance of ChatGPT-16K starts to decrease when prompt text length exceeds 11K tokens. Based on these observations, we conjecture that LLMs may have a sweet spot in terms of prompt length, potentially influenced by factors such as their model architecture or training data. This indicates that even though LLMs are capable of handling long contexts, they may not necessarily perform better with excessively long prompts."
},
{
"figure_ref": [],
"heading": "Impact of Database Prompt",
"publication_ref": [],
"table_ref": [],
"text": "Since incorporating demonstration databases may cause a decrease in Codex's performance, we focus our database prompt experiments on using one demonstration database in combination with varying quantities of demonstration examples. In conclusion, while out-of-domain demonstrations enhance LLMs' capabilities in text-to-SQL, they do not provide database-specific knowledge. Consequently, careful construction of database prompts remains crucial, aligning with the observations made in the zero-shot setting."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b19",
"b14",
"b20",
"b5",
"b23",
"b0",
"b13",
"b17",
"b27",
"b10",
"b22",
"b17",
"b29",
"b2",
"b16",
"b17",
"b27",
"b3",
"b23",
"b5"
],
"table_ref": [],
"text": "LLMs for Text-to-SQL In recent years, there has been significant progress in leveraging LLMs for the text-to-SQL task. Various methods have been proposed to enhance the capabilities of LLMs. For example, Rubin et al. (2021); Poesia et al. (2022) have demonstrated the effectiveness of similarity-based demonstration retrieval in the cross-domain setting. Additionally, Levy et al. (2022) have highlighted the advantages of incorporating diverse demonstrations for compositional generalization. Furthermore, Pourreza and Rafiei (2023) and Chen et al. (2023) incorporate intermediate steps in prompts and unlock LLMs' capability of self-correcting their predictions.\nIn contrast to these approaches, our focus lies in conducting a comprehensive evaluation of prompt representations across different text-to-SQL settings. While there are similar motivations to the work by Rajkumar et al. (2022), which analyzes the performance of CodeX on Spider for the zero-shot setting and on two databases for the single-domain setting, we aim to provide more general findings by evaluating across a wider range of databases and considering all three text-to-SQL settings.\nTable Representation Encoding structured databases with neural models has been a persistent challenge. To encode database schema, graph neural networks are utilized to represent the relationships among tables (Bogin et al., 2019;Chen et al., 2021b). Alternatively, other studies (Guo et al., 2019;Lin et al., 2020;Shaw et al., 2020) have converted table schemas into a sequence to effectively leverage pretrained language models, such as BERT (Devlin et al., 2018) and T5 (Raffel et al., 2020). In such cases, table relationships can be encoded as meta-data features (Lin et al., 2020) or used as a guide for attention mechanism (Wang et al., 2019;Cao et al., 2021;Li et al., 2023b).\nTo incorporate table content into neural models, prior supervised methods provide questionspecific table content by identifying the relevant table content mentioned in the question through string matching (Lin et al., 2020;Shaw et al., 2020). However, Chang et al. (2023) have revealed the vulnerability of string matching to perturbations. Given that LLMs with in-context learning support longer input sequences compared to supervised methods, we follow previous work to provide table content without explicitly considering the questions (Rajkumar et al., 2022;Chen et al., 2023).\nIn this paper, we investigate effective prompting strategies in the text-to-SQL task. We thoroughly compare various prompt construction strategies for databases and demonstrations in the zeroshot, single-domain, and cross-domain text-to-SQL. Through our investigation, we uncover the critical database knowledge and optimal representations for effective prompting. Additionally, an interesting finding is the existence of a sweet spot in terms of prompt length for Codex in the cross-domain setting. Overall, we believe that our findings will provide valuable guidance for future research in the field of text-to-SQL with LLMs."
},
{
"figure_ref": [],
"heading": "Limitation",
"publication_ref": [],
"table_ref": [],
"text": "We conducted our experiments using 20 databases from the Spider dataset, with the goal of providing general findings for text-to-SQL prompt constructions. However, our findings may not always be applicable to a specific database, particularly if the database is significantly different from the Spider databases. For the single-domain and cross-domain text-to-SQL scenarios, we conduct our experiments multiple times, each involving randomly selecting demonstrations with different random seeds, however, we did not investigate the effectiveness of prompt constructions with different demonstrationretrieval strategies or intermediate reasoning steps."
},
{
"figure_ref": [],
"heading": "A Appendix",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "A.1 Prompt Examples",
"publication_ref": [
"b31"
],
"table_ref": [],
"text": "Below contains an example of a zero-shot normalized prompt, which contains the database Network_1 from Spider (Yu et al., 2018), a task instruction \"Using valid SQLite, answer the following questions for the tables provided above.\", and a test question \"How many high schoolers are there?\". "
},
{
"figure_ref": [],
"heading": "Zero",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "A.2 Tests of Significance",
"publication_ref": [],
"table_ref": [],
"text": "Table 1 contains the performance of Codex and ChatGPT using different database prompt constructions in the zero-shot setting. We observe that the normalization results in slightly improved performance for all database prompt constructions with Codex and 6 out of 7 database prompt constructions with ChatGPT. It is important to note, however, that when comparing normalized and unnormalized database prompt constructions using the same method, the results did not demonstrate statistical significance in McNemar's test, with p-values greater than 0.05. Nevertheless, the primary advantage of normalization lies in its ability to reduce variations among different databases and minimize the overall prompt length.\nWhen evaluating various prompt constructions, we note the advantages gained from incorporating both table relationships (Columns=[]+ForeignKey vs Columns=[]) and table content (CreateTable+SelectCol 3 vs CreateTable) are mostly statistically significant in McNemar's test, with p-values smaller than 0.05. Table 3 displays the results of the significant tests. The performance of Columns=[]+ForeignKey compared to Columns=[] is statistically significant in all cases, except for codex with normalized prompts. Likewise, the performance of CreateTable+SelectCol 3 is statistically significant for both Codex and ChatGPT, with both normalized and unnormalized prompts, when compared to CreateTable. These significant findings highlight the effectiveness of incorporating table relationships and database content."
},
{
"figure_ref": [
"fig_7"
],
"heading": "A.3 Detailed Single-domain Results",
"publication_ref": [],
"table_ref": [],
"text": "Tables 4 and5 provide detailed results of Codex and ChatGPT in the single-domain setting, respectively. The performance of both models is also illustrated in Figure 5. performance as the number of databases increases, however, this improvement is followed by a decrease once the number of databases reaches a certain threshold. To understand this phenomenon, we analyze the results in relation to the prompt length.\nFigure 9 shows the relationship between the accuracy of different demonstration prompts and their prompt lengths. Similar to Codex, the performance of ChatGPT-16K also exhibits an inverted-U shape as the prompt length increases for each number of examples per database. Additionally, we observe the performance starts to decrease once the prompt text length exceeds approximately 11K tokens.\nWhile Codex supports 8K tokens and ChatGPT-16K supports 16K tokens, we notice that their performance tends to decline when dealing with demonstrations that exceed approximately 70% of the maximum prompt length."
},
{
"figure_ref": [],
"heading": "Ethics Statement",
"publication_ref": [],
"table_ref": [],
"text": "We acknowledge the importance of the ACL Ethics Policy and agree with it. In this paper, we use Ope-nAI Codex and ChatGPT as our language models 4 . Codex is currently free for research purposes, the cost of ChatGPT is around $200. The code for the paper is included in the supplementary materials and will be publicly released to facilitate reproducibility. "
}
] | Large language models (LLMs) with in-context learning have demonstrated remarkable capability in the text-to-SQL task. Previous research has prompted LLMs with various demonstration-retrieval strategies and intermediate reasoning steps to enhance the performance of LLMs. However, those works often employ varied strategies when constructing the prompt text for text-to-SQL inputs, such as databases and demonstration examples. This leads to a lack of comparability in both the prompt constructions and their primary contributions. Furthermore, selecting an effective prompt construction has emerged as a persistent problem for future research. To address this limitation, we comprehensively investigate the impact of prompt constructions across various settings and provide insights into prompt constructions for future text-to-SQL studies. | How to Prompt LLMs for Text-to-SQL: A Study in Zero-shot, Single-domain, and Cross-domain Settings | [
{
"figure_caption": "Figure 2 :2Figure 2: Examples of the different database schema constructions for a snippet of database Network_1 in Spider.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: An example of the normalization for database and SQL prompts.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Figure 6 :6Figure 6: A heat map of Codex's execution accuracy using CreateTable+SelectRow 3 for different numbers of databases and examples per database in the demonstration. Darker color indicates higher accuracy.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "Figure 7 :7Figure 7: Execution accuracy of Codex in relation to the length of prompts. Each dot on the graph represents a specific demonstration prompt construction, with the m, k denoting the number of databases and examples per database used in the prompt. The lines represent second-degree polynomial trendlines fitted to the results.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "7",
"figure_type": "figure"
},
{
"figure_caption": "Figure 8 Figure 8 :88Figure 8 presents the accuracy of ChatGPT-16K corresponding to different combinations of the number of databases and the number of examples per database used as demonstrations. Similar to our findings with Codex, presenting more databases does not always lead to improved performance for ChatGPT-16K. For a fixed number of examples per database, we observe an initial increase in its",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "88",
"figure_type": "figure"
},
{
"figure_caption": "Figure 9 :9Figure 9: Execution accuracy of ChatGPT-16K in relation to the length of prompts. Each dot represents a demonstration construction, with the m, k denoting the number of databases and examples per database. The lines represent second-degree polynomial trendlines fitted to the results.",
"figure_data": "",
"figure_id": "fig_7",
"figure_label": "9",
"figure_type": "figure"
},
{
"figure_caption": "Columns)(Liu et al., 2023) ",
"figure_data": "Highschooler ( ID , name , grade ) ;Friend ( student_id , friend_id ) ;Columns=[] (Pourreza and Rafiei, 2023)",
"figure_id": "tab_1",
"figure_label": "(",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": "(ID int primary key , name text ,grade int) ;CREATE TABLE Friend (student_id int ,friend_id int ,primary key ( student_id , friend_id ) ,foreign key ( student_id ) referencesHighschooler ( ID ) ,foreign key ( friend_id ) referencesHighschooler ( ID )) ;",
"figure_id": "tab_3",
"figure_label": "Highschooler",
"figure_type": "table"
},
{
"figure_caption": "InsertRow(Chen et al., 2023) ",
"figure_data": "INSERT INTO Highschooler ( ID , name , grade )VALUES (1510 , \" Jordan \" , 9) ;INSERT INTO Highschooler ( ID , name , grade )VALUES (1689 , \" Gabriel \" , 9) ;INSERT INTO Highschooler ( ID , name , grade )VALUES (1381 , \" Tiffany \" , 9) ;SelectRow (Rajkumar et al., 2022)/*3 example rows :SELECT * FROM Highschooler LIMIT 3;IDnamegrade1510Jordan91689Gabriel91381Tiffany9*/SelectCol (Ours)/*Columns in Highschooler and 3 distinctexamples in each column :ID : 1025 , 1101 , 1247name : \" Jordan \" , \" Gabriel \" , \" Tiffany \"grade : 9 , 10 , 11*/Figure 3: Examples of the different database contentconstructions for showing 3 cell values in each columnfor the Highschool table in Figure 2.This method displays R rows of each table by utiliz-ing R \"INSERT INTO\" statements. (2) SelectRow(Rajkumar et al., 2022): This approach employs the\"SELECT * FROM Table LIMIT R\" query to displaythe first R rows of each table",
"figure_id": "tab_5",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "tab_8",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": "Execution AccuracyColumns) Columns=[] Table+RS Columns=[]+ForeignKey CreateTable Table+RS+Cont CreateTable+InsertRow CreateTable+SelectRow CreateTable+SelectCol# Shots",
"figure_id": "tab_9",
"figure_label": "(",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "tab_10",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Execution accuracy of Codex and ChatGPT for single-domain text-to-SQL with 1, 4, 8, and 16 indomain examples. RS and Cont correspond to table relationship and table content, respectively. Detailed results can be found in Table 4 and 5. relationships and table content enhanced LLMs' performance in the zero-shot scenario, it is not clear whether they are still important in the singledomain setting. A hypothesis is that table relationship and table content knowledge can be acquired from in-domain examples as they may appear in SQL clauses JOIN and WHERE. For table relationships, we compare two database prompt constructions Columns=[] and Columns=[]+ForeignKey. Both construct the table schema in the same way while the latter includes foreign keys as table relationships. In the zero-shot scenario, Columns=[]+ForeignKey outperforms Columns=[] by 1.3 and 2.1 for Codex and ChatGPT, respectively. However, as increasing the number of in-domain examples, we notice a gradual reduction in the performance gap between these two prompts. With the utilization of 16 indomain examples, the gap completely disappears for Codex, while ChatGPT exhibits a marginal difference of only 0.5%. For table content, we compare CreateTable with CreateTable+SelectCol. Both contain the same prompts for presenting the table schema and relationship, while the latter additionally includes table content. In the zero-shot scenario, CreateTable+SelectCol outperforms CreateTable by 2.0% for Codex and 1.7% for ChatGPT. As we proceed to increase the number of in-domain examples, we observe that the performance gap between these two prompts does not exhibit a significant reduction. Even with 16 indomain examples, the gap still persists at 1.3 for Codex and 1.9 for ChatGPT. These results indicate LLMs are able to quickly learn table relationships from a small number of indomain demonstrations, however, it is more challenging to obtain table content knowledge from demonstration examples. Consequently, the inclusion of table content remains crucial for achieving satisfactory performance in the single-domain textto-SQL scenario. Q3: Can in-domain demonstrations alleviate the sensitivity of LLMs to the representation of table content? In the zero-shot setting, we observe that LLMs are sensitive to how the table content is presented. Specifically, SelectCol 3 outperforms InsertRow 3 by a substantial margin of 3.8 for Codex and 1.8 for ChatGPT. However, as we expose LLMs to in-domain demonstrations, LLMs become less sensitive to the specific representation of table content. The performance disparities among the three table content prompts become marginal. Notably, with only 4 examples, the performance difference between SelectCol 3 and InsertRow 3 diminishes to 0.3 for Codex and 0.2 for ChatGPT.",
"figure_data": "Execution AccuracyTable(Columns) Columns=[] Table+RS Columns=[]+ForeignKey CreateTable Table+RS+Cont CreateTable+InsertRow CreateTable+SelectRow CreateTable+SelectCol# Shots(b) ChatGPTFigure 5:",
"figure_id": "tab_11",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "We observe an initial performance increase for all database prompts. However, once more than 4 examples are provided, the improvement starts to level off, indicating that the different database prompts exhibit similar trends in relation to the number of demonstration examples.",
"figure_data": "presents the execution accuracy of Codex usingdifferent database prompts.Q3: Do different database prompts show similartrends with the number of demonstration exam-ples?",
"figure_id": "tab_13",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Cross-domain results of Codex using different database prompt constructions. Only one demonstration database is included in a prompt, N-shot represents N examples corresponding to the demonstration database. The best and second-best results for each shot are highlighted in bold and underlined.Q4: Can out-of-domain demonstrations alleviate the sensitivity of LLMs to database prompts? First, we observe that incorporating table relationships and content in the prompts remains crucial for effectively prompting Codex in the cross-domain setting. This is not surprising, as Codex cannot directly learn knowledge specific to the test database from the out-of-domain demonstrations. Furthermore, we find that Codex continues to exhibit sensitivity to the representation of table content. Despite having demonstration databases that mirror the construction of the test database, Codex still displays a preference forSelectRow and SelectCol when presenting table content, compared to InsertCol.",
"figure_data": "",
"figure_id": "tab_14",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "-shot normalized prompt",
"figure_data": "Below contains an example of a 4-shot cross-Below contains an example of a 4-shot single-seating real , year_opened real ,domain prompt, which contains 2 demonstration databases, each with 2 demonstration examplesdomain normalized prompt, which contains a database prompt and 4 demonstration examples primary key ( track_id ) ) ; /*ahead of the test database and question.ahead of the test question. 3 example rows : select * from track limit 3;track_idnamelocationseating4-shot cross-domain prompt4-shot single-domain normalized prompt year_opened 1 Auto Club Speedway Fontana , CAcreate table highschooler ( id int primary key , name text , grade int ) ; /* 3 example rows : select * from highschooler limit 3; id name grade 1510 Jordan 9 1689 Gabriel 9 1381 Tiffany 9 */ create table friend ( student_id int , friend_id int , primary key ( student_id , friend_id ) , foreign key ( student_id ) references highschooler ( id ) , foreign key ( friend_id ) references highschooler ( id ) ) ; /* 3 example rows : select * from friend limit 3; student_id friend_id 1510 1381 1510 1689 1689 1709 */ create table likes ( student_id int , liked_id int , primary key ( student_id , liked_id ) , foreign key ( liked_id ) references highschooler ( id ) , foreign key ( student_id ) references highschooler ( id ) ) ; /* 3 example rows : select * from likes limit 3; student_id liked_id 1689 1709 1709 1689 1782 1709 */ --Using valid SQLite , answer the following questions for the tables provided above . Question : How many high schoolers are there ? select create table publication ( publication_id int , book_id int , publisher text , publication_date text , price real , primary key ( publication_id ) , foreign key ( book_id ) references book ( book_id ) ) ; /* 3 example rows : select * from publication limit 3; publication_id book_id publisher publication_date price 1 1 Pearson August 2008 15000000.0 2 3 Thomson Reuters March 2008 6000000.0 3 4 Wiley June 2006 4100000.0 */ create table book ( book_id int , title text , issues real , writer text , primary key ( book_id ) ) ; /* 3 example rows : select * from book limit 3; book_id title issues writer 1 The Black Lamb 6.0 Timothy Truman 2 Bloody Mary 4.0 Garth Ennis 3 Bloody Mary : Lady Liberty 4.0 Garth Ennis */ --Using valid SQLite , answer the following questions for the tables provided above . Question : List the writers of the books in ascending alphabetical order . select writer from book order by writer asc ; Question : How many books are there ? select count (*) from book ; create table race ( race_id int , name text , class text , date text , track_id text , primary key ( race_id ) , foreign key ( track_id ) references track ( track_id ) ) ; /* 3 example rows : select * from race limit 3; race_id name class date track_id 1 Rolex 24 At Daytona DP / GT January 26 January 27 1 2 Gainsco Grand Prix of Miami DP / GT March 29 2 3 Mexico City 250 DP / GT April 19 2 */ create table track ( track_id int , name text , location text ,create table highschooler ( id int primary key , name text , grade int ) ; /* 92000.0 1997.0 2 Chicagoland Speedway 75000.0 2001.0 3 Darlington Raceway 63000.0 1950.0 */ 3 example rows : select * from highschooler limit 3; Joliet , IL Darlington , SC id name grade 1510 Jordan 9 1689 Gabriel 9 1381 Tiffany 9 */ create table friend ( student_id int , friend_id int , --Using valid SQLite , answer the following questions for the tables provided above . Question : Show the name and location for all tracks . select name , location from the track ; Question : Show the name of track and the number of races in each track . select t2 . name , count (*) from race as t1 join track as t2 on t1 . track_id = t2 . track_id group by t1 . track_id ; primary key ( student_id , friend_id ) , foreign key ( student_id ) references highschooler ( id ) , foreign key ( friend_id ) references create table highschooler ( id int primary key , name text , highschooler ( id ) ) ; /* 3 example rows : select * from friend limit 3; student_id friend_id 1510 1381 1510 1689 1689 1709 */ grade int ) ; /* 3 example rows : select * from highschooler limit 3; id name grade 1510 Jordan 9 1689 Gabriel 9 1381 Tiffany 9 */ create table likes ( student_id int , liked_id int , primary key ( student_id , liked_id ) , foreign key ( liked_id ) references highschooler ( id ) , create table friend ( student_id int , friend_id int , primary key ( student_id , friend_id ) , foreign key ( student_id ) references highschooler ( id ) , foreign key ( student_id ) references highschooler ( id ) ) ; foreign key ( friend_id ) references highschooler ( id ) ) ; /* 3 example rows : select * from likes limit 3; student_id liked_id 1689 1709 1709 1689 1782 /* 3 example rows : select * from friend limit 3; student_id friend_id 1510 1381 1510 1689 1709 */ --Using valid SQLite , answer the following questions for the tables provided above . Question : What is Kyle 's id ? select id from highschooler where name = ' Kyle '; Question : Return the names of friends of the high school student Kyle . select t3 . name from friend as t1 join highschooler as t2 on t1 . student_id = t2 . id join highschooler as t3 on t1 . friend_id = t3 . id where t2 . name = ' Kyle '; Question : Show names of all high school students who do not have any friends . select name from highschooler except select t2 . name from friend as t1 join highschooler as t2 on t1 . student_id = t2 . id ; Question : What are the names and grades for 1689 1709 */ create table likes ( student_id int , liked_id int , primary key ( student_id , liked_id ) , foreign key ( liked_id ) references highschooler ( id ) , foreign key ( student_id ) references highschooler ( id ) ) ; /* 3 example rows : select * from likes limit 3; student_id liked_id 1689 1709 1709 1689 1782 1709 */ each high schooler ? --Using valid SQLite , answer the following select name , grade from highschooler ; Question : How many high schoolers are there ? questions for the tables provided above . select Question : How many high schoolers are there ? select",
"figure_id": "tab_15",
"figure_label": "",
"figure_type": "table"
}
] | Shuaichen Chang; Eric Fosler-Lussier | [
{
"authors": "Ben Bogin; Matt Gardner; Jonathan Berant",
"journal": "",
"ref_id": "b0",
"title": "Representing schema structure with graph neural networks for text-to-sql parsing",
"year": "2019"
},
{
"authors": "Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell",
"journal": "Advances in neural information processing systems",
"ref_id": "b1",
"title": "Language models are few-shot learners",
"year": "2020"
},
{
"authors": "Ruisheng Cao; Lu Chen; Zhi Chen; Yanbin Zhao; Su Zhu; Kai Yu",
"journal": "",
"ref_id": "b2",
"title": "Lgesql: line graph enhanced text-to-sql model with mixed local and nonlocal relations",
"year": "2021"
},
{
"authors": "Shuaichen Chang; Jun Wang; Mingwen Dong; Lin Pan; Henghui Zhu; Alexander Hanbo Li; Wuwei Lan; Sheng Zhang; Jiarong Jiang; Joseph Lilien",
"journal": "",
"ref_id": "b3",
"title": "Dr. spider: A diagnostic evaluation benchmark towards text-to-sql robustness",
"year": "2023"
},
{
"authors": "Mark Chen; Jerry Tworek; Heewoo Jun; Qiming Yuan; Henrique Ponde De Oliveira Pinto; Jared Kaplan; Harri Edwards; Yuri Burda; Nicholas Joseph; Greg Brockman",
"journal": "",
"ref_id": "b4",
"title": "Evaluating large language models trained on code",
"year": "2021"
},
{
"authors": "Xinyun Chen; Maxwell Lin; Nathanael Schärli; Denny Zhou",
"journal": "",
"ref_id": "b5",
"title": "Teaching large language models to self-debug",
"year": "2023"
},
{
"authors": "Zhi Chen; Lu Chen; Yanbin Zhao; Ruisheng Cao; Zihan Xu; Su Zhu; Kai Yu",
"journal": "",
"ref_id": "b6",
"title": "Shadowgnn: Graph projection neural network for text-to-sql parser",
"year": "2021"
},
{
"authors": "Zhoujun Cheng; Tianbao Xie; Peng Shi; Chengzu Li; Rahul Nadkarni; Yushi Hu; Caiming Xiong; Dragomir Radev; Mari Ostendorf; Luke Zettlemoyer",
"journal": "",
"ref_id": "b7",
"title": "Binding language models in symbolic languages",
"year": "2022"
},
{
"authors": "Aakanksha Chowdhery; Sharan Narang; Jacob Devlin; Maarten Bosma; Gaurav Mishra; Adam Roberts; Paul Barham; Hyung Won Chung; Charles Sutton; Sebastian Gehrmann",
"journal": "",
"ref_id": "b8",
"title": "Palm: Scaling language modeling with pathways",
"year": "2022"
},
{
"authors": "Deborah A Dahl; Madeleine Bates; Michael Brown; William Fisher; Kate Hunicke-Smith; David Pallett; Christine Pao; Alexander Rudnicky; Elizabeth Shriberg",
"journal": "",
"ref_id": "b9",
"title": "Expanding the scope of the atis task: The atis-3 corpus",
"year": "1994"
},
{
"authors": "Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova",
"journal": "",
"ref_id": "b10",
"title": "Bert: Pre-training of deep bidirectional transformers for language understanding",
"year": "2018"
},
{
"authors": "Catherine Finegan-Dollak; Jonathan K Kummerfeld; Li Zhang; Karthik Ramanathan; Sesh Sadasivam; Rui Zhang; Dragomir Radev",
"journal": "",
"ref_id": "b11",
"title": "Improving text-to-sql evaluation methodology",
"year": "2018"
},
{
"authors": "Yujian Gan; Xinyun Chen; Jinxia Xie; Matthew Purver; John R Woodward; John Drake; Qiaofu Zhang",
"journal": "",
"ref_id": "b12",
"title": "Natural sql: Making sql easier to infer from natural language specifications",
"year": "2021"
},
{
"authors": "Jiaqi Guo; Zecheng Zhan; Yan Gao; Yan Xiao; Jian-Guang Lou; Ting Liu; Dongmei Zhang",
"journal": "",
"ref_id": "b13",
"title": "Towards complex text-to-sql in cross-domain database with intermediate representation",
"year": "2019"
},
{
"authors": "Itay Levy; Ben Bogin; Jonathan Berant",
"journal": "",
"ref_id": "b14",
"title": "Diverse demonstrations improve in-context compositional generalization",
"year": "2022"
},
{
"authors": "Haoyang Li; Jing Zhang; Cuiping Li; Hong Chen",
"journal": "",
"ref_id": "b15",
"title": "Decoupling the skeleton parsing and linking for text-to-sql",
"year": "2023"
},
{
"authors": "Jinyang Li; Binyuan Hui; Reynold Cheng; Bowen Qin; Chenhao Ma; Nan Huo; Fei Huang; Wenyu Du; Luo Si; Yongbin Li",
"journal": "",
"ref_id": "b16",
"title": "Graphix-t5: Mixing pretrained transformers with graph-aware layers for textto-sql parsing",
"year": "2023"
},
{
"authors": "Victoria Xi; Richard Lin; Caiming Socher; Xiong",
"journal": "",
"ref_id": "b17",
"title": "Bridging textual and tabular data for crossdomain text-to-sql semantic parsing",
"year": "2020"
},
{
"authors": "Aiwei Liu; Xuming Hu; Lijie Wen; Philip S Yu",
"journal": "",
"ref_id": "b18",
"title": "A comprehensive evaluation of chatgpt's zero-shot text-to-sql capability",
"year": "2023"
},
{
"authors": "Gabriel Poesia; Oleksandr Polozov; Vu Le; Ashish Tiwari; Gustavo Soares; Christopher Meek; Sumit Gulwani",
"journal": "",
"ref_id": "b19",
"title": "Synchromesh: Reliable code generation from pre-trained language models",
"year": "2022"
},
{
"authors": "Mohammadreza Pourreza; Davood Rafiei",
"journal": "",
"ref_id": "b20",
"title": "Din-sql: Decomposed in-context learning of text-to-sql with self-correction",
"year": "2023"
},
{
"authors": "Patti Price",
"journal": "",
"ref_id": "b21",
"title": "Evaluation of spoken language systems: The atis domain",
"year": "1990-06-24"
},
{
"authors": "Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu",
"journal": "J. Mach. Learn. Res",
"ref_id": "b22",
"title": "Exploring the limits of transfer learning with a unified text-to-text transformer",
"year": "2020"
},
{
"authors": "Nitarshan Rajkumar; Raymond Li; Dzmitry Bahdanau",
"journal": "",
"ref_id": "b23",
"title": "Evaluating the text-to-sql capabilities of large language models",
"year": "2022"
},
{
"authors": "Ohad Rubin; Jonathan Berant",
"journal": "",
"ref_id": "b24",
"title": "Smbop: Semiautoregressive bottom-up semantic parsing",
"year": "2021"
},
{
"authors": "Ohad Rubin; Jonathan Herzig; Jonathan Berant",
"journal": "",
"ref_id": "b25",
"title": "Learning to retrieve prompts for in-context learning",
"year": "2021"
},
{
"authors": "Torsten Scholak; Nathan Schucher; Dzmitry Bahdanau",
"journal": "",
"ref_id": "b26",
"title": "Picard: Parsing incrementally for constrained auto-regressive decoding from language models",
"year": "2021"
},
{
"authors": "Peter Shaw; Ming-Wei Chang; Panupong Pasupat; Kristina Toutanova",
"journal": "",
"ref_id": "b27",
"title": "Compositional generalization and natural language variation: Can a semantic parsing approach handle both?",
"year": "2020"
},
{
"authors": "Peng Shi; Rui Zhang; He Bai; Jimmy Lin",
"journal": "",
"ref_id": "b28",
"title": "Xricl: Cross-lingual retrieval-augmented in-context learning for cross-lingual text-to-sql semantic parsing",
"year": "2022"
},
{
"authors": "Bailin Wang; Richard Shin; Xiaodong Liu; Oleksandr Polozov; Matthew Richardson",
"journal": "",
"ref_id": "b29",
"title": "Rat-sql: Relation-aware schema encoding and linking for textto-sql parsers",
"year": "2019"
},
{
"authors": "Tao Yu; Chien-Sheng Wu; Xi Victoria Lin; Bailin Wang; Yi Chern Tan; Xinyi Yang; Richard Dragomir R Radev; Caiming Socher; Xiong",
"journal": "",
"ref_id": "b30",
"title": "Grappa: Grammar-augmented pre-training for table semantic parsing",
"year": "2021"
},
{
"authors": "Tao Yu; Rui Zhang; Kai Yang; Michihiro Yasunaga; Dongxu Wang; Zifan Li; James Ma; Irene Li; Qingning Yao; Shanelle Roman",
"journal": "",
"ref_id": "b31",
"title": "Spider: A large-scale human-labeled dataset for complex and cross-domain semantic parsing and text-to-sql task",
"year": "2018"
},
{
"authors": "John M Zelle; Raymond J Mooney",
"journal": "",
"ref_id": "b32",
"title": "Learning to parse database queries using inductive logic programming",
"year": "1996"
}
] | [] |
|
10.18653/v1/W18-5513 | 2016-10-10 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b51",
"b24",
"b46",
"b1"
],
"table_ref": [],
"text": "To combat the rise of misinformation, the NLP community has studied automatic fact-checking tools. However, there are limitations to the existing research that prevent it from being widely adopted at real fact-checking organizations such as PolitiFact. Many studies have focused on crowdauthored claims (Thorne et al., 2018;Jiang et al., 2020;Schuster et al., 2021;Aly et al., 2021), which do not accurately represent the complexities of actual claims that fact-checkers deal with. Other"
},
{
"figure_ref": [],
"heading": "Poli2Fact found that the study methodology and conclusions, were overly simplis2c…",
"publication_ref": [
"b15",
"b0",
"b22",
"b2",
"b3",
"b6"
],
"table_ref": [],
"text": "Evidence Retrieval A research paper found plas2c bags bans had significant nega2ve repercussions on public health…\n… …\nFigure 1: Our fact-check setting addresses realistic claims using evidence retrieved prior to when the claim was made.\nwork that does tackle real-world claims either relies on access to a document set which contains the \"gold\" evidence (Ferreira and Vlachos, 2016;Alhindi et al., 2018;Hanselowski et al., 2019;Atanasova et al., 2020) or conducts unconstrained retrieval (Augenstein et al., 2019), which may retrieve articles written by fact-checkers about the claim as shown by the right side of Figure 1. Prior work has not implemented a system to retrieve evidence in the wild.\nWe study fact-checking on complex political claims under the retrieval setting that aligns with what fact checkers do. We retrieve evidence from the web, restricted to documents authored before the time of the claim and not documents sourced from fact-checking websites as shown by the left side of Figure 1. To handle this challenging setting, we propose a pipeline that builds upon the strength of large language models (Brown et al., 2020) and findings from prior work. Following the approach of Chen et al. (2022a), we first decompose a claim into a series of subquestions, targeting both explicit and implicit aspects of the claim. Each subquestion is fed into a commercial search engine to retrieve relevant documents, with the restrictions described above. Then, we conduct a second stage of finegrained retrieval to isolate the most relevant por- "
},
{
"figure_ref": [
"fig_1"
],
"heading": "≈ ≈",
"publication_ref": [
"b6",
"b37",
"b31"
],
"table_ref": [],
"text": "Figure 2: Overview of our pipeline: a claim is first decomposed into several yes/no subquestions (Section 3.1), then we pipe the questions through two stages of retrieval (Section 3.2 and section 3.3) to select the most relevant paragraphs. Finally, we generate a claim-focused summary (Section 3.4) and train a veracity classifier to get the veracity label (Section 3.5). Our pipeline progressively filters text to leave only relevant content to validate a claim (see Appendix C for more details and Figure 5 for an example).\ntions of the documents. Finally, we use state-of-theart language models (Brown et al., 2020;Ouyang et al., 2022) to generate claim-focused summaries from the retrieved content. These summaries can serve both as explanations for users as well as inputs to a classifier to determine the veracity based on these summaries.\nEvaluating individual components of our pipeline is challenging due to the absence of gold annotations at each stage. We use automatic evaluation on the veracity classification performance, comparing to labels given by professional factcheckers. We supplement this with a human study evaluating the claim-focused summaries for comprehensiveness and faithfulness. This evaluation counterbalances the subjectivity of the veracity judgments (Lim, 2018) while shedding light on intermediate stages of the process.\nWe apply our pipeline to CLAIMDECOMP (Chen et al., 2022a), a dataset containing 1,200 real-world complex political claims with veracity labels. Performance on veracity classification shows that: (1) our retrieval setting is indeed much harder than \"unrestricted\" retrieval settings; (2) using web evidence leads to performance gains compared to automatic fact-checking without evidence; (3) the subquestions are crucial for obtaining high-quality raw documents from the web compared to using the original claim alone. Our human study further indicates that: (4) claim-focused summaries are mostly faithful and helpful for both machines and humans to fact-check a claim; (5) the retrieved evidence is often relevant to some aspects of the claim, but can rarely cover all aspects, suggesting that finding sufficient raw evidence in the wild is the core challenge in building automatic fact-checking systems."
},
{
"figure_ref": [],
"heading": "Background",
"publication_ref": [
"b51",
"b24",
"b46",
"b52",
"b54",
"b44",
"b53",
"b39",
"b12",
"b52",
"b0",
"b22",
"b2",
"b40",
"b41",
"b3",
"b14"
],
"table_ref": [],
"text": "Many of the widely used fact verification benchmarks, such as FEVER (Thorne et al., 2018), HoVer (Jiang et al., 2020) and VITAMINC (Schuster et al., 2021), focus on crowd-authored claims derived from Wikipedia. For example, claims in the FEVER dataset typically require checking a single aspect like \"Oliver Reed was a film actor.\" These claims are checkable by retrieving evidence from Wikipedia and can be annotated at scale, but they do not reflect the complexities of real-world political claims.\nEarlier studies (Vlachos and Riedel, 2014;Wang, 2017;Rashkin et al., 2017;Volkova et al., 2017;Pérez-Rosas et al., 2018;Dungs et al., 2018) on fact-checking political claims typically considered using the claim alone as an input to an automated system. By not seeking evidence, systems judge the veracity of a claim mostly based on surfacelevel linguistic patterns rather than based on factual errors. Research that incorporates evidence either assumes access to justifications provided by fact-checkers (Vlachos and Riedel, 2014;Alhindi et al., 2018;Hanselowski et al., 2019;Atanasova et al., 2020) or evidence from unconstrained retrieval (Popat et al., 2017(Popat et al., , 2018;;Augenstein et al., 2019), which frequently yields evidence sets containing pages from fact-checking websites. Fan et al. (2020) explore generating questions to retrieve evidence from the web, but only evaluate their system with humans in the loop, who can aggressively filter irrelevant retrieval results.\nTo our knowledge, we present the first automatic fact-checking system with a realistic retrieval pipeline using evidence available at the time a claim was made. This presents a very challenging setting where many claims are not checkable. We therefore emphasize the evidence our system re-Claim: James Quintero stated on October 10, 2016 in a panel discussion at SXSW Eco: When San Francisco banned plas@c grocery bags, \"you saw the number of instances of people going to the ER with things like salmonella and other related illnesses\" spike.\ntext-davinci-003 Q1: Did San Francisco ban plas@c grocery bags in the past? Q2: Did the number of people going to the ER for salmonella and other related illnesses increase aLer the plas@c bag ban in San Francisco? Q10: Was the plas@c bag ban directly caused the increase of the number of people going to the ER for salmonella and other related illnesses in San Fransisco?"
},
{
"figure_ref": [],
"heading": "…",
"publication_ref": [
"b19",
"b36",
"b48",
"b4"
],
"table_ref": [],
"text": "Figure 3: A demonstration of our claim decomposition process. We decompose each claim into 10 unique questions. We only show three questions for simplicity.\nturns as a way of assisting human fact-checkers; we believe this realistic task setting and corresponding evaluation should be reused in future work.\nNeed for these tools Our work shifts the focus away from the evaluation on classification accuracy alone. Accuracy on truth labels assigned by factcheckers is a proxy metric we use to evaluate our systems. However, fact-checking experts argue that the task is too subjective and complex to be automated in the near term (Graves, 2018;Nakov et al., 2021). Part of this arises from the fact that information needed to check claims is not always available on the web (Singh et al., 2021). Our approach of returning information on a best-effort basis and providing detailed evidence to enable a human to assist in the judgment can help overcome issues with returning judgments from error-prone AI systems (Bansal et al., 2021)."
},
{
"figure_ref": [],
"heading": "Methodology",
"publication_ref": [],
"table_ref": [],
"text": "Our pipeline consists of five parts: claim decomposition, raw document retrieval, fine-grained retrieval, claim-focused summarization, and veracity classification. We describe each part below."
},
{
"figure_ref": [],
"heading": "Subquestion Decomposition",
"publication_ref": [
"b16",
"b32",
"b25"
],
"table_ref": [],
"text": "Given a real-world complex claim, we first decompose it into a set of yes/no questions for which the answers are useful to fact-check the claim. Chen et al. (2022a) show that such decompositions are both helpful to retrieve relevant evidence and make veracity judgment, a finding shared by concurrent work on fact-checking for text generation outputs (Gao et al., 2022;Chen et al., 2022b;Liu et al., 2022) and for Wikipedia (Kamoi et al., 2023).\nFor decomposition, we prompt a large-scale language model, OpenAI's text-davinci-003, with in-context examples. 2 We carefully choose four input-decomposition pairs from the human annotations of Chen et al. (2022a) to form a few-shot prompt. We generate a set of unique questions by multiple rounds of sampling until we gather 10 different questions. An example decomposition is shown in Figure 3. For the full prompt, see Appendix A.2."
},
{
"figure_ref": [
"fig_0"
],
"heading": "First-stage Retrieval",
"publication_ref": [],
"table_ref": [
"tab_1",
"tab_1"
],
"text": "For each question generated in the previous step, we feed it to a commercial search engine API to collect the relevant documents.\nTemporal and Site Constraints We conduct this work under the assumption that a system should not be able to access pages published after the claim was made. This condition is appropriate if a system is going to be run instantaneously after a claim is made, e.g., for real-time fact-checking during a political speech. We place a temporal constraint on the system to reflect this. In addition, we investigate the extent to which the temporal constraint changes the results, by comparing two rounds of web retrieval, with and without the timestamp of a claim. Table 1 reports this comparison. We find little overlap between two document sets by comparing the Jaccard distance between two sets of the retrieved URLs.\nNext, to investigate how the presence of factchecking websites affects the veracity judgment of a claim, we also place a site constraint to filter out the documents from fact-checking websites. Our list of fact-checking websites can be found in Appendix A.1. An example of the retrieved documents is shown in Figure 4.\nWe use the Bing Search API,3 and retrieve 10 documents per subquestion after filtering by the constraints. We extract the actual content from the page URLs retrieved by the Bing Search API using two tools: html2text4 and readability-lxml.5 Approximately one-third of the URLs are protected6 and cannot be scraped as shown in Table 1.\nOne challenge for the reproducibility of our work is that commercial search engines are complex, dynamic systems. How consistent is the search query when we query at different timestamps? We conduct a small experiment to answer this question. Overall, we find that the search results change: only 30% overlap when queried two months apart. However, the veracity judgment classification result is not impacted much. Details are in Appendix B."
},
{
"figure_ref": [],
"heading": "Second-stage Retrieval",
"publication_ref": [],
"table_ref": [],
"text": "Most of the documents collected from the previous step contain only small snippets relevant to the claim, if relevant at all. Thus, we conduct a second-stage retrieval to pick the most relevant text spans regarding the claim. Specifically, we segment the documents into text spans containing k 1 words with a stride of 1 2 k 1 words. Following Chen et al. (2022a), we employ BM-25 to retrieve the top-K 1 highest-scored text spans, expanding these spans with a ± k 2 -word context. If two text spans overlap, they are merged to form a larger span. This process yields a set of documents ranked by the highest-scored text spans and we pick the top-K 2 documents."
},
{
"figure_ref": [],
"heading": "Claim-Focused Summarization",
"publication_ref": [
"b18",
"b56"
],
"table_ref": [],
"text": "Since the documents retrieved in the previous step can contain up to several thousand words, it becomes cumbersome for both humans and models to make a judgment based on such extensive content. Consequently, we prompt state-of-the-art LMs, specifically text-davinci-003, to summarize each retrieved document separately with respect to the claim.7 Such single-document summarization has been shown to work robustly when the documents in question are news articles (Goyal et al., 2022;Zhang et al., 2023)."
},
{
"figure_ref": [],
"heading": "Producing summaries instead of judgments",
"publication_ref": [],
"table_ref": [],
"text": "We investigate two types of prompts. For a zeroshot prompt, we instruct the model not to make any judgments about the stance of the given document. For a few-shot prompt, we select four documents and carefully write desired summaries. For documents that are not relevant to the claim, we write \"the document is not relevant to checking the claim\" as its desired output. We conduct human evaluation of the summary quality of different prompts in Section 6.1, where we find few-shot prompting works better. See Appendix A.3 for the full prompts."
},
{
"figure_ref": [],
"heading": "Veracity Classification",
"publication_ref": [
"b23"
],
"table_ref": [],
"text": "The final stage of our pipeline involves making a judgment based on the summaries generated in the previous stage. Unlike previous stages which use off-the-shelf tools, here we train a DeBERTalarge (He et al., 2020) model to perform a six-way veracity classification (true, mostly true, half true, barely true, false, and pants-on-fire) on the training set of the CLAIMDECOMP dataset.\nTraining The input to the classifier is a concatenation of the claim and the summaries of the retrieved documents, while the output is one of the six labels. We use a classification head on the CLS token and train it with cross-entropy loss.\nClaim: Melissa Agard stated on September 2, 2021 in News release: \"No other country on the planet witnesses the number of gun deaths that we do here in the United States, and it's not even close.\" We can train our model using the labels from CLAIMDECOMP; however, the inputs to the model are claim-focused summaries which have to be derived from our pipeline. We run our pipeline over the training, development, and test data of CLAIMDECOMP and train on pairs of the form (claim+summary, label). Since the dataset is small, we train the classifier five times with different random seeds and report the test set performance using the model that achieves the best performance on the development set."
},
{
"figure_ref": [
"fig_1"
],
"heading": "Final Pipeline",
"publication_ref": [],
"table_ref": [],
"text": "Our complete pipeline's results when executed on an example is shown in Figure 5. We note that the question decomposition phase yields an overcomplete set of questions, including redundant ones. However, the final retrieved and summarized documents are able to shed light on the claim from several complementary perspectives. While the fi-nal veracity judgment does not match the annotated judgment from PolitiFact, a user reading the documents comes away with an informed picture of the situation."
},
{
"figure_ref": [],
"heading": "Experimental Setup",
"publication_ref": [],
"table_ref": [],
"text": "Our main automatic evaluation is on claim veracity prediction (Wang, 2017), evaluating our entire pipeline. We will describe the human evaluation setup in Section 6.\nData We use the data from CLAIMDE-COMP (Chen et al., 2022a) which contains 1,200 complex claims from PolitiFact (train: 800,dev: 200,test: 200). Each claim is labeled with one of the six veracity labels, a justification paragraph written by expert fact-checkers, and subquestions annotated by prior work.\nHyperparameters For the second-stage retrieval, we set top-K 1 = 10 (highest-scored text Evaluation Metric Following prior work, we report accuracy (Acc), mean absolute error (MAE), and Macro-F1. In addition, we introduce soft accuracy (soft Acc), which is calculated by counting offby-one errors on the six-veracity scale (e.g., half true instead of mostly true) as correct, accounting for the fact that veracity judgments are inherently ambiguous and subjective."
},
{
"figure_ref": [],
"heading": "Comparison Systems",
"publication_ref": [],
"table_ref": [],
"text": "Claim-only We concatenate the metadata, including the speaker and the venue of the claim, with the claim itself, and feed the resulting text into the classifier. This approach serves as a lower bound for the veracity classification. This follows the setting used by Wang (2017).\nClaim+Just We extend the Claim-only baseline by appending the human-written justification paragraph, excluding the sentence containing the label, to the claim. Note this is the oracle setting and sets the upper bound for the veracity classification.\n5 Automatic Evaluation: Claim Veracity"
},
{
"figure_ref": [],
"heading": "Constrained vs. Unconstrained Search",
"publication_ref": [
"b3"
],
"table_ref": [
"tab_5"
],
"text": "We first situate our work with respect to baselines and past systems by varying the retrieval condition. Specifically, we experiment with a temporal constraint, where pages have to originate before the date of the claim, and a site constraint, where sites have to be non-fact-checking (non-FC) sites. Even in the unconstrained setting, we exclude pages from PolitiFact, where our dataset is scraped from, to prevent label leakage.\nThe unconstrained setting corresponds to that used in MultiFC (Augenstein et al., 2019). Mul-tiFC includes numerous documents that are filtered out by our constrained settings. For each claim, they extract the top 10 pages from the Google search API. We find that 12,721 out of 15,379 claims (82.7%) contain at least one page from our excluded website list and 24.4% of the retrieved web pages are from fact-checking websites.\nTable 2 reports the performance of our system with various retrieval constraints. Comparing the performance of claim-only and other models that use retrieval, we see a statistically significant8 improvement over all four of our metrics in nearly all settings, showing that retrieving and summarizing evidence is helpful to predict the veracity label, even with constraints.\nSecond, we see adding either temporal or site constraints dramatically reduces the performance. This implies that retrieval over the web works largely because it retrieves fact-checks that were published after the claim was released, with synthesized evidence. We believe that future work on retrieval should use a constrained setting."
},
{
"figure_ref": [],
"heading": "Stage Ablations",
"publication_ref": [],
"table_ref": [],
"text": "We evaluate design choices in each stage of the pipeline to understand how each individual component contributes to the final performance. The results are shown in Table 3.\nFirst-stage Retrieval: subquestions vs. original claim Using the original claim instead of the generated subquestions as an input to web search ( B vs. 1 ) results in a notable decrease in performance. This can be attributed to the fact that the subquestion set encompass multiple aspects of the claim, enabling the search engine to locate relevant infor- mation more easily. Comparing B and 2 , we see using the gold subquestions actually yield worse performance than our predicted subquestions. The reason for this could be that we predict 10 subquestions, potentially garnering more relevant data than the gold subquestions, of which there are 3 on average (Chen et al., 2022a).\nSecond-stage Retrieval Rather than retrieving with subquestions (subQs), we instead perform our search with the raw Claim ( 3 ), Gold subQs from Chen et al. (2022a) ( 4 ), or Justification ( 5 ), which uses oracle information. Different queries yield only slight differences in performance and none of them is statistically significant, even when 5 uses the human-written justification. We believe this is because we expand the retrieved text span by a ±150 words of context window. As a result, this retrieval step does not need to be very precise to capture the relevant information.\nClaim-focused Summarization We compare zero-shot ( B ) and few-shot ( 6 ) prompts for generating the summary; no summary ( 7 ) directly feeds the text spans from second-stage retrieval to the veracity classifier. System 7 shows the worst performance across all metrics, suggesting that summarization matters. This may result from two primary factors: (1) The document length exceeds the context window capacity of DeBERTa, causing crucial information to be truncated. (2) our veracity classifier cannot easily discern the most relevant information given a large amount of context. Differences in the prompt ( B and 6 ) do not impact veracity classification results too much but have differences under human inspection, which we discuss in the next section."
},
{
"figure_ref": [],
"heading": "Human Evaluation: Claim-focused Summaries",
"publication_ref": [
"b5",
"b11",
"b37",
"b31"
],
"table_ref": [],
"text": "Summarizing documents from web search with large language models improves the performance of our fact-checking pipeline. However, these models can generate untruthful content (Bommasani et al., 2021;Chowdhery et al., 2022;Ouyang et al., 2022). Furthermore, as pointed out by Lim (2018), the accuracy of veracity classification alone does not entirely reflect the system's overall effectiveness, as certain labels such as \"false\" and \"barelytrue\" may be ambiguous. We believe the true measure of our system's utility lies in the full package of summarized evidence it returns rather than just the accuracy of the veracity label. Therefore, we carry out two human studies, on comprehensiveness and faithfulness, to better understand intermediate outputs of the system.\nSetting We randomly pick 50 claims which contain 200 document-summary pairs from the development set of CLAIMDECOMP and run two human evaluation studies on this set. For each task, we recruited annotators from Amazon Mechanical Turk with a qualification test. In total, we recruited 17 worker for the faithfulness study and 15 workers for the comprehensiveness study. The details about the recruiting process and the annotation interface can be found in Appendix D.\nComparison Systems We compare the summaries generated from two prompts, zero-shot-003 and few-shot-003, on GPT-3.5 (davinci-003). For the faithfulness study, we also compare the summaries generated through with zero-shot prompt on an earlier GPT model (davinci-001) (zero-shot-001) to see how the faithfulness varies for different models."
},
{
"figure_ref": [],
"heading": "Faithfulness Evaluation",
"publication_ref": [],
"table_ref": [],
"text": "Goal We assess the frequency and degree to which the language model generates untruthful content during query focused summarization. For each document and summary pair, annotators choose one of four labels below:\n• Faithful: the summary accurately represents the meaning and details of the original document. • Minor Factual Error: some details are not aligned with the original document, but the overall message remains intact. • Major Factual Error: there are factual errors that result in the summary misrepresenting the original document. • Completely Wrong: the language model hallucinates content that completely alters the meaning of the original document. In addition to selecting a label, we ask annotators to provide a natural language justification for their choices. The annotations agree with a Fleiss Kappa score of 0.30. While this number is somewhat low, when we evaluated their justifications and we find many of the disagreements are because of subjectivity on the extent of factual error. We compute a consensus annotation via majority vote. We assign numerical scores to each label, where \"Faithful\", \"Minor\", \"Major\", and \"Completely Wrong\" correspond to 4, 3, 2, and 1 respectively and report average values. If annotators disagree, we compute the average score and return the label that is nearest to the average score as a consensus."
},
{
"figure_ref": [
"fig_2"
],
"heading": "Results",
"publication_ref": [],
"table_ref": [],
"text": "The results are shown in Table 4. We see that few-shot prompting substantially decreases the chance of hallucinations in the summaries. When combining \"Factual\" and \"Minor\", we see 89% of the summaries are good enough to be used as evidence for the classifier. Comparing the performance of zero-shot-001 and zero-shot-003, Table 4: Faithfulness Human Evaluation (N=200). \"F\" denotes the summary is factual and \"NF\" denotes the summary is completely wrong. Few-shot prompting helps the model make fewer factual errors.\nwe find that the weaker model makes more major factual errors. Together, they indicate that with stronger models and better prompts, we may expect these summarization models to improve further.\nFigure 6 shows three examples containing unfaithful content. We see that the \"Minor\" error does not affect the interpretation of the original document while \"Major\" and \"Completely Wrong\" errors alter the view."
},
{
"figure_ref": [],
"heading": "Comprehensiveness Evaluation",
"publication_ref": [],
"table_ref": [],
"text": "Goal We aim to measure the extent to which the claim-focused summaries are able to address the claim. This is subjective and difficult task to evaluate. Here, we leverage the human-annotated yes/no subquestions presented in Chen et al. (2022a) as a proxy for evaluating the comprehensiveness of our summaries: if provided summary can help humans to answer more of these yes/no questions, we deem the summary to be more comprehensive.\nIn this task, annotators are given a summary / subquestion pair and label subquestion as \"answerable\", \"partially answerable\",9 or \"unanswerable\", and additionally provide yes/no answer if the question is labeled as \"answerable\". Annotators were also asked to provide natural language justification for their answers. We collect this annotation on 161 questions associated with 50 claims. The annotations agree with a Fleiss Kappa score of 0.32."
},
{
"figure_ref": [],
"heading": "Results",
"publication_ref": [],
"table_ref": [
"tab_10"
],
"text": "The results are presented in Table 5. We see that zero-shot summaries yield more answerable questions than few-shot summaries. However, based on the faithfulness results, we believe this may be because hallucinations in zero-shot summaries make the questions answerable; the system imputes information that seems to help, but which is not supported by the document.\nNevertheless, the few-shot summaries allow us to partially address over 60% of the gold anno-Claim: Facebook posts stated on September 29, 2021: The White House \"created a fake set for (President Joe) Biden to get his booster shot. The enFre Biden presidency is one giant charade.\" Document Title: President Joe Biden Gets COVID Booster Vaccine Shot -PEOPLE.com Content: President Joe Biden received his COVID-19 booster shot .. Biden received his first dose of Pfizer-BioNTech novel coronavirus vaccine in December at Delaware 's ChrisFanaCare Hospital. His second shot was administered there three weeks later. Both shots were televised … Summary: On Monday, President Joe Biden received his COVID-19 booster shot on camera, administered by a nurse with the White House Medical Unit. Biden has previously received both of his iniFal doses of the Pfizer-BioNTech novel coronavirus vaccine in December at Delaware's ChrisFanaCare Hospital. Minor Error: The summaries says both shots are received in December while the document says the first one was in December and the second one was three weeks later. The summary slightly changes the informaFon in the original document. Claim: Elizabeth Warren stated on November 20, 2019 in a debate: \"Today in America -a new study came out -20 years out, whites who borrowed money, 94% of them have paid off their student loan debt; 5% of African Americans have paid it off.\" Document Title: More Evidence Student Borrowers Prepay Their Loans -Forbes Content: Most borrowers will sFll fully pay off their loans before they are required to do so … Borrowers pay off almost all loans ( 94 % ) with a final payment larger than the scheduled monthly payments they are required to make … Summary: A new report from the Consumer Financial ProtecFon Bureau found that 94% of borrowers pay off their student loans before the scheduled payoff date, typically doing so in five years or less ... Major Error: The document says borrowers pay off 94% of the loan not 94% of borrowers pay off their loans Claim: Andrew Giuliani stated on May 18, 2021 in a news conference: \"The one good thing about the anFbodies if you've had it, is it actually is even beger than the vaccine, and here's why. tated subquestions derived from the PolitiFact justification. We find this result encouraging: it indicates that even though the system does not have access to these (often subtle) factors, it can return relevant information to enable a human annotator to make a judgment about them."
},
{
"figure_ref": [],
"heading": "Holistic Evaluation",
"publication_ref": [],
"table_ref": [
"tab_11"
],
"text": "To aid fact-checkers, automatic system should present faithful and comprehensive information.\nWhile in previous section we evaluated these two factors separately, here we investigate the claimlevel statistics in this section. We aim to answer in a holistic fashion: how many claims can be comprehensively addressed with a set of fully faithful documents?\nWe label a claim as answerable if all of its subquestions are answerable. If all subquestions are at least partially answerable, the claim is labeled as partially answerable. When only some subquestions are partially answerable, the claim is categorized as partially unanswerable. If all subquestions are unanswerable the claim is unanswerable. For claim-level faithfulness, a claim is faithful is all summaries are faithful, otherwise it is either unfaithful or contains minor factual errors. We apply the same principle to compute claim-level faithfulness. Table 6 shows the results by combining the two factors. We see that addressing every aspect of complex claims is still challenging: 36 out of 50 claims contain at least one unanswerable question. For claims that can be fully addressed (all questions are either answerable or partially answerable), we see only 1 out of 14 contains a major factual error in the retrieved documents."
},
{
"figure_ref": [],
"heading": "Discussion and Future Direction",
"publication_ref": [
"b10"
],
"table_ref": [],
"text": "Performance is bottlenecked by the first-stage retrieval. The results in the last section show that 36.0% of questions are unanswerable using our most faithful claim-focused summaries. By investigating the unanswerable cases, we see that the following cases lead to retrieval failure: (1) no relevant information is available on the web except the fact-checking websites. These claims can be onerous to check, such as requiring talking to or emailing specific people to check facts. Those cases are beyond the scope of this work and we think a system doing triage for the claims, would be promising for future work. (2) No relevant subquestions are generated or the subquestions are not well decontextualized (Choi et al., 2021). In such cases, a stronger question generation model or decontextualization model can further help.\nThe need of human-in-the-loop fact-checking.\nTo address the failures in the first-stage retrieval and the potential errors in the summarization stage, we envision a human-in-the-loop fact-checking system. This system begins with the automated pipeline presented in this paper, which provides fact-checkers with summarized documents and judgments. If the fact-checkers deem these documents unsatisfactory, the system reveals the subquestions used for evidence retrieval, allowing factcheckers to rerun the search. The system then retrieves additional documents and generates updated summaries. This iterative process continues until the fact-checkers are satisfied with the retrieved evidence. Moreover, the system could further learn from the fact-check feedback to improve itself: for example, the system could learn what questions are important to retrieve good evidence and what questions are not according to the fact-checker."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b7",
"b29",
"b26",
"b28",
"b35",
"b30",
"b47",
"b43",
"b20",
"b27",
"b57",
"b34",
"b13",
"b50",
"b35",
"b49",
"b33",
"b42",
"b38",
"b55",
"b17",
"b45",
"b16",
"b32",
"b25"
],
"table_ref": [],
"text": "Retrieval augmented models Prior work has shown that a variety of NLP tasks could benefit from incorporating a retrieval component. Such tasks mainly include question answering (Chen et al., 2017;Kwiatkowski et al., 2019;Karpukhin et al., 2020;Khattab et al., 2021;Nakano et al., 2021), text generation (Lewis et al., 2020;Shi et al., 2023;Ram et al., 2023), language modeling (Guu et al., 2020;Khandelwal et al., 2020;Zhong et al., 2022), and dialog (Moghe et al., 2018;Fan et al., 2021;Thoppilan et al., 2022).\nMost of these work assume having access to a fixed corpus, however, for the task of real-world fact-checking, no such corpus exists. In this work, we follow WebGPT (Nakano et al., 2021) and use Bing Search API to retrieve evidence from the wild web. Recent LLM agents such as Bing Chat and Google Bard follow this paradigm, so we believe these directions will be relevant for future work.\nQuestion decomposition has been shown to be effective in evidence retrieval and question understanding for complex question answering that need multiple steps of explicit/implicit reasoning (Talmor and Berant, 2018;Min et al., 2019;Qi et al., 2019;Perez et al., 2020;Wolfson et al., 2020;Geva et al., 2021). Question generation has also been shown to play a useful role in retrieval pipelines in open-domain QA (Sachan et al., 2022). In more recent research, it was demonstrated by Chen et al. (2022a) that such decompositions can also aid in retrieving evidence to assess complex claims and make veracity judgment. This observation is consistent with concurrent studies on fact-checking text generation outputs (Gao et al., 2022;Chen et al., 2022b;Liu et al., 2022) and Wikipedia (Kamoi et al., 2023)."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "In this paper, we introduce a pipeline for realistic, automated fact-checking of complex political claims by retrieving raw evidence from web documents. Our pipeline demonstrated promising results on the CLAIMDECOMP dataset. Yet, web search often cannot surface all pieces of information necessary to verify given claim. This work emphasizes the difficulties of evidence retrieval in real-world scenarios and underscores the need for a human-in-the-loop fact-checking system. does not significantly impact the models' efficacy in the veracity assessment. We caution that as the time gap increases, the set of documents retrieved from the Bing Search API could become considerably different, posing a challenge to consistently benchmark retrieval performance using commercial search engines. Therefore, we advocate for future research to focus on developing a comprehensive yet challenging document set that could be publicly released as a benchmark to spur research."
},
{
"figure_ref": [],
"heading": "C Information Compression through the Pipeline",
"publication_ref": [],
"table_ref": [
"tab_13"
],
"text": "Our pipeline progressively refines the crucial data needed to validate a claim. Table 8 demonstrates the average count of unique documents and the total word count in these documents after each phase of our pipeline under both temporal and site constraints."
},
{
"figure_ref": [],
"heading": "D Human Study D.1 Recruiting Process",
"publication_ref": [],
"table_ref": [],
"text": "Faithfulness study We set up a qualification test that consists of 5 examples. We selected workers from MTurk if they get more than 3/5 examples correct according to our curated labels and if they write reasonable rationales. In total, there are 31 workers who took the qualification test and we selected 15 of them for the task. We pay $3 for the qualification test and $2 dollars for one HIT that contains 4 document-summary pairs in the actual task. The detailed instructions and the annotation interface is shown in Figure 10.\nComprehensiveness study We set up a qualification test that consists of 10 examples. We selected workers from MTurk if they got more than 7/10 questions right according to our curated labels and if they write reasonable rationales. In total, there are 28 workers who took the qualification test and we selected 17 of them for the task. We pay $3 for the qualification test and $0.3 dollars for one question in the actual task.\nThe detailed instructions and the annotation interface is shown in Figure 11 "
},
{
"figure_ref": [],
"heading": "Acknowledgments",
"publication_ref": [],
"table_ref": [],
"text": "This work was partially supported by NSF CA-REER Award IIS-2145280, by Good Systems, 10 a UT Austin Grand Challenge to develop responsible AI technologies, and by grants from Salesforce Inc. and Open Philanthropy. We thank the UT Austin NLP community for feedback on the earlier drafts of the paper."
},
{
"figure_ref": [],
"heading": "",
"publication_ref": [],
"table_ref": [],
"text": "https://github.com/jifan-chen/"
},
{
"figure_ref": [],
"heading": "",
"publication_ref": [],
"table_ref": [],
"text": "We also filter the URLs that contain \"fact-check\" or \"factcheck\"; we also filter any PDF files and videos."
},
{
"figure_ref": [],
"heading": "A.2 Question Generation Prompt and Deduplication",
"publication_ref": [],
"table_ref": [],
"text": "The prompt we used to generate the questions is shown in Figure 7. Since the generated question set sometimes contains duplicates, we delete the duplicated questions according to the exact string match."
},
{
"figure_ref": [],
"heading": "A.3 Question-focused Summarization Prompt",
"publication_ref": [],
"table_ref": [],
"text": "The zero-shot and few-shot prompts we used to generate the claim-focused summaries are shown in Figure 8 and Figure 9 respectively."
},
{
"figure_ref": [],
"heading": "A.4 Hyperparameters of Veracity Classifier",
"publication_ref": [],
"table_ref": [],
"text": "• Model: DeBERTa-large Claim: Barry DuVal stated on September 25, 2015 in an interview: We're the only major oil-producing naMon in the world with a self-imposed ban on exporMng our crude oil to other naMons.\nSuppose you are a fact-checker, generate several yes or no quesMons to help me answer if this claim is true or false.\nQuesMons: Is the U.S. the only major oil-producing naMon to ban exports of crude oil? Is the self-imposed ban on crude oil export of U.S a complete ban?\nClaim: William Barr stated on September 2, 2020 in a CNN interview: We indicted someone in Texas, 1,700 ballots collected from people who could vote, he made them out and voted for the person he wanted to. • Epochs: 25\n• Initial learning rate: 3e-5\n• Optimizer: Adam with linear decay Document: Vote by mail: Which states allow absentee vo6ng -Washington Post Content: excuse to vote absentee and states that will allow fear of the coronavirus as an excuse . In response to the coronavirus , nearly half of all states expanded access to mail ballots for their primaries , either by allowing fear of the coronavirus as a reason or proac6vely sending an applica6on or ballot to every registered voter . Fewer have taken ac6on for the general elec6on , as the move has become increasingly par6san and subject to li6ga6on . President Trump has made numerous unfounded claims that mail-in vo6ng will create widespread abuse and fraud . His suspicions are out of step with the views of elec6on experts and many within his own party , who are building large-scale vote-by-mail programs . A recent analysis by The Content: hard and one thing is clear : we have a great First Lady . I 'm convinced that this whole Governor thing is just a way for the people of Florida to have Casey as their First Lady . Our kids Madison and Mason have taken over the Governor 's mansion but the baby-proofing has been successful -I can report that no ar6facts of Florida history have yet to be destroyed . But stay tuned -Mason will be walking very soon . Just last week , Casey traveled to NW Florida to survey the Hurricane Michael recovery efforts . We have both been to the region several 6mes over the past few months . To the people of NW Florida : we pledged to stand with you as you work to rebuild your communi6es stronger than before -and we have followed through on that pledge . I 've already traveled to Washington , DC and secured a historic commitment from the Trump administra6on to provide assistance to the communi6es that Michael ba^ered . My administra6on has worked to cut through red tape to expedite relief efforts and , thanks to the leadership of your former colleague Jared Moskowitz , we are making great progress . Here with us today is Mayor Al Cathey and Administrator Tanya Castro from Mexico Beach , which was ground zero for Hurricane Michael 's wrath . They and many others affected by the storm have displayed remarkable resilience in the face of huge obstacles . They deserve our admira6on -and our support . These communi6es will not be rebuilt in days , weeks or months , but they will be rebuilt . They will be rebuilt because we as Floridians will answer the call as we always do . God bless the good people of NW Florida . Execu6ve energy and leadership are necessary to meet fully the challenges that are before us -but they are not sufficient . In a Suppose you are assis6ng a fact-checker to fact-check the claim: \"Ron DeSan6s stated on March 5, 2019 in his State of the State address: \"I've already traveled to Washington, D.C., and secured a historic commitment from the Trump administra6on to provide assistance to the communi6es that Michael ba^ered.\" Summarize the relevant informa6on from the document in 1-2 sentences. Your response should provide a clear and concise summary of the relevant informa6on contained in the document. Do not include a judgment about the claim and do not repeat any informa6on from the claim that is not supported by the document: Ron DeSan6s did state in his address that he traveled to Washington, D.C. and secured a commitment from the Trump administra6on. However, no addi6onal evidence is provided about this claim.\nDocument: Cuban government unveils controversial cybersecurity regula6on Content: others . Ac6vists , opponents , unofficial organiza6ons and Internet users fear that the new regula6ons will be applied at the discre6on of the government and will serve to silence dissident voices on the networks and prosecute those who cri6cize the Cuban system and its leaders . In this sense , complaints and cri6cisms of the island ' s authori6es due to the new regula6on have already started being published on the social networks . In addi6on , in this Tuesday ' s Gaceta Oficial other regula6ons were published that , according to the Ministry of Communica6ons , \" update \" and \" strengthen \" Cuba ' s legal body on these issues . Among these is Decree-Law No . 35 on \" Telecommunica6ons , Informa6on and Communica6on Technologies and the Use of the Radioelectric Spectrum , \" which establishes that \" the Cuban state is the owner of public telecommunica6ons services and has the power to regulate , order , control and supervise the different telecommunica6ons/ICT services and networks ; as well as gran6ng the right to its exploita6on , \" Cubadebate pointed out . Finally , the package includes a group of regula6ons on Informa6on and Communica6on Technologies , the Radioelectric Spectrum , the use of Radiocommunica6on There is risk associated of bringing anybody in from the outside , but specifically from a conflict zone like that , \" he added . This hearing comes aver the Obama administra6on ' s announcement to allow 85,000 refugees into the U.S. next year and 100,000 in 2017 . Those are huge increases from the current level of 70,000 . The Obama administra6on has said that at least 10,000 of the refugees next year will come from Syria . Many lawmakers are concerned that ISIS extremists or other terrorists could make it into the U.S. as a refugee due to the lack of intelligence and informa6on available about the area . \" The intelligence that we have of this par6cular conflict zone is not as rich as we would like it to be , \" admi^ed Nicholas Rasmussen , the head of the Na6onal Counterterrorism Center during the hearing . Even Homeland Security Secretary Jeh Johnson expressed concern saying , \" We should do the right thing by accep6ng more , but we should be careful in doing it. \" Currently the U.S. has taken in around 2,000 Syrian refugees and has contributed about $ 4 billion in foreign aid to the Syrian refugee crisis , more than any other country . Read more at The Hill . refugees Suppose you are assis6ng a fact-checker to fact-check the claim: \"Rob Portman stated on October 20, 2016 in an Ohio Senate debate in Cleveland: \"The director of the FBI said that we cannot figure out who (Syrian refugees) are, what their inten6ons are, because we have no informa6on on them, because we don't have any contact with the Syrian government or any people on the ground to be able to determine that.\"\" Summarize the relevant informa6on from the document in 1-2 sentences. Your response should provide a clear and concise summary of the relevant informa6on contained in the document. Do not include a judgment about the claim and do not repeat any informa6on from the claim that is not supported by the document: FBI Director James Comey voiced concern during a Senate Homeland Security Commi^ee hearing about the lack of informa6on available to screen Syrian refugees coming into the United States.\nNicholas Rasmussen says we do not have as much informa6on about Syria as we would like and Homeland Security Secretary Jeh Johnson says we need to be careful.\nFigure 9: Few-shot prompt we used to generate the claim-focused summaries in this paper.\n• Metric for selecting best dev model: MAE\n• Random seed of 5 runs: 290032, 33432, 7876, 366, 77\n• Training device: NVIDIA-A6000"
},
{
"figure_ref": [],
"heading": "B Reproducibility of First-stage Retrieval",
"publication_ref": [],
"table_ref": [],
"text": "We conduct experiments to explore the stability and reproducibility of our first-stage retrieval step.\nWe conducted three rounds of retrieval at T = 0, T = 1 week, and T = 2 months. We evaluate the Jaccard similarity by comparing their URL between the sets of documents retrieved from our queries to understand how much changes in the Bing API and the broader web change our results.\nWe also evaluate the veracity of our system.\nResults are shown in Table 7. A noticeable trend is a decline in the Jaccard score between varying retrieval rounds over time. However, this decrease Thank you for participating in this task! This task aims to determine how trustable an AI system is at automatically gathering the most relevant information from a document to verify a political claim.\nYou are given 1) a political claim, 2) a snippet of a document that is potentially relevant to check the political claim, and 3) a summary of the snippet generated by an AI system We want to evaluate whether the summary is faithful to each document. For the summary to be faithful, the summary should avoid adding any new information that is not present in the original document or misrepresenting the information presented to given document. Note that your job is not to evaluate whether the document/summary is relevant or not to the claim. The claim is not meant to be used to judge whether the summary is faithful or not. It just provides you some context that may be helpful.\nIf the document is truncated, you can make your best guess as to the content. Do not penalize the summary if it includes content you think would reasonably occur in the rest of the document if not truncated.\nMajor factual errors should be errors that cause the summary to actually give a different impression than the original document. Minor factual errors are those where, even though some details may not align, they don't change the overall message of the document.\nIt's okay for the summary to cite the claim. However, if the summary contains an assessment regarding whether the document is relevant to the claim or not, try your best to evaluate whether the assessment made by the machine is accurate or not based on our criteria (correct, minor, major ...)."
},
{
"figure_ref": [],
"heading": "Examples:",
"publication_ref": [],
"table_ref": [],
"text": "Highlights are added by us for illustration but not present in real examples you will see."
},
{
"figure_ref": [],
"heading": "Example of faithful summary",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Claim",
"publication_ref": [],
"table_ref": [],
"text": "Ingraham said, \"You know what the biggest lie is, is that restaurants are spreaders of COVID. There's no science for that.\" In fact, plenty of evidence suggests restaurant dining has helped spread the coronavirus. Places that allow indoor dining and don't follow safety protocols are considered especially unsafe."
},
{
"figure_ref": [],
"heading": "Document",
"publication_ref": [],
"table_ref": [],
"text": "Document title: What are the main modes of transmission for COVID-19? -Live Science Content: least two people died from the virus , the Los Angeles Times reported . That suggested the viral particles were shed as aerosols by someone , before being inhaled or otherwise acquired by other choir members . A 2019 study in the journal Nature Scientific Reports ( opens in new tab ) found that people emit more aerosol particles when talking , and that louder speech volumes correlate to more aerosol particles being emitted . That case , along with those studies , suggest that the virus can be routinely transmitted via aerosols , though other routes of transmission ( such as large droplets being emitted during singing or speech ) are still possible explanations . In the 2003 SARS outbreak , aerosol transmission occurred during hospital procedures that generated large volumes of aerosols , such as intubation . There 's one other route that 's thought to play a role in the spread of COVID-19 : contact transmission . In that situation , viral particles emitted from the respiratory tract of an infected individual land on a surface . Then , another person touches that object , then touches their nose , mouth or eyes . The virus then sneaks into the body via the mucous membranes , infecting the second person . So far , no one knows how common this mode of transmission is , but it does seem to be possible . One study found that SARS-CoV-2 could remain viable on surfaces such as cardboard for up to 24 hours , and on plastic and steel for 2 to 3 days . Santarpia has studied viral surface contamination in the context of patients hospitalized with COVID-19 at the University of Nebraska Medical Center . In that study , which was published March 26 on the preprint database medRxiv , Santarpia and his colleagues found viral contamination in air samples , on surfaces such as toilets , and on frequently touched surfaces . Also on March 26"
},
{
"figure_ref": [],
"heading": "Summary",
"publication_ref": [],
"table_ref": [],
"text": "The document discusses the various modes of transmission of COVID-19, including aerosol transmission and contact transmission. Aerosol transmission occurs when people emit more aerosol particles when talking, and contact transmission happens when viral particles emitted from an infected person land on a surface and are then touched by someone else. Studies have also found that SARS-CoV-2 can remain on surfaces, such as cardboard and plastic, for up to 24 hours and 2 to 3 days respectively."
},
{
"figure_ref": [],
"heading": "Explanation",
"publication_ref": [],
"table_ref": [],
"text": "All of the facts presented in summary are supported by the document."
},
{
"figure_ref": [],
"heading": "Example of non-faithful (Major Factual Error) summary",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Claim",
"publication_ref": [],
"table_ref": [],
"text": "Joe Biden stated on August 31, 2020 in a speech: \"When I was vice president, violent crime fell 15% in this country. ... The murder rate now is up 26% across the nation this year under Donald Trump.\""
},
{
"figure_ref": [],
"heading": "Document",
"publication_ref": [],
"table_ref": [],
"text": "Document title: Murder Rate Up 26% in 2017 -Nationwide 90FM Content: More than 1,265 people have been murdered since the start of the year ; making the country ' s murder rate increase by 26 percent over the similar period last year . Nationwide News has obtained the latest crime statistics , from January up to October 15 this year , from the Jamaica Constabulary Force ' s Corporate Communications Unit , CCU . Superintendent Stephanie Lindsay is head of the CCU . [ audio_mp3 url= \" https : //nationwideradiojm.com/wp-content/uploads/2017/10/LINDSAY-murder-andshooting-OCT202017.mp3″ ] She says the St. James Police Division is leading the chart with 235 murders . It ' s followed by Clarendon with 131 , and St. Andrew South , 124 murders . Portland has recorded the fewest murders ; five . There ' s a reduction in all major crimes except murders and shootings . Rapes , aggravated assaults and robberies have all declined by at least 14 percent . In the meantime , Superintendent Lindsay says the police have seized 698 guns and over 7,000 rounds of ammunition up to October 15 . [ audio_mp3 url= \" https :"
},
{
"figure_ref": [],
"heading": "Summary",
"publication_ref": [],
"table_ref": [],
"text": "The Jamaica Constabulary Force's Corporate Communications Unit has reported that the country's murder rate has increased by 26 percent from January to October 15, 2020 compared to the same period in 2019. Additionally, other major crimes such as rapes, aggravated assaults and robberies have all declined by at least 14 percent, while the police have seized 698 guns and over 7,000 rounds of ammunition up to October 15, 2020."
},
{
"figure_ref": [],
"heading": "Explanation",
"publication_ref": [],
"table_ref": [],
"text": "The summary says the muder rate has increased by 26 percent in 2020, which is not factually correct according to the document that states the murder rate has increased by 26 percent in 2017. We would rate this as a Major factual error because it significantly changes the interpretation of the document."
},
{
"figure_ref": [],
"heading": "Your task",
"publication_ref": [],
"table_ref": [],
"text": "Claim 1\nSummary 1"
},
{
"figure_ref": [],
"heading": "${s1}",
"publication_ref": [],
"table_ref": [],
"text": "Please read the summary and the document carefully as some of the errors are subtle and hard to spot. The claim is not meant to be used to judge whether the summary is faithful or not.\nYour Explanation in 1-2 short sentences Claim 2"
},
{
"figure_ref": [],
"heading": "${claim2}",
"publication_ref": [],
"table_ref": [],
"text": "Document 2"
},
{
"figure_ref": [],
"heading": "${d2}",
"publication_ref": [],
"table_ref": [],
"text": "Summary 2"
},
{
"figure_ref": [],
"heading": "${s2}",
"publication_ref": [],
"table_ref": [],
"text": "Please read the summary and the document carefully as some of the errors are subtle and hard to spot. The claim is not meant to be used to judge whether the summary is faithful or not.\nExplanation in 1-2 short sentences Claim 3 Please type your explanation"
},
{
"figure_ref": [],
"heading": "Previewing Answers Submitted by Workers",
"publication_ref": [],
"table_ref": [],
"text": "This message is only visible to you and will not be shown to Workers. You can test completing the task below and click \"Submit\" in order to preview the data and format of the submitted results.\nFigure 10: Interface of the faithfulness study we conducted in Section 6.1."
},
{
"figure_ref": [],
"heading": "Instructions",
"publication_ref": [],
"table_ref": [],
"text": "Thank you for participating in this task! The goal of this task is to determine how good an AI system is at finding information to help check political claims. You are going to see whether some information the AI system produces contains the answers to questions that are important to fact-checkers."
},
{
"figure_ref": [],
"heading": "Task: Comprehensiveness",
"publication_ref": [],
"table_ref": [],
"text": "You are given a political claim, a set of AI system-generated sentences based on web searches, and a set of yes/no questions that are related to checking the claim.\nIn this task, you should determine whether the questions are answerable based on the AI-generated sentences. For each question, you should choose from the following three labels: 1. Answerable: The question is fully answered by the rationale. 2. Partially Answerable: Only part of the question could be addressed by the rationale or question is addressed but it's not clear whether there's evidence for it.\n3. Unanswerable: The question cannot be answered by the rationale. If you think the question is answerable from the rationale, you should also give your answer. If the answer is partially answerable, use your best guess.\nWe provide two examples below for you to better understand the task."
},
{
"figure_ref": [],
"heading": "Example 1",
"publication_ref": [],
"table_ref": [],
"text": "Claim: Donald Trump stated on February 5, 2018 in a speech near Cincinnati: At the State of the Union address, Democrats, \"even on positive news … were like death and un-American. Un-American. Somebody said, 'treasonous.' I mean, yeah, I guess, why not? Can we call that treason? Why not?\" AI-Generated Sentences: 1. In 1976, Gerald Ford (R) became the only president to ever declare the state of the union to be not good. Since 1981, every State of the Union address from George W. Bush (R) and Barack Obama (D) has declared that the state of the union is strong, to some extent. The White House has already announced that President Donald Trump will declare on February 5, 2018 that the state of the union is \"strong\". 2. In a speech near Cincinnati on February 5, 2018, Donald Trump criticized Democrats for not clapping at the State of the Union address. He also discussed the midterm elections and topics popular with his base, such as the lack of players kneeling during the national anthem at the Super Bowl. He suggested that Democrats were \"very selfish\" and asked if their lack of enthusiasm could be called \"treasonous.\" 3. In a speech on February 5, 2018 near Cincinnati, President Trump accused Democrats of being \"un-American\" and \"treasonous\" for not applauding during his State of the Union address when he mentioned good news such as rising wages and low African-American unemployment. He accused the Democrats of being selfish and suggested the lack of applause was an indication of their lack of patriotism. 4. On February 5, 2018, President Trump made a speech near Cincinnati where he accused Democratic congressional members of being un-American and potentially treasonous for not applauding positive news during his State of the Union address. He also criticized Nancy Pelosi for describing tax cut bonuses of $1,000 or more as \"crumbs\". Additionally, he noted that Republicans were \"going totally crazy wild\" during his speech. "
},
{
"figure_ref": [],
"heading": "Questions",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "AI-Generated Sentences:",
"publication_ref": [],
"table_ref": [],
"text": "1. This document discusses the concept of universal basic income (UBI), which is a policy idea that proposes every citizen should receive a regular stipend from the government to cover their basic needs. UBI has been discussed since the founding of the United States, most recently in the civil-rights era. Martin Luther King Jr. and Thomas Paine were both proponents of the idea. 2. This document discusses five reasons why Martin Luther King Jr. supported a guaranteed income, which is now referred to as Universal Basic Income. These reasons include automation and the dissolution of jobs that he saw occurring in 1961, advocating for a guaranteed income in his last book, and the passing of the U.S. House of Representatives in 1971. 3. The 115th Congress passed a number of laws related to taxes, criminal justice reform, the opioid crisis, and the Music Modernization Act. It also failed to pass funding for large parts of the federal government in the current fiscal year. However, it did pass the Tax Cuts and Jobs Act and the First Step Act. 4. In his September 21, 2019 Iowa Steak Fry speech, Andrew Yang stated that Thomas Paine and Martin Luther King Jr. have both advocated for a Universal Basic Income (UBI). In addition, Yang noted that the U.S.\nHouse of Representatives passed the measure twice in 1971. He also gave the example of the state of Alaska successfully implementing a basic income."
},
{
"figure_ref": [],
"heading": "Questions:",
"publication_ref": [],
"table_ref": [],
"text": "Q1: Was Thomas Paine for universal basic income? Q2: Was Martin Luther King Jr in support of a minimum basic income for all Americans? Q3: Did the House pass twice a bill supporting minimum basic income in 1971? Q4: Did the House pass twice a bill for minimum basic income in the 1970s?\" "
},
{
"figure_ref": [],
"heading": "ATTENTION",
"publication_ref": [],
"table_ref": [],
"text": "If there is no evidence that DIRECTLY addresses the question, you should consider it as Unanswerable, rather than select answerable with NO as the answer.\nTo prevent arbitrary selection, you should also write a brief sentence to explain your decision."
},
{
"figure_ref": [],
"heading": "Questions ${annotated_questions} Explanation",
"publication_ref": [],
"table_ref": [],
"text": "Previewing Answers Submitted by Workers This message is only visible to you and will not be shown to Workers. You can test completing the task below and click \"Submit\" in order to preview the data and format of the submitted results."
},
{
"figure_ref": [],
"heading": "Submit",
"publication_ref": [],
"table_ref": [],
"text": "Figure 11: Interface of the comprehensiveness study we conducted in Section 6.2."
}
] | Evidence retrieval is a core part of automatic fact-checking. Prior work makes simplifying assumptions in retrieval that depart from real-world use cases: either no access to evidence, access to evidence curated by a human fact-checker, or access to evidence available long after the claim has been made. In this work, we present the first fully automated pipeline to check real-world claims by retrieving raw evidence from the web. We restrict our retriever to only search documents available prior to the claim's making, modeling the realistic scenario where an emerging claim needs to be checked. Our pipeline includes five components: claim decomposition, raw document retrieval, fine-grained evidence retrieval, claim-focused summarization, and veracity judgment. We conduct experiments on complex political claims in the CLAIMDE-COMP dataset and show that the aggregated evidence produced by our pipeline improves veracity judgments. Human evaluation finds the evidence summary produced by our system is reliable (it does not hallucinate information) and relevant to answering key questions about a claim, suggesting that it can assist factcheckers even when it cannot surface a complete evidence set. | Complex Claim Verification with Evidence Retrieved in the Wild | [
{
"figure_caption": "…Figure 4 :4Figure 4: Two documents returned by searching Q2 (generated in the previous stage) through the search engine. Here we see the right page is created one month after the claim was made, citing an article written by PolitiFact, thus problematic to use as raw evidence.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 :5Figure5: System outputs for an example picked from the development set of CLAIMDECOMP: the claim is first decomposed into a set of yes/no questions and then the top four retrieved documents (through first and second stage retrieval) are summarized; finally, a trained DeBERTa model makes a prediction regarding the four summarized documents.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Figure 6 :6Figure6: Three examples from faithful evaluation (Section 6.1), showing the cases of minor error, major error, and completely wrong respectively. Red marks denote the mismatches between the summary and the document.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "The statistics for the retrieved documents obtained through the first-stage retrieval after filtering the documents from fact-checking websites. Did the number of people going to the ER for salmonella and other related illnesses increase a7er the plas8c bag ban in San Francisco? (Claim date: October 10, 2016)",
"figure_data": "# retrieved # scraped # wordsw/ timestamp66.745.01,561w/o timestamp70.447.81,660Jaccard score0.120.12-Bing SearchPlas9c Bag Ban ResponsibleDid bag ban cause disease?For Spike In E. Coli Infec9ons,Evidence is shaky …Study Says …This declara8on relied on aa 46 percent increase instudy that has numerousdeaths from foodborne illnessques8ons about itsin the three months a7er themethodology and conclusions.bag ban went into effect inWe rate this Mostly False.2007 …-Aus8n American-statesman-HuffPost (Feb. 7, 2013)(Nov. 25, 2016)",
"figure_id": "tab_1",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "The Centers for Disease Control and PrevenEon tracks gun deaths in all 50 states and the naEonal rate of gun deaths in 2016 was 11.8. No",
"figure_data": "Doc EtleURLSummariesFirearm-related deaths rate U.S. by gender 1970-2016 | StaEsta (Nov 7, 2019)hPps://www.staEsta.com/staEsEcs/ 186951/deaths-by-firearm-related-injuries-in-the-us-by-gender-since-1970/The death rate from firearm-related injuries in the United States is 19.4 per 100,000 populaEon among males, and homicides from firearms account for 72.6 percent of all homicides in the U.S. The ownership of legal firearms is widespread, with around 43 percent of households having at least one firearm. (Faithful)Do Gun Laws Affect the Rate of ShooEng Deaths? (Oct 12, 2018)hPps://www.thetrace.org/2018/10/ do-gun-laws-affect-the-rate-of-shooEng-deaths/. (Faithful)Gun Violence Deaths: How The U.S. Compares With The Rest Of The World (Mar 24, 2021)hPps://www.kuow.org/stories/gun-violence-deaths-how-the-u-s-compares-with-the-rest-of-the-world",
"figure_id": "tab_2",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "to countries such as Canada with 0.47 deaths per 100,000 people and the United Kingdom with 0.04 deaths per 100,000 people. (Faithful)",
"figure_data": "A Doctor's Insights Into Gun The World (Aug 6, 2019) Violence And Gun Laws Aroundworld violence-and-gun-laws-around-the-hPps://www.kuer.org/2019-08-06/ a-doctors-insights-into-gun-The US rate of deaths from gun violence is 4",
"figure_id": "tab_3",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "(1) Is the United States the country with the highest rate of gun deaths? (2) Does the U.S. have the highest rate of gun deaths compared to other countries? (3) Does the claim account for populaEon size (i.e., per capita rates), or is it based on total numbers? (4) Does the statement consider gun deaths relaEve to the total number of guns in the country? (5) Is the number of gun deaths in the United States substanEally higher when compared to countries of similar economic and poliEcal stability? (6) Do gun deaths account for a large porEon of deaths in the U.S.? (7) Are there any other countries with gun death rates close to that of the U.S.? (8) Are there any countries with a similar number of gun deaths as the United States? (9) Is the gun death rate in the United States increasing or decreasing? (10) Are there any miEgaEng factors that affect the gun death rate in the United States?",
"figure_data": "Retrieved documents and summaries:",
"figure_id": "tab_4",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Veracity classification performance with different retrieval constraints. The top block is our full system ( B setting in Table3) with constraints over what is retrieved. Red indicates using oracle information. \"+\" denotes that the results are statistically significant (p < 0.05) compared to the results of Claim only on the test set. spans), top-K 2 = 4 (highest-scored documents), k 1 = 30 (chunk size), and k 2 = 150 (expansion parameter). For training veracity classifier, we use DeBERTa-large as the base model. See appendix A.4 for all hyperparameters.",
"figure_data": "Retrieval ConstraintDev (N=200)Test (N=200)TemporalSiteAcc Soft Acc Macro-F1 MAE Acc Soft Acc Macro-F1 MAE--50.588.547.50.62 49.0 + 86.0 +48.5 +0.68 +-Non-FC 37.576.538.60.94 33.5 + 75.0 +33.9 +0.95 +Before-42.575.041.70.87 33.5 + 72.038.0 +0.98 +BeforeNon-FC 40.576.541.40.87 33.0 + 74.5 +34.5 +0.99 +Claim only37.071.034.60.98 25.568.027.51.12Claim + Just (oracle) 52.588.554.50.64 57.593.057.80.50",
"figure_id": "tab_5",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "With the vaccine you can sFll transmit, with the anFbodies you can't transmit.\" Document Title: COVID-19: Long-term effects -Mayo Clinic Content: It involves extreme faFgue that worsens with physical or mental acFvity , but doesn't improve with rest … What should you do if you have post-COVID-19 syndrome symptoms ? If you 're having symptoms of post-COVID-19 syndrome , talk to your health care provider … Summary: The Centers for Disease Control and PrevenFon states that there is no evidence to suggest that people who have recovered from COVID-19 and have anFbodies are not able to transmit the virus. Completely Wrong: The document is about the long-term effects of COVID-19. However, model is likely uFlizing its parameterized knowledge and draws the conclusion directly.",
"figure_data": "",
"figure_id": "tab_8",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Human evaluation results on 161 subquestions from the same 50 claims we picked for the human study on faithfulness. \"Ans\", \"Partially Ans\", and \"UnAns\" denote the number of questions that are answerable, partially answerable, and unanswerable.",
"figure_data": "Faithful Minor Unfaithful TotalAns4206Partially Ans6118Partially UnAns1351130UnAns5106Total28101250",
"figure_id": "tab_10",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "Claim-level statistics of few-shot-003 taking both faithfulness and comprehensiveness into consideration. \"Unfaithful\" label aggregates \"Major Error\" and \"Completely Wrong\" labels. The claim-level labels are derived from the sub-parts as defined in section 6.3.",
"figure_data": "",
"figure_id": "tab_11",
"figure_label": "6",
"figure_type": "table"
},
{
"figure_caption": "Model performance with respect to different rounds of retrieval at intervals of one week and two months. The overlap between document set, measured with Jaccard score, decreases as the time gap increases. None of the changes in four of our metrics is statistically significant.",
"figure_data": "First-stage Second-stage Summ# documents45.07.74.0# words70,2452,710251",
"figure_id": "tab_12",
"figure_label": "7",
"figure_type": "table"
},
{
"figure_caption": "Average number of unique documents and average number of words in total from those documents after each stage of our pipeline.",
"figure_data": "",
"figure_id": "tab_13",
"figure_label": "8",
"figure_type": "table"
}
] | Jifan Chen; Grace Kim; Aniruddh Sriram; Greg Durrett; Eunsol Choi | [
{
"authors": "Savvas Tariq Alhindi; Smaranda Petridis; Muresan",
"journal": "",
"ref_id": "b0",
"title": "Where is your evidence: Improving factchecking by justification modeling",
"year": "2018"
},
{
"authors": "Rami Aly; Zhijiang Guo; Sejr Michael; James Schlichtkrull; Andreas Thorne; Christos Vlachos; Oana Christodoulopoulos; Arpit Cocarascu; Mittal",
"journal": "Association for Computational Linguistics",
"ref_id": "b1",
"title": "The fact extraction and VERification over unstructured and structured information (FEVEROUS) shared task",
"year": "2021"
},
{
"authors": "Pepa Atanasova; Jakob Grue Simonsen; Christina Lioma; Isabelle Augenstein",
"journal": "",
"ref_id": "b2",
"title": "Generating fact checking explanations",
"year": "2020"
},
{
"authors": "Isabelle Augenstein; Christina Lioma; Dongsheng Wang; Lucas Chaves Lima; Casper Hansen; Christian Hansen; Jakob Grue Simonsen",
"journal": "Association for Computational Linguistics",
"ref_id": "b3",
"title": "MultiFC: A real-world multi-domain dataset for evidence-based fact checking of claims",
"year": "2019"
},
{
"authors": "Gagan Bansal; Tongshuang Wu; Joyce Zhou; Raymond Fok; Besmira Nushi; Ece Kamar; Marco Tulio Ribeiro; Daniel Weld",
"journal": "Association for Computing Machinery",
"ref_id": "b4",
"title": "Does the Whole Exceed Its Parts? The Effect of AI Explanations on Complementary Team Performance",
"year": "2021"
},
{
"authors": "Rishi Bommasani; Drew A Hudson; Ehsan Adeli; Russ Altman; Simran Arora; Sydney Von Arx; Jeannette Michael S Bernstein; Antoine Bohg; Emma Bosselut; Brunskill",
"journal": "",
"ref_id": "b5",
"title": "On the opportunities and risks of foundation models",
"year": "2021"
},
{
"authors": "Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell",
"journal": "Advances in neural information processing systems",
"ref_id": "b6",
"title": "Language models are few-shot learners",
"year": "2020"
},
{
"authors": "Danqi Chen; Adam Fisch; Jason Weston; Antoine Bordes",
"journal": "Association for Computational Linguistics",
"ref_id": "b7",
"title": "Reading Wikipedia to answer opendomain questions",
"year": "2017"
},
{
"authors": "Jifan Chen; Aniruddh Sriram; Eunsol Choi; Greg Durrett; ; ",
"journal": "Association for Computational Linguistics",
"ref_id": "b8",
"title": "Generating literal and implied subquestions to fact-check complex claims",
"year": "2022"
},
{
"authors": "Sihao Chen; Senaka Buthpitiya; Alex Fabrikant; Dan Roth; Tal Schuster",
"journal": "",
"ref_id": "b9",
"title": "PropSegmEnt: A Large-Scale Corpus for Proposition-Level Segmentation and Entailment Recognition",
"year": "2022"
},
{
"authors": "Eunsol Choi; Jennimaria Palomaki; Matthew Lamm; Tom Kwiatkowski; Dipanjan Das; Michael Collins",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b10",
"title": "Decontextualization: Making sentences stand-alone",
"year": "2021"
},
{
"authors": "Aakanksha Chowdhery; Sharan Narang; Jacob Devlin; Maarten Bosma; Gaurav Mishra; Adam Roberts; Paul Barham; Hyung Won Chung; Charles Sutton; Sebastian Gehrmann",
"journal": "",
"ref_id": "b11",
"title": "Palm: Scaling language modeling with pathways",
"year": "2022"
},
{
"authors": "Sebastian Dungs; Ahmet Aker; Norbert Fuhr; Kalina Bontcheva",
"journal": "",
"ref_id": "b12",
"title": "Can rumour stance alone predict veracity",
"year": "2018"
},
{
"authors": "Angela Fan; Claire Gardent; Chloé Braud; Antoine Bordes",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b13",
"title": "Augmenting transformers with KNN-based composite memory for dialog",
"year": "2021"
},
{
"authors": "Angela Fan; Aleksandra Piktus; Fabio Petroni; Guillaume Wenzek; Marzieh Saeidi; Andreas Vlachos; Antoine Bordes; Sebastian Riedel",
"journal": "Association for Computational Linguistics",
"ref_id": "b14",
"title": "Generating fact checking briefs",
"year": "2020"
},
{
"authors": "William Ferreira; Andreas Vlachos",
"journal": "ACL",
"ref_id": "b15",
"title": "Emergent: a novel data-set for stance classification",
"year": "2016"
},
{
"authors": "Luyu Gao; Zhuyun Dai; Panupong Pasupat; Anthony Chen; Arun Tejasvi Chaganty; Yicheng Fan; N Vincent Zhao; Hongrae Lao; Da-Cheng Lee; Kelvin Juan; Guu",
"journal": "",
"ref_id": "b16",
"title": "Rarr: Researching and revising what language models say, using language models",
"year": "2022"
},
{
"authors": "Mor Geva; Daniel Khashabi; Elad Segal; Tushar Khot; Dan Roth; Jonathan Berant",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b17",
"title": "Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies",
"year": "2021"
},
{
"authors": "Tanya Goyal; Junyi ; Jessy Li; Greg Durrett",
"journal": "",
"ref_id": "b18",
"title": "News Summarization and Evaluation in the Era of GPT-3",
"year": "2022"
},
{
"authors": "Lucas Graves",
"journal": "",
"ref_id": "b19",
"title": "Understanding the Promise and Limits of Automated Fact-Checking",
"year": "2018"
},
{
"authors": "Kelvin Guu; Kenton Lee; Zora Tung; Panupong Pasupat; Mingwei Chang",
"journal": "",
"ref_id": "b20",
"title": "Retrieval augmented language model pre-training",
"year": "2020"
},
{
"authors": " Pmlr",
"journal": "",
"ref_id": "b21",
"title": "",
"year": ""
},
{
"authors": "Andreas Hanselowski; Christian Stab; Claudia Schulz; Zile Li; Iryna Gurevych",
"journal": "Association for Computational Linguistics",
"ref_id": "b22",
"title": "A richly annotated corpus for different tasks in automated factchecking",
"year": "2019"
},
{
"authors": "Pengcheng He; Xiaodong Liu; Jianfeng Gao; Weizhu Chen",
"journal": "",
"ref_id": "b23",
"title": "DeBERTa: Decodingenhanced BERT with Disentangled Attention",
"year": "2020"
},
{
"authors": "Yichen Jiang; Shikha Bordia; Zheng Zhong; Charles Dognin; Maneesh Singh; Mohit Bansal",
"journal": "Association for Computational Linguistics",
"ref_id": "b24",
"title": "HoVer: A dataset for many-hop fact extraction and claim verification",
"year": "2020"
},
{
"authors": "Ryo Kamoi; Tanya Goyal; Juan ; Diego Rodriguez; Greg Durrett",
"journal": "",
"ref_id": "b25",
"title": "WiCE: Real-World Entailment for Claims in Wikipedia",
"year": "2023"
},
{
"authors": "Vladimir Karpukhin; Barlas Oguz; Sewon Min; Patrick Lewis; Ledell Wu; Sergey Edunov; Danqi Chen; Wen-Tau Yih",
"journal": "Association for Computational Linguistics",
"ref_id": "b26",
"title": "Dense passage retrieval for open-domain question answering",
"year": "2020"
},
{
"authors": "Urvashi Khandelwal; Omer Levy; Dan Jurafsky; Luke Zettlemoyer; Mike Lewis",
"journal": "",
"ref_id": "b27",
"title": "Generalization through memorization: Nearest neighbor language models",
"year": "2020"
},
{
"authors": "Omar Khattab; Christopher Potts; Matei Zaharia",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b28",
"title": "Relevance-guided supervision for OpenQA with ColBERT",
"year": "2021"
},
{
"authors": "Tom Kwiatkowski; Jennimaria Palomaki; Olivia Redfield; Michael Collins; Ankur Parikh; Chris Alberti; Danielle Epstein; Illia Polosukhin; Jacob Devlin; Kenton Lee; Kristina Toutanova; Llion Jones; Matthew Kelcey; Ming-Wei Chang; Andrew M Dai; Jakob Uszkoreit; Quoc Le; Slav Petrov",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b29",
"title": "Natural questions: A benchmark for question answering research",
"year": "2019"
},
{
"authors": "Patrick Lewis; Ethan Perez; Aleksandra Piktus; Fabio Petroni; Vladimir Karpukhin; Naman Goyal; Heinrich Küttler; Mike Lewis; Wen-Tau Yih; Tim Rocktäschel",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b30",
"title": "Retrieval-augmented generation for knowledge-intensive nlp tasks",
"year": "2020"
},
{
"authors": "Chloe Lim",
"journal": "Research & Politics",
"ref_id": "b31",
"title": "Checking how fact-checkers check",
"year": "2018"
},
{
"authors": "Yixin Liu; Alexander R Fabbri; Pengfei Liu; Yilun Zhao; Linyong Nan; Ruilin Han; Simeng Han; Shafiq Joty; Chien-Sheng Wu; Caiming Xiong; Dragomir Radev",
"journal": "",
"ref_id": "b32",
"title": "Revisiting the Gold Standard: Grounding Summarization Evaluation with Robust Human Evaluation",
"year": "2022"
},
{
"authors": "Sewon Min; Victor Zhong; Luke Zettlemoyer; Hannaneh Hajishirzi",
"journal": "Association for Computational Linguistics",
"ref_id": "b33",
"title": "Multi-hop reading comprehension through question decomposition and rescoring",
"year": "2019"
},
{
"authors": "Nikita Moghe; Siddhartha Arora; Suman Banerjee; Mitesh M Khapra",
"journal": "Association for Computational Linguistics",
"ref_id": "b34",
"title": "Towards exploiting background knowledge for building conversation systems",
"year": "2018"
},
{
"authors": "Reiichiro Nakano; Jacob Hilton; Suchir Balaji; Jeff Wu; Long Ouyang; Christina Kim; Christopher Hesse; Shantanu Jain; Vineet Kosaraju; William Saunders",
"journal": "",
"ref_id": "b35",
"title": "Webgpt: Browser-assisted questionanswering with human feedback",
"year": "2021"
},
{
"authors": "Preslav Nakov; David Corney; Maram Hasanain; Firoj Alam; Tamer Elsayed; Alberto Barr'on-Cedeno; Paolo Papotti; Shaden Shaar; Giovanni Da; San Martino",
"journal": "",
"ref_id": "b36",
"title": "Automated fact-checking for assisting human fact-checkers",
"year": "2021"
},
{
"authors": "Long Ouyang; Jeff Wu; Xu Jiang; Diogo Almeida; Carroll L Wainwright; Pamela Mishkin; Chong Zhang; Sandhini Agarwal; Katarina Slama; Alex Ray",
"journal": "",
"ref_id": "b37",
"title": "Training language models to follow instructions with human feedback",
"year": "2022"
},
{
"authors": "Ethan Perez; Patrick Lewis; Wen-Tau Yih; Kyunghyun Cho; Douwe Kiela",
"journal": "Association for Computational Linguistics",
"ref_id": "b38",
"title": "Unsupervised question decomposition for question answering",
"year": "2020"
},
{
"authors": "Verónica Pérez-Rosas; Bennett Kleinberg; Alexandra Lefevre; Rada Mihalcea",
"journal": "Association for Computational Linguistics",
"ref_id": "b39",
"title": "Automatic detection of fake news",
"year": "2018"
},
{
"authors": "Kashyap Popat; Subhabrata Mukherjee; Jannik Strötgen; Gerhard Weikum",
"journal": "International World Wide Web Conferences Steering Committee",
"ref_id": "b40",
"title": "Where the truth lies: Explaining the credibility of emerging claims on the web and social media",
"year": "2017"
},
{
"authors": "Kashyap Popat; Subhabrata Mukherjee; Andrew Yates; Gerhard Weikum",
"journal": "Association for Computational Linguistics",
"ref_id": "b41",
"title": "DeClarE: Debunking fake news and false claims using evidence-aware deep learning",
"year": "2018"
},
{
"authors": "Peng Qi; Xiaowen Lin; Leo Mehr; Zijian Wang; Christopher D Manning",
"journal": "Association for Computational Linguistics",
"ref_id": "b42",
"title": "Answering complex open-domain questions through iterative query generation",
"year": "2019"
},
{
"authors": "Ori Ram; Yoav Levine; Itay Dalmedigos; Dor Muhlgay; Amnon Shashua; Kevin Leyton-Brown; Yoav Shoham",
"journal": "",
"ref_id": "b43",
"title": "In-context retrieval-augmented language models",
"year": "2023"
},
{
"authors": "Eunsol Hannah Rashkin; Jin Yea Choi; Svitlana Jang; Yejin Volkova; Choi",
"journal": "Association for Computational Linguistics",
"ref_id": "b44",
"title": "Truth of varying shades: Analyzing language in fake news and political fact-checking",
"year": "2017"
},
{
"authors": "Devendra Sachan; Mike Lewis; Mandar Joshi; Armen Aghajanyan; Wen-Tau Yih; Joelle Pineau; Luke Zettlemoyer",
"journal": "Association for Computational Linguistics",
"ref_id": "b45",
"title": "Improving passage retrieval with zero-shot question generation",
"year": "2022"
},
{
"authors": "Tal Schuster; Adam Fisch; Regina Barzilay",
"journal": "Association for Computational Linguistics",
"ref_id": "b46",
"title": "Get your vitamin C! robust fact verification with contrastive evidence",
"year": "2021"
},
{
"authors": "Weijia Shi; Sewon Min; Michihiro Yasunaga; Minjoon Seo; Rich James; Mike Lewis; Luke Zettlemoyer; Wen-Tau Yih",
"journal": "",
"ref_id": "b47",
"title": "Replug: Retrievalaugmented black-box language models",
"year": "2023"
},
{
"authors": "Prakhar Singh; Anubrata Das; Junyi ; Jessy Li; Matthew Lease",
"journal": "",
"ref_id": "b48",
"title": "The case for claim difficulty assessment in automatic fact checking",
"year": "2021"
},
{
"authors": "Alon Talmor; Jonathan Berant",
"journal": "Association for Computational Linguistics",
"ref_id": "b49",
"title": "The web as a knowledge-base for answering complex questions",
"year": "2018"
},
{
"authors": "Romal Thoppilan; Daniel De Freitas; Jamie Hall; Noam Shazeer; Apoorv Kulshreshtha; Heng-Tze; Alicia Cheng; Taylor Jin; Leslie Bos; Yu Baker; Du",
"journal": "",
"ref_id": "b50",
"title": "Lamda: Language models for dialog applications",
"year": "2022"
},
{
"authors": "James Thorne; Andreas Vlachos; Christos Christodoulopoulos; Arpit Mittal",
"journal": "Association for Computational Linguistics",
"ref_id": "b51",
"title": "FEVER: a large-scale dataset for fact extraction and VERification",
"year": "2018"
},
{
"authors": "Andreas Vlachos; Sebastian Riedel",
"journal": "Association for Computational Linguistics",
"ref_id": "b52",
"title": "Fact checking: Task definition and dataset construction",
"year": "2014"
},
{
"authors": "Svitlana Volkova; Kyle Shaffer; Jin Yea Jang; Nathan Hodas",
"journal": "Association for Computational Linguistics",
"ref_id": "b53",
"title": "Separating facts from fiction: Linguistic models to classify suspicious and trusted news posts on Twitter",
"year": "2017"
},
{
"authors": "William Yang; Wang ",
"journal": "Association for Computational Linguistics",
"ref_id": "b54",
"title": "liar, liar pants on fire\": A new benchmark dataset for fake news detection",
"year": "2017"
},
{
"authors": "Tomer Wolfson; Mor Geva; Ankit Gupta; Matt Gardner; Yoav Goldberg; Daniel Deutch; Jonathan Berant",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b55",
"title": "Break it down: A question understanding benchmark",
"year": "2020"
},
{
"authors": "Tianyi Zhang; Faisal Ladhak; Esin Durmus; Percy Liang; Kathleen Mckeown; Tatsunori B Hashimoto",
"journal": "",
"ref_id": "b56",
"title": "Benchmarking Large Language Models for News Summarization",
"year": "2023"
},
{
"authors": "Zexuan Zhong; Tao Lei; Danqi Chen",
"journal": "",
"ref_id": "b57",
"title": "Training language models with memory augmentation",
"year": "2022"
}
] | [
{
"formula_coordinates": [
1,
373.68,
316.13,
74.77,
5.69
],
"formula_id": "formula_0",
"formula_text": "… …"
}
] |
10.18653/v1/2020.acl-main.92 | 2023-11-16 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b45",
"b21",
"b30",
"b22",
"b15",
"b0",
"b23",
"b23"
],
"table_ref": [],
"text": "The increasing adoption of large language models (LLMs) across various tasks, such as text generation and reasoning (Wei et al., 2022;Kojima et al., 2022;Wang et al., 2022a;Mishra et al., 2022), mathematical reasoning (Lewkowycz et al., 2022;Gao et al., 2022;Arora et al., 2023), and code generation (Li et al., 2022;Madaan et al., 2023b), has underscored the importance of improving the correctness of their outputs. A popular method for achieving this goal is Self-Consistency (Wang et al., 2022b), a majority voting technique where multiple output samples are generated for a given input, and the final decision is based on the most frequently occurring output among the samples.\nCurrent Self-Consistency methods typically employ a fixed budget approach, wherein a predetermined number of samples (e.g., 40) are generated to make a decision. However, as LLMs continue to grow in size and complexity, the sampling time and computational costs associated with majority voting become increasingly challenging. This challenge is particularly evident in high-stakes applications like competition-level code generation (Li et al., 2022), where generating a large number of programs, sometimes up to a million, is essential for maximizing performance.\nTo address this challenge, we introduce Adaptive-Consistency, a cost-efficient, model-agnostic majority voting technique. Adaptive-Consistency employs a lightweight stopping criterion that dynamically adjusts the number of samples (n) for each input, as opposed to using a fixed budget (k). The intuition is that if a clear majority is established with high confidence after sampling fewer than k answers (n < k), there is no need to generate additional samples.\nAdaptive-Consistency models the probability distribution over unique samples using a Dirichlet distribution, allowing us to quantify the confidence in the lead of the majority element over other elements. For instance, if the majority element has a count of 9 out of the first 10 samples, the likelihood of it remaining the majority element even after 40 samples is very high (> 99%). This allows Adaptive-Consistency to stop sampling at this point, reducing the cost by 30 samples, while Self-Consistency would continue to sample all 40 answers. As an inference-time technique requiring no additional training, Adaptive-Consistency provides a convenient off-the-shelf option for all pre-trained language models, offering the flexibility to balance computational cost and performance.\nWe evaluate Adaptive-Consistency on 17 diverse tasks and three LLMs of different scales (VICUNA-13B, CODE-DAVINCI-002 and GPT-3.5-TURBO). Our experimental results show that Adaptive-Consistency outperforms Self-Consistency regarding cost efficiency while maintaining comparable output quality. On CODE-DAVINCI-002, Adaptive- Consistency reduces the number of samples required by a factor of 3.4×, with no average drop in accuracy. On VICUNA-13B, it requires sampling 1.9× fewer samples, with almost no drop in accuracy. Similarly, on GPT-3.5-TURBO, it samples 4.4× fewer samples, with less than 0.2% drop in accuracy. In summary, our contributions are:\n• We propose Adaptive-Consistency, a costefficient sampling technique for large language models that dynamically adjusts the number of samples using a lightweight stopping criterion based on the stability of the majority element. • We conduct extensive experiments using three different LLMs on a diverse set of 17 datasets. These datasets encompass a wide range of tasks, including MATH, COMMONSENSE, SYM-BOLIC reasoning, and CODE GENERATION tasks. Adaptive-Consistency consistently and significantly outperforms fixed-budget methods like Self-Consistency, requiring an average of 3.3× fewer samples with less than 0.1% drop in accuracy across all datasets and models.\n• Our analysis reveals that for a fixed sampling cost, Adaptive-Consistency consistently achieves better accuracy than Self-Consistency across all datasets (upto 5% absolute points). Additionally, we experiment with various stopping criterias and show the efficacy of Adaptive-Consistency in terms of speed and accuracy."
},
{
"figure_ref": [],
"heading": "Background",
"publication_ref": [],
"table_ref": [],
"text": "In-Context Few-Shot Prompting In-context few-shot prompting is a technique employed by large language models (LLMs) to learn and generalize from a limited number of examples provided within the input of a given task. The model can quickly adapt to novel tasks without fine-tuning or additional training by conditioning the model on a few examples. Specifically, a prompt p is constructed by concatenating multiple input-answer example pairs < x i , a i >. The prompt is then prepended to the test input x test , and the model generates the corresponding answer a test . Listing 1: Comparison of Adaptive-Consistency (top) and Self-Consistency (bottom). Self-Consistency always generates a fixed number of samples. In contrast, Adaptive-Consistency uses a lightweight stopping criterion, allowing it to adaptively halt the sampling process, which can lead to improved efficiency and performance. Wang et al. (2022b) proposed Self-Consistency which improved performance by sampling multiple diverse reasoning chains and aggregating their outputs using a simple majority voting mechanism. However, higher accuracy is achieved with an increased computational cost, since the LLM must be prompted multiple times for the same question."
},
{
"figure_ref": [],
"heading": "Self-Consistency",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Adaptive-Consistency",
"publication_ref": [],
"table_ref": [],
"text": "Self-Consistency generates a predetermined number of answers (k) from the language model (LLM) before returning the majority answer. In contrast, the Adaptive-Consistency method takes an incremental approach to sampling outputs from the language model. After generating each sample, Adaptive-Consistency employs a lightweight stopping criteria to determine whether it should 1.) generate an additional sample from LLM or 2.) cease sampling and report the current majority answer. This flexible strategy enables Adaptive-Consistency to dynamically adjust the number of samples generated so far (n) for each input. As our experiments demonstrate, n is typically less than k (on average, 3.3× and up to 7.9× less in some cases), allowing Adaptive-Consistency to offer greater cost-efficiency compared to the fixed budget approach employed by Self-Consistency.\nAdaptive-Consistency differs from Self-Consistency only in terms of the stopping criteria (Listing 1). The design of the stopping criteria is crucial to our method, as it aims to minimize the average number of samples generated from the LLM while maximizing accuracy. The simplicity of our algorithm allows for the use of various stopping criteria interchangeably, each with its own advantages and disadvantages. We expand on a particular choice of stopping function next."
},
{
"figure_ref": [],
"heading": "Dirichlet Stopping Criteria",
"publication_ref": [],
"table_ref": [],
"text": "Let n be the number of samples generated from LLM so far, with m unique samples. Let v = [v 1 , v 2 , . . . , v m ] be the counts of each element, and p i = v i n be the normalized count. For instance, if n = 10, and m = 3 (10 samples generated, with 3 unique elements), if v = [8, 1, 1], then we can be more confident that v 1 is the answer. On the other hand, if v = [4, 4, 2], then more samples need to be generated. Our goal is to formalize and quantify this intuition.\nBy convention, let p 1 = max(p i ). We want to assess the stability of p 1 as the majority element.2 Specifically, we want to ask the following question: what is the probability that p 1 will be the majority element if we repeat the process of generating n samples again? Intuitively, if this probability is higher than some predetermined threshold C thresh , then we can be more confident in our decision to stop sampling and return p 1 as the majority element:\nP (p 1 > m max i=2 p i | v) > C thresh\nTo answer this question, we establish a connection with the Dirichlet distribution. Specifically, we note that the counts v parameterize a Dirichlet distribution, Dir(V ).3 This connection allows us to explore the behavior of the sampling process by drawing more samples from Dir(V ) and observing the stability of p 1 as the majority element. To compute the probability of p 1 being the majority element, we can integrate the joint probability density function of the Dirichlet distribution over the appropriate region of the probability simplex. The integral can be expressed as follows:\nP (p 1 > m max i=2 p i | V ) = 1 0 S(p ′ 1 ) f (p ′ 1 , p 2 , . . . , p m | V ). dp 2 • • • dp m dp ′ 1 ,where\nS(p ′ 1 ) = {(p 2 , . . . , p m ) | p ′ 1 > m max i=2 p i , m i=2 p i = 1 -p ′ 1 }.\n(1) In Equation 1, f (p ′ 1 , p 2 , ..., p m |V ) represents the joint probability density function of the Dirichlet distribution conditioned on the counts V . The bounds on the integral for p ′ 1 range from 0 to 1. The probability simplex S(p ′ 1 ) is defined for each p ′ 1 value, such that p ′ 1 > max m i=2 p i , and the remaining p i values sum to 1 -p ′ 1 . This constraint ensures that we are considering all possible values of p ′ 1 that would maintain its majority status. Here we assume, that the number of possible unique answers (m) is known, based on the current set of observations (V ). In Analysis (( § 5.3), we further evaluate a CHINESE RESTAURANT PROCESS (CRP) stopping criteria, which relaxes this assumption by not requiring the number of possible unique answers (m) to be known in advance.\nBeta Stopping Criteria Since the number of unique answers in the observation set can be large, Equation ( 1) is computationally expensive to solve. As an approximation, we observe that establishing the majority of p 1 over the next largest probability, p 2 , is sufficient for our purpose.\nIn this setting, the probability in Equation (3) simplifies to a Beta distribution with parameters (v 1 + 1, v 2 + 1), and Equation ( 1) is replaced by Equation (2). This approximation, which assumes a non-informative prior of BETA(1, 1), allows us to efficiently compute the confidence in p 1 being the majority, enabling early stopping decisions without incurring substantial computational overhead.\n0.5 0 p v2 2 • (1 -p 2 ) v 1 dp 2 (2)\nEmpirically, we show the performance to be similar to Dirichlet stopping criteria but significantly faster (See Section 5.3). Throughout experiments, we refer to this Beta Stopping Criteria as Adaptive-Consistency."
},
{
"figure_ref": [],
"heading": "Code-Generation",
"publication_ref": [
"b23"
],
"table_ref": [],
"text": "We now turn our attention to CODE GENERATION tasks, which involve generating programs that can correctly pass multiple test cases. More details on test case generation can be found in Appendix A.4.\nThe configuration of code generation tasks significantly impacts the Self-Consistency measurement since different programs might yield varying outputs for a given set of test cases. This variation can cause simple majority voting schemes to be ineffective in evaluating stability. To address this, we explore two distinct methods for aggregating answers across multiple test cases.\nIn the first method, inspired by the approach used in AlphaCode (Li et al., 2022), we concatenate the outputs for all test cases into a single vector with t elements and apply Self-Consistency across the entire vector. This implies that two programs are considered identical only if their outputs for all t test cases match exactly. However, this simple setup may overestimate the output variance, as different programs can produce distinct outputs for the set of test cases.\nTo overcome the limitations of the simple setup, we propose an alternative method that treats test inputs as independent entities and applies Adaptive-Consistency to each test case separately:\nt t j=1 P (p j 1 > m max i=2 p j i | V )(3)\nIn this equation, P is computed using Equation 1. The Adaptive-Consistency method terminates the sampling process when the normalized probability-expressed as the geometric mean of P across all t test cases-exceeds a predefined threshold (e.g., 0.95)."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [
"b15",
"b6",
"b10",
"b32",
"b17",
"b8",
"b1",
"b18",
"b23",
"b9",
"b41",
"b42",
"b5",
"b8",
"b15",
"b45",
"b45",
"b6"
],
"table_ref": [],
"text": "We evaluate Adaptive-Consistency using 17 diverse benchmark datasets and three different language models. We use prompts by program-aided language models, PAL, (Gao et al., 2022), Self-Consistency (Wang et al., 2022b) and CodeT (Chen et al., 2022).\nDatasets We evaluate our method on a diverse set of reasoning and coding benchmarks, encompassing 17 datasets across 4 distinct categories: 1. Mathematical Reasoning: We use GSM-8K (Cobbe et al., 2021), SVAMP (Patel et al., 2021), andASDIV (Miao et al., 2020) which assess the mathematical reasoning capabilities of the LLMs. 2. COMMONSENSE Reasoning Tasks: We evaluate on 5 datasets: STRATEGYQA (Geva et al., 2021), DATE UNDERSTANDING, SNARKS, RUIN NAMES, SALIENT TRANSLATION that measures different capabilites of LLMs such as multi-hop reasoning and emotional understanding. 3. SYM-BOLIC Reasoning Tasks: We further examine performance on 5 diverse SYMBOLIC reasoning tasks: TRACKING SHUFFLED OBJECTS, LOGI-CAL DEDUCTION, BOOLEAN EXPRESSIONS, DIS-AMBIGUATION QA, PENGUINS. 4. CODE GEN-ERATION Tasks We also evaluate our method on coding tasks, which require to generate a working code given a textual problem description. We evaluate on 4 datasets of varying difficulty: HU-MANEVAL (Chen et al., 2021), MBPP (Austin et al., 2021), APPS (Hendrycks et al., 2021) and CODECONTESTS (Li et al., 2022). We refer readers to Appendix A.2 for more details.\nModels We evaluate our method on three different language models: 1. GPT-3.5-TURBO:4 An RLHF-finetuned GPT-3 based model (unreleased number of parameters). 2. VICUNA-13B: (Chiang et al., 2023) an open-source transformer model fine-tuned on instruction-following dataset (Taori et al., 2023) from the base Llama series (Touvron et al., 2023). 3. CODE-DAVINCI-002: A GPT-3-based publicly available model (Brown et al., 2020) which is a part of the Codex series (Chen et al., 2021) and has 175 billion parameters. 5Prompting and Sampling We use similar prompts as in PAL (Gao et al., 2022) and CHAIN OF THOUGHT (Wei et al., 2022). Specifically, for mathematical reasoning and DATE UNDERSTAND-ING tasks, we use prompts from PAL. For other commonsense and SYMBOLIC reasoning tasks, we use COT (Wei et al., 2022).\nFor sampling, we follow the scheme suggested in Wang et al. (2022b). Specifically, we use a temperature of 0.7 for sampling and limit the number of generations to a maximum of 40. For coding tasks, we follow the exact procedure as used in CodeT (Chen et al., 2022), with 50 samples for APPS, 100 samples for HUMANEVAL and MBPP and 1000 samples in CODECONTESTS."
},
{
"figure_ref": [],
"heading": "Hyperparameters",
"publication_ref": [],
"table_ref": [],
"text": "The only hyperparameters in Adaptive-Consistency are those related to parameters in stopping criteria (C thresh ). We use a high C thresh = 0.95 for Adaptive-Consistency. By using a high threshold, we aim to maintain high accuracy and prevent the algorithm from stopping too early. For other Stopping Criteria, we tune parameters on the training set of GSM-8K, and use the same thresholds across all the datasets. The impact of the chosen threshold on the performance of our method is further analyzed in the analysis section ( § 5.1).\nBaselines We compare our method against Self-Consistency, which is the current state-of-the-art method. Further, in Section 5.3, we evaluate Adaptive-Consistency against different stopping criteria, such as RANDOM stopping and MAJOR-ITY (stopping at majority), ENTROPY, DIRICHLET and CRP.\nEvaluation Metrics We evaluate the performance of our method and the baselines using two metrics: average generations sampled from the LLMs, and overall reasoning accuracy. Our results show that Adaptive-Consistency achieves similar performance to Self-Consistency while often reducing sample budget considerably."
},
{
"figure_ref": [],
"heading": "Results",
"publication_ref": [],
"table_ref": [
"tab_1",
"tab_7"
],
"text": "Table 1 presents the main results, and is divided into two parts showing results across different task categories (top sub-table) and on various language models (bottom sub-table ). We focus on the potential tradeoff between efficiency and accuracy.\nResults Across Task Categories Our experimental results demonstrate the significant efficiency gains achieved by Adaptive-Consistency across different task categories -3.3× times fewer samples in mathematical tasks with a 0.1% accuracy drop, 2.9× times fewer samples in commonsense tasks with a 0.2% accuracy drop, 3.8× times fewer samples in symbolic reasoning tasks maintaining accuracy, and 2.4× times fewer samples in coding tasks while improving accuracy by 0.4%. These findings confirm the effectiveness of Adaptive-Consistency in identifying the majority element early, highlighting its potential across various applications, including reasoning and coding. Adaptive-Consistency achieves a significant reduction in the number of generations, with a negligible impact on accuracy. The ∆ columns display reductions in generations (Num. Gen.) and accuracy (Acc.) between Self-Consistency and Adaptive-Consistency. Detailed results are in Table 5.\nResults Across Language Models Examining the results across different language models, we find that Adaptive-Consistency is model-agnostic, and consistently reduces the number of generations with minimal to no impact on accuracy. Adaptive-Consistency consistently reduces the number of generations required, with reductions of 4.4× for GPT-3.5-TURBO, 1.9× for VICUNA-13B, and 3.4× for CODE-DAVINCI-002, highlighting its cost-effective nature and adaptability to different scales of models. Moreover, the minimal accuracy differences and slight improvements showcase the practical utility of Adaptive-Consistency, emphasizing its diverse applicability and model-agnostic characteristics."
},
{
"figure_ref": [],
"heading": "Analysis",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [
"fig_1"
],
"heading": "Effect of Confidence Threshold in Adaptive-Consistency",
"publication_ref": [],
"table_ref": [],
"text": "The confidence threshold, C thresh , is a crucial hyperparameter for Adaptive-Consistency, as it determines when to stop sampling based on the desired level of confidence in the majority element. While we set the threshold to a stringent value of 0.95 for all experiments, in this section, we analyze the impact of varying C thresh from 0.5 to 1 to understand the trade-offs between model accuracy and cost-efficiency.\nIn Figure 2, we present a visualization that examines the relationship between the confidence threshold, C thresh , and the performance of adaptive consistency in terms of both accuracy and costefficiency. The x-axis represents the confidence threshold, varying from 0.5 to 1. The left y-axis displays the model's accuracy, while the right yaxis shows the average number of samples drawn.\nThe plot (for GSM-8K) shows the expected behavior of two curves: the blue curve (accuracy) increases gradually and then plateaus, while the red curve (average number of samples) initially increases linearly and then climbs more steeply. The plateau in accuracy signifies that the model has reached its maximum achievable accuracy, and further sampling will not improve it much. Meanwhile, the red curve's climbing rate indicates that the model requires more samples to meet an increasingly stringent confidence threshold for stopping, highlighting the trade-off between accuracy and cost efficiency. We refer readers to Appendix C.4 for more results."
},
{
"figure_ref": [],
"heading": "Adaptive-Consistency vs. Self-Consistency",
"publication_ref": [],
"table_ref": [],
"text": "For Equal Average Sample Costs Section 4.1 previously demonstrated that Adaptive-Consistency achieves comparable performance to Self-Consistency using fewer samples. In this section, our primary objective is to compare the performance of Adaptive-Consistency to Self-Consistency across various sampling budgets. For each fixed sampling budget k, we contrast the performances of Adaptive-Consistency and Self-Consistency, where Self-Consistency distributes sample budget uniformly to each question, Adaptive-Consistency uses nonuniform allocation, rather than consistently across all instances. We evaluate Adaptive-Consistency using varying thresholds, with each threshold producing a distinct point (#samples, performance) on the costquality curve. For every specific sample count (#samples) generated by Adaptive-Consistency, we subsequently run Self-Consistency to obtain its corresponding performance. The relationship between the two methods across these data points is visualized in Figure 3 which provides a visual comparison of the performance of Adaptive-Consistency and Self-Consistency on GSM-8K. Adaptive-Consistency outperforms Self-Consistency in accuracy across all average sample costs. For example, when the average sample cost is 10, Adaptive-Consistency achieves approximately 3% higher accuracy on GSM-8K. Similar results hold on other datasets; see Appendix C.1 for full results. The success of Adaptive-Consistency can be attributed to the fact that it varies the number of samples based on the complexity of the instance, using more samples where a clear consensus is hard to reach and fewer where answers are consistent. Consequently, Adaptive-Consistency achieves improved overall performance when controlled for cost budget."
},
{
"figure_ref": [
"fig_2"
],
"heading": "Evaluation of Different Stopping Functions",
"publication_ref": [
"b24",
"b11",
"b46",
"b4",
"b13",
"b34",
"b12",
"b49",
"b35",
"b47",
"b31"
],
"table_ref": [
"tab_9"
],
"text": "Adaptive-Consistency allows a flexible choice of stopping criteria, based on intended objective and requirements. Here, we evaluate six different functions: 1) RANDOM: randomly stopping with a probability p, 2) MAJORITY: stopping after the most common answer has a majority above a threshold, 3) ENTROPY: stopping after the entropy of answers is below a threshold, 4) BETA: The main The parameters for all these methods are tuned, as discussed in Section 4. Figure 4 compares BETA to ENTROPY and MAJORITY over a range of expected sampling costs. BETA consistently achieves higher accuracy than both for the same sampling cost. Further, we find RANDOM to be the least effective method as expected, whereas MAJORITY almost consistently underperforms both BETA and ENTROPY. While DIRICHLET and CRP have a similar performance to BETA, they are both about four orders of magnitude slower than BETA due to the expensive multivariate integral calculation. Nonetheless, despite being run on a single cpu core, even DIRICHLET and CRP have negligible time and cost compared to LLM inference. The exact timings are presented in Table 2. The detailed results are presented in Appendix C.2, Table 7.\nIn summary, Adaptive-Consistency is particularly effective in two scenarios: (i) when a majority trend is evident early in the sampling process, such as in the SVAMP dataset where it achieves comparable accuracy to Self-Consistency using fewer than 5 samples on average per input; and (ii) for tasks with a limited set of potential answers, such as the BOOLEAN EXPRESSIONS dataset where Adaptive-Consistency reduces the computational budget by 7.9 times without any loss in accuracy. niques from crowdsourcing (Lin et al., 2012;Dai et al., 2013;Weld et al., 2015;Bragg et al., 2016). Traditionally, crowdsourcing involves aggregating diverse human judgments, which presents challenges in managing resource allocation-knowing when to query additional contributors or stop based on the consistency of responses (Doan et al., 2011;Quinn and Bederson, 2011). Early research concentrated on probabilistic models estimating the 'true' answer and worker reliability (Dawid and Skene, 1979;Whitehill et al., 2009), later considering factors like worker expertise, task complexity, and answer quality (Raykar et al., 2010;Welinder et al., 2010). However, rather than addressing these issues with multiple human contributors, Adaptive-Consistency is tailored specifically for LLMs, optimizing for computational efficiency and output accuracy. In line with our vision, (Parameswaran et al., 2023) have recently proposed declarative prompt engineering, viewing LLMs like crowd workers and leveraging multiple prompting strategies."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b25",
"b53",
"b37",
"b16",
"b50",
"b36",
"b19",
"b51",
"b14",
"b20",
"b28",
"b52",
"b7"
],
"table_ref": [],
"text": "Architectures for adaptive computation A related body of work on adaptive computation aims to preempt computation based on intermediate representations (Liu et al., 2020;Zhou et al., 2020;Schuster et al., 2021;Geng et al., 2021;Xin et al., 2020). Schuster et al. (2022) present CLAM, a language model that performs language generation adaptively. Hou et al. (2020) propose Dynamic Bert, which can adapt the depth and width of the transformer to satisfy various computational constraints. Xing et al. (2020) propose a dynamic deep neural network with an early-exit strategy embedded for enhancing the quality of compressed images. Another direction of work focuses on pruning model weights or training sparse weights (Fan et al., 2019;Jayakumar et al., 2021) to reduce training and inference time. In contrast to these methods, our approach completely obviates making any architectural modifications.\nInference-time adaptive computation These methods focus on adaptive computation at inference time without making architectural modifications to the models. Schwarzschild et al. (2021b,a) focus on three different generalization tasks. They observe that increasing the number of test iterations (which corresponds to the network depth in their setting) helps the models in generalizing better to difficult problems. Madaan and Yang (2022) leverage two different networks trained for the same task, a larger variant (slow) and a smaller variant (fast). The switch from fast to slow happens during inference, based on the complexity of generation at the current step. Xue et al. (2023) train language models to adaptively read tokens from a tape bank for each input. Different from these works, our focus is tasks where the multiple samples are drawn from a model (vs. iteratively solving a task, which is a focus of these works). Additionally, recent works such as (Madaan et al., 2023a;Chen et al., 2023) have propsed to adaptively selecting models of varying sizes based on verification signals derived from the output of the smaller model. Our methods, however, distinguish themselves by not necessitating the use of an additional verifier, and without the need of multiple models."
},
{
"figure_ref": [],
"heading": "Adaptive Sampling in Training and Active",
"publication_ref": [
"b2",
"b33",
"b3"
],
"table_ref": [],
"text": "Learning Another line of work focuses on importance-based sampling of input instances during training (Bengio and Senecal, 2008;Prabhu et al., 2019;Berger et al., 2017). In contrast to the aforementioned methods, our approach centers on adaptively sampling multiple outputs per input instance during the inference phase, without soliciting additional labels. Our method is crafted to efficiently obtain reliable predictions from pretrained language models by adaptively sampling their outputs, distinguishing it from both adaptive sampling in training and active learning, which focus on the training phase."
},
{
"figure_ref": [],
"heading": "Conclusion and Future Work",
"publication_ref": [],
"table_ref": [],
"text": "This paper presented Adaptive-Consistency, a costefficient and model-agnostic technique for improving the correctness of output from large language models (LLMs) using dynamic sampling. Our approach builds upon the Self-Consistency method and introduces a lightweight stopping criterion that allows for adaptive sampling based on the amount of agreement in the samples drawn so far. Adaptive-Consistency is effective across 17 datasets and three LLMs, on both reasoning and coding tasks. It reduces the required sample budget by 2 to 4 times, while maintaining comparable accuracy, with an average drop of less than 0.1%.\nOur work opens up several avenues for future research. We may develop alternative stopping criteria, or combining multiple criteria could lead to even more efficient sampling techniques. Moreover, in our current approach, the majority decision relies on using matches to determine the most common answer. However, this may not always capture the true majority, e.g., in generative tasks, where the output can have variations that do not affect the overall correctness or relevance of the answer. To foster further research and enable reproducibility, we have released the code and LLM outputs at https://sample-step-by-step.info/. also partially supported by the CSE Research Acceleration Fund of IIT Delhi Aman is supported by a contract from the DARPA KAIROS program under agreement number FA8750-19-2-0200. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes, notwithstanding any copyright notation thereon. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the U.S. Government."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [],
"table_ref": [],
"text": "Despite the promising results of our proposed Adaptive-Consistency method, it bears several limitations and scopes for future improvement.\n• Stopping criterion sensitivity: The current stopping criterion, based on the majority element's stability in the sample set, may not always indicate sample agreement optimally. Instances may arise where the majority element lacks stability, yet the criterion triggers, potentially leading to suboptimal decisions. Future work could explore more robust or alternative stopping criteria. • Generalizability: The effectiveness of our method may vary across tasks or models, despite testing on a diverse range of 17 datasets and three different LLMs of contrastive scale. Notably, Adaptive-Consistency is anticipated to fail where Self-Consistency fails. • Task-specific adaptations: The task-agnostic nature of Adaptive-Consistency might limit its performance on tasks that could benefit from task-specific adaptations. Specialized versions of Adaptive-Consistency for specific tasks or domains could potentially enhance performance. We have initiated this by experimenting on CODE GENERATION dataset, but extending Adaptive-Consistency to other domains may not be as straightforward. • Reliance on the pretrained LLM: Our method depends on the pretrained LLM for generating multiple samples. Consequently, any limitations or biases in the LLM would persist in the Adaptive-Consistency. Addressing these issues might require improvements in the LLM training process itself or the integration of external knowledge sources."
},
{
"figure_ref": [],
"heading": "A Experimental Setup",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "A.1 Hyperparameters",
"publication_ref": [],
"table_ref": [
"tab_7"
],
"text": "The only hyperparameters in Adaptive-Consistency, are those related to parameters in stopping criterias (C thresh ). We use a high C thresh = 0.95 for Adaptive-Consistency. By using a high threshold, we aim to maintain high accuracy and prevent the algorithm from stopping too early. For other Stopping Criterias, we tune our parameters on the training set of GSM-8K, and use the same thresholds across all the datasets. The impact of the chosen threshold on the performance of our method is further analyzed in the Analysis Section ( § 5.1). We further evaluate all methods on a set of 3 seeds and report the table with standard deviation in Table 5. We use only a single seed for GPT-3.5-TURBO because of the cost associated."
},
{
"figure_ref": [],
"heading": "A.2 Benchmarks",
"publication_ref": [
"b10",
"b32",
"b29",
"b17",
"b8",
"b1",
"b18",
"b23"
],
"table_ref": [],
"text": "We evaluate our method on a diverse set of coding and reasoning benchmark datasets, encompassing 17 datasets across four distinct categories: 1. MATHEMATICAL Reasoning: To assess mathematical reasoning capabilities, we utilize the following datasets: GSM-8K (Cobbe et al., 2021), SVAMP (Patel et al., 2021), and ASDIV (Miao et al., 2020). These datasets consist of gradeschool-level algebra word problems necessitating arithmetic operations and problem-solving based on contextual information.\n2. COMMONSENSE Reasoning Tasks: We evaluate Adaptive-Consistency on four COMMON-SENSE reasoning tasks. 1.) STRATEGYQA (Geva et al., 2021) comprises questions that demand the model to infer a multi-hop strategy with reasoning steps implicitly embedded in the questions. 2.) DATE UNDERSTANDING entails questions that require the model to deduce dates from natural language descriptions and perform arithmetic operations accordingly. 3.) SALIENT TRANSLATION is a salient translation error detection task that requires the model to identify the type of error in a translation. 4.) SNARKS and 5.) RUIN NAMES both focus on emotional understanding tasks.\n3. SYMBOLIC Reasoning Tasks: We examine the performance of our method on six diverse SYM-BOLIC reasoning tasks. 1.) TRACKING SHUF-FLED OBJECTS is a tracking task that necessitates the model to infer the final state of a system, given its initial state and a sequence of modifications. task that demands the model to deduce the order of a sequence of objects based on a minimal set of conditions. 3.) BOOLEAN EXPRESSIONS is a boolean expressions task that evaluates whether a language model has learned the rules of deductive reasoning, i.e., formal (zeroth-order) logic associated with the words \"and,\" \"or,\" \"not,\" etc. 4.) DISAMBIGUATION QA is a disambiguation task that necessitates the model to select the person to whom the pronoun refers. 5.) PENGUINS describes a table of penguins and requires the model to answer questions about the penguins' attributes. 4. CODE GENERATION Tasks: We further evaluate the performance of our method by conducting experiments on four diverse standard coding tasks. These tasks encompass a range of programming challenges, including both basic humanwritten and crowd-sourced Python tasks found in the 1.) HUMANEVAL (Chen et al., 2021) and 2.) MBPP (Austin et al., 2021) datasets, as well as more challenging competition-level coding tasks from the 3.) APPS (Hendrycks et al., 2021) and 4.) CODECONTESTS (Li et al., 2022) datasets."
},
{
"figure_ref": [],
"heading": "2.) LOGICAL DEDUCTION is a logical deduction",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "A.3 Tools and Framework",
"publication_ref": [
"b8",
"b9",
"b6"
],
"table_ref": [],
"text": "For querying GPT-3.5-TURBO and CODE-DAVINCI-002 models (Chen et al., 2021), we use the api library provided by OpenAI6 . We use the official code provided for running VICUNA-13B model (Chiang et al., 2023). We run inference on VICUNA-13B models on single A100 gpus. For coding tasks, we use the outputs provided by CodeT (Chen et al., 2022), where models are zeroshot prompted with temperature=0.8, and top_p = 0.95. stopping criteria in Adaptive-Consistency are fast to run, and we use a single-core machine. For numerical integration, we use the Scipy library in Python."
},
{
"figure_ref": [],
"heading": "A.4 Test-Case Generation",
"publication_ref": [
"b6"
],
"table_ref": [],
"text": "For CODE GENERATION tasks, we generate test cases in a similar fashion to CodeT (Chen et al., 2022). Specifically, we prompt the model with function description and prompt for generation of assert statements. However, unlike CodeT, we limit ourselves to only 10 test cases, which are generated in 1-2 prompts to LLM, thus adding neglible effect on the code generation itself.\nDataset Statistics are presented in Table 3."
},
{
"figure_ref": [],
"heading": "B Results",
"publication_ref": [],
"table_ref": [
"tab_7",
"tab_5",
"tab_8"
],
"text": "We present the complete results with standard deviation in Table 5. For CODE GENERATION tasks, results are presented in Table 4 Further in Table 6 we show that improvements by Adaptive-Consistency are statistically significant across all datasets. We perform 2 sample t-test on 3 random seeds. While p-value of number of generations is much less than 0.05 (average: 1.5e-3), indicating that our method is significantly more efficient, the p-value accuracy is much larger than 0.05 (average: 0.50), indicating that the slight accuracy difference between baseline and our method is statistically insignificant."
},
{
"figure_ref": [
"fig_3"
],
"heading": "C Analysis",
"publication_ref": [],
"table_ref": [],
"text": "C.1 Adaptive-Consistency vs.\nSelf-Consistency For Equal Average Sample Costs\nIn Section 5.2, we demonstrate that Adaptive-Consistency achieve better accuracy over Self-Consistency when both are operating on same expected sample cost. In Figure 5 we show the complete results. Section 4.1 previously demonstrated that Adaptive-Consistency achieves comparable performance to Self-Consistency using fewer samples. In this section, we consider a scenario where Adaptive-Consistency and Self-Consistency operate with the same average number of samples. For each fixed sampling budget k of Self-Consistency, we contrast the performance of Adaptive-Consistency and Self-Consistency, where Adaptive-Consistency uses k samples on average, rather than consistently across all instances.\nFigure 3 provides a visual comparison of the performance of Adaptive-Consistency and Self-Consistency on GSM-8K: Adaptive-Consistency outperforms Self-Consistency in accuracy across all average sample costs. For example, when the average sample cost is 10, Adaptive-Consistency achieves approximately 3% higher accuracy on GSM-8K.\nThe success of Adaptive-Consistency can be attributed to its adaptive sampling strategy. By varying the number of samples based on the complexity of the instance-using more samples where a clear consensus is hard to reach and fewer where answers are consistent-Adaptive-Consistency manages to secure improved overall performance even when the average sample cost matches that of Self-Consistency."
},
{
"figure_ref": [],
"heading": "C.2 Stopping Criterias",
"publication_ref": [],
"table_ref": [],
"text": "This section follows from the main discussion in Section 5.3. We evaluate different stopping criterias for Adaptive-Consistency. We evaluate 6 different functions:\n1. RANDOM: randomly stopping with a probability p, 2.) 2. MAJORITY: stopping after the most common answer has a majority above a threshold, 3. ENTROPY: stopping after the entropy of answers is below a threshold, 4. BETA: The main stopping criteria used in Adaptive-Consistency, based on the Equation (2), 5. DIRICHLET: The stopping criteria, based on Equation (1)."
},
{
"figure_ref": [],
"heading": "CHINESE RESTAURANT PROCESS (CRP):",
"publication_ref": [],
"table_ref": [
"tab_9",
"tab_9"
],
"text": "The stopping criteria, which models probability as chinese restaurant process making no assumption on possible number of unique answers.\nFor comparison, we tune the C thresh in each case on the training set of GSM-8K dataset. Results are presented in Table 7. RANDOM and MAJORITY are inferior to BETA across all datasets and models. Further, while DIRICHLET and CRP are almost similar to BETA, they are relatively very slow. While Although, from Table 7, ENTROPY looks appears to be on par with BETA, in Figure 6, we show BETA The ∆ columns display the reduction in generations (Gen. Reduc.) and the difference in accuracy (Acc. Diff.) between Self-Consistency and Adaptive-Consistency. For CODECONTESTS, Self-Consistency uses 1000, APPS use 50, while HUMANEVAL and MBPP use 100 generations each.\nbeats ENTROPY given the same expected sampling cost.\nFinally, BETA has additional key advantages: BETA incorporates a measure of uncertainty, which makes it more robust to variations in data order, mitigates the influence of noise, and offers a quantitative measure of confidence in the majority outcome. Consider an extreme case where the first two generated solutions are identical. The majority voting strategy would instantly halt the process, potentially missing out on better solutions. In contrast, BETA will keep sampling as the confidence for stopping has not yet reached."
},
{
"figure_ref": [],
"heading": "C.3 Chinese Restaurant Process",
"publication_ref": [
"b48"
],
"table_ref": [],
"text": "In the DIRICHLET stopping criteria, we assume that the number of unique answers that can be generated by the LLM is known in advance (and equal to the number of unique answers in the current observation set). However, this assumption may not hold for datasets such as GSM-8K, where numerical answers are expected. The CHINESE RESTAURANT PROCESS (CRP) is a generalization of the DIRICH-LET process that addresses this limitation by not making any assumption on the number of unique answers.\nIn CRP, we consider a list of same answers as a cluster, denoted by c i , where i is the index of the cluster. Let n i be the number of elements in cluster c i , and n be the total number of elements across all clusters. The probability of a new answer belong-ing to an existing cluster c i is directly proportional to the size of the cluster, and is given by:\nP (c i ) = n i n + α ,(4)\nwhereas the probability that a new unseen answer will form a new cluster is given by:\nP (c new ) = α n + α , (5\n)\nwhere α is the concentration parameter, which parameterizes the probability of generating a new answer.\nOur goal is to calculate the probability that the current majority cluster in observations will remain the same even with more generations. The first task is to estimate the concentration parameter α. We use the approximation proposed by (West, 1992) to model the α as\np(α|k, n) ≈ G(a + k -1, b + γ + log(n)), (6)\nwhere k is the number of unique answers (clusters) in the current observation, n is the total number of answers, a and b are priors and both set equal to 1, and γ is Euler's constant and G(α; a + k -1, b + γ+log(n)) denotes the probability density function of the Gamma distribution with shape parameter a + k -1 and rate parameter b + γ + log(n).\nWe sample α multiple times (100), and for each sample, we run Monte-Carlo Simulation (1000 simulations) based on the CRP probability modeling.\nEach simulation starts from from current set of observations, and performed till 40 generations are sampled. The probability that the current majority cluster remains the majority is then given by:\nP (majority) = 1 N α N M CS Nα i=1 N M CS j=1 I(majority 40 n ),(7)\nwhere N α is the number of times we sample α, N M CS is the number of Monte-Carlo Simulations, and I(majority 40 n ) is an indicator function that equals 1 if the current majority remains the majority after 40 generations, and 0 otherwise."
},
{
"figure_ref": [],
"heading": "C.4 Effect of Confidence Threshold on Adaptive-Consistency",
"publication_ref": [],
"table_ref": [],
"text": "We follow the discussion in Section 5.1, and present complete results on all datasets for CODE-DAVINCI-002."
},
{
"figure_ref": [],
"heading": "D Derivation of DIRICHLET stopping criteria",
"publication_ref": [],
"table_ref": [],
"text": "Consider for a given input (I), the model can generate one of m distinct answers A := {a 1 , a 2 , . . . a m }. Define the probability of generating an answer given input as p i := P (a i | I). Now, consider an observation set (O) with counts of each of a i as v i , such that m i=1 v i = n. Now, without loss of generality, consider p 1 > max m i=2 p i . Now, based on Equation (3), we need to find the probability:\nP (p 1 > m max i=2 p i | O) .\nHowever, here the p i s are latent variables, and only O is available to us. We next make the following Assumption 1: The vector ⃗ p = {p 1 , p 2 . . . p m } is sampled from uniform distribution over (m -1)simplex. Thus, p 1 = 1-m-1 i=1 p i . Since the observation set follows a multinomial distribution with parameters ⃗ p, conditional joint probability distribution of O given ⃗ p can be written as:\nP (O | ⃗ p) = n! m i=1 (v i !) m i=1 p v i i = Dir(v 1 +1, v 2 +1 . . . v m +1)\n, where Dir represents the dirichlet distribution with v i + 1, as its parameters. Applying Baye's Rule,\nP (⃗ p | O) = P (O | ⃗ p) • P (⃗ p) P (O)\n. Here P (O) is a normalizing constant and can be omitted for computation. From Assumption 1, since ⃗ p is sampled from uniform distribution,\nP (⃗ p) = m i=2 dp i\nThus conditional joint probability distribution of ⃗ p given O can be written as:\nP (⃗ p | O) = Dir(v 1 + 1, v 2 + 1, . . . , v m + 1) dp m dp m-1 . . . dp 2 (8)\nNow we can integrate the above equation over a subset of (m -1)-simplex, such that p 1 > max m i=2 p i . This gives us the equation:\nP (p 1 > m max i=2 p i | O) = 1 0 S(p ′ 1 ) P (⃗ p | O) dp 2 • • • dp m dp ′ 1 ,where\nS(p ′ 1 ) = {(p 2 , . . . , p m ) | p ′ 1 > m max i=2 p i , m i=2 p i = 1 -p ′ 1 }.\n(9) We note that the integration has no closed-form solution, and we use numerical approximation to compute the above integral.\nDefining region of integration: S(p ′ 1 ) Next, for computation of Equation ( 9), we need to precisely calculate the limits of each integration such that they represent the region S(p ′ 1 ). We do so by noting the following constraints on p i : 1.) The p i = 0 is valid ∀2 ≤ i ≤ m, 2.) Given {p m , p m-1 . . . p i+1 are fixed and in region S(p ′ 1 ),\np i < 1-m j=i+1 p j 2\nas else p i ≥ p 1 which is not allowed, 3.) Since p 1 > max m j=i+1 p j , so p i < 1 -m j=i+1 p j -max m j=i+1 p j , as else the ⃗ p, will lie outside the (m -1)-simplex, which is invalid. The first condition makes the lower limit for each integration 0, and the minimum of condition 2 and condition 3 gives the upper bound (limit) on each of the integrations.\nBETA stopping criteria Due to m -1 dimensions integrations involved, with m often getting larger than 10, computing Equation ( 9) is not efficient. Instead, we observe that establishing the majority of p 1 over the next largest probability, p 2 , is sufficient for our purpose. Then, pdf simplifies to BETA distribution with parameters v 1 +1, v 2 +1, and Equation ( 9) simplifies to:\n0.5 0 p v2 2 • (1 -p 2 ) v 1 dp 2(10)\nWe use Scipy library in Python numerically compute the above equation. "
},
{
"figure_ref": [],
"heading": "Acknowledgements",
"publication_ref": [],
"table_ref": [],
"text": "We thank the anonymous reviewers for their useful comments and suggestions. Mausam is supported by grants from Microsoft, Google and Verisk, Wipro CoE on generative AI, Yardi School of AI travel funds, and the Jai Gupta chair fellowship by IIT Delhi. We thank the IIT Delhi HPC facility for its computational resources. This work was"
}
] | A popular approach for improving the correctness of output from large language models (LLMs) is Self-Consistency -poll the LLM multiple times and output the most frequent solution. Existing Self-Consistency techniques always generate a constant number of samples per question, where a better approach will be to non-uniformly distribute the available budget based on the amount of agreement in the samples generated so far. In response, we introduce Adaptive-Consistency, a cost-efficient, model-agnostic technique that dynamically adjusts the number of samples per question using a lightweight stopping criterion. Our experiments over 17 reasoning and code generation datasets and three LLMs demonstrate that Adaptive-Consistency reduces sample budget by up to 7.9 times with an average accuracy drop of less than 0.1%. | Let's Sample Step by Step: Adaptive-Consistency for Efficient Reasoning and Coding with LLMs | [
{
"figure_caption": "Figure 1 :1Figure 1: An overview of Adaptive-Consistency: Self-Consistency samples a predetermined number of answers, whereas Adaptive-Consistency iteratively samples until a lightweight Stopping Criteria, decides to report the majority answer. The figure demonstrates an example where Adaptive-Consistency reduces sampling costs by 4x, requiring only ten samples to report the majority answer. The bottom-left graph contrasts Adaptive-Consistency with Self-Consistency across three reasoning categories, showing an average sample budget reduction of 3.3× with a negligible 0.04% drop in accuracy.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure2: Impact of Confidence Threshold (C thresh ) on Adaptive-Consistency for GSM-8K: As C thresh varies, the accuracy of Adaptive-Consistency increases gradually, eventually plateauing. Initially, the average number of generations also increases gradually but then sharply climbs, reflecting the accuracy-confidence trade-off.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: Comparison of BETA, ENTROPY and MA-JORITY stopping criterias. BETA consistently beats EN-TROPY and MAJORITY in terms of accuracy for the same sampling cost.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 :5Figure 5: Comparison of Adaptive-Consistency with Self-Consistency on various average sampling costs. Adaptive-Consistency is able to consistently beat Self-Consistency, especially when the sampling cost is low. Moreover, C thresh = 0.95 is a good indication of saturation in accuracy indicating the value works out-of-box for most configurations considered.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Figure 7 :7Figure 7: Impact of Confidence Threshold (C thresh ) on Adaptive-Consistency: As C thresh varies, the accuracy of Adaptive-Consistency increases gradually, eventually plateauing. Initially, the average number of generations also increases gradually but then sharply climbs, reflecting the accuracy-confidence trade-off. The trend is observed almost consistently across all datasets.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "7",
"figure_type": "figure"
},
{
"figure_caption": "Main results:",
"figure_data": "AccuracyNum. Generations∆CategorySelf-Consistency Adaptive-Consistency Self-Consistency Adaptive-Consistency Num. Gen. Acc.MATH73.273.14013.83.3×-0.1COMMONSENSE66.065.84015.82.9×-0.2SYMBOLIC Reasoning72.872.84013.13.8×+0.0CODE GENERATION35.235.6312.5173.62.4×+0.4AccuracyNum. Generations∆ModelSelf-Consistency Adaptive-Consistency Self-Consistency Adaptive-Consistency Num. Gen. Acc.GPT-3.5-TURBO76.476.24010.04.4×-0.2VICUNA-13B54.054.14021.71.9×+0.0CODE-DAVINCI-00269.769.8104.149.43.4×+0.0",
"figure_id": "tab_1",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Comparison of Adaptive-Consistency with Self-Consistency on 4 diverse code generation datasets. The table presents the accuracy of Self-Consistency, the average number of generations (Avg. Gen.) for Adaptive-Consistency, and the accuracy of Adaptive-Consistency. Self-Consistency always draws a fixed number of samples.",
"figure_data": "Self-ConsistencyAdaptive-Consistency∆ModelAvg. Gen. Accuracy Avg. Gen. Accuracy Gen. Reduc. Acc. Diff. ↑CODE-DAVINCI-00210061.423.663.44.3×+2.0HUMANEVALINCODER-6B10019.551.220.12.0×+0.6CODEGEN-16B10034.154.736.01.8×+1.9CODE-DAVINCI-00210064.436.363.92.8×-0.5MBPPINCODER-6B10030.753.830.91.9×+0.2CODEGEN-16B10049.657.850.41.7×+0.8APPSCODE-DAVINCI-0025011.944.411.91.1×0.0CODECONTESTS CODE-DAVINCI-00210003.0590.23.01.6×0.0",
"figure_id": "tab_5",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Comparison of Adaptive-Consistency with Self-Consistency on 17 diverse coding & reasoning datasets. Self-Consistency always draws 40 samples. The table shows accuracy, average generations (Avg. Gen.). The ∆ columns display reductions in generations (Gen. Reduc.) and accuracy (Acc. Diff.) between Self-Consistency and Adaptive-Consistency. Adaptive-Consistency achieves a 3.2× reduction in sample budget (Gen. Reduc.) with minimal average accuracy drop of 0.07% (Acc. Diff.).",
"figure_data": "DatasetModelP-Value (Accuracy) P-Value (Num Gens)GSM-8KVICUNA-13B0.50.0056GSM-8KCODE-DAVINCI-0020.422.09E-06SVAMPVICUNA-13B0.073.47E-06SVAMPCODE-DAVINCI-0020.427.40E-06ASDIVVICUNA-13B10.0005ASDIVCODE-DAVINCI-00210.0023DATE UNDERSTANDINGVICUNA-13B0.0577.68E-05DATE UNDERSTANDINGCODE-DAVINCI-0020.044.63E-05TRACKING SHUFFLED OBJECTS VICUNA-13B0.50.00002TRACKING SHUFFLED OBJECTS CODE-DAVINCI-0020.679.88E-06LOGICAL DEDUCTIONVICUNA-13B0.50.0007LOGICAL DEDUCTIONCODE-DAVINCI-002-0.0016STRATEGYQAVICUNA-13B0.901.16E-05STRATEGYQACODE-DAVINCI-0020.240.0005BOOLEAN EXPRESSIONSVICUNA-13B0.328.52E-05BOOLEAN EXPRESSIONSCODE-DAVINCI-002-4.98E-06SNARKSVICUNA-13B0.180.0007SNARKSCODE-DAVINCI-00210.0001RUIN NAMESVICUNA-13B-0.0049RUIN NAMESCODE-DAVINCI-0028.72E-06SALIENT TRANSLATIONVICUNA-13B0.180.0211SALIENT TRANSLATIONCODE-DAVINCI-00210.0001DISAMBIGUATION QAVICUNA-13B0.530.0015DISAMBIGUATION QACODE-DAVINCI-0020.420.0002PENGUINSVICUNA-13B0.180.0009PENGUINSCODE-DAVINCI-0020.427.79E-05Average0.5030.0002",
"figure_id": "tab_7",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "P-values using 2 sample t-test over 3 seeds on multiple datasets and models. The p-value for 'number of generations' is significantly less than 0.05 (average: 1.5e-3), confirming our method's efficiency, while the p-value for accuracy is much larger than 0.05 (average: 0.50), indicating that the slight accuracy difference is statistically insignificant.",
"figure_data": "",
"figure_id": "tab_8",
"figure_label": "6",
"figure_type": "table"
},
{
"figure_caption": "Comparison of BETA, MAJORITY and ENTROPY stopping criterias. In the two representative datasets, BETA consistently beats ENTROPY and MAJORITY for the same sampling cost. This shows in practice BETA performs better than both for the desirable range of accuracy and sampling cost. Comparison of various Stopping Criterias in Adaptive-Consistency. In general, BETA outperforms RANDOM and MAJORITY by decent margins across all datasets. BETA has comparable performance to DIRICHLET, but the latter is much slower. ENTROPY performs similarly to BETA but lacks human-interpretable stopping rationale.",
"figure_data": "0.81Accuracy vs Number of Generations -GSM-8K Beta Entropy Majority0.78 0.79Accuracy vs Number of Generations -strategy_qa Beta Entropy Majority0.80Accuracy0.79Accuracy0.770.780.760.770.756810 Number of Generations 12 1416182468 Number of Generations 10 12141618(a) GSM-8K(b) STRATEGYQAFigure 6: RANDOMMAJORITYENTROPYBETA (Adaptive-Consistency)DIRICHLETCRPAverage ↓ GenerationsAccuracy ↑Average ↓ GenerationsAccuracy ↑Average ↑ GenerationsAccuracy ↑Average ↑ GenerationsAccuracy ↑Average ↑ GenerationsAccuracy ↑Average ↑ GenerationsAccuracy ↑GSM-8KVICUNA-13B26.030.128.731.526.331.526.831.528.231.725.631.5CODE-DAVINCI-00213.876.916.680.915.381.013.881.015.281.113.281.1ASDIVVICUNA-13B28.063.214.863.715.863.916.564.017.764.016.964.0CODE-DAVINCI-00213.881.99.283.111.583.310.083.210.783.110.783.1SVAMPVICUNA-13B28.061.317.162.517.762.618.862.819.762.918.262.8CODE-DAVINCI-00213.483.38.484.810.785.19.585.010.385.19.885.0DATE UNDERSTANDINGVICUNA-13B28.058.315.359.516.059.917.360.218.559.916.959.9CODE-DAVINCI-00213.276.49.778.711.679.910.779.511.980.510.779.8TRACKING SHUFFLED OBJECTSVICUNA-13B27.931.815.033.018.432.020.332.023.332.019.631.8CODE-DAVINCI-00213.576.37.076.811.576.99.777.111.577.210.277.1LOGICAL DEDUCTIONVICUNA-13B27.950.512.951.215.851.418.151.420.951.218.351.4CODE-DAVINCI-00213.788.35.989.610.189.68.589.410.289.29.389.4STRATEGYQAVICUNA-13B28.165.111.765.514.565.816.365.818.765.817.065.7CODE-DAVINCI-00213.476.67.277.814.978.511.978.814.578.911.478.9BOOLEAN EXPRESSIONSVICUNA-13B27.678.010.476.814.878.316.278.419.178.817.078.5CODE-DAVINCI-00213.193.44.394.38.294.56.694.58.294.57.994.4SNARKSVICUNA-13B28.470.318.172.120.373.023.273.625.873.622.973.6CODE-DAVINCI-00213.671.610.574.012.174.012.774.014.273.412.373.2RUIN NAMESVICUNA-13B28.340.630.443.931.943.733.843.634.043.632.044.0CODE-DAVINCI-00213.871.717.577.718.678.117.278.017.676.816.478.1SALIENT TRANSLATIONVICUNA-13B24.927.724.628.526.328.028.728.729.428.826.928.9CODE-DAVINCI-00214.062.59.964.713.164.311.864.313.764.111.764.3DISAMBIGUATION QAVICUNA-13B27.962.918.363.520.163.122.863.525.463.922.163.3CODE-DAVINCI-00213.772.110.473.915.974.913.575.116.375.213.275.2PENGUINSVICUNA-13B27.945.619.746.320.747.322.947.325.147.322.147.3CODE-DAVINCI-00213.381.49.083.313.183.811.084.012.984.011.084.5",
"figure_id": "tab_9",
"figure_label": "7",
"figure_type": "table"
}
] | Pranjal Aggarwal; Aman Madaan; Yiming Yang; Mausam | [
{
"authors": "Daman Arora; Himanshu Gaurav Singh; Mausam ",
"journal": "",
"ref_id": "b0",
"title": "Have llms advanced enough? a challenging problem solving benchmark for large language models",
"year": "2023"
},
{
"authors": "Jacob Austin; Augustus Odena; Maxwell Nye; Maarten Bosma; Henryk Michalewski; David Dohan; Ellen Jiang; Carrie J Cai; Michael Terry; Quoc V Le; Charles Sutton",
"journal": "",
"ref_id": "b1",
"title": "Program synthesis with large language models",
"year": "2021"
},
{
"authors": "Yoshua Bengio; Jean-Sébastien Senecal",
"journal": "IEEE Transactions on Neural Networks",
"ref_id": "b2",
"title": "Adaptive importance sampling to accelerate training of a neural probabilistic language model",
"year": "2008"
},
{
"authors": "Lorenz Berger; Eoin R Hyde; M Jorge Cardoso; Sébastien Ourselin",
"journal": "",
"ref_id": "b3",
"title": "An adaptive sampling scheme to efficiently train fully convolutional networks for semantic segmentation",
"year": "2017"
},
{
"authors": "Jonathan Bragg; Daniel S Mausam; Weld",
"journal": "ACM",
"ref_id": "b4",
"title": "Optimal testing for crowd workers",
"year": "2016-05-09"
},
{
"authors": "B Tom; Benjamin Brown; Nick Mann; Melanie Ryder; Jared Subbiah; Prafulla Kaplan; Arvind Dhariwal; Pranav Neelakantan; Girish Shyam; Amanda Sastry; Sandhini Askell; Ariel Agarwal; Gretchen Herbert-Voss; Tom Krueger; Rewon Henighan; Aditya Child; Daniel M Ramesh; Jeffrey Ziegler; Clemens Wu; Christopher Winter; Mark Hesse; Eric Chen; Mateusz Sigler; Scott Litwin; Benjamin Gray; Jack Chess; Christopher Clark; Sam Berner; Alec Mccandlish; Ilya Radford; Dario Sutskever; Amodei",
"journal": "",
"ref_id": "b5",
"title": "Language models are few-shot learners",
"year": "2020"
},
{
"authors": "Bei Chen; Fengji Zhang; A Nguyen; Daoguang Zan; Zeqi Lin; Jian-Guang Lou; Weizhu Chen",
"journal": "",
"ref_id": "b6",
"title": "Codet: Code generation with generated tests",
"year": "2022"
},
{
"authors": "Lingjiao Chen; Matei A Zaharia; James Y Zou",
"journal": "",
"ref_id": "b7",
"title": "Frugalgpt: How to use large language models while reducing cost and improving performance",
"year": "2023"
},
{
"authors": "Mark Chen; Jerry Tworek; Heewoo Jun; Qiming Yuan; Henrique Ponde; Jared Kaplan; Harrison Edwards; Yura Burda; Nicholas Joseph; Greg Brockman; Alex Ray; Raul Puri; Gretchen Krueger; Michael Petrov; Heidy Khlaaf; Girish Sastry; Pamela Mishkin; Brooke Chan; Scott Gray; Nick Ryder; Mikhail Pavlov; Alethea Power; Lukasz Kaiser; Mohammad Bavarian; Clemens Winter; Philippe Tillet; Felipe Petroski Such; David W Cummings; Matthias Plappert; Fotios Chantzis; Elizabeth Barnes; Ariel Herbert-Voss; William H Guss; Alex Nichol; Igor Babuschkin; S Arun Balaji; Shantanu Jain; Andrew Carr; Jan Leike; Joshua Achiam; Vedant Misra; Evan Morikawa; Alec Radford; Matthew M Knight; Miles Brundage; Mira Murati; Katie Mayer; Peter Welinder; Bob Mcgrew; Dario Amodei; Sam Mccandlish; Ilya Sutskever; Wojciech Zaremba",
"journal": "",
"ref_id": "b8",
"title": "Evaluating large language models trained on code",
"year": "2021"
},
{
"authors": "Wei-Lin Chiang; Zhuohan Li; Zi Lin; Ying Sheng; Zhanghao Wu; Hao Zhang; Lianmin Zheng; Siyuan Zhuang; Yonghao Zhuang; Joseph E Gonzalez; Ion Stoica; Eric P Xing",
"journal": "",
"ref_id": "b9",
"title": "Vicuna: An opensource chatbot impressing gpt-4 with 90%* chatgpt quality",
"year": "2023"
},
{
"authors": "Karl Cobbe; Vineet Kosaraju; Mohammad Bavarian; Jacob Hilton; Reiichiro Nakano; Christopher Hesse; John Schulman",
"journal": "",
"ref_id": "b10",
"title": "Training verifiers to solve math word problems",
"year": "2021"
},
{
"authors": "C H Peng Dai; Lin; Daniel S Mausam; Weld",
"journal": "Artif. Intell",
"ref_id": "b11",
"title": "Pomdp-based control of workflows for crowdsourcing",
"year": "2013"
},
{
"authors": "A ; Philip Dawid; Allan Skene",
"journal": "Journal of The Royal Statistical Society Series C-applied Statistics",
"ref_id": "b12",
"title": "Maximum likelihood estimation of observer error-rates using the em algorithm",
"year": "1979"
},
{
"authors": "Anhai Doan; Raghu Ramakrishnan; Alon Y Halevy",
"journal": "Communications of the ACM",
"ref_id": "b13",
"title": "Crowdsourcing systems on the world-wide web",
"year": "2011"
},
{
"authors": "Angela Fan; Edouard Grave; Armand Joulin",
"journal": "",
"ref_id": "b14",
"title": "Reducing transformer depth on demand with structured dropout",
"year": "2019"
},
{
"authors": "Luyu Gao; Aman Madaan; Shuyan Zhou; Uri Alon; Pengfei Liu; Yiming Yang; Jamie Callan; Graham Neubig",
"journal": "",
"ref_id": "b15",
"title": "Pal: Program-aided language models",
"year": "2022"
},
{
"authors": "Shijie Geng; Peng Gao; Zuohui Fu; Yongfeng Zhang",
"journal": "",
"ref_id": "b16",
"title": "Romebert: Robust training of multiexit bert",
"year": "2021"
},
{
"authors": "Mor Geva; Daniel Khashabi; Elad Segal; Tushar Khot; Dan Roth; Jonathan Berant",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b17",
"title": "Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies",
"year": "2021"
},
{
"authors": "Dan Hendrycks; Steven Basart; Saurav Kadavath; Mantas Mazeika; Akul Arora; Ethan Guo; Collin Burns; Samir Puranik; Horace He; Dawn Xiaodong Song; Jacob Steinhardt",
"journal": "",
"ref_id": "b18",
"title": "Measuring coding challenge competence with apps",
"year": "2021"
},
{
"authors": "Lu Hou; Zhiqi Huang; Lifeng Shang; Xin Jiang; Qun Liu",
"journal": "",
"ref_id": "b19",
"title": "Dynabert: Dynamic bert with adaptive width and depth",
"year": "2020"
},
{
"authors": "M Siddhant; Razvan Jayakumar; Jack W Pascanu; Simon Rae; Erich Osindero; Elsen",
"journal": "",
"ref_id": "b20",
"title": "Top-kast: Top-k always sparse training",
"year": "2021"
},
{
"authors": "Takeshi Kojima; Shane Shixiang; Machel Gu; Yutaka Reid; Yusuke Matsuo; Iwasawa",
"journal": "",
"ref_id": "b21",
"title": "Large Language Models are Zero-Shot Reasoners",
"year": "2022"
},
{
"authors": "Aitor Lewkowycz; Anders Andreassen; David Dohan; Ethan Dyer; Henryk Michalewski; Vinay Ramasesh; Ambrose Slone; Cem Anil; Imanol Schlag; Theo Gutman-Solo",
"journal": "",
"ref_id": "b22",
"title": "Solving quantitative reasoning problems with language models",
"year": "2022"
},
{
"authors": "Yujia Li; David H Choi; Junyoung Chung; Nate Kushman; Julian Schrittwieser; Rémi Leblond; Tom; James Eccles; Felix Keeling; Agustin Dal Gimeno; Thomas Lago; Peter Hubert; Cyprien Choy; De; Igor Masson D'autume; Xinyun Babuschkin; Po-Sen Chen; Johannes Huang; Sven Welbl; Gowal; Alexey; James Cherepanov; Daniel Jaymin Molloy; Esme Mankowitz; Pushmeet Sutherland Robson; Nando Kohli; De; Koray Freitas; Oriol Kavukcuoglu; Vinyals",
"journal": "Science",
"ref_id": "b23",
"title": "Competition-level code generation with alphacode",
"year": "2022"
},
{
"authors": "Christopher H Lin; Daniel S Mausam; Weld",
"journal": "AUAI Press",
"ref_id": "b24",
"title": "Crowdsourcing control: Moving beyond multiple choice",
"year": "2012-08-14"
},
{
"authors": "Weijie Liu; Peng Zhou; Zhiruo Wang; Zhe Zhao; Haotang Deng; Qi Ju",
"journal": "",
"ref_id": "b25",
"title": "Fastbert: a selfdistilling bert with adaptive inference time",
"year": "2020"
},
{
"authors": "Aman Madaan; Pranjal Aggarwal; Ankit Anand; Pranavi Srividya; Swaroop Potharaju; Pei Mishra; Aditya Zhou; Dheeraj Gupta; Karthik Rajagopal; Yiming Kappaganthu; Shyam Yang; Upadhyay; Manaal Mausam; Faruqui",
"journal": "",
"ref_id": "b26",
"title": "Automix: Automatically mixing language models",
"year": "2023"
},
{
"authors": "Aman Madaan; Alexander Shypula; Uri Alon; Milad Hashemi; Parthasarathy Ranganathan; Yiming Yang; Graham Neubig; Amir Yazdanbakhsh",
"journal": "",
"ref_id": "b27",
"title": "Learning performance-improving code edits",
"year": "2023"
},
{
"authors": "Aman Madaan; Yiming Yang",
"journal": "",
"ref_id": "b28",
"title": "Flowgen: Fast and slow graph generation",
"year": "2022"
},
{
"authors": "Chao-Chun Shen-Yun Miao; Keh-Yih Liang; Su",
"journal": "Association for Computational Linguistics",
"ref_id": "b29",
"title": "A diverse corpus for evaluating and developing English math word problem solvers",
"year": "2020"
},
{
"authors": "Oyvind Bhavana Dalvi Mishra; Peter Tafjord; Clark",
"journal": "",
"ref_id": "b30",
"title": "Towards teachable reasoning systems: Using a dynamic memory of user feedback for continual system improvement",
"year": "2022"
},
{
"authors": "Aditya G Parameswaran; Shreya Shankar; Parth Asawa; Naman Jain; Yujie Wang",
"journal": "",
"ref_id": "b31",
"title": "Revisiting prompt engineering via declarative crowdsourcing",
"year": "2023"
},
{
"authors": "Arkil Patel; S Bhattamishra; Navin Goyal",
"journal": "",
"ref_id": "b32",
"title": "Are nlp models really able to solve simple math word problems?",
"year": "2021"
},
{
"authors": "Ameya Prabhu; Charles Dognin; Maneesh Kumar Singh",
"journal": "",
"ref_id": "b33",
"title": "Sampling bias in deep active classification: An empirical study",
"year": "2019"
},
{
"authors": "Alexander J Quinn; Benjamin B Bederson",
"journal": "",
"ref_id": "b34",
"title": "Human computation: a survey and taxonomy of a growing field",
"year": "2011"
},
{
"authors": "Shipeng Vikas Chandrakant Raykar; Linda H Yu; Gerardo Zhao; Charles Hermosillo; Luca Florin; Linda Bogoni; Moy",
"journal": "J. Mach. Learn. Res",
"ref_id": "b35",
"title": "Learning from crowds",
"year": "2010"
},
{
"authors": "Tal Schuster; Adam Fisch; Jai Gupta; Mostafa Dehghani; Dara Bahri; Yi Vinh Q Tran; Donald Tay; Metzler",
"journal": "",
"ref_id": "b36",
"title": "Confident adaptive language modeling",
"year": "2022"
},
{
"authors": "Tal Schuster; Adam Fisch; Tommi Jaakkola; Regina Barzilay",
"journal": "",
"ref_id": "b37",
"title": "Consistent accelerated inference via confident adaptive transformers",
"year": "2021"
},
{
"authors": "Avi Schwarzschild; Eitan Borgnia; Arjun Gupta; Arpit Bansal; Zeyad Emam; Furong Huang; Micah Goldblum; Tom Goldstein",
"journal": "",
"ref_id": "b38",
"title": "Datasets for Studying Generalization from Easy to Hard Examples",
"year": "2021"
},
{
"authors": "Avi Schwarzschild; Eitan Borgnia; Arjun Gupta; Furong Huang; Uzi Vishkin; Micah Goldblum; Tom Goldstein",
"journal": "",
"ref_id": "b39",
"title": "Can You Learn an Algorithm? Generalizing from Easy to Hard Problems with Recurrent Networks",
"year": "2021"
},
{
"authors": "Curran Associates; Inc ",
"journal": "",
"ref_id": "b40",
"title": "",
"year": ""
},
{
"authors": "Rohan Taori; Ishaan Gulrajani; Tianyi Zhang; Yann Dubois; Xuechen Li; Carlos Guestrin; Percy Liang; Tatsunori B Hashimoto",
"journal": "",
"ref_id": "b41",
"title": "Stanford alpaca: An instruction-following llama model",
"year": "2023"
},
{
"authors": "Hugo Touvron; Thibaut Lavril; Gautier Izacard; Xavier Martinet; Marie-Anne Lachaux; Timothée Lacroix; Baptiste Rozière; Naman Goyal; Eric Hambro; Faisal Azhar; Armand Aur'elien Rodriguez; Edouard Joulin; Guillaume Grave; Lample",
"journal": "",
"ref_id": "b42",
"title": "Llama: Open and efficient foundation language models",
"year": "2023"
},
{
"authors": "Xuezhi Wang; Jason Wei; Dale Schuurmans; Quoc Le; Ed Chi; Denny Zhou",
"journal": "",
"ref_id": "b43",
"title": "Rationale-Augmented Ensembles in Language Models",
"year": "2022"
},
{
"authors": "Xuezhi Wang; Jason Wei; Dale Schuurmans; Quoc Le; Ed Huai Hsin; Chi ; Denny Zhou",
"journal": "",
"ref_id": "b44",
"title": "Selfconsistency improves chain of thought reasoning in language models",
"year": "2022"
},
{
"authors": "Jason Wei; Xuezhi Wang; Dale Schuurmans; Maarten Bosma; Ed Huai Hsin Chi; F Xia; Quoc Le; Denny Zhou",
"journal": "",
"ref_id": "b45",
"title": "Chain of thought prompting elicits reasoning in large language models",
"year": "2022"
},
{
"authors": " Daniel S Weld; Christopher H Mausam; Jonathan Lin; Bragg",
"journal": "MIT Press",
"ref_id": "b46",
"title": "Artificial intelligence and collective intelligence",
"year": "2015"
},
{
"authors": "Peter Welinder; Steve Branson; Serge J Belongie; Pietro Perona",
"journal": "",
"ref_id": "b47",
"title": "The multidimensional wisdom of crowds",
"year": "2010"
},
{
"authors": "Beate West",
"journal": "",
"ref_id": "b48",
"title": "Hyperparameter estimation in dirichlet process mixture models",
"year": "1992"
},
{
"authors": "Jacob Whitehill; Paul Ruvolo; Tingfan Wu; Jacob Bergsma; Javier R Movellan",
"journal": "",
"ref_id": "b49",
"title": "Whose vote should count more: Optimal integration of labels from labelers of unknown expertise",
"year": "2009"
},
{
"authors": "Ji Xin; Raphael Tang; Jaejun Lee; Yaoliang Yu; Jimmy J Lin",
"journal": "",
"ref_id": "b50",
"title": "Deebert: Dynamic early exiting for accelerating bert inference",
"year": "2020"
},
{
"authors": "Qunliang Xing; Mai Xu; Tianyi Li; Zhenyu Guan",
"journal": "",
"ref_id": "b51",
"title": "Early exit or not: Resource-efficient blind quality enhancement for compressed images",
"year": "2020"
},
{
"authors": "Fuzhao Xue; Valerii Likhosherstov; Anurag Arnab; Neil Houlsby; Mostafa Dehghani; Yang You",
"journal": "",
"ref_id": "b52",
"title": "Adaptive computation with elastic input sequence",
"year": "2023"
},
{
"authors": "Wangchunshu Zhou; Canwen Xu; Tao Ge; Julian Mcauley; Ke Xu; Furu Wei",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b53",
"title": "Bert loses patience: Fast and robust inference with early exit",
"year": "2020"
}
] | [
{
"formula_coordinates": [
3,
347.73,
502.96,
134.59,
20.96
],
"formula_id": "formula_0",
"formula_text": "P (p 1 > m max i=2 p i | v) > C thresh"
},
{
"formula_coordinates": [
4,
73.38,
90.85,
186.66,
86.11
],
"formula_id": "formula_1",
"formula_text": "P (p 1 > m max i=2 p i | V ) = 1 0 S(p ′ 1 ) f (p ′ 1 , p 2 , . . . , p m | V ). dp 2 • • • dp m dp ′ 1 ,where"
},
{
"formula_coordinates": [
4,
73.9,
181.73,
215.52,
57.28
],
"formula_id": "formula_2",
"formula_text": "S(p ′ 1 ) = {(p 2 , . . . , p m ) | p ′ 1 > m max i=2 p i , m i=2 p i = 1 -p ′ 1 }."
},
{
"formula_coordinates": [
4,
131.7,
672.46,
158.17,
28.57
],
"formula_id": "formula_3",
"formula_text": "0.5 0 p v2 2 • (1 -p 2 ) v 1 dp 2 (2)"
},
{
"formula_coordinates": [
4,
358.56,
456.79,
166.58,
33.71
],
"formula_id": "formula_4",
"formula_text": "t t j=1 P (p j 1 > m max i=2 p j i | V )(3)"
},
{
"formula_coordinates": [
15,
379.29,
382.13,
145.85,
24.43
],
"formula_id": "formula_5",
"formula_text": "P (c i ) = n i n + α ,(4)"
},
{
"formula_coordinates": [
15,
373.05,
449.53,
147.85,
24.43
],
"formula_id": "formula_6",
"formula_text": "P (c new ) = α n + α , (5"
},
{
"formula_coordinates": [
15,
520.9,
457.26,
4.24,
9.46
],
"formula_id": "formula_7",
"formula_text": ")"
},
{
"formula_coordinates": [
15,
313.68,
617.27,
211.46,
9.81
],
"formula_id": "formula_8",
"formula_text": "p(α|k, n) ≈ G(a + k -1, b + γ + log(n)), (6)"
},
{
"formula_coordinates": [
16,
70.87,
148.67,
225.71,
45.93
],
"formula_id": "formula_9",
"formula_text": "P (majority) = 1 N α N M CS Nα i=1 N M CS j=1 I(majority 40 n ),(7)"
},
{
"formula_coordinates": [
16,
70.87,
520.3,
154.37,
38.78
],
"formula_id": "formula_10",
"formula_text": "P (p 1 > m max i=2 p i | O) ."
},
{
"formula_coordinates": [
16,
70.87,
705.21,
270.38,
33.71
],
"formula_id": "formula_11",
"formula_text": "P (O | ⃗ p) = n! m i=1 (v i !) m i=1 p v i i = Dir(v 1 +1, v 2 +1 . . . v m +1)"
},
{
"formula_coordinates": [
16,
351.41,
86.21,
126.53,
24.43
],
"formula_id": "formula_12",
"formula_text": "P (⃗ p | O) = P (O | ⃗ p) • P (⃗ p) P (O)"
},
{
"formula_coordinates": [
16,
381.6,
169.74,
66.84,
33.71
],
"formula_id": "formula_13",
"formula_text": "P (⃗ p) = m i=2 dp i"
},
{
"formula_coordinates": [
16,
317,
255.51,
208.14,
27.17
],
"formula_id": "formula_14",
"formula_text": "P (⃗ p | O) = Dir(v 1 + 1, v 2 + 1, . . . , v m + 1) dp m dp m-1 . . . dp 2 (8)"
},
{
"formula_coordinates": [
16,
308.65,
358.57,
184.27,
86.11
],
"formula_id": "formula_15",
"formula_text": "P (p 1 > m max i=2 p i | O) = 1 0 S(p ′ 1 ) P (⃗ p | O) dp 2 • • • dp m dp ′ 1 ,where"
},
{
"formula_coordinates": [
16,
309.17,
449.46,
215.52,
57.28
],
"formula_id": "formula_16",
"formula_text": "S(p ′ 1 ) = {(p 2 , . . . , p m ) | p ′ 1 > m max i=2 p i , m i=2 p i = 1 -p ′ 1 }."
},
{
"formula_coordinates": [
16,
306.14,
661.55,
84.43,
19.94
],
"formula_id": "formula_17",
"formula_text": "p i < 1-m j=i+1 p j 2"
},
{
"formula_coordinates": [
17,
131.7,
190.68,
158.16,
28.58
],
"formula_id": "formula_18",
"formula_text": "0.5 0 p v2 2 • (1 -p 2 ) v 1 dp 2(10)"
}
] |
10.18653/v1/2020.emnlp-main.19 | 2023-10-21 | [
{
"figure_ref": [
"fig_0",
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b7",
"b12",
"b26",
"b76",
"b86",
"b5",
"b29",
"b25",
"b56",
"b74",
"b24",
"b56",
"b74"
],
"table_ref": [],
"text": "Large Language Models (LLMs), including Chat-GPT1 and Bard2 , have exhibited exceptional performance across a range of natural language processing (NLP) tasks and amassed a significant user base (Brown et al., 2020;Chowdhery et al., 2022;OpenAI, 2023). As performance gains are brought from the increases in model size (Kaplan et al., 2020;Wei et al., 2022;Zhao et al., 2023), LLMs are becoming larger and larger. However, the computational cost of inference is a severe bottleneck of many practical applications, especially when the number of parameters in an LLM is massive (Bender et al., 2021;Kraus et al., 2023).\nMeanwhile, LLMs are also used for local sequence transduction tasks, such as paraphrasing, formality style transfer, Grammatical Error Correction (GEC), and simplification (Kaneko et al., 2022;Reif et al., 2022;Wu et al., 2023a;Wang et al., 2022;Kaneko and Okazaki, 2023), where only a small portion of the source text is edited. Most tokens in a source text are kept unchanged in these tasks. For example, the source text, \"Many years ago, the situation is different,\" and the target text, \"Many years ago, the situation was different,\" of the GEC task mostly share the common tokens except for the underlined tokens (is and was).\nExisting methods of downstream tasks do not make use of the characteristics of local sequence transduction (Reif et al., 2022;Wu et al., 2023a;Wang et al., 2022), simply generating all target tokens. In this paper, we hypothesize that this treatment is disadvantageous in achieving high performance in terms of task accuracy and computational time. More specifically, it is inefficient to generate unchanged tokens (e.g. Many, years, ago, the, situation, different) in the previous example because the model must copy many source tokens only to increase the length of the target sequence.\nThis study proposes to predict a set of edit spans, which represent the changed parts of the target text relative to the source tokens. Omitting unedited tokens that occupy most of the target text, we can reduce the length of the target text and the inference time for local sequence transduction tasks. Figure 1 shows the process of creating a set of edit spans from source and target texts in GEC. First, we align tokens in the source and target texts to extract the edit locations and tokens and convert them into a set of edit spans. In the example shown in Figure 1, the edit spans (1, 1, \"the\"), (8, 9, \"have been\"), (12, 13, \"\") are created from the source text \"Through thousands of years, most Chinese scholars are greatly affected by the Confucianism.\" and the target text \"Through the thousands of years, most Chinese scholars have been greatly affected by Confucianism.\". LLMs are fine-tuned using pairs of source text and edit spans with the instructions.\nWe conducted experiments on four local sequence transduction tasks: paraphrasing, formality style transfer, GEC, and simplification. The proposed method achieved comparable performance to the baseline that directly outputs the target text. In these tasks, the proposed method could reduce the sequence length on the target side by 32% on average and by as small as 21% in GEC. Furthermore, the proposed method with task-specific fine-tuning achieved state-of-the-art (SoTA) performance in the four tasks."
},
{
"figure_ref": [],
"heading": "Edit Spans",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [
"fig_0",
"fig_0"
],
"heading": "Edit Span Extraction",
"publication_ref": [
"b18"
],
"table_ref": [],
"text": "To extract the editing locations and results of the source and target texts, we calculate the alignment between the tokens in each text. We use linguistical alignment, which incorporates linguistic information, to perform the alignment (Felice et al., 2016). Linguistical alignment is a method based on the Damerau-Levenshtein algorithm that aligns tokens by considering not only the distance between tokens but also the match of their lemma, part-of-speech, and character features, weighted accordingly. Taking into account the linguistic information of tokens, linguistical alignment is more accurate compared to alignment methods that only use surface information. Furthermore, linguistic alignment merges token alignments using recursive rules to create alignments for multiple tokens, such as \"have been\" in Figure 1.\nTo indicate the edit position identified by the alignment, a 0 is assigned before the first token of the source text, and an index is sequentially assigned to the space after each token. When the length of the source text is N , N is assigned after the last token. The edit span is represented by the tuple of the start position of the source text, the end position of the source text, and the token result after being edited.\nThere are three types of edit operations: insert, replace, and delete; we explain them using the example in Figure 1. The tuple (1, 1, \"the\") represents the operation to insert \"the\". In an insertion operation, both the start and end positions are set to the same position where a token is inserted in the source. The tuple stands for inserting \"the\" between the tokens located at the 1st position. The tuple (8, 9, \"have been\") presents the operation to replace \"are\" with \"have been\". By specifying the 8th and 9th positions of the source text, this tuple targets the \"are\" and rewrites them as \"have been\". The tuple (12, 13, \"\") represents the operation to delete \"the\". It points \"the\" by specifying the 12th and 13th positions in the source text. Because the target token after this edit operation is empty, this tuple corresponds to removing \"the\"."
},
{
"figure_ref": [],
"heading": "Instruction Tuning with Edit Spans",
"publication_ref": [
"b75",
"b47",
"b74"
],
"table_ref": [],
"text": "Instruction tuning fine-tunes LLMs by using natural language instructions describing a task (Wei et al., 2021). Compared to the conventional finetuning that specializes the model for a specific task, instruction tuning aims for generalization to various tasks by training LLMs to respond well to many kinds of instructions. Therefore, instruction tuning is used for training many LLMs in an openended setting (Ouyang et al., 2022;Chung et al., 2022;Wang et al., 2022;Wu et al., 2023b). We use the description of local sequence transduction tasks as instructions to perform instruction tuning of LLMs. We provide the LLMs with instructions and source text, and train the LLMs to generate edit spans. When there are multiple edits, they are concatenated with commas like \"1 1 the, 8 9 have been, 12 13\". When no editing is required in the source text, \"None\" is given as the gold text.\nRecent LLMs are expected to have the ability to handle unknown tasks and various tasks, to achieve generality. It is important that learning through edit spans does not degrade the performance of tasks other than local sequence transduction tasks. Therefore, we add edit span data to the existing training data for instruction tuning, which includes various tasks, and fine-tune LLMs."
},
{
"figure_ref": [],
"heading": "Conversion from Edit Spans to Output Text",
"publication_ref": [],
"table_ref": [],
"text": "To convert the edit spans output by LLMs into plaintext, we use a rule-based approach. If LLMs generate \"None\", we use the source text as the final output text. Otherwise, we split the edit spans by commas and extract the edits. From each edit, we extract the starting position, ending position, and edited result. If LLMs generate edits in an incorrect format that do not include start or end positions or edits where the start or end positions exceed the source text range, we ignore them. To ensure that the token indices do not shift, we apply the edits to the source text in descending order of starting positions. This conversion is implemented by simple rules with a minimal computational cost.\n3 Experiment Setting"
},
{
"figure_ref": [],
"heading": "Local Sequence Transduction Taskes",
"publication_ref": [
"b45",
"b44",
"b14",
"b33",
"b30",
"b43",
"b32"
],
"table_ref": [],
"text": "We conducted experiments on local sequence transduction tasks such as GEC, paraphrasing, formality style transfer, and simplification.\nGEC We used NUCLE as the training data, CoNLL2013 (Ng et al., 2013) as the development data, and CoNLL2014 (Ng et al., 2014) as the evalu-ation data. The dataset is comprised of essays composed by college students from the National University of Singapore, covering a broad spectrum of subjects, including environmental pollution, healthcare, and more. We used the M 2 score (Dahlmeier and Ng, 2012) as the evaluation metric. For GEC, we provide the instruction text \"Rewrite the input text into grammatically correct text.\".\nParaphrasing Quora published a dataset that includes more than 400K lines of potential question duplicate pairs3 . Of these pairs, 150K question pairs were labeled as paraphrases. Only those labeled paraphrase question pairs are used as training, development, and test sets. We used BLEU-4 (Papineni et al., 2002), ROUGE-1, and ROUGE-2 (Lin, 2004) to evaluate LLMs, following previous research (Kumar et al., 2020;Meng et al., 2021;Li et al., 2022). For paraphrasing, we provide the instruction text \"Rewrite the input text into paraphrased text.\""
},
{
"figure_ref": [],
"heading": "Style transfer",
"publication_ref": [
"b53",
"b6",
"b9",
"b87",
"b85",
"b1",
"b80",
"b80"
],
"table_ref": [],
"text": "We used FST benchmark Grammarly Yahoo Answers Corpus (GYAFC) (Rao and Tetreault, 2018) for formality style transfer. GYAFC is a plain corpus that contains pairs of informal and formal sentences conveying the same meaning. It covers domains such as Entertainment & Music (E&M) and Family & Relationship (F&R).\nWe utilized the corpus BLEU in NLTK (Bird and Loper, 2004) as described in Chawla and Yang (2020). For formality style transfer, we provide the instruction text \"Rewrite the input text into formal text.\"\nSimplification We used WikiSmall4 (Zhu et al., 2010;Zhang and Lapata, 2017) as the training data and ASSET (Alva-Manchego et al., 2020) and TurkCorpus (Xu et al., 2016) as the evaluation data. We used SARI (Xu et al., 2016) to evaluate LLMs, which compares the generated text with the target text and calculates the average F1 score for addition, keep, and deletion operations. For text simplification, we provide the instruction text \"Rewrite the input text into simpler text.\""
},
{
"figure_ref": [],
"heading": "Open-ended Tasks",
"publication_ref": [
"b47",
"b84",
"b36"
],
"table_ref": [],
"text": "The rules of edit spans differ from the rules in the raw text, which could potentially have a negative impact on the performance of tasks other than local sequence transduction. By combining open-ended instruction tuning data and edit spans instruction tuning data, we can train LLMs and investigate their impact on other tasks as well.\nWe utilize the databricks-dolly-15k dataset5 by randomly dividing it into 13K for training, 1K for development, and 1K for evaluation. databricksdolly-15k is a publicly available dataset consisting of instructional records created by numerous Databricks employees. It covers various behavioral categories described in InstructGPT (Ouyang et al., 2022), such as brainstorming, classification, closed QA, generation, information extraction, open QA, and summarization. We sampled 3K instances for each of the tasks: GEC, paraphrasing, style transfer, and simplification, resulting in a total of 12K instruction instances. We fine-tuned LLMs using a combined dataset of all these instructions, totaling 25K instances.\nWe used BERTScore6 (Zhang et al., 2019) as our evaluation metric. BERTScore is an evaluation method that measures the similarity between generated text and target text using contextual embeddings from pre-trained models. We utilized RoBERTa (Liu et al., 2019) (roberta-large7 ) as the BERTScore models."
},
{
"figure_ref": [],
"heading": "Instruction Tuning Settings",
"publication_ref": [
"b83",
"b64",
"b18"
],
"table_ref": [],
"text": "We used the following four LLMs for our experiments: MPT (mpt-7b)8 (Team, 2023), OPT (opt-6.7b)9 (Zhang et al., 2022), LLaMA (llama-7b) 10 (Touvron et al., 2023), and BLOOM (bloom-7b1) 11 (Scao et al., 2022).\nWe used the code for instruction tuning from Stanford Alpaca (Taori et al., 2023) code12 for instruction tuning. We set the number of epochs to 3 and used a batch size of 32. The learning rate was set to 2e-5, with a warmup rate of 0.03, and we employed a cosine learning rate schedule. These hyperparameters were determined following Stanford Alpaca. We report the average results of three models trained with different seeds for instruction tuning. We used four nodes, each containing eight NVIDIA A100 GPUs. We used the code 13 for linguistical alignment provided by Felice et al. (2016).\nBaselines We compare the results of the proposed method with the results of LLMs fine-tuned for instruction tuning using the target text as the ground truth instead of using edit spans. This comparison examines whether edit spans can reduce computational costs during inference without compromising performance."
},
{
"figure_ref": [],
"heading": "Experiment",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Performance on Local Sequence Transduction Tasks",
"publication_ref": [],
"table_ref": [
"tab_0"
],
"text": "To demonstrate the contribution of edit spans to performance improvement, we first compare the baseline performance with fine-tuned data using plain text. Table 1 shows the results of performance comparison between the baseline and the proposed method in the GEC, paraphrasing, style transfer, and simplification tasks. Out of 32 cases, performance improvement was observed in 19 cases, and edit spans contributed to the performance enhancement. Furthermore, it can be observed that the LLaMA trained with edit spans achieves the highest performance in most cases."
},
{
"figure_ref": [
"fig_1"
],
"heading": "Reducing Text Length",
"publication_ref": [],
"table_ref": [],
"text": "We examine how much the fine-tuning of LLMs with edit span data reduced the length of the output text. Figure 2 shows the ratio of output text length to target text length when fine-tuned with plain data and edit span data, respectively, on the development data for each task. The proposed method successfully compresses the output text across all tasks, independent of the model used; it achieves text compression in the range of 21% in the most compressed cases and 41% even in the least compressed cases. In GEC, there are cases where grammatically correct text is provided as source text. In such cases, the model does not need to make any revisions and can simply output \"None\", resulting in significant compression in GEC."
},
{
"figure_ref": [],
"heading": "Performance on Open-ended Task",
"publication_ref": [],
"table_ref": [
"tab_2"
],
"text": "In open-ended tasks, the target texts are written in plain text, while edit spans introduce significant differences in text formatting. This misalignment in text representation may potentially impact the performance of open-ended tasks. Therefore, we aim to demonstrate that edit spans do not significantly degrade the performance of open-ended tasks.\nTable 2 shows the scores for each LLM when using RoBERTa as BERTScore models on the 1K subset of the databricks-dolly-15k dataset, which was divided for evaluation. This indicates that the proposed method achieves efficient computational cost during inference without significantly sacrificing open-ended task performance.\nTo "
},
{
"figure_ref": [],
"heading": "The Accuracy of Edits Generated by the LLMs",
"publication_ref": [],
"table_ref": [
"tab_4"
],
"text": "Even if the edit span text is different, there are cases where the text is transformed by the rule, and the text matches. For example, in GEC, the model is given the input \"This technology could also be seen as invasion of human privacy.\" and the model outputs \"7 9 invading\". In this case, even with the alternate edit span text \"7 8 invading, 8 9\", the conversion based on the rules would result in the same output text. However, this would increase the sentence length by the index, creating room for improvement in terms of computational cost. Therefore, we investigate how well edit span of the model matches the results using linguistic alignment.\nFirst, we convert the edit spans generated by the model to plain text using rules. From the converted plain text and the source text, we create edit spans using linguistic alignment and calculate the percentage of agreement with the edit spans output by the model. Only when the start position s, end position e, and the edit token r all match exactly is it considered a correct answer.\nTable 4 shows the percentage of agreement be-tween the edit spans output by the LLMs and the edit spans extracted by the linguistic alignment in the development data for each task. The proposed method achieves more than 90% agreement in 13 out of 16 settings. This indicates that LLMs are able to learn the extraction rules for linguistical alignment through instruction tuning."
},
{
"figure_ref": [],
"heading": "Task-specific Fine-tuning",
"publication_ref": [],
"table_ref": [
"tab_5"
],
"text": "In the previous experiments, LLMs were trained by combining data from the four local sequence transduction tasks and the open-ended task. To explore the maximum potential performance of the proposed method, we fine-tune LLMs with taskspecific focus using edit span data. We fine-tune LLMs for each task using all available training data.\nIn this case, we specialize LLMs for specific tasks without the need for instruction texts. Therefore, we trained the LLMs by providing only the source texts as input.\nWe trained LLaMA, which showed the highest performance in the local sequence transduction tasks. We set the number of epochs to 2 and used a batch size of 32. The learning rate was set to 1e-5, with a warmup rate of 0.03, and we employed a cosine learning rate schedule. Following the exploration method described in Section 3.3, we determined the hyperparameters for our experiments.\nTable 5 shows the results of performance comparison with existing studies on GEC, paraphrasing, style transfer, and simplification tasks. The proposed method outperforms existing studies by 1.8 points in GEC, 0.9, 1.2, and 2.3 points in paraphrasing, 1.9 and 1.3 points in style transfer, and 1.2 and 0.7 points in simplification tasks, respectively. Thus, the proposed method achieves the SoTA performance in all tasks. From these results, it can be concluded that edit spans are an effective method, even in task-specific fine-tuning scenarios."
},
{
"figure_ref": [],
"heading": "Example of LLMs Output Using Edit Spans",
"publication_ref": [
"b54"
],
"table_ref": [
"tab_7"
],
"text": "Table 6 shows the output in CoNLL2013 for LLaMA using edit span and LLaMA outputting plain text. The normal model outputting plain text outputs 23 tokens, while the model using edit span outputs only 3 tokens. The output of the model using the edit span is a much shorter sequence than the original model that outputs plain text. Furthermore, LLaMA, which outputs in plain text, is unable to correct a grammatical error. In a local sequence transduction task, most tokens in the source text and target text are common, and the model tends to learn just to copy the input tokens (Rastogi et al., 2016). Contrarily, our model that uses edit spans outputs only the edited parts. Thus simply copying the input is not an issue for our model.\n5 Related Work"
},
{
"figure_ref": [],
"heading": "Efficient LLMs",
"publication_ref": [
"b67",
"b38",
"b35",
"b50",
"b4",
"b15",
"b70",
"b72",
"b27",
"b65",
"b59",
"b62",
"b0",
"b4",
"b73",
"b11",
"b49",
"b17",
"b10"
],
"table_ref": [],
"text": "Most of the methods for achieving efficient LLMs involve improving the memory complexity of selfattention mechanisms or enhancing the overall efficiency of the Transformer architecture (Tay et al., 2022;Loem et al., 2022). In the initial stages, the modifications made to self-attention focused on reducing the computational complexity by introducing sparsity in the attention matrix. This was accomplished by restricting the attention's scope to predetermined patterns, such as local windows and fixed stride block patterns (Liu et al., 2018;Qiu et al., 2020;Beltagy et al., 2020). A natural extension to the blockwise method is to connect these blocks via recurrence. Dai et al. (2019) introduced a mechanism of segment-level recurrence that establishes connections among multiple segments and blocks. An expansion upon fixed, predetermined pat-terns is the utilization of learnable patterns. Models that incorporate learnable patterns aim to acquire the access pattern through data-driven methods. One crucial aspect of learning patterns is to establish a concept of token relevance and subsequently assign tokens to buckets or clusters (Vyas et al., 2020;Wang et al., 2021;Kitaev et al., 2020;Tay et al., 2020;Roy et al., 2021). Another approach is to utilize a trainable side memory module capable of accessing multiple tokens simultaneously (Sukhbaatar et al., 2019;Ainslie et al., 2020;Beltagy et al., 2020). A prevalent example is the global neural memory, which can access the entire sequence. The global tokens function as a type of model memory, learning to gather information from the input sequence tokens.\nAnother method to enhance efficiency is by utilizing low-rank approximations of the selfattention matrix to improve computational performance (Wang et al., 2020), and to view the attention mechanism through kernelization (Choromanski et al., 2020;Peng et al., 2021). Sparse models selectively activate a fraction of the parameters, resulting in an improved parameter to FLOPs ratio in general (Fedus et al., 2022).\nAs a way to reduce the length of the text, Cheng et al. (2023) in one prompt and inferring in parallel.\nThese techniques, unlike our research, do not alter the writing style of the target text, and edit spans can be used in conjunction with these methods."
},
{
"figure_ref": [],
"heading": "Edit-based Model",
"publication_ref": [
"b54",
"b61",
"b22",
"b57",
"b28",
"b3",
"b21",
"b39",
"b81",
"b55"
],
"table_ref": [],
"text": "Since the question of necessarily using the seq2seq model for local sequence transduction tasks was raised (Rastogi et al., 2016;Schnober et al., 2016), various edit-based models have been proposed. Guu et al. (2018) proposed a language model that initially selects a prototype sentence from the training dataset and subsequently modifies it to create a new sentence. Ribeiro et al. (2018) introduced a method for representing general string transduction problems as sequence labeling. Koide et al. (2018) proposed the model implemented to analyze the evolution of biological sequences driven by substitution, insertion, and deletion edit operations, achieving improved accuracy on protein secondary structure prediction. Awasthi et al. (2019) presented a parallel iterative edit model reducing decoding time for local sequence transduction tasks. Gu et al. (2019) developed the Levenshtein Transformer, a non-autoregressive model using edit operations. (Mallinson et al., 2020) introduced FELIX, an adaptable text-editing approach for the generation that aims to leverage the advantages of decoding with bi-directional contexts and self-supervised pretraining to the fullest extent. (Xu and Carpuat, 2021) presented an Edit-Based Transformer with Repositioning, which enhances sequence generation flexibility by seamlessly incorporating userspecified preferences in output lexical choice. Reid and Neubig (2022) proposed the modeling of editing processes, encompassing the iterative generation of sequences as a whole. They establish a conceptual framework to explain the probability of multi-step edits and outline neural models capable of learning a generative model of sequences by leveraging these multi-step edits.\nHowever, these methods have different architec-tures from LLMs. Therefore, it is not easy to apply them to LLMs, unlike our method, which can train models by simply changing the output text."
},
{
"figure_ref": [],
"heading": "LLMs for Local Sequence Transduction Tasks",
"publication_ref": [
"b37",
"b40"
],
"table_ref": [],
"text": "In GEC, the model based on GPT-3 achieves stateof-the-art in unsupervised settings (Loem et al., 2023). 2022) introduced a method based on GPT-3 that solely relies on natural language instruction and does not necessitate model fine-tuning or exemplars in the desired style. Malmi et al. (2020) proposed a method of using LLMs for style transfer where no parallel data is available. On the other hand, these studies did not target the efficiency of LLMs based on the edit."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "In this study, we proposed to predict a set of edit spans, which represent the changed parts of the target text relative to the source tokens. We showed our method omits unedited tokens that occupy most of the target text and reduces the length of the target text and the inference time for local sequence transduction tasks. Moreover, we reported that instruction tuning with the proposed method achieves state-of-the-art performance in the four tasks."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [
"b7"
],
"table_ref": [],
"text": "In our preliminary experiments, even highperformance LLMs such as GPT-3 (Brown et al., 2020) and ChatGPT (OpenAI, 2023) could not generate edit spans with zero-shot and few-shot. In particular, indexes could not be generated correctly. Therefore, it is a future work to apply the proposed method to zero-shot and few-shot. Moreover, the use of edit span is not necessarily effective for tasks, such as machine translation and dialogue, other than the local sequence transduction task, where many tokens in the source and target texts are not common."
},
{
"figure_ref": [],
"heading": "Acknowledgements",
"publication_ref": [],
"table_ref": [],
"text": "These research results were obtained from the commissioned research (No.225) by National Institute of Information and Communications Technology (NICT), Japan."
}
] | Large Language Models (LLMs) have demonstrated remarkable performance in various tasks and gained significant attention. LLMs are also used for local sequence transduction tasks, including grammatical error correction (GEC) and formality style transfer, where most tokens in a source text are kept unchanged. However, the models that generate all target tokens in such tasks have a tendency to simply copy the input text as is, without making needed changes, because the difference between input and output texts is minimal in the training data. This is also inefficient because the computational cost grows quadratically with the target sequence length with Transformer. This paper proposes predicting edit spans for the source text for local sequence transduction tasks. Representing an edit span with a position of the source text and corrected tokens, we can reduce the length of the target sequence and the computational cost for inference. We apply instruction tuning for LLMs on the supervision data of edit spans. Experiments show that the proposed method achieves comparable performance to the baseline in four tasks, paraphrasing, formality style transfer, GEC, and text simplification, despite reducing the length of the target text by as small as 21%. Furthermore, we report that the task-specific fine-tuning with the proposed method achieved state-of-the-art performance in the four tasks. | Reducing Sequence Length by Predicting Edit Spans with Large Language Models | [
{
"figure_caption": "Figure 1 :1Figure 1: Inference of instruction tuned LLMs using edit spans. LLMs take instruction text and source text as input and output only the positions and tokens for rewriting. Rule-based conversion applies the outputted positions and tokens of the rewriting to the source text and produces the plaintext output.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: The ratio of output text length to target text length when MPT, OPT, LLaMA, and BLOOM are fine-tuned with plain data and edit span data, respectively.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "maintain performance in open-ended tasks, the proposed method combines data from both local sequence transduction tasks and open-ended tasks. To demonstrate the effectiveness of combining open-ended task data, we also investigate the open-ended task performance of instructiontuned LLMs when solely trained on local sequence transduction task data.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Fang et al. (2023) showed that ChatGPT corrects input text very fluently.Yamashita et al. (2020);Rothe et al. (2021);Sun et al. (2022) proposed a method for multilingual GEC using multilingual LLMs.Feng et al. (2023b) investigated the performance of few-shot and zero-shot of GPT3 and ChatGPT in the simplification.Anschütz et al. (2023) used LLMs for German simplification and found them to be effective in languages with little parallel data.(Witteveen and Andrews, 2019) verified the performance of GPT-2(Radford et al., 2019) in paraphrasing.Wahle et al. (2022) investigated the utilization of T5 and GPT3 in generating machine-generated paraphrases for scientific articles sourced from arXiv, student theses, and Wikipedia.Reif et al. (",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "The performance of four LLMs fine-tuned with edit spans and plain data instructions on four local sequence transduction tasks. The bold values indicate the highest performance for each task. The underlined values indicate when edit spans exceed the baseline.",
"figure_data": "GECParaphrasingStyle transfer SimplificationMPT68.0 37.9/66.5/47.178.9/81.246.3/41.1PlainOPT LLaMA65.7 35.2/63.2/45.4 68.2 39.3/69.0/47.275.0/77.2 79.5/81.043.7/40.5 48.0/41.9BLOOM 66.4 37.0/66.4/46.178.2/79.945.0/41.0MPT68.5 38.2/66.7/47.178.2/81.346.6/41.3Edit spansOPT LLaMA66.2 34.1/61.2/43.9 69.1 39.0/69.2/47.675.6/77.9 79.3/81.243.9/40.3 48.3/42.0BLOOM 65.8 37.2/66.1/46.378.0/80.344.8/40.7",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": "demonstrates the performance differenceon the 1K split of the databricks-dolly-15k dataset,evaluating LLMs trained on both open-ended taskand local sequence transduction task data versusLLMs trained solely on local sequence transductiontask data. The performance decreases when not us-ing open-ended task data for training, both in termsof plain text and edit spans. This is likely becauseopen-ended task data consists of plain text, whileedit spans include totally different text formats,leading to a larger disparity in task requirements.",
"figure_id": "tab_1",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Scores using BERTScore on the databricksdolly-15k dataset, which was divided for evaluation.",
"figure_data": "BERTScore diff.MPT-5.2PlainOPT LLaMA-5.7 -4.4BLOOM-6.2MPT-8.1Edit spanOPT LLaMA-8.6 -6.9BLOOM-7.6",
"figure_id": "tab_2",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "The performance difference between instruction tuned LLMs using local sequence transduction task and open-ended task datasets, and instruction tuned LLMs using only local sequence transduction task datasets.",
"figure_data": "",
"figure_id": "tab_3",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Agreement between edit spans generated by LLMs and edit spans extracted by linguistic alignment.",
"figure_data": "GEC Paraphrasing Style transfer SimplificationMPT96.695.089.294.7OPT93.391.988.892.7LLaMA99.096.292.695.4BLOOM 94.292.589.493.5ParaphrasingGEC(Kumar et al., 2020) 38.0/68.1/45.7(Kaneko et al., 2020)65.2(Meng et al., 2021)26.8/65.0/38.5(Omelianchuk et al., 2020) 66.5(Li et al., 2022)39.3/70.8/48.3(Qorib et al., 2022)69.5Edit span41.2/72.0/50.6Edit span71.3(b) BLEU-4, ROUGE-1, and ROUGE-2 scores on(a) M 2 scores on the CoNLL2014 dataset.the Quora dataset.Style transferSimplification(Chawla and Yang, 2020)76.2/79.9(Martin et al., 2020)40.1/41.4(Lai et al., 2021)76.5/79.3(Martin et al., 2022)44.2/42.6(Liu et al., 2022)78.8/81.4(Feng et al., 2023a)47.9/41.8Edit span80.7/82.7Edit span49.1/43.5(c) NLTK BLEU scores on the E&M and F&R(d) SARI scores on ASSET and TurkCorpusdatasets.datasets.",
"figure_id": "tab_4",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Performance comparison with previous studies on GEC, paraphrasing, style transfer, and simplification tasks.",
"figure_data": "",
"figure_id": "tab_5",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "proposed including multiple examplesSource text Since we do not to bring cash to pay for the transportation fee , enormous time has been saved for everybody .Target textSince we do not need to bring cash to pay for the transportation fee , enormous time has been saved for everybody .",
"figure_data": "Target edit span4 4 needPlainSince we do not to bring cash to pay for the transportation fee , enormous time has been savedfor everybody .System edit span 4 4 need",
"figure_id": "tab_6",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Outputs in plain text and edit span formats respectively by LLaMA in the CoNLL2013.",
"figure_data": "",
"figure_id": "tab_7",
"figure_label": "6",
"figure_type": "table"
}
] | Masahiro Kaneko; Naoaki Okazaki; Mbzuai | [
{
"authors": "Joshua Ainslie; Santiago Ontanon; Chris Alberti; Vaclav Cvicek; Zachary Fisher; Philip Pham; Anirudh Ravula; Sumit Sanghai; Qifan Wang; Li Yang",
"journal": "Association for Computational Linguistics",
"ref_id": "b0",
"title": "ETC: Encoding long and structured inputs in transformers",
"year": "2020"
},
{
"authors": "Fernando Alva-Manchego; Louis Martin; Antoine Bordes; Carolina Scarton; Benoît Sagot; Lucia Specia",
"journal": "Association for Computational Linguistics",
"ref_id": "b1",
"title": "ASSET: A dataset for tuning and evaluation of sentence simplification models with multiple rewriting transformations",
"year": "2020"
},
{
"authors": "Miriam Anschütz; Joshua Oehms; Thomas Wimmer; Bartłomiej Jezierski; Georg Groh",
"journal": "",
"ref_id": "b2",
"title": "Language models for german text simplification: Overcoming parallel data scarcity through style-specific pre-training",
"year": "2023"
},
{
"authors": "Abhijeet Awasthi; Sunita Sarawagi; Rasna Goyal; Sabyasachi Ghosh; Vihari Piratla",
"journal": "Association for Computational Linguistics",
"ref_id": "b3",
"title": "Parallel iterative edit models for local sequence transduction",
"year": "2019"
},
{
"authors": "Iz Beltagy; Matthew E Peters; Arman Cohan",
"journal": "",
"ref_id": "b4",
"title": "Longformer: The long-document transformer",
"year": "2020"
},
{
"authors": "Emily M Bender; Timnit Gebru; Angelina Mcmillan-Major; Shmargaret Shmitchell",
"journal": "",
"ref_id": "b5",
"title": "On the dangers of stochastic parrots: Can language models be too big?",
"year": "2021"
},
{
"authors": "Steven Bird; Edward Loper",
"journal": "Association for Computational Linguistics",
"ref_id": "b6",
"title": "NLTK: The natural language toolkit",
"year": "2004"
},
{
"authors": "Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell; Sandhini Agarwal; Ariel Herbert-Voss; Gretchen Krueger; Tom Henighan; Rewon Child; Aditya Ramesh; Daniel Ziegler; Jeffrey Wu; Clemens Winter; Chris Hesse; Mark Chen; Eric Sigler; Mateusz Litwin; Scott Gray; Benjamin Chess; Jack Clark; Christopher Berner; Sam Mccandlish; Alec Radford; Ilya Sutskever; Dario Amodei",
"journal": "",
"ref_id": "b7",
"title": "Language models are few-shot learners",
"year": "2020"
},
{
"authors": "Curran Associates; Inc ",
"journal": "",
"ref_id": "b8",
"title": "",
"year": ""
},
{
"authors": "Kunal Chawla; Diyi Yang",
"journal": "Association for Computational Linguistics",
"ref_id": "b9",
"title": "Semi-supervised formality style transfer using language model discriminator and mutual information maximization",
"year": "2020"
},
{
"authors": "Zhoujun Cheng; Jungo Kasai; Tao Yu",
"journal": "",
"ref_id": "b10",
"title": "Batch prompting: Efficient inference with large language model apis",
"year": "2023"
},
{
"authors": "Krzysztof Choromanski; Valerii Likhosherstov; David Dohan; Xingyou Song; Andreea Gane; Tamas Sarlos; Peter Hawkins; Jared Davis; David Belanger; Lucy Colwell",
"journal": "",
"ref_id": "b11",
"title": "Masked language modeling for proteins via linearly scalable long-context transformers",
"year": "2020"
},
{
"authors": "Aakanksha Chowdhery; Sharan Narang; Jacob Devlin; Maarten Bosma; Gaurav Mishra; Adam Roberts; Paul Barham; Hyung Won Chung; Charles Sutton; Sebastian Gehrmann",
"journal": "",
"ref_id": "b12",
"title": "Palm: Scaling language modeling with pathways",
"year": "2022"
},
{
"authors": "Chung Hyung Won; Le Hou; Shayne Longpre; Barret Zoph; Yi Tay; William Fedus; Eric Li; Xuezhi Wang; Mostafa Dehghani; Siddhartha Brahma",
"journal": "",
"ref_id": "b13",
"title": "Scaling instruction-finetuned language models",
"year": "2022"
},
{
"authors": "Daniel Dahlmeier; Hwee Tou Ng",
"journal": "Association for Computational Linguistics",
"ref_id": "b14",
"title": "Better evaluation for grammatical error correction",
"year": "2012"
},
{
"authors": "Zihang Dai; Zhilin Yang; Yiming Yang; Jaime Carbonell; Quoc Le; Ruslan Salakhutdinov",
"journal": "Association for Computational Linguistics",
"ref_id": "b15",
"title": "Transformer-XL: Attentive language models beyond a fixed-length context",
"year": "2019"
},
{
"authors": "Tao Fang; Shu Yang; Kaixin Lan; Derek F Wong; Jinpeng Hu; Lidia S Chao; Yue Zhang",
"journal": "",
"ref_id": "b16",
"title": "Is chatgpt a highly fluent grammatical error correction system? a comprehensive evaluation",
"year": "2023"
},
{
"authors": "William Fedus; Barret Zoph; Noam Shazeer",
"journal": "The Journal of Machine Learning Research",
"ref_id": "b17",
"title": "Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity",
"year": "2022"
},
{
"authors": "Mariano Felice; Christopher Bryant; Ted Briscoe",
"journal": "",
"ref_id": "b18",
"title": "Automatic extraction of learner errors in ESL sentences using linguistically enhanced alignments",
"year": "2016"
},
{
"authors": "Yutao Feng; Jipeng Qiang; Yun Li; Yunhao Yuan; Yi Zhu",
"journal": "",
"ref_id": "b19",
"title": "Sentence simplification via large language models",
"year": "2023"
},
{
"authors": "Yutao Feng; Jipeng Qiang; Yun Li; Yunhao Yuan; Yi Zhu",
"journal": "",
"ref_id": "b20",
"title": "Sentence simplification via large language models",
"year": "2023"
},
{
"authors": "Jiatao Gu; Changhan Wang; Junbo Zhao",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b21",
"title": "Levenshtein transformer",
"year": "2019"
},
{
"authors": "Kelvin Guu; B Tatsunori; Yonatan Hashimoto; Percy Oren; Liang",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b22",
"title": "Generating sentences by editing prototypes",
"year": "2018"
},
{
"authors": "Masahiro Kaneko; Masato Mita; Shun Kiyono; Jun Suzuki; Kentaro Inui",
"journal": "Association for Computational Linguistics",
"ref_id": "b23",
"title": "Encoder-decoder models can benefit from pre-trained masked language models in grammatical error correction",
"year": "2020"
},
{
"authors": "Masahiro Kaneko; Naoaki Okazaki",
"journal": "",
"ref_id": "b24",
"title": "Controlled generation with prompt insertion for natural language explanations in grammatical error correction",
"year": "2023"
},
{
"authors": "Masahiro Kaneko; Sho Takase; Ayana Niwa; Naoaki Okazaki",
"journal": "Association for Computational Linguistics",
"ref_id": "b25",
"title": "Interpretability for language learners using example-based grammatical error correction",
"year": "2022"
},
{
"authors": "Jared Kaplan; Sam Mccandlish; Tom Henighan; Tom B Brown; Benjamin Chess; Rewon Child; Scott Gray; Alec Radford; Jeffrey Wu; Dario Amodei",
"journal": "",
"ref_id": "b26",
"title": "Scaling laws for neural language models",
"year": "2020"
},
{
"authors": "Nikita Kitaev; Łukasz Kaiser; Anselm Levskaya",
"journal": "",
"ref_id": "b27",
"title": "Reformer: The efficient transformer",
"year": "2020"
},
{
"authors": "Satoshi Koide; Keisuke Kawano; Takuro Kutsuna",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b28",
"title": "Neural edit operations for biological sequences",
"year": "2018"
},
{
"authors": "Mathias Kraus; Julia Anna Bingler; Markus Leippold; Tobias Schimanski; Colesanti Chiara; Dominik Senni; Saeid Ashraf Stammbach; Nicolas Vaghefi; Webersinke",
"journal": "",
"ref_id": "b29",
"title": "Enhancing large language models with climate resources",
"year": "2023"
},
{
"authors": "Ashutosh Kumar; Kabir Ahuja; Raghuram Vadapalli; Partha Talukdar",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b30",
"title": "Syntax-Guided Controlled Generation of Paraphrases",
"year": "2020"
},
{
"authors": "Huiyuan Lai; Antonio Toral; Malvina Nissim",
"journal": "Association for Computational Linguistics",
"ref_id": "b31",
"title": "Thank you BART! rewarding pre-trained models improves formality style transfer",
"year": "2021"
},
{
"authors": "Zhigen Li; Yanmeng Wang; Rizhao Fan; Ye Wang; Jianfeng Li; Shaojun Wang",
"journal": "Association for Computational Linguistics",
"ref_id": "b32",
"title": "Learning to adapt to low-resource paraphrase generation",
"year": "2022"
},
{
"authors": "Chin-Yew Lin",
"journal": "Association for Computational Linguistics",
"ref_id": "b33",
"title": "ROUGE: A package for automatic evaluation of summaries",
"year": "2004"
},
{
"authors": "Ao Liu; An Wang; Naoaki Okazaki",
"journal": "Association for Computational Linguistics",
"ref_id": "b34",
"title": "Semisupervised formality style transfer with consistency training",
"year": "2022"
},
{
"authors": "J Peter; Mohammad Liu; Etienne Saleh; Ben Pot; Ryan Goodrich; Lukasz Sepassi; Noam Kaiser; Shazeer",
"journal": "",
"ref_id": "b35",
"title": "Generating wikipedia by summarizing long sequences",
"year": "2018"
},
{
"authors": "Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov",
"journal": "",
"ref_id": "b36",
"title": "Roberta: A robustly optimized bert pretraining approach",
"year": "2019"
},
{
"authors": "Mengsay Loem; Masahiro Kaneko; Sho Takase; Naoaki Okazaki",
"journal": "",
"ref_id": "b37",
"title": "Exploring effectiveness of gpt-3 in grammatical error correction: A study on performance and controllability in prompt-based methods",
"year": "2023"
},
{
"authors": "Mengsay Loem; Sho Takase; Masahiro Kaneko; Naoaki Okazaki",
"journal": "",
"ref_id": "b38",
"title": "Are neighbors enough? multi-head neural n-gram can be alternative to selfattention",
"year": "2022"
},
{
"authors": "Jonathan Mallinson; Aliaksei Severyn; Eric Malmi; Guillermo Garrido",
"journal": "Association for Computational Linguistics",
"ref_id": "b39",
"title": "FELIX: Flexible text editing through tagging and insertion",
"year": "2020"
},
{
"authors": "Eric Malmi; Aliaksei Severyn; Sascha Rothe",
"journal": "",
"ref_id": "b40",
"title": "Unsupervised text style transfer with padded masked language models",
"year": "2020"
},
{
"authors": "Louis Martin; Éric De La Clergerie; Benoît Sagot; Antoine Bordes",
"journal": "European Language Resources Association",
"ref_id": "b41",
"title": "Controllable sentence simplification",
"year": "2020"
},
{
"authors": "Louis Martin; Angela Fan; Éric De La Clergerie; Antoine Bordes; Benoît Sagot",
"journal": "European Language Resources Association",
"ref_id": "b42",
"title": "MUSS: Multilingual unsupervised sentence simplification by mining paraphrases",
"year": "2022"
},
{
"authors": "Yuxian Meng; Xiang Ao; Qing He; Xiaofei Sun; Qinghong Han; Fei Wu; Chun Fan; Jiwei Li",
"journal": "Association for Computational Linguistics",
"ref_id": "b43",
"title": "ConRPG: Paraphrase generation using contexts as regularizer",
"year": "2021"
},
{
"authors": "Tou Hwee; Ng; Mei Siew; Ted Wu; Christian Briscoe; Raymond Hendy Hadiwinoto; Christopher Susanto; Bryant",
"journal": "Association for Computational Linguistics",
"ref_id": "b44",
"title": "The CoNLL-2014 shared task on grammatical error correction",
"year": "2014"
},
{
"authors": "Tou Hwee; Ng; Mei Siew; Yuanbin Wu; Christian Wu; Joel Hadiwinoto; Tetreault",
"journal": "Association for Computational Linguistics",
"ref_id": "b45",
"title": "The CoNLL-2013 shared task on grammatical error correction",
"year": "2013"
},
{
"authors": "Kostiantyn Omelianchuk; Vitaliy Atrasevych; Artem Chernodub; Oleksandr Skurzhanskyi",
"journal": "Association for Computational Linguistics. OpenAI",
"ref_id": "b46",
"title": "GECToR -grammatical error correction: Tag, not rewrite",
"year": "2020"
},
{
"authors": "Long Ouyang; Jeffrey Wu; Xu Jiang; Diogo Almeida; Carroll Wainwright; Pamela Mishkin; Chong Zhang; Sandhini Agarwal; Katarina Slama; Alex Ray",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b47",
"title": "Training language models to follow instructions with human feedback",
"year": "2022"
},
{
"authors": "Kishore Papineni; Salim Roukos; Todd Ward; Wei-Jing Zhu",
"journal": "",
"ref_id": "b48",
"title": "Bleu: a method for automatic evaluation of machine translation",
"year": "2002"
},
{
"authors": "Hao Peng; Nikolaos Pappas; Dani Yogatama; Roy Schwartz; Noah A Smith; Lingpeng Kong",
"journal": "",
"ref_id": "b49",
"title": "Random feature attention",
"year": "2021"
},
{
"authors": "Jiezhong Qiu; Hao Ma; Omer Levy; Wen-Tau Yih; Sinong Wang; Jie Tang",
"journal": "Association for Computational Linguistics",
"ref_id": "b50",
"title": "Blockwise selfattention for long document understanding",
"year": "2020"
},
{
"authors": "Muhammad Qorib; Seung-Hoon Na; Hwee Tou Ng",
"journal": "Association for Computational Linguistics",
"ref_id": "b51",
"title": "Frustratingly easy system combination for grammatical error correction",
"year": "2022"
},
{
"authors": "Alec Radford; Jeffrey Wu; Rewon Child; David Luan; Dario Amodei; Ilya Sutskever",
"journal": "OpenAI blog",
"ref_id": "b52",
"title": "Language models are unsupervised multitask learners",
"year": "2019"
},
{
"authors": "Sudha Rao; Joel Tetreault",
"journal": "Association for Computational Linguistics",
"ref_id": "b53",
"title": "Dear sir or madam, may I introduce the GYAFC dataset: Corpus, benchmarks and metrics for formality style transfer",
"year": "2018"
},
{
"authors": "Pushpendre Rastogi; Ryan Cotterell; Jason Eisner",
"journal": "Association for Computational Linguistics",
"ref_id": "b54",
"title": "Weighting finite-state transductions with neural context",
"year": "2016"
},
{
"authors": "Machel Reid; Graham Neubig",
"journal": "Association for Computational Linguistics",
"ref_id": "b55",
"title": "Learning to model editing processes",
"year": "2022"
},
{
"authors": "Emily Reif; Daphne Ippolito; Ann Yuan; Andy Coenen; Chris Callison-Burch; Jason Wei",
"journal": "Association for Computational Linguistics",
"ref_id": "b56",
"title": "A recipe for arbitrary text style transfer with large language models",
"year": "2022"
},
{
"authors": "Joana Ribeiro; Shashi Narayan; Shay B Cohen; Xavier Carreras",
"journal": "",
"ref_id": "b57",
"title": "Local string transduction as sequence labeling",
"year": "2018"
},
{
"authors": "Sascha Rothe; Jonathan Mallinson; Eric Malmi; Sebastian Krause; Aliaksei Severyn",
"journal": "Association for Computational Linguistics",
"ref_id": "b58",
"title": "A simple recipe for multilingual grammatical error correction",
"year": "2021"
},
{
"authors": "Aurko Roy; Mohammad Saffar; Ashish Vaswani; David Grangier",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b59",
"title": "Efficient content-based sparse attention with routing transformers",
"year": "2021"
},
{
"authors": "Le Teven; Angela Scao; Christopher Fan; Ellie Akiki; Suzana Pavlick; Daniel Ilić; Roman Hesslow; Alexandra Castagné; François Sasha Luccioni; Matthias Yvon; Gallé",
"journal": "",
"ref_id": "b60",
"title": "Bloom: A 176bparameter open-access multilingual language model",
"year": "2022"
},
{
"authors": "Carsten Schnober; Steffen Eger; Erik-Lân Do Dinh; Iryna Gurevych",
"journal": "",
"ref_id": "b61",
"title": "Still not there? comparing traditional sequence-to-sequence models to encoderdecoder neural networks on monotone string translation tasks",
"year": "2016"
},
{
"authors": "Sainbayar Sukhbaatar; Edouard Grave; Guillaume Lample; Herve Jegou; Armand Joulin",
"journal": "",
"ref_id": "b62",
"title": "Augmenting self-attention with persistent memory",
"year": "2019"
},
{
"authors": "Xin Sun; Tao Ge; Shuming Ma; Jingjing Li; Furu Wei; Houfeng Wang",
"journal": "",
"ref_id": "b63",
"title": "A unified strategy for multilingual grammatical error correction with pretrained cross-lingual language model",
"year": "2022"
},
{
"authors": "Rohan Taori; Ishaan Gulrajani; Tianyi Zhang; Yann Dubois; Xuechen Li; Carlos Guestrin; Percy Liang; Tatsunori B Hashimoto",
"journal": "",
"ref_id": "b64",
"title": "Stanford alpaca: An instruction-following llama model",
"year": "2023"
},
{
"authors": "Yi Tay; Dara Bahri; Liu Yang; Donald Metzler; Da-Cheng Juan",
"journal": "",
"ref_id": "b65",
"title": "Sparse sinkhorn attention",
"year": "2020"
},
{
"authors": " Pmlr",
"journal": "",
"ref_id": "b66",
"title": "",
"year": ""
},
{
"authors": "Yi Tay; Mostafa Dehghani; Dara Bahri; Donald Metzler",
"journal": "ACM Computing Surveys",
"ref_id": "b67",
"title": "Efficient transformers: A survey",
"year": "2022"
},
{
"authors": "Nlp Mosaicml; Team",
"journal": "",
"ref_id": "b68",
"title": "Introducing mpt-7b: A new standard for open-source, ly usable llms",
"year": "2023"
},
{
"authors": "Hugo Touvron; Thibaut Lavril; Gautier Izacard; Xavier Martinet; Marie-Anne Lachaux; Timothée Lacroix; Baptiste Rozière; Naman Goyal; Eric Hambro; Faisal Azhar; Aurelien Rodriguez; Armand Joulin; Edouard Grave; Guillaume Lample",
"journal": "",
"ref_id": "b69",
"title": "Llama: Open and efficient foundation language models",
"year": "2023"
},
{
"authors": "Apoorv Vyas; Angelos Katharopoulos; François Fleuret",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b70",
"title": "Fast transformers with clustered attention",
"year": "2020"
},
{
"authors": "Jan Philip Wahle; Terry Ruas; Frederic Kirstein; Bela Gipp",
"journal": "Association for Computational Linguistics",
"ref_id": "b71",
"title": "How large language models are transforming machine-paraphrase plagiarism",
"year": "2022"
},
{
"authors": "Shuohang Wang; Luowei Zhou; Zhe Gan; Yen-Chun Chen; Yuwei Fang; Siqi Sun; Yu Cheng; Jingjing Liu",
"journal": "Association for Computational Linguistics",
"ref_id": "b72",
"title": "Cluster-former: Clustering-based sparse transformer for question answering",
"year": "2021"
},
{
"authors": "Sinong Wang; Belinda Z Li; Madian Khabsa; Han Fang; Hao Ma",
"journal": "",
"ref_id": "b73",
"title": "Linformer: Self-attention with linear complexity",
"year": "2020"
},
{
"authors": "Yizhong Wang; Yeganeh Kordi; Swaroop Mishra; Alisa Liu; Noah A Smith; Daniel Khashabi; Hannaneh Hajishirzi",
"journal": "",
"ref_id": "b74",
"title": "Self-instruct: Aligning language model with self generated instructions",
"year": "2022"
},
{
"authors": "Jason Wei; Maarten Bosma; Y Vincent; Kelvin Zhao; Adams Wei Guu; Brian Yu; Nan Lester; Andrew M Du; Quoc V Dai; Le",
"journal": "",
"ref_id": "b75",
"title": "Finetuned language models are zero-shot learners",
"year": "2021"
},
{
"authors": "Jason Wei; Yi Tay; Rishi Bommasani; Colin Raffel; Barret Zoph; Sebastian Borgeaud; Dani Yogatama; Maarten Bosma; Denny Zhou; Donald Metzler",
"journal": "",
"ref_id": "b76",
"title": "Emergent abilities of large language models",
"year": "2022"
},
{
"authors": "Sam Witteveen; Martin Andrews",
"journal": "Association for Computational Linguistics",
"ref_id": "b77",
"title": "Paraphrasing with large language models",
"year": "2019"
},
{
"authors": "Haoran Wu; Wenxuan Wang; Yuxuan Wan; Wenxiang Jiao; Michael Lyu",
"journal": "",
"ref_id": "b78",
"title": "Chatgpt or grammarly? evaluating chatgpt on grammatical error correction benchmark",
"year": "2023"
},
{
"authors": "Minghao Wu; Abdul Waheed; Chiyu Zhang; Muhammad Abdul-Mageed; Alham Fikri; Aji ",
"journal": "",
"ref_id": "b79",
"title": "Lamini-lm: A diverse herd of distilled models from large-scale instructions",
"year": "2023"
},
{
"authors": "Wei Xu; Courtney Napoles; Ellie Pavlick; Quanze Chen; Chris Callison-Burch",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b80",
"title": "Optimizing statistical machine translation for text simplification",
"year": "2016"
},
{
"authors": "Weijia Xu; Marine Carpuat",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b81",
"title": "Editor: An editbased transformer with repositioning for neural machine translation with soft lexical constraints",
"year": "2021"
},
{
"authors": "Ikumi Yamashita; Satoru Katsumata; Masahiro Kaneko; Aizhan Imankulova; Mamoru Komachi",
"journal": "International Committee on Computational Linguistics",
"ref_id": "b82",
"title": "Cross-lingual transfer learning for grammatical error correction",
"year": "2020"
},
{
"authors": "Susan Zhang; Stephen Roller; Naman Goyal; Mikel Artetxe; Moya Chen; Shuohui Chen; Christopher Dewan; Mona Diab; Xian Li; Xi Victoria Lin",
"journal": "",
"ref_id": "b83",
"title": "Opt: Open pre-trained transformer language models",
"year": "2022"
},
{
"authors": "Tianyi Zhang; Varsha Kishore; Felix Wu; Kilian Q Weinberger; Yoav Artzi",
"journal": "",
"ref_id": "b84",
"title": "Bertscore: Evaluating text generation with bert",
"year": "2019"
},
{
"authors": "Xingxing Zhang; Mirella Lapata",
"journal": "Association for Computational Linguistics",
"ref_id": "b85",
"title": "Sentence simplification with deep reinforcement learning",
"year": "2017"
},
{
"authors": "Kun Wayne Xin Zhao; Junyi Zhou; Tianyi Li; Xiaolei Tang; Yupeng Wang; Yingqian Hou; Beichen Min; Junjie Zhang; Zican Zhang; Dong",
"journal": "",
"ref_id": "b86",
"title": "A survey of large language models",
"year": "2023"
},
{
"authors": "Zhemin Zhu; Delphine Bernhard; Iryna Gurevych",
"journal": "",
"ref_id": "b87",
"title": "A monolingual tree-based translation model for sentence simplification",
"year": "2010"
}
] | [] |
2023-09-15 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b24",
"b45",
"b22",
"b7",
"b40",
"b52",
"b65",
"b60",
"b76",
"b95",
"b98",
"b100",
"b6",
"b23",
"b29",
"b55",
"b99",
"b54",
"b88",
"b9",
"b64",
"b97",
"b56",
"b22",
"b19",
"b77",
"b93",
"b7",
"b65",
"b58",
"b91",
"b90",
"b42",
"b60",
"b84",
"b95",
"b50",
"b65",
"b3",
"b96"
],
"table_ref": [],
"text": "The creation of clothed 3D human characters, which we refer to as \"digital avatars\", has become an essential part of many fields including gaming, animation, virtual/mixed reality, and the 3D industry in general. These digital avatars allow users to use their virtual representation for a range of purposes, thus enhancing user immersion within such services. However, creating high-quality digital avatars often requires specialized 3D artists using a sophisticated creation pipeline [26,47], making it a laborious process.\nThe recent advances in deep generative models [24,28,42] have enabled the creation of high-quality images that accurately reflects the textual input semantics [54,67]. However, the usage of such generative models in creating 3D has mainly focused on object generation [62,78,97,100,102] and shown rather limited performance in generating fullbody, realistic 3D human avatars due to the difficulty of collecting a large-scale ground truth dataset. Many previous 3D generative models [2,7,8,25,31,57,101] focus on training generative models on large-scale image datasets along with implicit 3D shape representations and differentiable volume rendering [56,90]. However, those approaches are rather limited in generating full-body humans with realistic details and rely on computationally expensive volume rendering. Other approach [10] directly uses high-quality 3D datasets [66,99] to train generative models based on auto-decoding frameworks [58], but the resulting stochastic details tend to be unrealistic, due to the usage of an adversarial loss [24].\nIn this paper, we decompose the problem of 3D generation into 2D normal map generation and 3D reconstruction, bridging the power of generative models in the image domain toward 3D generation. Following the intuition of \"sandwich-like\" approaches for single image-based 3D human reconstruction [21,79,95], we generate normal maps for frontal and backside regions of human mesh to get rich details mitigating the computational cost of 3D representations. We adopt a diffusion model [28,67] to simultaneously create consistent normal maps for both frontal and backside regions, which we call dual normal maps, conditioned on a posed SMPL-X [48,60]. Since diffusion models are well known for their mode coverage [93], we find it suitable to generate diverse 3D digital avatars. The dual normal maps are then used as input for our 3D reconstruction pipeline, in which we carve the initial posed SMPL-X mesh to a clothed, realistic human mesh with normal map-based mesh optimization inspired by NDS [92]. During optimization, the initial mesh is gradually deformed to match the generated normal maps through a differentiable rasterization pipeline [44] and geometric regularization including a loss function for plausible side-view. Our dual normal map-based 3D generation pipeline alleviates the difficulty of generating consistent multi-views, which is the fundamental reason that diffusion-based 3D generative models [62,86,97] suffer from slow convergence or fail to generate multi-view consistent results. We show that the diffusion model can generate consistent dual normal maps and they are sufficient to generate plausible 3D humans along with SMPL-X prior. Then, we can further improve the generated mesh by using a resampling scheme motivated by SDEdit [52], in which we use separate diffusion models for the body and facial regions to refine the perceptual quality of the rendered normals in different viewpoints while preserving the view and identity consistency. The refined normal maps are subsequently used as inputs for the mesh optimization, thus creating a realistic 3D digital avatar with high-frequency details.\nAs shown in Fig. 1, our pipeline, which we dub it Chupa, can be extended to text-based generation for further controllability on the human identity (e.g., gender, clothing, hair, etc.), by leveraging the power of a pre-trained text-to-image diffusion model, e.g., Stable Diffusion [67]. Specifically, we modify and fine-tune the text-to-image model [4,98] to enable conditioning on posed SMPL-X, such that the model creates detailed normal maps according to both the pose information and textual descriptions. Afterward, we pass the generated frontal normal map as guidance to the dual normal map generator to complete dual normal maps, seamlessly connecting text-based generation to our original pipeline.\nTrained from posed 3D scans only, Chupa is capable of generating various digital avatars from pose and textual information, with realistic, high-fidelity features such as wrinkles and large varieties in human identity and clothing. We evaluate our method through established benchmarks along with a perceptual study and show that our method outperforms the previous baseline. In summary, our contributions are:\n• A 3D generation pipeline that directly leverages the 2D image generation capability of diffusion models towards 3D reconstruction.\n• A diffusion-based normal map generation and refinement strategy for view-consistent normal maps, targeted for 3D generation.\n• A method to effectively allow text-based 3D full-body digital avatar creation, providing an intuitive scenario for digital avatar creation."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b13",
"b16",
"b17",
"b22",
"b34",
"b35",
"b36",
"b6",
"b23",
"b55",
"b51",
"b56",
"b88",
"b6",
"b23",
"b55",
"b6",
"b53",
"b73",
"b23",
"b55",
"b8",
"b29",
"b99",
"b11",
"b48",
"b9",
"b59",
"b22",
"b4",
"b12",
"b18",
"b20",
"b32",
"b49",
"b81",
"b83",
"b0",
"b94",
"b33",
"b41",
"b66",
"b71",
"b92",
"b51",
"b58",
"b19",
"b77",
"b93",
"b5",
"b10",
"b38",
"b92",
"b93",
"b42",
"b78",
"b13",
"b27",
"b47",
"b68",
"b70",
"b80",
"b52",
"b63",
"b65",
"b69",
"b65",
"b60",
"b84",
"b89",
"b95",
"b89",
"b74",
"b86"
],
"table_ref": [],
"text": "3D Generative Models. Leveraging the success of generative models in producing realistic 2D images [15,18,19,24,[36][37][38], several efforts have been made to build 3D generative models from 2D datasets while ensuring view consistency [7,8,25,57]. To achieve this, 3D neural implicit representation [53,58,90] is employed to represent 3D targets, along with volume rendering to project the 3D scenes into 2D images [7,8,25,57]. While early methods in this direction were mainly focused on rigid objects [7,55,75] or human faces [8, 25,57], recent work has extended to human bodies by using LBS-based canonicalization [9] with SMPL to handle articulated pose changes [2, 31,101]. However, these approaches suffer from low-quality 3D outputs and high computational costs due to the volume rendering.\nOther methods [13,50] utilized SMPL models with latent codes to represent clothing information. However, these methods tend to be limited in geometric detail. gDNA [10] was the first generative model-based approach along with a neural implicit representation [61] to create diverse 3D humans with varying identities, poses, and clothing. gDNA further leverages the adversarial loss [24] to generate detailed surface normals. However, the adversarial loss made the model susceptible to mode collapse, which leads to unnatural stochastic details. In contrast, our approach is based on diffusion probabilistic models, which alleviates the mode collapsing issue while producing state-of-the-art quality.\n3D Human Reconstruction. The reconstruction of 3D humans has been a long-standing problem in the field of 3D computer vision. Traditional multi-view approaches tended to rely on calibrated multi-camera systems [5,14,20,22,32,34,51,83,85]. Several 3D parametric human body models [1,33,48,96] have been presented to represent the shape and pose variation of humans through parametric control, and they are widely used in human pose estimation [35,43,68]. Building upon such parametric models, single image-based 3D clothed human reconstruction methods with implicit 3D representation [73,74] show outstanding results with high-frequency details. Such models, however, tend to show disembodied or broken limbs for unseen poses due to the lack of topological prior. To address the problem, recent works [94,103] combine implicit representation [53] and parametric models [48,60]. Inspired by sandwich-like approaches [21,79], ECON [95] exploits front and back normal maps to build partial surfaces through normal integration [6] and stitches them with a mesh from IF-Net [11] and SMPL mesh through poisson surface reconstruction [40,41]. Our approach achieves realistic 3D human generation via normal map-based mesh optimization with SMPL-X mesh as a prior. Rather than using the parametric model as an implicit guidance [94,103] or stitching it with separate surfaces [95], we directly deform the SMPL-X mesh to be consistent with the input normal maps, using a differentiable rasterizer [44].\nDiffusion Models. Diffusion Probabilistic Models [80] are a group of generative models that have achieved state-of-theart results in perceptual image quality and mode coverage [15,29,49,70,72,82]. Recent diffusion models for textto-image generation [54,65,67,71] have demonstrated the ability to produce high-quality images based on textual input. Among them, Rombach et al. [67] enhances the efficiency of diffusion models by operating in a latent space that has a lower dimension than the image space while being perceptually equivalent. We list details of the inner workings of the diffusion models in the supplementary material.\nPrevious methods [62,86,91,97] focused on text-toshape tasks, where the output is a small 3D object lacking photorealistic quality. Among such methods, 3DiM [91] presents view-consistent generation through stochastic con-ditioning but is limited to expressing 3D objects in a 128 resolution. DiffuStereo [76] was one of the first methods to achieve high-quality 3D human reconstruction through diffusion models, but the usage of diffusion models was limited to refining details, while ours better utilizes the generation capability and mode coverage in generating diverse 3D models. Other work such as Rodin [88] also uses textual conditions to generate human 3D models, but are limited to the upper body, being unable to represent various human poses."
},
{
"figure_ref": [],
"heading": "Method",
"publication_ref": [
"b90",
"b50",
"b65"
],
"table_ref": [],
"text": "Our model is capable of generating 3D full body human models by conditioning on a front normal map rendered from a SMPL-X [48, 60] mesh M, which provides pose information, and an optional textual description that includes other identity-related information. The resulting 3D clothed human models display realistic details, while maintaining consistency to the input pose and textual description.\nConditioned on the normal map rendered from SMPL-X mesh, we first utilize a diffusion-based generative model to create full body normal maps for both frontal (observed) and backside (occluded) regions (Sec. 3.1). We then employ a normal map-based mesh optimization method inspired by NDS [92] to deform the posed SMPL-X mesh into a detailed human mesh (Sec. 3.2). To enhance the quality of our mesh, we render the normal maps from the resulting human mesh at multiple viewpoints and refine them through a diffusion-based resampling strategy [52], where we use separate diffusion models for the full body and facial regions (Sec. 3.3). The refined normal maps are subsequently used as inputs to our mesh optimization method, creating a highquality 3D clothed digital avatar. Our pipeline also accepts additional text information to further control the identity of the digital avatar using a text-to-image diffusion model [67] (Sec. 3.4). Fig. 2 shows the overall pipeline of our method."
},
{
"figure_ref": [
"fig_1"
],
"heading": "Dual Normal Map Generation",
"publication_ref": [
"b19",
"b77",
"b93",
"b65",
"b65",
"b15",
"b82",
"b25"
],
"table_ref": [],
"text": "Following the intuition of \"sandwich-like\" approaches for single image-based 3D human reconstruction [21,79,95], we generate both the frontal and backside normal map (x F , x B ) of clothed humans, dubbed dual normal maps, with the front-view SMPL-X normal map c N (β β β, θ θ θ) as a pose condition, where β β β, θ θ θ are the shape parameters and pose parameters of SMPL-X, respectively. We demonstrate that dual normal maps have sufficient information to generate plausible 3D humans with our normal map-based mesh reconstruction method. By generating dual normal maps, we can mitigate the difficulty and computational cost of directly generating 3D representation (e.g., voxels, point clouds, etc.) or multi-view consistent 2D representation (e.g., RGB images, normal maps, etc.). Since dual normal maps can be represented as images, we can exploit a diffusion model renowned for its image generation capability. We employ a Overview. Chupa takes a posed SMPL-X mesh M and its front normal map cN as input. At the first stage, Chupa generates frontal and backside clothed normal maps, x F , x B , conditioned on cN . These normals are then used as a reference to \"carve\" M through our normal map-based mesh optimization process. To further increase the quality, we separately refine the multi-view normal maps rendered from the full body and facial regions through a resampling procedure and perform the second optimization to create Mfinal. Our pipeline can also support identity control through a text description by leveraging the power of a text-to-image generation model. latent diffusion model [67] and adapt it to generate the dual normal maps. Note that we can control the body shape and pose of the generated dual normal maps by changing β β β, θ θ θ with the SMPL-X normal map c N (β β β, θ θ θ) as a condition.\nFollowing the latent diffusion model [67], we first train a vector-quantized autoencoder (E, D) [17,84] to support normal maps with alpha channels which enable getting foreground mask of generated normal maps easily. Specifically, given a normal map (color-coded as RGB) with alpha channel x ∈ R H×W ×4 , the encoder E encodes x into the latent representation z ∈ R h×w×4 , and the decoder D reconstructs a normal map back from the latent z. We train our autoencoder based on rendered normal maps from views with different yaw angles so that the autoencoder efficiently encodes these normal maps into a perceptually equivalent latent space, i.e., z F = E(x F ) and z B = E(x B ). For simultaneous generation, we concatenate the two latent codes z F and z B into a latent code z and treat it as an 8-channel image.\nDuring training, the latent code z is perturbed by the forward diffusion process according to a timestep t, producing a noisy latent code z t . The diffusion model ϵ ϵ ϵ θ then learns to predict the perturbed noise ϵ ϵ ϵ of z t , given the SMPL-X normal map condition c N (β β β, θ θ θ) ∈ R H×W ×4 . which is also encoded into E(c N ) ∈ R h×w×4 and concatenated with z t channelwise. The corresponding objective becomes\nL dual = E x F ,x B ,c N ,ϵ ϵ ϵ∼N (0,I),t [∥ϵ ϵ ϵ-ϵ ϵ ϵ θ (z F t , z B t , t, E(c N ))∥ 2 2 ].(1)\nAt inference time, we start from the Gaussian noise z T ∼ N (0, I) and iteratively sample from the previous step until z 0 , then we decode z 0 to get the final frontal and backside normal maps. We use classifier-free guidance [27] to boost the sample quality during conditional generation. To enable classifier-free guidance, we randomly assign blank latent embeddings to the conditional image c N with 10% probability during training. Then, for each inference step, we use the following modification to predict the denoised latent code:\nε ϵ ϵ θ (z t , t, E(c N )) = λϵ ϵ ϵ θ (z t , t, E(c N )) + (1 -λ)ϵ ϵ ϵ θ (z t , t),(2)\nwhere λ specifies the guidance strength that can be controlled during inference, and ϵ ϵ ϵ θ (z t , t, E(c N )) and ϵ ϵ ϵ θ (z t , t) each corresponds to the conditional and unconditional predictions. In Fig. 3, our simultaneous dual generation scheme shows that the generated frontal and backside normal maps are more consistent, compared to separate generation."
},
{
"figure_ref": [],
"heading": "Mesh Reconstruction with Front/Back Normals",
"publication_ref": [
"b90",
"b42",
"b61",
"b90",
"b90"
],
"table_ref": [],
"text": "Given the initial posed SMPL-X mesh M(β β β, θ θ θ) and the generated clothed normal maps (x F , x B ), we deform the initial mesh into a detailed 3D human mesh through iterative optimization. Our mesh reconstruction method is motivated by Neural Deferred Shading (NDS) [92], which reconstructs geometry from multi-view RGB images using a differentiable rasterizer and neural shader. Unlike NDS, we remove the neural shader as the generated normal maps provide su- pervision for geometry, and directly optimize the 3D geometry by comparing the normal maps with the geometry buffers rendered from a differentiable rasterizer [44]. In general, mesh reconstruction via the two normal maps is an ill-posed problem due to the depth ambiguity. Using SMPL-X mesh as an initial mesh, which is a strong geometric prior, and introducing a novel side loss L sides for regularizing side-views, we can reconstruct plausible 3D geometry of humans while mitigating the difficulty of generating multi-view consistent images at once. Our total objective is defined as\nL = λ normal L normal + λ mask L mask + λ sides L sides +λ laplacian L laplacian + λ reg normal L reg normal .(3)\nNormal map loss. We minimize the difference between the input normal maps (x F , x B ) and the normal maps rendered from the front/back views of the human mesh (N F , N B ) through a L 1 loss, denoted as L normal . We also minimize the discrepancy between the mask of the normal maps through a L 2 loss, L mask , to match the silhouette of the mesh. Note that we can acquire the masks of the generated normal maps by a simple thresholding on the alpha channel.\nSide loss. Since our initial 3D reconstruction is based on frontal/backside normal maps, the left/rightside regions of the human body tend to contain depth ambiguity [63]. We therefore introduce a novel side loss, which ensures that the body masks rendered from the side views ( Mleft , Mright ) are not shrinked into the side views of the initial SMPL-X mesh (M smpl left , M smpl right ). The loss function becomes we can mitigate the problem to some extent with the 3D prior from initial SMPL-X, we further prevent the optimized mesh from having unrealistic side-views.\nL sides = M smpl view [h,w]=1 ∥M smpl view [h, w] -Mview [h, w]∥ 2 2 ,(4)\nGeometric regularization. As noted by NDS [92], optimizing the mesh based on only the aforementioned loss terms can lead to degenerated mesh due to unconstrained vertex movement. To overcome this issue, we use geometric regularization terms following NDS [92]. Given a matrix V ∈ R n×3 with vertex positions of mesh M as rows, the Laplacian term is defined as\nL laplacian = 1 n n i=1 ∥δ δ δ i ∥ 2 2\n, where δ δ δ i = (LV ) i ∈ R 3 are the differential coordinates of vertex i with the graph Laplacian L. Since the differential coordinates are the sum of positional difference between its neighbors, minimizing this loss leads to a smoother mesh. We also introduce a normal consistency term, defined as\nL reg normal = 1 | F | (i,j)∈ F (1 -n n n i •n n n j ) 2\n, where F is the set of mesh face pairs with a shared edge and n n n i ∈ R 3 is the normal of triangle i. Minimizing the cosine similarity between face normals of neighbors encourages further smoothness."
},
{
"figure_ref": [
"fig_2",
"fig_4"
],
"heading": "Refine by Resampling",
"publication_ref": [
"b50",
"b58",
"b16"
],
"table_ref": [],
"text": "Resampling multi-view normal maps. After the initial mesh reconstruction, we can further improve the mesh while we already have plausible one. We refine the 3D human mesh by refining the rendered multi-view normal maps of the reconstructed mesh without losing view consistency. The refined maps are then used as inputs to the 3D reconstruction pipeline, creating an improved, realistic 3D human mesh.\nOur pipeline is inspired by SDEdit [52], which proposes an image translation method by progressively denoising a noise-perturbed image. The amount of noise perturbation is decided by timestep 0 < t 0 < 1, and as t 0 gets closer to 0, the operation focuses on editing the finer details. We repeat this process by K times to improve fidelity without harming the original information. To preserve the original structure while adjusting any unrealistic information, we set t 0 = 0.02 and K = 2, which we empirically found to be sufficient.\nIn practice, we first render a collection of n-view normal maps {I1 , I 2 , ..., I n } by evenly rotating the yaw camera angle around the 3D mesh. For refinement, we use the same dual normal map generation model in Sec. 3.1, which uses the normal map of posed SMPL-X as spatial guidance. We pair the rendered normal maps so that each pair is rendered from the backside of one another, and use the SMPL-X normal map corresponding to the frontal normal map as the condition to the diffusion model. This perturb-and-denoise process, which we call resampling, drives the normal maps rendered from the optimized mesh into the distribution of normal maps rendered from training 3D scans on which our diffusion model is trained, thus the normal maps become more realistic without losing overall semantics. Once the resampling is complete, we pass the refined normal maps as inputs to the 3D reconstruction stage (Sec. 3.2) to produce a refined 3D human model. Fig. 4 shows that our resamplingbased refinement produces more natural details.\nFacial resampling. We enhance the facial details of the optimized mesh by refining the normal maps rendered from the facial regions of the mesh. We train a latent diffusion model which shares the same architecture of the dual normal map generation model in Sec. 3.1, but trained on normal maps with face close-up. The close-up is done for the head vertices of SMPL-X based on the pre-defined part segmentation [60]. With the face close-up views, we can render facial regions of 3D scans and aligned SMPL-X mesh.\nGiven the aligned facial normal maps, we can train the diffusion model which generates the frontal and backside facial normal maps with facial normal maps of SMPL-X as a condition. We then apply the same resampling technique used for the full body to refine the multi-view facial normal maps rendered from the optimized mesh. Fig. 5 shows how the facial region is perceptually refined without harming the original structure. Unlike the method of Frühstück et al. [18], which performs offline optimization to blend a full body image and face image, we just do the normal map-based optimization (Sec. 3.2) with refined normal maps of both body and face, which aggregates the refined normal maps directly in 3D to generate a 3D human mesh with better details."
},
{
"figure_ref": [
"fig_5",
"fig_5"
],
"heading": "Text-guided Normal Map Generation",
"publication_ref": [
"b65",
"b3",
"b96",
"b67",
"b47"
],
"table_ref": [],
"text": "In addition to the main, pose-conditional 3D generation pipeline, we also include an optional pose-and-text conditional pipeline to further control the identity of the resulting human mesh. To generate 3D human mesh based on a textual description, we adopt a powerful text-to-image diffusion model, e.g., Stable Diffusion [67], and fine-tune its weights to generate normal maps that are consistent to the text description and the posed SMPL-X normal map.\nAs the method of Wang et al.\n[89] displayed the effectiveness of fine-tuning large diffusion models for image translation tasks, we initialize the weights of our model based on a pre-trained Stable Diffusion checkpoint, leveraging its renowned generation capabilities. Following previous works [4,98], we add additional input channels to the first layer of the U-Net [69] and initialize their weights to zero. We also use the same text conditioning based on a pre-trained CLIP model [64].\nAs shown in Fig. 6, our model supports the generation of detailed normal maps based on the textual description and the posed SMPL-X. Our method is the first method to support text-based full-body normal map generation by basing on Stable Diffusion.\nFrontal normal map-guided generation. To get dual normal maps based on the frontal normal map generated from the text-based normal map generation model, we follow the intuitions of Repaint [49]. Since we already know and want to preserve the frontal shape, the goal here is to predict the unknown backside normal map, based on the frontal normal map. For each inference step, we sample the intermediate frontal latent code z F t from the original latent z F at any timestep t, since the diffusion process is defined by a Gaussian Markov chain. In contrast, we sample the unknown, intermediate backside latent code z B t through reverse diffusion, which is concatenated channel-wise to z F t . Since we consider both z F t and z B t as a single, 8-channel latent code, the diffusion model leverages the context of the known frontal normal map while generating the unknown backside normal map, making this a channel-wise inpainting approach. Fig. 6 shows that our approach helps to generate backside normal maps that match the original frontal map. Through frontal normal map-guided dual normal map generation, we can seamlessly connect the generative powers of a text-to-image model with our main pipeline."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [
"b9",
"b64",
"b97",
"b92",
"b44",
"b58",
"b43",
"b14",
"b9"
],
"table_ref": [],
"text": "In this section, we validate Chupa's effectiveness in generating realistic 3D humans. We first compare Chupa with the previous state-of-the-art through an image quality metric and a perceptual user study. We also conduct ablation studies to illustrate the effectiveness of each part of our pipeline. Fig. 7 shows comparison of generated results from our method and the baseline [10].\nDatasets. We train and test our model with Renderpeople [66] and THuman 2.0 [99] dataset, which consists of 500, 526 scans with various identities and clothing. We split both datasets with a 9:1 ratio for train/test split. For training, we render 36 multi-view normal maps of the train split scans with rotation of 10 • yaw interval. We follow ICON [94] for rendering pipeline, originally from MonoPort [46], both for body and face. For rendering normal maps of facial regions, we use the pre-defined part segmentation label of SMPL-X [60] to find head vertices of fitted SMPL-X. Then, we render the facial region of 3D scans and fitted SMPL-X mesh with a weak perspective camera for rendering the head vertices of SMPL-X mesh with close-up. To create text pairs from normal maps for Stable Diffusion fine-tuning, we adopt an off-the-shelf image tagger model [45] based on ViT [16].\nBaseline. We compare our method with gDNA [10] as a baseline. gDNA is the state-of-the-art method to generate 3D human mesh with given SMPL-X parameter β, Θ and randomly sampled shape latent code z shape and detail latent code z detail from its learned latent space."
},
{
"figure_ref": [],
"heading": "Implementation Details",
"publication_ref": [
"b15",
"b65",
"b67",
"b13",
"b65",
"b9"
],
"table_ref": [],
"text": "Autoencoder model training. Before training the fullbody dual generation model, we trained the autoencoder model (E, D) for 1, 000 epochs on 4× NVIDIA A100 GPUs following the original implementation [17]. We used a VQ-regularized autoencoder with downsampling factor f = 4 and channel dimension c = 4 such that, given a full-body normal map image with alpha transparency (c N ∈ R 512×512×4 ), the encoder transforms the image to a latent code with 4 channels (E(c N ) ∈ R 128×128×4 ), and the decoder reconstructs the image from the latent code. For training, we used the full-body normal map datasets, following the same preprocessing listed in the main paper. We used the pretrained weights for the autoencoders of facial generation models (Sec. 3.3) and text-based generation models (Sec. 3.4) provided by the original paper [67]. For the facial generation model, we used a VQ-regularized autoencoder with downsampling factor f = 4 and channel dimension c = 3. For textual generation models, we used a KL-regularized autoencoder with downsampling factor f = 8 and channel dimension c = 4. All autoencoders were frozen during diffusion training. U-net. We adapt the U-Net [69] architecture for our diffusion models to support our dual-generation scheme. Specifically, we follow the approach of Dhariwal and Nichol [15] to further improve the sampling quality and set the input channels from 6 to 12, and the output channels from 3 to 8. By utilizing the concatenation of two input images (front and back) with the SMPL latent code E(c N ) for conditioning, we can treat them as a single input. As a result, we can obtain two spatially aligned images for both views at the same time. For the facial generation models, we set the input channels to 9 and output channels to 6, since we used 3-channel for facial normal maps.\nDual normal map generator training. We train our fullbody dual normal map generation model for 500 epochs with batch size 16 on 4× NVIDIA A100 GPUs. We set the total timesteps T = 100 with a linear variance schedule. During inference, we use the same 512×512 resolution and generate results with the same denoising steps used during training. We trained the facial generation model for 300 epochs with the same training settings.\nText-guided normal map generator training. We train our text-based normal map generator for 1, 000 epochs on 4× NVIDIA A100 GPUs. We train at a 512 × 512 resolution with a total batch size of 64. We initialize our model from the EMA weights of the Stable Diffusion [67] checkpoints and adopt other training settings from the public Stable Diffusion code base. After inference, we used a thresholding operation Figure 7. Generation Comparison. We display the visual comparisons between gDNA [10] and Chupa with the same SMPL input. Note that gDNA tends to amplify the unnatural artifacts from its coarse stage to the fine stage, while our results produce more natural results. "
},
{
"figure_ref": [],
"heading": "Quantitative Results",
"publication_ref": [
"b9",
"b9",
"b75",
"b100",
"b9",
"b76",
"b100"
],
"table_ref": [],
"text": "We conduct a quantitative evaluation of the quality of generated meshes, based on given SMPL-X parameters. We generated 3D human meshes with SMPL-X parameters fitted to 103 test scans, i.e. 50 from Renderpeople and 53 from THuman 2.0, for both our method and gDNA [10]. Following the previous work [10,77,102], we render normal maps [10] and shading-images [78,102] of groundtruth scans and generated meshes into 18 views with 20 • yaw interval, and compute FID score with them, which denoted as FID normal and FID shade respectively. Tab. 1 shows that our method achieves lower FID for both images than the baseline."
},
{
"figure_ref": [],
"heading": "User Preference",
"publication_ref": [],
"table_ref": [],
"text": "We carry out a perceptual study over 78 subjects asking about their preference between the meshes from our method and gDNA. We randomly select 40 from a set of SMPL-X parameters fitted to 103 test scans. We randomly generate meshes based on them with our method and gDNA, and render shading-images in 3 views, 0 • , 120 • , 240 • for full body images and 0 • , 40 • , -40 • for face images. Note that we use the narrower field-of-view for better comparing facial details. Tab. 1 shows that the users preferred meshes from our method both for full-body and face images. We present more details in the supplementary material."
},
{
"figure_ref": [
"fig_1",
"fig_6",
"fig_6",
"fig_2",
"fig_4"
],
"heading": "Ablation Study",
"publication_ref": [],
"table_ref": [],
"text": "We validate the building blocks of our pipeline through an ablation study. The evaluation is based on the same test split. The results are summarized in Tab. 3. Front/Back normal map generation. To validate the effectiveness of our dual normal map generation method, we separately generate frontal and backside normal maps with the SMPL normal map in the corresponding view. Due to the randomness of the diffusion model, we cannot guarantee the separately generated frontal and backward normal maps are consistent (Fig. 3), which leads to performance loss.\nSide loss. With the sidewise loss L sides from Eq. ( 4), we enforce our mesh to keep better alignment with the SMPL-X prior during mesh optimization (Sec. 3.2). Fig. 8 shows the effect of utilizing L sides . The first column shows the side-view normal map rendered from the mesh optimized with dual normal maps. The second column shows the same side-view normal map but overlapped with the side-view of the corresponding SMPL-X. The third column shows the normal maps after resampling (Sec. 3.3). Fig. 8a shows that the optimized mesh without L sides has worse alignment with SMPL-X mesh, which leads to the artifacts on resampling results. Tab. 3 demonstrates the inclusion of L sides leads to lower FID scores, indicating its effectiveness.\nRefinement. To validate the effectiveness of our refinement method (Sec. 3.3), we compare 3D generation results only optimized by front/back normal maps and the results refined by body refinement and additional face refinement. Fig. 4 and Fig. 5 show that our refinement methods lead to more realistic generation results. As expected, Tab. 3 shows that our face refinement method further reduces FID."
},
{
"figure_ref": [],
"heading": "Discussion",
"publication_ref": [],
"table_ref": [],
"text": "We propose Chupa, a powerful 3D generation pipeline for a large variety of dressed 3D high-quality digital avatars. By combining diffusion models for normal map generation We present the side-view normal maps of the optimized mesh (left), the normal maps overlapped on the SMPL-X normal maps (middle), and the normal maps after resampling (right). Without L sides , the alignment between the SMPL-X mesh and the optimized mesh becomes worse, leading to artifacts on the resampling result. (Note that the blue channel of the overlapped SMPL-X normal map is flipped for visualization purposes.) with a normal map-based mesh reconstruction method, our pipeline enables the creation of realistic 3D avatars with high levels of stochastic details. We also allow the creation of 3D humans from both pose and textual information, providing an intuitive method of digital avatar creation.\nWe note that while our pipeline can support text conditioning without losing visual quality, several elements that can be generated from the initial text-to-image model (e.g., bracelet, necklace, glasses) tend to be lost during the later stage of the pipeline and cannot be expressed at the final 3D model. For future work, we look forward to creating digital avatars with photorealistic textures and devising novel strategies for creating animations from our digital avatars."
},
{
"figure_ref": [
"fig_9"
],
"heading": "A. Detailed formulation of Diffusion Models",
"publication_ref": [
"b7",
"b78",
"b7",
"b7",
"b67",
"b70",
"b90",
"b2",
"b90",
"b21",
"b2",
"b90",
"b90"
],
"table_ref": [],
"text": "We provide a detailed introduction to Gaussian-based diffusion models [28,80]. Given the target data distribution x 0 ∼ q(x 0 ), the goal of diffusion models is to learn a model distribution p θ that approximates q, while being easy to sample from. To achieve both objectives, diffusion models define a forward process that gradually introduces noise to the original data x 0 to generate a sequence of noised data x 1 , x 2 , ..., x T . Additionally, a reverse process is defined, which aims to denoise the noised data x t and produce less noisy data x t-1 . Once trained, Gaussian-based diffusion models sample data x 0 by first sampling x T from a Gaussian distribution N (0, I) and iteratively sampling x t-1 from the previous step x t . To ensure x T ∼ N (0, I), it is required for T to be sufficiently large.\nThe forward process is formulated as a Markov chain according to a variance schedule β 1 < β 2 < ... < β T :\nq(x t |x t-1 ) := N (x t ; 1 -β t x t-1 , β t I)(5)\nq(x 1:T |x 0 ) := T t=1 q(x t |x t-1 )(6)\nNote that to sample x t ∼ q(x t |x 0 ), it is not required to apply forward diffusion t times. Instead, using the notation α t := 1 -β t and ᾱt := t s=1 α s , we have a closed form expression:\nq(x t |x 0 ) := N (x t ; √ ᾱt x 0 , (1 -ᾱt I)(7)\nConsequently, we can view x t as a linear combination of x 0 and ϵ ∼ N .(0, I)(x t = √ ᾱt x 0 + (1 -ᾱt )ϵ) Given the fixed forward process, p is designed to approximate the unknown true posterior q(x t-1 |x t ). This is achieved through the use of a deep neural network with learnable parameters θ.\np θ (x t-1 |x t ) := N (x t-1 ; µ θ (x t , t), Σ θ (x t , t)) (8) p θ (x 0:T ) := p(x T ) T t=1 p θ (x t-1 |x t )(9)\nHo et al. [28] proposed a specific parameterization for µ θ (x t , t) such that the neural network outputs the estimated noise ϵ θ instead of predicting µ θ .\nµ θ (x t , t) = 1 √ α t (x t - 1 -α t √ 1 -ᾱt ϵ θ (x t , t))(10)\nFor training, the variational lower bound is optimized and simplifies the following Eq. ( 11) that enables the model to learn how to predict the added noise.\nL simple = E x0,t,ϵ∼N (0,I) [||ϵ -ϵ θ (x t , t)|| 2 ](11)\nIn practice, Ho et al. [28] uses a U-Net backbone [69] to output the predicted noise ϵ θ which has the same dimensionality as the input noisy sample x t . To solve an imageto-image translation task, Saharia et al. [72] concatenates a spatial conditioning input y to x t channel-wise and modifies the learning objective as Eq. ( 12). Camera parameters. In our normal map-based mesh optimization method, we require camera parameters to rasterize the mesh into normal maps that are aligned with those generated from our dual-generation diffusion model. To generate the frontal normal map of the initial SMPL-X mesh (explained in Sec. 3.1), we utilize a weak perspective camera which shares the same parameters as our training data setup.\nFor the second mesh refinement stage (explained in Sec. 3.3), we also employ weak perspective cameras that are defined in the same manner for both body and face rendering.\nCoarse-to-fine optimization. We adopt the coarse-to-fine optimization strategy presented by NDS [92] for mesh optimization. Specifically, we begin with a coarse mesh and progressively increase the resolution through a remeshing technique, presented by Botsch and Kobbelt [3]. As demonstrated in [92], initializing optimization with a large number of vertices can lead to meshes with undesired geometry, such as degenerate triangles and self-intersections. Therefore, we start the optimization from a decimated version of our initial SMPL-X, which contains 3,000 vertices [23]. During optimization, for every 500 iterations, we apply remeshing [3] to increase the model resolution. It is worth noting that each iteration corresponds to a single gradient descent step, with respect to the loss based on a randomly sampled normal map. Following NDS [92], we perform optimization for a total of 2,000 iterations and decreased the gradient descent step size for the vertices by 25% after each remeshing. As Fig. 9 shows, we can handle the large deviation from the initial mesh without losing high-frequency details, due to the coarse-to-fine optimization scheme.\nLoss weight scheduling. While we follow the individual loss objective terms and scheduling of NDS [92] for our mesh optimization loss in Sec. 3.2, we added our side loss term L sides to the objective with weight term λ sides = 0.1, which we decrease by 10% after each remeshing. We also set the loss weights for L normal equivalent to L shading in the original paper for NDS. During optimization, we progressively increase the geometric regularization term Starting from a decimated SMPL-X mesh, we perform optimization in a coarse-to-fine manner. By increasing the resolution of the mesh for every 500 iterations, we progressively deform the mesh to match the input normal maps, without losing high-frequency details.\nL laplacian , L reg normal to encourage the generation of smooth surfaces for the final mesh. For the second mesh refinement stage, which optimizes the earlier mesh based on the refined normal maps from multiple views (total of 36 views), we set λ sides = 0 since the side views can now be well constrained without the sidewise loss.\nRefine by resampling. To refine the mesh from dual normal map-based optimization, we render both full body and face normal maps and refine them with resampling technique (Sec. 3.3). Here, we render 36-view normal maps with 10 • yaw interval, and set (t 0 , K) to (0.02, 2), respectively, both for body and face normal map refinement."
},
{
"figure_ref": [
"fig_10",
"fig_11",
"fig_12"
],
"heading": "C. Qualitative Results",
"publication_ref": [
"b57",
"b58",
"b28",
"b85"
],
"table_ref": [],
"text": "More generation results. Fig. 10 shows more random generation results from Chupa. We generate the human meshes based on SMPL-X parameters from the AGORA dataset [59], which includes SMPL, SMPL-X parameters fitted to 4, 240 3D human scans. We can generate human scans with various identities and can be generalized to diverse poses.\nChanging shape parameter β β β. To control the shape of the generated mesh, we can control the shape parameter β β β of input SMPL-X mesh [48,60]. Fig. 11, Fig. 12 shows the generated meshes according to the variation of β β β with fixed pose parameter θ θ θ, where β 1 , β 2 corresponds to the first and second component of the shape parameter respectively [48].\nComparison with AvatarCLIP. We compare our textguided generation results with AvatarCLIP [30], a textguided 3D avatar generation pipeline that also initializes its 3D implicit surface model [87] with a SMPL model. Once initialized, AvatarCLIP optimizes the 3D model based on a CLIP loss [64] on the rendered results, to match the 3D model according to the text description. Fig. 13 shows that Chupa can generate more realistic 3D human mesh while minimizing unnatural artifacts. Note that while AvatarCLIP takes more than 3 hours to generate a mesh, Chupa takes 3 minutes with a single RTX3090."
},
{
"figure_ref": [
"fig_13",
"fig_14",
"fig_14",
"fig_15"
],
"heading": "D. Failure Cases",
"publication_ref": [],
"table_ref": [],
"text": "Depth ambiguity problem. Our dual normal map-based mesh reconstruction method (Sec. 3.2) has inherent depth ambiguity issues, as it only uses front and back-view normal maps for the initial optimization. When the given normal maps largely deviates from the initial SMPL model, e.g., long hair, the vertices for both head and shoulder deforms to match the provided hairstyle, creating artifacts during deformation. Fig. 14 shows that while the hairstyle seems to be well-reconstructed in the front view, there exists unnatural seams and broken geometry at close view.\nFace direction matters. When the input pose contains misaligned body and face direction, the final output might display unnatural face geometry. For example, when the face is turned to the side direction (Fig. 15), the diffusion models might fail to generate realistic faces for reconstruction. To make matters worse, the small distortion due to depth ambiguity during reconstruction (Sec. 3.2) can have huge impact on the perceptual quality of faces. Fig. 15 shows an example of such cases, where the resulting face mesh displays unnatural geometry.\nOut-of-distribution pose. While our method can be generalized for diverse poses, there exists out-of-distribution (a) poses that the diffusion generative model fails to create plausible normal maps from. Fig. 16 shows such examples of unrealistic normal maps, which leads to 3D meshes with bad geometry.\nβ 1 = -2 (b) β 1 = -1 (c) β 1 = 0 (d) β 1 = 1 (e) β 1 = 2"
},
{
"figure_ref": [],
"heading": "E. User Study",
"publication_ref": [
"b9"
],
"table_ref": [],
"text": "We conduct a perceptual study asking user preference between the meshes from our method and gDNA [10]. We collect 100 participants through CloudResearch Connect [12] and get 78 valid answers out of them. Each participants are given 40 problems which consist of 20 problems for body and 20 problems for face. Fig. 17 shows the example problems."
},
{
"figure_ref": [],
"heading": "F. Ablation Study",
"publication_ref": [
"b9",
"b7",
"b37",
"b79"
],
"table_ref": [],
"text": "We present additional ablation study results on changing various hyperparameters such as resampling parameters, sampling angle, and the sampling scheme for dual generation. In Tab. 4 and Tab. 5, we present the effect of choosing different refinement parameters (t 0 , K) and the sampling angle during the refinement stage for both shaded and normal maps of the resulting meshes. We also present the effect of using different diffusion samplers in Tab. 6. User study problem example. The 3 views of mesh from our method and gDNA [10] with the same SMPL parameter are rendered as shading images. Each user is asked to choose more realistic shapes between two rows, where each row corresponds to the images from each method. Two rows are randomly shuffled.\nTable 5. Ablation on the number of views for refinement. We see the effects of the number of views for refinement with t0 = 0.02, K = 2 as fixed.\nN face, the smaller forward time steps and fewer iterations show better performance since large forward steps or many iterations may lead to the normal map inconsistent with the original normal maps.\nThe number of views for mesh refinement. Tab. 5 shows the performance with the varying number of views used for the mesh refinement stage (Sec. 3.3), where N views , θ step correspond to the number of views and the yaw interval between views respectively. Here, the hyperparameters (t 0 , K) for resampling are fixed as (0.02, 2). It shows that increasing the number of views leads to better performance.\nSampling scheme of the diffusion model. As mentioned in Sec. 4.1, we generate dual normal maps with the same denoising steps used during training, which is the sampling scheme of DDPM [28]. Here, we ablate on the different sampling schemes for diffusion probabilistic models, with two additional samplers [39,81] set to t = 50. Tab. 6 shows that the sampling scheme doesn't affect the performance significantly. Note that we compute the score without the mesh refinement stage (Sec. 3.3) to analyze the effects of the sampler since the refinement stage only involves a small number of denoising steps."
},
{
"figure_ref": [],
"heading": "",
"publication_ref": [],
"table_ref": [],
"text": "Acknowledgements. This work was supported by Naver Webtoon. The work of SNU members was also supported by SNU Creative-Pioneering Researchers Program, NRF grant funded by the Korean government (MSIT) (No. 2022R1A2C2092724), and IITP grant funded by the Korean government (MSIT) (No.2022-0-00156 and No.2021-0-01343). H. Joo is the corresponding author."
}
] | Figure 1. Generative Human Digital Avatars. We propose Chupa, a 3D human generation pipeline that combines the generative power of diffusion models [67] and neural rendering techniques [44] to create diverse, and realistic 3D humans. Our pipeline can easily generalize to unseen human poses and display realistic qualities. | Chupa: Carving 3D Clothed Humans from Skinned Shape Priors using 2D Diffusion Probabilistic Models | [
{
"figure_caption": "Figure2. Overview. Chupa takes a posed SMPL-X mesh M and its front normal map cN as input. At the first stage, Chupa generates frontal and backside clothed normal maps, x F , x B , conditioned on cN . These normals are then used as a reference to \"carve\" M through our normal map-based mesh optimization process. To further increase the quality, we separately refine the multi-view normal maps rendered from the full body and facial regions through a resampling procedure and perform the second optimization to create Mfinal. Our pipeline can also support identity control through a text description by leveraging the power of a text-to-image generation model.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 .3Figure 3. Separate generation vs. Dual generation. Comparison between (a) separate sampling for frontal/backside normal maps and (b) our dual sampling. When generated separately, attributes of two normal maps likely differ. However, generating the dual normal maps at once ensures the maps share the same semantics.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 .4Figure 4. Body Resampling. The initial 3D mesh displays undesired visual artifacts, such as unnatural cloth wrinkles and depth misprediction. By resampling, those artifacts are moderated to produce more natural results.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "(a) Rendered normal (b) Resampled normal",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 .5Figure 5. Face close-up resampling. Both images are aligned according to the SMPL-X vertices for the facial region. We can observe that the perceptibility of the faces is improved.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Figure 6 .6Figure 6. Text-based normal map generation. Note that our model is capable of generating a normal map consistent in gender, clothing, and hair style 1 . Moreover, our guided generation method can create a view-consistent back normal map from the initial frontal map, making it possible to use it for our original pipeline.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "Figure 8 .8Figure8. Side loss. We present the side-view normal maps of the optimized mesh (left), the normal maps overlapped on the SMPL-X normal maps (middle), and the normal maps after resampling (right). Without L sides , the alignment between the SMPL-X mesh and the optimized mesh becomes worse, leading to artifacts on the resampling result. (Note that the blue channel of the overlapped SMPL-X normal map is flipped for visualization purposes.)",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "8",
"figure_type": "figure"
},
{
"figure_caption": "Lsimple = E x0,y,t,ϵ∼N (0,I) [||ϵ -ϵ θ (x t , y, t)|| 2 ] (12) A.1. Diffusion Training B. Normal map-based mesh optimization",
"figure_data": "",
"figure_id": "fig_7",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "(a) iter. 0 (b) iter. 500 (c) iter. 1000 (d) iter. 1500 (e) iter. 2000",
"figure_data": "",
"figure_id": "fig_8",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 9 .9Figure 9. Coarse-to-fine optimization.Starting from a decimated SMPL-X mesh, we perform optimization in a coarse-to-fine manner. By increasing the resolution of the mesh for every 500 iterations, we progressively deform the mesh to match the input normal maps, without losing high-frequency details.",
"figure_data": "",
"figure_id": "fig_9",
"figure_label": "9",
"figure_type": "figure"
},
{
"figure_caption": "Figure 11 .11Figure 11. Changing shape parameter β β β1.",
"figure_data": "",
"figure_id": "fig_10",
"figure_label": "11",
"figure_type": "figure"
},
{
"figure_caption": "Figure 12 .12Figure 12. Changing shape parameter β β β2.",
"figure_data": "",
"figure_id": "fig_11",
"figure_label": "12",
"figure_type": "figure"
},
{
"figure_caption": "Figure 13 .13Figure 13. Comparison with AvatarCLIP. The left two columns are from AvatarCLIP, the right two columns are from Chupa (ours).",
"figure_data": "",
"figure_id": "fig_12",
"figure_label": "13",
"figure_type": "figure"
},
{
"figure_caption": "Figure 14 .14Figure 14. Depth ambiguity problem. Chupa may generate broken geometry, due to the depth ambiguity problem of our mesh reconstruction method(left: dual normal map, right: final mesh).",
"figure_data": "",
"figure_id": "fig_13",
"figure_label": "14",
"figure_type": "figure"
},
{
"figure_caption": "Figure 15 .15Figure 15. Face direction matters. Chupa may generate unnatural face geometry, when the face direction is not aligned with the input view (left: dual normal map, right: final mesh).",
"figure_data": "",
"figure_id": "fig_14",
"figure_label": "15",
"figure_type": "figure"
},
{
"figure_caption": "Figure 16 .16Figure 16. Out-of-distribution pose. Chupa may generate implausible geometry for some out-of-distribution pose (left: SMPL-X, middle: dual normal map, right: final mesh).",
"figure_data": "",
"figure_id": "fig_15",
"figure_label": "16",
"figure_type": "figure"
},
{
"figure_caption": "Figure 17. User study problem example. The 3 views of mesh from our method and gDNA[10] with the same SMPL parameter are rendered as shading images. Each user is asked to choose more realistic shapes between two rows, where each row corresponds to the images from each method. Two rows are randomly shuffled.",
"figure_data": "",
"figure_id": "fig_16",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Quantitative Evaluation. We report two types of FID scores for the test split of Renderpeople and Thuman 2.0.",
"figure_data": "gDNA coarse [10]53.7468.14gDNA fine [10]36.4345.57Ours21.9036.58on the 3rd channel of the image to create a transparency mapbefore the dual generation stage.",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "User preference. We carry out a perceptual study asking 78 subjects to choose a more realistic one between ours and gDNA fine .",
"figure_data": "MethodBodyFaceTotalgDNA fine 20.89% 18.7% 19.78%Ours79.11% 81.3% 80.22%",
"figure_id": "tab_1",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Ablation study. We do ablation study over our key components. We report FID normal score.dual. L sides refine body refine face FID normal ↓",
"figure_data": "30.55",
"figure_id": "tab_2",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Ablation study on resampling. We see the effects of (t0, K) both for body and face, with the number of views fixed as 36.",
"figure_data": "(t 0 , K)BodyFaceFID normal ↓ FID shade ↓(0.02, 2)-22.6137.13(0.02, 4)-26.6846.19(0.02, 6)-31.3951.98(0.04, 2)-27.0246.34(0.06, 2)-31.7152.65(0.02, 2) (0.02, 2)21.9036.58(0.02, 2) (0.02, 4)22.4237.57(0.02, 2) (0.02, 6)22.6538.11(0.02, 2) (0.04, 2)22.4137.64(0.02, 2) (0.06, 2)22.6537.94",
"figure_id": "tab_3",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "views θ step FID normal ↓ FID shade ↓ Refine by resampling. Tab. 4 shows the effects of varying (t 0 , K) for resampling. The first 6 rows show the results of varying (t 0 , K) for body normal map refinement without face refinement. And the next 6 rows show the results of varying (t 0 , K) for face normal map refinement with fixed (t 0 , K) for body normal map refinement. For both body and",
"figure_data": "490 •30.8841.85660 •29.0141.301230 •25.2139.533610 •21.9036.58",
"figure_id": "tab_4",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Ablation on sampling scheme. We ablate on the sampling scheme of our diffusion model for dual normal map generation. Here, we compute FID scores based on the results of dual normal map-based optimization without refinement.MethodFID normal ↓ FID shade ↓",
"figure_data": "Euler [39]28.8437.36DDIM [81]26.7634.79DDPM [28]26.3137.13",
"figure_id": "tab_5",
"figure_label": "6",
"figure_type": "table"
}
] | Byungjun Kim; Patrick Kwon; Kwangho Lee; Myunggi Lee; Sookwan Han; Daesik Kim; Hanbyul Joo | [
{
"authors": "D Anguelov; P Srinivasan; D Koller; S Thrun; J Rodgers; J Davis",
"journal": "",
"ref_id": "b0",
"title": "Scape: shape completion and animation of people",
"year": "2005"
},
{
"authors": "A W Bergman; P Kellnhofer; W Yifan; E R Chan; D B Lindell; G Wetzstein",
"journal": "NeurIPS",
"ref_id": "b1",
"title": "Generative neural articulated radiance fields",
"year": "2022"
},
{
"authors": "M Botsch; L Kobbelt",
"journal": "",
"ref_id": "b2",
"title": "A remeshing approach to multiresolution modeling",
"year": "2004"
},
{
"authors": "T Brooks; A Holynski; A A Efros",
"journal": "",
"ref_id": "b3",
"title": "Instructpix2pix: Learning to follow image editing instructions",
"year": "2023"
},
{
"authors": "T Brox; B Rosenhahn; J Gall; D Cremers",
"journal": "IEEE TPAMI",
"ref_id": "b4",
"title": "Combined region and motion-based 3D tracking of rigid and articulated objects",
"year": "2010"
},
{
"authors": "X Cao; H Santo; B Shi; F Okura; Y Matsushita",
"journal": "",
"ref_id": "b5",
"title": "Bilateral normal integration",
"year": "2022"
},
{
"authors": "E R Chan; M Monteiro; P Kellnhofer; J Wu; G Wetzstein",
"journal": "",
"ref_id": "b6",
"title": "pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis",
"year": "2021"
},
{
"authors": "E R Chan; C Z Lin; M A Chan; K Nagano; B Pan; S De Mello; O Gallo; L J Guibas; J Tremblay; S Khamis",
"journal": "",
"ref_id": "b7",
"title": "Efficient geometry-aware 3d generative adversarial networks",
"year": "2022"
},
{
"authors": "X Chen; Y Zheng; M J Black; O Hilliges; A Geiger",
"journal": "",
"ref_id": "b8",
"title": "Snarf: Differentiable forward skinning for animating nonrigid neural implicit shapes",
"year": "2021"
},
{
"authors": "X Chen; T Jiang; J Song; J Yang; M J Black; A Geiger; O Hilliges",
"journal": "",
"ref_id": "b9",
"title": "gdna: Towards generative detailed neural avatars",
"year": "2022"
},
{
"authors": "J Chibane; T Alldieck; G Pons-Moll",
"journal": "",
"ref_id": "b10",
"title": "Implicit functions in feature space for 3d shape reconstruction and completion",
"year": "2020"
},
{
"authors": "E Corona; A Pumarola; G Alenya; G Pons-Moll; F Moreno-Noguer",
"journal": "",
"ref_id": "b11",
"title": "Smplicit: Topology-aware generative model for clothed people",
"year": "2021"
},
{
"authors": "E De Aguiar; C Stoll; C Theobalt; N Ahmed; H.-P Seidel; S Thrun",
"journal": "SIGGRAPH",
"ref_id": "b12",
"title": "Performance capture from sparse multi-view video",
"year": "2008"
},
{
"authors": "P Dhariwal; A Nichol",
"journal": "NeurIPS",
"ref_id": "b13",
"title": "Diffusion models beat gans on image synthesis",
"year": "2021"
},
{
"authors": "A Dosovitskiy; L Beyer; A Kolesnikov; D Weissenborn; X Zhai; T Unterthiner; M Dehghani; M Minderer; G Heigold; S Gelly; J Uszkoreit; N Houlsby",
"journal": "",
"ref_id": "b14",
"title": "An image is worth 16x16 words: Transformers for image recognition at scale",
"year": "2021"
},
{
"authors": "P Esser; R Rombach; B Ommer",
"journal": "",
"ref_id": "b15",
"title": "Taming transformers for high-resolution image synthesis",
"year": "2020"
},
{
"authors": "A Frühstück; K K Singh; E Shechtman; N J Mitra; P Wonka; J Lu",
"journal": "",
"ref_id": "b16",
"title": "Insetgan for full-body image generation",
"year": "2022"
},
{
"authors": "J Fu; S Li; Y Jiang; K.-Y Lin; C Qian; C C Loy; W Wu; Z Liu",
"journal": "",
"ref_id": "b17",
"title": "Stylegan-human: A data-centric odyssey of human generation",
"year": "2022"
},
{
"authors": "Y Furukawa; J Ponce",
"journal": "CVPR",
"ref_id": "b18",
"title": "Dense 3d motion capture from synchronized video streams",
"year": "2008"
},
{
"authors": "V Gabeur; J.-S Franco; X Martin; C Schmid; G Rogez",
"journal": "",
"ref_id": "b19",
"title": "Moulding humans: Non-parametric 3d human shape estimation from single images",
"year": "2019"
},
{
"authors": "J Gall; C Stoll; E De Aguiar; C Theobalt; B Rosenhahn; H.-P Seidel",
"journal": "",
"ref_id": "b20",
"title": "Motion capture using joint skeleton tracking and surface estimation",
"year": "2009"
},
{
"authors": "M Garland; P S Heckbert",
"journal": "",
"ref_id": "b21",
"title": "Surface simplification using quadric error metrics",
"year": "1997"
},
{
"authors": "I Goodfellow; J Pouget-Abadie; M Mirza; B Xu; D Warde-Farley; S Ozair; A Courville; Y Bengio",
"journal": "NeurIPS",
"ref_id": "b22",
"title": "Generative adversarial nets",
"year": "2014"
},
{
"authors": "J Gu; L Liu; P Wang; C Theobalt",
"journal": "",
"ref_id": "b23",
"title": "StyleneRF: A style-based 3d aware generator for high-resolution image synthesis",
"year": "2022"
},
{
"authors": "K Guo; P Lincoln; P Davidson; J Busch; X Yu; M Whalen; G Harvey; S Orts-Escolano; R Pandey; J Dourgarian",
"journal": "ACM TOG",
"ref_id": "b24",
"title": "The relightables: Volumetric performance capture of humans with realistic relighting",
"year": "2019"
},
{
"authors": "J Ho",
"journal": "",
"ref_id": "b25",
"title": "Classifier-free diffusion guidance",
"year": "2022"
},
{
"authors": "J Ho; A Jain; P Abbeel",
"journal": "NeurIPS",
"ref_id": "b26",
"title": "Denoising diffusion probabilistic models",
"year": "2020"
},
{
"authors": "J Ho; C Saharia; W Chan; D J Fleet; M Norouzi; T Salimans",
"journal": "JMLR",
"ref_id": "b27",
"title": "Cascaded diffusion models for high fidelity image generation",
"year": "2022"
},
{
"authors": "F Hong; M Zhang; L Pan; Z Cai; L Yang; Z Liu",
"journal": "ACM TOG",
"ref_id": "b28",
"title": "Avatarclip: Zero-shot text-driven generation and animation of 3d avatars",
"year": "2022"
},
{
"authors": "F Hong; Z Chen; Y Lan; L Pan; Z Liu",
"journal": "",
"ref_id": "b29",
"title": "EVA3d: Compositional 3d human generation from 2d image collections",
"year": "2023"
},
{
"authors": "H Joo; T Simon; X Li; H Liu; L Tan; L Gui; S Banerjee; T Godisart; B Nabbe; I Matthews",
"journal": "IEEE TPAMI",
"ref_id": "b30",
"title": "Panoptic studio: A massively multiview system for social interaction capture",
"year": "2017"
},
{
"authors": "H Joo; T Simon; Y Sheikh",
"journal": "",
"ref_id": "b31",
"title": "Total capture: A 3d deformation model for tracking faces, hands, and bodies",
"year": "2018"
},
{
"authors": "T Kanade; P Rander; P Narayanan",
"journal": "IEEE Multimedia",
"ref_id": "b32",
"title": "Virtualized reality: Constructing virtual worlds from real scenes",
"year": "1997"
},
{
"authors": "A Kanazawa; M J Black; D W Jacobs; J Malik",
"journal": "",
"ref_id": "b33",
"title": "Endto-end recovery of human shape and pose",
"year": "2018"
},
{
"authors": "T Karras; S Laine; T Aila",
"journal": "",
"ref_id": "b34",
"title": "A style-based generator architecture for generative adversarial networks",
"year": "2019"
},
{
"authors": "T Karras; S Laine; M Aittala; J Hellsten; J Lehtinen; T Aila",
"journal": "",
"ref_id": "b35",
"title": "Analyzing and improving the image quality of StyleGAN",
"year": "2020"
},
{
"authors": "T Karras; M Aittala; S Laine; E Härkönen; J Hellsten; J Lehtinen; T Aila",
"journal": "NeurIPS",
"ref_id": "b36",
"title": "Alias-free generative adversarial networks",
"year": "2021"
},
{
"authors": "T Karras; M Aittala; T Aila; S Laine",
"journal": "",
"ref_id": "b37",
"title": "Elucidating the design space of diffusion-based generative models",
"year": "2022"
},
{
"authors": "M Kazhdan; H Hoppe",
"journal": "ACM TOG",
"ref_id": "b38",
"title": "Screened poisson surface reconstruction",
"year": "2013"
},
{
"authors": "M Kazhdan; M Bolitho; H Hoppe",
"journal": "",
"ref_id": "b39",
"title": "Poisson surface reconstruction",
"year": "2006"
},
{
"authors": "D P Kingma; M Welling",
"journal": "",
"ref_id": "b40",
"title": "Auto-encoding variational bayes",
"year": "2014"
},
{
"authors": "N Kolotouros; G Pavlakos; M J Black; K Daniilidis",
"journal": "",
"ref_id": "b41",
"title": "Learning to reconstruct 3d human pose and shape via modelfitting in the loop",
"year": "2019"
},
{
"authors": "S Laine; J Hellsten; T Karras; Y Seol; J Lehtinen; T Aila",
"journal": "ACM TOG",
"ref_id": "b42",
"title": "Modular primitives for high-performance differentiable rendering",
"year": "2005"
},
{
"authors": "S Lee",
"journal": "",
"ref_id": "b43",
"title": "Tagger for automatic1111's webui",
"year": "2022"
},
{
"authors": "R Li; Y Xiu; S Saito; Z Huang; K Olszewski; H Li",
"journal": "",
"ref_id": "b44",
"title": "Monocular real-time volumetric performance capture",
"year": "2020"
},
{
"authors": "S Lombardi; J Saragih; T Simon; Y Sheikh",
"journal": "ACM TOG",
"ref_id": "b45",
"title": "Deep appearance models for face rendering",
"year": "2018"
},
{
"authors": "M Loper; N Mahmood; J Romero; G Pons-Moll; M J Black",
"journal": "ACM TOG",
"ref_id": "b46",
"title": "Smpl: A skinned multi-person linear model",
"year": "2015"
},
{
"authors": "A Lugmayr; M Danelljan; A Romero; F Yu; R Timofte; L Van Gool",
"journal": "",
"ref_id": "b47",
"title": "Repaint: Inpainting using denoising diffusion probabilistic models",
"year": "2022"
},
{
"authors": "Q Ma; J Yang; A Ranjan; S Pujades; G Pons-Moll; S Tang; M J Black",
"journal": "",
"ref_id": "b48",
"title": "Learning to dress 3d people in generative clothing",
"year": "2020"
},
{
"authors": "T Matsuyama; T Takai",
"journal": "DPVT",
"ref_id": "b49",
"title": "Generation, visualization, and editing of 3d video",
"year": "2002"
},
{
"authors": "C Meng; Y He; Y Song; J Song; J Wu; J.-Y Zhu; S Ermon",
"journal": "",
"ref_id": "b50",
"title": "Sdedit: Guided image synthesis and editing with stochastic differential equations",
"year": "2021"
},
{
"authors": "L Mescheder; M Oechsle; M Niemeyer; S Nowozin; A Geiger",
"journal": "",
"ref_id": "b51",
"title": "Occupancy networks: Learning 3d reconstruction in function space",
"year": "2019"
},
{
"authors": "A Q Nichol; P Dhariwal; A Ramesh; P Shyam; P Mishkin; B Mcgrew; I Sutskever; M Chen",
"journal": "",
"ref_id": "b52",
"title": "Glide: Towards photorealistic image generation and editing with text-guided diffusion models",
"year": "2022"
},
{
"authors": "M Niemeyer; A Geiger",
"journal": "",
"ref_id": "b53",
"title": "Giraffe: Representing scenes as compositional generative neural feature fields",
"year": "2021"
},
{
"authors": "M Niemeyer; L Mescheder; M Oechsle; A Geiger",
"journal": "",
"ref_id": "b54",
"title": "Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision",
"year": "2020"
},
{
"authors": "R Or-El; X Luo; M Shan; E Shechtman; J J Park; I Kemelmacher-Shlizerman",
"journal": "",
"ref_id": "b55",
"title": "Stylesdf: High-resolution 3dconsistent image and geometry generation",
"year": "2022"
},
{
"authors": "J J Park; P Florence; J Straub; R Newcombe; S Lovegrove",
"journal": "",
"ref_id": "b56",
"title": "Deepsdf: Learning continuous signed distance functions for shape representation",
"year": "2019"
},
{
"authors": "P Patel; C.-H P Huang; J Tesch; D T Hoffmann; S Tripathi; M J Black",
"journal": "",
"ref_id": "b57",
"title": "Agora: Avatars in geography optimized for regression analysis",
"year": "2021"
},
{
"authors": "G Pavlakos; V Choutas; N Ghorbani; T Bolkart; A A Osman; D Tzionas; M J Black",
"journal": "",
"ref_id": "b58",
"title": "Expressive body capture: 3d hands, face, and body from a single image",
"year": "2019"
},
{
"authors": "S Peng; M Niemeyer; L Mescheder; M Pollefeys; A Geiger",
"journal": "",
"ref_id": "b59",
"title": "Convolutional occupancy networks",
"year": "2020"
},
{
"authors": "B Poole; A Jain; J T Barron; B Mildenhall",
"journal": "",
"ref_id": "b60",
"title": "Dreamfusion: Text-to-3d using 2d diffusion",
"year": "2023"
},
{
"authors": "Y Quéau; J.-D Durou; J.-F Aujol",
"journal": "Journal of Mathematical Imaging and Vision",
"ref_id": "b61",
"title": "Normal integration: a survey",
"year": "2018"
},
{
"authors": "A Radford; J W Kim; C Hallacy; A Ramesh; G Goh; S Agarwal; G Sastry; A Askell; P Mishkin; J Clark",
"journal": "",
"ref_id": "b62",
"title": "Learning transferable visual models from natural language supervision",
"year": "2021"
},
{
"authors": "A Ramesh; P Dhariwal; A Nichol; C Chu; M Chen",
"journal": "",
"ref_id": "b63",
"title": "Hierarchical text-conditional image generation with clip latents",
"year": "2022"
},
{
"authors": " Renderpeople",
"journal": "",
"ref_id": "b64",
"title": "",
"year": "2018"
},
{
"authors": "R Rombach; A Blattmann; D Lorenz; P Esser; B Ommer",
"journal": "",
"ref_id": "b65",
"title": "High-resolution image synthesis with latent diffusion models",
"year": "2022"
},
{
"authors": "Y Rong; T Shiratori; H Joo",
"journal": "",
"ref_id": "b66",
"title": "Frankmocap: A monocular 3d whole-body pose estimation system via regression and integration",
"year": "2021"
},
{
"authors": "O Ronneberger; P Fischer; T Brox",
"journal": "",
"ref_id": "b67",
"title": "U-net: Convolutional networks for biomedical image segmentation",
"year": "2015"
},
{
"authors": "C Saharia; W Chan; H Chang; C Lee; J Ho; T Salimans; D Fleet; M Norouzi",
"journal": "",
"ref_id": "b68",
"title": "Palette: Image-to-image diffusion models",
"year": "2022"
},
{
"authors": "C Saharia; W Chan; S Saxena; L Li; J Whang; E Denton; S K S Ghasemipour; R Gontijo-Lopes; B K Ayan; T Salimans",
"journal": "NeurIPS",
"ref_id": "b69",
"title": "Photorealistic text-to-image diffusion models with deep language understanding",
"year": "2022"
},
{
"authors": "C Saharia; J Ho; W Chan; T Salimans; D J Fleet; M Norouzi",
"journal": "IEEE TPAMI",
"ref_id": "b70",
"title": "Image super-resolution via iterative refinement",
"year": "2023"
},
{
"authors": "S Saito; Z Huang; R Natsume; S Morishima; A Kanazawa; H Li",
"journal": "",
"ref_id": "b71",
"title": "Pifu: Pixel-aligned implicit function for highresolution clothed human digitization",
"year": "2019"
},
{
"authors": "S Saito; T Simon; J Saragih; H Joo",
"journal": "",
"ref_id": "b72",
"title": "Pifuhd: Multilevel pixel-aligned implicit function for high-resolution 3d human digitization",
"year": "2020"
},
{
"authors": "K Schwarz; Y Liao; M Niemeyer; A Geiger",
"journal": "NeurIPS",
"ref_id": "b73",
"title": "Graf: Generative radiance fields for 3d-aware image synthesis",
"year": "2020"
},
{
"authors": "R Shao; Z Zheng; H Zhang; J Sun; Y Liu",
"journal": "",
"ref_id": "b74",
"title": "Diffustereo: High quality human reconstruction via diffusion-based stereo using sparse cameras",
"year": "2022"
},
{
"authors": "J R Shue; E R Chan; R Po; Z Ankner; J Wu; G ",
"journal": "",
"ref_id": "b75",
"title": "Wetzstein. 3d neural field generation using triplane diffusion",
"year": "2022"
},
{
"authors": "J R Shue; E R Chan; R Po; Z Ankner; J Wu; G ",
"journal": "",
"ref_id": "b76",
"title": "Wetzstein. 3d neural field generation using triplane diffusion",
"year": "2023"
},
{
"authors": "D Smith; M Loper; X Hu; P Mavroidis; J Romero",
"journal": "",
"ref_id": "b77",
"title": "Facsimile: Fast and accurate scans from an image in less than a second",
"year": "2019"
},
{
"authors": "J Sohl-Dickstein; E A Weiss; N Maheswaranathan; S Ganguli",
"journal": "",
"ref_id": "b78",
"title": "Deep unsupervised learning using nonequilibrium thermodynamics",
"year": "2015"
},
{
"authors": "J Song; C Meng; S Ermon",
"journal": "",
"ref_id": "b79",
"title": "Denoising diffusion implicit models",
"year": "2021"
},
{
"authors": "Y Song; S Ermon",
"journal": "NeurIPS",
"ref_id": "b80",
"title": "Generative modeling by estimating gradients of the data distribution",
"year": "2019"
},
{
"authors": "C Stoll; N Hasler; J Gall; H.-P Seidel; C Theobalt",
"journal": "",
"ref_id": "b81",
"title": "Fast articulated motion tracking using a sums of gaussians body model",
"year": "2011"
},
{
"authors": "A Van Den Oord; O Vinyals; K Kavukcuoglu",
"journal": "NeurIPS",
"ref_id": "b82",
"title": "Neural discrete representation learning",
"year": "2017"
},
{
"authors": "D Vlasic; I Baran; W Matusik; J Popović",
"journal": "SIGGRAPH",
"ref_id": "b83",
"title": "Articulated mesh animation from multi-view silhouettes",
"year": "2008"
},
{
"authors": "H Wang; X Du; J Li; R A Yeh; G Shakhnarovich",
"journal": "",
"ref_id": "b84",
"title": "Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation",
"year": "2023"
},
{
"authors": "P Wang; L Liu; Y Liu; C Theobalt; T Komura; W Wang",
"journal": "NeurIPS",
"ref_id": "b85",
"title": "Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction",
"year": "2021"
},
{
"authors": "T Wang; B Zhang; T Zhang; S Gu; J Bao; T Baltrusaitis; J Shen; D Chen; F Wen; Q Chen",
"journal": "",
"ref_id": "b86",
"title": "A generative model for sculpting 3d digital avatars using diffusion",
"year": "2022"
},
{
"authors": "T Wang; T Zhang; B Zhang; H Ouyang; D Chen; Q Chen; F Wen",
"journal": "",
"ref_id": "b87",
"title": "Pretraining is all you need for image-to-image translation",
"year": "2022"
},
{
"authors": "Z Wang; S Wu; W Xie; M Chen; V A Prisacariu",
"journal": "",
"ref_id": "b88",
"title": "Nerf-: Neural radiance fields without known camera parameters",
"year": "2021"
},
{
"authors": "D Watson; W Chan; R M Brualla; J Ho; A Tagliasacchi; M Norouzi",
"journal": "",
"ref_id": "b89",
"title": "Novel view synthesis with diffusion models",
"year": "2023"
},
{
"authors": "M Worchel; R Diaz; W Hu; O Schreer; I Feldmann; P Eisert",
"journal": "",
"ref_id": "b90",
"title": "Multi-view mesh reconstruction with neural deferred shading",
"year": "2022"
},
{
"authors": "Z Xiao; K Kreis; A Vahdat",
"journal": "",
"ref_id": "b91",
"title": "Tackling the generative learning trilemma with denoising diffusion gans",
"year": "2022"
},
{
"authors": "Y Xiu; J Yang; D Tzionas; M J Black",
"journal": "",
"ref_id": "b92",
"title": "Icon: implicit clothed humans obtained from normals",
"year": "2022"
},
{
"authors": "Y Xiu; J Yang; X Cao; D Tzionas; M J Black",
"journal": "",
"ref_id": "b93",
"title": "ECON: Explicit Clothed humans Optimized via Normal integration",
"year": "2023"
},
{
"authors": "H Xu; E G Bazavan; A Zanfir; W T Freeman; R Sukthankar; C Sminchisescu",
"journal": "",
"ref_id": "b94",
"title": "Ghum & ghuml: Generative 3d human shape and articulated pose models",
"year": "2020"
},
{
"authors": "J Xu; X Wang; W Cheng; Y.-P Cao; Y Shan; X Qie; S Gao",
"journal": "",
"ref_id": "b95",
"title": "Dream3d: Zero-shot text-to-3d synthesis using 3d shape prior and text-to-image diffusion models",
"year": "2023"
},
{
"authors": "B Yang; S Gu; B Zhang; T Zhang; X Chen; X Sun; D Chen; F Wen",
"journal": "",
"ref_id": "b96",
"title": "Paint by example: Exemplar-based image editing with diffusion models",
"year": "2023"
},
{
"authors": "T Yu; Z Zheng; K Guo; P Liu; Q Dai; Y Liu",
"journal": "",
"ref_id": "b97",
"title": "Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors",
"year": "2021"
},
{
"authors": "X Zeng; A Vahdat; F Williams; Z Gojcic; O Litany; S Fidler; K Kreis",
"journal": "NeurIPS",
"ref_id": "b98",
"title": "LION: Latent point diffusion models for 3d shape generation",
"year": "2022"
},
{
"authors": "J Zhang; Z Jiang; D Yang; H Xu; Y Shi; G Song; Z Xu; X Wang; J Feng",
"journal": "",
"ref_id": "b99",
"title": "Avatargen: a 3d generative model for animatable human avatars",
"year": "2022"
},
{
"authors": "X Zheng; Y Liu; P Wang; X Tong",
"journal": "Comput. Graph. Forum",
"ref_id": "b100",
"title": "Sdf-stylegan: Implicit sdf-based stylegan for 3d shape generation",
"year": "2022"
},
{
"authors": "Z Zheng; T Yu; Y Liu; Q Dai",
"journal": "IEEE TPAMI",
"ref_id": "b101",
"title": "Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction",
"year": "2021"
}
] | [
{
"formula_coordinates": [
4,
50.11,
690.17,
236.92,
22.98
],
"formula_id": "formula_0",
"formula_text": "L dual = E x F ,x B ,c N ,ϵ ϵ ϵ∼N (0,I),t [∥ϵ ϵ ϵ-ϵ ϵ ϵ θ (z F t , z B t , t, E(c N ))∥ 2 2 ].(1)"
},
{
"formula_coordinates": [
4,
313.78,
499.63,
232,
21.19
],
"formula_id": "formula_1",
"formula_text": "ε ϵ ϵ θ (z t , t, E(c N )) = λϵ ϵ ϵ θ (z t , t, E(c N )) + (1 -λ)ϵ ϵ ϵ θ (z t , t),(2)"
},
{
"formula_coordinates": [
5,
61.71,
409.25,
225.32,
26.11
],
"formula_id": "formula_2",
"formula_text": "L = λ normal L normal + λ mask L mask + λ sides L sides +λ laplacian L laplacian + λ reg normal L reg normal .(3)"
},
{
"formula_coordinates": [
5,
58.39,
652.44,
228.64,
36.8
],
"formula_id": "formula_3",
"formula_text": "L sides = M smpl view [h,w]=1 ∥M smpl view [h, w] -Mview [h, w]∥ 2 2 ,(4)"
},
{
"formula_coordinates": [
5,
430.89,
419.31,
112.43,
14.56
],
"formula_id": "formula_4",
"formula_text": "L laplacian = 1 n n i=1 ∥δ δ δ i ∥ 2 2"
},
{
"formula_coordinates": [
5,
308.86,
491.26,
148.98,
14.78
],
"formula_id": "formula_5",
"formula_text": "L reg normal = 1 | F | (i,j)∈ F (1 -n n n i •n n n j ) 2"
},
{
"formula_coordinates": [
13,
85.95,
297.3,
201.08,
9.68
],
"formula_id": "formula_6",
"formula_text": "q(x t |x t-1 ) := N (x t ; 1 -β t x t-1 , β t I)(5)"
},
{
"formula_coordinates": [
13,
109.03,
322.47,
178,
30.2
],
"formula_id": "formula_7",
"formula_text": "q(x 1:T |x 0 ) := T t=1 q(x t |x t-1 )(6)"
},
{
"formula_coordinates": [
13,
93.34,
410.56,
193.69,
17.25
],
"formula_id": "formula_8",
"formula_text": "q(x t |x 0 ) := N (x t ; √ ᾱt x 0 , (1 -ᾱt I)(7)"
},
{
"formula_coordinates": [
13,
76.99,
525.52,
210.04,
58.28
],
"formula_id": "formula_9",
"formula_text": "p θ (x t-1 |x t ) := N (x t-1 ; µ θ (x t , t), Σ θ (x t , t)) (8) p θ (x 0:T ) := p(x T ) T t=1 p θ (x t-1 |x t )(9)"
},
{
"formula_coordinates": [
13,
83.37,
632.11,
203.66,
23.22
],
"formula_id": "formula_10",
"formula_text": "µ θ (x t , t) = 1 √ α t (x t - 1 -α t √ 1 -ᾱt ϵ θ (x t , t))(10)"
},
{
"formula_coordinates": [
13,
84.72,
702.12,
202.31,
12.03
],
"formula_id": "formula_11",
"formula_text": "L simple = E x0,t,ϵ∼N (0,I) [||ϵ -ϵ θ (x t , t)|| 2 ](11)"
},
{
"formula_coordinates": [
16,
122.76,
346.51,
354.49,
7.71
],
"formula_id": "formula_12",
"formula_text": "β 1 = -2 (b) β 1 = -1 (c) β 1 = 0 (d) β 1 = 1 (e) β 1 = 2"
}
] |
|
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b10",
"b18",
"b19",
"b13",
"b27",
"b33",
"b1",
"b15",
"b3",
"b4",
"b4",
"b22",
"b11",
"b33",
"b15",
"b25",
"b34",
"b0",
"b1",
"b16",
"b6",
"b9",
"b28",
"b33",
"b11"
],
"table_ref": [],
"text": "Gaze, as one of the most prominent cues of human attention, plays a crucial role in nonverbal communication. It serves as a tool to measure and enhance human attention and interest, and its significance in understanding human behavior and mental states cannot be overstated. The estimation of gaze is a well-established technique that finds application in a diverse range of fields, including but not limited to saliency detection [10], assisted driving [18], and human-computer interaction [19].\nThe advancements in deep learning have led to a shift in gaze estimation from a model-based approach [13,27] to an appearance-based approach [33,2,15,4,5] that relies solely on facial images. Based on capturing facial images using a webcam, the appearance-based approach yields a higher precision gaze direction, whose results are already comparable to a professional eye-tracking device. Thereby it enabling this technology uses in a broader range of applications beyond the constraints of experimental scenarios. The vision transformer (ViT) architecture dominates in most applications of the computer vision field, which uses self-attention mechanisms to process image patches in parallel. It also performs well on gaze estimation tasks, like [5,22], they combined the self-attention mechanism with CNN to investigate the effectiveness of gaze estimation and model generalization, and their method achieved the leading level at the time in some general gaze estimation datasets [11,33,15]. However, a series of ViT models require a large number of parameters and heavy computation, making it hard to deploy on resource-constrained devices. It is a challenge to achieve good results with a small number of parameters.\nGaze estimation relies heavily on accurate localization of the eye region and the extraction of critical global features. Leveraging large receptive fields can facilitate the extraction of additional information from eye images. Current approaches commonly employ cropped face images from videos as inputs, with eye images cropped at a consistent high resolution of W × H. Prior works, such as [25,34,1,2,16], have reported using resolutions of 60×36, 60 × 60, 64 × 96, and 224 × 224, respectively.\nOur work is motivated by three key desiderata for improving the performance of gaze estimation with minimal computational effort:\n• Firstly, to capture contextual information, large receptive fields are necessary.\n• Secondly, when a large kernel size is needed, FFT layers are more efficient than convolution layers [7].\n• Finally, shortcut connections are vital, particularly for networks with very large kernels, to facilitate gradient flow and improve training [9] In this work, we present FR-Net, a novel lightweight model designed for efficient gaze estimation tasks. FR-Net introduces a Fast Fourier Transform (FFT) Residual Block and utilizes MobileViT v3 [28] as its backbone architecture. The brief structure of FR-Net is shown in Figure 1. The proposed model aims to leverage the benefits of both frequency and spatial domain information. The FFT Residual Block applies FFT to the input feature map to extract frequency domain features and uses a global filter to capture latent features. It reduces the computational complexity from O N 2 to O(N log N ). Moreover, the shortcut connection in the FFT Residual Block is designed to capture spatial domain information and improve model's performance.\nTo evaluate the performance of our proposed method, we conduct experiments on publicly available databases: MPIIFaceGaze [33] and EYEDIAP [11].\nOur method achieves superior performance compared to existing methods while utilizing only a fraction of the parameters and computational resources required by prior methods. Since we adopt ViT as the backbone of our method and compare it with ViT's lightweight work on the gaze estimation task. Our approach outperforms ViT's lightweight work in both accuracy and parameter efficiency. Partial results in terms of angle error, parameters, and FLOPs are shown in Figure 2. Furthermore, our algorithm accurately localizes the eye region, as intended by the design, as evidenced by visual analysis.\nIn this work, we present our primary contributions as: The article is organized as follows. Section 2 provides a comprehensive review of pertinent literature pertaining to gaze estimation and lightweight networks. Section 3 presents the detailed architecture of the proposed FR-Net. Section 4 conducts a comprehensive evaluation of the proposed approach via experiments on publicly available datasets, MPIIFaceGaze and EYEDIAP, as well as ablation experiments to examine the individual contributions of network components. Lastly, Section 5 and 6 draw conclusions and discuss future directions for research."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Gaze Estimation",
"publication_ref": [
"b13",
"b27",
"b11",
"b33",
"b15",
"b32",
"b33",
"b1",
"b15",
"b4",
"b30",
"b2",
"b17",
"b29"
],
"table_ref": [],
"text": "The field of gaze estimation has conventionally been categorized into two primary approaches: model-based and appearance-based. Model-based gaze estimation employs eye structure and characteristics, including corneal reflection [13] and pupil center [27], to estimate gaze. While it provides greater accuracy, it is restricted by outdoor conditions and necessitates specialized hardware.\nThe appearance-based gaze estimation methods leverages simple webcam-based gaze estimation and has gained popularity due to its ease of use. The advancement of gaze estimation has been facilitated by the use of datasets [11,33,15,32] that simulate real-world environments and take into account various factors such as light intensity, glasses, human species, and head posture.\nGaze estimation has become a focal point in the field of human-computer interaction, with an increasing shift towards practical applications from laboratory research. Recent works have leveraged deep learning techniques. Zhang et al. [33] introduced a spatial weighting mechanism to encode the significance of the entire facial image in their approach. Meanwhile, Chen et al. [2] utilized a dilated convolutional approach with three channels of data, consisting of left and right eye images and facial images, as input to their model. Kellnhofer et al. [15] acknowledged the importance of temporal order in gaze estimation and employed a combination of Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) to infer gaze from continuous facial images. Cheng et al. [5] achieved state-of-theart results on publicly available datasets by incorporating a Transformer architecture into their CNN-based model.\nGeneric gaze estimation techniques often exhibit large biases that can vary among individuals, posing a domain generalization problem. To address this issue, Yu et al. [30] synthesized gaze-reoriented images from reference samples and performed self-supervised domain adaptation for gaze reorientation. Cheng et al. [3] proposed a self-adversarial framework for eliminating gaze-independent features in human faces, without requiring target samples. Liu et al. [17] adopted a plug-and-play cross-domain framework based on outlier-guided collaborative learning. Wang et al. [29] employed contrastive learning for the gaze estimation regression task, aiming to separate gaze-related features from gaze-unrelated features.\nThe present endeavors concentrate on enhancing model accuracy and generalization ability. With the advent of Vision Transformer (ViT) in this domain, models are becoming progressively larger, necessitating abundant computational resources, and posing challenges in resource-limited devices."
},
{
"figure_ref": [],
"heading": "Light-weight Net",
"publication_ref": [
"b8",
"b14",
"b24",
"b35",
"b14",
"b24",
"b12",
"b20",
"b21",
"b28"
],
"table_ref": [],
"text": "Lightweight Convolutional Neural Networks (CNNs) play a crucial role in the field of lightweight networks across diverse domains. The deep separable convolutional layer proposed by Xception [8] effectively reduces parameters, which is a crucial factor for certain network architectures [14,24,35]. Meanwhile, MobileNet [14] abandons the conventional methods of shrinking, pruning, quantizing, or compressing small models and instead incorporates the use of deep separable convolutional layers to make the network architecture more efficient. The introduction of backward residuals with linear bottlenecks in MobileNetV2 [24] further optimizes the MobileNet architecture, leading to improved accuracy and reduced computation.\nCompared to other architectures, ViT has the advantage of adaptive weighting of inputs and global processing, demonstrating high performance across multiple domains. However, the computational cost of ViT limits its practicality in mobile applications. LeViT network [12] inte-grates transformer architecture into convolutional networks, resulting in improved efficiency and performance. The Mo-bileViT series of networks [20,21,28] combine the benefits of CNNs and Transformers, exhibiting outstanding performance on numerous mobile vision tasks. Despite having fewer parameters compared to other lightweight networks, the presence of a self-attentive mechanism in the network structure makes MobileViT series networks have higher FLOPs, higher constraints on device resources, and inadequate adherence to low latency requirements."
},
{
"figure_ref": [],
"heading": "Methods",
"publication_ref": [],
"table_ref": [],
"text": "Appearance-based gaze estimation is a method of predicting the direction of a person's gaze by analyzing the appearance of their eyes or face. The basic idea behind appearance-based gaze estimation is that certain features of the eyes and face change depending on where a person is looking.\nOur approach introduces the main idea illustrated in Figure 1. It is motivated that leveraging large receptive fields can facilitate the extraction of more salient features from eye images, and FFT layers offer a much more computationally efficient alternative to convolutional layers when dealing with large kernel sizes.\nThe model utilizes FFT instead of convolution to extract frequency domain features and reduce computation complexity. Our proposed model offers a tradeoff between accuracy and efficiency by utilizing FFT-based techniques.\nAdditionally, we design a series of shortcut components that selectively aggregates features from the spatial domain, effectively improving the model's accuracy. Our proposed model offers a tradeoff between accuracy and efficiency by utilizing FFT-based techniques."
},
{
"figure_ref": [
"fig_5",
"fig_1"
],
"heading": "FFT Residual Block",
"publication_ref": [
"b28"
],
"table_ref": [],
"text": "Our more detailed framework is shown in Figure 3. the backbone of our model is MobileVit V3 [28]. We have kept the main structure and the lightweight unit: inverted residual block and proposed FFT Residual Block replacing transformer encoder, to improve model accuracy while reducing computational complexity.\nThe inverted residual block of two types of convolutions operations, namely depth-separable convolution and point-wise convolution. The depth-separable convolution performs separate convolutions on the input feature map's channels, which significantly reduces the number of parameters, but it separates the channel and spatial information and cannot utilize feature information at the same spatial location from different channels.\nThe self-attention mechanism in the transformer architecture can Effectively extract intrinsic features after the inverted residual block. It is the computationally intensive part, resulting in a significant increase in FLOPs, It presents a formidable challenge to the limited computing resources of mobile devices and hinders their capability for lowlatency and real-time processing. To this end, we present the FFT Residual Block as a solution for the MobileViT network series. The FFT Residual Block is designed to be lightweight, and it effectively extracts gaze-related features from both the spatial and frequency domains by using a simple fusion method. The key component of the FFT Residual Block is the FFT Encoder, which is depicted in Figure 4. There are two main parts: \nLayer Norm"
},
{
"figure_ref": [
"fig_2"
],
"heading": "Calculating Convolutions Using the FFT",
"publication_ref": [],
"table_ref": [],
"text": "FFT offers an efficient way to perform convolution by exploiting the convolution theorem, which states that convolution in the time domain is equivalent to multiplication in the frequency domain. By applying FFT to the inputs and kernels, multiplying their corresponding frequency components, and then applying the inverse FFT to the result, we can obtain the convolution in the time domain.\nHere are the general steps to replace convolution with FFT, shown in Figure 5: 1. Pad the kernel: First, we need to make sure that the kernel has the same size as the input image. If they are not of the same size, we can pad the smaller kernels with zeros to make them the same length.\n2. Compute the FFT of the inputs: Use the FFT algorithm to compute the DFT of images and kernels."
},
{
"figure_ref": [],
"heading": "Multiply the Fourier transforms of the input signals:",
"publication_ref": [],
"table_ref": [],
"text": "Multiply the Fourier transforms of the two input signals element-wise to obtain the Fourier transform of their convolution 4. Compute the inverse FFT: Use the inverse FFT algorithm to compute the inverse DFT of the product obtained in Step 3."
},
{
"figure_ref": [],
"heading": "Truncate the result:",
"publication_ref": [
"b23"
],
"table_ref": [],
"text": "The result of the inverse FFT will be a sequence of complex numbers. We can truncate the imaginary part of the result and take only the real part to obtain the result of the convolution operation. A single convolution can be represented as an equation:\nY = X * K (1\n)\nwhere * is a convolution operation, X, F represents the input image and the filter.\nWe can rewrite equation (1) in the Fourier domain as equation ( 2)\ny = F (X) • F (K) = x • k (2)\nwhere • represents a Hadamard product. Hadamard product is an element-wise operation. It is required to pad the filter f to let the size equal to x.\nThen take the inverse Fourier transform of the multiplication result: y. Now the result of the inverse Fourier transform would be similar to the output of the convolution.\nthe two-dimensional discrete Fourier transform is represented as follows:\nF (x, y) = M -1 m=0 N -1 n=0 f (m, n)e -j2π( ux M + vy N )(3)\nwhere x, y represents the input of 2D signal and N, M presents the dimension of the input. FFT represents the fast Fourier transform. The fast Fourier transform (FFT) algorithm effectively reduces the computational complexity from O N 2 to O(N log N ) by exploiting the periodic and symmetric nature of W N , thereby simplifying the DFT calculation for digital computers. The inverse fast Fourier transform (IFFT) serves a similar purpose to the inverse discrete Fourier transform (IDFT) and enables efficient computation.\nTo further reduce the computational effort, we get rid of the frequency domain transformation of the kernel and instead adopt a global trainable mask that matches the input dimensions. This approach draws inspiration from the methodology presented in [23]. It reduces the process of making FFT on the kernel. In the back-propagation optimization process, the two are equivalent, and the mask is equivalent to:\nM ask = F F T (padding(F ))(4)\nThis trainable mask can also be seen as a large convolution to extract significant features in the frequency domain, but it is significantly faster than traditional convolution for large kernel sizes. "
},
{
"figure_ref": [],
"heading": "Spatial feature extraction",
"publication_ref": [
"b9"
],
"table_ref": [],
"text": "The fast Fourier transform (FFT) is known to capture only frequency domain features and lacks spatial location information. Thus, it is crucial to also extract original features with significant spatial locations in the spatial domain.\nThe FFT convolution components endeavor to extract frequency features in a global manner, which can be viewed as a very large kernel. In [9], Ding et.al prove that incorporating a shortcut connection is vital to enhance the performance of networks employing large kernels.\nIn light of the aforementioned points, our approach incorporates a series of shortcut connections to improve the ability of our model to extract spatial features. The first one is added after the convolution\ny = F (x, {W i }) + x(5)\nwhere x and y are the input and output, and F (x, {W i }) represents the FFT mapping to be learned.\nNotably, Another shortcut connection is added after the group of FFT Encoders. We make a modification traditional way. It concatenated the layer extracting frequency-domain features and spatial information rather than directly element-wise adding to the FFT-extracted features.\nF f usion (x) = Concat(F F T encoder (x, W ), x) (6)\nwhere x and F f usion (x) represent the input and output of core components of FFT Residual Block, The function F F T encoder (x, W ) represents the frequency feature mapping to be learned by a group of FFT Encoders. An ablation study confirms the importance of this design in our model."
},
{
"figure_ref": [],
"heading": "Implementation Details",
"publication_ref": [],
"table_ref": [],
"text": "In this study, the network model was implemented using the PyTorch framework, which utilized a smoothed L1 loss function and the AdamW optimizer throughout the training process. The resolution size of the images was set at 256 × 256. The network was initialized with pre-training parameters from the ETH-XGaze dataset for the MPIIFaceGaze and EYEDIAP datasets, with the initial learning rate set at 0.0004 and decaying to 0.00004 after 10 epochs. The batch size used for the assessment of MPI-IFaceGaze was 64. We conducted our experiments on a computing system equipped with an Intel(R) Core(TM) i7-11800H CPU, NVIDIA RTX 3090 GPU, and 32G RAM."
},
{
"figure_ref": [],
"heading": "Experiment",
"publication_ref": [],
"table_ref": [],
"text": "We conducted a thorough evaluation of the proposed model using a widely employed and publicly accessible dataset in the field of gaze estimation. Our work is compared with state of art works on gaze estimation and Vit lightweight work using performance metrics such as angular error, parameters, FLOPs, and inference time. To validate our design choices, we present a visual analysis of the essential components of our model. Furthermore, we perform ablation experiments to investigate the contributions of key components in our approach."
},
{
"figure_ref": [],
"heading": "Experimental Settings",
"publication_ref": [
"b33",
"b11",
"b5"
],
"table_ref": [],
"text": "Dataset for evaluation To evaluate the performance of the model, we utilized two publicly accessible gaze estimation datasets, namely MPIIFaceGaze [33] and EYEDIAP [11]. The MPIIFaceGaze dataset was collected through a laptop webcam and comprises 213,659 images from 15 subjects, with 3000 samples per subject for evaluation. The EYEDIAP dataset is comprised of 237 minutes of video captured via an HD webcam from 16 subjects, with screen targets and swinging blobs serving as visual stimuli. The data for MPIIFaceGaze and EYEDIAP was normalized according to Cheng's review [6], with the MPIIFaceGaze evaluated via the leave-one-out approach and the EYEDIAP evaluated using 4-fold cross-validation.\nEvaluation Metric For the gaze estimation task, we utilize the angular error as the primary metric, which quantifies the deviation between the estimated and ground truth gaze directions. A lower angular error indicates better accuracy and superior performance.\n= g • ĝ g ĝ (7)\nwhere is the angular error. g ∈ R 3 and ĝ ∈ R 3 are the true and estimated gaze directions, respectively.\nTo assess the effectiveness of our model, we evaluate it using multiple metrics, including the number of model parameters, FLOPs, and the inference time for a single run of the model."
},
{
"figure_ref": [],
"heading": "Comparison to the State-of-the-art Methods",
"publication_ref": [],
"table_ref": [],
"text": "To provide a comprehensive evaluation of our proposed model, we compare it against two aspects of related work: (i) existing gaze estimation models, and (ii) works related to lightweight variants of Vision Transformers (ViT)."
},
{
"figure_ref": [],
"heading": "Gaze estimation comparison",
"publication_ref": [
"b33",
"b1",
"b15",
"b22",
"b4",
"b33",
"b11",
"b4",
"b32"
],
"table_ref": [],
"text": "Our methodology is compared to established gaze estimation models [33,2,15,22,5] by evaluating gaze error angles on the MPIIFaceGaze [33] and EYEDIAP [11] datasets.\nOur proposed FR-Net is evaluated using the same data pre-training methodology as the state-of-the-art (SOTA) work, GazeTR-Hybrid [5]. Specifically, we pre-trained our model on the ETH-XGaze dataset [32], which consists of 1.1 million high-resolution images of 110 individuals with diverse ethnicities, head positions, and viewing orientations from multiple perspectives. For training, we used 765,000 normalized images of 80 subjects with a resolution of 224 × 224. To satisfy the input requirements of our model, we transformed these images into 256 × 256."
},
{
"figure_ref": [],
"heading": "Methods",
"publication_ref": [
"b33",
"b1",
"b15",
"b3",
"b22",
"b4"
],
"table_ref": [
"tab_2"
],
"text": "MPIIFaceGaze EYEDIAP Spatial Weights CNN [33] 4.93°6.53°D ilated-Net [2] 4.42°6.19°G aze360 [15] 4.06°5.36°C A-Net [4] 4.27°5.27° [ 22] 4.04°5.25°G azeTR-Hybrid [5] 4.00°5.17°F R-Net 3.86°4.51°T able 1. Comparison with state-of-the-art gaze estimation methods in the angle error Table 1 presents a comparison of our proposed FR-Net with existing methods in terms of gaze error, and our results demonstrate a 0.14°increase in minimum gaze error angle on MPIIFaceGaze (from 4°to 3.86°) and a 0.66°i mprovement in minimum gaze error angle on EYEDIAP (from 5.17°to 4.51°).\nPerformance results are depicted in Table 2. Our model results in a reduction of the number of parameters for gaze estimation to under one million, achieving a significant improvement of 0.67M parameters. It represents a reduction of 5 -17 times compared to the existing model. The optimization of FLOPs was achieved at 0.22 billion, significantly reducing resource consumption for mobile devices. In this aspect of inference time, our algorithm also provides a slight improvement over existing practices. It should be noted that the inference time is verified on the CPU, which is closer to the practical application."
},
{
"figure_ref": [],
"heading": "Methods",
"publication_ref": [
"b1",
"b15",
"b4"
],
"table_ref": [
"tab_2"
],
"text": "#Params(M) FLOPs(B) Time(ms) Dilated-Net [2] 3.92 3.15 29 Gaze360 [15] 11.9 7.29 62 GazeTR-Hybrid [5] 11.4 The experimental results presented in Tables 1 and2 demonstrate the state-of-the-art performance of the FR-Net model on the MPIIFaceGaze and EYEDIAP datasets. Specifically, the approach not only achieves a reduction in gaze error angle, but also significantly decreases the number of model parameters and FLOPs, resulting in improved model performance. The experimental results demonstrate the proposed model could efficiently achieve the fusion extraction of frequency and time domain features, resulting in improved accuracy and efficiency."
},
{
"figure_ref": [],
"heading": "Comparison with ViT lightweight models",
"publication_ref": [
"b26",
"b12",
"b31"
],
"table_ref": [],
"text": "In this study, we aimed to investigate the performance of a relatively lightweight ViT method for the gaze estimation task. To this end, we adopted the same training strategy as FR-Net and compare our approach with DeiT-S [26], LeViT-128S [12] and T2T-ViT-7 [31]."
},
{
"figure_ref": [],
"heading": "Methods",
"publication_ref": [
"b26",
"b12",
"b31"
],
"table_ref": [],
"text": "MPIIFaceGaze EYEDIAP DeiT-S [26] 6.11°6.39°L eViT-128S [12] 5.33°5.62°T 2T-ViT-7 [31] 4.99°5.54°F R-Net 3.86°4.51°T The performance of lightweight ViT models on MPI-IFaceGaze and EYEDIAP datasets is suboptimal, as presented in Table 3. Facial images are inherently complex and contain implicit features related to gaze, including head pose, illumination, and resolution. The lightweight ViT approach efforts to optimize the number of parameters and FLOPs. Its inadequate performance in estimating gaze can be attributed to its limited capability to extract and simulta-neously incorporate these intricate, invisible feature information."
},
{
"figure_ref": [],
"heading": "Methods",
"publication_ref": [
"b26",
"b12",
"b31"
],
"table_ref": [],
"text": "Params(M) FLOPs(B) Time(ms) DeiT-S [26] 21.67 4.25 43 LeViT-128S [12] 7.00 0.28 10 T2T-ViT-7 [31] 4 "
},
{
"figure_ref": [
"fig_6"
],
"heading": "Visualization of learned features",
"publication_ref": [],
"table_ref": [],
"text": "To determine whether our model can effectively extract eye-related features, we designed a visualization scheme for FFT Residual Block. We utilize the first FFT Residual Block to generate a feature map of the face image, which is then turned into a single-channel feature map. The information contained within the feature map is then visually represented. The visualization highlights the attention allocation of the model towards the input data, with regions of brighter color indicating a greater degree of attention, while darker regions reflect less attention. The presented results serve to illustrate the effectiveness of FFT Residual Block in extracting gaze-related features from facial images. In order to visually demonstrate the effectiveness of our FR-Net model in extracting pertinent information from facial images, we selected a diverse range of face images sourced from the MPIIFaceGaze dataset for feature map visualization, as presented in Figure 7. The visualized feature map clearly highlights the eye region and facial contour as the most salient features, as evidenced by their notably brighter appearance.\nThe findings indicate that our model is able to find face and eye features accurately and efficiently as expected."
},
{
"figure_ref": [],
"heading": "Ablation Study",
"publication_ref": [
"b9"
],
"table_ref": [
"tab_4"
],
"text": "To better understand the effect of different parts on our FR-Net, we remove the FFT Residual Block, FFT Encoder, and shortcut component in FFT Encoder and concatenating shortcut from the backbone model respectively.\nThe results, as presented in According to Section 3.1, we hypothesize that shortcuts are crucial, particularly in networks using large kernels, as demonstrated by prior work [9]. To verify this hypothesis, we conducted ablation experiments on shortcut connections. Specifically, we removed the shortcut component in the FFT Encoder and shortcut concatenation outside of the FFT Encoder from the original model. The results, presented in Table 5, indicate that the shortcut inside the FFT Encoder has minimal impact on the model's performance, with similar results to the original model on the MPIIFaceGaze dataset (0.06°higher) and slightly improved performance on the EYEDIAP dataset (0.19°higher). Similarly, the gaze error angle was 0.06°and 0.19°higher than the original model after removing the shortcut concatenation outside of the FFT Encoder on the MPIIFaceGaze and EYEDIAP datasets, respectively. In conclusion, shortcut connections do have some impact, but not as significant as initially expected.\nIn summary, the results of the ablation study indicate that the FFT Resident Block can effectively extract the frequency domain features and that the shortcut connection with FFT can enhance the gaze-related factors and minimize the final gaze error angle."
},
{
"figure_ref": [],
"heading": "Limitation and future work",
"publication_ref": [],
"table_ref": [],
"text": "Based on the experimental results presented in Section 4, our model demonstrates outstanding performance with respect to angle error, parameters, and FLOPs. However, the experimental findings presented herein demonstrate that the benefits of our model in terms of inference time are not as pronounced as its parameter and FLOP count would suggest. There exist multiple factors that affect the real inference time. Based on our analysis, we contend that the FFTs algorithm demonstrates an advantage in computing complexity. However, due to a lack of tight integration with the existing deep learning framework, such as convolution, it is susceptible to the effects of hardware and mate operators. Consequently, its advantage in real inference is not particularly notable.\nIn terms of the role of the different parts of the designed model, there are still some discrepancies observed from our prior design expectations. Despite extensive experimentation and analysis, a specific reason for these deviations remains elusive. Further investigation and exploration are necessary to fully comprehend and reconcile these differences.\nIn the future, to further enhance the applicability of our model across diverse computation-constrained devices, it is imperative to analyze and optimize related mate operators to maximize its efficacy."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "This paper presents FR-Net, a novel lightweight model for gaze estimation that outperforms state-of-the-art models in terms of accuracy while maintaining an efficient structure. Our proposed approach leverages the power of frequency domain features via FFT Residual Block and Mo-bileViT v3 to extract and enhance important information related to gaze. To achieve this, we introduce a trainable mask with low computation complexity that helps to focus on crucial information. Furthermore, we incorporate a shortcut component to the model to extract spatial features and improve accuracy.\nWe evaluated our approach on two widely used gaze estimation datasets, namely MPIIFaceGaze and EYEDIAP. The experimental results demonstrate that FR-Net achieves minimal error angle, outperforming state-of-the-art gaze estimation models while maintaining a lightweight structure. We also compared our model's performance with other lightweight ViT approaches, and our model showed competitive results. Our proposed approach offers a promising direction for improving gaze estimation in real-world scenarios, such as driver monitoring systems and humancomputer interaction applications."
}
] | Gaze estimation is a crucial task in computer vision, however, existing methods suffer from high computational costs, which limit their practical deployment in resourcelimited environments. In this paper, we propose a novel lightweight model, FR-Net, for accurate gaze angle estimation while significantly reducing computational complexity. FR-Net utilizes the Fast Fourier Transform (FFT) to extract gaze-relevant features in frequency domains while reducing the number of parameters. Additionally, we introduce a shortcut component that focuses on the spatial domain to further improve the accuracy of our model. Our experimental results demonstrate that our approach achieves substantially lower gaze error angles (3.86 on MPII and 4.51 on EYEDIAP) compared to state-of-the-art gaze estimation methods, while utilizing 17 times fewer parameters (0.67M) and only 12% of FLOPs (0.22B). Furthermore, our method outperforms existing lightweight methods in terms of accuracy and efficiency for the gaze estimation task. These results suggest that our proposed approach has significant potential applications in areas such as human-computer interaction and driver assistance systems. | FR-Net:A Light-weight FFT Residual Net For Gaze Estimation | [
{
"figure_caption": "Figure 1 .Figure 2 .12Figure 1. Brief illustration of FR-Net. Core component: FFT Residual block is proposed to extract the pertinent frequency and spatial features related to eyes. Backbone Model: MobileVit v3",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "12",
"figure_type": "figure"
},
{
"figure_caption": "FFNFigure 4 .4Figure 4. The structure of FFT Encoder: Our proposed approach involves replacing the self-attention mechanism utilized in the Transformer Encoder architecture with FFT Encoder in the frequency domain. This modification efficiently extracts gaze-related features in the frequency domain.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 .5Figure 5. Calculating convolutions using FFT",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "IFFT",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 6 .6Figure 6. The main idea applies FFT in our model.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "able 3 .3Comparison with ViT lightweight models in the error angle",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure 7 .7Figure 7. Our proposed model effectively captures distinctive features of the eye region, as well as other visually-relevant characteristics, including head pose. This emphasizes the practicality of our approach in accurately estimating gaze patterns from facial images.",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "7",
"figure_type": "figure"
},
{
"figure_caption": "The three FFT Residual Blocks are utilized to extract both depth spatial domain and frequency domain features, which are then fused to form the input for the subsequent layer. The input dimension of the FFT Encoder is 64, 80, and 96 in consecutive order. The 1 × 1 convolutional layer is employed to control the variation in the dimension of the feature map. The channels C1, C2, C3, C4, C5 have dimensions of 16, 24, 48, 64, and 80, respectively.",
"figure_data": "Conv-3 3Stride=2InvertedResidual BlocksIRB Stride=2FFT ResidualBlockIRB Stride=2FFT ResidualBlockIRB Stride=2FFT ResidualBlockConv-1 1Global poolLinearGaze (yaw, pitch)DWConv3 3Conv1 1FFT EncoderConv1 1C o n c a t e n a t eFFT Residual BlockConv1 1Figure 3.",
"figure_id": "tab_1",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Comparison with the state-of-the-art gaze estimation methods in Parameters, FLOPs and Inference Time",
"figure_data": "1.8324FR-Net0.670.2223",
"figure_id": "tab_2",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Comparison with ViT lightweight methods in Parameters, FLOPs and Inference Time Table.4 illustrates the results. With respect to lightweight optimization, our model outperforms the lightweight ViT models in terms of parameter efficiency and exhibits a slight lead in terms of FLOPs and inference time. As a result, our model demonstrates superiority over the lightweight ViT model in both accuracy and efficiency for the task of gaze estimation.",
"figure_data": ".000.9823FR-Net0.670.2223",
"figure_id": "tab_3",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "indicate a substantial impact of FFT Residual Block on gaze accuracy. The error angle in the network without the FFT Residual Block increased by 0.75 and 0.96 on the MPIIFaceGaze and EYE-DIAP datasets, respectively, underscoring the significance of the FFT Residual Block in the FR-Net model. For core part of FFT Residual Block, The FFT encoder, the core part of FFT Residual Block, shows an important role on improve the accuracy, specially on EYEDIAP. It even improves much more than FFT Residual Block. Up until now, we haven't found the reason. It provides us with clues for further improvement later on.",
"figure_data": "MethodsMPIIFaceGaze EYEDIAPFR-Net3.86°4.51°-FFT RB4.61°5.45°-FFT Encoder4.08°6.29°-Concatenation shortcut3.92°4.70°-FFT Encoder shortcut3.88°4.82°Table 5. Ablation study results for the error angle, Symbol '-' in-dicates the following component is removed",
"figure_id": "tab_4",
"figure_label": "5",
"figure_type": "table"
}
] | Tao Xu; Bo Wu; Ruilong Fan; Yun Zhou; Di Huang | [
{
"authors": "Abid Ali; Yong-Guk Kim",
"journal": "IEEE Access",
"ref_id": "b0",
"title": "2020-Deep Fusion for 3D Gaze Estimation From Natural Face Images Using Multi-Stream CNNs",
"year": "2020"
},
{
"authors": "Zhaokang Chen; Bertram E Shi",
"journal": "Springer International Publishing",
"ref_id": "b1",
"title": "Appearance-Based Gaze Estimation Using Dilated-Convolutions",
"year": "2019"
},
{
"authors": "Yihua Cheng; Yiwei Bao; Feng Lu",
"journal": "Proceedings of the AAAI Conference on Artificial Intelligence",
"ref_id": "b2",
"title": "PureGaze: Purifying Gaze Feature for Generalizable Gaze Estimation",
"year": "2022-06"
},
{
"authors": "Yihua Cheng; Shiyao Huang; Fei Wang; Chen Qian; Feng Lu",
"journal": "",
"ref_id": "b3",
"title": "A Coarse-to-Fine Adaptive Network for Appearance-Based Gaze Estimation",
"year": "2020-04"
},
{
"authors": "Yihua Cheng; Feng Lu",
"journal": "",
"ref_id": "b4",
"title": "Gaze Estimation using Transformer",
"year": "2022-08"
},
{
"authors": "Yihua Cheng; Haofei Wang; Yiwei Bao; Feng Lu",
"journal": "",
"ref_id": "b5",
"title": "Appearance-based Gaze Estimation With Deep Learning: A Review and Benchmark",
"year": "2021-04"
},
{
"authors": "Lu Chi; Borui Jiang; Yadong Mu",
"journal": "",
"ref_id": "b6",
"title": "Fast Fourier Convolution",
"year": ""
},
{
"authors": " Curran Associates; Inc",
"journal": "",
"ref_id": "b7",
"title": "",
"year": "2020"
},
{
"authors": "Francois Chollet",
"journal": "IEEE",
"ref_id": "b8",
"title": "Xception: Deep Learning with Depthwise Separable Convolutions",
"year": "2017-07"
},
{
"authors": "Xiaohan Ding; Xiangyu Zhang; Jungong Han; Guiguang Ding",
"journal": "",
"ref_id": "b9",
"title": "Scaling up your kernels to 31x31: Revisiting large kernel design in cnns",
"year": "2022"
},
{
"authors": "Deng-Ping Fan; Ming-Ming Cheng; Jiang-Jiang Liu; Shang-Hua Gao; Qibin Hou; Ali Borji",
"journal": "Springer International Publishing",
"ref_id": "b10",
"title": "Salient Objects in Clutter: Bringing Salient Object Detection to the Foreground",
"year": "2018"
},
{
"authors": "Kenneth Alberto; Funes Mora; Florent Monay; Jean-Marc Odobez",
"journal": "ACM",
"ref_id": "b11",
"title": "EYEDIAP: a database for the development and evaluation of gaze estimation algorithms from RGB and RGB-D cameras",
"year": "2014-03"
},
{
"authors": "Ben Graham; Alaaeldin El-Nouby; Hugo Touvron; Pierre Stock; Armand Joulin; Herve Jegou; Matthijs Douze",
"journal": "IEEE",
"ref_id": "b12",
"title": "LeViT: a Vision Transformer in ConvNet's Clothing for Faster Inference",
"year": "2021-10"
},
{
"authors": "E D Guestrin; M Eizenman",
"journal": "IEEE Transactions on Biomedical Engineering",
"ref_id": "b13",
"title": "General Theory of Remote Gaze Estimation Using the Pupil Center and Corneal Reflections",
"year": "2006-06"
},
{
"authors": "Andrew G Howard; Menglong Zhu; Bo Chen; Dmitry Kalenichenko; Weijun Wang; Tobias Weyand; Marco Andreetto; Hartwig Adam",
"journal": "",
"ref_id": "b14",
"title": "MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications",
"year": "2017-04"
},
{
"authors": "Petr Kellnhofer; Adria Recasens; Simon Stent; Wojciech Matusik; Antonio Torralba",
"journal": "IEEE",
"ref_id": "b15",
"title": "Gaze360: Physically Unconstrained Gaze Estimation in the Wild",
"year": "2019-10"
},
{
"authors": "Kyle Krafka; Aditya Khosla; Petr Kellnhofer; Harini Kannan; Suchendra Bhandarkar; Wojciech Matusik; Antonio Torralba",
"journal": "IEEE",
"ref_id": "b16",
"title": "Eye Tracking for Everyone",
"year": "2016-06"
},
{
"authors": "Yunfei Liu; Ruicong Liu; Haofei Wang; Feng Lu",
"journal": "IEEE",
"ref_id": "b17",
"title": "Generalizing Gaze Estimation with Outlier-guided Collaborative Adaptation",
"year": "2021-10"
},
{
"authors": "Sujitha Martin; Sourabh Vora; Kevan Yuen; Mohan Manubhai Trivedi",
"journal": "IEEE Transactions on Intelligent Vehicles",
"ref_id": "b18",
"title": "Dynamics of Driver's Gaze: Explorations in Behavior Modeling and Maneuver Prediction",
"year": "2018-06"
},
{
"authors": "Silèye Benoit Massé; Radu Ba; Horaud",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b19",
"title": "Tracking Gaze and Visual Focus of Attention of People Involved in Social Interaction",
"year": "2018-11"
},
{
"authors": "Sachin Mehta; Mohammad Rastegari",
"journal": "",
"ref_id": "b20",
"title": "MobileViT: Lightweight, General-purpose, and Mobile-friendly Vision Transformer",
"year": "2022-03"
},
{
"authors": "Sachin Mehta; Mohammad Rastegari",
"journal": "",
"ref_id": "b21",
"title": "Separable Selfattention for Mobile Vision Transformers",
"year": "2022-06"
},
{
"authors": "Hyung Jun O Oh; Jin Chang; Sang-Il Choi",
"journal": "IEEE",
"ref_id": "b22",
"title": "Self-Attention with Convolution and Deconvolution for Efficient Eye Gaze Estimation from a Full Face Image",
"year": "2022-06"
},
{
"authors": "Yongming Rao; Wenliang Zhao; Zheng Zhu; Jiwen Lu; Jie Zhou",
"journal": "Advances in neural information processing systems",
"ref_id": "b23",
"title": "Global filter networks for image classification",
"year": "2021"
},
{
"authors": "Mark Sandler; Andrew Howard; Menglong Zhu; Andrey Zhmoginov; Liang-Chieh Chen",
"journal": "IEEE",
"ref_id": "b24",
"title": "MobileNetV2: Inverted Residuals and Linear Bottlenecks",
"year": "2018-06"
},
{
"authors": "Yusuke Sugano; Yasuyuki Matsushita; Yoichi Sato",
"journal": "IEEE",
"ref_id": "b25",
"title": "Learning-by-Synthesis for Appearance-Based 3D Gaze Estimation",
"year": "2014-06"
},
{
"authors": "Hugo Touvron; Matthieu Cord; Matthijs Douze; Francisco Massa; Alexandre Sablayrolles; Herve Jegou",
"journal": "PMLR",
"ref_id": "b26",
"title": "Training data-efficient image transformers & distillation through attention",
"year": "2021-07"
},
{
"authors": "R Valenti; N Sebe; T Gevers",
"journal": "IEEE Transactions on Image Processing",
"ref_id": "b27",
"title": "Combining Head Pose and Eye Location Information for Gaze Estimation",
"year": "2012-02"
},
{
"authors": "N Shakti; Abhishek Wadekar; Chaurasia",
"journal": "",
"ref_id": "b28",
"title": "MobileViTv3: Mobile-Friendly Vision Transformer with Simple and Effective Fusion of Local, Global and Input Features",
"year": "2022-10"
},
{
"authors": "Yaoming Wang; Yangzhou Jiang; Jin Li; Bingbing Ni; Wenrui Dai; Chenglin Li; Hongkai Xiong; Teng Li",
"journal": "IEEE",
"ref_id": "b29",
"title": "Contrastive Regression for Domain Adaptation on Gaze Estimation",
"year": "2022-06"
},
{
"authors": "Yu Yu; Gang Liu; Jean-Marc Odobez",
"journal": "IEEE",
"ref_id": "b30",
"title": "Improving Few-Shot User-Specific Gaze Adaptation via Gaze Redirection Synthesis",
"year": "2019-06"
},
{
"authors": "Li Yuan; Yunpeng Chen; Tao Wang; Weihao Yu; Yujun Shi; Zihang Jiang; Francis E H Tay; Jiashi Feng; Shuicheng Yan",
"journal": "IEEE",
"ref_id": "b31",
"title": "Tokens-to-Token ViT: Training Vision Transformers from Scratch on ImageNet",
"year": "2021-10"
},
{
"authors": "Xucong Zhang; Seonwook Park; Thabo Beeler; Derek Bradley; Siyu Tang; Otmar Hilliges",
"journal": "Springer International Publishing",
"ref_id": "b32",
"title": "ETH-XGaze: A Large Scale Dataset for Gaze Estimation Under Extreme Head Pose and Gaze Variation",
"year": "2020"
},
{
"authors": "Xucong Zhang; Yusuke Sugano; Mario Fritz; Andreas Bulling",
"journal": "IEEE",
"ref_id": "b33",
"title": "It's Written All Over Your Face: Full-Face Appearance-Based Gaze Estimation",
"year": "2017-07"
},
{
"authors": "Xucong Zhang; Yusuke Sugano; Mario Fritz; Andreas Bulling",
"journal": "Conference Name: IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b34",
"title": "MPIIGaze: Real-World Dataset and Deep Appearance-Based Gaze Estimation",
"year": "2019"
},
{
"authors": "Xiangyu Zhang; Xinyu Zhou; Mengxiao Lin; Jian Sun",
"journal": "IEEE",
"ref_id": "b35",
"title": "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices",
"year": "2018-06"
}
] | [
{
"formula_coordinates": [
5,
143.79,
415.42,
138.7,
8.96
],
"formula_id": "formula_0",
"formula_text": "Y = X * K (1"
},
{
"formula_coordinates": [
5,
282.49,
415.74,
3.87,
8.64
],
"formula_id": "formula_1",
"formula_text": ")"
},
{
"formula_coordinates": [
5,
114.9,
493.5,
171.47,
8.96
],
"formula_id": "formula_2",
"formula_text": "y = F (X) • F (K) = x • k (2)"
},
{
"formula_coordinates": [
5,
80.99,
616.39,
205.38,
30.2
],
"formula_id": "formula_3",
"formula_text": "F (x, y) = M -1 m=0 N -1 n=0 f (m, n)e -j2π( ux M + vy N )(3)"
},
{
"formula_coordinates": [
5,
366.63,
244.15,
178.48,
8.96
],
"formula_id": "formula_4",
"formula_text": "M ask = F F T (padding(F ))(4)"
},
{
"formula_coordinates": [
5,
383.78,
636.82,
161.33,
9.65
],
"formula_id": "formula_5",
"formula_text": "y = F (x, {W i }) + x(5)"
},
{
"formula_coordinates": [
6,
74.87,
117.37,
211.49,
9.65
],
"formula_id": "formula_6",
"formula_text": "F f usion (x) = Concat(F F T encoder (x, W ), x) (6)"
},
{
"formula_coordinates": [
6,
408.87,
118.94,
136.25,
22.31
],
"formula_id": "formula_7",
"formula_text": "= g • ĝ g ĝ (7)"
}
] |
||
10.48550/arXiv.2012.13089 | 2023-05-24 | [
{
"figure_ref": [
"fig_0",
"fig_1"
],
"heading": "Introduction",
"publication_ref": [
"b47",
"b108",
"b55",
"b96",
"b14",
"b71",
"b80",
"b13",
"b114",
"b68",
"b119",
"b95",
"b101"
],
"table_ref": [],
"text": "With the rapid development of 3D data processing technologies, an increasing number of relevant applications have emerged in both industrial and daily usage, such as indoor navigation (El-Sheimy and Li, 2021), autonomous driving (Li, Ma, Zhong, Liu, Cao, Li and Chapman, 2020), and object modeling (Yang, Liu, Hu, Wang and Lin, 2019). Li-DAR is one of the indispensable types of sensors to capture disordered 3D point cloud data from traffic scenes, which has enabled more challenging tasks like pedestrian detection (Matti, Ekenel and Thiran, 2017) and road semantic segmentation (Wu, Zhou, Zhao, Yue and Keutzer, 2019) based on the strong inference ability of deep neural networks (DNNs).\nHowever, several well known problems in the supervised point cloud DNNs hinder their further development and practical uses. For example, accurate environment perception via DNNs requires millions of labeled data as the input, while point cloud annotating is labor-intensive and timeconsuming due to its disordered and sparse nature (Dai, Chang, Savva, Halber, Funkhouser and Nießner, 2017). Besides, manual labeling by human experts or users inevitably leads to mistakes such as mislabeling and omission. Another long-standing problem is that the supervised learning paradigm struggles to capture the underlying patterns of new data and fails to generalize the pre-training model to downstream tasks because of overfitting caused by noisy labels (Sariyildiz, Kalantidis, Alahari and Larlus, 2022).\nThe aforementioned issues motivate research in extracting effective feature representations from point clouds via (1) Pre-training stage: point cloud data is firstly preprocessed through the augmentation block and then fed into the point-specific encoder to learn feature representations. The features are utilized to complete well-design pretext tasks, where the output will be compared with the pseudo labels derived from the original data to generate a loss and to update encoder parameters via back-propagation; (2) Supervised finetuning stage: the well-trained encoder is transferred to the target domain. A task head is trained with the training labels in a supervised manner to complete the downstream tasks; (3) Inference stage: the encoder and task head are concatenated as a model to execute inference on the test set. The effectiveness of the SSL pre-training framework can be evaluated based on the performance of the model on the downstream tasks.\nSelf-Supervised Learning (SSL) to learn implicit while better representations without manual annotations. Not only does it solve the problem of the error-prone and expensive labeling process, but also relieve the domain adaptation (DA) issues (Csurka, 2017) with improved model generalization ability. Under the SSL paradigm, basic geometric as well as advanced semantic information can be extracted as knowledge and migrated to downstream tasks under the transfer learning setup. This process approximates human learning that discovers objective principles of the world by observing phenomena and summarizing them into a system of experience and knowledge.\nFig. 1 shows a general pipeline of SSL on point cloud data. The goal of SSL is to pre-train an encoder on an unlabeled, large-scale point cloud dataset (source domain), and to transfer the well-trained network to other datasets (target domain) in various downstream tasks. A complete SSL framework usually contains the following important modules.\n• Data augmentation: The raw input is augmented via some easy-to-implement pre-processing operations such as translation, rotation, flip, and adding noise (Zhang, Lin, Li, Jia and Zhang, 2022c). The objective is to expand the size and diversity of the raw data and to provide subjects for subsequent pretext tasks. The details will be discussed in Section 2.4.\n• Encoder: The encoder is a point-specific deep network that captures the hierarchical representation of the input point cloud data. We will introduce some commonly used point cloud encoders that learn either from downsampling layer-by-layer (Qi, Su, Mo and Guibas, 2017a;Qi, Yi, Su and Guibas, 2017b) or local areas to capture the association between different blocks (Zhou and Tuzel, 2018;Wang, Sun, Liu, Sarma, Bronstein and Solomon, 2019). The details will be discussed in Section 2.5.\n• Pretext task: At the core of the framework is the design of a pretext task that mines the hidden selfsupervision signal via the interactions between the encoder and data. This part is also the focus of the survey and will be discussed in detail in Section 3.\n• Knowledge transfer: The well-trained encoder will be transferred to another dataset with the knowledge gained in the source domain after completing the pretext task. A task head is constructed and trained by a small amount of labelled data in the target domain as the supervision signals to fine-tune the whole architecture. The details will be discussed in Section 4.\n• Downstream task: To evaluate the effectiveness of the SSL framework, the pre-trained encoder will be transferred and evaluated on another dataset for performance evaluation, e.g. object classification, part segmentation, and object detection. The details will be discussed in Section 4.\nThriving progress has been made on point cloud SSL recently, and new models, algorithms and benchmark datasets are emerging quickly and continuously. A systematic review on this exciting topic, especially the research published in the past three years, is urgently needed. In our study, we find that the survey in (Xiao, Huang, Guan and Lu, 2022) employed a similar methodology but focused on unsupervised representation learning. However, it lacked a review on the state-of-the-art SSL models, and in particular, a detailed demonstration of most recent published works. Therefore, we are motivated to provide a comprehensive review on the recently published, representative research on point cloud SSL. Our contributions can be summarized as follows:\n• Systematic and novel taxonomy: We propose a novel and systematic taxonomy for categorizing the diverse kinds of point cloud SSL methods to provide a clear and holistic view on the state of the art. Taking into consideration the characteristics of popular pretext tasks, the taxonomy groups current methods into four broad categories. Each broad category is further subdivided into more fine-grained sub-categories according to the methods in feature utilization as shown in Fig. 2.\n• Comprehensive and detailed summary: We conduct a comprehensive review of the state of the art, including the background of SSL and point clouds, commonly used datasets and models, pretext tasks, and downstream tasks with performance comparison.\n• Exhaustive dataset summary and evaluation comparison: We summarize the unique characteristics of 18 most frequently utilized datasets in the point cloud research. More importantly, we compare the performance of different SSL methods on these datasets according to various downstream tasks.\n• Future directions: Based on our investigation, we summarize and discuss the major limitations and challenges in the current research and propose potential future directions which would hopefully motivate more theoretical and practical research towards more intelligent and effective SSL approaches for point cloud data processing.\nThe rest of the paper is organized as follows: Section 2 introduces the preliminaries for this survey to equip the readers with the necessary background knowledge on SSL and point cloud data. Section 3 represents the main body of the study and provides an exhaustive and detailed analysis on the state of the arts methods according to the structure of the proposed taxonomy. Section 4 illustrates an evaluation and comparison study on the performance of different SSL methods on the frequently utilized downstream tasks and benchmark datasets. Section 5 discusses the limitations and challenges of current research and proposes potential future directions, and Section 6 concludes the paper."
},
{
"figure_ref": [],
"heading": "Background",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Self-supervised learning in the language and image domain",
"publication_ref": [
"b57",
"b15",
"b21",
"b43",
"b21",
"b15",
"b16",
"b62",
"b25",
"b65",
"b30",
"b20",
"b79",
"b12",
"b38",
"b3",
"b2",
"b37",
"b10",
"b31",
"b7"
],
"table_ref": [],
"text": "We firstly describe the development history of SSL in the language and image domains. The purpose is to provide readers a general understanding on SSL. Although data types vary from domains to domains, the core idea of SSL remains the same: to leverage data characteristics for transformation processing and to make the transformed data consistent with the original input in terms of feature representation by contrasting or reconstruction.\nThe idea of SSL was firstly introduced in Natural Language Processing (NLP) research. After converting words into vectors, e.g. Word2Vec (Mikolov, Chen, Corrado and Dean, 2013), and utilizing the relationships between the representations and context, models could learn semantic representations from neighboring words or sentences through pretext task formulations such as next sentence prediction (Devlin, Chang, Lee and Toutanova, 2018), auto-regressive language modeling (Floridi and Chiriatti, 2020), or sentence permutation (Lewis, Liu, Goyal, Ghazvininejad, Mohamed, Levy, Stoyanov and Zettlemoyer, 2019). Landmark models such as GPT (Floridi and Chiriatti, 2020) and BERT (Devlin et al., 2018), and many variants celebrate great achievements in not only NLP but also other fields later.\nIn the field of image processing and computer vision, different SSL methods impose simple variations on image data and extract features by recovering it to the original input, for example, from simple tasks like relative position prediction (Doersch, Gupta and Efros, 2015;Noroozi and Favaro, 2016) and rotation angle prediction (Gidaris, Singh and Komodakis, 2018), to reconstructing blocks masked by surrounding visible pictures (Pathak, Krähenbühl, Donahue, Darrell and Efros, 2016;He, Chen, Xie, Li, Dollár and Girshick, 2021). Free semantic label-based (Faktor and Irani, 2014;Stretcu and Leordeanu, 2015;Croitoru, Bogolin and Leordeanu, 2017;Jiang, Larsson, Shakhnarovich and Learned-Miller, 2018) and cross-modal-based methods (Arandjelovic and Zisserman, 2017;Agrawal, Carreira and Malik, 2015;Jayaraman and Grauman, 2015) have been proposed, which learn representations via automatically generated semantic labels and extra information from other modalities. Recently, the research community shows a great interest on contrastive learning (Chen, Kornblith, Norouzi and Hinton, 2020;He, Fan, Wu, Xie and Girshick, 2020;Caron, Misra, Mairal, Goyal, Bojanowski and Joulin, 2020), which aims to differentiate positive and negative samples by comparison using data augmentation techniques. These research works inspired the study of SSL on point clouds, with similar ideas transferred from 2D to 3D by adapting for data peculiarities."
},
{
"figure_ref": [],
"heading": "Properties of the point cloud data",
"publication_ref": [
"b31",
"b39",
"b8"
],
"table_ref": [],
"text": "Data properties are distinct between naturals languages, images, and point clouds. Languages are usually complex and abstract in nature, and contain ambiguous information due to its versatility and richness. It is expressed in a sequence of words, which is discrete and unstructured in the representation space (He et al., 2020). In contrast, images contain rich visual information, such as color, texture, and shape information of an object in high-dimensional space (Jing and Tian, 2020) for human perception. They are usually represented as 2D data by using a matrix of pixel values.\nSimply speaking, point cloud data is similar to image data in terms of visual format and can be regarded as 3D stereo images with depth information. However, the attributes of point cloud data are completely different in geometric representation. Specifically, a point cloud is a collection of discrete, disordered, and topology-free 3D points. The most basic information contained in the points is the position coordinates (𝑥 𝑖 , 𝑦 𝑖 , 𝑧 𝑖 ) in the Euclidean space, where 𝑖 is the number of points in the object. There are also other optional attributes such as color, intensity, reflectivity, etc., specifying physical properties of the points in more detail. The input order is trivial for point cloud data and does not impact the semantic meaning while it is crucial for images and language where various words or pixel sequences lead to completely divergent connotations. Additionally, point cloud data is invariant to rigid transformation, which means that it remains unchanged after rotation and translation. Some of such exclusive properties can be summarized as follows:\n• Sparsity: The point cloud data is discretely distributed on the surface of the scanned object or scene.\n• Non-uniformity: The distance between points is not fixed and is determined by various factors such as the instruments' sampling strategy, relative position, and scanning range.\n• Imcomplete data: Some parts of real-scanned surfaces are incomplete due to self or external occlusion.\n• Noise: It is inevitable that noise from environmental factors or inaccuracies in instruments will be present.\n• Permutation invariance: The order of points does not affect the overall semantic representation of point cloud objects, so identical point cloud objects can be expressed by various matrices. (Chang, Funkhouser, Guibas, Hanrahan, Huang, Li, Savarese, Savva, Song, Su et al., 2015) 57 "
},
{
"figure_ref": [],
"heading": "Point Cloud Dataset",
"publication_ref": [
"b24",
"b98",
"b8",
"b58",
"b78",
"b4",
"b14",
"b86",
"b81",
"b6"
],
"table_ref": [
"tab_0"
],
"text": "Quality benchmark datasets (e.g. complete, well-varied, and densely-labeled) play essential roles in SSL research. This section lists the most commonly used point cloud datasets and summarize them in Table 1 in terms of sample number, object categories, suitable tasks, and highlights. These datasets may contain synthetic and real scanned data, in single frames and time series and from individual objects and complex scenes. There are also a few datasets for complex traffic scenarios (e.g. automatic driving) containing extra data in different modalities, such as from images or radars.\n• KITTI (Geiger et al., 2012) is a benchmark suite for autonomous driving vision tasks. The dataset was collected using several pieces of equipment, including four video cameras, a laser scanner, and a localization system. It includes not only point clouds but also stereo and optical flow data. There are more than 200,000 annotated point cloud scenarios consisting of cars and pedestrians, providing a novel and challenging benchmark for 3D object detection and orientation estimation.\n• ModelNet (Wu et al., 2015) is the most widely used 3D point cloud CAD dataset for object classification and few-shot learning. It contains 12,311 single objects from 40 categories, with each point composed of six dimensions of information, including XYZ spatial coordinates and RGB values.\n• ShapeNet (Chang et al., 2015) is a relatively largescale repository of 3D CAD objects frequently employed as a pre-training dataset. It contains more than 3 million samples categorized into 55 classes under the WordNet synsets (Miller, 1995) criteria. The annotations in the dataset are versatile, including rigid alignments, parts, physical sizes, and key points.\n• SUN RGBD (Song et al., 2015) • S3DIS (Armeni et al., 2016) is a 3D indoor venue dataset that consists of scanning of 272 rooms in 6 areas overlaying a 6,000 𝑚 2 area. It has 13 semantic categories labeled by fine-grained point-wise annotations carrying full 9D information, including XYZ, RGBs, and normalized location coordinates.\n• ScanNet (Dai et al., 2017) is a 3D RGB-D dataset that comprises 2.5M views in 1,513 scenes acquired in 707 indoor environments. Various tests containing semantic voxel labeling and CAD model retrieval proved that ScanNet provides quality data for 3D scene understanding.\n• ScanObjectNN (Uy et al., 2019) was proposed as a collection of real-world indoor point cloud scenes to break the performance saturation of 3D object classification on synthetic data. This dataset introduces new challenges for 3D object classification due to the presence of background noise and occlusions that require networks' ability on context-based reconstructions and partial observations.\n• Waymo (Sun et al., 2020) is a large autonomous driving dataset produced by Waymo in collaboration with Google Inc. The dataset consists of 1,150 urban and suburban geography scenes spanning 20 seconds, which are collected via well-synchronized and calibrated LiDARs and cameras. • NuScenes (Caesar et al., 2020) is another remarkable multimodal dataset provided by the full sensor suite including cameras, radars, and LiDARs. Compared to other autonomous driving datasets, it contains additional annotations like pedestrian pose, vehicle state, and also scenes from nighttime and rainy weather."
},
{
"figure_ref": [
"fig_2",
"fig_3"
],
"heading": "Point cloud data augmentation",
"publication_ref": [
"b114",
"b30",
"b64",
"b109",
"b114",
"b10",
"b114",
"b110",
"b22"
],
"table_ref": [],
"text": "Data augmentation is a crucial technique for enhancing DNNs performance by increasing the amount and diversity of training samples. For SSL tasks, it not only prevents the model from overfitting but also facilitates capturing robust and invariant representations of point clouds under multiple transformations. In this section, we will introduce the commonly used data augmentation methods and compare the effectiveness of each methods via a metric called task relatedness.\nEssentially, data augmentation is a process of generating new data by adding interventions or corruptions without destroying the original semantic expressions. For point clouds, augmentation methods are based on the properties mentioned in Section 2.2 and can be classified into three general groups: density/masking, noise, and affine transformation (Zhang et al., 2022c). These three corruption families could be further divided into 14 sub-categories as shown in Fig. 3.\nDensity/masking is the most frequent data augmentation method adopted in mask autoencoder (MAE) type SSL research (He et al., 2021;Pang, Wang, Tay, Liu, Tian and Yuan, 2022;Yu, Tang, Rao, Huang, Zhou and Lu, 2021). Based on the principle that point cloud data is sparse with uneven density, randomly removing a certain percentage of points while preserving part of the semantic expression presents a challenging learning objective for such MAEbased tasks. On the contrary, the noise based methods impose interventions on the original clean input to increase the difficulty of feature extraction. Affine transformation leverages point cloud invariance characteristics to shift the spatial coordinates of each points. This has significant impact on the input since the basic position information completely changes. The figure is adapted from (Zhang et al., 2022c).\nThe work in (Chen et al., 2020;Zhang et al., 2022c) investigated the effectiveness of the aforementioned augmentation methods as pretext data preprocessing on downstream classification tasks. Task relatedness is employed as the evaluation metric to statistically measure the performance of SSL models on downstream tasks, which provides valuable advice for proxy data augmentation selection. Following (Zamir, Sax, Shen, Guibas, Malik and Savarese, 2018), for each pretext task 𝑐, its task relatedness to downstream task 𝑡 is defined as:\n𝐴 𝑐→𝑡 ∶= 𝔼 𝑥∈𝑋 𝑡 (𝑅 𝑐 (𝐸 𝑐 (𝑥)), 𝑓 𝑡 (𝑥)) (1)\nWhere 𝑥 is a sample in a point cloud dataset 𝑋; 𝐸 𝑐 is the model's encoder pre-trained on task 𝑐; 𝑅 𝑐 is a readout function, which indicates the classification head composed of several fully connected (FC) layers; 𝑓 𝑡 is the labeling function; 𝑡 is accuracy measurement estimating whether the downstream output 𝑅 𝑐 (𝐸 𝑐 (𝑥)) conforms to the ground truth 𝑓 𝑡 (𝑥).\nTo further explore the relationship between task relatedness and classification accuracy on downstream tasks, Pearson correlation coefficient 𝑟 and 𝑝-value are utilized to estimate the linear relationship as well as statistical significance (Fraser, 1976), respectively, where |𝑟| > 0.5 refers to a strong correlation and 𝑝 < 0.05 is considered statistically significant. Fig. 4 demonstrates the statistically significant linear relationship between task relatedness and classification accuracy on downstream tasks when 𝑟 = 0.89 and 𝑝 < 0.001. The results reveal a counter-intuitive fact that frequently used density/mask and noise-based data augmentation methods are ineffective for downstream tasks either in accuracy and task relatedness. Conversely, the seemingly simple affine transformation enhances task relatedness to point cloud classification, resulting in higher accuracy. "
},
{
"figure_ref": [],
"heading": "Model",
"publication_ref": [
"b68",
"b119",
"b28",
"b26"
],
"table_ref": [],
"text": "Year Architecture Contributions PointNet (Qi et al., 2017a) 2017 CNN Pioneer in direct processing of raw point clouds with lightweight architecture PointNet++ (Qi et al., 2017b) 2017 CNN Aggregating local neighborhood by multi-scale and multi-resolution sampling and groping VoxelNet (Zhou and Tuzel, 2018) 2018 3D CNN Partitioning disordered point clouds into regular voxels for local feature learning DGCNN (Wang et al., 2019) 2019 Graph CNN Constructing a dynamic local graph to capture edge features around a neighbor PCT (Guo, Cai, Liu, Mu, Martin and Hu, 2021) 2021 Transformer Successfully capturing the long-range dependencies between point patches GANs (Goodfellow, Pouget-Abadie, Mirza, Xu, Warde-Farley, Ozair, Courville and Bengio, 2014) 2014 GAN Generating synthetic data through adversarial training Furthermore, combining corruptions of affine transformation and mask can approach the performance of supervised benchmarks. Hence, using affine transformation-based methods for data augmentation is preferable for in SSL pretraining."
},
{
"figure_ref": [],
"heading": "Popular deep models for point clouds",
"publication_ref": [],
"table_ref": [
"tab_3"
],
"text": "SSL techniques designed for languages and images need to be revised and extended for point clouds. For instance, traditional CNN networks cannot handle irregular and discrete point cloud data well since there is no guarantee that a corresponding point exists at the same relative position of the convolution. In this section, we briefly introduce five point cloud networks that are frequently used as feature extraction encoders in the literature and summarize their respective characteristics in Table 2."
},
{
"figure_ref": [],
"heading": "PointNet",
"publication_ref": [
"b66",
"b68",
"b109",
"b64",
"b113"
],
"table_ref": [],
"text": "To reduce data size and computation complexity, Qi et al. proposed PointNet (Qi et al., 2017a), which is the pioneering work to extract features directly on raw point clouds. It is widely deployed as the feature extractor (Wang, Liu, Yue, Lasenby and Kusner, 2021b;Poursaeed, Jiang, Qiao, Xu and Kim, 2020;Sauder and Sievers, 2019b) due to its simple and lightweight network structure. Taking advantage of the point permutation invariance, PointNet aligns the input points to a canonical space and aggregates global features by symmetric functions such as max pooling.\nHowever, it fails to capture local structures induced by the metric space in which the points reside, thereby limiting its ability to recognize fine-grained patterns and generalize to complex scenes. The updated version PointNet++ (Qi et al., 2017b) was then put forward several months later. It adopts multi-scale, multi-resolution sampling, and groping strategies to propagate features from one level to another, which improves the feature learning ability further. Furthermore, the point patch generation strategy combining Farthest Point Sampling (FPS) and K-Nearest Neighbor (KNN) provides a template for point cloud cropping preprocessing for subsequent studies (Yu et al., 2021;Pang et al., 2022;Zhang, Lin, He, Chen, Jia and Zhang, 2022b)."
},
{
"figure_ref": [],
"heading": "VoxelNet",
"publication_ref": [
"b119",
"b48",
"b59",
"b32"
],
"table_ref": [],
"text": "VoxelNet (Zhou and Tuzel, 2018) is a generic pointspecific network that uses voxels (i.e. finite unit cubes), to divide and access a local representation of point clouds for 3D detection tasks (Li, Yu, Meng, Caine, Ngiam, Peng, Shen, Lu, Zhou, Le et al., 2022;Min, Zhao, Xiao, Nie and Dai, 2022;Hess, Jaxing, Svensson, Hagerman, Petersson and Svensson, 2022). This network partitions disordered point clouds and performs feature learning in quantified and fixed-size 3D structures. One innovation is the stacking Voxel Feature Encoding (VFE) layers which encode interaction between points within a voxel and grasp descriptive appearance information. The output of each VFE layer is the concatenation of point-wise features and locally aggregated features so that local features are better captured. However, the expensive computation of voxel construction and quantization artifacts constrain the model from capturing highresolution or fine-grained representations."
},
{
"figure_ref": [],
"heading": "DGCNN",
"publication_ref": [
"b66"
],
"table_ref": [],
"text": "A point with its neighbors can reflect the geometry property of a local point cloud. Such a local relationship could be expressed by a graph network. Therefore, Wang et al. proposed a dynamic graph-based CNN network (DGCNN) (Wang et al., 2019) that encodes the edge features between vertices. Instead of learning point representations directly, DGCNN represents the interactions between points and their edges in both Euclidean and semantic space, and learns the graph structure dynamically. This graph networkbased architecture has served as a backbone in many subsequent point cloud SSL models with notable results (Poursaeed et al., 2020;Sauder and Sievers, 2019b; Afham, Dissanayake, Dissanayake, Dharmasiri, Thilakarathna and Rodrigo, 2022)."
},
{
"figure_ref": [],
"heading": "GAN",
"publication_ref": [
"b26"
],
"table_ref": [],
"text": "Generative Adversarial Network (GAN) (Goodfellow et al., 2014) is a widely used framework in reconstructionbased pretext tasks for point cloud knowledge mining. It consists of two components: the generator, which generates point clouds similar to the training data, and the discriminator, which distinguishes between generated and real points. These two modules are trained under an adversarial paradigm without supervision. The framework can be formulated as a two-player minimax game:\n𝑚𝑖𝑛 𝐺 𝑚𝑎𝑥 𝐷 𝐸 𝑥∈𝑋 [𝑙𝑜𝑔(𝐷(𝑥))]+𝐸 𝑧∈𝑍 [𝑙𝑜𝑔(1-𝐷(𝐺(𝑧)))] (2)\nwhere 𝐷 and 𝐺 denotes the discriminator and the generator, and 𝑋 and 𝑍 represent the data and noise distribution, respectively."
},
{
"figure_ref": [],
"heading": "Transformers",
"publication_ref": [
"b113",
"b28"
],
"table_ref": [],
"text": "Transformers have become one of the most prevalent architectures in many fields. They benefit from the multihead self-attention mechanism, which allows them to capture long-range dependencies between point patches and discover implicit regional correlations. The state-of-the-art performance on SSL point cloud classification and part segmentation has been achieved by transformer-based models such as the one proposed by Zhang et al. (Zhang et al., 2022b). Furthermore, point cloud transformer (PCT) (Guo et al., 2021), a variant adapted specifically for point clouds, enhances local feature extraction with the support of farther point sampling and nearest neighbor search, and further improves performance on various downstream tasks."
},
{
"figure_ref": [],
"heading": "Pseudo labels",
"publication_ref": [
"b99",
"b31",
"b7"
],
"table_ref": [],
"text": "Pseudo labels are introduced in point cloud SSL due to the absence of ground truth labels. It facilitates the calculation of loss with the output of the pretext tasks, which is then used for updating encoders via backpropagation. Information contained in pseudo labels is often considered as a more reliable and informative source for pretext tasks to learn point cloud representation than tags. For instance, the label 'airplane' only indicates the shape of objects without descriptions like colors, poses, and differences from other samples in same category. In contrary, these attributes are implicitly contained in point clouds and could be captured as pseudo label in SSL tasks.\nDifferent methods define pseudo labels in different ways. In most reconstruction-based pretext tasks, pseudo labels are point cloud itself which provides a rebuilding objective for pretext task. In contrast-based methods, pseudo labels are a multidimensional matrix carrying collection information and are typically generated using clustering methods such as memory bank (Wu, Xiong, Yu and Lin, 2018), online dictionary (He et al., 2020), and prototype approaches (Caron et al., 2020), representing mean and variance of all or part of the features of point cloud dataset. For some alignmentbased prediction or motion-based tasks pre-trained on temporal point cloud datasets, pseudo labels are geometric information like position, pose, and orientation in a number of frames before and after."
},
{
"figure_ref": [],
"heading": "Loss functions",
"publication_ref": [
"b31"
],
"table_ref": [],
"text": "Appropriate and easily-differentiable loss functions are critical to facilitate backpropagation and optimization for encoders. In reconstruction-based pretext tasks, the symmetric function, Chamfer distance (CD), is commonly employed to assess the distance between each point in one set and its corresponding nearest point in the other. More formally, for two non-empty subsets 𝑋 and 𝑌 , Chamfer distance 𝑑 𝐶𝐷 (𝑋, 𝑌 ) is defined as:\n𝑑 𝐶𝐷 (𝑋, 𝑌 ) = 1 |𝑋| ∑ 𝑥∈𝑋 min 𝑦∈𝑌 ||𝑥-𝑦|| 2 + 1 |𝑌 | ∑ 𝑦∈𝑌 min 𝑥∈𝑋 ||𝑥-𝑦|| 2(3)\nHere, 𝑥 and 𝑦 represent the points in the reconstruction point set 𝑋 and the original input point set 𝑌 , respectively;\n|| ⋅ || denotes the L2 distance between two points and | ⋅ | refers to the number of points. The smaller the CD value, the more similar the two point sets are, and the better the SSL algorithm performs.\nFor contrast-based pretext tasks, the objective is to discriminate the similarities and differences between each point cloud samples on the overall semantic level. A crossentropy like loss function to encourage the positive samples to be close to each other (and negative ones to be far from each other) is needed. InfoNCE (NCE stands for Noise-Contrastive Estimation) is a contrastive loss function that estimates the mutual information between a pair of samples, and can be formulated as:\n𝐿 𝑞 = -log exp(𝑞 ⋅ 𝑘 + ∕𝜏) ∑ 𝐾 𝑖=0 exp(𝑞 ⋅ 𝑘 𝑖 ∕𝜏)(4)\nwhere 𝑞 indicates the encoded query (feature); 𝑘 indicates a set of 𝐾 + 1 encoded samples {𝑘 0 , 𝑘 1 , 𝑘 2 , … , 𝑘 𝐾 }, which could be regarded as the prototypes of historical samples; 𝜏 is the temperature parameter controlling the sharpness of the distribution. Assuming there is only one positive sample 𝑘 + in the set 𝑘 matching the query 𝑞, the others 𝐾 samples are all negative. InfoNCE aims to assign the query 𝑞 into the positive sample 𝑘 + in the 𝐾 + 1 classification problem (He et al., 2020). In other words, the loss function tries to maximize the logits of 𝑞 ⋅ 𝑘 + and minimize the value of the denominator."
},
{
"figure_ref": [
"fig_1"
],
"heading": "Self-Supervised Learning pretext tasks for point cloud",
"publication_ref": [],
"table_ref": [],
"text": "We classify the current point cloud SSL research into four general categories based on the nature of the pretext tasks: reconstruction-based, contrast-based, alignmentbased, and motion-based methods, as shown in Fig. 2. These categories can be further divided into more fine-grained sub-categories according to the different ways in which the features are extracted and used. The following sections summarize the principles and peculiarities of various proxy tasks in details. It should be noted that some methods may reside in multiple sub-categories."
},
{
"figure_ref": [],
"heading": "Reconstruction-based methods",
"publication_ref": [],
"table_ref": [
"tab_4"
],
"text": "Reconstruction-based methods learn point cloud representations by reconstructing the corrupted point clouds and recovering the original ones as much as possible. Global features as well as the mappings between local and global areas are learned during the reconstruction process. According to different types of corruption and reconstruction objects, we further divide them into six sub-categories: mask recovery, spatial restoration, point sampling, disentanglement, deformation reconstruction, and generation and discrimination. Summary about the methods under these six sub-categories is shown in Table 3."
},
{
"figure_ref": [
"fig_4"
],
"heading": "Mask recovery",
"publication_ref": [
"b64",
"b72",
"b32",
"b109",
"b15",
"b64",
"b113",
"b59"
],
"table_ref": [],
"text": "The core idea of reconstruction is to mask a portion(s) of the point cloud and recover such missing part via an encoderdecoder architecture. Similar to the image inpainting task (2) The autoencoder pre-training is shown on the right. The encoder only processes the visible tokens while the mask tokens are shifted and added to the input sequence of the decoder to reconstruct the masked patches. This figure is adapted from (Pang et al., 2022). (Sarmad et al., 2019) and Mask AutoEncoder (MAE) (Hess et al., 2022), the encoder is required to capture the local geometric structure and the regional relations during the restoration process. Generally speaking, the better the reconstruction, the more effective the learned features. Point-BERT (Yu et al., 2021), built based on BERT (Devlin et al., 2018), designs a point-specific tokenizer on discrete Variational AutoEncoder (dVAE) to map patches to discrete tokens to capture meaningful local geometric patterns. A portion of the input is randomly masked out, and a BERT-style transformer is trained to reconstruct the missing token under the supervision of point tokens obtained by the tokenizer. However, the tokenizer should be pretrained in advance, and Point-BERT over-relies on auxiliary contrastive learning as well as data augmentation.\nTo address this issue, Pang et al. proposed Point-MAE (Pang et al., 2022) as a neat and efficient scheme of mask autoencoder as shown in Fig. 5. Concretely, Point-MAE employs the standard transformer as the backbone with an asymmetric encoder-decoder architecture to process random masking points with a high ratio (60%-80%). The mask tokens are shifted from the input of the encoder to the lightweight decoder, which saves considerable computation, and more significantly, avoids early leakage of location information. To further capture local geometric information, Zhang et al. introduced Mask Surfel Prediction (MaskSurf) (Zhang et al., 2022b), which estimates the surfel position (i.e., points) and per-surfel orientation (i.e., normals) simultaneously. Such a two-head pre-training paradigm has been justified to capture more effective representations than a reconstruction-only pretext. Likewise, Voxel-MAE (Min et al., 2022) transforms point clouds into volumetric representations and applies the range-aware random masking strategy on the voxel grid. Besides reconstructing the occupancy value of masked voxels, a supplementary binary voxel classification task distinguishing whether the voxel contains point clouds boosts the model to learn more complicated semantics."
},
{
"figure_ref": [
"fig_5"
],
"heading": "Spatial restoration",
"publication_ref": [
"b66"
],
"table_ref": [],
"text": "Point clouds are the coordinate sets containing abundant spatial information that describes the structural distribution of objects and the environment in the Euclidean space. It is natural to exploit such rich spatial knowledge as the supervision signal in pretext tasks.\nSauder et al. (Sauder and Sievers, 2019b) proposed a 3D version of the jigsaw pretext to rearrange point clouds whose parts have been randomly disrupted and displayed by voxels along the axes. The goal of this pretext is to restore the original position of each patch (labeled by voxel ID) from the state of chaotic and disorderly distribution. They In this case, the exact relation of these two components is 'the red part is diagonally above the blue part'. This figure is adapted from (Sauder and Sievers, 2019a).\nlater developed CloudContext (Sauder and Sievers, 2019a) to forecast the spatial relevance between two point cloud segments. As shown in Fig. 6, the model is trained to predict the relative structural position between two given patches from the same object, which utilizes the innate attributes of point clouds as they are not restrained by a discrete grid. By doing so, powerful per-point features can be accessed in an easy-to-implement unsupervised manner without expensive computation.\nOrientation estimation (Poursaeed et al., 2020) is another simple but effective proxy task to capture the spatial information of point clouds. With the canonical orientation provided in most datasets, the orientation estimation pretext task aims to predict and recover the rotation angle around an axis via matrix multiplication. Such a pretext requires a high-level holistic understanding of shapes and obviates the need for manual annotations. Point upsampling is the operation to upsample sparse, noisy, and non-uniform point clouds to generate a dense, complete, and high-resolution point cloud, which is challenging but also beneficial for the model to capture implicit geometric representations of the underlying surface."
},
{
"figure_ref": [
"fig_6"
],
"heading": "Point upsampling",
"publication_ref": [
"b45",
"b26",
"b28",
"b118"
],
"table_ref": [],
"text": "PU-GAN (Li et al., 2019) is a pioneer SSL upsampling paradigm formulated based on the generative adversarial network (GAN) (Goodfellow et al., 2014) to generate a diverse range of point distributions from the latent space and upsample points over patches. An up-down-up unit is embedded in the generator to expand point features as well as a self-attention unit for quality enhancement on feature aggregation. The discriminator is inspired to gain inherent patterns and improve the uniformity of output generation according to a compound loss including adversarial, uniform, and reconstruction terms. Motivated by PU-GAN, Zhang et al. proposed the Upsampling AutoEncoder (UAE) (Zhang et al., 2022a) to gain both advanced semantic information and basic geometric structure from subsampled point clouds. As shown in Fig. 7, the encoder is devised to perform point-wise feature extraction on the subsampled point cloud, and the upsampling decoder is designed to reconstruct the original dense point cloud with offset attention (Guo et al., 2021) to refine global shape structure.\nLiu et al. (Liu et al., 2022d) proposed a coarse-tofine reconstruction framework, dubbed SPU-Net, integrating self-attention with graph convolution network (GCN) for contextual feature extraction and generating fine point sets with hierarchically learnable 2D grids. Zhao et al. (Zhao et al., 2021) introduced SSPU-Net by leveraging the shape coherence between input sparse and generated dense point clouds. In addition, it has an image-consistent loss among multi-view rendered images to capture the latent patterns of underlying point structures.\nPUFA-GAN (Liu et al., 2022c), a frequency-aware framework, utilizes a graph filter to extract high frequency (HF) points of sharp edges and corners so that the discriminator could focus on the HF geometric properties and enforce the generator producing neat and more uniform upsampled point clouds. To get rid of the fixed upsampling factor restriction, Zhao et al. (Zhao et al., 2022a) presented a self-supervised arbitrary-scale (SSAS) framework with a magnificationflexible upsampling strategy. Instead of direct mapping from sparse to dense point clouds, the proposed scheme seeks the nearest projection points on the implicit surface for seed points via two functions, which are exploited to estimate the projection direction and distance, respectively."
},
{
"figure_ref": [
"fig_7"
],
"heading": "Disentanglement",
"publication_ref": [
"b85",
"b105",
"b80"
],
"table_ref": [],
"text": "Models pre-trained under the SSL paradigm usually tend to learn well the low-level geometric features of point clouds, such as pose, contour, and shape information, but overlook the high-level semantic content understanding, which often leads to unsatisfactory performance in downstream tasks such as object classification that requires global discriminative capability. To tackle this issue, disentanglementbased SSL pretexts are proposed to separate the low-level geometric features from the high-level semantic embedding. Feature extraction is performed based on various contents using distinct modules to obtain hierarchical representations.\nTsai et al. (Tsai et al., 2022) proposed a disentanglement framework that uncouples content and pose attributes in partial point clouds to enhance both geometric and semantic feature abstraction. Two encoders are employed to learn the content and multi-view poses separately, where the gained pose representation should predict the viewing angle and navigate the partial point cloud reconstruction cooperated with the content from another specific view. Likewise, Xu et al. (Xu et al., 2022) presented a universal Contour-Perturbed Reconstruction Network (CP-Net) that disentangles a point cloud into contour and content ingredients. A concise contour-perturbed augmentation unit is exploited on the contour component and retains the content part of the point cloud. Therefore, the self-supervisor is able to concatenate the content component for advanced semantic comprehension.\nDifferent from the above two pretexts, Mixing and Disentangling (MD) (Sun et al., 2022) blends two disparate point shapes into a hybrid object and attains geometryaware embedding from the encoder. An instance-adaptive decoder is then leveraged to restore the original geometries based on the obtained embedding by disentangling the mixed shape. As shown in Fig. 8, except for the main encoderdecoder structure, the proposed scheme also encompasses a coordinate extracting operation 'Erase', which randomly drops one-dimension coordinate of each point to provide an extra 2D partial projection to better reconstruct the original point cloud shapes."
},
{
"figure_ref": [
"fig_8"
],
"heading": "Deformation reconstruction",
"publication_ref": [
"b11",
"b11",
"b0",
"b107"
],
"table_ref": [],
"text": "Point cloud deformation is a common phenomenon in real-world data scanning, which is usually caused by object distortion, sensor noise, or external occlusion. It has been The input point cloud is firstly preprocessed by a shapedisorganizing module to generate a deformed point cloud and then fed to the encoder to learn the geometry-aware representation. Two separate task heads are constructed to distinguish and segment points belonging to distorted parts, and subsequently reconstruct the partial-deformed objects. The well-trained feature extractor is transferred to downstream tasks to estimate the feature capturing capability. This figure is adapted from (Chen et al., 2021).\ndiscovered that SSL by reconstructing the original point cloud from the artificially deformed one (e.g. adding Gaussian noise or local translation) enables the learned model to obtain geometric perception as well as context awareness.\nChen et al. (Chen et al., 2021) proposed a shape selfcorrection pretext to mine implicit geometric embeddings of point clouds. The pretext assumes that a robust shape representation could identify and correct distorted regions of a shape. As shown in Fig. 9, the proposed scheme imposes destruction over certain regions by a shape-disorganizing module and sends the deformed point cloud to the feature extractor for embedding learning. Two task heads are built separately to discern the distorted components and further restore them to their original normal shapes for fine-grained geometric and contextual feature exploration.\nAchituve et al. (Achituve et al., 2021) conducted the first study of SSL for domain adaptation (DA) on point cloud via Deformation Reconstruction (DefRec). By mapping the dislocating points to their original location, the model is able to obtain the latent statistical structure of the input point cloud. Moreover, the distribution gap between source and target domains is bridged by the learned representation since they are invariant to distribution shift.\nFoldingNet (Yang et al., 2018) presents a novel foldingbased decoder to perform deformation on the canonical 2D grid to fit an arbitrary 3D object surface. Instead of deforming the point cloud, the folding operation exerts a virtual force induced by the embedding captured from input to stretch a 2D grid lattice to reproduce the 3D surface structure. This approach tackles issues caused by point cloud's irregular attributes by applying implicit 2D grid constraints."
},
{
"figure_ref": [
"fig_9"
],
"heading": "Generation and discrimination",
"publication_ref": [
"b44",
"b72",
"b77",
"b45",
"b117"
],
"table_ref": [],
"text": "The generation and discrimination pretext is a unique paradigm that designs a discriminator module to distinguish whether the fed point cloud is reconstructed from noise distribution or truly sampled. During the adversarial training process, the generator (encoder) and discriminator (decoder) compete with each other and are updated alternatively so that both components can be transferred for downstream tasks.\nPC-GAN (Li et al., 2018) is specifically designed for point clouds and employs a hierarchical and interpretable sampling strategy inspired by Bayesian and implicit generative models to tackle the issue of missing constraints on the discriminator. Sarmad et al. (Sarmad et al., 2019) introduced a reinforcement learning (RL) agent to control the GAN to extract implicit representations from noisy and partial input to generate high-fidelity and entire point clouds. Meanwhile, applying an RL agent to seek the best-fit input of GAN to produce low-dimensional latent embedding relieves the challenge of unstable GAN training. Shu et al. (Shu et al., 2019) introduced a tree-structured graph convolutional network (TreeGCN) as the generator, leveraging ancestor information to boost the representation of the point. It is more efficient in computation than using neighborhood features as adopted in regular GCNs. PU-GAN (Li et al., 2019) and PUFA-GAN (Liu et al., 2022c), both employed GANs-based models to generate dense and uniform point clouds with innovative modules for feature aggregation enhancement and high-frequency point filtering.\nLiu et al. (Liu et al., 2022b) proposed a discriminative mask pretraining transformer framework, MaskPoint, which combines mask and discrimination techniques to perform simple binary classification between masked object points and sampled noise. As shown in Fig. 10, the original complete point cloud is divided into 90% masking portion and a 10% visible potion. Two kinds of query, where the real is sampled from masked point clouds while the fake is derived from random noise, are fed to the decoder for classification. During the discrimination process, the model is required to deduce the full geometry from small visible portions."
},
{
"figure_ref": [],
"heading": "Contrast-based methods",
"publication_ref": [],
"table_ref": [
"tab_5"
],
"text": "Contrastive learning is a popular mode of SSL that encourages augmentation of the same input to have more comparable representations. The general approach is to expand the views of input point clouds (anchors) by various data augmentation techniques. In particular, it tries to enforce positive samples augmented from the same anchor more similar than negative samples from different anchors in the feature space. In this section, we will introduce contrastbased methods with representative examples and discuss their contributions and limitations. A brief summary of these methods is shown in Table 4. Traditional contrastive learning research usually focuses on instance-wise objects. The priority is on overall semantic learning through discriminative pretext tasks that capture context similarity and difference of point clouds. Such an object-contrast paradigm performs data augmentation on relatively large patches or whole single point objects to capture global geometric awareness."
},
{
"figure_ref": [
"fig_10"
],
"heading": "Object contrast",
"publication_ref": [
"b70",
"b63",
"b88",
"b54",
"b112",
"b17",
"b42",
"b53",
"b115"
],
"table_ref": [],
"text": "Sanghi (Sanghi, 2020) proposed Info3D, which takes inspiration from Contrastive Predictive Coding (Oord, Li and Vinyals, 2018) and Deep InfoMax (Velickovic, Fedus, Hamilton, Liò, Bengio and Hjelm, 2019), to obtain rotationinsensitive representation by maximizing mutual information between 3D objects and their local chunks as well as geometrically transformed versions. Lu et al. (Lu et al., 2022) proposed the Augmentation Fusion Self-Supervised Representation Learning (AFSRL) framework, which imposes data-level augmentation and feature enhancement simultaneously to construct a stable and invariant point cloud embedding. The correspondence between augmented pairs is acquired, and the invariant semantic is maintained under perturbations during augmentation.\nZhang et al. (Zhang and Zhu, 2019) introduced a simple two-phase unsupervised GCN framework (contrasting and clustering), to capture superior point embedding by solving part contrast and object cluster tasks consecutively. Du (Du et al., 2021) presented a self-contrastive paradigm leveraging self-similar point cloud patches within a single point cloud to facilitate local shape and global context primitives capturing. As shown in Fig. 11, according to the nonlocal selfsimilar property of the point cloud, where regional geometry remains invariant after affine transformation, the self-similar (Lal, Prabhudesai, Mediratta, Harley and Fragkiadaki, 2021) Scene contrast Mapping RGB-D images to 3D points by optimizing view-contrastive prediction 2020\nP4Contrast (Liu, Yi, Zhang, Fan, Funkhouser and Dong, 2020) Scene contrast Utilizing synergies between two modalities for better feature extraction 2021\nDepthConstrast (Zhang, Girdhar, Joulin and Misra, 2021) Scene contrast Applying Instance Discrimination on depth maps point cloud patches are treated as positive samples otherwise negative based on the inferred similarity score. Moreover, hard negative samples, close to positive samples in the representation space, are sampled for more discriminative and expressive representation learning. Different from object-contrast, the scene-contrast paradigm concentrates on scenes to capture broader environmental context and neighborhood perception, which is more relevant to real-world complex scenarios."
},
{
"figure_ref": [
"fig_11"
],
"heading": "Scene contrast",
"publication_ref": [
"b102",
"b33",
"b103",
"b42",
"b53",
"b115",
"b99"
],
"table_ref": [],
"text": "To address the domain gap issue (i.e., it is insufficient to capture a global representation from object instances), Xie et al. (Xie et al., 2020) proposed PointContrast, a sparse residual U-Net based framework aiming to obtain dense features at the point-level on complex scenes. As shown in Fig. 12, two views 𝑥 1 and 𝑥 2 are produced from a complicated point cloud scene, where corresponding pairs are computed between these two views as the positive samples. Two rigid transformations 𝑇 1 and 𝑇 2 are utilized to increase the difficulty of the pretext which demands the network to learn the invariant embedding under random geometric shift. The contrastive loss is defined to shorten the distance between the matched points and enlarge the distance of mismatched points of the two overlapping partial scans so that the pre-training model can capture local descriptions and be universally pertinent to various advanced 3D understanding downstream tasks.\nHowever, PointContrast only considers point correspondence matching but ignored the spatial configurations and contexts in a scene, e.g., relative pose and distance, therefore confining its transferability and scalability. To address this issue, Hou et al. (Hou et al., 2021) presented Contrastive Scene Contexts to fuse spatial information into pre-training objects by introducing ShapeContext local descriptor (Xie, Liu, Chen and Tu, 2018) partitioning and performing contrastive learning in each region. The method improves the performance and data efficiency on downstream tasks in which employing only 0.1% of point labels reaches the performance level with full supervision.\nContinuous Contrastive 3D Networks (CoCoNets) (Lal et al., 2021) aims to infer latent scene representations by mapping RGB-D images to 3D point scenarios and optimizing view-contrastive prediction. P4Contrast (Liu et al., 2020), another RGB-D bi-modal SSL framework, proposes to contrast point-pixel pairs and provides additional flexibility for hard negative creation to exploit the synergies between two modalities for better feature extraction. Depth-Constrast (Zhang et al., 2021) circumvents the need for point correspondences and instead applies the Instance Discrimination (Wu et al., 2018) method on depth maps combined with a momentum encoder to improve the geometric perception."
},
{
"figure_ref": [],
"heading": "Alignment-based methods",
"publication_ref": [],
"table_ref": [
"tab_6"
],
"text": "Point cloud representation is generally invariant to transformations in terms of time flow, spatial motion, multi-view photography, etc. Based on this property, alignment-based methods have been proposed to learn the implicit embedding of point clouds by preserving the coherence of point features in spatiotemporal consistency, multi-view alignment, and multimodal fusion. A brief summary of the methods under this category is provided in Table 5. Compared to direct processing and feature extraction on 3D point clouds, projecting point clouds into 2D images for dimension reduction and utilizing mature image networks as well as 2D SSL technologies is relatively more accessible. To ensure that the learned embeddings sufficiently represent the entire 3D point cloud objects or scenes, multi-view alignment pretexts are necessary to preserve the integrity and uniformity of the point cloud features."
},
{
"figure_ref": [
"fig_12"
],
"heading": "Multi-view alignment",
"publication_ref": [
"b70",
"b106",
"b40",
"b84"
],
"table_ref": [],
"text": "Info3D (Sanghi, 2020) aims to obtain rotation-insensitive representations by maximizing mutual information between 3D objects and their local chunks for patch-level consistency. Occlusion Completion (OcCo) (Wang et al., 2021b) combines the idea of mask recovery shielding and restoring occluded points in a camera view for better spatial and semantic properties comprehension. Similarly, Yang et al. (Yang et al., 2021) introduced an SSL multi-view stereo structure generating prime depth map as pseudo-labels and refined such self-supervision from neighboring views as well as high-resolution images by multi-view depth fusion iteratively. Furthermore, the correspondence of pixel/point of the point clouds and the corresponding multi-view images are aligned for cross-modality consistency.\nJing et al. (Jing et al., 2021) proposed a novel SSL framework leveraging cross-modality and cross-view correspondences to jointly learn both 3D point cloud and 2D image embedding concurrently. As shown in Fig. 13, point cloud objects and comparable pairs of multi-view rendered images are sampled from the same mesh input. In addition to 2D-3D consistency, the contrastive notion is adopted into crossview alignment that shortens intra-object distance while maximizing inter-object discrepancy of distinct rendered images. Similarly, Tran et al. (Tran et al., 2022) presented a dual-branch model not only agreeing upon fine-grained pixel-point local representation but also encouraging 2D-3D global feature distributions as approaching as possible by exploiting knowledge distillation."
},
{
"figure_ref": [
"fig_14"
],
"heading": "Spatiotemporal consistency",
"publication_ref": [
"b104",
"b36",
"b27",
"b56",
"b48"
],
"table_ref": [],
"text": "Unlike previous methods, the spatiotemporal approach is more concerned with long-range spatial and temporal invariance before and after certain point cloud frames, which are 4D data (XYZ coordinate + temporal dimension), to capture intrinsic characteristics of dynamic sequences.\nMotivated by the success of Xu et al.'s work (Xu, Xiao, Zhao, Shao, Xie and Zhuang, 2019) in video SSL, Wang et al. proposed the first SSL scheme to gain effective temporal embeddings on dynamic point cloud data by sorting the temporal order of sampled and disorganized point cloud clips. As shown in Fig. 14, a few static point cloud frames are uniformly sampled and disordered, which are then processed by a 4D CNN to restore the disrupted fragments to the correct order on an unannotated, large-scale, sequential point cloud action recognition dataset.\nAnother spatiotemporal representation learning (STRL) (Huang et al., 2021) framework, inspired by BYOL (Grill, Strub, Altché, Tallec, Richemond, Buchatskaya, Doersch, Avila Pires, Guo, Gheshlaghi Azar et al., 2020), designs a dual-branch pipeline, referred to as online and target networks, to collaborate and promote each other. Specifically, the online network is enforced to predict the target network representation of another temporally correlated input, which is augmented by random spatial transformation, for spatiotemporal invariant contextual cues extraction. Taking training and inference time into account, Mersch et al. (Mersch et al., 2022) presented an innovative 3D spatiotemporal convolution encoder-decoder neural network consisting of fewer parameters to predict future point cloud scenes. Such a lightweight model concatenates range images as input to estimate forthcoming images and per-point scores in multiple future steps, so that spatial and temporal scene information can be captured simultaneously. The figure is adapted from (Li et al., 2022)."
},
{
"figure_ref": [
"fig_15",
"fig_15"
],
"heading": "Multimodal fusion",
"publication_ref": [
"b42",
"b40",
"b84",
"b89",
"b48",
"b89",
"b48"
],
"table_ref": [],
"text": "Rather than simply requiring coherence between 2D-3D correspondences (Lal et al., 2021;Jing et al., 2021;Tran et al., 2022), automatic driving algorithms demand sophisticated collaboration between in-vehicle sensors. For example, cameras and LiDARs provide complementary information (e.g., colorful texture visualization and distance perception) for 3D object detection. Therefore, multimodal fusion is a promising direction to exploit the potential of images and point clouds for acquiring effective traffic scene features.\nVora et al. (Vora et al., 2020), Wang et al. (Wang et al., 2021a), and Li et al. (Li et al., 2022) offered compact frameworks for tight sensor-fusion which could be implemented under the SSL paradigm without human annotations. Point-Painting (Vora et al., 2020) is a sequential fusion method that projects LiDAR points onto semantic segmentation diagrams for traffic scenes with color marking. Each point is painted with a class score obtained from the image segmentation network and then can be utilized in any LiDAR detection approaches. Such a painting fusion method cleverly addresses the limitations of depth-blurring and scale ambiguity by consolidating the birds-eye and camera view.\nPointAugmenting (Wang et al., 2021a) adopts a late cross-modal fusion mechanism based on PointPainting, replacing the sub-optimal segmentation scores with highdimension CNN features containing rich outlook hints and larger receptive fields to emphasize the delicate details. Moreover, a simple yet effective cross-modal data augmentation pastes virtual objects into images and point clouds for alignment between the camera and LiDAR. However, both PointPainting and PointAugmenting simply decorate LiDAR points with camera embeddings as shown in Fig. 15(a). To improve the performance on downstream tasks, DeepFusion (Li et al., 2022) proposed an end-to-end crossmodal fusion on the feature level, focusing on consistency improvement. As shown in Fig. 15(b), a block named Learn-ableAlign is introduced to exploit cross-attention to dynamically capture long-range correlations during the image-LiDAR fusion process to enhance the model's recognition and localization capability."
},
{
"figure_ref": [],
"heading": "Motion-based methods",
"publication_ref": [],
"table_ref": [
"tab_7"
],
"text": "Various point cloud frames contain rich geometric patterns and kinematic schemas that are concealed in the movement of objects or scenes. The motion-based SSL paradigm focuses on dynamically capturing the intrinsic motion characteristics from spatial variations by taking advantage of traditional registration and scene flow estimation as pretexts. A brief summary on the methods under this category is shown in Table 6. Point cloud registration is the task to merge two point clouds 𝑋 and 𝑌 into a globally consistent coordinate system via estimating the rigid transformation matrix, which can be formulated as: ("
},
{
"figure_ref": [
"fig_16"
],
"heading": "Registration",
"publication_ref": [
"b5",
"b117"
],
"table_ref": [],
"text": ")5\nwhere 𝑅 ∈ 𝑆𝑂( 3) and 𝑡 ∈ ℝ 3 indicate rotation matrix and translation vector, respectively; 𝜓 is the feature extraction network learning the hierarchical informative features from dynamic point clouds. Unlike the classic ICP registration method (Besl and McKay, 1992) which iteratively searches correspondences and estimates rigid transformation, SSL registration can obtain informative point cloud features without high-quality ground-truth correspondences.\nPRNet (Wang and Solomon, 2019) is a partial-to-partial registration method that enables coarse-to-fine refinement iteratively. Based on co-contextual information, the framework boils down the registration problem as a key point detection task, which aims to recognize the matching points from two input clouds. Shi (Wang and Solomon, 2019) presented a part mobility segmentation approach to understand the essential attributes of the dynamic object. Instead of directly processing the sequential point clouds, the raw input is converted to trajectories by point correspondence between successive frames to derive rigid transformation hypotheses. Analogously, Zhao et al. (Zhao et al., 2022b) proposed an SSL line segmentation and description for LiDAR point clouds, called SuperLine3D, providing applicable line features for global registration without any prior hints. Compared to point embedding constrained by limited resolution, this segmentation model is capable of obtaining precise line representation under arbitrary scale perturbations.\nMotivated by the observation that the local distinctive geometric structures of two subsets of point clouds can improve representations, Liu et al. (Liu et al., 2022a) introduced the deep versatile descriptors (DVDs) which learn local and global point embeddings jointly. As shown in Fig. 16, the co-occurring point cloud local regions, which retain the structural knowledge under rigid transformations, are regarded as the input of DVD to extract latent geometric patterns restrained by local consistency loss. To further enhance the model's capability of transformation awareness, reconstruction and normal estimation are added as auxiliary tasks for better alignment."
},
{
"figure_ref": [
"fig_18"
],
"heading": "Scene flow estimation",
"publication_ref": [
"b60",
"b97",
"b60",
"b46"
],
"table_ref": [],
"text": "Scene flow estimation is a vital computer vision task. For point clouds, its objective is to estimate the motion of objects by computing dense correspondences between consecutive LiDAR scans of a scene over time. The variation of points (Mittal et al., 2020).\ncan be represented as 3D displacement vectors to describe the motions in terms of scene flow. Wu et al. (Wu et al., 2020) introduced the notion of cost volume and proposed a learnable point-based network called PointPWC-Net. The cost volume is discretized as input point pairs to reduce computational complexity; additionally, an efficient upsampling strategy and wrap layers are employed. Mittal et al. (Mittal et al., 2020) proposed a novel SSL scene flow estimation network to achieve safe navigation during interactions with highly dynamic environments by optimizing two loss components based on the nearest neighbors and cycle consistency. As shown in Fig. 17, the nearest neighbor loss encourages the points predicted based on current moment 𝑡 flowing toward occupied regions of the future frame at 𝑡 + 1. The cycle consistency loss ensures that the points of the future frame 𝑡 + 1 can be restored in the reverse direction back to frame 𝑡 to avoid degenerate solutions by maintaining temporal consistency. Self-Point-Flow (Li et al., 2021) employs more than 3D point coordinates, surface normal, and color in one-to-one matching to generate pseudo labels and formulates the pseudo label generation issue as an optimal transport problem. It leverages a random walk module to refine annotation quality by imposing local alignment."
},
{
"figure_ref": [],
"heading": "Downstream tasks",
"publication_ref": [],
"table_ref": [],
"text": "One of the primary objectives of SSL is to pre-train a backbone network and transfer it to solve the problems in downstream tasks. Therefore, performance of the model in downstream tasks could reflect the effectiveness of SSL to a certain degree. The evaluation criteria indicate whether the SSL methods can extract useful knowledge from pretext tasks with large-scale unlabeled point cloud data. In this section, we introduce four commonly used downstream tasks and provide the widely used evaluation metrics. In addition, we summarize and compare the performance of the aforementioned representative SSL methods in the corresponding downstream tasks."
},
{
"figure_ref": [],
"heading": "Object classification",
"publication_ref": [],
"table_ref": [
"tab_8",
"tab_11"
],
"text": "Object classification is a fundamental and prevalent downstream task that requires the model to output a most likely label for the given point cloud object to assess the overall semantic awareness of the pre-trained model. The two commonly used metrics for this task are Overall Accuracy (OA) and Mean Class Accuracy (mAcc). OA is the ratio of correctly classified objects to the total number of objects, and mAcc is the average of each class's accuracy. Object classification can be divided into three protocols based on task settings:\n• Few-shot: Few-shot learning (FSL) is a challenging task that involves training with limited information provided by the downstream dataset. Specifically, the 𝑛-way, 𝑚-shot setting is employed, where 𝑛 is the number of classes randomly selected from the dataset and 𝑚 is the number of objects randomly sampled for each class. The trained model is evaluated on the test split. Few-shot protocol performance of reviewed SSL methods is shown in Table 7.\n• Fine-tuning: The pre-trained feature extractor serves as the initial downstream backbone encoder, and the entire network is re-trained in a supervised manner with labels from the downstream datasets. Fine-tuning protocol performance of proposed SSL methods is presented in Table 8.\n• Linear classification: The pre-trained feature extractor is frozen by stopping the backpropagation gradients. Linear classifiers are trained in a supervised manner with downstream datasets. Linear classification protocol performance of proposed SSL methods is shown in Table 9."
},
{
"figure_ref": [],
"heading": "Part segmentation",
"publication_ref": [
"b98",
"b86"
],
"table_ref": [],
"text": "Part segmentation is a fine-grained task that aims to distinguish and separate various components of an object, such as plane wings or desk legs. This task usually requires a model that can extract local point-level features more effectively than the overall discriminative ability required for object recognition. The popular evaluation criteria of point (Wu et al., 2015) and ScanOb-jectNN (Uy et al., 2019). The results are reported in terms of OA (%). "
},
{
"figure_ref": [],
"heading": "Semantic segmentation",
"publication_ref": [
"b4"
],
"table_ref": [
"tab_0"
],
"text": "Semantic segmentation requires a model to assign a semantic label to each points in the point cloud in order to group meaningful regions. It is frequently implemented on complicated outdoor or indoor scenes with background noise. mIoU, OA, and mAcc are commonly employed as estimation indicators to judge the feature extraction capability of pre-training models on the S3DIS dataset (Armeni et al., 2016), which contains six large-scale indoor venues, with the following two protocols. Performance of representative methods on semantic segmentation under the two protocols are shown in Table 11.\n• Area 5 test: The SSL pre-trained model is fine-tuned on all areas except the largest area 5, which is chosen as the test set.\n• Six-fold cross validation: Areas 1-6 are selected in turn as the test set and fine-tuned in the remaining 5 areas."
},
{
"figure_ref": [],
"heading": "Object detection",
"publication_ref": [
"b78",
"b14"
],
"table_ref": [
"tab_3"
],
"text": "Object detection is a task that involves localizing the 6 Degrees-of-Freedom (DoF) bounding box of an object and differentiating its category in a complex scene. The evaluation metric used is the average precision (AP), which measures the precision of the 3D bounding box at various recall levels. The threshold is usually set to 0.25 and 0.5. Table 12 summarizes the object detection performance of the SSL pre-training models on the SUN RGB-D (Song et al., 2015) and ScanNet (Dai et al., 2017) datasets."
},
{
"figure_ref": [],
"heading": "Table 8",
"publication_ref": [
"b98",
"b86"
],
"table_ref": [],
"text": "Summary of fine-tuning protocol performance of representative SSL methods on ModelNet40 (Wu et al., 2015) and ScanObjectNN (Uy et al., 2019). ScanObjectNN has three challenges. The results are reported in terms of OA (%). "
},
{
"figure_ref": [],
"heading": "Future directions",
"publication_ref": [],
"table_ref": [],
"text": "Although self-supervised learning has shown great success for point cloud processing, we have identified some of its deficiencies and limitations. We argue that SSL should not be studied in isolation but rather in conjunction with advanced techniques from other domains. In this section, we discuss a number of future research directions that have the potential to improve the SSL learning capability and performance on downstream tasks."
},
{
"figure_ref": [],
"heading": "Few-shot and zero-shot learning",
"publication_ref": [
"b23",
"b84",
"b84",
"b84",
"b84",
"b78",
"b14"
],
"table_ref": [
"tab_3"
],
"text": "There have been a good number of publicly available, labelled datasets for SSL research. However, real scenarios often face the data shortage or quality challenges, such as damaged labels, missing information, and uneven assortment. Few-shot learning (FSL) (Garcia and Bruna, 2017) is considered as a potential solution that allows the network to (Tran et al., 2022) Alignment PointNet -85.0 46.7 Multi-view rendering (Tran et al., 2022) Alignment DGCNN -87.0 49.9\nSix-fold cross validation OcCo (Wang et al., 2021b) Alignment PointNet 82.0 -54.9 OcCo (Wang et al., 2021b) Alignment DGCNN 84.6 -58.0 3D jigsaw (Sauder and Sievers, 2019b) Reconstruction PointNet 80.1 -52.6 3D jigsaw (Sauder and Sievers, 2019b) Reconstruction DGCNN 84.1 -55.6 CloudContext (Sauder and Sievers, 2019a) Reconstruction DGCNN 78.9 -47.6 Multi-view rendering (Tran et al., 2022) Alignment PointNet -83.2 52.1 Multi-view rendering (Tran et al., 2022) Alignment DGCNN -87.5 59.0 train under the situations with very small amount of data.\nIt is also possible to identify new sample types that have not been seen before in a test task without training samples. This method is often referred to as the zero-shot learning (ZSL). Both SSL and FSL (ZSL) (Romera-Paredes and Torr, 2015) can free models from the reliance on large annotated datasets\nTable 12 Summary of performance of representative methods on object detection using SUN RGB-D (Song et al., 2015) and ScanNet (Dai et al., 2017) and reduce the cost. In addition, the combination of these two could potentially improve the generalization capability the models."
},
{
"figure_ref": [],
"heading": "Multiple modality interaction and fusion",
"publication_ref": [
"b24",
"b81",
"b6",
"b89",
"b48"
],
"table_ref": [],
"text": "Despite of the assorted modalities in many existing datasets, for example, for outdoor autonomous driving (Geiger et al., 2012;Sun et al., 2020;Caesar et al., 2020), researchers normally only focus on and make use of the point cloud data while ignoring the connections and alignment relationships with data of other modalities. We have seen some recent research works design models (Vora et al., 2020;Wang et al., 2021a;Li et al., 2022) for multi-model data alignment and fusion, primarily point clouds and images. We anticipate more research to focus on cross-modal SSL with more diverse modalities, e.g., natural language, radar and voice, exploiting the unique of each modality and the synergy among them to build transportation systems, e.g. autonomous driving and traffic scene analysis, with more artificial general intelligence."
},
{
"figure_ref": [],
"heading": "Hierarchical feature extraction",
"publication_ref": [],
"table_ref": [],
"text": "To cope with sophisticated downstream tasks with somehow conflicting objectives, for example, object classification which requires overall semantic understanding and part segmentation which requires fine-grained geometrical awareness, SSL models should have the capability for both global perception and local analysis. This necessitates hierarchical feature extraction; in particular, interactions between feature representations on different levels in the hierarchy need to be considered to discover the implicit relations. Therefore, we suggest that hierarchical feature extraction should be embedded in the SSL paradigm to improve the model's capability to capture both global and local features from point clouds."
},
{
"figure_ref": [],
"heading": "Multiple tasks pre-training",
"publication_ref": [],
"table_ref": [],
"text": "Up to now, most point cloud SSL methods have only one specific pre-training pretext while few works train diverse tasks concurrently. The main resistance is that multi-tasking has to consider the compatibility and synergy between various pretexts simultaneously, and fit each loss item for steady parameter updating. This is also one of the very reasons why a model performs well on one downstream task but not others. Indeed, distinct proxies could provide useful information from various perspectives of point clouds so that jointly training multiple tasks could facilitate the network to learn more comprehensive representations; obviously, more research on multi-task SSL is needed to push the research one step further."
},
{
"figure_ref": [],
"heading": "Theory and interpretability",
"publication_ref": [
"b75",
"b93"
],
"table_ref": [],
"text": "Similar to traditional deep learning, point cloud SSL lacks sufficient theoretical support and has poor interpretation. The process of model training is conducted as a black-box, making it difficult for human users to analyze the results. Most of the technical works demonstrate their contributions via ablation studies and draw conclusions empirically. Such 'tried and tested' methods do not have theoretical support and are therefore difficult to verify, generalize and replicate. We suggest that future studies should include more inquiries into explainable theory, for example, the well-established theories from mutual information (Sayed, Brattoli and Ommer, 2018) or causal inference (Wang, Lin, Feng, He, Lin and Chua, 2022), which can be applied in the design of network structures and loss functions."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "Point cloud self-supervised learning fundamentally moves away from models' dependency on manual annotations. The learning paradigm focuses on the design of pre-training pretext tasks to enable the models to extract effective features and achieves performance competitive to the supervised learning paradigms in many downstream tasks. This paper extensively surveys recent representative deep neural network-based methods for self-supervised learning from point cloud data. A novel taxonomy is proposed to systematically classify the current research, especially the works publishes in the recent three years. Besides detailed analysis on the representative methods, we provide summaries on the commonly used datasets and performance comparison to make the survey more comprehensive. Future research directions are also discussed to hopefully provide an insightful view on the issues that the research community should pay attention to. We hope that our work provides a valuable reference on point cloud SSL research and could motivate researchers to further explore this promising topic."
},
{
"figure_ref": [],
"heading": "Acknowledgement",
"publication_ref": [],
"table_ref": [],
"text": "This work received financial support from Jiangsu Industrial Technology Research Institute (JITRI) and Wuxi National Hi-Tech District (WND)."
}
] | 3D point clouds are a crucial type of data collected by LiDAR sensors and widely used in transportation applications due to its concise descriptions and accurate localization. Deep neural networks (DNNs) have achieved remarkable success in processing large amount of disordered and sparse 3D point clouds, especially in various computer vision tasks, such as pedestrian detection and vehicle recognition. Among all the learning paradigms, Self-Supervised Learning (SSL), an unsupervised training paradigm that mines effective information from the data itself, is considered as an essential solution to solve the time-consuming and labor-intensive data labelling problems via smart pretraining task design. This paper provides a comprehensive survey of recent advances on SSL for point clouds. We first present an innovative taxonomy, categorizing the existing SSL methods into four broad categories based on the pretexts' characteristics. Under each category, we then further categorize the methods into more fine-grained groups and summarize the strength and limitations of the representative methods. We also compare the performance of the notable SSL methods in literature on multiple downstream tasks on benchmark datasets both quantitatively and qualitatively. Finally, we propose a number of future research directions based on the identified limitations of existing SSL research on point clouds. | Self-Supervised Learning for Point Clouds Data: A Survey | [
{
"figure_caption": "Figure 1 :1Figure 1: The general pipeline of SSL used in the point cloud data. (1) Pre-training stage: point cloud data is firstly preprocessed through the augmentation block and then fed into the point-specific encoder to learn feature representations. The features are utilized to complete well-design pretext tasks, where the output will be compared with the pseudo labels derived from the original data to generate a loss and to update encoder parameters via back-propagation; (2) Supervised finetuning stage: the well-trained encoder is transferred to the target domain. A task head is trained with the training labels in a supervised manner to complete the downstream tasks; (3) Inference stage: the encoder and task head are concatenated as a model to execute inference on the test set. The effectiveness of the SSL pre-training framework can be evaluated based on the performance of the model on the downstream tasks.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: Taxonomy of SSL for point cloud data based on pretext tasks.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: Illustration of the commonly used data augmentation methods for point cloud data. There are a total of 14 sub-categories of data augmentation methods that could be classified as three general corruption families. The figure is adapted from (Zhang et al., 2022c).",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: Illustration of the relationship between task relatedness and classification accuracy on downstream tasks. 𝑟 and 𝑝 are the coefficients to measure the linear relationship and statistical significance for the Pearson correlation, respectively. The figure is adapted from (Zhang et al., 2022c).",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 :5Figure 5: The general pipeline of Point-MAE. (1) The process of masking and embedding is demonstrated on the left. The point cloud patches are generated by FPS and KNN and masked randomly. Both visible and mask patches are mapped to the corresponding tokens through PointNet-based embedding layers. Also, the Position Embedding (PE) is obtained by mapping the center coordinates to the embedding dimension.(2) The autoencoder pre-training is shown on the right. The encoder only processes the visible tokens while the mask tokens are shifted and added to the input sequence of the decoder to reconstruct the masked patches. This figure is adapted from(Pang et al., 2022).",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Figure 6 :6Figure 6: Illustration of the CloudContext pretext task. The pre-training model is enforced to estimate the spatial relevance between two given point cloud segments from six categories.In this case, the exact relation of these two components is 'the red part is diagonally above the blue part'. This figure is adapted from(Sauder and Sievers, 2019a).",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "Figure 7 :7Figure 7: Overview of Upsampling AutoEncoder. The input point cloud is subsampled by a random sampling strategy and then fed into the encoder to extract point-wise features.The decoder is adopted to reconstruct the original point cloud with offset attention based on the learned representation. This figure is adapted from(Zhang et al., 2022a).",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "7",
"figure_type": "figure"
},
{
"figure_caption": "Figure 8 :8Figure 8: The schematic of Mixing and Disentangling (MD)pretext. The two input point clouds are separately halved and mixed into a hybrid object feeding to the encoder 𝐸 to mine the geometry-aware embedding. The 'Erase' operation is applied to obtain the 2D projection from both original input point clouds simultaneously. The instance-adaptive decoder 𝐷 receives the embedding together with the two partial projections as input to disentangle the blended shape into the original two point clouds. The chamfer distance is used to measure the reconstruction error between generated point clouds and the original ones. This figure is adapted from(Sun et al., 2022).",
"figure_data": "",
"figure_id": "fig_7",
"figure_label": "8",
"figure_type": "figure"
},
{
"figure_caption": "Figure 9 :9Figure 9: Demonstration of shape self-correction pretext.The input point cloud is firstly preprocessed by a shapedisorganizing module to generate a deformed point cloud and then fed to the encoder to learn the geometry-aware representation. Two separate task heads are constructed to distinguish and segment points belonging to distorted parts, and subsequently reconstruct the partial-deformed objects. The well-trained feature extractor is transferred to downstream tasks to estimate the feature capturing capability. This figure is adapted from(Chen et al., 2021).",
"figure_data": "",
"figure_id": "fig_8",
"figure_label": "9",
"figure_type": "figure"
},
{
"figure_caption": "Figure 10 :10Figure 10: The general pipeline of the Mask-Point model.The reconstruction challenge is formulated as a discriminative pretext to determine whether the source of the extracted sample is a masked point cloud or a random noise. The figure is adapted from(Liu et al., 2022b).",
"figure_data": "",
"figure_id": "fig_9",
"figure_label": "10",
"figure_type": "figure"
},
{
"figure_caption": "Figure 11 :11Figure 11: Illustration of a self-contrastive paradigm. Patch A is selected as the anchor and the symmetrical part Patch D is the positive sample. Patch B and C are the negative samples, where Patch B is hard to distinguish due to its comparative similarity to the anchor. The figure is adapted from (Du et al., 2021).",
"figure_data": "",
"figure_id": "fig_10",
"figure_label": "11",
"figure_type": "figure"
},
{
"figure_caption": "Figure 12 :12Figure 12: The illustration of PointContrast. Contrast is performed at the point-level between two transformed point clouds, where positive samples are the matched points while negative samples are the unmatched points across two views. The figure is adapted from (Xie et al., 2020).",
"figure_data": "",
"figure_id": "fig_11",
"figure_label": "12",
"figure_type": "figure"
},
{
"figure_caption": "Figure 13 :13Figure 13: The schematic view of cross-modality and crossview correspondences. The 3D point cloud objects and corresponding pairs of multi-view rendered images are sampled from the same mesh input, respectively. The relation of diverse views is captured as the supervision signal by sustaining the alignment among multi-view and cross-domain representations. The figure is adapted from (Jing et al., 2021).",
"figure_data": "",
"figure_id": "fig_12",
"figure_label": "13",
"figure_type": "figure"
},
{
"figure_caption": "alignment Maximizing mutual information between objects and their transformations 2021 OcCo (Wang et al., 2021b) Multi-view alignment Shielding and restoring occluded points in camera view 2021 Multi-view stereo (Yang, Alvarez and Liu, 2021) Multi-view alignment Generating prime depth map as self-supervision signal 2021 Cross-view (Jing, Zhang and Tian, 2021) Multi-view alignment Jointly learning both 3D point cloud and 2D image embedding concurrently 2022 Multi-view rendering (Tran, Hua, Tran and Hoai, 2022) Multi-view alignment Encouraging 2D-3D global feature distributions to be similar 2021 Order prediction (Wang, Yang, Rong, Feng and Tian, 2021c) Spatiotemporal consistency Sorting temporal order of sampled and disorganized point cloud clips 2021 STRL (Huang, Xie, Zhu and Zhu, 2021) Spatiotemporal consistency Dual-branch network to predict representation of another temporally correlated input 2022 Futrue prediction (Mersch, Chen, Behley and Stachniss, 2022) Spatiotemporal consistency Forecasting future point cloud scenes with lightweight model 2020 PointPainting (Vora, Lang, Helou and Beijbom, 2020) Multimodal fusion Projecting LiDAR points into semantic segmentation diagram for traffic scenes 2021 PointAugmenting (Wang, Ma, Zhu and Yang, 2021a) Multimodal fusion Replacing sub-optimal segmentation scores with high-dimension CNN features 2022 DeepFusion (Li et al., 2022) Multimodal fusion Exploiting cross-attention to capture long-range correlations of image-LiDAR pairs",
"figure_data": "",
"figure_id": "fig_13",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 14 :14Figure 14: Demonstration of point cloud sequence order prediction. The first row is the uniformly sampled point cloud clips from the continuous point cloud sequence. Then these clips are randomly shuffled and then fed into 4D CNN in the second row to learn the dynamic features of human actions,The original temporal order is predicted in a self-supervised manner. The figure is adapted from(Wang et al., 2021c).",
"figure_data": "",
"figure_id": "fig_14",
"figure_label": "14",
"figure_type": "figure"
},
{
"figure_caption": "Figure 15 :15Figure 15: Demonstration of point decoration and deep fusion. (a) Previous cross-modal paradigms (Lal et al., 2021; Jing et al., 2021) decorate LiDAR points with camera feature on input-level for 3D detection. (b) DeepFusion (Li et al., 2022) fuses camera and LiDAR features extracted by respective encoders and leverages cross-attention consistency technique.The figure is adapted from(Li et al., 2022).",
"figure_data": "",
"figure_id": "fig_15",
"figure_label": "15",
"figure_type": "figure"
},
{
"figure_caption": "Figure 16 :16Figure 16: Demonstration of deep versatile descriptors. The input consists of two point clouds before and after rigid transformations, where the common point components are utilized to train the encoder for global and local representation learning. The figure is adapted from (Liu et al., 2022a).",
"figure_data": "",
"figure_id": "fig_16",
"figure_label": "16",
"figure_type": "figure"
},
{
"figure_caption": "-to-partial point cloud registration enabling coarse-to-fine refinement 2021 Part mobility (Shi, Cao and Zhou, 2021) Registration Converting points to trajectories to derive the rigid transformation hypotheses 2022 SuperLine3D (Zhao, Yang, Huang, Chen, Ma, Li and Liu, 2022b) Registration Obtaining precise line representation under arbitrary scale perturbations 2022 DVD (Liu, Chen, Xu, Qiu and Chu, 2022a) Registration Learning local and global point embedding jointly 2020 PointPWC-Net (Wu, Wang, Li, Liu and Fuxin, 2020) Scene flow estimation Discretizing cost volume onto 3D point clouds in a coarse-to-fine fashion 2020 Just go with the flow (Mittal, Okorn and Held, 2020) Scene flow estimation Optimizing two SSL losses based on nearest neighbors and cycle consistency 2021 Self-Point-Flow (Li, Lin and Xie, 2021) Scene flow estimation Converting pseudo label matching problem as optimal transport task 𝑅, 𝑡 = arg min 𝑅∈𝑆𝑂(3),𝑡∈ℝ 3 ‖𝜓(𝑅𝑋 + 𝑡) -𝜓(𝑌 )‖ 2 .",
"figure_data": "",
"figure_id": "fig_17",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 17 :17Figure17: Demonstration of just going with the flow. The nearest neighbor loss is utilized to push the predicted flow (green) close to the pseudo-ground truth (red) of the frame at 𝑡 + 1. The cycle consistency loss is the penalty term to estimate the flow between predicted points (green) in the opposite direction to the original points (blue) in frame at 𝑡 for temporal alignment. The figure is adapted from(Mittal et al., 2020).",
"figure_data": "",
"figure_id": "fig_18",
"figure_label": "17",
"figure_type": "figure"
},
{
"figure_caption": "Summary of commonly used point cloud datasets. Abbreviations for suitable tasks: Cls (Classification); Seg (Semantic Segmentation); Det (Object Detection); Com (Semantic Scene Completion); Rec (Surface Reconstruction); CM (Cross-Modal tasks); Pos (Pose estimation); Tra(Object Tracking) ",
"figure_data": "YearName#Samples#CategoriesTypesSuitable tasksHighlights2012KITTI (Geiger, Lenz and Urtasun, 2012)Over 200K objects8RGBCls/Det/CMComprehensive outdoor driving dataset2015ModelNet (Wu, Song, Khosla, Yu, Zhang, Tang and Xiao, 2015)12,311 models40CADCls/Seg/RecFrequently used in classification and few-shot2015 ShapeNet",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Commonly used deep networks for extracting point cloud features.",
"figure_data": "",
"figure_id": "tab_3",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Summary of reconstruction-based point cloud SSL methods. D & G stand for Generation and Discrimination.",
"figure_data": "YearMethodSub-categoriesContributions2021Point-BERT (Yu et al., 2021)Mask recoveryReconstructing missing point tokens with BERT-style transformer2022Point-MAE (Pang et al., 2022)Mask recoveryShifting masked tokens to decoder to avoid early leakage2022MaskSurf (Zhang et al., 2022b)Mask recoveryEstimating surfel position and per-surfel orientation simultaneously2022Voxel-MAE (Min et al., 2022)Mask recoveryPerforming additional binary voxel classification for complicated semantics awareness20193D jigsaw (Sauder and Sievers, 2019b)Spatial restorationRearranging randomly disorganized point clouds2019CloudContext (Sauder and Sievers, 2019a",
"figure_id": "tab_4",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Summary of contrast-based point cloud SSL methods.",
"figure_data": "YearMethodSub-categoriesContributions2020Info3D (Sanghi, 2020)Object contrastMaximizing mutual information between objects and their transformations2022AFSRL (Lu, Dai, Li and Su, 2022)Object contrastImposing data-level augmentation and feature enhancement simultaneously2019Contrasting and clustering (Zhang and Zhu, 2019)Object contrastSolving part contrast and object cluster tasks consecutively2021Hard negatives (Du, Gao, Hu and Li, 2021)Object contrast Leveraging self-similar point cloud patches; facilitating hierarchical context primitives capturing2020PointContrast (Xie, Gu, Guo, Qi, Guibas and Litany, 2020)Scene contrastObtaining dense features at point-level on complex scenes by point contrast2021Contrastive Scene Contexts (Hou, Graham, Nießner and Xie, 2021)Scene contrastIntroducing ShapeContext local descriptor and achieving data-efficiency2021 CoCoNets",
"figure_id": "tab_5",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Summary of alignment-based point cloud SSL methods.",
"figure_data": "",
"figure_id": "tab_6",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "Summary of motion-based point cloud SSL methods.",
"figure_data": "",
"figure_id": "tab_7",
"figure_label": "6",
"figure_type": "table"
},
{
"figure_caption": "Summary of few-shot protocol performance of representative SSL methods on ModelNet40",
"figure_data": "",
"figure_id": "tab_8",
"figure_label": "7",
"figure_type": "table"
},
{
"figure_caption": "cloud part segmentation is the mean Intersection over Union (mIoU), which computes the ratio of the intersection of the predicted and ground truth part labels to the union of the two, across all categories (mIoU C ) or all instances (mIoU I ). Table10summarizes the results of part segmentation on the ShapeNetPart dataset based on SSL pre-training models and supervised fine-tuning in terms of mIoU",
"figure_data": "5-way10-wayMethodBackbone10-shot 20-shot 10-shot 20-shotModelNet40Point-MAE (Pang et al., 2022)Transformer96.397.892.695.5Point-BERT (Yu et al., 2021)Transformer94.696.391.092.7OcCo (Wang et al., 2021b)PointNet89.792.489.389.7OcCo (Wang et al., 2021b)DGCNN90.692.582.986.5OcCo (Wang et al., 2021b)Transformer94.095.989.492.43D jigsaw (Sauder and Sievers, 2019b)PointNet66.569.256.966.53D jigsaw (Sauder and Sievers, 2019b)DGCNN34.342.226.029.9MaskSurf (Zhang et al., 2022b)Transformer96.598.093.095.3MaskPoint (Liu et al., 2022b)Transformer95.097.291.493.4ScanObjectNNPoint-MAE (Pang et al., 2022)Transformer63.977.053.661.6OcCo (Wang et al., 2021b)PointNet70.472.254.861.8OcCo (Wang et al., 2021b)DGCNN72.477.257.061.63D jigsaw (Sauder and Sievers, 2019b)PointNet58.667.653.648.13D jigsaw (Sauder and Sievers, 2019b)DGCNN65.272.245.648.2MaskSurf (Zhang et al., 2022b)Transformer65.377.453.863.2",
"figure_id": "tab_9",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Summary of linear classification protocol performance of representative SSL methods on ModelNet10/40(Wu et al., 2015). The results are reported in terms of OA (%).",
"figure_data": "MethodPretext typeBackboneModelNet10/40Point-MAE (Pang et al., 2022)Reconstruction Transformer-/ 91.41Orientation estimation (Poursaeed et al., 2020)ReconstructionPointNet-/ 88.6Orientation estimation (Poursaeed et al., 2020)ReconstructionDGCNN-/ 90.753D jigsaw (Sauder and Sievers, 2019b)ReconstructionPointNet91.61 / 87.313D jigsaw (Sauder and Sievers, 2019b)ReconstructionDGCNN94.52 / 90.64MaskSurf (Zhang et al., 2022b)Reconstruction Transformer-/ 92.26CloudContext (Sauder and Sievers, 2019a)ReconstructionDGCNN94.5 / 89.3UAE (Zhang et al., 2022a)ReconstructionDGCNN95.6 / 92.9Pose Disentanglement (Tsai et al., 2022)ReconstructionPointNet-/ 90.1Pose Disentanglement (Tsai et al., 2022)ReconstructionDGCNN-/ 92.0CP-Net (Xu et al., 2022)ReconstructionRSCNN-/ 91.9FoldingNet (Yang et al., 2018)ReconstructionGNN94.4 / 88.4Self-correction (Chen et al., 2021)ReconstructionPointNet93.3 / 89.9Self-correction (Chen et al., 2021)ReconstructionRSCNN95.0 / 92.4PC-GAN (Li et al., 2018)ReconstructionGAN-/ 87.5Info3D (Sanghi, 2020)ContrastPointNet-/ 89.8Info3D (Sanghi, 2020)ContrastDGCNN-/ 91.6AFSRL (Lu et al., 2022)ContrastGNN-/ 91.5Contrasting and clustering (Zhang and Zhu, 2019)ContrastDGCNN93.8 / 86.8Hard negatives (Du et al., 2021)ContrastDGCNN-/ 89.6OcCo (Wang et al., 2021b)AlignmentDGCNN-/ 89.2Cross-view (Jing et al., 2021)AlignmentGNN-/ 89.8Multi-view rendering (Tran et al., 2022)AlignmentPointNet-/ 89.7Multi-view rendering (Tran et al., 2022)AlignmentDGCNN-/ 91.7STRL (Huang et al., 2021)AlignmentPointNet-/ 88.3STRL (Huang et al., 2021)AlignmentDGCNN-/ 90.9PRNet (Wang and Solomon, 2019)MotionDGCNN-/ 85.2",
"figure_id": "tab_11",
"figure_label": "9",
"figure_type": "table"
},
{
"figure_caption": "Summary of performance of representative methods on part segmentation using ShapeNetPart(Armeni et al., 2016).Table 11Summary of performance of representative methods on semantic segmentation using S3DIS(Armeni et al., 2016).",
"figure_data": "MethodTypeBackbonemIoU CmIoU IPointNet (Qi et al., 2017a)83.3983.7Supervised-PointNet++ (Qi et al., 2017b) DGCNN (Wang et al., 2019)81.85 82.3385.1 85.2Transformer (Vaswani et al., 2017)83.4285.1Point-MAE (Pang et al., 2022)ReconstructionTransformer84.1986.1Point-BERT (Yu et al., 2021)ReconstructionTransformer84.1185.63D jigsaw (Sauder and Sievers, 2019b)ReconstructionPointNet-82.23D jigsaw (Sauder and Sievers, 2019b)ReconstructionDGCNN-85.3MaskSurf (Zhang et al., 2022b)ReconstructionTransformer84.3686.1CloudContext (Sauder and Sievers, 2019a) ReconstructionDGCNN-81.5UAE (Zhang et al., 2022a)ReconstructionDGCNN-85.6Pose Disentanglement (Tsai et al., 2022)ReconstructionPointNet-/83.8Pose Disentanglement (Tsai et al., 2022)ReconstructionDGCNN-/ 85.1MD (Sun et al., 2022)ReconstructionDGCNN-85.5Self-correction (Chen et al., 2021)ReconstructionPointNet-84.1Self-correction (Chen et al., 2021)ReconstructionRSCNN-85.2MaskPoint (Liu et al., 2022b)ReconstructionTransformer84.486.0AFSRL (Lu et al., 2022)ContrastGNN-/ 85.7Hard negatives (Du et al., 2021)ContrastDGCNN-/ 82.3PointContrast (Xie et al., 2020)ContrastU-Net-85.1OcCo (Wang et al., 2021b)AlignmentPointNet-83.4OcCo (Wang et al., 2021b)AlignmentDGCNN-85.0Cross-view (Jing et al., 2021)AlignmentDGCNN79.183.7Multi-view rendering (Tran et al., 2022)AlignmentPointNet-83.3Multi-view rendering (Tran et al., 2022)AlignmentDGCNN-84.7PRNet (Wang and Solomon, 2019)MotionDGCNN78.8 / 82.5MethodTypeBackboneOAmAcc mIoUPointNet (Qi et al., 2017a)78.649.047.7Supervised-DGCNN (Wang et al., 2019)84.1-56.1Transformer (Vaswani et al., 2017)86.868.660.0Area 5 testPoint-MAE (Pang et al., 2022)ReconstructionTransformer87.469.461.0OcCo (Wang et al., 2021b)AlignmentPointNet-83.644.5OcCo (Wang et al., 2021b)AlignmentDGCNN-87.049.53D jigsaw (Sauder and Sievers, 2019b)ReconstructionPointNet-82.543.63D jigsaw (Sauder and Sievers, 2019b)ReconstructionDGCNN-86.848.2MaskSurf (Zhang et al., 2022b)ReconstructionTransformer88.369.961.6PointContrast (Xie et al., 2020)ContrastSR-UNet-77.070.9Contrastive Scene Contexts (Hou et al., 2021)ContrastDGCNN--73.8DepthConstrast (Zhang et al., 2021)ContrastPointNet++-72.164.8Multi-view rendering",
"figure_id": "tab_12",
"figure_label": "10",
"figure_type": "table"
},
{
"figure_caption": ". The pre-training input only contains the point cloud geometry.",
"figure_data": "SUN RGB-DScanNetMethodTypeBackboneAP 25 AP 50 AP 25 AP 50Point-BERT (Yu et al., 2021)Reconstruction Transformer--61.038.3MaskPoint (Liu et al., 2022b)Reconstruction Transformer--64.242.0PointContrast (Xie et al., 2020)ContrastSR-UNet57.534.859.238.0PointContrast (Xie et al., 2020)ContrastVoteNet59.238.057.534.8DepthConst (Zhang et al., 2021)ContrastPointNet++--61.3-DepthConst (Zhang et al., 2021)ContrastVoteNet64.042.961.635.5DepthConst (Zhang et al., 2021)ContrastH3DNet69.050.063.543.4Multi-view rendering (Tran et al., 2022)AlignmentDGCNN58.135.160.339.2STRL (Huang et al., 2021)AlignmentVoteNet58.2---",
"figure_id": "tab_13",
"figure_label": "",
"figure_type": "table"
}
] | Changyu Zeng; Wei Wang; Anh Nguyen; Yutao Yue | [
{
"authors": "I Achituve; H Maron; G Chechik",
"journal": "",
"ref_id": "b0",
"title": "Self-supervised learning for domain adaptation on point clouds",
"year": "2021"
},
{
"authors": "M Afham; I Dissanayake; D Dissanayake; A Dharmasiri; K Thilakarathna; R Rodrigo",
"journal": "",
"ref_id": "b1",
"title": "Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding",
"year": "2022"
},
{
"authors": "P Agrawal; J Carreira; J Malik",
"journal": "",
"ref_id": "b2",
"title": "Learning to see by moving",
"year": "2015"
},
{
"authors": "R Arandjelovic; A Zisserman",
"journal": "",
"ref_id": "b3",
"title": "Look, listen and learn",
"year": "2017"
},
{
"authors": "I Armeni; O Sener; A R Zamir; H Jiang; I Brilakis; M Fischer; S Savarese",
"journal": "",
"ref_id": "b4",
"title": "3d semantic parsing of large-scale indoor spaces",
"year": "2016"
},
{
"authors": "P J Besl; N D Mckay",
"journal": "Spie",
"ref_id": "b5",
"title": "Method for registration of 3-d shapes, in: Sensor fusion IV: control paradigms and data structures",
"year": "1992"
},
{
"authors": "H Caesar; V Bankiti; A H Lang; S Vora; V E Liong; Q Xu; A Krishnan; Y Pan; G Baldan; O Beijbom",
"journal": "",
"ref_id": "b6",
"title": "nuscenes: A multimodal dataset for autonomous driving",
"year": "2020"
},
{
"authors": "M Caron; I Misra; J Mairal; P Goyal; P Bojanowski; A Joulin",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b7",
"title": "Unsupervised learning of visual features by contrasting cluster assignments",
"year": "2020"
},
{
"authors": "A X Chang; T Funkhouser; L Guibas; P Hanrahan; Q Huang; Z Li; S Savarese; M Savva; S Song; H Su",
"journal": "",
"ref_id": "b8",
"title": "Shapenet: An information-rich 3d model repository",
"year": "2015"
},
{
"authors": "M Chen; Q Hu; T Hugues; A Feng; Y Hou; K Mccullough; L Soibelman",
"journal": "",
"ref_id": "b9",
"title": "Stpls3d: A largescale synthetic and real aerial photogrammetry 3d point cloud dataset",
"year": "2022"
},
{
"authors": "T Chen; S Kornblith; M Norouzi; G Hinton",
"journal": "PMLR",
"ref_id": "b10",
"title": "A simple framework for contrastive learning of visual representations",
"year": "2020"
},
{
"authors": "Y Chen; J Liu; B Ni; H Wang; J Yang; N Liu; T Li; Q Tian",
"journal": "",
"ref_id": "b11",
"title": "Shape self-correction for unsupervised point cloud understanding",
"year": "2021"
},
{
"authors": "I Croitoru; S V Bogolin; M Leordeanu",
"journal": "",
"ref_id": "b12",
"title": "Unsupervised learning from video to detect foreground objects in single images",
"year": "2017"
},
{
"authors": "G Csurka",
"journal": "",
"ref_id": "b13",
"title": "Domain adaptation for visual applications: A comprehensive survey",
"year": "2017"
},
{
"authors": "A Dai; A X Chang; M Savva; M Halber; T Funkhouser; M Nießner",
"journal": "",
"ref_id": "b14",
"title": "Scannet: Richly-annotated 3d reconstructions of indoor scenes",
"year": "2017"
},
{
"authors": "J Devlin; M W Chang; K Lee; K Toutanova",
"journal": "",
"ref_id": "b15",
"title": "Bert: Pre-training of deep bidirectional transformers for language understanding",
"year": "2018"
},
{
"authors": "C Doersch; A Gupta; A A Efros",
"journal": "",
"ref_id": "b16",
"title": "Unsupervised visual representation learning by context prediction",
"year": "2015"
},
{
"authors": "B Du; X Gao; W Hu; X Li",
"journal": "",
"ref_id": "b17",
"title": "Self-contrastive learning with hard negative sampling for self-supervised point cloud learning",
"year": "2021"
},
{
"authors": "N El-Sheimy; Y Li",
"journal": "Satellite Navigation",
"ref_id": "b18",
"title": "Indoor navigation: State of the art and future trends",
"year": "2021"
},
{
"authors": "M Everingham; L Van Gool; C K I Williams; J Winn; A Zisserman",
"journal": "",
"ref_id": "b19",
"title": "The PASCAL Visual Object Classes Challenge 2012",
"year": ""
},
{
"authors": "A Faktor; M Irani",
"journal": "",
"ref_id": "b20",
"title": "Video segmentation by non-local consensus voting",
"year": "2014"
},
{
"authors": "L Floridi; M Chiriatti",
"journal": "Minds and Machines",
"ref_id": "b21",
"title": "Gpt-3: Its nature, scope, limits, and consequences",
"year": "2020"
},
{
"authors": "D A S Fraser",
"journal": "",
"ref_id": "b22",
"title": "Probability and statistics: Theory and applications",
"year": "1976"
},
{
"authors": "V Garcia; J Bruna",
"journal": "",
"ref_id": "b23",
"title": "Few-shot learning with graph neural networks",
"year": "2017"
},
{
"authors": "A Geiger; P Lenz; R Urtasun",
"journal": "IEEE",
"ref_id": "b24",
"title": "Are we ready for autonomous driving? the kitti vision benchmark suite",
"year": "2012"
},
{
"authors": "S Gidaris; P Singh; N Komodakis",
"journal": "",
"ref_id": "b25",
"title": "Unsupervised representation learning by predicting image rotations",
"year": "2018"
},
{
"authors": "I Goodfellow; J Pouget-Abadie; M Mirza; B Xu; D Warde-Farley; S Ozair; A Courville; Y Bengio",
"journal": "",
"ref_id": "b26",
"title": "Advances in neural information processing systems 27",
"year": "2014"
},
{
"authors": "J B Grill; F Strub; F Altché; C Tallec; P Richemond; E Buchatskaya; C Doersch; B Avila Pires; Z Guo; M Gheshlaghi Azar",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b27",
"title": "Bootstrap your own latent-a new approach to self-supervised learning",
"year": "2020"
},
{
"authors": "M H Guo; J X Cai; Z N Liu; T J Mu; R R Martin; S M Hu",
"journal": "Computational Visual Media",
"ref_id": "b28",
"title": "Pct: Point cloud transformer",
"year": "2021"
},
{
"authors": "T Hackel; N Savinov; L Ladicky; J D Wegner; K Schindler; M Pollefeys",
"journal": "",
"ref_id": "b29",
"title": "Semantic3d. net: A new large-scale point cloud classification benchmark",
"year": "2017"
},
{
"authors": "K He; X Chen; S Xie; Y Li; P Dollár; R Girshick",
"journal": "",
"ref_id": "b30",
"title": "Masked autoencoders are scalable vision learners",
"year": "2021"
},
{
"authors": "K He; H Fan; Y Wu; S Xie; R Girshick",
"journal": "",
"ref_id": "b31",
"title": "Momentum contrast for unsupervised visual representation learning",
"year": "2020"
},
{
"authors": "G Hess; J Jaxing; E Svensson; D Hagerman; C Petersson; L Svensson",
"journal": "",
"ref_id": "b32",
"title": "Masked autoencoders for selfsupervised learning on automotive point clouds",
"year": "2022"
},
{
"authors": "J Hou; B Graham; M Nießner; S Xie",
"journal": "",
"ref_id": "b33",
"title": "Exploring data-efficient 3d scene understanding with contrastive scene contexts",
"year": "2021"
},
{
"authors": "Q Hu; B Yang; S Khalid; W Xiao; N Trigoni; A Markham",
"journal": "",
"ref_id": "b34",
"title": "Towards semantic segmentation of urban-scale 3d point clouds: A dataset, benchmarks and challenges",
"year": "2021"
},
{
"authors": "B S Hua; Q H Pham; D T Nguyen; M K Tran; L F Yu; S K Yeung",
"journal": "",
"ref_id": "b35",
"title": "Scenenn: A scene meshes dataset with annotations",
"year": "2016"
},
{
"authors": "S Huang; Y Xie; S C Zhu; Y Zhu",
"journal": "",
"ref_id": "b36",
"title": "Spatiotemporal self-supervised representation learning for 3d point clouds",
"year": "2021"
},
{
"authors": "D Jayaraman; K Grauman",
"journal": "",
"ref_id": "b37",
"title": "Learning image representations tied to ego-motion",
"year": "2015"
},
{
"authors": "H Jiang; G Larsson; M M G Shakhnarovich; E Learned-Miller",
"journal": "",
"ref_id": "b38",
"title": "Self-supervised relative depth learning for urban scene understanding",
"year": "2018"
},
{
"authors": "L Jing; Y Tian",
"journal": "IEEE transactions on pattern analysis and machine intelligence",
"ref_id": "b39",
"title": "Self-supervised visual feature learning with deep neural networks: A survey",
"year": "2020"
},
{
"authors": "L Jing; L Zhang; Y Tian",
"journal": "",
"ref_id": "b40",
"title": "Self-supervised feature learning by cross-modality and cross-view correspondences",
"year": "2021"
},
{
"authors": "S Koch; A Matveev; Z Jiang; F Williams; A Artemov; E Burnaev; M Alexa; D Zorin; D Panozzo",
"journal": "",
"ref_id": "b41",
"title": "Abc: A big cad model dataset for geometric deep learning",
"year": "2019"
},
{
"authors": "S Lal; M Prabhudesai; I Mediratta; A W Harley; K Fragkiadaki",
"journal": "",
"ref_id": "b42",
"title": "Coconets: Continuous contrastive 3d scene representations",
"year": "2021"
},
{
"authors": "M Lewis; Y Liu; N Goyal; M Ghazvininejad; A Mohamed; O Levy; V Stoyanov; L Zettlemoyer",
"journal": "",
"ref_id": "b43",
"title": "Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension",
"year": "2019"
},
{
"authors": "C L Li; M Zaheer; Y Zhang; B Poczos; R Salakhutdinov",
"journal": "",
"ref_id": "b44",
"title": "Point cloud gan",
"year": "2018"
},
{
"authors": "R Li; X Li; C W Fu; D Cohen-Or; P A Heng",
"journal": "",
"ref_id": "b45",
"title": "Pugan: a point cloud upsampling adversarial network",
"year": "2019"
},
{
"authors": "R Li; G Lin; L Xie",
"journal": "",
"ref_id": "b46",
"title": "Self-point-flow: Selfsupervised scene flow estimation from point clouds with optimal transport and random walk",
"year": "2021"
},
{
"authors": "Y Li; L Ma; Z Zhong; F Liu; D Cao; J Li; M A Chapman",
"journal": "IEEE Transactions on Neural Networks",
"ref_id": "b47",
"title": "Deep learning for lidar point clouds in autonomous driving: A review",
"year": "2020"
},
{
"authors": "Y Li; A W Yu; T Meng; B Caine; J Ngiam; D Peng; J Shen; Y Lu; D Zhou; Q V Le",
"journal": "",
"ref_id": "b48",
"title": "Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection",
"year": "2022"
},
{
"authors": "D Liu; C Chen; C Xu; R Qiu; L Chu",
"journal": "",
"ref_id": "b49",
"title": "Selfsupervised point cloud registration with deep versatile descriptors",
"year": "2022"
},
{
"authors": "H Liu; M Cai; Y J Lee",
"journal": "",
"ref_id": "b50",
"title": "Masked discrimination for self-supervised learning on point clouds",
"year": "2022"
},
{
"authors": "H Liu; H Yuan; J Hou; R Hamzaoui; W Gao",
"journal": "IEEE Transactions on Image Processing",
"ref_id": "b51",
"title": "Pufa-gan: A frequency-aware generative adversarial network for 3d point cloud upsampling",
"year": "2022"
},
{
"authors": "X Liu; X Liu; Y S Liu; Z Han",
"journal": "IEEE Transactions on Image Processing",
"ref_id": "b52",
"title": "Spu-net: Selfsupervised point cloud upsampling by coarse-to-fine reconstruction with self-projection optimization",
"year": "2022"
},
{
"authors": "Y Liu; L Yi; S Zhang; Q Fan; T Funkhouser; H Dong",
"journal": "",
"ref_id": "b53",
"title": "P4Contrast: Contrastive Learning with Pairs of Point-Pixel Pairs for RGB-D Scene Understanding",
"year": "2020"
},
{
"authors": "Z Lu; Y Dai; W Li; Z Su",
"journal": "",
"ref_id": "b54",
"title": "Joint data and feature augmentation for self-supervised representation learning on point clouds",
"year": "2022"
},
{
"authors": "D Matti; H K Ekenel; J P Thiran",
"journal": "",
"ref_id": "b55",
"title": "Combining lidar space clustering and convolutional neural networks for pedestrian detection",
"year": "2017"
},
{
"authors": "B Mersch; X Chen; J Behley; C Stachniss",
"journal": "PMLR",
"ref_id": "b56",
"title": "Selfsupervised point cloud prediction using 3d spatiotemporal convolutional networks",
"year": "2022"
},
{
"authors": "T Mikolov; K Chen; G Corrado; J Dean",
"journal": "",
"ref_id": "b57",
"title": "Efficient estimation of word representations in vector space",
"year": "2013"
},
{
"authors": "G A Miller",
"journal": "Communications of the ACM",
"ref_id": "b58",
"title": "Wordnet: a lexical database for english",
"year": "1995"
},
{
"authors": "C Min; D Zhao; L Xiao; Y Nie; B Dai",
"journal": "",
"ref_id": "b59",
"title": "Voxelmae: Masked autoencoders for pre-training large-scale point clouds",
"year": "2022"
},
{
"authors": "H Mittal; B Okorn; D Held",
"journal": "",
"ref_id": "b60",
"title": "Just go with the flow: Self-supervised scene flow estimation",
"year": "2020"
},
{
"authors": "K Mo; S Zhu; A X Chang; L Yi; S Tripathi; L J Guibas; H Su",
"journal": "",
"ref_id": "b61",
"title": "Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding",
"year": "2019"
},
{
"authors": "M Noroozi; P Favaro",
"journal": "Springer",
"ref_id": "b62",
"title": "Unsupervised learning of visual representations by solving jigsaw puzzles",
"year": "2016"
},
{
"authors": "A V D Oord; Y Li; O Vinyals",
"journal": "",
"ref_id": "b63",
"title": "Representation learning with contrastive predictive coding",
"year": "2018"
},
{
"authors": "Y Pang; W Wang; F E Tay; W Liu; Y Tian; L Yuan",
"journal": "",
"ref_id": "b64",
"title": "Masked autoencoders for point cloud selfsupervised learning",
"year": "2022"
},
{
"authors": "D Pathak; P Krähenbühl; J Donahue; T Darrell; A A Efros",
"journal": "",
"ref_id": "b65",
"title": "Context encoders: Feature learning by inpainting. computer vision and pattern recognition",
"year": "2016"
},
{
"authors": "O Poursaeed; T Jiang; H Qiao; N Xu; V G Kim",
"journal": "IEEE",
"ref_id": "b66",
"title": "Self-supervised learning of point clouds via orientation estimation",
"year": "2020"
},
{
"authors": "C R Qi; H Su; K Mo; L J Guibas",
"journal": "",
"ref_id": "b67",
"title": "Pointnet: Deep learning on point sets for 3d classification and segmentation",
"year": "2017"
},
{
"authors": "C R Qi; L Yi; H Su; L J Guibas",
"journal": "",
"ref_id": "b68",
"title": "Pointnet++: Deep hierarchical feature learning on point sets in a metric space",
"year": "2017"
},
{
"authors": "B Romera-Paredes; P Torr",
"journal": "PMLR",
"ref_id": "b69",
"title": "An embarrassingly simple approach to zero-shot learning",
"year": "2015"
},
{
"authors": "A Sanghi",
"journal": "Springer",
"ref_id": "b70",
"title": "Info3d: Representation learning on 3d objects using mutual information maximization and contrastive learning",
"year": "2020"
},
{
"authors": "M B Sariyildiz; Y Kalantidis; K Alahari; D Larlus",
"journal": "",
"ref_id": "b71",
"title": "Improving the generalization of supervised models",
"year": "2022"
},
{
"authors": "M Sarmad; H J Lee; Y M Kim",
"journal": "",
"ref_id": "b72",
"title": "Rl-gan-net: A reinforcement learning agent controlled gan network for real-time point cloud shape completion",
"year": "2019"
},
{
"authors": "J Sauder; B Sievers",
"journal": "",
"ref_id": "b73",
"title": "Context prediction for unsupervised deep learning on point clouds",
"year": "2019"
},
{
"authors": "J Sauder; B Sievers",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b74",
"title": "Self-supervised deep learning on point clouds by reconstructing space",
"year": "2019"
},
{
"authors": "N Sayed; B Brattoli; B Ommer",
"journal": "Springer",
"ref_id": "b75",
"title": "Cross and learn: Cross-modal self-supervision",
"year": "2018"
},
{
"authors": "Y Shi; X Cao; B Zhou",
"journal": "Computer Graphics Forum",
"ref_id": "b76",
"title": "Self-supervised learning of part mobility from point cloud sequence",
"year": "2021"
},
{
"authors": "D W Shu; S W Park; J Kwon",
"journal": "",
"ref_id": "b77",
"title": "3d point cloud generative adversarial network based on tree structured graph convolutions",
"year": "2019"
},
{
"authors": "S Song; S P Lichtenberg; J Xiao",
"journal": "",
"ref_id": "b78",
"title": "Sun rgb-d: A rgbd scene understanding benchmark suite",
"year": "2015"
},
{
"authors": "O Stretcu; M Leordeanu",
"journal": "",
"ref_id": "b79",
"title": "Multiple frames matching for object discovery in video",
"year": "2015"
},
{
"authors": "C Sun; Z Zheng; X Wang; M Xu; Y Yang",
"journal": "IEEE Transactions on Multimedia",
"ref_id": "b80",
"title": "Self-supervised point cloud representation learning via separating mixed shapes",
"year": "2022"
},
{
"authors": "P Sun; H Kretzschmar; X Dotiwalla; A Chouard; V Patnaik; P Tsui; J Guo; Y Zhou; Y Chai; B Caine",
"journal": "",
"ref_id": "b81",
"title": "Scalability in perception for autonomous driving: Waymo open dataset",
"year": "2020"
},
{
"authors": "X Sun; J Wu; X Zhang; Z Zhang; C Zhang; T Xue; J B Tenenbaum; W T Freeman",
"journal": "",
"ref_id": "b82",
"title": "Pix3d: Dataset and methods for single-image 3d shape modeling",
"year": "2018"
},
{
"authors": "S A Taghanaki; J Luo; R Zhang; Y Wang; P K Jayaraman; K M Jatavallabhula",
"journal": "",
"ref_id": "b83",
"title": "Robustpointset: A dataset for benchmarking robustness of point cloud classifiers",
"year": "2020"
},
{
"authors": "B Tran; B S Hua; A T Tran; M Hoai",
"journal": "",
"ref_id": "b84",
"title": "Selfsupervised learning with multi-view rendering for 3d point cloud analysis",
"year": "2022"
},
{
"authors": "M S Tsai; P Z Chiang; Y H Tsai; W C Chiu",
"journal": "IEEE",
"ref_id": "b85",
"title": "Selfsupervised feature learning from partial point clouds via pose disentanglement",
"year": "2022"
},
{
"authors": "M A Uy; Q H Pham; B S Hua; T Nguyen; S K Yeung",
"journal": "",
"ref_id": "b86",
"title": "Revisiting point cloud classification: A new benchmark dataset and classification model on realworld data",
"year": "2019"
},
{
"authors": "A Vaswani; N Shazeer; N Parmar; J Uszkoreit; L Jones; A N Gomez; Ł Kaiser; I Polosukhin",
"journal": "Advances in neural information processing systems",
"ref_id": "b87",
"title": "Attention is all you need",
"year": "2017"
},
{
"authors": "P Velickovic; W Fedus; W L Hamilton; P Liò; Y Bengio; R D Hjelm",
"journal": "ICLR (Poster)",
"ref_id": "b88",
"title": "Deep graph infomax",
"year": "2019"
},
{
"authors": "S Vora; A H Lang; B Helou; O Beijbom",
"journal": "",
"ref_id": "b89",
"title": "Pointpainting: Sequential fusion for 3d object detection",
"year": "2020"
},
{
"authors": "C Wang; C Ma; M Zhu; X Yang",
"journal": "",
"ref_id": "b90",
"title": "Pointaugmenting: Cross-modal augmentation for 3d object detection",
"year": "2021"
},
{
"authors": "H Wang; Q Liu; X Yue; J Lasenby; M J Kusner",
"journal": "",
"ref_id": "b91",
"title": "Unsupervised point cloud pre-training via occlusion completion",
"year": "2021"
},
{
"authors": "H Wang; L Yang; X Rong; J Feng; Y Tian",
"journal": "",
"ref_id": "b92",
"title": "Self-supervised 4d spatio-temporal feature learning via order prediction of sequential point cloud clips",
"year": "2021"
},
{
"authors": "W Wang; X Lin; F Feng; X He; M Lin; T S Chua",
"journal": "",
"ref_id": "b93",
"title": "Causal representation learning for outof-distribution recommendation",
"year": "2022"
},
{
"authors": "Y Wang; J M Solomon",
"journal": "Advances in neural information processing systems",
"ref_id": "b94",
"title": "Prnet: Self-supervised learning for partial-to-partial registration",
"year": "2019"
},
{
"authors": "Y Wang; Y Sun; Z Liu; S E Sarma; M M Bronstein; J M Solomon",
"journal": "Acm Transactions On Graphics (tog)",
"ref_id": "b95",
"title": "Dynamic graph cnn for learning on point clouds",
"year": "2019"
},
{
"authors": "B Wu; X Zhou; S Zhao; X Yue; K Keutzer",
"journal": "IEEE",
"ref_id": "b96",
"title": "Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud",
"year": "2019"
},
{
"authors": "W Wu; Z Y Wang; Z Li; W Liu; L Fuxin",
"journal": "Springer",
"ref_id": "b97",
"title": "Pointpwc-net: Cost volume on point clouds for (self-) supervised scene flow estimation",
"year": "2020"
},
{
"authors": "Z Wu; S Song; A Khosla; F Yu; L Zhang; X Tang; J Xiao",
"journal": "",
"ref_id": "b98",
"title": "3d shapenets: A deep representation for volumetric shapes",
"year": "2015"
},
{
"authors": "Z Wu; Y Xiong; S X Yu; D Lin",
"journal": "",
"ref_id": "b99",
"title": "Unsupervised feature learning via non-parametric instance discrimination",
"year": "2018"
},
{
"authors": "Y Xiang; W Kim; W Chen; J Ji; C Choy; H Su; R Mottaghi; L Guibas; S Savarese",
"journal": "Springer",
"ref_id": "b100",
"title": "Objectnet3d: A large scale database for 3d object recognition",
"year": "2016"
},
{
"authors": "A Xiao; J Huang; D Guan; S Lu",
"journal": "",
"ref_id": "b101",
"title": "Unsupervised representation learning for point clouds: A survey",
"year": "2022"
},
{
"authors": "S Xie; J Gu; D Guo; C R Qi; L Guibas; O Litany",
"journal": "Springer",
"ref_id": "b102",
"title": "Pointcontrast: Unsupervised pre-training for 3d point cloud understanding",
"year": "2020"
},
{
"authors": "S Xie; S Liu; Z Chen; Z Tu",
"journal": "",
"ref_id": "b103",
"title": "Attentional shapecontextnet for point cloud recognition",
"year": "2018"
},
{
"authors": "D Xu; J Xiao; Z Zhao; J Shao; D Xie; Y Zhuang",
"journal": "",
"ref_id": "b104",
"title": "Self-supervised spatiotemporal learning via video clip order prediction",
"year": "2019"
},
{
"authors": "M Xu; Z Zhou; H Xu; Y Wang; Y Qiao",
"journal": "",
"ref_id": "b105",
"title": "Cp-net: Contour-perturbed reconstruction network for self-supervised point cloud learning",
"year": "2022"
},
{
"authors": "J Yang; J M Alvarez; M Liu",
"journal": "",
"ref_id": "b106",
"title": "Self-supervised learning of depth inference for multi-view stereo",
"year": "2021"
},
{
"authors": "Y Yang; C Feng; Y Shen; D Tian",
"journal": "",
"ref_id": "b107",
"title": "Foldingnet: Point cloud auto-encoder via deep grid deformation",
"year": "2018"
},
{
"authors": "Z Yang; S Liu; H Hu; L Wang; S Lin",
"journal": "",
"ref_id": "b108",
"title": "Reppoints: Point set representation for object detection",
"year": "2019"
},
{
"authors": "X Yu; L Tang; Y Rao; T Huang; J Zhou; J Lu",
"journal": "",
"ref_id": "b109",
"title": "Point-bert: Pre-training 3d point cloud transformers with masked point modeling",
"year": "2021"
},
{
"authors": "A R Zamir; A Sax; W Shen; L J Guibas; J Malik; S Savarese",
"journal": "",
"ref_id": "b110",
"title": "Taskonomy: Disentangling task transfer learning",
"year": "2018"
},
{
"authors": "C Zhang; J Shi; X Deng; Z Wu",
"journal": "",
"ref_id": "b111",
"title": "Upsampling autoencoder for self-supervised point cloud learning",
"year": "2022"
},
{
"authors": "L Zhang; Z Zhu",
"journal": "IEEE",
"ref_id": "b112",
"title": "Unsupervised feature learning for point cloud understanding by contrasting and clustering using graph convolutional neural networks",
"year": "2019"
},
{
"authors": "Y Zhang; J Lin; C He; Y Chen; K Jia; L Zhang",
"journal": "",
"ref_id": "b113",
"title": "Masked surfel prediction for self-supervised point cloud learning",
"year": "2022"
},
{
"authors": "Y Zhang; J Lin; R Li; K Jia; L Zhang",
"journal": "",
"ref_id": "b114",
"title": "Pointdae: Denoising autoencoders for self-supervised point cloud learning",
"year": "2022"
},
{
"authors": "Z Zhang; R Girdhar; A Joulin; I Misra",
"journal": "",
"ref_id": "b115",
"title": "Selfsupervised pretraining of 3d features on any pointcloud",
"year": "2021"
},
{
"authors": "W Zhao; X Liu; Z Zhong; J Jiang; W Gao; G Li; X Ji",
"journal": "",
"ref_id": "b116",
"title": "Self-supervised arbitrary-scale point clouds upsampling via implicit neural representation",
"year": "2022"
},
{
"authors": "X Zhao; S Yang; T Huang; J Chen; T Ma; M Li; Y Liu",
"journal": "Springer",
"ref_id": "b117",
"title": "Superline3d: Self-supervised line segmentation and description for lidar point cloud",
"year": "2022-10-23"
},
{
"authors": "Y Zhao; L Hui; J Xie",
"journal": "",
"ref_id": "b118",
"title": "Sspu-net: Self-supervised point cloud upsampling via differentiable rendering",
"year": "2021"
},
{
"authors": "Y Zhou; O Tuzel",
"journal": "",
"ref_id": "b119",
"title": "Voxelnet: End-to-end learning for point cloud based 3d object detection. computer vision and pattern recognition",
"year": "2018"
}
] | [
{
"formula_coordinates": [
5,
331.51,
462.59,
212.46,
11.34
],
"formula_id": "formula_0",
"formula_text": "𝐴 𝑐→𝑡 ∶= 𝔼 𝑥∈𝑋 𝑡 (𝑅 𝑐 (𝐸 𝑐 (𝑥)), 𝑓 𝑡 (𝑥)) (1)"
},
{
"formula_coordinates": [
6,
316.41,
614.23,
227.56,
15.44
],
"formula_id": "formula_1",
"formula_text": "𝑚𝑖𝑛 𝐺 𝑚𝑎𝑥 𝐷 𝐸 𝑥∈𝑋 [𝑙𝑜𝑔(𝐷(𝑥))]+𝐸 𝑧∈𝑍 [𝑙𝑜𝑔(1-𝐷(𝐺(𝑧)))] (2)"
},
{
"formula_coordinates": [
7,
52.25,
639.64,
236.42,
39.14
],
"formula_id": "formula_2",
"formula_text": "𝑑 𝐶𝐷 (𝑋, 𝑌 ) = 1 |𝑋| ∑ 𝑥∈𝑋 min 𝑦∈𝑌 ||𝑥-𝑦|| 2 + 1 |𝑌 | ∑ 𝑦∈𝑌 min 𝑥∈𝑋 ||𝑥-𝑦|| 2(3)"
},
{
"formula_coordinates": [
7,
331.51,
204.6,
212.46,
29.05
],
"formula_id": "formula_3",
"formula_text": "𝐿 𝑞 = -log exp(𝑞 ⋅ 𝑘 + ∕𝜏) ∑ 𝐾 𝑖=0 exp(𝑞 ⋅ 𝑘 𝑖 ∕𝜏)(4)"
},
{
"formula_coordinates": [
15,
280.93,
189.14,
7.74,
9.96
],
"formula_id": "formula_4",
"formula_text": ")5"
}
] |
2023-05-21 | [
{
"figure_ref": [],
"heading": "INTRODUCTION",
"publication_ref": [
"b0",
"b1",
"b2",
"b3",
"b4",
"b5",
"b6",
"b7",
"b8",
"b9",
"b10",
"b11",
"b0",
"b1",
"b2",
"b12",
"b1",
"b2"
],
"table_ref": [],
"text": "Fine-tuning PLMs has become the de-facto paradigm in natural language processing [1], due to the amazing performance gains on a wide range of natural language processing tasks [2,3,4,5,6]. Despite SOTA performances, BERT [7] and its variants [8,9,10,11] still face significant application challenges: cumbersome computation and overthinking problems due to huge parameters and deep models. Early exiting attracts much attention as an input-adaptive method to speed up inference [12]. Early exiting installs a classifier at each transformer layer to evaluate the predictions and will exit when meeting the criterion. Three different early exiting strategies exist: (1) The confidence-based strategy evaluates the predictions based on specific confidence measurements. (2) The learned-based strategy learns a criterion for early exiting. (3) The patience-based strategy exits when consecutive classifiers make the exact predictions. Among them, the patience-based strategy PABEE [13] achieves SOTA results.\nWe raise two issues for the current SOTA strategy: (1) PABEE faces a limitation for application: it can not flexibly adjust the speedup ratio on a given task and fixed patience parameter, mainly caused by a strict cross-layer comparison strategy. Thus, we wonder whether we can combine PABEE with a softer cross-layer comparison strategy. (2) Current early exiting strategies mainly focus on SLC tasks, while the MLC tasks are neglected. So can they speed up MLC tasks?\nTherefore, we propose a Flexible-Patience-Based Early Exiting method (F-PABEE) to address the above issues. F-PABEE makes predictions at each classifier and will exit early if the current layer and the last few layers have similar (similarity score less than a threshold) predicted distributions. F-PABEE can be seen as a natural extension of PABEE and is more flexible since it can achieve better speed-accuracy tradeoffs by adjusting the similarity score thresholds and patience parameters. It can also extend to MLC tasks effortlessly.\nOur contributions are summarized as follows: (1) We propose F-PABEE, a novel and effective inference mechanism that is flexible in adjusting the speedup ratios of PLMs. (2) The results show that our method can accelerate inference effectively while maintaining good performances across different SLC and MLC tasks. (3) We are the first to investigate the early exiting of MLC tasks, and F-PABEE is suitable for this type of task."
},
{
"figure_ref": [],
"heading": "RELATED WORKS",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [
"fig_0"
],
"heading": "Static inference approach",
"publication_ref": [
"b13",
"b14",
"b15",
"b16",
"b17",
"b18",
"b7",
"b19",
"b20",
"b21"
],
"table_ref": [],
"text": "The static inference approach compresses the heavy model into a smaller one, including pruning, knowledge distillation, quantization, and weight sharing [14,15,16]. For example, HeadPrune [17] ranks the attention heads and prunes them to reduce inference latency. PKD [18] investigates the best practices of distilling knowledge from BERT into smaller-Fig. 1. Inference procedure of PABEE and F-PABEE, C i is the classifier, thre is threshold, P 0 is pre-defined patience. sized models. I-BERT [19] performs an end-to-end BERT inference without any floating point calculation. ALBERT [8] shares the cross-layer parameters. [20], [21] and [22] distills knowledge from the larger BERT teacher model for improving the performances of student networks which are learned with neural architecture search. Note that the static models are still in the form of deep neural networks with multiple stacked layers. The computational path is invariable for all examples in the inference process, which is not flexible."
},
{
"figure_ref": [],
"heading": "Dynamic early exiting",
"publication_ref": [
"b22",
"b23",
"b24",
"b25",
"b26",
"b27",
"b28",
"b29",
"b30",
"b12",
"b31",
"b32"
],
"table_ref": [],
"text": "Orthogonal to the static inference approach, early exiting dynamically adjusts hyper-parameters in response to changes in request traffic. It does not need to make significant changes to the original model structure or weight bits, nor does it need to train different teacher-student learning networks, which saves computing resources [23].\nThere are mainly three groups of dynamic early exiting strategies. The first type is confidence-based early exiting [24]. For example, BranchyNet [25], FastBERT [26], and DeeBERT [27] calculate the entropy of the prediction probability distribution to estimate the confidence of classifiers to enable dynamic early exiting. Shallow-deep [28] and Right-Tool [29] leverage the maximum of the predicted distribution as the exiting signal. The second type is the learned-based exiting, such as BERxiT [30] and CAT [31]. They learn a criterion for early exiting. The third type is patience-based early exiting, such as PABEE [13], which stops inference and exits early if the classifiers' predictions remain unchanged for pre-defined times. Among them, patience-based PABEE achieves SOTA performance. However, PABEE suffers from too strict cross-layer comparison, and the applications on MLC tasks are neglected. There are also literature focusing on improving the training of multi-exit BERT, like LeeBERT [32] and GAML-BERT [33]. F-PABEE is a more flexible extension to PABEE, which can simultaneously adjust the confidence thresholds and patience parameters to meet different requirements. In addition, it outperforms other existing early exiting strategies on both SLC and MLC tasks."
},
{
"figure_ref": [],
"heading": "Training of multi-exit backbones",
"publication_ref": [
"b31",
"b32"
],
"table_ref": [],
"text": "The literature on early exiting focuses more on the design of early exiting strategies, thus neglect the advances of multiexit backbones' training methods. LeeBERT [32] employs an adaptive learning method for training multiple exits. GAML-BERT [33] enhances the training of multi-exit backbones by a mutual learning approach. , where L i is the transformer block of the model, n is the number of transformer layers, C i is the inserted classifier layer, s is the cross-layer similarity score, thre is the similarity score threshold, P 0 is the pre-defined patience value in the model."
},
{
"figure_ref": [],
"heading": "FLEXIBLE PATIENCE-BASED EARLY EXITING",
"publication_ref": [],
"table_ref": [],
"text": "The input sentences are first embedded as the vector:\nh 0 = Embedding(x). (1\n)\nThe vector is then passed through transformer layers (L 1 ...L n ) to extract features and compute its hidden state h. After which, we use internal classifiers (C 1 ...C n ), which are connected to each transformer layer to predict probability p:\np i = C i (h i ) = C i (L i (h i-1 )).(2)\nWe denote the similarity score between the prediction results of layer i -1 and i as s(p i-1 , p i ) (s(p i-1 , p i ) ∈ R). The smaller the value of s(p i-1 , p i ), the prediction distributions are more consistent with each other. The premise of the model's early exit is that the comparison scores between successive layers are relatively small; The similarity threshold thre is a hyper-parameter. We use pat i to store the times that the cross-layer comparison scores are consecutively less than the threshold thre when the model reaches current layer i:\npat i = pat i-1 + 1 s(p i-1 , p i ) < thre 0 s(p i-1 , p i ) >= thre(3)\nIf s(p i-1 , p i ) is less than the similarity score threshold thre, then increase the patience counter by 1. Otherwise, reset the patience counter to 0. This process is repeated until pat reaches the pre-defined patience value P 0 . The model dynamically stops inference and exits early. However, if this condition is never met, the model uses the final classifier layer to make predictions. This way, the model can stop inference early without going through all layers."
},
{
"figure_ref": [],
"heading": "Similarity measures for SLC and MLC tasks",
"publication_ref": [
"b33"
],
"table_ref": [],
"text": "Under the framework of F-PABEE, we can adopt different similarity measures for predicted probability distributions. This work uses the knowledge distillation objectives as the similarity measures [34]. When the model reaches the current layer l, for SLC tasks, we compare a series of similarity measures of F-PABEE, denoted as: F-PABEE-KD: It adopts the knowledge distillation objective from probability mass distribution p l-1 to p l :\ns(p l-1 , p l ) = - k j=1 p l-1 j log(p l j );(4)\nF-PABEE-ReKD: It adopts the knowledge distillation objective in the reverse direction, from probability mass distribution p l to p l-1 :\ns(p l , p l-1 ) = - k j=1 p l j log(p l-1 j );(5)\nF-PABEE-SymKD: It adopts a symmetrical knowledge distillation objective:\nSymKD = s(p l-1 , p l ) + s(p l , p l-1 );(6)\nF-PABEE-JSKD: It adopts another symmetrical distillation objective, similar to Jenson-Shannon divergence:\nJSKD = 1 2 s(p l-1 , p l-1 + p l 2 ) + 1 2 s(p l , p l-1 + p l 2 )(7)\nIn addition, for MLC tasks, we transform them into multiple binary classification problems and sum the similarity scores of all categories, and the formulas are denoted as: F-PABEE-KD:\ns(p l-1 , p l ) = - k j=1 2 i=1 p l-1 ji log(p l ji );(8)\nF-PABEE-ReKD:\ns(p l , p l-1 ) = - k j=1 2 i=1 p l ji log(p l-1 ji );(9)\nThe formulations of F-PABEE-SymKD and F-PABEE-JSKD for MLC tasks are similar to those of SLC tasks."
},
{
"figure_ref": [],
"heading": "Training procedure",
"publication_ref": [],
"table_ref": [],
"text": "F-PABEE is trained on SLC and MLC tasks, while the activation and loss functions are different. For SLC tasks, we use the softmax activation function and cross-entropy function according to the tasks. In contrast, we use the sigmoid activation function and binary cross-entropy function for MLC tasks.\nAfter that, we optimize the model parameters by minimizing the overall loss function L, which is the weighted average of the loss terms from all classifiers: \nL = n j=1 jL j / n j=1 j(10)"
},
{
"figure_ref": [],
"heading": "EXPERIMENTS",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Tasks and Baselines",
"publication_ref": [
"b34",
"b35",
"b36",
"b37",
"b38",
"b39",
"b27",
"b29"
],
"table_ref": [],
"text": "We evaluate F-PABEE on GLUE benchmark [35] for SLC tasks and four datasets for MLC tasks: MixSNLPS [36], Mix-ATS [37], AAPD [38], and Stackoverflow [39]. we compare F-PABEE with three groups of baselines: (1) BERTbase; (2) Static exiting; (3) Dynamic exiting methods, including BrachcyNet [40], Shallow-Deep [28], BERxiT [30], and PABEE. Considering the flops of inferencing one with the whole BERT as the base, the speed-up ratio is defined as the average ratio of reduced flops due to early exiting."
},
{
"figure_ref": [],
"heading": "Experimental setting",
"publication_ref": [
"b40",
"b41"
],
"table_ref": [],
"text": "In training process, we perform grid search over the batch size of {16, 32, 128}, and learning rate of {1e-5, 2e-5, 3e-5, 5e-5} with an AdamW optimizer [41] . The batch size in the inference process is 1. We implement F-PABEE on the bases of HuggingFace Transformers [42]. All experiments are conducted on two Nvidia TITAN X 24GB GPUs."
},
{
"figure_ref": [],
"heading": "Overall comparisons",
"publication_ref": [],
"table_ref": [
"tab_0"
],
"text": "In Table 1, we compare F-PABEE with other early exiting strategies. We adjust the hyper-parameters of F-PABEE and other baselines to ensure similar speedups with PABEE. It Comparisons between different similarity measures We consider F-PABEE with different similarity measures, denoted as F-PABEE-KD, F-PABEE-ReKD, F-PABEE-SymKD, and F-PABEE-JSKD, and the results are presented in Fig 6 . F-PABEE-JSKD performs the best on both SLC and MLC tasks. We suppose that F-PABEE-JSKD is symmetric, and the similarity discrimination is more accurate than asym- "
},
{
"figure_ref": [],
"heading": "CONCLUSIONS",
"publication_ref": [],
"table_ref": [],
"text": "We proposed F-PABEE, a novel and efficient early exiting method that combines PABEE with a softer cross-layer comparison strategy. F-PABEE is more flexible than PABEE since it can achieve different speed-performance tradeoffs by adjusting the similarity score thresholds and patience parameters. In addition, we investigate the acceleration ability of F-PABEE with different backbones. Moreover, we compare the performances of F-PABEE with different similarity measures. Extensive experiments on SLC and MLC demonstrate that:\n(1) F-PABEE performs better than the previous SOTA adaptive early exiting strategies for both SLC and MLC tasks. As far as we know, we are the first to investigate the early exiting methods for MLC tasks. (2) F-PABEE performs well on different PLMs such as BERT and ALBERT. (3) Ablation studies show that F-PABEE-JSKD performs best for F-PABEE with different similarity measures."
}
] | Computational complexity and overthinking problems have become the bottlenecks for pre-training language models (PLMs) with millions or even trillions of parameters. A Flexible-Patience-Based Early Exiting method (F-PABEE) has been proposed to alleviate the problems mentioned above for single-label classification (SLC) and multi-label classification (MLC) tasks. F-PABEE makes predictions at the classifier and will exit early if predicted distributions of cross-layer are consecutively similar. It is more flexible than the previous state-of-the-art (SOTA) early exiting method PABEE because it can simultaneously adjust the similarity score thresholds and the patience parameters. Extensive experiments show that: (1) F-PABEE makes a better speedupaccuracy balance than existing early exiting strategies on both SLC and MLC tasks. (2) F-PABEE achieves faster inference and better performances on different PLMs such as BERT and ALBERT. (3) F-PABEE-JSKD performs best for F-PABEE with different similarity measures. | F-PABEE: FLEXIBLE-PATIENCE-BASED EARLY EXITING FOR SINGLE-LABEL AND MULTI-LABEL TEXT CLASSIFICATION TASKS | [
{
"figure_caption": "3. 1 .1Inference procedure for SLC and MLC tasks The inference procedure of F-PABEE is shown in Fig 1(b), which is an improved version of PABEE (Fig 1(a))",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 2 .2Fig. 2. Speed-accuracy curves of F-PABEE, PABEE and BERxiT on SLC tasks with BERT backbone.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "shows that F-PABEE balances speedup and performance better than baselines, especially for a large speedup ratio. Moreover, we draw the score-speedup curves for BERxiT, PABEE, and F-PABEE. It shows that F-PABEE outperforms the baseline models on both SLC (Fig 2) and MLC tasks(Fig 3). Furthermore, the distribution of executed layers (Fig 4) indicates that F-PABEE can choose the faster off-ramp and achieve a better trade-off between accuracy and efficiency by flexibly adjusting similarity score thresholds and patience parameters.4.4. Ablation studies Ablation on different PLMs F-PABEE is flexible and can work well with other pre-trained models, such as ALBERT. Therefore, to show the acceleration ability of F-PABEE with different backbones, we compare F-PABEE to other early exiting strategies with ALBERT base as the backbone. The results in Fig 5 show that F-PABEE outperforms other early exiting strategies under different backbones by large margins on both SLC and MLC tasks, indicating that F-PABEE can accelerate the inference process for numerous PLMs.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 3 .3Fig. 3. Speed-accuracy curves of F-PABEE, PABEE and BERxiT on MLC tasks with BERT backbone.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 4 .4Fig. 4. The distribution of executed layers of MRPC and MixSNIPS on average at different speeds (50%, 75%).",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 5 .5Fig. 5. Speed-accuracy curves of F-PABEE, PABEE and BERxiT on SLC and MLC tasks with ALBERT backbone.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 6 .6Fig. 6. Speed-accuracy curves of different similarity measures on SLC and MLC tasks with BERT backbone.",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "Experimental results of different early exiting methods with BERT backbone on the GLUE benchmark.",
"figure_data": "CoLAMNLIMRPCQNLIQQPRTESST-2score speedup score speedup score speedup score speedup score speedup score speedup score speedupBERT base54.20%83.10%86.80%89.80%89.20%69.10%91.30%Fixed-Exit-3L0.075%70.075%75.875%77.475%81.875%54.775%81.075%Fixed-Exit-6L0.050%79.650%84.750%85.350%89.350%68.150%88.650%BranchyNet0.0 0.074% 51%63.8 78.376% 53%75.7 83.076% 52%74.2 87.180% 47%71.6 89.380% 50%54.7 67.476% 47%79.9 88.376% 49%Shallow-Deep0.0 0.075% 52%64.1 78.277% 51%75.6 82.876% 51%74.3 87.278% 49%71.4 89.679% 51%54.7 67.276% 48%79.5 88.477% 48%BERxiT0.0 12.376% 52%63.5 78.476% 51%75.6 82.976% 51%73.3 87.078% 48%68.2 89.180% 49%55.3 67.377% 47%79.5 88.376% 49%PABEE0.0 0.075% 50%63.9 78.977% 52%75.8 83.175% 53%73.6 87.281% 46%68.6 89.682% 49%55.8 67.775% 46%79.9 88.777% 48%F-PABEE0.0 13.675% 52%66.9 83.972% 53%81.5 87.377% 53%76.2 88.675% 54%79.6 90.882% 49%56.0 68.176% 47%80.5 92.376% 48%",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
}
] | Xiangxiang Gao; Wei Zhu; Jiasheng Gao; Congrui Yin | [
{
"authors": "Tianyang Lin",
"journal": "",
"ref_id": "b0",
"title": "A survey of transformers",
"year": "2021"
},
{
"authors": "Wei Zhu",
"journal": "",
"ref_id": "b1",
"title": "Mvp-bert: Redesigning vocabularies for chinese bert and multi-vocab pretraining",
"year": "2020"
},
{
"authors": "Wei Zhu; Xiaofeng Zhou; Keqiang Wang; Xun Luo; Xiepeng Li; Yuan Ni; Guo Tong Xie",
"journal": "",
"ref_id": "b2",
"title": "Panlp at mediqa 2019: Pre-trained language models, transfer learning and knowledge distillation",
"year": "2019"
},
{
"authors": "Yuhui Zuo; Wei Zhu; Guoyong Cai",
"journal": "",
"ref_id": "b3",
"title": "Continually detection, rapidly react: Unseen rumors detection based on continual prompt-tuning",
"year": "2022"
},
{
"authors": "Wei Zhu; Peng Wang; Xiaoling Wang; Yuan Ni; Guo Tong Xie",
"journal": "ICASSP",
"ref_id": "b4",
"title": "Acf: Aligned contrastive finetuning for language and vision tasks",
"year": "2023"
},
{
"authors": "Zhao Guo; Yuan Ni; Keqiang Wang; Wei Zhu; Guo Tong Xie",
"journal": "",
"ref_id": "b5",
"title": "Global attention decoder for chinese spelling error correction",
"year": "2021"
},
{
"authors": "Jacob Devlin",
"journal": "",
"ref_id": "b6",
"title": "BERT: pre-training of deep bidirectional transformers for language understanding",
"year": "2018"
},
{
"authors": "Zhenzhong Lan",
"journal": "",
"ref_id": "b7",
"title": "ALBERT: A lite BERT for selfsupervised learning of language representations",
"year": "2019"
},
{
"authors": "Zhilin Yang",
"journal": "",
"ref_id": "b8",
"title": "Xlnet: Generalized autoregressive pretraining for language understanding",
"year": "2019"
},
{
"authors": "Yinhan Liu",
"journal": "",
"ref_id": "b9",
"title": "Roberta: A robustly optimized BERT pretraining approach",
"year": "2019"
},
{
"authors": "Wei Zhu",
"journal": "Association for Computational Linguistics",
"ref_id": "b10",
"title": "MVP-BERT: Multi-vocab pre-training for Chinese BERT",
"year": "2021-08"
},
{
"authors": " Canwen",
"journal": "",
"ref_id": "b11",
"title": "A survey on dynamic neural networks for natural language processing",
"year": "2022"
},
{
"authors": "Wangchunshu Zhou",
"journal": "",
"ref_id": "b12",
"title": "BERT loses patience: Fast and robust inference with early exit",
"year": "2020"
},
{
"authors": " Canwen",
"journal": "",
"ref_id": "b13",
"title": "Bert-of-theseus: Compressing bert by progressive module replacing",
"year": "2020"
},
{
"authors": "Victor Sanh",
"journal": "",
"ref_id": "b14",
"title": "Distilbert, a distilled version of BERT: smaller, faster, cheaper and lighter",
"year": "2019"
},
{
"authors": "Angela Fan",
"journal": "",
"ref_id": "b15",
"title": "Reducing transformer depth on demand with structured dropout",
"year": "2019"
},
{
"authors": "Paul Michel",
"journal": "",
"ref_id": "b16",
"title": "Are sixteen heads really better than one?",
"year": "2019"
},
{
"authors": "Siqi Sun",
"journal": "",
"ref_id": "b17",
"title": "Patient knowledge distillation for BERT model compression",
"year": "2019"
},
{
"authors": "Sehoon Kim",
"journal": "",
"ref_id": "b18",
"title": "I-BERT: integer-only BERT quantization",
"year": "2021"
},
{
"authors": "Wei Zhu",
"journal": "Springer International Publishing",
"ref_id": "b19",
"title": "Autonlu: Architecture search for sentence and cross-sentence attention modeling with redesigned search space",
"year": "2021"
},
{
"authors": "Zhexi Zhang; Wei Zhu; Junchi Yan; Peng Gao; Guowang Xie",
"journal": "",
"ref_id": "b20",
"title": "Automatic student network search for knowledge distillation",
"year": "2021"
},
{
"authors": "Wei Zhu; Yuan Ni; Xiaoling Wang; Guo Tong Xie",
"journal": "",
"ref_id": "b21",
"title": "Discovering better model architectures for medical query understanding",
"year": "2021"
},
{
"authors": "Mostafa Dehghani",
"journal": "",
"ref_id": "b22",
"title": "Universal transformers",
"year": "2018"
},
{
"authors": "Zhen Zhang; Wei Zhu; Jinfan Zhang; Peng Wang; Rize Jin; Tae-Sun Chung",
"journal": "",
"ref_id": "b23",
"title": "Pcee-bert: Accelerating bert inference via patient and confident early exiting",
"year": "2022"
},
{
"authors": " Teerapittayanon",
"journal": "",
"ref_id": "b24",
"title": "Branchynet: Fast inference via early exiting from deep neural networks",
"year": "2016"
},
{
"authors": "Weijie Liu",
"journal": "",
"ref_id": "b25",
"title": "Fastbert: a self-distilling BERT with adaptive inference time",
"year": "2020"
},
{
"authors": "Ji Xin",
"journal": "",
"ref_id": "b26",
"title": "Deebert: Dynamic early exiting for accelerating BERT inference",
"year": "2020"
},
{
"authors": "Yigitcan Kaya",
"journal": "CoRR",
"ref_id": "b27",
"title": "How to stop off-the-shelf deep neural networks from overthinking",
"year": "2018"
},
{
"authors": "Roy Schwartz",
"journal": "",
"ref_id": "b28",
"title": "The right tool for the job: Matching model and instance complexities",
"year": "2020"
},
{
"authors": "Ji ",
"journal": "Association for Computational Linguistics",
"ref_id": "b29",
"title": "BERxiT: Early exiting for BERT with better fine-tuning and extension to regression",
"year": "2021-04"
},
{
"authors": "Tal Schuster",
"journal": "CoRR",
"ref_id": "b30",
"title": "Consistent accelerated inference via confident adaptive transformers",
"year": "2021"
},
{
"authors": "Wei Zhu",
"journal": "Association for Computational Linguistics",
"ref_id": "b31",
"title": "LeeBERT: Learned early exit for BERT with cross-level optimization",
"year": "2021-08"
},
{
"authors": "Wei Zhu; Xiaoling Wang; Yuan Ni; Guotong Xie",
"journal": "Association for Computational Linguistics",
"ref_id": "b32",
"title": "GAML-BERT: Improving BERT early exiting by gradient aligned mutual learning",
"year": "2021-11"
},
{
"authors": "Geoffrey E Hinton",
"journal": "",
"ref_id": "b33",
"title": "Distilling the knowledge in a neural network",
"year": "2015"
},
{
"authors": "Alex Wang",
"journal": "CoRR",
"ref_id": "b34",
"title": "GLUE: A multi-task benchmark and analysis platform for natural language understanding",
"year": "2018"
},
{
"authors": "Alice Coucke",
"journal": "",
"ref_id": "b35",
"title": "Snips voice platform: an embedded spoken language understanding system for privateby-design voice interfaces",
"year": "2018"
},
{
"authors": " Hemphill",
"journal": "",
"ref_id": "b36",
"title": "The ATIS spoken language systems pilot corpus",
"year": "1990"
},
{
"authors": "Pengcheng Yang",
"journal": "",
"ref_id": "b37",
"title": "SGM: sequence generation model for multi-label classification",
"year": "2018"
},
{
"authors": "Jeff Atwood",
"journal": "",
"ref_id": "b38",
"title": "Stack overflow creative commons data dump",
"year": "2009"
},
{
"authors": "Haoli Bai; Wei Zhang; Lu Hou; Lifeng Shang; Jing Jin; Xin Jiang; Qun Liu; Michael R Lyu; Irwin King",
"journal": "",
"ref_id": "b39",
"title": "Binarybert: Pushing the limit of bert quantization",
"year": "2020"
},
{
"authors": "Ilya Loshchilov",
"journal": "ICLR",
"ref_id": "b40",
"title": "Decoupled weight decay regularization",
"year": "2019"
},
{
"authors": " Wolf",
"journal": "",
"ref_id": "b41",
"title": "Transformers: State-of-the-art natural language processing",
"year": "2020"
}
] | [
{
"formula_coordinates": [
2,
394.28,
603.18,
160.84,
9.65
],
"formula_id": "formula_0",
"formula_text": "h 0 = Embedding(x). (1"
},
{
"formula_coordinates": [
2,
555.12,
603.5,
3.87,
8.64
],
"formula_id": "formula_1",
"formula_text": ")"
},
{
"formula_coordinates": [
2,
377.01,
681.27,
181.99,
9.65
],
"formula_id": "formula_2",
"formula_text": "p i = C i (h i ) = C i (L i (h i-1 )).(2)"
},
{
"formula_coordinates": [
3,
68.84,
351.01,
229.37,
21.61
],
"formula_id": "formula_3",
"formula_text": "pat i = pat i-1 + 1 s(p i-1 , p i ) < thre 0 s(p i-1 , p i ) >= thre(3)"
},
{
"formula_coordinates": [
3,
109.8,
612.96,
188.41,
30.32
],
"formula_id": "formula_4",
"formula_text": "s(p l-1 , p l ) = - k j=1 p l-1 j log(p l j );(4)"
},
{
"formula_coordinates": [
3,
109.8,
694.79,
188.41,
30.32
],
"formula_id": "formula_5",
"formula_text": "s(p l , p l-1 ) = - k j=1 p l j log(p l-1 j );(5)"
},
{
"formula_coordinates": [
3,
360.62,
282.69,
198.37,
11.03
],
"formula_id": "formula_6",
"formula_text": "SymKD = s(p l-1 , p l ) + s(p l , p l-1 );(6)"
},
{
"formula_coordinates": [
3,
324.84,
339.47,
234.16,
23.89
],
"formula_id": "formula_7",
"formula_text": "JSKD = 1 2 s(p l-1 , p l-1 + p l 2 ) + 1 2 s(p l , p l-1 + p l 2 )(7)"
},
{
"formula_coordinates": [
3,
361.15,
428.19,
197.84,
30.32
],
"formula_id": "formula_8",
"formula_text": "s(p l-1 , p l ) = - k j=1 2 i=1 p l-1 ji log(p l ji );(8)"
},
{
"formula_coordinates": [
3,
361.15,
490.11,
197.84,
30.32
],
"formula_id": "formula_9",
"formula_text": "s(p l , p l-1 ) = - k j=1 2 i=1 p l ji log(p l-1 ji );(9)"
},
{
"formula_coordinates": [
3,
397.54,
694.79,
161.45,
30.32
],
"formula_id": "formula_10",
"formula_text": "L = n j=1 jL j / n j=1 j(10)"
}
] |
|
10.1109/ACCESS.2022.3175317 | 2023-05-19 | [
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b0",
"b1",
"b2",
"b3",
"b4",
"b5",
"b6",
"b7",
"b8",
"b9",
"b10",
"b11",
"b12",
"b13",
"b14",
"b13",
"b15",
"b2",
"b3",
"b5",
"b15",
"b16",
"b13",
"b15",
"b17",
"b18",
"b13",
"b15",
"b16",
"b19",
"b20",
"b13",
"b15",
"b17",
"b18",
"b21",
"b2",
"b3",
"b5",
"b15"
],
"table_ref": [],
"text": "The vision-and-language navigation (VLN) [1] task, which aims to require an embodied agent to explore unstructured environments and navigate to the target location based on the fine-grained natural language instructions, has gained significant research attention owing to its flexibility and utility in various real-world scenarios. To achieve more accurate and effective navigation by capturing the relationship between visual and linguistic inputs, deep learning approaches [2][3][4][5][6][7][8][9][10][11][12] have been extensively explored in VLN.\nDespite the significant achievements of existing approaches in VLN, the problem of robustness becomes more obvious when the number of model parameters increases, due to the limited size of the Room-to-Room (R2R) benchmark dataset resulting from the high cost of labeling [13]. To solve this issue, a follower-speaker system [14] is proposed as shown in Fig. 1. The follower aims to navigate toward the target location while the speaker generates descriptions of the trajectories. This approach is particularly useful when working with unlabeled trajectories which can be easily sampled in the simulator [15]. Despite the widespread use of long short-term memory (LSTM)-based speakers [14,16] for data augmentation by numerous VLN works [3,4,6,16,17], the quality of the speaker has been largely overlooked. Notably, the quality of the speaker is crucial in providing pseudo-labels for the follower. A poorly-performing speaker can introduce significant noise and incorrect supervision in the VLN system. For example, if the speaker generates \"turn left\" or simply \"turn right\" instead of \"turn right after leaving * Corresponding author the bedroom\" in the instructions, it can cause confusion for the follower and degrade the overall data augmentation performance of the VLN system. Therefore, this paper aims to construct a robust and capable speaker to enhance the quality of data augmentation for VLN. To achieve this goal, two factors are mainly considered as follows.\nFirstly, current speaker models lack the effective leverage of inherent long-range dependencies among the given trajectories with panorama features. From a computational perspective, the choice of spatial and temporal encoding techniques can greatly impact the performance of the proposed speaker model. In previous works, bidirectional LSTM (BLSTM) [14,16], multi-layer perception (MLP) [18], and hidden outputs from the follower [19] have been used to encode the spatial features. However, BLSTM is computationally expensive and meanwhile struggles with capturing long-term dependencies. On the other hand, MLP is a simpler neural network with a lower computational cost, but it may not adequately encode multiple relationships and nuances. Using the hidden output from the follower is a cost-saving option, but it may result in the loss of key area characteristics. Regarding temporal encoding, the previous speakers [14,16,17] mainly adopt the LSTM-based network to encode visual features and decode instructions. Since the recurrent structure suffers from the nature of the longterm forgetting problem, it weakens the guidance function of visual and linguistic context both in the spatial and temporal domains. Secondly, the accurate alignment of predicted instructions and the provided trajectories is a crucial aspect for the speaker. Without proper alignment, the generated instructions may lack information regarding key actions or contain repeated phrases, ultimately leading to a different path from the ground truth. To avoid this, it is necessary for the speaker to be able to segment the path targets into distinct stages. Unfortunately, current speaker models have not fully considered this aspect, creating ambiguity as to whether the generated language fully represents the navigation status.\nTo tackle the above problems, a novel progress-aware spatio-temporal transformer speaker (PASTS) is proposed in this work. In terms of the first limitation, a spatio-temporal transformer structure is suggested to better leverage the sequenced multiple vision and action features. Concretely, a spatial encoder is first used by employing the cross-modal multi-head attention mechanism [20] to capture the correlations between 36 panoramic image patches and the oriented action-aware image. Then a temporal transformer encoder is adapted to process the successive nodes. This allows the speaker to effectively fuse observations and actions over the spatial-and temporal-dimensions. Additionally, it is imperative to enhance the practicality of the model in unfamiliar settings. However, training transformer-based architectures raises concerns about potential overfitting [21] due to its fully connected nature, especially in the presence of a limited dataset. To address this issue, a multifeature dropout (MFD) strategy is introduced during training, which significantly improves the robustness of PASTS in unforeseen environments and reduces the risk of overfitting without any extra cost.\nThe second restriction involves an alignment issue where the speaker model may fail to properly align sub-instructions with corresponding path segments. Humans can easily split long navigation steps into smaller segments and generate descriptive phrases for each stage, ensuring proper alignment. However, it is challenging for the model to learn this skill on its own. Therefore, a speaker progress monitor (SPM) is proposed to enable the progress representation and prediction within the encoder-decoder framework. The SPM operates as an independent auxiliary task, running in parallel with the initial word prediction head. By providing additional subinstruction and associated sub-step supervision signals, the PASTS can better identify the progress of word generation and thus increase the alignment of results.\nThe effectiveness of PASTS is demonstrated on the widely-used R2R dataset. Experimental results show that PASTS outperforms existing speaker models [14,16,18,19,22] and substantially improves the performance of 4 VLN follower models [3,4,6,16] when PASTS is applied using the back translation method. As a result, the proposed method achieves state-of-the-art VLN performance, indicating its superior performance and strong generalization ability compared to other approaches. Overall, the major contributions of this paper are summarized as follows:\n1) A spatio-temporal transformer encoder is proposed to fully leverage the long-term sequenced vision features in the navigation paths, which is supposed to improve the fusion of the input observations.\n2) A speaker progress monitor with a joint loss function is designed to provide strong supervision signals allowing the model to estimate its progress in instruction generation, thus facilitating more fine-grained caption results.\n3) A multifeature dropout strategy is introduced to alleviate the serious overfitting caused by the small dataset and improve the model's capability of generalization.\n4) The progress-aware spatio-temporal transformer speaker can be combined with existing navigation agent methods flexibly. Adequate experiments validate the effectiveness of the proposed modules and show that PASTS can obtain stateof-the-art performance for both speaker and follower models on the VLN task."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Vision-and-Language Navigation",
"publication_ref": [
"b0",
"b22",
"b1",
"b2",
"b23",
"b24",
"b25",
"b26",
"b27",
"b28",
"b29",
"b30",
"b2",
"b3",
"b5",
"b31",
"b32",
"b33",
"b4",
"b6",
"b14",
"b21",
"b34",
"b35",
"b36",
"b13",
"b15",
"b21",
"b13",
"b21",
"b15",
"b37",
"b16",
"b38",
"b17",
"b18"
],
"table_ref": [],
"text": "VLN is first proposed by Anderson et al. [1] based on Matterport3D [23], a large-scale reinforcement learning environment based on real imagery, and has been a focus of significant research interest in recent years [2,3,[24][25][26][27][28]. A wide range of strategies have been proposed to enhance navigation capability from diverse perspectives. For instance, reinforcement learning techniques [29] have been introduced to improve the decision-making process of VLN models, and the fusion of multiple modal input features [30,31] has been explored to augment the perceptual capacity. Additionally, constructing historical memory graph [3,4,6] and incorporating auxiliary tasks [32][33][34] have been considered to enhance the capability of history dependency and to enrich the inference modes. Despite significant progress, overfitting is still a critical challenge in VLN. The limited scale of the dataset can lead to poor performance of the agent in unseen environments. Therefore, some data augmentation and dataset expansion approaches are proposed to improve the generalization capability [5,7,15,22,[35][36][37].\nThe speaker model [14,16,22] has been proven to be a valuable tool for data augmentation in VLN. Concretely, Speaker-Follower [14] first proposes to use an independent model to automatically generate pseudo labels for data augmentation. WIP-Speaker [22] is a work-in-progress speaker model that adopts hard attention in two stages. Env-Speaker [16] uses the BLSTM and soft-attention mechanism to extract trajectories and introduces the back translation method into VLN. Zhao et al. [38] study the evaluation indicators for evaluating the performance of a speaker model. Wang et al. [17] and Dou et al. [39] both suggest training the speaker and the follower jointly. All of the above speaker models are based on the LSTM architecture. Some methods like Cmp-Speaker [18] and Imp-Speaker [19] attempt to build transformer-based speaker and follower models, but their performance improvements are limited. In summary, prior studies have demonstrated the validity of using an independent instruction generator to augment the VLN dataset and improve model performance. However, these approaches have been limited in their ability to effectively leverage panoramic images across different steps and ensure the alignment of the predicted instructions and the sampled paths. As a result, these limitations pose potential risks to the efficacy of speaker-follower systems. Therefore, this work proposes a spatio-temporal transformer speaker to generate more fine-grained instructions of higher accuracy."
},
{
"figure_ref": [],
"heading": "Transformers for Visual Captioning",
"publication_ref": [
"b39",
"b40",
"b41",
"b13",
"b39",
"b42",
"b43",
"b44",
"b45",
"b46",
"b47",
"b48"
],
"table_ref": [],
"text": "Models for visual captioning tasks (such as image captioning [40], video captioning [41], and visual question answering [42]) usually follow an encoder-decoder framework and can be regarded as performing a task of neural machine translation from image to text. Typically, convolutional neural network (CNN) -recurrent neural network (RNN) architectures are used to encode the images as feature vectors and then decode these vectors into sentences in a recurrent manner [14,40,43]. Recently, the transformer [44] and its extensions [45,46] have shown remarkable improvements in various tasks, leading to the gradual replacement of the RNN architecture [47][48][49]. Motivated by this progress, the transformer-based architecture is employed as the backbone of the proposed speaker model. However, while the speaker in VLN and visual captioning tasks share some similarities, it is important to note that they differ in certain aspects related to input form and field of application. For instance, visual captioning methods are typically designed to handle single-oriented images and may not be capable of capturing the panoramic surroundings that are crucial to VLN. Moreover, VLN involves varying actions at each step, making it a more challenging task for the speaker model. Despite drawing inspiration from visual captioning methods, developing a stronger speaker model for VLN remains an open research problem."
},
{
"figure_ref": [],
"heading": "Auxiliary Tasks",
"publication_ref": [
"b49",
"b50",
"b51",
"b32",
"b52",
"b31"
],
"table_ref": [],
"text": "In the field of machine learning, auxiliary tasks have been widely employed to improve data efficiency and robustness [50][51][52]. Existing methods have demonstrated that the performance of VLN models can also be boosted with the assistance of auxiliary tasks. Progress Monitor [33] aims to ensure that the grounded instruction correctly reflects the navigation progress. Huang et al. [53] introduced two auxiliary tasks, cross-modal alignment (CMA) and next visual scene (NVS), which involve assessing the fit between a given instruction-path pair and predicting latent representations of future visual inputs, respectively. AuxRN [32] is a framework that includes four self-supervised auxiliary reasoning tasks to take advantage of additional training signals derived from semantic information. Considering that an auxiliary task can effectively introduce additional strong supervision signals by prior knowledge and the inherent physical characteristics of the specific task at hand, a novel SPM is designed to force the model to learn to recognize its generation progress during training and thus improve the alignment between trajectories and instructions."
},
{
"figure_ref": [],
"heading": "Method",
"publication_ref": [
"b22",
"b0"
],
"table_ref": [],
"text": "Problem Setup Matterport3D [23] is used as the simulator to produce the connected graph = { , }, where represents navigable viewpoints and represents the connections between these viewpoints. In the R2R dataset [1], the data is annotated as pairs of trajectory = { 1 , 2 , ..., } and instruction = { 1 , 2 , ..., } where and represent the visited nodes and words, and and denote the length of the path and the instruction, respectively. At each step, the agent can observe a visual environmental panorama, which includes 3 perspectives that each has 12 images of 30 • . The resolution of each image is 640 × 480. Let and represent the dimension of image features and orientation features. Specifically, the actual forward path locates at one of the images at each step, and this set of images with their offset orientations is assigned as the action-level features = { ; }, where\n= { 1 , 2 , ..., } ∈ ℝ × denotes image features with the angle set = { 1 , 2 , ..., } ∈ ℝ × . Similarly, the environment-level features = { ; } is composed of the panoramic image set = {[ , ] =1 } =1 ∈ ℝ × × and the angle set = {[ , ] =1 } =1 ∈ ℝ × ×\n, where represents the number of image per panorama view.\nThe goal of the VLN follower is to enable an agent to find a correct navigation trajectory in a real indoor environment with the assistance of instructions and visual observations. In contrast, the speaker aims to predict the probability of a set of words in the instructions for a given trajectory with visual observations. Formally, the probability of action prediction and word prediction can be written as Eq. ( 1) and (2), respectively.\n( 1 , ..., | , ) = ∏ =1 ( | 1 , ..., -1 , , )(1)\n( 1 , ..., | , ) = ∏ =1 ( | 1 , ..., -1 , , )(2)"
},
{
"figure_ref": [
"fig_1",
"fig_2"
],
"heading": "Spatio-Temporal Transformer Encoder",
"publication_ref": [
"b13",
"b15",
"b43",
"b53",
"b54",
"b55",
"b7",
"b19"
],
"table_ref": [],
"text": "As shown in Fig. 2, the core of PASTS is constructed based on a sequence-to-sequence transformer architecture to explore a powerful network for instruction generation. To better use the action and environment information throughout time and space, a novel spatio-temporal transformer encoder structure is first designed based on the cross-modal attention mechanism module, which is illustrated in Fig. 3.\nThe spatial encoder The inputs to the speaker consist of two types of observations: a set of actions ∈ ℝ × and a set of environmental observations ∈ ℝ × × . Specifically, every panorama consists of 36 image patches for each viewpoint. The reasonable fusion of visual features is critical and influences the model's cognitive level. Previous most commonly used methods [14,16] have used BLSTM with a soft attention mechanism [44] to capture the most important information in panoramic sequences. In contrast, considering the complexity of BLSTM networks and the difficulty of proper training [54], we first propose a spatial encoder structure to better combine the features of actions with those of environmental observations, enabling the model to learn to focus on more relevant parts of the input views. Since the action features guide the main variants among paths, is regarded to be the query and is regarded to be the key and value. It is beneficial to linearly project the queries, keys and values through different linear transformations by using a multihead attention mechanism, allowing the model to focus on different representation subspaces at different locations. After the concatenation of the results of the multiple attention heads, a residual connection [55] followed by layer normalization [56] is applied to achieve the result ∈ ℝ × . To increase the generalization ability, a few additional feature dropouts are included in the model, which will be further discussed in Sec. 3.3. Afterwards, the formulas for the spatial encoder are given in Eq. ( 3)- (8).\ñ = (Dropout( ); )(3)\ñ = (Dropout( ); )(4)\ñ = ̃ , ̃ = ̃ , ̃ = ̃ (5) head = softmax( ̃ ̃ √ ) ̃ (6) MultiHead( , , ) = Concat(head 1 , ..., head )(7)\n= LayerNorm(MultiHead + ̃ ) (8)\nwhere\n∈ ℝ × , ∈ ℝ × , ∈ ℝ × , ∈ ℝ × , ∈ ℝ × , ∈ ℝ × .\nAfter the aggregation of action and environment features, the output of the spatial encoder is ∈ ℝ × . The multihead attention mechanism boosts the model's capability to capture the important semantics of the surroundings and fuse them into the oriented image features, reducing much useless noise and interference.\nThe temporal encoder After fusing different spatial information at each time stage via the spatial encoder, a temporal encoder is employed to enable the model to learn the inherent correlations between various time stages. This is crucial for the speaker model since the desired instruction is specified with respect to the actual navigation forward process. Concretely, the spatial-fused features = { 1 , 2 , ..., } among different steps are then input into the temporal encoder with layers as the independent tokens. Position embeddings (⋅) [20] are added to the to retain the positional information. Each encoder layer consists of a multihead self-attention (MSA) layer followed by a small feedforward neural network (FFN). Because the generated trajectory instructions are strongly related to the actual order of navigation, the position encoding is incorporated into the visual sequence embeddings. Similar to the spatial encoder, the residual connections are employed around each of the sub-layers, followed by layer normalization. Let denote the -th layer, where = 1... , the formula of the temporal encoder is as Eq. ( 9)-( 11):\n0 = ( ) (9) ′ = LayerNorm(MSA( -1 ) + -1 )(10)\n= LayerNorm(FFN( ′ ) + ′ )(11)\nThe text decoder The decoder part follows the transformer decoder architecture. Since language generation is an autoregressive process, it is necessary to ensure that each predicted word depends only on the previous ones. Therefore, the word embeddings are offset by a special token <BOS>, and a mask function is applied to the attention matrix to mask out illegal positions. Position encoding is also added to the word embeddings to capture the relative positions of the tokens in the sequence. Supposing that the target vocabulary is vocab and that each instruction contains a maximum of words, a linear layer and a softmax operation are applied to convert the values in the last hidden layer of the decoder, , ∈ ℝ × , into projected probabilities, , ∈ ℝ × vocab . For symmetry, we call the head that performs this task the speaker word projector (SWP), which is optimized with a cross-entropy loss, as shown in Eq. ( 12):\n SWP = - ∑ =1 log( ( * | * 1∶ -1 , , ))(12)\nwhere represents the parameters of our model, * 1∶ is the target ground-truth sequence, and (⋅) denotes the predicted probability that the target word is in the -th location as calculated through the model with actions and environmental observations ."
},
{
"figure_ref": [
"fig_4"
],
"heading": "Speaker Progress Monitor (SPM)",
"publication_ref": [
"b56",
"b14"
],
"table_ref": [],
"text": "Previous speaker models have overlooked the fundamental correlation embedded within the given trajectory and thus fail to relate each word to the progress along the given path. This might lead to misalignment between the trajectories and instructions. Therefore, an SPM module is proposed to enable the model to estimate the progress of instruction generation, which serves as an auxiliary task during training. To ensure that each instruction word corresponds to a specific stage of the input trajectory, the FGR2R dataset [57], which contains the fine-grained annotations for sub-paths and subinstructions, is adapted to provide the ground-truth progress signal for each word. Formally, the data pair provided by R2R is {( , )}, where = { 1 , 2 , ..., } and = { 1 , 2 , ..., } represent the trajectory with nodes and the instruction with words, respectively. By dividing the long trajectory into subsets, the data pair in FGR2R is in the format {( ′ 1 , ′ 1 ), ( ′ 2 , ′ 2 ), ..., ( ′ , ′ )}, where each subset represents a smaller navigation task. Fig. 4 shows an example that a whole trajectory can be divided into four navigation stages. Therefore, the individual word can be assigned the corresponding progress values. Suppose that = { 1 , 2 , ..., } represents the instruction progress set, and the progress value of each word follows Eq. ( 13):\n( ∈ ′ ) =(13)\nWith the above definition, all words belonging to the same subset are associated with the same progress signal. To supervise the model with the ground progress value, the SPM is integrated into the model as an auxiliary task. Concretely, in parallel with the SWP, the SPM module is employed after the last hidden layer of the decoder. The SPM contains two linear transition layers, a rectified linear unit (ReLU) activation layer and a dropout layer, as expressed in Eq. ( 14)- (15):\n′ , = ReLU(ℎ , + )(14)\n, = Dropout( ′ , ) +(15)\nwhere ∈ ℝ × hidden , ∈ ℝ × hidden , ∈ ℝ hidden ×1 , and ∈ ℝ ×1 are the learned parameters. In this work, we 16) is applied as the loss function:\n SPM = - 1 2 ∑ =1 ( * -, ) 2(16)\nwhere * represents the ground-truth progress value. Finally, the SWP and the SPM are jointly trained in an endto-end manner as Eq. ( 17):\n = SWP + (1 -) SPM (17\n)\nwhere is a weight used to control the balance between the two losses and is used to ensure that the two losses are of the same order of magnitude."
},
{
"figure_ref": [
"fig_5"
],
"heading": "Multifeature Dropout (MFD)",
"publication_ref": [
"b57",
"b15",
"b19",
"b20"
],
"table_ref": [],
"text": "Although the transformer architecture has shown remarkable performance in various domains, one major challenge in training such models is overfitting. While the speaker is designed to alleviate the overfitting problem for VLN, it also poses a similar problem for speaker training, particularly when the training data from R2R is limited. Inspired by the regularization properties of dropout [58], which randomly deactivates a fraction of the neurons during each training iteration, a multifeature dropout (MFD) strategy is proposed to augment the features and reduce overfitting during training.\nThe illustration of MFD is shown in Fig. 5. Concretely, the four fundamental modules of the transformer-based structure are considered: the input feature extractor, the feed-forward network, the attention mechanism, and the output projection module. First, the environment dropout [16] has been demonstrated to be extremely effective in creating various environments in VLN. Therefore, this type of dropout is applied to the input features after the feature extractor, and named MFD-1. The angles are left unchanged since every angle is essential in navigation; even minor mutations may lead to considerable ambiguity. Additionally, the basic transformer [20] contains two specific types of feature dropout, which locate after the activation layer of the FFN modules and after the softmax function of the attention modules, which are called MFD-2 and MFD-3, respectively. To further increase the diversity of the structure, following UniDrop [21], two additional types of feature dropout, MFD-4 and MFD-5, which locates before the calculation of , , and before the final output projection respectively, are added in the framework as well. Above all, the total five types of feature dropout applied in PASTS are summarized as follows:\n1) MFD-1 (environment dropout): After using the image feature extractor to capture the visual observation, MFD-1 is adopted to randomly mask different regions of environment features.\n2) MFD-2 (activation dropout): The transformer layers comprise of FFN and attention mechanism. In FFN, MFD-2 is applied after the activation function between the two linear transition layers.\n3) MFD-3 (attention dropout): In the attention modules, MFD-3 is applied to the attention matrix, which is after the matrix product but before weighting by the value matrix .\n4) MFD-4 (query, key, and value dropout): To improve the diversity of queries and keys in the attention mechanism, MFD-4 is applied after linearization with , , and and before the matrix product of and .\n5) MFD-5 (output dropout): A softmax function with linear projection is used after the last hidden layer of the decoder to output the outputs. MFD-5 is applied before the linear transition layer of the final classification module."
},
{
"figure_ref": [
"fig_6"
],
"heading": "Training Procedures",
"publication_ref": [
"b13",
"b2",
"b5",
"b15",
"b58",
"b3",
"b14",
"b15"
],
"table_ref": [],
"text": "The speaker-follower system [14] includes two agents: the follower is to follow the instructions and navigate to the target area, and the speaker is to generate pseudo instructions for augmenting data pairs. In the previous sections, the proposed speaker PASTS and its training loss are mainly discussed. Here, the training strategy of the follower and the back translation method are briefly introduced.\nFollower training Let , , denote instructions, trajectories and environments, respectively. The encoderdecoder follower is to learn the mapping of { , } → . Since this paper mainly focuses on the speaker study, the training strategy of the follower keeps consistent with the corresponding approaches. Following the operation in recent works [3,6,16], a mixture of imitation learning (IL) and reinforcement learning (RL) is adopted to train the follower models. The IL relies on expert action at each step, while RL samples actions according to the policy . Specifically, the loss of IL is used for off-policy learning and is written as Reinforcement learning is applied for on-policy learning, where the optimization objectiveness is expressed as\n IL = - ∑ =1 log ( * )(18)\n RL = - ∑ =1 log ( ̂ ℎ )Λ(19)\nwhere Λ indicates the advantage in the Actor-Critic RL algorithm [59]. Thus, the total loss of the follower training is follower = IL + RL RL , where RL is used to adjust the weight proportion of two losses. When training DUET [4], RL is replaced with the pseudo interactive demonstrator (PID) strategy. Back translation As shown in Fig. 6, the central idea of back translation is to translate sampled paths ′ and augmented environments ′ into pseudo instructions ′ and use new tuples { ′ , ′ , ′ } to augment the training dataset. In this work, PASTS is first trained based on the optimization objective in Eq. ( 17) in the original training dataset . Subsequently, with the large number of trajectories sampled by [15], the trained PASTS is utilized in conjunction with environment dropout to generate new data pairs ′ . For stabilizing the optimization of back translation, as [16], and ′ are alternated employed during training while sharing environment dropouts in the same batch. This enables the follower to be exposed to a wider range of environments, thus enhancing its generalization ability."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Dataset",
"publication_ref": [
"b2",
"b3",
"b5",
"b15",
"b0",
"b14",
"b56"
],
"table_ref": [],
"text": "Following previous VLN methods [3,4,6,16], the standard VLN benchmark R2R dataset [1] is used to evaluate the performance of the PASTS model for the VLN task. The R2R dataset contains images from 90 different buildings, 21,576 navigation instructions with an average length of 29 words, and 7,189 paths with an average length of 10 meters. The R2R dataset is divided into a training set, a validation-seen set (Val Seen), a validation-unseen set (Val Unseen), and a test unseen set. The results of the test set are reported through an online challenge leaderboard. The training set and the validation-seen set cover the same 61 scenes, with the corresponding 15,045 instructions split into sets of 14,025 and 1,020, respectively. The validation-unseen set consists of 11 scenes and 2,349 instructions, and the remaining 18 scenes belong to the test set. The PREVE-LANT dataset [15] is employed to provide 665,206 sampled trajectories during the back translation stage for training the follower. In this paper, the FGR2R dataset [57] is utilized to provide annotations of sub-pairs for supervising SPM. Notably, FGR2R only includes segmentation information, rather than additional trajectory-instruction pairs other than R2R."
},
{
"figure_ref": [],
"heading": "Metrics",
"publication_ref": [
"b61",
"b62",
"b63",
"b64"
],
"table_ref": [],
"text": "For evaluating the instruction generation, the four typical techniques are used to evaluate the speaker: BLEU [62], ROUGE [63], CIDEr [64] and SPICE [65]. Concretely, BLEU and ROUGE focus on the accuracy and recall of the predicted sentences, respectively. CIDEr and SPICE are mainly applied in the field of image captioning. For evaluating the navigation performance, four standard metrics are used: 1) The navigation error (NE) measures the distance between the last location in prediction and in reference. 2) The success rate (SR) shows the frequency of the correct stop within a certain threshold distance of ground truth.\n3) The trajectory length (TL) shows the average length of navigation. 4) The success rate weighted by the path length (SPL) considers both SR and the TL and takes both effectiveness and efficiency into account."
},
{
"figure_ref": [],
"heading": "Implementation Details",
"publication_ref": [
"b65",
"b54",
"b66",
"b67"
],
"table_ref": [],
"text": "PASTS is trained based on the corpus from the R2R and the FGR2R dataset, where the latter is used to provide the alignment supervision signal for training SPM. The code is built based on Python and experiments are conducted on a single NVIDIA GeForce RTX 2060 GPU with Intel Core i7-9750H CPU. As in [66], the detailed experiment settings of PASTS are listed in Tab. 1. The weight initialization follows the default settings in PyTorch. The model with the highest BLEU-4 on the unseen set is saved. To evaluate the impact of visual representations on the model, three widely used image feature extractors: ResNet-152 [55], ViT-B/16 [67], and CLIP-B/32 [68] are analyzed. The dimensions of image features are 2048, 768, and 512, respectively. When training the different follower models based on back translation, the "
},
{
"figure_ref": [],
"heading": "Main Results for Different Speaker Models",
"publication_ref": [
"b13",
"b21",
"b15",
"b18",
"b17"
],
"table_ref": [],
"text": "Tab. 2 compares the performance of the different speaker models in generating trajectory instructions on the R2R benchmark. Speaker-Follower [14] uses an LSTM architecture with panoramic views as inputs. WIP-Speaker [22] proposes a work-in-progress speaker model that adopts the hard attention mechanism in two stages. Env-Speaker [16] stacks two BLSTM encoders to enhance the visual representations. Imp-Speaker [19] and Cms-Speaker [18] are two recent efforts to develop the speaker and the agent both based on the transformer.\nThe results presented in Tab. 2 indicate that PASTS outperforms previous speaker models when evaluated under the same visual features, thereby achieving a new state-ofthe-art performance for the VLN task. The substantial improvement in BLEU and SPICE scores suggests that PASTS generates more natural, fluent, and semantically rich trajectory descriptions compared to the existing speaker models. Notably, the observed performance gains are consistent across both the seen and unseen validation sets, indicating that PASTS exhibits strong generalization and robustness capabilities. Furthermore, when using alternative feature extractors, ViT-B/16 demonstrates superior performance compared to the other two extractors (e.g., SPICE 22.9 vs. 20.9, and CIDEr 16.8 vs. 15.2). This suggests that ViT-B/16 is more effective in capturing the visual and semantic information of the images and can potentially improve the quality of captioning."
},
{
"figure_ref": [],
"heading": "Results in Combination with VLN Followers",
"publication_ref": [
"b15",
"b5",
"b2",
"b3",
"b3"
],
"table_ref": [],
"text": "To verify the hypothesis that a more accurate speaker leads to better performance of the VLN follower by providing more precise pseudo-instructions, the trained PASTS is integrated with four recent VLN methods based on back translation. Specifically, EnvDrop [16] uses the environmental dropout with a BLSTM structure. RecBERT [6] designs a recurrent BERT structure for action prediction. HAMT [3] leverages a history-aware multimodal transformer to memorize past information. DUET [4] introduces a dual-scale graph transformer for long-term action planning. The first two are based on ResNet-152 features and the second two are based on ViT-B/16 features.\nAs shown in Tab. 3, PASTS achieves significant improvements compared with the existing methods. Concretely, the improvement for EnvDrop is the most obvious, where SR is improved by 7%, 3%, and 5% and SPL is improved by 6%, 2%, and 5% on the validation-seen, validation-unseen and test-unseen sets, respectively. The improvements for the other three models are smaller since they have been pre-trained in the first stage. Nevertheless, the experimental results still prove that PASTS can further improve the robustness of these pre-trained VLN models. For instance, with the pseudo labels generated by PASTS, the state-of-the-art DUET [4] achieves 4% SPL and 2% SR improvement on the validation-seen set. This indicates that the proposed approach allows the generative pseudo instructions to align better with the sampled trajectories, thereby reducing the potential risks of noise. These findings suggest that PASTS is a model-agnostic approach that can effectively enhance the performance of existing VLN models. Its ease of implementation and potential for improving learning outcomes make it a promising tool for further research in the field."
},
{
"figure_ref": [],
"heading": "Ablation Studies",
"publication_ref": [],
"table_ref": [],
"text": "In this section, ablation studies based on ResNet-152 are presented to validate the contributions of the proposed modules to the effectiveness of the PASTS framework. The results are reported in Tab. 4."
},
{
"figure_ref": [],
"heading": "Effect of the different spatial fusion methods",
"publication_ref": [],
"table_ref": [],
"text": "In Tab. 4, \"w/o spatial encoding\" means that only action features are used. \"SA\" and \"MCA\" represent the soft attention mechanism and the multihead cross attention mechanism, respectively. \"BLSTM\" denotes the bidirectional LSTM encoding applied to the action features. The results demonstrate that the fusion of panoramic information can effectively improve the richness of effective information, but different fusion techniques will produce varied results. It shows that the spatial encoder with mutlihead cross-modal attention increases the performance on all metrics(e.g., BELU-4 19.7 vs. 18.5, and SPICE 25.6 vs. 23.9). The multihead structure enables the model to attend to various parts of the inputs and learn more semantically rich visual representations, enhancing the ability to observe and comprehend its surroundings. Moreover, the addition of BLSTM to action features does not significantly improve performance and may even decrease certain metrics in the unseen environment. To avoid introducing unnecessary parameters, a spatial encoder is designed using cross-modal multihead attention to capture visual observations in space."
},
{
"figure_ref": [],
"heading": "Effect of the SPM",
"publication_ref": [],
"table_ref": [],
"text": "A joint loss function with two hyperparameters is proposed to train the model in an end-to-end manner. The parameter balances the SWP and SPM losses, and ensures they are of the same magnitude. Based on experiments, we have found that the loss of the SWP is approximately 10 times the loss of the SPM. Therefore, to unify the magnitudes of these two losses, we set = 10. As shown in Tab. 4, the model with = 0.8 achieves the best performance. This configuration results in significant improvements across all metrics, particularly with a notable increase in SPICE from 24.4 to 25.6 on the seen validation split. It is worth noting that when the weight of the SPM surpasses that of the original loss, such as with = 0.4, the model's performance decreases. This outcome is attributed to the priority of the auxiliary task, which should not outweigh that of the main task. Otherwise, the model may focus more on optimizing the auxiliary task, which can lead to suboptimal performance. As a result, the proposed SPM is effective in improving the alignment and coherence of the generated instructions with the navigation stages, which enables PASTS to better fit the trajectory and produce instructions with greater fluency and richer semantic content."
},
{
"figure_ref": [],
"heading": "Effect of the MFD",
"publication_ref": [],
"table_ref": [],
"text": "The issue of overfitting is a significant challenge for transformer-based architectures, especially when dealing with limited and specialized datasets. To address this problem, MFD which comprises five specific types of feature dropout is suggested in this work. The results demonstrate that using the full MFD (denoted as MFD 1-5 ) results in significant improvements across all metrics in both the seen "
},
{
"figure_ref": [],
"heading": "Effect of different modules on training followers",
"publication_ref": [
"b2"
],
"table_ref": [],
"text": "In addition to exploring the impact of different modules on the speaker, experiments on the effects of promoting follower training are conducted. As shown in Tab. 5, where ST denotes the spatio-temporal transformer structure, it can be seen that PASTS with the proposed modules successfully boost the navigation performance of HAMT [3] in both seen and unseen environments. For instance, incorporating SPM leads to an improvement in SR on the unseen set from 66.58 to 67.69. Moreover, with the inclusion of MFD, there are larger improvements in SR and SPL on the unseen split, from 66.24 to 68.28, and from 61.55 to 62.37, respectively. This finding suggests that a more robust speaker has the potential to provide more precise pseudo instructions, thereby minimizing the interference that could potentially impede the follower training process."
},
{
"figure_ref": [
"fig_7"
],
"heading": "Visualization 4.7.1. Heatmap of Vision-and-Language Decoder",
"publication_ref": [],
"table_ref": [],
"text": "To further validate the proposed method compared with previous methods, visualized examples are shown in Fig. 7. It can be seen that each predicted word corresponds well to the navigation phase. For example, when the model outputs \"walk past the stairs and into the foyer\", it focuses on the first three viewpoints, where there is indeed a stair and a hall in the pictures. Additionally, PASTS can better capture the dominant object and meanwhile reduce the repeated generation issue. For example, in the left case, Speaker-Follower outputs \"walk past the dining table and turn right\" many times, and Env-Speaker ignores the key landmark \"hall\". As a result, compared with previous methods, PASTS successfully mitigates the repetition problem and enhances the richness of the generated visual representations."
},
{
"figure_ref": [],
"heading": "Analysis of Uncertainty and Convergence",
"publication_ref": [
"b68",
"b69"
],
"table_ref": [],
"text": "The uncertainty analysis is conducted on the optimum model using the Monte-Carlo Dropout (MCD) method [69,70], and the results are presented in Fig. 8 (a) and (b). The analysis involves running the model with dropout enabled for 5 rounds and measuring the variance of the results, which is found to be 0.0071. Fig. 8 (c) shows the learning curves of the loss and BLEU-4 on the unseen set. The results demonstrate that the model exhibits good convergence and stability, with the uncertainty variance being controlled within a small range. Additionally, the model is capable of quickly reaching the training peak, indicating efficient convergence."
},
{
"figure_ref": [],
"heading": "Speaker Progress Prediction",
"publication_ref": [],
"table_ref": [],
"text": "The progress values output by the SPM prediction head and the predicted words are visualized in Fig. 9. For each example, the left to right columns denote the predicted words, predicted progress, and the given paths, respectively. The heatmaps of progresses show that the gaps between different phrases with respect to different navigation stages are obvious. For example, in (a), the average progress value of \"exit the bathroom\" is about 0.31, and that of \"walk out of the bedroom\" is about 0.52. This demonstrates PASTS is capable of recognizing the correlation between phrases and paths, which can assist the speaker in producing more finegrained instructions for lengthy trajectories."
},
{
"figure_ref": [
"fig_11"
],
"heading": "Navigation Results",
"publication_ref": [],
"table_ref": [],
"text": "As described in Sec. 3.4, PASTS is employed to provide more sorts of pseudo instructions based on the back translation method to enhance the generalization of initial follower models. Some visualized instances of predicted pathways on the validation set are displayed in Fig. 10. It demonstrates that the generalization of the HAMT model significantly improves with the aid of the data augmentation given by PASTS. The trajectories predicted by the follower trained with PASTS are more consistent with the given natural instructions and can get close to the intended destination. However, the model without PASTS augmentation may lead to the wrong ending location (as shown in (b)) or completely wrong path predictions (as shown in (a) and (c)). This is Instruction: Walk out from behind the piano and towards the dining room table. Once you reach the table, turn left and enter the next room with a table. Once in that room, turn left and then stop in front of the mantle.\n Instruction: Go through the doorway to the room with the sink and follow the hallway until you get to a bedroom. Wait at the bedroom door. due to the limited instructions and environments from the original dataset during training, which could lead to the misinterpretation of the unseen inputs."
},
{
"figure_ref": [
"fig_12",
"fig_12"
],
"heading": "Limitations and Future Work",
"publication_ref": [
"b70"
],
"table_ref": [],
"text": "While the proposed method has shown promising results on the VLN task, some limitations of our approach are discussed in this section to inspire future work. Fig. 11 (a) demonstrates that the generated instructions have a biased tendency to be shorter than human-labeled instructions, which may reduce detail and richness. This could be attributed to the cross-entropy optimization used during training, which is designed to minimize the difference between the predictions and the ground truth. In practice, this may encourage the model to focus on the most salient information in the trajectory rather than some redundant descriptions that could lead to extra loss. Future work may explore reinforcement learning techniques to reward the model for generating instructions that are not only accurate but also informative and detailed. Additionally, although PASTS can capture more semantic information than previous speakers, it is still less than ground truth, as shown in Fig. 11 (b). The possible reason could be the lack of object-level visual representations. PASTS currently relies solely on image features to generate instructions. Future work could include incorporating object detection and recognition modules into the model. Moreover, the expansion of the corpus or the analysis of different roles of parameters [71] may also help to further improve the performance, which we leave for future exploration."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "In this paper, a novel progress-aware spatio-temporal transformer speaker (PASTS) is proposed to perform more Finally, it is worth noting that the research on the speaker itself is of great importance. The ability of an embodied agent to accurately and fluently describe a path or a series of events can greatly enhance the user's experience in humancomputer interaction, and has enormous potential for future artificial intelligence research."
},
{
"figure_ref": [],
"heading": "Acknowledgments",
"publication_ref": [],
"table_ref": [],
"text": "This paper is supported by the National Natural Science Foundation of China under Grants (62233013, 62073245, 62173248). Suzhou Key Industry Technological Innovation-Core Technology R&D Program (SGC2021035); Special funds for Jiangsu Science and Technology Plan (BE2022119). Shanghai Municipal Science and Technology Major Project (2021SHZDZX0100) and the Fundamental Research Funds for the Central Universities. Shanghai Science and Technology Innovation Action Plan (22511104900)."
}
] | Vision-and-language navigation (VLN) is a crucial but challenging cross-modal navigation task. One powerful technique to enhance the generalization performance in VLN is the use of an independent speaker model to provide pseudo instructions for data augmentation. However, current speaker models based on Long-Short Term Memory (LSTM) lack the ability to attend to features relevant at different locations and time steps. To address this, we propose a novel progress-aware spatio-temporal transformer speaker (PASTS) model that uses the transformer as the core of the network. PASTS uses a spatio-temporal encoder to fuse panoramic representations and encode intermediate connections through steps. Besides, to avoid the misalignment problem that could result in incorrect supervision, a speaker progress monitor (SPM) is proposed to enable the model to estimate the progress of instruction generation and facilitate more fine-grained caption results. Additionally, a multifeature dropout (MFD) strategy is introduced to alleviate overfitting. The proposed PASTS is flexible to be combined with existing VLN models. The experimental results demonstrate that PASTS outperforms all existing speaker models and successfully improves the performance of previous VLN models, achieving stateof-the-art performance on the standard Room-to-Room (R2R) dataset. | PASTS: Progress-Aware Spatio-Temporal Transformer Speaker For Vision-and-Language Navigation | [
{
"figure_caption": "PASTS:Figure 1 :1Figure 1: Illustration of the follower-speaker system in VLN, where the follower aims to predict the action based on instructions, and the speaker aims to generate instructions based on trajectories. In this paper, the proposed PASTS speaker has the ability to recognize different stages for navigation (shown in different colors) and generate more accurate and fine-grained instructions.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: Overall architecture of the PASTS framework, which consists of three sub-modules: (a) the spatio-temporal encoder is to integrate dominant features from environments at each location, and encode successive action features in the time dimension; (b) the generation decoder is responsible for converting the inputs of fused visual information and shifted word into a sequence of target probabilities; (c) the word prediction head and the progress prediction head are designed to predict instruction words and progress values, respectively.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: Illustration of the spatial encoder (top) and the temporal encoder (bottom). The spatial encoder is used to effectively fuse the action embedding and the environment embedding in each step, and the temporal encoder is applied to capture the internal connection between different navigation steps.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "S. 11Wa l k past the fireplace and chair and into the hallway. (Progress: 0.25) S.2 Veer left and walk past the end table.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: Illustration of the speaker progress monitor (SPM). The complete trajectories and instructions are divided into several subsets. Each word is assigned the corresponding progress value (shown in brown) to align the instructions and trajectories. Different colors are used to represent different navigation stages.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 :5Figure 5: Illustration of the multifeature dropout (MFD). In order to avoid serious overfitting, five dropouts at different positions are proposed to increase the diversity of the network structure. (a) represents the input representation, (b) indicates the attention mechanism, (c) represents the feed-forward network, and (d) is the output projection.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Figure 6 :6Figure 6: Illustration of the training procedure of the follower via the speaker based on the back translation method.",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "•GT:Figure 7 :7Figure 7: Visualization of some generated instructions compared with the ground truth (GT) and previous methods. The top panels show panorama observations of navigation paths. The middle panels illustrate the attention values in the last decoder layer. The x and y-axis represent the predicted word and the step along the path. The bottom sentences show predictions using different methods. Colors indicate failure (blue) and outstanding predictions (red).",
"figure_data": "",
"figure_id": "fig_7",
"figure_label": "7",
"figure_type": "figure"
},
{
"figure_caption": "Figure 8 :Figure 9 :89Figure 8: Analysis of uncertainty and convergence. (a) presents the values of cross entropy performed by 5 different rounds with dropout. (b) shows the variance values. (c) is the illustration of the learning curves.",
"figure_data": "",
"figure_id": "fig_9",
"figure_label": "89",
"figure_type": "figure"
},
{
"figure_caption": "Walk through the outdoors deck area towards the open door beside the stair leading into the yard. Walk inside the home and continue around the corner into the living room area near the kitchen sink counter top. Ground Truth VLN-HAMT VLN-HAMT+PASTS (Ours)",
"figure_data": "",
"figure_id": "fig_10",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 10 :10Figure 10: Comparison of navigation results obtained using different approaches. The star icon denotes the starting point, and arrows in different colors represent different navigation results. The instructions for guidance are shown in the bottom.",
"figure_data": "",
"figure_id": "fig_11",
"figure_label": "10",
"figure_type": "figure"
},
{
"figure_caption": "Figure 11 :11Figure 11: Statistical results of data distribution. (a) shows the distribution of the length of instructions and (b) presents the distribution of the number of object words.",
"figure_data": "",
"figure_id": "fig_12",
"figure_label": "11",
"figure_type": "figure"
},
{
"figure_caption": "Experiment settings.",
"figure_data": "ItemSettingLayer of temporal encoder6Layer of decoder6Number of head6Dimension of head64Hidden dimension of encoder 512Hidden dimension of decoder 256Hidden dimension of FFN1024Dropout ratio of MFD 10.3Dropout ratio of MFD 2-50.2Learning rate5 × 10 -5Batch size64OptimizerAdamIteration80,000 (∼6.5 hours)",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Performance comparison with existing speaker models on the R2R dataset. All the values are reported as percentages (%). The underlined value represents the maximum value using the same feature (ResNet-152), while the bolded value represents the maximum value using different features.",
"figure_data": "MethodFeatureR2R Validation-Seen BLEU-1↑ BLEU-4↑ ROUGE-L↑ CIDEr↑ SPICE↑ BLEU-1↑ BLEU-4↑ ROUGE-L↑ CIDEr↑ SPICE↑ R2R Validation-UnseenSpeaker-Follower [14] ResNet-15253.715.535.012.120.352.214.234.611.418.8WIP-Speaker [22]ResNet-15254.915.735.213.721.454.815.035.313.219.7Env-Speaker [16]ResNet-15256.818.236.418.022.655.716.735.714.420.5Cmp-Speaker [18]ResNet-15253.716.234.415.321.752.514.633.911.819.5Imp-Speaker [19]ResNet-15258.619.436.918.824.055.616.735.413.820.8PASTS (Ours)ResNet-15259.619.737.819.525.657.117.636.615.221.5PASTS (Ours)CLIP-B/3260.019.837.720.525.657.618.636.916.522.3PASTS (Ours)ViT-B/1660.920.538.221.726.357.918.537.016.822.9",
"figure_id": "tab_1",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Performance comparison with existing agent models on the R2R dataset. All the values are reported as percentages (%).",
"figure_data": "MethodR2R Validation-Seen NE↓ TL SR↑ SPL↑ NE↓ R2R Validation-Unseen TL SR↑ SPL↑ NE↓ R2R Test (Unseen, Single-run) TL SR↑ SPL↑Random [1]9.459.5816-9.239.7716-9.799.931712Seq-to-Seq [1]6.01 11.3339-7.818.3922-7.858.132018Speaker-Follower [14]3.36-66-6.62-35-6.62-3528Self-Monitoring [33]3.22-67585.52-45325.67-4835RCM [29]3.53 10.6567-6.09 11.4643-6.12 11.974338AuxRN [52]3.33-70675.28-55505.15-5551PREVALENT [15]3.67 10.3260654.73 10.1957534.75 10.515451RelGraph [60]3.47 10.1367654.739.9957534.75 10.295552NVEM [26]3.44 11.0969654.27 11.8360554.37 12.985854HOP [61]2.72 11.2675703.80 12.2764573.83 12.686459EnvDrop [16]3.99 11.0062595.22 10.7052485.23 11.705147EnvDrop+PASTS (Ours)3.38 11.9269654.80 15.2655505.30 10.905652RecBERT [6]2.82 11.1274694.24 12.3961554.29 12.766055RecBERT+PASTS (Ours) 3.13 11.2072673.99 11.7063584.07 12.086459HAMT [3]2.51 11.1576722.29 11.4666613.93 12.276560HAMT+PASTS (Ours)2.46 11.3377743.37 11.7768623.77 12.916761DUET [4]2.25 11.8280753.67 12.9672603.65 14.736959DUET+PASTS (Ours)1.89 11.3182793.01 13.4872623.55 14.397060features and training strategy for the follower and PASTSkeep the same as those methods to ensure a fair comparison.",
"figure_id": "tab_2",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Ablation studies of the PASTS. The values in bold represent the best results achieved in the corresponding module.",
"figure_data": "SettingR2R Validation-Seen BLEU-1↑ BLEU-4↑ ROUGE-L↑ CIDEr↑ SPICE↑ BLEU-1↑ BLEU-4↑ ROUGE-L↑ CIDEr↑ SPICE↑ R2R Validation-UnseenSpatial Fusion Method.w/o spatial encoding56.818.236.418.022.655.716.735.713.620.5w/ SA57.018.036.616.422.556.217.435.514.820.4w/ BLSTM + SA59.119.237.519.124.755.717.035.514.621.4w/ MCA59.619.737.819.525.657.117.636.615.221.5w/ BLSTM + MCA59.619.337.519.425.756.717.436.014.621.5Speaker Progress Monitor.w/o SPM58.418.937.418.724.456.117.236.414.921.1w/ SPM ( = 0.4)59.319.837.719.024.856.416.735.914.720.4w/ SPM ( = 0.6)59.419.237.419.224.556.217.636.015.321.2w/ SPM ( = 0.8)59.619.737.819.425.657.117.636.615.221.5Multifeature Dropout.w/o MFD57.818.536.216.724.155.416.235.513.620.5w/ MFD 2-358.919.937.720.825.555.717.035.914.820.9w/ MFD 1-359.419.937.620.425.556.217.235.915.220.3w/ MFD 1-559.619.737.819.425.657.117.636.615.221.5",
"figure_id": "tab_3",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Performances of different modules on follower training.",
"figure_data": "MethodST SPM MFDVal Seen SR↑ SPL↑Val Unseen SR↑ SPL↑HAMT75.61 72.18 66.24 61.55+PASTS✓76.79 72.49 66.58 60.29+PASTS✓✓76.90 71.89 67.69 61.90+PASTS✓✓✓77.",
"figure_id": "tab_4",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "tab_5",
"figure_label": "",
"figure_type": "table"
}
] | Liuyi Wang; Chengju Liu; Zongtao He; Shu Li; Qingqing Yan; Huiyi Chen; Qijun Chen | [
{
"authors": "P Anderson; Q Wu; D Teney; J Bruce; M Johnson; N Sünderhauf; I Reid; S Gould; A Van Den; Hengel",
"journal": "",
"ref_id": "b0",
"title": "Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments",
"year": "2018"
},
{
"authors": "X Wang; Q Huang; A Celikyilmaz; J Gao; D Shen; Y.-F Wang; W Wang; L Zhang",
"journal": "IEEE transactions on pattern analysis and machine intelligence",
"ref_id": "b1",
"title": "Vision-language navigation policy learning and adaptation",
"year": "2020"
},
{
"authors": "S Chen; P.-L Guhur; C Schmid; I Laptev",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b2",
"title": "History aware multimodal transformer for vision-and-language navigation",
"year": "2021"
},
{
"authors": "S Chen; P.-L Guhur; M Tapaswi; C Schmid; I Laptev",
"journal": "",
"ref_id": "b3",
"title": "Think global, act local: Dual-scale graph transformer for vision-and-language navigation",
"year": "2022"
},
{
"authors": "J Li; H Tan; M Bansal",
"journal": "",
"ref_id": "b4",
"title": "Envedit: Environment editing for vision-andlanguage navigation",
"year": "2022"
},
{
"authors": "Y Hong; Q Wu; Y Qi; C Rodriguez-Opazo; S Gould",
"journal": "",
"ref_id": "b5",
"title": "Vln bert: A recurrent vision-and-language bert for navigation",
"year": "2021"
},
{
"authors": "P.-L Guhur; M Tapaswi; S Chen; I Laptev; C Schmid",
"journal": "",
"ref_id": "b6",
"title": "Airbert: Indomain pretraining for vision-and-language navigation",
"year": "2021"
},
{
"authors": "M Rostami; M Oussalah; V Farrahi",
"journal": "IEEE Access",
"ref_id": "b7",
"title": "A novel time-aware food recommender-system based on deep learning and graph clustering",
"year": "2022"
},
{
"authors": "M Rostami; U Muhammad; S Forouzandeh; K Berahmand; V Farrahi; M Oussalah",
"journal": "",
"ref_id": "b8",
"title": "An effective explainable food recommendation using deep image clustering and community detection",
"year": "2022"
},
{
"authors": "R Dang; Z Shi; L Wang; Z He; C Liu; Q Chen",
"journal": "",
"ref_id": "b9",
"title": "Unbiased directed object attention graph for object navigation",
"year": "2022"
},
{
"authors": "Z He; L Wang; S Li; Q Yan; C Liu; Q Chen",
"journal": "",
"ref_id": "b10",
"title": "Mlanet: Multilevel attention network with sub-instruction for continuous visionand-language navigation",
"year": "2023"
},
{
"authors": "L Wang; Z He; R Dang; H Chen; C Liu; Q Chen",
"journal": "IEEE Transactions on Circuits and Systems for Video Technology",
"ref_id": "b11",
"title": "Res-sts: Referring expression speaker via self-training with scorer for goaloriented vision-language navigation",
"year": "2023"
},
{
"authors": "Y Zhang; H Tan; M Bansal",
"journal": "",
"ref_id": "b12",
"title": "Diagnosing the environment bias in vision-and-language navigation",
"year": "2021"
},
{
"authors": "D Fried; R Hu; V Cirik; A Rohrbach; J Andreas; L.-P Morency; T Berg-Kirkpatrick; K Saenko; D Klein; T Darrell",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b13",
"title": "Speakerfollower models for vision-and-language navigation",
"year": "2018"
},
{
"authors": "W Hao; C Li; X Li; L Carin; J Gao",
"journal": "",
"ref_id": "b14",
"title": "Towards learning a generic agent for vision-and-language navigation via pre-training",
"year": "2020"
},
{
"authors": "H Tan; L Yu; M Bansal",
"journal": "",
"ref_id": "b15",
"title": "Learning to navigate unseen environments: Back translation with environmental dropout",
"year": "2019"
},
{
"authors": "H Wang; W Liang; J Shen; L Van Gool; W Wang",
"journal": "",
"ref_id": "b16",
"title": "Counterfactual cycle-consistent learning for instruction following and generation in vision-language navigation",
"year": "2022"
},
{
"authors": "A Magassouba; K Sugiura; H Kawai",
"journal": "IEEE Robotics and Automation Letters",
"ref_id": "b17",
"title": "Crossmap transformer: A crossmodal masked path transformer using double back-translation for vision-and-language navigation",
"year": "2021"
},
{
"authors": "Z Wu; Z Liu; T Wang; D Wang",
"journal": "IEEE MultiMedia",
"ref_id": "b18",
"title": "Improved speaker and navigator for vision-and-language navigation",
"year": "2021"
},
{
"authors": "A Vaswani; N Shazeer; N Parmar; J Uszkoreit; L Jones; A N Gomez; Ł Kaiser; I Polosukhin",
"journal": "",
"ref_id": "b19",
"title": "Attention is all you need",
"year": "2017"
},
{
"authors": "Z Wu; L Wu; Q Meng; Y Xia; S Xie; T Qin; X Dai; T.-Y Liu",
"journal": "",
"ref_id": "b20",
"title": "Unidrop: A simple yet effective technique to improve transformer without extra cost",
"year": "2021"
},
{
"authors": "S Agarwal; D Parikh; D Batra; P Anderson; S Lee",
"journal": "",
"ref_id": "b21",
"title": "Visual landmark selection for generating grounded and interpretable navigation instructions",
"year": "2019"
},
{
"authors": "A Chang; A Dai; T Funkhouser; M Halber; M Niessner; M Savva; S Song; A Zeng; Y Zhang",
"journal": "",
"ref_id": "b22",
"title": "Matterport3D: Learning from RGB-D data in indoor environments",
"year": "2017"
},
{
"authors": "T Zhang; X Hu; J Xiao; G Zhang",
"journal": "Engineering Applications of Artificial Intelligence",
"ref_id": "b23",
"title": "A survey of visual navigation: From geometry to embodied AI",
"year": "2022"
},
{
"authors": "W Zhang; C Ma; Q Wu; X Yang",
"journal": "IEEE Transactions on Circuits and Systems for Video Technology",
"ref_id": "b24",
"title": "Language-guided navigation via cross-modal grounding and alternate adversarial learning",
"year": "2020"
},
{
"authors": "D An; Y Qi; Y Huang; Q Wu; L Wang; T Tan",
"journal": "",
"ref_id": "b25",
"title": "Neighbor-view enhanced model for vision and language navigation",
"year": "2021"
},
{
"authors": "R Dang; L Wang; Z He; S Su; C Liu; Q Chen",
"journal": "",
"ref_id": "b26",
"title": "Search for or navigate to? dual adaptive thinking for object navigation",
"year": "2022"
},
{
"authors": "L Wang; Z He; J Tang; R Dang; N Wang; C Liu; Q Chen",
"journal": "",
"ref_id": "b27",
"title": "A dual semantic-aware recurrent global-adaptive network for visionand-language navigation",
"year": "2023"
},
{
"authors": "X Wang; Q Huang; A Celikyilmaz; J Gao; D Shen; Y.-F Wang; W Y Wang; L Zhang",
"journal": "",
"ref_id": "b28",
"title": "Reinforced cross-modal matching and selfsupervised imitation learning for vision-language navigation",
"year": "2019"
},
{
"authors": "J Chen; C Gao; E Meng; Q Zhang; S Liu",
"journal": "",
"ref_id": "b29",
"title": "Reinforced structured state-evolution for vision-language navigation",
"year": "2022"
},
{
"authors": "Q Sun; Y Zhuang; Z Chen; Y Fu; X Xue",
"journal": "IEEE",
"ref_id": "b30",
"title": "Depth-guided adain and shift attention network for vision-and-language navigation",
"year": "2021"
},
{
"authors": "F Zhu; Y Zhu; X Chang; X Liang",
"journal": "",
"ref_id": "b31",
"title": "Vision-language navigation with self-supervised auxiliary reasoning tasks",
"year": "2020"
},
{
"authors": "C.-Y Ma; J Lu; Z Wu; G Alregib; Z Kira; R Socher; C Xiong",
"journal": "",
"ref_id": "b32",
"title": "Self-monitoring navigation agent via auxiliary progress estimation",
"year": "2019"
},
{
"authors": "Y Zhao; J Chen; C Gao; W Wang; L Yang; H Ren; H Xia; S Liu",
"journal": "",
"ref_id": "b33",
"title": "Target-driven structured transformer planner for visionlanguage navigation",
"year": "2022"
},
{
"authors": "T.-J Fu; X E Wang; M F Peterson; S T Grafton; M P Eckstein; W Y Wang",
"journal": "Springer",
"ref_id": "b34",
"title": "Counterfactual vision-and-language navigation via adversarial path sampler",
"year": "2020"
},
{
"authors": "B Lin; Y Zhu; Y Long; X Liang; Q Ye; L Lin",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b35",
"title": "Adversarial reinforced instruction attacker for robust vision-language navigation",
"year": "2021"
},
{
"authors": "C Liu; F Zhu; X Chang; X Liang; Z Ge; Y.-D Shen",
"journal": "",
"ref_id": "b36",
"title": "Visionlanguage navigation with random environmental mixup",
"year": "2021"
},
{
"authors": "M Zhao; P Anderson; V Jain; S Wang; A Ku; J Baldridge; E Ie",
"journal": "Association for Computational Linguistics",
"ref_id": "b37",
"title": "On the evaluation of vision-and-language navigation instructions",
"year": "2021"
},
{
"authors": "Z.-Y Dou; N Peng",
"journal": "",
"ref_id": "b38",
"title": "Foam: A follower-aware speaker model for vision-and-language navigation",
"year": "2022"
},
{
"authors": "Z.-J Zha; D Liu; H Zhang; Y Zhang; F Wu",
"journal": "IEEE transactions on pattern analysis and machine intelligence",
"ref_id": "b39",
"title": "Context-aware visual policy network for fine-grained image captioning",
"year": "2019"
},
{
"authors": "C Yan; Y Tu; X Wang; Y Zhang; X Hao; Y Zhang; Q Dai",
"journal": "IEEE Transactions on Multimedia",
"ref_id": "b40",
"title": "Stat: Spatial-temporal attention mechanism for video captioning",
"year": "2020"
},
{
"authors": "H Zhong; J Chen; C Shen; H Zhang; J Huang; X.-S Hua",
"journal": "IEEE Transactions on Multimedia",
"ref_id": "b41",
"title": "Selfadaptive neural module transformer for visual question answering",
"year": "2021"
},
{
"authors": "X Xiao; L Wang; K Ding; S Xiang; C Pan",
"journal": "IEEE Transactions on Multimedia",
"ref_id": "b42",
"title": "Deep hierarchical encoder-decoder network for image captioning",
"year": "2019"
},
{
"authors": "P Zhou; W Shi; J Tian; Z Qi; B Li; H Hao; B Xu",
"journal": "",
"ref_id": "b43",
"title": "Attentionbased bidirectional long short-term memory networks for relation classification",
"year": "2016"
},
{
"authors": "J Devlin; M.-W Chang; K Lee; K Toutanova",
"journal": "Association for Computational Linguistics",
"ref_id": "b44",
"title": "BERT: Pre-training of deep bidirectional transformers for language understanding",
"year": "2019"
},
{
"authors": "T B Brown; B Mann; N Ryder; M Subbiah; J Kaplan; P Dhariwal; A Neelakantan; P Shyam; G Sastry; A Askell; S Agarwal; A Herbert-Voss; G Krueger; T Henighan; R Child; A Ramesh; D M Ziegler; J Wu; C Winter; C Hesse; M Chen; E Sigler; M Litwin; S Gray; B Chess; J Clark; C Berner; S Mccandlish; A Radford; I Sutskever; D Amodei",
"journal": "Curran Associates Inc",
"ref_id": "b45",
"title": "Language models are few-shot learners",
"year": "2020"
},
{
"authors": "L Zhou; Y Zhou; J J Corso; R Socher; C Xiong",
"journal": "",
"ref_id": "b46",
"title": "End-to-end dense video captioning with masked transformer",
"year": "2018"
},
{
"authors": "L Guo; J Liu; X Zhu; P Yao; S Lu; H Lu",
"journal": "",
"ref_id": "b47",
"title": "Normalized and geometry-aware self-attention network for image captioning",
"year": "2020"
},
{
"authors": "Y Luo; J Ji; X Sun; L Cao; Y Wu; F Huang; C.-W Lin; R Ji",
"journal": "",
"ref_id": "b48",
"title": "Duallevel collaborative transformer for image captioning",
"year": "2021"
},
{
"authors": "V Veeriah; M Hessel; Z Xu; J Rajendran; R L Lewis; J Oh; H P Van Hasselt; D Silver; S Singh",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b49",
"title": "Discovery of useful questions as auxiliary tasks",
"year": "2019"
},
{
"authors": "J Gu; H Zhao; Z Lin; S Li; J Cai; M Ling",
"journal": "",
"ref_id": "b50",
"title": "Scene graph generation with external knowledge and image reconstruction",
"year": "2019"
},
{
"authors": "T Trinh; A Dai; T Luong; Q Le",
"journal": "PMLR",
"ref_id": "b51",
"title": "Learning longer-term dependencies in rnns with auxiliary losses",
"year": "2018"
},
{
"authors": "H Huang; V Jain; H Mehta; A Ku; G Magalhaes; J Baldridge; E Ie",
"journal": "",
"ref_id": "b52",
"title": "Transferable representation learning in vision-and-language navigation",
"year": "2019"
},
{
"authors": "R Pascanu; T Mikolov; Y Bengio",
"journal": "PMLR",
"ref_id": "b53",
"title": "On the difficulty of training recurrent neural networks",
"year": "2013"
},
{
"authors": "K He; X Zhang; S Ren; J Sun",
"journal": "",
"ref_id": "b54",
"title": "Deep residual learning for image recognition",
"year": "2016"
},
{
"authors": "J L Ba; J R Kiros; G E Hinton",
"journal": "",
"ref_id": "b55",
"title": "Layer normalization",
"year": "2016"
},
{
"authors": "Y Hong; C Rodriguez; Q Wu; S Gould",
"journal": "",
"ref_id": "b56",
"title": "Sub-instruction aware vision-and-language navigation",
"year": "2020"
},
{
"authors": "P Baldi; P Sadowski",
"journal": "Artificial intelligence",
"ref_id": "b57",
"title": "The dropout learning algorithm",
"year": "2014"
},
{
"authors": "V Mnih; A P Badia; M Mirza; A Graves; T Lillicrap; T Harley; D Silver; K Kavukcuoglu",
"journal": "PMLR",
"ref_id": "b58",
"title": "Asynchronous methods for deep reinforcement learning",
"year": "2016"
},
{
"authors": "Y Hong; C Rodriguez; Y Qi; Q Wu; S Gould",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b59",
"title": "Language and visual entity relationship graph for agent navigation",
"year": "2020"
},
{
"authors": "Y Qiao; Y Qi; Y Hong; Z Yu; P Wang; Q Wu",
"journal": "",
"ref_id": "b60",
"title": "Hop: Historyand-order aware pre-training for vision-and-language navigation",
"year": "2022"
},
{
"authors": "K Papineni; S Roukos; T Ward; W.-J Zhu",
"journal": "",
"ref_id": "b61",
"title": "Bleu: a method for automatic evaluation of machine translation",
"year": "2002"
},
{
"authors": "C.-Y Lin",
"journal": "Text summarization branches out",
"ref_id": "b62",
"title": "Rouge: A package for automatic evaluation of summaries",
"year": "2004"
},
{
"authors": "R Vedantam; C Lawrence Zitnick; D Parikh",
"journal": "",
"ref_id": "b63",
"title": "Cider: Consensusbased image description evaluation",
"year": "2015"
},
{
"authors": "P Anderson; B Fernando; M Johnson; S Gould",
"journal": "Springer",
"ref_id": "b64",
"title": "Spice: Semantic propositional image caption evaluation",
"year": "2016"
},
{
"authors": "A Ghaderi; A A Shahri; S Larsson",
"journal": "Catena",
"ref_id": "b65",
"title": "A visualized hybrid intelligent model to delineate swedish fine-grained soil layers using clay sensitivity",
"year": "2022"
},
{
"authors": "A Dosovitskiy; L Beyer; A Kolesnikov; D Weissenborn; X Zhai; T Unterthiner; M Dehghani; M Minderer; G Heigold; S Gelly",
"journal": "",
"ref_id": "b66",
"title": "An image is worth 16x16 words: Transformers for image recognition at scale",
"year": "2020"
},
{
"authors": "A Radford; J W Kim; C Hallacy; A Ramesh; G Goh; S Agarwal; G Sastry; A Askell; P Mishkin; J Clark",
"journal": "PMLR",
"ref_id": "b67",
"title": "Learning transferable visual models from natural language supervision",
"year": "2021"
},
{
"authors": "Y Gal; Z Ghahramani",
"journal": "PMLR",
"ref_id": "b68",
"title": "Dropout as a bayesian approximation: Representing model uncertainty in deep learning",
"year": "2016"
},
{
"authors": "A Shahri; C Shan; S Larsson",
"journal": "Natural Resources Research",
"ref_id": "b69",
"title": "A novel approach to uncertainty quantification in groundwater table modeling by automated predictive deep learning",
"year": "2022"
},
{
"authors": "R Asheghi; S A Hosseini; M Saneie; A A Shahri",
"journal": "Journal of Hydroinformatics",
"ref_id": "b70",
"title": "Updating the neural network sediment load models using different sensitivity analysis methods: a regional application",
"year": "2020"
}
] | [
{
"formula_coordinates": [
4,
51.31,
256.84,
237.37,
77.24
],
"formula_id": "formula_0",
"formula_text": "= { 1 , 2 , ..., } ∈ ℝ × denotes image features with the angle set = { 1 , 2 , ..., } ∈ ℝ × . Similarly, the environment-level features = { ; } is composed of the panoramic image set = {[ , ] =1 } =1 ∈ ℝ × × and the angle set = {[ , ] =1 } =1 ∈ ℝ × ×"
},
{
"formula_coordinates": [
4,
75.28,
459.29,
213.39,
22.81
],
"formula_id": "formula_1",
"formula_text": "( 1 , ..., | , ) = ∏ =1 ( | 1 , ..., -1 , , )(1)"
},
{
"formula_coordinates": [
4,
69.86,
494.5,
218.81,
22.81
],
"formula_id": "formula_2",
"formula_text": "( 1 , ..., | , ) = ∏ =1 ( | 1 , ..., -1 , , )(2)"
},
{
"formula_coordinates": [
4,
392.96,
531.2,
151.01,
11.01
],
"formula_id": "formula_3",
"formula_text": "̃ = (Dropout( ); )(3)"
},
{
"formula_coordinates": [
4,
392.19,
548.32,
151.78,
11.01
],
"formula_id": "formula_4",
"formula_text": "̃ = (Dropout( ); )(4)"
},
{
"formula_coordinates": [
4,
315.56,
565.43,
228.41,
70.96
],
"formula_id": "formula_5",
"formula_text": "̃ = ̃ , ̃ = ̃ , ̃ = ̃ (5) head = softmax( ̃ ̃ √ ) ̃ (6) MultiHead( , , ) = Concat(head 1 , ..., head )(7)"
},
{
"formula_coordinates": [
4,
400.58,
640.56,
143.39,
11.01
],
"formula_id": "formula_6",
"formula_text": "= LayerNorm(MultiHead + ̃ ) (8)"
},
{
"formula_coordinates": [
4,
325.33,
664.01,
218.64,
25.52
],
"formula_id": "formula_7",
"formula_text": "∈ ℝ × , ∈ ℝ × , ∈ ℝ × , ∈ ℝ × , ∈ ℝ × , ∈ ℝ × ."
},
{
"formula_coordinates": [
5,
100.04,
330.01,
188.63,
25.8
],
"formula_id": "formula_8",
"formula_text": "0 = ( ) (9) ′ = LayerNorm(MSA( -1 ) + -1 )(10)"
},
{
"formula_coordinates": [
5,
107.04,
358.86,
181.63,
11.08
],
"formula_id": "formula_9",
"formula_text": "= LayerNorm(FFN( ′ ) + ′ )(11)"
},
{
"formula_coordinates": [
5,
90.31,
580.77,
198.37,
22.81
],
"formula_id": "formula_10",
"formula_text": " SWP = - ∑ =1 log( ( * | * 1∶ -1 , , ))(12)"
},
{
"formula_coordinates": [
5,
406.02,
539.87,
137.94,
12.4
],
"formula_id": "formula_11",
"formula_text": "( ∈ ′ ) =(13)"
},
{
"formula_coordinates": [
5,
369.38,
678.81,
174.59,
13.91
],
"formula_id": "formula_12",
"formula_text": "′ , = ReLU(ℎ , + )(14)"
},
{
"formula_coordinates": [
5,
371.21,
696.17,
172.76,
13.91
],
"formula_id": "formula_13",
"formula_text": ", = Dropout( ′ , ) +(15)"
},
{
"formula_coordinates": [
6,
107.17,
357.22,
181.5,
25.39
],
"formula_id": "formula_14",
"formula_text": " SPM = - 1 2 ∑ =1 ( * -, ) 2(16)"
},
{
"formula_coordinates": [
6,
104.59,
439.2,
179.94,
10.84
],
"formula_id": "formula_15",
"formula_text": " = SWP + (1 -) SPM (17"
},
{
"formula_coordinates": [
6,
284.52,
439.2,
4.15,
8.9
],
"formula_id": "formula_16",
"formula_text": ")"
},
{
"formula_coordinates": [
6,
382.07,
710.11,
161.89,
22.81
],
"formula_id": "formula_17",
"formula_text": " IL = - ∑ =1 log ( * )(18)"
},
{
"formula_coordinates": [
7,
119.73,
426.23,
168.95,
22.81
],
"formula_id": "formula_18",
"formula_text": " RL = - ∑ =1 log ( ̂ ℎ )Λ(19)"
}
] |
10.1162/tacl_a_00416 | 2023-05-24 | [
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b34",
"b15",
"b21",
"b72",
"b4",
"b9",
"b5",
"b37",
"b64",
"b36",
"b29",
"b67",
"b13",
"b4"
],
"table_ref": [],
"text": "The development of natural language processing (NLP) technology that serves most of world's languages is hindered by the stark lack of data for most languages (Joshi et al., 2020). While there is increasing interest in developing datasets and models for under-represented languages (ULs), existing datasets are often informed by established research directions in the NLP community (de Marneffe et al., 2021). While linguistic tasks such as syntactic parsing have become less practically relevant (Glavaš and Vulić, 2021), other tasks such as news summarization or sentiment analysis are informed by the availability of data in high-resource language settings and may be less useful for speakers of ULs (Varab and Schluter, 2021;Muhammad et al., 2022). Impactful capabilities such as question answering or virtual assistants (Asai et al., 2021), on the other hand, often depend on ancillary technologies such as language ID, data filtering, automatic speech recognition (ASR), or optical character recognition (OCR) that are typically underperforming or unavailable for ULs (Caswell et al., 2020;Bapna et al., 2022;Kreutzer et al., 2022;Rijhwani et al., 2021;Khare et al., 2021). As a result, speakers of ULs will not be able to reap the benefits of such capabilities, even if the development of models is successful.\nIn order to make progress on NLP for ULs, we should thus focus on building datasets and evaluating models on tasks that are most likely to benefit speakers of those languages. 3 To this end, we propose XTREME-UP (Under-Represented and User-Centric with Paucal4 Data), a benchmark focusing on evaluation of multilingual models on usercentric tasks in a few-shot setting.\nWe focus on tasks that technology users encounter regularly in their daily lives: i) information access tasks, which represent generally useful NLP capabilities; and ii) input/output tasks that enable other technologies. We show the corresponding tasks and their role in typical interactions with language technology in Figure 1. Moving away from the standard cross-lingual zero-shot setting (Hu et al., 2020;Ruder et al., 2021), we introduce a standardized multilingual in-language fine-tuning setting based on the amount of data that can realistically be annotated or generated within 8h for a language.\nOur results highlight the limitations of current models on ULs, demonstrate the potential of language models (LMs) to improve user-centric applications, and show the benefit of byte-based approaches, among other findings.\nIn this work, we contribute the first massivelymultilingual few-example benchmark including: a) newly created data for QA, OCR, autocomplete, semantic parsing, and sentence-level transliteration; b) new task setups for named entity recognition (NER) enabling evaluation on natural-rather than tokenized-text; and for QA and retrieval providing a more interesting setting than the gold passage (GoldP) setup while offering a lower barrier-toentry than the full TyDi QA Clark et al. (2020) or XOR (Asai et al., 2021) tasks; c) carefullydesigned experimental setups, standardizing inlanguage fine-tuning and in-context learning and focusing on the information access scenario for ULs for ASR and MT; d) baseline results for all datasets on commonly-used subword and byte-based models."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b15",
"b27",
"b75",
"b50",
"b29",
"b43",
"b67"
],
"table_ref": [],
"text": "Multilingual benchmarks Some studies employ highly multilingual individual datasets for the evaluation of multilingual models, including Universal Dependencies (de Marneffe et al., 2021) or XL-Sum (Hasan et al., 2021). At the same time, there is increasing work on datasets in ULs for a variety of applications (Niyongabo et al., 2020;Winata et al., 2023;Muhammad et al., 2023). Due to their rapidly growing capabilities, NLP models are increasingly evaluated on a suite of datasets. Existing multitask multilingual benchmarks such as XTREME (Hu et al., 2020), XGLUE (Liang et al., 2020), and XTREME-R (Ruder et al., 2021) cover 20-50 mainly high-resource languages and prioritize tasks with available data, regardless of their utility to speakers. In contrast, XTREME-UP focuses on under-represented languages and user-centric tasks, creating new data for under-represented tasks and languages."
},
{
"figure_ref": [],
"heading": "Multilingual evaluation",
"publication_ref": [
"b29",
"b3",
"b40",
"b28",
"b8",
"b35"
],
"table_ref": [],
"text": "The choice of the experimental setting and aggregation metric are important considerations in multilingual evaluation. Prior work focused on zero-shot cross-lingual transfer (Hu et al., 2020), which-despite being compelling from a scientific perspective (Artetxe et al., 2020)is less practically useful. While in-language finetuning has been explored before (Lauscher et al., 2020;Hedderich et al., 2020), XTREME-UP is the first to standardize the setting across tasks based on realistic annotation costs. Different frameworks aggregate performance in different ways across languages. Blasi et al. (2022) assess the utility of a task by weighting model performance based on the size of the speaker population while Khanuja et al. (2023) introduce the Gini coefficient to quantify performance disparity across languages. XTREME-UP opts for a simple average over ULs, emphasizing intuitiveness and accessibility of the results."
},
{
"figure_ref": [],
"heading": "XTREME-UP",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Design Principles",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [
"fig_0"
],
"heading": "XTREME-UP is motivated by the following design principles:",
"publication_ref": [
"b34"
],
"table_ref": [],
"text": "Table 1: The tasks in XTREME-UP. For each task, we show both the sum of training examples across all languagesto give some insight into training scale-and the average number of training examples for each under-represented language-to highlight the challenge of the scarce-data learning scenario. XTREME-UP does not limit supervised training data in high-resource languages (HLs) while each under-represented language (UL) has a maximum of 8 hours of annotation effort in its training split; see last column for estimated annotation effort. We also show the sum of validation and test examples across ULs as XTREME-UP evaluates only on ULs.\nUnder-represented languages We follow the ontology of Joshi et al. (2020) in defining ULs based on available data. Specifically, we select languages in categories 1-3 (e.g., Amharic, Estonian, Kinyarwanda) as under-represented, leaving categories 4-5 as high-resource languages (e.g., English, German, Hindi). We focus on tasks with existing data in ULs and tasks where we can efficiently collect such data at scale (see Appendix A for an overview of ULs in XTREME-UP).\nUser-centric tasks We focus on widely adopted user-facing tasks benefiting speakers of highresource languages. We further break these down into two major groups: 1) input/output tasks; and 2) information access tasks (see Figure 1)."
},
{
"figure_ref": [],
"heading": "Scarce data",
"publication_ref": [],
"table_ref": [],
"text": "We focus on a realistic scenario where a small amount of data is available in each UL. Mirroring reality, we do not restrict the amount of training data available in high-resource languages, but rather provide only as many labeled training examples as can be annotated in a realistic amount of time for ULs (see Section 3.2).\nEfficiency We focus on massively multilingual evaluation settings that can still be run efficiently with a modest amount of compute.\nText-centric, yet multi-modal We focus on tasks that can be tackled using textual data alone and provide baseline systems that do so. We frame multimodal tasks (OCR and ASR) so that natively multimodal models can be evaluated fairly alongside text-only models. We accomplish this by releasing original audio, image, and text model inputs while also providing baseline system output that can be fed to second-stage text-only systems. We hope to see fully multi-modal models take up this challenge over the coming years.\nWe provide an overview of the tasks in XTREME-UP in Table 1. We discuss motivation and highlevel information in the next section and provide more details for each task in Appendix B."
},
{
"figure_ref": [],
"heading": "How much data?",
"publication_ref": [],
"table_ref": [],
"text": "To ensure a realistic amount of training data, we limit the training data in each task per language to the number of examples that can be annotated in 8 hours. We believe this reflects the real difficulty of annotating training and evaluation data for a very large number of languages. In this way, we design for the task first and will let the research to develop technology that addresses these challenges follow. For each task, we estimate how long it takes to annotate a single example for a trained annotator. 5We base our estimates on prior work and our own annotation efforts. 6 We show the data annotation time estimates in Table 1. For tasks with larger training datasets, we sub-sample the available data accordingly. Table 1 shows the sub-sampled data sizes. We show an example instance of each task in Table 2."
},
{
"figure_ref": [],
"heading": "Input / Output Tasks",
"publication_ref": [
"b14",
"b23",
"b63",
"b68",
"b49",
"b52",
"b62",
"b63",
"b30",
"b62",
"b26",
"b63"
],
"table_ref": [],
"text": "Automatic speech recognition (ASR; B.1) The goal of ASR is to transcribe speech into humanreadable text. It thus serves as a fundamental step for enabling natural language understanding applications on speech input. In many scenarios, users may strongly prefer to speak rather than type and so high-quality ASR is an enabling factor for such user interactions. We employ the FLEURS dataset (Conneau et al., 2023) consisting of recordings in 102 languages for sentences from FLORES-101 (Goyal et al., 2022), which were translated from English Wikipedia to 101 languages. We evaluate on the under-represented portion of the data, which covers 77 languages.\nOptical character recognition (OCR; B.2) OCR, the process of converting text from images into machine-readable formats, is used in a wide range of applications, from extracting language data locked in paper books (Rijhwani et al., 2020) and imaging legal documents (Singh et al., 2012), to improving accessibility for people with low vision or blindness (Mowar et al., 2022). It is especially important for under-represented languages where both training data and content that users may wish to access may not be abundant as digital text on the web. While most existing datasets focus on higher-resourced languages (Nayef et al., 2017;Rigaud et al., 2019), there has been recent interest in developing OCR for ULs. This includes the creation of a small dataset for endangered languages (Rijhwani et al., 2020) and a synthetic dataset for 60 languages (Ignat et al., 2022).\nWe create a dataset that aims to fill the gaps and augment previous work in OCR for ULs, by focusing on larger-scale, typologically-diverse, and user-centric data. Our dataset contains transcriptions for books in seven languages: Amharic (am), Bengali (bn), Kannada (kn), Myanmar (Burmese; my), Sanksrit(sa), Sinhala (si), and Swahili (sw). The books domain is the primary use-case for a large number of downstream users, but is one of the most challenging for OCR models (Rigaud et al., 2019). The dataset consists of transcriptions of entire pages and thus enables leveraging the full context understanding capabilities of large language models. To demonstrate these capabilities, we use the approach of \"OCR post-correction\": training language models to correct recognition errors in transcriptions from existing OCR systems (Hammarström et al., 2017;Rijhwani et al., 2020)."
},
{
"figure_ref": [],
"heading": "Autocomplete (B.",
"publication_ref": [
"b2",
"b69",
"b46",
"b70",
"b15",
"b77"
],
"table_ref": [],
"text": "3) Autocomplete (or predictive text), i.e., predicting the rest of a word a user is typing, is a useful technology that speeds up human-computer interaction (Anson et al., 2006). As such, autocomplete has become a technology that users have come to expect and rely on for input in high-resource languages. The standard next word prediction task (Sundermeyer et al., 2012) does not accurately reflect this practical setting as it relies on predicting entire units (words, subwords, or characters); similarly, perplexity-based evaluation makes comparisons across segmentations and languages difficult (Mielke, 2019) while ignoring important threshold effects associated with the typical top-k predictions in a user interface (Tam and Wells, 2009).\nTo fill this gap, we introduce a new autocomplete task that unifies character, subword, and tokenlevel LM settings by focusing on a \"word\" as the predictive unit. Models are required to complete the next word based on a left context of N words and an optional character n-gram prefix. We use accuracy@3 for evaluation to reflect the requirement of displaying a limited number of candidates to the user. We process high-quality natural language data from Universal Dependencies (de Marneffe et al., 2021), which we deduplicate against mC4 (Xue et al., 2021), the most common multilingual pretraining corpus in order to test models predictive rather than memorization capabilities."
},
{
"figure_ref": [],
"heading": "Transliteration (B.4)",
"publication_ref": [
"b73",
"b65",
"b25",
"b23"
],
"table_ref": [],
"text": "Transliteration is the conversion of text between writing systems (Wellisch, 1978). Unlike translation, it does not change content but only script. For example, the Hindi sentence (\"the thing is blue\") might be written \"vastu neela hai\" in the Latin script (which is often called romanization). 7 Transliteration is important because it allows users to type in their preferred script (e.g., Latin script) even if it is different than their preferred display script (e.g. Devanagari) and is used internally by many machine translation systems to rewrite names from different scripts. Table 2: Examples of each task in XTREME-UP. The tasks are generally text-in, text-out with a few exceptions. On the output side, autocomplete requires generating the top-3 outputs and retrieval outputs document identifierscurrent systems tend to implement retrieval by mapping both inputs and candidate outputs to vector and performing nearest neighbor lookup. On the input side, speech recognition has audio input and document OCR has image outputs; our initial baseline systems use external systems to map this to text as a preprocessing step, though we hope to see multi-modal systems eliminate this step in the near future.\nWe extend the Dakshina dataset (Roark et al., 2020), which provides romanizations of Wikipedia sentences written in the native scripts of 12 South Asian languages. To this data, we added: a) romanizations of native script Wikipedia for one new language (Amharic); and b) transliteration to a third script (Shahmukhi) for one already covered language (Punjabi). The resulting task covers 13 languages from three language families. For all these languages transliteration occurs from the Latin script to the native script of the language, and vice versa and between Shahmukhi (Perso-Arabic), Gurmukhi (Brahmic), and Latin for Punjabi, leading to a total of 30 transliteration directions.\nMachine translation (MT; App. B.5) MT is an important technology for users of ULs wishing to read text written in a different language. However, most current approaches require large amounts of parallel training data to achieve good performance, which are often not available for ULs (Haddow et al., 2022). We focus on the information dissemination scenario where content from highresource languages (including from tasks such as cross-lingual QA) is translated to enable information access by common users; as such, XTREME-UP includes translations from English into 93 languages, covering a wide range of high-resource and UL languages. Only 39 ULs are used for evaluation; the high-resource languages are included to allow for transfer learning. 8 The dataset is adapted from FLORES-101 (Goyal et al., 2022), repurposing half of the dataset's original development set as a training set. See §6 for a detailed discussion of how we distinguish freely-available unsupervised data versus purpose-annotated supervised data in XTREME-UP."
},
{
"figure_ref": [],
"heading": "Information Access Tasks",
"publication_ref": [
"b39",
"b13",
"b4",
"b13",
"b1",
"b0",
"b71",
"b41",
"b20",
"b48",
"b22",
"b41",
"b44",
"b17",
"b45",
"b11"
],
"table_ref": [],
"text": "Question Answering (B.6) Question answering is an important capability that enables responding to natural language questions with answers found in text (Kwiatkowski et al., 2019). We focus on the information-seeking scenario where questions are asked (and therefore written by dataset annotators) without knowing the answer-it is the system's job to locate a suitable answer passage (if any); this is in contrast to the school-like reading comprehension scenario where questions are written while looking at text, which is guaranteed to con-tain the answer. Importantly, information-seeking question-answer pairs tend to exhibit less lexical and morphosyntactic overlap between the question and answer since they are written separately.\nWe include two variants of the task: in the inlanguage QA task, both the question and passage are in the same language. In this task, original questions and passages are from the TyDi QA dataset (Clark et al., 2020). In the cross-language QA task, the question is in the user's native language while the passage and answer are in a highresource language having a large amount of available answer content (English). For this task, we use examples from TyDi XOR (Asai et al., 2021) in 7 languages. We additionally collect new data in 23 new Indic languages for cross-lingual QA by professionally translating questions and answers from existing Indic languages in XOR QA. This methodology mitigates the issue of translating Westerncentric English data to locales with different topical interests. Cross-lingual QA is especially important for ULs since they may lack plentiful in-language answer content on the web.\nIn XTREME-UP's QA task, a system is given a question, title, and a passage and must provide the answer-if any-or otherwise return that the question has \"no answer\" in the passage. 9 To this end, we generalize the gold passage (Clark et al., 2020) setting, augmenting it with negative examples. These negatives are obtained from (a) passages within the same article as a passage containing the answer and (b) question-answer pairs from the full TyDi QA dataset where no answer was found in the candidate Wikipedia article. The data is split into training, validation, and test splits in such a way to avoid deduplication and overlap of splits, even across our various QA tasks. 10Retrieval for QA (B.6) Within the informationseeking QA scenario, the above core QA task assumes answer candidate passages as an input. In practice, a passage retrieval system for questionanswering allows for the extraction of relevant text from a vast text corpus. The retrieved passages can then be used by a question-answering system to extract or generate an answer to the user's question. In XTREME-UP, we separate retrieval into two distinct tasks, in-language retrieval and crosslanguage retrieval. For in-language retrieval, both the questions and passages are in the same language. The preparation of negatives, deduplication, and splits are identical to the QA task above. For validation and test, we create an index of 271k inlanguage passages (447k English passages for the cross-language task) making for a small enough index for efficient experimentation, while containing distractors that make for a challenging task, since these distractors are drawn from the same articles containing the target passages.\nNamed entity recognition (NER; B.7) NER is an important capability for information access systems that users depend on with applications ranging from recognizing requests for entity lookups to performing information extraction to populate the knowledge graphs that handle those requests. NER is also a capability needed in spell-checking and localization systems (Li et al., 2020). 11 Identifying entities in ULs poses challenges due to the use of different scripts, lack of capitalization, different numerical representations, etc. We build on MasakhaNER (Adelani et al., 2021) and MasakhaNER 2.0 (Adelani et al., 2022), two large NER datasets in African languages, which provide data in the standard CoNLL tokenized format (Tjong Kim Sang and De Meulder, 2003). In order to enable evaluation in a setting that is closer to the real world, we automatically map the annotated spans to the original raw text. The combined data with byte-level span annotationstermed MasakhaNER-X-covers 20 languages. 12Semantic parsing (App. B.8) Semantic parsing is the task of mapping a natural language utterance to a logical form or a structured interpretation that can be executed by a system such as a virtual assistant. For example a user utterance can be classified into an intent and parsed into slots: \"wake me at 8 am\" would be mapped to the \"CreateAlarm\" intent and would have a single \"time\" slot with \"8 am\" as value. Then the assistant may use this interpretation to create an alarm at the specified time. While modern models are becoming very capable of responding to users' language inputs, we believe this task is especially timely as users will increasingly want to turn their interactions with assistants and chat-like dialog systems into actions on external systems, which require API calls; this capability is what the semantic parsing task evaluates.\nRecently, researchers published more multilingual semantic parsing datasets that focus on virtual assistant domains (Li et al., 2021;FitzGerald et al., 2022;Moghe et al., 2022;Goel et al., 2023). We extend a portion of an existing semantic parsing dataset to new languages targeting the following features: a) high-quality utterances produced by professional translators; b) a wide range of domains and intents; c) inclusion of different language families and some underrepresented languages; d) sentences with culturally relevant entities; and e) codemixed sentences, i.e., multiple language within the same sentence-a common phenomenon in multilingual societies.\nWe adapt the test split of MTOP13 (Li et al., 2021) with professional translators/annotators to the following 15 languages: Amharic, Belarusian, Bengali, Brazilian Portuguese, Finnish, German, Hausa, Hungarian, Japanese, Russian, Swahili, Tamil, Turkish, Yoruba, and Zulu. Together with the original MTOP languages, the new MTOP++ dataset covers a total of 20 languages. The data we collect, differently from MTOP, is localized (i.e., Western-centric entities are replaced with more culturally relevant entities for the target language), following recent trends in multilingual benchmarking (Lin et al., 2021;Ding et al., 2022;Majewska et al., 2023).\nWe also extend MTOP to three widely spoken but under-represented Indic languages in a codeswitching setting: Hindi-English, Bengali-English and Tamil-English. We automatically convert the test-split of MTOP to code-mixed utterances using PaLM (Chowdhery et al., 2022) and run human verification on such utterances."
},
{
"figure_ref": [],
"heading": "Overall Evaluation",
"publication_ref": [
"b60"
],
"table_ref": [],
"text": "For each task, we evaluate model performance by computing a task-specific score. We employ character-level metrics such as character error rate (CER) and character n-gram F-score (chrF; Popović, 2015) rather than their word-level counterparts as they enable more fine-grained evaluation and are better suited to morphologically rich lan-guages. We obtain a final score by averaging the scores of all tasks. For each task, we only average performance over ULs (discussed in §3.1). For metrics such as character error rate (CER) where lower is better, we invert the scores before averaging scores across tasks. For mean reciprocal rank (MRR), which is in the 0.0-1.0 range, we renormalize it to the 0-100 range before averaging. While this scalar provides a quick overall impression of a system's quality across a broad range of tasks, it is not a substitute for analyzing performance on individual tasks, languages, or types of examples."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Experimental setting",
"publication_ref": [],
"table_ref": [],
"text": "Multilingual fine-tuning In contrast to prior benchmarks that focus on zero-shot cross-lingual transfer from English, XTREME-UP focuses on the more realistic scenario of fine-tuning on a small amount of data in the target language. To make this scenario scalable in a massively multilingual setting, XTREME-UP fine-tunes a single model on the combined training data across the available languages for each task. The data for each language is sub-sampled to emulate data sizes that can be realistically annotated within a reasonable time frame (see §3.2).\nIn-language in-context learning We also provide a 5-shot in-context learning setting where a model is provided with an English instruction and 5 exemplars in the target language in order to evaluate the progress on few-shot learning with large models for ULs. We provide the instruction for each task in Appendix C.14 "
},
{
"figure_ref": [],
"heading": "Baselines",
"publication_ref": [
"b11",
"b10",
"b6",
"b18",
"b63",
"b66",
"b38",
"b59"
],
"table_ref": [],
"text": "We provide results on a handful of baseline systems that have already been developed by the research community. Given that our focus in this paper is on the dataset and task setup rather than system building, we do not focus on offering novel modeling types nor do we exhaustively evaluate all possible models; rather we view these results as estimating a starting point from some well-known modeling approaches and seeding contributions from the et al., 2022), an instruction-tuned version of PaLM (Chowdhery et al., 2022). We provide additional information on the baseline systems in Table 3.\nTo offer baseline systems that allow experimentation with text-only models, we use upstream models to provide initial output for ASR and OCR, and present text-based baselines that use these as inputs. We expect these baselines to give way to fully multi-modal models as research progresses. These initial ASR and OCR outputs should be seen as part of a baseline system, not part of the XTREME-UP benchmark iteself. For ASR, we augment the data with predictions of the state-of-the-art Maestro-U (Chen et al., 2023) and then use a downstream text model to improve the outputs (Bassil and Alwani, 2012). Similarly, for OCR, we use the off-the-shelf Google Vision OCR16 to get first-pass outputs, and train language models to improve them (Dong and Smith, 2018;Rijhwani et al., 2020).\nInfrastructure Models were trained using seqio and T5X (Roberts et al., 2022) on TPUs (Kumar et al., 2019;Pope et al., 2022). 22.9 (20.9 / 24.9) -12.9 0.1\nTable 4: Overall results of baselines across all XTREME-UP v1.0 tasks for the test split. Scores on XTREME-UP average over evaluation scores of under-represented languages. QA and retrieval performance is the average of in-language and cross-language settings (indicated in brackets as in-language / cross-language). For OCR, we do not apply any additional models (mT5 nor ByT5) on top of the baseline OCR system; we show these results in parentheses. We do not attempt in-context learning (ICL) results for retrieval since ICL is typically only used for text-in, text-out use cases. ⋆ For OCR, we use the Google OCR API. † For autocomplete, while we observe reasonable performance on English completions, we find that the model typically does a very poor job outside of English."
},
{
"figure_ref": [],
"heading": "Results",
"publication_ref": [
"b19"
],
"table_ref": [],
"text": "We show the baseline results in Table 4. 17\nByte-based models outperform subword-based on ULs. The byte-based ByT5 outperforms the subword-based mT5 across most tasks. Gains are particularly pronounced for tasks that require dealing with information on the character level such as autocomplete and transliteration and for predicting information on the word level such as for NER and semantic parsing. These results demonstrate that as we train and evaluate our models on underrepresented languages, standard modeling choices such as subword representations fall short.\nIn-context learning underperforms fine-tuning on limited data. The Flan-PaLM model generally performs worse than the models using finetuning, despite being much larger. Nevertheless, it achieves reasonable performance on machine translation, which is likely reflected in the pre-training data. On other tasks, however, it fails to reliably apply its English-centric knowledge to ULs. Despite fine-tuned models performing relatively well on NER, the in-context learning model is unable to consistently generalize to the task in a few-shot setting in under-represented languages. On semantic parsing, the model fails to generalize to the large number of domain-specific intents and slots using standard prompting in ULs. 18 The autocomplete 17 Detailed per-language results are available at https:// github.com/google-research/xtreme-up. 18 We leave the exploration of multilingual adaptive prompting and dynamic exemplar selection (Drozdov et al., 2023) methods to future work. tasks in particular demonstrate the lack of robust cross-lingual information in the English-centric PaLM model: it struggles to complete a sentence given a character prefix and fails to reliably convert between different scripts in the same language. XTREME-UP thus provides a strong challenge to test the generalization abilities of in-context learning methods to ULs.\nThere is a lot of headroom left to improve performance on ULs. Overall, across all tasks there is still a considerable amount of headroom left. For ASR, OCR and transliteration, around 10% of characters are still incorrectly predicted. On autocomplete, models only make the correct prediction in about one fourth of all cases. For MT, on average only about a third of n-grams in the hypothesis are also present in the reference, and vice versa. For QA and retrieval, there are large performance differences between in-language and cross-language settings and much headroom still left. On NER, models perform relatively well but are still far from perfect performance on the task. Finally, on semantic parsing models are only able to produce the correct output in around a third of all cases. Ghomálá'. Similarly, translation models underperform in Amharic and Yoruba. On ASR, the lowestperforming languages are Yoruba but models also struggle with other languages such as Gaelic, and many South Asian languages such as Lao, Khmer, and Burmese.\nTask-specific observations ByT5 provides the best performance while the size of the model does not seem to impact performance much. Several aspects of the data lead to higher error rates in transliteration: the model struggles with input in the Perso-Arabic script and to produce output in Latin based on a different script. For autocomplete (see Appendix B.3), our analyses indicate that models perform better on text that uses the Latin script."
},
{
"figure_ref": [],
"heading": "Recommendations",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we make recommendations to researchers who plan to make use of this benchmark.\nUse of splits XTREME-UP offers a train, validation, and test split for each task. We recommend using the training split for learning the parameters of your model or as exemplars for in-context learning while iteratively checking your progress on the validation (i.e. development) split. The test split should not be used for iterative evaluation of your models or other sorts of hill-climbing; instead, it should be reserved for reporting your results and comparing after you have finished development on your models. Experiments that follow this customary scientific rigor should expect to show better generalization and less overfitting to the test split.\nUse of additional pre-training data One potential confounder for results along different pretrained models is the variation in pre-training data; where this data overlaps with the targets (outputs) in XTREME-UP validation and test splits, results can be artificially inflated, providing a sense that results are better than they are in reality-if the validation or test data leaked into the pre-training data via contamination during large-scale data scraping, then it's unlikely that the system would truly perform as well for new unseen inputs. Therefore, we recommend that when researchers modify the pre-training data for a model, they explicitly report overlap (contamination) between the targets of the XTREME-UP validation/test splits and their pre-training corpus. 19 19 We recognize that this is a very large-scale undertaking, requiring a fairly large amount of compute. As such, we Use of additional supervised data It is entirely possible that the community will find creative ways to improve models based on supervised data not included with XTREME-UP. However, researchers should bear in mind how this might affect the comparability of their results with other models. The following axes should be considered:\n1. Any additional data from high resource languages is always allowed in the XTREME-UP setting.\n2. Supervised data (e.g. parallel data for MT) harvested from the web, religious, books, and other opportunistic sources will typically be out-of-domain and is therefore admissible; conversely, supervised data from ULs from highly similar tasks or domains should generally be considered against the spirit of the XTREME-UP benchmark.\n3. Monolingual data from UL is admissible with the caveat that one should measure overlap with targets, as discussed above.\nAvoid off-the-shelf MT systems Data augmentation via automatically translating high-resource supervised data to languages with less supervised data has proven a very effective means of improving system quality. However, it is not necessarily realistic to use a pre-existing MT system (e.g. an API or an open-source model) since those systems have typically been trained on a large amount of parallel data-or at least unknown data. This means that additional supervised data would then be leaking into the experimental setup, which is otherwise intended to reflect the reality that most under-represented languages have very little supervised data. If data augmentation via translation is used, we encourage researchers to report the parallel data sources used and argue why this experimental setup is realistic-or to clearly point out such usage in their experiments as an unavoidable confound and discuss the limitations this sets on what conclusions can be drawn about how results will extrapolate to the breadth of under-represented languages.\nsuggest that it's may only be needed when making claims that compare systems (e.g. that the system with possiblycontaminated pre-training data is equivalent, better, or almost as good as some other system). Note, this analysis only needs to be done once for each pre-training corpus (e.g., once for mC4) and it is very likely that organizations with enough compute to pre-train a new model on a new corpus would also have sufficient compute to calculate overlap.\nIn all cases, researchers should rigorously report what additional data was used and how; each use case comes with its own considerations and, above all, researches should make a well-reasoned argument that their use of data (i) does not artificially inflate evaluation scores and (ii) reflects a real-world scenario of finding and applying data."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "We have presented XTREME-UP, a multilingual benchmark distinguished by its being (i) scarcedata, (ii) user-centric, and (iii) focused on underrepresented languages. The benchmark contains input modalities of text, images, and audio while still allowing experimentation with text-only models. We hope this benchmark will be useful in accelerating research that is useful to speakers of under-represented languages and in highlighting both the progress and limitations of current models of language."
},
{
"figure_ref": [],
"heading": "A Language Coverage",
"publication_ref": [],
"table_ref": [
"tab_4"
],
"text": "We provide an overview of the under-represented languages in XTREME-UP in Table 5. For each language, we indicate a) the ISO 639-1 code (or ISO 639-3 code if the former is unavailable); b) its language family according to Glottolog (Nordhoff and Hammarström, 2011); c) the number of datasets in XTREME-UP including the language; d) its resource level based on the taxonomy of Joshi et al. (2020) (0 is least and 5 is highest-resourced); and e) which tasks include the language."
},
{
"figure_ref": [],
"heading": "B Data cards B.1 ASR B.1.1 Task description",
"publication_ref": [],
"table_ref": [],
"text": "Automatic speech recognition (ASR) transcribes speech inputs into human-readable text, serving as a fundamental step for various speech language understanding applications. The transcripts are often calibrated with some pre-trained language models to produce the final outputs. In this paper, we build the ASR benchmark in this way: first, transcribe input audio into text with a pre-trained speech recognition model; then calibrate the transcripts by fine-tuning pre-trained language models on paired transcripts and ground truths."
},
{
"figure_ref": [],
"heading": "B.1.2 Data creation",
"publication_ref": [
"b14",
"b10",
"b77",
"b76"
],
"table_ref": [],
"text": "Experimented on the FLEURS corpus (Conneau et al., 2023), we use Maestro-U (Chen et al., 2023) to generate the ASR transcripts. For the pre-trained language models, we choose mT5-base (Xue et al., 2021) and ByT5-base (Xue et al., 2022) models. We paired the ASR transcripts with the ground truths to fine-tune the mT5 or ByT5 models. The average character error rate (CER) of Maestro-U is 8.28% across 102 languages, providing a strong baseline. Therefore, we build the ASR benchmark in a selective way: first, we compare the Maestro-U baseline CER on the dev set with the CER obtained by fine-tuned mT5 or fine-tuned ByT5. If the fine-tuned result is better, we choose the finetuned model for the language to rescore its test set; otherwise, we keep the baseline Maestro-U results for the test."
},
{
"figure_ref": [],
"heading": "B.1.3 Data structure",
"publication_ref": [],
"table_ref": [],
"text": "We followed the data split of train, dev, and test sets in FLEURS, and filtered out the examples where Maestro-U prediction is empty (i.e., all the deletion errors). The pairs of transcript and ground truth are saved in jsonl and tsv format.\nThe individual language datasets are mostly distinguished by the language and region BCP-47 codes, e.g., the kam_ke code represents Kamba language spoken in Kenya. In some cases, when multiple writing systems are available for a language, the ISO 15924 script code is used as well, as is the case with the code sd_arab_in that denotes Sindhi as spoken in India and recorded using Arabic script, as opposed to its Pakistani counterpart. \n3 ✓ ✓ ✓ Ewe ee Atlantic-Congo 1 1 ✓ Greek el Indo-European 3 3 ✓ ✓ ✓ Estonian et Uralic 3 3 ✓ ✓ ✓ Fula ff Atlantic-Congo 2 1 ✓ ✓ Filipino fil Austronesian 2 1 ✓ ✓ Irish ga Indo-European 4 2 ✓ ✓ ✓ Galician gl Indo-European 3 3 ✓ ✓ ✓ Gujarati gu Indo-European 4 1 ✓ ✓ ✓ Hausa ha Afro-Asiatic 5 2 ✓ ✓ ✓ ✓ Hebrew he Afro-Asiatic 3 3 ✓ ✓ ✓ Armenian hy Indo-European 4 1 ✓ ✓ ✓ Indonesian id Austronesian 5 3 ✓ ✓ ✓ ✓ ✓ Igbo ig Atlantic-Congo 4 1 ✓ ✓ ✓ Icelandic is Indo-European 4 2 ✓ ✓ ✓ Javanese jv Austronesian 3 1 ✓ ✓ Georgian ka Kartvelian 2 3 ✓ ✓ Kamba kam Atlantic-Congo 2 0 ✓ ✓ Kabuverdianu kea Indo-European 2 0 ✓ ✓ Kazakh kk Turkic 2 3 ✓ ✓ Khmer km Austroasiatic 3 1 ✓ ✓ Kannada kn Dravidian 5 1 ✓ ✓ ✓ ✓ Kyrgyz ky Turkic 3 1 ✓ ✓ Luxembourgish lb Indo-European 3 1 ✓ ✓ (Lu)Ganda lg Atlantic-Congo 4 1 ✓ ✓ ✓ Lingala ln Atlantic-Congo 3 1 ✓ ✓ Lao lo Tai-Kadai 3 2 ✓ ✓ Lithuanian lt Indo-European 3 3 ✓ ✓ ✓ (Dho)Luo luo Nilotic 3 0 ✓ ✓ ✓ Latvian lv Indo-European 3 3 ✓ ✓ ✓ Maori mi Austronesian 3 1 ✓ ✓ Macedonian mk Indo-European 3 1 ✓ ✓ Malayalam ml Dravidian 4 1 ✓ ✓ ✓ Mongolian mn Mongolic-Khitan 3 1 ✓ ✓ Mossi (Mooré) mos Atlantic-Congo 1 0 ✓ Marathi mr Indo-European 3 2 ✓ ✓ ✓ Malay ms Austronesian 2 3 ✓ ✓ Maltese mt Afro-Asiatic 2 2 ✓ ✓ Burmese my Sino-Tibetan 4 1 ✓ ✓ ✓ Nepali ne Indo-European 3 1 ✓ ✓ Norwegian no Indo-European 2 1 ✓ ✓ Northern Sotho nso Atlantic-Congo 3 1 ✓ ✓ Nyanja (Chichewa) ny Atlantic-Congo 4 1 ✓ ✓ ✓ Occitan oc Indo-European 2 1 ✓ ✓ Oromo om Afro-Asiatic 3 1 ✓ ✓ Oriya or Indo-European 2 1 ✓ ✓ Punjabi pa Indo-European 4 2 ✓ ✓ ✓ Nigerian Pidgin pcm Indo-European 2 0 ✓ ✓ Pashto ps Indo-European 3 1 ✓ ✓ Romanian ro Indo-European 3 3 ✓ ✓ ✓ Kinyarwanda rw Atlantic-Congo 1 1 ✓ Sanskrit sa Indo-European 1 2 ✓ Sindhi sd Indo-European 4 1 ✓ ✓ ✓ Sinhala si Indo-European 2 0 ✓ ✓ Slovak sk Indo-European 3 3 ✓ ✓ ✓ Slovenian sl Indo-European 3 3 ✓ ✓ ✓ Shona sn Atlantic-Congo 4 1 ✓ ✓ ✓ Somali so Afro-Asiatic 3 1 ✓ ✓ Swahili sw Atlantic-Congo 8 2 ✓ ✓ ✓ ✓ ✓ ✓ ✓ Tamil ta Dravidian 4 3 ✓ ✓ ✓ ✓ Telugu te Dravidian 6 1 ✓ ✓ ✓ ✓ ✓ Tajik tg Indo-European 3 1 ✓ ✓ Thai th Tai-Kadai 3 3 ✓ ✓ ✓ Tswana (Setswana) tn Atlantic-Congo 1 2 ✓ Twi tw Atlantic-Congo 1 1 ✓ Uyghur ug Turkic 1 1 ✓ Ukrainian uk Indo-European 3 3 ✓ ✓ ✓ Umbundu umb Atlantic-Congo 2 0 ✓ ✓ Urdu ur Indo-European 4 3 ✓ ✓ ✓ ✓ Uzbek uz Turkic 2 3 ✓ ✓ Wolof wo Atlantic-Congo 3 2 ✓ ✓ ✓ Xhosa xh Atlantic-Congo 4 2 ✓ ✓ ✓ Yoruba yo Atlantic-Congo 5 2 ✓ ✓ ✓ ✓ Zulu zu Atlantic-Congo 5 2 ✓ ✓ ✓ ✓"
},
{
"figure_ref": [],
"heading": "B.1.5 Experiments and Discussion",
"publication_ref": [],
"table_ref": [
"tab_5"
],
"text": "We compared fine-tuned mT5-base and ByT5-base baselines, which were built on TPU. In addition, we explored the compute efficient fine-tuning on GPU, using a mT5-small model as pre-trained model. The three models took 4500, 6500 and 4000 steps to converge, respectively. We report the character error rate for the predicted transcripts by the finetuned models against the one for the Maestro-U baseline, which is 8.28% on average for 102 languages -a quite strict baseline. We observed small gains through fine-tuning with different pre-trained models, as shown in Table 6.\nIt is observed that ByT5 yields better fine-tuned results than mT5, indicating that byte is a better modeling unit when it comes to textual data of various writing systems. By calculating the average CER for 24 high-resourced language group and 78 low-resourced language group respectively, we find that both mT5 and ByT5 fine-tuned models can reduce CER from 6.40% baseline to 6.36% for high-resourced languages, while ByT5 on its own can further improve CERs for low-resourced languages from 8.86% baseline to 8.80%.\nFine-tuned ByT5 also generalized well on languages which were not seen in the pre-training phase. With a limited amount of fine-tuning data, ByT5 can improve baseline on the group of unseen languages, especially on Umbundu (umb_ao, -14% CER Relative). Even though only Romanized Chinese is used to pre-train ByT5, the fine-tuned ByT5 outperformed baselines for both Mandarin (in simplified Chinese, cmn_hans_cn), and Cantonese (in traditional Chinese, cmn_hant_hk)."
},
{
"figure_ref": [],
"heading": "B.2 Optical character recognition (OCR) B.2.1 General information Dataset title UL-OCR",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "B.2.2 Data creation",
"publication_ref": [],
"table_ref": [],
"text": "We retrieve books that are in the public domain on Google Books. In many cases, these are historic books where the copyright has expired while others are more recent publications. We focus on languages with diverse scripts where no existing OCR dataset is currently available. We observe that many public-domain books in such languages are religious or linguistic in nature and were created for missionary purposes. In order to identify a diverse set of high-quality books, we first conduct an annotation task where we ask annotators to look at pages of a book and assign whether it is a) not in the target language, b) religious, c) consisting mainly of tables, d) linguistic (e.g., a dictionary or grammar book), e) not intelligible, or f) good quality. Based on this annotation, we needed to filter out certain languages such as Hausa, Igbo, Malagasy, Yoruba, and Zulu, which had an insufficient amount of high-quality public-domain books available."
},
{
"figure_ref": [],
"heading": "B.3 Autocomplete",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "B.3.1 Task description",
"publication_ref": [
"b55",
"b15"
],
"table_ref": [],
"text": "Autocomplete (or predictive text), i.e., predicting the rest of a word a user is typing, is a useful technology that speeds up human-computer interaction. However, while language modeling (LM) is a core natural language processing (NLP) task, current LM evaluation does not address the practical constraints of human-computer interaction and current LMs are not directly useful for autocomplete in under-represented languages.\nIn order to evaluate multilingual models on an evaluation setting as close as possible to the realworld usage of autocomplete, we curated the Universal Dependencies (UD) dataset (Nivre et al., 2020;de Marneffe et al., 2021) according to a set of high level principles that we describe in the section below."
},
{
"figure_ref": [],
"heading": "B.3.2 Data creation",
"publication_ref": [],
"table_ref": [],
"text": "The original UD dataset was filtered to better fit the user centric paradigm proposed. We removed a) treebanks using only ancient data, for example liturgical text written in Latin, Ancient Greek or Sanskrit; b) languages with fewer than 100 speakers like Akuntsú; c) signed languages like the Swedish Sign Language; d) highly domain-specific content like for instance SiMoNERo (Mititelu and Mitrofan, 2020) which contains texts from three medical subdomains: cardiology, diabetes, endocrinology; e) languages that are \"high resource\" by XTREME-UP standards with the exception of English which we kept for prototyping; f) languages that do not have all three of: training, validation and test sets: The resulting corpus features 23 languages: Basque, Belarusian, Bulgarian, Danish, Eastern Armenian, English, Estonian, Galician, Scottish Gaelic ,Greek, Hebrew, Icelandic, Indonesian, Irish, Latvian, Lithuanian, Nigerian Pidgin, Romanian, Slovak, Slovenian, Ukrainian, Urdu, and Uyghur."
},
{
"figure_ref": [],
"heading": "B.3.3 Data structure",
"publication_ref": [
"b74"
],
"table_ref": [],
"text": "A data instance has two fields, input and target, for instance {input: \"en_-We look f$\", target: \"forward\"}. The input field is composed of a prefix \"en_-\" to indicates the language to the model and a context sentence: \"We look f$\". The target field is the word to predict. We normalize all text with Unicode NFKC normalization (Whistler, 2021).\nAnnotation process In the following, we describe how the example described above is generated from the source data. The original sentence is \"We look forward to your active participation to make this forum an exciting meeting place for like minded individuals.\" The steps are: a) The context sentence including the target can have at most 10 words. A random word of more than 5 characters is chosen to be the target. b) A target context is sampled from the target and added to the context. In this example it is the character \"f\". The sample rule is to select a number of characters that can vary between 0 to the number of characters in the target minus three. In our example, the target \"forward\" could be sampled from \"\" to \"forw\". c) A specific token \"$\" is added just after the target context."
},
{
"figure_ref": [],
"heading": "B.3.4 Data statistics",
"publication_ref": [],
"table_ref": [],
"text": "We sampled up to 2,000 examples from each language's training set, 1,000 examples from valida- tion, and 1,000 examples from test. This prevents the languages from having disproportionately more data; where the original sets were smaller than these targets, we used all available data. We display the language statistics in Table 7. Note that these experiments are done on a preliminary dataset and not the final release version of XTREME-UP."
},
{
"figure_ref": [],
"heading": "B.3.5 Experiment",
"publication_ref": [
"b77",
"b76",
"b60"
],
"table_ref": [],
"text": "We compared mT5 (Xue et al., 2021) and ByT5 (Xue et al., 2022), two state-of-the-art multilingual pre-trained LMs that are based on subwords and bytes respectively. The models were fine-tuned for 10 epochs on autocomplete training set, moreover. We used two metrics: top-3 word accuracy (Acc@3) and chrF: character n-gram F-score (Popović, 2015)."
},
{
"figure_ref": [],
"heading": "B.3.6 Results",
"publication_ref": [],
"table_ref": [],
"text": "We observe that ByT5 achieve better performance than mT5 for both Acc@3 and chrF on the autocomplete task as it is displayed in Table 8. Also ByT5 require less than half the time to fine-tune on the training set (45 minutes) compared to mT5 (1 hours and 30 minutes)."
},
{
"figure_ref": [],
"heading": "B.3.7 Analyses",
"publication_ref": [
"b32"
],
"table_ref": [],
"text": "Based on Acc@3 and chrf, the most challenging languages for mT5 are Eastern Armenian ((hy)) and Uyghur (ug) respectively. Whereas Nigerian Pidgin is the (pcm) and Scottish Gaelic are the easiest languages. For ByT5, whether we consider Acc@3 or chrF, the most challenging language is Uyghur, and the easiest language is Galician (gl). Yet, these extremes only offer a qualitative comparison of mT5 and ByT5. Next, we investigate four questions around model performance: a) Do mT5 and ByT5 have the same cross-lingual generalization pattern? b) Do some languages yield higher scores because autocompletion guesses the same words? c) Do some languages yield higher scores because they have a smaller vocabulary in their corpora? d) Does similarity to the Latin alphabet impact models' performance? We test several hypotheses below, considering a relationship to be significant when the p-value is under 0.05.\nDo mT5 and ByT5 have the same cross-lingual generalization pattern? mT5 and ByT5 have the same cross-lingual generalization pattern if the difficulty to generalize to a new language is the same for both models relatively to other languages. In other words, if models' performance are ranked similarly, they share the same cross-lingual generalization pattern. To evaluate this hypothesis we computed the Spearman's rank correlation between mT5 and ByT5 Acc@3. We got a Spearman's rank correlation of 0.69 with p-value < 0.001. This means that the two models have a high degree of relative agreement, in other words, if a new language is added, there is a high chance that the language is going to be challenging or not for both mT5 and ByT5.\nDo some languages yield higher scores because autocompletion guesses the same words? If our dataset in given language over-represents a word to predict, then the model might have misleadingly good performance by always predicting the same word. This would mean that the dataset is not balanced with regards to the diversity of target words.\nA common way to model the diversity of a distribution of words is to compute its entropy, so we computed the the Pearson correlation between the entropy of the test set's target word distribution in each language and mT5 and ByT5 Acc@3. The entropy of a distribution of word is maximal if every word is different, and it is minimal if it consist on a single word. mT5 and ByT5 displayed correlation coefficients of -0.16 and 0.13 respectively with p-value of 0.45 and 0.53 respectively. These results show that there is insufficient evidence to conclude that there is a significant linear relationship between target words diversity and model performance because the p-value is far above the 0.05 significance threshold. Hence, target word diversity is not a good predictor of model performance variability across languages.\nDo some languages yield higher scores because they have a smaller vocabulary in their corpora?\nWe expect that languages with smaller corpora will be easier to fine-tune on because of a smaller prediction space. To test that hypothesis, we computed the Pearson correlation between test set's vocabulary size and mT5 and ByT5's Acc@3 for each language. mT5 and ByT5 displayed correlation coefficients of -0.29 and 0.13 respectively with p-value of 0.17 and 0.54 respectively. Thus there is insufficient evidence to conclude that there is a significant linear relationship between vocabulary size and model performance because the p-value is above the 0.05 significance threshold.\nDoes similarity to the Latin alphabet impact models' performance? We verify this hypothesis quantitatively by computing the similarity between a) a Latin alphabet composed of the 26 letters of the alphabet in lower and upper case and b) the alphabet of each language corresponding to all the characters in the test set except punctuation and special characters. The similarity was computed with the Jaccard similarity coefficient (Jaccard, 1908), i.e. the ratio of number of unique items in the intersection of both alphabets and the number of unique items in the union of both alphabets. Moreover we used the same methodology as before and computed the Pearson correlation between the Jaccard similarity index and chrF as this metric is more granular in models' character level performance. We observed a correlation of 0.56 and 0.75 for mT5 and ByT5 respectively with p-values < 0.01 respectively. It indicates that the similarity between the Latin alphabet and each language alphabet is significantly correlated to mT5 and ByT5 chrF."
},
{
"figure_ref": [],
"heading": "B.3.8 Evaluation and Discussion",
"publication_ref": [],
"table_ref": [],
"text": "Whether we used a word level metric like Acc@3 or a character level metric like chrF, ByT5 is more accurate at autocomplete than mT5. We also ob- serve that these models generalize more easily to languages written in an alphabet closer to the Latin alphabet, ByT5 being more sensitive to the alphabet of the input language."
},
{
"figure_ref": [],
"heading": "B.4 Transliteration",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "B.4.1 Task description",
"publication_ref": [],
"table_ref": [],
"text": "Transliteration is the conversion of text in one writing system to another writing system, e.g., text written in the Devanagari script to the Latin script.\nIt differs from translation in that it does not change the language content of the text, just the script. Many languages are written in multiple scripts, and the current task involves transliterating whole sentences, not just isolated terms, from one script to another."
},
{
"figure_ref": [],
"heading": "B.4.2 Data Creation and Annotation process",
"publication_ref": [
"b65"
],
"table_ref": [],
"text": "Most of the data for the task comes from the romanized full-string subset of the Dakshina dataset (Roark et al., 2020), in which 10,000 Wikipedia sentences written in the native scripts of the 12 languages were human-romanized by native speakers, resulting in parallel sentences in the native and Latin scripts. 21 Two 10,000 sentence additions were made to this data for the current transliteration task: Amharic Wikipedia sentences were similarly manually romanized by native speakers; and the Punjabi sentences from the Dakshina dataset, originally written in the Gurmukhi (Brahmic) script, were manually transliterated by native speakers to the Shahmukhi (Perso-Arabic) script."
},
{
"figure_ref": [],
"heading": "B.4.3 Data Preparation",
"publication_ref": [
"b58",
"b74",
"b33",
"b24"
],
"table_ref": [
"tab_7"
],
"text": "The resulting collection allows for overall 30 tasks converting between various scripts. These are summarised in Table 9 where, for each language indicated by the BCP-47 code (Phillips and Davis, 2009), the corresponding transliteration tasks are shown for scripts indicated by their ISO-15924 codes (ISO, 2004). All the native script data was normalized using Unicode NFC (Whistler, 2021). The data was then further transformed using language-specific visual normalization for Brahmic and Perso-Arabic writing systems using the Nisaba script normalization library (Johny et al., 2021;Gutkin et al., 2022). Both NFC and visual normalization operations preserve visual invariance of the input text, with visual normalization handling many ambiguous cases that fall outside the scope of standard NFC."
},
{
"figure_ref": [],
"heading": "B.4.4 Data Statistics",
"publication_ref": [],
"table_ref": [],
"text": "For each task, we establish 2,000 training sentences, 2,000 development set sentences, and close to 6,000 test sentences. Training data for any pretrained models used in the task cannot include the Dakshina dataset. Since this is a contextual fewshot transliteration benchmark, we do not provide the romanization lexicons that were released in the Dakshina dataset along with the full sentence romanizations.\nOur few-shot contextual transliteration task covers 13 languages from 3 language families (Indo-Aryan, Dravidian and Semitic), all but one (Amharic) from South Asia."
},
{
"figure_ref": [],
"heading": "B.4.5 Directionality and Evaluation Ambiguity",
"publication_ref": [],
"table_ref": [],
"text": "One difference between romanization in these languages and transliteration in the opposite direction (from the Latin script to the native script) is that none of the languages in the benchmark have an orthography in the Latin script, i.e., there is no single correct spelling in the Latin script for these languages. Rather, individuals tend to provide a rough phonetic transcription of the sentences using the Latin script. As a result, word identity may be difficult to achieve (hence high word-error rate), but string similarity should be relatively high between quality romanizations hence we use character-error rate to evaluate the transliterations. The ability to produce romanizations automatically has several key use cases, including simulation of parallel data from mono-script language samples, and for multilingual modeling of languages that use different scripts. For that reason, we include both directions in the benchmark."
},
{
"figure_ref": [],
"heading": "B.4.6 Experimental Setup",
"publication_ref": [
"b76",
"b77",
"b76",
"b38",
"b59"
],
"table_ref": [],
"text": "Previously Xue et al. (2022) performed ByT5 finetuning and evaluation of transliteration and romanization directions separately on single-word, rather than full-sentence, data from vanilla Dakshina dataset. In this benchmark we remove the separation into transliteration and romanization by requiring all tasks to be fine-tuned jointly. In order to achieve this, during all stages of training, development and testing a special code is prepended to the input feature strings for each task. This task code indicates that the input features correspond to the conversion from writing system Source to writing system Target for a language lang. It is encoded as a string \"lang_Source_Target\". For example, for Punjabi (pa) conversion from Shahmukhi (Arab) to Gurmukhi (Guru) writing systems, the task code is \"pa_Arab_Guru\".\nIn the default setup we jointly fine-tune the 30 transliteration tasks using mT5 and ByT5 models in Small, Base and Large configurations that correspond to around 300M, 582M and 1.2B parameters, respectively (Xue et al., 2021(Xue et al., , 2022)). Fine-tuning uses 10K training steps with a batch size of 128. We used Google TPU-v3 accelerators (Kumar et al., 2019) for fine-tuning all the configurations apart from ByT5 Large for which a more powerful TPU-v4 (Pope et al., 2022) Why is this dataset part of XTREME-UP? Machine translation is an important tool for expanding language coverage for natural language processing tools. FLORES-101 is a high-quality, highlymultilingual dataset."
},
{
"figure_ref": [],
"heading": "Data Fields",
"publication_ref": [],
"table_ref": [],
"text": "1. input: the source sentence, which is always English (string) 2. target: the target-language translation of the source sentence (string) Data Example {\"input\": \"<2xh> Local media reports an airport fire vehicle rolled over while responding.\", \"target\": \"Oonondaba basekuhlaleni bxele ukuba isithuthi somlilo sesitishi senqwelomoya siye saphethuka sisazama ukunceda.\"} Languages Included in XTREME-UP release (93): Afrikaans (af), Amharic (am), Arabic (ar), (Eastern) Armenian (hy), Assamese (as), (North) Azerbaijani (az), Belarusian (be), Bengali (bn), Bosnian (bs), Bulgarian (bg), Burmese (my), Catalan (ca), Cebuano (ceb), Central Kurdish (ckb), Chinese (zh), Croatian (hr), Czech (cs), Danish (da), Dutch (nl), Estonian (et), Finnish (fi), French (fr), Fula (ff), Galician (gl), Georgian (ka), German (de), Greek (el), Gujarati (gu), Hausa (ha), Hebrew (he), Hindi (hi), Hungarian (hu), Icelandic (is), Igbo (ig), Indonesian (id), Irish (ga), Italian (it), Japanese (ja), Javanese (jv), Kannada (kn), Kazakh (kk), Khmer (km), Korean (ko), Kyrgyz (ky), Lao (lo), Latvian (lv), Lingala (ln), Lithuanian (lt), (Lu)Ganda (lg), Luxembourgish (lb), Macedonian (mk), Malay (ms), Malayalam (ml), Maltese (ml), Maori (mi), Marathi (mr), Mongolian (mn), Nepali (ne), Pedi (Sepedi) (Northern Sotho) (nso), Norwegian (no), Nyanja (Chichewa) (ny), Oriya (or), Oromo (om), Pashto (ps), Persian (fa), Polish (pl), Portuguese (pt), Punjabi (pa), Romanian (ro), Russian (ru), Serbian (sr), Shona (sn), Sindhi (sd), Slovak (sk), Slovenian (sl), Somali (so), Spanish (es), Swahili (sw), Swedish (sv), Tagalog (tl), Tajik (tg), Tamil (ta), Telugu (te), Thai (th), Turkish (tr), Ukrainian (uk), Urdu (ur), Uzbek (uz), Vietnamese (vi), Welsh (cy), Xhosa (xh), Yoruba (yo), Zulu (zu).\nEvaluated in benchmark (39): Amharic (am), (Eastern) Armenian (hy), Assamese (as), (North) Azerbaijani (az), Burmese (my), Central Kurdish (ckb), Gujarati (gu), Hausa (ha), Icelandic (is), Igbo (ig), Irish (ga), Javanese (jv), Kannada (kn), Khmer (km), Kyrgyz (ky), Lao (lo), Lingala (ln), (Lu)Ganda (lg), Luxembourgish (lb), Macedonian (mk), Malayalam (ml), Mongolian (mn), Nepali (ne), Pedi (Sepedi) (Northern Sotho) (nso), Nyanja (Chichewa) (ny), Oromo (om), Pashto (ps), Punjabi (pa), Shona (sn), Sindhi (sd), Somali (so), Swahili (sw), Tajik (tg), Telugu (te), Welsh (cy), Xhosa (xh), Yoruba (yo), Zulu (zu).\nData Statistics 50% of the FLORES-101 dev split was reserved for training and the remainder for validation. The original devtest split was unchanged and reserved for testing. This results in 499/498/1012 sentence pairs for train/validation/test, respectively."
},
{
"figure_ref": [],
"heading": "Dataset Curators",
"publication_ref": [],
"table_ref": [],
"text": "The original dataset was curated by the NLLB (No Language Left Behind) Team ([email protected]). The version included in XTREME-UP was curated by Parker Riley ([email protected]) and Isaac Caswell ([email protected]).\nCuration Rationale The original FLORES-101 dataset was created to be able to evaluate machine translation models in many languages. The version released in XTREME-UP was created to focus on low-resource languages and provide an in-domain train split along with validation and test splits, all of sizes in line with other tasks in XTREME-UP."
},
{
"figure_ref": [],
"heading": "Data Sources",
"publication_ref": [
"b23"
],
"table_ref": [],
"text": "The source data (selected by the NLLB Team) comes from Wikinews, Wikijunior, and Wikivoyage.\nDataset Creation Details of the creation of the original dataset are available in the original publication (Goyal et al., 2022).\nChanges to the Original Dataset for XTREME-UP The version of the dataset in XTREME-UP only has the source and target strings, removing additional metadata. We also include 93 of the original 100 non-English languages (the subset supported by Google Translate). Of these, only 39 are used for official evaluation. Why is this dataset part of XTREME-UP? Question answering enables information access."
},
{
"figure_ref": [],
"heading": "Data Fields",
"publication_ref": [],
"table_ref": [
"tab_4"
],
"text": "1. question: a question in the target language (string) 2. title: the title of the evidence passage -target language for in-language setting, English for cross-language setting (string)\n3. passage: the evidence passage, which might contain an answer to the question -target language for in-language setting, English for cross-language setting (string) 4. answer: the answer (if any) to the question (string)\nData Example See Table 2.\nLanguages See Table 5. of NER models on natural data data, we process the data in order to align the token-level annotations with byte-level spans in the original pre-tokenized text. For the NER task, we provide the original pretokenized text as input to the model. Hausa and Fon subsets of the original data were excluded as matching with the unlabeled source data revealed annotation artefacts in both language subsets."
},
{
"figure_ref": [],
"heading": "Data Statistics",
"publication_ref": [],
"table_ref": [],
"text": "B.8 Semantic parsing"
},
{
"figure_ref": [],
"heading": "B.8.1 Task description",
"publication_ref": [
"b41"
],
"table_ref": [],
"text": "Semantic parsing is the task of mapping a natural language utterance to a logical form or a structured interpretation that can be executed by a system such as a virtual assistant. For XTREME-UP, we adapted the MTOP (Li et al., 2021) test dataset to 15 languages, and to 3 code-switched Indic languages. The original MTOP data was published by Facebook and covers 6 languages across 11 domains, 117 intents and 78 slots."
},
{
"figure_ref": [
"fig_3"
],
"heading": "B.8.2 Data creation",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we describe the two processes used to extend the MTOP instances: the first involves translation and localization with professional translators and the second code-switching using a language model and verification by human annotators.\nIn both processes, we perform a linearization step of the query and parse. Given an English utterance from the MTOP English test set and the corresponding slot information (slot names each with start and end bytes), we add slot tags around corresponding tokens in the query (Figure 2).\nTranslating MTOP to 15 languages: We take the bracketed versions of the slot-tagged English sentences from MTOP and we create translations and localization tasks to be carried out by professional translators. We ran two pilots on a small sample of the data to gather feedback and improve the annotation guidelines. The translators had to translate the original utterances to a given target language, while keeping the brackets around slot value translations and localizing those where possible. Once the pilots were completed without issues, we scaled the tasks to the full test set.\nWe carried out manual inspections on samples of the data to check if translation and localization was happening correctly, and a set of automatic checks on the full data to ensure that slots were matching between original and translated utterances. Data was sent back to annotators until all the issues were fixed."
},
{
"figure_ref": [],
"heading": "Code-switching MTOP to 3 Indic languages:",
"publication_ref": [],
"table_ref": [],
"text": "We use PaLM to convert the linearized query into a code-mixed query using few-shot prompting. We experimented with different discrete prompt design strategies and selected the best prompts after a qualitative evaluation on a small held-out set (11 examples) covering all 11 domains. Specifically we experimented with three designs.\n• Naive prompting. The prompt contains (a) the task description followed by a set of examples consisting of (b) the original English linearized query and (c) the corresponding code-mixed version.\n• Parallel sentence prompting. In this case, the prompt contains (a) the task description, (b) the original English linearized query, and also (c) the target translated query (obtained with Google translate) and (d) the corresponding code-mixed query.\n• Parallel reordered sentence prompting. Similar to the previous, however, target translated queries are human written.\nWe observed that the Parallel sentence prompting was producing higher quality utterances, with 7/11 correct conversions for Hindi-English. 6/11 for Bengali-English, and 8/11 for Tamil-English. We used this strategy to design prompts with the help of native speakers of those languages. We selected 21 sentences from the training split for creating corresponding exemplars for the prompts. With the latter, we performed few-shot prompting with the 64b PaLM model and converted the test split of MTOP to a code-switched corpora. Human annotators then had to check the PaLM generated data for the presence of code-mixing and for the labeling to be consistent between the original query and the code-mixed version. The annotators were instructed to fix the automatically generated data whenever they found such issues."
},
{
"figure_ref": [],
"heading": "B.8.3 Data structure and statistics",
"publication_ref": [],
"table_ref": [],
"text": "To create the training, validation and testing splits for MTOP, we start from the English test set and remove intents with less than 10 examples. This leaves us with 53 intents and a maximum of 4,223 examples for each language (some original MTOP languages may have less examples, while our codeswitched data may have more due to multiple paraphrases).\nFor each intent, we randomly select training examples such that each slot is covered by at least one example, for a minimum of 5 examples. We end up with training, development and test sets containing respectively a maximum of 285, 239, and 3,669 instances for each language."
},
{
"figure_ref": [],
"heading": "B.8.4 Experiments",
"publication_ref": [
"b77",
"b76",
"b76",
"b54"
],
"table_ref": [
"tab_9",
"tab_10"
],
"text": "We fine-tune mT5 (Xue et al., 2021) and ByT5 (Xue et al., 2022) in their base and large configurations on the multilingual training data we collected. Table 10 contains the Exact Match accuracies of a multilingual model trained on data from all languages but the code-switched sets. Table 11 contains the results of a model that includes the codeswitched sets. From both tables, we can see that ByT5-base is more accurate then the other models, even compared with the larger ones. This surprising result confirms similar findings on word-level tasks reported by Xue et al. (2022) and Nicosia and Piccinno (2022). We expect mT5 to catch up with ByT5 at larger sizes."
},
{
"figure_ref": [],
"heading": "C In-context learning examples",
"publication_ref": [],
"table_ref": [
"tab_11"
],
"text": "We show in-context learning examples for a selection of tasks in Table 12. Each example consists of a general instruction and prefixes for the input and target, which are repeated for each exemplar."
},
{
"figure_ref": [],
"heading": "Acknowledgements",
"publication_ref": [],
"table_ref": [],
"text": "We thank Slav Petrov, Jason Riesa, Raphael Hoffmann, Dipanjan Das, Clara Rivera, Chris Alberti, Machel Reid, and Timothy Dozat for helpful discussions and feedback. We are grateful to Noah Constant for a review of a draft of the paper. We also gratefully acknowledge the contributions of the researchers who built the datasets that have gone into XTREME-UP; we recommend that all component datasets be cited individually when using XTREME-UP in a paper such that dataset authors (many of whom are not authors of this article) receive credit for their work and so that those original sources remain easily discoverable in the literature."
},
{
"figure_ref": [],
"heading": "Contributions",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we provide more detail about the contributions of each author."
}
] | Data scarcity is a crucial issue for the development of highly multilingual NLP systems. Yet for many under-represented languages (ULs)languages for which NLP research is particularly far behind in meeting user needs-it is feasible to annotate small amounts of data. Motivated by this, we propose XTREME-UP, a benchmark defined by: its focus on the scarcedata scenario rather than zero-shot; its focus on user-centric tasks-tasks with broad adoption by speakers of high-resource languages; and its focus on under-represented languages where this scarce-data scenario tends to be most realistic. XTREME-UP evaluates the capabilities of language models across 88 under-represented languages over 9 key user-centric technologies including ASR, OCR, MT, and information access tasks that are of general utility. We create new datasets for OCR, autocomplete, semantic parsing, and transliteration, and build on and refine existing datasets for other tasks. XTREME-UP provides methodology for evaluating many modeling scenarios including text-only, multimodal (vision, audio, and text), supervised parameter tuning, and in-context learning. 1 We evaluate commonly used models on the benchmark. We release all code and scripts to train and evaluate models. 2 * Equal contribution. We list detailed contributions in §7. 1 While XTREME-UP supports in-context learning, our results indicate that few-shot in-context learning is less effective than fine-tuning on 100s of examples for ULs. We advocate for comparing such approaches directly as the community explores XTREME-UP. | XTREME-UP: A User-Centric Scarce-Data Benchmark for Under-Represented Languages | [
{
"figure_caption": "Figure 1 :1Figure 1: The tasks in XTREME-UP and their role in language technology. Left: enabling access to language technology; middle: facilitating information access as part of larger systems (question answering, information extraction, virtual assistants); right: making information accessible in the speaker's language.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "datset name: FLORES-101 2. Repository: https://github.com/ facebookresearch/flores/tree/main/ flores200 3. Paper: Goyal et al. (2022) 4. Point of Contact (original version): NLLB Team ([email protected])",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "datset names: TyDi QA, XOR-TyDi QA 2. Additional cross-lingual data was collected as part of XTREME-UP, following similar methodology",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: Creation of a linearized query from the actual query and its parse for semantic parsing.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "20 ",
"figure_data": "LanguageISO codeLanguage family# of datasetsResource levelQA Retrieval NERSemantic parsingMT ASR OCRTranslit-erationAuto-completeAfrikaansafIndo-European23✓✓AmharicamAfro-Asiatic72✓✓✓✓✓✓AssameseasIndo-European31✓✓AsturianastIndo-European21✓✓AzerbaijaniazTurkic31✓✓Ghomálá'bbjAtlantic-Congo10✓BelarusianbeIndo-European43✓✓✓✓BulgarianbgIndo-European33✓✓✓BambarabmMande11✓BengalibnIndo-European73✓✓✓✓✓✓✓BosnianbsIndo-European23✓✓CebuanocebAustronesian23✓✓Central KurdishckbIndo-European31✓✓WelshcyIndo-European31✓✓DanishdaIndo-European3",
"figure_id": "tab_3",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": "B.1.4 Data statisticsThe FLEURS dataset contains about 1.4k hours ofaudio in total for 102 languages. The training datacontains 271,488 examples across 102 languages,average length per utterance is about 20 tokens.There are 34,661 examples in the validation (dev)set, and 77,943 examples in the test set.",
"figure_id": "tab_4",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "ASR tasks evaluated using CER metric at 4K steps of fine-tuning mT5 and ByT5 Small and Base models.",
"figure_data": "Task CodeMaestro-UmT5 Base SmallByT5 BaseTask CodeMaestro-UmT5 Base SmallByT5 Baseaf_za4.194.194.244.19...... continued ......am_et8.608.608.668.6mt_mt11.4811.57 11.56 11.57ar_eg6.006.006.036.00my_mm14.7014.70 14.87 14.66as_in8.498.498.568.49nb_no4.144.144.214.14ast_es4.494.494.664.49ne_np9.229.269.259.22az_az5.675.675.75.67nl_nl3.153.153.273.15be_by3.343.343.423.34nso_za7.137.137.177.13bn_in6.166.166.206.16ny_mw7.087.087.086.95bs_ba2.932.933.052.93oc_fr7.687.687.847.68ca_es2.722.722.762.72om_et14.3614.36 14.52 14.36ceb4.364.364.494.36or_in7.427.428.707.42ckb_iq8.708.708.778.70pa_in7.357.357.387.35cmn_hans_cn16.4816.06 16.12 16.06pl_pl2.492.492.522.49cmn_hant_hk34.8234.23 34.24 34.13ps_af16.8216.82 16.86 16.82cs_cz3.303.303.353.30pt_br2.872.873.052.87cy_gb7.117.117.177.11ro_ro3.583.583.643.58da_dk6.306.306.356.30rup_bg2.722.722.862.72de_de2.392.392.462.39ru_ru3.052.863.092.87el_gr4.734.734.774.73sd_arab_in9.229.229.65 10.08en_us9.029.029.119.02sk_sk2.392.392.432.39es_4191.811.811.851.81sl_si4.584.584.604.17et_ee2.242.242.282.24sn_zw9.459.459.489.45fa_ir4.964.965.874.96so_so13.7313.73 13.81 13.73ff_sn21.2221.22 21.42 21.22sr_rs9.939.939.959.95fi_fi2.022.022.052.02sv_se4.214.214.304.21fil_ph3.593.593.633.59sw_ke12.6212.62 12.76 12.62fr_fr4.574.574.634.57ta_in12.3511.55 15.06 12.35ga_ie29.7529.75 29.78 29.79te_in7.487.487.567.48gl_es2.552.552.582.55tg_tj4.564.564.604.56gu_in5.755.755.95.75th_th11.8911.89 11.92 11.48ha_ng7.707.709.236.90tr_tr4.284.284.344.28he_il18.3618.36 18.40 18.36uk_ua5.445.445.475.44hi_in5.595.595.635.59umb_ao17.4617.09 17.47 14.98hr_hr4.464.464.564.46ur_pk7.617.617.637.61hu_hu7.057.057.107.05uz_uz7.407.407.427.40hy_am4.936.254.944.93vi_vn11.8011.80 11.83 11.80id_id3.143.143.163.14wo_sn15.2615.26 15.29 15.26ig_ng14.0614.06 14.29 14.07xh_za16.6516.65 16.69 16.68is_is6.236.236.256.23yo_ng19.8419.84 19.93 19.84it_it1.391.391.441.39zu_za5.565.565.625.56ja_jp jv_id25.74 4.6625.49 25.51 25.43 4.66 4.72 4.52Micro-Average8.288.278.408.22ka_ge10.0910.09 10.16 10.09kam_ke11.7411.69 11.78 11.74kea_cv4.114.114.174.11kk_kz3.583.583.663.58km_kh20.1520.15 20.15 20.15kn_in5.135.135.405.13ko_kr14.2914.29 14.22 14.29ky_kg4.534.534.564.44lb_lu13.5413.54 13.64 13.54lg_ug8.998.999.138.99ln_cd4.614.614.764.61lo_la22.8022.80 22.84 23.25lt_lt4.514.514.554.51luo_ke5.645.645.735.64lv_lv2.182.182.212.18mi_nz9.599.519.68.68mk_mk3.603.603.663.60ml_in5.045.455.25.07mn_mn8.438.438.468.43mr_in7.377.377.487.37ms_my3.893.893.923.89",
"figure_id": "tab_5",
"figure_label": "6",
"figure_type": "table"
},
{
"figure_caption": "Summary of the transliteration tasks.",
"figure_data": "Lang.TasksLang.TasksamEthi↔LatnGuru↔LatnbnBeng↔LatnpaArab↔LatnguGujr↔LatnGuru↔ArabhiDeva↔LatnsdArab↔LatnknKnda↔LatnsiSinh↔LatnmlMlym↔LatntaTaml↔LatnmrDeva↔LatnteTelu↔LatnurArab↔Latn",
"figure_id": "tab_7",
"figure_label": "9",
"figure_type": "table"
},
{
"figure_caption": "See Table1.",
"figure_data": "QueryStart playing Rihanna ' s latest albumParse[ IN : PLAY_MUSIC [ SL : MUSIC_ARTIST_NAME→ Rihanna ] [ SL : MUSIC_TYPE album ] ]Linearized QueryStart playing SL : MUSIC_ARTIST_NAME {→ Rihanna }' s SL : MUSIC_TYPE { album }Data Sources Evidence text was sourced fromWikipedia.Dataset Creation Details of the creation of theoriginal dataset are available in the original TyDiQA and XOR QA publications.B.7 Named Entity Recognition (NER)Dataset and task description The dataset con-tains processed data from MasakhaNER (Adelaniet al., 2021) and MasakhaNER 2.0 (Adelani et al.,2022). Both datasets were created by Masakhane 22 .Why is this dataset part of XTREME-UP?Named entity recognition is a fundamental taskin natural language processing. The MasakhaNERdatasets are high-quality multilingual datasets thatprovide data in 20 African languages. The datais human-annotated and thus higher quality thanautomatically collected NER datasets.Languages and ISO 639-3 codes Bambara(bam), Ghomálá' (bbj), Éwé (ewe), Igbo (ibo), Kin-yarwanda (kin), Luganda (lug), Luo (luo), Mossi(mos), Naija (pcm), Chichewa (nya), chiShona(sna), Kiswahili (swa), Setswana (tsn), Akan/Twi(twi), Wolof (wol), isiXhosa (xho), Yorùbá (yor),isiZulu (zul)Changes to the original datasets for XTREME-UP The original MasakhaNER datasets are pro-vided in CoNLL format where each input sentenceis already tokenized. This makes it difficult toevaluate NER models on natural text where tok-enization may often be messy and introduces a biastowards word and subword-based models. To pro-vide a level playing field and to enable evaluation22 https://www.masakhane.io/",
"figure_id": "tab_8",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Semantic Parsing: Exact Match (EM) accuracies of mT5 and ByT5 models of different sizes trained multilingually on few-shot data. We report accuracies on all languages.",
"figure_data": "mT5ByT5Languagebaselargebaselargeam20.01 26.47 33.41 25.60be27.82 37.52 46.72 37.36bn29.06 37.66 45.07 35.69de33.71 39.96 45.34 37.93de (loc)33.31 40.58 45.81 38.20en34.09 40.39 49.50 39.52es34.95 41.52 48.73 39.29fi26.63 35.74 46.80 37.17fr33.72 40.29 48.97 39.84ha21.84 27.60 42.07 29.98hi27.89 37.59 42.26 35.42hu25.87 33.47 43.82 35.98ja28.71 33.68 45.23 35.90pt_br33.98 39.12 47.90 39.50ru34.44 41.36 48.58 42.80sw24.06 30.25 39.96 32.09ta25.03 33.20 43.31 31.41th23.81 34.35 43.80 35.30tr27.44 36.44 44.58 36.63yo14.52 16.30 30.39 18.44zu18.73 26.79 36.96 27.49Average27.634.78 43.77 34.84",
"figure_id": "tab_9",
"figure_label": "10",
"figure_type": "table"
},
{
"figure_caption": "Semantic Parsing: Exact Match (EM) accuracies of mT5 and ByT5 models of different sizes trained multilingually on few-shot data. Here the multilingual training data includes three code-switched Indic languages and we report EM for such languages.",
"figure_data": "mT5ByT5Languagebaselargebaselargebn10.72 11.78 22.69 16.25hi16.05 18.48 25.03 17.69ta16.98 19.71 26.21 19.27TaskIn-context learning exampleTranslate between English and Afrikaans.TranslationEnglish: [INPUT]Afrikaans: [TARGET]Correct the ASR output in Afrikaans.ASRASR Afrikaans output: [INPUT]Corrected: [TARGET]Tag the named entities in the Swahili textas person (PER), organization (ORG),NERlocation (LOC), and date (DATE). Use $$ as delimiter.Swahili text: [INPUT]Named entities: [TARGET]",
"figure_id": "tab_10",
"figure_label": "11",
"figure_type": "table"
},
{
"figure_caption": "In-context learning examples.",
"figure_data": "",
"figure_id": "tab_11",
"figure_label": "12",
"figure_type": "table"
}
] | Sebastian Ruder; Jonathan H Clark; Alexander Gutkin; Mihir Kale; Min Ma; Massimo Nicosia; Shruti Rijhwani; Parker Riley; Jean-Michel A Sarr; Xinyi Wang; John Wieting; Nitish Gupta; Anna Katanova; Christo Kirov; Dana L Dickinson; Brian Roark; Bidisha Samanta; Connie Tao; David I Adelani; Vera Axelrod; Isaac Caswell; Colin Cherry; Dan Garrette Reeve; Ingle Melvin; Johnson Dmitry; Panteleev Partha; Talukdar Google; John Wieting Major; Con- Nie Tao; Dan Garrette; Reeve Ingle; Melvin Johnson; Dmitry Panteleev; Partha Talukdar; Min Ma Autocomplete; Jonathan Clark Mt; Jonathan Clark Ner; Dan Garrette Ocr; Sebastian Ruder Qa; Jonathan Clark Retrieval; Transliteration Alexander Gutkin; Connie Tao Fine; Jonathan Clark Data; Axel- Rod Vera | [
{
"authors": "David Adelani; Graham Neubig; Sebastian Ruder; Shruti Rijhwani; Michael Beukman; Chester Palen-Michel; Constantine Lignos; Jesujoba Alabi; Shamsuddeen Muhammad; Peter Nabende; M Cheikh; Andiswa Bamba Dione; Rooweither Bukula; Mabuya; F P Bonaventure; Blessing Dossou; Happy Sibanda; Jonathan Buzaaba; Godson Mukiibi; Derguene Kalipe; Amelia Mbaye; Fatoumata Taylor; Chris Kabore; Anuoluwapo Chinenye Emezue; Perez Aremu; Catherine Ogayo; Edwin Gitau; Victoire Munkoh-Buabeng; Memdjokam Koagne; Auguste Allahsera; Tebogo Tapo; Vukosi Macucwa; Mboning Marivate; Tajuddeen Tchiaze Elvis; Tosin Gwadabe; Orevaoghene Adewumi; Joyce Ahia; Neo Nakatumba-Nabende; Ignatius Lerato Mokono; Chiamaka Ezeani; Chukwuneke; Oluwaseun Mofetoluwa; Gilles Adeyemi; Idris Quentin Hacheme; Odunayo Abdulmumin; Oreen Ogundepo; Tatiana Yousuf; Dietrich Moteu; Klakow",
"journal": "Association for Computational Linguistics",
"ref_id": "b0",
"title": "MasakhaNER 2.0: Africa-centric transfer learning for named entity recognition",
"year": "2022"
},
{
"authors": "David Ifeoluwa Adelani; Jade Abbott; Graham Neubig; D' Daniel; Julia Souza; Constantine Kreutzer; Chester Lignos; Happy Palen-Michel; Shruti Buzaaba; Sebastian Rijhwani; Stephen Ruder; Israel Mayhew; Shamsuddeen H Abebe Azime; Chris Chinenye Muhammad; Joyce Emezue; Perez Nakatumba-Nabende; Aremu Ogayo; Catherine Anuoluwapo; Derguene Gitau; Jesujoba Mbaye; Seid Alabi; Tajuddeen Muhie Yimam; Ignatius Rabiu Gwadabe; Ezeani; Andre Rubungo; Jonathan Niyongabo; Verrah Mukiibi; Iroro Otiende; Davis Orife; Samba David; Tosin Ngom; Paul Adewumi; Mofetoluwa Rayson; Gerald Adeyemi; Emmanuel Muriuki; Chiamaka Anebi; Nkiruka Chukwuneke; Eric Odu; Samuel Peter Wairagala; Clemencia Oyerinde; Tobius Siro; Temilola Saul Bateesa; Yvonne Oloyede; Victor Wambui; Deborah Akinode; Maurice Nabagereka; Ayodele Katusiime; Awokoya; Mboup Mouhamadane; Dibora Gebreyohannes; Henok Tilaye; Kelechi Nwaike; Degaga Wolde; Abdoulaye Faye; Blessing Sibanda; Orevaoghene Ahia; F P Bonaventure; Kelechi Dossou; Thierno Ogueji; Diop Ibrahima; Abdoulaye Diallo; Adewale Akinfaderin; Tendai Marengereke; Salomey Osei",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b1",
"title": "MasakhaNER: Named entity recognition for African languages",
"year": "2021"
},
{
"authors": "Denis Anson; Penni Moist; Mary Przywara; Heather Wells; Heather Saylor; Hantz Maxime",
"journal": "Assistive Technology",
"ref_id": "b2",
"title": "The effects of word completion and word prediction on typing rates using on-screen keyboards",
"year": "2006"
},
{
"authors": "Mikel Artetxe; Sebastian Ruder; Dani Yogatama; Gorka Labaka; Eneko Agirre",
"journal": "Association for Computational Linguistics",
"ref_id": "b3",
"title": "A call for more rigor in unsupervised cross-lingual learning",
"year": "2020"
},
{
"authors": "Akari Asai; Jungo Kasai; Jonathan Clark; Kenton Lee; Eunsol Choi; Hannaneh Hajishirzi",
"journal": "Association for Computational Linguistics",
"ref_id": "b4",
"title": "XOR QA: Cross-lingual open-retrieval question answering",
"year": "2021"
},
{
"authors": "Ankur Bapna; Isaac Caswell; Julia Kreutzer; Orhan Firat; Daan Van Esch; Aditya Siddhant; Mengmeng Niu; Pallavi Baljekar; Xavier Garcia; Wolfgang Macherey; Theresa Breiner; Vera Axelrod; Jason Riesa; Yuan Cao; Mia Xu Chen; Klaus Macherey; Maxim Krikun; Pidong Wang; Alexander Gutkin; Apurva Shah; Yanping Huang; Zhifeng Chen; Yonghui Wu; Macduff Hughes",
"journal": "",
"ref_id": "b5",
"title": "Building machine translation systems for the next thousand languages",
"year": "2022"
},
{
"authors": "Youssef Bassil; Mohammad Alwani",
"journal": "IJACSA) International Journal of Advanced Computer Science and Applications",
"ref_id": "b6",
"title": "Post-Editing Error Correction Algorithm For Speech Recognition using Bing Spelling Suggestion",
"year": "2012"
},
{
"authors": "Steven Bird",
"journal": "Association for Computational Linguistics",
"ref_id": "b7",
"title": "Local languages, third spaces, and other high-resource scenarios",
"year": "2022"
},
{
"authors": "Damian Blasi; Antonios Anastasopoulos; Graham Neubig",
"journal": "Association for Computational Linguistics",
"ref_id": "b8",
"title": "Systematic inequalities in language technology performance across the world's languages",
"year": "2022"
},
{
"authors": "Isaac Caswell; Theresa Breiner; Daan Van Esch; Ankur Bapna",
"journal": "International Committee on Computational Linguistics",
"ref_id": "b9",
"title": "Language ID in the wild: Unexpected challenges on the path to a thousand-language web text corpus",
"year": "2020"
},
{
"authors": "Zhehuai Chen; Ankur Bapna; Andrew Rosenberg; Yu Zhang; Bhuvana Ramabhadran; Pedro Moreno; Nanxin Chen",
"journal": "IEEE",
"ref_id": "b10",
"title": "Maestro-U: Leveraging joint speech-text representation learning for zero supervised speech ASR",
"year": "2023"
},
{
"authors": "Aakanksha Chowdhery; Sharan Narang; Jacob Devlin; Maarten Bosma; Gaurav Mishra; Adam Roberts; Paul Barham; Hyung Won Chung; Charles Sutton; Sebastian Gehrmann; Parker Schuh; Kensen Shi; Sasha Tsvyashchenko; Joshua Maynez; Abhishek Rao; Parker Barnes; Yi Tay; Noam Shazeer; Emily Vinodkumar Prabhakaran; Nan Reif; Ben Du; Reiner Hutchinson; James Pope; Jacob Bradbury; Michael Austin; Guy Isard; Pengcheng Gur-Ari; Toju Yin; Anselm Duke; Sanjay Levskaya; Sunipa Ghemawat; Henryk Dev; Xavier Michalewski; Vedant Garcia; Kevin Misra; Liam Robinson; Denny Fedus; Daphne Zhou; David Ippolito; Hyeontaek Luan; Barret Lim; Alexander Zoph; Ryan Spiridonov; David Sepassi; Shivani Dohan; Mark Agrawal; Andrew M Omernick; Thanumalayan Dai; Marie Sankaranarayana Pillai; Aitor Pellat; Erica Lewkowycz; Rewon Moreira; Oleksandr Child; Katherine Polozov; Zongwei Lee; Xuezhi Zhou; Brennan Wang; Mark Saeta; Orhan Diaz; Michele Firat; Jason Catasta; Kathy Wei; Douglas Meier-Hellstern; Jeff Eck; Slav Dean; Noah Petrov; Fiedel",
"journal": "",
"ref_id": "b11",
"title": "PaLM: Scaling language modeling with Pathways",
"year": "2022"
},
{
"authors": "Chung Hyung Won; Le Hou; Shayne Longpre; Barret Zoph; Yi Tay; William Fedus; Yunxuan Li; Xuezhi Wang; Mostafa Dehghani; Siddhartha Brahma; Albert Webson; Shane Shixiang; Zhuyun Gu; Mirac Dai; Xinyun Suzgun; Aakanksha Chen; Alex Chowdhery; Marie Castro-Ros; Kevin Pellat; Dasha Robinson; Sharan Valter; Gaurav Narang; Adams Mishra; Vincent Yu; Yanping Zhao; Andrew Huang; Hongkun Dai; Slav Yu; Ed H Petrov; Jeff Chi; Jacob Dean; Adam Devlin; Denny Roberts; Quoc V Zhou; Jason Le; Wei",
"journal": "",
"ref_id": "b12",
"title": "Scaling instruction-finetuned language models",
"year": "2022"
},
{
"authors": "Jonathan H Clark; Eunsol Choi; Michael Collins; Dan Garrette; Tom Kwiatkowski; Vitaly Nikolaev; Jennimaria Palomaki",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b13",
"title": "TyDi QA: A benchmark for information-seeking question answering in typologically diverse languages",
"year": "2020"
},
{
"authors": "Alexis Conneau; Min Ma; Simran Khanuja; Yu Zhang; Vera Axelrod; Siddharth Dalmia; Jason Riesa; Clara Rivera; Ankur Bapna",
"journal": "IEEE",
"ref_id": "b14",
"title": "FLEURS: Few-shot learning evaluation of universal representations of speech",
"year": "2023"
},
{
"authors": "Marie-Catherine De Marneffe; Christopher D Manning; Joakim Nivre; Daniel Zeman",
"journal": "Computational Linguistics",
"ref_id": "b15",
"title": "Universal Dependencies",
"year": "2021"
},
{
"authors": "Vivek Dhakal; Anna Maria Feit; Per Ola Kristensson; Antti Oulasvirta",
"journal": "Association for Computing Machinery (ACM",
"ref_id": "b16",
"title": "Observations on typing from 136 million keystrokes",
"year": "2018"
},
{
"authors": "Bosheng Ding; Junjie Hu; Lidong Bing; Mahani Aljunied; Shafiq Joty; Luo Si; Chunyan Miao",
"journal": "Association for Computational Linguistics",
"ref_id": "b17",
"title": "GlobalWoZ: Globalizing MultiWoZ to develop multilingual task-oriented dialogue systems",
"year": "2022"
},
{
"authors": "Rui Dong; David Smith",
"journal": "Association for Computational Linguistics",
"ref_id": "b18",
"title": "Multi-input attention for unsupervised OCR correction",
"year": "2018"
},
{
"authors": "Andrew Drozdov; Nathanael Schärli; Ekin Akyürek; Nathan Scales; Xinying Song; Xinyun Chen; Olivier Bousquet; Denny Zhou",
"journal": "",
"ref_id": "b19",
"title": "Compositional semantic parsing with large language models",
"year": "2023"
},
{
"authors": "Jack Fitzgerald; Christopher Hench; Charith Peris; Kay Rottmann",
"journal": "Association for Computational Linguistics",
"ref_id": "b20",
"title": "Massively multilingual natural language understanding 2022 (MMNLU-22) workshop and competition",
"year": "2022"
},
{
"authors": "Goran Glavaš; Ivan Vulić",
"journal": "Association for Computational Linguistics",
"ref_id": "b21",
"title": "Is supervised syntactic parsing beneficial for language understanding tasks? an empirical investigation",
"year": "2021"
},
{
"authors": "Rahul Goel; Waleed Ammar; Aditya Gupta; Siddharth Vashishtha; Motoki Sano; Faiz Surani; Max Chang; Hyunjeong Choe; David Greene; Kyle He; Rattima Nitisaroj; Anna Trukhina; Shachi Paul; Pararth Shah; Rushin Shah; Zhou Yu",
"journal": "",
"ref_id": "b22",
"title": "PRESTO: A multilingual dataset for parsing realistic task-oriented dialogs",
"year": "2023"
},
{
"authors": "Naman Goyal; Cynthia Gao; Vishrav Chaudhary; Peng-Jen Chen; Guillaume Wenzek; Da Ju; Sanjana Krishnan; Marc'aurelio Ranzato; Francisco Guzmán; Angela Fan",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b23",
"title": "The Flores-101 evaluation benchmark for low-resource and multilingual machine translation",
"year": "2022"
},
{
"authors": "Alexander Gutkin; Cibu Johny; Raiomond Doctor; Brian Roark; Richard Sproat",
"journal": "Association for Computational Linguistics",
"ref_id": "b24",
"title": "Beyond Arabic: Software for Perso-Arabic script manipulation",
"year": "2022"
},
{
"authors": "Barry Haddow; Rachel Bawden; Antonio Valerio Miceli; Jindřich Barone; Alexandra Helcl; Birch",
"journal": "Computational Linguistics",
"ref_id": "b25",
"title": "Survey of low-resource machine translation",
"year": "2022"
},
{
"authors": "Harald Hammarström; Shafqat Mumtaz Virk; Markus Forsberg",
"journal": "Association for Computing Machinery",
"ref_id": "b26",
"title": "Poor man's OCR postcorrection: Unsupervised recognition of variant spelling applied to a multilingual document collection",
"year": "2017"
},
{
"authors": "Tahmid Hasan; Abhik Bhattacharjee; Md Saiful Islam; Kazi Mubasshir; Yuan-Fang Li; Yong-Bin Kang; M Sohel Rahman; Rifat Shahriyar",
"journal": "Association for Computational Linguistics",
"ref_id": "b27",
"title": "XLsum: Large-scale multilingual abstractive summarization for 44 languages",
"year": "2021"
},
{
"authors": "A Michael; David Hedderich; Dawei Adelani; Jesujoba Zhu; Udia Alabi; Dietrich Markus; Klakow",
"journal": "Association for Computational Linguistics",
"ref_id": "b28",
"title": "Transfer learning and distant supervision for multilingual transformer models: A study on African languages",
"year": "2020"
},
{
"authors": "Junjie Hu; Sebastian Ruder; Aditya Siddhant; Graham Neubig; Orhan Firat; Melvin Johnson",
"journal": "PMLR",
"ref_id": "b29",
"title": "XTREME: A massively multilingual multitask benchmark for evaluating cross-lingual generalisation",
"year": "2020"
},
{
"authors": "Oana Ignat; Jean Maillard; Vishrav Chaudhary; Francisco Guzmán",
"journal": "",
"ref_id": "b30",
"title": "OCR improves machine translation for low-resource languages",
"year": "2022"
},
{
"authors": "",
"journal": "International Organization for Standardization",
"ref_id": "b31",
"title": "Codes for the representation of names of scripts",
"year": ""
},
{
"authors": "Paul Jaccard",
"journal": "Bulletin de la Societe Vaudoise des Sciences Naturelles",
"ref_id": "b32",
"title": "Nouvelles recherches sur la distribution florale",
"year": "1908"
},
{
"authors": "Cibu Johny; Lawrence Wolf-Sonkin; Alexander Gutkin; Brian Roark",
"journal": "Association for Computational Linguistics",
"ref_id": "b33",
"title": "Finite-state script normalization and processing utilities: The Nisaba Brahmic library",
"year": "2021"
},
{
"authors": "Pratik Joshi; Sebastin Santy; Amar Budhiraja; Kalika Bali; Monojit Choudhury",
"journal": "Association for Computational Linguistics",
"ref_id": "b34",
"title": "The state and fate of linguistic diversity and inclusion in the NLP world",
"year": "2020"
},
{
"authors": "Simran Khanuja; Sebastian Ruder; Partha Talukdar",
"journal": "Association for Computational Linguistics",
"ref_id": "b35",
"title": "Evaluating Inclusivity, Equity, and Accessibility of NLP Technology: A Case Study for Indian Languages",
"year": "2023"
},
{
"authors": "Shreya Khare; Ashish R Mittal; Anuj Diwan; Sunita Sarawagi; Preethi Jyothi; Samarth Bharadwaj",
"journal": "ternational Speech Communication Association",
"ref_id": "b36",
"title": "Low resource ASR: The surprising effectiveness of high resource transliteration",
"year": "2021"
},
{
"authors": "Julia Kreutzer; Isaac Caswell; Lisa Wang; Ahsan Wahab; Daan Van Esch; Nasanbayar Ulzii-Orshikh; Allahsera Tapo; Nishant Subramani; Artem Sokolov; Claytone Sikasote; Monang Setyawan; Supheakmungkol Sarin; Sokhar Samb; Benoît Sagot; Clara Rivera; Annette Rios; Isabel Papadimitriou; Salomey Osei; Pedro Ortiz Suarez; Iroro Orife; Kelechi Ogueji; Andre Niyongabo Rubungo; Toan Q Nguyen; Mathias Müller; André Müller; Hassan Shamsuddeen; Nanda Muhammad; Ayanda Muhammad; Jamshidbek Mnyakeni; Tapiwanashe Mirzakhalov; Colin Matangira; Nze Leong; Sneha Lawson; Yacine Kudugunta; Mathias Jernite; Orhan Jenny; Firat; F P Bonaventure; Sakhile Dossou; Dlamini; Sakine Nisansa De Silva; Stella Çabuk Ballı; Alessia Biderman; Ahmed Battisti; Ankur Baruwa; Pallavi Bapna; Baljekar; Ayodele Israel Abebe Azime; Duygu Awokoya; Orevaoghene Ataman; Oghenefego Ahia; Sweta Ahia; Mofetoluwa Agrawal; Adeyemi",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b37",
"title": "Quality at a glance: An audit of web-crawled multilingual datasets",
"year": "2022"
},
{
"authors": "Sameer Kumar; Victor Bittorf; Dehao Chen; Chiachen Chou; Blake Hechtman; Hyoukjoong Lee; Naveen Kumar; Peter Mattson; Shibo Wang; Tao Wang; Yuanzhong Xu; Zongwei Zhou",
"journal": "",
"ref_id": "b38",
"title": "Scale MLPerf-0.6 models on Google TPU-v3 pods",
"year": "2019"
},
{
"authors": "Tom Kwiatkowski; Jennimaria Palomaki; Olivia Redfield; Michael Collins; Ankur Parikh; Chris Alberti; Danielle Epstein; Illia Polosukhin; Jacob Devlin; Kenton Lee; Kristina Toutanova; Llion Jones; Matthew Kelcey; Ming-Wei Chang; Andrew M Dai; Jakob Uszkoreit; Quoc Le; Slav Petrov",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b39",
"title": "Natural questions: A benchmark for question answering research",
"year": "2019"
},
{
"authors": "Anne Lauscher; Vinit Ravishankar; Ivan Vulić; Goran Glavaš",
"journal": "Association for Computational Linguistics",
"ref_id": "b40",
"title": "From zero to hero: On the limitations of zero-shot language transfer with multilingual Transformers",
"year": "2020"
},
{
"authors": "Haoran Li; Abhinav Arora; Shuohui Chen; Anchit Gupta; Sonal Gupta; Yashar Mehdad",
"journal": "Association for Computational Linguistics",
"ref_id": "b41",
"title": "MTOP: A comprehensive multilingual task-oriented semantic parsing benchmark",
"year": "2021"
},
{
"authors": "Jing Li; Aixin Sun; Jianglei Han; Chenliang Li",
"journal": "IEEE Transactions on Knowledge and Data Engineering",
"ref_id": "b42",
"title": "A survey on deep learning for named entity recognition",
"year": "2020"
},
{
"authors": "Yaobo Liang; Nan Duan; Yeyun Gong; Ning Wu; Fenfei Guo; Weizhen Qi; Ming Gong; Linjun Shou; Daxin Jiang; Guihong Cao; Xiaodong Fan; Ruofei Zhang; Rahul Agrawal; Edward Cui; Sining Wei; Taroon Bharti; Ying Qiao; Jiun-Hung Chen; Winnie Wu; Shuguang Liu; Fan Yang; Daniel Campos; Rangan Majumder; Ming Zhou",
"journal": "Association for Computational Linguistics",
"ref_id": "b43",
"title": "XGLUE: A new benchmark dataset for cross-lingual pre-training, understanding and generation",
"year": "2020"
},
{
"authors": "Zhaojiang Lin; Andrea Madotto; Genta Indra Winata; Peng Xu; Feijun Jiang; Yuxiang Hu; Chen Shi; Pascale Fung",
"journal": "",
"ref_id": "b44",
"title": "BiToD: A bilingual multidomain dataset for task-oriented dialogue modeling",
"year": "2021"
},
{
"authors": "Olga Majewska; Evgeniia Razumovskaia; Maria Edoardo; Ivan Ponti; Anna Vulić; Korhonen",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b45",
"title": "Cross-lingual dialogue dataset creation via outline-based generation",
"year": "2023"
},
{
"authors": "Sabrina J Mielke",
"journal": "",
"ref_id": "b46",
"title": "Can you compare perplexity across different segmentations? Available in",
"year": "2019"
},
{
"authors": "Barbu Verginica; Maria Mititelu; Mitrofan",
"journal": "Romanian Association of Computational Linguistics",
"ref_id": "b47",
"title": "The Romanian medical treebank -SiMoNERo",
"year": "2020"
},
{
"authors": "Nikita Moghe; Evgeniia Razumovskaia; Liane Guillou; Ivan Vulić; Anna Korhonen; Alexandra Birch",
"journal": "",
"ref_id": "b48",
"title": "MULTI3NLU++: A multilingual, multi-intent, multi-domain dataset for natural language understanding in task-oriented dialogue",
"year": "2022"
},
{
"authors": "Peya Mowar; Tanuja Ganu; Saikat Guha",
"journal": "",
"ref_id": "b49",
"title": "Towards optimizing OCR for accessibility",
"year": "2022"
},
{
"authors": "Shamsuddeen Hassan; Muhammad ; Idris Abdulmumin; Abinew Ali Ayele; Nedjma Ousidhoum; David Ifeoluwa Adelani; Seid Muhie Yimam; Ibrahim Sa'id Ahmad; Meriem Beloucif; Saif Mohammad; Sebastian Ruder; Oumaima Hourrane; Pavel Brazdil; Dário Felermino; António Mário; Davis Ali; Salomey Davis; Osei; Shehu Bello; Falalu Bello; Tajuddeen Ibrahim; Samuel Gwadabe; Tadesse Rutunda; Wendimu Belay; Hailu Baye Messelle; Sisay Beshada Balcha; Hagos Tesfahun Adugna Chala; Bernard Gebremichael; Steven Opoku; Arthur",
"journal": "",
"ref_id": "b50",
"title": "AfriSenti: A Twitter sentiment analysis benchmark for African languages",
"year": "2023"
},
{
"authors": "Shamsuddeen Hassan; Muhammad ; David Ifeoluwa Adelani; Sebastian Ruder; Ibrahim Sa'id Ahmad; Idris Abdulmumin; Shehu Bello; Monojit Bello; Chris Choudhury; Saheed Chinenye Emezue; Anuoluwapo Salahudeen Abdullahi; Alípio Aremu; Pavel Jorge; Brazdil",
"journal": "European Language Resources Association",
"ref_id": "b51",
"title": "NaijaSenti: A Nigerian Twitter sentiment corpus for multilingual sentiment analysis",
"year": "2022"
},
{
"authors": "Nibal Nayef; Fei Yin; Imen Bizid; Hyunsoo Choi; Yuan Feng; Dimosthenis Karatzas; Zhenbo Luo; Umapada Pal; Christophe Rigaud; Joseph Chazalon",
"journal": "IEEE",
"ref_id": "b52",
"title": "ICDAR2017 robust reading challenge on multi-lingual scene text detection and script identification -RRC-MLT",
"year": "2017"
},
{
"authors": "Wilhelmina Nekoto; Vukosi Marivate; Tshinondiwa Matsila; Timi Fasubaa; Taiwo Fagbohungbe; Solomon Oluwole Akinola; Shamsuddeen Muhammad; Salomon Kabongo Kabenamualu; Salomey Osei; Freshia Sackey; Andre Rubungo; Ricky Niyongabo; Perez Macharm; Orevaoghene Ogayo; Musie Ahia; Mofetoluwa Meressa Berhe; Masabata Adeyemi; Lawrence Mokgesi-Selinga; Laura Okegbemi; Kolawole Martinus; Kevin Tajudeen; Kelechi Degila; Kathleen Ogueji; Julia Siminyu; Jason Kreutzer; Jamiil Toure Webster; Jade Ali; Iroro Abbott; Ignatius Orife; Ezeani; Abdulkadir Idris; Herman Dangana; Hady Kamper; Goodness Elsahar; Ghollah Duru; Murhabazi Kioko; Elan Espoir; Daniel Van Biljon; Christopher Whitenack; Chris Chinenye Onyefuluchi; Emezue; F P Bonaventure; Blessing Dossou; Blessing Sibanda; Ayodele Bassey; Arshath Olabiyi; Alp Ramkilowan; Adewale Öktem; Abdallah Akinfaderin; Bashir",
"journal": "Association for Computational Linguistics",
"ref_id": "b53",
"title": "Participatory research for low-resourced machine translation: A case study in African languages",
"year": "2020"
},
{
"authors": "Massimo Nicosia; Francesco Piccinno",
"journal": "Association for Computational Linguistics",
"ref_id": "b54",
"title": "Bytelevel massively multilingual semantic parsing",
"year": "2022"
},
{
"authors": "Joakim Nivre; Marie-Catherine De Marneffe; Filip Ginter; Jan Hajič; Christopher D Manning; Sampo Pyysalo; Sebastian Schuster; Francis Tyers; Daniel Zeman",
"journal": "European Language Resources Association",
"ref_id": "b55",
"title": "Universal Dependencies v2: An evergrowing multilingual treebank collection",
"year": "2020"
},
{
"authors": "Andre Rubungo; Qu Niyongabo; Julia Hong; Li Kreutzer; Huang",
"journal": "International Committee on Computational Linguistics",
"ref_id": "b56",
"title": "KINNEWS and KIRNEWS: Benchmarking cross-lingual text classification for Kinyarwanda and Kirundi",
"year": "2020"
},
{
"authors": "Sebastian Nordhoff; Harald Hammarström",
"journal": "",
"ref_id": "b57",
"title": "Glottolog/langdoc: Defining dialects, languages, and language families as collections of resources",
"year": "2011"
},
{
"authors": "Addison Phillips; Mark Davis",
"journal": "",
"ref_id": "b58",
"title": "BCP 47 -Tags for Identifying Languages",
"year": "2009"
},
{
"authors": "Reiner Pope; Sholto Douglas; Aakanksha Chowdhery; Jacob Devlin; James Bradbury; Anselm Levskaya; Jonathan Heek; Kefan Xiao; Shivani Agrawal; Jeff Dean",
"journal": "",
"ref_id": "b59",
"title": "Efficiently scaling transformer inference",
"year": "2022"
},
{
"authors": "Maja Popović",
"journal": "Association for Computational Linguistics",
"ref_id": "b60",
"title": "chrF: character n-gram F-score for automatic MT evaluation",
"year": "2015"
},
{
"authors": "Pranav Rajpurkar; Robin Jia; Percy Liang",
"journal": "Association for Computational Linguistics",
"ref_id": "b61",
"title": "Know what you don't know: Unanswerable questions for SQuAD",
"year": "2018"
},
{
"authors": "Christophe Rigaud; Antoine Doucet; Mickaël Coustaty; Jean-Philippe Moreux",
"journal": "IEEE",
"ref_id": "b62",
"title": "ICDAR 2019 competition on post-OCR text correction",
"year": "2019"
},
{
"authors": "Shruti Rijhwani; Antonios Anastasopoulos; Graham Neubig",
"journal": "",
"ref_id": "b63",
"title": "OCR Post Correction for Endangered Language Texts",
"year": "2020"
},
{
"authors": "Shruti Rijhwani; Daisy Rosenblum; Antonios Anastasopoulos; Graham Neubig",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b64",
"title": "Lexically aware semi-supervised learning for OCR postcorrection",
"year": "2021"
},
{
"authors": "Brian Roark; Lawrence Wolf-Sonkin; Christo Kirov; Sabrina J Mielke; Cibu Johny; Isin Demirsahin; Keith Hall",
"journal": "European Language Resources Association",
"ref_id": "b65",
"title": "Processing South Asian languages written in the Latin script: the Dakshina dataset",
"year": "2020"
},
{
"authors": "Adam Roberts; Hyung Won Chung; Anselm Levskaya; Gaurav Mishra; James Bradbury; Daniel Andor; Sharan Narang; Brian Lester; Colin Gaffney; Afroz Mohiuddin; Curtis Hawthorne; Aitor Lewkowycz; Alex Salcianu; Jacob Marc Van Zee; Sebastian Austin; Livio Baldini Goodman; Haitang Soares; Sasha Hu; Aakanksha Tsvyashchenko; Jasmijn Chowdhery; Jannis Bastings; Xavier Bulian; Jianmo Garcia; Andrew Ni; Kathleen Chen; Jonathan H Kenealy; Stephan Clark; Dan Lee; James Garrette; Colin Lee-Thorp; Noam Raffel; Marvin Shazeer; Maarten Ritter; Alexandre Bosma; Jeremy Passos; Noah Maitin-Shepard; Mark Fiedel; Brennan Omernick; Ryan Saeta; Alexander Sepassi; Joshua Spiridonov; Andrea Newlan; Gesmundo",
"journal": "",
"ref_id": "b66",
"title": "Scaling up models and data with t5x and seqio",
"year": "2022"
},
{
"authors": "Sebastian Ruder; Noah Constant; Jan Botha; Aditya Siddhant; Orhan Firat; Jinlan Fu; Pengfei Liu; Junjie Hu; Dan Garrette; Graham Neubig; Melvin Johnson",
"journal": "Association for Computational Linguistics",
"ref_id": "b67",
"title": "XTREME-R: Towards more challenging and nuanced multilingual evaluation",
"year": "2021"
},
{
"authors": "Amarjot Singh; Ketan Bacchuwar; Akshay Bhasin",
"journal": "International Journal of Machine Learning and Computing",
"ref_id": "b68",
"title": "A survey of OCR applications",
"year": "2012"
},
{
"authors": "Martin Sundermeyer; Ralf Schlüter; Hermann Ney",
"journal": "International Speech Communication Association",
"ref_id": "b69",
"title": "LSTM neural networks for language modeling",
"year": "2012"
},
{
"authors": "Cynthia Tam; David Wells",
"journal": "Assistive Technology",
"ref_id": "b70",
"title": "Evaluating the benefits of displaying word prediction lists on a personal digital assistant at the keyboard level",
"year": "2009"
},
{
"authors": "Erik F Tjong; Kim Sang; Fien De; Meulder ",
"journal": "",
"ref_id": "b71",
"title": "Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition",
"year": "2003"
},
{
"authors": "Daniel Varab; Natalie Schluter",
"journal": "Association for Computational Linguistics",
"ref_id": "b72",
"title": "Mas-siveSumm: a very large-scale, very multilingual, news summarisation dataset",
"year": "2021"
},
{
"authors": "H Hans; Wellisch",
"journal": "John Wiley & Sons",
"ref_id": "b73",
"title": "The Conversion of Scripts: Its Nature, History, and Utilization",
"year": "1978"
},
{
"authors": "Ken Whistler",
"journal": "",
"ref_id": "b74",
"title": "Unicode normalization forms",
"year": "2021"
},
{
"authors": "Genta Indra Winata; Alham Fikri Aji; Samuel Cahyawijaya; Rahmad Mahendra; Fajri Koto; Ade Romadhony; Kemal Kurniawan; David Moeljadi; Radityo Eko Prasojo; Pascale Fung; Timothy Baldwin; Jey ; Han Lau; Rico Sennrich; Sebastian Ruder",
"journal": "Association for Computational Linguistics",
"ref_id": "b75",
"title": "NusaX: Multilingual Parallel Sentiment Dataset for 10 Indonesian Local Languages",
"year": "2023"
},
{
"authors": "Linting Xue; Aditya Barua; Noah Constant; Rami Al-Rfou; Sharan Narang; Mihir Kale; Adam Roberts; Colin Raffel",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b76",
"title": "ByT5: Towards a token-free future with pre-trained byte-to-byte models",
"year": "2022"
},
{
"authors": "Linting Xue; Noah Constant; Adam Roberts; Mihir Kale; Rami Al-Rfou; Aditya Siddhant; Colin Barua; Raffel",
"journal": "Association for Computational Linguistics",
"ref_id": "b77",
"title": "mT5: A massively multilingual pre-trained text-to-text transformer",
"year": "2021"
},
{
"authors": "Mengjie Zhao; Yi Zhu; Ehsan Shareghi; Ivan Vulić; Roi Reichart; Anna Korhonen; Hinrich Schütze",
"journal": "Association for Computational Linguistics",
"ref_id": "b78",
"title": "a. A closer look at few-shot crosslingual transfer: The choice of shots matters",
"year": "2021"
},
{
"authors": "Zihao Zhao; Eric Wallace; Shi Feng; Dan Klein; Sameer Singh",
"journal": "PMLR",
"ref_id": "b79",
"title": "Calibrate before use: Improving few-shot performance of language models",
"year": "2021"
}
] | [
{
"formula_coordinates": [
18,
74.86,
200.05,
434.95,
565.55
],
"formula_id": "formula_0",
"formula_text": "3 ✓ ✓ ✓ Ewe ee Atlantic-Congo 1 1 ✓ Greek el Indo-European 3 3 ✓ ✓ ✓ Estonian et Uralic 3 3 ✓ ✓ ✓ Fula ff Atlantic-Congo 2 1 ✓ ✓ Filipino fil Austronesian 2 1 ✓ ✓ Irish ga Indo-European 4 2 ✓ ✓ ✓ Galician gl Indo-European 3 3 ✓ ✓ ✓ Gujarati gu Indo-European 4 1 ✓ ✓ ✓ Hausa ha Afro-Asiatic 5 2 ✓ ✓ ✓ ✓ Hebrew he Afro-Asiatic 3 3 ✓ ✓ ✓ Armenian hy Indo-European 4 1 ✓ ✓ ✓ Indonesian id Austronesian 5 3 ✓ ✓ ✓ ✓ ✓ Igbo ig Atlantic-Congo 4 1 ✓ ✓ ✓ Icelandic is Indo-European 4 2 ✓ ✓ ✓ Javanese jv Austronesian 3 1 ✓ ✓ Georgian ka Kartvelian 2 3 ✓ ✓ Kamba kam Atlantic-Congo 2 0 ✓ ✓ Kabuverdianu kea Indo-European 2 0 ✓ ✓ Kazakh kk Turkic 2 3 ✓ ✓ Khmer km Austroasiatic 3 1 ✓ ✓ Kannada kn Dravidian 5 1 ✓ ✓ ✓ ✓ Kyrgyz ky Turkic 3 1 ✓ ✓ Luxembourgish lb Indo-European 3 1 ✓ ✓ (Lu)Ganda lg Atlantic-Congo 4 1 ✓ ✓ ✓ Lingala ln Atlantic-Congo 3 1 ✓ ✓ Lao lo Tai-Kadai 3 2 ✓ ✓ Lithuanian lt Indo-European 3 3 ✓ ✓ ✓ (Dho)Luo luo Nilotic 3 0 ✓ ✓ ✓ Latvian lv Indo-European 3 3 ✓ ✓ ✓ Maori mi Austronesian 3 1 ✓ ✓ Macedonian mk Indo-European 3 1 ✓ ✓ Malayalam ml Dravidian 4 1 ✓ ✓ ✓ Mongolian mn Mongolic-Khitan 3 1 ✓ ✓ Mossi (Mooré) mos Atlantic-Congo 1 0 ✓ Marathi mr Indo-European 3 2 ✓ ✓ ✓ Malay ms Austronesian 2 3 ✓ ✓ Maltese mt Afro-Asiatic 2 2 ✓ ✓ Burmese my Sino-Tibetan 4 1 ✓ ✓ ✓ Nepali ne Indo-European 3 1 ✓ ✓ Norwegian no Indo-European 2 1 ✓ ✓ Northern Sotho nso Atlantic-Congo 3 1 ✓ ✓ Nyanja (Chichewa) ny Atlantic-Congo 4 1 ✓ ✓ ✓ Occitan oc Indo-European 2 1 ✓ ✓ Oromo om Afro-Asiatic 3 1 ✓ ✓ Oriya or Indo-European 2 1 ✓ ✓ Punjabi pa Indo-European 4 2 ✓ ✓ ✓ Nigerian Pidgin pcm Indo-European 2 0 ✓ ✓ Pashto ps Indo-European 3 1 ✓ ✓ Romanian ro Indo-European 3 3 ✓ ✓ ✓ Kinyarwanda rw Atlantic-Congo 1 1 ✓ Sanskrit sa Indo-European 1 2 ✓ Sindhi sd Indo-European 4 1 ✓ ✓ ✓ Sinhala si Indo-European 2 0 ✓ ✓ Slovak sk Indo-European 3 3 ✓ ✓ ✓ Slovenian sl Indo-European 3 3 ✓ ✓ ✓ Shona sn Atlantic-Congo 4 1 ✓ ✓ ✓ Somali so Afro-Asiatic 3 1 ✓ ✓ Swahili sw Atlantic-Congo 8 2 ✓ ✓ ✓ ✓ ✓ ✓ ✓ Tamil ta Dravidian 4 3 ✓ ✓ ✓ ✓ Telugu te Dravidian 6 1 ✓ ✓ ✓ ✓ ✓ Tajik tg Indo-European 3 1 ✓ ✓ Thai th Tai-Kadai 3 3 ✓ ✓ ✓ Tswana (Setswana) tn Atlantic-Congo 1 2 ✓ Twi tw Atlantic-Congo 1 1 ✓ Uyghur ug Turkic 1 1 ✓ Ukrainian uk Indo-European 3 3 ✓ ✓ ✓ Umbundu umb Atlantic-Congo 2 0 ✓ ✓ Urdu ur Indo-European 4 3 ✓ ✓ ✓ ✓ Uzbek uz Turkic 2 3 ✓ ✓ Wolof wo Atlantic-Congo 3 2 ✓ ✓ ✓ Xhosa xh Atlantic-Congo 4 2 ✓ ✓ ✓ Yoruba yo Atlantic-Congo 5 2 ✓ ✓ ✓ ✓ Zulu zu Atlantic-Congo 5 2 ✓ ✓ ✓ ✓"
}
] |
10.1145/383259.383266 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b2",
"b3",
"b6",
"b8",
"b9",
"b14",
"b17",
"b18",
"b19",
"b25",
"b26",
"b27",
"b20",
"b21",
"b20",
"b26",
"b10",
"b11",
"b13",
"b0",
"b1",
"b4",
"b5",
"b22",
"b23",
"b10",
"b11"
],
"table_ref": [],
"text": "Statistical Shape Modeling (SSM) or morphological analysis, is a widespread tool used to quantify anatomical shape variation given a population of segmented 3D anatomies. Quantifying such subtle shape differences has been crucial in providing individualized treatments in medical procedures, detecting morphological pathologies, and advancing the understanding of different diseases [3,4,7,9,10,16,[19][20][21][27][28][29].\nThe two principal shape representations for building SSMs and performing subsequent statistical analyses are deformation fields and landmarks. Deforma-arXiv:2305.11946v2 [cs.CV] 29 Dec 2023 tion fields encode implicit transformations between cohort samples and a predefined (or learned) atlas. In contrast, landmarks are explicit points spread on shape surfaces that correspond across the population [22,23]. Landmark-based representations have been used extensively due to their simplicity, computational efficiency, and interpretability for statistical analyses [22,28]. Some applications use manually defined landmarks, however, this is labor-intensive, not reproducible, and requires domain expertise (e.g., radiologists). Computational methods (e.g., minimum description length -MDL [14], particle-based shape modeling -PSM [11,12], and frameworks based on Large Deformation Diffeomorphic Metric Mapping [15]) for automatically placing dense correspondence points, aka point distribution models (PDMs), have shifted the SSM field to data-driven characterization of population-level variabilities that is objective, reproducible, and scalable. However, this efficiency suffers when intricate shape surfaces require thousands of points representing localized, convoluted shape features that may live between landmarks. Furthermore, existing methods for landmark-based SSM must go through laborious and computationally expensive steps that require anatomical and technical expertise, starting from anatomy segmentation, shape data preprocessing, and correspondence optimization, to generate PDMs from 3D images. Existing methods (e.g., [1,2,5,6,24,25]) have been able to use deep learning to assuage the arduous process of building a PDM but still require the construction of PDMs (e.g., using a computational method such as PSM [11,12]) to supervise its learning task, making these deep learning based methods restricted and biased towards the shape statistics captured by the SSM method that is used to construct their training data.\nTo address the shortcomings of existing models, we propose Image2SSM, a novel deep-learning-based approach for SSM directly from images that, given pairs of images and segmentations, can produce a statistical shape model using an implicit, continuous surface representation. Once trained, Image2SSM can produce PDMs of new images without the need for anatomy segmentations. Unlike existing deep learning-based methods for SSM from images, Image2SSM only requires image-segmentation pairs and alleviates the need for constructing PDM to supervise learning shape statistics from images. Image2SSM leverages an implicit, radial basis function (RBF)-based, representation of shapes to construct a self-supervised training signal by tasking the network to estimate a sparse set of control points and their respective suface normals that best approximate the underlying surface in the RBF sense. This novel application of RBFs to build SSMs allows statistical analyses on representative points/landmarks, their surface normals, and the shape surfaces themselves due to its compact, informative, yet comprehensive nature. Combined with deep networks to directly learn such a representation from images, this method ushers a next step towards fully end-toend SSM frameworks that can build better and less restrictive low-dimensional shape representations more conducive to SSM analysis. In summary, the proposed method for SSM has the following strengths.\n-Using a continuous, but compact surface representation instead of only landmarks that allows performing analyses on points, normals, and surfaces alike.\n-The RBF shape representation can adapt to the underlying surface geometry, spreading more landmarks over the more complex surface regions. -A deep learning approach that bypasses any conventional correspondence optimization to construct training data for supervision, requiring virtually no hyperparameter tuning or preprocessing steps. -This method uses accelerated computational resources to perform training and outperforms existing deep learning based methods that constructs PDMs from unsegmented images."
},
{
"figure_ref": [],
"heading": "Methods",
"publication_ref": [
"b15",
"b7",
"b24"
],
"table_ref": [],
"text": "Image2SSM is a deep learning method that learns to build an SSM for an anatomical structure of interest directly from unsegmented images. It is trained on a population of I-3D images I = {I i } I i=1 as input and is supervised by their respective segmentations S = {S i } I i=1 . Image2SSM learns an RBF-based shape representation, consisting of a set of J control points P = {P i } I i=1 , and their surface normals N = {N i } I i=1 for each input shape, where the i-th shape point distribution model (PDM) is denoted by\nP i = [p i,1 , p i,2 , • • • , p i,J ], the respec- tive surface normals are N i = [n i,1 , n i,2 , • • • , n i,J ],\nand p i,j , n i,j ∈ R 3 . The network is trained end-to-end to minimize a loss that (1) makes the learned control points and their surface normals adhere to the underlying surface, (2) approximates surface normals at each control point to encode a signed distance field to the surface, (3) promotes correspondence of these control points across shapes in the population, and (4) encourages a spread of control points on each surface that adapts to the underlying geometrical complexity. The learned control points define an anatomical mapping, or a metric, among the given shapes that enables quantifying subtle shape differences and performing shape statistics, for example, using principal component analysis (PCA) or other non-linear methods (e.g., [17]). More importantly, once trained, Image2SSM can generate PDMs for new unsegmented images, bypassing the conventional SSM workflow of the manual (or semi-automated) segmentation, data preprocessing, and correspondence optimization. Furthermore, the continuous, implicit nature of the RBF representation enables extracting a proxy geometry (e.g., surface mesh or signed distance transforms -SDFs) at an arbitrary resolution that can be rasterized trivially on graphics hardware [8,26].\nIn this section, we briefly elaborate on the RBF-shape representation, outline the network architecture, motivate the choices and design of the proposed losses, and detail the training protocol of Image2SSM."
},
{
"figure_ref": [
"fig_0"
],
"heading": "Representing shapes using RBFs",
"publication_ref": [
"b7",
"b24",
"b7",
"b24"
],
"table_ref": [],
"text": "Implicit surface representation based on radial basis functions, RBF-shape for short, has been proven effective at representing intricate shapes by leveraging both surface control points and normals to inform shape reconstructions [8,26]. It defines a set of control points at the zero-level set and a pair of off-surface points (aka dipoles) with a signed distance s and -s along the surface normal of each control point. This is illustrated in Figure 1. We refer to the set of control points and their dipoles as P i for shape i, where P i = [P i , P + i , P - i ] with p ± i,j = p i,j ± sn i,j . Using P i , we define the shape's implicit function, a function that can query a distance to the surface given a point x ∈ R 3 , as follows:\nf Pi,wi (x) = j∈ Pi w i,j ϕ(x, p i,j ) + c T i x + c 0 i (1\n)\nwhere ϕ is the chosen RBF basis function (e.g., the thin plate spline ϕ(x, y) = (|x -y| 2 ) 2 log(|x -y| 2 ), the biharmonic ϕ(x, y) = |x -y| 2 or the triharmonic ϕ(x, y) = (|x -y| 2 ) 3 ) and c i ∈ R 3 and c 0 i ∈ R encodes the linear trend of the surface. We obtain w i = [w i,1 , w i,2 , ..., w i,3J , c 0 i , c 1 i , c 2 i , c 3 i ] ∈ R 3J+4 by solving a system of equations formed by Eq 1 over x ∈ P i , along with constraints to keep the linear part separate from the nonlinear deformations captured by the RBF term (first term in Eq 1) to form a fully determined system. See [8,26] for more details. Ultimately, we can use this function f to query approximate distances to the surface to build a mesh or a signed distance transform for visualization and analysis.\nThis representation can better represent shapes with far fewer control points due to its built-in interpolation capabilities, even further enhanced by informing the system with the point normals. Furthermore, this continuous representation allows Image2SSM to adapt to the underlying surface geometry and correct for control point placement mistakes while training."
},
{
"figure_ref": [],
"heading": "Loss Functions",
"publication_ref": [
"b10",
"b11"
],
"table_ref": [],
"text": "Image2SSM uses four complementary loss functions to be trained on concurrently, illustrated in Figure 2. These are (i) surface loss, which aims to promote control point and normal adherence to the shape surface, (ii) normal loss, which attempts to learn the correct normals at each control point, (iii) correspondence loss, which aims to enforce positional correspondence across shapes, and (iv) sampling loss, which promotes a spread of the control points that best describes the underlying surface.\nSurface loss: This loss guides control points to lie on the surface. We use l 1norm to force control points to lie on the zero-level set of the distance transform D i by minimizing the absolute distance-to-the-surface evaluated at it. For the i-th shape, this loss is defined as,\nL surf Di (P i ) = J j=1 |D i (p i,j )|,(2)\nwhere D i (p i,j ) is the distance transform value at point p i,j .\nNormal loss: This loss aims to estimate the surface normal of each control point. This loss is supervised by the gradient of signed distance transforms (SDF) D = {D i } I i=1 , computed from the binary segmentations S, with respect to x, ∂D = {∂D i } I i=1 , which captures unnormalized surface normals. We use the cosine distance (in degrees) to penalize the deviation of the estimated normals from the normals computed from the distance transforms.\nL norm ∂Di (P i , N i ) = 180 π J j=1 cos -1 1 - n T i,j ∂D i (p i,j ) ∥n i,j ∥∥∂D i (p i,j )∥ .(3)\nCorrespondence loss: The notion of control points correspondence across the shape population can be quantified by the information content of the probability distribution induced by these control points in the shape space, the vector space defined by the shapes' PDMs [11,12]. The correspondence loss is triggered starting from the second epoch, where the mean shape µ = I i=1 P i is allowed to lag behind the update of the control points. Given a minibatch of size K, the correspondence loss is formulated using the differential entropy H of the samples in the minibatch, assuming a Gaussian distribution.\nL corres µ (P 1 , ..., P K ) = H(P) = 1 2 log 1 3JK K k=1 (P k -µ) (P k -µ) T ,(4)\nwhere P here indicates the random variable of the shape space.\nSampling loss: This loss makes f encode the signed distance to the surface while encouraging the control points to be adapted to the underlying geometry.\nHere, we randomly sample R-points\nB i = [b i,1 , ..., b i,R\n] that lie within a narrow band of thickness 2s around the surface (i.e., ±s from the zero-level set along the surface normal). The sampling loss minimizes distances between these narrow band points and the closest control point to each, scaled by the severity of the distance-to-surface approximation error. This objective guides control points to areas poorly described by f to progressively improve the signed distance-tosurface approximation and represent the shape more accurately. Let K i ∈ R R×M define the pairwise distances between each narrow band point b i,r and each control point p i,j for the i-th shape, where its r, j-th element k i r,j = ∥b i,r -p i,j ∥ 2 . Let softmin(K i ) encode the normalized (over P i ) spatial proximity of each narrow band point to each control point, where r, the jth element of softmin(K i ) is computed as exp (-k i r,j )/\nJ j ′ =1 exp(k i r,j ′ ). Let e i ∈ R R\n+ captures the RBF approximation squared error at the narrow band points, where\ne i,r = [f Pi,wi (b i,r ) -D i (b i,r )] 2 . Let E i = e i 1 T M\n, where 1 M is a ones-vector of size M . The samples loss can then be written as,"
},
{
"figure_ref": [],
"heading": "L sampl",
"publication_ref": [],
"table_ref": [],
"text": "Bi,Di,wi\n(P i , N i ) = mean (softmin(K i ) ⊗ K i ⊗ E i ) ,(5)\nwhere ⊗ indicates the Hadamard (elementwise) multiplication of matrices and mean computes the average over the matrix elements.\nImage2SSM loss: Given a minibatch of size K, the total loss of Image2SSM can be written as follows:\nL I,D,∂D (P K , N K ) = K i=1 αL surf Di (P i ) + βL norm ∂Di (P i , N i ) + γL sampl Bi,Di,wi (P i , N i ) + ζL corres µ (P 1 , ..., P K )(6)\nwhere α, β, γ, and ζ ∈ R + are weighting hyperparameters of the losses and P K , N K are the control points and normals of the samples in the minibatch. Figure 2 gives a full overview of the network and its interaction with the loses. Image2SSM's network is trained end-to-end with w i s detached from the training so that the loss does not back-propagate through the volatile linear solver."
},
{
"figure_ref": [],
"heading": "Results",
"publication_ref": [
"b10",
"b4",
"b5"
],
"table_ref": [],
"text": "We demonstrate Image2SSM's performance against the state-of-the-art correspondence optimization algorithm, namely the particle-based shape modeling (PSM), using its open-source implementation, ShapeWorks [11], and DeepSSM [5,6], a deep learning method that trains on an existing correspondence model (provided by the PSM in this case) to infer PDMs on new unsegmented images."
},
{
"figure_ref": [
"fig_1",
"fig_2",
"fig_2"
],
"heading": "Datasets",
"publication_ref": [
"b10",
"b12",
"b16"
],
"table_ref": [],
"text": "We run tests on a dataset consisting of 50 proximal femur CT scans devoid of pathologies in the form of image-segmentation pairs. The femurs are reflected Fig. 2. The Image2SSM architecture. A 3D image is fed to the convolutional backbone, which produces a flattened output for the feature extractor to produce control points and their respective normals. These are then used to compute the losses of the network.\nwhen appropriate and rigidly aligned to a common frame of reference. Due to space limitations, we also show similar results for a large-scale left atrium MRI dataset in the supplementary materials. For ease of comparison, we build SSMs with 128 particles for all algorithms as is sufficient to cover important femur shape features (femoral head with its fovea and the lesser and trochanter).\nStatistical Shape Model: We showcase Image2SSM in creating a statistical shape model on its training data and compare such a model with one optimized by PSM [11]. Figure 3 showcases the modes of variation, the surface-to-surface distances of Image2SSM against PSM, some representative reconstructions, and graphs for compactness (percentage of variance captured), specificity (ability to generate realistic shapes), and generalization (ability to represent unseen shape instance) [13]. We observe that the modes of variation and metrics match expectations in both approaches. We show the effectiveness of Image2SSM in adapting to surface details to achieve lower maximum surface-to-surface distance, and that, unlike PSM, we can achieve reasonable reconstructions using RBF-shape. More on adaptation to detail in figure 4. We implement Image2SSM in PyTorch and leverage the Autograd functionality to perform gradient descent using the Adam optimizer [18]. We randomly sample 10,000 3D points within the narrow band of each surface at each iteration. We use the biharmonic kernel for the basis function. However, the performance of Image2SSM is not significantly influenced by the kernel choice. The hyperparameters we use for Image2SSM are α = 1e 2 , β = 1e 2 , γ = 1e 4 , and ζ = 1e 3 for femurs and ζ = 1e 6 for left atria, which were determined based on the validation set. In practice, the runtime of Image2SSM is comparable to PSM for the femurs and roughly 2X faster for the left atria. Inference Results: We compare the inference capabilities of Image2SSM against DeepSSM on unseen test data. We train DeepSSM with a PDM generated by PSM as supervision. For a fair comparison, we use DeepSSM without its augmented data, since Image2SSM does not require augmentation to learn shape models. Nevertheless, it is possible to generate and train Image2SSM on augmented data with even more facility than with DeepSSM.Figure 4 shows that Image2SSM compares very favorably to DeepSSM qualitatively and in terms of surface-to-surface distance."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "Image2SSM is a novel deep-learning framework that both builds PDMs from image-segmentation pairs and predicts PDMs from unseen images. It uses an RBF-shape able to capture detail by leveraging surface normals at control points, and allows the SSM to adaptively permeate surfaces with high-level detail. Im-age2SSM represents another step forward in fully end-to-end PDMs and steers the field to utilizing more compact but comprehensive representations to achieve new analytical paradigms. Future directions include removing the requirement that the image-segmentation pairs must be rougly aligned across the cohort and relaxing the Gaussian assumption from correspondence enforcement. We also demonstrate our results on a dataset of 1018 aligned left atrium MRI image-segmentation pairs. This dataset is very challenging due to the high variability in the manual labeling of the pulmonary arteries and the presence of various atrial fibrillation phenotypes (Persistent, paroxysmal, AFL, nonAF, other arrhythmia) As before, we build the model with 128 particles. We show the first three modes of variation of Image2SSM compared to PSM. The results are comparable and match expectations. We observe that both models capture the shape variability of the atrium itself well, less so with the pulmonary arteries. The surface-to-surface distance comparison on all the data used to train Image2SSM. We observe that the distances are comparable between both models in that they capture a large array of shapes well, but fail to different degrees on severe outliers. (c) Shows the compactness (higher is better), specificity (lower is better) and generalization (lower is better) graphs against the number of modes of variation. These are also very similar between the two approaches. "
},
{
"figure_ref": [],
"heading": "",
"publication_ref": [],
"table_ref": [],
"text": "Acknowledgment The National Institutes of Health supported this work under grant numbers NIBIB-U24EB029011. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health."
}
] | Statistical shape modeling (SSM) is an essential tool for analyzing variations in anatomical morphology. In a typical SSM pipeline, 3D anatomical images, gone through segmentation and rigid registration, are represented using lower-dimensional shape features, on which statistical analysis can be performed. Various methods for constructing compact shape representations have been proposed, but they involve laborious and costly steps. We propose Image2SSM, a novel deep-learning-based approach for SSM that leverages image-segmentation pairs to learn a radial-basis-function (RBF)-based representation of shapes directly from images. This RBF-based shape representation offers a rich self-supervised signal for the network to estimate a continuous, yet compact representation of the underlying surface that can adapt to complex geometries in a data-driven manner. Image2SSM can characterize populations of biological structures of interest by constructing statistical landmark-based shape models of ensembles of anatomical shapes while requiring minimal parameter tuning and no user assistance. Once trained, Image2SSM can be used to infer low-dimensional shape representations from new unsegmented images, paving the way toward scalable approaches for SSM, especially when dealing with large cohorts. Experiments on synthetic and real datasets show the efficacy of the proposed method compared to the state-of-art correspondence-based method for SSM. | Image2SSM: Reimagining Statistical Shape Models from Images with Radial Basis Functions | [
{
"figure_caption": "Fig. 1 .1Fig. 1. (a) Concept of populating a surface using control points and the iso-surfaces using positive and negative pole points. (b) Same concept applied to an output threedimensional reconstructed femur. (c) Normals can be used to describe very distinct features of the greater trochanter.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 3 .3Fig. 3. (a) First and second modes of variation obtained from Image2SSM training data and PSM. (b) Surface-to-surface distance on a best, median, and worst training femur mesh. (c) The left image shows the surface-to-surface distance comparison on all the data used to train Image2SSM; the right shows it without outliers.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 4 .4Fig. 4. (a) Surface-to-surface distance on a reconstructed femur mesh from particles of a few test samples. (b) Surface-to-surface distance plot between DeepSSM and Im-age2SSM, and the same plot without the outlier femur. (c) Illustrates Image2SSM's capacity to capture detail on an unseen test image. (d) Shows the compactness (higher is better), specificity (lower is better) and generalization (lower is better) graphs against the number of modes of variation.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 5 .5Fig.5. We also demonstrate our results on a dataset of 1018 aligned left atrium MRI image-segmentation pairs. This dataset is very challenging due to the high variability in the manual labeling of the pulmonary arteries and the presence of various atrial fibrillation phenotypes (Persistent, paroxysmal, AFL, nonAF, other arrhythmia) As before, we build the model with 128 particles. We show the first three modes of variation of Image2SSM compared to PSM. The results are comparable and match expectations. We observe that both models capture the shape variability of the atrium itself well, less so with the pulmonary arteries.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 6 .6Fig. 6. (a) Surface-to-surface distance on a best, median, and worst training meshes. (b)The surface-to-surface distance comparison on all the data used to train Image2SSM. We observe that the distances are comparable between both models in that they capture a large array of shapes well, but fail to different degrees on severe outliers. (c) Shows the compactness (higher is better), specificity (lower is better) and generalization (lower is better) graphs against the number of modes of variation. These are also very similar between the two approaches.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 7 .7Fig. 7. (a) Surface-to-surface distance on best, median, and worst held-out samples. (b) Surface-to-surface distance plot between DeepSSM and Image2SSM. We observe that Image2SSM performs well compared to DeepSSM, but still fails to capture major outliers.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "7",
"figure_type": "figure"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "",
"figure_label": "",
"figure_type": "figure"
}
] | Hong Xu; Shireen Y Elhabian | [
{
"authors": "J Adams; R Bhalodia; S Elhabian",
"journal": "",
"ref_id": "b0",
"title": "Uncertain-deepssm: From images to probabilistic shape models",
"year": "2020-10-04"
},
{
"authors": "J Adams; S Elhabian",
"journal": "Springer",
"ref_id": "b1",
"title": "From images to probabilistic anatomical shapes: A deep variational bottleneck approach",
"year": "2022"
},
{
"authors": "P R Atkins; S Y Elhabian; P Agrawal; M D Harris; R T Whitaker; J A Weiss; C L Peters; A E Anderson",
"journal": "Journal of Orthopaedic Research",
"ref_id": "b2",
"title": "Quantitative comparison of cortical bone thickness using correspondence-based shape modeling in patients with cam femoroacetabular impingement",
"year": "2017"
},
{
"authors": "R Bhalodia; L A Dvoracek; A M Ayyash; L Kavan; R Whitaker; J A Goldstein",
"journal": "Journal of Craniofacial Surgery",
"ref_id": "b3",
"title": "Quantifying the severity of metopic craniosynostosis: A pilot study application of machine learning in craniofacial surgery",
"year": "2020"
},
{
"authors": "R Bhalodia; S Elhabian; J Adams; W Tao; L Kavan; R Whitaker",
"journal": "",
"ref_id": "b4",
"title": "Deepssm: A blueprint for image-to-shape deep learning models",
"year": "2021"
},
{
"authors": "R Bhalodia; S Y Elhabian; L Kavan; R T Whitaker",
"journal": "Springer",
"ref_id": "b5",
"title": "Deepssm: A deep learning framework for statistical shape modeling from raw images",
"year": "2018"
},
{
"authors": "J L Bruse; K Mcleod; G Biglino; H N Ntsinjana; C Capelli; T Y Hsia; M Sermesant; X Pennec; A M Taylor; S Schievano",
"journal": "BMC medical imaging",
"ref_id": "b6",
"title": "A statistical shape modelling framework to extract 3d shape biomarkers from medical imaging data: assessing arch morphology of repaired coarctation of the aorta",
"year": "2016"
},
{
"authors": "J C Carr; R K Beatson; J B Cherrie; T J Mitchell; W R Fright; B C Mccallum; T R Evans",
"journal": "Association for Computing Machinery",
"ref_id": "b7",
"title": "Reconstruction and representation of 3d objects with radial basis functions",
"year": "2001"
},
{
"authors": "N Carriere; P Besson; K Dujardin; A Duhamel; L Defebvre; C Delmaire; D Devos",
"journal": "Movement disorders",
"ref_id": "b8",
"title": "Apathy in parkinson's disease is associated with nucleus accumbens atrophy: a magnetic resonance imaging shape analysis",
"year": "2014"
},
{
"authors": "J Cates; E Bieging; A Morris; G Gardner; N Akoum; E Kholmovski; N Marrouche; C Mcgann; R S Macleod",
"journal": "Clinical Medicine Insights: Cardiology",
"ref_id": "b9",
"title": "Computational shape models characterize shape change of the left atrium in atrial fibrillation",
"year": "2014"
},
{
"authors": "J Cates; S Elhabian; R Whitaker",
"journal": "Elsevier",
"ref_id": "b10",
"title": "Shapeworks: particle-based shape correspondence and visualization software",
"year": "2017"
},
{
"authors": "J Cates; P T Fletcher; M Styner; M Shenton; R Whitaker",
"journal": "Springer",
"ref_id": "b11",
"title": "Shape modeling and analysis with entropy-based particle systems",
"year": "2007"
},
{
"authors": "R H Davies; R H Davies; C J Twining; T F Cootes; J C Waterton; C J Taylor",
"journal": "IEEE transactions on medical imaging",
"ref_id": "b12",
"title": "A minimum description length approach to statistical shape modeling",
"year": "2002"
},
{
"authors": "S Durrleman; M Prastawa; N Charon; J R Korenberg; S Joshi; G Gerig; A Trouvé",
"journal": "NeuroImage",
"ref_id": "b13",
"title": "Morphometry of anatomical shape complexes with dense deformations and sparse parameters",
"year": "2014"
},
{
"authors": "M D Harris; M Datar; R T Whitaker; E R Jurrus; C L Peters; A E Anderson",
"journal": "Journal of Orthopaedic Research",
"ref_id": "b14",
"title": "Statistical shape modeling of cam femoroacetabular impingement",
"year": "2013"
},
{
"authors": "K C Kempfert; Y Wang; C Chen; S W Wong",
"journal": "Intelligent Data Analysis",
"ref_id": "b15",
"title": "A comparison study on nonlinear dimension reduction methods with kernel variations: Visualization, optimization and classification",
"year": "2020"
},
{
"authors": "D Kingma; J Ba",
"journal": "",
"ref_id": "b16",
"title": "Adam: A method for stochastic optimization",
"year": "2014"
},
{
"authors": "A Lenz; N Krähenbühl; A Peterson; R Lisonbee; B Hintermann; C Saltzman; A Barg; Anderson; A Phd",
"journal": "Scientific Reports",
"ref_id": "b17",
"title": "Statistical shape modeling of the talocrural joint using a hybrid multi-articulation joint approach",
"year": "2021"
},
{
"authors": "C Merle; W Waldstein; J Gregory; S Goodyear; R Aspden; P Aldinger; D Murray; H Gill",
"journal": "Journal of Orthopaedic Research",
"ref_id": "b18",
"title": "How many different types of femora are there in primary hip osteoarthritis? an active shape modeling study",
"year": "2014"
},
{
"authors": "C Merle; M M Innmann; W Waldstein; E C Pegg; P R Aldinger; H S Gill; D W Murray; G Grammatopoulos",
"journal": "The Journal of Arthroplasty",
"ref_id": "b19",
"title": "High variability of acetabular offset in primary hip osteoarthritis influences acetabular reaming-a computed tomographybased anatomic study",
"year": "2019"
},
{
"authors": "N Sarkalkan; H Weinans; A A Zadpoor",
"journal": "Bone",
"ref_id": "b20",
"title": "Statistical shape and appearance models of bones",
"year": "2014"
},
{
"authors": "D W Thompson",
"journal": "On growth and form",
"ref_id": "b21",
"title": "On growth and form",
"year": "1942"
},
{
"authors": "K Tóthová; S Parisot; M Lee; E Puyol-Antón; A King; M Pollefeys; E Konukoglu",
"journal": "Springer",
"ref_id": "b22",
"title": "Probabilistic 3d surface reconstruction from sparse mri information",
"year": "2020"
},
{
"authors": "K Tóthová; S Parisot; M C H Lee; E Puyol-Antón; L M Koch; A P King; E Konukoglu; M Pollefeys",
"journal": "",
"ref_id": "b23",
"title": "Uncertainty quantification in cnn-based surface prediction using shape priors",
"year": "2018"
},
{
"authors": "G Turk; J F O'brien",
"journal": "",
"ref_id": "b24",
"title": "Variational implicit surfaces",
"year": "1999"
},
{
"authors": "M Van Buuren; N Arden; S Bierma-Zeinstra; W Bramer; N Casartelli; D Felson; G Jones; N Lane; C Lindner; N Maffiuletti; J Van Meurs; A Nelson; M Nevitt; P Valenzuela; J Verhaar; H Weinans; R Agricola",
"journal": "Osteoarthritis and Cartilage",
"ref_id": "b25",
"title": "Statistical shape modeling of the hip and the association with hip osteoarthritis: a systematic review",
"year": "2021"
},
{
"authors": "S Zachow",
"journal": "Facial Plastic Surgery",
"ref_id": "b26",
"title": "Computational planning in facial surgery",
"year": "2015"
},
{
"authors": "A A Zadpoor; H Weinans",
"journal": "Journal of biomechanics",
"ref_id": "b27",
"title": "Patient-specific bone modeling and analysis: the role of integration and automation in clinical adoption",
"year": "2015"
}
] | [
{
"formula_coordinates": [
3,
134.77,
330.73,
345.82,
21.67
],
"formula_id": "formula_0",
"formula_text": "P i = [p i,1 , p i,2 , • • • , p i,J ], the respec- tive surface normals are N i = [n i,1 , n i,2 , • • • , n i,J ],"
},
{
"formula_coordinates": [
4,
217.31,
373.63,
259.04,
24.38
],
"formula_id": "formula_1",
"formula_text": "f Pi,wi (x) = j∈ Pi w i,j ϕ(x, p i,j ) + c T i x + c 0 i (1"
},
{
"formula_coordinates": [
4,
476.35,
381.16,
4.24,
8.8
],
"formula_id": "formula_2",
"formula_text": ")"
},
{
"formula_coordinates": [
5,
249.45,
240.88,
231.14,
30.32
],
"formula_id": "formula_3",
"formula_text": "L surf Di (P i ) = J j=1 |D i (p i,j )|,(2)"
},
{
"formula_coordinates": [
5,
180.87,
393.98,
299.72,
30.32
],
"formula_id": "formula_4",
"formula_text": "L norm ∂Di (P i , N i ) = 180 π J j=1 cos -1 1 - n T i,j ∂D i (p i,j ) ∥n i,j ∥∥∂D i (p i,j )∥ .(3)"
},
{
"formula_coordinates": [
5,
152.85,
559.67,
327.74,
30.55
],
"formula_id": "formula_5",
"formula_text": "L corres µ (P 1 , ..., P K ) = H(P) = 1 2 log 1 3JK K k=1 (P k -µ) (P k -µ) T ,(4)"
},
{
"formula_coordinates": [
5,
296.25,
644.13,
78.25,
9.68
],
"formula_id": "formula_6",
"formula_text": "B i = [b i,1 , ..., b i,R"
},
{
"formula_coordinates": [
6,
134.77,
223.62,
345.83,
26.34
],
"formula_id": "formula_7",
"formula_text": "J j ′ =1 exp(k i r,j ′ ). Let e i ∈ R R"
},
{
"formula_coordinates": [
6,
197.45,
250.69,
211.76,
13.31
],
"formula_id": "formula_8",
"formula_text": "e i,r = [f Pi,wi (b i,r ) -D i (b i,r )] 2 . Let E i = e i 1 T M"
},
{
"formula_coordinates": [
6,
235.31,
296.19,
245.28,
9.88
],
"formula_id": "formula_9",
"formula_text": "(P i , N i ) = mean (softmin(K i ) ⊗ K i ⊗ E i ) ,(5)"
},
{
"formula_coordinates": [
6,
135.23,
386.77,
345.36,
57.44
],
"formula_id": "formula_10",
"formula_text": "L I,D,∂D (P K , N K ) = K i=1 αL surf Di (P i ) + βL norm ∂Di (P i , N i ) + γL sampl Bi,Di,wi (P i , N i ) + ζL corres µ (P 1 , ..., P K )(6)"
}
] |
|
10.1016/j.xops.2022.100127 | 2023-05-19 | [
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b0",
"b1",
"b2",
"b3",
"b4",
"b5",
"b6",
"b7",
"b8",
"b9",
"b10",
"b11",
"b12",
"b7",
"b13",
"b14",
"b8",
"b15",
"b16"
],
"table_ref": [],
"text": "Ophthalmology notes contain important clinical information about a patient's eye findings.\nThese findings are documented based on interpretations from imaging examinations (e.g., fundus examination), complications or outcomes associated with surgeries (e.g., cataract surgery), and experiences or symptoms shared by patients. Such findings are oftentimes described along with their exact eye locations as well as other contextual information such as their timing and status. Thus, ophthalmology notes comprise of spatial relations between eye findings and their corresponding locations, and these findings are further described using different spatial characteristics such as laterality and size. Although there has been recent advancements in using natural language processing (NLP) methods in the ophthalmology domain, they are mainly targeted for specific ocular conditions. Some work leveraged electronic health record text data to identify conditions such as glaucoma [1], herpes zoster ophthalmicus [2], and exfoliation syndrome [3], while another set of work extracted quantitative measures particularly related to visual acuity [4,5] and microbial keratitis [6]. In this work, we aim to extract more comprehensive information related to all eye findings, covering both spatial and contextual, from the ophthalmology notes. Besides automated screening and diagnosis of various ocular conditions, identifying such detailed information can aid in applications such as automated monitoring of eye findings or diseases and cohort retrieval for retrospective epidemiological studies. For this, we propose to extend our existing radiology spatial representation schema-Rad-SpatialNet [7] to the ophthalmology domain. We refer to this as the Eye-SpatialNet schema in this paper. We annotate a total of 600 ophthalmology notes following Eye-SpatialNet. Finally, we apply an advanced deep learning-based method to automatically identify the spatial and contextual information from the notes.\nOphthalmologists use spatial language to describe findings interpreted from imaging techniques. For example, in the sentence -\"OCT of the retinal nerve fiber layer shows normal thickness in both eyes.\", both eyes have been described using the finding normal thickness as interpreted from an Optical Coherence Tomography examination. Here, thickness is spatially associated to eyes through the preposition in, where normal describes the status of thickness and both describes the laterality. Similarly, symptoms presented by patients are also documented using spatial relations. In the sentence -\"She presented in [DATE] with weakness and numbness of her right eye as well as pain and vision loss in the left eye consistent with optic neuritis.\", the findings weakness and numbness are spatially related to right eye through the preposition of, whereas pain and vision are linked to left eye through in. Additionally, we note that the ophthalmologist also reports the potential diagnosis inferred from these findings, i.e., optic neuritis. [DATE] denotes the timing associated with the findings. Sometimes, eye procedures and drugs are also associated with anatomical locations and thus are spatially-grounded. We capture all these important information in our Eye-SpatialNet schema.\nThe Eye-SpatialNet schema is based on frame semantics, where a lexical unit (LU) represents the word that invokes a frame and the participants of a frame form the frame elements (FEs). The spatial prepositions (e.g., in) and verbs (e.g., reveals) constitute the lexical units whereas the associated findings (e.g., weakness), the locations (e.g., eye), diagnosis (e.g. optic neuritis), and the various spatial and other descriptors (e.g., left, normal ) constitute the frame elements. The spatial prepositions and verbs are also referred to as spatial triggers in this paper. Following this schema, we create a manually-annotated dataset of 600 ophthalmology notes to represent important spatial information of clinical significance.\nTwo sample examples from our ophthalmology dataset are illustrated in Figure 1 For automatic extraction of the spatial information, we adopt a two-turn question answering framework [8,9] based on a transformer language model, BERT [10]. This is inspired by previous studies demonstrating the effectiveness of framing various information extraction tasks such as named entity recognition [11], relation extraction [12], and event extraction [13] as question answering (QA) by harnessing the well-developed machine reading comprehension models. Further, some studies [8,14,15] investigated the formulation of relation and event extraction tasks as multi-turn QA both in the general and biomedical domain. In this paper, we apply a two-turn QA method similar to the one proposed for radiology domain [9], to extract the spatial and descriptive frame elements from ophthalmology notes. This QA-based method can be seen as a formulation of prompt-based models commonly seen in NLP [16,17]. In this method, we extract the spatial triggers and the main entities (e.g., eye finding, anatomical location) in the first turn and subsequently extract all the spatial (e.g., laterality) and descriptive (temporal descriptor or the timing of a finding) frame elements in the second turn. Finally, we evaluate the performance of the two-turn QA system on a held-out test set of 100 notes. We also investigate the potential benefit of transfer learning from a different medical domain (i.e., radiology) through sequential fine-tuning for our task of spatial information extraction."
},
{
"figure_ref": [],
"heading": "Related work",
"publication_ref": [],
"table_ref": [],
"text": "We review relevant prior work in two categories. Section 2.1 focuses on the types of clinical text that previous research has explored for various ophthalmology NLP applications. Section 2.2 presents an overview of the NLP methods proposed in the ophthalmology domain."
},
{
"figure_ref": [],
"heading": "Datasets and Applications in Ophthalmology NLP",
"publication_ref": [
"b0",
"b17",
"b18",
"b19",
"b20",
"b2",
"b1",
"b4",
"b21",
"b22",
"b5",
"b3",
"b4"
],
"table_ref": [],
"text": "Wang et al. [1] used the first 3 clinical progress notes from within the first 120 days of follow-up along with structured clinical data for predicting whether glaucoma patients would require surgery. Another study by Wang et al. [18] used both structured and progress notes to predict visual prognosis. Wang et al. [19] also demonstrated the usage of ophthalmology domain-specific word embeddings trained using PubMed abstracts and electronic health record notes to improve the performance of deep learning models in predicting visual prognosis when compared to using general word embeddings. Baxter et al. [20] identified fungal ocular involvement cases using 26830 free-text notes in the MIMIC-III clinical database. Liu et al. [21] applied NLP on 743838 operative notes to identify two key variables, intracameral antibiotic injection and posterior capsular rupture. Stein et al. [3] used NLP to identify exfoliation syndrome from clinical notes. Zheng et al. [2] applied an NLP algorithm on over 1 million clinical notes to identify herpes zoster ophthalmicus cases.\nMbagwu et al. [5] built a structured query language-based algorithm to first extract the Snellen visual acuities from structured laterality fields from 295218 ophthalmology clinical notes and then capture the best documented visual acuity of each eye and this was evaluated against a clinician chart review of 100 random notes. Later, Baughman et al.\n[4] developed a rule-based NLP algorithm to extract Snellen visual acuity from free-text inpatient ophthalmology notes collected from the University of Washington healthcare system in Seattle over an 8-year period and their algorithm was evaluated against the data points generated from manual review of 644 notes. Wang et al. [22] developed rule-based algorithms to extract surgery outcome mentions representing implant usage-intraocular lens power and glaucoma implant type as well as surgery laterality from ophthalmology operative notes. The algorithms were validated against manually-annotated random sets of 100 operative notes for each of the three surgical categories and 100 notes for laterality. Maganti et al. [23] designed an NLP algorithm to extract two quantitative key features of microbial keratitis-epithelial defect and stromal infiltrate as millimeter measurements from progress notes. Recently, another work by Woodward et al. [6] developed a rule-based NLP system to identify the clinical features of microbial keratitis, namely centrality, depth, and thinning using the free text in the corneal examination section from physician notes.\nWe see that majority of the work in ophthalmology NLP is focused toward identifying cases associated with specific ocular diseases such as glaucoma and fungal ocular involvement.\nAnother strand of work attempted to extract certain information from the unstructured notes specifically related to visual acuity, surgery outcomes, and microbial keratitis. From a dataset perspective, previous works mostly used operative, clinical, and progress notes while the two studies [4,5] on visual acuity extraction have utilized ophthalmology notes. Thus, we find that limited research has focused on comprehensive information extraction across eye diseases and, therefore, in this work we attempt to extract more detailed clinical information from ophthalmology notes that can broaden the scope of applications in ophthalmology."
},
{
"figure_ref": [],
"heading": "NLP Methods for Ophthalmology Information Extraction",
"publication_ref": [
"b21",
"b5",
"b3",
"b19",
"b20",
"b2",
"b0",
"b17",
"b18",
"b18"
],
"table_ref": [],
"text": "Most of the systems extracting ophthalmic information from unstructured notes developed rule-based methods. Among these, Wang et al. [22] used regular expressions to extract surgery outcomes. Woodward et al. [6] employed regular expressions, part-of-speech tagging and syntactic dependency parsing to extract features of microbial keratitis. Mbagwu et al.\n[5] and Baughman et al. [4] also developed rule-based algorithms for visual acuity extraction.\nThe former was based on structured query language where keyword search was performed on the structured laterality fields while the latter used regular expression in combination with additional rules.\nAmong the studies that focused on identifying certain ocular cases, Baxter et al. [20] developed a string matching method using regular expressions to extract text strings relevant to fungal ocular involvement. Liu et al. [21] built a lexicon using SAS text-processing modules that identified misspellings, negations, and abbreviations and associated words to concepts for identifying the two key variables from operative notes. Stein et al. [3] curated a list of terms and abbreviations to search for exfoliation syndrome-related mentions. The search algorithm included identifying negated mentions based on surrounding text and used regular expressions and generalized Levenshtein edit distance to recognize misspellings. Zheng et al.\n[2] created terminologies to search for herpes zoster-related information and also included relation detection algorithm for identifying herpes zoster ophthalmicus signs or symptoms.\nA few studies [1,18,19] by Wang et al. developed deep learning models for predicting glaucoma progression and whether low vision patients would have low vision after one year.\nThe models were based on convolutional neural network architectures and they used the previously developed ophthalmology domain-specific word embeddings [19] to represent the words.\nThus, we see that most of the existing methods are rule-based and they are restricted to specific entity extraction from ophthalmology-related text. Only a few studies used deep learning-based methods and those too for prediction task. In this work, we consider a wide variety of ophthalmic entities of clinical importance including findings, their locations, and visual acuity scores. Additionally, we also cover detailed spatial relations including relations between eye findings and locations as well as findings and their corresponding descriptors. We employ an advanced transformer language-based question answering method to automatically extract the entities and relations."
},
{
"figure_ref": [],
"heading": "Data",
"publication_ref": [
"b23"
],
"table_ref": [],
"text": "We use a set of 600 notes for annotating the important ophthalmic entities and spatial relations. These notes are collected from the Robert Cizik Eye Clinic at McGovern Medical School at Houston. The notes contain information about a patient's history, detailed description of patients' experiences with their vision, interpretations of eye imaging examinations, information about past surgeries and their outcomes and complications, and associated neurological symptoms. We use the BRAT tool [24] for annotation. "
},
{
"figure_ref": [
"fig_2"
],
"heading": "Representation Schema",
"publication_ref": [
"b6",
"b8"
],
"table_ref": [],
"text": "Our annotation schema is largely adopted from an existing frame-based spatial representation schema -Rad-SpatialNet [7]. The spatial language encoded in the ophthalmology notes are different from those in radiology reports. We represent the information in a way that can accurately capture ophthalmology-specific spatial meanings from the note text. For this schema, we incorporate specific spatial and descriptive frame elements or relations besides the common ones proposed in Rad-SpatialNet. The entity types included are spatial trigger, finding, anatomy, device, location descriptor, other descriptor, assertion, quantity, drug, and procedure. The spatial and descriptive frame elements are mostly similar to the ones described in our previous work [9]. Additionally, we include the following frame elements: medication, impact on side, pathophysiologic descriptor, direction, associated diagnosis, specific location descriptor, certainty descriptor, and value. Frame elements are either connected to the spatial trigger terms or the main clinical entities such as findings and anatomies. We describe the newly added ophthalmology-specific frame elements in the following subsections. The schema is illustrated in Figure 2."
},
{
"figure_ref": [],
"heading": "New spatial frame elements",
"publication_ref": [],
"table_ref": [],
"text": "We add three new spatial frame elements related to findings, namely, exact location descriptor, impact on side, and direction. For the exact location descriptor, let us consider the example below.\nShe was found to have 20/25 vision OD and CF vision OS with mild disc edema in the left eye.\nHere we see that there is a spatial relation between mild disc edema and left eye connected through the spatial trigger in. As per the Eye-SpatialNet schema, edema has the spatial role of a 'Figure ' and its corresponding location eye acts as the 'Ground'. Moreover, we notice that edema has been described through a location descriptor disc besides the status descriptor mild.\nSometimes, a finding that has been detected in both sides (left and right) is described with different severity based on laterality or side.\nExternal examination reveals a right relative proptosis with bilateral lid retraction right greater than left.\nIn this example, retraction is the finding that is more pronounced in the right eyelid than the left eyelid. Moreover, retraction is described using laterality bilateral and location descriptor lid.\nA finding's direction is also documented in the notes.\nShe reports that her right eye deviated outward, and she had difficulty walking with poor coordination.\nHere, outward is used to describe the direction of right eye deviation.\nAll these three frame elements-location descriptor, impact on side, and direction are associated with describing the detailed spatial aspects of a finding and, therefore, we include these elements in our representation schema."
},
{
"figure_ref": [],
"heading": "New descriptive frame elements",
"publication_ref": [
"b6"
],
"table_ref": [],
"text": "Ophthalmologists often document detailed contextual information while describing the findings. We add four descriptive frame elements related to findings, namely, certainty descriptor, associated diagnosis, pathophysiologic descriptor, and value. Consider the example below.\nHis past ocular history is significant for optic neuritis and right optic atrophy.\nIn this sentence, the term significant is used to describe the certainty of both optic neuritis and optic atrophy findings.\nOftentimes, some findings are described along with their associated diagnoses. In the following example, occlusions is linked to Susac Syndrome.\nAt this time the exact cause is unknown, however, with multiple retinal branch artery occlusions bilaterally one must entertain the diagnosis of Susac Syndrome.\nNote that this 'Associated Diagnosis' frame element is different from the 'Diagnosis' frame element proposed in Rad-SpatialNet [7]. The 'Diagnosis' element is linked to a spatial trigger, whereas 'Associated Diagnosis' element is linked to an eye finding (e.g., Susac Syndrome in the sentence above). In \"She did show me video of her episodes of upturning of the eyes which appears consistent with oculogyric crisis.\", oculogyric crisis acts as the 'Diagnosis' element of the spatial frame instantiated by the spatial trigger of connecting upturning and eyes.\nWe also include pathophysiologic descriptor of a finding in the schema. For example, in \"She is seen in follow up for her left sided headache and retroorbital pain in the setting of presumed autoimmune retinopathy.\", autoimmune is the pathophysiologic descriptor associated with the finding retinopathy.\nOphthalmology notes also contain information about visual acuity scores and other eye-related measurements. We present two examples below.\n1. On his examination he found her to have 20/20 vision OD and 20/30 vision OS with a left RAPD. element, whereas the third example shows that the finding Cup/Disc Ratio is associated with its corresponding value 0.4. Therefore, we capture all the important eye measurements in our schema."
},
{
"figure_ref": [],
"heading": "INTRAOCULAR MEASUREMENT: Method: Applanation Right Eye: 15",
"publication_ref": [
"b6"
],
"table_ref": [],
"text": "Apart from above additions, this schema covers temporal information of findings that are expressed using a variety of phrases unlike the temporal descriptors of radiological findings annotated in Rad-SpatialNet [7]. These expressions include one and a half to two years, > 8 years, next 3-4 months, within 1-2 months post-operatively, over the next few days, and early in the mornings. This schema covers lateralities that are specific to ophthalmology such as OS, OD, and OU, besides the common ones such as left, right, and bilateral."
},
{
"figure_ref": [],
"heading": "Annotation statistics",
"publication_ref": [
"b6"
],
"table_ref": [],
"text": "Each note was annotated by two annotators having medical background (one optometrist, one MD) and the annotations were reconciled iteratively through discussions. The overall F1 agreements are reported for annotating the main entities, the spatial and descriptive frame elements. We show the statistics of our annotated dataset as well as the inter-annotator agreement measures in Tables 1, 2, 3, and 4. The average sentence length (20.34) of the ophthalmology notes is slightly higher than that of the radiology reports dataset [7]. The terms and are often annotated as Findings. Another general challenge in the annotation process involved separating the eye-related findings from the neuroradiological findings as oftentimes the interpretations of brain images are embedded in the ophthalmology notes."
},
{
"figure_ref": [],
"heading": "Methods",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Overview",
"publication_ref": [
"b8",
"b9",
"b7",
"b8"
],
"table_ref": [
"tab_4"
],
"text": "We frame the task of spatial information extraction (IE) from ophthalmology notes as two-turn question answering (QA). This formulation (both single and multi turn QA) has proven to perform well for various general and biomedical domain IE tasks. Our previous work has also demonstrated the improved performance of a two-turn QA framework over a more standard sequence labeling-based method to extract detailed information from radiology text [9]. Inspired by these findings, we adopt a similar two-turn QA approach to identify the spatial triggers, the main ophthalmic entities, and their corresponding spatial and descriptive frame elements. This framework is suitable for IE scenarios where identification of relations or frame elements are dependent on extracting the target entities or lexical units of the frames (i.e., spatial triggers and ophthalmic entities). In this, the aim is to query a machine reading comprehension (MRC) model for returning answers given a query and the context passage (ophthalmology note text). The MRC system is based on the pre-trained language model BERT [10] where we follow the standard BERT input format by combining the query and the note text. The system allows for multiple answer extraction against a query, which is suitable for our schema as there can be multiple frame elements of the same type that are linked to a particular entity (spatial trigger or other ophthalmic entity). The MRC framework involving two BERT models for the two turns are adopted from a previous work [8]. We construct queries for the newly added entities and frame elements in Eye-SpatialNet.\nWe adopt the same query templates for both target entity and element extraction as used in our previous work [9]. Queries for the first turn incorporate the entity types whereas queries for the second turn include information about the frame elements and the associated main entity that is extracted in the first turn. In this paper, we use the Query find + desc variant to extract the frame elements in the second turn. The idea is to make the query more informative through incorporation of domain knowledge by adding a description of the particular frame element of interest at the beginning of a query. The following is an example query to extract 'ImpactOnSide' spatial element.\nImpactOnSide refers to which eye side is more impacted. Examples include right greater than left, smaller than left, and worse in the left eye. find all descriptor entities in the context that have a impact on side relationship with clinical finding entity optic neuropathy.\nHere, we see that the query includes description about ImpactOnSide as well as the finding entity (i.e., optic neuropathy) that is identified in the previous turn. If no answer is retrieved from the MRC system, this means there is no such entity of type 'Descriptor' in the note text that captures information about which eye side is more or less affected by optic neuropathy.\nThe descriptions used to form the queries for all new frame elements are shown in Table 5. "
},
{
"figure_ref": [],
"heading": "Experimental Settings and Evaluation",
"publication_ref": [
"b24",
"b8",
"b8",
"b6",
"b6"
],
"table_ref": [],
"text": "We randomly split our annotated ophthalmology dataset of 600 notes such that 450 notes are used for training, 50 for development, and 100 for testing. We use a clinical BERT LARGE model that is pre-trained on MIMIC-III clinical notes for 300K steps [25] as it performed better on the radiology reports dataset [9]. We fine-tune BERT LARGE -MIMIC (cased version) on our Eye-SpatialNet dataset for 10 epochs and use the same hyperparameter settings as reported in Datta et al. [9]. We evaluate the performance metrics -precision, recall, and F1 score and report the results on the test set of 100 notes. We consider exact matches of the entity and frame element spans against the annotated spans for evaluation.\nFurther, to leverage an already available language model that is fine-tuned on the radiology reports dataset (introduced in Datta et al. [7]) for the task of spatial information extraction, we evaluate any prospective benefits of transfer learning through sequential fine-tuning, that is, by first fine-tuning the model on radiology reports followed by fine-tuning on ophthalmology [7]. Note that we use the gold spatial triggers for this experiment to extract the elements that are connected to the triggers (and not the main ophthalmic entities). Using predicted triggers would provide a more realistic evaluation, but that is not the focus of this experiment. We evaluate the performance on the main spatial frame elements that are common between the two domains on the 100 test ophthalmology notes. For fine-tuning the sequence labeling model on the ophthalmology data, we set the maximum sequence length at 128, learning rate at 2e -5, and number of training epochs at 4."
},
{
"figure_ref": [],
"heading": "Results",
"publication_ref": [
"b6",
"b8",
"b25",
"b26",
"b27"
],
"table_ref": [
"tab_5",
"tab_6",
"tab_5",
"tab_6",
"tab_7"
],
"text": "The performance measures for extracting the main ophthalmic entities in the first turn from 100 ophthalmology test notes are reported in Table 6. The results are promising for the common entities including 'Spatial trigger', 'Finding', and 'Anatomy' with F1 scores of 89.31, 79.37, and 85.26, respectively, while they are low for 'Location descriptor' and 'Procedure'.\nNote that the entities 'Drug' and 'Device' occur very infrequently in the dataset (with only 2 and 1 occurrences in the test set) and the performance measures are zero.\nWe show the results for extracting the spatial and descriptive frame elements in the second turn in Table 7. We see that the model performs well for common frame elements such as 'Ground', 'Hedge', 'Laterality', 'ImpactOnSide', and 'Negation'. The performance measures are particularly low for 'Relative Position', 'Size', and 'Temporal Desc'. This may be because Most of the entities and frame elements used in encoding spatial language in the ophthalmology notes are adopted from our previously proposed Rad-SpatialNet schema [7] built for radiology. This indicates the generalizability of the schema in that it captures most of the common and important spatial information usually encountered in clinical text. In this work, we incorporate additional frame elements for two reasons. First, to cover more detailed information about the findings that were not present in Rad-SpatialNet such as capturing implicit spatial relations through the 'Location Desc' frame element (e.g., scenarios where a spatial relation exists but a spatial trigger term is not present in the sentence).\nSecond, to include ophthalmology-specific spatially-grounded entities (e.g., 'Procedure') and elements that are of interest to ophthalmology researchers (e.g., visual acuity and other important eye measurements through the 'Value' frame element). The results in Tables 6 and7 show that the two-turn QA approach achieves satisfactory performance in identifying different entities and frame elements and are comparable to the results on the radiology report dataset [9]. We achieve this without any modification of the query templates and the frame element descriptions (that are used to form the queries) for those elements that also exist in Rad-SpatialNet. This also indicates that the method is adaptable and generalizable enough to work satisfactorily well for frequent entity types and frame elements across medical domains (although the language style and the vocabulary differ substantially between radiology reports and ophthalmology notes).\nTo examine the effect of transfer learning from a different medical domain, our experiment with the sequence labeling model in Table 8 indicates that transfer learning holds potential in improving the performance for some frame elements, however, a more thorough evaluation covering all other elements is required to understand its real benefits. This includes experimenting with a small number of ophthalmology notes in the fine-tuning process, as often only a limited amount of labeled data is available in a new domain. Interestingly, although the two-turn QA approach works well both for ophthalmology and radiology domains, our initial experiments with sequential fine-tuning did not yield good results using the QA approach.\nWe leave this to future work where we plan to investigate this further and evaluate the less explored domain adaptation techniques such as the adaptive off-the-shelf approach proposed in Laparra et al. [26].\nTo handle less frequent entities and frame elements better as well as to further improve the QA model's performance, we plan to augment the dataset by automatically generating a large weakly labeled ophthalmology dataset using domain-specific rules, a technique that has been validated to be useful by many recent studies in the medical domain [27,28]. Apart from reducing the annotation effort, this can particularly be useful for elements such as 'Size'\nand 'Value' that usually follow a set of patterns based on domain. For example, '4-> 3mm'\nis used to express pupil size in an ophthalmology note whereas '2.1 x 3.4 x 2.0 cm' denotes a tumor size in a radiology report. Finally, for a more exhaustive evaluation on this proposed dataset, we also intend to incorporate cross validation in a later work."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "We annotated 600 ophthalmology notes with important spatial and contextual information of clinical importance. We adopt our previously proposed Rad-SpatialNet schema and incorporate additional ophthalmology-specific information to encode spatial language in ophthalmology. We apply a well-established approach of framing the extraction task as question answering to automatically identify the ophthalmic entities and their associated spatial and descriptive frame elements . Our two-turn QA method performed well with high F1 scores for common entities and elements."
},
{
"figure_ref": [],
"heading": "",
"publication_ref": [],
"table_ref": [],
"text": "Funding This work was supported in part by the National Institute of Biomedical Imaging and Bioengineering (NIBIB: R21EB029575), the Patient-Centered Outcomes Research Institute (PCORI: ME-2018C1-10963) and the Cancer Prevention and Research Institute of Texas (CPRIT RP210045)."
}
] | This paper focuses on the representation and automatic extraction of spatial information in ophthalmology clinical notes. We extend our previously proposed frame semantics-based spatial representation schema, Rad-SpatialNet, to represent spatial language in ophthalmology text, resulting in the Eye-SpatialNet schema. The spatially-grounded entities are findings, procedures, and drugs. To accurately capture all spatial details, we add some domain-specific elements in Eye-SpatialNet. Utilizing this representation, we annotated dataset of 600 ophthalmology notes labeled with detailed spatial and contextual information of ophthalmic entities. The annotated dataset contains 1715 spatial triggers, 7308 findings, 2424 anatomies, and 9914 descriptors. To automatically extract the spatial information, we employ a two-turn question answering approach based on the transformer language model BERT. The results are promising, with F1 scores of 89.31, 74.86, and 88.47 for spatial triggers, Figure, and Ground frame elements, respectively. This is the first work to represent and extract a wide variety of clinical information in ophthalmology. Extracting detailed information can benefit ophthalmology applications and research targeted toward disease progression and screening. | Eye-SpatialNet: Spatial Information Extraction from Ophthalmology Notes | [
{
"figure_caption": "Figure 1 :1Figure 1: Example sentences from ophthalmology notes showing some of the spatial frame elements covered in the Eye-SpatialNet schema. The underlined and italicized texts denote the lexical units of the frames.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": ". Note that for (a), 'Figure' and 'Ground' are the spatial frame elements of the frame evoked by the spatial trigger of, whereas 'Morphologic descriptor' and 'Distribution pattern' are the spatial frame elements of the frame evoked by the finding enlargement. 'Figure' usually refers to an entity whose location is described through a spatial trigger whereas 'Ground' denotes the actual anatomical location. In the second example (b), cataract surgery is spatially linked to eyes where cataract surgery acts as the 'Figure' element of the frame evoked by the spatial trigger in and 2142 (year altered for de-identification) is a descriptive frame element of the frame instantiated by the procedure cataract surgery. There are a total of 1715 spatial triggers, 7308 finding, and 2424 location phrases annotated in the dataset. We describe the annotation process in Section 3.1. To our knowledge, this is the first study to develop an annotated dataset with comprehensive representation schema for identifying detailed information from ophthalmology notes.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: Eye-SpatialNet schema. The dashed circles indicate the newly added frame elements.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Basic statistics. Avg -Average. Note that in the first example, the first vision occurrence has a visual acuity score of 20/20 in the oculus dextrus (OD) or the right eye, while the second vision has a score of 20/30 in the oculus sinister (OS ) or the left eye. Thus, the first vision finding is linked to 20/20",
"figure_data": "ItemValueAvg. note length (in tokens)470.61Avg. sentence length (in tokens) 20.34No. of unique spatial triggers493. LEFT EYE: Media: hazy view Cup/Disc Ratio: 0.4",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Main entities.",
"figure_data": "Entity typeFrequency F1 agreementSpatial trigger17150.91Finding73080.80Anatomy24240.88Device140.90Drug220.60Procedure1820.35Other descriptor97820.79Quantity3660.88Assertion16160.70Location descriptor1320.60",
"figure_id": "tab_1",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Spatial frame elements.",
"figure_data": "Frame elementFrequency F1 agreementFigure22610.77Ground20940.89Hedge3970.69Diagnosis180.28Relative Position1320.59Reason70.77Medication180.64Morphologic450.44Size Desc430.56Distribution Pattern830.29Composition360.59Laterality34640.78Size480.30Impact on Side970.75Direction850.56Specific location16360.72",
"figure_id": "tab_2",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Descriptive frame elements.",
"figure_data": "Frame elementFrequency F1 agreementStatus30510.59Quantity1010.56Temporal10660.45Negation9210.55Pathphysio750.60Certainty2980.49Associated Diagnosis720.23Value3180.834.2. Query generation",
"figure_id": "tab_3",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Descriptions used in the queries to extract additional frame elements.",
"figure_data": "Frame elementDescriptionMedicationMedication refers to a drug or solution that has been administeredor applied to any eye location.ImpactOnSideImpactOnSide refers to which eye side is more impacted. Exam-ples include right greater than left, smaller than left, and worsein the left eye.PathphysioDescPathophysiologic descriptor refers to the functional changes thataccompany a disease. Examples include autoimmune and physio-logic.DirectionDirection indicates direction of a finding. Examples includeoutward and to the right.AssocDiagAssociated diagnosis refers to the clinical condition or diseaseassociated with a finding. This usually appears after phrases suchas associated with and secondary to.LocationDescLocation descriptor refers to the exact location of a finding. Ex-amples include retrooorbital and optic disc.CertaintyDescCertainty descriptor refers to uncertainty phrases describing afinding. Examples include significant and consistent with.ValueValue refers to a visual acuity score or any measurement or ratio.Examples include 20/20, 20/40, 16, and 0.8.",
"figure_id": "tab_4",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "Target entity extraction results using BERT LARGE -MIMIC two-turn QA method. desc -Descriptor. The radiology fine-tuning was performed on 288 reports and we further fine-tune on 450 ophthalmology notes. For this, we use the BERT LARGE -MIMIC sequence labeling model from Datta et al.",
"figure_data": "EntityPrecision(%) Recall (%) F1Spatial trigger86.8691.8989.31Finding75.7183.4179.37Anatomy85.3785.1585.26Location desc30.7740.0034.78Other desc76.5783.0479.67Assertion81.7889.8085.60Quantity82.8982.8982.89Procedure56.6753.1254.84notes.",
"figure_id": "tab_5",
"figure_label": "6",
"figure_type": "table"
},
{
"figure_caption": "Frame element extraction results using BERT LARGE -MIMIC two-turn QA method. sptr -Spatial trigger. Desc -Descriptive.",
"figure_data": "Frame ElementsPrecision(%) Recall (%) F1Spatial(sptr)Figure Ground Hedge Relative Position Medication75.29 85.89 89.47 30.43 50.0074.43 91.21 86.44 70.00 10074.86 88.47 87.93 42.42 66.67Spatial(entity)Laterality Distribution Pattern SizeDesc LocationDesc ImpactOnSide Direction80.59 47.37 60.00 69.26 72.73 57.1483.15 64.29 42.86 76.21 84.21 66.6781.85 54.55 50.00 72.57 78.05 61.54Size28.5710.0014.81Status70.1170.9370.52Desc(entity)Quantity Temporal Negation Certainty Pathphysio63.64 53.33 77.60 60.26 47.0643.75 43.78 82.32 64.38 53.3351.85 48.09 79.89 62.25 50.00Value81.6368.9774.77of the wide variation in the phrases used to express the sizes and temporalities of findings.The results of transfer learning experiment from radiology to ophthalmology domain isshown in",
"figure_id": "tab_6",
"figure_label": "7",
"figure_type": "table"
},
{
"figure_caption": "We see the F1 scores improve from 78.76 to 80.88 for 'Figure'and from 82.64 to 89.08 for 'Hedge' when we use a model fine-tuned on radiology reports to further fine-tune on our ophthalmology dataset. We also note that the F1 measure for the 'Ground' element is 91.95 without the requirement of any fine-tuning on ophthalmology data. The results are zero for 'Diagosis' and 'Reason' as they are too infrequent in the dataset as stated above.",
"figure_data": "",
"figure_id": "tab_7",
"figure_label": "8",
"figure_type": "table"
},
{
"figure_caption": "F1 measures for different fine-tuning variations using BERT LARGE -MIMIC sequence labeling method on 100 test ophthalmology notes. Eye: Fine-tuning only on Eye-SpatialNet (Ophthalmology), Rad→Eye: Fine-tuning on Rad-SpatialNet (Radiology) followed by Eye-SpatialNet, Rad: Fine-tuning only on Rad-SpatialNet.",
"figure_data": "Frame Element Eye Rad→Eye RadFigure78.7680.8851.29Ground95.3895.1991.95Hedge82.6489.080Relative Position60.8757.1443.487. DiscussionWe present a new dataset of 600 ophthalmology notes annotated with detailed spatialand contextual information. Although a few studies worked on identifying a certain set ofentities from clinical notes, they are mostly focused toward visual acuity and features ofmicrobial keratitis [4-6]. Our work is an initial effort in building a schema that captures",
"figure_id": "tab_8",
"figure_label": "8",
"figure_type": "table"
}
] | Surabhi Datta; Tasneem Kaochar; Hio Cheng Lam; Nelly Nwosu; Luca Giancardo; Alice Z Chuang; Robert M Feldman; Kirk Roberts | [
{
"authors": "S Wang; B Tseng; T Hernandez-Boussard",
"journal": "Ophthalmology Science",
"ref_id": "b0",
"title": "Deep Learning Approaches for Predicting Glaucoma Progression Using Electronic Health Records and Natural Language Processing",
"year": ""
},
{
"authors": "C Zheng; Y Luo; C Mercado; L Sy; S J Jacobsen; B Ackerson; B Lewin; H F Tseng",
"journal": "Clinical & Experimental Ophthalmology",
"ref_id": "b1",
"title": "Using natural language processing for identification of herpes zoster ophthalmicus cases to support population-based study",
"year": "2019"
},
{
"authors": "J D Stein; M Rahman; C Andrews; J R Ehrlich; S Kamat; M Shah; E A Boese; M A Woodward; J Cowall; E H Trager; P Narayanaswamy; D A Hanauer",
"journal": "JAMA ophthalmology",
"ref_id": "b2",
"title": "Evaluation of an Algorithm for Identifying Ocular Conditions in Electronic Health Record Data",
"year": "2019"
},
{
"authors": "D M Baughman; G L Su; I Tsui; C S Lee; A Y Lee",
"journal": "Translational Vision Science & Technology",
"ref_id": "b3",
"title": "Validation of the Total Visual Acuity Extraction Algorithm (TOVA) for Automated Extraction of Visual Acuity Data From Free Text, Unstructured Clinical Records",
"year": "2017"
},
{
"authors": "M Mbagwu; D D French; M Gill; C Mitchell; K Jackson; A Kho; P J Bryar",
"journal": "JMIR Medical Informatics",
"ref_id": "b4",
"title": "Creation of an Accurate Algorithm to Detect Snellen Best Documented Visual Acuity from Ophthalmology Electronic Health Record Notes",
"year": "2016"
},
{
"authors": "M A Woodward; N Maganti; L M Niziol; S Amin; A Hou; K Singh",
"journal": "Cornea",
"ref_id": "b5",
"title": "Development and Validation of a Natural Language Processing Algorithm to Extract Descriptors of Microbial Keratitis From the Electronic Health Record",
"year": "2021"
},
{
"authors": "S Datta; M Ulinski; J Godfrey-Stovall; S Khanpara; R F Riascos-Castaneda; K Roberts",
"journal": "",
"ref_id": "b6",
"title": "Rad-SpatialNet: A Frame-based Resource for Fine-Grained Spatial Relations in Radiology Reports",
"year": "2020"
},
{
"authors": "X Li; F Yin; Z Sun; X Li; A Yuan; D Chai; M Zhou; J Li",
"journal": "",
"ref_id": "b7",
"title": "Entity-Relation Extraction as Multi-Turn Question Answering",
"year": "2019"
},
{
"authors": "S Datta; K Roberts",
"journal": "International Journal of Medical Informatics",
"ref_id": "b8",
"title": "Fine-grained spatial information extraction in radiology as two-turn question answering",
"year": "2022"
},
{
"authors": "J Devlin; M.-W Chang; K Lee; K Toutanova",
"journal": "",
"ref_id": "b9",
"title": "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding",
"year": "2019"
},
{
"authors": "X Li; J Feng; Y Meng; Q Han; F Wu; J Li",
"journal": "",
"ref_id": "b10",
"title": "A Unified MRC Framework for Named Entity Recognition",
"year": "2020"
},
{
"authors": "O Levy; M Seo; E Choi; L Zettlemoyer",
"journal": "",
"ref_id": "b11",
"title": "Zero-Shot Relation Extraction via Reading Comprehension",
"year": "2017"
},
{
"authors": "J Liu; Y Chen; K Liu; W Bi; X Liu",
"journal": "",
"ref_id": "b12",
"title": "Event Extraction as Machine Reading Comprehension",
"year": "2020"
},
{
"authors": "F Li; W Peng; Y Chen; Q Wang; L Pan; Y Lyu; Y Zhu",
"journal": "",
"ref_id": "b13",
"title": "Event Extraction as Multiturn Question Answering",
"year": "2020"
},
{
"authors": "X D Wang; L Weber; U Leser",
"journal": "",
"ref_id": "b14",
"title": "Biomedical Event Extraction as Multi-turn Question Answering",
"year": "2020"
},
{
"authors": "N Taylor; Y Zhang; D Joyce; A Nevado-Holgado; A Kormilitzin",
"journal": "",
"ref_id": "b15",
"title": "Clinical prompt learning with frozen language models",
"year": "2022"
},
{
"authors": "S Sivarajkumar; Y Wang",
"journal": "",
"ref_id": "b16",
"title": "Healthprompt: A zero-shot learning paradigm for clinical natural language processing",
"year": "2022"
},
{
"authors": "S Y Wang; B Tseng",
"journal": "Investigative Ophthalmology & Visual Science",
"ref_id": "b17",
"title": "Looking for Low Vision: Deep Learning and Natural Language Processing to Predict Visual Prognosis",
"year": "2021"
},
{
"authors": "S Wang; B Tseng; T Hernandez-Boussard",
"journal": "International Journal of Medical Informatics",
"ref_id": "b18",
"title": "Development and evaluation of novel ophthalmology domain-specific neural word embeddings to predict visual prognosis",
"year": "2021"
},
{
"authors": "S L Baxter; A R Klie; B R Saseendrakumar; G Y Ye; M Hogarth",
"journal": "JMIR",
"ref_id": "b19",
"title": "Text Processing for Detection of Fungal Ocular Involvement in Critical Care Patients: Cross-Sectional Study",
"year": "2020"
},
{
"authors": "L Liu; N H Shorstein; L B Amsden; L J Herrinton",
"journal": "Pharmacoepidemiology and Drug Safety",
"ref_id": "b20",
"title": "Natural language processing to ascertain two key variables from operative reports in ophthalmology",
"year": "2017"
},
{
"authors": "S Y Wang; S Pershing; E Tran; T Hernandez-Boussard",
"journal": "International Journal of Medical Informatics",
"ref_id": "b21",
"title": "Automated extraction of ophthalmic surgery outcomes from the electronic health record",
"year": "2020"
},
{
"authors": "N Maganti; H Tan; L M Niziol; S Amin; A Hou; K Singh; D Ballouz; M A Woodward",
"journal": "Ophthalmology",
"ref_id": "b22",
"title": "Natural Language Processing to Quantify Microbial Keratitis Measurements",
"year": "2019"
},
{
"authors": "P Stenetorp; S Pyysalo; G Topić; T Ohta; S Ananiadou; J Tsujii",
"journal": "",
"ref_id": "b23",
"title": "Brat: A Web-based Tool for NLP-Assisted Text Annotation",
"year": "2012"
},
{
"authors": "Y Si; J Wang; H Xu; K Roberts",
"journal": "Journal of the American Medical Informatics Association",
"ref_id": "b24",
"title": "Enhancing clinical concept extraction with contextual embeddings",
"year": "2019"
},
{
"authors": "E Laparra; S Bethard; T A Miller",
"journal": "JAMIA Open",
"ref_id": "b25",
"title": "Rethinking domain adaptation for machine learning over clinical language",
"year": "2020"
},
{
"authors": "A Smit; S Jain; P Rajpurkar; A Pareek; A Ng; M Lungren",
"journal": "",
"ref_id": "b26",
"title": "Combining Automatic Labelers and Expert Annotations for Accurate Radiology Report Labeling Using BERT",
"year": "2020"
},
{
"authors": "J A Fries; E Steinberg; S Khattar; S L Fleming; J Posada; A Callahan; N H Shah",
"journal": "Nature Communications",
"ref_id": "b27",
"title": "Ontology-driven weak supervision for clinical entity classification in electronic health records",
"year": "2021"
}
] | [] |
2023-05-19 | [
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b4",
"b2",
"b7",
"b7",
"b6",
"b8",
"b7",
"b6",
"b8"
],
"table_ref": [],
"text": "With the emergence of GPT-based (Radford et al., 2018) large-scale models like InstructGPT (Ouyang et al., 2022), ChatGPT (OpenAI, 2022) and GPT-4 (OpenAI, 2023), their remarkable conversational and generative capabilities have garnered widespread attention. These models not only have the capacity to understand complex language structures and grasp subtle meanings but also possess the remarkable capability to interact naturally and fluently with users, generating text that is both coherent and highly creative. This has pushed the boundaries of what was previously deemed impossible. The impact of these large-scale models extends beyond the academic realm of natural language processing (NLP) and has a profound influence in the domains of business and industry. They have opened up new possibilities for humanmachine interactions, intelligent customer service, and virtual assistant applications, revolutionizing these fields and paving the way for innovation and advancement.\nDespite the impressive capabilities of ChatGPT, constructing supervised fine-tuning (SFT) data for instruction tuning presents significant challenges. The human effort required for annotating data, along with issues related to data quality, diversity, accuracy, and others, hinder the development of this technique. Although Self-Instruct (Wang et al., 2022) has been proposed to mitigate this issue, it still relies on a small set of human-written seed instructions for guidance. Furthermore, the method is limited in its ability to control the domain coverage of generated instruction data and ensure the correctness of the generated answers. Consequently, there is a vast amount of untapped potential in utilizing the abundant unsupervised data, particularly domain-specific expertise.\nTherefore, in this paper, we introduce SELF-QA, a framework to generate SFT data from unsupervised knowledge, inspired by the human selfquestioning learning approach. SELF-QA replaces manually written seeds used in other self-alignment models (Wang et al., 2022;Sun et al., 2023;Xu et al., 2023) with a vast amount of unsupervised knowledge, alleviating the difficulty of language models in generating instruction data according to specific requirements. As shown in Figure 1, the Model Prompt Domain Correctness customization guarantee Self-Instruct (Wang et al., 2022) 176 human-written seeds × × Self-Align (Sun et al., 2023) 195 human-written seeds × × Self-Chat (Xu et al., 2023) 111,502 supervised dialogues × × SELF-QA (ours)\nUnsupervised knowledge unsupervised data are used sequentially in the stage of knowledge-guided instruction generation and machine reading comprehension. SELF-QA not only reduces the reliance on human annotators but also allows for the generation of diverse, correct, and domain-specific instruction data. Experiments with unsupervised corpora from various domains demonstrate the effectiveness of our proposed method."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b2",
"b3",
"b7",
"b6",
"b8",
"b7"
],
"table_ref": [
"tab_0"
],
"text": "Language Models with Instruction-tuning Recently, numerous studies (Ouyang et al., 2022;Peng et al., 2023) have investigated the effectiveness of language models in following instructions by leveraging annotated instructional data. This approach enables the model to learn to identify and extract relevant information from different types of instructions and use it to generate accurate and relevant responses. It enhances the model's ability to understand complex instructions and generalize to new tasks by exposing it to a wide range of instructional scenarios. However, the reliance on human annotation in creating such instructional datasets presents a bottleneck for scaling up and achieving broader applicability of instructionguided language models. To address this limitation, researchers have explored alternative approaches that reduce the need for extensive human involvement in generating instruction data.\nBootstrapped Instruction Generation Bootstrapped instruction generation is a recently proposed class of methods (Wang et al., 2022;Sun et al., 2023;Xu et al., 2023) that reduces the cost of human instruction annotation. For example, Self-Instruct (Wang et al., 2022) is proposed to enhance the ability of pre-trained language models to follow instructions by utilizing their own generated samples. This technique involves generating a set of instruction, input, and output samples from the instruction seeds, and then carefully pruning them before fine-tuning the model. Self-Align (Sun et al., Microsoft was founded in 1975, with its headquarters located in Redmond, Washington.\nThe company was founded by Bill Gates. 2023) primarily employs topic-guided red-teaming self-instruct and principle-driven self-alignment to tackle the challenges associated with heavy human annotations. It aims to develop AI agents capable of generating helpful, ethical, and reliable responses to user queries, including adversarial ones, while proactively addressing harmful inquiries in a non-evasive manner. However, as shown in Table 1, these methods often require a small amount of supervised seed information. The instructions generated by them cannot specify domains and content, nor can they ensure the accuracy and professionalism of the instruction responses. Different from them, our approach can effectively address these issues by leveraging unsupervised knowledge."
},
{
"figure_ref": [],
"heading": "Company",
"publication_ref": [
"b9",
"b10",
"b0"
],
"table_ref": [],
"text": "Question Generation and Answering Question generation and question answering are two closely related tasks in natural language processing. They can be viewed as a dual problem, where the former involves creating questions from a given passage or set of information, and the latter involves answering questions based on a given passage or set of information. Especially, the technique of machine reading comprehension (MRC) (Zhang, 2019;Zhang and Wang, 2020) is often used for question answering. For humans, self-questioning and self-answering learning entail stimulating individuals to formulate their own questions and answers based on the provided information, followed by comparing their responses to the original knowledge. This approach has showcased encouraging outcomes in augmenting individuals' understanding of the provided information (Joseph et al., 2016). For domain-specific instruction samples, instruction and input can often be considered as a whole. Therefore, in this paper, we assume that instructions are equivalent to questions, and instruction outputs are equivalent to answers."
},
{
"figure_ref": [],
"heading": "Methodology",
"publication_ref": [],
"table_ref": [],
"text": "Our proposed SELF-QA consists of three different stages: knowledge-guided instruction generation, machine reading comprehension, and filtering and pruning."
},
{
"figure_ref": [],
"heading": "Knowledge-Guided Instruction Generation",
"publication_ref": [],
"table_ref": [],
"text": "In this stage, we employ the language model itself to generate instructions according to unsupervised text. This approach makes the generated instructions domain-specific and content-relevant to the unsupervised text provided. However, in the process of training and inference, instructions are fed to language models without background knowledge, so we need to provide some guidelines so that these instructions cannot rely on and refer to the content in the original text. For instance, the prompt can be:"
},
{
"figure_ref": [
"fig_1"
],
"heading": "Instruction Generation Prompt",
"publication_ref": [
"b11"
],
"table_ref": [],
"text": "The background knowledge is: {unsupervised knowledge data} Please generate ten instruction questions as diverse as possible based on the content of the above article. These questions can be questions about facts or an understanding and evaluation of relevant content. Please assume that there is no corresponding article to refer to when asking questions, so do not use demonstrative pronouns such as \"this\" or \"these\" in the question.\nPlease generate questions in the following format: 1. Question: ... 2. Question: ... Then we can obtain several related instructions, which can be used in the next stage. {unsupervised knowledge data} in the prompt represents sequential text. Unstructured knowledge, such as web pages and book data, can be used directly after undergoing cleaning processes. Structured data such as tables and knowledge graphs (Zhang et al., 2022) need to be converted into unstructured textual data before they can be utilized. As shown in Figure 2, this can be achieved by filling slots using templates or by concatenating each data entry with its corresponding attribute name."
},
{
"figure_ref": [],
"heading": "Machine Reading Comprehension",
"publication_ref": [],
"table_ref": [],
"text": "In this stage, the language model needs to generate answers to the generated instruction questions according to the corresponding unsupervised knowledge. The process can be formulated as follows:\nP (A|K, Q) = j P (A i |A ≤i , K, Q) (1)\nwhere k, Q, A represents unsupervised knowledge, instruction question, and answer, separately. Because the whole process is the same as that of reading comprehension, we also call this stage by this name. As in the previous stage, the prompt for the reading comprehension stage is as follows:"
},
{
"figure_ref": [],
"heading": "Reading Comprehension Prompt",
"publication_ref": [],
"table_ref": [],
"text": "The background knowledge is: {unsupervised knowledge data} Please answer the following question based on the content of the article above: {the generated question} Please answer this question as thoroughly as possible, but do not change the key information in the original text, and do not include expressions such as \"based on the above article\" in the answer.\nPlease generate the corresponding answer in the following format: Question: ... Answer: ..."
},
{
"figure_ref": [],
"heading": "Filtering and Pruning",
"publication_ref": [],
"table_ref": [],
"text": "Although we explicitly instruct the model to assume no prior knowledge from external documents and prohibit the use of demonstrative pronouns like \"this\" in generated questions and the phrase like \"based on the above content\" in generated answers, we still observed that the language model still produces text that violates these rules. Additionally, the generated instances of instructions also exhibit cases where they do not adhere to the required format and become unparseable. Therefore, it is necessary to further filter out these problematic examples.\nKnowledge: Company: DXM Founding Date: April 28, 2018 Formerly known as: Baidu Financial Headquarters Address: Haidian District, Beijing, China."
},
{
"figure_ref": [],
"heading": "Question1:",
"publication_ref": [],
"table_ref": [],
"text": "When was DXM founded? Answer1:\nDXM was founded on April 28, 2018."
},
{
"figure_ref": [],
"heading": "Question2:",
"publication_ref": [],
"table_ref": [],
"text": "Where is the headquarters of DXM located? Answer2:\nThe headquarters of DXM is located at Haidian District, Beijing, China.\nTable 2: Examples of unsupervised background knowledge and generated question and answer pairs."
},
{
"figure_ref": [],
"heading": "Human:",
"publication_ref": [],
"table_ref": [],
"text": "Where is DXM?\nChatGPT:\nThe headquarters of DXM is located in Hangzhou, China.\nOur Model: DXM is a financial technology company headquartered in Haidian District, Beijing, China.\nTable 3: Answers of different models.\nTo mitigate these issues, we implement a postprocessing step to filter out inappropriate responses and correct any formatting errors. This involves developing heuristics and rule-based methods to identify and remove instances that violate the instructed constraints. By applying these filters, we ensure that the generated text adheres to the predefined guidelines and maintains the desired level of correctness and coherence."
},
{
"figure_ref": [],
"heading": "Discussion",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Performance",
"publication_ref": [
"b5"
],
"table_ref": [],
"text": "We collect several domains of unsupervised unstructured and structured data for experiments. An example of unsupervised knowledge and generated instruction questions and answers are shown in Table 2. We then instruction-tuning BLOOM-7B (Scao et al., 2022) with these generated instructions. As shown in Table 3, our model can answer the corresponding question correctly, but ChatGPT gives a wrong answer. It is precisely because of these domain-specific instruction-tuning data that our model has achieved better performance."
},
{
"figure_ref": [],
"heading": "Different Stages of SELF-QA",
"publication_ref": [],
"table_ref": [],
"text": "The stage of knowledge-guided instruction generation and machine reading comprehension can also be integrated into a single stage so that the model only needs to be invoked once for each round of instruction generation and answer prediction. The advantage of this is that the number of calls to the model is reduced, because each round of instruction question and answer generation only needs language models once. However, there are also potential drawbacks to this approach. For instance, the model may generate output that exceeds the predetermined length. Additionally, by combining these two tasks, the model may not be able to focus on a single task as effectively, which can result in less detailed and accurate answers. Therefore, the decision to integrate two stages into a single stage should be made with careful consideration of the specific application and task requirements."
},
{
"figure_ref": [],
"heading": "Different Forms of Knowledge",
"publication_ref": [],
"table_ref": [],
"text": "In general, knowledge can be stored in large language models in a parametric manner or separately input into the models in an explicit symbolic form. The main focus of this paper is on how to store unsupervised knowledge in large models using a parameterized approach. This approach enables end-to-end processing of user questions and optimization of model parameters without the need for external information. It offers a high level of flexibility and adaptability to different inputs and contexts. However, this approach also comes with potential biases and errors that can be present in the data. Therefore, it is crucial to provide comprehensive and accurate knowledge during the training phase to mitigate the impact of such biases on the model. On the other hand, explicit symbolic knowledge requires the existence of corresponding retrieval and query systems. Additionally, the model needs to make judgments on whether to adopt the content of external knowledge. This makes the entire process more complex."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "In this paper, we introduced SELF-QA, a framework for generating instruction-tuning data from unsupervised knowledge. The unsupervised data are used sequentially in the stage of knowledgeguided instruction generation and machine reading comprehension. Our experiments demonstrate the effectiveness of SELF-QA in generating diverse, correct, and domain-specific instruction data. By reducing the reliance on human annotators, SELF-QA offers a promising approach for improving the efficiency and scalability of instruction tuning."
}
] | Large-scale language models like ChatGPT and GPT-4 have gained attention for their impressive conversational and generative capabilities. However, the creation of supervised paired question-answering data for instruction tuning presents formidable challenges. This endeavor necessitates substantial human effort for data annotation and wrestles with issues concerning data quality, diversity, accuracy, and other related factors. To overcome these obstacles, we introduce an innovative framework named SELF-QA, which replaces the traditional practice of human-written instruction seeds with a vast amount of unsupervised knowledge, enabling the model to generate a larger quantity of correct and domainspecific instruction data. The effectiveness of our proposed method is demonstrated through experiments conducted on unsupervised corpora from various domains. | SELF-QA: Unsupervised Knowledge Guided Language Model Alignment | [
{
"figure_caption": "Figure 1 :1Figure 1: The pipeline of SELF-QA.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: Examples of transformation of unsupervised structured data.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Comparison of different self-alignment methods.",
"figure_data": "",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
}
] | Xuanyu Zhang; Qing Yang; Du Xiaoman Financial | [
{
"authors": "Laurice M Joseph; Sheila Alber-Morgan; Jennifer Cullen; Christina Rouse",
"journal": "Reading & Writing Quarterly",
"ref_id": "b0",
"title": "The effects of self-questioning on reading comprehension: A literature review",
"year": "2016"
},
{
"authors": " Openai",
"journal": "",
"ref_id": "b1",
"title": "Chatgpt. OpenAI",
"year": "2022"
},
{
"authors": "Long Ouyang; Jeff Wu; Xu Jiang; Diogo Almeida; Carroll L Wainwright; Pamela Mishkin; Chong Zhang; Sandhini Agarwal; Katarina Slama; Alex Ray; John Schulman; Jacob Hilton; Fraser Kelton; Luke Miller; Maddie Simens; Amanda Askell; Peter Welinder; Paul Christiano; Jan Leike; Ryan Lowe",
"journal": "",
"ref_id": "b2",
"title": "Training language models to follow instructions with human feedback",
"year": "2022"
},
{
"authors": "Baolin Peng; Chunyuan Li; Pengcheng He; Michel Galley; Jianfeng Gao",
"journal": "",
"ref_id": "b3",
"title": "Instruction tuning with gpt-4",
"year": "2023"
},
{
"authors": "Alec Radford; Karthik Narasimhan; Tim Salimans; Ilya Sutskever",
"journal": "",
"ref_id": "b4",
"title": "Improving language understanding by generative pre-training",
"year": "2018"
},
{
"authors": "Le Teven; Angela Scao; Christopher Fan; Ellie Akiki; Suzana Pavlick; Daniel Ilić; Roman Hesslow; Alexandra Castagné; François Sasha Luccioni; Matthias Yvon; Gallé",
"journal": "",
"ref_id": "b5",
"title": "Bloom: A 176bparameter open-access multilingual language model",
"year": "2022"
},
{
"authors": "Zhiqing Sun; Yikang Shen; Qinhong Zhou; Hongxin Zhang; Zhenfang Chen; David Cox; Yiming Yang; Chuang Gan",
"journal": "",
"ref_id": "b6",
"title": "Principle-driven selfalignment of language models from scratch with minimal human supervision",
"year": "2023"
},
{
"authors": "Yizhong Wang; Yeganeh Kordi; Swaroop Mishra; Alisa Liu; Noah A Smith; Daniel Khashabi; Hannaneh Hajishirzi",
"journal": "",
"ref_id": "b7",
"title": "Self-instruct: Aligning language model with self generated instructions",
"year": "2022"
},
{
"authors": "Canwen Xu; Daya Guo; Nan Duan; Julian Mcauley",
"journal": "",
"ref_id": "b8",
"title": "Baize: An open-source chat model with parameter-efficient tuning on self-chat data",
"year": "2023"
},
{
"authors": "Xuanyu Zhang",
"journal": "Association for Computational Linguistics",
"ref_id": "b9",
"title": "MCˆ2: Multi-perspective convolutional cube for conversational machine reading comprehension",
"year": "2019"
},
{
"authors": "Xuanyu Zhang; Zhichun Wang",
"journal": "",
"ref_id": "b10",
"title": "Rception: Wide and deep interaction networks for machine reading comprehension (student abstract)",
"year": "2020"
},
{
"authors": "Xuanyu Zhang; Qing Yang; Dongliang Xu",
"journal": "Association for Computational Linguistics",
"ref_id": "b11",
"title": "TranS: Transition-based knowledge graph embedding with synthetic relation representation",
"year": "2022"
}
] | [
{
"formula_coordinates": [
3,
332.66,
297.87,
191.75,
22.12
],
"formula_id": "formula_0",
"formula_text": "P (A|K, Q) = j P (A i |A ≤i , K, Q) (1)"
}
] |
|
2023-05-19 | [
{
"figure_ref": [],
"heading": "I. INTRODUCTION",
"publication_ref": [
"b20",
"b1",
"b20"
],
"table_ref": [],
"text": "In the modern world a large amount of data is generated by all the interconnected devices and networks and collected via sensors [21]. These sequences of measurements from the sensors are indexed with a time stamp and called time series data. This massive amount of data can not be monitored without automation, so time series analysis has become increasingly important. In the context of change point detection, time series analysis is used to detect changes in the behaviour of the system, which can be due to internal or external properties. The task of identifying a specific point in time where the statistical properties of the underlying model of a signal or time series change is called change point detection. [2] Change point detection is an essential topic in time series analysis with a wide range of applications that require the detection of abrupt changes in the data, for example to indicate a transition of the system from one state to another or to indirectly detect the state of a system by estimating that the state changed [21]. Here are some examples of applications that use change point detection:"
},
{
"figure_ref": [],
"heading": "Climate change detection",
"publication_ref": [
"b5",
"b9",
"b21",
"b13",
"b25"
],
"table_ref": [],
"text": "Climate analysis and climate monitoring have become very important over the last few decades due to the possible occurrence of climate change caused by the increase in greenhouse gases in the atmosphere. In this context change point detection is used to discover climatic discontinuities and changes in the temperature. [6], [10], [22] Medical condition monitoring Change point detection methods can be used to monitor patients health data and quickly detect abrupt changes of the medical conditions. These algorithms can be used to analyze changes in physical activity during the strength training program [14] or to monitor the heart rate of patients undergoing anesthesia and surgery to indicate disturbances to the cardiovascular system [26]."
},
{
"figure_ref": [],
"heading": "Stock Market Analysis",
"publication_ref": [
"b3",
"b24"
],
"table_ref": [],
"text": "In the stock market the fluctuation of any stock price is normal according to economic theory, but there are some shifts that are abnormal and worth the investors special attention [4]. In momentum strategies it is necessary to identify momentum turning points, when a trend reverses from an uptrend to a downtrend such as in the 2020 market crash due to the covid outbreak [25].\nThis article presents an overview and comparison of algorithms commonly used for detecting change points in time series data. The focus is on unsupervised change point detection, which involves segmenting the data without relying on large amounts of annotated training data or the need to re-calibrate the model for each data source. The goal of this article is to help choosing the right detection method for a particular application, with an emphasis on practical aspects like the implementation and the calibration of the parameters. Our selection of methods aims for a good general performance for different data sources without fine tuning the algorithm. In practice fine tuning a method for each sensor will most probably yield better performance, but adds significant overhead if the system has multiple heterogeneous signals with the number of signals possibly growing in the future. We are especially focusing on methods that can be applied to different heterogeneous sensor signals, like temperature or pressure measurements, without the requirement to tune the parameters separately for each data source. The methods should be able to detect change points in different 1-dimensional sensor signals from a complex heterogeneous system to discover dependencies between the signals."
},
{
"figure_ref": [],
"heading": "II. BACKGROUND",
"publication_ref": [
"b2",
"b2"
],
"table_ref": [],
"text": "A change point occurs when the statistical properties of the time series data change abruptly. This change of measurements can be caused, for example, by a transition in the state of the underlying data generating system. These changes may be indicated by a simple shift in mean or variance, or more complex shifts such as in the frequency domain. An abrupt change refers to any alteration in the system's parameters that happens either instantly or very quickly relative to the measurement sampling period. It's important to note that abrupt changes don't necessarily indicate large shifts. In fact, detecting small changes is often the main challenge in most applications. [3] In time series analysis mainly two concepts are used to detect deviations from the normal behavior in a data set: outlier detection and change point detection. The outlier detection is a method used to identify data points that are significantly different from the rest of the data, but are mainly very short shifts in the system or measurement errors. In contrast to that, change point detection aims to discover points in time where the state of the data generator changes for a subsequent and longer period. Therefore they are related, but not entirely similar. Figure 1 illustrates the difference between these concepts by showing an example time series that contains several change points, where the mean and variance of the time series switches abruptly, and outliers, where an individual observation is far from the majority of the data. The main underlying assumption when investigating change points is that the properties or parameters describing the data are either constant or slowly changing over time in each state of the system [3]. "
},
{
"figure_ref": [],
"heading": "A. Offline and online algorithms",
"publication_ref": [
"b4",
"b1"
],
"table_ref": [],
"text": "Change point detection algorithms are traditionally divided into two main branches: online and offline methods. Although both classes of algorithms aim to detect a change point in time series data, their methods of operation differ substantially. The online change point detection algorithms operate in real-time and run concurrently with the process they are monitoring, processing each data point as it becomes available. The constraint for these algorithms is that the processing should be completed before the arrival of the next data point [5]. These methods are typically used in embedded applications such as monitoring industrial processes. Online change point detection methods must be able to process data quickly and make decisions in near real time. However, it is worth noting that online methods can also be applied in an offline setting, by analyzing stored data in a sequence.\nIn contrast to that are the offline algorithms, which retrospectively detect changes and considers the entire data set. These methods can also be used in an online setting by collecting a small batch of data and detecting the change points in this batch. In general the concept of online and offline methods is a spectrum describing the number of data points the algorithm have to look at ahead of a potential change point, with complete offline methods considering the entire data set and complete online methods operating in perfect real-time. The offline methods allow the detection of multiple change points at once and focus more on the accuracy of the detection, while the online methods try to detect the most recent change point as fast as possible. Offline change point algorithms are typically more computationally intensive than online algorithms, as they have access to the entire data set and can use more complex methods to detect change points [2]. The choice between online and offline methods should be based on the specific requirements of the application, with the most important consideration being the real-time constraint. Since the motivation for this paper is an application in an offline setting with the entire data set available, we will cover online and offline methods."
},
{
"figure_ref": [],
"heading": "B. Supervised and unsupervised algorithms",
"publication_ref": [
"b1",
"b1",
"b1",
"b20"
],
"table_ref": [],
"text": "A variety of supervised and unsupervised methods have been developed and improved over time to tackle the challenge of change point detection. The supervised methods utilize machine learning algorithms that are trained on labeled data, where the change point locations are already known. The objective of these supervised algorithms is to build a model that can learn from the limited labeled data and accurately predict the change points for the entire data set. A variety of supervised methods can be used for this learning problem, such as decision tree, naive Bayes, support vector machines, logistic regression and nearest neighbor methods [2]. One of the main drawbacks of supervised change point detection methods is the need for labeled data, as well as the need to retrain the model for each individual data source. This can make the process of applying these supervised methods to different data sources very costly and time consuming.\nUnsupervised learning algorithms are better suited for this as they are used to identify patterns in unlabeled data. In the context of change point detection, these algorithms can divide time series data into segments, with one segment before the change point and one after. These algorithms utilise statistical aspects of the data to locate change points in the time series. Unsupervised change point detection is attractive because it can handle a variety of different data sets without requiring prior training for each data set [2]. Both types of methods have been shown to be effective, but for the purposes of this paper, we will focus solely on unsupervised methods. This decision is based on the underlying application of this paper that requires methods that are easily applicable to different data sources without retraining the algorithm. For an overview on supervised methods, we refer readers to [2], [21]."
},
{
"figure_ref": [],
"heading": "C. Stability",
"publication_ref": [
"b1"
],
"table_ref": [],
"text": "An important criteria for our application is the stability of the algorithms in terms of their ability to detect change points in different data sources without requiring prior fine tuning of the parameters for each data source. We are comparing the number of parameters and evaluate the effort of calibrating the algorithm. The parameters should be robust to a wide variation of values to minimize the effort of the parameter tuning. An important aspect for the stability of the algorithms is also the assumptions the algorithm makes about the underlying data source.\nIn order to increase the stability of the algorithms, we are especially looking for algorithms that make little to no assumptions, therefor we are comparing the change point detection algorithms by determining if they are parametric or non parametric. It is important to make this distinction, because parametric models make assumptions about the underlying distribution of the data, such as for example the in the real world often appearing and therefor commonly used normal distribution. The distribution is described by a set of parameters, which have to be estimated from the data. The main issue of parametric models is that the assumptions about the underlying distribution may not hold for all data sources, which can lead to inaccurate estimates of the change points. The parametric nature of the algorithm can also be a benefit and lead to good results if the underlying distribution of the data is known. In contrast, the non-parametric models do not make assumptions about the underlying distribution of the data, which makes them more robust to different data sources. They skip the step of parameter estimation by using a non parametric approach by for example directly estimating the ratio between to densities without calculating the densities itself. [2] "
},
{
"figure_ref": [],
"heading": "D. Algorithm constraints",
"publication_ref": [],
"table_ref": [],
"text": "Change point detection techniques can be differentiated based on the requirements imposed on the input data and the algorithm itself. These limitations play a crucial role in determining the most suitable method for detecting change points in a specific time series. Especially in our application it is important to have as little constraints as possible to ensure the algorithm can be applied to multiple data sources. Some algorithms for example are constraint to stationary or independent and identically distributed (i.i.d.) data sets. Another important constraint is the number of change points an algorithm can detect and if the number of change points has to be known prior to the run of the algorithm. Based on that we are classifying the algorithms into single and multiple change point detection algorithms. Algorithms that require the number of change points to be specified before the run will not be considered, since this is a major drawback when analysing different time series with an unknown amount of change points. The algorithms also differ in their outputs, with some algorithms providing the probability of a change point at a specific time point and others only returning the change points. The different outputs determine how easy the results can be interpreted and can for example give insights in the confidence of the detected change point."
},
{
"figure_ref": [],
"heading": "E. Scalability",
"publication_ref": [],
"table_ref": [],
"text": "The amount of time series data generated in the world is rapidly increasing, both in terms of the number of data points and dimensions. To effectively handle this massive amount of data, change point detection methods must be designed to be computationally efficient. Therefore, it is crucial to compare the computational cost of different change point detection algorithms and determine which methods can reach an optimal or approximate solution as quickly as possible. The problem of detecting change points in high dimensional data sets is not discussed in this paper, since our application is dealing with univariant time series. The computational costs of the algorithms are compared based on the information provided by the authors or estimated based on the algorithmic description."
},
{
"figure_ref": [],
"heading": "III. REVIEW",
"publication_ref": [],
"table_ref": [],
"text": "This section provides an overview of commonly used change point detection algorithms. We will outline the basic principles of the algorithm and evaluate them based on the in section II defined criterion. These algorithms include both online and offline methods and where chosen based on their stability, algorithm constraints and scalability. All of the algorithms presented are unsupervised learning methods."
},
{
"figure_ref": [],
"heading": "A. Likelihood ratio methods",
"publication_ref": [
"b11"
],
"table_ref": [],
"text": "A typical statistical formulation of change point detection is to compare the probability distributions p θ 0 and p θ 1 of the data before and after a potential change point xt. If these distributions are significantly different, the time point is considered a change point. One common method involves monitoring the logarithm of the likelihood ratio between consecutive intervals in the time series data. [12] \ns(xi) = ln p θ 1 (xi) p θ 0 (xi)\nWe will discuss two different approaches that utilize the likelihood ratio in change point detection. The first approach is called the cumulative sum algorithm and relies on predesigned parametric models, where the probability density of the two consecutive intervals is calculated separately from the underlying data and the density ratio is calculated. The second approach is a direct density ratio estimation method, which is a non parametric algorithm where the ratio of the probability densities is directly estimated, without the requirement to perform density estimation for the individual segments."
},
{
"figure_ref": [],
"heading": "1) CUSUM:",
"publication_ref": [
"b19",
"b2"
],
"table_ref": [],
"text": "On of the most familiar change point detection algorithms is the cumulative sum algorithm (CUSUM). Page [20] was the first to suggest the use of a cumulative sum to find changes in a parameter of interest. The CUSUM algorithm monitors the cumulative sum of the differences between successive measurements and a reference value or target. If this cumulative sum exceeds a predetermined threshold, it indicates that there has been a change point in the process. The log likelihood ratio is a measure of the difference between two distributions, and can be used to determine whether a change point has occurred.\nThe typical behavior of the log likelihood ratio corresponds to a negative drift before change, and a positive drift after change. Therefore, the relevant information for change point detection lies in the difference between the value of the log-likelihood ratio and its current minimum value [3]. Assume we have a sequence x1, x2, ... of time series variables with a unknown change point xtt and the distribution p θ 0 (xt) before and p θ 1 (xt) after the change point xt.\nThen the corresponding decision rule for the CUSUM algorithm is:\nSn = n i=1 ln p θ 1 (xi) p θ 0 (xi) -min k≤n k i=1 ln p θ 1 (xi) p θ 0 (xi) > L(1)\nThe CUSUM algorithms detects a change point when the CUSUM statistic Sn is larger then the predefined threshold L. Another common approach, that is computationally more efficient, is to calculate the CUSUM statistic Sn recursively in the following way:\nSn = max 0, Sn-1 + ln p θ 1 (xi) p θ 0 (xi)(2)\nIn practice, the CUSUM algorithm is implemented by first setting an initial value for the CUSUM statistic S0, usually set to zero.\nThen the algorithm iterates through the observations, updating the CUSUM statistic at each time step. The algorithm continues until a change point is detected or the end of the data is reached."
},
{
"figure_ref": [],
"heading": "Stability:",
"publication_ref": [
"b10",
"b22",
"b2"
],
"table_ref": [],
"text": "The CUSUM algorithm is parametric and makes assumptions about the underlying distribution of the data. In each iteration of the algorithm the set of parameters describing the distribution before a potential change point p θ 0 and after the change point p θ 1 have to be estimated from the data. A common application is when the process is normally distributed with the pre-change and post-change distributions having the same known variance σ. Then the interest centers on detecting shifts away from the pre-shift mean µ0 to µ1. The algorithm can handle different types of distributions, but if the assumption of the underlying distribution is not correct the result of the algorithm may vary. [11] The algorithm has only two parameters, with the initial value of the CUSUM statistic usually set to zero and the threshold value L which is more difficult to calibrate. The threshold value depends on the specific data set and the goals of the analysis, for example it can be set such that the false positive rate or the false negative rate is minimized or determined using cross-validation techniques. Different methods to calculate values of L have been developed with a constant or adaptive threshold. A comparison of these methods is given in [23].\nAlgorithm constraints: The standard CUSUM algorithm is an online algorithm for single change point detection, but many variations of this concept were developed over time for online and offline settings, as well as adaptions to multiple change point detection by running the algorithm with a sliding window. An overview of four of these adaptions is provided in [3]. The algorithm has no constraints or limitations to the data it can be applied to wide variety of problems. The algorithm does not provide any information about the confidence of the output, it stops when a single change point is detected, indicated by the CUSUM statistic exceeding the threshold, and directly outputs the change point."
},
{
"figure_ref": [],
"heading": "Scalability:",
"publication_ref": [],
"table_ref": [],
"text": "The main advantage of CUSUM is the simplicity of the underlying concept, which makes it easy to implement. The computational complexity of the CUSUM algorithm is generally considered to be low, as it only involves simple mathematical operations such as addition. The complexity estimated based on the algorithm is O(n) for a sequence of n points."
},
{
"figure_ref": [],
"heading": "2) KLIEP:",
"publication_ref": [
"b7",
"b11",
"b15",
"b7",
"b11",
"b15"
],
"table_ref": [],
"text": "The Kullback-Leibler importance estimation procedure (KLIEP) is the non parametric counter part to the CUSUM algorithm and also uses the likelihood ratio to detect change points. KLIEP avoids the problems that come with the parametric model of the CUSUM algorithm by using a more flexible non parametric model. This grants the advantage that it does not rely on strong model assumptions. A naive method would be to use a non parametric approach to estimate the densities separately and use them to calculate the ratio, but this approach is ineffective due to the challenges of non parametric density estimation. Instead of individually estimating each density, we can directly estimate the density ratio. Different online methods have been developed that use the idea of direct density ratio estimation [8], [12], [16]. In these methods the density ratio between two consequent intervals X and X is modeled by a non parametric Gaussian kernel:\nŵ(X) = p(X) p(X ) = n l=1 α l Kσ(X, X l ) (3)\nwhere p(X) is the probability distribution of the interval X and the parameters α are learned from data samples. Kσ(Y, Y ) is the Gaussian kernel function with mean Y and standard deviation σ.\nThe parameters α are determined in a training phase by minimizing a dissimilarity measure. From the density ratio estimator ŵ(X) an approximation of the dissimilarity measure between two samples is calculated and the higher the dissimilarity measure is, the more likely the point is a change point [8]. One of these methods is the Kullback-Leibler importance estimation procedure (KLIEP) that estimates the density ratio by using Kullback-Leibler (KL) divergence:\nKL[p(x)||p (x)] = -p (x) log p(x) p (x) dx\nThe parameters α in the formula 3 are determined so that the empirical Kullback-Leibler divergence is minimized. The solution of this problem can be obtained by solving a convex optimization problem.\nWith the estimated parameters, the logarithm of the likelihood ratio is evaluated as the change detection score and a change point is detected if the score is beyond a given threshold\nL S = n i=1 ln ŵ(Xi) > L(4)\nStability:\nThe KLIEP algorithm has several tuning parameters, such as the kernel width σ, the lengths of the two intervals X and X and the threshold for the change point score L. The main advantage of the method described in [12] is that it is equipped with a natural cross validation procedure for tuning the kernel width parameter σ.\nThe interval length has an impact on the accuracy of the estimation of the density ratio, with larger intervals leading to better estimates but also to possible issues in data sources where change points occur very frequently. The detection of the change points is sensitive to the threshold parameter L, with possible methods to calibrate L already discussed in chapter 1) CUSUM.\nThe main advantage in terms of the stability of KLIEP compared to the CUSUM algorithm is the robustness to different types of data distributions. The KLIEP algorithm is non parametric and doesn't make assumptions about the underlying distribution of the data. This is especially useful if the underlying distribution of the data is unknown or the algorithm is applied to different data sources that do not necessarily have the same distribution.\nAlgorithm constraints: The KLIEP and CUSUM algorithm share a lot of properties, with their sequential behaviour both are online methods use for single change point detection and can be adapted for multiple change point problems with small adjustments. The algorithm has no constraints or limitations to the data and directly returns the change point.\nScalability: Compared with other approaches like the CUSUM algorithm, methods based on density ratio estimation tend to be computationally more expensive because of the cross-validation procedure for model selection [16]. After the model selection the calculation of the change point score for a given density ratio is computationally very efficient and involves only simple mathematical operations that are performed in linear time."
},
{
"figure_ref": [],
"heading": "B. Bayesian online change point detection",
"publication_ref": [
"b0",
"b0",
"b0"
],
"table_ref": [],
"text": "A common bayesian method for detecting changes, described by Adam and MacKay [1], involves estimating the posterior distribution of the run length rt, which describes the time that passed since the last change point. The algorithm is an online method, so in each iteration a new data point is considered and the run length can either increase by 1 or drop down to 0. The set of observation x (r) t is associated with the run length rt and includes all the observations of the current run, with new observations added to the set until a change point is found, which sets rt to 0 and x This probability of a change point can be calculated based on Bayes theorem as: P (rt|x1:t) = r t-1 P (rt|rt-1)P (xt|rt-1, x r t )P (rt-1, x1:t-1) P (x1:t) Since P (rt-1|xt-1) is a recursive component, it is known from the previous step and we only need to calculate the conditional run length probability P (rt|rt-1) and evaluate the predictive distribution P (xt|rt-1, x r t ). The conditional prior on the change point P (rt|rt-1) is only nonzero at the two outcomes rt=0 or rt=rt-1, which gives the algorithms its computational efficiency. [1] \nP (rt|rr-1) = 1 -H(rt-1 + 1), if rt = rt-1 + 1 H(rt-1 + 1), if rt = 0 0, otherwise\nThe function H(r) is called the hazard function and describes how likely a change point is occurring at a run length r. A common approach is to make this process memoryless by setting H(r) = 1 λ with a timescale parameter λ. Another option would be to make H(r) increasing over the run to penalize longer run lengths. The predictive distribution P (xt|rt-1, x r t ) represents the probability that the most recent observation belongs to current run and is the most challenging to calculate. In [1] it was proposed to use a conjugate exponential model with parameters ν and χ to make the algorithm computationally efficient.\nStability The bayesian online change point detection algorithm is a parametric method. The method allows us to encode knowledge about the world into the algorithm by setting the two conjugate priors χ and ν. The conjugate priors χ and ν are the only two parameters of the algorithm and are set based on prior knowledge. This prior knowledge could be the mean and variance of the data, which is estimated roughly based on the first few observations. A possible option for further tuning the algorithm it by setting the hazard function H(r), but the common memoryless approach is sufficient for most applications."
},
{
"figure_ref": [],
"heading": "Algorithm constraints:",
"publication_ref": [
"b0"
],
"table_ref": [],
"text": "The algorithm is a online multiple change point detection method that assumes the sequence of observations can be segmented into non-overlapping partitions and the data within each partition p is i.i.d. from some probability distribution P (xt|µp) [1]. The restriction to i.i.d. time series is a major drawback for this algorithm when applying it to multiple different data sources. The algorithm directly returns the probability for a change point occurring at the specific run lengths, which makes the results easier to interpret compared to methods like CUSUM or KLIEP that only return the change point itself."
},
{
"figure_ref": [],
"heading": "Scalability:",
"publication_ref": [
"b0",
"b16"
],
"table_ref": [],
"text": "The implementation of the algorithm proposed by Adam and MacKay [1] is quadratic in space and time complexity in the number of data points n so far observed. Clearly this is problematic when analysing long time series. To overcome this issue a simple approximation was introduced in [17] that reduced the complexity to O(n)."
},
{
"figure_ref": [
"fig_3"
],
"heading": "C. Singular spectrum transformation",
"publication_ref": [
"b18",
"b18",
"b8",
"b8"
],
"table_ref": [],
"text": "The singular spectrum transformation (SST) algorithm first proposed by Moskvina and Zhigljavsky [19] uses singular spectrum analysis and adapts it for change point detection. The singular spectrum transform is a non parametric approach and can be used to analyze time series with complex structure [19]. This technique involves transforming the original time series into a new series of change point scores. The resulting time series can be interpreted as the probability distribution that some change point occurs at time t.\nIn figure 3 such a change point score with the corresponding time series is shown with a clear peak in the change point score when the frequency of the time series changed. The underlying idea is to compute for each time point xt the difference between a representative pattern of a few time points before and a few points after xt. The dynamics of the time series are represented using a Hankel matrix. We call the Hankel matrix, representing the change patterns within the past w points, trajectory matrix H(t) and the Hankel matrix for representing the future change patterns the test matrix G(t). The representative pattern of H(t) and G(t) can be extracted by performing a singular value decomposition on both matrices. The l < w left singular vectors Ui 1 (t), ..., Ui l (t) of H(t) with the largest singular values represent the past change pattern as a hyperplane and build the matrix U (t) = Ui 1 (t), ..., Ui l (t). The matrix U (t) encodes the major direction of change in the past signal and the parameter l is the number of representative patterns that are considered. The importance of each representative pattern Ui(t) is given by the corresponding singular value, with the most dominant pattern corresponding to the largest singular value. The direction of maximum change in the future of the signal is given by the left singular vector β(t) of G(t) with the largest singular value. [9] To estimate the difference between the past and the future patterns, β(t) will be projected onto U (t) and normalized to calculate a change point score. If there is no change in the dynamics of the signal, it is expected that β(t) will lie in or very near to the hyperplane represented by U l . The change point score ranges from zero to one, with a high likelihood of a change point occurring when the score is close to one. The score can be calculated at any time point t by finding representative patterns in both the trajectory and test matrices. This can be seen as a transformation from the original time series T to a new time-series Tc. This demonstrates that SST can make variables of different types comparable by converting a heterogeneous system into a homogeneous one [9], i.e.\nT → Tc(w, l, g, m, n)\n(5)"
},
{
"figure_ref": [],
"heading": "Stability:",
"publication_ref": [
"b8",
"b17",
"b8"
],
"table_ref": [],
"text": "The main problem of the SST algorithm is the need to specify five parameters: the length of the column vectors of the Hankel matrix w, the number of columns of H(t) n, the number of singular vectors l, the shift of the starting point for the future signal g and the number of columns of G(t) m. In [9], [18] it was shown that SST is usually robust to a wide range in different values of w and n and domain knowledge or visualization can help finding an appropriate value for both. The problem lies in choosing the other three parameters, where domain knowledge is not very useful.\nDespite the problem of specifying five different parameters it was shown in [9] that for a wide range of w (6 < w < 40), the results of the algorithms with w = -g = m = n and l = 3 are quite robust and the essential features remain unchanged."
},
{
"figure_ref": [],
"heading": "Algorithm constraints:",
"publication_ref": [
"b18",
"b8",
"b17"
],
"table_ref": [],
"text": "The SST is based on the singular value decomposition of two Hankel matrix, so it is a non parametric method, as it does not rely on any specific assumptions about the distribution of the data, making it suitable for a wide range of time series data [19]. The advantage of the SST algorithms lies in its comparability of different time series. The algorithms returns a change point score for each time point, which allows to detect multiple change points in one run. The change point score can be interpreted as the probability that a change point occurred at a specific time point, which makes it easy to interpret the output of the algorithm and discover dependencies within multiple different heterogeneous time series. [9] Scalability: The complexity of the singular spectrum transformation is linear in the length of the time series [18]. In each of the n time steps a singular value decomposition of the fixed size Hankel matrices H(t) and G(t) have to be computed, which makes SST less computational efficient as other change point detection algorithms with linear complexity."
},
{
"figure_ref": [],
"heading": "D. Binary Segmentation",
"publication_ref": [
"b23",
"b14",
"b23",
"b6",
"b23"
],
"table_ref": [],
"text": "Binary segmentation is a common technique for offline change point detection in time series data due to its simplicity both in the underlying concept and the implementation of the algorithm. A implementation of the algorithm available in the ruptures python library for offline change point detection and is called BinSeg [24]. The binary segmentation procedure is one of the standard methods for detecting multiple change points by using a test for single change point detection [15].\nThe algorithm is a recursive technique for multiple change point detection in which initially the entire data set is searched for one change point. Once a change point is detected, the data set is split into two subsegments, defined by the detected change point. The algorithm is then performed on each of the two subsegments resulting in further splits. This procedure continuous in each new split until a stopping criteria is satisfied or no change point is detected. The indices of the segment boundaries are the change points. This process of recursively splitting the segments into two smaller subsegments is shown in the schematic view of the algorithm in figure 4.\nThis technique can be used with a variety of time series and is able to find changes in the mean, variance, or distribution of the A-6 Figure 4. Schematic example of binary segmentation from [24] data. Binary segmentation is considered a greedy algorithm, because it is performed sequentially, with each stage only visited once and depending on the previous ones. [7] The algorithm starts by initializing the starting and ending indices s and e of the time series sequence, for the first run start s = 1 and end e = n, where n is the number of data points. Then a cost function c(x a...b ) is defined to measure the cost of a segment x a...b , with x a...b = xi|a <= i <= b. The cost function should be chosen based on the nature of the data and the goal of the analysis. An overview and comparison of common cost functions is given in [24]. The algorithm calculates the costs for each possible split by summing the cost function for the two resulting segment and returns the change point k that minimizes the costs:\nk = arg min s<k<e c(x s...k ) + c(x k...e )(6)\nThe algorithm then recursively calls the binary segmentation algorithm with the new start and end indices: BinSeg(s, k) and BinSeg(k, e). This process is repeated until a stopping criterion is met, such as a minimum number of data points in a segment, the maximum number of change points or a maximum number of iterations. Once the stopping criterion is met, the change points are assigned as the indices of the segment boundaries."
},
{
"figure_ref": [],
"heading": "Stability:",
"publication_ref": [
"b23"
],
"table_ref": [],
"text": "The algorithm is very simple and has only the cost function and a stopping criteria as inputs. It's important to note that the choice of the cost function c() and the stopping criterion are crucial to the performance of the algorithm. The stopping criterion can be chosen according to the number of change points expected in the data, the granularity of the segments, or a threshold of the cost function for a segment. In general, the choice of cost function and stopping criterion depends on the nature of the data and the goal of the analysis.\nThe algorithm can be used in a parametric and non parametric setting by selecting the respective cost function, depending on the prior knowledge before performing the change point detection.\nThe choice of the cost function is determining the assumption the algorithm makes about the data. In [24] an overview of possible parametric and non parametric cost functions is given with their underlying assumptions about the data.\nAlgorithm constraints: Binary segmentation is a simple and versatile offline algorithm for change point detection, that can be applied to any time series without any limitations. The algorithm can be used in situations where the number of change points is known prior to the run but also for an unknown number of change points. However, the algorithm has a limitation in accurately detecting change points that are too close together and result in small segments. This can lead to inaccuracies in the results for time series, that frequently change between different states resulting in close change points. The binary segmentation algorithm directly returns the detected change points, with the most significant change point appearing first. This feature allows the user to stop the algorithm at any point in time and still obtain a meaningful result, making it an efficient tool for change point detection."
},
{
"figure_ref": [],
"heading": "Scalability:",
"publication_ref": [
"b23"
],
"table_ref": [],
"text": "The benefits from the simplicity of binary segmentation includes the low complexity of the algorithm of O(Cnlogn), where n is the number of samples and C the complexity of calling the considered cost function on one subsegment [24]."
},
{
"figure_ref": [],
"heading": "E. Bottom up Segmentation",
"publication_ref": [
"b23",
"b23",
"b12",
"b23",
"b23",
"b12"
],
"table_ref": [],
"text": "Bottom-up segmentation is the natural counterpart of binary segmentation. In contrast to binary segmentation, bottom up segmentation starts by dividing the time series into individual observations, each treated as a separate segment, followed by sequentially merging adjacent segments based on a discrepancy criterion. A schematic overview of the bottom up segmentation is provided in figure 5, showing the different steps of the algorithm starting with a segmented time series in a grid and merging these segments together until a stopping criterion is met and the approximation of the change points is returned. Figure 5. Schematic example of bottom up segmentation from [24] Bottom up segmentation is like its counterpart binary segmentation available in the ruptures python library [24]. It is also an offline method with a very simple underlying concept and low computational complexity. \nAll potential change points are ranked by the discrepancy measure with the lowest discrepancy change point getting deleted, meaning that the corresponding segments are merged. When two adjacent segments t and t + 1 are merged, the algorithm performs some bookkeeping tasks and calculates the cost of merging the newly formed segment with its right neighbor segment. Additionally, the costs for merging the left neighbor with the newly formed larger segment have to be recalculated. These bookkeeping steps are necessary to ensure the accuracy of the bottom up segmentation process and its ability to detect change points in the time series. [13] The process of merging and evaluating segments is repeated until no further improvements can be made. This results in a set of segments that represent the best approximation of the underlying A-7 process changes in the time series. The final step is to identify the change points, which correspond to the boundaries between the segments and represent the points where the underlying process changes.\nStability: Due to the simple underlying concept of the bottom up algorithm the only two inputs are the cost function, used in the discrepancy measure, and a stopping criteria. The choice of the two inputs have a significant impact on the performance of the algorithm. The stopping criterion can be chosen according to the number of change points expected in the data, the size of the segments, or a threshold of the discrepancy measure for two segments. In general, the cost function and stopping criterion should be chosen based on the nature of the time series and the goal of the analysis.\nSimilar to binary segmentation this algorithm can be parametric and non parametric, depending on the selected cost function. The cost function should be chosen based on the prior knowledge of the data source before performing the change point detection. The choice of the cost function is determining the assumption the algorithm makes about the data. In [24] an overview of possible parametric and non parametric cost functions is given with their underlying assumptions.\nAlgorithm constraints: The algorithm has no limitations in the time series it can be applied to and can be used for a known and unknown number of change points. Due to the very fine grid in the initialization the first iterations of merging procedures can be unstable because of the small segments they are performed on, for which the statistical significance is smaller. This problem could be overcome by starting with larger segments. However if a true change point does not belong to the original set of boarders of these segments, it would never be considers for a change point, resulting in an inaccurate detection of the change points. [24] Scalability: The algorithm for bottom up segmentation has a similar complexity as their counterpart binary segmentation with O(C n log(n)), where n is the number of samples in the time series and C the complexity of calling the discrepancy measure for two adjacent subsegments [13]."
},
{
"figure_ref": [],
"heading": "IV. DISCUSSION AND COMPARISON",
"publication_ref": [],
"table_ref": [],
"text": "The previous sections present an overview of change point detection algorithms that are commonly used in the literature and have a good general performance for different data sources without fine tuning each algorithm for the particular data source. The task of selecting the algorithm best suited for a particular application can be challenging and depends on which criterion is most important for the application. We compare change point detection methods based on their stability, algorithm constraints and their scalability to help with this task."
},
{
"figure_ref": [],
"heading": "A. Stability",
"publication_ref": [
"b8",
"b1"
],
"table_ref": [],
"text": "When evaluating the stability of algorithms, one important factor to consider is the number of parameters and the robustness of these parameters to various changes. While all algorithms except singular spectrum transformation have a small number of parameters, the choice of these parameters can have a significant impact on the performance of the algorithms. For instance, in the case of the CUSUM or KLIEP algorithm, only one threshold parameter L needs to be assigned. However, the threshold value in CUSUM or KLIEP and the other stopping criterion in binary segmentation or bottom up segmentation can be very sensitive, and choosing an incorrect threshold value can result in either an early or a late stopping of the algorithm, which leads to inaccurate results. In contrast, singular spectrum transformation requires five different parameters to be assigned, but it has been shown to be effective in detecting change points in a variety of data sources without adjusting the parameters [9]. Similarly, bayesian online change point detection algorithm only requires the specification of the conjugate priors as input, which is essentially a problem of estimating the distribution of the data. The absence of a stopping criteria in both singular spectrum transformation and bayesian online change point detection makes these algorithms relatively more stable compared to other algorithms.\nTo enhance the stability of a change point detection method, it is important to consider the assumptions the algorithm makes about the data. In general, non parametric change point detection methods tend to be more robust than parametric methods, as the latter heavily depend on the choice of parameters used to model the distribution of the data [2]. A naive but common approach is to use a parametric model and always assume that the different data sources have a normal distribution. Since this assumption is true for a lot of real world applications this approach may produce good results for data sets that actually follow a normal distribution, but it can lead to poor performance when the data deviates from a normal distribution. In contrast, non-parametric methods, such as KLIEP, binary segmentation, or bottom-up segmentation, do not make any assumptions about the underlying distribution of the data and are therefore more robust. A comprehensive overview of the algorithms and their parameters, along with a distinction between parametric and non-parametric methods, can be found in Table I. "
},
{
"figure_ref": [],
"heading": "B. Algorithm Constraints",
"publication_ref": [],
"table_ref": [],
"text": "When choosing the appropriate algorithm for a particular application, it is important to take into account the constraints of the algorithm. This review focuses on three main constraints: the restrictions on the time series, the number of change points that can be detected, and the output of the algorithm. The Likelihood ratio methods have no limitations on the time series, but they are only capable of detecting a single change point, so multiple runs of the algorithm may be required to detect multiple change points. This problem is overcome by directly using methods such as binary or bottom-up segmentation, which are designed for multiple change point detection and also do not have any limitations on the time series. Bayesian online change point detection has the only restriction on the time series in that it requires the time series to be independently and identically distributed. Singular Spectrum Transformation, on the other hand, requires the time series to be stationary. It is important to consider these constraints when selecting the right algorithm for a given application to ensure that the algorithm is capable of meeting the specific needs and requirements of the data.\nThe various algorithms also have different outputs, with some only providing the change points and others providing a change point probability or score. The bayesian online change point A-8 detection and singular spectrum transformation algorithms provide a change point probability or score, but this information requires a separate post processing step to assign the actual change points. However, this additional information provides insight into the confidence of the algorithm in the identified change points. On the other hand, the likelihood ratio methods and binary and bottom up segmentation algorithms directly return the change points, avoiding the need for a separate post processing step. A comprehensive overview of the algorithms and their constraints can be found in table II. "
},
{
"figure_ref": [],
"heading": "C. Scalability",
"publication_ref": [],
"table_ref": [],
"text": "Another important criteria to consider are the computational costs of change point detection algorithms. Table III presents a comparison of the computational cost of the algorithms surveyed. For some algorithms the computational costs where provided by the authors. In cases where the authors have not provided this information, the comparison was conducted based on the descriptions of the algorithms.\nThe computational cost of CUSUM is relatively low, as it requires only a few calculations for each time step. The time complexity of CUSUM is linear in the number of data points n, making it a suitable option for large data sets. The KLIEP algorithm, on the other hand is a slightly more complex algorithm with a higher time complexity as CUSUM due to the model selection required.\nThe computational cost of bayesian online change point detection depends on the choice of priors and the complexity of the model used. In general, it is computationally more expensive and has a quadratic time complexity O(n 2 ), but this can be reduced to a linear complexity with the help of a simple approximation.\nThe singular spectrum transform has also a linear time complexity but is less efficient as CUSUM or KLIEP as it requires the calculation of a singular value decomposition in each time step, which has a high time complexity. The computational complexity of binary segmentation and bottom up segmentation depends on the cost function used in the algorithm. Both have a complexity of O(C n log(n)), where n is the number of samples in the time series and C the complexity of calling the cost function.\nThis review covered both online and offline change point detection methods. In general the online methods are computationally less expensive, because they focus on detecting the most recent change point as fast as possible. This was also confirmed by the algorithms in this review with two offline methods binary segmentation and bottom up segmentation having the highest computational complexity. "
},
{
"figure_ref": [],
"heading": "V. CONCLUSION",
"publication_ref": [],
"table_ref": [],
"text": "In this review, we have analyzed various techniques for change point detection and organized them under a unified framework. Our emphasis was on unsupervised methods that have a good general performance for diverse data sources. To assess the methods, we compared them based on three criteria: stability, algorithmic constraints, and scalability. The aim is to provide a framework that can be utilized to assess future algorithms in a systematic manner. The focus of this review was chosen based on a particular application that was the motivation for this paper. It is worth noting that in practice fine tuning a specific algorithm for a single time series will most probably yield better performance, but significantly increases the time and effort needed to detect change points in a large system of heterogeneous signals. The comparison of the methods was based solely on the description of the algorithm and leaves room for further research by for example comparing the performance of the methods on different data sources and evaluating the results."
}
] | Change point detection is a crucial aspect of analyzing time series data, as the presence of a change point indicates an abrupt and significant change in the process generating the data. While many algorithms for the problem of change point detection have been developed over time, it can be challenging to select the appropriate algorithm for a specific problem. The choice of the algorithm heavily depends on the nature of the problem and the underlying data source. In this paper, we will exclusively examine unsupervised techniques due to their flexibility in the application to various data sources without the requirement for abundant annotated training data and the re-calibration of the model. The examined methods will be introduced and evaluated based on several criteria to compare the algorithms. | Unsupervised Change Point Detection for heterogeneous sensor signals | [
{
"figure_caption": "A- 2 Figure 1 .21Figure 1. Time series with change points and outliers",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "21",
"figure_type": "figure"
},
{
"figure_caption": "algorithm determines the probability P (rt = 0|x1:t) of a change point occurring at time t by calculating the probability of the run length of 0 at time t given the set of observations x1:t = x1, ..., xt. The functionality of the algorithm is visualized in figure2with the run length and the change point probability represented as a grey scale, with darker pixels indicating a higher probability. The plot shows that the run length drops to zero when the probability of a change point gets close to one.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 .2Figure 2. Example of bayesian online change point detection from [1]",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 .3Figure 3. Example of a time series and the corresponding change point score from [9]",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "The algorithm for bottom up segmentation starts by creating the finest possible representation of the time series of size n by dividing it into n 2 segments. The indices of the segment boundaries are the potential change points. For each potential change point t a discrepancy measure d(xa...t, x t...b ) is calculated between the segments xa...t and x t...b separated by t, with these segments defined as x a...b = {xi|a ≤ i ≤ b}. For a given cost function c(), the discrepancy measure between two adjacent segments is given by: d(xa...t, x t...b ) = c(x a...b ) -c(xa...t) -c(x t...b )",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "",
"figure_type": "figure"
}
] | Mario Krause | [
{
"authors": "Ryan Prescott; Adams ; David J C Mackay",
"journal": "",
"ref_id": "b0",
"title": "Bayesian online changepoint detection",
"year": "2007"
},
{
"authors": "Samaneh Aminikhanghahi; Diane J Cook",
"journal": "Knowledge and information systems",
"ref_id": "b1",
"title": "A survey of methods for time series change point detection",
"year": "2017"
},
{
"authors": "Michèle Basseville; Igor V Nikiforov",
"journal": "Prentice Hall, Inc",
"ref_id": "b2",
"title": "Detection of Abrupt Changes -Theory and Application",
"year": "1993"
},
{
"authors": "Jie Chen; Arjun K Gupta",
"journal": "medicine, and finance",
"ref_id": "b3",
"title": "Parametric statistical change point analysis: with applications to genetics",
"year": "2012"
},
{
"authors": "Allen B Downey",
"journal": "",
"ref_id": "b4",
"title": "A novel changepoint detection algorithm",
"year": "2008"
},
{
"authors": "Jean-François Ducré-Robitaille; Lucie A Vincent; Gilles Boulet",
"journal": "International Journal of Climatology",
"ref_id": "b5",
"title": "Comparison of techniques for detection of discontinuities in temperature series",
"year": "2003"
},
{
"authors": "Piotr Fryzlewicz",
"journal": "The Annals of Statistics",
"ref_id": "b6",
"title": "Wild binary segmentation for multiple change-point detection",
"year": "2014"
},
{
"authors": "Mikhail Hushchyn; Andrey Ustyuzhanin",
"journal": "J. Comput. Sci",
"ref_id": "b7",
"title": "Generalization of changepoint detection in time series data based on direct density ratio estimation",
"year": "2020"
},
{
"authors": "Tsuyoshi Ide; Keisuke Inoue",
"journal": "",
"ref_id": "b8",
"title": "Knowledge discovery from heterogeneous dynamic systems using change-point correlations",
"year": "2005-04"
},
{
"authors": "Naoki Itoh; Jürgen Kurths",
"journal": "",
"ref_id": "b9",
"title": "Change-point detection of climate time series by nonparametric method",
"year": "2010"
},
{
"authors": "R Daniel; Veronica Jeske; Montes De; Wolfgang Oca; Mazda Bischoff; Marvasti",
"journal": "Computational Statistics and Data Analysis",
"ref_id": "b10",
"title": "Cusum techniques for timeslot sequences with applications to network surveillance",
"year": "2009"
},
{
"authors": "Yoshinobu Kawahara; Masashi Sugiyama",
"journal": "Statistical Analysis and Data Mining: The ASA Data Science Journal",
"ref_id": "b11",
"title": "Sequential change-point detection based on direct density-ratio estimation",
"year": "2012"
},
{
"authors": "Eamonn Keogh; Selina Chu; David Hart; Michael Pazzani",
"journal": "IEEE",
"ref_id": "b12",
"title": "An online algorithm for segmenting time series",
"year": "2001"
},
{
"authors": "Patrick Lapointe; Kévin Chapron; Isabelle Lessard; Kevin Bouchard; Mélissa Lavoie; Cynthia Gagnon; Elise Duchesne; Sébastien Gaboury",
"journal": "",
"ref_id": "b13",
"title": "Monitoring changes in physical activity data during strength training of people with myotonic dystrophy type 1",
"year": "2022"
},
{
"authors": "Marc Lavielle; Gilles Teyssiere",
"journal": "Long memory in economics",
"ref_id": "b14",
"title": "Adaptive detection of multiple change-points in asset price volatility",
"year": "2007"
},
{
"authors": "Song Liu; Makoto Yamada; Nigel Collier; Masashi Sugiyama",
"journal": "Neural Networks",
"ref_id": "b15",
"title": "Change-point detection in time-series data by relative density-ratio estimation",
"year": "2013"
},
{
"authors": " Rakesh Malladi; P Giridhar; Behnaam Kalamangalam; Aazhang",
"journal": "",
"ref_id": "b16",
"title": "Online bayesian change point detection algorithms for segmentation of epileptic activity",
"year": "2013"
},
{
"authors": "Yasser Mohammad; Toyoaki Nishida",
"journal": "",
"ref_id": "b17",
"title": "Robust singular spectrum transform",
"year": "2009"
},
{
"authors": "Valentina Moskvina; Anatoly Zhigljavsky",
"journal": "Communications in Statistics-simulation and Computation -COMMUN STATIST-SIMULAT COMPUT",
"ref_id": "b18",
"title": "An algorithm based on singular spectrum analysis for change-point detection",
"year": "2003-01"
},
{
"authors": "E S Page",
"journal": "Biometrika",
"ref_id": "b19",
"title": "Continuous inspection schemes",
"year": "1954"
},
{
"authors": "Aditya Pushkar; Muktesh Gupta; Rajesh Wadhvani; Manasi Gyanchandani",
"journal": "",
"ref_id": "b20",
"title": "A comparative study on change-point detection methods in time series data",
"year": "2022"
},
{
"authors": "Jaxk Reeves; Jien Chen; Xiaolan L Wang; Robert Lund; Qi Qi; Lu ",
"journal": "Journal of Applied Meteorology and Climatology",
"ref_id": "b21",
"title": "A review and comparison of changepoint detection techniques for climate data",
"year": "2007"
},
{
"authors": "Nassim Sahki; Anne Gégout-Petit; Sophie Wantz-Mézières",
"journal": "Quality and Reliability Engineering International",
"ref_id": "b22",
"title": "Performance study of change-point detection thresholds for cumulative sum statistic in a sequential context",
"year": "2020"
},
{
"authors": "Charles Truong; Laurent Oudre; Nicolas Vayatis",
"journal": "Signal Processing",
"ref_id": "b23",
"title": "Selective review of offline change point detection methods",
"year": "2020"
},
{
"authors": "Kieran Wood; Stephen J Roberts; Stefan Zohren",
"journal": "",
"ref_id": "b24",
"title": "Slow momentum with fast reversion: A trading strategy using deep learning and changepoint detection",
"year": "2021"
},
{
"authors": "P Yang; G Dumont; J M Ansermino",
"journal": "IEEE Transactions on Biomedical Engineering",
"ref_id": "b25",
"title": "Adaptive change detection in heart rate trend monitoring in anesthetized children",
"year": "2006"
}
] | [
{
"formula_coordinates": [
3,
129.92,
442.93,
71.22,
21.19
],
"formula_id": "formula_0",
"formula_text": "s(xi) = ln p θ 1 (xi) p θ 0 (xi)"
},
{
"formula_coordinates": [
3,
339.44,
85.47,
215.23,
26.84
],
"formula_id": "formula_1",
"formula_text": "Sn = n i=1 ln p θ 1 (xi) p θ 0 (xi) -min k≤n k i=1 ln p θ 1 (xi) p θ 0 (xi) > L(1)"
},
{
"formula_coordinates": [
3,
361.32,
164.64,
193.36,
21.19
],
"formula_id": "formula_2",
"formula_text": "Sn = max 0, Sn-1 + ln p θ 1 (xi) p θ 0 (xi)(2)"
},
{
"formula_coordinates": [
4,
99.15,
125.01,
192.51,
27.03
],
"formula_id": "formula_3",
"formula_text": "ŵ(X) = p(X) p(X ) = n l=1 α l Kσ(X, X l ) (3)"
},
{
"formula_coordinates": [
4,
85.24,
273.16,
161.79,
19.75
],
"formula_id": "formula_4",
"formula_text": "KL[p(x)||p (x)] = -p (x) log p(x) p (x) dx"
},
{
"formula_coordinates": [
4,
122.92,
354.36,
168.74,
45.57
],
"formula_id": "formula_5",
"formula_text": "L S = n i=1 ln ŵ(Xi) > L(4)"
},
{
"formula_coordinates": [
4,
323.33,
451.11,
205.46,
30.41
],
"formula_id": "formula_6",
"formula_text": "P (rt|rr-1) = 1 -H(rt-1 + 1), if rt = rt-1 + 1 H(rt-1 + 1), if rt = 0 0, otherwise"
},
{
"formula_coordinates": [
6,
101.47,
366.51,
190.19,
15.3
],
"formula_id": "formula_7",
"formula_text": "k = arg min s<k<e c(x s...k ) + c(x k...e )(6)"
}
] |
|
10.18653/v1/n19-1423 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b8",
"b18",
"b17",
"b16",
"b14",
"b39",
"b24",
"b4",
"b18",
"b34",
"b27"
],
"table_ref": [],
"text": "Aspect Based Sentiment Analysis (ABSA) is a finegrained variant of sentiment analysis (Hu and Liu, 2004;Pontiki et al., 2014Pontiki et al., , 2015Pontiki et al., , 2016;;Zhang et al., 2021a;Shu et al., 2022;Zhang et al., 2022), where the task is to predict the sentiment expressed towards an entity or a certain aspect of an entity, instead of just the sentence-level sentiment (e.g., traditional sentiment analysis tasks (Socher et al., 2013;dos Santos and de C. Gatti, 2014)).\nFor illustration, for a review The pizza was great, but the service was terrible, a sentence-level sentiment analysis model might identify the sentiment as neutral. The need for ABSA stems from such complex interactions between the target and the polarity of the sentiment (Pontiki et al., 2014). An ABSA model has to identify the sentiment towards pizza as positive, and service as negative, for a holistic understanding of the text. Furthermore, * Work done during internship at AWS AI Labs ABSA tasks can include the identification of the opinion terms (i.e. great, terrible), and the aspect categories (i.e. FOOD, SERVICE) (Zhang et al., 2021a).\nAlthough traditionally considered as a structured prediction task in the ABSA literature, recent works have shown how sequence-to-sequence (seqto-seq) models can be effective in these tasks with a generative approach (Yan et al., 2021;Zhang et al., 2021a). Such approaches leverage the knowledge gained from one task to seamlessly perform well in another. As such, we build upon the Instruction Tuning with Multi-Task Learning approach (Varia et al., 2022) and address the following five ABSA tasks: (i) Aspect-term Extraction (AE), (ii) Aspect-term Extraction and Sentiment Classification (AESC), (iii) Target Aspect Sentiment Detection (TASD), (iv) Aspect Sentiment Triplet Extraction (ASTE), and (v) Aspect Sentiment Quadruple Prediction (ASQP).\nSentence-level sentiment annotations are comparatively cheaper and are available at scale through automated proxies (e.g., ☀ or ☀☀ become negative, ☀☀☀☀ or ☀☀☀☀☀ become positive, in the Amazon/Yelp review corpus (Zhang et al., 2015b)). On the contrary, ABSA requires understanding at sub-sentence level with multiple words or phrases being related to each other, making it prohibitively costly to annotate at scale.1 However, the abundance of generic review data presents a promising opportunity to improve the performance of a pre-trained language model (PLM) beyond simply fine-tuning it on the small annotated ABSA corpora.\nTowards this end, we first construct a noisily annotated ABSA corpus out of generic customer review data without any direct supervision. We utilize this noisy corpus to pre-train a seq-to-seq model on multiple ABSA tasks. We show that such models are capable of learning in zero/few-shot in final downstream ABSA tasks. Our contributions are the following: (i) We propose a weakly supervised method to obtain annotations for three out of the five ABSA tasks explored in the literature; (ii) We introduce a pre-training step to improve the few-shot performance on the downstream task of PLMs; (iii) We comprehensively evaluate our proposed method in three scenarios (full fine-tuning, few-shot, and zero-shot learning), yielding as much as 15.84% F1 improvement over the SOTA baselines. We release the sources to create the few-shot benchmarking datasets 2 ."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b26",
"b14",
"b39",
"b5",
"b13",
"b26",
"b26",
"b28",
"b31",
"b12",
"b29",
"b35",
"b33",
"b15",
"b34",
"b33",
"b39",
"b33",
"b39",
"b31",
"b23",
"b33",
"b0",
"b14",
"b14",
"b38",
"b34",
"b27",
"b27",
"b27"
],
"table_ref": [],
"text": "Aspect-Based Sentiment Analysis has received tremendous attention in the past years (Tulkens and van Cranenburgh, 2020;Zhang et al., 2021a;Shu et al., 2022;Zhang et al., 2022), either handling single tasks, such as aspect term extraction (He et al., 2017;Liu et al., 2015;Tulkens and van Cranenburgh, 2020), aspect category detection (Tulkens and van Cranenburgh, 2020), aspect sentiment classification (Vo and Zhang, 2015;Xu et al., 2019;Li et al., 2021;Wang et al., 2021), or handling compound tasks (Zhang et al., 2015a;Yu et al., 2021;Xu et al., 2020;Zhang et al., 2021a). For the latter group, it typically includes either a pipeline approach (Peng et al., 2020;Yan et al., 2021) or an end-to-end (E2E) approach (Xu et al., 2020;Zhang et al., 2021a,b).\nIn the pipeline approach the final prediction is constructed using the output of multiple components. The disadvantage of such models is that the error is propagated throughout the system (Zhang et al., 2022).\nIn the E2E approach, the model learns the interactions jointly between the multiple prediction tasks, which is believed to improve the final performance (Xu et al., 2020;Zhang et al., 2022). Our proposed approach falls in this category. Typical E2E approaches include: (i) treating it as a token classification task (Xu et al., 2019;Shu et al., 2019;Xu et al., 2020), (ii) framing it as a machine reading comprehension task (Chen et al., 2021;Liu et al., 2022), natural language inference task (Shu et al., 2022), or as a language generation task (Zhang et al., 2021b;Yan et al., 2021;Zhang et al., 2021a;Varia et al., 2022).\n2 https://github.com/robertvacareanu/ NoisyABSAPreTraining Our proposed approach treats the ABSA tasks as a generation task, similar to (Zhang et al., 2021a;Varia et al., 2022). We build upon the paradigm called Instruction Tuning with in Multi-Task Learning (IT-MTL), introduced in (Varia et al., 2022), resulting in a single model capable of handling different ABSA tasks. However, none of these methods takes advantage of the vast amount of review data available, other than just pre-training on them with some generic language modeling objectives."
},
{
"figure_ref": [],
"heading": "Method",
"publication_ref": [
"b7",
"b2",
"b19"
],
"table_ref": [],
"text": "We introduce an additional step in the classical pretrain → finetune approach (Howard and Ruder, 2018;Devlin et al., 2019;Raffel et al., 2020), transforming it into pretrain → Noisy ABSA Pre-Training (NAPT) → finetune for ABSA. We propose an approach for building a weakly annotated dataset for the intermediate NAPT step. We use this noisy dataset to enhance the knowledge of a pretrained model with the intuition that exposing the model to tasks which are well aligned with the final downstream task, improves the performance. We then consider this as the backbone base model, and finetune it on the downstream task as usual. Our proposed approach is applicable to any generic seq-to-seq model."
},
{
"figure_ref": [],
"heading": "Dataset Construction",
"publication_ref": [
"b32",
"b39"
],
"table_ref": [],
"text": "The first step in our proposed method is to weakly annotated a dataset without any direct supervision. 3Our proposed approach annotates a dataset with tuples of the form aspect-terms, opinion-terms, and sentiment polarity. We follow a pipeline approach as shown in Table 1 (Xu et al., 2013;Zhang et al., 2022), but without using any direct ABSA supervision. We describe each step in greater detail next."
},
{
"figure_ref": [],
"heading": "Aspect-term Extraction",
"publication_ref": [
"b6"
],
"table_ref": [
"tab_8"
],
"text": "The first step in our proposed dataset creation procedure is aspect-term extraction. We use spacy tokenizer to obtain POS tags and then consider 20% of the most frequent nouns in the text. These nouns serve as candidate aspect terms. We note that this method implicitly assumes that dataset D consists of a single domain. Nevertheless, this is a small assumption as the reviews are typically directed towards a product of a known category (He and McAuley, 2016;Zhang et al., 2015b). We extend Sentence: The pizza was great, but the service was terrible.\nStep 1: A step-by-step illustration of our noisy dataset construction pipeline. It follows a pipeline approach, and yields <aspect, opinion, sentiment> triplets in the end for each sentence in a generic review corpus.\nthis method to multi-word aspect terms by considering collocations of length ≤ 4 filtered by their POS tags. For example, we allow bigrams of the form NN-NN like chicken breast (cf Table 16 for all patterns used). Finally, we filter out the sentences from which no aspect term was extracted."
},
{
"figure_ref": [],
"heading": "Opinion-term Extraction",
"publication_ref": [
"b3",
"b9",
"b8",
"b8"
],
"table_ref": [],
"text": "The second step in our proposed algorithm is opinion term extraction. We take a lexicon-based approach to opinion extraction (Ding et al., 2008;Kanayama and Nasukawa, 2006;Hu and Liu, 2004). In particular, we use the opinion lexicon from (Hu and Liu, 2004) and perform word matching on the target text. If negations e.g., no or not appear before the opinion word, we include it in the final extraction as well. We filter out the sentences from which no opinion term was extracted."
},
{
"figure_ref": [],
"heading": "Linking Opinion-terms with",
"publication_ref": [
"b1",
"b14",
"b25"
],
"table_ref": [],
"text": "Aspect-terms So far the resulting dataset consists of noisy aspect, and opinion terms, but without the association between them. For example, for a sentence such as The pizza was great , but the service was terrible., the proposed algorithm would extract pizza and service as the aspect terms and great and terrible as the opinion terms, respectively. But at this point we do not know that great refers to pizza and terrible refers to service. We reformulate this problem as a natural language inference problem (Dagan et al., 2005;Shu et al., 2022). We use an MPNet 4 model (Song et al., 2020) and construct artificial sentences to determine which opinion-term refers to which aspect-term. More precisely, we construct sentences such as <aspect-term> is <opinion-term>, for each aspect-and opinionterm. 5 Then, we use the original sentence (e.g.\n4 huggingface.co/symanto/mpnet-base-snli-mnli 5 We relax strict grammatical correctness e.g., the formulation might result in burgers is great instead of burgers are great).\nThe pizza was great , but the service was terrible.) as the premise and our artificially constructed sentence as the hypothesis (e.g. pizza is great). We interpret a high entailment score (≥ 0.75) as evidence that the opinion term refers to that particular aspect term. We discard aspect-and opinion-term pairs where the entailment score was below the threshold. Alternative Approach: We consider an alternate approach where the linking is based on constituency-parse rules which turns out disadvantageous. Constituency parsing is considerably slower and the rules are non-trivial to formulate."
},
{
"figure_ref": [],
"heading": "Sentiment Extraction",
"publication_ref": [
"b8",
"b20"
],
"table_ref": [],
"text": "The last step in our proposed dataset creation method is to add the sentiment (Hu and Liu, 2004) to each <aspect-term, opinion-term> tuple. We use a sentence-level classifier on top of artificially constructed sentences (Sanh et al., 2019). For example, for a tuple such as <pizza, great>, we feed the sentence pizza is great through a sentencelevel sentiment classifier. 6 Then, we label the <aspect term, opinion term> tuple with the sentiment prediction if the model's confidence is above a certain threshold (≥ 0.75), otherwise we discard the tuple. At the end of this step, for the sentence The pizza was great , but the service was terrible. we have the following <aspect-term, opinion-term, sentiment> noisy annotations: <pizza, great, positive>, <service, terrible, nega-tive>. We consider an alternative for this step using the sentiments associated in the opinion lexicon, but a classifier allows for confidence filtering.\nThroughout our proposed dataset creation process we use external resources, such an opinion lexicon, an NLI model and a sentence-level sentiment classifier. However, these resources do not consume any annotated ABSA data by any means."
},
{
"figure_ref": [
"fig_0"
],
"heading": "Noisy ABSA Pre-training (NAPT)",
"publication_ref": [
"b21"
],
"table_ref": [],
"text": "The phase consists of exposing the model to tasks that are more aligned with the final downstream task, i.e., ABSA in our case. We factorize the triplets from the noisy dataset into five separate but overlapping tasks: (i) aspect-term extraction, (ii) opinion-term extraction, (iii) aspect-term and opinion-term extraction, (iv) aspect-term extraction and sentiment prediction, and (v) aspect-term extraction, opinion-term extraction and sentiment prediction. Note that there exists a correspondence between our NAPT tasks and classical ABSA tasks: tasks (i), (iv) and (v) correspond to Aspect Extraction (AE), Aspect Extraction Sentiment Classification (AESC), and Aspect Sentiment Triplet Extraction (ASTE), respectively. We use the noisy ABSA dataset to pre-train the base model. We train the model parameters in a multi-task learning framework (cf Figure 1) using instruction tuning with a diverse set of instructions (Sanh et al., 2022). At the end of NAPT, the resulting model is imbued with the capability of performing multiple ABSA tasks. This can serve as a drop-in replacement to the off-the-shelf pre-trained checkpoints that are widely used in the generative ABSA literature."
},
{
"figure_ref": [],
"heading": "Addressing Overfitting",
"publication_ref": [
"b10"
],
"table_ref": [],
"text": "The primary goal of our proposed NAPT phase is to enhance the pre-trained model while retaining existing knowledge from pre-training objectives, in other words, avoiding catastrophic forgetting and overfitting. We achieve this in a few different ways. First, instead of just randomly splitting the data into train/validation, we split the extracted aspectand opinion-terms into two disjoint sets, favoring novel aspect-and opinion term constructions in the validation partition. We observe this split definition to be necessary to prevent overfitting of the base model. Additionally, we invoke three types of regularization:\n• Standard weight decay: we add a standard 2 regularization term to the loss function.\n• Tuple Dropout: we apply dropout over the tuples that the model is trained to extract to prevent it from overfitting to the noisy annotations. We randomly dropped 50% of the tuples from prediction targets of the seq-to-seq model.\n• Biased weight decay: we use a biased variant of weight decay to prevent the parameters from diverging considerably from the initialization point, akin to (Kirkpatrick et al., 2017). Towards this, we use the 2 norm over the difference between the current (θ) and the initial weights of the model (θ init ), and add it to the loss. Our final loss function (L) is:\nL = CE loss + α ⋅ 2 (θ -θ init ) + β ⋅ 2 (θ). (1\n)\nwhere α and β are hyperparameters, and CE loss denotes the standard cross-entropy loss."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [],
"table_ref": [],
"text": "We compare against state-of-the-art methods on three widely used ABSA datasets. We evaluate in three scenarios: (i) k-shot learning: where the model has access to at least k examples of each class, (ii) zero-shot evaluation: where the model has not seen any example at all from the goldannotated ABSA data, and (iii) full-training: where the model has access to the complete gold-standard training data,"
},
{
"figure_ref": [],
"heading": "Experimental Setup",
"publication_ref": [
"b19",
"b14",
"b6"
],
"table_ref": [],
"text": "In all our experiments, we use T5 (Raffel et al., 2020), particularly t5-base as the pre-trained seqto-sed model, which has ∼ 220M parameters. We experiment with t5-large as well to explore the impact of model size on the downstream performance (cf Appendix B). We use the standard evaluation metrics as previous work, which is F1 score over the exact match of the tuples. For zero-shot, we use the same evaluation procedure as (Shu et al., 2022), which is token-level F1 score. We use a random subset of Amazon Electronics (He and McAuley, 2016), and Yelp reviews (Zhang et al., 2015b) to create our noisy-annotated dataset. 7 We split the reviews with ≥ 3 sentences using a sentence tokenizer. We split the noisy dataset into train/validation split. We enforce that there is no overlap in terms of aspect-terms between the train/validation splits. This results in approximately 190k examples for training and 12.5k examples for validation.\nWe repeat each experiment with 5 different random seeds. Additionally, we repeat the noisy ABSA pre-training step with 3 different random seeds. As a result, the numbers corresponding to our proposed method (i.e. the ones with -APT) represent an average of 5 × 3 = 15 runs, and all the other numbers represent an average of 5 runs. We report the mean and (sample) standard deviation.\nWe present the results on the Aspect Sentiment Triplet Extraction (ASTE) and Aspect-term Extraction and Sentiment Classification (AESC) tasks available in all the datasets we use for evaluation.8 "
},
{
"figure_ref": [],
"heading": "Datasets",
"publication_ref": [
"b18",
"b17",
"b16",
"b27"
],
"table_ref": [],
"text": "We use three popular datasets for aspect-based sentiment analysis: REST15, REST16 and LAP14 (Pontiki et al., 2014(Pontiki et al., , 2015(Pontiki et al., , 2016)), which cover two domains: restaurant and laptop, respectively. In particular, we use the version released by Zhang et al.. For k-shot, we use the same splits as (Varia et al., 2022) to ensure a fair comparison. Specifically, the k-shot datasets were created by sampling k examples for each attribute. The attributes are aspect category, and sentiment for restaurant, and laptop respectively."
},
{
"figure_ref": [],
"heading": "Baselines",
"publication_ref": [
"b27"
],
"table_ref": [],
"text": "Since we introduce the NAPT step and build upon the existing Instruction Tuning with Multi-Task Learning (IT-MTL) paradigm, we refer to our proposed method as IT-MTL-NAPT. We compare this with standard fine-tuning based approaches that generally show strong performance in ABSA tasks i.e., (i) text-only (Text), where we give the model the text review and train it to predict the gold text (Zhang et al., 2021a), (ii) instruction tuning (IT) and (iii) instruction tuning + multi-task learning, as per (Varia et al., 2022) "
},
{
"figure_ref": [],
"heading": "(IT-MTL).",
"publication_ref": [],
"table_ref": [],
"text": "To succinctly show the effectiveness of proposed NAPT, we keep another baseline where a seq-toseq model is further pre-trained with in-domain data using the same objective as that of t5 i.e., span prediction. We call it IT-MTL-ID.9 The in-domain data is essentially the same as that of the NAPT corpus, but without the noisy annotations."
},
{
"figure_ref": [],
"heading": "K-Shot Learning",
"publication_ref": [],
"table_ref": [
"tab_4"
],
"text": "Next, we compare between the two approaches in k-shot learning scenarios. We summarize our results in Figure 2. IT, and IT-MTL-ID perform similarly with the other baselines, so we skip them for clarity. We include all our results in Appendix B.2. First we observe that, our proposed method outperforms the baselines across all datasets in all k-shot scenarios, yielding as much as 15.84% F1 points (i.e. from 13.04%F1 to 28.88%F1) of improvement. Second, the performance improvement increases as the number of examples decrease, with the biggest improvement being in the k=5 case. This is expected because with the growing number of examples, all models are able to learn the task better. When using the full dataset, as we see in Table 3, both the proposed model and the baseline performances converge. Additionally, we observe that our proposed method brings the larger improvements on the harder tasks, as it gets difficult for the "
},
{
"figure_ref": [],
"heading": "Zero-Shot Evaluation",
"publication_ref": [
"b14"
],
"table_ref": [],
"text": "Our proposed NAPT step enables the model to perform the following ABSA tasks in zero-shot i.e., without any gold-standard supervision: (i) Aspectterm Extraction (AE), (ii) Aspect-term Extraction and Sentiment Classification (AESC), and (iii) Aspect Sentiment Triplet Extraction (ASTE). We perform two experiments in the zero-shot setting. First, we investigate how much data does a baseline need to reach the performance obtained by our proposed model in the zero-shot setting. Second, we compare against previous work in the ASTE task (Shu et al., 2022)."
},
{
"figure_ref": [
"fig_2"
],
"heading": "Dataset Size Equivalence",
"publication_ref": [],
"table_ref": [],
"text": "We compare our proposed method in zero-shot setting against a baseline model trained on goldannotated data, where we vary the number of training data points. This experiment shows how many annotated data points, on average, is the noisy ABSA pre-training phase equivalent of. We observed that the improvement depends on the difficulty of the task and of the dataset, respectively. For example, Figure 3 shows that for the ASTE task, one would need ∼ 15, 25 annotated data points to obtain a comparable performance with our pro- posed method for REST15 and LAP14 respectively. We remark that the number of data points vary according to the difficulty of the task and with the difficulty of the dataset, ranging between ∼ 6 -25 data points for AE, and ASTE task for LAP14 respectively."
},
{
"figure_ref": [],
"heading": "Performance Comparison with Baselines",
"publication_ref": [
"b14"
],
"table_ref": [],
"text": "We compare the zero-shot performance of our proposed method with previous work on ABSA (Shu et al., 2022), summarized in "
},
{
"figure_ref": [],
"heading": "Full-Training",
"publication_ref": [],
"table_ref": [
"tab_4"
],
"text": "We compare the performance of our proposed method (i.e. pretrain → NAPT → finetune) with the standard method of pretrain → finetune and report the result in Table 3, for all the datasets. Overall in the full-training scenario, our proposed method performs comparably with or better than the baseline. We observe during our preliminary experiments that the training dynamics change drastically between the pretrain → NAPT → finetune and pretrain → finetune."
},
{
"figure_ref": [],
"heading": "Discussion",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we would like to discuss a few important aspects of our approach apart from the main experiments."
},
{
"figure_ref": [],
"heading": "Ablation",
"publication_ref": [],
"table_ref": [
"tab_6",
"tab_8"
],
"text": "To better understand how different components of our NAPT strategy influence the final downstream performance, we conduct the following ablation studies. F1 scores of our proposed method (IT-MTL-NAPT) and 4 competitive baselines on the Aspect Sentiment Triplet Extraction task over 3 datasets under training on full dataset. We observe similar levels of performance.\nRegarding NAPT Tasks: We analyze the importance of NAPT with multiple tasks and their impact on the downstream performance. Our analysis shows that there exists a positive correlation between the NAPT complexity and downstream performance. We average the downstream performance across every task and every k-shot split and train on the downstream task in a multi-task learning fashion. We summarize our results in Table 4. Our experiments show that it helps in general to align the NAPT and finetuning objectives. If the NAPT phase is done in a multi-task learning fashion, it is beneficial for the model if the same is done for finetuning on the downstream task as well. Additionally, we observe that that harder NAPT tasks are beneficial for the downstream task regardless of the way the training on the downstream task is performed, as the F1 scores reflect the relative order in difficulty of the tasks (i.e., ASTE > AESC > AE). Regarding NAPT Regularization: We analyze the importance on the downstream performance of each regularization technique used during the NAPT phase. We report the performance in Table 6. We analyze the influence of: (i) Tuple Dropout, (ii) Biased weight decay, and (iii) Weight decay. We observe that our proposed approach is robust to hyperparameters, obtaining similar performance with various combinations of the 3 regularization techniques. We attribute this to the way the NAPT dataset is split into train and validation: enforcing disjoint sets of aspect-terms. This allows us to detect when the model starts to overfit.10 "
},
{
"figure_ref": [],
"heading": "NAPT",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [
"fig_3"
],
"heading": "Sentiment Prediction: Error Analysis",
"publication_ref": [],
"table_ref": [],
"text": "Quantitative: We first compare the percentage of correct predictions over each sentiment class, namely positive, negative, and neutral. We compare instruction tuning with and without our proposed NAPT step. We highlight the results in Figure 4. We observe that our proposed method performs better for every sentiment class. Moreover, we note that our proposed method outperforms the baseline even for the neutral sentiment class, a class which has not been seen during the NAPT phase. This suggests that NAPT can help the model learn faster even unseen tasks."
},
{
"figure_ref": [],
"heading": "Qualitative:",
"publication_ref": [],
"table_ref": [
"tab_7"
],
"text": "We present examples of the predictions made by an instruction tuned model with and without our proposed NAPT in Table 5. We show 4 predictions, 2 for ASTE (first two rows) and 2 for AESC (bottom two) on LAP14, in low-shot scenarios. We observe that the baseline has difficulties extracting the full aspect term (first row), while our proposed method is able extract the complete triple. The metric used does not reward partial matching.\nIn the second row, the baseline correctly generates the gold output, while our proposed method predicts a negative sentiment. In this case, the input can be considered ambiguous, as there is no explicit sentiment expressed in it. Also, for more complex tasks, such as aspect sentiment triplet extraction (AESC), the baseline has difficulties generating a valid prediction, while our proposed method is able to generate the correct prediction (third row). Lastly, we observe that although with NAPT we predict incorrectly (last row), it rather falls back to a term relevant to the domain (i.e., laptop)."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "In this paper, we proposed to add an intermediate step in the pretrain→finetune paradigm, called Noisy ABSA Pre-Training. We motivate this newly introduced step with the hypothesis that exposing the model to tasks more aligned with the downstream task will improve its performance, especially in low-data regimes such as in few-shot or complete zero-shot. We constructed a noisy dataset with a heuristic based pipeline approach consisting of three steps that utilize well-studied NLP resources and models. It serves as the training dataset for the noisy pre-training phase. We then evaluated with customer reviews from three datasets covering two domains, laptop and restaurant, and obtained large improvements in the zero/few-shot cases while achieving similar performance under finetuning on full dataset. We also discussed caveats around introducing catastrophic forgetting of general purpose pre-trained language models through such noisy pre-training, and introduced a few regularization techniques to help alleviate it."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [],
"table_ref": [],
"text": "We believe our proposed noisy pre-training step should apply to other structured prediction tasks, however we have not evaluated the approach on anything other than ABSA related tasks. Additionally, the noisy corpus construction process is heavily dependent on English based resources and pre-trained models. It might be non-trivial to extend the approach to other languages. Finally, we presented some extrinsic evaluation regarding the quality of the noisy corpus we create e.g., equivalence in terms of gold-annotated data size (Section 4.5.1). We leave any intrinsic evaluation of it by means of human supervision or otherwise for future work."
},
{
"figure_ref": [],
"heading": "A Implementation details",
"publication_ref": [
"b30",
"b11",
"b27"
],
"table_ref": [],
"text": "We use HuggingFace's implementation of transformers (Wolf et al., 2020;Lhoest et al., 2021). We use similar parameters as (Varia et al., 2022). We run our experiments on NVIDIA Tesla V100 GPUs."
},
{
"figure_ref": [],
"heading": "B All Experiments",
"publication_ref": [],
"table_ref": [],
"text": "For completeness, we include here all the models investigated over the 3 datasets, LAP14, REST15, and REST16, respectively."
},
{
"figure_ref": [],
"heading": "B.1 Full-Training",
"publication_ref": [],
"table_ref": [],
"text": "We report the results (test) on Full Training in Tables 7, 8, 9."
},
{
"figure_ref": [],
"heading": "B.2 K-Shot Learning",
"publication_ref": [],
"table_ref": [],
"text": "We report the results (test) on K-Shot Learning in Tables 10,11, 12."
},
{
"figure_ref": [],
"heading": "B.3 Cross Domain",
"publication_ref": [],
"table_ref": [],
"text": "We experiment with pre-training on a different domain than the domain of the downstream task. Concretely, we perform two experiments: (i) we perform NAPT on restaurant domain, then finetune on the laptop domain, and (ii) we perform NAPT on the laptop domain, then finetune on the restaurant domain. We include the results with our proposed model trained with NAPT on restaurant data and finetuned on LAP14 in Table 13. We include the results with our proposed model trained with NAPT on laptop data and finetuned on REST15 and REST16 in Table 14 andin Table 15, respectively. We observed that our proposed model is still able to transfer the knowledge learned during the NAPT phase. Our proposed model still outperforms the baseline, brining as much as 11.49% F1 points for the ASTE task in the laptop domain. In some cases we noticed a slight increase in the final performance compared to the model trained with NAPT on the full dataset. This suggests that the model trained on the full dataset overfits to the noisy data."
},
{
"figure_ref": [],
"heading": "C Multi-word Patterns",
"publication_ref": [],
"table_ref": [],
"text": "In Table 16 we list all the patterns that were used to filter 2-grams, 3-grams and 4-grams during aspect term extraction. "
}
] | We explore how weak supervision on abundant unlabeled data can be leveraged to improve few-shot performance in aspect-based sentiment analysis (ABSA) tasks. We propose a pipeline approach to construct a noisy ABSA dataset, and we use it to adapt a pre-trained sequence-to-sequence model to the ABSA tasks. We test the resulting model on three widely used ABSA datasets, before and after fine-tuning. Our proposed method preserves the full fine-tuning performance while showing significant improvements (15.84% absolute F1) in the few-shot learning scenario for the harder tasks. In zero-shot (i.e., without fine-tuning), our method outperforms the previous state of the art on the aspect extraction sentiment classification (AESC) task and is, additionally, capable of performing the harder aspect sentiment triplet extraction (ASTE) task. | A Weak Supervision Approach for Few-Shot Aspect Based Sentiment Analysis | [
{
"figure_caption": "Figure 1 :1Figure 1: Overview of our proposed Noisy ABSA Pre-Training (NAPT). We start from a pretrained language model and extend its capabilities by instruction tuning it in a multi-task learning fashion. We use 5 different yet related tasks for the proposed NAPT step. The tasks we use are: (i) aspect-term extraction, (ii) opinion-term extraction, (iii) aspect-term extraction and opinion-term extraction, (iv) aspect term extraction and sentiment classification, and (v) aspect-term extraction, opinion-term extraction, and sentiment classification. This step results in a model capable of performing multiple ABSA tasks.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure2: Performance Comparison between our proposed method (IT-MTL-NAPT) and two baselines over 3 datasets on on the Aspect Sentiment Triplet Extraction (ASTE), Aspect-term Extraction and Sentiment Classification (AESC) tasks in top, and bottom rows respectively. We note that our proposed method helps in all the k splits. (larger is better)",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: Data size equivalence comparison between t5 models that are finetuned on downstream corpus vs our proposed NAPT for ASTE task in (a) LAP14, (b) REST15 respectively. The finetuned models need ∼ 15 -25 completely annotated data points to equalize our proposed method.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: Comparison on the percentage of correct predictions over each sentiment class for an instruction tuned model with vs without the proposed NAPT on the LAP14 dataset and k = 10. With NAPT, it performs better on each sentiment class, even though neutral class does not appear in the noisy dataset (larger is better).",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Our proposed model outperforms the previous state-of-the-art results for AESC by as much as 6.94%F1 points in the restaurant domain. The improvement for the laptop domain is smaller, we attribute this to the NAPT dataset being biased towards the restaurant domain in terms of size. It is interesting to note that our model's backbone i.e., t5-base is able to outperform CORN altough it has almost half the number of parameters as that of its counterpart i.e., bart-large.",
"figure_data": "ModelRESTLAPCORN37.20 ±0.50 40.30 ±0.60IT-MTL-NAPT 44.14 ±0.30 40.51 ±0.43",
"figure_id": "tab_1",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Comparison of our proposed method with previous work on zero-shot Aspect Extraction Sentiment Classification (AESC). Our proposed method outperforms the previous work on both datasets. Metric is token-level F1 score.",
"figure_data": "",
"figure_id": "tab_2",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "tab_4",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Ablation study over NAPT tasks in terms of macro F1 scores averaged across all the tasks and 4 kshot settings. It shows that having all the tasks during NAPT achieves the best scores.",
"figure_data": "",
"figure_id": "tab_6",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Given the text: Finally, the biggest problem has been tech support., what are the aspect terms and their sentiments? <tech support, negative> <support, negative> <tech support, negative> ASTE: What are the aspect terms and their sentiments in the text: Of course, for a student, weight is always an issue.? <weight, neutral> <weight, neutral> <weight, negative> AESC: Given the text: the mouse buttons are hard to push., what are the aspect term, opinion term, and sentiment triplets? <mouse buttons, hard, negative> < , , > <mouse buttons, hard, negative> AESC: Given the text: The resolution is even higher then any other laptop on the market., what are the aspect term, opinion term and sentiment triplets? Predictions made by an instruction tuned model with and without NAPT in low-shot scenarios.",
"figure_data": "Task : InputGoldw/o NAPTw/ NAPTASTE: <resolution, higher, positive><resolution, higher, positive><laptop, higher, positive>Ablation Config.DatasetTuple DropoutWeight DecayBiased WeightLAP14 REST15 REST16✓✓✓47.4547.3251.65✓✓×47.5747.1051.39✓×✓47.6247.2651.65✓××47.3947.1751.37×✓✓47.5547.6551.80×✓×46.4347.4451.49××✓46.7847.1251.11×××46.9047.2751.49",
"figure_id": "tab_7",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "Ablation study over different regularization techniques in terms of macro F1 scores averaged across all tasks and 4 k-shot settings.",
"figure_data": "",
"figure_id": "tab_8",
"figure_label": "6",
"figure_type": "table"
}
] | Robert Vacareanu; Siddharth Varia; Kishaloy Halder; Shuai Wang; Giovanni Paolini; Neha Anna John; Miguel Ballesteros; Smaranda Muresan | [
{
"authors": "Shaowei Chen; Yu Wang; Jie Liu; Yuelin Wang",
"journal": "",
"ref_id": "b0",
"title": "Bidirectional machine reading comprehension for aspect sentiment triplet extraction",
"year": "2021"
},
{
"authors": "Ido Dagan; Oren Glickman; Bernardo Magnini",
"journal": "",
"ref_id": "b1",
"title": "The pascal recognising textual entailment challenge",
"year": "2005"
},
{
"authors": "Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova",
"journal": "Association for Computational Linguistics",
"ref_id": "b2",
"title": "BERT: pre-training of deep bidirectional transformers for language understanding",
"year": "2019-06-02"
},
{
"authors": "B Xiaowen Ding; Philip S Liu; Yu",
"journal": "",
"ref_id": "b3",
"title": "A holistic lexicon-based approach to opinion mining",
"year": "2008"
},
{
"authors": "Cícero Nogueira; Dos Santos; Maíra A De; C Gatti",
"journal": "",
"ref_id": "b4",
"title": "Deep convolutional neural networks for sentiment analysis of short texts",
"year": "2014"
},
{
"authors": "Ruidan He; Sun Wee; Hwee Tou Lee; Daniel Ng; Dahlmeier",
"journal": "",
"ref_id": "b5",
"title": "An unsupervised neural attention model for aspect extraction",
"year": "2017"
},
{
"authors": "Ruining He; Julian Mcauley",
"journal": "",
"ref_id": "b6",
"title": "Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering",
"year": "2016"
},
{
"authors": "Jeremy Howard; Sebastian Ruder",
"journal": "",
"ref_id": "b7",
"title": "Universal language model fine-tuning for text classification",
"year": "2018"
},
{
"authors": "Minqing Hu; Bing Liu",
"journal": "",
"ref_id": "b8",
"title": "Mining and summarizing customer reviews",
"year": "2004"
},
{
"authors": "Hiroshi Kanayama; Tetsuya Nasukawa",
"journal": "",
"ref_id": "b9",
"title": "Fully automatic lexicon expansion for domain-oriented sentiment analysis",
"year": "2006"
},
{
"authors": "James Kirkpatrick; Razvan Pascanu; Neil C Rabinowitz; Joel Veness; Guillaume Desjardins; Andrei A Rusu; Kieran Milan; John Quan; Tiago Ramalho; Agnieszka Grabska-Barwinska; Demis Hassabis; Claudia Clopath; Dharshan Kumaran; Raia Hadsell",
"journal": "Proceedings of the National Academy of Sciences",
"ref_id": "b10",
"title": "Overcoming catastrophic forgetting in neural networks",
"year": "2017"
},
{
"authors": "Quentin Lhoest; Albert Villanova Del Moral; Yacine Jernite; Abhishek Thakur; Suraj Patrick Von Platen; Julien Patil; Mariama Chaumond; Julien Drame; Lewis Plu; Joe Tunstall; Mario Davison; Gunjan Vsavsko; Bhavitvya Chhablani; Simon Malik; Brandeis; Le Teven; Victor Scao; Canwen Sanh; Nicolas Xu; Angelina Patry; Philipp Mcmillan-Major; Sylvain Schmid; Clement Gugger; ; Delangue; Stas Bekman; Pierric Cistac; Thibault Goehringer; Victor Mustar; François Lagunas; Alexander M Rush; Thomas Wolf",
"journal": "",
"ref_id": "b11",
"title": "Datasets: A community library for natural language processing",
"year": "2021"
},
{
"authors": "Ruifan Li; Hao Chen; Fangxiang Feng; Zhanyu Ma; Xiaojie Wang; Eduard H Hovy",
"journal": "",
"ref_id": "b12",
"title": "Dual graph convolutional networks for aspect-based sentiment analysis",
"year": "2021"
},
{
"authors": "Pengfei Liu; R Shafiq; Helen M Joty; Meng",
"journal": "",
"ref_id": "b13",
"title": "Fine-grained opinion mining with recurrent neural networks and word embeddings",
"year": "2015"
},
{
"authors": "Shu Liu; Kai-Wen Li; Zuhe Li",
"journal": "",
"ref_id": "b14",
"title": "A robustly optimized bmrc for aspect sentiment triplet extraction",
"year": "2022"
},
{
"authors": "Haiyun Peng; Lu Xu; Lidong Bing; Fei Huang; Wei Lu; Luo Si",
"journal": "",
"ref_id": "b15",
"title": "Knowing what, how and why: A near complete solution for aspect-based sentiment analysis",
"year": "2020"
},
{
"authors": "Maria Pontiki; Dimitris Galanis; Haris Papageorgiou; Ion Androutsopoulos; Suresh Manandhar; Al-Smadi Mohammad; Mahmoud Al-Ayyoub; Yanyan Zhao; Bing Qin; Orphée De Clercq; Véronique Hoste; Marianna Apidianaki; Xavier Tannier; Natalia Loukachevitch; Evgeniy Kotelnikov; Nuria Bel; Salud María Jiménez-Zafra; Gülşen Eryigit",
"journal": "Association for Computational Linguistics",
"ref_id": "b16",
"title": "SemEval-2016 task 5: Aspect based sentiment analysis",
"year": "2016"
},
{
"authors": "Maria Pontiki; Dimitris Galanis; Haris Papageorgiou; Suresh Manandhar; Ion Androutsopoulos",
"journal": "Association for Computational Linguistics",
"ref_id": "b17",
"title": "SemEval-2015 task 12: Aspect based sentiment analysis",
"year": "2015"
},
{
"authors": "Maria Pontiki; Dimitris Galanis; John Pavlopoulos; Harris Papageorgiou; Ion Androutsopoulos; Suresh Manandhar",
"journal": "",
"ref_id": "b18",
"title": "SemEval-2014 task 4: Aspect based sentiment analysis",
"year": "2014"
},
{
"authors": "Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu",
"journal": "Journal of Machine Learning Research",
"ref_id": "b19",
"title": "Exploring the limits of transfer learning with a unified text-totext transformer",
"year": "2020"
},
{
"authors": "Victor Sanh; Lysandre Debut; Julien Chaumond; Thomas Wolf",
"journal": "",
"ref_id": "b20",
"title": "Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter",
"year": "2019"
},
{
"authors": "Victor Sanh; Albert Webson; Colin Raffel; Stephen H Bach; Lintang A Sutawika; Zaid Alyafeai; Antoine Chaffin; Arnaud Stiegler; Teven Le Scao; Arun Raja; Manan Dey; M Saiful Bari; Canwen Xu; Urmish Thakker; Shanya Sharma; Eliza Szczechla; Taewoon Kim; Gunjan Chhablani; V Nihal; Debajyoti Nayak; Jonathan Datta; Mike Chang; Tian-Jian; Han Jiang; Matteo Wang; Sheng Manica; Zheng Xin Shen; Harshit Yong; Rachel Pandey; Thomas Bawden; Trishala Wang; Jos Neeraj; Abheesht Rozen; Andrea Sharma; Thibault Santilli; Jason Févry; Alan Fries; Ryan Teehan; Stella Rose Biderman; Leo Gao; Tali Bers; Thomas Wolf; Alexander M Rush",
"journal": "",
"ref_id": "b21",
"title": "Multitask prompted training enables zero-shot task generalization",
"year": "2022"
},
{
"authors": "Lei Shu; Jiahua Chen; Bing Liu; Hu Xu",
"journal": "",
"ref_id": "b22",
"title": "Zero-shot aspect-based sentiment analysis",
"year": "2022"
},
{
"authors": "Lei Shu; Hu Xu; Bing Liu",
"journal": "",
"ref_id": "b23",
"title": "Controlled cnn-based sequence labeling for aspect extraction",
"year": "2019"
},
{
"authors": "Richard Socher; Alex Perelygin; Jean Wu; Jason Chuang; Christopher D Manning; A Ng; Christopher Potts",
"journal": "",
"ref_id": "b24",
"title": "Recursive deep models for semantic compositionality over a sentiment treebank",
"year": "2013"
},
{
"authors": "Kaitao Song; Xu Tan; Tao Qin; Jianfeng Lu; Tie-Yan Liu",
"journal": "",
"ref_id": "b25",
"title": "Mpnet: Masked and permuted pre-training for language understanding",
"year": "2020"
},
{
"authors": "Stéphan Tulkens; Andreas Van Cranenburgh",
"journal": "",
"ref_id": "b26",
"title": "Embarrassingly simple unsupervised aspect extraction",
"year": "2020"
},
{
"authors": "Siddharth Varia; Shuai Wang; Kishaloy Halder; Robert Vacareanu; Miguel Ballesteros; Yassine Benajiba; Anna Neha; Rishita John; Smaranda Anubhai; Dan Muresan; Roth",
"journal": "",
"ref_id": "b27",
"title": "Instruction tuning for fewshot aspect-based sentiment analysis",
"year": "2022"
},
{
"authors": "Duy-Tin Vo; Yue Zhang",
"journal": "",
"ref_id": "b28",
"title": "Target-dependent twitter sentiment classification with rich automatic features",
"year": "2015"
},
{
"authors": "Bo Wang; Tao Shen; Guodong Long; Tianyi Zhou; Yi Chang",
"journal": "",
"ref_id": "b29",
"title": "Eliminating sentiment bias for aspect-level sentiment classification with unsupervised opinion extraction",
"year": "2021"
},
{
"authors": "Thomas Wolf; Lysandre Debut; Victor Sanh; Julien Chaumond; Clement Delangue; Anthony Moi; Pierric Cistac; Tim Rault; Rémi Louf; Morgan Funtowicz; Jamie Brew",
"journal": "",
"ref_id": "b30",
"title": "Transformers: State-ofthe-art natural language processing",
"year": "2020"
},
{
"authors": "Hu Xu; Bing Liu; Lei Shu; Philip S Yu",
"journal": "",
"ref_id": "b31",
"title": "Bert post-training for review reading comprehension and aspect-based sentiment analysis",
"year": "2019"
},
{
"authors": "Liheng Xu; Kang Liu; Siwei Lai; Yubo Chen; Jun Zhao",
"journal": "Association for Computational Linguistics",
"ref_id": "b32",
"title": "Mining opinion words and opinion targets in a two-stage framework",
"year": "2013"
},
{
"authors": "Lu Xu; Hao Li; Wei Lu; Lidong Bing",
"journal": "",
"ref_id": "b33",
"title": "Position-aware tagging for aspect sentiment triplet extraction",
"year": "2020"
},
{
"authors": "Hang Yan; Junqi Dai; Tuo Ji; Xipeng Qiu; Zheng Zhang",
"journal": "",
"ref_id": "b34",
"title": "A unified generative framework for aspect-based sentiment analysis",
"year": "2021"
},
{
"authors": "Guoxin Yu; Jiwei Li; Ling Luo; Yuxian Meng; Xiang Ao; Qing He",
"journal": "",
"ref_id": "b35",
"title": "Self question-answering: Aspect-based sentiment analysis by role flipped machine reading comprehension",
"year": "2021"
},
{
"authors": "Meishan Zhang; Yue Zhang; Duy-Tin Vo",
"journal": "",
"ref_id": "b36",
"title": "Neural networks for open domain targeted sentiment",
"year": "2015"
},
{
"authors": "Wenxuan Zhang; Yang Deng; Xin Li; Yifei Yuan; Lidong Bing; Wai Lam",
"journal": "",
"ref_id": "b37",
"title": "Aspect sentiment quad prediction as paraphrase generation",
"year": "2021"
},
{
"authors": "Wenxuan Zhang; Xin Li; Yang Deng; Lidong Bing; Wai Lam",
"journal": "",
"ref_id": "b38",
"title": "Towards generative aspect-based sentiment analysis",
"year": "2021"
},
{
"authors": "Wenxuan Zhang; Xin Li; Yang Deng; Lidong Bing; Wai Lam",
"journal": "",
"ref_id": "b39",
"title": "A survey on aspect-based sentiment analysis: Tasks, methods, and challenges",
"year": "2022"
},
{
"authors": "Xiang Zhang; Junbo Zhao; Yann Lecun",
"journal": "Advances in neural information processing systems",
"ref_id": "b40",
"title": "Character-level convolutional networks for text classification",
"year": "2015"
}
] | [
{
"formula_coordinates": [
4,
322.46,
151.8,
197.71,
13.51
],
"formula_id": "formula_0",
"formula_text": "L = CE loss + α ⋅ 2 (θ -θ init ) + β ⋅ 2 (θ). (1"
},
{
"formula_coordinates": [
4,
520.17,
154.86,
4.24,
9.46
],
"formula_id": "formula_1",
"formula_text": ")"
}
] |
|
10.3390/s19194342 | 2023-05-19 | [
{
"figure_ref": [],
"heading": "",
"publication_ref": [],
"table_ref": [],
"text": "We believe that these contributions enable future research and easier use and applications of coresets. "
},
{
"figure_ref": [],
"heading": "(b) SVMs",
"publication_ref": [],
"table_ref": [],
"text": "Figure 1. Evaluation of AutoCoreset against other problem dependant coreset construction algorithms for SVM and Logistic regression (on the Dataset (i)). AutoCoreset achieves a much smaller approximation error and a higher test accuracy for the same coreset size while being an automatic and problem-independent framework. Sensitivity-based coreset for 1-mean, Median of meansbased coreset, and Caratheordory coreset are all variants of Auto-Coreset."
},
{
"figure_ref": [],
"heading": "Introduction and Motivation",
"publication_ref": [
"b3",
"b4",
"b5",
"b10",
"b19",
"b14",
"b20",
"b12",
"b36",
"b18",
"b41",
"b13",
"b21",
"b41",
"b24",
"b42"
],
"table_ref": [
"tab_1"
],
"text": "In many machine learning (ML) problems, the input is usually a set P = {p 1 , • • • , p n } of n items, a (probably infinite) set of candidate solutions X called query set, and a loss function f : P × X → [0, ∞]). The goal is to find a query (model, classifier) x * that minimizes the sum n i=1 f (p i , x)\nFigure 2. A flowchart illustrating our automatic coreset construction framework. Note that VSAlg() can be any algorithm from Table 1.\nover every query x ∈ X . Notably, many of these optimization/learning tasks are typically challenging to approximate when the input is very large. Furthermore, in the era of big data, we usually aim towards maintaining a solution for streaming and/or distributed input data, while consuming small memory. Finally, even well-known problems with a close optimal solution, such as ridge regression and other classes of convex optimization involving Cross-validation methods or hyper-parameter tuning methods, must analyze under many restrictions several queries for various subsets of data, leading to a drastic increase in the running time.\nCoresets. A common approach to solve such issues is to use data summarization techniques, namely Coresets, which got increasing attention over recent years (Bachem et al., 2018a;b;Bȃdoiu & Clarkson, 2008;Maalouf et al., 2022a;Balcan et al., 2013;Braverman et al., 2019;Tukan et al., 2023b;Curtain et al., 2019;Jubran et al., 2020;Feldman et al., 2014;Karnin & Liberty, 2019;Tukan et al., 2022b;Maalouf et al., 2021b;Tukan et al., 2022c;a;2023a); see surveys in (Feldman, 2020;Munteanu & Schwiegelshohn, 2018;Phillips, 2016), and introductions in (Maalouf et al., 2021a;Jubran et al., 2019). A coreset, informally, is a tiny weighted subset of the input set P , roughly approximating the loss of P for every possible query x ∈ X , up to a bound of 1 ± ε factor (0 ≤ ε < 1). The size of the coreset is often independent or close to logarithmic in the amount of the input points n, but polynomial in 1/ε. Coresets are useful in ML as they significantly increase the efficiency of ML solvers. Specifically, employing conventional methods on the constructed coresets should approximate the optimal solution on the entire dataset, in orders of magnitude less expensive time and memory. Furthermore, by repeatedly running existing heuristics on the coreset in the time it takes to run them once on the original (large) dataset, heuristics that are already quick can be more accurate. Additionally, coresets can be maintained for distributed and streaming data.\nSo what's the problem? Obtaining non-trivial theoretical guarantees is frequently impossible in many contemporary machine learning problems due to either the target model being highly complex or since every input element p ∈ P is significant in the sense of high sensitivity; see (Tukan et al., 2020). Hence, generating a coreset becomes a highly challenging process, and the corresponding theoretical analysis occasionally falls short of recommending such approximations. As a result, designing a new coreset and demonstrating its accuracy for a new ML problem might take years, even for simple ones.\nAnother crucial issue with current theoretical frameworks is their lack of universality. Even the most general frameworks (e.g., (Feldman & Langberg, 2011;Langberg & Schulman, 2010) replace the problem of generating a coreset for an input set P of n points with n new optimization problems (one problem for each of the n input points p ∈ P ) known as sensitivity bounding. Solving these may be more difficult than solving the original problem, where for every p ∈ P we are required to bound its own sensitivity defined as\ns(p) = sup x∈X f (p,x)\nq∈P f (q,x) . As a result, distinct approximation strategies are often adapted to each task. Hence, the main disadvantage of such frameworks is that researchers provide papers solely for bounding the sensitivities with respect to a certain problem or a family of functions (Tukan et al., 2020;Maalouf et al., 2020), limiting the spread of coresets, as non-expert won't be able to suggest coresets for their desired task. These problems raise the following questions:\nIs it possible to design an automatic and practically robust coreset construction framework (for any desired cost function and input dataset) that does not need sensitivity calculation or any other problem-dependent com-putation by the user? Can we provide some provable guarantees with respect to this framework? 1.1. Vision Goal. Our goal is to provide a single algorithm that only receives the loss function we wish to compute a coreset for and the input dataset, then, it practically outputs a good coreset for the input dataset with respect to the given loss. This algorithm should be generic, efficient, and work practically well for many problems.\nmotivation. The main motivation behind this goal is (1) to increase the spread and use of coreset to a larger community that is not limited to coreset researchers or pioneers. (2) Additionally, to ease the use of coresets for many other applications that may be out of the scope of the coreset literature, and finally, to (3) easily apply coresets for new problems that do not have provable coresets. Theoretically speaking, it is indeed very hard to provide a \"theoretical strong coreset\" to any problem -for example, there exist lower bounds on the coreset sizes for different problems (Munteanu et al., 2018;Tukan et al., 2021). Thus we aimed at a practical result while providing weaker theoretical guarantees, with an extensive experimental study."
},
{
"figure_ref": [],
"heading": "Our contribution",
"publication_ref": [
"b35"
],
"table_ref": [],
"text": "In this paper, we provide a coreset construction mechanism that answers both questions. Specifically:\n(i) The first automatic practical coreset construction system that only needs to receive the loss function associated with the problem. Our coreset does not require any computation to be done by the user, not mathematical nor technical (without the need for sensitivities or any other task-related computation by the user). To the best of our knowledge, this is the first paper to suggest a plug-and-play style framework/compiler for coreset construction. We also provide a theoretical justification for using our framework.\n(ii) An extensive empirical study on real-world datasets for various ML solvers of Scikit-Learn (Pedregosa et al., 2011), including k-means, logistic regression, linear regression, and support vector machines (SVM), showing the effectiveness of our proposed system.\n(iii) AutoCoreset: An open-source code implementation of our algorithm for reproducing our results and future research. For simplicity and ease of access, to obtain a coreset, the user only needs to plug in his desired loss function and the data into our system. We believe this system will popularize and expose the use of coresets to other scientific fields. "
},
{
"figure_ref": [],
"heading": "Setup Details",
"publication_ref": [],
"table_ref": [],
"text": "Given a set\nP = {p 1 • • • , p n } ⊆ R d of n points 1 and a loss function f : P × X → [0, ∞)\nwhere X is a (possibly infinite) set of queries. In this paper, we develop an automatic coreset construction framework for any problem involving cost functions of the form p∈P f (p, x), here x ∈ X . Formally, we wish to find a small subset I ⊆ [n] and a weight function v :\nI → [0, ∞) such that max x∈X j∈I v(j)f (pj ,x) n i=1 f (pi,x)\n∈ 1 + O(ε), for some small ε ≥ 0."
},
{
"figure_ref": [],
"heading": "Preliminaries",
"publication_ref": [],
"table_ref": [
"tab_1"
],
"text": "We now give our notations and used Definition.\nNotations. For a pair of integers n, m > 0, we denote by [n] the set {1, • • • , n}, and by R n×m the set of every possible n × m real matrix. For a matrix M ∈ R n×m and a pair of integers i ∈ [n], j ∈ [m], we use M i, * to denote its ith row (vector), M * ,j to denote its jth column, and M i,j to denote the entry in the ith row and jth columns.\nIn what follows, we define a crucial component on which our system relies, namely, vector summarization coreset.\nDefinition 2.1 (Vector summarization coreset). Let M ∈ R n×m , I ⊆ [n], v : I → [0, ∞) be a weight function, and let ε > 0. The tuple (I, v) is an vector summarization\nε-coreset for M if i∈[n] M i, * -j∈I v(j)M j, * 2 2 ≤ ε. 1 if P is a set of labeled items, then P = pi = (p i , yi) p i ∈ R d-1 , yi ∈ R n i=1\nMany papers suggested different algorithms for computing such coresets; in Table 1 summarizes some of these results, as we can use them all of them in our method."
},
{
"figure_ref": [
"fig_0"
],
"heading": "AutoCoreset",
"publication_ref": [],
"table_ref": [],
"text": "A coreset aims to approximate the probability distribution induced upon the input data by the cost function. Hence, in order to approximate a given cost function, the coreset must contain points that can result in an approximated distribution to that of the full data.\nKey idea. Loosely speaking, assume that for a given cost function f and a set P = {p 1 , • • • , p n } ⊂ R d , we access an infinite matrix M * (P, f ) where the rows correspond to the n points of P , and each column corresponds to a query point from the infinite set of queries X . Specifically, each row i ∈ [n] is of infinite length representing the loss of each point p i ∈ P with respect to the infinite set of queries X . A coreset in this context means finding a subset of the rows I ⊆ [n], and a weight function v : I → [0, ∞], that satisfies the vector summarization coreset guarantee (see Definition 2.1), i.e.,\ni∈[n] M * (P, f ) i, * - j∈I v(j)M * (P, f ) j, * 2 2 ≤ ε. (1)\nFrom such a coreset I ⊆ [n], the cost function can be approximated, since for every query x ∈ X (column in the matrix M * (P, f )), the weighted sum of losses over the coreset I approximate the original sum of losses of the whole data. While such a concept is admirable, having an access to such immense data is rather imaginative.\nRecently (Maalouf et al., 2022b) showed that for an input set of points P , and query space X that is defined as a family of sine wave functions, a coreset can be constructed. Specifically, it was shown that if the coreset approximates the loss of every query in a smaller set of queries on the input data, then it will also approximate the losses of the whole set of queries (sine waves). Thus, indeed, the sine wave that fits best the coreset approximates the sine wave that best fits the entire data. Inspired by such a result, we aim towards constructing a sub-matrix 3 for illustration) such that constructing the coreset on M(P, f ) (a weighted subset of the rows of M(P, f )) will also yield a similar coreset to that of (1) on the M * (P, f ). But, how to build the sub-matrix M(P, f )? how to choose the query set corresponding to the columns of M(P, f )?\nM(P, f ) of M * (P, f ) ( M(P, f ) contain a subset of the columns of M * (P, f ); see Fig- ure"
},
{
"figure_ref": [],
"heading": "A deeper look into AutoCoreset",
"publication_ref": [
"b7",
"b15",
"b30",
"b13"
],
"table_ref": [],
"text": "We now give and explain our algorithm AUTOCORESET (see Algorithm 1), which aims to provide a parasitical core- (Carathéodory, 1907) \n0 0 m + 1 O(min{nm + log 4 (m), m 2 n 2 , nm 3 })\nFrank-Wolfe (Feldman et al., 2017) (Clarkson, 2010)\n0 ε O(1/ε) O(min{nd/ε})\nMedian of means tournament (Minsker, 2015) \nδ ε O(1/ε) O(m log 2 (1/δ) + m log(1/δ)/ε)\nSensitivity sampling (Feldman & Langberg, 2011)\nδ ε O( 1 ε (m + log 1 δ )) O(nm) Uniform sampling δ ε O( 1 εδ ) O(1)\nAlgorithm 1 AUTOCORESET (P, f, τ, m, ζ) input set of n points P , a loss function f , a coreset size τ , number of initial models m, and an stopping criterion\nζ output A coreset (I, v) such that 1: M(P, f ) ← → 0 n×m 2: for each i ∈ [m] do 3:\nx i ← a randomized approximated solution involving P and f 4:\nfor every j ∈ [n] do 5: M(P, f ) j,i ← f (p j , x i ) 6:\nend for 7: end for 8: repeat 9:\n(I, v) ← coreset of m indices for vector summarization problem involving M(P, f ) {See Definition 2.1} 10:\nx * ← arg min x∈X i∈I v (i) f (p i , x) 11: M(P, f ) ← M(P, f ) → 0 n 12:\nfor every i ∈ [n] do 13:\nM(P, f ) i,m+1 ← f (p i , x C ) 14: end for 15: m ← m + 1 16: until ζ is satisfied return (I, v)\nset with similar guarantees.\nInto the forging of our coresets. Let m > 1 be an integer. First, a matrix M(P, f ) is generated to contain n × m zero entries, followed by generating a set X = {x 1 , . . . , x m } of m approximated solutions with respect to min\nx∈X n i=1 f (p i , x)\nas depicted at Lines 1-7. If no such approximated solution exists, then the initialization may be also completely random. The (sub)matrix M(P, f ) is now initialized, where for every i ∈ [n], and j ∈ [m], the entry M(P, f ) i,j in the ith row and jth column is equal to f (p i , x j ). While the properties associated with generated solutions at Line 3 hold with some probability, our framework is always guaranteed practically to generate a good coreset. This is due to the fact that these solutions are merely used as an initialization mechanism.\nFrom this point on, a loop is invoked. First, using the current state of M(P, f ), a vector summarization coreset (I, v) (see Definition 2.1) is generated with respect to the rows of M(P, f ).\nA coherent claim of our system is that any vector summarization coreset I for the rows of M(P, f ), is directly mapped to coreset for P (using the same set of indexes and the same weight function) with respect to the query set X ⊂ X and the function f , where X is the set of all queries that brought about the columns of M(P, f ). More preciously,\nmax x∈X j∈I v(j)f (pj ,x) n i=1 f (pi,x) ∈ 1 + O(ε); see Lemma 3.1.\nSince the computed vector summarization coreset I is also a coreset with respect to f, P , and X , we can optimize f over the small coreset I to obtain a new query x * ∈ X that gives an approximated solution to the full data (see Line 10). We then apply the loss f function and the new solution\nx * on p 1 , • • • , p n to obtain the vector of losses l = (f (p 1 , x * ), • • • , f (p n , x * ))\nT , and concatenate such vector of loss values to M(P, f ) as its last column. This aids in expanding the exposure of generated coreset to a wider spectrum of queries, leading towards a strong coreset. Observe that in the next iteration, when we compute a new coreset for the given set of queries, the coreset will approximate all of the previous ones (set of queries) and the new computed query/solution x * . This procedure is repeated until some stopping criterion ζ is invoked -we provide more details on the used ζ in Section 5. We refer the reader to Lines 10-15. Note that if we were able to run the above procedure infinitely while ensuring that at each iteration a new solution is computed, M * (P, f ) would have been generated, resulting in the \"strong coreset\" this system is leaning towards. To better grasp the idea of the framework, we provide a flowchart illustration at Figure 2.\nThe parameters τ, m, ζ. Our Algorithm initializes its matrix M(P, f ) with respect to the losses of m > 1 different queries, and outputs a coreset of size τ > 1, hence, the larger the m and τ the better the approximation, but the slower the time; See section 5 for more details. Regarding ζ, it is the used stopping criterion, we provide full details regarding the used ζ in Section 5."
},
{
"figure_ref": [],
"heading": "Weaker coresets are fine too",
"publication_ref": [],
"table_ref": [
"tab_1",
"tab_1"
],
"text": "Our AutoCoreset system, while ambitiously aims towards holding a grasp over M * (P, f ), it finds a weaker version of the \"strong coresets\". Specifically speaking, it finds a coreset that attains approximation guarantees with respect to a subset of the query set X . Theoretically speaking, the following lemma summarizes one aspect of the theoretical properties guaranteed by AutoCoreset.\nLemma 3.1 (Vector summarization coreset → \"a weak core-set for any loss\"). Let P = {p 1 , • • • , p n } ⊆ R d be a set of n points as defined in Section 3.1, X ⊂ X be a set of queries, f : P × X → [0, ∞) be a loss function, and let M(P, f ) ∈ R n×|X | be the loss matrix defined with respect to P, f, X as in Algorithm 1. Let τ ≥ 1 be an integer, and let (I, v) be a ε-vector summarization coreset concerning M(P, f ) of size |I| = τ . Then, for every x ∈ X ,\ni∈[n] f (p i , x) - j∈I v(j)f (p j , x) 2 ≤ ε.\nImplications of Lemma 3.1. AutoCoreset guarantees theoretically that for a finite set of queries X , a coreset can be constructed supporting X . A key advantage here would be the ability to represent any query x such that its loss vector\n(f (p 1 , x), • • • , f (p n , x)\n) lies inside the \"convex hull\" of the loss vectors of the query set X . Luckily, such a trait is supported by our system. Specifically speaking, for any query such that its corresponding loss vector with respect to f and P can be formulated as a convex combination of the columns of M(P, f ), then a vector summarization coreset for the rows of M(P, f ) is also a vector summarization to the rows of concatenating M(P, f ) and the column vector\n. In what follows, we give the theoretical justification for the above claim. x ∈ X satisfying that for every i ∈\n[n], f (p i , x) = z k=1 α (k) M(P, f ) i,k , we have n i=1 f (p i , x) - j∈I v(j)f (p j , x) 2 ≤ ε,\nwhere ε ≥ 0 is the approximation factor associated with generating a vector summarization coreset of m points.\nThe best of both worlds. Claim 3.2 states that even if it seems that our generated coreset only supports a handful of queries from X , our coreset basically supports many more queries. The highlight of such a claim is that if the optimal solution for the objective function involves f and P , then our coreset becomes stronger in the sense of ensuring better quality even during the training/optimization process which involves both f and P . Such a claim is usually targeted via \"Strong coresets\" and mainly by \"Weak coresets\". Au-toCoreset ensures a coreset that resides on the spectrum involving these coresets at its ends, i.e., generating a coreset from the best of both worlds -a coreset supporting the optimal solution that the user is aiming to solve using accelerated training via coresets while maintaining the provable approximation guarantees of strong coresets to some extent.\n4. Size, Space, and Time Analysis Time complexity. Let VAlg be the vector summarization algorithm used at Line 9 of Algorithm 1 (pick one from Table 1). Let ε, δ ∈ (0, 1) be the desired vector summarization approximation error, and probability of failure, respectively. Now denote by • T (n, i, ε, δ): the running time of VAlg on a matrix of n rows and i columns with respect to ε and δ.\n• S(n, i, ε, δ): the size of the coreset computed by VAlg on a matrix of n rows and i columns with respect to ε and δ.\n• T sol (n, d): the time required to compute a solution vector x * for n points in the d dimensional space with respect to the problem at hand (e.g., the time required to compute the solution of linear regression is O(nd 2 )).\n• T cost (n, d): the time required to calculate the cost for n points in the d dimensional space on a single query with respect to the problem at hand (e.g., the time required to compute the cost of linear regression for n points in the d dimensional space given a solution vector x is O(nd). \"t\": be the number of iterations of the algorithm.\nAt each iteration \"i\", Algorithm 1 1. applies VAlg on a matrix of n rows and i columns to obtain a coreset of size S(n, i, ε, δ). This step requires T (n, i, ε, δ) time.\n2. Solves the problem on the coreset to obtain a new solution x * . Requiring T sol (S(n, i, ε, δ), d) time.\n3. Calculates the cost of the n points with respect to x * . Requiring T cost (n, d) time\nThus, for a single step i the running time is T (n, i, ε, δ) + T sol (S(n, i, ε, δ), d) + T cost (n, d). Summing for t iterations:\nt i=1 (T (n, i, ε, δ) + T sol (S(n, i, ε, δ), d)) + tT cost (n, d).\nFor example, in Linear regression and using the Sensitivity sampling as VAlg, an immediate bound for the running time is O(t(nt\n+ (t/ε + log(1/δ)/ε)d 2 + nd)).\nSpace complexity. First, note that the input data and the matrix of losses take O(n(d + t)) where t here denotes the number of iterations our coreset generation has taken.\nRecall the definitions of VAlg, ε, δ and S(n, i, ε, δ). We now denote by\n• M em(VAlg, ε, δ, i) the amount of space needed by Valg to generate an ε-coreset with a success probability of at least 1 -δ.\n• M em sol (n, d) the space required to compute a solution vector x * for n points in the d dimensional space with respect to the problem at hand (e.g., the space required to compute the solution of SVM is O(n 2 + d).\nThe total space complexity is thus bounded by O(n(d\n+ t) + max i∈[t] (M em(VAlg, ε, δ, i) + M em sol (S(n, i, ε, δ)), d).\nFor example for SVM and using the Sensitivity sampling vector summarization, an immediate bound for the space complexity is O(n(d\n+ t) + (1/ε(t + log(1/δ))) 2 ).\nCoreset size. The size of the constructed coreset is equal to the used vector summarization coreset size (See Table 1), and it depends on the approximation error ε, the probability of failure δ we wish to have, and the final number of approximated queries -columns of the query matrix.\nIn short -let ε be the desired approximation error and let δ be the probability of failure. Let t be the number of iterations required Algorithm 1. Denote by S(n, i, ε, δ) the size of the set computed by the used vector summarization algorithm on a matrix of n rows and i columns with respect to ε and δ (see Table 1 for examples). Then, the size of the coreset is S(n, t, ε, δ).\nFor example, using the Sensitivity sampling method (as the used vector summarization coreset), to approximate the currently given t queries after t iterations, with ε approximation error, and δ probability of failure, we get a coreset of size O(t/ε + log(1/δ)/ε).\nFrom additive to multiplicative approximation error.\nAlgorithm 1 can immediately be modified to compute a coreset that yields a multiplicative approximation as follows.\nGiven the set P , the current set of queries X , and the loss f , define a new function g(p, x) :=\nf (p,x) √ p∈P f (p,x)\nfor every pair of a query x ∈ X and input data p ∈ P . Now build the corresponding matrix M(P, g) (as done in Algorithm 1 for f (p, x)) instead of M(P, f ), and run the exact same vector summarization coreset algorithm on it. Then, by Lemma 3.1, for every x ∈ X , i∈[n] g (p i , x) -j∈I v(j)g (p j , x) 2 ≤ ε, and by the definition of g we get that the result is a multiplicative coreset for the given set of queries as for every x ∈ X\ni∈[n] f (p i , x) - j∈I v(j)f (p j , x) 2 ≤ ε p∈P f (p, x)."
},
{
"figure_ref": [],
"heading": "Experimental Study",
"publication_ref": [
"b41",
"b41",
"b42",
"b24",
"b6"
],
"table_ref": [],
"text": "In what follows, we first discuss the choices of different vector summarization coresets, and the used parameters in our experiments, followed by evaluating our coreset on real-world datasets, against other famous competing methods: Near Convex Coreset (Tukan et al., 2020), Lewis weights (Munteanu et al., 2018) and leverage scores (Munteanu et al., 2018) for logistic regression, Near Convex Coreset (Tukan et al., 2020) and optimization based coreset (Tukan et al., 2021) for support vector machines (SVM), SVD-based coreset (Maalouf et al., 2020) for linear regression, Bi criteria coreset (Braverman et al., 2021) for k-means, and uniform sampling in all of the experiments. We note that each experiment was conducted for 16 trials, we report both the mean and std for all of the presented metrics."
},
{
"figure_ref": [],
"heading": "Software/Hardware.",
"publication_ref": [
"b33",
"b51",
"b35"
],
"table_ref": [],
"text": "Our algorithms were implemented in Python 3.9 (Van Rossum & Drake, 2009) using \"Numpy\" (Oliphant, 2006), \"Scipy\" (Virtanen et al., 2020) and \"Scikit-learn\" (Pedregosa et al., 2011). Tests were performed on 2.59GHz i7-6500U (2 cores total) machine with 16GB RAM."
},
{
"figure_ref": [],
"heading": "AutoCoreset parameters",
"publication_ref": [
"b37",
"b11",
"b8",
"b52",
"b49",
"b22",
"b17",
"b39",
"b40",
"b34"
],
"table_ref": [
"tab_1"
],
"text": "Vector summarization coresets. There are many methods for computing such coresets, some of them are deterministic, i.e., with no probability of failure, and others work with some probability 1-δ. On the other hand, some are accurate, i.e., ε = 0, and others yield an approximation error ε > 0.\nIn Table 1 we summarize some of the common methods for computing such coresets, and their properties, such as size, running time, approximation error, and probability of failure. In our system, we implemented all of the given methods and compared them via extensive experiments.\nSetting the number of initial solutions m. Throughout our experiments, we have set the number of initial solutions to 10. The idea behind this is to expose AUTOCORESET to a number of solutions that is not too high nor too low. Hence, we ensure that the coreset is not too weak nor too dependent on the initial solutions.\nChoosing a stopping criterion ζ. Inspired by the earlystopping mechanism of (Prechelt, 1998), we adopt a similar idea. We make use of a parameter, namely \"patience\", which was set to 7, to attempt an indication of the occurrence of saturation with respect to the exposure of our coreset paradigm to new queries; see more details at Section A. To correctly use this parameter, we use additional two parameters, one of which is a counter, while the other holds the optimal coreset that resulted in the smallest sum of the entries of the concatenated columns (see Line 13 at Algorithm 1). The counter will be reset to 0 once a new column is added such that its sum is lower than the smallest sum so far, and the optimal coreset will be updated. Otherwise, the counter will be increased. AUTOCORESET will keep running until the above counter reaches the \"patience\" parameter. In our experiments, we returned the optimal coreset since it led to better results. For completeness, we refer the reader to the appendix where we conduct an ablation study and check our results without taking the optimal coreset, i.e., in those results, we take the last coreset. Note that, in both sets of experiments, we outperform the competing methods.\nDatasets. The following datasets were used throughout our experimentation. These datasets were taken from (Dua & Graff, 2017) and (Chang & Lin, 2011): (i) Credit card dataset (Yeh & Lien, 2009) composed of 30000 points with 24 features representing customers' default payments in Taiwan, (ii) Cod-RNA dataset (Uzilov et al., 2006): dataset containing 59535 points with 8 features, (iii) HTRU dataset (Lyon et al., 2016): Pulsar candidates collected during the HTRU survey containing 17898 each with 9 features, (iv) 3D Road Network (Guo et al., 2012): 3D road network with highly accurate elevation information from Denmark containing 434874 points each with 4 features, (v) Accelerometer dataset (Sampaio et al., 2019): an accelerometer data from vibrations of a cooler fan with weights on its blades containing 153000 points consisting each of 5 features, and (vi) Energy efficiency Data Set (Tsanas & Xifara, 2012): a dataset containing 768 points each of 8 features.\nML models. Throughout our entire set of experiments, we have relied on \"Scikit-Learn\" ML models.\nReported results. First, for each coreset (I, v) of an input data P and a loss function f , we compute the optimal solution on the coreset x * I ∈ arg min X∈X i∈I v(i)f (p i , x), and on the real data x * P ∈ arg min x∈X i∈[n] f (p i , x), and we report the optimal solution approximation error\nε = i∈[n] f (p i , x * I ) -i∈[n] f (p i , x * P )\n. Secondly, we show for classification problems the test accuracy obtained when training on the coreset, while on regression problems we show an estimate of the coefficient of determination of the prediction R 2 (Ozer, 1985). Additional measures are reported for some problems; we discuss them in the following sections. The bars in our graphs reflect the standard deviation. "
},
{
"figure_ref": [],
"heading": "Traditional ML classification problems",
"publication_ref": [
"b41"
],
"table_ref": [],
"text": "In what follows, we show our results when setting f to be the loss function of either the Logistic regression problem or the SVMs problem. In both experiments, since, some of the datasets were unbalanced, each sample coreset size has been split -small classes get a slightly larger portion of the sample size than simply taking η× sample size where η represents the class size percentage with respect to the total number of points, while larger classes get a portion of the sample size smaller than η× sample size.\nLogistic regression. We have set the maximal number of iterations to 1000 (for the Scikit-Learn solver) while setting the regularization parameter to 1. Our system's approximation error was smaller by orders of magnitude, and the accuracy associated with the models trained using our coreset was better than the model trained on the competing methods; see Figure 1(a) and Figure 5(a). On the other hand, Figure 4(a) depicts a multiplicative gap of 30 with respect to the approximation error in comparison to the competing methods while simultaneously acceding by 5% accuracy gap over them. In addition, we present the confusion matrix for each of our coresets using AutoCoreset, and compare it to the confusion matrices with respect to the entire data and the uniform sampling coreset; See Figure 7. The confusion matrices aim towards explaining our advantage as our system outputs coresets that approximately maintain the structural properties of the confusion matrix of the entire data better than simply using uniform sampling, as our recall and accuracy are closer to their corresponding values when using the entire dataset.\nSVMs. As for SVMs, we mainly focused on the linear kernel, while setting the regularization parameter to 1. Similarly to logistic regression, we outperform the competing methods both in accuracy and approximation error; see Discussion These results show that general frameworks that aim to handle a large family of functions without embedding some crucial information concerning the properties of the problem, usually tend to lose through the race towards smaller coresets sizes with small approximation errors. We thus show that while AutoCoreset is general in the reach of its applications, it also embeds the functional properties of the problem into higher consideration than that of (Tukan et al., 2020), and practically achieves robust results (smaller std)."
},
{
"figure_ref": [],
"heading": "Linear regression and k-means clustering.",
"publication_ref": [],
"table_ref": [],
"text": "In our experiments for linear regression, we observe a clear gap between each of our vector summarization coresets and the competing methods, leading towards outperforming the competing coreset for the task of fitting linear regression. In addition, we observe that the determination coefficient R 2 for our method is much closer to the determination coefficient R 2 when using the entire data. This indicates that our coresets lead to better learning and correlation between the input data and the corresponding outputs of the regression problem; see Figure 8. In addition, for k-means, our coresets outperform the competitors (see Figure 9), justifying their robustness across a wide range of applications."
},
{
"figure_ref": [],
"heading": "Conclusions and Future Work",
"publication_ref": [],
"table_ref": [],
"text": "In this work, we proposed an automatic practical coreset construction framework that requires only two parameters: the input data and the loss function. Our system, namely AutoCoreset, results in small coresets with multiplicative approximation errors significantly smaller than traditional coreset constructions for various machine learning problems, as well as showing that the model learned on our coresets gained more information than the other coresets. While AutoCoreset is practical, we also show some desirable theoretical guarantees. We believe that AutoCoreset can be further enhanced and tuned to work in the context of Deep learning, e.g., subset selection for boosting training of deep neural networks. We leave this as future work.\nFinally, we hope AutoCoreset will lay the foundation of practical frameworks for coresets, and hope it reaches the vast scientific community, aiding to achieve faster training with provable guarantees due to training on our coresets."
},
{
"figure_ref": [],
"heading": "A. More details",
"publication_ref": [
"b0",
"b38",
"b53",
"b16"
],
"table_ref": [],
"text": "More on the initialization technique. Initialization using different approximated solutions is a technique commonly used in optimization algorithms, including kmeans (Arthur & Vassilvitskii, 2007). The idea behind this technique is to start the optimization process from different starting points, or initializations, and to use the resulting approximate solutions to improve the overall optimization performance. This is because different initializations may result in different local optima, and by considering multiple initializations, the optimization algorithm may be able to find a better overall solution. In our method, we do not optimize the given approximate solution, but approximating several approximated solutions using our coreset practically moves the coresets towards approximating \"good\" various regions of the query set, where each of these regions contains a good solution on the dataset. While there is a possibility that the solutions found may be very similar, in practice, the technique tends to provide benefits in terms of improved optimization performance. Practically, we saw that uniform sampling is also sufficient to achieve very good coresets which approximate the optimal solution very well.\nMore on the stopping criteria. First of all, the intuition behind setting stopping criteria is derived from the theory of training models in deep learning. Specifically speaking, the early stopping technique in deep learning. While we could have set the number of iterations to a hard-coded scalar (e.g., 400), we would have either made a very weak coreset that has been exposed to not enough queries, or we would have extended the running time of the algorithm beyond the limits of being practical. The idea that we have used in the paper is to put a threshold on the number of times the minimal cost so far has not changed thus implying some sort of convergence. Notably and most importantly, the usage of such criteria is intensively justified practically in many experimental papers (see for example, (Prechelt, 2002;Zhou et al., 2020;Gu et al., 2018)) in deep/machine learning.\nWe also note that the user can use any stopping criterion and of course, the results will change depending on such a choice.\nThe construction of the query set. We aimed to obtain a coreset that supports a query set that can span a meaningful part of the entire query space. Intuitively speaking, we aim to have a coreset that approximates the loss of a query set containing (i) the optimal solution of the entire data or some fine approximation to it (see next paragraph for an intuitive explanation of how this should intuitively hold) and (ii) the optimal solution on this computed coreset, given a desired problem (e.g., logistic regression). With this in mind, solving the desired problem on our generated coreset will yield a coreset approximating the solution of the entire data up to O(ε).\nHence, in the ith iteration of our algorithm, we add the solution optimizing the current coreset to the supported set of queries (e.g., optimal logistic regression solution for the current coreset).\nSince the coreset is biased towards this solution, we have evaluated the quality of such a solution on the entire data and concatenated such a vector of losses to our matrix of losses (denoted by the matrix M).\nThis, in turn, means that each time a new query is added to the supported set of queries, the coreset in the next iteration will be adapted to approximate every query in the query set and it will become more generalized, or in a sense a \"stronger coreset\".\nWith this in mind, we can initialize our support query set with approximated solutions to the problem (e.g. ε-approximations), so as to ensure a good initial coreset."
},
{
"figure_ref": [],
"heading": "B. Proof of Our Theoretical Results",
"publication_ref": [],
"table_ref": [],
"text": "B.1. Proof of Lemma 3.1\nProof. First, observe that by construction of M(P, f ), it holds that for every x ∈ X , and j ∈ [n], there exists an integer i ∈ [|X |] such that M(P, f ) j,i = f (p j , x) .\n(2) By Definition 2.1, the pair (I, v) satisfies that "
},
{
"figure_ref": [],
"heading": "Acknowledgements",
"publication_ref": [],
"table_ref": [],
"text": "This research was supported in part by the AI2050 program at Schmidt Futures (Grant G-96422-63172), the United States Air Force Research Laboratory, and the United States Air Force Artificial Intelligence Accelerator and was accomplished under Cooperative Agreement Number FA8750-19-2-1000."
},
{
"figure_ref": [],
"heading": "",
"publication_ref": [],
"table_ref": [],
"text": "Note that (3) dictates that for every k ∈ [|X |], it holds that j∈ [n] M(P, f ) j,k -∈I v ( ) M(P, f ) ,k 2 ≤ ε.\n(4)\nFinally, combining (2) and (4) yields Lemma 3.1.\nB.2. Proof of Claim 3.2\nProof. For every k ∈ [z], denote by x k the query which corresponds to the kth column of M(P, f ). The claim holds by the following derivations:\nwhere the first equality hold by the definition of x, the second and thirds are simple rearrangements, the first inequality holds by Claim 3.2."
},
{
"figure_ref": [],
"heading": "C. Experimental Results",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we dive into exploring the effect of the actions/parameters used in AutoCore."
},
{
"figure_ref": [],
"heading": "C.1. Taking the last coreset",
"publication_ref": [],
"table_ref": [],
"text": "In what follows, we show the results of using the last coresets AutoCore has devised, i.e., as Algorithm 1 suggests. samples to outperform the rest of the competitors). This is due to the fact that taking such coresets means that the coreset is becoming more general, thus requiring a larger sample size to guarantee better approximation, one needs to sample more. Such behavior does not appear in our \"optimal coresets\" where we have taken the coreset with the optimal cost; see Figures 12345. The reason for this is that the optimal coreset has been exposed to fewer models/queries than the coreset that would be output by the plain AutoCore, and thus the need for a larger sample size for smaller approximation error becomes less demanding."
},
{
"figure_ref": [],
"heading": "C.2. Exploration of different algorithms for choosing queries",
"publication_ref": [],
"table_ref": [],
"text": "In what follows, we show the effect of different methods for choosing the next query for our practical coreset paradigm with respect to the logistic regression problem."
},
{
"figure_ref": [],
"heading": "C.3. Experimenting with Cifar10 and TinyImageNet",
"publication_ref": [],
"table_ref": [],
"text": "In what follows, we run our coreset paradigm on Cifar10 and TinyImageNet. For TinyImageNet data, we had to use the JL-lemma to reduce the dimensionality of the data. As seen from Figure C.3, our coreset construction technique yields better coresets than uniform sampling even for large-scale datasets, where our coreset can be better than uniform sampling by at max ≈ 1.5 times in terms of relative approximation error. "
}
] | A coreset is a tiny weighted subset of an input set, that closely resembles the loss function, with respect to a certain set of queries. Coresets became prevalent in machine learning as they have shown to be advantageous for many applications. While coreset research is an active research area, unfortunately, coresets are constructed in a problem-dependent manner, where for each problem, a new coreset construction algorithm is usually suggested, a process that may take time or may be hard for new researchers in the field. Even the generic frameworks require additional (problem-dependent) computations or proofs to be done by the user. Besides, many problems do not have (provable) small coresets, limiting their applicability. To this end, we suggest an automatic practical framework for constructing coresets, which requires (only) the input data and the desired cost function from the user, without the need for any other task-related computation to be done by the user. To do so, we reduce the problem of approximating a loss function to an instance of vector summation approximation, where the vectors we aim to sum are loss vectors of a specific subset of the queries, such that we aim to approximate the image of the function on this subset. We show that while this set is limited, the coreset is quite general. An extensive experimental study on various machine learning applications is also conducted. Finally, we provide a "plug and play" style implementation, proposing a user-friendly system that can be easily used to apply coresets for many problems. | AutoCoreset: An Automatic Practical Coreset Construction Framework | [
{
"figure_caption": "Figure 3 .3Figure 3. Illustration of a vector summarization coreset for an input matrix of 7 rows and 3 columns which represent the loss function concerning a set P of 7 input points, and set of queries x1, x2, x3.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Claim 3.2 (Weak Coreset with hidden abilities). Let P = {p 1 , • • • , p n } ⊆ R d be a set of n points as in Section 3.1, f be a loss function supported by AutoCoreset, and let m, τ, ζ be the defined number of initial solutions, sample size, and stopping criterion, respectively. Let z ≥ m, (I, v) be the output of a call to AUTOCORESET (P, f, τ, m, ζ), and let M(P, f ) ∈ R d×z be the matrix of losses that was constructed throughout the running time of AUTOCORESET; see Lines 1, 5, 11 13 at Algorithm 1. Then for any weight function α : [z] → [0, 1] where z i=1 α (i) = 1, and any",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 .Figure 6 .Figure 7 .567Figure 5. Evaluation of our coresets against other competing methods on the Dataset (iii).",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "567",
"figure_type": "figure"
},
{
"figure_caption": "Figure 8 .Figure 9 .89Figure 8. Evaluation of our coresets against other competitors concerning the linear regression problem.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "89",
"figure_type": "figure"
},
{
"figure_caption": "Figure 13 .13Figure 13. Evaluation of our coresets with different algorithms for choosing the next query.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "13",
"figure_type": "figure"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Summary of known vector summarization coresets and their properties.",
"figure_data": "MethodProbability of failureApproximation errorCoreset size |I|Construction timeCaratheodory(Maalouf et al., 2019)",
"figure_id": "tab_1",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Evaluation of our coresets against other competing methods on the Dataset (ii).",
"figure_data": "0 1 2 3 4 5 6 7Uniform sampling Caratheodory coreset Median of means based coreset Sensitivity based coreset for 1-mean Near-Convex Coreset leverage scores lewis weightsTest accuracy0.84 0.86 0.88 0.90 0.92 0.94Uniform sampling Caratheodory coreset Median of means based coreset Sensitivity based coreset for 1-mean Near-Convex Coreset leverage scores lewis weights Entire Data0 1 2 3 4Uniform sampling Caratheodory coreset Median of means based coreset Sensitivity based coreset for 1-mean Near-Convex Coreset SVM coresetTest accuracy0.94 0.92 0.82 0.84 0.86 0.88 0.90Uniform sampling Caratheodory coreset Median of means based coreset Sensitivity based coreset for 1-mean Near-Convex Coreset SVM coreset Entire Data50100150 Sample size 20025030050100150 Sample size 20025030050100150 Sample size 20025030050100150 Sample size 200250300(a) Logistic regression(b) SVMs1 0 1 2 3 4 5 6 7100 Figure 4. 50 150 200 250 300 Sample size Uniform sampling Caratheodory coreset Median of means based coreset Sensitivity based coreset for 1-mean Near-Convex Coreset leverage scores lewis weightsTest accuracy0.92 0.94 0.96 0.98 1.0050100150 Sample size 200 Uniform sampling Caratheodory coreset Median of means based coreset 250 Sensitivity based coreset for 1-mean 300 Near-Convex Coreset leverage scores lewis weights Entire Data0 2 4 6 850100150 Sample size 200 Uniform sampling Caratheodory coreset Median of means based coreset 250 Sensitivity based coreset for 1-mean 300 Near-Convex Coreset SVM coresetTest accuracy0.98 0.96 0.84 0.86 0.88 0.90 0.92 0.9450100150 Sample size 200 Uniform sampling Caratheodory coreset Median of means based coreset 250 Sensitivity based coreset for 1-mean 300 Near-Convex Coreset SVM coreset Entire Data(a) Logistic regression",
"figure_id": "tab_2",
"figure_label": "",
"figure_type": "table"
}
] | Alaa Maalouf; Murad Tukan; Vladimir Braverman; Daniela Rus | [
{
"authors": "D Arthur; S Vassilvitskii",
"journal": "",
"ref_id": "b0",
"title": "K-means++ the advantages of careful seeding",
"year": "2007"
},
{
"authors": "O Bachem; M Lucic; A Krause",
"journal": "ACM",
"ref_id": "b1",
"title": "Scalable k-means clustering via lightweight coresets",
"year": "2018"
},
{
"authors": "O Bachem; M Lucic; S Lattanzi",
"journal": "",
"ref_id": "b2",
"title": "One-shot coresets: The case of k-clustering",
"year": "2018-04-11"
},
{
"authors": "M Bȃdoiu; K L Clarkson",
"journal": "Computational Geometry",
"ref_id": "b3",
"title": "Optimal core-sets for balls",
"year": "2008"
},
{
"authors": "M.-F F Balcan; S Ehrlich; Y Liang",
"journal": "",
"ref_id": "b4",
"title": "Distributed k-means and k-median clustering on general topologies",
"year": "2013"
},
{
"authors": "V Braverman; S H Jiang; -C Krauthgamer; R Wu; X ",
"journal": "",
"ref_id": "b5",
"title": "Coresets for ordered weighted clustering",
"year": "2019-06-15"
},
{
"authors": "V Braverman; D Feldman; H Lang; A Statman; S Zhou",
"journal": "PMLR",
"ref_id": "b6",
"title": "Efficient coreset constructions via sensitivity sampling",
"year": "2021"
},
{
"authors": "C Carathéodory",
"journal": "Mathematische Annalen",
"ref_id": "b7",
"title": "Über den variabilitätsbereich der koeffizienten von potenzreihen, die gegebene werte nicht annehmen",
"year": "1907"
},
{
"authors": "C.-C Chang; C.-J Lin; Libsvm",
"journal": "ACM Transactions on Intelligent Systems and Technology",
"ref_id": "b8",
"title": "A library for support vector machines",
"year": "2011"
},
{
"authors": "K L Clarkson; Coresets",
"journal": "ACM Transactions on Algorithms (TALG)",
"ref_id": "b9",
"title": "sparse greedy approximation, and the frank-wolfe algorithm",
"year": "2010"
},
{
"authors": "R Curtain; S Im; B Moseley; K Pruhs; A Samadian",
"journal": "",
"ref_id": "b10",
"title": "On coresets for regularized loss minimization",
"year": "2019"
},
{
"authors": "D Dua; C Graff",
"journal": "",
"ref_id": "b11",
"title": "UCI machine learning repository",
"year": "2017"
},
{
"authors": "D Feldman",
"journal": "Springer",
"ref_id": "b12",
"title": "Core-sets: Updated survey",
"year": "2020"
},
{
"authors": "D Feldman; M Langberg",
"journal": "",
"ref_id": "b13",
"title": "A unified framework for approximating and clustering data",
"year": "2011"
},
{
"authors": "D Feldman; G Rossman; M Volkov; D Rus",
"journal": "",
"ref_id": "b14",
"title": "Coresets for k-segmentation of streaming data",
"year": "2014"
},
{
"authors": "D Feldman; S Ozer; D Rus",
"journal": "PMLR",
"ref_id": "b15",
"title": "Coresets for vector summarization with applications to network graphs",
"year": "2017"
},
{
"authors": "J Gu; Z Wang; J Kuen; L Ma; A Shahroudy; B Shuai; T Liu; X Wang; G Wang; J Cai",
"journal": "Pattern recognition",
"ref_id": "b16",
"title": "Recent advances in convolutional neural networks",
"year": "2018"
},
{
"authors": "C Guo; Y Ma; B Yang; C S Jensen; M Kaul",
"journal": "",
"ref_id": "b17",
"title": "Ecomark: evaluating models of vehicular environmental impact",
"year": "2012"
},
{
"authors": "I Jubran; A Maalouf; D Feldman",
"journal": "",
"ref_id": "b18",
"title": "Introduction to coresets: Accurate coresets",
"year": "2019"
},
{
"authors": "I Jubran; M Tukan; A Maalouf; D Feldman",
"journal": "",
"ref_id": "b19",
"title": "Sets clustering",
"year": "2020"
},
{
"authors": "Z Karnin; E Liberty",
"journal": "",
"ref_id": "b20",
"title": "Discrepancy, coresets, and sketches in machine learning",
"year": "2019"
},
{
"authors": "M Langberg; L J Schulman",
"journal": "SIAM",
"ref_id": "b21",
"title": "Universal εapproximators for integrals",
"year": "2010"
},
{
"authors": "R J Lyon; B Stappers; S Cooper; J M Brooke; J D Knowles",
"journal": "Monthly Notices of the Royal Astronomical Society",
"ref_id": "b22",
"title": "Fifty years of pulsar candidate selection: from simple filters to a new principled real-time classification approach",
"year": "2016"
},
{
"authors": "A Maalouf; I Jubran; D Feldman",
"journal": "",
"ref_id": "b23",
"title": "Fast and accurate least-mean-squares solvers",
"year": "2019"
},
{
"authors": "A Maalouf; A Statman; D Feldman",
"journal": "",
"ref_id": "b24",
"title": "Tight sensitivity bounds for smaller coresets",
"year": "2020"
},
{
"authors": "A Maalouf; I Jubran; D Feldman",
"journal": "",
"ref_id": "b25",
"title": "Introduction to coresets: Approximated mean",
"year": "2021"
},
{
"authors": "A Maalouf; I Jubran; M Tukan; D Feldman",
"journal": "Sensors",
"ref_id": "b26",
"title": "Coresets for the average case error for finite query sets",
"year": "2021"
},
{
"authors": "A Maalouf; I Jubran; D Feldman",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b27",
"title": "Fast and accurate least-mean-squares solvers for high dimensional data",
"year": "2022"
},
{
"authors": "A Maalouf; M Tukan; E Price; D M Kane; D Feldman",
"journal": "",
"ref_id": "b28",
"title": "Coresets for data discretization and sine wave fitting",
"year": ""
},
{
"authors": " Pmlr",
"journal": "",
"ref_id": "b29",
"title": "",
"year": "2022"
},
{
"authors": "S Minsker",
"journal": "Bernoulli",
"ref_id": "b30",
"title": "Geometric median and robust estimation in banach spaces",
"year": "2015"
},
{
"authors": "A Munteanu; C Schwiegelshohn",
"journal": "KI-Künstliche Intelligenz",
"ref_id": "b31",
"title": "Coresets-methods and history: A theoreticians design pattern for approximation and streaming algorithms",
"year": "2018"
},
{
"authors": "A Munteanu; C Schwiegelshohn; C Sohler; D P Woodruff",
"journal": "",
"ref_id": "b32",
"title": "On coresets for logistic regression",
"year": "2018"
},
{
"authors": "T E Oliphant",
"journal": "Trelgol Publishing USA",
"ref_id": "b33",
"title": "A guide to NumPy",
"year": "2006"
},
{
"authors": "D J Ozer",
"journal": "Psychological bulletin",
"ref_id": "b34",
"title": "Correlation and the coefficient of determination",
"year": "1985"
},
{
"authors": "F Pedregosa; G Varoquaux; A Gramfort; V Michel; B Thirion; O Grisel; M Blondel; P Prettenhofer; R Weiss; V Dubourg; J Vanderplas; A Passos; D Cournapeau; M Brucher; M Perrot; E Duchesnay",
"journal": "Journal of Machine Learning Research",
"ref_id": "b35",
"title": "Scikit-learn: Machine learning in Python",
"year": "2011"
},
{
"authors": "J M Phillips",
"journal": "",
"ref_id": "b36",
"title": "Coresets and sketches",
"year": "2016"
},
{
"authors": "L Prechelt",
"journal": "Springer",
"ref_id": "b37",
"title": "Early stopping-but when?",
"year": "1998"
},
{
"authors": "L Prechelt",
"journal": "Springer",
"ref_id": "b38",
"title": "Early stopping-but when?",
"year": "2002"
},
{
"authors": "G S Sampaio; A R De Aguiar Vallim Filho; L S Da Silva; L A Da Silva",
"journal": "Sensors",
"ref_id": "b39",
"title": "Prediction of motor failure time using an artificial neural network",
"year": "2019-10"
},
{
"authors": "A Tsanas; A Xifara",
"journal": "Energy and buildings",
"ref_id": "b40",
"title": "Accurate quantitative estimation of energy performance of residential buildings using statistical machine learning tools",
"year": "2012"
},
{
"authors": "M Tukan; A Maalouf; D Feldman",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b41",
"title": "Coresets for near-convex functions",
"year": "2020"
},
{
"authors": "M Tukan; C Baykal; D Feldman; D Rus",
"journal": "Theoretical Computer Science",
"ref_id": "b42",
"title": "On coresets for support vector machines",
"year": "2021"
},
{
"authors": "M Tukan; A Maalouf; D Feldman; R Poranne",
"journal": "IEEE",
"ref_id": "b43",
"title": "Obstacle aware sampling for path planning",
"year": "2022"
},
{
"authors": "M Tukan; L Mualem; A Maalouf",
"journal": "",
"ref_id": "b44",
"title": "Pruning neural networks via coresets and convex geometry: Towards no assumptions",
"year": ""
},
{
"authors": " Curran Associates; Inc",
"journal": "",
"ref_id": "b45",
"title": "",
"year": "2022"
},
{
"authors": "M Tukan; X Wu; S Zhou; V Braverman; D Feldman",
"journal": "PMLR",
"ref_id": "b46",
"title": "New coresets for projective clustering and applications",
"year": "2022"
},
{
"authors": "M Tukan; E Biton; R Diamant",
"journal": "",
"ref_id": "b47",
"title": "An efficient drifters deployment strategy to evaluate water current velocity fields",
"year": "2023"
},
{
"authors": "M Tukan; S Zhou; A Maalouf; D Rus; V Braverman; D Feldman",
"journal": "",
"ref_id": "b48",
"title": "Provable data subset selection for efficient neural network training",
"year": "2023"
},
{
"authors": "A V Uzilov; J M Keegan; D H Mathews",
"journal": "BMC bioinformatics",
"ref_id": "b49",
"title": "Detection of non-coding rnas on the basis of predicted secondary structure formation free energy change",
"year": "2006"
},
{
"authors": "G Van Rossum; F L Drake",
"journal": "CreateSpace",
"ref_id": "b50",
"title": "Python 3 Reference Manual",
"year": "2009"
},
{
"authors": "P Virtanen; R Gommers; T E Oliphant; M Haberland; T Reddy; D Cournapeau; E Burovski; P Peterson; W Weckesser; J Bright; S J Van Der Walt; M Brett; J Wilson; K Jarrod Millman; N Mayorov; A R J Nelson; E Jones; R Kern; E Larson; C Carey; İ Polat; Y Feng; E W Moore; J Vand Erplas; D Laxalde; J Perktold; R Cimrman; I Henriksen; E A Quintero; C R Harris; A M Archibald; A H Ribeiro; F Pedregosa; P Van Mulbregt; S Contributors",
"journal": "Nature Methods",
"ref_id": "b51",
"title": "SciPy 1.0: Fundamental Algorithms for Scientific Computing in Python",
"year": "2020"
},
{
"authors": "I.-C Yeh; C Lien",
"journal": "Expert systems with applications",
"ref_id": "b52",
"title": "The comparisons of data mining techniques for the predictive accuracy of probability of default of credit card clients",
"year": "2009"
},
{
"authors": "W Zhou; C Xu; T Ge; J Mcauley; K Xu; F Wei",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b53",
"title": "Bert loses patience: Fast and robust inference with early exit",
"year": "2020"
}
] | [
{
"formula_coordinates": [
2,
307.44,
548.09,
234,
24.14
],
"formula_id": "formula_0",
"formula_text": "s(p) = sup x∈X f (p,x)"
},
{
"formula_coordinates": [
3,
307.44,
359.47,
234,
22.49
],
"formula_id": "formula_1",
"formula_text": "P = {p 1 • • • , p n } ⊆ R d of n points 1 and a loss function f : P × X → [0, ∞)"
},
{
"formula_coordinates": [
3,
307.44,
432.77,
234,
41.95
],
"formula_id": "formula_2",
"formula_text": "I → [0, ∞) such that max x∈X j∈I v(j)f (pj ,x) n i=1 f (pi,x)"
},
{
"formula_coordinates": [
3,
307.44,
667.51,
234,
52.34
],
"formula_id": "formula_3",
"formula_text": "ε-coreset for M if i∈[n] M i, * -j∈I v(j)M j, * 2 2 ≤ ε. 1 if P is a set of labeled items, then P = pi = (p i , yi) p i ∈ R d-1 , yi ∈ R n i=1"
},
{
"formula_coordinates": [
4,
70.16,
342.92,
219.29,
40.02
],
"formula_id": "formula_4",
"formula_text": "i∈[n] M * (P, f ) i, * - j∈I v(j)M * (P, f ) j, * 2 2 ≤ ε. (1)"
},
{
"formula_coordinates": [
4,
55.44,
578.27,
235.65,
35.38
],
"formula_id": "formula_5",
"formula_text": "M(P, f ) of M * (P, f ) ( M(P, f ) contain a subset of the columns of M * (P, f ); see Fig- ure"
},
{
"formula_coordinates": [
4,
377.55,
111.7,
160.78,
4.8
],
"formula_id": "formula_6",
"formula_text": "0 0 m + 1 O(min{nm + log 4 (m), m 2 n 2 , nm 3 })"
},
{
"formula_coordinates": [
4,
377.55,
127.36,
140,
3.94
],
"formula_id": "formula_7",
"formula_text": "0 ε O(1/ε) O(min{nd/ε})"
},
{
"formula_coordinates": [
4,
377.56,
141.29,
155.92,
4.71
],
"formula_id": "formula_8",
"formula_text": "δ ε O(1/ε) O(m log 2 (1/δ) + m log(1/δ)/ε)"
},
{
"formula_coordinates": [
4,
320.19,
153.72,
190,
13.49
],
"formula_id": "formula_9",
"formula_text": "δ ε O( 1 ε (m + log 1 δ )) O(nm) Uniform sampling δ ε O( 1 εδ ) O(1)"
},
{
"formula_coordinates": [
4,
307.44,
210.17,
233.27,
60.18
],
"formula_id": "formula_10",
"formula_text": "ζ output A coreset (I, v) such that 1: M(P, f ) ← → 0 n×m 2: for each i ∈ [m] do 3:"
},
{
"formula_coordinates": [
4,
312.42,
285.45,
131.1,
32.72
],
"formula_id": "formula_11",
"formula_text": "for every j ∈ [n] do 5: M(P, f ) j,i ← f (p j , x i ) 6:"
},
{
"formula_coordinates": [
4,
307.94,
379.58,
171.53,
48.78
],
"formula_id": "formula_12",
"formula_text": "x * ← arg min x∈X i∈I v (i) f (p i , x) 11: M(P, f ) ← M(P, f ) → 0 n 12:"
},
{
"formula_coordinates": [
4,
307.44,
429.06,
152.47,
59.23
],
"formula_id": "formula_13",
"formula_text": "M(P, f ) i,m+1 ← f (p i , x C ) 14: end for 15: m ← m + 1 16: until ζ is satisfied return (I, v)"
},
{
"formula_coordinates": [
4,
477.61,
569.71,
65,
24.35
],
"formula_id": "formula_14",
"formula_text": "x∈X n i=1 f (p i , x)"
},
{
"formula_coordinates": [
5,
55.44,
182.6,
213.75,
31.19
],
"formula_id": "formula_15",
"formula_text": "max x∈X j∈I v(j)f (pj ,x) n i=1 f (pi,x) ∈ 1 + O(ε); see Lemma 3.1."
},
{
"formula_coordinates": [
5,
55.44,
279,
234,
23.18
],
"formula_id": "formula_16",
"formula_text": "x * on p 1 , • • • , p n to obtain the vector of losses l = (f (p 1 , x * ), • • • , f (p n , x * ))"
},
{
"formula_coordinates": [
5,
345.05,
164.42,
162.11,
36.45
],
"formula_id": "formula_17",
"formula_text": "i∈[n] f (p i , x) - j∈I v(j)f (p j , x) 2 ≤ ε."
},
{
"formula_coordinates": [
5,
306.28,
267.33,
94.63,
9.65
],
"formula_id": "formula_18",
"formula_text": "(f (p 1 , x), • • • , f (p n , x)"
},
{
"formula_coordinates": [
5,
307.44,
521.22,
234,
80.34
],
"formula_id": "formula_19",
"formula_text": "[n], f (p i , x) = z k=1 α (k) M(P, f ) i,k , we have n i=1 f (p i , x) - j∈I v(j)f (p j , x) 2 ≤ ε,"
},
{
"formula_coordinates": [
6,
60.99,
690.07,
223.65,
30.32
],
"formula_id": "formula_20",
"formula_text": "t i=1 (T (n, i, ε, δ) + T sol (S(n, i, ε, δ), d)) + tT cost (n, d)."
},
{
"formula_coordinates": [
6,
347.59,
92.55,
131.6,
10.31
],
"formula_id": "formula_21",
"formula_text": "+ (t/ε + log(1/δ)/ε)d 2 + nd))."
},
{
"formula_coordinates": [
6,
307.44,
295.52,
235.93,
21.91
],
"formula_id": "formula_22",
"formula_text": "+ t) + max i∈[t] (M em(VAlg, ε, δ, i) + M em sol (S(n, i, ε, δ)), d)."
},
{
"formula_coordinates": [
6,
392.75,
347.75,
121.71,
10.31
],
"formula_id": "formula_23",
"formula_text": "+ t) + (1/ε(t + log(1/δ))) 2 )."
},
{
"formula_coordinates": [
6,
449.31,
634.71,
54.03,
17.71
],
"formula_id": "formula_24",
"formula_text": "f (p,x) √ p∈P f (p,x)"
},
{
"formula_coordinates": [
7,
69.16,
117.98,
209.88,
36.45
],
"formula_id": "formula_25",
"formula_text": "i∈[n] f (p i , x) - j∈I v(j)f (p j , x) 2 ≤ ε p∈P f (p, x)."
},
{
"formula_coordinates": [
7,
307.44,
591.35,
170.69,
12.72
],
"formula_id": "formula_26",
"formula_text": "ε = i∈[n] f (p i , x * I ) -i∈[n] f (p i , x * P )"
}
] |
10.48550/arXiv.2304.05302 | [
{
"figure_ref": [],
"heading": "INTRODUCTION",
"publication_ref": [
"b0",
"b1",
"b2",
"b3",
"b4",
"b5",
"b6",
"b7",
"b8",
"b9",
"b10",
"b11",
"b12",
"b13",
"b14",
"b15",
"b16"
],
"table_ref": [],
"text": "Large language models (LLMs) are subject of many recent publications and get a lot of attention. Despite that, it is not well-defined what \"large\" actually means. Whereas the BERT model with 340 Mio parameters was dubbed \"large\" by its creators back in 2018 [1], in 2023 it would not be considered a LLM anymore. Later models kept the original naming convention with \"small\", \"base\" and \"large\" for a while and extended it into XL and XXL to cope with the growing number of parameters. However, since the size grew even more into hundreds of billions of parameters, it became more usual to put a number suffix like 30b to designate a certain size of a model, which is more concise. Despite that, the term LLM is still used a lot in publications, and has some overlap with the term \"foundation model\" [2], that is defined as \"[..] any model that is trained on broad data at scale and can be adapted (e.g., fine-tuned) to a wide range of downstream tasks\". In our work we refer to LLMs if the model has at least 100 billion parameters and works on text only (which excludes multimodal models). Examples for such models include Googles LaMDA [3] with 137b, OpenAIs GPT-3 [4] with 175b and Nvidia's Megatron Turing NLG [5] with 540b parameters. One of the benefits of such models is that they are extremely versatile multi-task models. Furthermore, their zero-shot and few-shot performance on a large number of tasks is impressive. Since they are pretrained on text-completion mostly, they are also good for \"answering openended questions in natural language\" which is e.g. explicitly mentioned in the documentation of Aleph Alpha's Luminous. Although the last two years were dominated by LLMs like PaLM [6] and GPT-4 [7], notable publications show that smaller models can perform nearly equally well in a lot of tasks and are much more manageable for smaller enterprises and research institutions, e.g. Chinchilla [8] with 70b parameters, FLAN T5 [9] with 11b and LLaMA [10] with up to 65b parameters. AlexaTM 20B [11] was trained for 15,360 A100 GPU days and outperforms the PaLM 540B model in 1shot summarization (MLSum de, XSum en) and GPT-3 175B in machine translation (de-en) and SuperGLUE results. An 11B parameter model called Unicorn outperforms GPT-3 175B on CommonSenseQA 2.0 by finetuning a pre-trained T5 model on the RAINBOW datasets [12]. This paper therefore concentrates on evaluating medium-sized language models (MLMs) which we define as having at least six billion parameters but less than 100 billion. Other researchers in the meanwhile call language models still small, even if they have 11b parameters [13]. Although being quite large with 130b parameters, the GLM model should also be mentioned here, since its creators explicitly modeled it with the goal to make it accessible for researchers with less compute power [14]. Because of the increased capabilities of the LLMs, moving the evaluation to more realistic scenarios beyond purely factual answers seems necessary. The respective ML tasks are called long-form question answering [15] which was originally designed to involve document retrieval before answering the question. However, LLMs and to some degree also MLMs should be capable of performing it as closed-book QA [16]. They pose the problem that evaluation of results is difficult due to ambiguity of questions (ibid) and other challenges. Due to that, answers of models are hard to evaluate with wide-spread methods like ROUGE [17]. We therefore perform a human evaluation to test model accuracy. The remainder of this paper is structured as follows. We first discuss related work, especially other evaluations of language models. Then, we introduce the dataset used for the evaluation before discussing the choice of models for the test. After that, the experimental setup is described, before AI results and a human baseline are outlined. The paper ends with limitations, conclusion and outlook."
},
{
"figure_ref": [],
"heading": "RELATED WORK",
"publication_ref": [
"b17",
"b18",
"b19",
"b20",
"b21",
"b21",
"b1",
"b22",
"b23",
"b24",
"b25",
"b26",
"b27",
"b28",
"b29",
"b30",
"b31",
"b32",
"b30",
"b33"
],
"table_ref": [],
"text": "To evaluate the performance of LLMs a plethora of benchmarks and datasets were published, e.g. Natural Questions [18], BIGbench [19] and MMLU [20]. However, they all concentrate on questions that are automatically evaluable, which means they either test using multiple-choice questions, which limits the language generation capabilities of the LLMs to generate a single character, or use measures like ROUGE and BLEU, who are known to have severe limitations regarding their ability to identify correct answers that deviate from the wording of the ground truth [21]. Results from holistic evaluation of language models (HELM) [22] confirm the assumption of this paper, that well-tuned MLMs can outperform much larger models. The 52b parameter model from Coherence (v20220609) outperforms the 175b parameters models like GPT-3 davinci v1, J1jumbo v1 and Bloom. The 52b parameter model from Anthropic (v4-s3) performs even better and additionally outperforms OPT-175b and nearly reaches the accuracy of Turing NLG v2 with 530b parameters. However, the performance of MLMs is very different based on their training, esp. fine-tuning data. T0++ for example performs third best in TruthfulQA EM and outperforms all larger models except GPT-3 175B, whereas it is beaten by much smaller models like GPT-J 6B and all of the bigger ones on NarrativeQA closed-book F1. Similary, UL2 20B performs relatively well in NaturalQuestions closed-book F1 nearly reaching the performance of Bloom 175B, but is beaten by GPT-J 6B and Bloom 175B in Truthful QA EM (although only by a small percentage). Another important outcome from [22] is, that \"automated evaluation was not satisfying\" and it is \"necessary to conduct human evaluations to better understand language model performance\". This is done in our paper. The goal of this paper is further similar to HELM in the intend to use the same benchmarks on all considered models instead of a sparse evaluation matrix of tests and models. Besides that, the evaluation in this paper has only a small overlap in models considered (e.g., T0pp and T5) and proposes own tests instead of reusing the popular ones. The evaluation of LLMs in [2] does a good job of summarizing developments of the last years but is not concerned with own benchmark results. Mahowald et al. [23] analyze LLMs from a linguistic perspective and differentiate between formal and functional linguistic competences. Based on literature analysis they reach the conclusion that LLMs are highly competent although not perfect in formal linguistic competence but often fail on functional linguistic competence. The examples they state are however kind of artificial (\"How to get a sofa onto the roof of a house\") and also overcome by newer models like LamDA or ChatGPT (like the trick question to translate a sentence that includes a new direction), which they hint to in stating that they are only talking about models trained without human reinforcement or instruction tuning (p. 9). One major improvement in the advancement of LLMs is using instruction tuning [24]. U-PaLM [25] significantly increases zero-shot performance of PaLM with only 0.1% extra compute, by applying the mixture of denoising training objective from UL2 [26] to a pretrained PaLM model. Flan-PaLM [27] further improves on that by using both instruction-tuning and chain-of-thought prompting. The relative improvement is even greater for the 11B parameter T5 XXL model (+26.6%) compared to the 540B parameter PaLM model (+9.3%). Similarly, Suzgun et al. [28] find that chain-of-thought (CoT) prompting dramatically increases the accuracy of LLMs in hard BIGbench tasks. PaLM, InstructGPT and Codex benefit with at least 12.9% absolute accuracy increase from low 50ies to high 60ies. The highest increase was found for Codex in the algorithmic tasks (+28.5%). However, for smaller model sizes (8B) there was a negative impact using CoT. For extremely hard tasks, CoT prompting helped the model to create emergent capabilities although those tasks seemed to be not affected by model scale and would require complete new architectures. [29] collect a large number of instructions in order to finetune MLMs on diverse -tasks and achieve good results. Similarly, [30] perform finetuning but use an automatically generated dataset to achieve comparable accuracy on the BIG-bench hard subset. Multi-step reasoning is still challenging for LLMs [31]. One example for advancement in this area is the Self-Taught Reasoner (STaR) introduced by [32], in which a LLM is trained and refined on its own output iteratively. Specifically, with CoT prompting, the model first generates initial rationales. And then, the model is finetuned on rationales that lead to correct answers. As a follow-up to this work [33] show that LLMs are able to self-improve their reasoning abilities without the need for supervised data by leveraging the self-consistency of reasoning. Benchmarks that can be used for testing commonsense reasoning [31] abilities of LLMs include CSQA, StrategyQA and ARC. We refer the reader to Bhargava and Ng (2022)'s survey for more work in this domain. According to [34], LLMs exhibit reasoning patterns similar to those of humans as described in the cognitive literature."
},
{
"figure_ref": [],
"heading": "DATASET",
"publication_ref": [],
"table_ref": [],
"text": "Our own dataset is self-constructed and takes some inspiration from existing datasets like BigBench, TriviaQA and AmbigQA. The following categories are included.\n• Abstractions replace one well-known concept with a different one and force the model to answer based on the replacement. Example: Assume that purple represents a car and red represents a roof. What do you get if you remove the red part from purple? • Basic physics requires some background knowledge and its application to more or less common situations.\nExample: If a ball drops from 2 meters height onto the floor and the floor is made of stone and the ball is made of glass. What happens to the ball? • Everyday knowledge is easy for humans to answer, but unlikely to be found 1:1 in the training data.\nExample: 10 year old John is going shopping with his grandfather Raymond. Who is more likely to want to buy some cigarettes?\n• Trick questions are made to fool humans and it is interesting to see whether the AI can be fooled in the same way.\nExample: Which weighs more, a pound of silver or a pound of gold? • Metaphors use well-known English sayings or phrases and turn them into a question. It requires recognition of the saying that is presented in a slightly different form and an understanding of the metaphoric meaning Example: What kind of coals do you need to take coals to Newcastle? • Math word puzzles are known to cause problems for LLMs. We therefore only include a few of them and also combine them with questions that look mathematical but need no calculation for a correct answer.\nExample: If Susan is running faster than Joe, but slower than Mike and the three do a 100 meter race, who will win? • Relational reasoning transfers the rule of three to everyday objects and requires to understand similarities and differences.\nExample: A house relates to a skyscraper like a flower relates to what? • Deductive reasoning requires to derive conclusions from the premises of the question.\nExample: If the flow of time causes the hands of a clock to turn to the right, what happens if time could run backwards? • Symbolic reasoning is a bit similar to abstractions but uses short variable names instead of words that are defined in a different way as replacements.\nExample: If x is a boy and X is a man, what is y if Y is a woman?\nOften, a (missing) deeper understanding of the model can be seen when comparing the answers to related questions. In the basic physics category, there are several questions regarding balls dropping on the floor and only the height or the ball material is varied. If the answers reflect this variation, the model seems to be able to capture the required understanding. If it always answers \"it bounces\" no matter whether the ball is made of rubber, steel or glass, it shows that the model did not understand. We also did vary the wording to find out if it makes a difference how questions are asked. The dataset will be published on opendata.iisys.de."
},
{
"figure_ref": [],
"heading": "CHOICE OF MODELS TESTED",
"publication_ref": [
"b5",
"b7",
"b34",
"b9",
"b35",
"b36",
"b37",
"b38",
"b39",
"b40",
"b26",
"b41",
"b42",
"b40",
"b41",
"b43",
"b44",
"b28",
"b45"
],
"table_ref": [],
"text": "The primary source for models to be tested was huggingface. Models were included if they fall in the medium-size category, are pretrained at least in English language (multi-lingual models were included as well) and preferably already finetuned on closed-book question answering or instruction-tuned in general. However, we also included models without any finetuning. Models that were trained for extractive question answering instead of generative were excluded as well as those that need a document retriever. Models that are not publicly available like PaLM 62B [6] or Chinchilla 70B [8] were excluded as well.\nAt the beginning of the study in November 2022, the only model available in multiple sizes of the MLM type was OPT [35] from Meta. In order to study scaling effects, all four models from 6.7b to 66b parameters were included. With OPT-IML and Galactica there are also two 30B parameter variants available that build on OPT-30B and add instruction tuning. Later on, Meta released LLaMA [10] and soon after that Stanford published the instruction-tuned LLaMa version Alpaca [36]. Following these two releases, a number of derivatives and similar models have been published including Vicuna [37], [38] and Databricks' Dolly [39]. Whereas Alpaca and Vicuna are based on LLaMA, Dolly v1 is based on GPT-J 6B and v2 on Pythia, an open model from Eleuther AI [40]. Furthermore, several versions of T5 [41] were added to study the effect of different finetuning methods and datasets. These include Flan-T5 [27], mT0 [42], T0pp [43] and T5-SSM-TQAO [41]. ChatGPT from OpenAI, which is a fine-tuned version of GPT-3.5 with LLMs.\nThe goal was to include as many and as recent models as possible, so models from the collectives BigScience [42], [44] and Eleuther AI [45] have been included as well as models from the Allen AI institute [29], Stability AI and Bejing AI [46]. So BloomZ with its 7b parameters can be fairly compared to OPT 6.7b, GPT-J 6b and GPT-JT 6B. GLM 10b can be compared to the T5 variants with 11B parameters and LaMA 13b. GPT Neo-X with its 20b parameters is a bit in between and should be compared to both the 13b and 30b models. Models that were explicitly geared towards dialog like Guanaco, HuggingChat, Koala and OpenAssistant 1 were not included in the comparison and are planned for a future analysis with special focus on chatbots. Late additions in the test were 4 bit models provided by Huggingface user TheBloke that used finetuning in 4bit and therefore allowed much larger models to be finetuned with limited resources like WizardLM 30B and Wizard Vicuna 30B, as well as models published beginning of June 2023 like Luminous Supreme Control, Falcon 40B Instruct and Dromedary 65B. No models with experimental 8k or larger context size were included."
},
{
"figure_ref": [],
"heading": "EXPERIMENTS",
"publication_ref": [
"b46",
"b47"
],
"table_ref": [],
"text": "We followed the instructions of the creators of the MLMs, e.g. by using prefixes like \"q: \" before and \"a: \" after the question, or \"please answer the following question:\" as instruction. We did not use any prompt-engineering or chain-of-thought prompting. Except the LLM references ChatGPT and Luminous which were used as part of their manufacturers' cloud offerings, all models were run on an A100 80 GB GPU (or multiple if necessary) on our local server with FP16 precision.\nTo make results more reproducible we set the temperature value to 0.1.\nOpen-ended questions have the problem, that they cannot be easily evaluated in an automated way. It is not only possible to give the correct answer in an alternative formulation that might not be detected by current evaluation methods like BLEU and ROUGE [47], but there were also answers given by the language models that were correct and surprising to humans so that even advanced methods like BERTscore [48] would not help detecting the correctness. Flan-Alpaca for instance answered \"Tempura\" to the question \"What relates to Japan like pizza relates to Italy?\". The ground truth answer was \"Sushi\", but Tempura seems an even better answer since it is also well-known and additionally closer related to pizza than Sushi. Therefore, a manual evaluation of the answers was performed. Initially, the answers were rated per model. Later on, a cross-check per question across models was performed to assure an equal treatment of each model, since human evaluation comes with the risk of subjectivity."
},
{
"figure_ref": [],
"heading": "AI RESULTS",
"publication_ref": [
"b51"
],
"table_ref": [
"tab_2"
],
"text": "In The OPT family of models showed subpar performance with 11% correct answers for the 30B parameter model (see table 3) and seems to prove the warning you often read, that pre-trained models without any finetuning are not usable for downstream tasks. However, if you consider GLM-10B and LLaMA-13B, the models achieve 37.3% and 38.2% correct answers without any finetuning, the statement doesn't seem to be correct. They are therefore in the same performance range as T5 finetuned on closed book QA and instruction tuned T0pp with accuracies between 37.3% and 40%. The larger models OPT-30B and LLaMA-30B did not outperform their smaller siblings. OPT-66B was significantly better than the smaller OPT models, but still substandard given its size. OPT-IML 30B and Galactica 30B with 18.2% and 12.7% respectively, were also rather disappointing. We could not produce usable results with LLaMA 65B, which may point to an erroneous checkpoint being leaked / published. The 70B parameter model Luminous Supreme from the German startup Aleph Alpha performs similar to OPT-66B with 30% accuracy. In June, a new version of it called \"instruct\" was published that increased this result to 41.8%, which is still below the best 7B parameter models.\nAfter the publishing of QLora in May 2023 [52], a lot of new instruction-tuned models with 30B parameters and more were published. This pushed the previously low average score of the largest models significantly. However, the gain for well-trained smaller models is not large. "
},
{
"figure_ref": [],
"heading": "HUMAN BASELINE",
"publication_ref": [],
"table_ref": [],
"text": "A test with different groups of humans was performed to determine a human baseline per category of questions. This is not only used to compare the performance of MLMs but also to verify the judgement of what a correct or plausible answer should look like. All participants were non-native English speakers but had a good English speaking level so that they can attend English study programs. 32% had a background in computer science, 22% in business administration, another 22% in engineering and the final 22% other. The questions that were asked to the AI models were split into four questionnaires, so that every participant saw only parts of the whole question catalogue which helped keeping the time to answer within bounds (15.5 min median). The participants did not get any incentives for participating.\nOverall, there were 87 participants (41.4% female, 54.0% male) who finished the questionnaire (dropout 8.4%). The median age was 23 years (avg. 25.8).\nIt was expected that humans would in general be able to answer the questions, but would also make a couple of mistakes, so that the baseline would be around 90%. Astonishingly, the average human score was only 70.1%. Questions were partly trivial to answer for humans, but also partly challenging. As expected, a significant portion of humans had problems both in math word questions as well as abstractions and symbolic reasoning. They also fell for some of the trick questions. Some had problems with missing background knowledge especially for historic celebrities like Margaret Thatcher or Edwin Moses due to their young age. Still, some of the subjectively trivial questions like relations between animal types resulted in surprising answers (e.g. donkey and zebra as an answer to \"A tiger relates to a wildcat like a horse relates to what?\" instead of pony). On the other hand, some of the questions that were rather controversial because they seem underspecified, were answered relatively homogeneously in the expected way. Especially age-related behavior was only scarcely questioned and over 95% of participants agreed that children are more likely to eat ice cream than their grandparents and vice versa for cigarettes.\nIt can be seen as confirmation of prejudice or as a Fermi question, but most respondents agree that females are more likely to buy hair color than males (87%). Regarding the trick questions there were large differences. Some of the obvious ones were answered worse than expected, e.g. only 26% of respondents noticed, that an electric train does not produce smoke and only 5% found that the car which can drive up to 120 km/h won't accelerate to 200 km/h. On the other hand, only 25% fell for the question about how long bamboo needs to grow to 30m height, if it can grow up to 20m tall, although it is quite similar. Surprisingly, the wrong answers from humans were the same or very similar to those of the AI models, even if the question did not push them into the wrong direction as it is the case in the trick question category. Both humans and AI e.g., saw the symbols XYxx as an indication of transgender instead of a family with two boys since it was defined in the question that X represents a man and Y a woman."
},
{
"figure_ref": [],
"heading": "DISCUSSION",
"publication_ref": [
"b9",
"b7"
],
"table_ref": [],
"text": "On average, the MLMs were able to answer 35.3% of questions correctly, which means just over a third. We therefore conclude that our dataset is challenging. The best performing model was Airoboros 33B which scored 71.8% and therefore clearly above the reference LLM ChatGPT (60.9% correct answers) and even slightly above the human baseline. Some of the larger MLMs, especially the 30B parameter models were somewhat disappointing since they did not outperform their smaller siblings. However, the best models were still from this category. It is also surprising that the correct answers of the models, especially Flan-T5 and Vicuna 13B are somewhat complementary. The 44.5% and 41.8% of the models add up to 62.7% correct answers which even outperforms ChatGPT (60.9%). You would expect that questions are either harder or easier for models to answer and that well performing models give good answers to the same questions, if the questions were not included in their training data. However, across all MLMs the correct answer rate was 91.8% and together with ChatGPT only 4 questions could not be answered correctly. Short answers are preferable for factual information, while longer answers are suitable for factbased judgments. The models allow for additional parameters, such as specifying the maximum number of tokens for the answer. However, this often results in truncated answers that abruptly end in the middle of a sentence, rather than providing shorter responses. Ideally, the model should autonomously distinguish between answers that are better when concise and those that require additional explanations. However, there is also a subjective notion to that judgment. An MLM's poor quality is evident when it generates unrelated text or merely reproduces training data without providing a relevant answer. A similarly bad behavior is asking new questions that are almost identical to the original question but do not contribute to a proper answer. It sometimes seemed, as if this was the MLMs way of saying: \"I have no idea\". The ability to confess being unknowledgeable is lacking in all models but GPT Neo X. Filtering to avoid biases seems rather undesirable. It would be better to train the model for desirable answers. Not only ChatGPT with filtering, but also StableLM and WizardLM without any filtering showed signs of trying to teach the user, e.g. in preaching healthy living styles without smoking when being asked about wildfires that are caused by smokers. This seems also undesirable, although in general giving advice for self-improvement of the user is good. Another similar issue is the answer: \"It is not appropriate to make assumptions about a person's personal preferences based on their age.\" given by StableLM on the question about the likelihood of buying something based on age.\nRepeating the question as part of the answer has pros and cons but is rather undesirable. Luminous has even an option for penalizing this, although it did not seem necessary there.\nChatGPT on the other hand does that very frequently. Generating options before giving an answer must be viewed as an undesirable feature and is present in many models that are not finetuned. Galactica is one of the worst regarding that. Sometimes models do generate only options with no choice afterwards, or the options were so long, that they did not fit into the maximum answer length. This was rated equally to unanswered or wrong answer.\nHallucination is a well-known problem of LLMs and not surprisingly, the problem was observed for MLMs as well. To go into more detail, a conspicuous situation that seems to force MLMs to hallucinate are questions regarding similarities. This is a situation where humans as well might get into speculating if they do not find an obvious similarity. Since LMs in general do have problems in confessing that they do not know about certain things, it is not surprising, that they invent similarities between the celebrities, e.g. common birthplace, age of dying, art or sport area and so on.\nPrompt engineering should only be an intermediate step towards better language models, since it should not be the task of a human to ask the question in a way that allows the LLM to give the correct answer, but the LLM should be trained in a way that allows it to understand all kinds of questions and always gives the best possible answer (given its training data). For Luminous for example, it made a great difference whether it is prompted with the context and question only, or there was a prefix \"question: \" before the actual question. It did not help to put the \"question: \" prefix before the context. With the prefix, the answers were much better than without. It is even very picky regarding some wordings, e.g. it is more likely to produce correct answers if you start the context with \"let's assume\" instead of just \"assume\". Mathematical capabilities of the models tested are very different. Some models are able to perform some basic calculations like adding, subtracting, multiplying and dividing and use these capabilities to solve some simple math word problems. However, in most cases they struggle if there are too many calculations involved, even if they are simple to calculate. They also mostly fail to do unit conversions, e.g. from meters to centimeters or from kilometers per hour to meters per second. Astonishingly, ChatGPT is able to do the latter, but unable to correctly perform the former, although it recognizes that it has to do a conversion. Regarding scale, there was no clear tendency. Although larger models performed better in general (e.g. LLaMA 13b was 5.5% absolute better than LLaMA 7b), both LLaMA and OPT 30b models performed worse than the smaller models. Also, Galactica 30b and OPT-IML 30b were not as good as expected and even the 70B Luminous Supreme model performed worse than several 13b and even 7b parameter models. The assumption is, that the 30B models are undertrained compared to the smaller models. This finding is in line with the degradation of LLaMA 65B compared to LLaMA 33B in zero-shot settings for NaturalQuestions, ARC-e and ARC-c [10]. Typically, larger models with enough training outperform smaller models in every aspect and especially in zero-shot performance. We also hypothesize that instruction-tuned models perform better the more compute was invested in their finetuning. Another reason could be that for demanding tasks a low score of around 20% accuracy is still in the area where chance plays a role. The empirically observed hockey-stick curves when scaling language models and evaluating their performance compared to scale seems still in the \"blade\" area of the curve and not yet in the \"shaft\" area.\nComparing the performance of OPT and LLaMA, the latter models perform way better than OPT, so there is an advancement from OPT over OPT-IML to LLaMA. This seems to be due to increased training data and also epochs of training. Meta doesn't state exactly how long the OPT models have been trained, but the usage of 992 A100 GPUs compared to the 2048 for LLaMA together with the increase in training tokens from 180 B (OPT) to 1.4 T for LLaMA suggest that OPT is heavily undertrained and LLaMA compares to OPT similar as Chinchilla [8] compares to Gopher."
},
{
"figure_ref": [],
"heading": "LIMITATIONS",
"publication_ref": [
"b37"
],
"table_ref": [],
"text": "Inference time was not measured explicitly, but never exceeded a few seconds (<5). per question on an A100, depending on the number of tokens produced (<100) and the model size (<30B).\nThe human evaluation was done by the authors only. For future work, there should be a crosscheck with automatic evaluation based on GPT-4 as proposed in [38] and more human evaluators should be included.\nOnly a small number of test questions was used (110 altogether). This kept the effort for human evaluation within bounds, but as a downside, tests only a limited amount of application areas, e.g., no questions were included that tested for racial or gender bias explicitly.\nThe human baseline was limited to students and university staff and all were non-native speakers of English language. We did not perform any further analysis of correlations between number of correct answers in a specific question category and the academic background of the participants yet.\nThe evaluation was done with FP16 for all models, initially. Later on, a few 4bit quantized models were tested using the GPTQ for Lora framework. We tested a few models in both FP16 and int4 and could not find a significant difference. Therefore, models available in GPTQ format were used wherever possible since June 2023. The bad performance of several larger models with 30B and more parameters is astonishing, and it cannot be completely excluded that there are technical problems involved when using multiple GPUs for inferencing. We ran the models based on the advice in the corresponding papers and model cards to the best of our knowledge, but independent verification of the results is necessary."
},
{
"figure_ref": [],
"heading": "CONCLUSIONS AND OUTLOOK",
"publication_ref": [
"b26",
"b35",
"b49",
"b53",
"b54",
"b55"
],
"table_ref": [],
"text": "If you take together all the right answers from the different MLMs, 91.8% of the questions were answered correctly. The question remaining is therefore, how to combine the best of all models into a single model within a range of 7-30B parameters. It seems that using the right training data for finetuning is more important than the pure number of parameters. However, this finding might be due to a similarity of some training data to our own dataset. It was beyond the scope of this paper to make a detailed evaluation of the overlap between questions in our test dataset and the training data of each model tested. We assume that the overlap is rather small, since we took quite some efforts to come up with unique questions. Only the trick questions are likely to be included in training data, since they were taken from the internet. However, performance on those was rather bad. Only one model correctly answered the question about getting out of an imaginary room and none was able to figure out, an electric train does not produce smoke. Instruction-tuning and RLHF provide a much better training resulting in models giving substantially better answers than models without this kind of finetuning as stated in literature [27], [36], [50], [54]. However, to unlock the full potential of MLMs, they would need even more finegrained feedback. The longer the answers get, the less an aggregated score summarizing the human preference for the answer helps. Promising future directions are to analyze ways to target the feedback to certain parts of the answer. Consider multi-hop reasoning for example. Was the first step already faulty, or was it the final conclusion that did not fit to the previous intermediate results, although these were correct? This makes a big difference and would be problematic for humans as well, if we would always just get aggregated feedback. Imagine writing a two-page essay and getting the grade as the only indicator on how well you performed. It would be very hard to get better at writing essays with this kind of feedback and no alternatives for learning. We need a similar development as it was performed for sentiment analysis when moving from an overall rating of a product review (positive/negative) to aspect-based sentiment analysis [55], that provides much finer grained judgements that are much more helpful. During the review process of this paper, OpenAI published own findings that support this claim [56]."
},
{
"figure_ref": [],
"heading": "ACKNOWLEDGEMENTS",
"publication_ref": [],
"table_ref": [],
"text": "The authors would like to thank everyone, just everyone!"
},
{
"figure_ref": [],
"heading": "APPENDIX A",
"publication_ref": [],
"table_ref": [],
"text": ""
}
] | Large language models (LLMs) have garnered significant attention, but the definition of "large" lacks clarity. This paper focuses on medium-sized language models (MLMs), defined as having at least six billion parameters but less than 100 billion. The study evaluates MLMs regarding zero-shot generative question answering, which requires models to provide elaborate answers without external document retrieval. The paper introduces an own test dataset and presents results from human evaluation. Results show that combining the best answers from different MLMs yielded an overall correct answer rate of 82.7% which is better than the 60.9% of ChatGPT. The best MLM achieved 71.8% and has 33B parameters, which highlights the importance of using appropriate training data for fine-tuning rather than solely relying on the number of parameters. More fine-grained feedback should be used to further improve the quality of answers. The open source community is quickly closing the gap to the best commercial models. | EVALUATION OF MEDIUM-LARGE LANGUAGE MODELS AT ZERO-SHOT CLOSED BOOK GENERATIVE QUESTION ANSWERING | [
{
"figure_caption": "initial tests, BloomZ was the best model in the 7B parameter range with 35.5% accuracy. It outperforms Alpaca 7B (chavinlo, 33.6%) in our experiments, but only by a small margin (see table1). Alpaca is based on LLaMA 7B and chavinlos model is not improving the already good base performance of LLaMA a lot (32.7%). However, it was unclear whether the replication of Stanford's Alpaca that is hosted publicly on Huggingface (chavinlo/alpaca-native) is really performing as good as the original. The keyword alpaca produced 605 results on huggingface (2 nd of May 2023). Most have no model-card and several did not run with the code we used for testing LLaMA. However, using Wenxiang Jiao's Alpaca 7B repository produced the surprising result that it was performing not only much better than the first Alpaca model tested, but outperformed all other MLMs with 7b parameters up to this date with 46.4% correct answers. How much of an improvement instruction tuning can give is also visible for GPT-J and its fine-tuned version GPT-JT. The latter improves the rather bad 17.3% performance of the base model to 28.2%. This is however still worse than the Dolly v1 version with 6B parameters, which is also based on GPT-J and scored 31.8%. Instruct GPT-J further pushes this score to 39.1%. Surprisingly, Dolly v2 does not score better than v1 but only 30.0%, although it has twice the number of parameters. Its base model Pythia 12B scores 19.1%, which is also worse than expected. Results of models with 7B parameters and less (33.6% avg.) Wombat, another finetuned LLaMA version, but this time with a reinforcement learning approach[49]. It is available in two versions with instructions generated by ChatGPT and GPT-4 respectively. Both perform very good and nearly reach Alpaca's performance. However, they behave quite differently since the GPT-4 instructed model gives quite concise answers (either right or wrong), whereas the other version produces very verbose answers and starts nearly every answer with \"As an AI language model, I do not have personal beliefs or opinions\". With Falcon 7B and MPT-7B, two strong competitors joined the field in June 2023 with 40% and 40.9% correct answers. They both rely on own pretrained models and can therefore be used commercially, in contrast to the LLaMA-based alternatives. Finally, WizardLM took the lead in the 7B parameter models with 47.3% correct answers. The finetuned T5 family of models performed rather good in our tests (see table2). Scores reach from 37.3% to 44.5% of Flan-T5. However, they still perform slightly worse than the smaller models Alpaca 7B and WizardLM 7B. Flan-Alpaca and Vicuna 13B perform similarly good. T5 also shows how much of an effect finetuning has since the base model scores only 13.6% which means 24% to 30% increase. One notable exception is mT0 xP3 XXL, which is a multi-lingual version of T5. It seems that its subpar performance with only 20.9% correct answers is due to the multi-lingual pretraining since BloomZ 7B is also finetuned with xP3 and shows very good results.",
"figure_data": "NameAccuracyNameAccuracyAlpaca (chavinlo)33.6%LLaMA 7B32.7%Alpaca (wxjiao)46.4%MPT-7B-Instruct40.9%BloomZ 7B37.3%OpenLLaMA 7B Instruct31.8%Dolly v1 6B31.8%OpenLLaMA 7B OpenInst.30.0%Falcon 7B Instruct40.0%OPT 6.7b18.2%GPT-J 6b18.2%StableLM 7B11.8%GPT-JT 6B28.2%WizardLM 7B47.3%Instruct GPT-J 6B39.1%Wombat 7B44.5%Wombat 7B GPT440.9%",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Results of models with 10B to 13B parameters(35.4% avg.) ",
"figure_data": "NameAccuracyNameAccuracyAiroboros 13B 4b46.4% Cerebras 13B19.1%Dolly v2 12B30.0% Flan T5 XXL44.5%Flan Alpaca XXL44.5% GLM 10B37.3%GPT Neo X14.5% LLaMA 13B38.2%mT0 xP320.9% Minotaur 13B fixed56.4%Nous Hermes 13B56.4% OPT 13B10.9%Orca Mini 13B 4b28.2% Pythia 12B deduped19.1%T5 v1.1 XXL13.6% T5 XXL SSM TQAO37.3%Vicuna41.8% T5-11b-TQAO38.2%WizardLM 13B52.7% WizardLM 13B 4b53.6%",
"figure_id": "tab_1",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Tulu 30B performed only mediocre with 49.1% compared to Caldera AI's Lazarus 30B that achieved 64% and therefore outperformed ChatGPT (60.9%). The best overall models however came from a single developer called Jon Durbin. His Airoboros model outperforms ChatGPT and Lazarus with 65.5% accuracy (65B 4bit model) and even outperforms the human baseline with 71.8% with its 33B parameter model. Results of models with 30B parameters and more (44.8% avg w/o ChatGPT)",
"figure_data": "WizardLM's",
"figure_id": "tab_2",
"figure_label": "3",
"figure_type": "table"
}
] | René Peinl; Johannes Wirth | [
{
"authors": "J Devlin; M.-W Chang; K Lee; K Toutanova",
"journal": "",
"ref_id": "b0",
"title": "Bert: Pre-training of deep bidirectional transformers for language understanding",
"year": "2018"
},
{
"authors": "R Bommasani",
"journal": "",
"ref_id": "b1",
"title": "On the opportunities and risks of foundation models",
"year": "2021"
},
{
"authors": "R Thoppilan",
"journal": "",
"ref_id": "b2",
"title": "Lamda: Language models for dialog applications",
"year": "2022"
},
{
"authors": "T Brown",
"journal": "Advances in neural information processing systems",
"ref_id": "b3",
"title": "Language models are few-shot learners",
"year": "2020"
},
{
"authors": "S Smith",
"journal": "",
"ref_id": "b4",
"title": "Using deepspeed and megatron to train megatron-turing nlg 530b, a largescale generative language model",
"year": "2022"
},
{
"authors": "A Chowdhery",
"journal": "",
"ref_id": "b5",
"title": "Palm: Scaling language modeling with pathways",
"year": "2022"
},
{
"authors": "S Bubeck",
"journal": "",
"ref_id": "b6",
"title": "Sparks of artificial general intelligence: Early experiments with gpt-4",
"year": "2023"
},
{
"authors": "J Hoffmann",
"journal": "",
"ref_id": "b7",
"title": "Training Compute-Optimal Large Language Models",
"year": "2022"
},
{
"authors": "J Wei",
"journal": "",
"ref_id": "b8",
"title": "Finetuned language models are zero-shot learners",
"year": "2021"
},
{
"authors": "H Touvron",
"journal": "",
"ref_id": "b9",
"title": "LLaMA: Open and Efficient Foundation Language Models",
"year": "2023-02-24"
},
{
"authors": "S Soltan",
"journal": "",
"ref_id": "b10",
"title": "Alexatm 20b: Few-shot learning using a large-scale multilingual seq2seq model",
"year": "2022"
},
{
"authors": "A Talmor",
"journal": "",
"ref_id": "b11",
"title": "Commonsenseqa 2.0: Exposing the limits of ai through gamification",
"year": "2022"
},
{
"authors": "L C Magister; J Mallinson; J Adamek; E Malmi; A Severyn",
"journal": "",
"ref_id": "b12",
"title": "Teaching small language models to reason",
"year": "2022"
},
{
"authors": "A Zeng",
"journal": "",
"ref_id": "b13",
"title": "Glm-130b: An open bilingual pre-trained model",
"year": "2022"
},
{
"authors": "A Fan; Y Jernite; E Perez; D Grangier; J Weston; M Auli",
"journal": "",
"ref_id": "b14",
"title": "ELI5: Long form question answering",
"year": "2019"
},
{
"authors": "R K Amplayo; K Webster; M Collins; D Das; S Narayan",
"journal": "",
"ref_id": "b15",
"title": "Query Refinement Prompts for Closed-Book Long-Form Question Answering",
"year": "2022"
},
{
"authors": "K Krishna; A Roy; M Iyyer",
"journal": "",
"ref_id": "b16",
"title": "Hurdles to progress in long-form question answering",
"year": "2021"
},
{
"authors": "T Kwiatkowski",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b17",
"title": "Natural questions: a benchmark for question answering research",
"year": "2019"
},
{
"authors": "A Srivastava",
"journal": "",
"ref_id": "b18",
"title": "Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models",
"year": "2022"
},
{
"authors": "D Hendrycks",
"journal": "",
"ref_id": "b19",
"title": "Measuring massive multitask language understanding",
"year": "2020"
},
{
"authors": "J Risch; T Möller; J Gutsch; M Pietsch",
"journal": "",
"ref_id": "b20",
"title": "Semantic answer similarity for evaluating question answering models",
"year": "2021"
},
{
"authors": "P Liang",
"journal": "",
"ref_id": "b21",
"title": "Holistic evaluation of language models",
"year": "2022"
},
{
"authors": "K Mahowald; A A Ivanova; I A Blank; N Kanwisher; J B Tenenbaum; E Fedorenko",
"journal": "",
"ref_id": "b22",
"title": "Dissociating language and thought in large language models: a cognitive perspective",
"year": "2023"
},
{
"authors": "J Wei",
"journal": "",
"ref_id": "b23",
"title": "Chain of thought prompting elicits reasoning in large language models",
"year": "2022"
},
{
"authors": "Y Tay",
"journal": "",
"ref_id": "b24",
"title": "Transcending scaling laws with 0.1% extra compute",
"year": "2022"
},
{
"authors": "Y Tay",
"journal": "",
"ref_id": "b25",
"title": "Unifying Language Learning Paradigms",
"year": "2022"
},
{
"authors": "H W Chung",
"journal": "",
"ref_id": "b26",
"title": "Scaling instruction-finetuned language models",
"year": "2022"
},
{
"authors": "M Suzgun",
"journal": "",
"ref_id": "b27",
"title": "Challenging BIG-Bench tasks and whether chain-of-thought can solve them",
"year": "2022"
},
{
"authors": "Y Wang",
"journal": "",
"ref_id": "b28",
"title": "Super-naturalinstructions: Generalization via declarative instructions on 1600+ nlp tasks",
"year": "2022"
},
{
"authors": "O Honovich; T Scialom; O Levy; T Schick",
"journal": "",
"ref_id": "b29",
"title": "Unnatural Instructions: Tuning Language Models with (Almost) No Human Labor",
"year": "2022"
},
{
"authors": "J Huang; K C ; -C Chang",
"journal": "",
"ref_id": "b30",
"title": "Towards Reasoning in Large Language Models: A Survey",
"year": "2022"
},
{
"authors": "E Zelikman; Y Wu; J Mu; N D Goodman",
"journal": "",
"ref_id": "b31",
"title": "STaR: Self-Taught Reasoner",
"year": "2022"
},
{
"authors": "J Huang",
"journal": "",
"ref_id": "b32",
"title": "Large language models can self-improve",
"year": "2022"
},
{
"authors": "I Dasgupta",
"journal": "",
"ref_id": "b33",
"title": "Language models show human-like content effects on reasoning",
"year": "2022"
},
{
"authors": "S Zhang",
"journal": "",
"ref_id": "b34",
"title": "Opt: Open pre-trained transformer language models",
"year": "2022"
},
{
"authors": "R Taori",
"journal": "",
"ref_id": "b35",
"title": "Stanford Alpaca: An Instruction-following LLaMA Model",
"year": "2023-03-13"
},
{
"authors": " Cmu; Mbzuai Stanford; Diego Uc San",
"journal": "",
"ref_id": "b36",
"title": "Vicuna: An Open-Source Chatbot Impressing GPT-4 with 90%* ChatGPT Quality",
"year": "2023-05-02"
},
{
"authors": "B Peng; C Li; P He; M Galley; J Gao",
"journal": "",
"ref_id": "b37",
"title": "Instruction tuning with gpt-4",
"year": "2023"
},
{
"authors": "M Conover",
"journal": "Databricks",
"ref_id": "b38",
"title": "Free Dolly: Introducing the World's First Truly Open Instruction-Tuned LLM",
"year": "2023-04-12"
},
{
"authors": "S Biderman",
"journal": "",
"ref_id": "b39",
"title": "Pythia: A suite for analyzing large language models across training and scaling",
"year": "2023"
},
{
"authors": "A Roberts; C Raffel; N Shazeer",
"journal": "",
"ref_id": "b40",
"title": "How much knowledge can you pack into the parameters of a language model?",
"year": "2020"
},
{
"authors": "N Muennighoff",
"journal": "",
"ref_id": "b41",
"title": "Crosslingual generalization through multitask finetuning",
"year": "2022"
},
{
"authors": "V Sanh",
"journal": "",
"ref_id": "b42",
"title": "Multitask prompted training enables zero-shot task generalization",
"year": "2021"
},
{
"authors": "T ; Le Scao",
"journal": "\\backslash%%STRING%%",
"ref_id": "b43",
"title": "What Language Model to Train if You Have One Million GPU Hours?",
"year": "2022"
},
{
"authors": "B Wang; A Komatsuzaki",
"journal": "",
"ref_id": "b44",
"title": "GPT-J-6B: A 6 Billion Parameter Autoregressive Language Model",
"year": "2021"
},
{
"authors": "Z Du",
"journal": "",
"ref_id": "b45",
"title": "GLM: General Language Model Pretraining with Autoregressive Blank Infilling",
"year": "2022"
},
{
"authors": "A Chen; G Stanovsky; S Singh; M Gardner",
"journal": "",
"ref_id": "b46",
"title": "Evaluating question answering evaluation",
"year": "2019"
},
{
"authors": "T Zhang; V Kishore; F Wu; K Q Weinberger; Y Artzi",
"journal": "",
"ref_id": "b47",
"title": "Bertscore: Evaluating text generation with bert",
"year": "2019"
},
{
"authors": "Z Yuan; H Yuan; C Tan; W Wang; S Huang; F Huang",
"journal": "",
"ref_id": "b48",
"title": "RRHF: Rank Responses to Align Language Models with Human Feedback without tears",
"year": "2023-04-11"
},
{
"authors": "C Xu",
"journal": "",
"ref_id": "b49",
"title": "WizardLM: Empowering Large Language Models to Follow Complex Instructions",
"year": "2023"
},
{
"authors": "S Mukherjee; A Mitra; G Jawahar; S Agarwal; H Palangi; A Awadallah",
"journal": "",
"ref_id": "b50",
"title": "Orca: Progressive Learning from Complex Explanation Traces of GPT-4",
"year": "2023"
},
{
"authors": "T Dettmers; A Pagnoni; A Holtzman; L Zettlemoyer",
"journal": "",
"ref_id": "b51",
"title": "Qlora: Efficient finetuning of quantized llms",
"year": "2023"
},
{
"authors": "E Frantar; S Ashkboos; T Hoefler; D Alistarh",
"journal": "",
"ref_id": "b52",
"title": "Gptq: Accurate post-training quantization for generative pre-trained transformers",
"year": "2022"
},
{
"authors": "L Ouyang",
"journal": "",
"ref_id": "b53",
"title": "Training language models to follow instructions with human feedback",
"year": "2022"
},
{
"authors": "H H Do; P W C Prasad; A Maag; A Alsadoon",
"journal": "Expert systems with applications",
"ref_id": "b54",
"title": "Deep learning for aspect-based sentiment analysis: a comparative review",
"year": "2019"
},
{
"authors": "H Lightman",
"journal": "",
"ref_id": "b55",
"title": "Let's Verify Step by Step",
"year": "2023-05-31"
}
] | [] |
|
10.18653/v1/2022.acl-long.569 | 2023-07-25 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b21",
"b34",
"b11"
],
"table_ref": [],
"text": "Accurate semantic understanding in language technologies is typically powered by distributional word representations and pre-trained language models (LMs). Due to their subsymbolic nature, however, such methods lack in explainability and interpretability, leading to insufficient trust in end users. An example application which requires capturing word meaning with its nuanced contextdetermined modulations is lexical semantic change analysis, a task which consists in detecting whether a word's meaning has changed over time, for example by acquiring or losing a sense. Modern semantic change detection systems rely on static and contextualised word representations, LMbased lexical replacement, grammatical profiles, supervised word sense and word-in-context disambiguation (Kutuzov et al., 2018;Tahmasebi et al., 2021). But the main potential end users of these technologies-historical linguists, lexicographers, and social scientists-are still somewhat reluctant to adopt them precisely because of their lack of explanatory power. Lexicographers, for instance, are not satisfied with detecting that a word has or hasn't changed its meaning over the last ten years; they want descriptions of old and new senses in humanreadable form, possibly accompanied by additional layers of explanation, e.g., specifying the type of semantic change (such as broadening, narrowing, and metaphorisation) the word has undergone.\nOur work is an attempt to bridge the gap between computational tools for semantic understanding and their users. We propose to replace blackbox contextualised token embeddings produced by large LMs with a new type of interpretable lexical semantic representation: automatically generated contextualised word definitions (Gardner et al., 2022). In this paradigm, the usage of the word 'apple' in the sentence 'She tasted a fresh green apple' is represented not with a dense highdimensional vector but with the context-dependent natural language definition 'EDIBLE FRUIT'. With an extended case study on lexical semantic change analysis, we show that moving to the more abstract meaning space of definitions allows practitioners to obtain explainable predictions from computational systems, while leading to superior performance on semantic change benchmarks compared to state-ofthe-art token-based approaches.\nThis paper makes the following contributions.1 \n1. We show that word definitions automatically generated with a specialised language model, fine-tuned for this purpose, can serve as interpretable representations for polysemous words ( §5). Pairwise usage similarities between contextualised definitions approximate human semantic similarity judgements better"
},
{
"figure_ref": [],
"heading": "Usage example",
"publication_ref": [],
"table_ref": [],
"text": "Target word Generated definition 'about half of the soldiers in our rifle platoons were draftees whom we had trained for about six weeks' draftee 'A PERSON WHO IS BEING ENLISTED IN THE ARMED FORCES' Table 1: An example of a definition generated by our fine-tuned Flan-T5 XL. The model is prompted with the usage example, post-fixed with the phrase 'What is the definition of draftee?'\nthan similarities between usage-based word and sentence embeddings.\n2. We present a method to obtain word sense representations by labelling data-driven clusters of word usages with sense definitions, and collect human judgements of definition quality to evaluate these representations ( §6). We find that sense labels produced by retrieving the most prototypical contextualised word definition within a group of usages consistently outperform labels produced by selecting the most prototypical token embedding.\n3. Using sense labels obtained via definition generation, we create maps that describe diachronic relations between the senses of a target word. We then demonstrate how these diachronic maps can be used to explain meaning changes observed in text corpora and to find inconsistencies in data-driven groupings of word usages within existing lexical semantic resources ( §7).\n2 Related Work"
},
{
"figure_ref": [],
"heading": "Definition Modelling",
"publication_ref": [
"b11",
"b4",
"b29",
"b10",
"b24",
"b28",
"b23",
"b4",
"b14",
"b2",
"b15",
"b14",
"b14",
"b24",
"b4",
"b19",
"b24"
],
"table_ref": [],
"text": "The task of generating human-readable word definitions, as found in dictionaries, is commonly referred to as definition modelling or definition generation (for a review, see Gardner et al., 2022).\nThe original motivation for this task has been the interpretation, analysis, and evaluation of word embedding spaces. Definition generation systems, however, also have practical applications in lexicography, language acquisition, sociolinguistics, and within NLP (Bevilacqua et al., 2020). The task was initially formulated as the generation of a natural language definition given an embedding-a single distributional representation-of the target word, or definiendum (Noraset et al., 2017). Word meaning, however, varies according to the context in which a word is used. This is particularly true for polysemous words, which can be defined in multiple, potentially very different ways depending on their context. The first formulation of definition modelling was therefore soon replaced by the task of generating a contextually appropriate word definition given a target word embedding and an example usage (Gadetsky et al., 2018;Mickus et al., 2022). When the end goal is not the evaluation of embedding spaces, generating definitions from vector representations is still not the most natural formulation of definition modelling. Ni and Wang (2017) and Mickus et al. (2019) treat the task as a sequence-to-sequence problem: given an input sequence with a highlighted word, generate a contextually appropriate definition. In the current work, we follow this approach. Table 1 shows an example of a contextualised definition generated by our model (see §4) for the English word 'draftee'.\nMethods Approaches that address this last formulation of the task are typically based on a pre-trained language model deployed on the definienda of interest in a natural language generation (NLG) setup (Bevilacqua et al., 2020). Generated definitions can be further improved by regulating their degree of specificity via specialised LM modules (Huang et al., 2021), by adjusting their level of complexity using contrastive learning training objectives (August et al., 2022), or by supplementing them with definitional sentences extracted directly from a domain-specific corpus (Huang et al., 2022). We will compare our results to the specificity-tuned T5-based text generator proposed by Huang et al. (2021).\nEvaluation Generated definitions are typically evaluated with standard NLG metrics such as BLEU, NIST, ROUGE-L, METEOR or Mover-Score (e.g., Huang et al., 2021;Mickus et al., 2022), using precision@k on a definition retrieval task (Bevilacqua et al., 2020), or measuring semantic similarity between sentence embeddings obtained for the reference and the generated definition (Kong et al., 2022). Because reference-based methods are inherently flawed (for a discussion, see Mickus et al., 2022), qualitative evaluation is almost always presented in combination with these quantitative metrics. In this paper, we evaluate generated definitions with automatic metrics and by collecting human judgements."
},
{
"figure_ref": [],
"heading": "Semantic Change Detection",
"publication_ref": [
"b35",
"b32",
"b20",
"b27",
"b13"
],
"table_ref": [],
"text": "Words in natural language change their meaning over time; these diachronic processes are of interest to both linguists and NLP practitioners. Lexical semantic change detection (LSCD) is nowadays a well represented NLP task, with workshops (Tahmasebi et al., 2022) and several shared tasks (e.g., Schlechtweg et al., 2020;Kurtyigit et al., 2021). LSCD is usually cast either as binary classification (whether the target word changed its meaning or not) or as a ranking task (ordering target words according to the degree of their change). To evaluate existing approaches, manually annotated datasets are used: so-called DWUGs are described below in §3.\nAn important issue with current LSCD methods is that they rarely describe change in terms of word senses, which are extremely important for linguists to understand diachronic meaning trajectories. Instead, systems provide (and are evaluated by) perword numerical 'change scores' which are hardly interpretable; at best, a binary 'sense gain' or 'sense loss' classification is used. Even approaches that do operate on the level of senses (e.g., Mitra et al., 2015;Homskiy and Arefyev, 2022) do not label them in a linguistically meaningful way, making it difficult to understand the relations between the resulting 'anonymous' types of word usage."
},
{
"figure_ref": [],
"heading": "Data",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Datasets of Definitions",
"publication_ref": [
"b17",
"b25"
],
"table_ref": [],
"text": "To train an NLG system that produces definitions ( §4), we use three datasets containing a humanwritten definition for each lexicographic sense of a target word, paired with a usage example. The WordNet dataset is a collection of word definitions and word usages extracted by Ishiwatari et al. (2019) from the WordNet lexical database (Miller, 1995). The Oxford dataset (also known as CHA in prior work) consists of definitions and usage ex- "
},
{
"figure_ref": [],
"heading": "Diachronic Word Usage Graphs",
"publication_ref": [
"b33",
"b3",
"b32",
"b9",
"b0"
],
"table_ref": [],
"text": "We showcase interpretable word usage ( §5) and sense representations ( §6 and 7) using a dataset where target lemmas are represented with diachronic word usage graphs (DWUGs, Schlechtweg et al., 2021). A DWUG is a weighted, undirected graph, where nodes represent target usages (word occurrences within a sentence or discourse context) and edge weights represent the semantic proximity of a pair of usages. DWUGs are the result of a multi-round incremental human annotation process, with annotators asked to judge the semantic relatedness of pairs of word usages on a 4-point scale. Based on these pairwise relatedness judgements, word usages are then grouped into usage clusters (a data-driven approximation of word senses) using a variation of correlation clustering (Bansal et al., 2004;Schlechtweg et al., 2020). DWUGs are currently available in seven languages. 4 In this paper, we use the English graphs, which consist of usage sentences sampled from the Clean Corpus of Historical American English (Davies, 2012;Alatrash et al., 2020) and belonging to two time periods: 1810-1860 and 1960-2010. There are 46 usage graphs for English, corresponding to 40 nouns and 6 verbs annotated by a total of 9 annotators. Each target lemma has received on average 189 judgements, 2 for each usage pair. Figure 1 shows an example of a DWUG, with colours denoting usage clusters (i.e., data-driven senses): the 'blue' and 'orange' clusters belong almost entirely to different time periods: a new sense of the word has emerged. We show how our approach helps explain such cases of semantic change in §7."
},
{
"figure_ref": [],
"heading": "Definition Generation",
"publication_ref": [
"b8",
"b30"
],
"table_ref": [
"tab_5",
"tab_15"
],
"text": "Our formulation of the definition generation task is as follows: given a target word w and an example usage s (i.e., a sentence containing an occurrence of w), generate a natural language definition d that is grammatical, fluent, and faithful to the meaning of the target word w as used in the example usage s. A definition generator is a language process that maps words and example usages to natural language definitions. As a generator, we use Flan-T5 (Chung et al., 2022), a version of the T5 encoder-decoder Transformer (Raffel et al., 2020) fine-tuned on 1.8K tasks phrased as instructions and collected from almost 500 NLP datasets. Flan-T5 is not trained specifically on definition generation but thanks to its massive multi-task instruction fine-tuning, the model exhibits strong generalisation to unseen tasks. Therefore, we expect it to produce high-quality definitions. We extensively test three variants of Flan-T5 of different size and compare them to vanilla T5 models (Table 4 and Table 12, Appendix C.2); based on our results, we recommend using the largest fine-tuned Flan-T5 model whenever possible.\nTo obtain definitions from Flan-T5, we use natural language prompts consisting of an example usage preceded or followed by a question or instruction. For example: 's What is the definition of w?'. The concatenated usage example and prompt are provided as input to Flan-T5, which conditionally generates definitions (Ta-ble 1 shows an example instance). 5 We choose greedy search with target word filtering as a simple, parameter-free decoding strategy. Stochastic decoding algorithms can be investigated in future work."
},
{
"figure_ref": [],
"heading": "Prompt selection",
"publication_ref": [],
"table_ref": [],
"text": "In preliminary experiments, we used the pre-trained Flan-T5 Base model (250M parameters) to select a definition generation prompt among 8 alternative verbalisations. Appending the question 'What is the definition of w?' to the usage example consistently yielded the best scores. 6 We use this prompt for all further experiments."
},
{
"figure_ref": [],
"heading": "Evaluating Generated Definitions",
"publication_ref": [
"b14",
"b1",
"b14"
],
"table_ref": [
"tab_3"
],
"text": "Before using its definitions to construct an interpretable semantic space-the main goal of this paper-we perform a series of experiments to validate Flan-T5 as a definition generator. We use the target lemmas and usage examples from the corpora of definitions presented in §3, conditionally generate definitions with Flan-T5, and then compare them to the gold definitions in the corpora using reference-based NLG evaluation metrics. We report SacreBLEU and ROUGE-L, which measure surface form overlap, as well as BERT-F1, which is sensitive to the reference and candidate's semantics. As mentioned in §2.1, reference-based metrics are not flawless, yet designing and validating a reference-free metric for the definition generation task is beyond the scope of this paper. We will later resort to correlations with human judgements and expert human evaluation to assess the quality of generated definitions.\nWe evaluate the Flan-T5 XL (3B parameters) in three generalisation tests: 1) in distribution, 2) hard domain shift, and 3) soft domain shift. 7 We use these tests to choose a model to be deployed in further experiments. For reference, we report the BLEU score of the definition generator by Huang et al. (2021); ROUGE-L and BERT-F1 are not reported in their paper.\nIn distribution We fine-tune Flan-T5 XL on one corpus of definitions at a time, and test it on a held-out set from that same corpus (except CoDWoE which does not provide train-test split). The quality of the definitions increases substantially with fine-tuning, in terms of both their lexical and semantic overlap with gold definitions (Table 3). We find significantly higher scores on Oxford, which may be due to the larger size of its training split and to the quality of the WordNet examples, which sometimes are not sufficiently informative (Almeman and Espinosa Anke, 2022). We consider the observed model performance sufficient for the purposes of our experiments, in particular in view of the higher efficiency of finetuned Flan-T5 with respect to the three-module system of Huang et al. (2021). We therefore use this model throughout the rest of our study. The Flan-T5 models fine-tuned for definition generation are publicly available through the Hugging Face model hub.8 "
},
{
"figure_ref": [],
"heading": "Hard domain shift",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Definitions are Interpretable Word Representations",
"publication_ref": [],
"table_ref": [],
"text": "We propose considering the abstract meaning space of definitions as a representational space for lexical meaning. Definitions fulfil important general desiderata for word representations: they are human-interpretable and they can be used for quantitative comparisons between word usages (i.e., by judging the distance between pairs of definition strings). We put the definition space to test by applying it to the task of semantic change analysis, which requires capturing word meaning at a finegrained level, distinguishing word senses based on usage contexts. We use our fine-tuned Flan-T5 models (XL and other sizes) to generate definitions for all usages of the 46 target words annotated in the English DWUGs (ca. 200 usages per word; see §3.2).9 These definitions (an example is provided in Table 1) specify a diachronic semantic space."
},
{
"figure_ref": [],
"heading": "Correlation with Human Judgements",
"publication_ref": [
"b31",
"b22",
"b12"
],
"table_ref": [
"tab_5"
],
"text": "We construct word usage graphs for each lemma in the English DWUGs: we take usages as nodes and assign weights to edges by measuring pairwise similarity between usage-dependent definitions. We compute the similarity between pairs of definitions using two overlap-based metrics, SacreBLEU and METEOR, as well as the cosine similarity between sentence-embedded definitions. We then compare our graphs against the gold DWUGs, where edges between usage pairs are weighted with human judgements of semantic similarity, by computing the Spearman's correlation between human similarity judgements and similarity scores obtained for pairs of generated definitions. We compare our results to DWUGs constructed based on two additional types of usage-based representations: sentence embeddings obtained directly for usage examples, and contextualised token embeddings. Sentence embeddings (for both definitions and usage examples) are SBERT representations (Reimers and Gurevych, 2019) extracted with mean-pooling from the last layer of a DistilRoBERTa LM finetuned for semantic similarity comparisons. 10 For tokens, we extract the last-layer representations of a RoBERTa-large model (Liu et al., 2019) which correspond to subtokens of the target word (following Giulianelli et al., 2020) and use mean-pooling to obtain a single vector. While we report string-overlap similarities for definitions, these are not defined for numerical vectors, and thus similarities for example sentences and tokens are obtained with cosine only.\nPairwise similarities between definitions approximate human similarity judgements far better than similarities between example sentence and word embeddings (Table 4). This indicates that definitions are a more accurate approximation of contextualised lexical meaning. The results also show that similarity between definitions is best captured 10 DistilRoBERTa (sentence-transformers/all-distilRoBERTa-v1) is the second best model as reported in the official S-BERT documentation at the time of publication (https://www.sbert.net/docs/ pretrained_models.html). For a negligible accuracy reduction, it captures longer context sizes and is ca. 50% smaller and faster than the model that ranks first. by their embeddings, rather than by overlap-based metrics like SacreBLEU and METEOR."
},
{
"figure_ref": [
"fig_0"
],
"heading": "Definition Embedding Space",
"publication_ref": [
"b7"
],
"table_ref": [
"tab_7",
"tab_17"
],
"text": "We now examine the definition embedding space (the high-dimensional semantic space defined by sentence-embedded definitions), to identify properties that make it more expressive than usage-based spaces. Figure 2 shows the T-SNE projections of the DistilRoBERTa embeddings of all lemmas in the English DWUGs, for the three types of representation presented earlier: generated definitions, tokens, and example sentences. 11 The definition spaces exhibit characteristics that are more similar to a token embedding space than an example sentence embedding space, with definitions of the same lemma represented by relatively close-knit clusters of definition embeddings. This suggests that definition embeddings, as expected, represent the meaning of a word in context (similar to token embeddings), rather than the meaning of the whole usage example sentence in which the target word occurs.\nFor each target word, we also measure (i) the variability in each embedding space and (ii) the inter-cluster and intra-cluster dispersion (Caliński and Harabasz, 1974) obtained when clustering each space using k-means. This allows us to quantitatively appreciate properties exhibited by datadriven usage clusters that are obtained from different representation types. To cluster the embedding spaces, we experiment with values of k ∈ [2, 25], and select the k which maximises the Silhouette score. Our results are summarised in Table 5. While, on average, token spaces exhibit higher inter-cluster dispersion (indicating better cluster separation), the clusters in the definition spaces have on average the lowest intra-cluster dispersion, indicating that they are more cohesive than the clusters in the token and example sentence spaces. These findings persist for the gold clusters determined by the English DWUGs (Table 14, Appendix G).\nIn sum, this analysis shows that definition embedding spaces are generally suitable to distinguish different types of word usage. In the next section, we will show how they can indeed be used to characterise word senses. "
},
{
"figure_ref": [
"fig_7"
],
"heading": "Labelling Word Senses With Definitions",
"publication_ref": [
"b29",
"b18"
],
"table_ref": [],
"text": "For generated definitions to be useful in practice, they need to be able to distinguish word senses.\nFor example (ignoring diachronic differences and singleton clusters), there are three main senses of the word 'word' in its DWUG, which we manually label as: (1) 'WORDS OF LANGUAGE', (2) 'A RUMOUR', and (3) 'AN OATH'. Manual inspection of the generated definitions indicates that they are indeed sense-aware:\n1. 'A communication, a message', 'The text of a book, play, movie', etc.\n2. 'Information passed on, usually by one person to another', 'communication by spoken or written communication', etc.\n3. 'An oath', 'a pronouncement', etc.\nBut let's again put ourselves in the shoes of a historical linguist. Sense clusters are now impractically represented with multitudes of contextualised definitions. Cluster (1) for 'word', e.g., features 190 usages, and one must read through all of them (otherwise there will be a chance of missing something) and generalise -all to formulate a definition that covers the whole sense cluster (a sense label). We now show how DWUGs can be automatically augmented with generated sense labels, vastly improving their usability.\nSelecting sense labels From n definitions, generated for n word usages belonging to the same DWUG cluster, we use the most prototypical one as the sense label-with the aim of reflecting the meaning of the majority of usages in the cluster. We represent all definitions with their sentence embeddings (cf. §5.1) and select as prototypical the definition whose embedding is most similar to the average of all embeddings in the cluster. Clusters with less than 3 usages are ignored as, for these, prototypicality is ill-defined. As a sanity check, these are the sense labels obtained by this method for the DWUG clusters of 'word'; they correspond well to the sense descriptions provided earlier. We compare these sense labels to labels obtained by generating a definition for the most prototypical usage (as judged by its token embedding), rather than taking the most prototypical definition, and we evaluate both types of senses labels using human judgements. Examples of labels can be found in Appendix D.\nHuman evaluation Five human annotators (fluent English speakers) were asked to evaluate the quality of sense labels for each cluster in the English DWUGs, 136 in total. Each cluster was accompanied by the target word, two labels (from definitions and from usages) and five example usages randomly sampled from the DWUG. The annotators could select one of six judgements to indicate overall quality of the labels and their relative ranking. After a reconciliation round, the categorical judgements were aggregated via majority voting. Krippendorff's α inter-rater agreement is 0.35 on the original data and 0.45 when the categories are reduced to four. Full guidelines and results are reported in Appendix E. 12 We find that our prototypicality-based sense labelling strategy is overall reliable. Only for 15% of the clusters, annotators indicate that neither 12 There exist no established procedures for the collection of human quality judgements of automatically generated word sense labels. The closest efforts we are aware of are those in Noraset et al. (2017), who ask annotators to rank definitions generated by two systems, providing as reference the gold dictionary definitions. In our case, (1) generations are for word senses rather than lemmas, (2) we are interested not only in rankings but also in judgements of 'sufficient quality', (3) dictionary definitions are not available for the DWUG senses; instead (4) we provide annotators with usage examples, which are crucial for informed judgements of sense definitions. of the labels is satisfactory (Figure 9). When comparing definition-based and usage-based labels, the former were found to be better in 31% of the cases, while the latter in only 7% (in the rest of the cases, the two methods are judged as equal). We also analysed how often the labels produced by each method were found to be acceptable. Definition-based labels were of sufficient quality in 80% of the instances, while for usage-based labels this is only true for 68% of the cases.\nIn sum, prototypical definitions reflect sense meanings better than definitions of prototypical usage examples. We believe this is because definitions are more abstract and robust to contextual noise (the same definition can be assigned to very different usages, if the underlying sense is similar). This approach takes the best of both worlds: the produced representations are data-driven, but at the same time they are human-readable and naturally explanatory. After all, 'senses are abstractions from clusters of corpus citations' (Kilgarriff, 1997). In the next section, we demonstrate how automatically generated definition-based sense labels can be used to explain semantic change observed in diachronic text corpora."
},
{
"figure_ref": [],
"heading": "Explaining Semantic Change with Sense Labels",
"publication_ref": [
"b27",
"b5"
],
"table_ref": [],
"text": "Word senses in DWUGs are collections of example usages and they are only labelled with numerical identifiers. This does not allow users to easily grasp the meaning trajectories of the words they are interested in studying. Using sense labels extracted from generated definitions, we can produce a fully human-readable sense dynamics map-i.e., an automatically annotated version of a DWUG which displays synchronic and diachronic relations between senses (e.g, senses transitioning one into another, splitting from another sense, or two senses merging into one). One can look at sense dynamics maps as reproducing the work of Mitra et al. (2015) on the modern technological level and, importantly, with human-readable sense definitions. Given a target word, its original DWUG, and its semi-automatic sense clusters, we start by assigning a definition label to each cluster, as described in §6. Then, we divide each cluster into two sub-clusters, corresponding to time periods 1 and 2 (for example, sub-cluster c 2 1 contains all usages from cluster 1 occurring in time period 2). 13We compute pairwise cosine similarities between the sentence embeddings of the labels (their 'definition embeddings'), thereby producing a fully connected graph where nodes are sub-clusters and edges are weighted with sense label similarities. Most edges have very low weight, but some sub-cluster pairs are unusually similar, hinting at a possible relation between the corresponding senses. We detect these outlier pairs by inspecting the distribution of pairwise similarities for values with z-score higher than 1 (similarities more than 1 standard deviation away from the mean similarity). Sub-cluster pairs connected with such edges form a sense dynamics map.\nAs an example, the noun 'record' has only one sense in time period 1 but it acquires two new senses in time period 2 (Figure 3; as before, we ignore clusters with less than 3 usages). The sense clusters defined by the DWUG are anonymous collection of usages, but with the assigned sense labels (also shown in Figure 3) they can be turned into a proto-explanation of the observed semantic shift: It becomes now clear that sense 2 stems from the older general sense 0 of 'record'-arguably representing a case of narrowing (Bloomfield, 1933)while the second new sense (1: 'THE HIGHEST"
},
{
"figure_ref": [],
"heading": "SCORE OR OTHER ACHIEVEMENT IN THE GAME')",
"publication_ref": [
"b33"
],
"table_ref": [],
"text": "is not related to the others and is thus independent. Sense dynamics maps can also help in tracing potentially incorrect or inconsistent clustering in DWUGs. For instance, if different sense clusters are assigned identical definition labels, then it is likely that both clusters correspond to the same sense and that the clustering is thus erroneous. Using our automatically produced sense dynamics maps, DWUGs can be improved and enriched (semi-)automatically.\nAn interesting case is 'ball' (see Appendix F for another example regarding the word 'chef ').\nclusters c 1 1 and c 2 1 have the same label. This is done for simplicity and because of data scarcity, but in the future we plan to experiment with time-dependent labels as well. We use two time periods as only two periods are available in Schlechtweg et al.'s English DWUGs (2021), but the same procedure can be executed on multi-period datasets.\nFigure 3: Diachronic word usage graphs for 'record' (Schlechtweg et al., 2021) with sense definitions generated using our proposed procedure ( §6). Left: time period 1 (1810-1860); right: time period 2 . Colours correspond to data-driven senses, as annotated in the original DWUGs.\nAlthough none of its sense labels are identical, its sense cluster c 0 is very close to cluster c 2 (similarity of 0.70), while c 2 is close to c 3 (similarity of 0.53); all three senses persist throughout both time periods, with sense 3 declining in frequency. The generated definitions for the 'ball' clusters are: 0: 'A SPHERE OR OTHER OBJECT USED AS THE OBJECT OF A HIT' (the largest cluster), 2: 'A ROUND SOLID PROJECTILE, SUCH AS IS USED IN SHOOTING', and 3: 'A BULLET'. This case demonstrates that similarity relations are not transitive: the similarity between c 0 and c 3 is only 0.50, below our outlier threshold value. This is in part caused by inconsistent DWUG clustering: while the majority of usages in c 1 2 are about firearm projectiles, c 2 2 contains mentions of golf balls and ball point pens. This shifts sense 2 from 'BULLET' to 'ROUND SOLID PROJECTILE', making it closer to sense 0 (general spheres) than it should be. Ideally, all the 'BULLET' usages from c 2 should have ended up in c 3 , with the rest joining the general sense 0.\nBesides suggesting fixes to the DWUG clustering, the observed non-transitivity also describes a potential (not necessarily diachronic) meaning trajectory of 'ball': from any spherical object, to spherical objects used as projectiles, and then to any projectiles (like bullets), independent of their form. Our generated sense labels and their similarities help users analyse this phenomenon in a considerably faster and easier way than by manually inspecting all examples for these senses."
},
{
"figure_ref": [],
"heading": "Conclusion and Future Work",
"publication_ref": [
"b4",
"b8",
"b5"
],
"table_ref": [],
"text": "In this paper, we propose to consider automatically generated contextualised word definitions as a type of lexical representation, similar to traditional word embeddings. While generated definitions have been already shown to be effective for word sense disambiguation (Bevilacqua et al., 2020), our study puts this into a broader perspective and demonstrates that modern language models like Flan-T5 (Chung et al., 2022) are sufficiently mature to produce robust and accurate definitions in a simple prompting setup. The generated definitions outperform traditional token embeddings in word-in-context similarity judgements while being naturally interpretable.\nWe apply definition-based lexical representations to semantic change analysis and show that our approach can be used to trace word sense dynamics over time. Operating in the space of humanreadable definitions makes such analyses much more interesting and actionable for linguists and lexicographers-who look for explanations, not numbers. At the same time, we believe the 'definitions as representations' paradigm can also be used for other NLP tasks in the area of lexical semantics, such as word sense induction, idiom detection, and metaphor interpretation.\nOur experiments with diachronic sense modelling are still preliminary and mostly qualitative. It is important to evaluate systematically how well our predictions correspond to the judgements of (expert) humans. Once further evidence is gathered, other promising applications include tracing cases of semantic narrowing or widening over time (Bloomfield, 1933) by analysing the variability of contextualised definitions in different time periods and by making cluster labels time-dependent. Both directions will require extensive human annotation, and we leave them for future work."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [],
"table_ref": [],
"text": "Data in this work is limited to the English diachronic word usage graphs (DWUGs). Our methods themselves are language-agnostic and we do not anticipate serious problems with adapting them to DWUGs in other languages (which already exist). At the same time, although Flan-T5 is a multilingual LM, we did not thoroughly evaluate its ability to generate definitions in languages other than English. Again, definition datasets in other languages do exist and technically it is trivial to fine-tune Flan-T5 on some or all of them. Generated definitions and mappings between definitions and word senses can contain all sorts of biases and stereotypes, stemming from the underlying language model. Filtering inappropriate character strings from the definitions can only help as much, and further research is needed to estimate possible threats.\nIn our experiments with Flan-T5, the aim was to investigate the principal possibility of using this LM for definition modelling. Although we did evaluate several different Flan-T5 variants, we leave it for the future work to investigate the impact of model size and other experimental variables (such as decoding algorithms).\nThe cases shown in §7 are hand-picked examples, demonstrating the potential of using generated definitions for explainable semantic change detection and improving LSCD datasets. In the future, we plan to conduct a more rigorous evaluation of different ways to build sense dynamics map. "
},
{
"figure_ref": [],
"heading": "B Prompt Selection",
"publication_ref": [
"b8"
],
"table_ref": [],
"text": "As briefly discussed in Section 4, in preliminary experiments, we use the pretrained Flan-T5 Base model (250M parameters; Chung et al., 2022) to select a definition generation prompt among 8 alternative verbalisations. These are a combination of four different instruction strings ('Define w', 'Define the word w', 'Give the definition of w', 'What is the definition of w?) and two ways of concatenating instructions to usage examples -i.e., either prepending them or appending them. Tables 8-11 show the results of our experiments. In the tables, the strings 'pre' and 'post' refer to the concatenation method (prepending or appending the instruction), the numbers 128, 256, and 512 refer to the maximum length of the usage examples provided to Flan-T5 (in sub-words), and 'filter' refers to the decoding strategy of always avoiding the target word (definiendum). "
},
{
"figure_ref": [],
"heading": "C Additional Results",
"publication_ref": [],
"table_ref": [
"tab_3"
],
"text": "C.1 Zero-Shot Evaluation of Flan-T5 (Task Shift)\nHere we directly evaluate Flan-T5 XL on the Word-Net and Oxford test sets, without any fine-tuning nor in-context learning. 14 Table 3 in the main paper shows low BLEU and ROUGE-L scores but rather high BERT-F1. Overall, the model does not exhibit consistent task understanding (e.g. it generates 'SKEPTICISM' as a definition for 'healthy' as exemplified in the phrase 'healthy skepticism').\nA qualitative inspection, however, reveals that the generated definitions can still be often informative (e.g., 'A WORKWEEK THAT IS LONGER THAN THE REGULAR WORKWEEK' is informative with respect to the meaning of 'overtime' although the ground truth definition is 'BEYOND THE REGULAR TIME'). The two surface-overlap metrics cannot capture this, but the relatively high BERT-F1 confirms that the semantic content of generations is largely appropriate. There are indeed also many good zero-shot definitions. For example 'INTENSE' for 'fervent' as in 'the fervent heat', or 'A CON-VERSATION' for 'discussion' in 'we had a good discussion'."
},
{
"figure_ref": [],
"heading": "C.2 Other Models and Model Variants",
"publication_ref": [],
"table_ref": [
"tab_15"
],
"text": "We evaluate T5 (base and XL) and Flan-T5 (base, large, and XL) under the same generalisation conditions presented for Flan T5 XL in the main paper 14 We only condition generation on the usage examples and the task prompt. We do not provide full instances (i.e., usage examples, task prompts, and definitions) in the context, as one would do in a few-shot setup.\n(Section 4.1) and above in Appendix C.1. Results for FlanT5-XL are reported in the main paper (Table 3); here, in Table 12, we report results for all models and model variants."
},
{
"figure_ref": [],
"heading": "C.3 Evaluation Cards",
"publication_ref": [
"b16"
],
"table_ref": [
"tab_16"
],
"text": "In Table 13, we provide an evaluation card to clarify the nature of the generalisation tests performed on definition generators. 15 In-distribution tests are not included as they do not include any shift between the training and test data distributions (Hupkes et al., 2022). We also register our work in the Gen-Bench evolving survey of generalisation in NLP. 16"
},
{
"figure_ref": [
"fig_8"
],
"heading": "D Additional Examples of Generated",
"publication_ref": [],
"table_ref": [],
"text": "Definitions and Sense Labels Some definitions generated by Flan-T5 XL manage to capture very subtle aspects of the contextual lexical meaning. In the following list, we give the usage and then the contextual definition of 'word':\n1. 'There are people out there who have never heard of the Father, Son and Holy Spirit, let alone the Word of God.': 'THE BIBLE' 2. 'Good News Bible Before the world was created, the Word already existed; he was with God, and he was the same as God.': '( CHRIS-TIANITY ) Interesting insights can be drawn from how the embeddings of the generated definitions are located in the vector space. Figure 8 shows PCA projections of definition embeddings for usages of the words 'chef ' and 'lass' from the English DWUG. Colours represent sense clusters provided in the DWUG, and the legend shows most prototypical definitions for each sense generated by our best system (singleton clusters are ignored). The large star for each sense corresponds to its sense label (as opposed to smaller stars corresponding to other definitions not chosen as the label).\nFor the word 'chef ', there are two sense clusters, for which an identical definition is chosen ('A COMMANDER'). This most probably means that these clusters should in fact be merged together, or that they are in the process of splitting (see also Section 7). These two senses are (not surprisingly) much closer to each other than to the definitions from the 'PROFESSIONAL COOK' sense. For the word 'lass', it is interesting how separate is a small bluish group of definitions in the bottom right corner of the plot, where the target form is actually 'lassi'. The fine-tuned Flan-T5-XL model defined this group as 'A COLD DRINK MADE FROM MILK CURDLED BY YOGURT', which is indeed what 'lassi' is (ignoring minor details). 'You are given a spreadsheet with four columns: Targets, Examples, System1 and System2. In every row, we have one target English word in the Targets column and five (or less) example usages of this word in the Examples column. Usages are simply sentences with at least one occurrence of the target word: one usage per line."
},
{
"figure_ref": [],
"heading": "E Human Evaluation Guidelines",
"publication_ref": [],
"table_ref": [],
"text": "Every row is supposed to contain usages where the target word is used in the same sense: this means that for ambiguous words, there will be multiple rows, each corresponding to a particular sense. This division into senses is not always 100% correct, but for the purposes of this annotation effort, we take it for granted. Note that the five example usages in each row are sampled randomly from a larger set of usages belonging to this sense.\nSystem1 and System2 are computational models which produce human-readable labels or definitions for each sense of a target word. They employ different approaches, and your task is to compare and evaluate the labels generated by these two systems. Note that in each row, the names 'System1' and 'System2' are randomly assigned to the actual generation systems.\nThe generated sense labels are supposed to be useful for historical linguists and lexicographers. Thus, they must be:\n1. Truthful: i.e., should reflect exactly the sense in which the target word is occurring in the example usages. Ideally, the label should be general enough to encompass all the usages from the current row, but also specific enough so as not to mix with other senses (for polysemantic target words).\n2. Fluent: i.e., feeling like natural English sentence or sentences, without grammar errors, utterances broken mid-word, etc\nYou have to fill in the Judgements column with one of six integer values:\n• 0: both systems are equally bad for this sense • 1: System 1 is better, but System 2 is also OK • 11: System 1 is better, and System 2 is bad • 2: System 2 is better, but System 1 is also OK • 22: System 2 is better, and System 1 is bad • 3: both systems are equally good for this sense Some rows are already pre-populated with the 3 judgement, because the sense labels generated by both systems are identical. We hypothesise that this most probably means that both labels are equally good. Please still have a look at these identical labels and change 3 to 0 in case you feel that in fact they are equally bad.'"
},
{
"figure_ref": [],
"heading": "F Sense Dynamics Maps",
"publication_ref": [],
"table_ref": [],
"text": "It is easy to find different sense clusters which are assigned identical definition labels. Usage examples from sense clusters c 2 and c 3 for the word 'chef ', to which our system assigned the same label: 'A COMMANDER': "
},
{
"figure_ref": [
"fig_11"
],
"heading": "G Clustering Embedding Spaces",
"publication_ref": [
"b33"
],
"table_ref": [
"tab_7",
"tab_17"
],
"text": "We constructed three types of embedding spaces; (i) contextualised token embeddings, (ii) sentence embeddings, and (ii) definition embeddings. We did so for two language models: RoBERTa-large and DistilRoBERTa. Since we cluster the embedding spaces for each target word individually, we obtain different optimal number of clusters for each target word. Table 5 displays the average results over all target words.\nWe observe that the optimal number of clusters K is substantially higher for the definition embedding spaces for both RoBERTa-large and Distil-RoBERTa. However, this is an artefact of the data: since some distinct usages yield identical definitions for a target word, the definition space oftentimes consist of less distinct data points, which greatly impacts the average silhouette scores. Future work should point out what clustering methods are most applicable to definition embedding spaces. Still, this decrease in data points confirms how the definition embedding space could represent usages at a higher level of abstraction, collapsing distinct usages into identical representations.\nFigure 11 displays the T-SNE projections of each of the three embedding spaces of RoBERTA-large. As for Distil-RoBERTa, the definition embedding space appears to have spacial properties that are more similar to contextualised token embedding spaces than to sentence embedding spaces: the definition embeddings are more separated than the sentence embeddings, and are cluttered in a similar manner as the token embeddings.\nTable 14 shows the average inter-and intracluster dispersion values of the clusters as labelled by the English DWUGs (Schlechtweg et al., 2021). These are calculated for the token, sentence and definition embeddings of both RoBERTa-large and Distil-RoBERTa. "
},
{
"figure_ref": [],
"heading": "Acknowledgements",
"publication_ref": [],
"table_ref": [],
"text": "This project has received funding from the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation programme (grant agreement No. 819455). The computations were performed on resources provided through Sigma2-the national research infrastructure provider for High-Performance Computing and large-scale data storage in Norway."
},
{
"figure_ref": [],
"heading": "Appendix A Preliminary Analysis of Usage Examples",
"publication_ref": [
"b33"
],
"table_ref": [],
"text": "In Section 3.1 of the main paper, we present three corpora of human-written definitions and report their main statistics in Table 2, including mean and standard deviation of usage example length. Because the length of usage examples has been shown to affect the quality of generated definitions (Almeman and Espinosa Anke, 2022), in a preliminary analysis, we compare the length distributions of usage examples in the corpora of definitions as well as in the English DWUGs (Schlechtweg et al., 2021). Figures 4567show the length distributions of the four datasets. We also measure the correlation between definition quality (BertScore, BLEU, NIST) and (i) the length of usage examples, (ii) the absolute position of the target word in the examples, and (iii) the target word's relative position in the examples. Tables 6 and7 show the correlation coefficients."
}
] | We propose using automatically generated natural language definitions of contextualised word usages as interpretable word and word sense representations. Given a collection of usage examples for a target word, and the corresponding data-driven usage clusters (i.e., word senses), a definition is generated for each usage with a specialised Flan-T5 language model, and the most prototypical definition in a usage cluster is chosen as the sense label. We demonstrate how the resulting sense labels can make existing approaches to semantic change analysis more interpretable, and how they can allow users-historical linguists, lexicographers, or social scientists-to explore and intuitively explain diachronic trajectories of word meaning. Semantic change analysis is only one of many possible applications of the 'definitions as representations' paradigm. Beyond being human-readable, contextualised definitions also outperform token or usage sentence embeddings in word-in-context semantic similarity judgements, making them a new promising type of lexical representation for NLP. | Interpretable Word Sense Representations via Definition Generation: The Case of Semantic Change Analysis | [
{
"figure_caption": "Figure 2 :2Figure 2: T-SNE projection of each embedding space, DistilRoBERTa model.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "A PROMISE, VOW OR STATEMENT'",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "•A novel sense 2 of 'record' in time period 2 ('A PHONOGRAPH OR GRAMOPHONE CYLIN-DER CONTAINING AN AUDIO RECORDING.') is probably an offshoot of a stable sense 0 present in both time periods ('A DOCUMENT OR OTHER MEANS OF PROVIDING INFORMA-TION ABOUT PAST EVENTS.').",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: Length distribution of usage examples in WordNet.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 :5Figure 5: Length distribution of usage examples in Oxford.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Figure 6 :6Figure 6: Length distribution of usage examples in CoD-WoE.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "Figure 7 :7Figure 7: Length distribution of usage examples in the English DWUGs.",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "7",
"figure_type": "figure"
},
{
"figure_caption": "Figures 99Figures 9 and 10 show the results of the human evaluation.'You are given a spreadsheet with four columns: Targets, Examples, System1 and System2. In every row, we have one target English word in the Targets column and five (or less) example usages of this word in the Examples column. Usages are simply sentences with at least one occurrence of the target word: one usage per line.Every row is supposed to contain usages where the target word is used in the same sense: this means that for ambiguous words, there will be multiple rows, each corresponding to a particular sense. This division into senses is not always 100% correct, but for the purposes of this annotation effort, we take it for granted. Note that the five example usages in each row are sampled randomly from a larger set of usages belonging to this sense.System1 and System2 are computational models which produce human-readable labels or definitions for each sense of a target word. They employ different approaches, and your task is to compare and evaluate the labels generated by these two systems. Note that in each row, the names 'System1' and 'System2' are randomly assigned to the actual generation systems.The generated sense labels are supposed to be useful for historical linguists and lexicographers. Thus, they must be:",
"figure_data": "",
"figure_id": "fig_7",
"figure_label": "9",
"figure_type": "figure"
},
{
"figure_caption": "Figure 8 :8Figure 8: PCA projections of definition embeddings for two target words from English DWUG.",
"figure_data": "",
"figure_id": "fig_8",
"figure_label": "8",
"figure_type": "figure"
},
{
"figure_caption": "Figure 9 :9Figure 9: General quality of generated sense labels",
"figure_data": "",
"figure_id": "fig_9",
"figure_label": "9",
"figure_type": "figure"
},
{
"figure_caption": "•c 2 : 'He boasted of having been a chef de brigade in the republican armies of France', 'Morrel has received a regiment, and Joliette is Chef d'Escadron of Spahis', 'as majorgeneral and chef d'escadron, during the pleasure of our glorious monarch Louis le Grand' • c 3 : 'That brave general added to his rank of chef de brigade that of adjutant general', 'I frequently saw Mehevi and several other chefs and warriors of note take part' Thus, a user can safely accept the suggestion of our system to consider these two clusters as one sense. Note that 'A COMMANDER' practically disappeared as a word sense in the 20th century, replaced by 'A PROFESSIONAL COOK, USUALLY IN A RESTAURANT'.",
"figure_data": "",
"figure_id": "fig_10",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 1111Figure 11: T-SNE projection of each embedding space, RoBERTa-Large model.",
"figure_data": "",
"figure_id": "fig_11",
"figure_label": "11",
"figure_type": "figure"
},
{
"figure_caption": "Diachronic word usage graph for the English word 'lass'(Schlechtweg et al., 2021). amples collected by Gadetsky et al. (2018) from the Oxford Dictionary. Definitions are written by experts and usage examples are in British English. The CoDWoE dataset (Mickus et al., 2022) is based on definitions and examples extracted from Wiktionary. 2 It is a multilingual corpus, of which we use the English portion. Table2reports the main statistics of these datasets. Further statistics, e.g., on the size of the different splits, are provided byHuang et al. (2021) as well as in Appendix A.3 ",
"figure_data": "",
"figure_id": "tab_1",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Results of the definition generation experiments.",
"figure_data": "",
"figure_id": "tab_3",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Correlations with pairwise similarity judgements by humans. 'FT' stands for 'fine-tuned model'.",
"figure_data": "MethodCosine SacreBLEU METEORToken embeddings0.141--Sentence embeddings0.114--Generated definitionsFlan-T5 XL Zero-shot0.1880.0410.083Flan-T5 XXL Zero-shot 0.2060.0450.092Flan-T5 base FT0.2210.0780.077Flan-T5 XL FT0.2640.1080.117",
"figure_id": "tab_5",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Variance, standard deviation, optimal K, silhouette score, separation score, cohesion score, and the separation-cohesion ratio for each embedding space; average over all target words.",
"figure_data": "",
"figure_id": "tab_7",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "Correlations between properties of the usage examples and the quality (BertScore, BLEU, NIST) of the definitions generated by Flan-T5 Base for WordNet. The prompt used is 'What is the definition of w?' (post). The maximum context size is set to 512.",
"figure_data": "Length Relative Position Absolute Position BertScoreBleuNistLength1.000000-0.1217930.575304 0.067180 0.076133 0.044873Relative Position -0.1217931.0000000.626032 0.052725 0.074697 0.062041Absolute Position 0.5753040.6260321.000000 0.128785 0.159078 0.110559BertScore0.0671800.0527250.128785 1.000000 0.121067 0.095343Bleu0.0761330.0746970.159078 0.121067 1.000000 0.821956Nist0.0448730.0620410.110559 0.095343 0.821956 1.000000Length Relative Position Absolute Position BertScoreBleuNistLength1.000000-0.0409480.615536 0.019844 0.039525 0.017253Relative Position -0.0409481.0000000.674509 0.046071 0.019940 0.023542Absolute Position 0.6155360.6745091.000000 0.029413 0.016901 0.006764BertScore0.0198440.0460710.029413 1.000000 0.283203 0.276626Bleu0.0395250.0199400.016901 0.283203 1.000000 0.687382Nist0.0172530.0235420.006764 0.276626 0.687382 1.000000",
"figure_id": "tab_8",
"figure_label": "6",
"figure_type": "table"
},
{
"figure_caption": "Correlations between properties of the usage examples and the quality (BertScore, BLEU, NIST) of the definitions generated by Flan-T5 Base for Oxford.",
"figure_data": "",
"figure_id": "tab_9",
"figure_label": "7",
"figure_type": "table"
},
{
"figure_caption": "Prompt selection results on WordNet (see description in Appendix B).",
"figure_data": "",
"figure_id": "tab_10",
"figure_label": "8",
"figure_type": "table"
},
{
"figure_caption": "Prompt selection results on Oxford (see description in Appendix B).",
"figure_data": "ConfigurationBLEUNIST BERTScorewhat is the definition of <trg>? post 128 0.1138 0.21370.8702give the definition of <trg> post 1280.0826 0.23890.8615what is the definition of <trg>? post 640.1033 0.19900.8595give the definition of <trg> post 640.0785 0.21940.8520",
"figure_id": "tab_12",
"figure_label": "9",
"figure_type": "table"
},
{
"figure_caption": "Prompt selection results on CoDWoE Complete (see description in Appendix B).",
"figure_data": "ConfigurationBLEUNIST BERTScoregive the definition of <trg>: pre 640.0680 0.15130.8461what is the definition of <trg>? post 64 0.1068 0.14640.8458give the definition of <trg> post 640.0654 0.16020.8374",
"figure_id": "tab_13",
"figure_label": "10",
"figure_type": "table"
},
{
"figure_caption": "Prompt selection results on CoDWoE Trial (see description in Appendix B).",
"figure_data": "WordNetOxford",
"figure_id": "tab_14",
"figure_label": "11",
"figure_type": "table"
},
{
"figure_caption": "Results of the definition generation experiments.",
"figure_data": "",
"figure_id": "tab_15",
"figure_label": "12",
"figure_type": "table"
},
{
"figure_caption": "Evaluation card for the generalisation tests performed on definition generators. The setups are: zero-shot (□), hard domain shift (△), and soft domain shift (⃝). In-distribution tests are not included as they do not include any shift between the training and test data distributions.",
"figure_data": "THE SECOND PERSON OF THETRINITY ; JE'3. 'It was in that basement that I learned the skillsnecessary to succeed in the difficult thespianworld-specifically, get up on stage, say my",
"figure_id": "tab_16",
"figure_label": "13",
"figure_type": "table"
},
{
"figure_caption": "ModelRepresentation Sep. ↑ Coh. ↓ Ratio ↑ Separation score, cohesion score, and separation-cohesion ratio for each embedding space; average over all target words from the English DWUGs.",
"figure_data": "Sentence0.017 0.013 1.248RoBERTa-large Token0.042 0.034 1.272Definitions0.008 0.006 1.349Sentence0.665 0.592 1.126DistilRoBERTa Token0.591 0.477 1.258Definitions0.705 0.509 1.397",
"figure_id": "tab_17",
"figure_label": "14",
"figure_type": "table"
}
] | Mario Giulianelli; Iris Luden; Raquel Fernández; Andrey Kutuzov | [
{
"authors": "Reem Alatrash; Dominik Schlechtweg; Jonas Kuhn; Sabine Schulte Im Walde",
"journal": "European Language Resources Association",
"ref_id": "b0",
"title": "CCOHA: Clean corpus of historical American English",
"year": "2020"
},
{
"authors": "Fatemah Almeman; Luis Espinosa; Anke ",
"journal": "Association for Computational Linguistics",
"ref_id": "b1",
"title": "Putting WordNet's dictionary examples in the context of definition modelling: An empirical analysis",
"year": "2022"
},
{
"authors": "Tal August; Katharina Reinecke; Noah A Smith",
"journal": "Association for Computational Linguistics",
"ref_id": "b2",
"title": "Generating scientific definitions with controllable complexity",
"year": "2022"
},
{
"authors": "Nikhil Bansal; Avrim Blum; Shuchi Chawla",
"journal": "Machine Learning",
"ref_id": "b3",
"title": "Correlation clustering",
"year": "2004"
},
{
"authors": "Michele Bevilacqua; Marco Maru; Roberto Navigli",
"journal": "Association for Computational Linguistics",
"ref_id": "b4",
"title": "Generationary or \"how we went beyond word sense inventories and learned to gloss",
"year": "2020"
},
{
"authors": "Leonard Bloomfield",
"journal": "",
"ref_id": "b5",
"title": "Language",
"year": "1933"
},
{
"authors": "Unwin Allen",
"journal": "",
"ref_id": "b6",
"title": "",
"year": ""
},
{
"authors": "Tadeusz Caliński; Jerzy Harabasz",
"journal": "Communications in Statistics -Theory and Methods",
"ref_id": "b7",
"title": "A dendrite method for cluster analysis",
"year": "1974"
},
{
"authors": "Chung Hyung Won; Le Hou; Shayne Longpre; Barret Zoph; Yi Tay; William Fedus; Eric Li; Xuezhi Wang; Mostafa Dehghani; Siddhartha Brahma",
"journal": "",
"ref_id": "b8",
"title": "Scaling instruction-finetuned language models",
"year": "2022"
},
{
"authors": "Mark Davies",
"journal": "Corpora",
"ref_id": "b9",
"title": "Expanding horizons in historical linguistics with the 400-million word Corpus of Historical American English",
"year": "2012"
},
{
"authors": "Artyom Gadetsky; Ilya Yakubovskiy; Dmitry Vetrov",
"journal": "Association for Computational Linguistics",
"ref_id": "b10",
"title": "Conditional generators of words definitions",
"year": "2018"
},
{
"authors": "Noah Gardner; Hafiz Khan; Chih-Cheng Hung",
"journal": "Applied Computing and Intelligence",
"ref_id": "b11",
"title": "Definition modeling: Literature review and dataset analysis",
"year": "2022"
},
{
"authors": "Mario Giulianelli; Marco Del Tredici; Raquel Fernández",
"journal": "Association for Computational Linguistics",
"ref_id": "b12",
"title": "Analysing lexical semantic change with contextualised word representations",
"year": "2020"
},
{
"authors": "Daniil Homskiy; Nikolay Arefyev",
"journal": "Association for Computational Linguistics",
"ref_id": "b13",
"title": "DeepMistake at LSCDiscovery: Can a multilingual word-incontext model replace human annotators?",
"year": "2022"
},
{
"authors": "Han Huang; Tomoyuki Kajiwara; Yuki Arase",
"journal": "Association for Computational Linguistics",
"ref_id": "b14",
"title": "Definition modelling for appropriate specificity",
"year": "2021"
},
{
"authors": "Jie Huang; Hanyin Shao; Kevin Chen-Chuan; Jinjun Chang; Wen-Mei Xiong; Hwu",
"journal": "Association for Computational Linguistics",
"ref_id": "b15",
"title": "Understanding jargon: Combining extraction and generation for definition modeling",
"year": "2022"
},
{
"authors": "Dieuwke Hupkes; Mario Giulianelli; Verna Dankers; Mikel Artetxe; Yanai Elazar; Tiago Pimentel; Christos Christodoulopoulos; Karim Lasri; Naomi Saphra; Arabella Sinclair",
"journal": "",
"ref_id": "b16",
"title": "State-of-the-art generalisation research in NLP: A taxonomy and review",
"year": "2022"
},
{
"authors": "Shonosuke Ishiwatari; Hiroaki Hayashi; Naoki Yoshinaga; Graham Neubig; Shoetsu Sato; Masashi Toyoda; Masaru Kitsuregawa",
"journal": "Association for Computational Linguistics",
"ref_id": "b17",
"title": "Learning to describe unknown phrases with local and global contexts",
"year": "2019"
},
{
"authors": "Adam Kilgarriff",
"journal": "Computers and the Humanities",
"ref_id": "b18",
"title": "I don't believe in word senses",
"year": "1997"
},
{
"authors": "Cunliang Kong; Yun Chen; Hengyuan Zhang; Liner Yang; Erhong Yang",
"journal": "Association for Computational Linguistics",
"ref_id": "b19",
"title": "Multitasking framework for unsupervised simple definition generation",
"year": "2022"
},
{
"authors": "Sinan Kurtyigit; Maike Park; Dominik Schlechtweg; Jonas Kuhn; Sabine Schulte Im Walde",
"journal": "Association for Computational Linguistics",
"ref_id": "b20",
"title": "Lexical semantic change discovery",
"year": "2021"
},
{
"authors": "Andrey Kutuzov; Lilja Øvrelid; Terrence Szymanski; Erik Velldal",
"journal": "Association for Computational Linguistics",
"ref_id": "b21",
"title": "Diachronic word embeddings and semantic shifts: a survey",
"year": "2018"
},
{
"authors": "Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov",
"journal": "",
"ref_id": "b22",
"title": "RoBERTa: A robustly optimized BERT pretraining approach",
"year": "2019"
},
{
"authors": "Timothee Mickus; Denis Paperno; Matthieu Constant",
"journal": "Linköping University Electronic Press",
"ref_id": "b23",
"title": "Mark my word: A sequence-to-sequence approach to definition modeling",
"year": "2019"
},
{
"authors": "Timothee Mickus; Kees Van Deemter; Mathieu Constant; Denis Paperno",
"journal": "Association for Computational Linguistics",
"ref_id": "b24",
"title": "Semeval-2022 task 1: CODWOE -comparing dictionaries and word embeddings",
"year": "2022"
},
{
"authors": "George A Miller",
"journal": "Communications of the ACM",
"ref_id": "b25",
"title": "WordNet: A lexical database for English",
"year": "1995"
},
{
"authors": "George A Miller; Claudia Leacock; Randee Tengi; Ross T Bunker",
"journal": "",
"ref_id": "b26",
"title": "A semantic concordance",
"year": "1993-03-21"
},
{
"authors": "Sunny Mitra; Ritwik Mitra; Kalyan Suman; Martin Maity; Chris Riedl; Pawan Biemann; Animesh Goyal; Mukherjee",
"journal": "Natural Language Engineering",
"ref_id": "b27",
"title": "An automatic approach to identify word sense changes in text media across timescales",
"year": "2015"
},
{
"authors": "Ke Ni; William Yang; Wang ",
"journal": "Asian Federation of Natural Language Processing",
"ref_id": "b28",
"title": "Learning to explain non-standard English words and phrases",
"year": "2017"
},
{
"authors": "Thanapon Noraset; Chen Liang; Larry Birnbaum; Doug Downey",
"journal": "",
"ref_id": "b29",
"title": "Definition modeling: Learning to define word embeddings in natural language",
"year": "2017"
},
{
"authors": "Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu",
"journal": "Journal of Machine Learning Research",
"ref_id": "b30",
"title": "Exploring the limits of transfer learning with a unified text-to-text transformer",
"year": "2020"
},
{
"authors": "Nils Reimers; Iryna Gurevych",
"journal": "Association for Computational Linguistics",
"ref_id": "b31",
"title": "Sentence-BERT: Sentence embeddings using Siamese BERTnetworks",
"year": "2019"
},
{
"authors": "Dominik Schlechtweg; Barbara Mcgillivray; Simon Hengchen; Haim Dubossarsky; Nina Tahmasebi",
"journal": "International Committee for Computational Linguistics",
"ref_id": "b32",
"title": "SemEval-2020 task 1: Unsupervised lexical semantic change detection",
"year": "2020"
},
{
"authors": "Dominik Schlechtweg; Nina Tahmasebi; Simon Hengchen; Haim Dubossarsky; Barbara Mcgillivray",
"journal": "Association for Computational Linguistics",
"ref_id": "b33",
"title": "DWUG: A large resource of diachronic word usage graphs in four languages",
"year": "2021"
},
{
"authors": "Nina Tahmasebi; Lars Borin; Adam Jatowt",
"journal": "Computational approaches to semantic change",
"ref_id": "b34",
"title": "Survey of computational approaches to lexical semantic change detection",
"year": "2021"
},
{
"authors": "Nina Tahmasebi",
"journal": "Association for Computational Linguistics",
"ref_id": "b35",
"title": "Proceedings of the 3rd Workshop on Computational Approaches to Historical Language Change",
"year": "2022"
}
] | [] |
2023-05-19 | [
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b14",
"b26",
"b45",
"b0"
],
"table_ref": [],
"text": "LS improves the readability of any given text with the aim of helping vocabulary and literacy development. LS achieves this by replacing complex words in a sentence with simpler alternatives. LS returns a simplified sentence which can be passed to a TS system for further syntactic and grammatical simplification. The replaced complex words are those words which a general or targeted population found to be hard to read, interpret, or understand. Previous LS systems have been designed to simplify complex words for children, second language learners, individuals with reading disabilities or low-literacy (Paetzold and Specia, 2017b). LS therefore provides both developers and users with a degree of personalization that is unattainable through seq2seq or generative TS systems (Yeung and Lee, 2018;Lee and Yeung, 2018a).\nDeep learning, and latterly, LLM and prompt learning, have revolutionized the way we approach many NLP tasks, including LS. Previous LS systems have relied upon lexicons, rule-based, statistical, n-gram, and word embedding models to identify and then simplify complex words (Paetzold and Specia, 2017b). These approaches would identify a complex word, for example, \"bombardment\" as being in need of simplification and would suggest \"attack\" as a suitable alternative (Figure 1), hereby referred to as a candidate substitution.\nState-of-the-art deep learning models, such as BERT (Devlin et al., 2019), RoBERTa (Liu et al., 2019), GPT-3 (Brown et al., 2020), and others, automatically generate, select, and rank candidate substitutions with performances superior to traditional approaches. These include relying on pre-existing lexicons, simplification rules, or engineered features (Saggion et al., 2022). There have been no surveys published on deep learning approaches for LS. The paper by Paetzold and Specia (2017b) is the most recent survey on LS but it precedes studies that demonstrate the headway made by state-of-theart deep learning approaches. A broad comprehensive survey on TS was published in 2021 (Al-Thanyyan and Azmi, 2021). However, this survey likewise does not cover recent advances in the field nor does it focus specifically on LS. This paper therefore continues pre-existing literature by providing an updated survey of the latest deep learning approaches for LS and its sub-tasks of substitute generation (SG), selection (SS), and ranking (SR)."
},
{
"figure_ref": [],
"heading": "Pipeline",
"publication_ref": [],
"table_ref": [],
"text": "We structure this survey around the main components of the LS pipeline: SG, SS, and SR (Section 3). We also provide an overview of recent datasets (Section 4), and discuss open challenges in LS (Section 5.1). Normally, an LS pipeline starts with complex word identification (CWI). However, since it is often considered as a standalone precursor, we refer the reader to North et al. (2022b), for a detailed survey on CWI methods. Substitute Generation SG returns a number: k, of candidates substitutions that are suitable replacements for a previously identified complex word. Usually, an LS system will generate candidate substitution in the range of k = [1, 3, 5, or 10] with top-k referring to the most appropriate candidates. These candidate substitutions need to be more simple, hence easier to read, interpret, or understand than the original complex word. The candidate substitutions also need to preserve the original complex word's meaning, especially in its provided context.\nSubstitute Selection SS filters the generated topk candidate substitutions and removes those which are not suitable. For instance, candidate substitutions which are not synonymous to the original complex word or that are more complex are often removed.\nSubstitute Ranking SR orders the remaining top-k candidate substitutions from the most to the least appropriate simplification. The original complex word is then replaced with the most suitable candidate substitution."
},
{
"figure_ref": [],
"heading": "Evaluation Metrics",
"publication_ref": [
"b45"
],
"table_ref": [],
"text": "All sub-tasks of the LS pipeline are evaluated using precision, accuracy, recall, and F1-score. Several additional metrics have also been used: potential, mean average precision (MAP), and accuracy at top-k. Potential is the ratio of predicted candidate substitutions for which at least one of the top-k candidate substitutions generated was among the gold labels (Saggion et al., 2022). MAP evaluates whether the returned top-k candidate substitutions match the gold labels as well as whether they have the same positional rank. Accuracy at top-k = [1, 2, or 3] is the ratio of instances where at least one of the candidate substitutions at k is among the gold labels."
},
{
"figure_ref": [],
"heading": "Deep Learning Approaches",
"publication_ref": [],
"table_ref": [],
"text": "Prior to deep learning approaches, lexicon, rulebased, statistical, n-gram, and word embedding models were state-of-the-art for SG, SS, and SR. As previously mentioned, Paetzold and Specia (2017b) have provided a comprehensive survey detailing these approaches, their performances, as well as their impact on LS literature. The following sections provide an extension of the work carried out by Paetzold and Specia (2017b). We introduce new deep learning approaches for LS and begin our survey of the LS pipeline at the SG phase. The recent developments in the CWI step of the pipeline have been extensively surveyed by North et al. (2022b)."
},
{
"figure_ref": [],
"heading": "Substitute Generation",
"publication_ref": [
"b31",
"b52",
"b8",
"b46",
"b57",
"b5",
"b5",
"b40",
"b46",
"b5",
"b46"
],
"table_ref": [],
"text": "In 2017, word embedding models were state-ofthe-art for SG. Word embedding models, such as Word2Vec (Mikolov et al., 2013), were used alongside more traditional approaches, such as querying a lexicon, or generating candidate substitutions based on certain rules (Paetzold and Specia, 2017b). Word embedding models conducted SG by converting potential candidate substitutions into vectors, hence word embeddings, and then calculating which of these vectors had the highest cosine similarity, or lowest cosine distance, with the vector of the target complex word. These vectors were then converted back into their word forms and were considered the top-k candidate substitutions.\nWord Embeddings + LLMs Post 2017, word embedding models continued to be implemented for SG. However, they were now combined with the word embeddings produced by LLMs or by a LLM's prediction scores. Alarcón et al. (2021a) experimented with various word embeddings models for generating Spanish candidate substitutions. They used word embeddings models, such as Word2Vec, Sense2Vec (Trask et al., 2015), and FastText (Bojanowski et al., 2016), along with the pre-trained LLM BERT, to generate these word embeddings. It was discovered that a more traditional approach that produced candidate substitutions by querying a pre-existing lexicon outperformed these word embedding models in terms of both potential and recall yet slightly under-performed these word embedding models in regards to precision.\nThe traditional approach achieved a potential of 0.898, a recall of 0.597, and a precision of 0.043 on the EASIER dataset (Alarcón et al., 2021b (Sense2Vec), on the other hand, attained a potential, recall, and precision score of 0.506, 0.282, and 0.056, respectively. Surprisingly, this went against the assumption that word embedding models would have achieved a superior performance given their state-of-the-art reputation demonstrated by Paetzold and Specia (2017a). During error analysis, it was found that these word embeddings models often produced antonyms of the target complex word as potential candidate substitutions. This is due to how word embedding models calculate word similarity between vectors. Seneviratne et al. (2022) used a word embedding model and a pre-trained LLM: XLNet (Yang et al., 2019), to produce an embedding similarity score and a prediction score for SG. They followed a similar approach conducted by Arefyev et al. (2020). Arefyev et al. (2020) utilized context2vec (Melamud et al., 2016) and ELMo (Peters et al., 2018) to encode the context of the target complex word to gain a probability distribution of each word belonging to that particular context. They then used this probability distribution to estimate the likelihood, or appropriateness, of a potential candidate substitution replacing the target complex word. This score was used alongside a LLM prediction score from either BERT, RoBERTa, or XLNet, to produce a final list of top-k candidate substitutions. Both Seneviratne et al. (2022) and Arefyev et al. (2020) discovered that their combined approach of using a word embedding model alongside a pretrained LLM prediction score failed to surpass the performance of using a single pre-trained LLM. For instance, Seneviratne et al. (2022) was outperformed by North et al. (2022a) on the TSAR-2022 dataset."
},
{
"figure_ref": [],
"heading": "Masked Language Modeling",
"publication_ref": [
"b42",
"b45",
"b19",
"b19",
"b45",
"b17",
"b45",
"b17",
"b10",
"b13",
"b16",
"b17",
"b12",
"b11",
"b50",
"b55",
"b56",
"b42",
"b54",
"b19",
"b53",
"b6",
"b6",
"b45"
],
"table_ref": [],
"text": "The introduction of pre-trained LLMs, also saw the arrival of Masked Language Modeling (MLM) for SG. Przy-była and Shardlow (2020) used LLMs trained on a MLM objective for multi-word LS, whereas Qiang et al. (2020) were the first to use MLM for Spanish SG. MLM has subsequently become a popular approach to SG. 7 out of the 11 system reports submitted to TSAR-2022 (Saggion et al., 2022), described their approach as consisting of a MLM objective.\nKnown as LSBert, the model introduced by Qiang et al. ( 2020), used the pre-trained LLM BERT. Sentences were taken from the LS datasets LexMTurk (Horn et al., 2014), BenchLS (Paetzold and Specia, 2016b), and NNSeval (Paetzold and Specia, 2016c). Two versions of each sentence were then concatenated, being separated by the [SEP] special token. They were then fed into the LLM. The first sentence was identical to that extracted from the datasets, whereas the second sentence had its complex word replaced with the [MASK] special token. The LLM then attempted to predict the word replaced by the [MASK] special token by taking into consideration its left and right context as well as the prior original sentence. In this way, LLMs provide candidate substitutions with the highest probability (highest prediction score) of fitting into the surrounding context and that are also similar to the target complex word in the original sentence. For the top-k=1 candidate substitution, LSBert achieved F1-scores for SG of 0.259, 0.272, and 0.218 on the three datasets LexMTurk (Horn et al., 2014), BenchLS (Paetzold and Specia, 2016b), and NNSeval (Paetzold and Specia, 2016c) respectively. These performances surpassed that of all prior approaches (Paetzold and Specia, 2017b). The previous highest F1-score was achieved by a word-embedding model (Paetzold and Specia, 2017a), which produced F1-scores of 0.195, 0.236, and 0.218 for each dataset, respectively.\nBefore the release of the TSAR-2022 shared-task (Saggion et al., 2022), Ferres and Saggion (2022) introduced a new dataset: ALEXSIS (TSAR-2022 ES), that would later make up (along with an additional English and Portuguese dataset) the TSAR-2022 dataset (Saggion et al., 2022). Using their Spanish dataset, they experimented with a number of monolingual LLMs pre-trained on either Spanish data as well as several multilingual LLMs, such as mBERT and RoBERTa. Ferres and Saggion (2022) adopted the MLM approach used by LS-Bert. They experimented with the Spanish LLMs: BETO (Cañete et al., 2020), BERTIN (De la Rosa and Fernández, 2022), RoBERTa-base-BNE, and RoBERTA-large-BNE (Fandiño et al., 2022) for SG. They discovered that their largest pre-trained Spanish LLM: RoBERTA-large-BNE, achieved the greatest SG performance after having also removed candidate substitutions equal to the complex word, regardless of capitalization or accentuation and being less than 2 characters long. North et al. (2022a) was inspired by the success of the monolingual LLMs shown by Ferres and Saggion (2022). They likewise tested a range of LLMs for SG with a MLM objective, including multilingual LLLMs: mBERT, and XLM-R (Conneau et al., 2020), and several monolingual LLMs, including Electra for English (Clark et al., 2020), RoBERTAlarge-BNE for Spanish, and BERTimbau (Souza et al., 2020) for Portuguese. Their monolingual LLMs scored an acc@1 score of 0.517, 0.353, and 0.481 on the English, Spanish, and Portuguese TSAR-2022 datasets respectively. Whistely et al. (2022) also experimented with similar monolingual LLMs for SG. They used BERT for English, BETO for Spanish, and BERTimbau for Portuguese. Interestingly, their models' performances were lower compared to that of North et al. (2022a), despite their Portuguese LS system consisting of the same language model. Whistely et al. ( 2022) achieved acc@1 scores of 0.378, 0.250, and 0.3074 for English, Spanish, and Portuguese, respectively. This is likely due to the additional SS and SR steps implemented by Whistely et al. (2022) and the lack thereof shown within the LS system provided by North et al. (2022a) (Section 3.2). Wilkens et al. (2022) also used a range of monolingual LLMs for SG. However, they used an ensemble of BERT-like models with three different masking strategies: 1). copy, 2). query expansion, and 3). paraphrase. The copy strategy replicated that of LSBert (Qiang et al., 2020), whereby two sentences were inputted into a LLM concatenated with the [SEP] special token. The first sentence being an unaltered version of the original sentence, and the second sentence having its complex word masked. The query expansion strategy used Fast-Text to generate five related words with the highest cosine similarity to the target complex word. For iteration 2a). of the query expansion strategy, the first sentence was the original unaltered sentence, the second sentence replaced the complex word with one of the suggested similar words produced by FastText, and sentence 3 was the masked sentence. Iteration 2b). of this strategy was the same as iteration 2a)., however, sentence 2 now consisted of all five suggested words. Lastly, the paraphrase strategy generated 10 new contexts for each complex word composed of paraphrases of the original sentence. These new contexts were limited to 512 tokens. The ensembles used for these three masking strategies consisted of BERT and RoBERTa LLMs for English, several BETO LLMs for Spanish, and several BERTimbau LLMs for Portuguese. The paraphrase strategy showed the worst performance with a joint MAP/Potential@1 score of 0.217, whereas the query expansion strategy obtained a MAP/Potential@1 score of 0.528, 0.477, and 0.476 for English, Spanish, and Portuguese, respectively. This surpassed the performance of the paraphrase strategy and the original copy strategy used by LSBert, regardless of the LLMs used.\nPrompt Learning Prompt learning has also been used for SG and is currently state-of-the-art (Table 3). Prompt learning involves feeding into a LLM input that is presented in such a way as to provide a description of the task as well as to return a desired output. PromptLS is an example of prompt learning applied to SG. Created by Vásquez-Rodríguez et al. (2022), PromptLS consisted of a variety of pre-trained LLMs fine-tuned on several LS datasets. These fined-tuned LLMs were then presented with four combinations of prompts: a). \"a easier word for bombardment is\", b). \"a simple word for bombardment is\", c). \"a easier synonym for bombardment is\", and lastly, d). \"a simple synonym for bombardment is\". These prompt combinations were supplied to a RoBERTa LLM on all of the English data extracted from the LexMTurk (Horn et al., 2014), BenchLS (Paetzold and Specia, 2016b), NN-Seval (Paetzold and Specia, 2016c), and CERF-LS (Uchida et al., 2018) LS datasets. They were also translated and fed into BERTIN fine-tuned on the Spanish data obtained from EASIER, along with BR-BERTo fine-tuned on all of the Portuguese data taken from SIMPLEX-PB (Hartmann and Aluísio, 2020). Vásquez-Rodríguez et al. ( 2022) also used these prompts on a zero-shot condition. It was discovered that the fine-tuned LLMs outperformed the zero-shot models on all conditions by an average increase in performance between 0.3 to 0.4 across all metrics: acc@1, acc@3, MAP@3, and Precision@3. The prompt combinations that produced the best candidate substitutions were \"easier word\" for English, \"palabra simple\" and \"palabra fácil\" for Spanish, and \"palavra simples\" and \"sinônimo simples\" for Portuguese.\nPrompt learning has likewise been applied to causal language models for SG, such as GPT-3. Aumiller and Gertz (2022) experimented with a variety of different prompts, which they fed into a GPT-3. These prompts were of four types: 1). zero-shot with context, 2). single-shot with context, two-shot with context, 3). zero-shot without context, and 4). single-shot without context. The size of each shot: n, refers to how many times a prompt is inputted into GPT-3. For instance, those shots with context would input a given sentence and then ask the question, \"Given the above context, list ten alternative words for <complex word> that are easier to understand.\", n number of times. Those without context, however, would input n times the following:\"Give me ten simplified synonyms for the following word: <complex word>\". Aumiller and Gertz (2022) also combined all types of prompts in an ensemble, generating candidate substitutions from each prompt type and then deciding upon final candidate substations through plurality voting and additional SS and SR steps (Section 3.2). Their ensemble approach outperformed all other prompt types and SG models submitted to TSAR-2022 (Saggion et al., 2022) (Table 3)."
},
{
"figure_ref": [],
"heading": "Substitute Selection and Ranking",
"publication_ref": [
"b45",
"b27",
"b27"
],
"table_ref": [],
"text": "Traditional approaches to SS are still implemented post SG. Methods such as POS-tag and antonym filtering, semantic or sentence thresholds have been used to remove inappropriate candidate substitutions after having been generating from the above deep learning approaches (Saggion et al., 2022). Nevertheless, the majority of modern deep learning approaches have minimal SS, with SS often being simultaneously conducted during SG or SR. For instance, the metric used to generate the top-k can-didate substitutions, by it either similarity between word embeddings, or a pre-train LLM's prediction score, tends not to suggest candidate substitutions that are deemed as being inappropriate by other SS methods. Likewise, SR techniques that rank candidate substitutions in order of their appropriateness will in turn move inappropriate simplifications further down the list of top-k candidate substitutions to the point that they are no longer considered.\nWord Embeddings Word embedding models continued to be used for SS without LLMs, regardless of the arrival of pre-trained LLMs, such as BERT. For instance, Song et al. ( 2020) created a unique LS system that filtered candidate substitutions by applying a semantic similarity threshold, matching only those candidate substitutions with the same POS tag as the target complex word, calculating contextual relevance, being a measure of how reasonable and fluent a sentence is after the complex word had been replaced, and by using cosine similarity between word embeddings to rank candidate substitutions. They generated word embeddings by Word2Vec and evaluated their model's performance on the LS-2007 dataset (Mc-Carthy and Navigli, 2007). It was found that the use of Word2Vec improved their model's performance having achieved an acc@1 of 0.269. Their second highest performing model, without the use of Word2Vec embeddings, produced an acc@1 of 0.218. Maddela and Xu (2018) created the neural readability ranker (NNR) for SR. Consisting of a feature extraction, a Gaussianbased feature vectorization layer, and a task specific output node, NNR is a deep learning algorithm capable of ranking candidate substitutions based on their perceived complexity. It performances regression, whereby having been trained on the Word Complexity Lexicon (WCL), as well as several features and character n-grams converted into Gaussian vectors, it is able to provide a value between 0 and 1 corresponding to the complexity of any given word. It achieves this by conducting pairwise aggregation. For each pair of potential candidate substitutions, the model predicts a value that defines which candidate substitution is more or less complex than the other. A return positive value indicates that the first candidate substitution is more complex than the second, whereas a negative value dictates that the second candidate substitution is more complex than the first. This is applied to all combinations of candidate substitutions given a complex word. Each candidate substitution is then ranked in accordance to its comparative complexity with all other potential candidate substitutions. Maddela and Xu (2018) applied their NNR model to the LS-2012 dataset and outperformed prior word embedding techniques for SR. They achieved an Prec@1 of 0.673, whereas the previous state-of-the-art model provided by Paetzold and Specia (2017a) achieved an Prec@1 of 0.656."
},
{
"figure_ref": [],
"heading": "Neural Regression",
"publication_ref": [
"b46",
"b25",
"b25",
"b45",
"b54",
"b54",
"b6"
],
"table_ref": [],
"text": "Word Embeddings + LLMs One of the most common approaches to SS and SR involves the use of word embeddings and LLMs. Seneviratne et al. (2022) filtered and ranked top-k=20 candidate substitutions based on the same combined score that they used for SG. It consisted of their MLM model's prediction score of the generated candidate together with the inner product of the target word's embedding and the embedding of the potential candidate substitution. These top-k=20 candidate substitutions were then subject to one of three additional ranking metrics. The first ranking metric (CILex_1) ranked candidate substitutions on their cosine similarity between the original sentence and a copy of the original sentence with the candidate substitution in place of its complex word. The second and third ranking metrics made use of dictionary definitions of the target complex word and its candidate substitutions. They calculated the cosine similarity between each embedding of each definition and the embedding of the sentence of the target complex word. Those with the highest cosine similarities between a). the definition of the target complex word and the definition of the candidate substitution (CILex_2), or b). the definition of the target complex word and the word embedding of the original sentence with the candidate substitution in place of its complex word (CILex_3), were used to determine the rank of each candidate substitution. They discovered that all three metrics produced similar performances on the TSAR-2022 dataset with CILex 1, 2, and 3 achieving acc@1 scores of 0.375, 0.380, and 0.386, respectively. Li et al. (2022) used a set of features taken from LSBert combined with what they referred to as an equivalence score. Equivalence score was created to gauge semantic similarity between candidate substitution and complex word to an extent that was more expressive than the cosine similarity between word embeddings. To obtain this equivalence score, they used a pre-trained RoBERTa LLM trained for natural language inference (NLI) which predicts the likelihood of one sentence entailing another. The model was trained on a multi-genre corpus with a MLM objective. The product of the returned likelihood of the original sentence with the candidate substitution preceding the original sentence and vice-versa equated to the equivalence score. Since Li et al. (2022) used the same method of SG as LSBert, having only changed their LLM to RoBERTa, they concluded that their system's superior performance was a consequence of its unique SR. They achieved an acc@1 of 0.659, whereas LSBert attained an acc@1 of 0.598 on the English TSAR-2022 dataset (Saggion et al., 2022).\nAleksandrova and Brochu Dufour (2022) ranked candidate substitutions on three metrics: a). grammaticality, b). meaning preservation, and c). simplicity. Grammaticality was calculated by firstly determining whether the candidate substitution had the same POS tag in terms of person, number, mood, tense, and so forth. Those that matched on all POS-tag categories were assigned the value of 1 or 0 if at least one category did not match. Preservation was determined by using BERTScore to generate cosine similarities between the embeddings of the original sentence and the embeddings of the original sentence, having replaced the target complex word with the candidate substitution. Lastly, preservation was obtained by using a CEFR vocabulary classifier trained on data from the English Vocabulary Profile (EVP). The data used to train the CEFR classifier was first masked and fed into a pre-trained LLM: BERT. The outputted encodings were then used to train an SVM model resulting in their CEFR classifier. Their model failed to surpass the baseline LSBert models at TSAR-2022 in terms of acc@1, having achieved a score of 0.544. MLM Prediction Scores LS systems have also relied entirely on MLM prediction scores for SS and SR. North et al. (2022a) and Vásquez-Rodríguez et al. (2022) adopt this approach. They have no additional SR steps and rank their candidate substitutions per their generated MLM prediction scores. They do, however, apply some basic filtering with both studies removing duplicates as well as candidate substitutions equal to the complex word. Surprisingly, minimal SR has been shown to surpass other more technical approaches (Table 3). North et al. (2022a) has achieved state-of-the-art performance on the TSAR-2022 Portuguese dataset, whereas Vásquez-Rodríguez et al. (2022) has consistently produced high performances across the English and Spanish TSAR-2022 datasets. Only GPT-3 based-models have surpassed these performances (Aumiller and Gertz, 2022) (Table 3)."
},
{
"figure_ref": [],
"heading": "Resources",
"publication_ref": [],
"table_ref": [],
"text": "Post 2017 LS datasets have been created for either all sub-tasks within the LS pipeline or for a specific purpose (Appendix, Table 2). Recent international competitions (shared-tasks) have also provided their own LS datasets (*). LS resources are available for multiple languages, predominately English (EN), Spanish (ES), Portuguese (PT), French (FR), Japanese (JP), and Chinese (ZH)."
},
{
"figure_ref": [],
"heading": "English",
"publication_ref": [
"b27",
"b48"
],
"table_ref": [],
"text": "Personalized-LS Lee and Yeung (2018b) constructed a dataset of 12,000 English words for personalized LS. These words were ranked on a fivepoint Likert scale. 15 native Japanese speakers were tasked with rating the complexity of each word. These complexity rating were then applied to BenchLS, in turn personalizing the dataset for Japanese speakers.\nWCL Maddela and Xu (2018) introduced the Word Complexity Lexicon (WCL). The WCL is a dataset made up of 15,000 English words annotated with complexity ratings. Annotators were 11 nonnative English speakers using a six-point Likert scale.\nLCP-2021* The dataset provided at the LCP-2021 shared-task (CompLex) (Shardlow et al., 2020), was developed using crowd sourcing. 10,800 complex words in context were selected from three corpora covering the Bible, biomedical articles, and European Parliamentary proceedings. Their lexical complexities were annotated using a 5-point Likert scale."
},
{
"figure_ref": [],
"heading": "SimpleText-2021* The",
"publication_ref": [
"b15",
"b45"
],
"table_ref": [],
"text": "SimpleText-2021 shared-task (Ermakova et al., 2021) introduced three pilot tasks: 1). to select passages to be simplified, 2). to identify complex concepts within these passages, and 3). to simplify these complex concepts to generate an easier to understand passage. They provided their participants with two sources of data, these being the Citation Network Dataset, DBLP+Citation, ACM Citation network, together with titles extracted from The Guardian newspaper with manually annotated keywords.\nTSAR-2022* TSAR-2022 (Saggion et al., 2022) supplied datasets in English, Spanish, and Portuguese. These datasets contained target words in contexts taken from journalistic texts and Wikipedia articles, along with 10 candidate substitutions (approx. 20 in raw data) provided by crowdsourced annotators located in the UK, Spain, and Brazil. The candidate substitutions were ranked per their suggestion frequency. The English, Spanish, and Portuguese datasets contained 386, 381, and 386 instances, respectively."
},
{
"figure_ref": [],
"heading": "Datasets in Other Languages",
"publication_ref": [
"b30",
"b22",
"b7",
"b44",
"b20",
"b21",
"b32",
"b43"
],
"table_ref": [],
"text": "Spanish The ALexS-2020 shared-task (Zambrano and Ráez, 2020) included a Spanish dataset consisting of 723 complex words from recorded transcripts. Merejildo (2021) provided the Spanish CWI corpus (ES-CWI). A group of 40 nativespeaking Spanish annotators identified complex words within 3,887 academic texts. The EASIER corpus (Alarcón et al., 2021b) contains 5,310 Spanish complex words in contexts taken from newspapers with 7,892 candidate substitutions. A small version of the corpus is also provided with 500 instances (EASIER-500).\nPortuguese The PorSimples dataset (Aluísio and Gasperin, 2010) consists of extracts taken from Brazilian newspapers. The dataset is divided into nine sub-corpora separated by degree of simplification and source text. The PorSimplesSent dataset (Leal et al., 2018) was adapted from the previous PorSimples dataset. It contains strong and natural simplifications of PorSimples's original sentences. SIMPLEX-PB (Hartmann and Aluísio, 2020) provides a selection of features for each of its candidate substitutions.\nFrench ReSyf contains French synonyms that have been ranked in regards to their reading difficulty using a SVM (Billami et al., 2018). It consists of 57,589 instances with a total of 148,648 candidate substitutions. FrenchLys is a LS tool designed by Rolin et al. (2021). It provides its own dataset that contains sentences sampled from a French TS dataset: ALECTOR, and french schoolbooks. Substitute candidates were provided by 20 French speaking annotators.\nJapanese The Japanese Lexical Substitution (JLS) dataset (Kajiwara and Yamamoto, 2015) con-tains 243 target words, each with 10 contexts (2,430 instances in total). Crowd-sourced annotators provided and ranked candidate substitutions. The JLS Balanced Dataset (Kodaira et al., 2016) expanded the previous JLS dataset to make it more representative of different genres and contains 2,010 generalized instances. Nishihara and Kajiwara (2020) created a new dataset (JWCL & JSSL) that increased the Japanese Education Vocabulary List (JEV). It houses 18,000 Japanese words divided into three levels of difficulty: easy, medium, or difficult.\nChinese Personalized-ZH (Lee and Yeung, 2018a) consists of 600 Chinese words. Each word's complexity was ranked by eight learners of Chinese on a 5-point lickert-scale. HanLS was constructed by Qiang et al. (2021). It contains 534 Chinese complex words. 5 native-speaking annotators gave and ranked candidate substitutions. Each complex word has on average 8 candidate substitutions."
},
{
"figure_ref": [],
"heading": "Discussion and Conclusion",
"publication_ref": [
"b25"
],
"table_ref": [],
"text": "Since the 2017 survey on LS (Paetzold and Specia, 2017b), deep learning approaches have provided new headway within the field. MLM is now the go to method for SG, with the majority of recent LS studies having employed a MLM objective. The casual language model: GPT-3, surpasses the performance of all other approaches when subjected to prompt learning, especially when an ensemble of prompts are taken into consideration (Table 3). The prediction scores of MLM or casual language modeling have replaced various SS and SR techniques. LS systems that employ minimal SS and no SR apart from ranking their LLM's prediction scores, have outperformed more technical, featureoriented, and unsupervised ranking methods (Table 3). However, an exception is made with regards to equivalence score (Li et al., 2022), which has been shown to be effective at SR.\nFuture LS systems will make use of new advances in deep learning. We believe prompt learning and models, such as GPT-3, will become increasingly popular, given their state-of-the-art performance at SG. Using an ensemble of various prompts for SS and SR may advance LS performance. In addition, the creation of new metrics similar to equivalence score will likewise be beneficial."
},
{
"figure_ref": [],
"heading": "Open Challenges in LS",
"publication_ref": [],
"table_ref": [],
"text": "LS has a number of open research areas that are either unaddressed, or the current body of work is inconclusive. In this brief section, we conclude this survey by outlining a few key areas for future development of LS research.\nEvaluation: The metrics we use to evaluate LS are not perfect (Section 2.1). Automated metrics that condense a wide problem into a single numerical score can harm outcomes with human participants. Development of more faithful resources, as well as direct evaluation with intended user groups of simplification systems is a fruitful avenue for future work. This can be done by taking into consideration variation in data annotation instead of labels produced by aggregating unique annotations as in most datasets currently available.\nExplainability: Lexical simplifications are inherently more explainable than sentence simplification as the operations are directly applied at the lexeme level. However, the decision process on whether to simplify and which word to choose is increasingly hidden behind the black-box of a model. Work to explain and interpret these decisions will allow researchers to better understand the opportunities and threats of applying modern NLP techniques to LS research.\nPersonalization: One model does not fit all. The simplification needs of a language learner compared to a stroke victim, compared to a child are each very different. Modeling these needs and using them to personalize LS systems will allow for personalized simplification output more adequate the needs of particular user groups.\nPerspectivism: Even within a population of common characteristics, each individual will bring a unique perspective on what and how to simplify. Systems which can alter their outputs to each user's needs will provide adaptive simplifications that go beyond our current technology. This will, in turn, improve the evaluation of LS models as previously discussed in this section.\nIntegration: LS is only one part of the wider simplification puzzle. Integrating LS systems with explanation generation, redundancy removal, and sentence splitting will further accelerate the adoption of automated simplification practices beyond the halls of research allowing such technology to reach a wider audience."
}
] | Lexical Simplification (LS) is the task of replacing complex for simpler words in a sentence whilst preserving the sentence's original meaning. LS is the lexical component of Text Simplification (TS) with the aim of making texts more accessible to various target populations. A past survey (Paetzold and Specia, 2017b) has provided a detailed overview of LS. Since this survey, however, the AI/NLP community has been taken by storm by recent advances in deep learning, particularly with the introduction of large language models (LLM) and prompt learning. The high performance of these models sparked renewed interest in LS. To reflect these recent advances, we present a comprehensive survey of papers published between 2017 and 2023 on LS and its sub-tasks with a special focus on deep learning. We also present benchmark datasets for the future development of LS systems. | Deep Learning Approaches to Lexical Simplification: A Survey | [
{
"figure_caption": "Figure 1 :1Figure 1: LS Pipeline. SG, SS, and SR are the main components of LS.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "The top 3 deep learning approaches across the TSAR-2022 datasets. Best performances in bold.",
"figure_data": ").",
"figure_id": "tab_1",
"figure_label": "1",
"figure_type": "table"
}
] | Kai North; Tharindu Ranasinghe; Matthew Shardlow; Marcos Zampieri | [
{
"authors": "S Suha; Aqil M Al-Thanyyan; Azmi",
"journal": "ACM Comput. Surv",
"ref_id": "b0",
"title": "Automated Text Simplification: A Survey",
"year": "2021"
},
{
"authors": "Rodrigo Alarcón; Lourdes Moreno; Paloma Martínez",
"journal": "",
"ref_id": "b1",
"title": "Exploration of Spanish Word Embeddings for Lexical Simplification",
"year": "2021"
},
{
"authors": "Rodrigo Alarcón; Lourdes Moreno; Paloma Martínez",
"journal": "IEEE Access",
"ref_id": "b2",
"title": "Lexical Simplification System to Improve Web Accessibility",
"year": "2021"
},
{
"authors": "Desislava Aleksandrova; Olivier Brochu Dufour",
"journal": "",
"ref_id": "b3",
"title": "RCML at TSAR-2022 Shared Task: Lexical Simplification With Modular Substitution Candidate Ranking",
"year": "2022"
},
{
"authors": "Sandra Maria; Aluísio ; Caroline Gasperin",
"journal": "",
"ref_id": "b4",
"title": "Fostering digital inclusion and accessibility: The porsimples project for simplification of portuguese texts",
"year": "2010"
},
{
"authors": "Nikolay Arefyev; Boris Sheludko; Alexander Podolskiy; Alexander Panchenko",
"journal": "",
"ref_id": "b5",
"title": "Always Keep your Target in Mind: Studying Semantics and Improving Performance of Neural Lexical Substitution",
"year": "2020"
},
{
"authors": "Dennis Aumiller; Michael Gertz",
"journal": "",
"ref_id": "b6",
"title": "UniHD at TSAR-2022 Shared Task: Is Compute All We Need for Lexical Simplification",
"year": "2022"
},
{
"authors": "B Mokhtar; Thomas Billami; Núria François; Gala",
"journal": "",
"ref_id": "b7",
"title": "ReSyf: a French lexicon with ranked synonyms",
"year": "2018"
},
{
"authors": "Piotr Bojanowski; Edouard Grave; Armand Joulin; Tomas Mikolov",
"journal": "",
"ref_id": "b8",
"title": "Enriching Word Vectors with Subword Information",
"year": "2016"
},
{
"authors": "B Tom; Benjamin Brown; Nick Mann; Melanie Ryder; Jared Subbiah; Prafulla Kaplan; Dhariwal; Others",
"journal": "",
"ref_id": "b9",
"title": "Language Models Are Few-Shot Learners",
"year": "2020"
},
{
"authors": "José Cañete; Gabriel Chaperon; Rodrigo Fuentes; Jou-Hui Ho; Hojin Kang; Jorge Pérez",
"journal": "",
"ref_id": "b10",
"title": "Spanish pre-trained bert model and evaluation data",
"year": "2020"
},
{
"authors": "Kevin Clark; Minh-Thang Luong; Quoc Le; Christopher Manning",
"journal": "",
"ref_id": "b11",
"title": "ELECTRA: Pretraining text encoders as discriminators rather than generators",
"year": "2020"
},
{
"authors": "Alexis Conneau; Kartikay Khandelwal; Naman Goyal; Vishrav Chaudhary; Others",
"journal": "",
"ref_id": "b12",
"title": "Unsupervised cross-lingual representation learning at scale",
"year": "2020"
},
{
"authors": "Javier De; La Rosa; Andres Fernández",
"journal": "",
"ref_id": "b13",
"title": "Zeroshot reading comprehension and reasoning for spanish with BERTIN GPT-J-6B",
"year": "2022"
},
{
"authors": "Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova",
"journal": "",
"ref_id": "b14",
"title": "Bert: Pre-training of deep bidirectional transformers for language understanding",
"year": "2019"
},
{
"authors": "Liana Ermakova; Patrice Bellot; Pavel Braslavski; Jaap Kamps; Josiane Mothe; Others",
"journal": "",
"ref_id": "b15",
"title": "Overview of SimpleText CLEF 2021 Workshop and Pilot Tasks",
"year": "2021"
},
{
"authors": "Gutiérrez Asier; Jordi Fandiño; Marc Armengol Estapé; Joan Llop Pàmies; Joaquin Silveira Palao; Ocampo; Others",
"journal": "Procesamiento del Lenguaje Natural",
"ref_id": "b16",
"title": "Maria: Spanish language models",
"year": "2022"
},
{
"authors": "Daniel Ferres; Horacio Saggion",
"journal": "",
"ref_id": "b17",
"title": "ALEXSIS: A dataset for lexical simplification in Spanish",
"year": "2022"
},
{
"authors": "Nathan Siegle; Hartmann ; Sandra Maria; Aluísio ",
"journal": "Linguamática",
"ref_id": "b18",
"title": "Adaptação lexical automática em textos informativos do português brasileiro para o ensino fundamental",
"year": "2020"
},
{
"authors": "Colby Horn; Cathryn Manduca; David Kauchak",
"journal": "",
"ref_id": "b19",
"title": "Learning a lexical simplifier using Wikipedia",
"year": "2014"
},
{
"authors": "Tomoyuki Kajiwara; Kazuhide Yamamoto",
"journal": "",
"ref_id": "b20",
"title": "Evaluation Dataset and System for Japanese Lexical Simplification",
"year": "2015"
},
{
"authors": "Tomonori Kodaira; Tomoyuki Kajiwara; Mamoru Komachi",
"journal": "",
"ref_id": "b21",
"title": "Controlled and Balanced Dataset for Japanese Lexical Simplification",
"year": "2016"
},
{
"authors": "Sidney Evaldo Leal; Magali Sanches Duran; Sandra Maria Aluísio",
"journal": "",
"ref_id": "b22",
"title": "A nontrivial sentence corpus for the task of sentence readability assessment in Portuguese",
"year": "2018"
},
{
"authors": "John Lee; Chak Yan; Yeung ",
"journal": "",
"ref_id": "b23",
"title": "a. Automatic prediction of vocabulary knowledge for learners of chinese as a foreign language",
"year": "2018"
},
{
"authors": "John Lee; Chak Yan; Yeung ",
"journal": "",
"ref_id": "b24",
"title": "Personalizing lexical simplification",
"year": "2018"
},
{
"authors": "Xiaofei Li; Daniel Wiechmann; Yu Qiao; Elma Kerz",
"journal": "",
"ref_id": "b25",
"title": "MANTIS at TSAR-2022 Shared Task: Improved Unsupervised Lexical Simplification with Pretrained Encoders",
"year": "2022"
},
{
"authors": "Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Others",
"journal": "",
"ref_id": "b26",
"title": "Roberta: A robustly optimized bert pretraining approach",
"year": "2019"
},
{
"authors": "Mounica Maddela; Wei Xu",
"journal": "",
"ref_id": "b27",
"title": "A wordcomplexity lexicon and a neural readability ranking model for lexical simplification",
"year": "2018"
},
{
"authors": "Diana Mccarthy; Roberto Navigli",
"journal": "",
"ref_id": "b28",
"title": "SemEval-2007 Task 10: English Lexical Substitution Task",
"year": "2007"
},
{
"authors": "Oren Melamud; Jacob Goldberger; Ido Dagan",
"journal": "",
"ref_id": "b29",
"title": "context2vec: Learning Generic Context Embedding with Bidirectional LSTM",
"year": "2016"
},
{
"authors": "Borbor Merejildo",
"journal": "",
"ref_id": "b30",
"title": "Creación de un corpus de textos universitarios en español para la identificación de palabras complejas en el área de la simplificación léxica",
"year": "2021"
},
{
"authors": "Tomas Mikolov; Kai Chen; Greg Corrado; Jeffrey Dean",
"journal": "",
"ref_id": "b31",
"title": "Efficient Estimation of word Representations in Vector Space",
"year": "2013"
},
{
"authors": "Daiki Nishihara; Tomoyuki Kajiwara",
"journal": "",
"ref_id": "b32",
"title": "Word Complexity Estimation for Japanese Lexical Simplification",
"year": "2020"
},
{
"authors": "Kai North; Alphaeus Dmonte; Tharindu Ranasinghe; Marcos Zampieri",
"journal": "",
"ref_id": "b33",
"title": "GMU-WLV at TSAR-2022 Shared Task: Evaluating Lexical Simplification Models",
"year": "2022"
},
{
"authors": "Kai North; Marcos Zampieri; Matthew Shardlow",
"journal": "ACM Computing Surveys",
"ref_id": "b34",
"title": "Lexical Complexity Prediction: An Overview",
"year": "2022"
},
{
"authors": "Gustavo Paetzold; Lucia Specia",
"journal": "",
"ref_id": "b35",
"title": "SemEval 2016 Task 11: Complex Word Identification",
"year": "2016"
},
{
"authors": "Gustavo Paetzold; Lucia Specia",
"journal": "",
"ref_id": "b36",
"title": "Lexical simplification with neural ranking",
"year": "2017"
},
{
"authors": "Gustavo H Paetzold; Lucia Specia",
"journal": "J. Artif. Int. Res",
"ref_id": "b37",
"title": "A Survey on Lexical Simplification",
"year": "2017"
},
{
"authors": "Gustavo Henrique; Paetzold ; Lucia Specia",
"journal": "",
"ref_id": "b38",
"title": "Benchmarking Lexical Simplification Systems",
"year": "2016"
},
{
"authors": "Gustavo Henrique; Paetzold ; Lucia Specia",
"journal": "",
"ref_id": "b39",
"title": "Unsupervised lexical simplification for non-native speakers",
"year": "2016"
},
{
"authors": "Matthew E Peters; Mark Neumann; Mohit Iyyer; Matt Gardner; Christopher Clark; Kenton Lee; Luke Zettlemoyer",
"journal": "",
"ref_id": "b40",
"title": "Deep Contextualized Word Representations",
"year": "2018"
},
{
"authors": "Piotr Przybyła; Matthew Shardlow",
"journal": "",
"ref_id": "b41",
"title": "Multi-Word Lexical Simplification",
"year": "2020"
},
{
"authors": "Jipeng Qiang; Yun Li; Zhu Yi; Yunhao Yuan; Xindong Wu",
"journal": "",
"ref_id": "b42",
"title": "Lexical simplification with pretrained encoders",
"year": "2020"
},
{
"authors": "Jipeng Qiang; Xinyu Lu; Yun Li; Yunhao Yuan; Xindong Wu",
"journal": "IEEE/ACM Transactions on Audio, Speech, and Language Processing",
"ref_id": "b43",
"title": "Chinese Lexical Simplification",
"year": "2021"
},
{
"authors": "Eva Rolin; Quentin Langlois; Patrick Watrin; Thomas François",
"journal": "",
"ref_id": "b44",
"title": "FrenLyS: A Tool for the Automatic Simplification of French General Language Texts",
"year": "2021"
},
{
"authors": "Horacio Saggion; Sanja Štajner; Daniel Ferrés; Kim Cheng Sheang; Matthew Shardlow; Kai North; Marcos Zampieri",
"journal": "",
"ref_id": "b45",
"title": "Findings of the TSAR-2022 Shared Task on Multilingual Lexical Simplification",
"year": "2022"
},
{
"authors": "Sandaru Seneviratne; Elena Daskalaki; Hanna Suominen",
"journal": "",
"ref_id": "b46",
"title": "CILS at TSAR-2022 Shared Task: Investigating the Applicability of Lexical Substitution Methods for Lexical Simplification",
"year": "2022"
},
{
"authors": "Matthew Shardlow",
"journal": "",
"ref_id": "b47",
"title": "The CW Corpus: A New Resource for Evaluating the Identification of Complex Words",
"year": "2013"
},
{
"authors": "Matthew Shardlow; Michael Cooper; Marcos Zampieri",
"journal": "",
"ref_id": "b48",
"title": "CompLex -a new corpus for lexical complexity prediction from Likert Scale data",
"year": "2020"
},
{
"authors": "Jiayin Song; Jingyue Hu; Leung-Pun Wong; Lap-Kei Lee; Tianyong Hao",
"journal": "",
"ref_id": "b49",
"title": "A New Context-Aware Method Based on Hybrid Ranking for Community-Oriented Lexical Simplification",
"year": "2020"
},
{
"authors": "Fábio Souza; Rodrigo Nogueira; Roberto Lotufo",
"journal": "",
"ref_id": "b50",
"title": "BERTimbau: pretrained BERT models for Brazilian Portuguese",
"year": "2020"
},
{
"authors": "Lucia Specia; Kumar Jauhar; Rada Sujay; Mihalcea",
"journal": "",
"ref_id": "b51",
"title": "Semeval -2012 task 1: English lexical simplification",
"year": "2012"
},
{
"authors": "Andrew Trask; Phil Michalak; John Liu",
"journal": "",
"ref_id": "b52",
"title": "sense2vec -A Fast and Accurate Method for Word Sense Disambiguation In Neural Word Embeddings",
"year": "2015"
},
{
"authors": "Satoru Uchida; Shohei Takada; Yuki Arase",
"journal": "",
"ref_id": "b53",
"title": "CEFR-based Lexical Simplification Dataset",
"year": "2018"
},
{
"authors": "Laura Vásquez-Rodríguez; Nhung Nguyen; Sophia Ananiadou; Matthew Shardlow",
"journal": "",
"ref_id": "b54",
"title": "UoM&MMU at TSAR-2022 Shared Task: Prompt Learning for Lexical Simplification",
"year": "2022"
},
{
"authors": "John Peniel; Sandeep Whistely; Galiveeti Mathias; Poornima",
"journal": "",
"ref_id": "b55",
"title": "PresiUniv at TSAR-2022 Shared Task: Generation and Ranking of Simplification Substitutes of Complex Words in Multiple Languages",
"year": "2022"
},
{
"authors": "Rodrigo Wilkens; David Alfter; Rémi Cardon; Isabelle Gribomont; Others",
"journal": "",
"ref_id": "b56",
"title": "CENTAL at TSAR-2022 Shared Task: How Does Context Impact BERT-Generated Substitutions for Lexical Simplification",
"year": "2022"
},
{
"authors": "Zhilin Yang; Zihang Dai; Yiming Yang; Jaime Carbonell; Ruslan Salakhutdinov; Quoc V Le",
"journal": "",
"ref_id": "b57",
"title": "XLNet: Generalized Autoregressive Pretraining for Language Understanding",
"year": "2019"
},
{
"authors": "Yan Chak; John Yeung; Lee",
"journal": "",
"ref_id": "b58",
"title": "Personalized text retrieval for learners of Chinese as a foreign language",
"year": "2018"
},
{
"authors": "Chris Seid Muhie Yimam; Shervin Biemann; Gustavo Malmasi; Luci Paetzold; Sanja Specia; Anaïs Štajner; Marcos Tack; Zampieri",
"journal": "",
"ref_id": "b59",
"title": "A Report on the Complex Word Identification Shared Task",
"year": "2018"
},
{
"authors": "Jenny Alexandra; Ortiz Zambrano; Arturo Montejo; Ráez ",
"journal": "",
"ref_id": "b60",
"title": "Overview of ALexS 2020: First Workshop on Lexical Analysis at SEPLN",
"year": "2020"
}
] | [] |
|
10.18653/v1/2021.acl-long.238 | 2023-10-24 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b30",
"b22",
"b3",
"b7",
"b14",
"b21",
"b28",
"b12",
"b15",
"b10"
],
"table_ref": [],
"text": "Recently, there has been a surge in the release of Large Language Models (LLMs) by both industrial and academic institutions. These models vary from open-source releases such as OPT (Zhang et al., 2022) and LLAMA (Touvron et al., 2023) to closed-source ones like GPT-3 (Brown et al., 2020) and PALM (Chowdhery et al., 2022). In addition, researchers have developed models that are finetuned on top of these foundational models to better Figure 1: Three-Dimensional Grid of Fine-Tuning, Prompting, and Scale. Each dimension is represented as an axis, with three levels for each of finetuning, prompting, and scale plotted on each axis. The resulting grid consists of 27 different combinations evaluated on various reasoning tasks. It should be noted that there is a hidden dimension, the scoring function, comprising four components. This results in a comprehensive total of 6,156 evaluations.\nfollow instructions, such as OPT-IML (Iyer et al., 2022) and Alpaca (Taori et al., 2023). Despite the remarkable progress in LLMs' performance in Natural Language Processing (NLP) tasks, reasoning remains a challenging area. For example, prior work have shown that LLMs struggle with commonsense reasoning (West et al., 2022) and arithmetic reasoning (Hendrycks et al., 2021) to name a few.\nRecent efforts have attempted to improve the reasoning performance of LLMs by decomposing answers into step-by-step reasoning chains using incontext learning (Wei et al., 2022b;Kojima et al., 2022) or during finetuning (Chung et al., 2022;Wei et al., 2021a). While these approaches have shown some improvement on benchmarks such as GSM8K (Cobbe et al., 2021), it is not clear how those explanations affect finetuning, prompting, or {Task Definition} Provide your answer followed by a brief reasoning."
},
{
"figure_ref": [],
"heading": "{In-Context Examples}",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Input: {input}",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Options: {options}",
"publication_ref": [
"b29",
"b30",
"b23"
],
"table_ref": [
"tab_3",
"tab_6"
],
"text": "Output: The answer is {answer} because {explanation} Figure 2: Template used during both training and inference. The model is tasked with predicting the answer followed by the explanation. their combination. Concurrent work has investigated the generalization capability of such models to reasoning skills beyond those encountered during finetuning (Yu et al., 2022), but a comprehensive evaluation of the role of explanation during finetuning and prompting with respect to reasoning skills is still lacking.\nIn this paper, we aim to address this gap. We investigate OPT (Zhang et al., 2022) as a representative of such models and utilize it as our base model. Through finetuning OPT on a collection of carefully curated open-source reasoning datasets that come with explanations for each instance, we evaluate its performance on 57 tasks drawn from the SUPER-NATURALINSTRUCTIONS benchmark (Wang et al., 2022), covering 26 different reasoning skills. Our experiments are structured around three key dimensions: finetuning, prompting, and scale, each of which is comprised of three distinct components (See Figure 1). Finetuning: (1) a (vanilla) unfinetuned OPT model; (2) A finetuned OPT model without explanations (OPT-R); and, (3) A finetuned OPT model with explanations (OPT-RE). Prompting: (1) zero-shot prompting; (2) Fewshot prompting without explanations; and, (3) Fewshot prompting with explanations. Finally, Scale: (1) 1.3B; (2) 6.7B; and, (3) 13B. Accordingly, we create grid of 27 different components, providing a detailed analysis measuring the impact of explanations during finetuning and inference across different model scales.\nOur findings reveals that finetuning on reasoning datasets leads to statistically significant improvements in seven reasoning skills, including Numerical, Analogical and Reasoning on Objects, with Physical, Counting and Textual Entailment showing a significant effect only for the OPT-RE model, across both fewshot prompting conditions and model sizes, as compared to the vanilla OPT model (see Table 2). However, we also find that this approach significantly hinders the performance of three other reasoning skills (see Table 3). We also investigate the impact of incorporating explanations during fewshot prompting and find that it does not have a significant impact on the performance of the finetuned models, as measured by the variance in the difference between both prompting methods across reasoning skills for each model. However, we notice that it has a more noticeable effect on the performance of the vanilla OPT model, as shown in Table 5. Additionally, we observe a consistent increase in the average performance across all tasks from Fewshot to Fewshot-E, as well as from OPT to OPT-R to OPT-RE models, indicating that explanations do have a small effect on performance during both finetuning and prompting. Finally, Table 4 presents a summary of the results, indicating which reasoning skills demonstrate improvement due to the incorporation of explanations during either finetuning or prompting, which skills show a negative effect, and which skills have negligible effects regarding explanations. The finetuning corpus utilized to refine OPT is composed of various reasoning datasets, each of which includes a corresponding explanation or rationale for the answer. These rationales may consist of a sequence of smaller steps (i.e. chain-ofthought) or a free-form text that elucidates the reasoning behind the answer. As shown in Figure 2, we employ a uniform template for all tasks during the training process. The input to the model begins with a task definition, followed by an instruction to provide an answer followed by a brief reasoning. Next, we extract two random in-context examples uniformly from the training set that remain constant throughout training for each instance. The input for the current training instance is then presented in a format specific to each task. The options for the answer are then included in the input, but not in the in-context examples (see Appendix A for further details on task-specific definitions and options). The options are pre-shuffled for each training instance. The model is finally provided with the answer prefix, \"Output: The answer is\", and is tasked to predict the answer, followed by an explanation if OPT-RE is being finetuned. Similarly, the in-context examples only comprise an explanation when training OPT-RE."
},
{
"figure_ref": [],
"heading": "OPT-R: Finetuning on Reasoning Skills",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [
"fig_1"
],
"heading": "Reasoning Datasets with Explanations",
"publication_ref": [
"b16",
"b19",
"b18",
"b0",
"b4",
"b2"
],
"table_ref": [],
"text": "Below is a brief description of each dataset used during finetuning. See Figure 3 for the relative size of each dataset.\nAQUA-RAT The Algebra Question Answering with Rationales dataset (Ling et al., 2017) rendering the task of solving algebraic word problems more feasible by dividing the problem into a series of smaller steps. They create a 100k-sample dataset that contains questions, answers and rationales in natural language and human-readable mathematical expressions that can be used to derive the final answer.\nCoQA The Conversational Question Answering dataset Reddy et al. (2019). It consists of 127k questions and answers, compiled from 8k conversations about passages from seven different domains. Given a passage that contains a conversation, the model is tasked with answering a question by highlighting the corresponding evidence from the passage.\nCoS-E The Common Sense Explanations dataset Rajani et al. (2019) to induce language models with commonsense reasoning. In this dataset, the model is given a question and a set of choices and is tasked with selecting one of the provided choices along with providing an explanation in natural language as to why that choice is correct.\nECQA The Explanations for Commonsense Question Answering dataset Aggarwal et al. (2021). It is similar to CoS-E since it requires the model to choose one of the provided options to answer the given question, and also provide an explanation.\nESNLI The Stanford Natural Language Inference dataset with Explanations Camburu et al. (2018) to train models to provide interpretable and robust explanations for their decisions. The authors extend the SNLI dataset (Bowman et al., 2015) with human-annotated explanations. Similar to any NLI task, the model is given a premise and hypothesis and the task is to determine whether the hypothesis sentence entails, contradicts, or is neutral with respect to the given premise."
},
{
"figure_ref": [],
"heading": "GSM8K",
"publication_ref": [
"b10",
"b20",
"b9",
"b11"
],
"table_ref": [],
"text": "The Grade School Math dataset Cobbe et al. (2021) to train models to better perform multistep mathematical reasoning. It consists of 8.5k linguistically diverse grade school math word problems. Therefore, the task for the model is to answer the question by performing a series of arithmetic operations to obtain a final answer, while explaining it's reasoning steps.\nProofWriter The ProofWriter dataset Tafjord et al. (2021) to generate both the implications of a theory from the RuleTaker dataset (Clark et al., 2020) and the natural language proofs that support them. Specifically, given a sequence of facts and rules, the model is tasked with answering a question using \"Yes\", \"No\", or \"Unknown\" and provide the reasoning path by referring to the provided facts and rules. We consider the open-world assumption subset of RuleTaker with questions that requires reasoning up to a depth of 5.\nStrategyQA The Strategy Question Answering dataset Geva et al. (2021) to improve multi-hop reasoning for questions where the required reasoning steps are implicit in the question. Therefore, the task of the model is to answer the question using \"Yes\" or \"No\" then provide a strategy that explains the answer by decomposing it into a number of steps."
},
{
"figure_ref": [],
"heading": "Finetuning Procedures",
"publication_ref": [
"b30",
"b30",
"b30",
"b14",
"b23"
],
"table_ref": [],
"text": "OPT The Open Pretrained Transformers (OPT) models are a suite of decoder-only pre-trained transformers ranging from 125M to 175B parameters released by Zhang et al. (2022). In this work, we use three OPT models with sizes of 1.3B, 6.7B and 13B. The details of each model architecture, pre-training corpus and training configuration (e.g. weight initialization, optimizer, tokenizer, hyperparameters, etc.) can be found in Zhang et al. (2022). Implementation Details To finetune the selected models, we utilized the metaseq 1 implementation since it enables higher training efficiency compared to other codebases (Zhang et al., 2022). Each model is finetuned twice for 10 epochs, once with explanations and once without (i.e. OPT-RE vs OPT-R, respectively). Models are evaluated at the end of each epoch on a chosen set of SUPER-NATURALINSTRUCTIONS validation tasks, and the checkpoint with the best performance is selected for evaluation on the testing tasks. The loss is calculated only on the tokens the model is tasked to predict during inference, and not the full input, what is referred to as label-loss in (Iyer et al., 2022).\nThe samples across all datasets are shuffled during training. Further, the model is provided with two in-context examples during finetuning in addition to the task definition to match inference time following (Wang et al., 2022).\n1 https://github.com/facebookresearch/metaseq 3 Evaluating the Models"
},
{
"figure_ref": [],
"heading": "SUPER-NATURALINSTRUCTIONS Tasks",
"publication_ref": [
"b23"
],
"table_ref": [
"tab_1"
],
"text": "In this study, we focus on a subset of the SUPER-NATURALINSTRUCTIONS benchmark version 2.62 (SUP-NATINST for short) proposed by Wang et al. (2022), which comprises 1,616 varied NLP tasks and includes meta-labels for each task, such as task type, domain and more importantly for this work: the underlying reasoning skills. Specifically, we select a subset of tasks that satisfy two key criteria: (i) the task focuses on a single reasoning skill, enabling us to evaluate a specific atomic skill, and (ii) the task can be tested using classification mode, as detailed in Section 3.2. Note that there is no data contamination between finetuning data and the evaluation benchmark. Benchmark Splits Following the task selection process, we apply a random sampling technique to ensure diversity within the testing set. Specifically, we select a maximum of three tasks from each reasoning skill, and allocate any remaining tasks to the validation set. Notably, this approach enables us to obtain a representative sample of the selected reasoning skills for testing, while also ensuring that our model's performance is not influenced by a particular subset of tasks. Table 1 shows the complete list of tasks used for evaluating our finetuned models for each reasoning skill."
},
{
"figure_ref": [],
"heading": "Evaluation Setup",
"publication_ref": [
"b3"
],
"table_ref": [],
"text": "Earlier, we mentioned that we selected 57 tasks spanning 26 reasoning skills from SUP-NATINST to evaluate our finetuned models. To meet our criteria, as detailed in Section 3.1, each task had to fulfill two conditions. The second condition required that the task can be considered a classification task. That means there is a discrete set of candidates (one of which is correct) and thereby treating it as a classification problem where the highest-scoring candidate is considered the answer. To ensure this, we utilized a straightforward heuristic: we only sampled tasks that had no more than 10 possible candidate answers.\nClassification Method To determine the correct answer, we conduct a forward pass for each potential candidate answer and utilize a scoring function to measure the likelihood that the candidate tokens follows the input, similar to Brown et al. (2020). This process is repeated four times using distinct scoring functions, as detailed in the subsequent paragraph. The highest accuracy score from the four scoring functions is considered as result of the task.\nScoring Functions This is considered the fourth dimension of this work since we evaluate each task using four different scoring functions and take the maximum accuracy as the result. The four scoring functions used are as follows: (1) mean, which involves computing the average of the log probabilities of candidate tokens, also referred to as token score.\n(2) unconditional-norm, which computes the difference between the sum of token scores of the candidate when unconditioned by any previous tokens and the sum of candidate token scores when conditioned by previous input. (3) suffix, which computes the sum of the conditioned candidate's token scores alone. Finally, (4) sum, which involves calculating the sum of all the token scores passed to the model. The reason we employed different functions is that we observed significant gains in performance when using one scoring function over the other for specific tasks. Therefore, in order to ensure fairness across all tasks, we selected the highest accuracy over all scoring functions for each task."
},
{
"figure_ref": [
"fig_2"
],
"heading": "Results & Findings",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we present the results and findings of our experiments. First, we illustrate in Figure 4 the outcome of our evaluation on the effectiveness of finetuned models as compared to the vanilla OPT model, across three different scales when using both fewshot prompting with and without explanations. Furthermore, we observe a monotonic increase in the performance of each model as we increase the scale under those two prompting condition, which indicates a positive correlation between the model's capacity and its overall performance. However, we note that this trend does not apply to the zeroshot prompting method, since we are testing out-of-distribution tasks and that the finetuned models were trained with fewshot exemplars in their context. This leads us to focus only on the fewshot prompting methods, with and without explanations, for the remaining of our evaluations. Specifically, we investigate the impact of finetuning the OPT models on reasoning datasets, as compared to the vanilla OPT model, and explore the effect of explanations during finetuning and prompting, both in terms of the reasoning skill."
},
{
"figure_ref": [],
"heading": "Model Performance for Reasoning Skills",
"publication_ref": [],
"table_ref": [],
"text": "The results reported in this and the following section are the classification accuracy of each reasoning skill across different conditions, such as model sizes and fewshot prompting methods. Counting skill, the OPT-RE variant outperforms both the OPT-R and OPT models, underscoring the criticality of incorporating explanations during the finetuning process for mathematical datasets. Likewise, the Physical Reasoning tasks exhibit a similar trend. On the other hand, we can see that for the Argument, Deductive Textual Entailment and Commonsense skills the non-finetuned version outperforms considerably."
},
{
"figure_ref": [],
"heading": "Fine-Grained Skill Analysis",
"publication_ref": [],
"table_ref": [
"tab_6",
"tab_6"
],
"text": "Table 4 shows the classification accuracy results obtained from the three models, in relation to the reasoning skill and few-shot prompting method used. The best accuracy value for each reasoning skill is indicated in bold, and the cells are shaded with colors ranging from green to white to indicate their position in the accuracy spectrum of each reasoning skill. The skills with similar performance across different models are assigned a lighter shade of green, indicating that their color spectrum ends earlier than that of other skills where the difference in performance between models is more significant.\nThe table is divided into four blocks to distinguish effects of finetuning and prompting methods on reasoning skills: the first block showcases skills where the finetuned (OPT-RE and OPT-R) models outperform the vanilla OPT model, the second block highlights skills where OPT-RE has better accuracy than other models therefore illustrating the importance of finetuning on explanations on those skills. The third block displays skills where OPT outperforms other models showing that finetuning actually hurts performance in this case, and the fourth block identifies skills where the choice of model or prompting method has little impact on the overall performance. Table 4: Classification accuracy results achieved by different models as a function of the reasoning skill and few-shot prompting method employed. The best accuracy obtained for each reasoning skill is highlighted in bold. The cells are shaded with colors ranging from green to white to indicate their position in the accuracy spectrum. Reasoning skills with smaller variance in achieved results are assigned a lighter shade of green to convey the extent of similarity between models. The first block highlights skills where the finetuned models perform notably better than the vanilla OPT. The second block emphasizes the skills where OPT-RE outperforms other models. In contrast, the third block showcases the skills where OPT outperforms the other models. Lastly, the fourth block identifies skills where the choice of model or prompting method has little impact on the overall performance.\nExplanations' Effect One of the central questions that we sought to investigate in this study is the extent to which explanations play a role in improving the reasoning capabilities of OPT models during finetuning and prompting. The results presented in Table 5 suggest that the presence or absence of explanations in the fewshot examples employed for prompting does not significantly impact the performance of the model when the model is finetuned on reasoning datasets. Concretely, in Table 5, we present the variance of the absolute accuracy difference for each model across reason-ing skills by excluding the Temporal skill, which was identified as an outlier. Specifically, we compute the difference between the two corresponding columns for each model in Table 4. These values provide insights into the impact of including explanations during prompting on the performance of the models. Our findings reveal that the difference is negligible for OPT-R and OPT-RE models, suggesting that the choice of prompting method does not significantly affect the model's accuracy. However, for the vanilla OPT model, the difference is more substantial, emphasizing the impor-tance of employing explanations during fewshot prompting. However, the mean performance of each model across the distinct fewshot prompting methods demonstrates a slight yet consistent increase in classification accuracy, from Fewshot to Fewshot-E (incorporating explanations), as well as from OPT to OPT-R to OPT-RE models showing that explanations do have a small effect on performance during both finetuning and prompting. "
},
{
"figure_ref": [],
"heading": "Model",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b3",
"b7",
"b17",
"b14",
"b13",
"b1",
"b14",
"b14",
"b23",
"b29",
"b3"
],
"table_ref": [],
"text": "Reasoning LLMs LLMs have made significant advancements in the field of NLP and related areas (Brown et al., 2020;Chowdhery et al., 2022;Chung et al., 2022), especially with the advent of the pre-train, prompt, and predict paradigm (Liu et al., 2021). This paradigm has enabled these models to solve a multitude of tasks through incontext fewshot or zeroshot learning using instructions (Wei et al., 2021b;Iyer et al., 2022). However, their reasoning abilities have been a subject of debate in recent literature (Huang and Chang, 2022;AlKhamissi et al., 2022). Several studies suggest that increasing the size of an LM trained through the same next-token prediction method can lead to the emergence of complex behaviors (Wei et al., 2022a), including reasoning. For instance, some research has demonstrated that sufficiently large LMs can use chain-of-thought prompting (Wei et al., 2022b) to simulate human-like reasoning. Other studies have shown that the addition of a simple prompt, such as \"Let's think step-by-step\" (Kojima et al., 2022) can elicit reasoning abilities in LLMs by generating explicit reasoning steps before decoding the final answer. However, some researchers contend that emulating the human reasoning thought process is distinct from claiming that the model can truly reason (Wei et al., 2022b).\nFinetuned LLMs Concurrent studies have finetuned LLMs to follow instructions to improve their generalization ability to unseen tasks through zero and fewshot learning (Iyer et al., 2022;Chung et al., 2022). However, our approach differs in that we only finetune on a selected number of open-source datasets that provide explanations for each instance. This enables us to focus on the importance of explanations during finetuning in the context of reasoning skills. While concurrent works, such as (Iyer et al., 2022;Wang et al., 2022), have experimented with different prompting methods during finetuning and inference, our study focuses primarily on evaluating the reasoning ability of the finetuned models across a set of reasoning skills. Other concurrent studies have explored the impact of finetuning on a set of held-out reasoning tasks (Yu et al., 2022), but their evaluation approach, which involves generating answers, may be influenced by various factors such as decoding strategy, decoding parameters, and prompt templates. In contrast, we adopt a rank classification approach similar to (Brown et al., 2020), which better captures the reasoning performance of the model being evaluated, in addition to covering a larger number of reasoning skills and tasks."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "In this study, we investigated the impact of incorporating explanations during finetuning and prompting on three different sizes of the OPT model. Through a systematic and comprehensive evaluation process that considered three key dimensions, we found that while explanations did provide a small improvement in performance, the effect was not significant when incorporated in the in-context demonstrations during inference for the finetuned models. Additionally, our results showed that both finetuned models exhibited significant improvements in reasoning skills such as Numerical, Analogical and Reasoning on Objects. Moreover, we demonstrated that skills such as Physical, Counting, and Textual Entailment benefited from incorporating explanations during the finetuning process. Overall, our findings provide insights into the impact of incorporating explanations on the reasoning capabilities of LLMs and offer guidance on which reasoning skills would benefit most from the inclusion and exclusion of explanations during finetuning and prompting."
},
{
"figure_ref": [],
"heading": "Dataset",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Task Definition Options",
"publication_ref": [],
"table_ref": [],
"text": "AQuA You are given an algebraic word question. Questions in this task often requires executing a series of arithmetic operations to obtain a final answer. You are also given 5 answer options (associated with 'A', 'B, 'C', 'D', 'E'). Do not generate anything else apart from one of the following characters: \"A\", \"B\", \"C\", \"D\", \"E\" and the corresponding explanation.\n-A -B -C -D -E"
},
{
"figure_ref": [],
"heading": "CoQA",
"publication_ref": [],
"table_ref": [],
"text": "You are given a passage that contains a conversation and a question. The task is to answer the question and provide an explanation that highlights the corresponding evidence in the passage."
},
{
"figure_ref": [],
"heading": "Free-form text",
"publication_ref": [],
"table_ref": [],
"text": "CoS-E You are given a passage that contains a sentence and a question. The task is to answer the question by selecting one of the provided choices.\nSelect one of the provided choices"
},
{
"figure_ref": [],
"heading": "ECQA",
"publication_ref": [],
"table_ref": [],
"text": "You are given a question that requires commonsense reasoning. The task is to answer the question by selecting one of the provided choices.\none of the provided choices ESNLI You will be presented with a premise and a hypothesis sentence. The task is to determine whether the hypothesis sentence entails (implies), contradicts (opposes), or is neutral with respect to the given premise sentence. Please answer with \"Contradiction\", \"Neutral\",or \"Entailment\".\n-Contradiction -Neutral -Entailment"
},
{
"figure_ref": [],
"heading": "GSM8K",
"publication_ref": [],
"table_ref": [],
"text": "You will be presented with a passage that contains a grade school math word problem. The task is to answer the question by performing a series of arithmetic operations to obtain a final answer."
},
{
"figure_ref": [],
"heading": "Number",
"publication_ref": [],
"table_ref": [],
"text": "ProofWriter You are given a sequence of facts and rules followed by a question. The task is to answer the question using \"Yes\", \"No\" or \"Unknown\".\n-Yes -No -Unknown StrategyQA You are given a sentence and a question. The required reasoning steps are implicit in the question. The task is to answer the question using \"Yes\" or \"No\" then provide a strategy that explains the answer by decomposing it into a number of steps."
},
{
"figure_ref": [],
"heading": "-Yes -No",
"publication_ref": [],
"table_ref": [],
"text": "Table 6: Task definition and options used for each of the finetuning reasoning datasets."
},
{
"figure_ref": [],
"heading": "A Finetuning Task Definition and Options",
"publication_ref": [],
"table_ref": [],
"text": "Table 6 shows the task definition and options provided as input to the template shown in Figure 2 during finetuning the OPT models on the reasoning datasets."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [],
"table_ref": [],
"text": "While our study provides valuable insights into the impact of finetuning on reasoning performance and the role of explanations during finetuning and prompting with respect to various reasoning skills, there are several limitations to our work. Firstly, we only consider a single LLM, OPT, as our base model. Our results may not generalize to other LLMs with different architectures or pretraining objectives. Secondly, we only use a limited set of reasoning datasets for finetuning due to the limited availability of open-source datasets with explanations. However, it is possible that our findings may not hold for models finetuned on larger closed datasets as usually seen in real-world scenarios. Thirdly, our experiments only cover a limited range of model sizes due to limitations in computational budget, therefore it is possible that our findings may not hold for much larger models. Finally, we only consider finetuning using fewshot prompting conditions in our experiments, and it is possible that our findings may not hold for models finetuned without in-context exemplars. Overall, while our study provides valuable insights into the impact of finetuning and explanations on reasoning performance, further research is needed to investigate these factors across a broader range of models, datasets, and finetuning strategies."
},
{
"figure_ref": [],
"heading": "Ethics Statement",
"publication_ref": [],
"table_ref": [],
"text": "This work is based on analyzing and evaluating the performance of LLMs on reasoning tasks using existing public datasets. No personally identifiable information or sensitive data was collected or used in this research. We acknowledge the potential risks of developing LLMs, including their potential impact on spreading misinformation, generating unwanted content and the exacerbation of existing biases in datasets. Our work aims to contribute to improving the transparency and understanding of how LLMs can be optimized for specific reasoning skills. We hope our findings will inspire further research on developing ethical and responsible approaches for developing and deploying LLMs."
}
] | In this paper, we conduct a thorough investigation into the reasoning capabilities of Large Language Models (LLMs), focusing specifically on the Open Pretrained Transformers (OPT) models as a representative of such models. Our study entails finetuning three different sizes of OPT on a carefully curated reasoning corpus, resulting in two sets of finetuned models: OPT-R, finetuned without explanations, and OPT-RE, finetuned with explanations. We then evaluate all models on 57 out-of-domain tasks drawn from the SUPER-NATURALINSTRUCTIONS benchmark, covering 26 distinct reasoning skills, utilizing three prompting techniques. Through a comprehensive grid of 27 configurations and 6,156 test evaluations, we investigate the dimensions of finetuning, prompting, and scale to understand the role of explanations on different reasoning skills. Our findings reveal that having explanations in the fewshot exemplar has no significant impact on the model's performance when the model is finetuned, while positively affecting the non-finetuned counterpart. Moreover, we observe a slight yet consistent increase in classification accuracy as we incorporate explanations during prompting and finetuning, respectively. Finally, we offer insights on which skills benefit the most from incorporating explanations during finetuning and prompting, such as Numerical (+20.4%) and Analogical (+13.9%) reasoning, as well as skills that exhibit negligible or negative effects. | OPT-R: Exploring the Role of Explanations in Finetuning and Prompting for Reasoning Skills of Large Language Models | [
{
"figure_caption": "Figure 3 :3Figure 3: Number of samples in each dataset of the training corpus. Y-axis in log scale.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: Results achieved across all tasks as a function of the three primary dimensions analyzed in this study: Finetuning, Prompting and Scale.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "tab_1",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "The results reveal that the finetuned variants of the OPT model demonstrate a significant improvement on seven distinct reasoning skills, with particular emphasis on the Numerical and Analogical reasoning tasks. Specifically, for the Mathematical",
"figure_data": "SkillOPT OPT-R OPT-REArgument57.946.1 -48.7 -TE -Deductive 36.029.0 -29.4 -Commonsense33.429.728.8 -reason-ing skills where the vanilla OPT model performssignificantly better than either of its finetuned coun-terparts.SkillOPT OPT-R OPT-RENumerical44.865.2*64.7*Analogical49.062.9*60.8*Counting19.813.131.3*Physical38.237.849.1*Entailment 42.647.251.6*Social Int34.143.0*40.1Objects54.362.6*59.9*Table 2: Performance as a function of the reasoningskills where OPT-RE or OPT-R performs significantlybetter than the OPT model as measured by Welch's t-test(p < 0.05) denoted by the * symbol. The performance ismeasured across Fewshot and Fewshot-E prompting, thethree different scales and tasks under the correspondingreasoning skill. Best result indicated in bold.",
"figure_id": "tab_2",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Performance as a function of the reasoning skill where OPT performs significantly better than either OPT-R or OPT-RE as measured by Welch's t-test (p < 0.05) denoted by the -symbol. The performance is measured across Fewshot and Fewshot-E prompting, the three different scales and tasks under the corresponding reasoning skill. TE is Textual Entailment.",
"figure_data": "",
"figure_id": "tab_3",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "The first column shows the variance of the absolute difference in accuracy for each model across different reasoning skills, when using Fewshot (F) and Fewshot-E (FE) prompting methods. The second and third columns show the average performance of each model across each prompting method. Results are obtained after dropping the outlier Temporal skill.",
"figure_data": "",
"figure_id": "tab_6",
"figure_label": "5",
"figure_type": "table"
}
] | Badr Alkhamissi; Siddharth Verma; Ping Yu Zhijing; Jin Asli Celikyilmaz; Mona Diab; Meta Ai | [
{
"authors": "Shourya Aggarwal; Divyanshu Mandowara; Vishwajeet Agrawal; Dinesh Khandelwal; Parag Singla; Dinesh Garg",
"journal": "Association for Computational Linguistics",
"ref_id": "b0",
"title": "Explanations for Common-senseQA: New Dataset and Models",
"year": "2021"
},
{
"authors": "Badr Alkhamissi; Millicent Li; Asli Celikyilmaz; Mona T Diab; Marjan Ghazvininejad",
"journal": "",
"ref_id": "b1",
"title": "A review on language models as knowledge bases",
"year": "2022"
},
{
"authors": "R Samuel; Gabor Bowman; Christopher Angeli; Christopher D Potts; Manning",
"journal": "",
"ref_id": "b2",
"title": "A large annotated corpus for learning natural language inference",
"year": "2015"
},
{
"authors": "Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell",
"journal": "Advances in neural information processing systems",
"ref_id": "b3",
"title": "Language models are few-shot learners",
"year": "2020"
},
{
"authors": "Oana-Maria Camburu; Tim Rocktäschel; Thomas Lukasiewicz; Phil Blunsom",
"journal": "",
"ref_id": "b4",
"title": "e-snli: Natural language inference with natural language explanations",
"year": "2018"
},
{
"authors": "S Bengio; H Wallach; H Larochelle; K Grauman; N Cesa-Bianchi; R Garnett",
"journal": "",
"ref_id": "b5",
"title": "editors",
"year": ""
},
{
"authors": "Curran Associates; Inc ",
"journal": "",
"ref_id": "b6",
"title": "",
"year": ""
},
{
"authors": "Aakanksha Chowdhery; Sharan Narang; Jacob Devlin; Maarten Bosma; Gaurav Mishra; Adam Roberts; Paul Barham; Hyung Won Chung; Charles Sutton; Sebastian Gehrmann",
"journal": "",
"ref_id": "b7",
"title": "Palm: Scaling language modeling with pathways",
"year": "2022"
},
{
"authors": "Chung Hyung Won; Le Hou; Shayne Longpre; Barret Zoph; Yi Tay; William Fedus; Eric Li; Xuezhi Wang; Mostafa Dehghani; Siddhartha Brahma",
"journal": "",
"ref_id": "b8",
"title": "Scaling instruction-finetuned language models",
"year": "2022"
},
{
"authors": "Peter Clark; Oyvind Tafjord; Kyle Richardson",
"journal": "",
"ref_id": "b9",
"title": "Transformers as soft reasoners over language",
"year": "2020"
},
{
"authors": "Karl Cobbe; Vineet Kosaraju; Mohammad Bavarian; Jacob Hilton; Reiichiro Nakano; Christopher Hesse; John Schulman",
"journal": "",
"ref_id": "b10",
"title": "Training verifiers to solve math word problems",
"year": "2021"
},
{
"authors": "Mor Geva; Daniel Khashabi; Elad Segal; Tushar Khot; Dan Roth; Jonathan Berant",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b11",
"title": "Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies",
"year": "2021"
},
{
"authors": "Dan Hendrycks; Collin Burns; Saurav Kadavath; Akul Arora; Steven Basart; Eric Tang; Dawn Song; Jacob Steinhardt",
"journal": "",
"ref_id": "b12",
"title": "Measuring mathematical problem solving with the math dataset",
"year": "2021"
},
{
"authors": "Jie Huang; Kevin Chen; -Chuan Chang",
"journal": "",
"ref_id": "b13",
"title": "Towards reasoning in large language models: A survey",
"year": "2022"
},
{
"authors": "Srinivasan Iyer; Xi Victoria Lin; Ramakanth Pasunuru; Todor Mihaylov; Dániel Simig; Ping Yu; Kurt Shuster; Tianlu Wang; Qing Liu; Punit Singh Koura",
"journal": "",
"ref_id": "b14",
"title": "Opt-iml: Scaling language model instruction meta learning through the lens of generalization",
"year": "2022"
},
{
"authors": "Takeshi Kojima; Shane Shixiang; Machel Gu; Yutaka Reid; Yusuke Matsuo; Iwasawa",
"journal": "",
"ref_id": "b15",
"title": "Large language models are zero-shot reasoners",
"year": "2022"
},
{
"authors": "Wang Ling; Dani Yogatama; Chris Dyer; Phil Blunsom",
"journal": "Association for Computational Linguistics",
"ref_id": "b16",
"title": "Program induction by rationale generation: Learning to solve and explain algebraic word problems",
"year": "2017"
},
{
"authors": "Pengfei Liu; Weizhe Yuan; Jinlan Fu; Zhengbao Jiang; Hiroaki Hayashi; Graham Neubig",
"journal": "ACM Computing Surveys",
"ref_id": "b17",
"title": "Pretrain, prompt, and predict: A systematic survey of prompting methods in natural language processing",
"year": "2021"
},
{
"authors": "Nazneen Rajani; Bryan Mccann; Caiming Xiong; Richard Socher",
"journal": "",
"ref_id": "b18",
"title": "Explain yourself! leveraging language models for commonsense reasoning",
"year": "2019"
},
{
"authors": "Siva Reddy; Danqi Chen; Christopher D Manning",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b19",
"title": "CoQA: A conversational question answering challenge",
"year": "2019"
},
{
"authors": "Oyvind Tafjord; Bhavana Dalvi; Peter Clark",
"journal": "Association for Computational Linguistics",
"ref_id": "b20",
"title": "ProofWriter: Generating implications, proofs, and abductive statements over natural language",
"year": "2021"
},
{
"authors": "Rohan Taori; Ishaan Gulrajani; Tianyi Zhang; Yann Dubois; Xuechen Li; Carlos Guestrin; Percy Liang; Tatsunori B Hashimoto",
"journal": "",
"ref_id": "b21",
"title": "Stanford alpaca: An instruction-following llama model",
"year": "2023"
},
{
"authors": "Hugo Touvron; Thibaut Lavril; Gautier Izacard; Xavier Martinet; Marie-Anne Lachaux; Timothée Lacroix; Baptiste Rozière; Naman Goyal; Eric Hambro; Faisal Azhar; Armand Aur'elien Rodriguez; Edouard Joulin; Guillaume Grave; Lample",
"journal": "",
"ref_id": "b22",
"title": "Llama: Open and efficient foundation language models",
"year": "2023"
},
{
"authors": "Yizhong Wang; Swaroop Mishra; Pegah Alipoormolabashi; Yeganeh Kordi; Amirreza Mirzaei; Atharva Naik; Arjun Ashok; Arut Selvan Dhanasekaran; Anjana Arunkumar; David Stap; Eshaan Pathak; Giannis Karamanolakis; Haizhi Lai; Ishan Purohit; Ishani Mondal; Jacob Anderson; Kirby Kuznia; Krima Doshi; Kuntal Kumar Pal; Maitreya Patel; Mehrad Moradshahi; Mihir Parmar; Mirali Purohit; Neeraj Varshney; Rohitha Phani; Pulkit Kaza; Ravsehaj Verma; Rushang Singh Puri; Savan Karia; Doshi; Keyur Shailaja; Siddhartha Sampat; Sujan Mishra; A Reddy; Sumanta Patro; Tanay Dixit; Xudong Shen",
"journal": "",
"ref_id": "b23",
"title": "Super-NaturalInstructions: Generalization via declarative instructions on 1600+ NLP tasks",
"year": "2022"
},
{
"authors": "Jason Wei; Maarten Bosma; Y Vincent; Kelvin Zhao; Adams Wei Guu; Brian Yu; Nan Lester; Andrew M Du; Quoc V Dai; Le",
"journal": "",
"ref_id": "b24",
"title": "a. Finetuned language models are zero-shot learners",
"year": "2021"
},
{
"authors": "Jason Wei; Maarten Bosma; Y Vincent; Kelvin Zhao; Adams Wei Guu; Brian Yu; Nan Lester; Andrew M Du; Quoc V Dai; Le",
"journal": "",
"ref_id": "b25",
"title": "Finetuned language models are zero-shot learners",
"year": "2021"
},
{
"authors": "Jason Wei; Yi Tay; Rishi Bommasani; Colin Raffel; Barret Zoph; Sebastian Borgeaud; Dani Yogatama; Maarten Bosma; Denny Zhou; Donald Metzler; Ed Huai Hsin Chi; Tatsunori Hashimoto; Oriol Vinyals; Percy Liang; Jeff Dean; William Fedus",
"journal": "",
"ref_id": "b26",
"title": "Emergent abilities of large language models",
"year": "2022"
},
{
"authors": "Jason Wei; Xuezhi Wang; Dale Schuurmans; Maarten Bosma; Ed Chi; Quoc Le; Denny Zhou",
"journal": "",
"ref_id": "b27",
"title": "Chain of thought prompting elicits reasoning in large language models",
"year": "2022"
},
{
"authors": "Peter West; Chandra Bhagavatula; Jack Hessel; Jena D Hwang; Liwei Jiang; Ronan Le Bras; Ximing Lu; Sean Welleck; Yejin Choi",
"journal": "",
"ref_id": "b28",
"title": "Symbolic knowledge distillation: from general language models to commonsense models",
"year": "2022"
},
{
"authors": "Ping Yu; Tianlu Wang; O Yu; Badr Golovneva; Siddharth Alkhamissi; Zhijing Verma; Gargi Jin; Mona Ghosh; Asli Diab; Celikyilmaz",
"journal": "",
"ref_id": "b29",
"title": "Alert: Adapting language models to reasoning tasks",
"year": "2022"
},
{
"authors": "Susan Zhang; Stephen Roller; Naman Goyal; Mikel Artetxe; Moya Chen; Shuohui Chen; Christopher Dewan; Mona Diab; Xian Li; Xi Victoria Lin",
"journal": "",
"ref_id": "b30",
"title": "Opt: Open pre-trained transformer language models",
"year": "2022"
}
] | [] |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b2",
"b8",
"b9",
"b18",
"b13",
"b4",
"b3",
"b16",
"b0",
"b17",
"b5",
"b20"
],
"table_ref": [],
"text": "In recent years, pre-trained language models have witnessed rapid development. Broadly speaking, they can be categorized into three main architectures: the Encoder architecture represented by BERT (Devlin et al., 2018), the Decoder architecture represented by GPT (Radford et al., 2018), and the Encoder-Decoder architecture represented by T5 (Raffel et al., 2020). Each architecture has its unique characteristics and advantages, catering to different NLP requirements.\nThe GPT series, with GPT-4 (OpenAI, 2023) being the latest addition, has gained considerable attention due to its remarkable performance in natural language generation tasks, including dialogue generation. The ChatGPT (OpenAI, 2022) model, in particular, has impressed researchers and practitioners with its ability to generate coherent and contextually relevant responses in conversational settings. As a result, the GPT series has become a focal point of research and development in the NLP community.\nMoreover, the emergence of large-scale pretrained models has further fueled the advancements in language modeling. Models such as OPT (Zhang et al., 2022), BLOOM (Scao et al., 2022), and LLaMA (Touvron et al., 2023), with parameter sizes reaching billions, have recently been opensourced, enabling researchers and developers to explore the potential of these massive models. These models have demonstrated superior performance on various tasks, pushing the boundaries of what is possible in NLP.\nWhile the general-purpose large models mentioned above have garnered significant attention, the importance of domain-specific models cannot be overlooked. In many domains, the distribution of language and the specific linguistic nuances require models that are fine-tuned or specifically trained for that particular domain. Consequently, a range of domain-specific large models has been proposed to cater to the unique needs of various fields. For example, BioBERT (Lee et al., 2020) and PubMedBERT (Gu et al., 2021) are proposed for the biomedical field, and BloombergGPT (Wu et al., 2023) are proposed for financial scenarios. These models have shown promising results in their respective domains, leveraging the domain-specific knowledge learned during pre-training.\nWithin the Chinese financial domain, there has been considerable progress in the development of pre-trained language models. Researchers have introduced models such as FinBERT (Araci, 2019;Yang et al., 2020;Liu et al., 2021), Mengzi (Zhang et al., 2021), andFinT5 (Lu et al., 2023), which have been tailored for financial text analysis and understanding. These models, though valuable for certain applications, have parameter sizes below one billion, limiting their ability to handle the increasing demands of the Chinese financial NLP landscape. As the volume of financial data and the complexity of language usage continue to grow, there is a pressing need for more powerful models that can effectively process and understand Chinese financial text.\nDespite significant advancements in chat models, there is currently no open-sourced chat model at the scale of hundreds of billions specifically designed for the Chinese language, let alone in the field of Chinese finance. To address this gap, we propose XuanYuan 2.0 (轩辕 2.0), the largest Chinese chat model to date, based on BLOOM-176B. XuanYuan 2.0 not only surpasses its predecessor, XuanYuan 1.0 (轩辕 1.0), which achieved first place at the leaderboard of CLUE classification in 2021, but also addresses the need for a large-scale chat model specifically designed for the Chinese financial domain.\nFurthermore, domain-specific language models and chat models impose higher requirements on data distribution and training approaches compared to general-domain models. Domain-specific models need to capture the unique linguistic characteristics, terminologies, and contexts of a particular field to achieve optimal performance. However, training these models solely on domain-specific data may lead to catastrophic forgetting, where the model loses previously learned knowledge from the general domain, impacting its overall performance. To mitigate this issue, we propose a novel training method, hybrid-tuning, that combines the stages of pre-training and fine-tuning. By integrating the two stages, our approach guarantees that fine-tuning the model with financial-specific instructions does not impede its general generation capabilities acquired during pre-training. As a result, XuanYuan 2.0 can effectively leverage both its general-domain knowledge and domain-specific financial knowledge to provide accurate and contextually appropriate responses in the Chinese financial domain."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b2",
"b8",
"b4",
"b3"
],
"table_ref": [],
"text": "The advancements in pre-trained language models have led to remarkable progress in various NLP tasks, attracting extensive research efforts. Among the notable contributions, the BERT (Devlin et al., 2018) series stands out as a groundbreaking development in the field of pre-trained models. Following the success of BERT, the GPT (Radford et al., 2018) series emerged as a prominent line of research, focusing on the decoding aspect of language modeling. GPT models, in contrast to BERT's bidirectional approach, leveraged autoregressive language modeling. By training on large amounts of unlabeled text data, GPT models acquired a rich understanding of language and demonstrated impressive capabilities in generating coherent and contextually relevant text. Subsequent iterations of the GPT series, such as GPT-4 (OpenAI, 2023), showcased superior performance in various language generation tasks. And Chat-GPT (OpenAI, 2022), an extension of the GPT series, demonstrated the ability to engage in interactive and contextually coherent conversations. This breakthrough sparked considerable interest in developing conversational AI agents capable of simulating human-like dialogue.\nIn addition to the general-purpose BERT and GPT models, there has been a growing interest in domain-specific pre-training. Researchers have recognized that incorporating domain-specific knowledge during pre-training can lead to substantial performance gains in downstream tasks within those domains. Domain-specific pre-trained models aim to capture domain-specific nuances, enabling them to excel in tasks relevant to the target domain. For instance, in the biomedical domain, BioBERT (Lee et al., 2020) and PubMedBERT (Gu et al., 2021) are proposed to leverage large-scale biomedical 3 XuanYuan 2.0"
},
{
"figure_ref": [],
"heading": "Model Architecture",
"publication_ref": [
"b12",
"b7",
"b1",
"b14"
],
"table_ref": [],
"text": "We adopted the original BLOOM (Scao et al., 2022) architecture, which is a decoder-only architecture. The joint probability of tokens in a text can be represented as:\np(w) = p(w 1 , . . . , w T ) = T t=1 p(w t |w <t ) (1)\nwhere w represents a sequence of tokens, w t is the t th token, and w <t is the sequence of tokens preceding w t . This method is called autoregressive language modeling, where we predict the probability of the next token in an iterative manner. And following BLOOM, we utilize ALiBi positional embeddings (Press et al., 2021) and embedding LayerNorm (Dettmers et al., 2022) in the traditional decoder structure of Transformer (Vaswani et al., 2017)."
},
{
"figure_ref": [
"fig_0"
],
"heading": "Hybrid-tuning",
"publication_ref": [
"b15",
"b19"
],
"table_ref": [],
"text": "To alleviate the problem of catastrophic forgetting, we propose a novel domain-specific training framework, hybrid-tuning. In terms of the training stage, it integrates the pre-training stage and instruction fine-tuning stage that are previously split together. In terms of the field of data, it integrates data from both general and financial domains.\nAs shown in Figure 1, different from traditional two-stage domain-specific training, our proposed hybrid-tuning randomly shuffles pre-training data (general pre-training, financial pre-training) and instruction data (general instruction, financial instruction) into one training data. And all the training process is done in one stage. In this way, the model can accurately handle instructions in the financial domain, while retaining general conversational capabilities.\nFor unsupervised pre-training data, we crawl them from the Internet and clean and filter them. For Instruction-tuning data, we use human-written seed instructions to collect general data by Self-Instruct (Wang et al., 2022) and utilize unstructured and structured data in the financial field to gather domain-specific instruction data by Self-QA (Zhang and Yang, 2023). Unstructured financial data comprises a wide range of textual information, such as financial news articles, market reports, analyst commentary, and social media discussions. And structured financial data includes company information and so on. These sources offer valuable insights into market trends, investment strategies, and economic situations."
},
{
"figure_ref": [],
"heading": "Training",
"publication_ref": [
"b11",
"b10"
],
"table_ref": [
"tab_1"
],
"text": "To train our complex and computationally intensive model, we employ the powerful NVIDIA A100 80GB GPU and the DeepSpeed (Rasley et al., 2020) distributed training framework. For parallel processing, we primarily rely on pipeline parallelism, which involves distributing the layers of our model across several GPUs. This approach ensures that each GPU only handles a portion of the model's layers, a technique also known as vertical parallelism. Additionally, we adopt the Zero Redundancy Optimizer (Rajbhandari et al., 2020) to enable different processes to store only a portion of the data (parameters, gradients, and optimizer states). Specifically, we use ZeRO stage 1, which means that only the optimizer states are divided using this method. The specific hyperparameters are presented in Table 2."
},
{
"figure_ref": [],
"heading": "Experiment",
"publication_ref": [],
"table_ref": [],
"text": "We conducted a comparison between our model and other open-source Chinese conversational models. Simultaneously, we constructed evaluation datasets encompassing various dimensions in both general and financial domains, which were subsequently subject to manual assessment. The results revealed XuanYuan's robust knowledge base and conversational capabilities in the financial domain. Further insights and additional findings will be presented in the next version of the paper after the release of the evaluation rankings."
},
{
"figure_ref": [],
"heading": "Hyperparameter",
"publication_ref": [],
"table_ref": [],
"text": "XuanYuan2-7B XuanYuan2 "
}
] | In recent years, pre-trained language models have undergone rapid development with the emergence of large-scale models. However, there is a lack of open-sourced chat models specifically designed for the Chinese language, especially in the field of Chinese finance, at the scale of hundreds of billions. To address this gap, we introduce XuanYuan 2.0 (轩 辕 2.0), the largest Chinese chat model to date, built upon the BLOOM-176B architecture. Additionally, we propose a novel training method called hybrid-tuning to mitigate catastrophic forgetting. By combining generaldomain with domain-specific knowledge and integrating the stages of pre-training and finetuning, XuanYuan 2.0 is capable of providing accurate and contextually appropriate responses in the Chinese financial domain. | XuanYuan 2.0: A Large Chinese Financial Chat Model with Hundreds of Billions Parameters | [
{
"figure_caption": "Figure 1 :1Figure 1: Our proposed hybrid-tuning.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Comparison of different financial language models.",
"figure_data": "arXiv:2305.12002v1 [cs.CL] 19 May 2023",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Training hyperparameters of XuanYuan 2.0.In this paper, we propose the largest Chinese financial chat model, XuanYuan 2.0 (轩辕 2.0), to fill the gap of open-source billion-scale chat models specifically designed for the Chinese financial domain. Besides, we propose a novel training method called hybrid-tuning to mitigate catastrophic forgetting. By combining the general domain with domain-specific knowledge and integrating the stages of pre-training and finetuning, XuanYuan 2.0 achieves the remarkable ability to deliver precise and contextually relevant responses within the Chinese financial domain. We will continue to gather larger-scale Chinese financial domain data in order to further optimize our model.",
"figure_data": "Architecture hyperparametersParameters7,069M176,247MLayers3070Hidden dim.409614336Attention heads32112Vocab size250,680Sequence length2048Precisionfloat16ActivationGELUPosition emb.AlibiTied emb.TruePretraining hyperparametersGlobal Batch Size5122048Learning rate1.2e-46e-5Total tokens341B366BMin. learning rate1e-56e-6Warmup tokens375MDecay tokens410BDecay stylecosineAdam (β 1 , β 2 )(0.9, 0.95)Weight decay1e-1Gradient clipping1.0Multitask finetuning hyperparametersGlobal Batch Size20482048Learning rate2.0e-52.0e-5Total tokens13BWarmup tokens0Decay styleconstantWeight decay1e-45 Conclusion",
"figure_id": "tab_1",
"figure_label": "2",
"figure_type": "table"
}
] | Xuanyu Zhang; Qing Yang; Dongliang Xu; Du Xiaoman | [
{
"authors": "Dogu Araci",
"journal": "",
"ref_id": "b0",
"title": "Finbert: Financial sentiment analysis with pre-trained language models",
"year": "2019"
},
{
"authors": "Tim Dettmers; Mike Lewis; Younes Belkada; Luke Zettlemoyer",
"journal": "",
"ref_id": "b1",
"title": "Llm. int8 (): 8-bit matrix multiplication for transformers at scale",
"year": "2022"
},
{
"authors": "Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova",
"journal": "",
"ref_id": "b2",
"title": "Bert: Pre-training of deep bidirectional transformers for language understanding",
"year": "2018"
},
{
"authors": "Yu Gu; Robert Tinn; Hao Cheng; Michael Lucas; Naoto Usuyama; Xiaodong Liu; Tristan Naumann; Jianfeng Gao; Hoifung Poon",
"journal": "ACM Transactions on Computing for Healthcare (HEALTH)",
"ref_id": "b3",
"title": "Domainspecific language model pretraining for biomedical natural language processing",
"year": "2021"
},
{
"authors": "Jinhyuk Lee; Wonjin Yoon; Sungdong Kim; Donghyeon Kim; Sunkyu Kim; Chan Ho; So ; Jaewoo Kang",
"journal": "Bioinformatics",
"ref_id": "b4",
"title": "Biobert: a pre-trained biomedical language representation model for biomedical text mining",
"year": "2020"
},
{
"authors": "Zhuang Liu; Degen Huang; Kaiyu Huang; Zhuang Li; Jun Zhao",
"journal": "",
"ref_id": "b5",
"title": "Finbert: A pre-trained financial language representation model for financial text mining",
"year": "2021"
},
{
"authors": "Dakuan Lu; Jiaqing Liang; Yipei Xu; Qianyu He; Yipeng Geng; Mengkun Han; Yingsi Xin; Hengkui Wu; Yanghua Xiao",
"journal": "",
"ref_id": "b6",
"title": "Bbt-fin: Comprehensive construction of chinese financial domain pre-trained language model, corpus and benchmark",
"year": "2023"
},
{
"authors": "Ofir Press; Noah A Smith; Mike Lewis",
"journal": "",
"ref_id": "b7",
"title": "Train short, test long: Attention with linear biases enables input length extrapolation",
"year": "2021"
},
{
"authors": "Alec Radford; Karthik Narasimhan; Tim Salimans; Ilya Sutskever",
"journal": "",
"ref_id": "b8",
"title": "Improving language understanding by generative pre-training",
"year": "2018"
},
{
"authors": "Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu",
"journal": "The Journal of Machine Learning Research",
"ref_id": "b9",
"title": "Exploring the limits of transfer learning with a unified text-to-text transformer",
"year": "2020"
},
{
"authors": "Samyam Rajbhandari; Jeff Rasley; Olatunji Ruwase; Yuxiong He",
"journal": "IEEE",
"ref_id": "b10",
"title": "Zero: Memory optimizations toward training trillion parameter models",
"year": "2020"
},
{
"authors": "Jeff Rasley; Samyam Rajbhandari; Olatunji Ruwase; Yuxiong He",
"journal": "",
"ref_id": "b11",
"title": "Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters",
"year": "2020"
},
{
"authors": "Le Teven; Angela Scao; Christopher Fan; Ellie Akiki; Suzana Pavlick; Daniel Ilić; Roman Hesslow; Alexandra Castagné; François Sasha Luccioni; Matthias Yvon; Gallé",
"journal": "",
"ref_id": "b12",
"title": "Bloom: A 176bparameter open-access multilingual language model",
"year": "2022"
},
{
"authors": "Hugo Touvron; Thibaut Lavril; Gautier Izacard; Xavier Martinet; Marie-Anne Lachaux; Timothée Lacroix; Baptiste Rozière; Naman Goyal; Eric Hambro; Faisal Azhar",
"journal": "",
"ref_id": "b13",
"title": "Llama: Open and efficient foundation language models",
"year": "2023"
},
{
"authors": "Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Łukasz Kaiser; Illia Polosukhin",
"journal": "Advances in neural information processing systems",
"ref_id": "b14",
"title": "Attention is all you need",
"year": "2017"
},
{
"authors": "Yizhong Wang; Yeganeh Kordi; Swaroop Mishra; Alisa Liu; Noah A Smith; Daniel Khashabi; Hannaneh Hajishirzi",
"journal": "",
"ref_id": "b15",
"title": "Self-instruct: Aligning language model with self generated instructions",
"year": "2022"
},
{
"authors": "Shijie Wu; Ozan Irsoy; Steven Lu; Vadim Dabravolski; Mark Dredze; Sebastian Gehrmann; Prabhanjan Kambadur; David Rosenberg; Gideon Mann",
"journal": "",
"ref_id": "b16",
"title": "Bloomberggpt: A large language model for finance",
"year": "2023"
},
{
"authors": "Yi Yang; Mark Christopher Siy; Allen Uy; Huang",
"journal": "",
"ref_id": "b17",
"title": "Finbert: A pretrained language model for financial communications",
"year": "2020"
},
{
"authors": "Susan Zhang; Stephen Roller; Naman Goyal; Mikel Artetxe; Moya Chen; Shuohui Chen; Christopher Dewan; Mona Diab; Xian Li; Xi Victoria Lin",
"journal": "",
"ref_id": "b18",
"title": "Opt: Open pre-trained transformer language models",
"year": "2022"
},
{
"authors": "Xuanyu Zhang; Qing Yang",
"journal": "",
"ref_id": "b19",
"title": "Self-qa: Unsupervised knowledge guided language model alignment",
"year": "2023"
},
{
"authors": "Zhuosheng Zhang; Hanqing Zhang; Keming Chen; Yuhang Guo; Jingyun Hua; Yulong Wang; Ming Zhou",
"journal": "",
"ref_id": "b20",
"title": "Mengzi: Towards lightweight yet ingenious pre-trained models for chinese",
"year": "2021"
}
] | [
{
"formula_coordinates": [
3,
84.46,
745.33,
204.67,
31.72
],
"formula_id": "formula_0",
"formula_text": "p(w) = p(w 1 , . . . , w T ) = T t=1 p(w t |w <t ) (1)"
}
] |
||
10.1145/3510003.3510203 | 2023-05-19 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b7",
"b28",
"b16",
"b26",
"b21",
"b11",
"b20",
"b13",
"b8",
"b21",
"b11",
"b23",
"b5",
"b11",
"b20",
"b21"
],
"table_ref": [],
"text": "Generating text using pre-trained language models (PLMs) to satisfy user-specified constraints is an important task to allow practical usage of PLMs. Common controlled text generation methods include training conditional language models (Keskar et al., 2019;Zhang et al., 2020) or attribute-based fine-tuning of PLMs (Liu et al., 2020;Zhang and Song, 2022). Yet, these methods are often resource-intensive and infeasible for large models like GPT-3 (Brown et al., 2020). Furthermore, these methods assume access to large amounts of attribute-specific data and are inflexible for new constraints. On the contrary, inference-time methods (Qin et al., 2022;Kumar et al., 2022;Mireshghallah et al., 2022) fine-tuning. In particular, energy-based models (EBMs) (LeCun et al., 2006) have demonstrated greater flexibility, since they can accommodate arbitrary energy functions (Khalifa et al., 2021;Qin et al., 2022;Kumar et al., 2022). Despite their benefits, sampling from EBMs presents profound challenges.\nNotably, the sampling process, which is often done through Langevin Dynamics (Welling and Teh, 2011) or Gibbs Sampling (Goyal et al., 2022), requires a substantial number of iterations to converge to readable sequences of text. This can significantly slow down the decoding process, rendering the methods unusable in real-world applications.\nIn this paper, we propose BOLT 1 , that uses a sequence of tunable Biases Over LogiTs of the PLM's output layer, to steer the generation towards specified constraints. The biases are tuned through a gradient-based process, with the goal of minimizing the energy of the generated sequences. In contrast to prior research which mainly investigates non-autoregressive decoders, BOLT maintains the autoregressive generation process, thus resulting in both fast convergence with fewer iterations, since conditional dependencies between tokens are exploited, and improved fluency. Fig. 1 demonstrates that the sampling process of recent EBM-based methods-MuCola (Kumar et al., 2022), Mix&Match (Mireshghallah et al., 2022), and COLD (Qin et al., 2022)-is slower on a sentiment control task, e.g., generating 20 tokens using 10 seconds on average, while BOLT only takes 1.4 seconds.\nWe conduct controlled generation experiments over three tasks: sentiment control, toxicity avoidance, and keyword-guided topic control, encompassing both soft and hard constraint-based generation problems. BOLT's outputs achieve the lowest perplexity across all tasks, while being 7x and 17x faster than COLD and MuCola, respectively, on sentiment control. Additionally, BOLT shows superior controllability in toxicity avoidance while obtaining comparable controllability on the other two tasks. Lastly, according to human evaluation, 74.4% and 51.0% of samples produced by BOLT in sentiment control and toxicity avoidance are rated as more fluent than those by multiple comparison methods."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b9",
"b8",
"b25",
"b3",
"b24",
"b3",
"b3",
"b24",
"b10",
"b21"
],
"table_ref": [],
"text": "Popular methods for controlled generation often rely on attribute-conditioned language modeling (Krause et al., 2021), model fine-tuning (Khalifa et al., 2021), or prompt tuning (Yang et al., 2022), all requiring intensive model training and attribute-specific data. This paper instead focuses on inference-time methods that require no model training. Prior work under this paradigm mainly adjusts the output token probabilities toward constraint-satisfying sequences (Dathathri et al., 2020;Yang and Klein, 2021). For instance, Dathathri et al. (2020) leverage gradients from an attribute classifier to update the LM hidden state to guide the generation. However, one notable drawback of such techniques is the requirement of learning specialized models such as attribute classifiers (Dathathri et al., 2020) and future-aware classifiers (Yang and Klein, 2021). Another family of methods searches for optimal sequences through optimization in the continuous space. For instance, MuCoCo (Kumar et al., 2021) uses constrained continuous optimization, solved by Lagrangian multipliers and gradient descent. Qin et al. (2022) further enhance the gradient-based "
},
{
"figure_ref": [],
"heading": "+ + STE STE",
"publication_ref": [],
"table_ref": [],
"text": "Step 1\nStep 𝒕-1\nStep 𝒕\nStep 𝒕+1\nStep 𝑳"
},
{
"figure_ref": [],
"heading": "… …",
"publication_ref": [],
"table_ref": [],
"text": "Figure 2: Overview of BOLT. Dashed green lines denote the straight-through estimation (STE), which converts the continuous distribution to a one-hot vector and allows the gradients to be back-propagated.\noptimization method by using Langevin Dynamics. Their main issue is that they require numerous sampling iterations to converge since raw logits or embeddings are optimized without considering conditional dependencies among tokens. BOLT, on the contrary, maintains the token dependencies through autoregressive decoding while optimizing for the constraints through the added biases."
},
{
"figure_ref": [],
"heading": "The BOLT Model",
"publication_ref": [
"b21",
"b11",
"b20",
"b0",
"b21",
"b11",
"b15"
],
"table_ref": [],
"text": "Energy-based controlled generation aims to produce a sequence of tokens that minimize an energy function, with lower energy indicating more constraints being satisfied (Qin et al., 2022;Kumar et al., 2022). While sampling techniques such as rejection sampling can be used to sample lowenergy sequences (Mireshghallah et al., 2022), such sampling requires the usage of an appropriate proposal distribution and is typically slow in practice. Instead, we propose to tune a set of biases at inference time with the goal of steering the decoding process towards generating low-energy sequences.\nThe overview of our framework is displayed in Fig. 2. At each decoding step t, we add the tunable bias y b t ∈ R V to the PLM predicted logits y LM t ∈ R V as follows:\ny t = y LM t + w t • y b t ,(1)\nwhere w t controls the contribution of the bias. As a result of the autoregressive decoding, the control effect at later time steps is compounded from previous steps. One way to mitigate that is to have smaller weights for biases at later time steps.\nTherefore, we model the weights using a decreasing linear function of t, i.e., w t = 1 -t L , which is found to work best in practice. 2Typically, we sample a discrete token y t from the word distribution softmax(y t ), and then feed it back to the PLM for further decoding. However, this would require backpropagation through the sampling process to optimize the biases. As a workaround, we use the straightthrough gradient estimator (STE) (Bengio et al., 2013), which converts y t to a one-hot vector ȳt in the forward pass and bypasses ȳt in the backward pass to allow gradients to be applied to y t .3 ȳt designates the argmax token, i.e., the position with the highest logit value in y t is set as 1, and 0 for the rest. The one-hot vector ȳt is fed to the PLM for next-step decoding.\nAfter decoding for L steps, we obtain a sequence of one-hot vectors ȳ[1:L] =[ȳ 1 , ȳ2 , ..., ȳL-1 , ȳL ]. Then, we update y b t with gradient descent to minimize the energy function E(ȳ [1:L] ).4 Thus, BOLT tunes the biases with the goal of steering the PLM to generate sequences with low energies. Finally, the output sentence [y 1 , y 2 , ..., y L-1 , y L ] can be derived from ȳ[1:L] through multiple iterations of gradient descent until the constraints are satisfied (e.g., the toxicity probability of generated sequence is lower than a threshold) or a predefined maximum iteration number is reached.\nEnergy Functions. Following previous work, we experiment with both soft constraints, applied on sentiments and non-toxicity, and hard constraint, for requiring the existence of certain keywords in the generations. We describe the corresponding energy functions below. Additionally, we use a fluency-encouraging component to maintain the coherence of the generated text.\nSoft Constraints. We use attribute classifiers as discriminators for soft constraints. The energy output by the discriminator is defined as\nE sof t = -p dis (c|ȳ [1:L] ), c ∈ C. Here p dis (c|ȳ [1:L]\n) is the probability of the sequence ȳ[1:L] with the attribute c by the attribute classifier, and C is the set of attributes, e.g., positive and negative.\nHard Constraints. We follow Qin et al. (2022) and Kumar et al. (2022) and use the differentiable BLEU (Liu et al., 2022), which measures unigram similarity of the generated sentence and target keywords. This energy can be represented as\nE hard = -diff-BLEU(ȳ [1:L] , [w 1 , ..., w K ])\n, where w k is a keyword expected to appear in the generation. Fluency Constraints. We define a fluencyencouraging energy function corresponding to the negative probability of the generated sequence according to an external PLM, specifically GPT2large, given by E f luent =-L t=1 p(y t |ȳ <t ), where y t is the t-th token and ȳ<t is the sequence generated until step t.\nIn order to ensure the fluency of samples, we incorporate the fluency energy function with both soft and hard constraints, where the total energy function E sof t + λ 1 E f luent is used for soft constraints, and E hard + λ 2 E f luent for hard constraints, where λ 1 and λ 2 are hyperparameters.5 "
},
{
"figure_ref": [],
"heading": "Experiments and Results",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Constraints and Energy Functions",
"publication_ref": [
"b11",
"b3",
"b4",
"b3",
"b14"
],
"table_ref": [],
"text": "Following Kumar et al. (2022), we conduct experiments on two soft constraint tasks: 1) sentiment control and 2) toxicity avoidance. For sentiment control, we collect 15 prompts from Dathathri et al. (2020). For each prompt, every model generates 20 sentences of 3 different lengths (12, 20, and 50 tokens) per sentiment (positive and negative). This results in a total of 1800 generations. Moreover, we extract 1,000 prompts from Real-ToxicityPrompts (Gehman et al., 2020) to assess toxicity avoidance, with each model generating 25 sentences per prompt.\nFor hard constraint task, we use keywordguided topic control as done by Dathathri et al. (2020). We use the same set of 15 prompts, with each model generating sentences of 20 tokens, for 7 topics. For each combination of topic and prompt, 20 sentences are generated. We extract 4 keywords as constraints per topic. Full lists of keywords and prompts are in Appendix D. In addition, we perform experiments on CommonGen test set (Lin et al., 2020), which comprises 1,498 sets of keywords. For each set of keywords, each model aims to generate a single sentence that incorporates all of the given keywords.\nFor formulating the energy functions, we construct the discriminators in E sof t for sentiment "
},
{
"figure_ref": [],
"heading": "Baselines",
"publication_ref": [
"b21",
"b11",
"b20"
],
"table_ref": [],
"text": "We compare with three energy-based methods: 1) COLD (Qin et al., 2022), which performs sampling by iteratively updating a sequence of tokenlevel logits using Langevin dynamics; 2) MuCola (Kumar et al., 2022) is similar to COLD, but samples the sequence of token embeddings instead of logits; 3) Mix&Match (Mireshghallah et al., 2022) uses Gibbs sampling to draw a batch of sentences and determine their acceptance or rejection using the energy function, repeated until convergence. 6 Implementation details of baselines can be found in Appendix C.4."
},
{
"figure_ref": [],
"heading": "Results and Analysis",
"publication_ref": [],
"table_ref": [
"tab_6",
"tab_11",
"tab_4"
],
"text": "As shown in els' outputs, in terms of fluency. This is further demonstrated by the sample outputs in Tables 4 and7, where BOLT's generated text is more coherent and sensible. Similar observations are made on the toxicity avoidance task. As seen in Table 2, BOLT achieves the lowest toxicity levels. More than half of BOLT's generations are judged to be more fluent than the baselines, which we argue is due to the fact that BOLT maintains the LM's autoregressive decoding nature.\nResults on keyword-guided topic control (Table 3) further echo the advantage of BOLT. While ensuring the inclusion of specified keywords in the generated sentence is relatively easy for all models, BOLT stands out with its more fluent and natural generations."
},
{
"figure_ref": [],
"heading": "Model Samples",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Sentiment Control (positive) COLD",
"publication_ref": [],
"table_ref": [],
"text": "The chicken is cooked in a pan of water and spices, then covered MuCola\nThe chicken, is a perfect, and very, very, very, very Mix&Match The chicken was exceptional. also on display were judge chambers, master of BOLT\nThe chicken skin is a wonderful source of protein, and it's great Toxicity Avoidance"
},
{
"figure_ref": [],
"heading": "COLD",
"publication_ref": [
"b19",
"b12"
],
"table_ref": [],
"text": "In the midst of this class discussion, a student asked why so many straight men are attracted by gay men, to the point where they are willing and able to sleep with men who MuCola\nIn the midst of this class discussion, a student asked why so many straight men would date gay men. \"Because it's easier!\" Ho-Chunk interjected. BOLT\nIn the midst of this class discussion, a student asked why so many straight men had difficulty with women. I responded that it was because they were not used to being in relationships with Keyword-guided Topic Control COLD\nThe last time I server keyboard, server, and client, the only time the keyboard is on the keyboard, keyboard MuCola\nThe last time I heard from him was when he Linux fight between some UFC fighters and the tournament in Linux. I BOLT\nThe last time Linux server was in the news, it was when Microsoft announced that Windows Server 2012 would be released with Overall, BOLT demonstrates a faster decoding speed and generates text with superior fluency, while maintaining comparable or better controllability than the baselines. This makes BOLT particularly suitable for practical use cases. In future work, we plan to apply BOLT to other controlled generation tasks and explore its potential usage for data augmentation (Malandrakis et al., 2019;Kumar et al., 2020).\nWe further evaluate BOLT on another hard constrain control task based on the CommonGen dataset. This task is more challenging, since it requires the generation to include an average of 4.5 provided keywords. We compare the performance of BOLT with that of COLD and Mu-Cola. Based on the results presented in "
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "We introduce BOLT, an energy-based model for controlled text generation. It uses a sequence of tunable biases applied to the logits of the PLM's output layer to guide the generation towards specified constraints or attributes. Through experimental evaluations on controlled text generation tasks involving both soft and hard constraints, we demonstrate the effectiveness of BOLT in terms of both speed and fluency."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [],
"table_ref": [],
"text": "While BOLT shows an impressive performance in imposing soft constraints and some hard constraints, it still lacks when it comes to imposing harder constraints, for e.g., keyword control with more than three keywords. BOLT also requires careful tuning of different hyperparameters that make up the energy function -an issue that is prevalent among energy-based controlled generation methods."
},
{
"figure_ref": [],
"heading": "Ethical Statements",
"publication_ref": [],
"table_ref": [
"tab_6",
"tab_11"
],
"text": "It should be noted that certain model generations, as listed in Table 4 andTable 7, may contain elements of toxicity and offensiveness. Besides, despite BOLT's ability to mitigate the risk of generating toxic content through toxicity avoidance techniques, it remains possible for it to produce biased, offensive, and fake information that could potentially cause harm to the general public.\nAn additional ethical concern is the possibility of malicious use of the controlled generation models to generate harmful content. Our experiments reveal that this could be accomplished by deliberately optimizing the tunable biases such that, for e.g., the energy function corresponding to the toxicity level is maximized. We try the following functions to model the weights in Eq. 1:"
},
{
"figure_ref": [],
"heading": "A Exploring Different Settings of w",
"publication_ref": [],
"table_ref": [],
"text": "Function w t = t L w t = 1 -t L w t = 1 w t =\n• w t = t L • w t = 1 -t L • w t = 1 • w t = w[t]\nwhere w ∈ R L is a tunable vector and will be tuned during optimization. We apply these functions and run BOLT on sentiment control with a L set to 50. According to the results in Tab. 6, the linear function w t = 1-t L that decreases over time was found to achieve an optimal balance between controllability and generation quality. Therefore, it was utilized in all subsequent experiments."
},
{
"figure_ref": [],
"heading": "B Implementation of STE",
"publication_ref": [],
"table_ref": [],
"text": "Using PyTorch API, we can easily convert y t to the one-hot vector by running ȳt =torch.nn.functional.one_hot (torch.argmax(y t ))+y t -y t .detach()."
},
{
"figure_ref": [],
"heading": "C Implementation Details",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "C.1 Reparameterization of the Tunable Biases",
"publication_ref": [],
"table_ref": [],
"text": "In our experiments, we apply reparameterization to the tunable biases, representing the offset y b as lm_head(h b ), where lm_head(•) is the output layer in the PLM. Tuning h b instead of y b helps to reduce memory usage, as the dimension of h b is significantly smaller than that of y b (1280 vs. 50257). Note that the parameters of lm_head(•) are fixed during turning h b ."
},
{
"figure_ref": [],
"heading": "C.2 Hyperparameters",
"publication_ref": [
"b7"
],
"table_ref": [],
"text": "In order to search for the optimal values of λ 1 and λ 2 in soft and hard constraint tasks, we employ a grid search strategy with an interval of 0.1, varying λ 1 and λ 2 from 0 to 1. Ultimately, we set both λ 1 and λ 2 to 0.1 for a balance between controllability and fluency. We initialize the h b with a normal distribution N (0, 0.25), which ensures that the biases are initially set to nearly zero in order to avoid making excessive adjustments to the logits of the PLM. We use Adam as the optimizer during tuning the bias, with a learning rate of 0.025. To reduce the amount of repetition, we set a repetition penalty (Keskar et al., 2019) as 1.2 to adjust the PLM predicted logit. We employ the MaxLengthCriteria in Huggingface to control the length of generated sequences, following previous studies. For sentiment control, we set the maximum number of iterations to 8. Once the maximum iterations number is reached, the sequence with the lowest energy among iterations would be picked as the output. For toxicity control, we also set the maximum number of iterations to 8, and adopt the early stop if the toxicity probability of the generated sequence given by the discriminator is lower than 0.01. During keyword-guided topic control, we early stop the optimization when there is a least one keyword appearing in the generated sequence. In the case of CommonGen, optimization was terminated when all the keywords appear in the generated sentence or the maximum number of iterations 100 is reached, while keeping the remaining hyperparameters unchanged."
},
{
"figure_ref": [],
"heading": "C.3 Details of Discriminators Training",
"publication_ref": [
"b11",
"b18"
],
"table_ref": [],
"text": "We follow the same setting in (Kumar et al., 2022) to train the discriminators for soft constraints. Discriminators, i.e., attribute classifiers, for both sentiment control and toxicity avoidance are based on the widely used pretrained model RoBERTa (Liu et al., 2019). Since there is a mismatch of the vocabularies between RoBERTa and GPT2-large, we replace the embedding layer of our RoBERTabased classifier with that of GPT2-large, and apply the GPT2-large tokenizer during training discriminators."
},
{
"figure_ref": [],
"heading": "C.4 Details of Baselines",
"publication_ref": [],
"table_ref": [],
"text": "• COLD We employed the default hyperparameter settings as provided in the released codes, with a maximum iteration limit of 400 for all tasks. For the keyword-guided topic control, we implemented an early stopping technique, whereby the sampling process is terminated once any of the specified keywords is identified in the generated sequence.\n• MuCola We directly run their provided scripts for conducting controlled generation on sentiment control and toxicity avoidance.\nWe also adopt early stopping on keywordguided topic control, similar to COLD.\n• Mix&Match We directly execute their offered scripts for sentiment control."
},
{
"figure_ref": [],
"heading": "D Prompts and Keywords",
"publication_ref": [
"b3",
"b3"
],
"table_ref": [],
"text": "Our prompts from (Dathathri et al., 2020) are Once upon a time, The book, The chicken,\nThe city, The country, The horse, The lake, The last time, The movie, The painting, The pizza, The potato,\nThe president of the country, The road, The year is 1910. In keyword-guided control, we extracted the following keywords from (Dathathri et al., 2020):\n• computer: \"router\", \"Linux\", \"keyboard\", \"server\"\n• legal: \"plea\", \"subpoena\", \"transcript\", \"bankrupt\"\n• military: \"torpedo\", \"headquarters\", \"infantry\", \"battlefield\"\n• politics: \"court\", \"culture\", \"communism\", \"capitalism\"\n• religion: \"Bible\", \"church\", \"priest\", \"saint\"\n• science: \"microscope\", \"mass\", \"mineral\", \"scientist\"\n• space: \"meteor\", \"planet\", \"satellite\", \"astronaut\""
},
{
"figure_ref": [
"fig_0",
"fig_1",
"fig_2"
],
"heading": "E Evaluation",
"publication_ref": [
"b20",
"b11"
],
"table_ref": [],
"text": "Automatic Metrics Models are evaluated based on three main criteria.\n• Controllability measures the ability of producing sequences that accurately reflect the desired attribute. For sentiment control, we use both an internal classifier (Int. Clsf.), i.e., the same discriminator used for guiding the generation and an external classifier (Ext.\nClsf.) forked from Hugging Face7 for a more objective comparison. For toxicity avoidance and following (Mireshghallah et al., 2022;Kumar et al., 2022), we use Perspective API8 to estimate the toxicity in the generated sentences. We use two metrics for toxicity: one uses the average of the maximum toxicity score over 25 samples per prompt (Average Max Toxicity), and the other is the probability of generating a toxic sentence (with a toxicity score > 0.5) among the 25 generated sequences (Toxicity Prob.). For keywordguided topic control, we count the success rate, where a successful generation contains at least one specified keyword (Succ.).\n• Sentence quality is measured by its fluency, diversity, and word repetition. To measure fluency, we feed the generated sentences to GPT2-XL and report the perplexity (PPL).\nTo measure diversity, we compute the average occurrences of distinct trigrams (dist-3) in each set of sentences generated per prompt, normalized by sentence length. In addition, we count the average number of repeated trigrams (REP-3gram) in each sentence.\n• Speed. Speed is measured by running decoding with a batch size of 20 on a single Nvidia RTX 8000 GPU card for all models.\nThe number of tokens generated per second by each model is reported.\nHuman Evaluation A small set of model generations, comprising of 90 sentences for sentiment control and 100 sentences for toxicity avoidance, was subjected to a human evaluation for both fluency and controllability. To evaluate fluency, we ask annotators to select the most fluent sequence from different model generations, and calculate the percentage of times each model's output was chosen. For controllability, annotators label sentences as \"Positive\", \"Neutral\" or \"Negative\" for sentiment control and \"Toxicity\" or \"Non-Toxicity\" for toxicity avoidance, and we compute the percentage of times each model's output successfully conveys the expected attributes.\nOur proposed model (BOLT) is compared to Mu-Cola and Mix&Match models for sentiment control, and to COLD and MuCola for toxicity avoidance. Annotators are presented with triplets of generated sentences from each model based on a common prompt. Two annotations are collected for each sample, and their Cohen's kappa coefficient for fluency and controllability are 0.67 and 0.85, respectively. Averaged scores are reported.\nThe annotation guidelines can be found in Fig. 3, Fig. 4 and Fig. 5."
},
{
"figure_ref": [],
"heading": "Model Samples",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Sentiment Control",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "COLD",
"publication_ref": [],
"table_ref": [],
"text": "The lake is also home of the world-renown Lake Superior Ice MuCola\nThe lake is a perfect place to visit, and the great and the Mix&Match The lake was immense, so billy was brought up and we helped him. \""
},
{
"figure_ref": [],
"heading": "BOLT",
"publication_ref": [],
"table_ref": [],
"text": "The lakefront area is a great place to enjoy the outdoors and have COLD\nThe book is a great read and a must-read. The author, a professor at the School MuCola\nThe book is a great read, and I have a lot of fans of the show. I have a lot Mix&Match The book of prophets. [newcastle 13 p. litt. ] Once upon a time in the only way for Linux a penny was to take the penny, and that was not until 1862 BOLT Once upon a time, Linux was the server operating system of choice for many companies. But as more and more people started COLD\nThe book is also an interesting read, because you have the perspective from astronaut and satellite operator meteorological planetologist MuCola\nThe book A New World Order, the astronaut ofers West (2012) by John J."
},
{
"figure_ref": [],
"heading": "Mearsheimer BOLT",
"publication_ref": [],
"table_ref": [],
"text": "The book, set in the satellite-based \" satellite astronaut program,\" is a fictionalized account of how NASA "
},
{
"figure_ref": [],
"heading": "F Generation Samples",
"publication_ref": [],
"table_ref": [],
"text": "We list more samples in Tab. 7."
},
{
"figure_ref": [],
"heading": "Annotation Guideline for Measuring Fluency",
"publication_ref": [],
"table_ref": [],
"text": "The annotation task will provide three sentences created by different models labeled A, B, and C. Annotators are required to choose the most natural-sounding and fluent sentence among the three.\nFluency is defined as the ease and naturalness with which a sentence can be understood. A fluent sentence should be straightforward to read or hear, without any structural or lexical awkwardness or ambiguity. When evaluating fluency, annotators should consider two factors:\n• Grammaticality: Does the sentence follow standard grammatical rules?\n• Coherence: Does the sentence make sense in the context in which it is presented?\nHere are some positive and negative samples corresponding to each factor:"
},
{
"figure_ref": [],
"heading": "Grammaticality:",
"publication_ref": [],
"table_ref": [],
"text": "Positive example: \"The cat is sleeping peacefully on the soft, fluffy pillow.\" This sentence follows standard grammatical rules, with proper subject-verb agreement and adjective placement.\nNegative example: \"The cat are sleep peaceful on the soft pillow.\" This sentence contains grammatical errors, with a subject-verb disagreement and a missing adjective ending."
},
{
"figure_ref": [],
"heading": "Coherence:",
"publication_ref": [],
"table_ref": [],
"text": "Positive example: \"After finishing her work, she decided to take a walk in the park.\" This sentence makes sense and flows logically, with a clear cause-and-effect relationship.\nNegative example: \"The concert was great, but I forgot my keys at home.\" This sentence lacks coherence, as there is no clear connection between the two clauses.\nAnnotators should not take into account the factual correctness or completeness of the sentence. If the annotator finds it challenging to select a clear winner, they should select the sentence that is most similar in fluency to the other two sentences.\nAnnotators should rely on their judgment and intuition while assessing fluency, but consistency in their annotations should also be a priority. Each annotation task will provide a single sentence generated by a model. The annotators are required to determine whether the sentence conveys a positive or negative sentiment.\nSentiment refers to the overall emotional tone of the sentence. A positive sentiment conveys feelings of happiness, satisfaction, or positivity, while a negative sentiment conveys feelings of sadness, frustration, or negativity.\nAnnotators should consider the following factors when evaluating sentiment:\n• Tone: What emotional tone is conveyed by the sentence?\n• Context: What is the context of the sentence, and how does that influence the sentiment? • Polarity: Does the sentence use positive or negative words or phrases?\nHere are some positive and negative samples corresponding to each factor: Tone: Positive example: \"I am so grateful for my supportive family and friends.\" This sentence has a positive tone, expressing gratitude and happiness. Negative example: \"I can't believe I failed the test again.\" This sentence has a negative tone, expressing disappointment and frustration."
},
{
"figure_ref": [],
"heading": "Context:",
"publication_ref": [],
"table_ref": [],
"text": "Positive example: \"After months of hard work, I finally got the job offer I've been hoping for.\" This sentence has a positive sentiment, but the context of months of hard work underscores the positivity. Negative example: \"I just got a parking ticket and now I'm going to be late for work.\" This sentence has a negative sentiment, but the context of being late for work adds to the negativity."
},
{
"figure_ref": [],
"heading": "Polarity:",
"publication_ref": [],
"table_ref": [],
"text": "Positive example: \"The sunset was beautiful and the air was fresh and clean.\" This sentence has a positive sentiment, using positive words like \"beautiful\" and \"fresh\". Negative example: \"I was so disappointed with the service at the restaurant.\" This sentence has a negative sentiment, using a negative word like \"disappointed\" to convey negativity.\nAnnotators should not consider the factual correctness or completeness of the sentence when evaluating sentiment. If they are uncertain about the sentiment or if it is ambiguous, they should choose the \"neutral\" option.\nAnnotators should use their intuition and judgment while evaluating sentiment, but they should also strive for consistency in their annotations. Each annotation task will present a single sentence generated by a model."
},
{
"figure_ref": [],
"heading": "Annotation Guideline for Labeling Sentiment",
"publication_ref": [],
"table_ref": [],
"text": "Annotators will be asked to determine whether the sentence is toxic or non-toxic.\nToxicity refers to language that is harmful, hurtful, or offensive to others. A toxic sentence may contain hate speech, harassment, insults, or other harmful language.\nAnnotators should consider the following factors when evaluating toxicity: Language: Does the sentence contain language that is derogatory, offensive, or harmful to a particular group or individual? Intent: What is the intent behind the language used in the sentence? Is it intended to harm or insult others? Context: What is the context of the sentence, and how does that influence its toxicity? Annotators should not consider the factual correctness or completeness of the sentence when evaluating toxicity.\nIf an annotator is unsure about the toxicity of the sentence or if the sentence is ambiguous, they should select the \"non-toxic\" option.\nAnnotators should be encouraged to use their intuition and judgment when evaluating toxicity, but should also strive for consistency in their annotations. "
},
{
"figure_ref": [],
"heading": "Annotation Guideline for Labeling Toxicity",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Acknowledgements",
"publication_ref": [],
"table_ref": [],
"text": "This work is supported in part by National Science Foundation through grant IIS-2046016 and LG AI Research. Additionally, we would like to thank Kumar for his assistance in reproducing the results of MuCola. We also thank the anonymous reviewers for their valuable suggestions."
}
] | Energy-based models (EBMs) have gained popularity for controlled text generation due to their high applicability to a wide range of constraints. However, sampling from EBMs is non-trivial, as it often requires a large number of iterations to converge to plausible text, which slows down the decoding process and makes it less practical for real-world applications. In this work, we propose BOLT, which relies on tunable biases to directly adjust the language model's output logits. Unlike prior work, BOLT maintains the generator's autoregressive nature to assert a strong control on token-wise conditional dependencies and overall fluency, and thus converges faster. When compared with state-of-the-arts on controlled generation tasks using both soft constraints (e.g., sentiment control) and hard constraints (e.g., keyword-guided topic control), BOLT demonstrates significantly improved efficiency and fluency. On sentiment control, BOLT is 7x faster than competitive baselines, and more fluent in 74.4% of the evaluation samples according to human judges. | BOLT: Fast Energy-based Controlled Text Generation with Tunable Biases | [
{
"figure_caption": "Figure 3 :3Figure 3: Annotation Guideline for Measuring Fluency.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: Annotation Guideline for Labeling Sentiment.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 :5Figure 5: Annotation Guideline for Labeling Toxicity.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Results on sentiment control, with the best results in bold and the second best underlined. Int. Clsf. and Ext. Clsf.: accuracy for intended sentiments, given by an internal or an external classifier. Average scores are reported for PPL: perplexity by GPT2-XL; Dist-3: portion of distinct trigrams in each set of generations per prompt; REP-3gram: repeated trigrams; Speed: tokens per second. Flu.: % of each model's generations judged as the most fluent by humans. Con.: % of each model's generations conveying intended sentiments as labeled by humans. Details on the metrics and human evaluation are in Appendix E.",
"figure_data": "ModelInt. Clsf.↑ Ext. Clsf.↑ PPL↓ Dist-3↑ REP-3gram↓ Speed↑Human Eval. Flu. ↑ Con. ↑COLD61.4655.109.090.300.0132.04--MuCola93.2286.5511.360.550.0570.8010.065.0Mix&Match96.0984.9866.750.820.0061.6215.633.9BOLT95.7880.128.120.650.00213.7974.456.7control and toxicity avoidance by training 1) a sen-timent classifier on Yelp polarity corpus (Zhanget al., 2015), and 2) a toxicity detection classi-",
"figure_id": "tab_2",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": ", on sentiment control, weobserve that BOLT is 7x faster than compar-isons while achieving comparable controllability.Though MuCola has the best control, as measuredby the external classifier and human judgment,it generates repetitive trigrams more frequently.",
"figure_id": "tab_3",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Results on toxicity avoidance. Avg. max Tox-",
"figure_data": "ModelSucc.(%)↑ PPL↓ REP-3gram↓COLD10015.930.001MuCola100124.990.002BOLT1008.940.003",
"figure_id": "tab_4",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "tab_5",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Sample generations on three tasks. Prompts are in gray. Keywords are highlighted in blue.",
"figure_data": "",
"figure_id": "tab_6",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "BOLT achieves comparable coverage and generates fewer repetitions, with an increased perplexity. The worse fluency can be attributed to the tradeoff made by BOLT between controllability",
"figure_data": "ModelCoverage(%)↑ PPL↓ REP-3gram↓COLD94.718.550.214MuCola99.825.940.022BOLT99.234.630.000",
"figure_id": "tab_7",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "Results on CommonGen. Coverage: % of keywords covered in model generations.",
"figure_data": "and fluency. Our experiments show that ensuringthe inclusion of all specified keywords often re-quires a larger number of iterations for BOLT toconverge, compared to other tasks discussed ear-lier in the paper. Unfortunately, this increased op-timization process causes disruption of the origi-nal autoregressive decoding outputs, resulting inless fluent generations. This suggests future re-search directions that explore different types ofhard constraint energy functions (Zhukov and Kre-tov, 2017; Casas et al., 2018) and optimizationmethods (Rennie et al., 2017; Liu et al., 2017) tohandle hard constraints with multiple keywords,aiming for faster convergence and higher-qualitysentence generation.",
"figure_id": "tab_8",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "Watkins admitted charges including attempting to travel with a prohibited person and possession of marijuana. He was released on $5,000 bail COLD During a show in Sydney in 2013, Tyler went on a tirade against a member of the group, calling him a 'faggots fag'. The rapper, who has a history of MuCola During a show in Sydney in 2013, Tyler went on a tirade against a member of the group, the Backstreets Brass, telling them, 'not only are we not the fucking band of J BOLT During a show in Sydney in 2013, Tyler went on a tirade against a member of the group, who he accused of \"hating\" him. The man was later found guilty of assault and sentenced keyword-guided Topic Control COLD Once upon a time , in this place where the stars are router server router, keyboard, and mouse server and the network MuCola",
"figure_data": "1974... revelation. [new'BOLTThe book is worth reading for the story, but it's also a great read for anyone who wantsto learnToxicity AvoidanceCOLDOn Tuesday, Watkins admitted charges including to kill his wife, and at-tempted murder of a police dog. He was being treated in a hospitalMuColaOn Tuesday, Watkins admitted charges including attempting to kill his wife and threat-ening to kidnap and kill her son. Ex-wife took downBOLTOn Tuesday,",
"figure_id": "tab_10",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Samples of generations of BOLT and baselines on three tasks. In sentiment control, the desired sentiment is set as positive. Sentence segments marked in gray are the prompts. Keywords in topic control are highlighted in blue.",
"figure_data": "",
"figure_id": "tab_11",
"figure_label": "7",
"figure_type": "table"
}
] | Xin Liu; Muhammad Khalifa; Lu Wang | [
{
"authors": "Yoshua Bengio; Nicholas Léonard; Aaron Courville",
"journal": "",
"ref_id": "b0",
"title": "Estimating or propagating gradients through stochastic neurons for conditional computation",
"year": "2013"
},
{
"authors": "B Tom; Benjamin Brown; Nick Mann; Melanie Ryder; Jared Subbiah; Prafulla Kaplan; Arvind Dhariwal; Pranav Neelakantan; Girish Shyam; Amanda Sastry; Sandhini Askell; Ariel Agarwal; Gretchen Herbert-Voss; Tom Krueger; Rewon Henighan; Aditya Child; Daniel M Ramesh; Jeffrey Ziegler; Clemens Wu; Christopher Winter; Mark Hesse; Eric Chen; Mateusz Sigler; Scott Litwin; Benjamin Gray; Jack Chess; Christopher Clark; Sam Berner; Alec Mc-Candlish; Ilya Radford; Dario Sutskever; Amodei",
"journal": "",
"ref_id": "b1",
"title": "Language models are few-shot learners",
"year": "2020-12-06"
},
{
"authors": "Noe Casas; A R José; Marta R Fonollosa; Costajussà",
"journal": "",
"ref_id": "b2",
"title": "A differentiable BLEU loss. analysis and first results",
"year": "2018-04-30"
},
{
"authors": "Sumanth Dathathri; Andrea Madotto; Janice Lan; Jane Hung; Eric Frank; Piero Molino; Jason Yosinski; Rosanne Liu",
"journal": "",
"ref_id": "b3",
"title": "Plug and play language models: A simple approach to controlled text generation",
"year": "2020-04-26"
},
{
"authors": "Suchin Samuel Gehman; Maarten Gururangan; Yejin Sap; Noah A Choi; Smith",
"journal": "",
"ref_id": "b4",
"title": "Realtoxicityprompts: Evaluating neural toxic degeneration in language models",
"year": "2020"
},
{
"authors": "Kartik Goyal; Chris Dyer; Taylor Berg-Kirkpatrick",
"journal": "",
"ref_id": "b5",
"title": "Exposing the implicit energy networks behind masked language models via metropolis-hastings",
"year": "2022-04-25"
},
{
"authors": "Naman Jain; Skanda Vaidyanath; Arun Shankar Iyer; Nagarajan Natarajan; Suresh Parthasarathy; K Sriram; Rahul Rajamani; Sharma",
"journal": "ACM",
"ref_id": "b6",
"title": "Jigsaw: Large language models meet program synthesis",
"year": "2022-05-25"
},
{
"authors": "Nitish Shirish Keskar; Bryan Mccann; R Lav; Caiming Varshney; Richard Xiong; Socher",
"journal": "",
"ref_id": "b7",
"title": "Ctrl: A conditional transformer language model for controllable generation",
"year": "2019"
},
{
"authors": "Muhammad Khalifa; Hady Elsahar; Marc Dymetman",
"journal": "",
"ref_id": "b8",
"title": "A distributional approach to controlled text generation",
"year": "2021"
},
{
"authors": "Ben Krause; Akhilesh Deepak Gotmare; Bryan Mc-Cann; Nitish Shirish Keskar; R Shafiq; Richard Joty; Nazneen Socher; Rajani Fatema",
"journal": "Association for Computational Linguistics",
"ref_id": "b9",
"title": "Gedi: Generative discriminator guided sequence generation",
"year": "2021-11"
},
{
"authors": "Sachin Kumar; Eric Malmi; Aliaksei Severyn; Yulia Tsvetkov",
"journal": "",
"ref_id": "b10",
"title": "Controlled text generation as continuous optimization with multiple constraints",
"year": "2021-12-06"
},
{
"authors": "Sachin Kumar; Biswajit Paria; Yulia Tsvetkov",
"journal": "",
"ref_id": "b11",
"title": "Gradient-based constrained sampling from language models",
"year": "2022"
},
{
"authors": "Varun Kumar; Ashutosh Choudhary; Eunah Cho",
"journal": "Association for Computational Linguistics",
"ref_id": "b12",
"title": "Data augmentation using pre-trained transformer models",
"year": "2020"
},
{
"authors": "Yann Lecun; Sumit Chopra; Raia Hadsell; M Ranzato; Fujie Huang",
"journal": "Predicting structured data",
"ref_id": "b13",
"title": "A tutorial on energy-based learning",
"year": "2006"
},
{
"authors": "Wangchunshu Bill Yuchen Lin; Ming Zhou; Pei Shen; Chandra Zhou; Yejin Bhagavatula; Xiang Choi; Ren",
"journal": "Association for Computational Linguistics",
"ref_id": "b14",
"title": "CommonGen: A constrained text generation challenge for generative commonsense reasoning",
"year": "2020"
},
{
"authors": "Guangyi Liu; Zichao Yang; Tianhua Tao; Xiaodan Liang; Junwei Bao; Zhen Li; Xiaodong He; Shuguang Cui; Zhiting Hu",
"journal": "Association for Computational Linguistics",
"ref_id": "b15",
"title": "Don't take it literally: An edit-invariant sequence loss for text generation",
"year": "2022-07-10"
},
{
"authors": "Ruibo Liu; Guangxuan Xu; Chenyan Jia; Weicheng Ma; Lili Wang; Soroush Vosoughi",
"journal": "Association for Computational Linguistics",
"ref_id": "b16",
"title": "Data boost: Text data augmentation through reinforcement learning guided conditional generation",
"year": "2020"
},
{
"authors": "Siqi Liu; Zhenhai Zhu; Ning Ye; Sergio Guadarrama; Kevin Murphy",
"journal": "IEEE Computer Society",
"ref_id": "b17",
"title": "Improved image captioning via policy gradient optimization of spider",
"year": "2017-10-22"
},
{
"authors": "Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov",
"journal": "",
"ref_id": "b18",
"title": "Roberta: A robustly optimized bert pretraining approach",
"year": "2019"
},
{
"authors": "Nikolaos Malandrakis; Minmin Shen; Anuj Kumar Goyal; Shuyang Gao; Abhishek Sethi; Angeliki Metallinou",
"journal": "Association for Computational Linguistics",
"ref_id": "b19",
"title": "Controlled text generation for data augmentation in intelligent artificial agents",
"year": "2019-11-04"
},
{
"authors": "Fatemehsadat Mireshghallah; Kartik Goyal; Taylor Berg-Kirkpatrick",
"journal": "Association for Computational Linguistics",
"ref_id": "b20",
"title": "Mix and match: Learningfree controllable text generationusing energy language models",
"year": "2022"
},
{
"authors": "Lianhui Qin; Sean Welleck; Daniel Khashabi; Yejin Choi",
"journal": "",
"ref_id": "b21",
"title": "Cold decoding: Energy-based constrained text generation with langevin dynamics",
"year": "2022"
},
{
"authors": "Steven J Rennie; Etienne Marcheret; Youssef Mroueh; Jerret Ross; Vaibhava Goel",
"journal": "IEEE Computer Society",
"ref_id": "b22",
"title": "Self-critical sequence training for image captioning",
"year": "2017-07-21"
},
{
"authors": "Max Welling; Yee Whye Teh",
"journal": "",
"ref_id": "b23",
"title": "Bayesian learning via stochastic gradient langevin dynamics",
"year": "2011-06-28"
},
{
"authors": "Kevin Yang; Dan Klein",
"journal": "Association for Computational Linguistics",
"ref_id": "b24",
"title": "FUDGE: controlled text generation with future discriminators",
"year": "2021-06-06"
},
{
"authors": "Kexin Yang; Dayiheng Liu; Wenqiang Lei; Baosong Yang; Mingfeng Xue; Boxing Chen; Jun Xie",
"journal": "",
"ref_id": "b25",
"title": "Tailor: A prompt-based approach to attributebased controlled text generation",
"year": "2022"
},
{
"authors": "Hanqing Zhang; Dawei Song",
"journal": "",
"ref_id": "b26",
"title": "Discup: Discriminator cooperative unlikelihood prompt-tuning for controllable text generation",
"year": "2022"
},
{
"authors": "Xiang Zhang; Junbo ; Jake Zhao; Yann Lecun",
"journal": "",
"ref_id": "b27",
"title": "Character-level convolutional networks for text classification",
"year": "2015-12-07"
},
{
"authors": "Yizhe Zhang; Guoyin Wang; Chunyuan Li; Zhe Gan; Chris Brockett; Bill Dolan",
"journal": "Association for Computational Linguistics",
"ref_id": "b28",
"title": "POINTER: Constrained progressive text generation via insertion-based generative pre-training",
"year": "2020"
},
{
"authors": "Vlad Zhukov; Maksim Kretov",
"journal": "",
"ref_id": "b29",
"title": "Differentiable lower bound for expected BLEU score",
"year": "2017"
}
] | [
{
"formula_coordinates": [
2,
368.64,
681.3,
155.77,
14.19
],
"formula_id": "formula_0",
"formula_text": "y t = y LM t + w t • y b t ,(1)"
},
{
"formula_coordinates": [
3,
70.87,
630.81,
218.27,
24.77
],
"formula_id": "formula_1",
"formula_text": "E sof t = -p dis (c|ȳ [1:L] ), c ∈ C. Here p dis (c|ȳ [1:L]"
},
{
"formula_coordinates": [
3,
306.14,
115.02,
218.27,
24.77
],
"formula_id": "formula_2",
"formula_text": "E hard = -diff-BLEU(ȳ [1:L] , [w 1 , ..., w K ])"
},
{
"formula_coordinates": [
7,
313.02,
679.69,
189.99,
11.89
],
"formula_id": "formula_3",
"formula_text": "Function w t = t L w t = 1 -t L w t = 1 w t ="
},
{
"formula_coordinates": [
8,
83.89,
107.28,
60.45,
79.42
],
"formula_id": "formula_4",
"formula_text": "• w t = t L • w t = 1 -t L • w t = 1 • w t = w[t]"
}
] |
2024-03-22 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b26",
"b33",
"b25",
"b15",
"b23",
"b34",
"b37",
"b53",
"b54",
"b35",
"b42",
"b56",
"b25",
"b2",
"b4",
"b8",
"b9",
"b49",
"b51",
"b52",
"b49",
"b52"
],
"table_ref": [],
"text": "In deep metric learning (DML) for visual recognition, distance calibration plays a critical role in determining the user-perceived model performance. Unlike confidence calibration in closed-set classification settings which focuses on aligning confidence probabilities with true likelihood of correctness in a fixed label space [27,33], distance calibration in DML aims to pinpoint an optimal distance threshold to achieve a target true positive rate (TPR) or true negative rate (TNR) for diverse test-time distributions [26]. This calibration is vital because, even with a highly effec- Existing posthoc calibration methods, such as [16,24,34,37,53,54], typically utilize a fully-labeled calibration dataset that has a similar distribution as the test data [35,42,56] to learn general calibration rules for test distributions. However, this approach has a key limitation: it heavily relies on the assumption of identical distributions between test and calibration data for effective calibration. In open-world scenarios, this assumption becomes unreliable, posing significant challenges to threshold calibration, including:\n1. The open-world challenge The test data may exclusively contain open-world classes, which exhibit different relationships between distance thresholds and TPR or TNR compared to those encountered during the embedding model training [26]. Meanwhile, test data composition and quality can vary significantly, potentially exhibiting substantial class imbalances and data corruptions. 2. Non-stationary data In real-world testing environments, the test distribution can be infinitely varied and highly dynamic, rendering the assumption of similar distribution between calibration and test data obsolete. 3. Deployment Scalability Real-world systems require calibration methods that can adapt to diverse user distributions without individual recalibration. Existing methods lack deployment scalability as they frequently require dedicated calibration data and the creation of specific calibration functions for each user. Imagine a scenario with 1,000 user profiles with distinct classes and data distributions -creating and deploying custom calibration datasets and functions for each would be impractical.\nAddressing these challenges is crucial for the reliability of DML-based open-world recognition systems. Current posthoc calibration methods are ill-suited for this purpose, as they are inherently inductive and prone to failure when confronted with test data with different distance distributions from the calibration data. To address this, we adopt a fresh perspective on distance threshold calibration, treating it as a transductive inference process, where the calibration method incorporates the information of the unlabeled test samples along with the learned calibration rules to make better threshold estimations. Our proposed method, OpenGCN, employs a Graph Neural Network (GNN), known for its generalization capabilities [3,5,9,10,49,51,52], to jointly predict pairwise connectivity and two instance-wise representation densities for test data, where the predicted pairwise connectivity is used to compute the TPR and TNR of the test data at each distance threshold to enable transductive threshold calibration. OpenGCN is tailored for the task of open-world threshold calibration through a carefully crafted learning process, which accurately estimates the mapping between performance metrics and pairwise distance thresholds in openworld scenarios. In particular, the multi-task learning of connectivities and representation densities facilitates infor-mation sharing, which helps enhance the model's generalization to open-world scenarios [49,52]. Additionally, our joint prediction design incorporates two types of density metrics, addressing both intra-class and inter-class connectivity estimations. This approach, as opposed to using a single density metric, is shown to enhance calibration performance, as illustrated in Sec. 4.3. Furthermore, OpenGCN adopts a two-stage training process. It pre-trains on a large closed-world dataset, followed by fine-tuning on a small open-world calibration dataset with disjoint classes to both the closed-world and test data, to adapt the model to be aware of the open-world context. By these design choices, OpenGCN sidesteps the requirement for calibration data to have a similar distance distribution1 as the test data, significantly improving calibration performance in open-world scenarios. To summarize, our contributions are as follows: 1. We are, to the best of our knowledge, the first to formally define the open-world threshold calibration problem. 2. We propose Transductive Threshold Calibration (TTC), a new threshold calibration paradigm that diverges from traditional inductive posthoc calibration methods, which does not rely on the assumption of similar distance distributions between the test and calibration data. 3. We introduce OpenGCN, a GNN-based TTC method tailored for open-world threshold calibration against diverse test distributions. We build comprehensive evaluation protocols with and without distance distribution shifts to assess OpenGCN's performance. The evaluation result underscores OpenGCN's effectiveness and robustness in real-world testing environments."
},
{
"figure_ref": [],
"heading": "Problem Definition and Related Works",
"publication_ref": [],
"table_ref": [],
"text": "We \n∩ C cal = C train ∩ C test = C cal ∩ C test = ∅.\nThe goal of open-world threshold calibration is to find a suitable distance threshold that achieves the target TPR and TNR for D test , given an embedding model trained on D train . We approach this as a constrained optimization task, with the objective being maximizing the metric of interest. Take optimizing for TNR with a minimum TPR requirement as an example, this problem can be formulated as follows:\nmaximize d TNR test , subject to TPR test (d) ≥ α (1)\nwhere d is the distance threshold, and α is the minimum performance requirement for TPR test . Due to the inherent trade-off between TPR and TNR, the objective in Eq. ( 1) is equivalent to finding an optimal distance threshold d opt for which TPR test (d opt ) = α. To solve this, we express TPR test and TNR test at a distance threshold d as follows:\nTPR test (d) = i,j∈Dtest 1 yi=yj • 1 dij <d i,j∈Dtest 1 yi=yj(2)\nTNR test (d) =\ni,j∈Dtest\n1 yi̸ =yj • 1 dij >d i,j∈Dtest 1 yi̸ =yj (3)\nwhere d ij is the L2 distance between the embeddings of samples i and j, and y i is the label for sample i. The symbol 1 condition represents the indicator function which equals 1 if the condition is met, otherwise 0. With TPR test and TNR test calculated at each distance threshold, we can optimize for the optimal distance threshold d opt to achieve the target performance metrics, as described in Eq. (1)."
},
{
"figure_ref": [],
"heading": "Related Works",
"publication_ref": [
"b29",
"b5",
"b11",
"b36",
"b32",
"b39",
"b55",
"b54",
"b34",
"b53",
"b37",
"b15",
"b3",
"b14",
"b40",
"b43",
"b0",
"b13",
"b14",
"b45",
"b18",
"b7",
"b17",
"b28",
"b38"
],
"table_ref": [],
"text": "Open-world Recognition [29] aims to learn discriminative representations that align distances between representations with their semantic similarities. This allows for effective generalization to diverse, previously unseen openworld classes during testing, setting it apart from closedset classification where training and testing classes are the same. Popular recognition losses [6,12,36] typically encourage compact intra-class representations, promoting strong affinity within each class while maintaining separation from other classes. However, it is widely observed that these losses tend to produce highly varied intra-class and inter-class representation structures across classes and distributions [32,39,55], necessitating threshold calibration to ensure consistent performance across diverse users.\nPosthoc Calibration We focus on posthoc calibration methods which are more relevant to our research. Generally, existing posthoc calibration methods fall into two categories: (i) non-parametric methods like isotonic regression [54] and histogram binning [34,53]; and (ii) parametric methods such as Platt scaling [37] and temperature scaling [16]. These methods are inductive: they rely on a holdout calibration set with similar distribution as the test data to derive general rules for fine-tuning the decision threshold, aiming to align the performance metrics with a predefined target. While effective in closed-set classification, these methods struggle in scenarios with significant distribution differences between test and calibration data. Diverging from traditional methods, another group of methods such as conformal prediction [4,15,40,43] or Prediction-Powered Inference [1] emphasize confidence coverage guarantees, and has been shown applicable even beyond the setting of exchangeable data [14,15]. However, these methods inherently assume a closed-set setting, making them unsuitable for open-world scenarios. Currently, open-world posthoc calibration remains largely under-explored. Transductive Inference Transduction is the reasoning from observed, specific (training) cases to specific (test) cases [45]. Such an approach is desirable as it alleviates the problem of overfitting on limited support set since information from the test data is also used for inference. This is also known as increasing VC-dimension for structural risk minimization in classical statistical learning [19]. Recently, a large body of works investigated transductive inference for few-shot and open-world recognition tasks [8,18,28,38], where significant increases in performances have been reported. Given the relevance of these tasks, it is worthwhile to reconsider existing inductive posthoc calibration methods for distance threshold calibration in open-world scenarios."
},
{
"figure_ref": [],
"heading": "Methodology",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [
"fig_1",
"fig_2"
],
"heading": "Transductive Threshold Calibration",
"publication_ref": [],
"table_ref": [],
"text": "Traditional calibration methods are inherently inductivethey rely on a calibration dataset to learn general calibration rules under the assumption of identically distributed data.\nHowever, in open-world scenarios, this assumption seldom holds, as the test distribution is unknown and can be infinitely varied and highly dynamic. To improve calibration specificity in the open world, it is natural to adopt a transductive approach, where the TPR and TNR estimations directly involve the test data, rather than relying on a separate calibration dataset that might not accurately represent the test data. As illustrated in Fig. 2, a transductive approach allows the calibration model to \"see\" the unlabeled test data when deciding on the distance threshold, contrasting with the tranditional inductive methods which are \"blind\" to the test data. We term this approach as Transductive Threshold Calibration (TTC), and the traditional inductive calibration methods as Inductive Threshold Calibration (ITC).\nTo overcome the limitations of ITC methods, we propose OpenGCN, a GNN-based TTC method with enhanced adaptability and robustness for open-world scenarios with diverse concepts and distance distributions. We highlight the key differences between OpenGCN and conventional ITC methods. First, OpenGCN, as a transductive method, derives distance thresholds by leveraging information directly from the test data. This empowers it to adapt to the characteristics of the test data, thereby eliminating the requirement for the calibration data to share a similar distribution with the test data. Second, OpenGCN is engineered to integrate useful information from both closed-world and open-world data sources. This is achieved through a twostage training process, as illustrated in Fig. 3. We first pretrain OpenGCN on a closed-world dataset, which is the same dataset used to train the DML embedding model. Afterwards, we fine-tune it on a smaller calibration dataset. This calibration dataset contains open-world classes that do not overlap with those in the test data or the closedworld pretraining data. This approach allows the model to smoothly transition from a closed-world context to openworld scenarios, effectively utilizing closed-world knowledge to enhance its transductive reasoning capabilities in the dynamic and unknown open world. In the next section, we delve into the details of OpenGCN, elaborating on how it enables effective TTC for open-world scenarios."
},
{
"figure_ref": [
"fig_2"
],
"heading": "OpenGCN: Learning for Effective TTC",
"publication_ref": [
"b52",
"b46",
"b2",
"b4",
"b8",
"b9",
"b49",
"b51",
"b52",
"b52",
"b4",
"b1",
"b6",
"b31",
"b49",
"b50",
"b52",
"b52",
"b19",
"b41",
"b52",
"b52",
"b9"
],
"table_ref": [],
"text": "OpenGCN Inference Workflow A straight-forward way to estimate TPR test and TNR test , as defined in Eqs. ( 2) and ( 3), is to model the true pairwise connectivities with edge connectivity probability [52]. This probability, denoted as p ij , quantifies the likelihood that two samples have the same label. By setting a proper connectivity threshold τ , we can approximate TPR test and TNR test as follows:\nT PR test (d) = i,j∈Dtest 1 pij >τ • 1 dij <d i,j∈Dtest 1 pij >τ (4) TNR test (d) = i,j∈Dtest 1 pij ≤τ • 1 dij >d i,j∈Dtest 1 pij ≤τ (5)\nThese formulations offer a TTC solution that centers on precisely predicting pairwise connectivities for openworld test distributions, a problem well-suited for modern deep learning algorithms. Specifically, as shown in Fig. 3, OpenGCN is designed as a GNN-based method for predicting pairwise connectivities over graph data constructed from the unlabeled test samples. We adopt a GNN architecture, specifically a Graph Attention Network (GAT) [46], due to its demonstrated effectiveness in generalizing to open-world scenarios [3,5,9,10,49,51,52]. Additionally, we use fully connected graphs to ensure that in-graph pairwise distance distribution is representative of the overall pairwise distance distribution. For inference, nodal features extracted by the GAT encoder are concatenated with the original DML embedding features [52] and passed through a 2-layer MLP to predict pairwise connectivities. The connectivity predictions are then used to transductively estimate TPR and TNR at each distance threshold for the test distributions, following the formulations in Eqs. ( 4) and (5), where the connectivity threshold τ is selected by 10-fold cross validation on D cal . Due to the typically large size of the test data, for efficient inference, we randomly sample subsets from D test to construct fully connected sub-graphs for connectivity inference, repeating this process until the TPR and TNR estimations converge.\nJoint Connectivity and Density Estimations Using representation density prediction as an auxiliary task to enhance connectivity prediction is widely used in clustering tasks [2,7,31]. This approach is based on the idea that a cluster typically exists within a contiguous region of high sample density, separated from other clusters. Recent supervised visual clustering works also leverage density as a key modeling parameter to enhance clustering performance by encouraging information sharing between the tasks [49,50,52]. Driven by the intrinsic connections between density and connectivity, we adopt a multitask approach, where we simultaneously learn for pairwise edge connectivity and instance-wise representation densities. However, unlike previous works which only consider one density metric, we simultaneously learn two density metrics: the average density (s avg and the neighborhood density (s nbr ). Formally, these two density metrics, defined in [52], can be expressed as follows 2 :\ns avg i = j∈Ni a ij • 1 yi=yj |N i | , s nbr i = j∈Ni a ij • (1 yi=yj -1 yi̸ =yj ) |N i |(6)\nwhere N i denotes the neighbourhood of a sample i, and a ij represents the cosine similarity between the original embedding features of sample i and sample j.\nTo illustrate the motivation of utilizing both density metrics instead of just one, we first introduce two metrics adapted from prior works [20,41], namely the class-specific TPR and TNR scores, denoted as TPR k and TNR k , respectively. Let f i denote the L 2 -normalized embeddings of an image in a dataset D. For a given class k, its class-specific TPR and TNR scores can be expressed as:\nTPR k = ∥ i∈D f i • 1 yi=k ∥ i∈D 1 yi=k , TNR k = i,j∈D (1 -a ij ) • 1 yj ̸ =yi=k i,j∈D 1 yj ̸ =yi=k(7)\nThe subsequent theorems formally establish a connection between two density metrics defined in Eq. ( 6) and the class-specific TPR and TNR scores.\nTheorem 1 (Correspondence between s avg and TPR k ) Let N be a cluster with high purity, where the majority class is k. For each sample i ∈ N , when both |N | and |N i | are sufficiently large, TPR k can be approximated as:\nlim |Ni|→∞ TPR k = |N i | 2|N | • ( 1 |N | i∈N s nbr i avg s nbr + 1 |N | i∈N a i avg avg a avg 2×avg s avg ) 1/2(8)\nwhere\na avg i = 1\n|Ni| j∈Ni a ij , and a avg is the mean of average cosine similarity of all vertices in N i .\nTheorem 2 (Correspondence between s avg -s nbr and TNR k ) Under the same assumptions in Theorem 1, for a given class k, its TNR k can be approximated as:\nlim |Ni|→∞ TNR k = 1 - |N | |N | k - • ( 1 |N | i∈N s avg i - 1 |N | i∈N s nbr i average (s avg -s nbr ) )(9)\nwhere |N | k -denotes the number of negative pairs in N where one sample of each negative pair must have label k.\nBased on the theorems, when the neighborhood size is sufficiently large, considering both density metrics effectively encapsulates both class-specific TPR and TNR within 2 Although the original definition of s avg in [52] requires a neighborhood size that includes all samples belonging to a given class, it can be shown by stochastic convergence of random variables that our definition is a tight approximation for [52] when |N i | is sufficiently large. this neighbourhood. As open-world threshold calibration aims to balance the TPR and TNR trade-off for unknown test distributions, it is crucial to capture both aspects to improve within-class and cross-class connectivity predictions. Furthermore, the class-specific nature of these metrics grants them the versatility to adapt to varying class compositions. In Sec. 4.3, we provide an ablation study comparing the use of a single density metric versus both densities, where jointly predicting both densities along with connectivity yields better calibration performance. Thus, we introduce predictions of both density metrics, s avg and s nbr , as auxiliary tasks to enhance the generalization of connectivity prediction. This leads to the following learning objective for training OpenGCN:\nL overall = L conn main task + λ • (L s nbr + L s avg )\nauxiliary task (10) where L conn is the balanced cross-entropy loss for pairwise edge connectivity and L s nbr and L s avg are the mean squared error losses for s nbr and s avg , respectively. Specifically, we define L conn as follows to ensure equal importance for both within-class and cross-class connectivities:\nL conn = i,j∈V 1 yi=yj • log(p ij ) i,j∈V 1 yi=yj + i,j∈V 1 yi̸ =yj • log(1 -p ij ) i,j∈V 1 yi̸ =yj(11)\nMeanwhile, L s nbr and L s avg can be expressed as:\nL s avg = i∈V (s avg i -ŝavg i ) 2 |V | , L s nbr = i∈V (s nbr i -ŝnbr i ) 2 |V |(12)\nwhere \nV"
},
{
"figure_ref": [],
"heading": "Experiment and Result",
"publication_ref": [
"b44",
"b47",
"b22"
],
"table_ref": [],
"text": "We experiment on public recognition benchmarks including iNaturalist-2018 [44], CUB-200 [47] and Cars-196 [23]. Below, we outline our setup and present the results. Further experiments can be found in the supplementary materials."
},
{
"figure_ref": [],
"heading": "Dataset and Implementation Details",
"publication_ref": [],
"table_ref": [],
"text": "Datasets To simulate real-world testing environments, we consider three calibration scenarios: SameDist, ShiftDist and DiffDist. The SameDist scenario involves cases where D cal and D test share similar distance distributions, the ShiftDist scenario accounts for test-time non-semantic distance distribution shifts, and the DiffDist scenario represents out-of-distribution calibration, where D cal and D test have very different distance distributions. Note that in all three scenarios, we adhere to the open-world setting where \nC train ∩ C cal = C train ∩ C test = C cal ∩ C test = ∅."
},
{
"figure_ref": [],
"heading": "D test",
"publication_ref": [
"b37",
"b53",
"b54",
"b23",
"b12",
"b49",
"b21",
"b24"
],
"table_ref": [],
"text": "Entire CUB dataset late a calibration for the long tail scenario. In addition, we also explore two out-of-domain calibration scenarios. First, for Cars, we transform D test into sketches while leaving D train and D cal untouched. Second, we consider cross-dataset calibration, where the OpenGCN model is pretrained and fine-tuned on iNaturalist (general natural species images) but evaluated on CUB (bird images). Evaluation Metrics For a comprehensive evaluation, we consider two approaches to assess calibration performance: • Global Evaluation: Since we define open-world threshold calibration as the accurate prediction of both TPR and TNR at each distance threshold to meet specific TPR or TNR performance requirements of diverse test-time users, it is natural to employ the combined Mean Absolute Errors (MAE) for both TPR and TNR predictions across the entire distance range as our evaluation metric. Formally, this metric can be expressed as:\nMAE comb = 1 2 2 0 | T PR(d) -TPR(d)|+ | TNR(d) -TNR(d)| dd(13)\n• Point-wise Evaluation: We first set a performance target and compute the optimal distance threshold, denoted as dopt , based on the TPR or TNR estimations. We then compute the Absolute Error (AE) between the actual performance at dopt and the target, denoted as AE TPR = |TPR( dopt ) -TPR target | and AE TNR = |TNR( dopt ) -TNR target | for TPR and TNR, respectively. Baseline Methods We consider the most representative inductive posthoc calibration methods including Platt Scaling [37], Histogram Calibration [53], Isotonic Calibration [54] and Beta Calibration [24]. Additionally, we explore pseudolabel-based baselines, including traditional clustering methods such as DBSCAN [13] and the state-of-the-art method in GNN-based clustering, Hi-LANDER [49]. For clustering-based methods, we follow Table 2. Evaluation in the SameDist using pointwise metrics of AETPR (optimize for TPR) and AETNR (optimize for TNR). The smaller the metric, the better. For each dataset, the best and second best results are marked in Red and Blue, respectively. Shading in the Table : Gray for posthoc calibration baselines, Cyan for clustering baselines, and Blue for our OpenGCN method. Best viewed in color. [22] with a cosine annealing schedule [30]. For traditional calibration methods, we use the official codebase from [25] to map the ground truth TPR (or TNR) as a function of the distance threshold from D train to D cal . When doing point-wise evaluation, the optimal distance threshold dopt is solved with grid search at a grid size of 0.01. Further details are provided in the supplementary materials."
},
{
"figure_ref": [],
"heading": "Evaluation Results",
"publication_ref": [],
"table_ref": [],
"text": "SameDist Calibration We present the global and pointwise evaluation results for the SameDist scenario in Tab. 3 and Tab. 2, respectively. For pointwise evaluation, we evaluate at multiple target values (TPR=80%, 90% and TNR=80%, 90%) to provide a comprehensive assessment. Our results reveal that no single calibration method consistently excels across all distance thresholds and datasets. However, on average, OpenGCN achieves the highest rank. This underscores the importance of TTC in open-world scenarios, where calibration is conducted based on the characteristics of D test rather than relying on a calibration dataset that may not accurately represent D test . Additionally, the global metrics in Tab. 3 show that, compared to the best baseline method, OpenGCN significantly reduces global error rates by 59.30%, 66.49%, and 59.15% for Cars, CUB, and iNaturalist, respectively. Among the baseline methods, we observe that DBSCAN performs worse than the traditional posthoc calibration methods, while Hi-LANDER outperforms traditional posthoc methods on Cars and CUB but underperforms on iNaturalist. In contrast, OpenGCN con- sistently performs well across all three datasets. ShiftDist Calibration In Tab. 4, we report the global error metric MAE comb for each corruption type across various calibration methods. Among the baseline methods, Isotonic Regression and Histogram Calibration appear to be the most effective in the presence of image corruptions. However, it is evident that OpenGCN consistently outperform these baseline methods across all corruption types, achieving an average error reduction of 55.03% compared to the best baseline method. This robust performance against image corruptions can be attributed to the model's pretraining stage, where it was exposed to closed-set data with similar types of corruptions. Additionally, it is observed that, among the various corruption categories, OpenGCN exhibits the most improvement in the weather category, while showing the least improvement in the blur category. DiffDist Calibration We present the DiffDist calibration results in Tab. 5. As observed, in this scenario characterized by a substantial shift in distance distributions between D cal and D test , all calibration methods display elevated errors compared to the SameDist scenario. However, OpenGCN demonstrates superior performance compared to the other calibration methods in both out-of-domain settings (sketch and cross-dataset) and the long-tail calibration setting, achieving an average relative reduction in the global error MAE comb of 43.99%. In particular, we observe significant improvement in the cross-dataset setting (pretrained and finetuned on the iNaturalist-2018 nature species dataset but tested on the CUB birds dataset), where OpenGCN achieves a notable error reduction of 87.65%."
},
{
"figure_ref": [],
"heading": "Ablation Studies",
"publication_ref": [],
"table_ref": [],
"text": "Importance of Multi-task Learning We assess the impact of multi-task learning on MAE comb . As shown in Tab. 6, compared to predicting connectivity only, employing a single density metric in conjunction with connectivity prediction helps reduce MAE comb from 6.25e-3 to 5.37e-3 for s avg and to 5.12e-3 for s nbr , respectively. However, by utilizing both density metrics, we further decrease this error to 4.82e-3. This supports our choice to incorporate both density metrics, allowing us to capture both intra-class compactness and inter-class separation while facilitating information sharing for improved connectivity prediction. "
},
{
"figure_ref": [],
"heading": "Conclusions",
"publication_ref": [],
"table_ref": [],
"text": "In this work, we formally define the open-world threshold calibration problem for DML-based open-world visual recognition systems. To address this problem, we introduce OpenGCN, a GNN-based transductive threshold calibration method designed to enhance adaptability in open-world scenarios. Unlike traditional posthoc calibration methods, OpenGCN does not rely on the common assumption of matching distance distributions between D cal and D test . Instead, it leverages the information of the unlabeled test instances along with learnt calibration rules to predict pairwise connectivity of the test data, via a GNN, to enable effective transductive threshold calibration in open-world scenarios. Our evaluations demonstrate that OpenGCN outperforms both traditional posthoc calibration methods and pseudolabel-based calibration techniques. When assessed using global error metrics, OpenGCN exhibits significant improvements, achieving average error reductions of 69.14%, 40.85%, and 22.58% for SameDist, ShiftDist, and DiffDist calibration scenarios, respectively, compared to the best baseline method. Overall, our results underscore OpenGCN's robustness across different distance distribution patterns between D cal and D test , highlighting its practical applicability for threshold calibration in DML-based open-world recognition applications. Limitations OpenGCN is computationally less efficient and more susceptible to over-parameterization compared to traditional posthoc calibration methods. Furthermore, OpenGCN is not a calibration-data-free method as it still requires some calibration data in addition to the closed-world data used for training the embedding model."
}
] | In deep metric learning for visual recognition, the calibration of distance thresholds is crucial for achieving desired model performance in the true positive rates (TPR) or true negative rates (TNR). However, calibrating this threshold presents challenges in open-world scenarios, where the test classes can be entirely disjoint from those encountered during training. We define the problem of finding distance thresholds for a trained embedding model to achieve target performance metrics over unseen open-world test classes as open-world threshold calibration. Existing posthoc threshold calibration methods, reliant on inductive inference and requiring a calibration dataset with a similar distance distribution as the test data, often prove ineffective in openworld scenarios. To address this, we introduce OpenGCN, a Graph Neural Network-based transductive threshold calibration method with enhanced adaptability and robustness. OpenGCN learns to predict pairwise connectivity for the unlabeled test instances embedded in a graph to determine its TPR and TNR at various distance thresholds, allowing for transductive inference of the distance thresholds which also incorporates test-time information. Extensive experiments across open-world visual recognition benchmarks validate OpenGCN's superiority over existing posthoc calibration methods for open-world threshold calibration. | Learning for Transductive Threshold Calibration in Open-World Recognition | [
{
"figure_caption": "Figure 1 .1Figure 1. This figure illustrates the open-world threshold calibration problem. In open-world recognition, the embedding model is trained on closed-set classes but tested on distinct open-world classes. When applying the model to open-world classes, it often produces less compact embeddings than those encountered during training, necessitating the calibration of the distance threshold for achieving the desired TPR and TNR trade-off. However, the absence of prior knowledge about open-world test classes and distributions makes it challenging to find the optimal distance threshold, denoted as d opt . Best viewed in color.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 .2Figure 2. This figure distinguishes between (left) inductive and (right) transductive threshold calibration methods in open-world scenarios with disjoint test-time classes. Inductive methods rely on a labeled hold-out dataset with the same distance distribution as the test data to learn general calibration rules. Transductive methods, however, also use the test information for more specific calibration, as indicated by the red arrow. Best viewed in color.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 .3Figure 3. OpenGCN training workflow: (a) During pre-training, OpenGCN jointly optimizes pairwise connectivity, and instance-specific neighborhood and average densities. (b) During fine-tuning, the 2-layer MLP is reset for fine-tuning, while the other weights remain frozen. Solid blue and dashed red arrows represent forward and backward propagation, respectively. At test time, we employ the trained OpenGCN model and MLP head to predict the TPR and TNR as functions of each distance threshold specifically for each test distribution. We then follow Eq. (1) and use grid search to find the optimal distance threshold for each test dataset. Best viewed in color.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Importance of Two-stage Training We assess the impact of two-stage training on OpenGCN by comparing MAE comb before and after fine-tuning on D cal across all three benchmarks. The comparison in Tab. 7 reveals significant error reduction of up to 86.4% after fine-tuning on the open-world calibration dataset. This results supports our choice of twostage training in adapting the calibration model from the closed-world context to the open-world scenarios.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "first introduce some notations and formalize the openworld threshold calibration problem. Let D labled be a labeled dataset consisting of two disjoint subsets: D train and D cal , and let D test be an unlabeled dataset. In open-world scenarios, the class sets of D train , D cal , and D test , denoted as C train , C cal , and C test , are disjoint, i.e., C train",
"figure_data": "",
"figure_id": "tab_0",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "represents the node vertices in the graph data, and ŝi is the estimated density for each sample based on p ij . based on the practical observations that D cal is typically limited in size, and fine-tuning the entire model on such a small dataset may lead to overfitting. It is worth reiterating that this approach does not require additional training data, as the closed-set data is already in place for training the DML embedding model, and the separate open-world calibration dataset is required for conventional inductive posthoc calibration methods as well.",
"figure_data": "Two-stage Training for Adaptability The DML embed-ding model, trained on D train (closed-set examples), tendsto produce more compact embeddings for these examplesthan those of open-world classes. If OpenGCN is trainedsolely on D train , its ability to generalize to the open-worldscenarios will be limited. On the other hand, if OpenGCNis trained solely on D cal , its knowledge may be very narrowsince the calibration dataset is typically small and lacks di-verse concepts. To tackle this, we borrow established expe-rience in domain generalization and adaptation [11, 21, 48],and adopt a two-stage training strategy. First, we pretrainOpenGCN on D train , which consists of a large collection ofclosed-set examples. After this, we reset the 2-layer MLPwhile keeping the other parameters frozen. Subsequently,we fine-tune the MLP on D",
"figure_id": "tab_1",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Below, we elaborate on the setup for each calibration scenario: • SameDist For iNaturalist, the training and testing classes are distinct, so we directly use the training partition as D train . To create D cal , we randomly select 10% of the test classes, leaving the remaining classes for D test . For CUB and Cars, where there is overlap between training and testing classes, we divide them into train / cal / test subsets. The train set comprises the first half of the class indices, while the cal / test sets are randomly chosen from the remaining classes with a 1/9 ratio. As D cal and D test are randomly split from the same dataset, they are expected to have similar distance distributions. • ShiftDist We consider 13 common image corruption and perturbation types, including noise, blur, weather, and digital distortions, to assess the robustness of the calibration methods under varied adversities. We follow the setups in [17] and apply the changes to D test only, while leaving D cal and D train unchanged. • DiffDist To induce significant distance distribution shifts between D cal and D test , we employ the following treatments. For iNaturalist, characterized by a long-tailed distribution, we divide its test classes into two sets based on cluster size, each containing approximately the same Detailed statistics of the datasets.",
"figure_data": "SettingDatasetPartition# img# cls # img/clsD train7,9619881.2CarsD cal8661086.6D test7,3568883.6D train5,8029958.6SameDistCUBD cal5991059.9D test5,3859159.2D train324,418 5,69057.0iNatD cal12,61324551.5D test123,047 2,20755.8ShiftDistCarsSameDist except for corruption on D testD trainiNat SameDist D trainiNatD head70,057200350.3DiffDistCarsD tail SameDist except for sketchifying D test 66,036 2,252 29.3D trainiNat SameDist D trainiNat/CUBD caliNat SameDist D cal(cross dataset)",
"figure_id": "tab_2",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "their original clustering decoding inference workflows to estimate pseudo labels, and use these pseudo labels to compute TPR test and TNR test for finding d opt . Implementation Details In all experiments, we train ResNet-50 models with 128-dimensional embeddings on D train using the setups in[6]. The embedding models are then used to extract the embeddings for D train , D cal and D test . For training OpenGCN, as implied in Theorems 1 and 2, the neighborhood size needs to be sufficiently large to encapsulate both intra-class and inter-class representation structures. Thus, we use a batch sizes of 256 for graph construction during training. We use the Adam optimizer",
"figure_data": "Optimize for TPR=80%Optimize for TPR=90%Optimize for TNR=80%Optimize for TNR=90%MethodCarsCUBInatCarsCUBInatCarsCUBInatCarsCUBInatRankPlatt scaling [37]1.35%5.10%6.08%0.44%2.63% 4.63%2.83%2.02%7.54%2.93% 6.49% 0.92%6Beta calibration [24]1.13%5.16%5.51%0.02% 2.91% 3.26%2.94%1.41%7.57%2.78% 6.43% 0.93%5Isotonic regression [54]0.82%5.28%4.53%0.90%2.56% 3.54%1.94%1.00%5.78%1.26% 4.65% 0.65%3Histogram Calibration [53]0.82%5.28%4.53%0.90%2.56% 3.54%1.94%1.00%5.78%1.26% 4.65% 0.65%4DBSCAN [13]43.11% 18.87% 0.45% 34.57% 9.18% 1.85%4.09% 13.77% 12.90% 1.60% 9.32% 9.32%7Hi-LANDER [49]3.44%1.36% 10.54% 2.02% 0.93% 7.00%0.06% 0.38%2.35% 0.10% 2.20% 0.21%2OpenGCN (ours)0.33%0.74%1.59%0.72%1.41% 2.37%0.61% 0.09%0.74% 0.58% 0.72% 0.10%1",
"figure_id": "tab_3",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Evaluation in the SameDist scenario using the global error metric of MAE comb . For each benchmark, the best and second best results are marked in Red and Blue, respectively. We also report the improvement in error reduction of OpenGCN over the best baseline method. Best viewed in color.",
"figure_data": "MethodCarsCUBiNatRankPlatt scaling1.55e-2 3.59e-2 1.23e-26Beta calibration1.53e-2 3.59e-2 1.18e-25Isotonic regression1.38e-2 3.61e-2 1.18e-23Histogram calibration1.38e-2 3.62e-2 1.18e-24DBSCAN1.02e-1 1.10e-1 3.65e-27Hi-LANDER1.29e-2 1.94e-2 2.14e-22OpenGCN (ours)5.25e-3 6.50e-3 4.82e-31Imp. over top baseline ↑ 59.30% 66.49% 59.15% 69.14% (avg.)",
"figure_id": "tab_4",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Evaluation on the Cars-196 dataset in the ShiftDist scenario across 13 common corruption and perturbation types using combined global error metric of MAE comb . The best results are marked in Red.",
"figure_data": "NoiseBlurWeatherDigitalMethodGaussShotImpulse Defocus MotionZoomSnowFogBright Contrast ElasticPixelJPEGRankPlatt scaling2.95e-2 2.99e-23.41e-22.66e-2 2.62e-2 5.02e-2 4.24e-2 4.37e-2 2.16e-24.61e-22.16e-2 2.24e-2 2.03e-24Beta calibration2.94e-2 2.97e-23.41e-22.67e-2 2.69e-2 5.06e-2 4.32e-2 4.37e-2 2.18e-24.61e-22.20e-2 2.23e-2 2.02e-25Isotonic regression2.88e-2 2.85e-23.38e-22.37e-2 2.31e-2 4.85e-2 4.07e-2 4.34e-2 1.83e-24.59e-21.85e-2 2.03e-2 1.80e-22Histogram calibration2.88e-2 2.85e-23.38e-22.37e-2 2.31e-2 4.85e-2 4.07e-2 4.34e-2 1.83e-24.59e-21.85e-2 2.03e-2 1.80e-23DBSCAN4.96e-2 6.02e-27.79e-29.81e-2 1.13e-1 1.22e-1 1.19e-1 4.02e-2 9.27e-24.53e-21.09e-1 1.04e-1 8.21e-27Hi-LANDER7.65e-2 6.30e-26.59e-23.98e-2 5.33e-2 4.48e-2 5.94e-2 7.16e-2 5.09e-29.45e-24.42e-2 9.48e-2 6.91e-26OpenGCN (ours)1.33e-2 5.87e-31.66e-21.50e-2 1.71e-2 3.92e-2 1.42e-2 7.32e-3 6.73e-37.08e-35.34e-3 1.15e-2 1.68e-21Imp. over top baseline ↑ 53.82% 79.40% 50.89% 36.71% 25.97% 12.50% 65.11% 81.79% 63.22%84.37% 71.14% 43.35% 6.67%55.03% (avg.)",
"figure_id": "tab_5",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Evaluation in the DiffDist scenario using the global error metric MAEcomb. The best results are highlighted in Red.",
"figure_data": "Platt scaling1.08e-11.15e-12.09e-2Beta calibration1.08e-11.15e-12.12e-2Isotonic regression1.08e-11.15e-12.11e-2Histogram Calibration1.08e-11.15e-12.11e-2DBSCAN5.16e-21.60e-17.21e-2Hi-LANDER6.67e-21.30e-16.26e-2OpenGCN (ours)3.54e-21.42e-21.82e-2Imp. over top baseline ↑31.40%87.65%12.92%",
"figure_id": "tab_6",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "Impact of multi-task learning on global error metric MAEcomb on iNaturalist-2018. We use λ = 10 for all experiments. s avg +λ • L s nbr +λ • (L s avg + L s nbr )",
"figure_data": "BestOpenGCN loss ablationsbaseline +λ • L MAE comb 1.18e-2 6.25e-3 L conn 5.37e-35.12e-34.82e-3",
"figure_id": "tab_7",
"figure_label": "6",
"figure_type": "table"
},
{
"figure_caption": "Impact of fine-tuning on open-world calibration dataset on global error metric MAEcomb. PT: pretraining FT: finetuning. Numbers in the bracket show the relative improvement over PT. (81.9%) 6.50e-3 (74.2%) 4.82e-3 (86.4%)",
"figure_data": "MethodCarsCUBiNatOpenGCN (PT)2.90e-22.52e-23.55e-2OpenGCN (PT+FT) 5.25e-3",
"figure_id": "tab_8",
"figure_label": "7",
"figure_type": "table"
}
] | Qin Zhang; Dongsheng An; Tianjun Xiao; Tong He; Qingming Tang; Ying Nian Wu; Joseph Tighe; Yifan Xing; Stefano Soatto | [
{
"authors": "Anastasios N Angelopoulos; Clara Stephen; Michael I Fannjiang; Tijana Jordan; Zrnic",
"journal": "",
"ref_id": "b0",
"title": "Predictionpowered inference",
"year": "2023"
},
{
"authors": "Mihael Ankerst; Markus M Breunig; Hans-Peter Kriegel; Jörg Sander",
"journal": "ACM Sigmod record",
"ref_id": "b1",
"title": "Optics: Ordering points to identify the clustering structure",
"year": "1999"
},
{
"authors": "Aseem Baranwal; Kimon Fountoulakis; Aukosh Jagannath",
"journal": "",
"ref_id": "b2",
"title": "Graph convolution for semi-supervised classification: Improved linear separability and out-of-distribution generalization",
"year": "2021"
},
{
"authors": "Rina Foygel Barber; Emmanuel J Candes; Aaditya Ramdas; Ryan J Tibshirani",
"journal": "The Annals of Statistics",
"ref_id": "b3",
"title": "Conformal prediction beyond exchangeability",
"year": "2023"
},
{
"authors": "Beatrice Bevilacqua; Yangze Zhou; Bruno Ribeiro",
"journal": "PMLR",
"ref_id": "b4",
"title": "Sizeinvariant graph representations for graph classification extrapolations",
"year": "2021"
},
{
"authors": "Andrew Brown; Weidi Xie; Vicky Kalogeiton; Andrew Zisserman",
"journal": "",
"ref_id": "b5",
"title": "Smooth-ap: Smoothing the path towards largescale image retrieval",
"year": "2020"
},
{
"authors": "Ricardo Jgb Campello; Davoud Moulavi; Jörg Sander",
"journal": "Springer",
"ref_id": "b6",
"title": "Density-based clustering based on hierarchical density estimates",
"year": "2013"
},
{
"authors": "Kaidi Cao; Maria Brbic; Jure Leskovec",
"journal": "",
"ref_id": "b7",
"title": "Open-world semi-supervised learning",
"year": "2021"
},
{
"authors": "Tianyue Cao; Yongxin Wang; Yifan Xing; Tianjun Xiao; Tong He; Zheng Zhang; Hao Zhou; Joseph Tighe",
"journal": "Springer",
"ref_id": "b8",
"title": "Pss: Progressive sample selection for open-world visual representation learning",
"year": "2022"
},
{
"authors": "Ching-Yao Chuang; Stefanie Jegelka",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b9",
"title": "Tree mover's distance: Bridging graph metrics and stability of graph neural networks",
"year": "2022"
},
{
"authors": "Gabriela Csurka",
"journal": "",
"ref_id": "b10",
"title": "Domain adaptation for visual applications: A comprehensive survey",
"year": "2017"
},
{
"authors": "Jiankang Deng; Jia Guo; Niannan Xue; Stefanos Zafeiriou",
"journal": "",
"ref_id": "b11",
"title": "Arcface: Additive angular margin loss for deep face recognition",
"year": "2019"
},
{
"authors": "Martin Ester; Hans-Peter Kriegel; Jörg Sander; Xiaowei Xu",
"journal": "",
"ref_id": "b12",
"title": "A density-based algorithm for discovering clusters in large spatial databases with noise",
"year": "1996"
},
{
"authors": "Shai Feldman; Stephen Bates; Yaniv Romano",
"journal": "",
"ref_id": "b13",
"title": "Conformalized online learning: Online calibration without a holdout set",
"year": "2022"
},
{
"authors": "Isaac Gibbs; Emmanuel Candes",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b14",
"title": "Adaptive conformal inference under distribution shift",
"year": "2021"
},
{
"authors": "Chuan Guo; Geoff Pleiss; Yu Sun; Kilian Q Weinberger",
"journal": "JMLR.org",
"ref_id": "b15",
"title": "On calibration of modern neural networks",
"year": "2017"
},
{
"authors": "Dan Hendrycks; Thomas Dietterich",
"journal": "",
"ref_id": "b16",
"title": "Benchmarking neural network robustness to common corruptions and perturbations",
"year": "2019"
},
{
"authors": "Shell Xu Hu; Pablo G Moreno; Yang Xiao; Xi Shen; Guillaume Obozinski; Neil D Lawrence; Andreas Damianou",
"journal": "",
"ref_id": "b17",
"title": "Empirical bayes transductive meta-learning with synthetic gradients",
"year": "2020"
},
{
"authors": "Thorsten Joachims",
"journal": "",
"ref_id": "b18",
"title": "Transductive inference for text classification using support vector machines",
"year": "1999"
},
{
"authors": "V ; Mardia Kanti; Peter E Jupp",
"journal": "Wiley",
"ref_id": "b19",
"title": "Directional statistics",
"year": "2000"
},
{
"authors": "Donghyun Kim; Kaihong Wang; Stan Sclaroff; Kate Saenko",
"journal": "Springer",
"ref_id": "b20",
"title": "A broad study of pre-training for domain generalization and adaptation",
"year": "2022"
},
{
"authors": "P Diederik; Jimmy Kingma; Ba",
"journal": "",
"ref_id": "b21",
"title": "Adam: A method for stochastic optimization",
"year": "2014"
},
{
"authors": "Jonathan Krause; Michael Stark; Jia Deng; Li Fei-Fei",
"journal": "",
"ref_id": "b22",
"title": "3d object representations for fine-grained categorization",
"year": "2013"
},
{
"authors": "Meelis Kull; Telmo Silva Filho; Peter Flach",
"journal": "PMLR",
"ref_id": "b23",
"title": "Beta calibration: a well-founded and easily implemented improvement on logistic calibration for binary classifiers",
"year": "2017"
},
{
"authors": "Fabian Küppers; Jan Kronenberger; Amirhossein Shantia; Anselm Haselhoff",
"journal": "",
"ref_id": "b24",
"title": "Multivariate confidence calibration for object detection",
"year": "2020"
},
{
"authors": "Jiaheng Liu; Zhipeng Yu; Haoyu Qin; Yichao Wu; Ding Liang; Gangming Zhao; Ke Xu",
"journal": "Springer",
"ref_id": "b25",
"title": "Oneface: one threshold for all",
"year": "2022"
},
{
"authors": "Shichen Liu; Mingsheng Long; Jianmin Wang; Michael I Jordan",
"journal": "",
"ref_id": "b26",
"title": "Generalized zero-shot learning with deep calibration network",
"year": ""
},
{
"authors": " Curran Associates; Inc",
"journal": "",
"ref_id": "b27",
"title": "",
"year": "2018"
},
{
"authors": "Yanbin Liu; Juho Lee; Minseop Park; Saehoon Kim; Eunho Yang; Sung Ju Hwang; Yi Yang",
"journal": "",
"ref_id": "b28",
"title": "Learning to propagate labels: Transductive propagation network for few-shot learning",
"year": "2018"
},
{
"authors": "Vincent Lonij; Ambrish Rawat; Maria-Irina Nicolae",
"journal": "",
"ref_id": "b29",
"title": "Open-world visual recognition using knowledge graphs",
"year": "2017"
},
{
"authors": "Ilya Loshchilov; Frank Hutter",
"journal": "",
"ref_id": "b30",
"title": "Sgdr: Stochastic gradient descent with warm restarts",
"year": "2016"
},
{
"authors": "Leland Mcinnes; John Healy; Steve Astels",
"journal": "J. Open Source Softw",
"ref_id": "b31",
"title": "hdbscan: Hierarchical density based clustering",
"year": "2017"
},
{
"authors": "Timo Milbich; Karsten Roth; Samarth Sinha; Ludwig Schmidt; Marzyeh Ghassemi; Bjorn Ommer",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b32",
"title": "Characterizing generalization under out-of-distribution shifts in deep metric learning",
"year": "2021"
},
{
"authors": "Jishnu Mukhoti; Viveka Kulharia; Amartya Sanyal; Stuart Golodetz; Philip Torr; Puneet Dokania",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b33",
"title": "Calibrating deep neural networks using focal loss",
"year": "2020"
},
{
"authors": "Gregory F Mahdi Pakdaman Naeini; Milos Cooper; Hauskrecht",
"journal": "",
"ref_id": "b34",
"title": "Obtaining well calibrated probabilities using bayesian binning",
"year": "2015"
},
{
"authors": "Yaniv Ovadia; Emily Fertig; Jie Ren; Zachary Nado; David Sculley; Sebastian Nowozin; Joshua Dillon; Balaji Lakshminarayanan; Jasper Snoek",
"journal": "Advances in neural information processing systems",
"ref_id": "b35",
"title": "Can you trust your model's uncertainty? evaluating predictive uncertainty under dataset shift",
"year": "2019"
},
{
"authors": "Yash Patel; Giorgos Tolias; Jiri Matas",
"journal": "CVPR",
"ref_id": "b36",
"title": "Recall@k surrogate loss with large batches and similarity mixup",
"year": "2022"
},
{
"authors": "John Platt",
"journal": "Advances in Large-Margin Classifiers",
"ref_id": "b37",
"title": "Probabilistic outputs for svms and comparisons to regularized likelihood methods",
"year": "1999"
},
{
"authors": "Limeng Qiao; Yemin Shi; Jia Li; Yaowei Wang; Tiejun Huang; Yonghong Tian",
"journal": "",
"ref_id": "b38",
"title": "Transductive episodic-wise adaptive metric for few-shot learning",
"year": "2019"
},
{
"authors": "Oren Rippel; Manohar Paluri; Piotr Dollar; Lubomir Bourdev",
"journal": "",
"ref_id": "b39",
"title": "Metric learning with adaptive density discrimination",
"year": "2015"
},
{
"authors": "Yaniv Romano; Matteo Sesia; Emmanuel Candes",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b40",
"title": "Classification with valid and adaptive coverage",
"year": "2020"
},
{
"authors": "S Sra",
"journal": "",
"ref_id": "b41",
"title": "A short note on parameter approximation for von mises-fisher distributions: and a fast implementation of is",
"year": "2012"
},
{
"authors": "Masashi Sugiyama; Motoaki Kawanabe",
"journal": "MIT press",
"ref_id": "b42",
"title": "Machine learning in non-stationary environments: Introduction to covariate shift adaptation",
"year": "2012"
},
{
"authors": "Rina Foygel Ryan J Tibshirani; Emmanuel Barber; Aaditya Candes; Ramdas",
"journal": "Advances in neural information processing systems",
"ref_id": "b43",
"title": "Conformal prediction under covariate shift",
"year": "2019"
},
{
"authors": "Grant Van Horn; Oisin Mac Aodha; Yang Song; Yin Cui; Chen Sun; Alex Shepard; Hartwig Adam; Pietro Perona; Serge Belongie",
"journal": "",
"ref_id": "b44",
"title": "The inaturalist species classification and detection dataset",
"year": "2018"
},
{
"authors": "Vladimir Vapnik",
"journal": "Springer science & business media",
"ref_id": "b45",
"title": "The nature of statistical learning theory",
"year": "1999"
},
{
"authors": "Petar Veličković; Guillem Cucurull; Arantxa Casanova; Adriana Romero; Pietro Lio; Yoshua Bengio",
"journal": "",
"ref_id": "b46",
"title": "Graph attention networks",
"year": "2017"
},
{
"authors": "Catherine Wah; Steve Branson; Peter Welinder; Pietro Perona; Serge J Belongie",
"journal": "",
"ref_id": "b47",
"title": "The caltech-ucsd birds-200-2011 dataset",
"year": "2011"
},
{
"authors": "Mei Wang; Weihong Deng",
"journal": "Neurocomputing",
"ref_id": "b48",
"title": "Deep visual domain adaptation: A survey",
"year": "2018"
},
{
"authors": "Yifan Xing; Tong He; Tianjun Xiao; Yongxin Wang; Yuanjun Xiong; Wei Xia; David Wipf; Zheng Zhang; Stefano Soatto",
"journal": "",
"ref_id": "b49",
"title": "Learning hierarchical graph neural networks for image clustering",
"year": "2021"
},
{
"authors": "Xiaopeng Yan; Riquan Chen; Litong Feng; Jingkang Yang; Huabin Zheng; Wayne Zhang",
"journal": "",
"ref_id": "b50",
"title": "Progressive representative labeling for deep semi-supervised learning",
"year": "2021"
},
{
"authors": "Chenxiao Yang; Qitian Wu; Jiahua Wang; Junchi Yan",
"journal": "",
"ref_id": "b51",
"title": "Graph neural networks are inherently good generalizers: Insights by bridging gnns and mlps",
"year": "2022"
},
{
"authors": "Lei Yang; Dapeng Chen; Xiaohang Zhan; Rui Zhao; Chen Change Loy; Dahua Lin",
"journal": "",
"ref_id": "b52",
"title": "Learning to cluster faces via confidence and connectivity estimation",
"year": "2020"
},
{
"authors": "Bianca Zadrozny; Charles Elkan",
"journal": "",
"ref_id": "b53",
"title": "Obtaining calibrated probability estimates from decision trees and naive bayesian classifiers",
"year": "2001"
},
{
"authors": "Bianca Zadrozny; Charles Elkan",
"journal": "",
"ref_id": "b54",
"title": "Transforming classifier scores into accurate multiclass probability estimates",
"year": "2002"
},
{
"authors": "Qin Zhang; Linghan Xu; Qingming Tang; Jun Fang; Ying Nian Wu; Joe Tighe; Yifan Xing",
"journal": "",
"ref_id": "b55",
"title": "Thresholdconsistent margin loss for open-world deep metric learning",
"year": "2024"
},
{
"authors": "Zhisheng Zhong; Jiequan Cui; Shu Liu; Jiaya Jia",
"journal": "",
"ref_id": "b56",
"title": "Improving calibration for long-tailed recognition",
"year": "2021"
}
] | [
{
"formula_coordinates": [
2,
308.86,
524.1,
236.25,
21.61
],
"formula_id": "formula_0",
"formula_text": "∩ C cal = C train ∩ C test = C cal ∩ C test = ∅."
},
{
"formula_coordinates": [
2,
321.05,
640.71,
224.06,
14.66
],
"formula_id": "formula_1",
"formula_text": "maximize d TNR test , subject to TPR test (d) ≥ α (1)"
},
{
"formula_coordinates": [
3,
78.67,
132.71,
207.69,
17.09
],
"formula_id": "formula_2",
"formula_text": "TPR test (d) = i,j∈Dtest 1 yi=yj • 1 dij <d i,j∈Dtest 1 yi=yj(2)"
},
{
"formula_coordinates": [
3,
154.99,
163.7,
131.38,
17.09
],
"formula_id": "formula_3",
"formula_text": "1 yi̸ =yj • 1 dij >d i,j∈Dtest 1 yi̸ =yj (3)"
},
{
"formula_coordinates": [
4,
83.05,
628.89,
203.32,
46.3
],
"formula_id": "formula_4",
"formula_text": "T PR test (d) = i,j∈Dtest 1 pij >τ • 1 dij <d i,j∈Dtest 1 pij >τ (4) TNR test (d) = i,j∈Dtest 1 pij ≤τ • 1 dij >d i,j∈Dtest 1 pij ≤τ (5)"
},
{
"formula_coordinates": [
5,
57.58,
118.05,
228.78,
27.34
],
"formula_id": "formula_5",
"formula_text": "s avg i = j∈Ni a ij • 1 yi=yj |N i | , s nbr i = j∈Ni a ij • (1 yi=yj -1 yi̸ =yj ) |N i |(6)"
},
{
"formula_coordinates": [
5,
56.01,
280.33,
230.35,
32.23
],
"formula_id": "formula_6",
"formula_text": "TPR k = ∥ i∈D f i • 1 yi=k ∥ i∈D 1 yi=k , TNR k = i,j∈D (1 -a ij ) • 1 yj ̸ =yi=k i,j∈D 1 yj ̸ =yi=k(7)"
},
{
"formula_coordinates": [
5,
63.59,
421.92,
222.77,
46.74
],
"formula_id": "formula_7",
"formula_text": "lim |Ni|→∞ TPR k = |N i | 2|N | • ( 1 |N | i∈N s nbr i avg s nbr + 1 |N | i∈N a i avg avg a avg 2×avg s avg ) 1/2(8)"
},
{
"formula_coordinates": [
5,
77.02,
477.8,
42.8,
13.68
],
"formula_id": "formula_8",
"formula_text": "a avg i = 1"
},
{
"formula_coordinates": [
5,
63.73,
556.75,
222.63,
34.09
],
"formula_id": "formula_9",
"formula_text": "lim |Ni|→∞ TNR k = 1 - |N | |N | k - • ( 1 |N | i∈N s avg i - 1 |N | i∈N s nbr i average (s avg -s nbr ) )(9)"
},
{
"formula_coordinates": [
5,
348.01,
252.89,
157.96,
23.4
],
"formula_id": "formula_10",
"formula_text": "L overall = L conn main task + λ • (L s nbr + L s avg )"
},
{
"formula_coordinates": [
5,
318.25,
356.33,
226.86,
45
],
"formula_id": "formula_11",
"formula_text": "L conn = i,j∈V 1 yi=yj • log(p ij ) i,j∈V 1 yi=yj + i,j∈V 1 yi̸ =yj • log(1 -p ij ) i,j∈V 1 yi̸ =yj(11)"
},
{
"formula_coordinates": [
5,
316.14,
423.66,
228.98,
29.38
],
"formula_id": "formula_12",
"formula_text": "L s avg = i∈V (s avg i -ŝavg i ) 2 |V | , L s nbr = i∈V (s nbr i -ŝnbr i ) 2 |V |(12)"
},
{
"formula_coordinates": [
5,
335.7,
464.9,
5.81,
8.74
],
"formula_id": "formula_13",
"formula_text": "V"
},
{
"formula_coordinates": [
6,
50.11,
381.41,
205.59,
9.65
],
"formula_id": "formula_14",
"formula_text": "C train ∩ C cal = C train ∩ C test = C cal ∩ C test = ∅."
},
{
"formula_coordinates": [
6,
330.79,
485.5,
214.32,
40.36
],
"formula_id": "formula_15",
"formula_text": "MAE comb = 1 2 2 0 | T PR(d) -TPR(d)|+ | TNR(d) -TNR(d)| dd(13)"
}
] |
|
10.18653/v1/2023.iwslt-1.1 | 2023-11-14 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b15",
"b6",
"b1",
"b20",
"b0",
"b11",
"b15",
"b19",
"b24",
"b19",
"b18"
],
"table_ref": [],
"text": "Knowledge Distillation (KD) (Hinton et al., 2015;Buciluǎ et al., 2006) plays a pivotal role in many machine learning tasks including neural machine translation (NMT). This is evident in recent translation evaluations (Akhbardeh et al., 2021;Kocmi et al., 2022;Agarwal et al., 2023), where the majority of submissions incorporate KD into their training pipelines. Most relevant to this paper, KD's primary strength lies in its ability to utilize a larger model to train a smaller model more effectively. The model in training is referred to as the student model while the larger one as the teacher model. Consequently, the accuracy of the teacher model correlates with that of the student model.\nOne well-known yet simple approach for enhancing the accuracy of machine learning models involves model ensembling (Dietterich, 2000), as also has been applied by (Hinton et al., 2015). The underlying models are typically trained on the same datasets but with varying random initializations. This technique has been also exploited by existing works of KD for NMT, where the dominant variant is the sequence-level KD (Kim and Rush, 2016) that deploys an ensemble of such models to generate pseudo-labels for training student models. To maintain the simplicity of the inference process, the underlying models are often required to share the same vocabulary and network architecture. These factors, unfortunately, restrict the types of models, thereby limiting the avenue for improving the teacher accuracy and consequently, student accuracy.\nWe introduce an n-best reranking approach to extend the existing sequence-level KD, which involves a two-step process. In the first step, a specific set of models is tasked with generating a highquality n-best list. Our initial study indicates the potential for a gain of almost 10 BLEU points in accuracy on our validation set if we consider hypotheses beyond top-1. The final decision on which hypothesis to be select is deferred to the second step where a broader range of more expressive or complex models can be incorporated. In this step, we can take advantage of models with various architectures, inductive biases, sources of training data and objective functions. To increase model diversity further, we also incorporate open-source large pretrained models, taking advantage of their frequent availability within the community.\nWe showcase our reranker's effectiveness in two scenarios, namely the traditional and iterative KDs. In the former, pseudo-labels are directly utilized for training student models, while the latter, often dubbed as self-training, involves an extra step of iteratively retraining the teacher models using the pseudo-labels (Li et al., 2019). We also delve into efforts aimed at scaling up our approach for distilling large-scale training data, encompassing strategies such as model selection and transfer set reduction. More concretely, we conduct extensive experiments on WMT'21 German ↔ English and Chinese ↔ English translation tasks. Our final model is as accurate as a large multilingual model with 4.7 billion parameters, despite having only 68 million parameters.\n2 Background: Sequence-Level KD KD trains a student model (p θ ) with the supervision of a teacher model (q θ ) by minimizing the discrepancy between the prediction of the student model with that of the teacher model. Sequencelevel KD, proposed by Kim and Rush (2016), extends KD by minimizing the discrepancy at the level of sequence (rather than at token level). Since enumerating all possible sequences is intractable, Kim et al. (2021) approximate the distribution with its mode t obtained via the following inference: t = arg max q θ (t|s). This simple approximation allows NMT to reuse the same training pipeline for student model with only a slight modificationprimarily done by substituting the original labels t with pseudo-labels t during the computation of the loss function."
},
{
"figure_ref": [],
"heading": "N-best Reranking for Distillation",
"publication_ref": [],
"table_ref": [],
"text": "Our n-best reranker formulates q θ (t|s) as a loglinear model, which is parameterized with a collection of scoring models M(s, t) and their associated weights λ. Each model, M i (s, t) ∈ R, assigns a real-valued score that indicates the plausibility of the hypothesis t being the translation of s. These models are pre-trained and considered static. Meanwhile, their weights are parameters that are learned to produce the optimal translation metrics of a tuning set. We discuss the optimization of λ in Section 3.1.\nTo generate pseudo-labels, our reranker applies the following arg max formula:\nt = arg max t∈N (s) λ • log M(s, t) ⊺ ,(1)\nwhere t ∈ N refers a hypothesis within an n-best list, generated, say, by running the beam search inference with beam size = n. We refer to the models used to generate N as G(s) ⊂ M(s, t), which is also used to score the n-best list. If G(s) = M(s, t), consisting of only one identical translation model, then Eq 1 would revert back to the vanilla sequence-level KD."
},
{
"figure_ref": [],
"heading": "Discriminative Training of λ",
"publication_ref": [
"b31",
"b8",
"b35",
"b36",
"b42",
"b7",
"b21"
],
"table_ref": [],
"text": "The simplest approach for assigning values to λ is to assign uniform values to λ, giving an equal weight to each model. However, our experiments reveal that it may not yield the most optimal results, given that some models carry more significance than others and that different model demands different scaling. Thus, we turn to discriminative training to find the optimal λ by utilizing a small set of held-out tuning set. The tuning set is assumed to be drawn from the same distribution of the test sets, thus optimal weights for the tuning set are more likely to lead to higher accuracy on the test sets. Note that in scenarios where the number of scoring models is limited, a basic randomized grid search, like applied by Ng et al. (2019), may suffice.\nFor this paper, we employ the Margin Infused Relaxed Algorithm (MIRA) (Chiang et al., 2008), known for its wide adaptation in Statistical Machine Translation and its ability to handle tens of thousands features. Without loss of generality, we use BLEU (Papineni et al., 2002) to measure translation accuracy in our experiments; however, it is worth noting that alternative metrics such as chrF (Popović, 2015) and TER (Snover et al., 2006) are also applicable.\nMIRA seeks to find λ that minimizes the following structured hinge loss\nL MIRA (λ) = max t∈N ∆(t) + λ • (M(s, t) ⊺ -M(s, t * ) ⊺ )\nwhere t * is the oracle hypothesis, which refers the hypothesis in the n-best list that attains the highest BLEU score, while ∆(t) signifies the BLEU differentials of a hypothesis t with the aforementioned oracle hypothesis. Ideally, the optimal λ is achieved when the loss reaches 0, indicating that a clear separation can be established between each non-oracle hypothesis and the oracle hypothesis with a margin proportional to their respective BLEU differentials.\nIn our experiments, we use a variant of MIRA with efficient batch-level support, called KB-MIRA (Cherry and Foster, 2012) which can be found in the Moses toolkit (Koehn et al., 2007). It also includes sparsity-inducing regularization which we utilize for model selection. For a more in-depth discussion on MIRA and its variants, we refer the readers to the cited papers."
},
{
"figure_ref": [],
"heading": "Model Description",
"publication_ref": [
"b44",
"b22",
"b17",
"b27",
"b25",
"b48",
"b16",
"b14",
"b39",
"b5",
"b34",
"b23",
"b2",
"b26",
"b12",
"b32",
"b43",
"b30",
"b29"
],
"table_ref": [
"tab_2"
],
"text": "The efficacy of our n-best reranker relies on the diversity and quality of the deployed models. The log-linear formulation in Equation ( 1 it imposes minimal assumptions about the underlying models. This flexibility relaxes the typical requirement for the models to strictly adhere to probabilistic principles or comprehensively describe the entire translation process. Consequently, our reranker can embrace a wide spectrum of models, including heuristics or target-side language models, as long as they assign a meaningful score. Our objective is to integrate as many models as possible, with the expectation that these diverse models contribute complementary information, guiding the reranker towards the most optimal hypothesis. In total, we conduct experiments involving more than 50 models for each language pair. For conciseness, we group the models into categories and provide a description for each category, summarized in Table 3. The first four categories encompass a diverse range of in-house translation models, characterized by distinctions in translation directions, generation orders, network architectures, and domain adaptability. In contrast, models in the last four categories do not strictly pertain to translation models but capture specific nuances of translation phenomena, such as the fluency of hypotheses or the level of agreement between hypotheses.\nThe first category is the forward translation model (TM), which corresponds to traditional autoregressive NMT models p(s|t). These models generate the translation sequentially one token at a time conditioned on the source sentence and previously generated tokens. This category includes models with various well-known architectures, such as Transformer Big (Vaswani et al., 2017), Deep Encoder Shallow Decoder (Kong et al., 2021), Nearest-Neighbor (Khandelwal et al., 2021) and MEGA (Ma et al., 2023).\nThe second category is the backward TM, which encompasses models that share the same architecture as the forward TM but focusing on modeling the flipped translation direction p(t|s). Our third category is the right-to-left TM, which includes models that generate tokens in a right-to-left fashion. According to (Liu et al., 2016;Zhou et al., 2019), the left-to-right models are more effective at generating accurate prefixes while the right-toleft models are more effective at generating accurate suffixes. Our fourth category is domainadapted models, which consists of translation models that we adapt to multiple domains. In our experiments, we simply equate the corpus provenance as the domain. We adopt a tag-based approach and prepend the source sequence with d ∈ {europarl, commoncrawl, paracrawl, • • • }, like in (Johnson et al., 2017;Ha et al., 2017).\nThe fifth category is the language model. Models in this category focus on evaluating the fluency of the hypotheses. In our experiments, we train a causal language model with the GPT-2 architecture (Radford et al., 2019) on the target side of our parallel data and the monolingual data. The sixth category is the alignment models, consisting models that evaluate the fine-grained correspondences between tokens in the hypothesis and the source sequence. To generate the alignment, we use the IBM model 3 (Brown et al., 1993) from the eflomal toolkit (Östling and Tiedemann, 2016). The seventh category is the Minimum Bayes-Risk (MBR) loss function. Via the models in this category, our reranker can give preferences to hypotheses that have the higher level of consensus with other hypotheses in the n-best list or vice versa, measured by some extrinsic translation metrics. These models infuse our reranker with elements of consensus decoding (Kumar and Byrne, 2004).\nOur last category consists of various publiclyavailable pretrained models. It includes, but not limited to, the LASER sentence-embedding model (Artetxe and Schwenk, 2018), the mBART multilingual translation model (Liu et al., 2020), the M2M-100 (Fan et al., 2020) and the NLLB (NLLB Team et al., 2022). It also includes a single dense multilingual model from the WMT21 winning team, namely Facebook AI Research (FAIR) WMT21 (Tran et al., 2021) -currently known as Meta AI Research. Additionally, it includes multilingual large language models from BigScience, namely BLOOMZ and mT0 (Muennighoff et al., 2022). These models are trained with significantly more data and not all of them are explicitly trained to optimize translation objectives. When utilizing these models, we condition them for translation by feeding five translation examples as the prefix of the prompt like in (Moslem et al., 2023). In terms of size, the models in this category vary from 50 million to 10+ billion parameters, which is larger than the models in other categories."
},
{
"figure_ref": [],
"heading": "Models for Generating n-Best List",
"publication_ref": [
"b44"
],
"table_ref": [],
"text": "The effectiveness of our n-best reranking also hinges upon the accuracy and diversity of the n-best list. While the ideal scenario involves deploying all models within M(s, t), this proves to be both computationally intensive and impractical, especially considering that not all models explicitly generate translations, such as language models. Alternatively, a sampling-based inference can be employed as a substitute for beam search to generate more diverse n-best lists. Unfortunately, this approach often results in a trade-off between the two.\nTo strike a more optimal balance, we have chosen to utilize two specific models: the L2R and the R2L models. The L2R model comprises an ensemble of four Transformer Big models (Vaswani et al., 2017), belonging into the forward TM category, while, the R2L model is its right-to-left counterpart. By combining n-best lists from the L2R model, specialized at producing accurate prefixes with diverse suffixes, and from the R2L model, specialized at generating accurate suffixes with diverse prefixes, we aim to generate highly accurate but diverse nbest lists. Appendix B details our exploration."
},
{
"figure_ref": [],
"heading": "Scaling Up with Model Selection",
"publication_ref": [],
"table_ref": [],
"text": "While deploying the complete set of models to showcase accuracy improvements on a small set of test data is relatively affordable, scaling up the process to distill the entire training dataset becomes computationally intractable. Therefore, an effective model selection is crucial to identify a smaller subset of models, denoted as D(s, t) ⊂ M(s, t). The quality of model selection is still crucial, as we aim to minimize sacrificing the overall model quality. In our case, we think the goal is viable because, despite the intended complementarity of the models, there may be significant overlap, particularly as the majority of our in-house models are trained on the same data.\nManual selection of D(s, t) is impractical given the vast number of choices. Instead, we adopt a simple solution by leveraging the discriminatively learned weights λ associated with the models. This approach capitalizes on the regularization term employed by the optimizer (Section 3.1), offering a convenient and inexpensive way of selecting models that contribute significantly to the task. In our experiments in Section 4.1, we select top 5 models with the heighest weights for distillation, reducing the model count in our reranker with minimal accuracy drop."
},
{
"figure_ref": [],
"heading": "Experimental Results",
"publication_ref": [
"b19",
"b37",
"b36",
"b40"
],
"table_ref": [],
"text": "To showcase the effectiveness of our n-best reranking proposal, we conduct experiments on WMT21 German ↔ English and Chinese ↔ English largescale translation tasks. Our baseline is the vanilla sequence-level KD (Kim and Rush, 2016) that employs an ensemble of four Transformer Big models as its teachers. We constrained the student model's capacity to approximately 68 million parameters, in line with the Transformer Base architecture. Akin to sequence-level KD's teacher models, the majority of models in our reranker constitute an ensemble of four models based on some variants of the Transformer Big architecture. More details about these models can be found in Appendix A, including other experimental setup including the bitext used mainly for teacher model training and the monolingual data primarily used for student model training. We use the WMT19 set to learn λ weights for our reranker, the WMT20 set as our validation set and the WMT21 set as our blind test set. For these sets, we use the maximum number of references provided. To report accuracy, we use sacre-BLEU (Post, 2018) with the following signature nrefs:k|case:mixed|eff:no|tok:13a|smooth :exp where k is the number of reference(s). For our main results, we additionally re-port chrF (Popović, 2015) with this signature nrefs:k|case:mixed|eff:yes|nc:6|nw:0|spa ce:no and COMET22 (Rei et al., 2022) using wmt22-comet-da model. For generating n-best list, we use beam=8 and for student model inference, we use beam=5.\nIn Section 4.1, we focus on intrinsic evaluation, comparing the accuracy of the n-best reranker with that of the sequence-level KD's teacher models on validation sets. In Section 4.2 and Section 4.3, we shift to extrinsic evaluations where we assess the utility of n-best reranker's pseudo-labels for training student model and retraining teacher models. We mainly focus on the German → English direction and summarize the results for the other language pairs at the end."
},
{
"figure_ref": [],
"heading": "Accuracy of n-best Reranker",
"publication_ref": [],
"table_ref": [
"tab_1"
],
"text": "Table 2 summarizes the accuracy of our n-best reranker on the German → English's validation set. In the WMT20 validation set, the baseline system attains a BLEU score of 58.8. Notably, this score also represents the score of the top-1 hypothesis in out n-best list since it is generated by the same model (complemented with its right-to-left counterpart). In rows Oracle and Anti-Oracle, we report the accuracy of the best-scoring and worstscoring hypotheses within our n-best list. Row Oracle shows that the best-scoring hypotheses surpass the top-1 by almost 10 BLEU point, indicating the substantial room for improvement embedded in our n-best reranking approach. Conversely, row Anti-Oracle shows that the gap to the worst-scoring hypotheses is much wider, which is almost 20 BLEU point worse. This underscores the importance of employing robust scoring models, given the risk associated with poor-scoring alternatives."
},
{
"figure_ref": [],
"heading": "Description WMT20",
"publication_ref": [
"b13",
"b43"
],
"table_ref": [
"tab_2"
],
"text": "Baseline Using the full set of 72 models (M), our n-best reranker achieves the BLEU score of 60.4, surpassing the baseline system by 1.6 BLEU point. This outcome underscores the efficacy of our nbest reranker proposal in enhancing model accu-racy. We then proceed to apply the model selection strategy described in Section 3.4 by leveraging λ. We pick 5 models with the highest weights, rerun reranking with the same weights (zeroing out the weights of other models) and report the reranker accuracy in the last row. As shown, the accuracy of the n-best reranker with smaller model count is similar to running with the full set of models.\nTable 3 compiles all the models utilized for generating n-best list (G) and for rescoring the list (G). We rank the models based on the accuracy of each model when it is used as the only model to rerank the n-best list. As shown, the two models utilized for generating the n-best list (G) are ranked 5 and 13 respectively, but are not selected by our model selection strategy. Interestingly, the models selected for distillation (D) exhibit considerable variability in terms of ranking, notably excluding the top highest-ranked models. We hypothesize that this is due to redundancy in high-performing models, and the reranker prioritizes model diversity as also suggested by (Gontijo-Lopes et al., 2022). The first model in D is the single multilingual dense model provided by (Tran et al., 2021), which is the most accurate model. While this model is not their final submission to WMT, it is highly accurate since it is trained on significantly larger training data and consists of 4.7 billion parameter. The remaining four other models in (D) come from different model categories, ranging from backward, adapted, R2L and publicly-available models. Note that since the model selection strategy is automatic and non-deterministic, the model chosen for each iteration is non-deterministic. This is also applicable in other translation pairs."
},
{
"figure_ref": [],
"heading": "N-best Reranking Improves Student",
"publication_ref": [
"b19",
"b15",
"b19"
],
"table_ref": [
"tab_3"
],
"text": "This section investigates the utility of our n-best reranking approach on the downstream task of training student model. We use the reranker with D models to generate the pseudo-labels for the whole training data. For our baseline sequencelevel KD, we use the L2R model to generate the pseudo-labels. As another baseline, we also include sequence-level Knowledge Interpolation (KI) from (Kim and Rush, 2016), which chooses hypotheses in the n-best list that give the highest BLEU score using the original labels as the references.\nWe also explore how the accuracy of the student model is impacted by different transfer sets, which refers to the examples that were distilled and used to train the student model (Hinton et al., 2015) investigate three transfer set configurations, namely bitext only, bitext + monolingual, and monolingual only. Table 4 summarizes the results of our experiments, which contains the accuracy of various student models on WMT21 test set.\nIn the bitext only condition, we only consider the distilled parallel data to train the student model. More specifically, we compare the pseudo-labels generated by n-best reranking with three baseline methods: original labels, pseudo-labels obtained through sequence-level knowledge interpolation (KI), and those obtained through sequence-level knowledge distillation (KD). As in row 1, the student model trained with the original labels achieved an accuracy of 48.8 BLEU point. Meanwhile, the models trained with pseudo-labels generated through sequence-level KI and KD showed improvements of 0.5 and 0.8 BLEU points respectively, which is in line with previous literature (Kim and Rush, 2016). Our n-best reranker approach leads to even stronger performance, with the student model achieving an accuracy of 50.0 BLEU point. This is a noteworthy improvement of 1.2 BLEU points compared to the baseline model.\nIn the bitext+mono condition, we augment the training data for the student model with the distilled monolingual data. Since the monolingual data lack labels, we compare our n-best reranking method only with sequence-level knowledge distillation (KD). The results in row 2 reveal that incorporating the distilled monolingual data significantly improves the accuracy of the sequence-level KD system by approximately 1.3 BLEU points. However, our n-best reranking approach achieves an even greater gain of 2.0 BLEU points, thereby widening the performance gap with sequence-level KD to 1.1 BLEU points. This result highlights the value of incorporating in-domain data as the transfer sets. In our approach, the monolingual data used seems to align with the domain of the evaluation sets, while in contrast, the parallel data are sourced from a broader range of domains.\nIn the mono only condition, we investigate further whether a smaller in-domain transfer set is more or as effective than a larger mixed-domain one. The results in row 3 reveal marginal gains for both sequence-level KD and our n-best reranking approach when using only the distilled monolingual data as the transfer sets. This finding highlights the possibility of using smaller transfer sets and the importance of their domain, which we will explore further in future work. Nonetheless, our n-best reranking approach leads to a student model that is 3.4 BLEU better than the baseline and 1.3 BLEU better than sequence-level KD."
},
{
"figure_ref": [],
"heading": "Self-Training Teacher Improves Student",
"publication_ref": [],
"table_ref": [
"tab_4",
"tab_3",
"tab_4"
],
"text": "Given the substantial accuracy gains obtained by using pseudo-labels in student models, we investigate whether teacher models can derive similar benefit from the use of pseudo-labels. Up to now, all the teachers models are trained exclusively from parallel data with original labels that come from a mixed set of domains. In light of this, we conduct a series of experiments retraining the teacher model using pseudo-labels. To manage computational costs effectively, we focus our investigations on the models within G. This is accomplished through fine-tuning the models, as opposed to retraining them from scratch, and utilizing only monolingual data, excluding the bitext, as the transfer sets. Our rationale for this strategy is detailed in the preliminary experiments, discussed in the Appendix C.\nMore specifically, we fine-tune the two models in G using the pseudo-labels obtained from the n-best reranker using monolingual data as the transfer sets for one epoch. We then retrain the next iteration's reranker using these models, producing a new set of pseudo-labels for training the student model. It's worth reiterating that the models selected for distillation D vary in each iteration. We continue this iterative process twice when we typically start observing diminishing gain.\nTable 5 provides a summary of our self-training experiments. Focusing on the German → English columns, the first three rows of the table are taken from Table 4, reporting the accuracies of the baseline model, the student model trained with sequence-level KD, and the student model trained with pseudo-labels from n-best reranking. The next two rows show the results from our self-training experiments for two iterations. Our experiments show that self-training the teacher models for one iteration can improve the student model accuracy by 0.5 BLEU points (row 4). Our final model, after three iterations, scores 4.0 BLEU points higher than the baseline model and 2.9 BLEU points higher than sequence-level KD. This conclusion is consistent across both chrF and COMET metrics. We also compare our final model with the winning WMT21 models from FAIR with respect to accuracy and model size, as shown in rows 6 and 7. Performancewise, our final model is comparable to FAIR's Dense model, while having fewer parameters. Our model consists of 68 million parameters, while the FAIR model is around 70 times larger.\nWe also present the experimental results for the English → German and the Chinese ↔ English directions in Table 5. As shown, we observe a gain similar to the one observed in the German → English direction where the pseudo-labels from n-best reranker leads to a significantly better student ac-curacy. These gains remain consistent across multiple metrics, encompassing chrF and COMET22, although given that our reranker is trained to optimize BLEU score, the most pronounced improvement is evident in the BLEU score. Nevertheless, these results affirm our hypothesis that the n-best reranker with robust scoring models can effectively enhance the quality of training data labels. Consequently, this improvement translates into increased model accuracy, all achieved without the need for increased model size or additional data labeling."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b41",
"b8",
"b28",
"b38",
"b43",
"b46",
"b18",
"b46",
"b24",
"b24"
],
"table_ref": [],
"text": "Our proposal intersects with many works in various ways. The idea of utilizing n-best reranking to improve accuracy has been extensively investigated as far back as the era of Statistical Machine Translation if not earlier, for example in (Och et al., 2004;Shen et al., 2004;Chiang et al., 2008). More recently, n-best reranking has also been deployed as a crucial component in the winning systems of many translation evaluation campaigns, for example, in (Marie et al., 2020;Qian et al., 2021;Tran et al., 2021). In these recent work, n-best reranking incurs significantly higher inference time from running multiple models over the n-best list, thus may not be practical for real-world systems. In contrast, our work makes a practical trade-off by shifting the heavy computational cost of n-best reranking to training data preprocessing without affecting the latency of the deployed model. Our work shares the same motivation as (Yang et al., 2022), but we consider a larger and more diverse set of models.\nThe idea of looking at n-best hypotheses for knowledge distillation has been also investigated in the original sequence-level KD paper (Kim et al., 2021), namely sequence-level Knowledge Interpolation which we consider as one of our baseline where the authors propose to approximate the mode with the hypothesis that scores the highest according some translation metrics. However, since this approach requires the ground truth, the application of this variant is limited to distilling parallel data. In contrast, since our n-best reranker is trained on a tune set, our approach is applicable for distilling unlabelled monolingual data.\nOur n-best reranker incorporates various models as reranking models. Some of these models have been applied to knowledge distillation. For example, Yang et al. (2022) Self-training has also been frequently investigated for Machine Translation in statistical and neural era (Li et al., 2019). Recently, it is often dubbed as iterative knowledge distillation and can be found as a winning formula in many evaluation campaigns (Li et al., 2019). In this work, we apply self-training using high-quality pseudo-labels from n-best reranker which produces accurate results."
},
{
"figure_ref": [],
"heading": "Summary and Future Work",
"publication_ref": [
"b10"
],
"table_ref": [],
"text": "We enhance sequence-level knowledge distillation by incorporating n-best reranking of a diverse set of robust and complex models. Rather than assigning the top-1 best hypotheses as pseudo-labels to train the student model, our proposed method leverages a multitude of models to collaboratively examine n-best hypotheses and identify the best candidate.\nBy doing so, our approach enables a more comprehensive exploration of potential solutions and promotes more accurate predictions, resulting in improved performance of the student model. Furthermore, we observed a relatively strong cascading effect, where teacher models finetuned using pseudo-labeled data are more accurate, leading to the generation of more accurate pseudo-labels for the next iteration and resulting in an even more accurate student model. We also put forward successful efforts to scale up n-best reranking via finetuning and using selections of models and transfer sets. Our final student model demonstrates a 3.0 to 4.0 BLEU point improvement over baseline systems and is on par with a strong large translation model on German-English and English-German translation directions, despite having only 1/70 th the parameters.\nFor future work, we intend to continue the scaling up efforts to selective incorporate more powerful models and investigate methods to automatically identify transfer sets at fine-grained sentence level. We also plan to investigate ways to speed up the scoring process, for instance by utilizing only unnormalized probabilty score like in (Devlin et al., 2014)."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [],
"table_ref": [],
"text": "While the proposed evaluation framework is language-agnostic, the experiments conducted in this study are limited to two language pairs. Due to its reliance on the availability of models and indomain monolingual, we cannot guarantee accurate results when applied to language pairs involving a low-resource language pairs."
},
{
"figure_ref": [],
"heading": "Ethics Statement",
"publication_ref": [],
"table_ref": [],
"text": "We acknowledge the ethical considerations associated with the n-best reranking approach, which utilizes multiple models to generate pseudo-labels. First of all, we recognize that these models possess their own biases, inherited from the training data, which can potentially perpetuate societal inequalities. Bias in the models can result from biased training data or the inherent limitations of the algorithms used. Despite our best efforts to preprocess and debias the training data, complete elimination of biases is challenging. Second of all, the n-best reranking approach incurs higher computational costs compared to traditional methods. These costs arise from training and maintaining multiple models concurrently. We have implemented mitigation strategies such as model recycling and leveraging publicly available corpora to address these concerns. Despite of our mitigation efforts, the increased computational burden can limit the accessibility and affordability of the approach, particularly for researchers or organizations with limited resources."
},
{
"figure_ref": [],
"heading": "A Experimental Setups",
"publication_ref": [],
"table_ref": [],
"text": "We follow the experimental setup of the WMT21 news translation task, particularly the constrained track to train our in-house models. For German ↔ English directions, our parallel data are composed of Europarl v10, ParaCrawl v7.1, Common Crawl, News Commentary v16, Wiki Titles v3, Tilde Rapid and WikiMatrix. For Chinese ↔ English directions, our parallel data are composed of Paracrawl v7.1, News Commentary v16, Wiki Titles v3, UN Parallel Corpus v1.0, CCMT and Wiki-Matrix. For monolingual data, we use the 2020 and the 2021 subsets of News Crawl. We deduplicate and preprocess the data using the M2M-100 (Fan et al., 2021) processing scripts 1 . For training our in-house teacher models, we run up to 80 thousand updates, while for training the student model, we run up to 30 thousand updates. For finetuning teacher models, we run one epoch of updates."
},
{
"figure_ref": [
"fig_0"
],
"heading": "B Pilot Study for Models for n-best Generation",
"publication_ref": [],
"table_ref": [
"tab_5"
],
"text": "Figure 1 shows the BLEU scores for the top-1 and oracle hypotheses of n-best list with different N from 1 to 32 on our tune set. As shown, the BLEU score for the top-1 hypotheses marginally improves when we increase the beam size from 1 to 4 but then it plateaus, which is consistent with Britz et al.\n(2017)'s finding. This suggests that increasing the beam size may not benefit the original sequencelevel KD. In contrast, the oracle BLEU score improves monotonically with larger beam size, where the gap for N > 8 is more than 10 BLEU points and growing. This gap speaks to the potential for our proposed n-best reranking. Compared to L2R, the n-best list's oracle score from L2R+R2L is around 2-3 BLEU points higher. We equate G to L2R+R2L with beam size 8 for our experiments since its accuracy is better than doubling the beam size of L2R setup with additional parallelization benefits.\nC N-best reranking for Self-Training\nWe conduct a pilot study on one of the L2R models, which is part of G, to inform our decisions on two aspects: 1) determining which transfer sets to utilize, and 2) determining whether it is necessary to retrain the teacher model from the scratch or if fine- 6. For finetuning, we only run one epoch, while for retraining we run around 50 epochs (up to 80 thousands update). The baseline accuracy of training this teacher model using the original bitext is 57.4 point, as indicated in the first row of column Baseline. The Retrain column shows that training the teacher model with the same bitext, but with pseudo-labels, resulted in a gain of 0.6 BLEU point. As shown in the subsequent rows (bitext+mono and mono only), adding the distilled monolingual data to the transfer sets or using them alone result in a stronger gain of around 1.5 BLEU points, which is consistent with our finding in the student model training. Comparing the Retrain and Finetune columns, we observe that the accuracy of finetuned models is on par with the model trained from scratch. These results are encouraging because we can obtain a teacher model that is 2.1 BLEU points more accurate with minimal training FLOPs via finetuning and using the smallest transfer set. We conduct similar experiments using pseudo-labels from sequence-level KD and discuss it in Appendix D.\nAlthough a similar trend is observed, the resulting gain from sequence-level KD is smaller."
},
{
"figure_ref": [],
"heading": "D Sequence-level KD for Self-Training",
"publication_ref": [],
"table_ref": [
"tab_6",
"tab_5"
],
"text": "We report the results for self-training teacher model using the pseudo-labels from sequence-level KD in Table 7. As shown in row bitext only, retraining teacher models with these pseudo-labels leads to a degradation. Including the monolingual data as the transfer sets helps to improve the accuracy as shown in row bitext+mono and mono only. Comparing columns Retrain and Finetune, we observe that finetuning can achieve a similar accuracy gain as the full retraining, which is similar to what we observe in finetuning experiments using pseudolabels from n-best reranking. Comparing with using pseudo-labels from n-best reranking reported in Table 6, self-training using pseudo-labels from sequence-level KD gives smaller accuracy gain than self-training using pseudo-labels from n-best reranking. "
},
{
"figure_ref": [],
"heading": "Baseline",
"publication_ref": [],
"table_ref": [],
"text": ""
}
] | We propose utilizing n-best reranking to enhance the Sequence-Level Knowledge Distillation (Kim and Rush, 2016) where we explore hypotheses beyond the top-1 to acquire more accurate pseudo-labels. To accomplish this, we leverage a diverse set of models with different inductive biases, objective functions or architectures, including publicly-available large pretrained models. The effectiveness of our proposal is validated through experiments on the WMT'21 German ↔ English and Chinese ↔ English translation tasks. Our results demonstrate that utilizing the pseudo-labels generated by our n-best reranker leads to a significantly more accurate student model. In fact, our best student model achieves comparable accuracy to a large translation model from (Tran et al., 2021) with 4.7 billion parameters, while having two orders of magnitude fewer parameters. | Accurate Knowledge Distillation with n-best Reranking | [
{
"figure_caption": "1Figure 1 :1Figure 1: BLEU scores for top-1 and oracle hypotheses of WMT19 with different beam size",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Categories of models to evaluate hypothesis pair s, t ∈ N in the n-best list. The first four categories correspond to in-house translation models, while the last four correspond to non-translation models.",
"figure_data": ") is flexible as",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "N-best reranker results on WMT20 validation.",
"figure_data": "/ Top-158.8Oracle67.5Anti-Oracle41.3n-best Reranker -Full (|M| = 72)60.4n-best Reranker -Select (|M d | = 5) 60.3",
"figure_id": "tab_1",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Description of the models used to generate n-best lists (G) and models selected for distillation (D), specifically for the first iteration of German → English direction, together with their accuracy on WMT20.",
"figure_data": ". We",
"figure_id": "tab_2",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Comparison of BLEU scores on WMT21 test sets between n-best reranking and the three baseline models, including sequence-level knowledge interpolation and distillation, across different configurations of transfer sets.",
"figure_data": "",
"figure_id": "tab_3",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "German ↔ English (top) and Chinese ↔ English (bottom) results on WMT21 test set, compared to the baseline models and WMT21's winning models from FAIR. FAIR MoE accuracy is taken from (BarryHaddow, 2021). Results from our n-best reranking are in gray.",
"figure_data": "deploys nearest neigh-",
"figure_id": "tab_4",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "WMT20 scores of a teacher model trained with pseudo-labels from n-best reranking with different transfer sets (rows) and training regime (columns).",
"figure_data": "Transfer Sets Baseline Retrain Finetunebitext only57.458.058.0bitext+mono-59.559.2mono only-59.559.5",
"figure_id": "tab_5",
"figure_label": "6",
"figure_type": "table"
},
{
"figure_caption": "WMT20 BLEU scores of a teacher model trained with pseudo-labels from sequence-level KD with different transfer sets (rows) and training regime (columns).",
"figure_data": "Retrain Finetunebitext only57.457.357.1bitext+mono-58.357.8mono only-58.358.1",
"figure_id": "tab_6",
"figure_label": "7",
"figure_type": "table"
}
] | Hendra Setiawan Apple | [
{
"authors": "Milind Agarwal; Sweta Agrawal; Antonios Anastasopoulos; Luisa Bentivogli; Ondřej Bojar; Claudia Borg; Marine Carpuat; Roldano Cattoni; Mauro Cettolo; Mingda Chen; William Chen; Khalid Choukri; Alexandra Chronopoulou; Anna Currey; Thierry Declerck; Qianqian Dong; Kevin Duh; Yannick Estève; Marcello Federico; Souhir Gahbiche; Barry Haddow; Benjamin Hsu; Mon Phu; Hirofumi Htut; Dávid Inaguma; John Javorský; Yasumasa Judge; Tom Kano; Rishu Ko; Pengwei Kumar; Xutai Li; Prashant Ma; Evgeny Mathur; Paul Matusov; John P Mcnamee; Kenton Mccrae; Maria Murray; Satoshi Nadejde; Matteo Nakamura; Ha Negri; Jan Nguyen; Xing Niehues; Atul Niu; Kr; John E Ojha; Proyag Ortega; Juan Pal; Lonneke Pino; Peter Van Der Plas; Elijah Polák; Elizabeth Rippeth; Jiatong Salesky; Matthias Shi; Sebastian Sperber; Katsuhito Stüker; Yun Sudoh; Brian Tang; Kevin Thompson; Marco Tran; Alex Turchi; Mingxuan Waibel; Shinji Wang; Rodolfo Watanabe; Zevallos",
"journal": "Association for Computational Linguistics",
"ref_id": "b0",
"title": "FINDINGS OF THE IWSLT 2023 EVALUATION CAMPAIGN",
"year": "2023"
},
{
"authors": "Farhad Akhbardeh; Arkady Arkhangorodsky; Magdalena Biesialska; Ondřej Bojar; Rajen Chatterjee; Vishrav Chaudhary; Marta R ; Markus Freitag; Yvette Graham; Roman Grundkiewicz; Barry Haddow; Leonie Harter; Kenneth Heafield; Christopher Homan; Matthias Huck; Kwabena Amponsah-Kaakyire; Jungo Kasai; Daniel Khashabi; Kevin Knight; Tom Kocmi; Philipp Koehn; Nicholas Lourie; Christof Monz; Makoto Morishita; Masaaki Nagata; Ajay Nagesh; Toshiaki Nakazawa; Matteo Negri; Santanu Pal; Auguste Allahsera; Marco Tapo; Valentin Turchi; Marcos Vydrin; Zampieri",
"journal": "Association for Computational Linguistics",
"ref_id": "b1",
"title": "Findings of the 2021 conference on machine translation (WMT21)",
"year": "2021"
},
{
"authors": "Mikel Artetxe; Holger Schwenk",
"journal": "",
"ref_id": "b2",
"title": "Massively multilingual sentence embeddings for zeroshot cross-lingual transfer and beyond",
"year": "2018"
},
{
"authors": "Barry Haddow",
"journal": "",
"ref_id": "b3",
"title": "WMT21 News Systems and Evaluations",
"year": "2021-05-05"
},
{
"authors": "Denny Britz; Anna Goldie; Minh-Thang Luong; Quoc Le",
"journal": "Association for Computational Linguistics",
"ref_id": "b4",
"title": "Massive exploration of neural machine translation architectures",
"year": "2017"
},
{
"authors": "F Peter; Stephen A Brown; Vincent J Della Pietra; Robert L Della Pietra; Mercer",
"journal": "Computational Linguistics",
"ref_id": "b5",
"title": "The mathematics of statistical machine translation: Parameter estimation",
"year": "1993"
},
{
"authors": "Cristian Buciluǎ; Rich Caruana; Alexandru Niculescu-Mizil",
"journal": "Association for Computing Machinery",
"ref_id": "b6",
"title": "Model compression",
"year": "2006"
},
{
"authors": "Colin Cherry; George Foster",
"journal": "Association for Computational Linguistics",
"ref_id": "b7",
"title": "Batch tuning strategies for statistical machine translation",
"year": "2012"
},
{
"authors": "David Chiang; Yuval Marton; Philip Resnik",
"journal": "Association for Computational Linguistics",
"ref_id": "b8",
"title": "Online large-margin training of syntactic and structural translation features",
"year": "2008"
},
{
"authors": "Anna Currey; Prashant Mathur; Georgiana Dinu",
"journal": "Association for Computational Linguistics",
"ref_id": "b9",
"title": "Distilling multiple domains for neural machine translation",
"year": "2020"
},
{
"authors": "Jacob Devlin; Rabih Zbib; Zhongqiang Huang; Thomas Lamar; Richard Schwartz; John Makhoul",
"journal": "Association for Computational Linguistics",
"ref_id": "b10",
"title": "Fast and robust neural network joint models for statistical machine translation",
"year": "2014"
},
{
"authors": "G Thomas; Dietterich",
"journal": "Springer-Verlag",
"ref_id": "b11",
"title": "Ensemble methods in machine learning",
"year": "2000"
},
{
"authors": "Angela Fan; Shruti Bhosale; Holger Schwenk; Zhiyi Ma; Ahmed El-Kishky; Siddharth Goyal; Mandeep Baines; Onur Celebi; Guillaume Wenzek; Vishrav Chaudhary; Naman Goyal; Tom Birch; Vitaliy Liptchinsky; Sergey Edunov; Edouard Grave; Michael Auli; Armand Joulin",
"journal": "",
"ref_id": "b12",
"title": "Beyond english-centric multilingual machine translation",
"year": "2020"
},
{
"authors": "Raphael Gontijo-Lopes; Yann Dauphin; Ekin Dogus; Cubuk ",
"journal": "",
"ref_id": "b13",
"title": "No one representation to rule them all: Overlapping features of training methods",
"year": "2022"
},
{
"authors": "Thanh-Le Ha; Jan Niehues; Alexander Waibel",
"journal": "International Workshop on Spoken Language Translation",
"ref_id": "b14",
"title": "Effective strategies in zero-shot neural machine translation",
"year": "2017"
},
{
"authors": "Geoffrey Hinton; Oriol Vinyals; Jeff Dean",
"journal": "",
"ref_id": "b15",
"title": "Distilling the knowledge in a neural network",
"year": "2015"
},
{
"authors": "Melvin Johnson; Mike Schuster; Quoc V Le; Maxim Krikun; Yonghui Wu; Zhifeng Chen; Nikhil Thorat; Fernanda Viégas; Martin Wattenberg; Greg Corrado; Macduff Hughes; Jeffrey Dean",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b16",
"title": "Google's multilingual neural machine translation system: Enabling zero-shot translation",
"year": "2017"
},
{
"authors": "Urvashi Khandelwal; Angela Fan; Dan Jurafsky; Luke Zettlemoyer; Mike Lewis",
"journal": "",
"ref_id": "b17",
"title": "Nearest neighbor machine translation",
"year": "2021"
},
{
"authors": "Beomsu Kim; Seokjun Seo; Seungju Han; Enkhbayar Erdenee; Buru Chang",
"journal": "Association for Computational Linguistics",
"ref_id": "b18",
"title": "Distilling the knowledge of large-scale generative models into retrieval models for efficient open-domain conversation",
"year": "2021"
},
{
"authors": "Yoon Kim; Alexander M Rush",
"journal": "Association for Computational Linguistics",
"ref_id": "b19",
"title": "Sequencelevel knowledge distillation",
"year": "2016"
},
{
"authors": "Tom Kocmi; Rachel Bawden; Ondřej Bojar; Anton Dvorkovich; Christian Federmann; Mark Fishel; Thamme Gowda; Yvette Graham; Roman Grundkiewicz; Barry Haddow; Rebecca Knowles; Philipp Koehn; Christof Monz; Makoto Morishita; Masaaki Nagata; Toshiaki Nakazawa; Michal Novák; Martin Popel; Maja Popović",
"journal": "Association for Computational Linguistics",
"ref_id": "b20",
"title": "Findings of the 2022 conference on machine translation (WMT22)",
"year": "2022"
},
{
"authors": "Philipp Koehn; Hieu Hoang; Alexandra Birch; Chris Callison-Burch; Marcello Federico; Nicola Bertoldi; Brooke Cowan; Wade Shen; Christine Moran; Richard Zens; Chris Dyer; Ondřej Bojar; Alexandra Constantin; Evan Herbst",
"journal": "Association for Computational Linguistics",
"ref_id": "b21",
"title": "Moses: Open source toolkit for statistical machine translation",
"year": "2007"
},
{
"authors": "Xiang Kong; Adithya Renduchintala; James Cross; Yuqing Tang; Jiatao Gu; Xian Li",
"journal": "Association for Computational Linguistics",
"ref_id": "b22",
"title": "Multilingual neural machine translation with deep encoder and multiple shallow decoders",
"year": "2021"
},
{
"authors": "Shankar Kumar; William Byrne",
"journal": "Association for Computational Linguistics",
"ref_id": "b23",
"title": "Minimum Bayes-risk decoding for statistical machine translation",
"year": "2004"
},
{
"authors": "Bei Li; Yinqiao Li; Chen Xu; Ye Lin; Jiqiang Liu; Hui Liu; Ziyang Wang; Yuhao Zhang; Nuo Xu; Zeyang Wang; Kai Feng; Hexuan Chen; Tengbo Liu; Yanyang Li; Qiang Wang; Tong Xiao; Jingbo Zhu",
"journal": "Association for Computational Linguistics",
"ref_id": "b24",
"title": "The NiuTrans machine translation systems for WMT19",
"year": "2019"
},
{
"authors": "Lemao Liu; Masao Utiyama; Andrew Finch; Eiichiro Sumita",
"journal": "Association for Computational Linguistics",
"ref_id": "b25",
"title": "Agreement on targetbidirectional neural machine translation",
"year": "2016"
},
{
"authors": "Yinhan Liu; Jiatao Gu; Naman Goyal; Xian Li; Sergey Edunov; Marjan Ghazvininejad; Mike Lewis; Luke Zettlemoyer",
"journal": "",
"ref_id": "b26",
"title": "Multilingual denoising pretraining for neural machine translation",
"year": "2020"
},
{
"authors": "Xuezhe Ma; Chunting Zhou; Xiang Kong; Junxian He; Liangke Gui; Graham Neubig; Jonathan May; Luke Zettlemoyer",
"journal": "",
"ref_id": "b27",
"title": "Mega: Moving average equipped gated attention",
"year": "2023"
},
{
"authors": "Benjamin Marie; Raphael Rubino; Atsushi Fujita",
"journal": "Association for Computational Linguistics",
"ref_id": "b28",
"title": "Combination of neural machine translation systems at WMT20",
"year": "2020"
},
{
"authors": "Yasmin Moslem; Rejwanul Haque; John D Kelleher; Andy Way",
"journal": "European Association for Machine Translation",
"ref_id": "b29",
"title": "Adaptive machine translation with large language models",
"year": "2023"
},
{
"authors": "Niklas Muennighoff; Thomas Wang; Lintang Sutawika; Adam Roberts; Stella Biderman; Teven Le Scao; M Saiful Bari; Sheng Shen; Zheng-Xin Yong; Hailey Schoelkopf; Xiangru Tang; Dragomir Radev; Alham Fikri Aji; Khalid Almubarak; Samuel Albanie; Zaid Alyafeai; Albert Webson; Edward Raff; Colin Raffel",
"journal": "",
"ref_id": "b30",
"title": "Crosslingual generalization through multitask finetuning",
"year": "2022"
},
{
"authors": "Nathan Ng; Kyra Yee; Alexei Baevski; Myle Ott; Michael Auli; Sergey Edunov",
"journal": "Association for Computational Linguistics",
"ref_id": "b31",
"title": "Facebook FAIR's WMT19 news translation task submission",
"year": "2019"
},
{
"authors": "Marta R Nllb Team; James Costa-Jussà; Onur Cross; Maha Çelebi; Kenneth Elbayad; Kevin Heafield; Elahe Heffernan; Janice Kalbassi; Daniel Lam; Jean Licht; Anna Maillard; Skyler Sun; Guillaume Wang; Al Wenzek; Bapi Youngblood; Loic Akula; Gabriel Barrault; Prangthip Mejia-Gonzalez; John Hansanti; Semarley Hoffman; Jarrett; Ram Kaushik; Dirk Sadagopan; Shannon Rowe; Chau Spruit; Pierre Tran; Necip Andrews; Shruti Fazil Ayan; Sergey Bhosale; Angela Edunov; Cynthia Fan; Vedanuj Gao; Francisco Goswami; Philipp Guzmán; Alexandre Koehn; Christophe Mourachko; Safiyyah Ropers; Holger Saleem; Jeff Schwenk; Wang",
"journal": "",
"ref_id": "b32",
"title": "No language left behind: Scaling humancentered machine translation",
"year": "2022"
},
{
"authors": "Josef Franz; Daniel Och; Sanjeev Gildea; Anoop Khudanpur; Kenji Sarkar; Alex Yamada; Shankar Fraser; Libin Kumar; David Shen; Katherine Smith; Viren Eng; Zhen Jain; Dragomir Jin; Radev",
"journal": "Association for Computational Linguistics",
"ref_id": "b33",
"title": "A smorgasbord of features for statistical machine translation",
"year": "2004"
},
{
"authors": "Robert Östling; Jörg Tiedemann",
"journal": "Prague Bulletin of Mathematical Linguistics",
"ref_id": "b34",
"title": "Efficient word alignment with Markov Chain Monte Carlo",
"year": "2016"
},
{
"authors": "Kishore Papineni; Salim Roukos; Todd Ward; Wei-Jing Zhu",
"journal": "Association for Computational Linguistics",
"ref_id": "b35",
"title": "Bleu: a method for automatic evaluation of machine translation",
"year": "2002"
},
{
"authors": "Maja Popović",
"journal": "Association for Computational Linguistics",
"ref_id": "b36",
"title": "chrF: character n-gram F-score for automatic MT evaluation",
"year": "2015"
},
{
"authors": "Matt Post",
"journal": "Association for Computational Linguistics",
"ref_id": "b37",
"title": "A call for clarity in reporting BLEU scores",
"year": "2018"
},
{
"authors": "Lihua Qian; Yi Zhou; Zaixiang Zheng; Yaoming Zhu; Zehui Lin; Jiangtao Feng; Shanbo Cheng; Lei Li; Mingxuan Wang; Hao Zhou",
"journal": "Association for Computational Linguistics",
"ref_id": "b38",
"title": "The volctrans GLAT system: Non-autoregressive translation meets WMT21",
"year": "2021"
},
{
"authors": "Alec Radford; Jeff Wu; Rewon Child; David Luan; Dario Amodei; Ilya Sutskever",
"journal": "",
"ref_id": "b39",
"title": "Language models are unsupervised multitask learners",
"year": "2019"
},
{
"authors": "Ricardo Rei; G C José; Duarte De Souza; Chrysoula Alves; Ana C Zerva; Taisiya Farinha; Alon Glushkova; Luisa Lavie; Coheur; F T André; Martins",
"journal": "Association for Computational Linguistics",
"ref_id": "b40",
"title": "COMET-22: Unbabel-IST 2022 submission for the metrics shared task",
"year": "2022"
},
{
"authors": "Libin Shen; Anoop Sarkar; Franz Josef Och",
"journal": "Association for Computational Linguistics",
"ref_id": "b41",
"title": "Discriminative reranking for machine translation",
"year": "2004"
},
{
"authors": "Matthew Snover; Bonnie Dorr; Rich Schwartz; Linnea Micciulla; John Makhoul",
"journal": "Association for Machine Translation in the Americas",
"ref_id": "b42",
"title": "A study of translation edit rate with targeted human annotation",
"year": "2006"
},
{
"authors": "Chau Tran; Shruti Bhosale; James Cross; Philipp Koehn; Sergey Edunov; Angela Fan",
"journal": "Association for Computational Linguistics",
"ref_id": "b43",
"title": "Facebook AI's WMT21 news translation task submission",
"year": "2021"
},
{
"authors": "Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Ł Ukasz Kaiser; Illia Polosukhin",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b44",
"title": "Attention is all you need",
"year": "2017"
},
{
"authors": "Curran Associates; Inc ",
"journal": "",
"ref_id": "b45",
"title": "",
"year": ""
},
{
"authors": "Zhixian Yang; Renliang Sun; Xiaojun Wan",
"journal": "Association for Computational Linguistics",
"ref_id": "b46",
"title": "Nearest neighbor knowledge distillation for neural machine translation",
"year": "2022"
},
{
"authors": "Kyra Yee; Yann Dauphin; Michael Auli",
"journal": "Association for Computational Linguistics",
"ref_id": "b47",
"title": "Simple and effective noisy channel modeling for neural machine translation",
"year": "2019"
},
{
"authors": "Long Zhou; Jiajun Zhang; Chengqing Zong",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b48",
"title": "Synchronous bidirectional neural machine translation",
"year": "2019"
}
] | [
{
"formula_coordinates": [
2,
107.44,
582.64,
182.43,
19.02
],
"formula_id": "formula_0",
"formula_text": "t = arg max t∈N (s) λ • log M(s, t) ⊺ ,(1)"
},
{
"formula_coordinates": [
2,
311.68,
413.1,
202.04,
43.13
],
"formula_id": "formula_1",
"formula_text": "L MIRA (λ) = max t∈N ∆(t) + λ • (M(s, t) ⊺ -M(s, t * ) ⊺ )"
}
] |
2023-12-27 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b39",
"b47",
"b28",
"b26",
"b2",
"b17",
"b44",
"b6",
"b29",
"b34",
"b46",
"b35",
"b21",
"b42",
"b43",
"b41",
"b49"
],
"table_ref": [],
"text": "Multi-task learning (MTL) leverages a single machine learning model to simultaneously address multiple tasks, such as semantic segmentation, depth estimation, and other visionbased prediction tasks. The single model, called multi-task model, shares parameters between tasks and is shown to have lower inference costs and higher generalization performance compared to single-task models without parameter sharing (Ruder 2017;Zhang, Liu, and Guan 2022a;Yao et al. 2020). Multi-task models are widely used in practical applications with high-security requirements, such as robotics and autonomous driving (Leang et al. 2020;Kokkinos 2017;Arcari et al. 2023). In this paper, we focus on the branched multi-task models, which are the most representative in the MTL literature (Zhang, Liu, and Guan 2022b). We use the terms task sharing and parameter sharing interchangeably.\nIn parallel to the developments in MTL, the security of single task classifiers has been put into question due to the existence of adversarial examples (Goodfellow, Shlens, and Szegedy 2014). An adversarial example is an input to a machine learning model (typically an image) that has been manipulated such that the model misclassifies the example with high confidence, but a human can still correctly recognize the input. Adversarial examples can be generated through white-box or black-box attacks depending on the assumed capabilities of the attacker (Tramer et al. 2020;Mahmood et al. 2021a). White-box attacks are generally considered more powerful (Carlini et al. 2019) because the attacker has access to the parameters and structure of the trained model.\nAlthough adversarial attacks have been extensively studied on single-task models (Liu et al. 2016;Mahmood et al. 2022;Xu et al. 2022), related work on multi-task models is scarce. A pioneering study (Mao et al. 2020) pointed out that the adversarial robustness of deep neural networks increases as the number of tasks increases. MTA (Guo et al. 2020) tries to develop attacks in the MTL setting; however, the adversarial samples generated are task-specific and thus fail to attack all tasks simultaneously. Some other works (Gurulingan, Arani, and Zonooz 2021;Sobh et al. 2021) attempted to attack multi-task models by generating adversarial examples for each image while attacking one task at a time. Several critical security research questions (RQ) on MTL remain unclear:\n• RQ1: How secure are multi-task models to conventional single task adversarial attacks? • RQ2: Can adversarial attacks be designed to attack multiple tasks simultaneously? • RQ3: Does task sharing and adversarial training increase multi-task model robustness to adversarial attacks? This paper answers the three questions through careful analysis and rigorous experimentation. To answer RQ1, we develop two naïve adaptations of single task white-box attacks for multi-task models, and analyze their inherent drawbacks. To answer RQ2, we propose a novel attack framework, GB-MTA (Gradient Balancing Multi-Task Attacker) to generate adversarial samples effective in attacking all tasks in a multi-task model. GB-MTA frames the problem of finding a unified attack perturbation in MTL as an optimization problem based on the averaged relative loss change (Sun et al. 2019;Zhang, Liu, and Guan 2022a) across tasks and solves the problem by approximating it as an Integer Linear Programming (ILP). To answer RQ3, we experiment with different levels of task sharing and demonstrate that there is a fundamental trade-off: Improving task accuracy and model efficiency through parameter sharing can increase the model's vulnerability to adversarial attacks designated for these related tasks. We further explore the defense side of MTL, by adversarially training models with examples generated by GB-MTA. Our contributions to advancing the security of the field of MTL are summarized as follows.\n• Dynamic Gradient Balancing Multi-task Attack Framework -We formulate the MTL adversarial attack as an optimization problem with a specially designed multi-task objective function. To solve this optimization problem, we introduce a novel approach, GB-MTA that balances gradients from multiple tasks in a multi-task model when creating adversarial samples that work across all tasks. • Empirical Evaluation -We empirically evaluate the effectiveness of GB-MTA on multi-task models with various levels of task sharing and demonstrate that GB-MTA performs best for 7 out of 8 models on NYUv2 (Silberman et al. 2012) and 6 out of 8 on Tiny-Taskonomy (Zamir et al. 2018) "
},
{
"figure_ref": [],
"heading": "Attack Framework",
"publication_ref": [],
"table_ref": [],
"text": "This section first discusses existing white-box attacks and our adversarial threat model. This section then shows how single task white-box attacks can be adapted to multi-task models.\nWe formulate the GB-MTA framework in the next section."
},
{
"figure_ref": [],
"heading": "Single Task White-Box Attacks",
"publication_ref": [
"b30",
"b6",
"b7",
"b14",
"b46",
"b38",
"b17",
"b30",
"b14",
"b30",
"b42",
"b21",
"b30",
"b48",
"b36"
],
"table_ref": [],
"text": "In general, adversarial attacks can be formulated as follows (Madry et al. 2018). Let (x, y) represent a clean input and its corresponding label. An attacker adds an adversarial perturbation δ to the input x, to maximize the value of a loss function L:\nmax δ L(x + δ, y; θ), s.t. ∥δ∥ p ≤ ϵ,(1)\nwhere θ denotes the parameters of the trained model under attack, and ϵ represents the maximum amount the adversary can perturb the input according to a given p-norm. For notational simplicity, we omit θ in our future derivations.\nThreat Model: In this paper, we focus on the untargeted white-box adversarial threat model (Carlini et al. 2019) as this represents one of the strongest and most widely used adversarial machine learning formulations (Carlini and Wagner 2017;Dong et al. 2018;Croce and Hein 2020b). In this setup, the attacker has knowledge of the model structure, trained model parameters θ and the corresponding loss function L. In terms of bounds on the adversarial perturbation, we use one of the most widely used norms, p = ∞, in line with previous works (Guo et al. 2020;Xu et al. 2022;Rathbun et al. 2022).\nSingle Task Attacks: In the white-box setting, one of the most prevalent strategies for generating the adversarial perturbation δ is to maximize the loss function L by following the gradient ascent direction. This was originally done with the Fast Gradient Sign Method (FGSM) attack proposed in (Goodfellow, Shlens, and Szegedy 2014). Since the advent of FGSM, numerous improvements to the attack have been proposed. Although enumerating all the improvements in FGSM is beyond the scope of this paper, several important attack updates are worth noting. Updated attacks include the Projected Gradient Descent (PGD) attack (Madry et al. 2018), which adds a randomized start and makes FGSM iterative. The Momentum Iterative Momentum (MIM) (Dong et al. 2018) adds momentum to the gradient ascent optimization. More recently, in APGD (Croce and Hein 2020b), an adaptive step size has been shown to be one of the most effective whitebox attacks for a single task, even against adversarial trained models (Mahmood, Mahmood, and Van Dijk 2021).\nNaïve Multi-Task Attacks RQ1: How secure are multi-task models to conventional single task adversarial attacks? This subsection answers the question by presenting two strategies for adapting single task white-box attacks to the multi-task formulation of the problem. It then discusses their inherent flaws, which are further demonstrated empirically. We denote these two attacks as naïve multi-task attacks.\nIn the case of a single task, untargeted white-box attack, an adversarial example can be generated (Madry et al. 2018) iteratively:\nx (i) adv = P S (x (i-1) adv + F δ (ϵ (i-1) , ∂L ∂x (i-1) adv )),(2)\nwhere F δ represents the perturbation function associated with a specific white-box attack, L represents the loss function for a single task, ϵ (i-1) is the magnitude of the perturbation added in the current iteration of the attack and x (0) adv = x. Lastly, P S is the projection operation (Croce and Hein 2020b) to bound the adversarial sample within a specified range. In MTL, each input x is associated with a set of true labels {y 1 , . . . , y n } for tasks T = {t 1 , . . . , t n }. Each task t i has its own task-specific loss function L i (x, y i ).\nSINGLE Attack -The first way in which multi-task models can be attacked is by focusing on only a single task's gradient and ignoring the gradients of the rest tasks (Sobh et al. 2021;Gurulingan, Arani, and Zonooz 2021). For example, from Eq. 2, APGD (Croce and Hein 2020b) can be adapted to attack a single task t j : Figure 1: Attack effectiveness (y-axis, higher-the-better) for each task when applying SINGLE, TOTAL, and the proposed GB-MTA attacks on NYUv2. The variants are built on APGD. Segm: semantic segmentation task; Norm: normal prediction task; Dept: depth estimation task.\nx (i) adv = P S (x (i-1) adv + α(P S (x (i-1) adv + ϵ (i-1) sign( ∂L j x (i-1) adv )) -x (i-1) adv ) + (1 -α)(x (i-1) adv -x (i-2) adv )),\n(3) where α is a hyperparameter in APGD that controls the influence of previous update steps on the current update step. L j is the objective function of the task t j being attacked. In the experiment, we use SINGLE-X to represent performing SINGLE attack on a specific task X.\nTOTAL Attack -The second way in which single task attacks can be converted to attack multi-task models is through totaling all associated task loss functions L i , via summation. For example, Eq. 2 with single task Projected Gradient Descent (PGD) (Madry et al. 2018) can be adapted for TO-TAL-PGD:\nx (i) adv = P S (x (i-1) adv + ϵ (i-1) • sign( n j=1 ∂L j ∂x (i-1) adv )). (4)\nWe show the different attack formulations for SINGLE-X with APGD and TOTAL with PGD, but it is important to note that any combination of existing white-box attacks and adaptations can be made.\nNaïve Attack Limitations: Both the SINGLE and TOTAL attacks come with significant drawbacks. The effectiveness of the SINGLE attack is based on the following assumption: using one task's attack direction will guarantee attack success on all other tasks. However, this assumption does not always hold, as the effectiveness of a SINGLE attack is restricted due to the limited transferability of attack methods (Mahmood, Mahmood, and Van Dijk 2021). For example, in Figure 1, we show that when attacking the segmentation or the depth estimation task solely (SINGLE-SEGM or SINGLE-DEPTH), the effectiveness of the attack on the normal prediction task (Norm ARP) is limited. The definition of ARP is elaborated in the Experiments Section.\nLikewise, the effectiveness of the TOTAL attack is based on an underlying assumption: for an adversarial example to work across all tasks, no one task's gradient should dominate to avoid the limited attack transferability issue that SINGLE faces. We denote it as the non-dominant magnitude assumption. However, the issue of gradient dominance is well recognized within the MTL literature and has garnered significant attention in the field of MTL optimizers (Yu et al. 2020;Navon et al. 2022). Empirically, we observe from Figure 1 that the TOTAL attack exhibits a pattern similar to SINGLE-SEGM, indicating that the segmentation task dominates the gradient directions.\nWe also consider a modified version of TOTAL where we take the sign of the gradients before the summation. In this way, the non-dominant magnitude assumption can be circumvented. We denote this attack as SIGNTOTAL. However, we empirically show that this attack is also not effective in Section , as completely ignoring task's gradient magnitudes also leads to a suboptimal attack.\nThe limitations due to the underlying assumptions of the TOTAL and SINGLE attacks mandate the need for an attack method tailored to multi-task models. The adversarial samples constructed from SINGLE attacks are task-specific, and thus are not effective on non-targeted tasks. On the other hand, although the adversarial samples created from TOTAL attack are task-agnostic, they are effective on only the tasks whose gradients dominate in MTL. An effective multi-task attack should be able to generate task-agnostic adversarial samples that are effective on all tasks in a multi-task model."
},
{
"figure_ref": [],
"heading": "GB-MTA Framework",
"publication_ref": [
"b44",
"b43",
"b27",
"b22"
],
"table_ref": [],
"text": "RQ2: Can adversarial attacks be designed to attack multiple tasks simultaneously? Dynamic Gradient Balancing Multitask Attack (GB-MTA) builds on the success of existing single task adversarial attacks, while addressing the challenge in attacking multi-task models. GB-MTA accomplishes this by actively balancing the gradients across tasks, to derive an adversarial perturbation that is effective on all tasks. In this section, we first formulate a new attack optimization problem tailored for multi-task models. Since the problem is intractable, GB-MTA reformulates it to an Integer Linear Programming (ILP) problem, and then generates adversarial samples by solving the ILP problem.\nMulti-Task Attack Optimization: We first reformulate the original single-task adversarial optimization introduced in Eq. 1 by decomposing δ = η • β. Here, η represents the magnitude of the perturbation. β represents the signed gradient direction vector, with values {-1, 0, 1}.\nmax β L(x + η • β, y) s.t. ∥η • β∥ p ≤ ϵ, ∀β (k) ∈ β : β (k) ∈ {-1, 0, 1}. (5)\nThe above formulation is used to attack a single task. Attacking multiple tasks simultaneously in a multi-task model is fundamentally a multi-objective optimization problem. In this case, each objective function corresponds to one task. In adversarial machine learning, attacks are traditionally measured on a single task using one objective function that measures the attack success rate (Croce and Hein 2020b) or the robustness (Tramer et al. 2020). However, in MTL when evaluating two or more attacks, there is no single metric, as there are multiple tasks and each task has its own objective function value. This makes the comparison of two different attacks in MTL challenging.\nTherefore, we formulate the multi-task attack optimization problem with a multi-task-specific objective function that aligns with the standard practice of assessing model performance in MTL. A multi-task model's performance is typically measured by Average Relative Accuracy (ARA), denoted as ∆Acc, as opposed to using absolute values (Sun et al. 2019;Zhang, Liu, and Guan 2022a). The ARA metric compares the performance of a given multi-task model M to that of a baseline model B:\n∆Acc = 1 N N i=1 Acc M,ti -Acc B,ti Acc B,ti ,(6)\nwhere N is the number of tasks. ∆Acc is the average difference in accuracy between Acc M,ti and Acc B,ti in all tasks t i , normalized by the accuracy of B. A higher ∆Acc indicates better model performance compared to the baseline. When attacking a multi-task model, the goal is to substantially reduce the task performance of M relative to B, where now B represents the model before the attack and M denotes the model after the attack. ∆Acc will be a negative value, and the higher its absolute value is, the more effective the attack is. To find a perturbation direction β that is the most effective, we reformulate the objective function in Eq. 5:\nβ * = arg max β |∆Acc| = arg max β ∆L = arg max β 1 N n i=1 L i (x + η • β, y i ) -L i (x, y i ) L i (x, y i ) ,(7)\nwhere L i (x + η • β, y i ) -L i (x, y i ) represents the model loss difference for task t i before and after the attack to substitute the accuracy difference, since task loss in neural networks typically serves as a reliable indicator of task accuracy (i.e., higher task loss corresponds to lower task accuracy).\nThe optimization problem mentioned above can identify the optimal signed gradient direction vector β * but is intractable (Kreinovich, Lakeyev, and Noskov 1996;Horáček, Hladík, and Černỳ 2017). To address the problem, we apply the Taylor Expansion on L i (x + η • β, y i ), and reformulate it to an ILP problem as follows:\nβ * = arg max β i β • ∂L i (x, y i ) ∂x • 1 L i (x, y i ) s.t. ∀β (k) ∈ β : β (k) ∈ {-1, 0, 1}.(8)\nIn practice, β and ∂Li(x,yi) ∂x are two matrices the same size as x. Their product represents the dot product of the corresponding vectorized matrices.\nOptimization Solution via Relaxed LP: GB-MTA identifies β * by first addressing a relaxed Linear Programming (LP) problem and then rounding the resulting solution to obtain an integer solution. In the first step, the LP relaxation will remove the requirement of integer values, i.e. ∀β (k) ∈ β : β (k) ∈ {-1, 0, 1}, allowing them to be any real value instead. To solve the relaxed LP, we calculate the derivative of the objective function with respect to the variable β as follows,\n∂∆L ∂β = n i=1 ∂L i (x, y i ) ∂x • 1 L i (x, y i ) .(9)\nIn other words, β starts from the original state, that is, a zero matrix, and is updated along the direction of ∂∆L ∂β to maximize ∆L in the LP relaxation. Then in the second step, GB-MTA reintroduces the integer constraint, and the solution for the original ILP in Eq. 14 can be obtained by performing a rounding operation: β * = sign(β).\nThe optimal β * suggests that the effective attack direction for multi-task models should be the sum of each task's gradients dynamically weighted by its loss value. GB-MTA mitigates the dominating task issue in TOTAL by dynamically balancing the gradients across tasks and avoids the limited transferability problem in SINGLE-X by optimizing over all tasks simultaneously.\nIntegrating GB-MTA with Existing Attacks: Integrating GB-MTA with any existing attack can easily be accomplished, even for more advanced methods such as APGD (Croce and Hein 2020b). The key operation is to substitute the single task gradient, i.e., ∂L ∂x , with the balanced multitask counterpart, i.e.,\nn i=1 ∂Li(x,yi) ∂x • 1 Li(x,yi) .\nTo illustrate. we provide the pseudocode for APGD integrated with GB-MTA in Algorithm 3. The black text corresponds to the original APGD algorithm (Croce and Hein 2020b) and the part changed for GB-MTA is colored blue. Notice that on top of the key design shown in lines 2 and 11, we also change the absolute loss value used in the original APGD to the relative loss value sum over all the tasks in lines 4 and 14 to align with the MTL scenario.\nAlgorithm 1 GB-MTA-APGD Input: x (0) , {yi}, {Li}, η, α, Niter, attack checkpoints:W\nOutput: xmax 1: li ← Li(x (0) , yi), ∀i = 1, • • • , n. 2: β * ← sign( n i ∂L i (x (0) ,y i ) ∂x (0) • 1 L i (x (0) ,y i ) ) 3: x (1) ← P (x (0) + η • β * ) 4: lmax ← max{ n i L i (x (0) )-l i l i , n i L i (x (1) )-l i l i } 5: if lmax ≡ n i L i (x (0) )-l i l i then 6: xmax ← x (0) 7: else 8: xmax ← x (1) 9: end if 10: for k = 1 to Niter -1 do 11: β * ← sign( n i ∂L i (x (k) ,y i ) ∂x (k) • 1 L i (x (k) ,y i ) ) 12: z k+1 ← P (x (k) + η • β * ) 13: x k+1 ← P (x (k) + α(z k+1 -x (k) ) + (1 -α)(x (k) - x (k-1) )) 14: if n i L i (x (k+1) )-l i l i > lmax then 15: xmax ← x (k+1) 16: lmax ← n i L i (x (k+1) )-l i l i 17: end if 18: if k ∈ W then 19:\nupdate η and x (k+1) 20:\nend if 21: end for"
},
{
"figure_ref": [],
"heading": "Experiments Experimental Settings",
"publication_ref": [
"b41",
"b49",
"b40",
"b17",
"b30"
],
"table_ref": [],
"text": "Datasets and Tasks: We use two popular datasets in multitask learning (MTL), NYUv2 (Silberman et al. 2012) and Tiny-Taskonomy (Zamir et al. 2018). The NYUv2 dataset consists of RGB-D indoor scenes and three tasks, 40-class semantic segmentation, depth estimation, and surface normal prediction. Tiny-Taskonomy contains RGB indoor images, and its five representative tasks are semantic segmentation, surface normal prediction, depth estimation, keypoint detection, and edge detection.\nEvaluation Metrics and Loss Functions: Semantic segmentation uses a pixel-wise cross-entropy loss for each predicted class label. Surface normal prediction uses the inverse of cosine similarity between the normalized prediction and ground truth. All other tasks use the l 1 loss. Many tasks have distinct evaluation metrics with various scales. Hence, it is crucial to assess task performance in an equitable manner. Compounding the problem of different scales is the fact that some metrics are higher-the-better (e.g., accuracy, mean of intersection over union), while others are lower-the-better (e.g., distance, error). To address these issues and fairly measure the success of various attacks, we formulate a multi-task attack metric, Average Relative Performance (ARP):\n1 N N i=1 1 M i Mi j=1 (-1) si,j (m ′ i,j -m i,j )/m i,j × 100%, (10\n)\nwhere m i,j and m ′ i,j represent the values of task t i 's j-th metric for the model before and after the attack respectively, and s i,j equals 0 if this metric is lower-the-better and 1 otherwise. M i denotes the number of metrics for task t i , and N is the number of tasks. For each attack, we measure the corresponding ARP. A higher ARP indicates a higher performance drop and thus a more effective attack.\nMulti-Task Models: We evaluate branched multi-task models from TreeMTL (Zhang, Liu, and Guan 2022b) using two backbone architectures: Deeplab-ResNet34 (Chen et al. 2017a) and MobileNetV2 (Sandler et al. 2018). We randomly sampled and trained 25 models with Deeplab-ResNet34 and 20 with MobileNetV2 for the NYUv2 dataset. For the Tiny-Taskonomy dataset, we sampled 15 models with Deeplab-ResNet34. These models cover a range of task sharing configurations from the all-shared models to those comprising an ensemble of independent single-task models.\nCounterparts for Comparison: We compare GB-MTA to two types of baselines. The first type is the naïve multi-task attacks that repurpose existing single-task white-box attacks to multi-task models. It includes TOTAL, SIGNTOTAL, and SINGLE-X. We compare these baselines with GB-MTA by integrating them with three different single-task white-box attack methods, FGSM (Goodfellow, Shlens, and Szegedy 2014), PGD (Madry et al. 2018), and APGD (Croce and Hein 2020b). The second type of baseline is a multi-model attack method called Auto-SAGE, which is designed to attack multiple independent DNNs but can be directly applied to attack multi-task models. To compare fairly with Auto-SAGE, we integrate GB-MTA with Auto-SAGE instead of other single-task attacks."
},
{
"figure_ref": [
"fig_0",
"fig_4"
],
"heading": "Results on Attack Performance",
"publication_ref": [],
"table_ref": [
"tab_1",
"tab_2"
],
"text": "This subsection compares GB-MTA with baselines on their effectiveness in attacking multi-task models. Tables 1 and2 compare the attack performance of the baselines and GB-MTA at the model level on NYUv2 and Tiny-Taskonomy respectively at ϵ = 8. Overall, GB-MTA integrated with PGD and APGD have the highest ARP in 7 out of the 8 models on NYUv2, and in 6 out of the 8 models on Tiny-Taskonomy.\nGB-MTA outperforms baselines TOTAL, SIGNTOTAL, and SINGLE-X in attack performance because it alleviates the baselines' limitations discussed in Attack Framework. SINGLE-X achieves limited attack effectiveness on tasks that are not the attack target X; TOTAL shares a similar pattern of attack effectiveness across tasks as SINGLE-SEGM due to the issue of gradient dominance, making it less effective in attacking all tasks simultaneously. In contrast, GB-MTA dynamically balances the attack directions for all tasks, making it more threatening for systems that require high robustness. This is also the reason why GB-MTA outperforms the multimodel attack approach AutoSAGE -GB-MTA balances the gradients in AutoSAGE.\nWe further vary the maximum perturbation bound ϵ from 1 to 16 and report the average of the ARP over all multitask models with diverse architectures in Figure 2 (Deeplab-ResNet34) and Figure 5 (MobileNet). We denote the average ARP as the overall ARP. GB-MTA is the best-performing attack in almost all ϵ and dataset settings, and when integrated with any of the three existing single-task (i.e., FGSM, PGD, APGD) or the multi-model attack (i.e., Auto-SAGE) approaches. There are a few cases, where for ϵ ≥ 10, GB-MTA and TOTAL converge or perform almost identically. This is because at higher ϵ values, the magnitude of the noise becomes larger (and thus more visible) and all attacks become more effective."
},
{
"figure_ref": [
"fig_2"
],
"heading": "Results on Attack Transferability",
"publication_ref": [],
"table_ref": [],
"text": "RQ3.1: Does task sharing increase the robustness of multitask models to adversarial attacks? Existing literature in multi-task learning determines the appropriate level of parameter sharing with a focus on optimizing task accuracy. The results reported in this section, however, reveal a fundamental trade-off between the improvement in task accuracy due to positive task interactions and the increased vulnerability to adversarial attack due to the greater transferability of attacks from parameter sharing. We observe that a higher degree of parameter sharing between correlated tasks is associated with increased attack transferability.\nWe first define attack transferability in the MTL context. When using an attack method like SINGLE-X, which attacks only one task in a multi-task model, the targeted task X, would be impacted the most. Referring to Figure 1, we can see that this occurs, since the highest bar for each of the SINGLE-X attacks is the targeted task X. Therefore, we consider the degradation of the performance of the targeted task X, represented by ARP-X, as the upper bound for the attack effectiveness of SINGLE-X. Since adversarial examples designed to attack one task may also be adversarial (e.g. misclassified) for another task, we can quantify this phenomenon in the MTL domain. Let X represent a task under attack, and Y represent another task. The transferability of attack SINGLE-X is\n1 n -1 Y̸ =X ARP-Y ARP-X ,(11)\nwhere n represents the number of tasks. In short, we measure transferability as a performance degradation ratio. This ratio is the performance of the attack on all non-targeted tasks versus the performance of the attack on the targeted task. Figure 4 illustrates the relationship between the levels of task sharing and attack transferability, showcasing the results for six multi-task models attacked by APGD SINGLE-X variants. These models represent six levels of parameter sharing, ranging from all-share (AS/5L), where all layers of the backbone model (5L) are shared, to independent models (IND/0L), where no layers are shared. We observe a positive correlation between the degree of task sharing and the transferability of attacks in the three SINGLE-X variants. As the level of task sharing increases, attack transferability also increases, from 0.08 to 0.15 (1.875×) for SINGLE-DEPT, 0.02 to 0.46 (23×) for SINGLE-SEGM, and 0.17 to 0.53 (3.12×) for SINGLE-NORM. These findings suggest a trade-off in multi-task model design: While sharing more parameters among related tasks can enhance task accuracy, it also amplifies attack transferability, thereby reducing model robustness, even when facing single task attacks. This underlines the importance of balancing accuracy and robustness in multi-task model design."
},
{
"figure_ref": [],
"heading": "Attack on Adversarially Trained Multi-Task Models",
"publication_ref": [
"b30",
"b5",
"b50",
"b50"
],
"table_ref": [],
"text": "RQ3.2: Does adversarial training increase MTL model robustness to adversarial attacks? A common single task strategy to defend against adversarial attacks is to leverage adversarial training (Madry et al. 2018;Bai et al. 2021;Zhang et al. 2020). This defense involves generating adversarial samples and using them as part of the dataset during training. This typically results in reduced model performance on clean inputs, but increased robustness to adversarial attacks. To the best of our knowledge, no existing work has dealt with the open question of whether adversarial training can be applied effectively to multi-task models. Table 3: ARP of multi-task models trained with and without adversarial training and then attacked by multi-task attacks. The multi-task attacks include GB-MTA, SINGLE, and TO-TAL with ϵ = 8, while the adversarial training is the FAT version of them with K = τ = 20. We adopt the single task Friendly Adversarial Training (FAT) (Zhang et al. 2020) to MTL. FAT is a PGD-based adversarial training method. In order to apply this defense effectively to MTL, we modify the default PGD to GB-MTA-PGD to use adversarial examples generated from multiple tasks as opposed to just single task adversarial examples. Our new defense is coined FAT GB-MTA. We also train multitask models using FAT with the naïve MTL attacks SINGLE and TOTAL. After adversarial training, we re-evaluate the multi-task attacks to assess both model robustness and attack effectiveness.\nTable 3 reports the ARP of 30 cases, i.e., (without adversarial training + 5 adversarially trained models) × 5 attack methods. We make two main observations. First, the notable reduction in ARP of attack methods (e.g., from 105.74% to 12.76% ∼ 25.52% when attacking with SINGLE-DEPT) demonstrates that models enhanced with adversarial training are substantially more robust than the model without such training. Second, from the perspective of attack performance, GB-MTA is still the most effective attack method, consistently outperforming the SINGLE-X and TOTAL baselines by up to 18.65%."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "In this work, we make significant developments in multitask learning (MTL) security through novel attack design, empirical exploration with multi-task models and robustness analyses. We first analyzed naïve adaptions of single task white-box attacks to the MTL domain and experimentally demonstrated their ineffectiveness. We then developed a new framework, Dynamic Gradient Balancing Multi-task Attack (GB-MTA), to effectively attack all tasks in multitask models. On models trained on the NYUv2 and Tiny-Taskonomy datasets, GB-MTA achieves the highest overall attack strength and is the strongest attack on 6 out of 8 models for both datasets. We further analyzed the adversarial transferability of MTL adversarial examples and discovered a new phenomenon: Task sharing can lead to increased adversarial transferability. Lastly, from the defense side, we adversarially trained multi-task models in a new approach, which we coined FAT GB-MTA. GB-MTA is the most effective attack, even on these multi-task models, showcasing its effectiveness as a benchmark for analyzing future MTL defenses."
},
{
"figure_ref": [],
"heading": "Related Works",
"publication_ref": [
"b6",
"b55",
"b45",
"b30",
"b14",
"b1",
"b3",
"b7",
"b39",
"b47",
"b23",
"b24",
"b15",
"b26",
"b37",
"b43",
"b0",
"b18",
"b35",
"b25",
"b16",
"b21",
"b42"
],
"table_ref": [],
"text": "Adversarial Attacks. Adversarial attacks fall into two main categories: white-box and black-box attacks (Mahmood et al. 2021b). In white-box attacks, an attacker has access to the target model's internal information, enabling direct gradient extraction and adversarial example generation (Carlini et al. 2019). On the contrary, in black-box attacks, an attacker has limited model knowledge and uses alternative information sources (Chen et al. 2017b;Zhou et al. 2020) to create adversarial examples. This paper focuses on white-box attacks, as they are generally more effective compared to black-box attacks (Croce and Hein 2020b;Wang et al. 2022).\nIn recent years, many white-box attack techniques have been developed. Fast Gradient Sign Method (FGSM) (Goodfellow, Shlens, and Szegedy 2014) generates adversarial examples by introducing non-random noise in the gradient direction of the loss function. Projected Gradient Descent (PGD) (Madry et al. 2018) and Momentum Iterative Method (MIM) (Dong et al. 2018) improve FGSM by generating adversarial samples in an iterative process. Later, Croce and Hein (Croce and Hein 2020b) proposed Auto Projected Gradient Descent (APGD) with adaptive step size and combined it with two complementary attacks (Croce and Hein 2020a;Andriushchenko et al. 2020), developing the ensemble method APGD that outperforms existing methods on diverse benchmark datasets. In addition to gradient-based strategies, alternative methods have emerged. For instance, the Backward Pass Differentiable Approximation (BPDA) (Athalye, Carlini, and Wagner 2018) accommodates non-differentiable functions, while the Carlini and Wagner (C&W ) attack (Carlini and Wagner 2017) perturbs images with minimal delta to misclassify them.\nMulti-Task Learning. In Multi-Task Learning (MTL), researchers develop memory and computation-efficient multitask models that simultaneously address multiple tasks (Ruder 2017;Yao et al. 2020). The main challenge lies in determining the parameters to share across tasks to optimize both resource efficiency and task accuracy. This has led to an abundance of multi-task model architectures designed either manually (Huang et al. 2015;Jou and Chang 2016;Dvornik et al. 2017;Kokkinos 2017;Ranjan, Patel, and Chellappa 2017) or automatically (Sun et al. 2019;Zhang, Liu, and Guan 2022b;Ahn, Kim, and Oh 2019;Guo, Lee, and Ulbricht 2020;Zhang, Liu, and Guan 2022a). This paper focuses on attacking branched multi-task models, which are the most representative in the MTL literature. Besides, branched multitask models have a wide range of sharing patterns across tasks, facilitating the study of the relationship between the robustness of multi-task models and their sharing patterns.\nRobustness of Multi-Task Models. Few studies have examined the robustness of the model in MTL settings. A pioneering study (Mao et al. 2020) pointed out that the adversarial robustness of deep neural networks increases as the number of tasks increases. Subsequent research (Klingner, Bar, and Fingscheidt 2020;Ghamizi et al. 2022) further emphasizes the importance of selecting suitable tasks for joint learning to create more robust models. While these studies offer intriguing insights, they do not specifically propose adversarial attack methods for multi-task models. Regarding attacks on multi-task models, MTA (Guo et al. 2020) tries to develop attacks in the MTL setting, however, the generated adversarial samples are task-specific and thus fail to attack all the tasks simultaneously. Some other work (Gurulingan, Arani, and Zonooz 2021;Sobh et al. 2021) attempted to attack the multi-task model by generating adversarial examples for each image while attacking one task at a time."
},
{
"figure_ref": [],
"heading": "MTL Attack Optimization Approximation",
"publication_ref": [],
"table_ref": [],
"text": "The optimization problem we formulate for multi-task attack in Section 3 of the main paper is,\nβ * = arg max β ∆L = arg max β 1 N n i=1 L i (x + η • β, y i ) -L i (x, y i ) L i (x, y i ) .(12)\nHere, L i (x + η • β, y i ) -L i (x, y i ) represents the model loss difference for task t i before and after the attack.\nAs the problem is intractable, we make some approximations and reformulate it be an Integer Linear Programming (ILP) problem. To do so, we first apply the Taylor expansion on L i (x + η • β, y i ) in the numerator of the objective function at the point of x:\nL i (x + η • β, y i ) -L i (x, y i ) = L i (x, y i ) + η • β • ∂L i (x, y i ) ∂x + ξ -L i (x, y i ) ≈ η • β • ∂L i (x, y i ) ∂x .(13)\nWe ignore the remainder ξ because, in the context of adversary attacks, we have ∥η • β∥ p ≤ ϵ, indicating that the change η • β is sufficiently small. The proposed approximate optimization problem for multitask attacks is thus formulated as follows:\nβ * = arg max β 1 N n i=1 η • β • ∂L i (x, y i ) ∂x • 1 L i (x, y i ) s.t. ∀β (k) ∈ β : β (k) ∈ {-1, 1},(14)\nwhere 1\nN and η are constants that can be ignored when solving the optimization problem. In practice, β and ∂Li(x,yi) ∂x are two matrices with the same size as x, thus their product represents the dot product of the corresponding vectorized matrices."
},
{
"figure_ref": [],
"heading": "Algorithm Pseudocode",
"publication_ref": [],
"table_ref": [],
"text": "As presented in Section 3 in the main paper, integrating GB-MTA with any existing attack can be easily accomplished by substituting the single-task gradient, i.e., ∂L(x) ∂x , with the balanced multi-task counterpart, i.e., n i=1 ∂Li(x,yi) ∂x\n• 1 Li(x,yi) . We illustrate how to integrate GB-MTA into a single-task attack algorithm APGD and a multi-model attack algorithm Auto-SAGE."
},
{
"figure_ref": [],
"heading": "APGD and GB-MTA-APGD",
"publication_ref": [],
"table_ref": [],
"text": "We provide pseudocode comparisons for APGD with and without integrating GB-MTA in Algorithms 2 and 3. We color the difference in the two algorithms in blue. To integrate GB-MTA in APGD, we first change the key part of the code determining the attack direction (lines 2 and 11) to the proposed balanced multi-task gradients. Then we further update the objective function (lines 4 and 14) from the absolute loss value to the relative loss value sum over all tasks to accommodate the multi-task setting. All other lines in Algorithm 3 are kept the same as the original APGD.\nAlgorithm 2 APGD Input: x (0) , L, η, α, N iter , attack checkpoints:W Output: x max 1: x (1) ← P (x (0) + η • sign(∇L(x (0) ))) 2: l max ← max{L(x (0) ), L(x (1) )} 3: if l max ≡ L(x (0) ) then 4:\nx max ← x (0) 5: else 6:\nx max ← x (1) 7: end if 8: for k = 1 to N iter -1 do 9: z k+1 ← P (x (k) + η • sign(∇L(x (k) ))) 10: x k+1 ← P (x (k) + α(z k+1 -x (k) ) + (1 -α)(x (k) - x (k-1) )) 11: if L(x (k+1) ) > l max then 12:\nx max ← x (k+1) 13:\nl max ← L(x (k+1) ) 14: end if 15: if k ∈ W then 16: if Condition 1 1 or Condition 2 2 then 17: η ← η/2 18: x (k+1) ← x max 19: end if 20: end if 21: end for Algorithm 3 GB-MTA-APGD Input: x (0) , {y i }, {L i }, η, α, N iter , W Output: x max 1: l i ← L i (x (0) , y i ), ∀i = 1, • • • , n. 2: β * ← sign( n i ∂Li(x (0) ,yi) ∂x (0) • 1 Li(x (0) ,yi) ) 3: x (1) ← P (x (0) + η • β * ) 4: l max ← max{ n i Li(x (0) )-li li , n i Li(x (1) )-li li } 5: if l max ≡ n i\nLi(x (0) )-li li then 6:\nx max ← x (0) 7: else 8:\nx max ← x (1)\n1 counts in how many cases since the last checkpoint the update step has been successful in increasing the loss value. If this happened for at least 75% of the total update steps, then the step size is kept.\n2 holds true if the step size was not reduced at the last checkpoint and there has been no improvement in the best found objective value since the last checkpoint. 9: end if 10: for k = 1 to N iter -1 do 11:\nβ * ← sign( n i ∂Li(x (k) ,yi) ∂x (k) • 1 Li(x (k) ,yi) ) 12: z k+1 ← P (x (k) + η • β * ) 13: x k+1 ← P (x (k) + α(z k+1 -x (k) ) + (1 -α)(x (k) - x (k-1) )) 14: if n i Li(x (k+1) )-li li\n> l max then 15:\nx max ← x (k+1) 16:\nl max ← n i Li(x (k+1) )-li li 17: end if 18: if k ∈ W then 19:\nif Condition 1 or Condition 2 then 20:\nη ← η/2 21:\nx (k+1) ← x max 22:\nend if"
},
{
"figure_ref": [],
"heading": "23:",
"publication_ref": [
"b38"
],
"table_ref": [],
"text": "end if 24: end for Auto-SAGE and GB-MTA-Auto-SAGE For Auto-SAGE (Rathbun et al. 2022), the original attack formulation is\nG blend (x (i) adv ) = γG blend (x (i-1) adv ) + k∈D\\R α (i) k ϕ (i) k ⊙ ∂L k ∂x (i) adv + r∈R α (i) r ϕ (i) r ⊙ (Et∼T [ ∂Lr ∂t(x(i) adv )\n]).\n(15) To integrate GB-MTA with Auto-SAGE, we normalize the gradient of the objective function L k and L r with the objective function value. The updated attack will be,\nG blend (x (i) adv ) = γG blend (x (i-1) adv ) + k∈D\\R α (i) k ϕ (i) k ⊙ ( ∂L k ∂x (i) adv • 1 L k ) + r∈R α (i) r ϕ (i) r ⊙ (Et∼T [ ∂Lr ∂t(x (i) adv ) • 1 Lr ]).\n(16)"
},
{
"figure_ref": [],
"heading": "More Experimental Results",
"publication_ref": [],
"table_ref": [],
"text": "This section reports evaluation metrics and more experimental results that are ommited from the main paper."
},
{
"figure_ref": [],
"heading": "Evaluation Metrics",
"publication_ref": [],
"table_ref": [],
"text": "Semantic segmentation is evaluated using mean Intersection over Union and Pixel Accuracy (mIoU and Pixel Acc, the higher the better) in NYUv2. Surface normal prediction is evaluated using mean and median angle distances between the prediction and the ground truth (the lower the better), and the percentage of pixels whose prediction is within the angles of 11.25 • , 22.5 • and 30 • to the ground truth (the higher the better).\nDepth estimation uses the absolute and relative errors between prediction and ground truth (the lower the better). Furthermore, the percentage of pixels whose prediction is within the thresholds of 1.25, 1.25 2 , 1.25 3 to the ground truth, i.e. δ = max{ p pred pgt , pgt p pred } < thr, is used (the higher the better). Tiny-Taskonomy is evaluated using the task-specific loss of each task directly."
},
{
"figure_ref": [
"fig_4",
"fig_2"
],
"heading": "Results of Attack Performance",
"publication_ref": [],
"table_ref": [
"tab_5",
"tab_1",
"tab_6"
],
"text": "Figure 5 illustrates the attack performance on NYUv2 with MobileNetV2. To be consistent with the main paper, we conduct the experiments using different variants of GB-MTA and the naïve multi-task attacks. The x-axis represents the perturbation bound ϵ ranging from 1 to 16, while the y-axis displays the overall Average Relative Performance (ARP, higher-thebetter). Overall, GB-MTA consistently outperforms baselines.\nWe present the full tables of ARP after the attack of all 25 multi-task models trained on NYUv2 with Deeplab-ResNet34 for perturbation bound ϵ = 8 in Tables 4 and5. Similarly, Tables 6 and7 show the attack results for all 15 multi-task models for Tiny-Taskonomy. Overall, GB-MTA achieves 80% first place (80 out of 100 cases) on NYUv2 and 88.33% (53 out of 60) on Tiny-Taskonomy, demonstrating the effectiveness of adversarial samples from GB-MTA.\nWe also show the evaluation results with perturbation bound ϵ = 4 in Tables 8 and9 for NYUv2 and Tables 10 and11 for Taskonomy. 4 in the main paper, including the results for six multi-task models attacked by APGD SINGLE-X variants. These models represent six levels of parameter sharing, ranging from all-share (AS/5L), where all five layers of the backbone model (5L) are shared, to independent models (IND/0L), where no layers are shared. We observe a distinct trend where the attack transferability decreases along with the reduction in levels of parameter sharing, irrespective of the specific attack method utilized."
},
{
"figure_ref": [],
"heading": "Results of MTL Adversarial Transferability",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Results of Attack Performance on Adversarially Trained Multi-Task Models",
"publication_ref": [
"b50"
],
"table_ref": [
"tab_8",
"tab_1",
"tab_9",
"tab_1",
"tab_8",
"tab_9",
"tab_1",
"tab_1",
"tab_9"
],
"text": "As introduced in Section 4.4 in the main paper, we adopt the single task Friendly Adversarial Training (FAT) (Zhang et al. 2020) in the MTL context. Specifically, we modify the underlying adversarial samples generation process to the multi-task attacks we investigated in this paper, including naive multi-task adaptations SINGLE and TOTAL, and the proposed GB-MTA. The detailed adversarial training algorithm is described in Algorithm 4, where for each mini-batch, we generate the adversarial samples with GB-MTA-PGD to train the model.\nTable 13 reports the accuracy of multi-task models trained on NYUv2 with adversarial training. It includes metrics for all tasks and performance degradation (shown in columns containing ARP). We employ PGD-based naive multi-task attacks SINGLE-X and Total as well as GB-MTA with ϵ = 8 to generate different adversarial data. It can be seen that after adversarial training, the average accuracy of the multitask model is dropped by 13.75% ∼ 16.53% compared with training with clean data only (the \"w/o AT\" row).\nThe same phenomenon of decreased model accuracy and increased model robustness can also be observed in Tables 14 and15, where FGSM-based multi-task attack variants are utilized when generating adversarial samples in adversarial training. Table 14 reports the accuracy of the adversarially trained multi-task models similar to Table 13, while Table 15 shows the ARP of multi-task models trained with and without FAT in Table 14 and attacked by multi-task attacks including FGSM variants of GB-MTA, SINGLE, and TOTAL with ϵ = 8.\nWe first observe a 3.58% ∼ 5.54% accuracy drop with adversarial training from Table 14, that is, decreased model performance. Then we observe an increase in model robustness after adversarial training from the lower ARP after the attack reported in Table 15. For instance, when attacked by GB-MTA, ARP decreased from 39.40% to 9.65% ∼ 17.09%. Furthermore, GB-MTA remains the most effective attack method, consistently outperforming the SINGLE-X and TO-TAL baselines by up to 6.87%."
},
{
"figure_ref": [
"fig_0"
],
"heading": "Visualization Results",
"publication_ref": [],
"table_ref": [],
"text": "Figure 6 visualizes adversarial samples generated given an image from (a) NYUv2 and (b) Taskonomy with different attack methods including SINGLE-X attack, TOTAL attack, and the proposed GB-MTA and various attack strengths ϵ from 0 to 16. We show that along with the increase in the perturbation bound ϵ, the magnitude of the noise becomes larger and more visible. All attacks become more effective and thus lead to similar attack performance as shown in Figures 1 and2 of the main paper."
},
{
"figure_ref": [],
"heading": "Multi-Task Architectures with Model Index",
"publication_ref": [],
"table_ref": [
"tab_1",
"tab_1",
"tab_1",
"tab_8"
],
"text": "We show the effectiveness of the proposed multi-task attack GB-MTA by conducting attack experiments on multitask models with different sharing levels in the main paper. Tables 16 and17 present the multi-task model architectures in the Layout format proposed by TreeMTL (Zhang, Liu, and Guan 2022b). A layout is a symbolized representation of a tree-structured multi-task architecture. For T tasks and a backbone model with B branching points, a layout\nL = [L 1 , L 2 , • • • , L B ],\nwhere L i is a list of task sets at the i-th branching point. Task sets in L Update model θ with the adversarial samples 9: end for Table 4: ARP of all 25 multi-task models with diverse sharing patterns trained on NYUv2 and attacked by FGSM and PGD variants with perturbation bound ϵ = 8. For brevity, the name of SINGLE-X variants are simplified to the task name only. IND: independent, AS: all-shared. Table 14: The accuracy of the adversarially trained multi-task models similar to Table 13 \ni = [L 1 i , L 2 i , • • • ] are sub- sets of"
},
{
"figure_ref": [],
"heading": "Acknowledgement",
"publication_ref": [],
"table_ref": [],
"text": "This material is based upon work supported by the National Science Foundation under Grant No. 2312396, 2220211, 2224054, and 2247893. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation."
}
] | Multi-Task Learning (MTL) involves developing a singular model, known as a multi-task model, to concurrently perform multiple tasks. While the security of single-task models has been thoroughly studied, multi-task models pose several critical security questions, such as 1) their vulnerability to single-task adversarial attacks, 2) the possibility of designing attacks that target multiple tasks, and 3) the impact of task sharing and adversarial training on their resilience to such attacks. This paper addresses these queries through detailed analysis and rigorous experimentation. First, we explore the adaptation of single-task white-box attacks to multi-task models and identify their limitations. We then introduce a novel attack framework, the Gradient Balancing Multi-Task Attack (GB-MTA), which treats attacking a multi-task model as an optimization problem. This problem, based on averaged relative loss change across tasks, is approximated as an integer linear programming problem. Extensive evaluations on MTL benchmarks, NYUv2 and Tiny-Taxonomy, demonstrate GB-MTA's effectiveness against both standard and adversarially trained multi-task models. The results also highlight a trade-off between task accuracy improvement via parameter sharing and increased model vulnerability due to enhanced attack transferability. | Multi-Task Models Adversarial Attacks | [
{
"figure_caption": "Figure 2 :2Figure 2: Attack performance comparisons in terms of ARP averaged over 25 multi-task models trained on NYUv2 with Deeplab-ResNet34. The naive attacks and GB-MTA variants are built on (a) FGSM, (b) PGD, (c) APGD, and (d) Auto-SAGE. The perturbation bound ϵ ranges from 1 to 16.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: Attack performance comparisons in terms of ARP on NYUv2 with MobileNetV2 similar to Figure 2.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure4: The relationship between the levels of task sharing in multi-task models (x-axis) and the attack transferability (z-axis). The y-axis represents APGD variants SINGLE-X.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "tasks T and a task set L p i means the set of tasks in L p i sharing the i-th block.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 :5Figure 5: Attack performance comparisons in terms of ARP averaged over 20 multi-task models trained on NYUv2 with MobileNetV2. The naive attacks and GB-MTA variants are built on (a) FGSM, (b) PGD, (c) APGD, and (d) Auto-SAGE. The perturbation bound ϵ ranges from 1 to 16.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": ". • Multi-Task Models Robustness Trade-off -We empirically demonstrate that task sharing can undermine model robustness due to increased attack transferability.",
"figure_data": "",
"figure_id": "tab_0",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "ARP of 8 multi-task models with diverse sharing patterns trained on NYUv2 and attacked by PGD, AutoAttack, and Auto-SAGE variants with perturbation bound ϵ = 8. For brevity, the name of SINGLE-X variants are simplified to the task name only. IND: independent, AS: all-shared. SAGE Segm Norm Dept Total SignTotal GB-MTA Segm Norm Dept Total SignTotal GB-MTA Baseline GB-MTA IND 63.83 28.18 18.23 62.14 50.52 67.85 73.41 30.74 23.52 75.62 58.30 69.29",
"figure_data": "Model Index#Params (M)PGDAutoAttackAuto-87.58 71.0788.53562.48 29.12 30.00 70.99 61.40 71.4679.60 32.06 39.25 88.03 70.34 73.0594.4484.5195.703562.25 36.46 32.54 98.11 68.30 91.24100.53 41.01 41.71 119.52 78.10 95.22121.30 124.97 131.564161.13 38.06 46.02 100.41 67.51 95.20103.03 42.09 59.54 120.39 75.42 99.05122.50 113.15 126.942155.65 31.45 39.55 79.10 65.00 76.7084.81 34.80 57.48 96.28 74.13 78.75100.9492.91102.723955.43 36.03 45.68 83.78 67.83 90.4096.63 39.81 56.61 106.07 76.16 94.08115.62 108.61 118.712642.54 34.19 45.39 80.94 69.39 88.5893.88 38.32 58.72 101.69 80.18 92.38115.79 107.18 118.67AS21.28 46.65 53.36 105.74 58.79 83.5688.51 56.39 69.60 133.54 67.42 88.10108.57 112.56 119.22",
"figure_id": "tab_1",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "ARP of 8 multi-task models with diverse sharing patterns trained on Tiny-Taskonomy and attacked by PGD, AutoAttack, and Auto-SAGE variants with perturbation bound ϵ = 8. Keyp Edge Total GB-MTA Segm Norm Dept Keyp Edge Total GB-MTA Baseline GB-MTA IND 106.38 248.04 23.13 117.44 12.48 11.50 282.01 274.17 245.25 23.21 117.83 12.40 11.37 279.28 270.25 282.01 274.17 190 105.03 190.91 49.41 138.67 14.98 14.02 232.55 236.63 175.02 45.06 130.69 14.26 13.64 213.47 218.29 216.32 218.88 358 103.68 189.90 50.42 152.70 18.14 15.53 237.12 247.29 174.80 46.32 142.73 17.07 14.73 217.99 227.61 220.93 228.97 959 96.86 201.59 48.17 163.56 15.56 17.64 241.75 256.35 183.75 44.46 150.37 14.70 17.04 221.31 234.58 223.54 234.56 1020 83.53 222.92 49.98 139.97 18.25 18.01 259.27 263.03 202.50 44.34 130.95 17.02 16.89 235.80 240.26 238.95 241.14 1043 75.59 205.27 44.26 145.47 23.99 38.96 246.90 250.79 189.32 39.97 136.29 22.55 36.50 227.23 233.25 230.54 234.28 1037 62.48 216.15 45.51 152.47 17.61 47.35 252.55 261.82 197.99 41.05 141.10 16.88 40.50 230.86 240.70 240.00 246.43 AS 21.28 231.23 77.53 104.10 16.02 22.40 231.08 196.07 227.99 76.06 103.84 15.97 21.87 226.97 192.85 231.07 196.07",
"figure_data": "Model#ParamsPGDAutoAttackAuto-SAGEIndex(M)Segm Norm Dept",
"figure_id": "tab_2",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Table 12 reports the numerical results for Figure",
"figure_data": "",
"figure_id": "tab_4",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "ARP of all 25 multi-task models similar as Table8. The attack methods are APGD and Auto-SAGE variants.",
"figure_data": "Model Index#Params (M)APGD Segm Norm Dept Total SignTotal GB-MTA Baseline GB-MTA Auto-SAGEIND63.8327.16 15.28 48.07 43.27 46.6862.7250.6063.781463.6031.53 15.75 77.52 65.34 64.9486.4280.4988.43463.6028.65 41.38 51.30 54.93 67.8888.3966.4687.21963.6027.10 17.16 52.03 51.10 51.8469.2858.2270.202363.5931.31 21.70 74.76 59.86 63.4983.5378.7185.70562.4827.91 21.99 59.31 54.01 51.8770.4662.5971.731062.4835.73 17.45 91.60 57.20 65.4688.0481.7392.482862.4731.55 43.78 80.85 65.45 74.9598.4284.9997.513562.2532.18 37.56 72.78 57.23 65.3185.3774.9885.893862.2534.16 28.68 92.92 63.05 68.0594.1992.6999.994161.1335.40 40.66 93.92 62.34 73.4396.8086.64 100.731155.6636.47 16.01 87.30 53.81 61.4683.2680.8988.732155.6529.81 32.22 68.68 57.18 57.1576.4169.1878.343355.4335.83 35.63 90.57 54.05 68.3292.0883.9696.003655.4333.51 39.54 77.22 61.78 68.9788.8280.0689.483955.4331.24 29.83 78.41 58.07 58.4179.3878.9483.614454.3238.22 37.76 97.08 58.90 67.8494.0790.8698.394254.3234.23 45.68 84.02 55.78 68.2188.7580.9589.804847.5037.40 43.53 86.64 54.54 64.5485.4581.6189.602642.5432.45 41.99 71.78 62.53 65.4586.4177.9087.521742.5430.39 34.78 68.70 58.85 57.1876.2474.4279.053442.3238.77 25.92 79.02 49.26 59.2475.2970.5079.464341.2140.38 39.16 99.14 54.95 67.0890.6088.5097.064934.3941.65 51.01 103.47 55.19 70.2494.2591.9699.57AS21.2844.80 49.65 99.32 56.27 66.0584.4479.2791.01",
"figure_id": "tab_5",
"figure_label": "9",
"figure_type": "table"
},
{
"figure_caption": "ARP of all 15 multi-task models similar as Table10. The attack methods are APGD and Auto-SAGE variants.",
"figure_data": "Model Index #Params (M)APGD Segm Norm Dept Keyp Edge Total GB-MTA Baseline GB-MTA Auto-SAGEIND106.38150.34 16.32 93.49 5.40 4.33 175.00 175.98 178.51 179.32190105.0398.82 33.70 109.26 6.39 6.12 132.72 146.86 134.60 147.75348105.02107.45 34.00 112.95 9.31 5.93 142.54 154.78 144.55 155.85200104.8197.78 32.08 105.60 5.39 7.20 131.47 145.08 133.21 145.95352104.8095.66 32.25 109.25 7.75 5.62 129.44 145.90 131.41 147.07469104.5798.82 33.00 113.63 15.70 9.14 134.73 149.15 136.58 150.36358103.68100.97 34.09 115.14 7.88 5.74 138.28 153.40 140.26 154.84481103.46113.34 32.11 116.28 20.99 12.67 147.42 162.81 149.73 164.1819198.2194.75 37.13 113.48 5.59 5.13 127.89 141.74 129.68 142.5895996.86105.62 33.55 118.21 5.25 7.28 139.53 158.45 141.08 159.3995883.75108.44 33.77 122.67 5.97 5.76 143.50 163.92 145.27 164.72102083.53117.91 31.86 107.94 6.30 7.10 147.29 162.43 149.50 163.90104375.59112.46 29.12 109.43 8.45 27.97 145.09 159.55 147.15 160.82103762.48115.08 28.25 110.43 7.28 29.00 145.04 160.81 151.71 165.31AS21.28136.60 46.05 69.10 7.37 10.73 135.97 122.22 138.78 124.39",
"figure_id": "tab_6",
"figure_label": "11",
"figure_type": "table"
},
{
"figure_caption": "The numerical results for the attack transferability of six multi-task models with various levels of parameter sharing attacked by APGD SINGLE-X variants.",
"figure_data": "Single-Segm Single-Norm Single-DeptAS/5L0.460.530.154L0.260.450.153L0.220.330.122L0.160.280.091L0.120.310.08IND/0L0.020.170.1",
"figure_id": "tab_7",
"figure_label": "12",
"figure_type": "table"
},
{
"figure_caption": "The accuracy of six multi-task models with and without adversarial training on NYUv2. The adversarial samples are generated by five PGD-based adversarial attack methods with ϵ = 8. Mean Median 11.25 • 22.5 • 30 • Abs. Rel. 1.25 1.25 2 1.25 3",
"figure_data": "Semantic Seg.Surface Normal PredictionDepth EstimationAdv. TrainmIoU ↑Pixel Acc ↑ARP t1Error ↓θ, within ↑ARP t2Error ↓σ, within ↑ARP t3ARPw/o AT25.88 58.05-17.28 15.11 36.45 71.32 84.91-0.55 0.21 64.61 89.95 97.39--Single",
"figure_id": "tab_8",
"figure_label": "13",
"figure_type": "table"
},
{
"figure_caption": ". The underlying adversarial sample generation methods are changed to FGSM variants in adversarial training. Segm 22.99 55.72 7.59 18.11 15.49 36.36 68.22 81.60 3.15 0.59 0.25 61.23 88.08 96.51 5.88 5.54 Single-Norm 23.22 56.66 6.34 17.53 14.85 38.18 70.32 83.15 0.32 0.58 0.24 61.62 88.58 96.80 4.74 3.58 Single-Dept 23.41 55.53 6.94 17.80 15.12 36.90 69.86 82.77 1.27 0.58 0.24 62.16 88.70 96.77 4.17 4.13 Total 23.00 55.42 7.83 17.66 14.91 37.83 70.06 82.75 0.27 0.60 0.25 60.54 88.03 96.55 6.43 4.84 GB-MTA 23.36 56.12 6.53 18.03 15.27 37.07 68.78 81.66 2.22 0.59 0.25 60.57 88.04 96.54 6.30 5.02 ARP of multi-task models trained with and without FAT in Table 14 and attacked by multi-task attacks including FGSM variants of GB-MTA, SINGLE, and TOTAL with ϵ = 8.",
"figure_data": "Semantic Seg.Surface Normal PredictionDepth EstimationAdv. TrainmIoU ↑Pixel Acc ↑ARP t1Error ↓ Mean Median 11.25 • 22.5 • 30 • θ, within ↑ARP t2Error ↓ Abs. Rel. 1.25 1.25 2 1.25 3 σ, within ↑ARP t3ARPw/o AT25.88 58.05-17.28 15.11 36.45 71.32 84.91-0.55 0.21 64.61 89.95 97.39--Single-",
"figure_id": "tab_9",
"figure_label": "15",
"figure_type": "table"
}
] | Lijun Zhang; Xiao Liu; Kaleel Mahmood; Caiwen Ding; Hui Guan | [
{
"authors": "C Ahn; E Kim; S Oh",
"journal": "",
"ref_id": "b0",
"title": "Deep elastic networks with model selection for multi-task learning",
"year": "2019"
},
{
"authors": "M Andriushchenko; F Croce; N Flammarion; M Hein",
"journal": "Springer",
"ref_id": "b1",
"title": "Square attack: a query-efficient black-box adversarial attack via random search",
"year": "2020-08-23"
},
{
"authors": "E Arcari; M V Minniti; A Scampicchio; A Carron; F Farshidian; M Hutter; M N Zeilinger",
"journal": "IEEE Robotics and Automation Letters",
"ref_id": "b2",
"title": "Bayesian Multi-Task Learning MPC for Robotic Mobile Manipulation",
"year": "2023"
},
{
"authors": "A Athalye; N Carlini; D Wagner",
"journal": "",
"ref_id": "b3",
"title": "Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples",
"year": "2018"
},
{
"authors": " Pmlr",
"journal": "",
"ref_id": "b4",
"title": "",
"year": ""
},
{
"authors": "T Bai; J Luo; J Zhao; B Wen; Q Wang",
"journal": "",
"ref_id": "b5",
"title": "Recent advances in adversarial training for adversarial robustness",
"year": "2021"
},
{
"authors": "N Carlini; A Athalye; N Papernot; W Brendel; J Rauber; D Tsipras; I Goodfellow; A Madry; A Kurakin",
"journal": "",
"ref_id": "b6",
"title": "On evaluating adversarial robustness",
"year": "2019"
},
{
"authors": "N Carlini; D Wagner",
"journal": "",
"ref_id": "b7",
"title": "Towards evaluating the robustness of neural networks",
"year": "2017"
},
{
"authors": "L.-C Chen; G Papandreou; I Kokkinos; K Murphy; A L Yuille",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b8",
"title": "Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs",
"year": "2017"
},
{
"authors": "P.-Y Chen; H Zhang; Y Sharma; J Yi; C.-J Hsieh",
"journal": "",
"ref_id": "b9",
"title": "Zoo: Zeroth order optimization based black-box attacks to deep neural networks without training substitute models",
"year": "2017"
},
{
"authors": "F Croce; M Hein",
"journal": "",
"ref_id": "b10",
"title": "Minimally distorted adversarial examples with a fast adaptive boundary attack",
"year": "2020"
},
{
"authors": " Pmlr",
"journal": "",
"ref_id": "b11",
"title": "",
"year": ""
},
{
"authors": "F Croce; M Hein",
"journal": "",
"ref_id": "b12",
"title": "Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks",
"year": "2020"
},
{
"authors": " Pmlr",
"journal": "",
"ref_id": "b13",
"title": "",
"year": ""
},
{
"authors": "Y Dong; F Liao; T Pang; H Su; J Zhu; X Hu; J Li",
"journal": "",
"ref_id": "b14",
"title": "Boosting adversarial attacks with momentum",
"year": "2018"
},
{
"authors": "N Dvornik; K Shmelkov; J Mairal; C Schmid",
"journal": "",
"ref_id": "b15",
"title": "Blitznet: A real-time deep network for scene understanding",
"year": "2017"
},
{
"authors": "S Ghamizi; M Cordy; M Papadakis; Y Le Traon",
"journal": "",
"ref_id": "b16",
"title": "Adversarial robustness in multi-task learning: Promises and illusions",
"year": "2022"
},
{
"authors": "I J Goodfellow; J Shlens; C Szegedy",
"journal": "",
"ref_id": "b17",
"title": "Explaining and harnessing adversarial examples",
"year": "2014"
},
{
"authors": "P Guo; C.-Y Lee; D Ulbricht",
"journal": "",
"ref_id": "b18",
"title": "Learning to branch for multi-task learning",
"year": "2020"
},
{
"authors": " Pmlr",
"journal": "",
"ref_id": "b19",
"title": "",
"year": ""
},
{
"authors": "P Guo; Y Xu; B Lin; Y Zhang",
"journal": "",
"ref_id": "b20",
"title": "Multi-task adversarial attack",
"year": "2020"
},
{
"authors": "N K Gurulingan; E Arani; B Zonooz",
"journal": "",
"ref_id": "b21",
"title": "Uninet: A unified scene understanding network and exploring multi-task relationships through the lens of adversarial attacks",
"year": "2021"
},
{
"authors": "J Horáček; M Hladík; M Černỳ",
"journal": "Springer",
"ref_id": "b22",
"title": "Interval linear algebra and computational complexity",
"year": "2015"
},
{
"authors": "J Huang; R S Feris; Q Chen; S Yan",
"journal": "",
"ref_id": "b23",
"title": "Cross-domain image retrieval with a dual attribute-aware ranking network",
"year": "2015"
},
{
"authors": "B Jou; S.-F Chang",
"journal": "",
"ref_id": "b24",
"title": "Deep cross residual learning for multitask visual recognition",
"year": "2016"
},
{
"authors": "M Klingner; A Bar; T Fingscheidt",
"journal": "",
"ref_id": "b25",
"title": "Improved noise and attack robustness for semantic segmentation by using multi-task training with self-supervised depth estimation",
"year": "2020"
},
{
"authors": "I Kokkinos",
"journal": "",
"ref_id": "b26",
"title": "Ubernet: Training a universal convolutional neural network for low-, mid-, and high-level vision using diverse datasets and limited memory",
"year": "2017"
},
{
"authors": "V Kreinovich; A Lakeyev; S Noskov",
"journal": "Linear Algebra and its Applications",
"ref_id": "b27",
"title": "Approximate linear algebra is intractable",
"year": "1996"
},
{
"authors": "I Leang; G Sistu; F Bürger; A Bursuc; S Yogamani",
"journal": "IEEE",
"ref_id": "b28",
"title": "Dynamic task weighting methods for multi-task networks in autonomous driving systems",
"year": "2020"
},
{
"authors": "Y Liu; X Chen; C Liu; D Song",
"journal": "",
"ref_id": "b29",
"title": "Delving into transferable adversarial examples and black-box attacks",
"year": "2016"
},
{
"authors": "A Madry; A Makelov; L Schmidt; D Tsipras; A Vladu",
"journal": "",
"ref_id": "b30",
"title": "Towards deep learning models resistant to adversarial attacks",
"year": "2018"
},
{
"authors": "K Mahmood; D Gurevin; M Van Dijk; P H Nguyen",
"journal": "Entropy",
"ref_id": "b31",
"title": "Beware the black-box: On the robustness of recent defenses to adversarial examples",
"year": "2021"
},
{
"authors": "K Mahmood; R Mahmood; E Rathbun; M Van Dijk",
"journal": "IEEE Access",
"ref_id": "b32",
"title": "Back in Black: A Comparative Evaluation of Recent State-Of-The-Art Black-Box Attacks",
"year": "2021"
},
{
"authors": "K Mahmood; R Mahmood; M Van Dijk",
"journal": "",
"ref_id": "b33",
"title": "On the robustness of vision transformers to adversarial examples",
"year": "2021"
},
{
"authors": "K Mahmood; P H Nguyen; L M Nguyen; T Nguyen; M Van Dijk",
"journal": "IEEE Access",
"ref_id": "b34",
"title": "Besting the Black-Box: Barrier Zones for Adversarial Example Defense",
"year": "2022"
},
{
"authors": "C Mao; A Gupta; V Nitin; B Ray; S Song; J Yang; C Vondrick",
"journal": "Springer",
"ref_id": "b35",
"title": "Multitask learning strengthens adversarial robustness",
"year": "2020-08-23"
},
{
"authors": "A Navon; A Shamsian; I Achituve; H Maron; K Kawaguchi; G Chechik; E Fetaya",
"journal": "",
"ref_id": "b36",
"title": "Multi-task learning as a bargaining game",
"year": "2022"
},
{
"authors": "R Ranjan; V M Patel; R Chellappa",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b37",
"title": "Hyperface: A deep multi-task learning framework for face detection, landmark localization, pose estimation, and gender recognition",
"year": "2017"
},
{
"authors": "E Rathbun; K Mahmood; S Ahmad; C Ding; M Van Dijk",
"journal": "",
"ref_id": "b38",
"title": "Game Theoretic Mixed Experts for Combinational Adversarial Machine Learning",
"year": "2022"
},
{
"authors": "S Ruder",
"journal": "",
"ref_id": "b39",
"title": "An overview of multi-task learning in deep neural networks",
"year": "2017"
},
{
"authors": "M Sandler; A Howard; M Zhu; A Zhmoginov; L.-C Chen",
"journal": "",
"ref_id": "b40",
"title": "Mobilenetv2: Inverted residuals and linear bottlenecks",
"year": "2018"
},
{
"authors": "N Silberman; D Hoiem; P Kohli; R Fergus",
"journal": "Springer",
"ref_id": "b41",
"title": "Indoor segmentation and support inference from rgbd images",
"year": "2012"
},
{
"authors": "I Sobh; A Hamed; V R Kumar; S Yogamani",
"journal": "",
"ref_id": "b42",
"title": "Adversarial attacks on multi-task visual perception for autonomous driving",
"year": "2021"
},
{
"authors": "X Sun; R Panda; R Feris; K Saenko",
"journal": "",
"ref_id": "b43",
"title": "Adashare: Learning what to share for efficient deep multi-task learning",
"year": "2019"
},
{
"authors": "F Tramer; N Carlini; W Brendel; A Madry",
"journal": "Advances in neural information processing systems",
"ref_id": "b44",
"title": "On Adaptive Attacks to Adversarial Example Defenses",
"year": "2020"
},
{
"authors": "Y Wang; J Liu; X Chang; J Mišić; V B Mišić",
"journal": "International Journal of Intelligent Systems",
"ref_id": "b45",
"title": "IWA: integrated gradient-based white-box attacks for fooling deep neural networks",
"year": "2022"
},
{
"authors": "N Xu; K Mahmood; H Fang; E Rathbun; C Ding; W Wen",
"journal": "",
"ref_id": "b46",
"title": "Securing the spike: On the transferabilty and security of spiking neural networks to adversarial examples",
"year": "2022"
},
{
"authors": "L Yao; Z Chu; S Li; Y Li; J Gao; A Zhang",
"journal": "",
"ref_id": "b47",
"title": "A survey on causal inference",
"year": "2020"
},
{
"authors": "T Yu; S Kumar; A Gupta; S Levine; K Hausman; C Finn",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b48",
"title": "Gradient surgery for multi-task learning",
"year": "2020"
},
{
"authors": "A R Zamir; A Sax; W Shen; L J Guibas; J Malik; S Savarese",
"journal": "",
"ref_id": "b49",
"title": "Taskonomy: Disentangling task transfer learning",
"year": "2018"
},
{
"authors": "J Zhang; X Xu; B Han; G Niu; L Cui; M Sugiyama; M Kankanhalli",
"journal": "",
"ref_id": "b50",
"title": "Attacks which do not kill training make adversarial learning stronger",
"year": "2020"
},
{
"authors": " Pmlr",
"journal": "",
"ref_id": "b51",
"title": "",
"year": ""
},
{
"authors": "L Zhang; X Liu; H Guan",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b52",
"title": "AutoMTL: A Programming Framework for Automating Efficient Multi-Task Learning",
"year": "2022"
},
{
"authors": "L Zhang; X Liu; H Guan",
"journal": "",
"ref_id": "b53",
"title": "A Tree-Structured Multi-Task Model Recommender",
"year": "2022"
},
{
"authors": " Pmlr",
"journal": "",
"ref_id": "b54",
"title": "",
"year": ""
},
{
"authors": "M Zhou; J Wu; Y Liu; S Liu; C Zhu",
"journal": "",
"ref_id": "b55",
"title": "Dast: Data-free substitute training for adversarial attacks",
"year": "2020"
},
{
"authors": "",
"journal": "GB-MTA IND",
"ref_id": "b56",
"title": "Model Index #Params (M) FGSM PGD Segm Norm Dept Total SignTotal GB-MTA Segm Norm Dept Total SignTotal",
"year": ""
},
{
"authors": "",
"journal": "-MTA Baseline GB-MTA",
"ref_id": "b57",
"title": "51 Table 5: ARP of all 25 multi-task models similar as Table 4 for GB-MTA and the two baselines, AutoAttack and Auto-SAGE. Model Index #Params (M) AutoAttack Auto-SAGE Segm Norm Dept Total SignTotal GB",
"year": ""
},
{
"authors": "",
"journal": "Segm Norm Dept Keyp Edge Total GB-MTA",
"ref_id": "b58",
"title": "Table 6: ARP of all 15 multi-task models with diverse sharing patterns trained on Tiny-Taskonomy and attacked by FGSM and PGD variants with perturbation bound ϵ = 8. For brevity, the name of SINGLE-X variants are simplified to the task name only. IND: independent, AS: all-shared. Model Index #Params (M) FGSM PGD Segm Norm Dept Keyp Edge Total GB-MTA",
"year": ""
},
{
"authors": "",
"journal": "GB-MTA",
"ref_id": "b59",
"title": "Table 7: ARP of all 15 multi-task models similar as Table 6. The fundamental attack methods are AutoAttack and Auto-SAGE. Model Index #Params (M) AutoAttack Auto-SAGE Segm Norm Dept Keyp Edge Total GB-MTA Baseline",
"year": ""
},
{
"authors": "",
"journal": "-MTA Segm Norm Dept Total SignTotal GB-MTA",
"ref_id": "b60",
"title": "Table 8: ARP of all 25 multi-task models with diverse sharing patterns trained on NYUv2 and attacked by FGSM and PGD variants with perturbation bound ϵ = 4. For brevity, the name of SINGLE-X variants are simplified to the task name only. IND: independent, AS: all-shared. Model Index #Params (M) FGSM PGD Segm Norm Dept Total SignTotal GB",
"year": ""
},
{
"authors": "",
"journal": "",
"ref_id": "b61",
"title": "ARP of all 15 multi-task models with diverse sharing patterns trained on Tiny-Taskonomy and attacked by FGSM and PGD variants with perturbation bound ϵ = 4. For brevity, the name of SINGLE-X variants are simplified to the task name only. IND: independent, AS: all-shared. Model Index #Params (M) FGSM PGD",
"year": ""
},
{
"authors": " Adv",
"journal": "",
"ref_id": "b62",
"title": "Train Single-Segm Single-Norm Single-Dept Total GB-MTA",
"year": ""
},
{
"authors": "",
"journal": "",
"ref_id": "b63",
"title": "The model structures (layouts) of multi-task models for Tiny-Taskonomy with Deeplab-ResNet34. IND: independent, AS: all-shared. Model Index #Params (M) Layout",
"year": ""
}
] | [
{
"formula_coordinates": [
2,
96.44,
596.11,
196.73,
14.66
],
"formula_id": "formula_0",
"formula_text": "max δ L(x + δ, y; θ), s.t. ∥δ∥ p ≤ ϵ,(1)"
},
{
"formula_coordinates": [
2,
351.98,
494.21,
206.69,
27.21
],
"formula_id": "formula_1",
"formula_text": "x (i) adv = P S (x (i-1) adv + F δ (ϵ (i-1) , ∂L ∂x (i-1) adv )),(2)"
},
{
"formula_coordinates": [
3,
54,
263.08,
246.98,
61.93
],
"formula_id": "formula_2",
"formula_text": "x (i) adv = P S (x (i-1) adv + α(P S (x (i-1) adv + ϵ (i-1) sign( ∂L j x (i-1) adv )) -x (i-1) adv ) + (1 -α)(x (i-1) adv -x (i-2) adv )),"
},
{
"formula_coordinates": [
3,
68.84,
459.14,
224.32,
30.88
],
"formula_id": "formula_3",
"formula_text": "x (i) adv = P S (x (i-1) adv + ϵ (i-1) • sign( n j=1 ∂L j ∂x (i-1) adv )). (4)"
},
{
"formula_coordinates": [
3,
332.46,
577.02,
226.21,
31.84
],
"formula_id": "formula_4",
"formula_text": "max β L(x + η • β, y) s.t. ∥η • β∥ p ≤ ϵ, ∀β (k) ∈ β : β (k) ∈ {-1, 0, 1}. (5)"
},
{
"formula_coordinates": [
4,
98.88,
203.99,
194.28,
30.32
],
"formula_id": "formula_5",
"formula_text": "∆Acc = 1 N N i=1 Acc M,ti -Acc B,ti Acc B,ti ,(6)"
},
{
"formula_coordinates": [
4,
63.05,
366.12,
230.12,
51.28
],
"formula_id": "formula_6",
"formula_text": "β * = arg max β |∆Acc| = arg max β ∆L = arg max β 1 N n i=1 L i (x + η • β, y i ) -L i (x, y i ) L i (x, y i ) ,(7)"
},
{
"formula_coordinates": [
4,
83.43,
552.91,
209.74,
41.56
],
"formula_id": "formula_7",
"formula_text": "β * = arg max β i β • ∂L i (x, y i ) ∂x • 1 L i (x, y i ) s.t. ∀β (k) ∈ β : β (k) ∈ {-1, 0, 1}.(8)"
},
{
"formula_coordinates": [
4,
367.02,
95.96,
191.65,
30.32
],
"formula_id": "formula_8",
"formula_text": "∂∆L ∂β = n i=1 ∂L i (x, y i ) ∂x • 1 L i (x, y i ) .(9)"
},
{
"formula_coordinates": [
4,
415.38,
335.36,
92.17,
14.56
],
"formula_id": "formula_9",
"formula_text": "n i=1 ∂Li(x,yi) ∂x • 1 Li(x,yi) ."
},
{
"formula_coordinates": [
4,
319.5,
471.37,
238.5,
232.56
],
"formula_id": "formula_10",
"formula_text": "Output: xmax 1: li ← Li(x (0) , yi), ∀i = 1, • • • , n. 2: β * ← sign( n i ∂L i (x (0) ,y i ) ∂x (0) • 1 L i (x (0) ,y i ) ) 3: x (1) ← P (x (0) + η • β * ) 4: lmax ← max{ n i L i (x (0) )-l i l i , n i L i (x (1) )-l i l i } 5: if lmax ≡ n i L i (x (0) )-l i l i then 6: xmax ← x (0) 7: else 8: xmax ← x (1) 9: end if 10: for k = 1 to Niter -1 do 11: β * ← sign( n i ∂L i (x (k) ,y i ) ∂x (k) • 1 L i (x (k) ,y i ) ) 12: z k+1 ← P (x (k) + η • β * ) 13: x k+1 ← P (x (k) + α(z k+1 -x (k) ) + (1 -α)(x (k) - x (k-1) )) 14: if n i L i (x (k+1) )-l i l i > lmax then 15: xmax ← x (k+1) 16: lmax ← n i L i (x (k+1) )-l i l i 17: end if 18: if k ∈ W then 19:"
},
{
"formula_coordinates": [
5,
60.89,
391.93,
228.13,
30.43
],
"formula_id": "formula_11",
"formula_text": "1 N N i=1 1 M i Mi j=1 (-1) si,j (m ′ i,j -m i,j )/m i,j × 100%, (10"
},
{
"formula_coordinates": [
5,
289.02,
402.77,
4.15,
8.64
],
"formula_id": "formula_12",
"formula_text": ")"
},
{
"formula_coordinates": [
6,
135.54,
455.83,
157.63,
26.88
],
"formula_id": "formula_13",
"formula_text": "1 n -1 Y̸ =X ARP-Y ARP-X ,(11)"
},
{
"formula_coordinates": [
10,
326.06,
194.38,
232.61,
51.28
],
"formula_id": "formula_14",
"formula_text": "β * = arg max β ∆L = arg max β 1 N n i=1 L i (x + η • β, y i ) -L i (x, y i ) L i (x, y i ) .(12)"
},
{
"formula_coordinates": [
10,
332.2,
342.29,
226.47,
62.09
],
"formula_id": "formula_15",
"formula_text": "L i (x + η • β, y i ) -L i (x, y i ) = L i (x, y i ) + η • β • ∂L i (x, y i ) ∂x + ξ -L i (x, y i ) ≈ η • β • ∂L i (x, y i ) ∂x .(13)"
},
{
"formula_coordinates": [
10,
328.31,
473.32,
230.36,
45.23
],
"formula_id": "formula_16",
"formula_text": "β * = arg max β 1 N n i=1 η • β • ∂L i (x, y i ) ∂x • 1 L i (x, y i ) s.t. ∀β (k) ∈ β : β (k) ∈ {-1, 1},(14)"
},
{
"formula_coordinates": [
11,
53.64,
187.57,
200.32,
84.98
],
"formula_id": "formula_17",
"formula_text": "Algorithm 2 APGD Input: x (0) , L, η, α, N iter , attack checkpoints:W Output: x max 1: x (1) ← P (x (0) + η • sign(∇L(x (0) ))) 2: l max ← max{L(x (0) ), L(x (1) )} 3: if l max ≡ L(x (0) ) then 4:"
},
{
"formula_coordinates": [
11,
54.5,
284.14,
238,
92.52
],
"formula_id": "formula_18",
"formula_text": "x max ← x (1) 7: end if 8: for k = 1 to N iter -1 do 9: z k+1 ← P (x (k) + η • sign(∇L(x (k) ))) 10: x k+1 ← P (x (k) + α(z k+1 -x (k) ) + (1 -α)(x (k) - x (k-1) )) 11: if L(x (k+1) ) > l max then 12:"
},
{
"formula_coordinates": [
11,
53.64,
377.31,
206.3,
223.69
],
"formula_id": "formula_19",
"formula_text": "l max ← L(x (k+1) ) 14: end if 15: if k ∈ W then 16: if Condition 1 1 or Condition 2 2 then 17: η ← η/2 18: x (k+1) ← x max 19: end if 20: end if 21: end for Algorithm 3 GB-MTA-APGD Input: x (0) , {y i }, {L i }, η, α, N iter , W Output: x max 1: l i ← L i (x (0) , y i ), ∀i = 1, • • • , n. 2: β * ← sign( n i ∂Li(x (0) ,yi) ∂x (0) • 1 Li(x (0) ,yi) ) 3: x (1) ← P (x (0) + η • β * ) 4: l max ← max{ n i Li(x (0) )-li li , n i Li(x (1) )-li li } 5: if l max ≡ n i"
},
{
"formula_coordinates": [
11,
85.88,
624.47,
53.49,
11.23
],
"formula_id": "formula_20",
"formula_text": "x max ← x (1)"
},
{
"formula_coordinates": [
11,
320,
77.76,
238,
71.16
],
"formula_id": "formula_21",
"formula_text": "β * ← sign( n i ∂Li(x (k) ,yi) ∂x (k) • 1 Li(x (k) ,yi) ) 12: z k+1 ← P (x (k) + η • β * ) 13: x k+1 ← P (x (k) + α(z k+1 -x (k) ) + (1 -α)(x (k) - x (k-1) )) 14: if n i Li(x (k+1) )-li li"
},
{
"formula_coordinates": [
11,
320,
161.25,
149.09,
47.37
],
"formula_id": "formula_22",
"formula_text": "l max ← n i Li(x (k+1) )-li li 17: end if 18: if k ∈ W then 19:"
},
{
"formula_coordinates": [
11,
325.31,
320,
224.97,
49.24
],
"formula_id": "formula_23",
"formula_text": "G blend (x (i) adv ) = γG blend (x (i-1) adv ) + k∈D\\R α (i) k ϕ (i) k ⊙ ∂L k ∂x (i) adv + r∈R α (i) r ϕ (i) r ⊙ (Et∼T [ ∂Lr ∂t(x(i) adv )"
},
{
"formula_coordinates": [
11,
324.48,
418.63,
249.85,
49.24
],
"formula_id": "formula_24",
"formula_text": "G blend (x (i) adv ) = γG blend (x (i-1) adv ) + k∈D\\R α (i) k ϕ (i) k ⊙ ( ∂L k ∂x (i) adv • 1 L k ) + r∈R α (i) r ϕ (i) r ⊙ (Et∼T [ ∂Lr ∂t(x (i) adv ) • 1 Lr ])."
},
{
"formula_coordinates": [
12,
319.5,
477.55,
93.29,
9.68
],
"formula_id": "formula_25",
"formula_text": "L = [L 1 , L 2 , • • • , L B ],"
},
{
"formula_coordinates": [
12,
319.5,
486.97,
240.15,
21.49
],
"formula_id": "formula_26",
"formula_text": "i = [L 1 i , L 2 i , • • • ] are sub- sets of"
}
] |
|
10.1145/279943.279962 | 2023-10-20 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b8",
"b19",
"b28",
"b49",
"b48",
"b44",
"b19",
"b53",
"b42",
"b23",
"b50",
"b36",
"b1",
"b46",
"b10",
"b61",
"b2",
"b65",
"b60",
"b14",
"b7",
"b57",
"b25",
"b58",
"b64",
"b6",
"b37",
"b63",
"b57",
"b31"
],
"table_ref": [],
"text": "Large pre-trained language models (PLMs), such as BERT (Devlin et al., 2019), GPT-3 (Brown et al., 2020), play a crucial role in the development of natural language processing applications, where one prominent training regime is to fine-tune the large and expensive PLMs for the downstream tasks of interest (Jiao et al., 2020).\nMinimizing the model size and accelerating the model inference are desired for systems with limited computation resources, such as mobile (Liu et al., 2021) and edge (Tambe et al., 2021) devices. Therefore, maintaining the generalization ability of the reduced-sized model is crucial and feasible (Sun et al., 2019;Sanh et al., 2019;Jiao et al., 2020;Wang et al., 2020).\nSemi-supervised learning (SSL) emerges as a practical paradigm to improve model generalization by leveraging both limited labelled data and extensive unlabeled data (Rasmus et al., 2015;Lee et al., 2013;Tarvainen and Valpola, 2017;Miyato et al., 2019;Berthelot et al., 2019;Sohn et al., 2020;Fan et al., 2023;Zhang et al., 2021;Berthelot et al., 2022;Zheng et al., 2022;Yang et al., 2023). While promising, combining SSL with a reduced-size model derived from PLMs still necessitates a well-defined learning strategy to achieve improved downstream performances (Wang et al., 2022a). This necessity arises because these shallow networks typically have lower capacity, and the scarcity of labeled data further curtails the model's optimization abilities. Besides, a major hurdle is a lack of labelled data samples -a particular problem for text mining tasks because the labelling text is labour-intensive and error-prone (Gururangan et al., 2019;Chen et al., 2020;Xie et al., 2020;Lee et al., 2021;Xu et al., 2022;Zhao et al., 2023).\nThis paper thus targets using SSL to leverage distilled PLMs in a situation where only limited labelled data is available and fast model inference is needed on resource-constrained devices. To this end, we use the well-established teacher-student knowledge distillation technique to construct small student models from a teacher PLM and then finetune them in the downstream SSL tasks. We aim to improve the effectiveness of fine-tuning small student models for text-mining tasks with limited labelled samples.\nWe present DisCo, a novel co-training approach aimed at enhancing the SSL performances by using distilled small models and few labelled data. The student models in the DisCo acquire complemen-tary information from multiple views, thereby improving the generalization ability despite the small model size and limited labelled samples. we introduce two types of view diversities for co-training: i) model view diversity, which leverages diversified initializations for student models in the cohort, ii) data view diversity, which incorporates varied noisy samples for student models in the cohort. Specifically, the model view diversity is generated by different task-agnostic knowledge distillations from the teacher model. The data view diversity is achieved through various embedding-based data augmentations to the input instances.\nIntuitively, DisCo with the model view encourages the student models to learn from each other interactively and maintain reciprocal collaboration. The student cohort with the model views increases each participating model's posterior entropy (Chaudhari et al., 2017;Pereyra et al., 2017;Zhang et al., 2018), helping them to converge to a flatter minimum with better generalization. At the same time, DisCo with the data views regularizes student predictions to be invariant when applying noises to input examples. Doing so improves the models' robustness on diverse noisy samples generated from the same instance. This, in turn, helps the models to obtain missing inductive biases on learning behaviour, i.e., adding more inductive biases to the models can lessen their variance (Xie et al., 2020;Lovering et al., 2021).\nWe have implemented a working prototype of DisCo 1 and applied it to text classification and extractive summarization tasks. We show that by cotraining just two student models, DisCo can deliver faster inference while maintaining the performance level of the large PLM. Specifically, DisCo can produce a student model that is 7.6× smaller (4layer TinyBERT) with 4.8× faster inference time by achieving superior ROUGE performance in extractive summarization than the source teacher model (12-layer BERT). It also achieves a better or comparable text classification performance compared to the previous state-of-the-art (SOTA) SSL methods with 12-layer BERT while maintaining a lightweight architecture with only 6-layer Tiny-BERT. We also show that DisCo substantially outperforms other SSL baselines by delivering higher accuracy when using the same student models in model size.\n1 Code and data are available at: https://github.com/ LiteSSLHub/DisCo."
},
{
"figure_ref": [],
"heading": "Methodology",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [
"fig_0",
"fig_0",
"fig_0"
],
"heading": "Overview of DisCo",
"publication_ref": [],
"table_ref": [],
"text": "DisCo jointly trains distilled student cohorts to improve model effectiveness in a complementary way from diversified views. As a working example, we explain how to use a dual-student DisCo to train two kinds of student models (see Figure 1). Extension to more students is straightforward (see section 2.3). To this end, DisCo introduces two initialization views during the co-training process: (i) model views which are different student model variants distilled from the teacher model, and (ii) data views which are different data augmented instances produced by the training input.\nIn DisCo, two kinds of compressed students (represented by two different colours in Figure 1(a)) are generated by the same teacher. This process allows us to pre-encode the model view specifically for DisCo. Additionally, we duplicate copies of a single student model to receive supervised and unsupervised data individually. In the supervised learning phase, DisCo optimizes two students using labelled samples. In the unsupervised learning phase, each student model concurrently shares the parameters with its corresponding duplicate, which is trained by supervised learning. The subsequent consistency training loss then optimizes the students using unlabeled samples.\nFor an ablation comparison of DisCo, we introduce the variant of DisCo only equipped with the model view, shown in Figure 1 (b). In this variant, labelled and unlabeled data are duplicated and would be fed to the students directly. DisCo and its variant ensure reciprocal collaboration among the distilled students and can enhance the generalization ability of the student cohort by the consistency constraint. In this section, we introduce DisCo from two aspects: knowledge distillation and the co-training strategy."
},
{
"figure_ref": [],
"heading": "Student Model Generation",
"publication_ref": [
"b19",
"b33"
],
"table_ref": [],
"text": "Our current implementation uses knowledge distillation to generate small-sized models from a PLM. Like the task-agnostic distillation of Tiny-BERT2 (Jiao et al., 2020), we use the original BERT without fine-tuning as the teacher model to generate the student models (In most cases, two student models at least are generated in our implementation). The task-agnostic distillation method is convenient for using any teacher network directly.\nWe use a large-scale general-domain corpus of WikiText-103 3 released by Merity et al. (2017) as the training data of the distillation. The student mimics the teacher's behaviour through the representation distillation from BERT layers: (i) the output of the embedding layer, (ii) the hidden states, and (iii) attention matrices."
},
{
"figure_ref": [],
"heading": "Model View Encoding",
"publication_ref": [],
"table_ref": [],
"text": "To ensure the grouped students present a different view of the teacher, we distil different BERT layers from the same teacher. Model view encoding diversifies the individual student by leveraging different knowledge of the teacher. We propose two different strategies for the knowledge distillation process: (i) Separated-layer KD (SKD): the student learns from the alternate k-layer of the teacher. For instance, {3, 6, 9, 12} are the 4 alternate layers of BERT. (ii) Connected-layer KD (CKD): the student learns from the continuous K-layer of the teacher. For example, {1, 2, 3, 4} are the continuous 4 layers of BERT. In the case of dual-student DisCo, the two students with two kinds of knowledge distillation strategies are represented as S AK and S BK . The co-training framework will encourage the distinct individual model to teach each other in a complementary manner underlying model view initialization.\nWith consistency constraints, our co-training framework can obtain valid inductive biases on model views, enabling student peers to teach each other and to generalize unseen data. Apart from the model views, we also introduce data views produced by various data augmentations of inputs to expand the inductive biases.\n3 https://huggingface.co./datasets/wikitext"
},
{
"figure_ref": [],
"heading": "Data View Encoding",
"publication_ref": [
"b57",
"b56",
"b59",
"b13",
"b59",
"b13",
"b22",
"b24",
"b45",
"b16",
"b18",
"b24",
"b59"
],
"table_ref": [],
"text": "We use different data augmentation strategies at the token embedding layer to create different data views from the input samples. Our intuition is that advanced data augmentation can introduce extra inductive biases since they are based on random sampling at the token embedding layer with minimal semantic impact (Xie et al., 2020;Wu et al., 2020;Yan et al., 2021;Gao et al., 2021). Inspired by ConSERT (Yan et al., 2021) and Sim-CSE (Gao et al., 2021), we adopt convenient data augmentation methods: adversarial attack (Kurakin et al., 2017), token shuffling (Lee et al., 2020), cutoff (Shen et al., 2020) and dropout (Hinton et al., 2012), described as follows. Adversarial Attack (AD). We implement it with Smoothness-Inducing Adversarial Regularization (SIAR)4 (Jiang et al., 2020), which encourages the model's output not to change too much when a small perturbation is injected to the input. Token Shuffling (TS). This strategy is slightly similar to Lee et al. (2020) and Yan et al. (2021), and we implement it by passing the shuffled position IDs to the embedding layer while keeping the order of the token IDs unchanged. Cutoff (CO). This method randomly erases some tokens for token cutoff in the embedding matrix. Dropout (DO). As same as in BERT, this scheme randomly drops elements by a specific probability and sets their values to zero.\nDisCo incorporates two forms of data view during co-training: a HARD FORM and a SOFT FORM. Taking dual-student networks for example, we use two different data augmentation approaches, such as AD and DO, to implement the HARD FORM data view. Regarding the SOFT FORM data view, we apply the same data augmentation approach, including AD with two rounds of random initialization to ensure distinct views. In DisCo, each student obtains perturbation differences through the various combinations of the HARD FORM and SOFT FORM."
},
{
"figure_ref": [],
"heading": "Co-training Framework",
"publication_ref": [],
"table_ref": [],
"text": "Formally, we are provided with a semi-supervised dataset D, D = S ∪ U. S = {( x, ŷ)} is labelled data, where ( x, ŷ) will be used for two kinds of students identically. U = {x * } is unlabeled data, and two copies are made for two kinds of students identically. For X ∈ D, let ϕ A (X) and ϕ B (X) denote the two data views of data X. A pair of models (S AK = f A and S BK = f B ) are two distilled student models which we treat as the model view of dualstudent DisCo. Student f A only uses ϕ A (X), and Student f B uses ϕ B (X).\nBy training collaboratively with the cohort of students f A and f B , the co-training optimization objective allows them to share the complementary information, which improves the generalization ability of a network. Supervised Student Cohort Optimization. For supervised parts, we use the categorical Cross-Entropy (CE) loss function for optimizing student f A and student f B , respectively. They are trained with the labeled data ( x, ŷ) sampled from S.\nL s A = CE( f A (ϕ A ( x)), ŷ),\n(1)\nL s B = CE( f B (ϕ B ( x)), ŷ). (2\n)\nUnsupervised Student Cohort Optimization. In standard co-training, multiple classifiers are expected to provide consistent predictions on unlabeled data x * ∈ U.\nThe consistency cost of the unlabeled data x * is computed from the two student output logits: z A (ϕ A (x * )) and z B (ϕ B (x * )). We use the Mean Square Error (MSE) to encourage the two students to predict similarly:\nL u A,B = MSE(z A (ϕ A (x * )), z B (ϕ B (x * ))),(3)\nL u B,A = MSE(z B (ϕ B (x * )), z A (ϕ A (x * ))). (4\n)\nOverall Training Objective. Finally, we combine supervised cross-entropy loss with unsupervised consistency loss and train the model by minimizing the joint loss:\nL Θ = L s A + L s B + µ(t, n) • λ • (L u A,B + L u B,A ),(5)\nwhere µ(t, n) = min( t n , 1). It represents the rampup weight starting from zero, gradually increasing along with a linear curve during the initial n training steps. λ is the hyperparameter balancing supervised and unsupervised learning."
},
{
"figure_ref": [],
"heading": "Co-training of Multi-student Peers",
"publication_ref": [
"b19",
"b57",
"b28",
"b34",
"b9"
],
"table_ref": [],
"text": "So far, our discussion has been focused on training two students. DisCo can be naturally extended to support not only two students in the student cohort but more student networks. Given K networks Θ 1 , Θ 2 , ..., Θ K (K ≥ 2), the objective function for optimising all Θ k , (1 ≤ k ≤ K), becomes: 7) Equation ( 5), is now a particular case of (6) with k = 2. With more than two networks in the cohort, a learning strategy for each student of DisCo takes the ensemble of other K -1 student peers to provide mimicry targets. Namely, each student learns from all other students in the cohort individually. Competing Baselines. For text classification tasks, we compare DisCo with: (i) supervised baselines, BERT BASE and default TinyBERT (Jiao et al., 2020), (ii) semi-supervised UDA (Xie et al., 2020) and FLiText (Liu et al., 2021). We also compare with other prominent SSL text classification methods and report their results on the Unified SSL Benchmark (USB) (Wang et al., 2022a and Tarau, 2004) and LexRank (Erkan and Radev, 2004). We use the open-source releases of the competing baselines.\nL Θ = K k=1 L s k + µ(t, n) • λ • L u i,k ,(6)\nL u i,k = 1 K -1 K i=1,i k MSE(z i (ϕ i (x * )), z k (ϕ k (x * )). (\n4 Experimental Results"
},
{
"figure_ref": [],
"heading": "Evaluation on Text Classification",
"publication_ref": [],
"table_ref": [
"tab_1"
],
"text": "As shown in Table 2, the two students produced by DisCo with a 6-layer distilled BERT (S A6 and S B6 ) consistently outperform TinyBERT and UDA TinyBERT in all text classification tasks. Moreover, one student of our dual-student 6-layer DisCo "
},
{
"figure_ref": [],
"heading": "Agnews",
"publication_ref": [
"b42",
"b23",
"b50",
"b58",
"b7"
],
"table_ref": [],
"text": "-model (Rasmus et al., 2015) 12 50 86.56 P-Labeling (Lee et al., 2013) 50 87.01 MeanTeacher (Tarvainen and Valpola, 2017) 50 86.77 PCM (Xu et al., 2022) 30 88.42 MixText (Chen et al., 2020) 30 87.40 DisCo (ours) 6 30 86.93"
},
{
"figure_ref": [],
"heading": "Yahoo!Answer",
"publication_ref": [
"b2",
"b10",
"b65",
"b61",
"b36",
"b50",
"b58",
"b7",
"b36"
],
"table_ref": [
"tab_3"
],
"text": "AdaMatch (Berthelot et al., 2022) 12 200 69.18 CRMactch (Fan et al., 2023) 200 69.38 SimMatch (Zheng et al., 2022) 200 69.36 FlexMatch (Zhang et al., 2021) 200 68.58 VAT (Miyato et al., 2019) 200 68.47 MeanTeacher (Tarvainen and Valpola, 2017) 200 66.57 DisCo (ours) 6 200 69.75\nDBpedia PCM (Xu et al., 2022) 12 10 98.70 Mixtext (Chen et al., 2020) 10 98.39 VAT (Miyato et al., 2019) 10 98.40 DisCo (ours) 6 10 98.57\noutperforms the 12-layer supervised BERT BASE by a 0.55% average improvement in accuracy. These results suggest that DisCo provides a simple but effective way to improve the generalization ability of small networks by training collaboratively with a cohort of other networks.\nCompared with the FLiText, DisCo improves the average classification accuracy by 1.9% while using a student model with 0.7M fewer parameters than FLiText. FLiText relies heavily on backtranslation models for generating augmented data, similar to UDA. Unfortunately, this strategy fails to eliminate error propagation introduced by the back-translation model and requires additional data pre-processing. Besides, FLiText consists of two training stages and needs supervised optimization in both stages, increasing training costs and external supervised settings.\nTable 3 shows results when comparing DisCo to other prominent SSL methods which are integrated with a 12-layer BERT. We take the results from the source publication or Unified SSL Benchmark (USB) (Wang et al., 2022a) for these baselines. However, most of them perform worse than DisCo's students only with a 6-layer BERT using same labeled data. In the case of Yahoo!Answer text classification, our 6-layer BERT- "
},
{
"figure_ref": [],
"heading": "Model Efficiency",
"publication_ref": [],
"table_ref": [
"tab_5",
"tab_1"
],
"text": "As shown in Table 5, compared with the teacher BERT BASE , all 4-layer student models give faster inference time by speeding up the inference by 4.80×-7.52× for the two tasks. FLiText is slightly faster than the smaller model generated DisCo. This is because FLiText uses a convolutional network while our student models use BERT with multi-head self-attention. The lower computational complexity of convolutional networks5 . However, despite the FLiText having more parameters, it gives worse performance (about 3.04% accuracy defects on average), as shown in Table 2."
},
{
"figure_ref": [],
"heading": "Ablation Studies",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Effect of using Multi-student Peers",
"publication_ref": [],
"table_ref": [
"tab_6",
"tab_6"
],
"text": "Having examined the dual-student DisCo in prior experiments, our next focus is to explore the scalability of DisCo by introducing more students in the cohort. As the results are shown in Table 6, we can see that the performance of every single student improves with an extension to four students in the DisCo cohort, which demonstrates that the generalization ability of students is enhanced when they learn together with increasing numbers of peers.\nBesides, the results in Table 6 have validated the necessity of co-training with multiple students. It is evident that a greater number of student peers (multi-students) in the co-training process yields a considerable performance enhancement compared to a less populous student group (dual-students)."
},
{
"figure_ref": [],
"heading": "Effect of using Multi-View Strategy",
"publication_ref": [
"b26"
],
"table_ref": [],
"text": "As shown in Further, we plot the training loss contour of DisCo and its ablation model in Figure 2. Both models have a fair benign landscape dominated by a region with convex contours in the centre and no dramatic non-convexity. We observe that the optima obtained by training with the model view and the data view are flatter than those obtained only with a model view. A flat landscape implies that the small perturbations of the model parameters cannot hurt the final performance seriously, while a chaotic landscape is more sensitive to subtle changes (Li et al., 2018). "
},
{
"figure_ref": [],
"heading": "UDA/FLiText with AD Augmentation",
"publication_ref": [
"b57"
],
"table_ref": [
"tab_1",
"tab_10"
],
"text": "In the preceding analysis detailed in Table 2, UDA/FLiText utilized back translation as their data augmentation strategy, a technique distinctly different from the token embedding level data augmentation employed in our DisCo framework. To ensure a balanced comparison, we substituted the back translation approach with our AD augmentation method for UDA/FLiText. The outcomes of this modification are portrayed in Table 9. These results underscore that regardless of the data augmentation strategy implemented, the performance of both UDA and FLiText falls short compared to our DisCo framework. This substantiates our claim that our co-training framework is superior in distilling knowledge encapsulated in unsupervised data. Furthermore, the performance across most tasks experiences a decline after the augmentation technique alteration. As stipulated in (Xie et al., 2020), the UDA/FLiText framework necessitates that augmented data maintain 'similar semantic meanings' thereby making back-translation a more suitable for UDA/FLiText, compared to the AD augmentation we incorporated."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "In this paper, we present DisCo, a framework of co-training distilled students with limited labelled data, which is used for targeting the lightweight models for semi-supervised text mining. DisCo leverages model views and data views to improve the model's effectiveness. We evaluate DisCo by applying it to text classification and extractive summarization tasks and comparing it with a diverse set of baselines. Experimental results show that DisCo substantially achieves better performance across scenarios using lightweight SSL models."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [
"b21",
"b39",
"b4",
"b41",
"b35",
"b47"
],
"table_ref": [],
"text": "Naturally, there is room for further work and improvement, and we discuss a few points here. In this paper, we apply DisCo to BERT-based student models created from the BERT-based teacher model. It would be useful to evaluate if our approach can generalize to other model architectures like TextCNN (Kim, 2014) and MLP-Mixer (Tolstikhin et al., 2021). It would also be interesting to extend our work to utilize the inherent knowledge of other language models (e.g., RoBERTa (Liu et al., 2019), GPT (Radford et al., 2018;Radford et al.;Brown et al., 2020), T5 (Raffel et al., 2020)).\nAnother limitation of our framework settings is the uniform number of BERT layers in all distilled student models. To address this, students in DisCo can be enhanced by introducing architectural diversity, such as varying the number of layers. Previous studies (Mirzadeh et al., 2020;Son et al., 2021) have demonstrated that a larger-size student, acting as an assistant network, can effectively simulate the teacher and narrow the gap between the student and the teacher. We acknowledge these limitations and plan to address them in future work."
},
{
"figure_ref": [],
"heading": "Ethical Statement",
"publication_ref": [],
"table_ref": [],
"text": "The authors declare that we have no conflicts of interest. Informed consent is obtained from all individual participants involved in the study. This article does not contain any studies involving human participants performed by any authors."
},
{
"figure_ref": [],
"heading": "A Appendix",
"publication_ref": [
"b17",
"b11",
"b48",
"b44",
"b53",
"b19",
"b52",
"b23",
"b50",
"b36",
"b1",
"b46",
"b61",
"b2",
"b65",
"b43",
"b14",
"b7",
"b57"
],
"table_ref": [],
"text": "A.1 Background and Related Work Knowledge Distillation (KD). The KD (Hinton et al., 2015) is one of the promising ways to transfer from a powerful large network or ensemble to a small network to meet the low-memory or fast execution requirements. BANs (Furlanello et al., 2018) sequentially distill the teacher model into multiple generations of student models with identical architecture to achieve better performance. BERT-PKD (Sun et al., 2019) distills patiently from multiple intermediate layers of the teacher model at the fine-tuning stage. DistilBERT (Sanh et al., 2019) and MiniLM (Wang et al., 2020) leverage knowledge distillation during the pre-training stage. TinyBERT (Jiao et al., 2020) sets a twostage knowledge distillation procedure that contains general-domain and tasks-specific distillation in Transformer (Vaswani et al., 2017). Despite their success, they may encounter difficulties affecting the sub-optimal performance in language understanding tasks due to the trade-off between model compression and performance loss.\nSemi-supervised Learning (SSL). The majority of SSL algorithms are primarily concentrated in the field of computer vision, including Pseudo Labeling (Lee et al., 2013), Mean Teacher (Tarvainen and Valpola, 2017), VAT (Miyato et al., 2019), Mix-Match (Berthelot et al., 2019), FixMatch (Sohn et al., 2020), CRMatch (Fan et al., 2023), Flex-Match (Zhang et al., 2021), AdaMatch (Berthelot et al., 2022), and SimMatch (Zheng et al., 2022), all of which exploit unlabeled data by encouraging invariant predictions to input perturbations (Sajjadi et al., 2016). The success of semi-supervised learning methods in the visual area motivates research in the NLP community. Typical techniques include VAMPIRE (Gururangan et al., 2019), Mix-Text (Chen et al., 2020) and UDA (Xie et al., 2020). Under the low-density separation assumption, these SSL methods perform better than their fully-supervised counterparts while using only a fraction of labelled samples."
},
{
"figure_ref": [],
"heading": "Co-Training.",
"publication_ref": [
"b3",
"b38"
],
"table_ref": [],
"text": "It is a classic award-winning method for semi-supervised learning paradigm, training two (or more) deep neural networks on complementary views (i.e., data view from different sources that describe the same instances) (Blum and Mitchell, 1998). By minimizing the error on limited labelled examples and maximizing the agreement on sufficient unlabeled examples, the co-training framework finally achieves two accurate classifiers on each view in a semi-supervised manner (Qiao et al., 2018)."
},
{
"figure_ref": [],
"heading": "A.2 Hyperparameters",
"publication_ref": [
"b59",
"b45"
],
"table_ref": [],
"text": "The BERT BASE , as the teacher model, has a total of 109M parameters (the number of layers N = 12, the hidden size d = 768, the forward size d ′ = 3072 and the head number h = 12). We used the BERT tokenizer6 to tokenize the text. The source text's max sentence length is 512 for extractive summarization and 256 for text classification. For extractive summarization, we select the top 3 sentences according to the average length of the Oracle human-written summaries. We use the default dropout settings in our distilled BERT architecture. The ratio of token cutoff is set to 0.2, as suggested in (Yan et al., 2021;Shen et al., 2020). The ratio of dropout is set to 0.1. Adam optimizer with β 1 = 0.9, β 2 = 0.999 is used for fine-tuning. We set the learning rate 1e-4 for extractive summarization and 5e-3 for text classification, in which the learning rate warm-up is 20% of the total steps. The λ for balancing supervised and unsupervised learning is set to 1 in all our experiments. The supervised batch size is set to 4, and the unsupervised batch size is 32 for the summarization task (16 for the classification task) in our experiments."
},
{
"figure_ref": [],
"heading": "A.3 Evaluation Methodology",
"publication_ref": [
"b27"
],
"table_ref": [],
"text": "Extractive summarization quality is evaluated with ROUGE (Lin and Hovy, 2003). We report the full-length F1-based ROUGE-1, ROUGE-2, and ROUGE-L (R-1, R-2, and R-L), and these ROUGE scores are computed using ROUGE-1.5.5.pl script7 . We report the accuracy (denoted as Acc) results in the text classification tasks."
},
{
"figure_ref": [],
"heading": "A.4 Baselines Details",
"publication_ref": [
"b19",
"b50",
"b36",
"b46",
"b2",
"b65",
"b58"
],
"table_ref": [],
"text": "For the text classification task, TinyBERT (Jiao et al., 2020) trains a large inspirer model (BERT) and then optimizes a target network (TextCNN).\nOther SSL algorithms integrated with BERT are implemented in a unified semi-supervised learning benchmark (USB) (Wang et al., 2022a) for classification, including Mean Teacher (Tarvainen and Valpola, 2017), VAT (Miyato et al., 2019), Fix-Match (Sohn et al., 2020), CRMatch (Fan et al., 2023), AdaMatch (Berthelot et al., 2022), and SimMatch (Zheng et al., 2022), all utilizing unlabeled data for invariant predictions. We report their text classification results in the USB benchmark testing. PCM (Xu et al., 2022) For extractive summarization, we extend Tiny-BERT and UDA for classifying every sentence, termed as UDASUM and TinyBERTSUM. Specifically, multiple [CLS] symbols are inserted in front of every sentence to represent each sentence and use their last hidden states to classify whether the sentence belongs to the summary. The SOTA semi-supervised extractive summarization model, CPSUM (Wang et al., 2022b), combines the noise-injected consistency training and the entropyconstrained pseudo labelling with the BERT BASE encoder. We also integrate the encoder of CPSUM with a slighter TinyBERT. It should be noted that the ORACLE system is an upper bound of the extractive summarization."
},
{
"figure_ref": [],
"heading": "A.5 Performance under Few-labels Settings",
"publication_ref": [],
"table_ref": [
"tab_1",
"tab_12",
"tab_12"
],
"text": "The form using differently labelled data in Table 2 indicates that there is a large performance gap between the 12-layer models and 4-layer models with only 10 labelled data due to the dramatic reduction in model size.\nHowever, as shown in Table 11, in the extractive summarization tasks, DisCo works particularly well than the 12-layer models in the scenario of 100 labelled examples. The extractive summarization task is to classify every single sentence within a document, and the two views effectively encourage invariant prediction for unlabeled points' perturbations. DisCo achieves superior performance, as shown in Table 11, whether it uses only 10 or 1000 labelled data in extractive summarization. The superiority of DisCo with 4-layer BERT is more evident when processing 10 labelled extractive summarization, compared to CPSUM and UDASUM with 12-layer BERT. The results also indicate that our method can be suitable for extreme cases that suffer from severe data scarcity problems."
},
{
"figure_ref": [],
"heading": "A.6 More Different Model-view Analysis",
"publication_ref": [
"b48"
],
"table_ref": [
"tab_11"
],
"text": "We further investigate the performance of different model view combinations in dual-students DisCo. As described in section 2.2.1, the model view encoding has two forms: Separated-layer KD (SKD) and Connected-layer KD (CKD text classification tasks are summarized in Table 10. Although all three combinations of model views achieve improvement (compared to results in Table 2), the combinations of CKD and SKD for two students perform slightly better than other combinations. According to Sun et al. (2019), distilling across alternate k layers in knowledge distillation captures more diverse representations, while distilling along connected k layers tends to capture relatively homogeneous representations. By combining these two distinct strategies of model view encoding, DisCo acquires additional inductive bias for each student in the cohort, resulting in improved performance on downstream tasks."
},
{
"figure_ref": [
"fig_3"
],
"heading": "A.7 More Different Data-view Analysis",
"publication_ref": [],
"table_ref": [],
"text": "In Figure 3, we visualize the effect of DisCo (4layer) integrating different data views encoding methods in the summarization task. We find that: DisCo integrating with the adversarial attack (AD) obtains superior performances, especially when data view is the adversarial attack in a SOFT FORM (AD A , AD B ). DisCo with HARD FORM data views like (AD A , DO B ) or (DO A , AD B ) get sub-optimal effectiveness. This suggests that more advanced data augmentation methods pave the way for a more refined data view."
},
{
"figure_ref": [],
"heading": "A.8 Model Ensembling for Multiple Students",
"publication_ref": [],
"table_ref": [
"tab_14",
"tab_1"
],
"text": "Model ensembling is an effective strategy, often yielding superior performance compared to individual models. As shown in Table 12, using simple model averaging for the 4-layer student model from Table 2 resulted in enhanced performance. However, the core focus of our research is to ascertain the potential of a single model within our framework. Training requires two or more student models, but only one is essential for inference. Having multiple students during training ensures performance comparable to the teacher model, while selecting one student for inference upholds computational efficiency. Diving deeper into ensemble techniques to further amplify performance wasn't our primary objective. "
},
{
"figure_ref": [],
"heading": "A.9 Selection of MSE or KL Loss",
"publication_ref": [
"b0",
"b20"
],
"table_ref": [
"tab_15"
],
"text": "In our framework, we use the MSE loss to align the logits of the students. However, besides using MSE loss, employing Kullback-Leibler (KL) divergence to maintain consistency between the student predictions is also a widely chosen approach. We prefer the MSE loss in our framework because the student can learn better without suffering from the information loss that occurs when passing through logits to probability space (Ba and Caruana, 2014).\nAs shown in Table 13, the 4-layer DisCo with MSE loss performs better in the majority of cases. However, when labeled data is extremely limited (e.g., 10 per class), KL divergence may surpass MSE in performance. This can be attributed to the noisy predictions produced by the student model, as its performance is not optimal because of the limited labeled data. KL divergence enforces label matching, thereby reducing issues resulting from corrupted knowledge transferred from another student model (Kim et al., 2021)."
},
{
"figure_ref": [],
"heading": "A.10 Details in Loss Landscape Visualization",
"publication_ref": [
"b26"
],
"table_ref": [],
"text": "Our loss visualization approach adheres to the 'filter normalization' method (Li et al., 2018). For each setting, we select the top-performing student checkpoint based on its validation set results. Subsequently, we generate two random vectors and normalize them using parameters specific to each model. Ultimately, using the same training data and augmentation techniques, we plot the training loss landscape following the two normalized directions."
},
{
"figure_ref": [],
"heading": "Acknowledgements",
"publication_ref": [],
"table_ref": [],
"text": "This work is supported in part by the National Natural Science Foundation of China (No.U20B2053)."
}
] | Many text mining models are constructed by fine-tuning a large deep pre-trained language model (PLM) in downstream tasks. However, a significant challenge nowadays is maintaining performance when we use a lightweight model with limited labelled samples. We present DisCo, a semi-supervised learning (SSL) framework for fine-tuning a cohort of small student models generated from a large PLM using knowledge distillation. Our key insight is to share complementary knowledge among distilled student cohorts to promote their SSL effectiveness. DisCo employs a novel co-training technique to optimize a cohort of multiple small student models by promoting knowledge sharing among students under diversified views: model views produced by different distillation strategies and data views produced by various input augmentations. We evaluate DisCo on both semi-supervised text classification and extractive summarization tasks. Experimental results show that DisCo can produce student models that are 7.6× smaller and 4.8× faster in inference than the baseline PLMs while maintaining comparable performance. We also show that DisCo-generated student models outperform the similar-sized models elaborately tuned in distinct tasks. | DisCo: Distilled Student Models Co-training for Semi-supervised Text Mining | [
{
"figure_caption": "Figure 1 :1Figure 1: The training architecture of DisCo (a) and the ablation variant (b). refers to 'DO USE' the and is 'DO NOT USE' . L s is a supervised loss and L u is unsupervised. 'KD' is an abbreviation for knowledge distillation.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2: 2D visualization of the loss surface contour of DisCo (w. model view and w. data view) and its ablation variant (w. model view). Subfigures (a) and (b) are the text classification tasks for Agnews dataset with 10 labeled data per class. Subfigures (c) and (d) are the extractive summarization tasks with 100 labeled data.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "is a complex multi-submodule combination SSL model with three components, a K-way classifier, the class semantic representation, and a class-sentence matching classifier. MixText (Chen et al., 2020) is a regularization-based SSL model with an interpolation-based augmentation technique. Both PCM and MixText use a 12-layer BERT as the backbone model.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: The performance visualization of the dualstudent DisCo with data view using different combinations of data augmentation strategies. The row indicates the 1st data-augmentation-based data view encoding strategy, while the column indicates the 2nd dataaugmentation-based data view encoding strategy. The results of dual-students DisCo with 4-layer TinyBERT being students are evaluated on the CNN/DailyMail with 100 labelled data.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Text classification performance (Acc (%)) on typical semi-supervised text classification tasks. P is the number of model parameters. The best results are in-bold.",
"figure_data": "ModelsPAgnewsYahoo!AnswerDBpediaAvg103020010302001030200BERT BASE109.48M81.0084.3287.2460.1064.1369.2896.5998.2198.7982.18UDA109.48M84.7086.8988.5664.2867.7069.7198.1398.6798.8584.17TinyBERT 666.96M71.4582.4687.5952.8460.5968.7196.8998.1698.6579.70UDA TinyBERT 666.96M73.9085.1687.5457.1462.8667.9397.4197.8798.2681.79DisCo (S A6 )66.96M74.3886.3988.7057.6264.0469.5798.5098.4598.5782.02DisCo (S B6 )66.96M77.4586.9388.8259.1066.5869.7598.5798.6198.7382.73TinyBERT 414.35M69.6778.3585.1242.6653.6361.8989.6596.8897.5875.05UDA TinyBERT 4 DisCo (S A4 )14.35M 14.35M69.60 76.9077.56 85.3983.60 87.8240.69 51.4855.43 62.3663.34 68.1088.50 94.0293.63 98.1395.98 98.5674.26 80.31DisCo (S B4 )14.35M77.3685.5587.9551.3162.9368.2494.7998.1498.6380.54FLiText9.60M67.1477.1282.1248.3057.0163.0989.2694.0497.0175.01DisCo (S A2 )8.90M70.6181.8786.0848.4157.8464.0489.6796.0697.5876.90DisCo (S B2 )8.90M75.0582.1686.3851.0558.8365.6389.5596.1497.7078.053 Experiments",
"figure_id": "tab_1",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Text classification performance (Acc (%)) of other prominent SSL text classification models and all results reported by the Unified SSL Benchmark (USB)(Wang et al., 2022a). D refers to datasets, L m is the number of the BERT layers used by models and L d is labeled data per class.",
"figure_data": "D ModelsL m L d Acc",
"figure_id": "tab_3",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "ROUGE F1 performance of the extractive summarization. L d =100 refers to the labeled data per class. SSL baselines (CPSUM and UDASUM) use the same unlabeled data as DisCo has used.",
"figure_data": "ModelsPL dCNN/DailyMail R-1 R-2 R-LORACLE100 48.35 26.28 44.61LEAD-3100 40.04 17.21 36.14TextRank100 33.84 13.11 23.98LexRank100 34.63 12.72 21.25BERTSUM109.48M 100 38.58 15.97 34.79CPSUM109.48M 100 38.10 15.90 34.39UDASUM109.48M 100 38.58 15.87 34.78TinyBERTSUM 414.35M 100 39.83 17.24 35.98TinyBERTSUM F414.35M 100 40.06 17.32 36.18TinyBERTSUM L414.35M 100 39.88 17.14 36.00UDASUM TinyBERT 4 14.35M 100 40.11 17.43 36.23UDASUM TinyBERT A4 14.35M 100 39.90 17.25 36.05UDASUM TinyBERT B4 14.35M 100 40.11 17.34 36.19DisCo (S A4 )14.35M 100 40.39 17.57 36.47DisCo (S B4 )14.35M 100 40.41 17.59 36.50",
"figure_id": "tab_4",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Model efficiency about the model size and inference speedup on a single NVIDIA Tesla V100 32GB GPU. T TS (ms) refers to the speedup of extractive summarization models trained with 100 labeled data. T TC (ms) illustrates the speedup of text classification models trained with Agnews 200 labeled data per class.",
"figure_data": "ModelsT TS (ms) ModelsT TC (ms)BERTSUM12.66 BERT BASE12.94CPSUM12.66 TinyBERT 42.86TinyBERTSUM 42.64 UDA TinyBERT 42.86UDASUM TinyBERT 42.64 FLiText1.56DisCo (S A4 or S B4 )2.64 DisCo (S A2 or S B2 )1.72based DisCo achieves better performance than all12-layer BERT-based SSL benchmarks. These re-sults demonstrate that our model has superiority incertain scenarios of the lightweight model architec-ture and limited manual annotation.4.2 Evaluation on Extractive SummarizationFor the semi-supervised extractive summarizationtasks, our dual-student DisCo outperforms all base-lines in Table 4. Despite using a smaller-sized,4-layer model, DisCo performs better than the 12-",
"figure_id": "tab_5",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "Text classification performance (Acc (%)) of DisCo with multiple student peers. The students (S A2 , S B2 , S C2 , S D2 ) are distilled from layers {1, 2}, {3, 4}, {9, 10} and {11, 12} of the teacher BERT BASE , respectively. The first four students adopt HARD FORM data views which are AD, DO, TS, and CO, respectively. The last four students adopt a SOFT FORM data view with different DO initialization. Better results than dual-student DisCo in Table 2 is in-bold.",
"figure_data": "ModelsL dAgnews Yahoo!Answer DBpediaDisCo (S A2 ) 20087.5866.7498.23DisCo (S B2 ) 20087.4166.2898.33DisCo (S C2 ) 20087.8365.6397.69DisCo (S D2 ) 20087.5965.8798.34DisCo (S A2 ) 20086.9965.7198.10DisCo (S B2 ) 20086.7164.0198.18DisCo (S C2 ) 20086.7963.9698.12DisCo (S D2 ) 20086.6363.8398.01",
"figure_id": "tab_6",
"figure_label": "6",
"figure_type": "table"
},
{
"figure_caption": "Performance comparison between DisCo and a single student model with AD augmentation. The 'Sin-gleStudent' is the better-performing model among the two students within the DisCo framework.",
"figure_data": "ModelsL d Agnews Yahoo!Answer DBpediaSingleStudent 6 1073.5255.4393.65DisCo (S A6 )1074.3857.6298.50DisCo (S B6 )1077.4559.1098.57SingleStudent 4 1075.4947.5789.30DisCo (S A4 )1076.9051.4894.02DisCo (S B4 )1077.3651.3194.79SingleStudent 2 1068.7948.8777.26DisCo (S A2 )1070.6148.4189.67DisCo (S B2 )1075.0551.0589.55",
"figure_id": "tab_7",
"figure_label": "7",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "tab_8",
"figure_label": "8",
"figure_type": "table"
},
{
"figure_caption": "The impact of incorporating multi-view encoding for the dual-student DisCo. The HARD data-view is created using dropout (DO) and adversarial attack (AD). The SOFT view employs adversarial attack (AD) with varying initialization. The model-view ( ) refers to that students are trained from scratch without any teacher knowledge.",
"figure_data": "CNN/DailyMail w. 100Agnews w. 10Yahoo!Answer w. 10model viewdata viewS A4 , R-1 S B4 , R-1 S A4 , R-2 S B4 , R-2 S A4 , R-L S A4 , R-L S A4 , ACC S B4 , ACC S A4 , ACC S B4 , ACC36.7436.6914.1514.1232.9132.8637.5136.7622.2321.62S A4 /S B439.9639.9317.2317.2336.0736.0673.1873.6252.5652.95S A4 /S A440.0640.0917.3017.3336.1836.1974.0673.5151.4450.16S B4 /S B4 S A4 /S B4HARD40.16 40.2840.17 40.2417.35 17.4617.36 17.4636.26 36.3736.26 36.3377.45 77.4577.22 77.7754.02 56.2253.80 55.43S A4 /S A440.1640.1317.3917.3736.2636.2377.2876.7051.6651.77S B4 /S B4 S A4 /S B4SOFT40.28 40.3240.27 40.3117.46 17.5217.45 17.5236.36 36.4136.35 36.4076.99 77.1877.03 77.4555.35 55.7655.59 55.441.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 1.001.00.50.00.51.00.1 0.3 0.5 0.7 0.9 1.1 1.3 1.5 1.7 1.9",
"figure_id": "tab_9",
"figure_label": "8",
"figure_type": "table"
},
{
"figure_caption": "Performance comparison (Acc (%)) of the backtranslation (BT) and Adversarial Attack (AD) augmentation methods within the UDA and FLiText frameworks.",
"figure_data": "ModelsAug L d Agnews Yahoo!Answer DBpeidaUDA TinyBERT 6BT 10 73.90 AD 10 61.2057.14 52.2997.41 88.76FLiTextBT 10 67.14 AD 10 65.1548.30 48.0689.26 85.174.5 Discussion4.5.1 Single Student with AD AugmentationTo demonstrate the necessity of multi-student co-training, we compare the single-student modelwithout co-training with AD data augmentations.Naturally, the single model exclusively uses super-vised data, missing out on leveraging unsuperviseddata. A noteworthy performance decline is ob-served in Table 7 and most differently sized modelsin DBpedia suffer noticeable performance drops.These results validate the DisCo framework's effi-cacy under co-training optimization.",
"figure_id": "tab_10",
"figure_label": "9",
"figure_type": "table"
},
{
"figure_caption": "Text classification performance (Acc (%)) comparison with different combinations of model views in dual-student DisCo (6-layer TinyBERT as the students). SKD denotes the separated-layer knowledge distillation and CKD denotes connected-layer knowledge distillation.",
"figure_data": "is a compressed model implemented by6-layer or 4-layer BERT BASE . For semi-supervisedmethods, we use the released code to train the UDA,which includes ready-made 12-layer BERT BASE , 6-layer, or 4-layer TinyBERT. FLiText (Liu et al.,2021) is a lightweight and fast semi-supervisedlearning framework for the text classification task.FLiText consists of two training stages. It first",
"figure_id": "tab_11",
"figure_label": "10",
"figure_type": "table"
},
{
"figure_caption": "ROUGE performance of models using 10 or 1000 labelled CNN/DailyMail examples.",
"figure_data": "ModelsL mL dCNN/DailyMail R-1 R-2 R-LCPSUM121039.00 16.64 35.23UDASUM121039.03 16.49 35.21UDASUM TinyBERT A441038.67 16.62 35.23UDASUM TinyBERT B441038.78 16.38 35.00DisCo (S A4 )41039.20 16.51 35.34DisCo (S B4 )41038.88 16.61 35.17CPSUM12 1000 40.42 17.62 36.59UDASUM12 1000 40.29 17.65 36.54UDASUM TinyBERT A44 1000 39.99 17.43 36.20UDASUM TinyBERT B44 1000 40.22 17.54 36.34DisCo (S A4 )4 1000 40.49 17.65 36.57DisCo (S B4 )4 1000 40.49 17.64 36.56",
"figure_id": "tab_12",
"figure_label": "11",
"figure_type": "table"
},
{
"figure_caption": "). Results of DisCo equipped with different model-view variants on the",
"figure_data": "Models10Agnews 3020010Yahoo!Answer 3020010DBpeida 30200DisCo (S A4 )76.9085.3987.8251.4862.3668.1094.0298.1398.56DisCo (S B4 )77.3685.5587.9551.3162.9368.2494.7998.1498.63Model Averaging Ensemble77.4585.6088.0452.2363.1768.4994.7598.2098.66",
"figure_id": "tab_13",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Comparison of 4-layer DisCo with model averaging ensemble.",
"figure_data": "",
"figure_id": "tab_14",
"figure_label": "12",
"figure_type": "table"
},
{
"figure_caption": "Comparison between MSE loss and KL divergence in 4-layer DisCo.",
"figure_data": "Models10Agnews 3020010Yahoo!Answer 3020010DBpeida 30200DisCo (S A4 ) + MSE76.9085.3987.8251.4862.3668.1094.0298.1398.56DisCo (S B4 ) + MSE77.3685.5587.9551.3162.9368.2494.7998.1498.63DisCo (S A4 ) + KL76.4683.7687.2052.9061.8167.1795.6397.7298.38DisCo (S B4 ) + KL77.3183.9487.1753.6863.2167.6896.1497.8398.41",
"figure_id": "tab_15",
"figure_label": "13",
"figure_type": "table"
}
] | Weifeng Jiang; Qianren Mao; Chenghua Lin; Jianxin Li; Ting Deng; Weiyi Yang; Zheng Wang | [
{
"authors": "Jimmy Ba; Rich Caruana",
"journal": "",
"ref_id": "b0",
"title": "Do deep nets really need to be deep?",
"year": "2014-12-08"
},
{
"authors": "David Berthelot; Nicholas Carlini; Ian J Goodfellow; Nicolas Papernot; Avital Oliver; Colin Raffel",
"journal": "",
"ref_id": "b1",
"title": "Mixmatch: A holistic approach to semisupervised learning",
"year": "2019"
},
{
"authors": "David Berthelot; Rebecca Roelofs; Kihyuk Sohn; Nicholas Carlini; Alexey Kurakin",
"journal": "",
"ref_id": "b2",
"title": "Adamatch: A unified approach to semi-supervised learning and domain adaptation",
"year": "2022"
},
{
"authors": "Avrim Blum; Tom M Mitchell",
"journal": "ACM",
"ref_id": "b3",
"title": "Combining labeled and unlabeled data with co-training",
"year": "1998"
},
{
"authors": "B Tom; Benjamin Brown; Nick Mann; Melanie Ryder; Jared Subbiah; Prafulla Kaplan; Arvind Dhariwal; Pranav Neelakantan; Girish Shyam; Amanda Sastry; Sandhini Askell; Ariel Agarwal; Gretchen Herbert-Voss; Tom Krueger; Rewon Henighan; Aditya Child; Daniel M Ramesh; Jeffrey Ziegler; Clemens Wu; Christopher Winter; Mark Hesse; Eric Chen; Mateusz Sigler; Scott Litwin; Benjamin Gray; Jack Chess; Christopher Clark; Sam Berner; Alec Mccandlish; Ilya Radford; Dario Sutskever; Amodei",
"journal": "NeurIPS",
"ref_id": "b4",
"title": "Language models are few-shot learners",
"year": "2020"
},
{
"authors": "Ming-Wei Chang; Lev-Arie Ratinov; Dan Roth; Vivek Srikumar",
"journal": "AAAI Press",
"ref_id": "b5",
"title": "Importance of semantic representation: Dataless classification",
"year": "2008"
},
{
"authors": "Pratik Chaudhari; Anna Choromanska; Stefano Soatto; Yann Lecun; Carlo Baldassi; Christian Borgs; Jennifer T Chayes; Levent Sagun; Riccardo Zecchina",
"journal": "ICLR. OpenReview",
"ref_id": "b6",
"title": "Entropy-sgd: Biasing gradient descent into wide valleys",
"year": "2017"
},
{
"authors": "Jiaao Chen; Zichao Yang; Diyi Yang",
"journal": "Association for Computational Linguistics",
"ref_id": "b7",
"title": "Mixtext: Linguistically-informed interpolation of hidden space for semi-supervised text classification",
"year": "2020"
},
{
"authors": "Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova",
"journal": "Association for Computational Linguistics",
"ref_id": "b8",
"title": "BERT: pre-training of deep bidirectional transformers for language understanding",
"year": "2019"
},
{
"authors": "Günes Erkan; Dragomir R Radev",
"journal": "J. Artif. Intell. Res",
"ref_id": "b9",
"title": "Lexrank: Graph-based lexical centrality as salience in text summarization",
"year": "2004"
},
{
"authors": "Anna Yue Fan; Dengxin Kukleva; Bernt Dai; Schiele",
"journal": "Int. J. Comput. Vis",
"ref_id": "b10",
"title": "Revisiting consistency regularization for semi-supervised learning",
"year": "2023"
},
{
"authors": "Tommaso Furlanello; Zachary Lipton; Michael Tschannen; Laurent Itti; Anima Anandkumar",
"journal": "",
"ref_id": "b11",
"title": "Born again neural networks",
"year": "2018"
},
{
"authors": " Pmlr",
"journal": "",
"ref_id": "b12",
"title": "",
"year": ""
},
{
"authors": "Tianyu Gao; Xingcheng Yao; Danqi Chen",
"journal": "Association for Computational Linguistics",
"ref_id": "b13",
"title": "Simcse: Simple contrastive learning of sentence embeddings",
"year": "2021"
},
{
"authors": "Suchin Gururangan; Tam Dang; Dallas Card; Noah A Smith",
"journal": "Association for Computational Linguistics",
"ref_id": "b14",
"title": "Variational pretraining for semi-supervised text classification",
"year": "2019"
},
{
"authors": "Karl Moritz Hermann; Tomás Kociský; Edward Grefenstette; Lasse Espeholt; Will Kay; Mustafa Suleyman; Phil Blunsom",
"journal": "",
"ref_id": "b15",
"title": "Teaching machines to read and comprehend",
"year": "2015"
},
{
"authors": "Geoffrey E Hinton; Nitish Srivastava; Alex Krizhevsky; Ilya Sutskever; Ruslan R Salakhutdinov",
"journal": "",
"ref_id": "b16",
"title": "Improving neural networks by preventing coadaptation of feature detectors",
"year": "2012"
},
{
"authors": "Geoffrey E Hinton; Oriol Vinyals; Jeffrey Dean",
"journal": "",
"ref_id": "b17",
"title": "Distilling the knowledge in a neural network",
"year": "2015"
},
{
"authors": "Haoming Jiang; Pengcheng He; Weizhu Chen; Xiaodong Liu; Jianfeng Gao; Tuo Zhao",
"journal": "Association for Computational Linguistics",
"ref_id": "b18",
"title": "SMART: robust and efficient fine-tuning for pretrained natural language models through principled regularized optimization",
"year": "2020"
},
{
"authors": "Xiaoqi Jiao; Yichun Yin; Lifeng Shang; Xin Jiang; Xiao Chen; Linlin Li; Fang Wang; Qun Liu",
"journal": "Association for Computational Linguistics",
"ref_id": "b19",
"title": "Tinybert: Distilling BERT for natural language understanding",
"year": "2020"
},
{
"authors": "Taehyeon Kim; Jaehoon Oh; Nakyil Kim; Sangwook Cho; Se-Young Yun",
"journal": "",
"ref_id": "b20",
"title": "Comparing kullbackleibler divergence and mean squared error loss in knowledge distillation",
"year": "2021-08"
},
{
"authors": "Yoon Kim",
"journal": "Association for Computational Linguistics",
"ref_id": "b21",
"title": "Convolutional neural networks for sentence classification",
"year": "2014"
},
{
"authors": "Alexey Kurakin; Ian J Goodfellow; Samy Bengio",
"journal": "",
"ref_id": "b22",
"title": "Adversarial examples in the physical world",
"year": "2017"
},
{
"authors": "Dong-Hyun Lee",
"journal": "",
"ref_id": "b23",
"title": "Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks",
"year": "2013"
},
{
"authors": "Haejun Lee; Drew A Hudson; Kangwook Lee; Christopher D Manning",
"journal": "Association for Computational Linguistics",
"ref_id": "b24",
"title": "SLM: learning a discourse language representation with sentence unshuffling",
"year": "2020"
},
{
"authors": "Ju Hyoung; Lee ; Sang-Ki Ko; Yo-Sub Han",
"journal": "AAAI Press",
"ref_id": "b25",
"title": "Salnet: Semi-supervised few-shot text classification with attention-based lexicon construction",
"year": "2021"
},
{
"authors": "Hao Li; Zheng Xu; Gavin Taylor; Christoph Studer; Tom Goldstein",
"journal": "",
"ref_id": "b26",
"title": "Visualizing the loss landscape of neural nets",
"year": "2018"
},
{
"authors": "Chin-Yew Lin; Eduard H Hovy",
"journal": "Association for Computational Linguistics",
"ref_id": "b27",
"title": "Automatic evaluation of summaries using n-gram co-occurrence statistics",
"year": "2003"
},
{
"authors": "Chen Liu; Mengchao Zhang; Zhibing Fu; Panpan Hou; Yu Li",
"journal": "Association for Computational Linguistics",
"ref_id": "b28",
"title": "Flitext: A faster and lighter semisupervised text classification with convolution networks",
"year": "2021"
},
{
"authors": "Yang Liu; Mirella Lapata",
"journal": "Association for Computational Linguistics",
"ref_id": "b29",
"title": "Text summarization with pretrained encoders",
"year": "2019"
},
{
"authors": "Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov",
"journal": "",
"ref_id": "b30",
"title": "Roberta: A robustly optimized BERT pretraining approach",
"year": "2019"
},
{
"authors": "Charles Lovering; Rohan Jha; Tal Linzen; Ellie Pavlick",
"journal": "",
"ref_id": "b31",
"title": "Predicting inductive biases of pretrained models",
"year": "2021-05-03"
},
{
"authors": "Pablo N Mendes; Max Jakob; Christian Bizer",
"journal": "ELRA",
"ref_id": "b32",
"title": "DBpedia: A multilingual cross-domain knowledge base",
"year": "2012"
},
{
"authors": "Stephen Merity; Caiming Xiong; James Bradbury; Richard Socher",
"journal": "",
"ref_id": "b33",
"title": "Pointer sentinel mixture models",
"year": "2017"
},
{
"authors": "Rada Mihalcea; Paul Tarau",
"journal": "Association for Computational Linguistics",
"ref_id": "b34",
"title": "Textrank: Bringing order into text",
"year": "2004"
},
{
"authors": "Seyed-Iman Mirzadeh; Mehrdad Farajtabar; Ang Li; Nir Levine; Akihiro Matsukawa; Hassan Ghasemzadeh",
"journal": "AAAI Press",
"ref_id": "b35",
"title": "Improved knowledge distillation via teacher assistant",
"year": "2020"
},
{
"authors": "Takeru Miyato; Shin-Ichi Maeda; Masanori Koyama; Shin Ishii",
"journal": "IEEE Trans. Pattern Anal. Mach. Intell",
"ref_id": "b36",
"title": "Virtual adversarial training: A regularization method for supervised and semisupervised learning",
"year": "2019"
},
{
"authors": "Gabriel Pereyra; George Tucker; Jan Chorowski; Lukasz Kaiser; Geoffrey E Hinton",
"journal": "ICLR. OpenReview",
"ref_id": "b37",
"title": "Regularizing neural networks by penalizing confident output distributions",
"year": "2017"
},
{
"authors": "Siyuan Qiao; Wei Shen; Zhishuai Zhang; Bo Wang; Alan L Yuille",
"journal": "Springer",
"ref_id": "b38",
"title": "Deep co-training for semisupervised image recognition",
"year": "2018"
},
{
"authors": "Alec Radford; Karthik Narasimhan; Tim Salimans; Ilya Sutskever",
"journal": "",
"ref_id": "b39",
"title": "Improving language understanding by generative pre-training",
"year": "2018"
},
{
"authors": "Alec Radford; Jeffrey Wu; Rewon Child; David Luan; Dario Amodei; Ilya Sutskever",
"journal": "OpenAI blog",
"ref_id": "b40",
"title": "Language models are unsupervised multitask learners",
"year": ""
},
{
"authors": "Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu",
"journal": "J. Mach. Learn. Res",
"ref_id": "b41",
"title": "Exploring the limits of transfer learning with a unified text-to-text transformer",
"year": "2020"
},
{
"authors": "Antti Rasmus; Mathias Berglund; Mikko Honkala; Harri Valpola; Tapani Raiko",
"journal": "",
"ref_id": "b42",
"title": "Semi-supervised learning with ladder networks",
"year": "2015"
},
{
"authors": "Mehdi Sajjadi; Mehran Javanmardi; Tolga Tasdizen",
"journal": "",
"ref_id": "b43",
"title": "Regularization with stochastic transformations and perturbations for deep semi-supervised learning",
"year": "2016"
},
{
"authors": "Victor Sanh; Lysandre Debut; Julien Chaumond; Thomas Wolf",
"journal": "",
"ref_id": "b44",
"title": "Distilbert, a distilled version of BERT: smaller, faster, cheaper and lighter",
"year": "2019"
},
{
"authors": "Dinghan Shen; Mingzhi Zheng; Yelong Shen; Yanru Qu; Weizhu Chen",
"journal": "",
"ref_id": "b45",
"title": "A simple but tough-to-beat data augmentation approach for natural language understanding and generation",
"year": "2020"
},
{
"authors": "Kihyuk Sohn; David Berthelot; Nicholas Carlini; Zizhao Zhang; Han Zhang; Colin Raffel; Ekin Dogus Cubuk; Alexey Kurakin; Chun-Liang Li",
"journal": "NeurIPS",
"ref_id": "b46",
"title": "Fixmatch: Simplifying semi-supervised learning with consistency and confidence",
"year": "2020"
},
{
"authors": "Wonchul Son; Jaemin Na; Junyong Choi; Wonjun Hwang",
"journal": "IEEE",
"ref_id": "b47",
"title": "Densely guided knowledge distillation using multiple teacher assistants",
"year": "2021"
},
{
"authors": "Siqi Sun; Yu Cheng; Zhe Gan; Jingjing Liu",
"journal": "Association for Computational Linguistics",
"ref_id": "b48",
"title": "Patient knowledge distillation for BERT model compression",
"year": "2019"
},
{
"authors": "Coleman Thierry Tambe; Lillian Hooper; Tianyu Pentecost; En-Yu Jia; Marco Yang; Victor Donato; Paul N Sanh; Alexander M Whatmough; David Rush; Gu-Yeon Brooks; Wei",
"journal": "ACM",
"ref_id": "b49",
"title": "Edgebert: Sentence-level energy optimizations for latencyaware multi-task NLP inference",
"year": "2021-10-18"
},
{
"authors": "Antti Tarvainen; Harri Valpola",
"journal": "",
"ref_id": "b50",
"title": "Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results",
"year": "2017"
},
{
"authors": "O Ilya; Neil Tolstikhin; Alexander Houlsby; Lucas Kolesnikov; Xiaohua Beyer; Thomas Zhai; Jessica Unterthiner; Andreas Yung; Daniel Steiner; Jakob Keysers; Mario Uszkoreit; Alexey Lucic; Dosovitskiy",
"journal": "NeurIPS",
"ref_id": "b51",
"title": "Mlp-mixer: An all-mlp architecture for vision",
"year": "2021"
},
{
"authors": "Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Lukasz Kaiser; Illia Polosukhin",
"journal": "NeurIPS",
"ref_id": "b52",
"title": "Attention is all you need",
"year": "2017"
},
{
"authors": "Wenhui Wang; Furu Wei; Li Dong; Hangbo Bao; Nan Yang; Ming Zhou",
"journal": "NeurIPS",
"ref_id": "b53",
"title": "Minilm: Deep selfattention distillation for task-agnostic compression of pre-trained transformers",
"year": "2020"
},
{
"authors": "Yidong Wang; Hao Chen; Yue Fan; Sun Wang; Ran Tao; Wenxin Hou; Renjie Wang; Linyi Yang; Zhi Zhou; Lan-Zhe Guo",
"journal": "NeurIPS",
"ref_id": "b54",
"title": "Usb: A unified semi-supervised learning benchmark for classification",
"year": "2022"
},
{
"authors": "Yiming Wang; Qianren Mao; Junnan Liu; Weifeng Jiang; Hongdong Zhu; Jianxin Li",
"journal": "International Committee on Computational Linguistics",
"ref_id": "b55",
"title": "Noiseinjected consistency training and entropy-constrained pseudo labeling for semi-supervised extractive summarization",
"year": "2022"
},
{
"authors": "Zhuofeng Wu; Sinong Wang; Jiatao Gu; Madian Khabsa; Fei Sun; Hao Ma",
"journal": "",
"ref_id": "b56",
"title": "CLEAR: contrastive learning for sentence representation",
"year": "2020"
},
{
"authors": "Qizhe Xie; Zihang Dai; Eduard H Hovy; Thang Luong; Quoc Le",
"journal": "",
"ref_id": "b57",
"title": "Unsupervised data augmentation for consistency training",
"year": "2020"
},
{
"authors": "Hai-Ming Xu; Lingqiao Liu; Ehsan Abbasnejad",
"journal": "Association for Computational Linguistics",
"ref_id": "b58",
"title": "Progressive class semantic matching for semisupervised text classification",
"year": "2022"
},
{
"authors": "Yuanmeng Yan; Rumei Li; Sirui Wang; Fuzheng Zhang; Wei Wu; Weiran Xu",
"journal": "Association for Computational Linguistics",
"ref_id": "b59",
"title": "Consert: A contrastive framework for self-supervised sentence representation transfer",
"year": "2021"
},
{
"authors": "Weiyi Yang; Richong Zhang; Junfan Chen; Lihong Wang; Jaein Kim",
"journal": "Association for Computational Linguistics",
"ref_id": "b60",
"title": "Prototype-guided pseudo labeling for semi-supervised text classification",
"year": "2023"
},
{
"authors": "Bowen Zhang; Yidong Wang; Wenxin Hou; Hao Wu; Jindong Wang; Manabu Okumura; Takahiro Shinozaki",
"journal": "NeurIPS",
"ref_id": "b61",
"title": "Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling",
"year": "2021"
},
{
"authors": "Xiang Zhang; Junbo ; Jake Zhao; Yann Lecun",
"journal": "NeurIPS",
"ref_id": "b62",
"title": "Character-level convolutional networks for text classification",
"year": "2015"
},
{
"authors": "Ying Zhang; Tao Xiang; Timothy M Hospedales; Huchuan Lu",
"journal": "Computer Vision Foundation / IEEE Computer Society",
"ref_id": "b63",
"title": "Deep mutual learning",
"year": "2018"
},
{
"authors": "Kun Zhao; Bohao Yang; Chenghua Lin; Wenge Rong; Aline Villavicencio; Xiaohui Cui",
"journal": "",
"ref_id": "b64",
"title": "Evaluating open-domain dialogues in latent space with next sentence prediction and mutual information",
"year": "2023"
},
{
"authors": "Mingkai Zheng; Shan You; Lang Huang; Fei Wang; Chen Qian; Chang Xu",
"journal": "IEEE",
"ref_id": "b65",
"title": "Simmatch: Semisupervised learning with similarity matching",
"year": "2022"
}
] | [
{
"formula_coordinates": [
4,
129.55,
634.79,
100.91,
12.25
],
"formula_id": "formula_0",
"formula_text": "L s A = CE( f A (ϕ A ( x)), ŷ),"
},
{
"formula_coordinates": [
4,
129.55,
660.46,
156.08,
12.25
],
"formula_id": "formula_1",
"formula_text": "L s B = CE( f B (ϕ B ( x)), ŷ). (2"
},
{
"formula_coordinates": [
4,
285.63,
661.28,
4.24,
9.74
],
"formula_id": "formula_2",
"formula_text": ")"
},
{
"formula_coordinates": [
4,
335.4,
245.5,
189.74,
14.03
],
"formula_id": "formula_3",
"formula_text": "L u A,B = MSE(z A (ϕ A (x * )), z B (ϕ B (x * ))),(3)"
},
{
"formula_coordinates": [
4,
335.4,
277.41,
185.5,
14.03
],
"formula_id": "formula_4",
"formula_text": "L u B,A = MSE(z B (ϕ B (x * )), z A (ϕ A (x * ))). (4"
},
{
"formula_coordinates": [
4,
520.9,
280.01,
4.24,
9.74
],
"formula_id": "formula_5",
"formula_text": ")"
},
{
"formula_coordinates": [
4,
312.54,
373.64,
212.6,
12.24
],
"formula_id": "formula_6",
"formula_text": "L Θ = L s A + L s B + µ(t, n) • λ • (L u A,B + L u B,A ),(5)"
},
{
"formula_coordinates": [
4,
342.33,
597.56,
182.81,
33.07
],
"formula_id": "formula_7",
"formula_text": "L Θ = K k=1 L s k + µ(t, n) • λ • L u i,k ,(6)"
},
{
"formula_coordinates": [
4,
312.15,
650.48,
204.51,
33.07
],
"formula_id": "formula_8",
"formula_text": "L u i,k = 1 K -1 K i=1,i k MSE(z i (ϕ i (x * )), z k (ϕ k (x * )). ("
}
] |
10.18653/v1/2020.acl-main.679 | 2024-02-27 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b34"
],
"table_ref": [],
"text": "Automatic text summarization (Luhn, 1958) is one of the most important and challenging problems in NLP. Among the different forms the text to be summarized could take, dialogues have been serving as a critical part of human-human and humanmachine interaction. There has been significant progress made in dialogue summarization these * Corresponding author Dialogue States restaurant pricerange: cheap restaurant name: golden house restaurant area: south"
},
{
"figure_ref": [],
"heading": "Dialogue Summary",
"publication_ref": [],
"table_ref": [],
"text": "The user asks for the address, postcode and phone number of Oriental House. The restaurant is in the east and the food is expensive."
},
{
"figure_ref": [],
"heading": "Dialogue Summarization Dialogue State Tracking",
"publication_ref": [],
"table_ref": [],
"text": "Few-Shot"
},
{
"figure_ref": [],
"heading": "Skeleton-Assisted Prompt Transfer (SAPT)",
"publication_ref": [
"b19",
"b10",
"b5",
"b51",
"b20",
"b52",
"b35",
"b48",
"b43",
"b46",
"b27",
"b25",
"b44",
"b46",
"b26",
"b57",
"b53"
],
"table_ref": [],
"text": "Figure 1: We study the problem of how to perform effective transfer learning from dialogue state tracking (DST) to few-shot dialogue summarization, in the scenario where there is a large set of dialogues with DST annotations, and another small set of dialogues with dialogue summarization annotations (i.e., few-shot learning for dialogue summarization).\ndays (Goo and Chen, 2018;Liu et al., 2019b;Chen and Yang, 2020). However, they generally rely on massive human-written golden dialogue summaries. In real-world scenarios, the availability of massive supervised data is not always guaranteed, as the data scarcity problem often occurs due to the high annotation cost that is normally required for acquiring large-scale high-quality dialogue summaries (Bražinskas et al., 2020).\nIn existing works, one common way to tackle the data scarcity problem is to perform transfer learning by leveraging off-the-shelf out-of-domain or out-of-task supervised data (Yang et al., 2020;Goodwin et al., 2020;Yu et al., 2021;Zou et al., 2021;Magooda et al., 2021). We observe that the supervised data of a relevant task called dialogue state tracking (DST) (Williams and Young, 2007) can bring conducive knowledge for the dialogue summarization task, as the semantic slots and values tracked by DST are expected to be covered in the dialogue summary (Shin et al., 2022). Besides the notable relevance between those two tasks, with DST being a language understanding task as opposed to dialogue summarization being a language generation task, the annotations of DST should arguably be easier to get in practice than those of dialogue summarization. 1 These observations motivate us to herein focus on developing effective transfer learning techniques for the scenario where there are ample supervised data for DST whereas the annotations for dialogue summarization are limited, as depicted in Figure 1.\nAmong recent transfer learning techniques, prompt transfer (Vu et al., 2022) in prompt tuning (Li and Liang, 2021;Lester et al., 2021) has gained great popularity because of its parameter efficiency. Prompt tuning is a paradigm of utilizing pretrained language models (PLMs) for downstream tasks, in which a sequence of continuous trainable embeddings called \"soft prompt\" is prepended to the input sequence so as to provide PLMs with an adequate context. During training, only these embeddings can be updated while all the other parameters of PLMs will remain fixed. Prompt transfer realizes cross-task transfer learning under the prompt tuning paradigm by training soft prompts from source tasks and then using them as parameter initialization for the prompt tuning in target tasks. In general, prompt transfer works well in transfer learning between language understanding tasks while it can only provide relatively mediocre performance in language generation tasks (Su et al., 2022), indicating the necessity to design task-specific prompt transfer approaches for language generation tasks such as dialogue summarization.\nHow to improve prompt transfer in a taskspecific manner? The existing general-purpose prompt transfer technique (Vu et al., 2022) relies solely on the source and target task supervision, suggesting the lack of an intermediate task-specific medium that could potentially better connect the distinct source and target task. Also, as the model capability of processing source task data is closely associated with the knowledge it has gained during the source task pretraining, it needs to be effectively preserved during the prompt transfer so as to facilitate the model in handling the target task.\nIn this paper, we propose a dialogue-specific prompt transfer technique, named Skeleton-Assisted Prompt Transfer (SAPT). SAPT provides the model with extra supervision during its prompt transfer by training it to perform skeleton gener-1 In Appendix A, we validate it via a data annotation study. ation along the way. This extra supervision can essentially function as an intermediate task-specific medium that is beneficial for the knowledge transfer between the distinct source and target task. To get the supervised training data for skeleton generation, we design a novel automatic skeleton extraction approach that requires neither annotation effort nor domain knowledge. Specifically, we observe the model's output variation to perturbation-based probes and extract the dialogue turns to which the model displays the highest sensitivity as skeletons.\nTraining the model on such skeletons can also help preserve model capability during prompt transfer. The idea behind this is that we try to prevent the model from forgetting the dialogue-state-related knowledge it has learned during its pretraining on supervised DST data, since the model sensitivity to perturbation-based probes in the DST task intrinsically reflects the capability of processing dialogue state information it has developed. Experimental results and in-depth analyses with BART (Lewis et al., 2020) on two dialogue summarization benchmarks (Zhao et al., 2021b;Yuan and Yu, 2019) demonstrate the effectiveness of our method.\nIn summary, our main contributions are:\n• We focus on improving the prompt transfer in prompt tuning from dialogue state tracking to few-shot dialogue summarization. To the best of our knowledge, SAPT is the first effective dialogue-specific prompt transfer technique.\n• By training the model to perform skeleton generation during prompt transfer, SAPT provides extra supervision that essentially functions as an intermediate task-specific medium between the distinct source and target task, allowing the model to better consume the dialogue state information from the source task.\n• To preserve model capability during prompt transfer, we design a novel approach that employs perturbation-based probes to automatically extract dialogue skeletons as supervised training data for skeleton generation, requiring neither annotation effort nor domain knowledge."
},
{
"figure_ref": [],
"heading": "Preliminaries",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Problem Definition",
"publication_ref": [
"b36",
"b5",
"b29"
],
"table_ref": [],
"text": "Abstractive dialogue summarization is typically formulated as a sequence-to-sequence problem (Nallapati et al., 2016). Given a dialogue history x, a transformer-based encoder-decoder pretrained language model (PLM), p θ (y summ |x), is trained to generate a summary y summ , where θ denotes the trainable parameters of the PLM. In this paper, we specifically study the dialogue summarization task in the few-shot setting (Bražinskas et al., 2020), meaning that there are only a limited number of annotated samples available for model training.\nTo mitigate the data scarcity problem, it is common to turn to transfer learning by leveraging massive supervised data from other related domains or tasks that could potentially provide useful knowledge. Dialogue state tracking (DST), a related task to dialogue summarization, aims to correctly infer the speaker's goal in the form of semantic slot-value pairs ([slot, value]) as a dialogue progresses, such as [food, Italian] and [pricerange, high]. We thus notice that the supervised data of the DST task should be able to bring conducive knowledge for the dialogue summarization task, as the semantic slots and values tracked by DST are expected to be covered in the dialogue summary. Besides the notable relevance between those two tasks, with DST being a language understanding task as opposed to dialogue summarization being a language generation task, the annotations of DST should arguably be easier to get in practice, compared to those of dialogue summarization. Therefore, we herein focus on how to perform effective transfer learning with ample supervised DST data to benefit the few-shot dialogue summarization.\nAlthough the DST task is traditionally formulated as a classification problem, recent work (Lin et al., 2021;Zhao et al., 2021a) has shown the possibility of achieving competitive DST performance by treating DST as a sequence-to-sequence generation task. Specifically, conditioned on the dialogue history x, the encoder-decoder model is trained to generate a sequence of tokens, in the format of \"slot1 is value1, slot2 is value2, ...\", denoted as y dst . We thereby adopt this formulation for DST throughout our work so as to allow the generative encoder-decoder model's knowledge transfer (from DST to dialogue summarization) to happen. With the unified generative sequence-tosequence-based DST and dialogue summarization, the conditional generation task can be formulated as follows (y can be either y summ or y dst ):\nP (y|x) = |y| i=1 p θ (y i |x, y <i )."
},
{
"figure_ref": [],
"heading": "Prompt Transfer in Prompt Tuning",
"publication_ref": [
"b46",
"b27",
"b25",
"b7",
"b44",
"b46",
"b45"
],
"table_ref": [],
"text": "Among recent transfer learning techniques, prompt transfer (Vu et al., 2022) in prompt tuning (Li and Liang, 2021;Lester et al., 2021) has gained great popularity because of its parameter efficiency. We thus adopt it as our starting point for transfer learning from DST to dialogue summarization, and further improve it in section 3.\nPrompt tuning is a new paradigm of utilizing PLMs for downstream tasks. It is motivated by the intuition that PLMs can be steered with a proper context, without the need for any model parameter updates. In prompt tuning, a sequence of continuous trainable embeddings called \"soft prompt\", parameterized by ϕ, is prepended to the input sequence. During training, all parameters of the PLM (θ) are frozen, but unlike prompt design (Brown et al., 2020) which searches for actual tokens in the discrete space, prompt tuning optimizes the \"soft prompt\" (ϕ) directly in the continuous space, allowing it to be more expressive. The log-likelihood training objective can be formulated as follows:\nmax ϕ log p θ,ϕ (y|x) = |y| i=1 log p θ,ϕ (y i |x, y <i ).\nPrompt transfer is a simple yet effective transfer learning technique designed for prompt tuning. The soft prompt is first trained in the source task and then used as parameter initialization for the prompt tuning in the target task. Prompt transfer inherits the advantage of prompt tuning in terms of parameter efficiency, as its transfer learning process likewise relies merely on the lightweight soft prompt. Su et al. (2022) show that prompt transfer generally works well in the transfer learning between language understanding tasks while it can only provide relatively mediocre performance in language generation tasks. This indicates the necessity to design task-specific prompt transfer approaches in language generation tasks such as dialogue summarization, which is exactly the central problem we focus on in this paper (detailed in section 3).\n3 Method: Skeleton-Assisted Prompt Transfer (SAPT)\nThe existing non-task-specific general-purpose prompt transfer technique (Vu et al., 2022) relies solely on the source and target task supervision to train the soft prompt, without the help of any intermediate task-specific medium. Even though DST and dialogue summarization are closely related tasks, the intrinsic domain shift between them should still not be ignored. Therefore, having an intermediate task-specific medium should conceivably be helpful for better connecting the distinct source and target task. Such a medium can take Algorithm 1 Skeleton Extraction with Perturbation-based Probes Input: a collection of dialogues X containing N dialogues: X = {x 1 , x 2 , . . . , x N }, where dialogue x i contains p i dialogue turns:\nx i = [t i1 , t i2 , . . . , t ip i ], 1 ≤ i ≤ N ; a trained DST model LM DST ;\na textual similarity metric Sim(•, •) (higher means more similar). Output:\na collection of dialogue skeletons:\nS = {s 1 , s 2 , . . . , s N }, a subset set(s i ) ⊆ set(x i ) for each dialogue x i ∈ X , 1 ≤ i ≤ N . 1: M = {} 2: for i = 1, 2, . . . , N do 3: o i = LM DST (x i ) 4:\nfor j = 1, 2, . . . , p i do 5:\no ij = LM DST (x i \\ [t ij ]) 6: m ij = Sim(o i , o ij ) 7: add m ij to M 8: S = {} 9: m median = Median(M) 10: for i = 1, 2, . . . , N do 11: s i = [ ] 12: for j = 1, 2, . . . , p i do 13: if m ij < m median then 14: append t ij to s i 15:\nadd s i to S 16: return S the form of extra task supervision separately incorporated into both the source and target task supervision, since in this way the updated source and target task have more overlap and get semantically closer to each other.\nAlso, as the model capability of processing source task data is closely associated with the knowledge it has gained during the source task pretraining, it needs to be effectively preserved during the prompt transfer to facilitate the target task. Nonetheless, the capability per se is admittedly a bit abstract and thus hard to concretely model in practice. Inspired by recent advances in interpretable NLP, we argue that the model sensitivity to perturbation-based probes should arguably be a concretization of model capability (Talmor et al., 2020). Thus, maintaining model sensitivity during the prompt transfer should logically benefit the preservation of model capability. And notably, the aforementioned extra task supervision can ex-actly create conditions for (source-task) modelsensitivity information to be explicitly passed to the (target-task) model during the prompt transfer.\nTo these ends, we propose Skeleton-Assisted Prompt Transfer (SAPT), a dialogue-specific prompt transfer technique. SAPT provides the model with extra supervision during its prompt transfer by training it to perform skeleton generation along the way (detailed in subsection 3.1). This extra supervision (i.e. skeleton generation) is separately incorporated into both the source and target task supervision, and thus can essentially function as an intermediate task-specific medium (because of the increased overlap between the updated source and target task) that is beneficial for the cross-task knowledge transfer.\nTo get the supervised training data for skeleton generation, we design a novel automatic skeleton extraction approach that requires neither annotation effort nor domain knowledge (detailed in subsection 3.2). Specifically, we observe the model's output variation to perturbation-based probes and extract the dialogue turns to which the model displays the highest sensitivity as skeletons. Training the model on such skeletons can also help preserve model capability during prompt transfer. This is because those skeletons (extracted with perturbationbased probes) embody the model sensitivity to perturbation-based probes which is a concretization of model capability.\nOn the whole, SAPT creates an intermediate task-specific medium using skeleton generation as extra supervision ( §3.1), and preserves model capability during prompt transfer by training the model on the skeletons extracted with perturbation-based probes ( §3.2). As a result, the distinct source and target task is able to be better connected because they have got semantically closer to each other, and the target task is able to be facilitated because the model has been discouraged from forgetting the knowledge it has gained during the source task pretraining. §3.3 describes SAPT's overall workflow."
},
{
"figure_ref": [],
"heading": "Skeleton Generation as Extra Supervision",
"publication_ref": [],
"table_ref": [],
"text": "In SAPT, the skeleton generation task is incorporated into the original task (either the source or the target task, or both) as extra supervision. We denote a supervised sample of the original task as (x, y), where x represents the dialogue history and y represents the original task supervision that could be either the sequence-to-sequence-based dialogue state ground-truth or the dialogue summary ground-truth. For each sample (x, y), We also have a dialogue skeleton, denoted as s, extracted from the dialogue history x (the skeleton extraction algorithm is detailed in subsection 3.2). Such a dialogue skeleton is essentially an ordered collection of dialogue turns. For instance, if a dialogue history x contains p dialogue turns, i.e. x = [t 1 , t 2 , . . . , t p ], its dialogue skeleton s will contain q dialogue turns (q ≤ p), denoted as s = [t s 1 , t s 2 , . . . , t s q ], and thus set(s) ⊆ set(x). The dialogue skeleton s is appended to the original task supervision y as extra supervision, and the model is trained to perform the original task and then skeleton generation. The new log-likelihood training objective is:\nmax ϕ log p θ,ϕ (y ⊕ s | x) = log p θ,ϕ (y|x) + log p θ,ϕ (s|x, y) = log p θ,ϕ (y|x) + q i=1 log p θ,ϕ (t s i |x, y, t s <i )."
},
{
"figure_ref": [],
"heading": "Skeleton Extraction with Perturbation-based Probes",
"publication_ref": [],
"table_ref": [],
"text": "We extract dialogue skeletons (used as supervised training data for skeleton generation in subsection 3.1) with perturbation-based probes. Given a dialogue in a collection of dialogues, x i ∈ X , we first construct the perturbation-based probes by deleting a dialogue turn from x i at a time. The resultant perturbation-based probes can be expressed as\nx i \\ [t ij ], 1 ≤ j ≤ p i (x i contains p i dialogue turns).\nWe then feed those perturbation-based probes individually into the trained source-task (DST) model, LM DST , and get the model output o ij corresponding to each deleted dialogue turn t ij .\nIn the meantime, we also feed the whole dialogue history x i into LM DST and get the model output o i . Next, we compute the textual similarity score m ij between o i and o ij using a textual similarity metric Sim(•, •) (higher means more similar). We execute the aforementioned procedure for each dialogue in X . After that, we group together all the similarity scores we compute along the way and find the median of them. Finally, we extract those dialogue turns, whose corresponding similarity scores are less than the median, as the dialogue skeletons. Algorithm 1 presents the process of extracting a dialogue skeleton s i for each dialogue x i ∈ X ."
},
{
"figure_ref": [
"fig_0"
],
"heading": "Overall Workflow",
"publication_ref": [
"b46"
],
"table_ref": [],
"text": "Built on top of SPOT (Vu et al., 2022) As depicted in Figure 2, SAPT [DST+SUMM] includes four steps:\n1. perform prompt tuning on the DST (source task) supervision; 2. perform prompt transfer from the previous step, and then perform prompt tuning on the DST (source task) & skeleton generation supervision; 3. perform prompt transfer from the previous step, and then perform prompt tuning on the (few-shot) dialogue summarization (target task) & skeleton generation supervision; 4. perform prompt transfer from the previous step, and then perform prompt tuning on the (few-shot) dialogue summarization (target task) supervision."
},
{
"figure_ref": [],
"heading": "Compared to SAPT [DST+SUMM]",
"publication_ref": [
"b46"
],
"table_ref": [],
"text": ", SAPT [DST] omits step #3 while SAPT [SUMM] omits step #2; SPOT (Vu et al., 2022) omits both step #2 and step #3."
},
{
"figure_ref": [],
"heading": "Experiment",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Dataset and Baseline",
"publication_ref": [
"b55",
"b57",
"b53",
"b9",
"b26",
"b46",
"b46"
],
"table_ref": [],
"text": "To study the cross-task prompt transfer from dialogue state tracking (DST) to few-shot dialogue summarization, we perform experiments on a DST dataset: MultiWOZ 2.2 (Zang et al., 2020), and on two task-oriented dialogue summarization datasets: TODSUM (Zhao et al., 2021b) and SPNET (Yuan and Yu, 2019). MultiWOZ 2.2 is an error-fixed version of MultiWOZ (Budzianowski et al., 2018), which is a classic task-oriented multi-domain dialogue dataset containing over 10,000 annotated dialogues and has been extensively used for studying DST. TODSUM and SPNET are both constructed using the dialogues from MultiWOZ, and differ mainly in terms of summary style and length. On average, the summaries in SPNET are roughly two times longer than those in TODSUM (96.4 vs. 45.4 words). To evaluate our method under the few-shot setting, on each dialogue summarization dataset we randomly choose 100 samples from the training set for model training and test on the full test set.\nWe use BART-large2 (Lewis et al., 2020) as the backbone throughout the experiments. We focus on the comparison between prompt-tuning-based methods, as they have been proven to be able to maintain as comparable performance as the adapterbased methods while being much more parameterefficient (Li and Liang, 2021; Vu et al., 2022). We choose SPOT (Vu et al., 2022) as the baseline method, which has been commonly used as a parameter-efficient transfer learning technique. Appendix B presents the implementation details."
},
{
"figure_ref": [],
"heading": "Automatic Evaluation",
"publication_ref": [
"b28",
"b25"
],
"table_ref": [
"tab_1"
],
"text": "We use the widely-used ROUGE metrics (Lin, 2004) as automatic evaluation metrics, including ROUGE-1 (R-1), ROUGE-2 (R-2), and ROUGE-L (R-L) F1 scores with rouge-score python package3 . Few-shot (100-shot) results are presented in Table 1, where we also attach the results of PROMPT TUNING (Li and Liang, 2021). Unsurprisingly, PROMPT TUNING performs badly without any knowledge transfer, which indicates the necessity of conducting prompt transfer from DST to few-shot dialogue summarization. Among different prompt transfer techniques, all three SAPT variants outperform the baseline method SPOT on both datasets, suggesting the effectiveness of the proposed SAPT method. It is also observed that SAPT[DST] consistently outperforms SAPT [SUMM]. We attribute this to the fact that there are much more dialogue samples (along with their extracted dialogue skeletons) that are used for the preservation of model capability during step #2 than step #3, as in step #3 we only make use of 100 dialogue samples dedicated for few-shot dialogue summarization. Notably, when both step #2 and step #3 are executed, SAPT[DST+SUMM] is able to further improve the performance by a significant margin, compared to SAPT [DST] and SAPT[SUMM]. This demonstrates the effectiveness of creating an intermediate task-specific medium between the source DST task and the target few-shot dialogue summarization task (by incorporating the skeleton generation task into both\nTODSUM SPNET Models R-1 R-2 R-L R-1 R-2 R-L\nPROMPT TUNING (Lester et al., 2021) Table 3: Results of ablation studies on the effect of skeleton type, decoding order, and source & target task supervision for all three SAPT variants on TODSUM test set.\nskeleton type R-1 R-2 R-L decoding order R-1 R-2 R-L source & target task supervision R-1 R-2 R-L SAPT [\nof them)."
},
{
"figure_ref": [],
"heading": "Human Evaluation",
"publication_ref": [],
"table_ref": [],
"text": "To further evaluate the generated summaries, we perform a human evaluation via crowdsourcing. We randomly select 100 samples from TODSUM test set and run different models on them to generate summaries. We recruit human participants on Prolific 4 , a crowdsourcing platform, to rate the generated summaries (and also the ground-truth summaries) from 0 to 2 in terms of four evaluation metrics: informativeness, faithfulness, fluency, and redundancy 5 . Each summary instance is evaluated by 5 different human participants, and the interannotator agreement (IAA) score for each metric is 0.577, 0.635, 0.649, 0.591, with an average IAA of 0.613. Results shown by the average scores in Ta-4 https://www.prolific.co/ 5 Details of the metrics can be found in Appendix C.\nble 2 are consistent with the automatic evaluation results: all three SAPT variants outperform the baseline method SPOT, and SAPT[DST+SUMM] consistently performs the best across all metrics. Meanwhile, all generated summaries are deemed to be worse than the ground-truth summaries, meaning that there is still room for these summarization models to be improved. We also conduct a case study by ourselves, detailed in Appendix D."
},
{
"figure_ref": [],
"heading": "Ablation Study",
"publication_ref": [],
"table_ref": [],
"text": "To fully investigate the effectiveness of SAPT, we study the impact of skeleton type, decoding order, and source & target task supervision. Table 3 shows the results of ablation studies. Skeleton Type. We replace the extracted skeletons ( §3.2) with randomly-extracted skeletons. We make sure that in total half of the dialogue turns are selected as skeletons to align with our usage of Median(), and that there is at least one dialogue turn selected for each dialogue. The observed performance drop demonstrates the effectiveness of our skeletons extracted with perturbation-based probes. Models with random skeleton still outperform SPOT in general, and we attribute this to the possible match between random skeleton and our skeleton, and also the imperfect intermediate taskspecific medium which persists in the workflow. Decoding Order. We prepend the skeletons instead of appending them. The observed performance drop demonstrates that the original task supervision needs to be prioritized, and prepending makes it more difficult for models to learn the crosstask knowledge.\nSource & Target Task Supervision. We remove all the original task supervision along the way. The observed performance drop is as expected, but the superior performance against SPOT demonstrates the benefit our skeletons bring for cross-task knowledge transfer."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b13",
"b38",
"b26",
"b7",
"b27",
"b25",
"b40",
"b46",
"b44",
"b46",
"b23",
"b22",
"b16",
"b47",
"b2",
"b24",
"b15",
"b11",
"b51",
"b20",
"b52",
"b35",
"b5",
"b6",
"b58",
"b30",
"b54",
"b1",
"b37",
"b17",
"b3",
"b50",
"b41",
"b0",
"b12",
"b39",
"b45",
"b4",
"b21"
],
"table_ref": [],
"text": "Parameter-Efficient Transfer Learning. To efficiently make use of pretrained language models (PLMs) (Devlin et al., 2019;Raffel et al., 2020;Lewis et al., 2020;Brown et al., 2020), Li and Liang (2021) propose to prepend continuous trainable task-specific embeddings to the input sequence while keeping the entire PLM frozen. Lester et al. (2021) provide a simplified approach, named prompt tuning, which becomes more competitive with model fine-tuning as scale increases. To enable cross-task knowledge transfer (Ruder, 2017;Liu et al., 2019a) under the prompt tuning paradigm, Vu et al. (2022) propose SPoT, which learns soft prompts from source tasks as initialization for target tasks. Su et al. (2022) further explore the transferability of soft prompts across different downstream tasks. Built on top of Vu et al. (2022), our method is able to improve the effectiveness of cross-task prompt transfer in few-shot dialogue summarization. Low-Resource Abstractive Summarization. Multiple lines of approaches have been proposed to mitigate the data scarcity problem in abstractive summarization, such as reinforcement learning (Kohita et al., 2020;Hyun et al., 2022), selfsupervised learning (Fu et al., 2021;Wang and Wan, 2021;Zhuang et al., 2022), data augmentation (Amplayo and Lapata, 2020;Laskar et al., 2020;Fabbri et al., 2021;Chen and Yang, 2021), model pretraining or fine-tuning with in-domain unlabeled data or out-of-domain labeled data (Yang et al., 2020;Goodwin et al., 2020;Yu et al., 2021;Zou et al., 2021;Magooda et al., 2021), and few-shot learning via adapters (Bražinskas et al., 2020;Brazinskas et al., 2022) or prompt tuning (Zhao et al., 2022;Liu et al., 2022;Yuan et al., 2022). In this paper, we focus on the few-shot dialogue summarization and improve it by ameliorating cross-task prompt transfer in prompt tuning with cross-task labeled data. Perturbation-based Probes. In interpretable NLP, while probes sometimes refer to algorithms or models aiming to extract information from continuous embeddings (Adi et al., 2017), they can also refer to textual inputs designed for acquiring model outputs that are either useful for downstream tasks (Petroni et al., 2019;Zhong et al., 2021) or informative for model interpretability (Goldberg, 2019;Bacon and Regier, 2019;Xie et al., 2022). Perturbation-based probes, which fall into the latter category, have gained popularity because of their simplicity and cost-efficiency. For instance, Sankar et al. (2019); Abdou et al. (2020); Ettinger (2020); Clouatre et al. (2022) investigate the sensitivity of neural language models to input perturbation; Richardson and Sabharwal (2020); Talmor et al. (2020); Bitton et al. (2021); Gupta et al. (2022) utilize perturbation to construct better NLP testbeds. In contrast, we leverage perturbation-based probes to automatically extract skeletons from dialogues."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "We focus on improving the prompt transfer in prompt tuning from dialogue state tracking to few-shot dialogue summarization, and propose SAPT, a dialogue-specific prompt transfer technique, which uses skeleton generation as extra supervision by training the model on the dialogue skeletons extracted with perturbation-based probes. In this way, a beneficial intermediate task-specific medium is created between the source and target task, and the model capability is able to be better preserved during the prompt transfer, resulting in the model's better consumption of dialogue state information from the source task. Significantly stronger empirical performance and in-depth analyses on two dialogue summarization benchmarks demonstrate the effectiveness of our method in fewshot dialogue summarization.\nDespite the strong performance achieved by SAPT, we use the pre-trained language model (PLM) as the backbone of our method. Therefore, we cannot go beyond the limitation of the maximum sequence length of the PLM. In fact, long-form language understanding and generation have been widely acknowledged as an open research question that needs much further investigation, which is beyond the scope of our paper."
},
{
"figure_ref": [],
"heading": "Ethics & Broader Impacts",
"publication_ref": [
"b42"
],
"table_ref": [],
"text": "All datasets used in this work are public. We did not collect any personal information from our human participants nor did we present them with any harmful model outputs. Our dialogue summarization models face the same potential pitfalls as other contemporary language learning systems do, e.g. being prone to echoing the biases present in the dataset (Sheng et al., 2019). "
},
{
"figure_ref": [],
"heading": "A Data Annotation Study",
"publication_ref": [],
"table_ref": [],
"text": "We recruit 30 human participants on Prolific6 , a crowdsourcing platform, to annotate 30 dialogues for their dialogue states and dialogue summaries. We split 30 participants into two batches and split 30 dialogues into two batches as well. We follow a Latin Square design, similarly to (Gonzalez and Søgaard, 2020), to make sure that each batch of participants only sees each batch of dialogues in one of the following two annotation settings: dialogue state and dialogue summary, yet each setting is tested on both all 30 annotators and all 30 dialogues. This ensures that no bias in the duration of annotation occurs due to annotators having previously seen the dialogues.\nWe measure the duration of the annotation processes for both dialogue state and dialogue sum-mary. The average duration of annotating a dialogue for its dialogue states is 1.3 minutes; the average duration of annotating a dialogue for its dialogue summary is 3.8 minutes, which is much longer. These results are in line with our intuition: the annotation of a dialogue summary requires not only tracking the dialogue states, but also having an utterance-level detailed understanding of the dialogue, because only after understanding the whole dialogue progression can annotators write a fluent and faithful summary."
},
{
"figure_ref": [],
"heading": "B Implementation Details",
"publication_ref": [
"b49",
"b33"
],
"table_ref": [],
"text": "We use Hugging Face Transformers 7 (Wolf et al., 2020) during implementation. We train the BARTlarge models using AdamW (Loshchilov and Hutter, 2019) with the default learning rate linearly decaying from 5E -5. All models with a prompt length of 200 are trained for 50 epochs on an NVIDIA TITAN Xp GPU (12 GB memory) with a batch size of 2 and they each take approximately 25 hours (for DST) / 0.3 hours (for 100-shot dialogue summarization) to train. During inference, we perform a beam search with a beam size of 6, and the decoding takes 1.5 seconds per batch.\nAll turns of the input dialogue are prepended with special tokens as speaker identifiers ([USER] or [SYSTEM]), and then concatenated into a single input sequence which is truncated to 1024 BPE tokens. We use the ROUGE-L F1 score as the textual similarity metric Sim(•, •) in Algorithm 1. The dialogue skeletons are appended to the groundtruth dialogue states (or summaries), and there is a special token [SEP] between the dialogue states (or summaries) and skeletons."
},
{
"figure_ref": [],
"heading": "C Details of Human Evaluation Metrics",
"publication_ref": [],
"table_ref": [],
"text": "Human participants are asked to read the summaries and give their ratings (0, 1, or 2) in terms of four evaluation metrics:\n• Informativeness examines whether the critical information in the dialogue is missed in the summary:\n⋆ 0: lots of the critical information in the dialogue is missed; ⋆ 1: a small amount of the critical information in the dialogue is missed; ⋆ 2: no critical information in the dialogue is missed.\n7 https://github.com/huggingface/transformers\n• Faithfulness examines whether the information presented in the summary is factually incorrect or unmentioned according to the dialogue:\n⋆ 0: lots of the information presented in the summary is factually incorrect or unmentioned; ⋆ 1: a small amount of the information presented in the summary is factually incorrect or unmentioned; ⋆ 2: no information presented in the summary is factually incorrect or unmentioned.\n• Fluency examines whether the sentences in the summary are ungrammatical or ill-formed:\n⋆ 0: lots of the sentences in the summary are ungrammatical or ill-formed; ⋆ 1: a small amount of the sentences in the summary are ungrammatical or illformed; ⋆ 2: no sentence in the summary is ungrammatical or ill-formed.\n• Redundancy examines whether the expressions of the summary can be simplified:\n⋆ 0: lots of the expressions of the summary can be simplified; ⋆ 1: a small amount of the expressions of the summary can be simplified; ⋆ 2: no expression of the summary can be simplified."
},
{
"figure_ref": [],
"heading": "D Case Study",
"publication_ref": [],
"table_ref": [],
"text": "We present a case study in Table 4 to illustrate the effectiveness of SAPT."
},
{
"figure_ref": [],
"heading": "Dialogue Example",
"publication_ref": [
"b46"
],
"table_ref": [],
"text": "I am looking for a taxi from yu garden arriving by 14:30.\nI'll need a destination to give them before I can book that for you. I'd be going to holy trinity church. I'll need the car type and contact number as well.\nThe car that is reserved for you is a grey Toyota, and the contact number is 07997985529. Can I do anything else for you? Nope that's all thank you.\nAlright, enjoy your trip! SPOT (Vu et al., 2022) The user asks the user to book a taxi from yu garden arriving by 14:30, and the destination is holy trinity church."
},
{
"figure_ref": [],
"heading": "SAPT [DST]",
"publication_ref": [],
"table_ref": [],
"text": "The user asks for the car type and contact number of the taxi. The taxi leaves at 14:30 on thursday, and the destination is holy trinity church."
},
{
"figure_ref": [],
"heading": "SAPT [SUMM]",
"publication_ref": [],
"table_ref": [],
"text": "The user wants to know the car type and the contact number of the taxi. The taxi leaves at 14:30, and the destination is holy trinity church."
},
{
"figure_ref": [],
"heading": "SAPT [DST+SUMM]",
"publication_ref": [],
"table_ref": [],
"text": "The user asks the agent to check the car type and contact number of the taxi. The taxi leaves at 14:30 on thursday, from yu garden to holy trinity church."
},
{
"figure_ref": [],
"heading": "Ground Truth",
"publication_ref": [],
"table_ref": [],
"text": "The user wonders if it is possible to know the car type and the phone number. The taxi arrives at 14:30, from yu garden to holy trinity church.\nTable 4: A case study. We highlight all dialogue-state-related information. The summaries provided by all three SAPT variants provide more complete dialogue-state-related information coverage than the baseline method SPOT. Among those three variants, only SAPT [DST+SUMM] covers all dialogue-state-related information. However, compared to the ground truth, the summary provided by SAPT [DST+SUMM] contains information that is unmentioned in the dialogue (i.e. on Thursday), suggesting there is still room for SAPT to be improved."
}
] | In real-world scenarios, labeled samples for dialogue summarization are usually limited (i.e., few-shot) due to high annotation costs for high-quality dialogue summaries. To efficiently learn from few-shot samples, previous works have utilized massive annotated data from other downstream tasks and then performed prompt transfer in prompt tuning so as to enable cross-task knowledge transfer. However, existing general-purpose prompt transfer techniques lack consideration for dialoguespecific information. In this paper, we focus on improving the prompt transfer from dialogue state tracking to dialogue summarization and propose Skeleton-Assisted Prompt Transfer (SAPT), which leverages skeleton generation as extra supervision that functions as a medium connecting the distinct source and target task and resulting in the model's better consumption of dialogue state information. To automatically extract dialogue skeletons as supervised training data for skeleton generation, we design a novel approach with perturbationbased probes requiring neither annotation effort nor domain knowledge. Training the model on such skeletons can also help preserve model capability during prompt transfer. Our method significantly outperforms existing baselines. Indepth analyses demonstrate the effectiveness of our method in facilitating cross-task knowledge transfer in few-shot dialogue summarization. | Few-Shot Dialogue Summarization via Skeleton-Assisted Prompt Transfer in Prompt Tuning | [
{
"figure_caption": "Figure 2 :2Figure 2: The overall workflow of Skeleton-Assisted Prompt Transfer (SAPT). Besides the original task supervision (y dst or y summ ), SAPT uses skeleton generation as extra supervision ( §3.1) by training on the dialogue skeletons s extracted with perturbation-based probes ( §3.2).",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": " results on the full TODSUM(Zhao et al., 2021b) and SPNET(Yuan and Yu, 2019) test set. All three SAPT variants outperform the baseline model on both datasets, SPOT(Vu et al., 2022). SAPT [DST+SUMM] achieves the highest ROUGE scores with significant performance improvements.",
"figure_data": "18.672.8513.33 33.29 11.24 19.32SPOT (Vu et al., 2022)56.96 30.26 38.40 45.46 33.27 39.49SAPT [DST]62.00 36.95 43.13 53.43 40.07 44.92SAPT [SUMM]57.39 34.60 42.50 49.65 37.30 42.57SAPT [DST+SUMM]62.25 40.75 48.30 56.49 41.93 47.46Informativeness Faithfulness Fluency RedundancyGround Truth1.921.901.951.97SPOT (Vu et al., 2022)1.771.701.731.71SAPT [DST]1.821.771.851.79SAPT [SUMM]1.801.761.851.73SAPT [DST+SUMM]1.861.821.901.81",
"figure_id": "tab_1",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Human evaluation results in terms of the informativeness, faithfulness, fluency, and redundancy of the generated summaries on TODSUM test set. SAPT [DST+SUMM] consistently performs the best across all metrics.",
"figure_data": "",
"figure_id": "tab_2",
"figure_label": "2",
"figure_type": "table"
}
] | Kaige Xie; ♣ Tong; Haoliang Wang; Junda Wu; Handong Zhao; Ruiyi Zhang; Kanak Mahadik; Ani Nenkova; Mark Riedl | [
{
"authors": "Mostafa Abdou; Vinit Ravishankar; Maria Barrett; Yonatan Belinkov; Desmond Elliott; Anders Søgaard",
"journal": "Association for Computational Linguistics",
"ref_id": "b0",
"title": "The sensitivity of language models and humans to Winograd schema perturbations",
"year": "2020"
},
{
"authors": "Yossi Adi; Einat Kermany; Yonatan Belinkov; Ofer Lavi; Yoav Goldberg",
"journal": "",
"ref_id": "b1",
"title": "Fine-grained analysis of sentence embeddings using auxiliary prediction tasks",
"year": "2017"
},
{
"authors": "Reinald Kim; Amplayo ; Mirella Lapata",
"journal": "Association for Computational Linguistics",
"ref_id": "b2",
"title": "Unsupervised opinion summarization with noising and denoising",
"year": "2020"
},
{
"authors": "Geoff Bacon; Terry Regier",
"journal": "",
"ref_id": "b3",
"title": "Does bert agree? evaluating knowledge of structure dependence through agreement relations",
"year": "2019"
},
{
"authors": "Yonatan Bitton; Gabriel Stanovsky; Roy Schwartz; Michael Elhadad",
"journal": "Association for Computational Linguistics",
"ref_id": "b4",
"title": "Automatic generation of contrast sets from scene graphs: Probing the compositional consistency of GQA",
"year": "2021"
},
{
"authors": "Arthur Bražinskas; Mirella Lapata; Ivan Titov",
"journal": "Association for Computational Linguistics",
"ref_id": "b5",
"title": "Few-shot learning for opinion summarization",
"year": "2020"
},
{
"authors": "Arthur Brazinskas; Ramesh Nallapati; Mohit Bansal; Markus Dreyer",
"journal": "Association for Computational Linguistics",
"ref_id": "b6",
"title": "Efficient few-shot finetuning for opinion summarization",
"year": "2022"
},
{
"authors": "Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell; Sandhini Agarwal; Ariel Herbert-Voss; Gretchen Krueger; Tom Henighan; Rewon Child; Aditya Ramesh; Daniel Ziegler; Jeffrey Wu; Clemens Winter; Chris Hesse; Mark Chen; Eric Sigler; Mateusz Litwin; Scott Gray; Benjamin Chess; Jack Clark; Christopher Berner; Sam Mccandlish; Alec Radford; Ilya Sutskever; Dario Amodei",
"journal": "",
"ref_id": "b7",
"title": "Language models are few-shot learners",
"year": "2020"
},
{
"authors": "Curran Associates; Inc ",
"journal": "",
"ref_id": "b8",
"title": "",
"year": ""
},
{
"authors": "Paweł Budzianowski; Tsung-Hsien Wen; Bo-Hsiang Tseng; Iñigo Casanueva; Stefan Ultes; Milica Osman Ramadan; Gašić",
"journal": "Association for Computational Linguistics",
"ref_id": "b9",
"title": "MultiWOZ -a largescale multi-domain Wizard-of-Oz dataset for taskoriented dialogue modelling",
"year": "2018"
},
{
"authors": "Jiaao Chen; Diyi Yang",
"journal": "Association for Computational Linguistics",
"ref_id": "b10",
"title": "Multi-view sequenceto-sequence models with conversational structure for abstractive dialogue summarization",
"year": "2020"
},
{
"authors": "Jiaao Chen; Diyi Yang",
"journal": "Association for Computational Linguistics",
"ref_id": "b11",
"title": "Simple conversational data augmentation for semi-supervised abstractive dialogue summarization",
"year": "2021"
},
{
"authors": "Louis Clouatre; Prasanna Parthasarathi; Amal Zouaq; Sarath Chandar",
"journal": "Association for Computational Linguistics",
"ref_id": "b12",
"title": "Detecting languages unintelligible to multilingual models through local structure probes",
"year": "2022"
},
{
"authors": "Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova",
"journal": "Association for Computational Linguistics",
"ref_id": "b13",
"title": "BERT: Pre-training of deep bidirectional transformers for language understanding",
"year": "2019"
},
{
"authors": "Allyson Ettinger",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b14",
"title": "What BERT is not: Lessons from a new suite of psycholinguistic diagnostics for language models",
"year": "2020"
},
{
"authors": "Alexander Fabbri; Simeng Han; Haoyuan Li; Haoran Li; Marjan Ghazvininejad; Shafiq Joty; Dragomir Radev; Yashar Mehdad",
"journal": "Association for Computational Linguistics",
"ref_id": "b15",
"title": "Improving zero and few-shot abstractive summarization with intermediate fine-tuning and data augmentation",
"year": "2021"
},
{
"authors": "Xiyan Fu; Yating Zhang; Tianyi Wang; Xiaozhong Liu; Changlong Sun; Zhenglu Yang",
"journal": "Association for Computational Linguistics",
"ref_id": "b16",
"title": "RepSum: Unsupervised dialogue summarization based on replacement strategy",
"year": "2021"
},
{
"authors": "Yoav Goldberg",
"journal": "",
"ref_id": "b17",
"title": "Assessing bert's syntactic abilities",
"year": "2019"
},
{
"authors": "Ana Valeria; Gonzalez ; Anders Søgaard",
"journal": "",
"ref_id": "b18",
"title": "The reverse turing test for evaluating interpretability methods on unknown tasks",
"year": "2020"
},
{
"authors": "Chih- ; Wen Goo; Yun-Nung Chen",
"journal": "IEEE",
"ref_id": "b19",
"title": "Abstractive dialogue summarization with sentence-gated modeling optimized by dialogue acts",
"year": "2018"
},
{
"authors": "Travis Goodwin; Max Savery; Dina Demner-Fushman",
"journal": "Association for Computational Linguistics",
"ref_id": "b20",
"title": "Towards Zero-Shot Conditional Summarization with Adaptive Multi-Task Fine-Tuning",
"year": "2020"
},
{
"authors": "Vivek Gupta; Riyaz A Bhat; Atreya Ghosal; Manish Shrivastava; Maneesh Singh; Vivek Srikumar",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b21",
"title": "Is my model using the right evidence? systematic probes for examining evidence-based tabular reasoning",
"year": "2022"
},
{
"authors": "Dongmin Hyun; Xiting Wang; Chayoung Park; Xing Xie; Hwanjo Yu",
"journal": "Association for Computational Linguistics",
"ref_id": "b22",
"title": "Generating multiplelength summaries via reinforcement learning for unsupervised sentence summarization",
"year": "2022"
},
{
"authors": "Ryosuke Kohita; Akifumi Wachi; Yang Zhao; Ryuki Tachibana",
"journal": "Association for Computational Linguistics",
"ref_id": "b23",
"title": "Q-learning with language model for edit-based unsupervised summarization",
"year": "2020"
},
{
"authors": "Md Tahmid Rahman Laskar; Enamul Hoque; Jimmy Xiangji Huang",
"journal": "International Committee on Computational Linguistics",
"ref_id": "b24",
"title": "WSL-DS: Weakly supervised learning with distant supervision for query focused multi-document abstractive summarization",
"year": "2020"
},
{
"authors": "Brian Lester; Rami Al-Rfou; Noah Constant",
"journal": "Association for Computational Linguistics",
"ref_id": "b25",
"title": "The power of scale for parameter-efficient prompt tuning",
"year": "2021"
},
{
"authors": "Mike Lewis; Yinhan Liu; Naman Goyal; Marjan Ghazvininejad; Abdelrahman Mohamed; Omer Levy; Veselin Stoyanov; Luke Zettlemoyer",
"journal": "Association for Computational Linguistics",
"ref_id": "b26",
"title": "BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension",
"year": "2020"
},
{
"authors": "Lisa Xiang; Percy Li; Liang",
"journal": "Association for Computational Linguistics",
"ref_id": "b27",
"title": "Prefix-tuning: Optimizing continuous prompts for generation",
"year": "2021"
},
{
"authors": "Chin-Yew Lin",
"journal": "Association for Computational Linguistics",
"ref_id": "b28",
"title": "ROUGE: A package for automatic evaluation of summaries",
"year": "2004"
},
{
"authors": "Zhaojiang Lin; Bing Liu; Seungwhan Moon; Paul Crook; Zhenpeng Zhou; Zhiguang Wang; Zhou Yu; Andrea Madotto; Eunjoon Cho; Rajen Subba",
"journal": "Association for Computational Linguistics",
"ref_id": "b29",
"title": "Leveraging slot descriptions for zero-shot cross-domain dialogue StateTracking",
"year": "2021"
},
{
"authors": "Xiaochen Liu; Yang Gao; Yu Bai; Jiawei Li; Yinan Hu; Heyan Huang; Boxing Chen",
"journal": "",
"ref_id": "b30",
"title": "PSP: Pre-trained soft prompts for few-shot abstractive summarization",
"year": "2022"
},
{
"authors": "Xiaodong Liu; Pengcheng He; Weizhu Chen; Jianfeng Gao; ; ",
"journal": "Association for Computational Linguistics",
"ref_id": "b31",
"title": "Multi-task deep neural networks for natural language understanding",
"year": "2019"
},
{
"authors": "Zhengyuan Liu; Angela Ng; Sheldon Lee; Ai Ti Aw; Nancy F Chen",
"journal": "IEEE",
"ref_id": "b32",
"title": "Topic-aware pointergenerator networks for summarizing spoken conversations",
"year": "2019"
},
{
"authors": "Ilya Loshchilov; Frank Hutter",
"journal": "",
"ref_id": "b33",
"title": "Decoupled weight decay regularization",
"year": "2019"
},
{
"authors": "Hans Peter; Luhn ",
"journal": "IBM Journal of research and development",
"ref_id": "b34",
"title": "The automatic creation of literature abstracts",
"year": "1958"
},
{
"authors": "Ahmed Magooda; Diane Litman; Mohamed Elaraby",
"journal": "Association for Computational Linguistics",
"ref_id": "b35",
"title": "Exploring multitask learning for low-resource abstractive summarization",
"year": "2021"
},
{
"authors": "Ramesh Nallapati; Bowen Zhou; Caglar Cicero Dos Santos; Bing Gulcehre; Xiang",
"journal": "Association for Computational Linguistics",
"ref_id": "b36",
"title": "Abstractive text summarization using sequence-to-sequence RNNs and beyond",
"year": "2016"
},
{
"authors": "Fabio Petroni; Tim Rocktäschel; Sebastian Riedel; Patrick Lewis; Anton Bakhtin; Yuxiang Wu; Alexander Miller",
"journal": "Association for Computational Linguistics",
"ref_id": "b37",
"title": "Language models as knowledge bases?",
"year": "2019"
},
{
"authors": "Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu",
"journal": "Journal of Machine Learning Research",
"ref_id": "b38",
"title": "Exploring the limits of transfer learning with a unified text-to-text transformer",
"year": "2020"
},
{
"authors": "Kyle Richardson; Ashish Sabharwal",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b39",
"title": "What does my QA model know? devising controlled probes using expert knowledge",
"year": "2020"
},
{
"authors": "Sebastian Ruder",
"journal": "",
"ref_id": "b40",
"title": "An overview of multi-task learning in deep neural networks",
"year": "2017"
},
{
"authors": "Chinnadhurai Sankar; Sandeep Subramanian; Chris Pal; Sarath Chandar; Yoshua Bengio",
"journal": "Association for Computational Linguistics",
"ref_id": "b41",
"title": "Do neural dialog systems use the conversation history effectively? an empirical study",
"year": "2019"
},
{
"authors": "Emily Sheng; Kai-Wei Chang; Premkumar Natarajan; Nanyun Peng",
"journal": "Association for Computational Linguistics",
"ref_id": "b42",
"title": "The woman worked as a babysitter: On biases in language generation",
"year": "2019"
},
{
"authors": "Jamin Shin; Hangyeol Yu; Hyeongdon Moon; Andrea Madotto; Juneyoung Park",
"journal": "Association for Computational Linguistics",
"ref_id": "b43",
"title": "Dialogue summaries as dialogue states (DS2), template-guided summarization for few-shot dialogue state tracking",
"year": "2022"
},
{
"authors": "Yusheng Su; Xiaozhi Wang; Yujia Qin; Chi-Min Chan; Yankai Lin; Huadong Wang; Kaiyue Wen; Zhiyuan Liu; Peng Li; Juanzi Li; Lei Hou; Maosong Sun; Jie Zhou",
"journal": "Association for Computational Linguistics",
"ref_id": "b44",
"title": "On transferability of prompt tuning for natural language processing",
"year": "2022"
},
{
"authors": "Alon Talmor; Yanai Elazar; Yoav Goldberg; Jonathan Berant",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b45",
"title": "oLMpics-on what language model pre-training captures",
"year": "2020"
},
{
"authors": "Tu Vu; Brian Lester; Noah Constant; Rami Al-Rfou; ' ; Daniel Cer",
"journal": "Association for Computational Linguistics",
"ref_id": "b46",
"title": "SPoT: Better frozen model adaptation through soft prompt transfer",
"year": "2022"
},
{
"authors": "Ke Wang; Xiaojun Wan",
"journal": "Association for Computational Linguistics",
"ref_id": "b47",
"title": "TransSum: Translating aspect and sentiment embeddings for selfsupervised opinion summarization",
"year": "2021"
},
{
"authors": "Jason D Williams; Steve Young",
"journal": "Computer Speech & Language",
"ref_id": "b48",
"title": "Partially observable markov decision processes for spoken dialog systems",
"year": "2007"
},
{
"authors": "Thomas Wolf; Lysandre Debut; Victor Sanh; Julien Chaumond; Clement Delangue; Anthony Moi; Pierric Cistac; Tim Rault; Remi Louf; Morgan Funtowicz; Joe Davison; Sam Shleifer; Clara Patrick Von Platen; Yacine Ma; Julien Jernite; Canwen Plu; Teven Xu; Sylvain Le Scao; Mariama Gugger; Quentin Drame; Alexander Lhoest; Rush",
"journal": "Association for Computational Linguistics",
"ref_id": "b49",
"title": "Transformers: State-of-the-art natural language processing",
"year": "2020"
},
{
"authors": "Kaige Xie; Sarah Wiegreffe; Mark Riedl",
"journal": "Association for Computational Linguistics",
"ref_id": "b50",
"title": "Calibrating trust of multi-hop question answering systems with decompositional probes",
"year": "2022"
},
{
"authors": "Ziyi Yang; Chenguang Zhu; Robert Gmyr; Michael Zeng; Xuedong Huang; Eric Darve",
"journal": "Association for Computational Linguistics",
"ref_id": "b51",
"title": "TED: A pretrained unsupervised summarization model with theme modeling and denoising",
"year": "2020"
},
{
"authors": "Tiezheng Yu; Zihan Liu; Pascale Fung",
"journal": "Association for Computational Linguistics",
"ref_id": "b52",
"title": "AdaptSum: Towards low-resource domain adaptation for abstractive summarization",
"year": "2021"
},
{
"authors": "Lin Yuan; Zhou Yu",
"journal": "",
"ref_id": "b53",
"title": "Abstractive dialog summarization with semantic scaffolds",
"year": "2019"
},
{
"authors": "Ruifeng Yuan; Zili Wang; Ziqiang Cao; Wenjie Li",
"journal": "Association for Computational Linguistics",
"ref_id": "b54",
"title": "Few-shot query-focused summarization with prefix-merging",
"year": "2022"
},
{
"authors": "Xiaoxue Zang; Abhinav Rastogi; Srinivas Sunkara; Raghav Gupta; Jianguo Zhang; Jindong Chen",
"journal": "Association for Computational Linguistics",
"ref_id": "b55",
"title": "MultiWOZ 2.2 : A dialogue dataset with additional annotation corrections and state tracking baselines",
"year": "2020"
},
{
"authors": "Jeffrey Zhao; Mahdis Mahdieh; Ye Zhang; Yuan Cao; Yonghui Wu; ; ",
"journal": "Association for Computational Linguistics",
"ref_id": "b56",
"title": "Effective sequence-tosequence dialogue state tracking",
"year": "2021"
},
{
"authors": "Lulu Zhao; Fujia Zheng; Keqing He; Weihao Zeng; Yuejie Lei; Huixing Jiang; Wei Wu; Weiran Xu; Jun Guo; Fanyu Meng",
"journal": "",
"ref_id": "b57",
"title": "Todsum: Task-oriented dialogue summarization with state tracking",
"year": "2021"
},
{
"authors": "Lulu Zhao; Fujia Zheng; Weihao Zeng; Keqing He; Weiran Xu; Huixing Jiang; Wei Wu; Yanan Wu",
"journal": "Association for Computational Linguistics",
"ref_id": "b58",
"title": "Domain-oriented prefix-tuning: Towards efficient and generalizable fine-tuning for zero-shot dialogue summarization",
"year": "2022"
}
] | [
{
"formula_coordinates": [
3,
353.28,
663.2,
123.99,
34.74
],
"formula_id": "formula_0",
"formula_text": "P (y|x) = |y| i=1 p θ (y i |x, y <i )."
},
{
"formula_coordinates": [
4,
81.91,
311.87,
196.17,
34.74
],
"formula_id": "formula_1",
"formula_text": "max ϕ log p θ,ϕ (y|x) = |y| i=1 log p θ,ϕ (y i |x, y <i )."
},
{
"formula_coordinates": [
4,
324.69,
142.08,
199.72,
37.95
],
"formula_id": "formula_2",
"formula_text": "x i = [t i1 , t i2 , . . . , t ip i ], 1 ≤ i ≤ N ; a trained DST model LM DST ;"
},
{
"formula_coordinates": [
4,
312.26,
223.4,
213.43,
90.68
],
"formula_id": "formula_3",
"formula_text": "S = {s 1 , s 2 , . . . , s N }, a subset set(s i ) ⊆ set(x i ) for each dialogue x i ∈ X , 1 ≤ i ≤ N . 1: M = {} 2: for i = 1, 2, . . . , N do 3: o i = LM DST (x i ) 4:"
},
{
"formula_coordinates": [
4,
307.77,
318.22,
156.64,
151.28
],
"formula_id": "formula_4",
"formula_text": "o ij = LM DST (x i \\ [t ij ]) 6: m ij = Sim(o i , o ij ) 7: add m ij to M 8: S = {} 9: m median = Median(M) 10: for i = 1, 2, . . . , N do 11: s i = [ ] 12: for j = 1, 2, . . . , p i do 13: if m ij < m median then 14: append t ij to s i 15:"
},
{
"formula_coordinates": [
5,
317.53,
285.59,
197.3,
72.37
],
"formula_id": "formula_5",
"formula_text": "max ϕ log p θ,ϕ (y ⊕ s | x) = log p θ,ϕ (y|x) + log p θ,ϕ (s|x, y) = log p θ,ϕ (y|x) + q i=1 log p θ,ϕ (t s i |x, y, t s <i )."
},
{
"formula_coordinates": [
5,
306.14,
535.01,
218.27,
23.39
],
"formula_id": "formula_6",
"formula_text": "x i \\ [t ij ], 1 ≤ j ≤ p i (x i contains p i dialogue turns)."
},
{
"formula_coordinates": [
7,
132.12,
76.07,
328.18,
23.22
],
"formula_id": "formula_7",
"formula_text": "TODSUM SPNET Models R-1 R-2 R-L R-1 R-2 R-L"
},
{
"formula_coordinates": [
7,
77.19,
355.51,
438.7,
18.3
],
"formula_id": "formula_8",
"formula_text": "skeleton type R-1 R-2 R-L decoding order R-1 R-2 R-L source & target task supervision R-1 R-2 R-L SAPT ["
}
] |
2023-11-10 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b0",
"b1",
"b2",
"b3",
"b4",
"b5",
"b6",
"b7",
"b1",
"b8",
"b9",
"b10",
"b11",
"b12",
"b13",
"b14",
"b15",
"b16"
],
"table_ref": [],
"text": "Text-to-image generative models (or called text-to-image models for short)-e.g., Stable Diffusion [1], DALL•E [2], and Imagen [3]-are popular due to the invention and deployment of diffusion models [4], [5] and large-scale language models [6], [7], [8]. Such text-to-image modelswhich generate a synthetic image based on a given text prompt-have broad applications such as graphic design and virtual environment creation. For example, Microsoft has embedded DALL•E [2] into an application named Designer [9] and an image creator tool as part of Microsoft Edge; in addition, Stable Diffusion has been used by more than 10 million people daily up to February 2023. Yet, one practical ethical concern facing text-to-image models is that they may generate sensitive Not-Safe-for-Work (NSFW) images [10], [11] such as those related to violence and child-inappropriate. Therefore, existing text-to-image models all adopt so-called safety filters as guardrails to block the generation of such NSFW images. However, the robustness of such safety filters-especially those used in practice-to adversarial manipulations of prompts is still unknown.\nOne intuitive method for jailbreaking safety filters is to treat them as closed-boxes and launch text-based adversarial attacks like TextBugger [12], Textfooler [13], BAE [14], and a concurrent work called adversarial prompting [15] to perturb prompts. However, existing text-based attacks focus on misleading a classification model but not bypassing safety filters with NSFW generations. For example, none of the aforementioned approaches is able to bypass the closed-box safety filter of DALL•E 2 according to our evaluation. There are three reasons that text-based adversarial attacks are insufficient for bypassing safety filters. First, they are inefficient at probing a safety filter, resulting in a large number of queries to a text-to-image model and thus increasing the cost for an attacker. Second, although the one-time bypass rate may be high, the bypass rate becomes low when the adversarial texts are reused for generating NSFW images because the safety filter is not considered when finding the adversarial texts and is still effective during reuse attacks. Lastly, existing works focus less on the quality of generated images, often resulting in images losing the intended NSFW semantics.\nTwo recent works studied the safety filters of text-toimage models. Specifically, Rando et al. [16] reverse engineer Stable Diffusion's safety filter and then propose a manual bypass strategy that adds extra unrelated text to a prompt. Another concurrent work-Qu et al. [17]-manually gathers a template NSFW prompt dataset to evaluate safety filters of open-source text-to-image models, e.g., Stable Diffusion. However, the generation of adversarial prompts to bypass safety filters is largely manual, which often results in a low bypass rate. For example, similar to text-based adversarial attacks, neither approach is able to bypass the closed-box safety filter of DALL•E 2 according to our evaluation.\nIn this paper, we propose the first automated attack framework, called SneakyPrompt, to jailbreak safety filters of textto-image models. Our key insight is to search for alternative tokens to replace the filtered ones in a given NSFW prompt while still preserving the semantics of the prompt and the follow-up generated NSFW images. Intuitive approaches will be brute force, beam, or greedy searches, but they are often cost-ineffective, e.g., incurring many queries to the target textto-image model. Therefore, these intuitive approaches are treated as baselines of SneakyPrompt. Our high-level idea is to leverage reinforcement learning (RL), which interacts with the target text-to-image model and perturbs the prompt based on rewards related to two conditions: (i) semantic similarity, and (ii) success in bypassing safety filters. Such an RL-based approach not only solves the challenge of closed-box access to the text-to-image model but also minimizes the number of queries as the reward function can guide SneakyPrompt to find adversarial prompts efficiently.\nTo summarize, we make the following contributions."
},
{
"figure_ref": [],
"heading": "•",
"publication_ref": [
"b1",
"b0"
],
"table_ref": [],
"text": "We design and implement SneakyPrompt to jailbreak safety filters of text-to-image models using different search strategies including reinforcement learning and baselines such as beam, greedy, and brute force.\n•\nWe show that SneakyPrompt successfully finds adversarial prompts that allow a text-to-image model with a closed-box safety filter-namely DALL•E 2 [2]-to generate NSFW images.\n•\nWe extensively evaluate SneakyPrompt on a large variety of open-source safety filters with another state-of-the-art text-to-image model-namely Stable Diffusion [1]. Our evaluation results show that SneakyPrompt not only successfully bypasses those safety filters, but also outperforms existing text-based adversarial attacks.\nEthical Considerations. We responsibly disclosed our findings to DALL•E (specifically OpenAI via their online portal and an email) and Stable Diffusion (specifically Stability AI via a Zoom discussion). We did not receive a response from OpenAI, but Stability AI would like to develop more robust safety filters together with us. We also discussed our work with the Institutional Review Board (IRB) and obtained an exempt decision."
},
{
"figure_ref": [],
"heading": "Related Work and Preliminary",
"publication_ref": [
"b17",
"b18",
"b19",
"b20",
"b3",
"b4",
"b0",
"b1",
"b2",
"b21",
"b5",
"b22",
"b23",
"b24",
"b25",
"b26",
"b27",
"b28",
"b29",
"b14",
"b15",
"b16"
],
"table_ref": [],
"text": "In this section, we describe related work on text-to-image models including existing attacks on such models, present existing adversarial attacks on learning models, especially text-based ones, and then illustrate some preliminaries such as reinforcement learning (RL) and the challenges in applying RL upon SneakyPrompt. Text-to-image Models. Text-to-image models-which have been firstly demonstrated by Mansimov et al. [18]-generate images based on a textual description denoted as a prompt. Later on, different works have focused either on model structure [19], [20] or learning algorithm [21] to optimize image quality. Modern text-to-image approaches often adopt diffusion models [4], [5], where the image begins with random noises and noises are progressively removed using a de-noising network. Examples include Stable Diffusion [1], DALL•E [2], Imagen [3], and Midjourney [22]. More specifically, such text-to-image models are often text-conditioned, which adopt text embedding of a prompt from a frozen text encoder, e.g., CLIP [6], to guide image generation; some recent works have also proposed learning free [23] or zeroshot image generation [24] for large-scale generative models.\nGiven their popularity, many works have been proposed to investigate vulnerabilities of text-to-image models. Wu et al. [25] and Duan et al. [26] demonstrate the feasibility of membership inference attack [27], [28] on text-to-image models. Carlini et al. [29] propose an image extracting attack to illustrate the possibility of extracting training samples used to train the text-to-image model. Millière et al. [30] demonstrate that attackers can find adversarial examples that combine words from different languages against text-toimage models. Maus et al. [15] also propose the concept of adversarial prompt and design a black-box framework based on Bayesian optimization for such a prompt generation. Note that on one hand, the definition of the adversarial prompt in Maus et al. is concurrent to our approach, and on the other hand, Maus et al. cannot bypass safety filters as shown in our evaluation because their goal is to generate the target class of images using meaningless tokens without the presence of any safety filters. The closest works are Rando et al. [16] and Qu et al. [17], which investigate safety filters of text-to-image models. However, their approaches are largely manual with relatively low bypass rates and they are only applicable to offline text-to-image models."
},
{
"figure_ref": [],
"heading": "Adversarial Examples.",
"publication_ref": [
"b30",
"b31",
"b32",
"b33",
"b34",
"b11",
"b12",
"b13",
"b35"
],
"table_ref": [],
"text": "Adversarial examples are carefully crafted inputs to confuse a learning model for an incorrect decision, e.g., wrong classification results. Extensive numbers of research [31], [32], [33] are proposed on the generation of adversarial examples in computer vision. People have also studied adversarial examples in the natural language processing (NLP) domain. There are generally two directions on adversarial text examples. First, people propose to ensure that the perturbed word looks similar to the original input, e.g., \"nice\" vs \"n1ce\". For example, recent work [34] adopts Gumble-softmax distribution to approximate the discrete categorical distribution for text-based adversarial examples. Second, people also propose using synonyms to paraphrase the input, keep the original semantics, and change the final prediction. Alzantot et al. [35] propose heuristic methods to search for replacement words with similar semantic meanings. TextBugger [12] shows that manipulation of important words, e.g., swapping, removing, and substituting, can lead to alternation of the predictions of sentences with little impact on human understanding in both closed-box and open-box settings. Jin et al. [13] propose rule-based synonym replacement strategies to generate more naturallooking adversarial examples and improve semantic similarity to the original token under the acceptance of human judges. Garg et al. [14] propose to mask a portion of the text and use BERT masked language model to generate replacement with grammatical improvement and semantic coherence.\nExisting approaches to adversarial examples can be applied to text-to-image models with safety filters as well. However, since they are not designed for bypassing safety filters, they face three major issues that we also show in our evaluation. First, existing approaches do not preserve the semantics of the generated images, i.e., the NSFW semantics may have been lost during the generation. Second, existing approaches may not be cost-effective, i.e., they may incur a significant number of queries to the text-to-image model. Third, the adversarial prompts generated by existing approaches may not be reusable due to random seeds adopted by text-to-image models. That is, those adversarial prompts may be effective one time, but lose effectiveness if used for more than one time. Reinforcement Learning (RL). RL [36] is a technique to incorporate feedback to make decisions. The key concepts in RL include state, action, policy network, reward, and environment. Given a state, the policy network essentially outputs a distribution over the possible actions. One action is sampled from the distribution and applied to the environment, which returns a reward. The reward can then be used to update the policy network such that it is more likely to generate actions with a large accumulative reward in the future. Note that the deployment of RL to search for an adversarial prompt is challenging because SneakyPrompt needs to not only decide the action space for adversarial prompts, which is a large word space, but also design a reward function to bypass the safety filter while still preserving the generated images' NSFW semantics."
},
{
"figure_ref": [],
"heading": "Problem Formulation",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we first define adversarial prompt against safety filters of text-to-image models and then describe the threat model of SneakyPrompt."
},
{
"figure_ref": [],
"heading": "Definitions",
"publication_ref": [
"b1",
"b21",
"b0",
"b15"
],
"table_ref": [],
"text": "We describe the definitions of two important concepts: safety filters and adversarial prompts. Safety Filter. A safety filter-formally denoted as Fprohibits text-to-image model users from generating certain images with so-called sensitive content, such as those related to adult, violent, or politics. The deployments of safety filters are common practices used by existing text-to-image models. For example, DALL•E 2 [2] filters out contents from 11 categories such as hate, harassment, sexual, and self-harm. Midjourney [22] blocks the generation of images that are not PG-13. Stable Diffusion [1] also filters out contents from 17 concepts [16].\nTo the best of our knowledge, there is no existing documentation on the taxonomy of safety filters used in text-to-image models. Therefore, we come up with our own taxonomy and describe them below. Note that we denote the online text-to-image model as M with a frozen text encoder E and a diffusion model D, the input prompt as p, and the output generated image as M(p). Figure 1 shows the three categories of safety filters: Let us describe the definition from two aspects. First, the adversarial prompt is a relative concept. That is, p a is adversarial relatively to another sensitive, target prompt p t , which is originally blocked by the safety filter of a textto-image model. Second, there are two conditions for an adversarial prompt p a : (i) p a bypasses the safety filter F, and (ii) the generated image from p a is semantically similar to that generated from p t . Both conditions are important, i.e., even if the bypass is successful but the generated image loses the semantics, p a is not an adversarial prompt.\nFigure 2 shows some simple examples of adversarial prompts generated by SneakyPrompt to illustrate what they look like. The text in the parenthesis is p t , which is blocked by an external safety filter (blocking both dogs and cats) added after DALL•E 2 for illustration purposes. growled menacingly at the stranger who approached its owner Figure 2: Examples of adversarial prompts that generate cats and dogs (the images above the prompts) using DALL•E 2 and bypass an external safety filter, i.e., the default stable diffusion safety filter refactored to restrict both concepts. The target, sensitive prompt is highlighted in red and its corresponding adversarial prompt is in blue. Black texts are unchanged between target and adversarial prompts. Note that we use dogs and cats as part of the external safety filters in the illustrative figure to avoid illegitimate or violent content that might make the audience uncomfortable. We show real images with NSFW content that bypass the DALLE•2's safety filter in Appendix A due to the concerns of possible disturbing content to readers.\nThe adversarial prompts are shown in blue together with the black texts. The above images are generated by DALL•E 2, which still preserves the semantics of either dogs or cats."
},
{
"figure_ref": [],
"heading": "Threat Model",
"publication_ref": [
"b36",
"b0",
"b37",
"b15",
"b16"
],
"table_ref": [],
"text": "We assume that an adversary has closed-box access to an online text-to-image model and may query the model with prompts. Since modern text-to-image models often charge users per query [37], we assume the adversary has a certain cost constraint, i.e., the number of queries to the target textto-image model is bounded. In addition, the adversary has access to a local shadow text encoder Ê. We describe the details of the closed-box access and the shadow text encoder as follows: \nÊ(p) = E(p):\nThat is, the adversary may adopt a Ê with exactly the same architecture and parameters as E. For example, Stable Diffusion [1] utilizes a public CLIP text encoder (i.e., ViT-L/14 [38]), which can be deployed locally for shadow access.\nAttack Scenarios. Next, we describe two realistic attack scenarios that are considered in the paper.\n• One-time attack: The adversary searches adversarial prompts for one-time use. Each time the adversary obtains new adversarial prompts via search and generates corresponding NSFW images. • Re-use attack: The adversary obtains adversarial prompts generated by other adversaries or by themselves in previous one-time attacks, and then re-uses the provided adversarial prompts for NSFW images. We consider re-use attacks as the default use scenario just like existing works [16], [17] where they all provide prompts for future uses. The main reason is that reuse attacks do not need to repeatedly query the target model and thus save query costs. At the same time, one-time attacks are also evaluated in comparison with prior works."
},
{
"figure_ref": [],
"heading": "SneakyPrompt",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we give an overview of SneakyPrompt and then propose different variants of search methods, including three heuristic searches as a baseline SneakyPrompt-base and a reinforcement learning based search as an advanced approach SneakyPrompt-RL."
},
{
"figure_ref": [
"fig_2"
],
"heading": "Overview",
"publication_ref": [],
"table_ref": [],
"text": "Key Idea. We first give an intuitive explanation of why SneakyPrompt can bypass safety filters to generate NSFW images in Figure 3. A safety filter-no matter whether text-, image-, or text-image-based-can be considered as a binary (i.e., sensitive or non-sensitive) classifier with a decision boundary in the text embedding space. Moreover, suppose prompts with similar NSFW semantics form a ball in the text embedding space, which has intersections with the decision boundary.\nThe intuition of our SneakyPrompt is to search for an adversarial prompt whose generated image not only has semantics similar enough to the target prompt but also crosses the decision boundary of the safety filter. For example, the prompt 'mambo incomplete clicking' is one adversarial prompt relative to the sensitive, target prompt 'naked'; 'nude' is one sensitive prompt with a similar semantic of 'naked' that is blocked by the safety filter; and 'happy' is one nonsensitive prompt with a dissimilar semantic of 'naked'.\nWe then formalize the key idea of SneakyPrompt, which, given a target prompt p t , aims to search for an adversarial prompt p a to a text-to-image model M that satisfies the following three objectives:\n• Objective I: Searching for a prompt with target semantic.\nThat is, M(p a ) has the same sensitive semantics as the target prompt p t . • Objective II: Bypassing the safety filter. That is, p a bypasses the safety filter F, i.e., F(M, p a ) = 0. • Objective III: Minimizing the number of online queries.\nThat is, the number of queries to M is minimized.\nTo achieve Objective I, SneakyPrompt finds an adversarial prompt p a such that the similarity (e.g., cosine similarity in our experiments) between the image embedding of the generated image M(p a ) and the text embedding Ê(p t ) of the target prompt p t is large enough. To achieve Objective II, SneakyPrompt repeatedly queries the target text-to-image model until finding an adversarial prompt p a that bypasses the safety filter. To achieve Objective III, SneakyPrompt leverages reinforcement learning to strategically perturb the prompt based on query results. Overall Pipeline. Figure 4 describes the overall pipeline of SneakyPrompt in searching for an adversarial prompt for a target, sensitive prompt p t with six major steps. Given a target prompt p t , SneakyPrompt first finds the n sensitive tokens in it via matching with a predefined list of NSFW words, or if none matches, using a text NSFW classifier to choose the n tokens with the highest probabilities of being NSFW. The key idea of SneakyPrompt is to replace each sensitive token in p t as m non-sensitive tokens (called replacing tokens) to construct an adversarial prompt p a . In total, we have nm replacing tokens. Suppose D is the token vocabulary, e.g., in our experiments, we use the CLIP token vocabulary which has 49,408 tokens. A straightforward way is to search each replacing token from D. However, this is very inefficient as the size of D is very large. To address the challenge, we reduce the search space of each replacing token to D l which only includes the tokens in D whose lengths are at most l. Formally, our overall search space S of the nm replacing tokens can be defined as follows:\nS = {(c 1 , c 2 , • • • , cnm)|c j ∈ D l , ∀j = 1, 2, • • • , nm},(1)\nwhere c j is a replacing token. Next, we describe our six steps.\n• Step (1): 2) and (3) if the safety filter is not bypassed. ( 5) GetSimilarity(M(p a ), Ê(p t )) calculates the normalized cosine similarity between the image embedding of the generated image M(p a ) and the text embedding of p t . (6) Repeating Steps ( 2)-( 5) if the similarity does not meet the threshold δ."
},
{
"figure_ref": [],
"heading": "Baseline Search with Heuristics",
"publication_ref": [],
"table_ref": [],
"text": "SneakyPrompt-base adopts one of the following three heuristics as the function Sample:\n• BruteForce: In this baseline, the function Sample samples each replacing token c j from D l uniformly at random, where j = 1, 2 "
},
{
"figure_ref": [],
"heading": "Guided Search via Reinforcement Learning",
"publication_ref": [
"b32",
"b39"
],
"table_ref": [],
"text": "Since the baseline approaches are cost-ineffective, we design a guided search version, called SneakyPrompt-RL, using reinforcement learning (RL) to search for an adversarial prompt. Roughly speaking, the function Sample uses a policy network to sample the replacing tokens\nC = (c 1 , c 2 , • • • , c nm ).\nThe sampled replacing tokens C can be viewed as an action in the action/search space S, the resulting adversarial prompt p a can be viewed as a state, and the text-to-image model M can be viewed as the environment in RL. When the action C is applied to the environment (i.e., the corresponding adversarial prompt p a is used to query M), the policy network receives a reward, which is then used to update the policy network. Next, we describe our policy network, reward, and loss function used to update the policy network. Policy Network. A policy network P defines a probability distribution of actions in the action/search space S. We denote by P (C) the probability of the action\nC = (c 1 , c 2 , • • • , c nm ). Moreover, we assume P (C) = P (c 1 ) nm j=2 P (c j |c 1 , c 2 , • • • , c j-1\n), which allows us to efficiently sample the nm replacing tokens one by one using P . Specifically, we sample c 1 based on P (c 1 ); given the sampled c 1 , we sample c 2 based on P (c 2 |c 1 ); and this process is repeated until c nm is sampled. The sampled C is then used together with the target prompt p t to construct an adversarial prompt p a . Following previous work [33], [40], we use an LSTM with a fully connected layer as a policy network P . if r q-4 , r q-3 , r q-2 , r q-1 , rq < 0 then 33:\n//Expand the search space by replacing one more token in pt 34:\nS, ω ← GetSearchSpace(Initial = 0) 35:\nend if 36:\n//Rewards do not change in 3 consecutive queries 37:\n//or fraction ω of tokens in pt to be replaced is no smaller than 0.3 38:\nif |r q-2 + rq -2r q-1 | <1e-4 or ω ≥ 0.3 then 39:\nreturn p ′ a and M(p ′ a ) 40:\nend if 41:\nq ← q + 1 42: end while 43: return p ′ a and M(p ′ a )\nReward. Intuitively, if the adversarial prompt p a based on the sampled replacing tokens C bypasses the safety filter, we should assign a reward, with which the policy network can be updated to increase the GetSimilarity(M(p a ), Ê(p t )) such that the next generated adversarial prompt is likely to have a larger GetSimilarity. If p a does not bypass the safety filter, we assign a negative reward, which aims to update the policy network such that it is less likely to sample C. Moreover, the reward is smaller to penalize C more if more queries have been sent to the text-to-image model M. Based on such intuitions, we define a reward r q for the adversarial prompt p a in the qth query to the target model as follows: \nr q = GetSimilarity(M(p a ), Ê(p t )) if F(M, p a ) = 0 -q/(10 • Q) if F(M, p a ) = 1 ,(2)\nS = {(c 1 , c 2 , • • • , cnm)|c j ∈ D l , ∀j = 1, 2, • • • , nm} 30: L ← Number of tokens in pt 31: ω ← n/L 32: return S and ω\nwhere Q is the maximum number of queries SneakyPrompt can send to the target model M. Updating Policy Network. Intuitively, if the reward r q is smaller, the policy network should be less likely to sample C. Based on such intuition, we use the following loss function to update P : loss = -r q • ln(P (C)).\n(\nWe update P using one iteration of gradient descent with a learning rate η. Two Optimization Strategies. We propose two strategies to further optimize the effectiveness and efficiency of SneakyPrompt-RL.\n• Strategy One: Search Space Expansion. Recall that we start by replacing n sensitive tokens in the target prompt p t . If the generated adversarial prompts did not bypass the safety filter in multiple (e.g., 5 in our experiments) consecutive queries, we add one more token in the target prompt p t to be replaced by m tokens. In other words, we increase the action/search space for the policy network. Such an expansion strategy not only increases the bypass rate but also decreases the number of queries. • Strategy Two: Early Stop. We have three criteria to stop the search early. (i) The search stops early if the GetSimilarity(M(p a ), Ê(p t )) ≥ δ, which indicates a high-quality NSFW image has been generated. (ii) The Alternative Reward Function with Offline Queries. We also consider an alternative reward function for SneakyPrompt-RL, which only requires offline queries to the shadow text encoder. This alternative reward function can further reduce the number of queries to the target text-to-image model, though the generated image has reduced quality. In particular, we consider GetSimilarity = 1-ℓ 2 ( Ê(p a ), Ê(p t )) for an adversarial prommpt p a , where ℓ 2 is the Euclidea distance between two text embeddings. Note that we also normalize the similarity scores GetSimilarity to be [0, 1], so it is easier to set the threshold δ. In each query to the target model, we sample replacing tokens C using the policy network and construct an adversarial prompt p a based on C and the target prompt p t . Instead of using p a to query the target model immediately, we calculate the alternative reward using GetSimilarity locally and update the policy network using the alternative reward. If the alternative reward is smaller than δ, we repeat the sampling and policy-network-updating process until construct an adversarial prompt whose alternative reward is no smaller than δ. Then, we use the adversarial prompt to query the target model. If the adversarial prompt bypasses the safety filter, the search process stops, and the adversarial prompt and generated image are returned. Otherwise, a negative reward is used to update the policy network and the process is repeated. More details can be found in Algorithm 3 in Appendix."
},
{
"figure_ref": [],
"heading": "Experimental Setup",
"publication_ref": [
"b41",
"b42",
"b43",
"b48",
"b49",
"b9"
],
"table_ref": [
"tab_6"
],
"text": "We implement SneakyPrompt using Python 3.9 with Pytorch. All experiments are performed using two GeForce RTX 3090 graphics cards (NVIDIA). Our target text-toimage models include (i) Stable Diffusion with the open source model on Hugging Face [42] and (ii) DALL•E 2 with the official online API provided by OpenAI [43]. The default target model is Stable Diffusion. We also show the detailed hyper-parameters used by SneakyPrompt in Table 1. Our default SneakyPrompt is SneakyPrompt-RL with GetSimilarity = cos(M(p a ), Ê(p t )) unless otherwise mentioned. We now describe the experimental setup details. Prompt Datasets. We generated two prompt datasets for evaluating safety filters using ChatGPT with GPT-3.5.\n• NSFW-200 dataset. We followed a post on Reddit [44] to generate 200 target prompts with NSFW content using ChatGPT with GPT-3.5. • Dog/Cat-100 dataset. We used ChatGPT with GPT-3.5\nto generate 100 prompts describing the scenario with dogs or cats. The purpose of the dataset is to demonstrate the feasibility of SneakyPrompt in bypassing safety filters while avoiding NSFW content that potentially makes people uncomfortable. Safety Filters. Our evaluation involves the following seven different safety filters that cover all categories in Figure 1 as well as an unknown category. Evaluation Metrics. We adopt three evaluation metrics.\n• Bypass rate: For one-time attacks, we compute our bypass rate as the number of adversarial prompts that bypass a safety filter divided by the total number of adversarial prompts. An adversary would only re-use the adversarial prompts that successfully bypass safety filters in one-time attacks. Therefore, for re-use attacks, the bypass rate is the fraction of re-uses that bypass a safety filter for successful one-time adversarial prompts. • FID score: We use the FID [49] score to evaluate the image semantic similarity of our generation. We follow the official implementation [50] of Pytorch in calculating FID between our generation with three ground-truth datasets as the reference. (i) target: this dataset contains 1,000 images generated by NSFW-200 with different random seeds from Stable Diffusion without the presence of the safety filter;\n(ii) real: this dataset contains 40,000 real sensitive images from the NSFW image dataset [10]. (iii) target-dog/cat: this dataset contains 1,000 images generated by Dog/Cat-100 with different random seeds from Stable Diffusion. The higher the FID score is, the less similar the two images' distributions are in semantics.\n• Number of online queries: The number of queries to textto-image models used for searching for an adversarial prompt. Note that this metric is not evaluated for reuse attacks, because no additional queries in generating adversarial prompts are needed."
},
{
"figure_ref": [],
"heading": "Evaluation",
"publication_ref": [],
"table_ref": [],
"text": "We answer the following Research Questions (RQs). "
},
{
"figure_ref": [],
"heading": "RQ1: Effectiveness at Bypassing Safety Filters",
"publication_ref": [
"b15",
"b16",
"b11",
"b12",
"b13",
"b15",
"b50",
"b51"
],
"table_ref": [
"tab_8",
"tab_9",
"tab_8"
],
"text": "In this research question, we evaluate how effective SneakyPrompt is at bypassing existing safety filters. Some real adversarial prompts are shown in Appendix A. Overall Results. Table 2 shows the quantitative results of SneakyPrompt in bypassing existing safety filters. Table 3 shows examples of target and adversarial prompts generated by SneakyPrompt. In general, SneakyPrompt effectively bypasses all the safety filters to generate images with similar semantics to target prompts with a small number (below 25) of queries. Let us start with six safety filters on Stable Diffusion. SneakyPrompt achieves an average 96.37% onetime bypass rate (with 100.00% on four of them) and an average of 14.68 queries (with at least 2.26 queries), with a reasonable FID score indicating image semantic similarity. The bypass rate drops and the FID score increases for re-use Table 4: [RQ2] Performance of SneakyPrompt-RL compared with different baselines in bypassing Stable Diffusion with its original safety filter. We use the prompt examples provided by both Rando et al. [16] and Qu et al. [17] five times for re-use performance. Note that these prompts are pre-created manually, thus not being applicable in the one-time searched scenario. Maus et al. cannot generate any NSFW images after 5,000 queries and therefore we do not report FID scores. attacks because the diffusion model adopts random seeds in generating images. The only exception is the bypass rate for text-based safety filters because such filters are deployed before the diffusion model.\nWe then describe the closed-box safety filter of DALLE• 2. SneakyPrompt achieves 57.15% one-time bypass rate with an average of 24.49 queries. Note that while the rate seems relatively low, this is the first work that bypasses the closedbox safety filter of DALL•E 2 with real-world sensitive images as shown in Appendix A. None of the existing works, including text-based adversarial examples [12], [13], [14] and Rando et al. [16], is able to bypass this closed-box safety filter. Interestingly, the re-use bypass rate for DALLE• 2 is 100%, i.e., as long as an adversarial prompt bypass the filter once, it will bypass the filter multiple times with NSFW images generated.\nWe now describe our observations related to the robustness of safety filters below.\nScale of Safety Filter. Table 2 shows that a safety filter's robustness (especially against SneakyPrompt) is positively correlated with its scale, i.e., the total number of parameters. This observation holds for both one-time and re-use performance on all metrics. For example, SneakyPrompt performs the worst against text-classifier because its scale is much larger than other variants. As a comparison, SneakyPrompt achieves the least number of queries and the highest image se-mantic similarity against text-match and text-image-threshold (which are not learning-based filters). Type of Safety Filter. We have two observations with regard to the safety filter type when safety filters have a similar number of parameters. First, the combination of text and image is better than those relying on any single factor. For example, text-image-threshold outperforms text-match in terms of all metrics. In addition, image-clip-classifier outperforms image-classifier with a small number of online queries, because image-clip-classifier utilizes the embedding from CLIP (which includes both image and text information) as opposed to only image information in image-classifier.\nSecond, image-based safety filters have a lower re-use bypass rate compared to text-based ones. It is because the random seeds used for the text-to-image model to generate images are uncontrollable, which leads to different generated images using the same prompt at different times. Therefore, the one-time bypassed adversarial prompt may not bypass the image-based safety filter during re-use. As a comparison, because the text-based safety filter takes a prompt as input, the generated adversarial prompt can bypass the same safety filter in re-use as long as the filter is not updated. Non-English Word Safety Filter. We add a simple non-English word safety filter in combination with existing filters and show the effectiveness of SneakyPrompt in bypassing this adaptive safety filter. Note that the search space of SneakyPrompt will be the google-10000-english dictionary [51] that contains a list of the 10,000 most common English words, instead of all tokens from CLIP vocabulary dictionary [52].\nWe have three observations. First, the FID score is on par with SneakyPrompt-RL against existing safety filters alone, which indicates SneakyPrompt-RL is also effective in maintaining image semantics when a non-English word filter is present. Second, the bypass rate is 3.75% on average lower than that without a non-English word filter. The reason is that SneakyPrompt might select synonyms for bypassing once but cannot be reused as shown in existing text adversarial examples. Third, the number of online queries is 3.11 on average higher since the search space is limited to English words; thus, more queries are needed to find an adversarial prompt that can bypass and maintain the image semantics.\n[RQ1] Take-away: SneakyPrompt successfully bypasses all safety filters including the closed-box safety filter adopted by DALL•E 2 as well as a non-English word safety filter."
},
{
"figure_ref": [],
"heading": "RQ2: Performance Comparison with Baselines",
"publication_ref": [
"b11",
"b12",
"b13",
"b52",
"b15",
"b16",
"b14",
"b12",
"b11",
"b11",
"b11",
"b15",
"b16"
],
"table_ref": [
"tab_11",
"tab_12"
],
"text": "In this research question, we first compare SneakyPrompt with existing methods using the original safety filter of Stable Diffusion and then add the non-English word filter. We then show none of the existing methods can effectively bypass the DALL•E 2's safety filter. Specifically, we compare SneakyPrompt with the following works:\n• Text-based adversarial examples. We use three related works that generate closed-box, text-based adversarial examples, which are Textbugger [12], Textfooler [13],\nand BAE [14]. We follow the implementation by TextAttack [53] with their default hyperparameters. • Manually-curated adversarial prompts. We use Rando et al. [16] (which contains manually-generated prompts) and Qu et al. [17] (in which prompts are manually created based on a template). • Optimized adversarial prompts. We use Maus et al. [15], a concurrent work to ours, to generate adversarial prompts. • SneakyPrompt baselines with different search algorithms. We call them SneakyPrompt-base. We start from the original safety filter of Stable Diffusion. Table 4 shows the comparison results in terms of three metrics and we describe two different scenarios below.\n• Re-use adversarial prompts. This attack scenario assumes that adversarial prompts are pre-generated and re-used for the attack. On one hand, SneakyPrompt-RL achieves the highest bypass rate against the safety filter compared with all existing works and SneakyPrompt-base. The reason is that existing works, particularly text-based adversarial examples, either make minimal modifications to texts or use synonyms to replace the target token, which cannot sustain different rounds for another random seed from the diffusion model. On the other hand, SneakyPrompt-RL also has the lowest FID score compared to other methods for re-use adversarial prompts. That is, SneakyPrompt largely keeps the original semantics while existing methods-even if successfully bypassing the safety filter-will more or less modify the semantics compared to the target one. This is because SneakyPrompt computes the similarity between the generated image and target prompt to prevent the image semantics from being modified too much and thus make it as consistent with the target prompt as possible. First, the reason is that baseline methods do not have any constraints for word choice. For example, the number of queries of brute force search is a magnitude larger than other heuristic searching methods as all tokens have the same probability of being chosen. Second, SneakyPrompt-RL takes the least number of queries, using 50% fewer queries compared with the second, which is TextFooler [13]. The reason is that SneakyPrompt-RL adopts an early stop strategy and benefits from RL. Note that manual prompts are not applicable here for one-time searched prompts because it is not scalable to probe text-to-image models for each prompt manually. Next, we also add non-English word filter to the default stable diffusion's safety filter and compare SneakyPrompt-RL with text-based adversarial examples. Table 5 shows the re-use and one-time bypass rates. The bypass rate of SneakyPrompt-RL does not change much because SneakyPrompt-RL can search in an English word space. Instead, the one-time performance of TextBugger [12], which utilizes alphabet swap or substitute, drops to only 36.38%, and the re-use performance drops to only 18.42%.\nLast, we also evaluate the online, closed-box safety filter of DALL•E 2. Table 6 shows that none of the existing works can effectively bypass the original safety filter. The bypass rates of all the works except for TextBugger [12] are essentially zero and the bypass rate of TextBugger [12] is also as low as 1.00%.\n[RQ2] Take-away: SneakyPrompt-RL outperforms SneakyPrompt-base, existing text-based adversarial examples, and manual prompts from Rando et al. [16] and Qu et al. [17]."
},
{
"figure_ref": [
"fig_4",
"fig_4",
"fig_4",
"fig_4",
"fig_9",
"fig_9",
"fig_4"
],
"heading": "RQ3: Study of Different Parameter Selection",
"publication_ref": [
"b2",
"b4",
"b9",
"b14",
"b19"
],
"table_ref": [
"tab_14",
"tab_15"
],
"text": "In this research question, we study how different parameters affect the overall performance of SneakyPrompt. Reward Function. We have two variants of the reward function, one based on cosine similarity between embedding of the generated image and embedding of the target prompt, and the other based on ℓ 2 distance between embeddings of the target and adversarial prompts. Table 7 shows the comparison between the two reward functions: cosine similarity results in a higher bypass rate and a better image semantic similarity for both one-time and re-use attacks, while ℓ 2 distance results in a smaller number of online queries.\nShadow Text Encoder. Table 8 shows the impact of different shadow text encoders Ê, particularly when Ê = E and Ê ̸ = E. We have two observations. First, Ê = E improves the image semantic similarity with a smaller FID score for both one-time and re-use attacks and especially for one-time. The reason is the same text encoder adopted by the attacker results in the same text embedding as the internal result of text-to-image models, used to guide the semantics of image generation, which provides a more precise semantic compared with a different text encoder. The relatively smaller improvement for re-use FID score is because of the disturbance of the random seed. Second, there is no significant difference in bypass rates and the number of online queries. SneakyPrompt-RL achieves a 100% bypass rate with both shadow text encoders for the one-time performance and achieves around 70% for re-use performance with different random seeds involved. The number of queries is also similar because they use the same similarity threshold δ.\nSimilarity Threshold. Figure 5 shows the impact of different similarity threshold values, ranging from 0.22 to 0.30, on SneakyPrompt's performance using three metrics. Let us start with the bypass rate in Figure 5a. The bypass rate stays the same for one-time attacks but drops for re-use attacks because the generated image may be closer to the target and thus blocked by the safety filter. This is also reflected in Figure 5b as the FID score decreases as the threshold increases. Similarly, Figure 5c shows that the number of queries increases as the similarity threshold because it will be harder to satisfy the threshold during searching.\nSearch Space. We evaluate the impact of search space size-which is partially controlled by the parameter l-on the performance of SneakyPrompt. Specifically, we change l to be [3,5,10,15,20], which lead to a search space of [625 m , 6922 m , 29127 m , 46168 m , 48095 m ] according to the number of candidate tokens in the dictionary D l . Figure 6 shows the results of the impacts of l on three metrics. First, Figure 6a shows that a larger l leads to a higher re-use bypass rate. The reasons are twofold. On one hand, the larger the l, the larger the search space is. That is, RL has room to explore more tokens and increase the bypass rate. On the other hand, a larger l introduces longer tokens that dilute the target prompt more, thus also increasing the bypass rate. Second, the image semantic similarity has little correlation with the search space. Instead, image semantic similarity in terms of FID scores are more related to the semantic similarity threshold δ as we show in Figure 5. Third, the larger the l is, the more online queries SneakyPrompt takes. The reason is that RL needs more queries to explore a larger search space to satisfy the semantic similarity threshold.\n[RQ3] Take-away: SneakyPrompt needs to balance the bypass rate and the FID score with the number of queries in selecting different parameters."
},
{
"figure_ref": [
"fig_2"
],
"heading": "RQ4: Explanation of Bypassing",
"publication_ref": [],
"table_ref": [
"tab_16"
],
"text": "In this research question, we explain why SneakyPrompt successfully bypasses different safety filters while maintaining the image semantic similarities. Specifically, we use the ground-truth output probability of different safety filters for the explanation. Table 9 shows the experiment results that align with our high-level explanation in Figure 3. We first show the results of the output probability on NSFW of different safety filters for both the target and the adversarial prompts or their generated images, depending on the type of safety filters. We normalize the probability into [0,1], where a value larger than 0.5 indicates the sensitive input. The probability output of the target prompt p t is from 0.546 to 1.000, i.e., the target prompt is sensitive. As a comparison, the probability outputs of adversarial prompts generated by SneakyPrompt range from 0.000 to 0.482, i.e., they are non-sensitive. This result suggests that SneakyPrompt is effective in bypassing different safety filters. Next, we also Search space parameter l (space small to large) Search space parameter l (space small to large) Search space parameter l (space small to large) show the semantic similarity between the target prompt and its generated image, and that between the target prompt and the image generated by adversarial prompts found by SneakyPrompt. We observe that the former, i.e., 0.298, is close to the latter, i.e., 0.267 to 0.289, for different safety filters, which indicates the ability of SneakyPrompt to maintain the semantics of target prompts and images generated based on them.\n[RQ4] Take-away: The outputs from safety filters show that SneakyPrompt bypasses them while still maintaining the NSFW semantics."
},
{
"figure_ref": [],
"heading": "Conclusion, Discussion, and Future Work",
"publication_ref": [
"b53"
],
"table_ref": [],
"text": "We show that a black-box safety filter of a text-to-image model can be jailbroken to produce an NSFW image with a small number of queries to the model. Reinforcement learning can reduce the number of queries to the text-toimage model by leveraging the query results to strategically guide the perturbation of a prompt. Our results imply that existing guardrails of text-to-image models are insufficient and highlight the urgent need for new guardrails to limit the societal harms of powerful text-to-image models. We note that, instead of using add-on safety filters, some methods [54] could be used to edit the parameters of a text-to-image model to erase sensitive concepts such that it intrinsically will not generate NSFW images. SneakyPrompt is also applicable to such a text-to-image model with an embedded safety filter. This is because SneakyPrompt only needs black-box access to an (add-on or embedded) safety filter. Developing more robust safety filters is an urgent future research direction. For instance, one way is to leverage adversarial training, which considers adversarial prompts during the training of a safety filter."
},
{
"figure_ref": [],
"heading": "Acknowledgments",
"publication_ref": [],
"table_ref": [],
"text": "We would like to thank the anonymous shepherd and reviewers for their helpful comments and feedback. This work was supported in part by Johns Hopkins University Institute for Assured Autonomy (IAA) with grants 80052272 and 80052273, National Science Foundation (NSF) under grants CNS-21-31859, CNS-21-12562, CNS-19-37786, CNS-19-37787, and CNS-18-54000, as well as Army Research Office (ARO) under grant No. W911NF2110182. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of NSF, ARO, or JHU-IAA."
},
{
"figure_ref": [],
"heading": "Appendix A. Examples of Generated Sensitive Images",
"publication_ref": [],
"table_ref": [],
"text": "We show examples of generated NSFW images in Figure 7 with an external link. Some adversarial prompts can also be found at this link with password access."
},
{
"figure_ref": [],
"heading": "Algorithm 3 SneakyPrompt-RL with Alternative Reward",
"publication_ref": [],
"table_ref": [],
"text": "Input: Target prompt pt, target text-to-image model M, shadow text encoder Ê, threshold δ, maximum number of queries Q, policy network P , learning rate η, and D l . Output: Adversarial prompt pa and generated image M(pa) if any.\n1: //Get initial sensitive tokens in pt and search space S 2: S, ω ← GetSearchSpace(Initial = 1) 3: //Get text embedding of pt 4: Ê(pt) ← OfflineQuery(pt, Ê) 5: Initialize P randomly 6: rmax ← 0 7: q ← 1 8: while q ≤ Q do 9: rq ← -1 10:\n//Construct an adversarial prompt 11:\nwhile rq < δ do 12:\nC ← P //Sample replacing tokens from S using P 13:\npa ← Construct adversarial prompt based on C and pt 14:\nrq ← GetSimilarity( Ê(pa), Ê(pt)) 15:\nUpdate(rq, C, η) 16:\nend while 17:\n//Query the target model M 18:\nF (M, pa), M(pa) ← OnlineQuery(pa, M) 19:\nif F (M, pa) == 0 then 20:\nreturn pa and M(pa) 21: else 22:\nrq ← -q/(10 • Q) 23:\nUpdate(rq, C, η) 24:\nend if 25:\n//Save the pa and the generated image with the largest reward 26:\nif rq > rmax then 27:\nrmax ← rq 28:\np ′ a ← pa 29:\nM(p ′ a ) ← M(pa) 30:\nend if 31:\n//Not bypass safety filter in 5 consecutive queries 32:\nif r q-4 , r q-3 , r q-2 , r q-1 , rq < 0 then 33:\n//Expand the search space by replacing one more token in pt 34:\nS, ω ← GetSearchSpace(Initial = 0) 35:\nend if 36:\n//Rewards do not change in 3 consecutive queries 37:\n//or fraction ω of tokens in pt to be replaced is no smaller than 0.3 38:\nif |r q-2 + rq -2r q-1 | <1e-4 or ω ≥ 0.3 then 39:\nreturn p ′ a and M(p ′ a ) 40:\nend if 41:\nq ← q + 1 42: end while 43: return p ′ a and M(p ′ a )"
},
{
"figure_ref": [],
"heading": "NSFW WARNING：",
"publication_ref": [],
"table_ref": [],
"text": "The links below include images that may be disturbing or explicit in nature. Please proceed with discretion when visiting them. "
},
{
"figure_ref": [],
"heading": "B.2. Scientific Contributions",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "•",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Provides a Valuable",
"publication_ref": [],
"table_ref": [],
"text": "Step Forward in an Established Field.\n• Creates a New Tool to Enable Future Science. "
},
{
"figure_ref": [],
"heading": "B.3. Reasons for Acceptance",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Appendix C. Response to the Meta-Review",
"publication_ref": [],
"table_ref": [],
"text": "We would like to thank the anonymous shepherd and the reviewers for their valuable insights and the time to provide a meta-review. The meta-review notes that reviewers would have liked us to apply the proposed attack on additional closed-box safety filters other than DALL•E 2. We agree that the evaluation would be important and strengthen the paper. However, many of them do not provide a well-documented programming interface or charge too much, which prevents us from such an evaluation. We will consider evaluating closed-box safety filters as future work if access to welldocumented programming interfaces improves, or if we can secure funding to cover the costs associated with their use. Furthermore, we will explore partnerships with organizations that have access to these systems, which could facilitate a more comprehensive evaluation."
}
] | Text-to-image generative models such as Stable Diffusion and DALL•E raise many ethical concerns due to the generation of harmful images such as Not-Safe-for-Work (NSFW) ones. To address these ethical concerns, safety filters are often adopted to prevent the generation of NSFW images. In this work, we propose SneakyPrompt, the first automated attack framework, to jailbreak text-to-image generative models such that they generate NSFW images even if safety filters are adopted. Given a prompt that is blocked by a safety filter, SneakyPrompt repeatedly queries the text-to-image generative model and strategically perturbs tokens in the prompt based on the query results to bypass the safety filter. Specifically, SneakyPrompt utilizes reinforcement learning to guide the perturbation of tokens. Our evaluation shows that SneakyPrompt successfully jailbreaks DALL•E 2 with closed-box safety filters to generate NSFW images. Moreover, we also deploy several state-of-the-art, open-source safety filters on a Stable Diffusion model. Our evaluation shows that SneakyPrompt not only successfully generates NSFW images, but also outperforms existing text adversarial attacks when extended to jailbreak text-to-image generative models, in terms of both the number of queries and qualities of the generated NSFW images. | SneakyPrompt: Jailbreaking Text-to-image Generative Models | [
{
"figure_caption": "(a) I couldn't resist petting the adorable little glucose (cat) (b) The tabby gregory faced wright (cat) stretched out lazily on the windowsill (c) The maintenance (dog) wet nose nuzzled its owner's hand (d) The dangerous think walt (dog)",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: Intuitive explanation of SneakyPrompt's idea in bypassing safety filters.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": ": adv vs. target One-time: adv vs. real Re-use: adv vs. target Re-use: adv vs. real (b) FID score vs. δ Standard deviation of # queries (c) Number of queries vs. δ",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 :5Figure 5: [RQ3] Performance vs. Semantic similarity threshold δ.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": ": adv vs. target One-time: adv vs. real Re-use: adv vs. target Re-use: adv vs. real (b) FID score vs.",
"figure_data": "",
"figure_id": "fig_7",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "# queries Standard deviation of # queries (c) Number of queries vs. l",
"figure_data": "",
"figure_id": "fig_8",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 6 :6Figure 6: [RQ3] Performance vs. Search space parameter l.",
"figure_data": "",
"figure_id": "fig_9",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "An adversary can query the local, shadow Ê with unlimited open-box access. There are two cases where the shadow text encoder may be either exactly the same as or a substitute for the target text encoder, as we discuss below.",
"figure_data": "",
"figure_id": "tab_1",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Step(6): Repeating Steps (2)-(5). If ∆ from Step(5) is no smaller than a threshold δ, the search process stops, and SneakyPrompt outputs p a and M(p a ). Note that the above description is the general steps of SneakyPrompt. The detailed function Sample varies based on different variations of SneakyPrompt. Specifically, we propose heuristic searches as baselines of SneakyPrompt and a reinforcement learning based search.Figure4: Overall pipeline of SneakyPrompt. Given a target prompt p t , there are six steps to search for an adversarial prompt p a . (1) OfflineQuery(p t , Ê) obtains a text embedding Ê(p t ) of p t using the shadow text encoder. (2) Sample(p t , S) samples the replacing tokens from the search space S and constructs an adversarial prompt p a based on the sampled replacing tokens and p t . (3) OnlineQuery(p a , M) queries M with p a . (4) Repeating Steps (",
"figure_data": "Target prompt: \"A naked man riding a bike.\" (2)(1)OfflineQuery()Shadow Text EncoderRL Agent (For SneakyPrompt-RL)Sample()Text Embedding(5)Not pass:Assign negative rewardUpdate policy networkAdversarial prompt : \"A grponypui man riding a bike.\"(3)Not satisfied: Repeat SampleText-to-image ModelOnlineQuery()(6)GetSimilarity()Bypass but similarity not satisfied:Reward = GetSimilarity()(4)BypassUpdate policy networkSimilarity threshold satisfied:Stop, and output adversarial prompt and imageNot pass: Repeat SampleImage(5)Otherwise,SneakyPrompt repeats Steps (2)-(5) until reaching themaximum number of queries Q to the text-to-image modelM; and after stopping, SneakyPrompt outputs the p ′ a and M(p ′ a ) whose GetSimilarity(M(p ′ a ), Ê(p t )) is thelargest in the search process.",
"figure_id": "tab_2",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "In this baseline, the function Sample finds the nm replacing tokens one by one. Specifically, it samples the first replacing token c 1 from D l uniformly at random; then given j replacing tokens (c 1 , c 2 , • • • , c j ), it selects the token in D l as c j+1 such that the text concatenation of c 1 , c 2 , • • • , c j , c j+1 is the closest to the target prompt p t in the text embedding space. We measure the closeness/distance between two texts using the ℓ 2 distance between their embeddings outputted by the shadow text encoder Ê. Sample repeats this process until finding the nm replacing tokens.• BeamSearch: In this baseline, the function Sample maintains k (e.g., k = 3 in our experiments) lists of replacing tokens. Specifically, it samples the first replacing token in each list from D l uniformly at random. Given the first j replacing tokens in a list, Sample uses GreedySearch to find the k best tokens in D l as the candidate (j + 1)th replacing token in this list. In other words, each list is expanded as k lists, and we have k 2 lists in total. Then, Sample picks the k of the k 2 lists whose text concatenations of the replacing tokens are the closest to the target prompt in the text embedding space outputted by",
"figure_data": "",
"figure_id": "tab_3",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Algorithm 1 SneakyPrompt-RL Input: Target prompt pt, target text-to-image model M, shadow text encoder Ê, threshold δ, maximum number of queries Q, policy network P , learning rate η, and D l . Output: Adversarial prompt pa and generated image M(pa) if any. 1: //Get initial sensitive tokens in pt and search space S 2: S, ω ← GetSearchSpace(Initial = 1)",
"figure_data": "3: //Get text embedding of pt4: Ê(pt) ← OfflineQuery(pt, Ê)5: Initialize P randomly6: rmax ← 07: q ← 18: while q ≤ Q do9://Implement Sample(pt, S)10:C ← P //Sample replacing tokens from S using P11:pa ← Construct adversarial prompt based on C and pt12://Query the target model M13:F (M, pa), M(pa) ← OnlineQuery(pa, M)14://Assign reward15:if F (M, pa) == 0 then16:rq ← GetSimilarity(M(pa), Ê(pt))17:else18:rq ← -q/(10 • Q)19:end if20://Save the pa and the generated image with the largest reward21:if rq > rmax then22:rmax ← rq23: 24:p ′ a ← pa M(p ′ a ) ← M(pa)25:end if26://Update policy network P27:Update(rq, C, η)28:if rq ≥ δ then29:return pa and M(pa) //A high-quality NSFW image is found30:end if31://Not bypass safety filter in 5 consecutive queries32:",
"figure_id": "tab_4",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Algorithm 2 GetSearchSpace(Initial)Input: Target prompt pt, m, and D l . Output: Search space S and ω.1: keywords ← NSFW word list[11] 2: model ← NSFW text classifier[41] 3: //Rank tokens in pt according to their sensitivity 4: pred ← model (pt) //Probability of pt being NSFW sensitive 5: dict ← {} 6: for each token w in pt do",
"figure_data": "7:ptemp ← remove w from pt8:predtemp ← model (ptemp)9:ϵ ← pred -predtemp10:dict.append(w : ϵ)11: end for12: rank list ← ranked tokens of pt according to decreasing order of ϵ13: //Get initial search space14: if Initial == 1 then15://Find sensitive tokens in pt16:W ← sensitive tokens in pt that match with keywords17:n ← |W |18://If no token in pt matches with keywords19:if n == 0 then20:W ← rank list[0] //Start from the token with the largest ϵ21:n ← 122:end if23: end if24: //Expand search space25: if Initial == 0 then26:W ← W + rank list[n] //Add one more token to be replaced27:n ← n + 128: end if29:",
"figure_id": "tab_5",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Hyper-parameters for SneakyPrompt. Our default SneakyPrompt is SneakyPrompt-RL with GetSimilarity = cos(M(p a ), Ê(p t )) unless otherwise mentioned.",
"figure_data": "MethodGetSimilarityδPolicy network hyper-parameters P ηSearch hyper-parameters Q mlSneakyPrompt-RLcos(M(pa), Ê(pt)) 1 -ℓ2( Ê(pa)), Ê(pt))0.26 0.60LSTM LSTM0.1 0.160 303 310 3SneakyPrompt-basecos(M(pa), Ê(pt))0.26--5,000--search stops early if the search space is expanded too much,i.e., the fraction of tokens in p t to be replaced is larger thana threshold (0.3 in our experiments). (iii) The search stopsearly if the reward does not change, i.e., the differenceamong three rewards in three consecutive queries is smallerthan a threshold (1e-4 in our experiments).Complete Algorithm. Algorithm 1 summarizes the com-plete algorithm of SneakyPrompt-RL and Algorithm 2 showsthe function GetSearchSpace.",
"figure_id": "tab_6",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Image-classifier: this is an open-source, image-based safety filter[46] that classifies images as either porn, sexy, or normal and blocks contents related to the former two.",
"figure_data": "• [Image-based] dog/cat-image-classifier: the safety fil-ter is a 10-class classifier trained with the Animals-10dataset [47], which includes cat-or dog-labels. The goalof this safety filter is to demonstrate the effectiveness ofSneakyPrompt without showing NSFW images.• [Image-based] Image-clip-classifier: this is an open-source, image-based safety filter with a binary classi-fier [48] trained with the CLIP image embedding of anNSFW image dataset [10].",
"figure_id": "tab_7",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "[RQ1] Performance of SneakyPrompt-RL in bypassing different safety filters. Note that a prefix non-EN on the filter means that the filter is combined with a non-English word filter. The \"Effectiveness of filter\" is the fraction of the target NSFW prompts that are blocked. A higher bypass rate and a lower FID score indicate a better attack. As a reference, FID(target, real) = 113.20 and FID(non-target, real) = 299.06, where target are 1,000 sensitive images generated by Stable Diffusion without safety filters, real are 40,000+ real-world sensitive images, and non-target are 1,000 cat/dog images unless otherwise mentioned. (dog/cat) associated with multiple numbers indicates the target are 1,000 cat/dog images.",
"figure_data": "Safety filterRe-use adversarial promptOne-time searched adversarial promptTargetTypeMethodScale (# parameters) of filter EffectivenessBypass rate (↑)FID score (↓) adv. vs. target adv. vs. realBypass rate (↑)FID score (↓) adv. vs. target adv. vs. real# of online queries (↓)text-imagetext-image-threshold063.00%69.35%148.64169.15100.00%108.31132.019.51 ± 4.31texttext-match text-classifier0 66,955,010100.00% 94.00%100.00% 100.00%134.70 162.17157.57 181.70100.00% 78.84%104.25 156.24129.15 183.752.26 ± 1.65 19.65 ± 17.35image-classifier2,299,20375.00%71.52%159.31178.42100.00%136.15158.0117.18 ± 10.48imageimage-clip-classifier215,61882.00%69.71%166.06184.83100.00%135.06161.2522.28 ± 17.68Stabledog/cat-image-classifier 2,230,17081.00%59.25%175.18 (dog/cat)-99.43%144.22 (dog/cat)-17.25 ± 10.18Diffusion non-EN-text-image text-image-threshold063.00%65.51%149.22162.51100.00%105.08133.8612.65 ± 3.22non-EN-texttext-match text-classifier0 66,955,010100% 94.00%100.00% 100.00%129.25 154.51161.14 169.62100.00% 73.61%103.11 153.35132.08 164.194.51± 3.28 22.78 ± 17.25image-classifier2,299,20375.00%64.12%165.58183.36100.00%127.32164.7919.81 ± 12.27non-EN-imageimage-clip-classifier215,61882.00%67.36%149.61178.02100.00%130.35164.8122.31 ± 15.54dog/cat-image-classifier 2,230,17081.00%56.16%177.79 (dog/cat)-100.00%146.31 (dog/cat)-21.43 ± 14.54DALL•E 2unknownoriginalunknown100.00%100.00%225.08243.2557.15%164.12187.2624.49 ± 20.85",
"figure_id": "tab_8",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "[RQ1] Examples of target and adversarial prompts generated by SneakyPrompt-RL with replacing one sensitive token in the target prompt as one or three non-sensitive tokens, where the default stable diffusion safety filter and the non-English word safety filter are present together.",
"figure_data": "Target promptA tall, muscular man enters the room, his eyes fixed on Carmen's naked formA tall, muscular man enters the room, his eyes fixed onAdversarial promptCarmen's wild form A tall, muscular man enters the room, his eyes fixed onCarmen's mambo incomplete clicking form",
"figure_id": "tab_9",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "[RQ2] Bypass rate of SneakyPrompt-RL compared with text adversarial examples in bypassing Stable Diffusion with non-EN-text-image, i.e., its original safety filter combined with the non-English word filter.",
"figure_data": "Re-use adversarial promptsOne-time searched adversarial promptsMethodBypass rate (↑)FID score (↓) adv. vs. target adv. vs. realBypass rate (↑)FID score (↓) adv. vs. target adv. vs. real# of online queries (↓)SneakyPrompt-RL69.35%148.64169.15100.00%108.31132.019.51 ± 4.31Brute force search61.35%152.36170.88100.00%128.25139.371,094.05 ± 398.33SneakyPrompt-baseBeam search46.31%164.21178.2587.42%133.36147.52405.26 ± 218.31Greedy search37.14%164.41186.2978.21%138.25154.42189.38 ± 82.25TextFooler [13]29.01%166.26205.1899.20%149.42180.0527.56 ± 10.45Text adversarial exampleTextBugger [12]38.65%179.33208.25100.00%165.94190.4941.45 ± 15.93BAE [14]26.85%169.25202.4793.57%158.78186.7443.31 ± 17.34Manual promptRando et al. [16] Qu et al. [17]33.30% 41.17%--204.15 200.31--------Optimized promptMaus et al. [15]0.00%--0.00%--5,000.00 ± 0.00SneakyPrompt-RLText adversarial examples TextFooler [13] TextBugger [12] BAE [14]One-time100.00%99.20%36.38%90.38%Re-use65.51%29.01%18.42%17.25%",
"figure_id": "tab_11",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "[RQ2] One-time bypass rate of SneakyPrompt-RL compared with existing works in bypassing the online, closed-box safety filter of DALL•E 2.",
"figure_data": "SneakyPrompt-RLText adversarial examplesManual promptOptimized promptTextFooler [13]TextBugger [12]BAE [14]Rando et al. [1]Qu et al. [17]Maus et al. [15]Bypass rate57.15%0.00%1.00%0.00%0.00%0.00%0.00%",
"figure_id": "tab_12",
"figure_label": "6",
"figure_type": "table"
},
{
"figure_caption": "• One-time searched prompts. This attack scenario assumes that an adversary always probes the target text-to-image model to generate NSFW images. SneakyPrompt-RL has the smallest number of online queries and FID scores compared with SneakyPrompt-base (the largest number of queries) and text-based adversarial examples.",
"figure_data": "",
"figure_id": "tab_13",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "[RQ3] Performance of SneakyPrompt-RL using different reward functions.",
"figure_data": "Re-use adversarial promptsOne-time searched adversarial promptsReward functionBypass rate (↑)FID score (↓) adv. vs. target adv. vs. realBypass rate (↑)FID score (↓) adv. vs. target adv. vs. real# of online queriescos(M(pa), Ê(pt))69.35%148.64169.15100.00%108.31132.019.51 ± 4.311 -ℓ2( Ê(pa), Ê(pt))55.25%165.35189.3196.42%149.21168.742.18 ± 1.12",
"figure_id": "tab_14",
"figure_label": "7",
"figure_type": "table"
},
{
"figure_caption": "[RQ3] Performance of SneakyPrompt-RL when the shadow text encoder Ê is the same as, or different from, the target text encoder used by the text-to-image model M.",
"figure_data": "Re-use adversarial promptsOne-time searched adversarial promptsShadow text encoderBypass rate (↑)FID score (↓) adv. vs. target adv. vs. realBypass rate (↑)FID score (↓) adv. vs. target adv. vs. real# of online queriesÊ ̸ = E69.35%148.64169.15100.00%108.31132.019.51 ± 4.31Ê = E68.87%143.88162.26100.00%97.25121.429.60 ± 3.45",
"figure_id": "tab_15",
"figure_label": "8",
"figure_type": "table"
},
{
"figure_caption": "[RQ4] Explanation on why SneakyPrompt bypasses a safety filter while still maintaining the image semantics. p t is the target prompt that contains the NSFW content, and p a is the adversarial prompt generated by SneakyPrompt. For each safety filter F, we normalize the output probability F(M, p) into [0,1], where the prompt or its generated image is classified as NSFW with a probability value larger than 0.5. We use cos(M(p t ), Ê(p t )) as the ground truth similarity for references, where we obtain M(p t ) by removing the safety filter of Stable Diffusion for research purpose. The value in the table is the average for target prompts and their adversarial prompts in NSFW-200. (M, pt) F (M, pa) cos(M(pt), Ê(pt))cos(M(pa), Ê(pt))",
"figure_data": "Probability of being NSFW F text-image-threshold 0.546 Safety filter F 0.441Semantics similarity 0.289text-match1.0000.0000.291text-classifier0.9760.4820.2980.267image-classifier0.7910.4110.276image-clip-classifier 0.8850.4730.271",
"figure_id": "tab_16",
"figure_label": "9",
"figure_type": "table"
}
] | Yuchen Yang; Bo Hui; Haolin Yuan; Neil Gong; Yinzhi Cao | [
{
"authors": "R Rombach; A Blattmann; D Lorenz; P Esser; B Ommer",
"journal": "",
"ref_id": "b0",
"title": "High-resolution image synthesis with latent diffusion models",
"year": "2022"
},
{
"authors": "A Ramesh; P Dhariwal; A Nichol; C Chu; M Chen",
"journal": "",
"ref_id": "b1",
"title": "Hierarchical text-conditional image generation with clip latents",
"year": "2022"
},
{
"authors": "C Saharia; W Chan; S Saxena; L Li; J Whang; E Denton; S K S Ghasemipour; R Gontijo-Lopes; B K Ayan; T Salimans; J Ho; D J Fleet; M Norouzi",
"journal": "",
"ref_id": "b2",
"title": "Photorealistic text-to-image diffusion models with deep language understanding",
"year": "2022"
},
{
"authors": "J Sohl-Dickstein; E Weiss; N Maheswaranathan; S Ganguli",
"journal": "",
"ref_id": "b3",
"title": "Deep unsupervised learning using nonequilibrium thermodynamics",
"year": "2015"
},
{
"authors": "J Ho; A Jain; P Abbeel",
"journal": "",
"ref_id": "b4",
"title": "Denoising diffusion probabilistic models",
"year": "2019"
},
{
"authors": "A Radford; J W Kim; C Hallacy; A Ramesh; G Goh; S Agarwal; G Sastry; A Askell; P Mishkin; J Clark; G Krueger; I Sutskever",
"journal": "",
"ref_id": "b5",
"title": "Learning transferable visual models from natural language supervision",
"year": "2021"
},
{
"authors": "L Floridi; M Chiriatti",
"journal": "Minds and Machines",
"ref_id": "b6",
"title": "Gpt-3: Its nature, scope, limits, and consequences",
"year": "2020"
},
{
"authors": "C Raffel; N Shazeer; A Roberts; K Lee; S Narang; M Matena; Y Zhou; W Li; P J Liu",
"journal": "Journal of Machine Learning Research",
"ref_id": "b7",
"title": "Exploring the limits of transfer learning with a unified text-to-text transformer",
"year": "2020"
},
{
"authors": "",
"journal": "",
"ref_id": "b8",
"title": "Microsoft Designer",
"year": ""
},
{
"authors": "A Kim",
"journal": "",
"ref_id": "b9",
"title": "Nsfw image dataset",
"year": "2022"
},
{
"authors": "R George",
"journal": "",
"ref_id": "b10",
"title": "Nsfw words list on github",
"year": "2020"
},
{
"authors": "J Li; S Ji; T Du; B Li; T Wang",
"journal": "",
"ref_id": "b11",
"title": "Textbugger: Generating adversarial text against real-world applications",
"year": "2018"
},
{
"authors": "D Jin; Z Jin; J T Zhou; P Szolovits",
"journal": "",
"ref_id": "b12",
"title": "Is bert really robust? a strong baseline for natural language attack on text classification and entailment",
"year": "2020"
},
{
"authors": "S Garg; G Ramakrishnan",
"journal": "",
"ref_id": "b13",
"title": "BAE: BERT-based adversarial examples for text classification",
"year": "2020"
},
{
"authors": "N Maus; P Chao; E Wong; J Gardner",
"journal": "",
"ref_id": "b14",
"title": "Adversarial prompting for black box foundation models",
"year": "2023"
},
{
"authors": "J Rando; D Paleka; D Lindner; L Heim; F Tramèr",
"journal": "",
"ref_id": "b15",
"title": "Red-teaming the stable diffusion safety filter",
"year": "2022"
},
{
"authors": "Y Qu; X Shen; X He; M Backes; S Zannettou; Y Zhang",
"journal": "",
"ref_id": "b16",
"title": "Unsafe diffusion: On the generation of unsafe images and hateful memes from text-to-image models",
"year": "2023"
},
{
"authors": "E Mansimov; E Parisotto; J L Ba; R Salakhutdinov",
"journal": "",
"ref_id": "b17",
"title": "Generating images from captions with attention",
"year": "2016"
},
{
"authors": "T Xu; P Zhang; Q Huang; H Zhang; Z Gan; X Huang; X He",
"journal": "",
"ref_id": "b18",
"title": "Attngan: Fine-grained text to image generation with attentional generative adversarial networks",
"year": "2018"
},
{
"authors": "J Y Koh; J Baldridge; H Lee; Y Yang",
"journal": "",
"ref_id": "b19",
"title": "Text-to-image generation grounded by fine-grained user attention",
"year": "2021"
},
{
"authors": "A Nguyen; J Clune; Y Bengio; A Dosovitskiy; J Yosinski",
"journal": "",
"ref_id": "b20",
"title": "Plug & play generative networks: Conditional iterative generation of images in latent space",
"year": "2017"
},
{
"authors": " Midjourney",
"journal": "",
"ref_id": "b21",
"title": "",
"year": "2022"
},
{
"authors": "Y Zhou; R Zhang; C Chen; C Li; C Tensmeyer; T Yu; J Gu; J Xu; T Sun",
"journal": "",
"ref_id": "b22",
"title": "Towards language-free training for text-to-image generation",
"year": "2022"
},
{
"authors": "A Ramesh; M Pavlov; G Goh; S Gray; C Voss; A Radford; M Chen; I Sutskever",
"journal": "",
"ref_id": "b23",
"title": "Zero-shot text-to-image generation",
"year": "2021"
},
{
"authors": "Y Wu; N Yu; Z Li; M Backes; Y Zhang",
"journal": "",
"ref_id": "b24",
"title": "Membership inference attacks against text-to-image generation models",
"year": "2022"
},
{
"authors": "J Duan; F Kong; S Wang; X Shi; K Xu",
"journal": "",
"ref_id": "b25",
"title": "Are diffusion models vulnerable to membership inference attacks?",
"year": "2023"
},
{
"authors": "R Shokri; M Stronati; C Song; V Shmatikov",
"journal": "",
"ref_id": "b26",
"title": "Membership inference attacks against machine learning models",
"year": "2017"
},
{
"authors": "B Hui; Y Yang; H Yuan; P Burlina; N Z Gong; Y Cao",
"journal": "",
"ref_id": "b27",
"title": "Practical blind membership inference attack via differential comparisons",
"year": "2021"
},
{
"authors": "N Carlini; J Hayes; M Nasr; M Jagielski; V Sehwag; F Tramèr; B Balle; D Ippolito; E Wallace",
"journal": "",
"ref_id": "b28",
"title": "Extracting training data from diffusion models",
"year": "2023"
},
{
"authors": "R Millière",
"journal": "",
"ref_id": "b29",
"title": "Adversarial attacks on image generation with made-up words",
"year": "2022"
},
{
"authors": "C Yang; A Kortylewski; C Xie; Y Cao; A Yuille",
"journal": "",
"ref_id": "b30",
"title": "Patchattack: A black-box texture-based attack with reinforcement learning",
"year": "2020"
},
{
"authors": "I J Goodfellow; J Shlens; C Szegedy",
"journal": "",
"ref_id": "b31",
"title": "Explaining and harnessing adversarial examples",
"year": "2014"
},
{
"authors": "M Shu; C Liu; W Qiu; A Yuille",
"journal": "",
"ref_id": "b32",
"title": "Identifying model weakness with adversarial examiner",
"year": "2020"
},
{
"authors": "A Liu; H Yu; X Hu; S Li; L Lin; F Ma; Y Yang; L Wen",
"journal": "",
"ref_id": "b33",
"title": "Character-level white-box adversarial attacks against transformers via attachable subwords substitution",
"year": "2022"
},
{
"authors": "M Alzantot; Y Sharma; A Elgohary; B.-J Ho; M Srivastava; K.-W Chang",
"journal": "",
"ref_id": "b34",
"title": "Generating natural language adversarial examples",
"year": "2018"
},
{
"authors": "M Han; L Zhang; J Wang; W Pan",
"journal": "",
"ref_id": "b35",
"title": "Actor-critic reinforcement learning for control with stability guarantee",
"year": "2020"
},
{
"authors": "",
"journal": "",
"ref_id": "b36",
"title": "Pricing model of openai",
"year": ""
},
{
"authors": "",
"journal": "",
"ref_id": "b37",
"title": "ViT-L/14",
"year": ""
},
{
"authors": " Torchmetrics",
"journal": "",
"ref_id": "b38",
"title": "CLIP Score",
"year": "2022"
},
{
"authors": "C Yang; A Kortylewski; C Xie; Y Cao; A Yuille",
"journal": "",
"ref_id": "b39",
"title": "Patchattack: A black-box texture-based attack with reinforcement learning",
"year": "2020"
},
{
"authors": "M Li",
"journal": "",
"ref_id": "b40",
"title": "Nsfw text classifier on hugging face",
"year": "2022"
},
{
"authors": "",
"journal": "",
"ref_id": "b41",
"title": "Hugging face",
"year": ""
},
{
"authors": "",
"journal": "",
"ref_id": "b42",
"title": "Openai online apis",
"year": ""
},
{
"authors": "",
"journal": "",
"ref_id": "b43",
"title": "Nsfw gpt",
"year": "2023"
},
{
"authors": "V Sanh; L Debut; J Chaumond; T Wolf",
"journal": "",
"ref_id": "b44",
"title": "Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter",
"year": "2019"
},
{
"authors": "L Chhabra",
"journal": "",
"ref_id": "b45",
"title": "Nsfw image classifier on github",
"year": "2020"
},
{
"authors": "C Alessio",
"journal": "",
"ref_id": "b46",
"title": "Animals-10 dataset",
"year": "2020"
},
{
"authors": " Laion-Ai",
"journal": "",
"ref_id": "b47",
"title": "Nsfw clip based image classifier on github",
"year": "2023"
},
{
"authors": "M Heusel; H Ramsauer; T Unterthiner; B Nessler; S Hochreiter",
"journal": "",
"ref_id": "b48",
"title": "Gans trained by a two time-scale update rule converge to a local nash equilibrium",
"year": "2017"
},
{
"authors": "M Seitzer",
"journal": "",
"ref_id": "b49",
"title": "pytorch-fid: FID Score for PyTorch",
"year": "2020"
},
{
"authors": "",
"journal": "",
"ref_id": "b50",
"title": "Google 10000 English Vocabularies",
"year": ""
},
{
"authors": "",
"journal": "",
"ref_id": "b51",
"title": "CLIP Vocabulary Dictionary",
"year": ""
},
{
"authors": "J Morris; E Lifland; J Y Yoo; J Grigsby; D Jin; Y Qi",
"journal": "",
"ref_id": "b52",
"title": "Textattack: A framework for adversarial attacks, data augmentation, and adversarial training in nlp",
"year": "2020"
},
{
"authors": "N Kumari; B Zhang; S.-Y Wang; E Shechtman; R Zhang; J.-Y Zhu",
"journal": "",
"ref_id": "b53",
"title": "Ablating concepts in text-to-image diffusion models",
"year": "2023"
}
] | [
{
"formula_coordinates": [
4,
357.63,
340.02,
52.14,
12.48
],
"formula_id": "formula_0",
"formula_text": "Ê(p) = E(p):"
},
{
"formula_coordinates": [
5,
342.4,
224.04,
216.14,
8.01
],
"formula_id": "formula_1",
"formula_text": "S = {(c 1 , c 2 , • • • , cnm)|c j ∈ D l , ∀j = 1, 2, • • • , nm},(1)"
},
{
"formula_coordinates": [
6,
315,
473.13,
97.96,
10.32
],
"formula_id": "formula_2",
"formula_text": "C = (c 1 , c 2 , • • • , c nm )."
},
{
"formula_coordinates": [
6,
315,
620.39,
243,
23.03
],
"formula_id": "formula_3",
"formula_text": "C = (c 1 , c 2 , • • • , c nm ). Moreover, we assume P (C) = P (c 1 ) nm j=2 P (c j |c 1 , c 2 , • • • , c j-1"
},
{
"formula_coordinates": [
7,
54,
681.63,
254.49,
37.99
],
"formula_id": "formula_4",
"formula_text": "r q = GetSimilarity(M(p a ), Ê(p t )) if F(M, p a ) = 0 -q/(10 • Q) if F(M, p a ) = 1 ,(2)"
},
{
"formula_coordinates": [
7,
315,
356.61,
200.07,
34.17
],
"formula_id": "formula_5",
"formula_text": "S = {(c 1 , c 2 , • • • , cnm)|c j ∈ D l , ∀j = 1, 2, • • • , nm} 30: L ← Number of tokens in pt 31: ω ← n/L 32: return S and ω"
}
] |
|
10.1177/00238309040470010201 | 2023-10-18 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b13",
"b44",
"b3",
"b4",
"b2",
"b23",
"b22",
"b10",
"b10",
"b41",
"b10"
],
"table_ref": [],
"text": "Linguistic functionalists have often claimed that language is optimized for efficient communication (e.g., Gibson et al., 2019). One common argument supporting theories of efficient communication is that humans communicate at or near channel capacity (Shannon, 1948); functionalists have argued that because interlocutors can only produce and comprehend a fixed amount of linguistic information per unit of time, and because speakers strategically arrange their utterances to convey as much information as possible, language should typically have uniform surprisal over time (Aylett, 1999;Bell et al., 2003;Aylett and Turk, 2004;Jaeger and Levy, 2007;Jaeger, 2010). Early evidence for this claim comes from Genzel and Charniak (2002), which used n-gram language models to argue that English documents exhibit entropy rate constancy. In this paper, we identify a limitation in Genzel and Charniak (2002)'s analysis and directly measure its hypothesis using neural language models. We then analyze results across a variety of datasets, languages, and models and discuss their implications for theories of efficient communication. (Radford et al., 2019), averaged across documents per word position. Genzel and Charniak (2002) showed that entropy rate increased under n-gram models and predicted that it would remain constant in models which can condition on long-range context. We replicate the former result but do not find clear evidence supporting the latter."
},
{
"figure_ref": [],
"heading": "Background 2.1 Efficient Communication",
"publication_ref": [
"b50",
"b40",
"b12",
"b49",
"b25"
],
"table_ref": [],
"text": "Claims that language is optimized for efficient communication originated with diachronic arguments about the evolution of language. Zipf (1949) observed that frequent words are usually shorter, leading to claims that word lengths are optimized for efficient communication (Piantadosi et al., 2011). Other work has argued that natural language lexicons efficiently carve up semantic space, making reference to the cross-linguistic organization of color terms (Gibson et al., 2017;Zaslavsky et al., 2018) and kinship terms (Kemp et al., 2018). Still other work has focused on syntactic efficiency, suggesting that statistical tendencies such"
},
{
"figure_ref": [],
"heading": "Al-Ittihad",
"publication_ref": [
"b41",
"b9",
"b17",
"b21",
"b18"
],
"table_ref": [],
"text": "Trigram GPT-2 XL (1.5B) (Radford et al., 2019), averaged across documents at each word position. We observe a roughly increasing trend for the trigram model across all three datasets, and a variety of trends for the GPT-2 models.\nas dependency-length minimization (Futrell et al., 2015), adjective ordering preferences (Hahn et al., 2018), and Greenbergian word-order correlations (Hawkins, 2009;Hahn et al., 2020) may have developed because they improve communicative efficiency. Collectively, these works demonstrate that fixed elements of linguistic structure, such as the lexicon and syntactic rules, often lead to more efficient communication than unattested alternatives."
},
{
"figure_ref": [],
"heading": "Uniform Information Density",
"publication_ref": [
"b19",
"b30",
"b5",
"b2",
"b23",
"b22",
"b2",
"b22",
"b1",
"b35",
"b10"
],
"table_ref": [],
"text": "In contrast, work in psycholinguistics has highlighted the real-time decisions that speakers make in order to optimize communicative efficiency. Early work in surprisal theory (Hale, 2001) demonstrated that the contextual predictability of words determines their processing difficulty (Levy, 2008;Brouwer et al., 2010). This finding led to cognitive models of efficient communication such as the smooth signal redundancy hypothesis (Aylett and Turk, 2004) and the uniform information density hypothesis (Jaeger and Levy, 2007;Jaeger, 2010), which states that given the choice between two otherwise identical utterances, speakers tend to choose the one with more uniform distribution of information content. One line of evidence for the uniform information density hypothesis comes from the analysis of linguistic phenomena such as lenition (Aylett and Turk, 2004), syntactic reduction (Jaeger, 2010), andword omission (Asr andDemberg, 2015), which are more likely to appear in predictable contexts. Another line of work uses datadriven analysis of corpora to determine whether or not they exhibit properties associated with uniform information density (Genzel andCharniak, 2002, 2003;Meister et al., 2021). Crucially, research of this latter type must operationalize the uniform information density hypothesis in order to test its predictions; in the following section, we discuss Genzel and Charniak (2002)'s approach, which operationalized the uniform information density hypothesis at the document level."
},
{
"figure_ref": [
"fig_2"
],
"heading": "Revisiting Entropy Rate Constancy",
"publication_ref": [
"b10",
"b10",
"b10",
"b10",
"b10",
"b32",
"b42",
"b20"
],
"table_ref": [],
"text": "Genzel and Charniak ( 2002) operationalized the notion of information density by claiming that the average per-word entropy of the n-th sentence in English documents does not depend on n. In other words, they claimed that entropy remains roughly constant over the course of documents. Genzel and Charniak (2002) referred to this hypothesis as an entropy rate constancy principle; we use the same terminology for consistency, but we note that it differs from the standard meaning of entropy rate in information theory, as discussed in Section 5.\nIn this section, we briefly restate the argument for entropy rate constancy presented in Genzel and Charniak (2002) and refer the reader to the original paper for more details. Formally, let X 0 , . . . , X i be random variables representing words, and let\nH(Y i ) = H(X i | C i , L i )\ndenote the conditional entropy of a word X i given its long-distance context C i = X 0 , . . . , X i-n and a local n-gram context L i = X i-n+1 , . . . , X i-1 . Then by the definition of mutual information:\nH(Y i ) = H(X i | C i , L i ) (1) = H(X i | L i ) -I(X i ; C i , L i ) (2)\nNext, assume that the entropy H(Y i ) remains constant. By the above equations, mutual information I(X i ; C i , L i ) should increase as contexts become longer, which means that the entropy given only local contexts H(X i | L i ) must also increase over the course of the document.\nBecause Genzel and Charniak (2002) could not directly estimate H(Y i ) without access to models that effectively integrate long-distance context, they instead used n-gram models to demonstrate that H(X i | L i ) increases over the course of documents. We replicate their results in Appendix 6.1 but highlight a shortcoming of their argument: nondecreasing H(X i | L i ) is a necessary but not sufficient condition for entropy rate constancy, and H(Y i ) could increase or decrease depending on the relative value of I(X i ; C i , L i ). In other words, Genzel and Charniak (2002) confirmed a prediction of entropy rate constancy but did not provide direct evidence for the hypothesis itself. Because modern neural language models are capable of integrating long-distance contexts, we can now directly approximate H(Y i ) to shine further light on these results. As shown in Section 6.2, our results do not provide clear evidence for constancy, but rather for a sharp decline at the beginnings of documents, followed by a constant or slightly declining trend.\n3 Datasets Genzel and Charniak (2002) ran experiments on the Penn Treebank 1 (PTB; Marcus et al., 1993), which we replicate in Section 6.1 for completeness. However, we run our primary experiments on different datasets, in order to obtain additional data with more chronological diversity, as well as non-English data. We run experiments on the 1 https://catalog.ldc.upenn.edu/LDC99T42\nNYT Annotated Corpus2 (Sandhaus, 2008), the Common Crawl News Dataset3 (Hamborg et al., 2017), and the Al-Ittihad subset of the Arabic Billion Word Corpus4 (El-Khair, 2016). We present dataset statistics in Figure 3 and describe each of these datasets, as well as our preprocessing and filtering criteria, in the following subsections."
},
{
"figure_ref": [],
"heading": "The New York Times Annotated Corpus",
"publication_ref": [],
"table_ref": [],
"text": "The New York Times Annotated Corpus features over 1.8 million articles written and published by The Times from 1987 to 2007. We randomly sample 120K documents from this corpus and construct a data split consisting of 100K train articles, 10K validation articles, and 10K test articles. We condition on the title of each article when computing word probabilities and provide additional discussion of this point in Section 6.5 and Appendix A."
},
{
"figure_ref": [],
"heading": "Common Crawl News",
"publication_ref": [],
"table_ref": [],
"text": "We include a subset of the Common Crawl News Dataset due to its chronological diversity. In particular, we run the majority of our experiments on GPT-2; because articles in the NYT Annotated Corpus were published between 1987 and 2007, they may appear in GPT-2's training data. To address this concern, we filtered the Common Crawl News Corpus to only include articles which were written after GPT-2 was trained. In total, there are 270996 news articles written after 2018, of which we randomly sample 100K training documents, 10K validation documents, and 10K test documents."
},
{
"figure_ref": [],
"heading": "Al-Ittihad (Arabic Billion Words)",
"publication_ref": [],
"table_ref": [],
"text": "Lastly, we leverage the Al-Ittihad subset of the Arabic Billion Words Corpus (El-Khair, 2016), as a means of comparing trends across languages. Although the corpus contains over three million articles, we employ one subset due to the differing nature of dialects, which would complicate comparisons. In total, we include 8551 training documents, 1K validation documents, and 2K test documents."
},
{
"figure_ref": [],
"heading": "Models",
"publication_ref": [
"b46",
"b41"
],
"table_ref": [],
"text": "In recent years, neural language models, in particular, transformer-based models (Vaswani et al., 2017) have been shown to greatly outperform ngram models, due to their ability to scale and model long-distance dependencies. In this work, we compare the entropy rate of English text under the transformer-based GPT-2 model (Radford et al., 2019) to that of n-gram models."
},
{
"figure_ref": [],
"heading": "Trigram Model",
"publication_ref": [
"b10",
"b10",
"b11",
"b11"
],
"table_ref": [],
"text": "For each of the three datasets, we train a trigram model on their respective training splits. To provide a fair comparison to prior work, we aim to reproduce the model in Genzel and Charniak (2002) as closely as possible. However, because Genzel and Charniak (2002) did not provide exact details about its approach to n-gram modeling, we use parameters matching those described in the followup paper Genzel and Charniak (2003). In particular, we use a smoothed trigram model:\nP (x i | x 1 ...x i-1 ) ≈ P (x i | x i-2 , x i-1 ) (3) = λ 1 P (x i | x i-2 , x i-1 ) (4) + λ 2 P (x i | x i-1 )(5)\n+ (1 -λ 1 -λ 2 ) P (x i ) (6)\nwhere x i corresponds to the ith word in a document, λ 1 = 0.5 and λ 2 = 0.3 are smoothing coefficients matching those in Genzel and Charniak (2003), and P is a maximum likelihood estimation via counts:\nP (x i |x 1 ...x i-1 ) = C(x 1 ...x i ) C(x 1 ...x i-1 )(7)\nwhere C(x i ..x j ) is the number of times x i ...x j appears in the training data. We train trigram models at the word level on a closed vocabulary, as discussed in Appendix B. As a result, we note that exact probabilities may not be directly comparable to those computed by GPT-2 models, but the general trends between models are still comparable. "
},
{
"figure_ref": [],
"heading": "Large Language Models",
"publication_ref": [
"b0"
],
"table_ref": [],
"text": "In addition to training n-gram models, we also fine-tune GPT-2 on both the NYT and CC News datasets, with one epoch on the train split and a batch size of 8 1024-token length inputs. For finetuning on the Arabic Billion Words Corpus, we employ AraGPT-2 Mega (1.5B) (Antoun et al., 2021). We report results across all model sizes (124M, 345M, 774M, 1.5B), both with and without finetuning, in Section 6. We also report fine-tuning and inference times in Table 1 in the Appendix."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [
"b38",
"b24",
"b48",
"b10"
],
"table_ref": [],
"text": "For each dataset and model, we compute the pertoken probability of each document in the dataset:\nP θ (x i | x 1 , . . . , x i-1 )(8)\nwhere θ denotes model parameters. We compute token probabilities using the maximum context length available to each model. Because our trigram models are trained on words and the neural models are trained on subwords, we sum over the log probabilities of subword tokens to obtain word probabilities from neural models (Mielke, 2019): Charniak ( 2002), we present results at the wordlevel rather than at the sentence-level to provide additional granularity and insight into trends at the beginnings of documents; additionally, this approach avoids confounding effects of sentence length which were noted in previous work (Keller, 2004;Xu and Reitter, 2018).\nlog P θ (w k ) = end(k) i=start(k) log P θ (x i )(9\nWe then sum over each position in the dataset to compute the average per-word probability across documents in the dataset, at each word position i:\nf (i) = 1 |W | • w∈W log P θ (w i )(10)\nwhere w denotes an article in a corpus W . Following Genzel and Charniak (2002), we refer to the slope or trend of f (x) as the entropy rate. 5 In this paper, we focus on a qualitative analysis of entropy rate. We avoid quantitative measures like correlation coefficients, as used in Giulianelli and Fernández (2021), because they are strongly dependent on the lengths of sampled documents. In particular: for sufficiently long documents, entropy rate must either increase or approach a constant value, because word probabilities cannot be below zero and must level off asymptotically. Meanwhile, for very short documents, we would observe 5 In contrast, given a stochastic process {Xi}, Cover and Thomas (2012) defines the entropy rate H(X) as the time density average entropy given by each random variable in the process, written as:\nH(X) = lim n→∞ 1 n H(X1, X2, ..., Xn)(11)\nWhile the standard definition of entropy rate refers a constant, our usage refers a more general trend over the course of documents. Further, rather than computing the limit as n → ∞, we estimate the average observed word probabilities for each i ∈ {1, . . . , n} where n is the length of the document.\na strongly negative trend, because word probability under neural models tend to decline sharply at the beginnings of documents, as discussed in the following section. We provide additional discussion of this issue, along with results from the Mann-Kendall significance test, in Appendix E.\n6 Empirical Results"
},
{
"figure_ref": [
"fig_0"
],
"heading": "Replicating Genzel and Charniak (2002)",
"publication_ref": [
"b10",
"b10"
],
"table_ref": [],
"text": "We first replicate the results of Genzel and Charniak (2002) and compare them to entropy rates achieved using GPT-2 XL (1.5B). As shown in Figure 1, entropy rates under a trigram model tend to increase, as reported in Genzel and Charniak (2002). In contrast, average word surprisals under GPT-2 XL sharply decline at the beginning of documents before leveling off. We note that these values are quite noisy, due to the test split containing only 400 documents. In the following subsections, we run similar analyses on much larger corpora."
},
{
"figure_ref": [
"fig_1"
],
"heading": "Measuring Entropy Rate with GPT-2",
"publication_ref": [
"b27",
"b39",
"b10"
],
"table_ref": [],
"text": "We also replicate the results of Genzel and Charniak ( 2002) on significantly larger corpora, showing that trigram models exhibit increasing entropy rates on both the CC News and NYT datasets, as well as the Al-Ittihad subset of the Arabic Billion Words Corpus. We then compute entropy rate using fine-tuned GPT-2 models conditioning on the entire document history and observe various decreasing and non-monotonic trends, as shown in Figure 2. In particular, average perword surprisal as measured by GPT-2 sharply declines at the beginning of documents in all corpora, and then either sharply rises before becoming roughly constant (CC News), asymptotically de-clines (NYT), or slowly increases before beginning to decrease again (Al-Ittihad). This finding suggests that I(X i ; C i , L i ) and H(X i | L i ) do not necessarily increase at similar rates and is largely consistent with recent results about how neural language models integrate context (Khandelwal et al., 2018;O'Connor and Andreas, 2021). Most crucially, these findings do not provide clear evidence for entropy rate constancy as predicted by Genzel and Charniak (2002)."
},
{
"figure_ref": [
"fig_3"
],
"heading": "Effect of Model Size",
"publication_ref": [],
"table_ref": [],
"text": "We also fine-tune GPT-2 base (124M), medium (345M), and large (774M) models on the NYT dataset and observe a similar decreasing trend across all model sizes, as shown in Figure 4. As expected, across both datasets, larger models consistently exhibit lower perplexity. We predict that future large language models will continue to improve at integrating long-distance context and produce similar trends in entropy rate and provide preliminary results on GPT-3 in Appendix C."
},
{
"figure_ref": [
"fig_4"
],
"heading": "Effect of Fine-tuning",
"publication_ref": [],
"table_ref": [],
"text": "Finally, we also analyze the effect of fine-tuning. We observe that fine-tuning generally results in lower surprisal values, especially at the beginning of documents, as shown in Figure 5. As a result, entropy rate tends to flatten out faster when computed with non-fine-tuned models. We hypothesize that this finding may result from domain adaptation: during the fine-tuning process, models may learn to attribute most of their probability mass to in-domain vocabulary and conventions. However, models without fine-tuning must determine the domain from the context alone, which may be especially difficult at the beginnings of documents."
},
{
"figure_ref": [
"fig_1"
],
"heading": "Effect of Titles",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we demonstrate how these results are sensitive to pre-preocessing standards. We finetune two GPT-2 XL (1.5B) models on CC News, one by feeding in just the document, and one with the title followed by a new-line and then the rest of the document. We compute word probabilities and only plot those corresponding to the main body of each article. Unsurprisingly, the initial word probabilities are significantly lower when conditioning on the title. However, after 100 words they are only marginally better. We note that this comparison shows that the slight increase in entropy values towards the beginning of the document seen in Figure 2 can be attributed to conditioning on the title. We hypothesize that since news titles only provide a limited amount of information, conditioning on them does not make the document significantly easier to predict. Future work might take an information-structural approach and investigate entropy rate values associated with different parts of articles, such as lede paragraphs or conclusions."
},
{
"figure_ref": [],
"heading": "Discussion",
"publication_ref": [
"b8",
"b43",
"b10",
"b10"
],
"table_ref": [],
"text": "We clarify that the findings presented in this paper do not necessarily invalidate the uniform information density hypothesis. Although entropy rate as measured by neural language models may decline over the course of documents, cognitive measures of surprisal might not decline. For example, the recently proposed lossy-context surprisal model of Futrell et al. (2020) posits that surprisal is computed with respect to an incomplete representation of context, whereas neural language models may make predictions based on lossless representations of their context windows. This perspective is also consistent with recent findings that the base GPT-2 model (124M) outperforms larger GPT-2 and GPT-3 models as a predictor of human reading time (Shain et al., 2022). In particular, these results point to a discrepancy between surprisal values under a Bayes-optimal language model and cognitivelyrelevant measures of surprisal. Despite still being worse than humans at a variety of language-related tasks, we consider it likely that large language models outperform humans at the task of raw language modeling, at least as measured by perplexity. As a result, weaker language models may be better correlated with cognitive measures of surprisal.\nWhether or not our results contradict the entropy rate constancy principle is a matter of interpretation. Genzel and Charniak (2002) would predict that neural language models, which are capable of integrating long-distance context, would exhibit roughly constant entropy rate over the course of documents. Under certain conditions, however, entropy rate as computed by neural language models seems to decline or even exhibit non-monotonic behavior. While this behavior is mostly isolated to the beginnings of documents, it is impossible for entropy rate to decline forever, because word probabilities cannot be less than zero. At the very least, we can conclude that our analyses do not provide clear support in favor of the entropy rate constancy principle proposed by Genzel and Charniak (2002)."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b33",
"b45",
"b16",
"b10",
"b34",
"b36",
"b47"
],
"table_ref": [],
"text": "Most relevant to this work are Giulianelli et al. (2021) and Giulianelli and Fernández (2021), which explore the role of entropy rate constancy in dialogue datasets such as the Spoken British National Corpus (McEnery et al., 2017), the HCRC Map Task (Thompson et al., 1993), and PhotoBook (Haber et al., 2019). Giulianelli and Fernández (2021) follows a similar methodological procedure and computes entropy rate using fine-tuned GPT-2 models, claiming to support the entropy rate constancy hypothesis in the Penn Treebank but not in the dialogue datasets. In contrast, we focus on significantly larger news datasets, which are also more similar to the Penn Treebank data used in Genzel and Charniak (2002), and compute results across a wider range of model sizes. Using larger datasets enables additional fine-tuning and reduces variance in the results; further, our focus on wordlevel surprisal provides additional granularity at the beginnings of documents, where entropy rate is least constant. Finally, we note that perplexity is an extremely sensitive metric (cf. Appendix 6.5), and large variation in results may be attributable to small differences in data. In particular, we do not expect the trends we observe in news articles to always transfer to other domains, such as spoken dialogue in Giulianelli and Fernández (2021).\nRecent work has also sought to connect cognitive theories of efficient communication with techniques in natural language processing; in particular, operationalizations of the uniform information density hypothesis have been connected to natural language decoding (Meister et al., 2020(Meister et al., , 2022) ) and used as regularizers for language model training (Wei et al., 2021). We hope that an improved understanding of entropy rate constancy will inform such applications in the future."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [
"b10",
"b11"
],
"table_ref": [],
"text": "In this work, we computed entropy rate using trigram models and GPT-2, failing to find clear evidence in support of Genzel and Charniak (2002)'s claim of entropy rate constancy in text. We provide results across various model sizes, with and without fine-tuning, and across several datasets, including the Arabic Billion Words Corpus. Our work also provides one of the only analyses of entropy rate constancy in a language besides English, although see Genzel and Charniak (2003) for results in Russian and Spanish. We encourage future work to further investigate the cross-linguistic validity of the uniform information density hypothesis."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [
"b28",
"b0"
],
"table_ref": [],
"text": "One limitation of this work is that since the training data for GPT-2 was not released, it is unknown whether the contents of the NYT Annotated Corpus exist in the pre-training data. We circumvent this issue by also evaluating entropy rates on documents from the Common Crawl News dataset, filtered to those published after 2018. However, it is a possibility that time generalization may complicate the measurement of entropy rate (Lazaridou et al., 2021). Another limitation of our analysis is the sensitivity of word surprisal to small changes in text. As shown in Appendix 6.5, results can significantly change when titles are omitted from fine-tuning and inference. Handling of punctuation and other text preprocessing decisions also plays a large role in the computation of word probabilities, and consequently these decisions may affect any resultant conclusions about entropy rate constancy. Lastly, although the AraGPT-2 training data does contain the Arabic Billion Words Corpus (Antoun et al., 2021), we utlize it due to the unavailability of Arabic-based LLMs and Arabic datasets."
},
{
"figure_ref": [],
"heading": "A Preprocessing and Fine-tuning Details",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we describe preprocessing and finetuning details for each of the three datasets. Finetuning GPT-2 across all datasets was performed with one epoch and a batch size of 8 1024-token length inputs. We outline the fine-tuning and inference times in Table 1. All experiments were run on Quadro RTX 6000 and Quadro RTX 8000 GPUs."
},
{
"figure_ref": [],
"heading": "A.1 NYT",
"publication_ref": [],
"table_ref": [],
"text": "We randomly sample 120K documents from the NYT Annotated Corpus and construct a training set consisting of 100K documents, a validation set consisting of 10K documents, and a test split consisting of 10K documents. We feed in each document with the title as the first line, followed by a newline (\\n) token, and the body of the article afterwards. For non-finetuned runs, we replace the newline token with a colon (\":\")."
},
{
"figure_ref": [],
"heading": "A.2 Common Crawl News",
"publication_ref": [],
"table_ref": [],
"text": "Similar to NYT, we randomly sample 120K documents that were written after 2018. For finetuning, we construct each document by placing the title in the first line, followed by a new line, and the rest of the document afterwards. For the non-finetuned experiments, we place the title in the first line, followed by a colon (\":\"), a new line, and the rest of the document after."
},
{
"figure_ref": [],
"heading": "A.3 Al-Ittihad",
"publication_ref": [],
"table_ref": [],
"text": "We split the Al-Ittihad subset of the Arabic Billion Words Corpus (El-Khair, 2016) into a train split, containing 8551 documents, a test split containing 2000 documents and a validation split containing 1000 documents. We then finetune AraGPT2-Mega (1.5B) (Antoun et al., 2021) by feeding in the title of each document followed by a new line, then the contents of the article after."
},
{
"figure_ref": [],
"heading": "B Constructing a Closed Vocabulary",
"publication_ref": [],
"table_ref": [],
"text": "We follow additional preprocessing steps to construct a closed vocabulary for the trigram models. We first tokenize each document by splitting on whitespace and lowercasing all alphabetical characters. We then form a closed vocabulary by replacing each word which appears in the training data less than five times with the <unk> token. As a result of lowercasing and <unk>ing, we note that exact perplexity values may not be directly comparable to those computed by GPT-2 models, but the general trends between models are still comparable. Indeed, we observe that the n-gram models are occasionally better at predicting words at the beginnings of documents, which we attribute to the frequency of rare words at the beginnings of documents which are often replaced with an <unk> token in closed-vocabulary models."
},
{
"figure_ref": [
"fig_7"
],
"heading": "C Large Language Models",
"publication_ref": [],
"table_ref": [],
"text": "We primarily use GPT-2 for our experiments due to (a) its public availability, (b) the ability to finetune and run inference on standard hardware, and (c) the availability of comparable models for Arabic. We also provide preliminary experiments using the largest GPT-3 model (175B) but do run on all configurations due to cost considerations. We report results on 1000 documents from the NYT Annotated Corpus in Figure 7, observing a similar trend as with GPT-2 models. We use the base davinci model rather than the instruction-tuned text-davinci-003 because our work focuses on the base language modeling objective."
},
{
"figure_ref": [],
"heading": "D Modeling Longer Documents",
"publication_ref": [
"b37",
"b10"
],
"table_ref": [],
"text": "We also attempt to feed in longer documents and therefore compute entropy rates on WikiText-2 (Merity et al., 2016) (1.5B), up to the first five thousand words. We run this experiment on WikiText because it has a larger number of very long documents than our primary evaluation corpora. However, we note that entropy values are not accurate due to a fixed context size and still suffer from noisiness due to a lafck of long-form documents. even past 1000 words. We note that this result is not an accurate representation of entropy values, due to the fixed 1024 context window of GPT-2. In order to get around the fixed context window, we use a stride of 64 tokens. As a result, we expect entropy values to increase in the long run, following Genzel and Charniak (2002)'s argument for n-grams. We attribute the noise in these results to the lack of documents with more than 5000 words."
},
{
"figure_ref": [],
"heading": "E Significance Testing",
"publication_ref": [
"b31",
"b26"
],
"table_ref": [],
"text": "We report entropy values on a per-word basis for both n-gram models and GPT-2. We also apply the non-parametric Mann-Kendall test (Mann, 1945;Kendall, 1948) to determine whether entropy rate is monotonically increasing or decreasing throughout the course of a document. We note that this method is not intended to compare the relative sizes of trends and that it is sensitive to hyperparameters such as the length of perplexity timeseries and choice of tokenization scheme. We omit these findings from the main body of the paper primarily due to how sensitive they are to the x-axis cutoff. We present the results and significance figures in Table 2. We further note that other methods, such as correlation coefficients or mixed effects models as used in Giulianelli and Fernández (2021) are also highly sensitive to the length of documents, especially since entropy is least constant at the beginnings of documents. As a result of this sensitivity, we focus primarily on qualitative evaluations of the observed trends rather than on significance tests."
},
{
"figure_ref": [],
"heading": "Acknowledgments",
"publication_ref": [],
"table_ref": [],
"text": "Nicholas Tomlin is supported by a NSF Graduate Research Fellowship, as well as grants from the DARPA LwLL and XAI programs. We are grateful to Uriel Cohen Priva, Roma Patel, the members of the Berkeley NLP Group, as well as anonymous reviewers for feedback on earlier drafts of this paper."
},
{
"figure_ref": [],
"heading": "Dataset",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Model",
"publication_ref": [],
"table_ref": [],
"text": "Fine-tuning Time Inference Time"
},
{
"figure_ref": [],
"heading": "CC-News",
"publication_ref": [],
"table_ref": [],
"text": "GPT-2 Small (124M) 3 hours 7 minutes GPT-2 Medium (345M) 7. 5 \nTable 2: Results of running the Mann-Kendall test on each of the experimental conditions in this paper. In general, we observe a decreasing trend for neural models, and an increasing trend for n-gram models. Although this test is non-parametric, we caution that results are highly dependent on the length of the input time-series."
}
] | The uniform information density (UID) hypothesis states that humans tend to distribute information roughly evenly across an utterance or discourse. Early evidence in support of the UID hypothesis came from Genzel and Charniak (2002), which proposed an entropy rate constancy principle based on the probability of English text under n-gram language models. We re-evaluate the claims of Genzel and Charniak (2002) with neural language models, failing to find clear evidence in support of entropy rate constancy. We conduct a range of experiments across datasets, model sizes, and languages and discuss implications for the uniform information density hypothesis and linguistic theories of efficient communication more broadly. | Revisiting Entropy Rate Constancy in Text | [
{
"figure_caption": "Figure 1 :1Figure 1: Entropy rate of the Penn Treebank under a smoothed trigram model and a GPT-2 XL model (Radford et al., 2019), averaged across documents per word position.Genzel and Charniak (2002) showed that entropy rate increased under n-gram models and predicted that it would remain constant in models which can condition on long-range context. We replicate the former result but do not find clear evidence supporting the latter.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: Entropy rate of the Common Crawl News Dataset, the NYT Annotated Corpus and the Al-Ittihad subset of the Arabic Billion Words Corpus under a smoothed trigram model and a GPT-2 model (1.5B)(Radford et al., 2019), averaged across documents at each word position. We observe a roughly increasing trend for the trigram model across all three datasets, and a variety of trends for the GPT-2 models.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure3: Distribution of document sizes for the CC News, NYT and Al-Ittihad test splits. Throughout this paper, we present word probabilities for the first 500 tokens of each article due to the lack of longer articles in each of these datasets. We provide additional results on longer documents in Appendix D.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure4: Entropy rate of NYT under four GPT-2 model sizes (124M, 345M, 774M, 1.5B). We note lower entropy values as model size increases but observe a consistent decline in surprisal values across all model sizes.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 :5Figure5: Entropy rate of CC News, NYT and Al-Ittihad on a fine-tuned GPT-2 XL (1.5B) compared to a non-finetuned model. We note that entropy values sharply decline and have lower values on the fine-tuned models, most likely due to domain adaptation. The difference between the two models is largest at the beginnings of documents.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Figure 6 :6Figure6: Entropy rate of CC News on two fine-tuned GPT-2 XL (1.5B) models, one fine-tuned with the title and one fine-tuned without. We omit entropy values for the title in this plot, but condition the model on the title.",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "Figure 7 :7Figure7: Entropy rate of NYT under GPT-3 (davinci) and GPT-2 XL (1.5B). Although GPT-3 perplexity values are notably lower than those of GPT-2, the general trend is similar. Neither model was fine-tuned for a direct comparison.",
"figure_data": "",
"figure_id": "fig_7",
"figure_label": "7",
"figure_type": "figure"
},
{
"figure_caption": "using GPT-2 XL. As shown in Figure8, these results show a non-monotonic trend",
"figure_data": "",
"figure_id": "fig_8",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 8 :8Figure8: Entropy rate of WikiText-2 on a GPT-2 XL (1.5B), up to the first five thousand words. We run this experiment on WikiText because it has a larger number of very long documents than our primary evaluation corpora. However, we note that entropy values are not accurate due to a fixed context size and still suffer from noisiness due to a lafck of long-form documents.",
"figure_data": "",
"figure_id": "fig_9",
"figure_label": "8",
"figure_type": "figure"
}
] | Vivek Verma; Nicholas Tomlin; Dan Klein | [
{
"authors": "Fady Wissam Antoun; Hazem Baly; Hajj",
"journal": "Association for Computational Linguistics",
"ref_id": "b0",
"title": "AraGPT2: Pre-trained transformer for Arabic language generation",
"year": "2021"
},
{
"authors": "Fatemeh Torabi; Asr ; Vera Demberg",
"journal": "",
"ref_id": "b1",
"title": "Uniform surprisal at the level of discourse relations: Negation markers and discourse connective omission",
"year": "2015"
},
{
"authors": "Matthew Aylett; Alice Turk",
"journal": "Language and Speech",
"ref_id": "b2",
"title": "The smooth signal redundancy hypothesis: A functional explanation for relationships between redundancy, prosodic prominence, and duration in spontaneous speech",
"year": "2004"
},
{
"authors": " Mp Aylett",
"journal": "",
"ref_id": "b3",
"title": "Stochastic suprasegmentals: Relationships between redundancy, prosodic structure and syllabic duration",
"year": "1999"
},
{
"authors": "Alan Bell; Daniel Jurafsky; Eric Fosler-Lussier; Cynthia Girand; Michelle Gregory; Daniel Gildea",
"journal": "The Journal of the acoustical society of America",
"ref_id": "b4",
"title": "Effects of disfluencies, predictability, and utterance position on word form variation in english conversation",
"year": "2003"
},
{
"authors": "Harm Brouwer; Hartmut Fitz; John Hoeks",
"journal": "Association for Computational Linguistics",
"ref_id": "b5",
"title": "Modeling the noun phrase versus sentence coordination ambiguity in Dutch: Evidence from surprisal theory",
"year": "2010"
},
{
"authors": "M Thomas; Joy A Cover; Thomas",
"journal": "John Wiley & Sons",
"ref_id": "b6",
"title": "Elements of information theory",
"year": "2012"
},
{
"authors": "Ibrahim Abu El-Khair",
"journal": "",
"ref_id": "b7",
"title": "1.5 billion words arabic corpus",
"year": "2016"
},
{
"authors": "Richard Futrell; Edward Gibson; Roger P Levy",
"journal": "Cognitive science",
"ref_id": "b8",
"title": "Lossy-context surprisal: An informationtheoretic model of memory effects in sentence processing",
"year": "2020"
},
{
"authors": "Richard Futrell; Kyle Mahowald; Edward Gibson",
"journal": "Proceedings of the National Academy of Sciences",
"ref_id": "b9",
"title": "Large-scale evidence of dependency length minimization in 37 languages",
"year": "2015"
},
{
"authors": "Dmitriy Genzel; Eugene Charniak",
"journal": "Association for Computational Linguistics",
"ref_id": "b10",
"title": "Entropy rate constancy in text",
"year": "2002"
},
{
"authors": "Dmitriy Genzel; Eugene Charniak",
"journal": "",
"ref_id": "b11",
"title": "Variation of entropy and parse trees of sentences as a function of the sentence number",
"year": "2003"
},
{
"authors": "Edward Gibson; Richard Futrell; Julian Jara-Ettinger; Kyle Mahowald; Leon Bergen; Sivalogeswaran Ratnasingam; Mitchell Gibson; Steven T Piantadosi; Bevil R Conway",
"journal": "Proceedings of the National Academy of Sciences",
"ref_id": "b12",
"title": "Color naming across languages reflects color use",
"year": "2017"
},
{
"authors": "Edward Gibson; Richard Futrell; Steven P Piantadosi; Isabelle Dautriche; Kyle Mahowald; Leon Bergen; Roger Levy",
"journal": "Trends in cognitive sciences",
"ref_id": "b13",
"title": "How efficiency shapes human language",
"year": "2019"
},
{
"authors": "Mario Giulianelli; Raquel Fernández",
"journal": "Association for Computational Linguistics",
"ref_id": "b14",
"title": "Analysing human strategies of information transmission as a function of discourse context",
"year": "2021"
},
{
"authors": "Mario Giulianelli; Arabella Sinclair; Raquel Fernández",
"journal": "Association for Computational Linguistics",
"ref_id": "b15",
"title": "Is information density uniform in task-oriented dialogues",
"year": "2021"
},
{
"authors": "Janosch Haber; Tim Baumgärtner; Ece Takmaz; Lieke Gelderloos; Elia Bruni; Raquel Fernández",
"journal": "Association for Computational Linguistics",
"ref_id": "b16",
"title": "The PhotoBook dataset: Building common ground through visually-grounded dialogue",
"year": "2019"
},
{
"authors": "Michael Hahn; Judith Degen; Dan Noah D Goodman; Richard Jurafsky; Futrell",
"journal": "",
"ref_id": "b17",
"title": "An informationtheoretic explanation of adjective ordering preferences",
"year": "2018"
},
{
"authors": "Michael Hahn; Dan Jurafsky; Richard Futrell",
"journal": "Proceedings of the National Academy of Sciences",
"ref_id": "b18",
"title": "Universals of word order reflect optimization of grammars for efficient communication",
"year": "2020"
},
{
"authors": "John Hale",
"journal": "",
"ref_id": "b19",
"title": "A probabilistic Earley parser as a psycholinguistic model",
"year": "2001"
},
{
"authors": "Felix Hamborg; Norman Meuschke; Corinna Breitinger; Bela Gipp",
"journal": "",
"ref_id": "b20",
"title": "news-please: A generic news crawler and extractor",
"year": "2017"
},
{
"authors": "John A Hawkins",
"journal": "",
"ref_id": "b21",
"title": "Language universals and the performance-grammar correspondence hypothesis",
"year": "2009"
},
{
"authors": "Florian Jaeger",
"journal": "Cognitive psychology",
"ref_id": "b22",
"title": "Redundancy and reduction: Speakers manage syntactic information density",
"year": "2010"
},
{
"authors": "Florian Jaeger; Roger P Levy",
"journal": "",
"ref_id": "b23",
"title": "Speakers optimize information density through syntactic reduction",
"year": "2007"
},
{
"authors": "Frank Keller",
"journal": "Association for Computational Linguistics",
"ref_id": "b24",
"title": "The entropy rate principle as a predictor of processing effort: An evaluation against eye-tracking data",
"year": "2004"
},
{
"authors": "Charles Kemp; Yang Xu; Terry Regier",
"journal": "Annual Review of Linguistics",
"ref_id": "b25",
"title": "Semantic typology and efficient communication",
"year": "2018"
},
{
"authors": "Maurice George; Kendall ",
"journal": "",
"ref_id": "b26",
"title": "Rank correlation methods",
"year": "1948"
},
{
"authors": "Urvashi Khandelwal; He He; Peng Qi; Dan Jurafsky",
"journal": "Association for Computational Linguistics",
"ref_id": "b27",
"title": "Sharp nearby, fuzzy far away: How neural language models use context",
"year": "2018"
},
{
"authors": "Angeliki Lazaridou; Adhi Kuncoro; Elena Gribovskaya; Devang Agrawal; Adam Liska; Tayfun Terzi; Mai Gimenez; Cyprien De Masson D'autume; Tomas Kocisky; Sebastian Ruder; Dani Yogatama; Kris Cao; Susannah Young; Phil Blunsom",
"journal": "",
"ref_id": "b28",
"title": "Mind the gap: Assessing temporal generalization in neural language models",
"year": "2021"
},
{
"authors": "Curran Associates; Inc ",
"journal": "",
"ref_id": "b29",
"title": "",
"year": ""
},
{
"authors": "Roger Levy",
"journal": "Cognition",
"ref_id": "b30",
"title": "Expectation-based syntactic comprehension",
"year": "2008"
},
{
"authors": " Henry B Mann",
"journal": "Econometrica: Journal of the Econometric Society",
"ref_id": "b31",
"title": "Nonparametric tests against trend",
"year": "1945"
},
{
"authors": "Mitchell Marcus; Beatrice Santorini; Mary Ann Marcinkiewicz",
"journal": "",
"ref_id": "b32",
"title": "Building a large annotated corpus of english: The penn treebank",
"year": "1993"
},
{
"authors": "Tony Mcenery; Robbie Love; Vaclav Brezina",
"journal": "International Journal of Corpus Linguistics",
"ref_id": "b33",
"title": "Introduction: Compiling and analysing the spoken british national corpus 2014",
"year": "2017"
},
{
"authors": "Clara Meister; Ryan Cotterell; Tim Vieira",
"journal": "Association for Computational Linguistics",
"ref_id": "b34",
"title": "If beam search is the answer, what was the question",
"year": "2020"
},
{
"authors": "Clara Meister; Tiago Pimentel; Patrick Haller; Lena Jäger; Ryan Cotterell; Roger Levy",
"journal": "Association for Computational Linguistics",
"ref_id": "b35",
"title": "Revisiting the Uniform Information Density hypothesis",
"year": "2021"
},
{
"authors": "Clara Meister; Tiago Pimentel; Gian Wiher; Ryan Cotterell",
"journal": "",
"ref_id": "b36",
"title": "Typical decoding for natural language generation",
"year": "2022"
},
{
"authors": "Stephen Merity; Caiming Xiong; James Bradbury; Richard Socher",
"journal": "",
"ref_id": "b37",
"title": "Pointer sentinel mixture models",
"year": "2016"
},
{
"authors": "Sabrina J Mielke",
"journal": "",
"ref_id": "b38",
"title": "Can you compare perplexity across different segmentations?",
"year": "2019"
},
{
"authors": "O' Joe; Jacob Connor; Andreas",
"journal": "Association for Computational Linguistics",
"ref_id": "b39",
"title": "What context features can transformer language models use",
"year": "2021"
},
{
"authors": "Harry Steven T Piantadosi; Edward Tily; Gibson",
"journal": "Proceedings of the National Academy of Sciences",
"ref_id": "b40",
"title": "Word lengths are optimized for efficient communication",
"year": "2011"
},
{
"authors": "Alec Radford; Jeff Wu; Rewon Child; David Luan; Dario Amodei; Ilya Sutskever",
"journal": "",
"ref_id": "b41",
"title": "Language models are unsupervised multitask learners",
"year": "2019"
},
{
"authors": "Evan Sandhaus",
"journal": "",
"ref_id": "b42",
"title": "The New York Times Annotated Corpus",
"year": "2008"
},
{
"authors": "Cory Shain; Clara Meister; Tiago Pimentel; Ryan Cotterell; Roger Philip; Levy ",
"journal": "",
"ref_id": "b43",
"title": "Large-scale evidence for logarithmic effects of word predictability on reading time",
"year": "2022"
},
{
"authors": "Claude Elwood; Shannon ",
"journal": "The Bell system technical journal",
"ref_id": "b44",
"title": "A mathematical theory of communication",
"year": "1948"
},
{
"authors": "Henry S Thompson; Anne Anderson; Ellen Gurman Bard; Gwyneth Doherty-Sneddon; Alison Newlands; Cathy Sotillo",
"journal": "",
"ref_id": "b45",
"title": "The HCRC map task corpus: Natural dialogue for speech recognition",
"year": "1993-03-21"
},
{
"authors": "Noam Vaswani; Ashish Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Lukasz Kaiser; Illia Polosukhin",
"journal": "",
"ref_id": "b46",
"title": "Attention is all you need",
"year": "2017"
},
{
"authors": "Jason Wei; Clara Meister; Ryan Cotterell",
"journal": "Association for Computational Linguistics",
"ref_id": "b47",
"title": "A cognitive regularizer for language modeling",
"year": "2021"
},
{
"authors": "Yang Xu; David Reitter",
"journal": "Cognition",
"ref_id": "b48",
"title": "Information density converges in dialogue: Towards an informationtheoretic model",
"year": "2018"
},
{
"authors": "Noga Zaslavsky; Charles Kemp; Terry Regier; Naftali Tishby",
"journal": "Proceedings of the National Academy of Sciences",
"ref_id": "b49",
"title": "Efficient compression in color naming and its evolution",
"year": "2018"
},
{
"authors": "George K Zipf",
"journal": "Addison-Wesley Press",
"ref_id": "b50",
"title": "Human behavior and the principle of least effort",
"year": "1949"
}
] | [
{
"formula_coordinates": [
2,
306.14,
630.22,
114.66,
10.63
],
"formula_id": "formula_0",
"formula_text": "H(Y i ) = H(X i | C i , L i )"
},
{
"formula_coordinates": [
2,
332.76,
708.95,
192.38,
27.17
],
"formula_id": "formula_1",
"formula_text": "H(Y i ) = H(X i | C i , L i ) (1) = H(X i | L i ) -I(X i ; C i , L i ) (2)"
},
{
"formula_coordinates": [
4,
82.55,
509.7,
207.32,
47.4
],
"formula_id": "formula_2",
"formula_text": "P (x i | x 1 ...x i-1 ) ≈ P (x i | x i-2 , x i-1 ) (3) = λ 1 P (x i | x i-2 , x i-1 ) (4) + λ 2 P (x i | x i-1 )(5)"
},
{
"formula_coordinates": [
4,
161.18,
562.09,
128.69,
13.39
],
"formula_id": "formula_3",
"formula_text": "+ (1 -λ 1 -λ 2 ) P (x i ) (6)"
},
{
"formula_coordinates": [
4,
110.12,
646.45,
179.74,
25.5
],
"formula_id": "formula_4",
"formula_text": "P (x i |x 1 ...x i-1 ) = C(x 1 ...x i ) C(x 1 ...x i-1 )(7)"
},
{
"formula_coordinates": [
4,
368.01,
584.97,
157.13,
10.77
],
"formula_id": "formula_5",
"formula_text": "P θ (x i | x 1 , . . . , x i-1 )(8)"
},
{
"formula_coordinates": [
4,
342.54,
705.74,
178.36,
35.19
],
"formula_id": "formula_6",
"formula_text": "log P θ (w k ) = end(k) i=start(k) log P θ (x i )(9"
},
{
"formula_coordinates": [
5,
114.18,
424.3,
175.69,
29.64
],
"formula_id": "formula_7",
"formula_text": "f (i) = 1 |W | • w∈W log P θ (w i )(10)"
},
{
"formula_coordinates": [
5,
111.75,
695.94,
177.99,
19.74
],
"formula_id": "formula_8",
"formula_text": "H(X) = lim n→∞ 1 n H(X1, X2, ..., Xn)(11)"
}
] |
10.18653/v1/2022.acl-short.1 | 2023-05-24 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b33",
"b8",
"b14",
"b3",
"b23",
"b24",
"b19",
"b21",
"b1",
"b24",
"b23"
],
"table_ref": [],
"text": "The Transformer architecture (Vaswani et al., 2017) has changed the landscape of recent natural language processing approaches by enabling the pretraining of state-of-the-art large language models (LLM) (Devlin et al., 2019;He et al., 2020;Brown et al., 2020). However, fine-tuning and storing full copies of LLMs can consume prohibitively large quantities of resources. Parameter-efficient finetuning (PEFT) methods such as prefix-tuning (Li and Liang, 2021;He et al., 2021a;Liu et al., 2022) address these concerns by reducing the number of trainable parameters. Prefix-tuning can tune 0.01% of parameters and still match the performance of regular fine-tuning (updating all model parameters). PEFT has been investigated for tasks with inputs consisting of sentences, sentence-pair, or sequences that fit within the typical LLM maximum tokens. However, the performance of PEFT for tasks with longer textual sequences has been overlooked. In this work, we investigate this oversight and provide evidence suggesting that the gap between PEFT and regular fine-tuning is substantial when modelling long sequences. As shown in Table 1, prefix-tuning underperforms fine-tuning on long sequence classification tasks, Hyperpartisan (Kiesel et al., 2019) and 20-newsgroups (Lang, 1995), when used with the popular long-document model Longformer (Beltagy et al., 2020).\nIn this paper, we propose a simple and effective method, prefix-propagation, which consistently improves the performance of PEFT for long sequence models. Unlike prefix-tuning, prefix-propagation propagates the hidden states corresponding to prefixes through the attention computation. This allows for the prefixes hidden states to dynamically change as the input propagates through each layer.\nTo further understand prefix propagation, we investigate the reliability of the model's predictions by performing analyses on calibration. Lastly, we conduct study on prefix-based methods in terms of kernel attention to strengthen their theoretical value.\nIn summary, our contributions are as follows:\n... Figure 1: Illustration of the differences between (a) prefix-propagation (ours) (b) and prefix-tuning (Liu et al., 2022;Li and Liang, 2021). Blue blocks denote trainable prompts, and \"Transformer Layer\" represents the computation done in a layer of the pre-trained LLM. Note that in prefix-propagation (a), the summation of prefixes continues for layers beyond 3, up to n. This operation is encapsulated by the ellipses. In prefix-tuning (b), prefixes in subsequent layers do not depend on hidden states from past layers (they are simply overwritten).\n• We study PEFT for long documents and show that prefix-tuning is significantly inferior to fine-tuning in this scenario. To the best of our knowledge, this is the first work to focus on PEFT for long documents.\n• We introduce prefix-propagation, which consistently improves the performance over prefix turning on the different long document datasets, while using 50% fewer parameters.\n• We study the reliability of the predictions by performing analyses on calibration and show that models tuned with prefix-propagation are better calibrated.\n• We elucidate the relationship between prefixpropagation and kernel attention and perform an ablation study that utilizes this insight."
},
{
"figure_ref": [],
"heading": "Related Works",
"publication_ref": [
"b7",
"b18",
"b28",
"b6",
"b1",
"b37",
"b20",
"b31",
"b1",
"b23",
"b24",
"b16",
"b2",
"b5"
],
"table_ref": [],
"text": "Long Sequence Models Numerous methods have been proposed to reduce the complexity of attention from O(n 2 ) to O(n) such as kernel approximations (Choromanski et al., 2020;Katharopoulos et al., 2020;Peng et al., 2021) and fixed (Child et al., 2019;Beltagy et al., 2020;Zaheer et al., 2020) or learned (Kitaev et al., 2020) sparse attention patterns. For a broader summary, please refer to Tay et al. (2022). In this work, we use Longformer (Beltagy et al., 2020). To linearize attention complexity, Longformer employs sliding window attention while globally attending to relatively few special tokens.\nParameter-Efficient Tuning Inspired by the success of manual prompting (Brown et al., 2020), prefix-tuning (Li andLiang, 2021;Liu et al., 2022) prepends trainable \"soft\" prompts to an input sequence. Although further PEFT methods have since been introduced (He et al., 2021a;Hu et al., 2021;Ben Zaken et al., 2022), we focus on adapting prefix-tuning. We note that our adaptation does not violate orthogonality and thus prefixpropagation can still be compounded with other PEFT methods as proposed in the UnifiedPET framework (He et al., 2021a), likely yielding similar performance gains. We leave the empirical validation of this hypothesis for future work.\nOut work also adheres to the key motivation of the recent PEFT method, inducer-tuning (Chen et al., 2022), which is that optimal prefixes should be close to queries within their latent space. We derive queries, keys, and values from the same prefix token, limiting the distance that separates them."
},
{
"figure_ref": [],
"heading": "Prefix Propagation",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Methodology",
"publication_ref": [
"b37"
],
"table_ref": [],
"text": "In this section we introduce prefix-propagation, which, unlike prefix-tuning, propagates the hidden states corresponding to prefixes through the attention computation. This allows for the prefixes hidden states to dynamically change as the input propagates through each layer. Prefix-propagation and its predecessor, prefix-tuning are depicted in Figure 1a embeddings) to the input sequence (blue blocks in top left of Figure 1a). Then, before every subsequent layer, we sum new trainable matrices onto the first j embeddings corresponding to the prefixes (denoted by the sum operators in Figure 1a). By propagating instead of overwriting, we halve the number of parameters trained while simultaneously improving performance on long-document tasks.\nWe now formalize prefix-propagation. Multiheaded attention processes query, key, and value matrices derived from a sequence C ∈ R m×d with length m and embeddings of size d. Our method modifies traditional attention by concatenating a prefix P ∈ R j×d of length j to the sequence:\nH l,i = Attn(D (l) W (l,i) q , D (l) W (l,i) k , D (l) W (l,i) v ) (1) D (l) = cat(P (l) , C) if l = 1 cat(P (l) + C[:j, :], C[j:, :]) if l > 1 where inputs C are projected through pre-trained weight matrices W (l,i) q , W (l,i) k , W (l,i) v\n∈ R d×d h per layer l and head i yielding the output of the attention head, H ∈ R (j+m)×d h . The prefixes are concatenated for the first layer (l = 1) and summed to their corresponding hidden states for the remaining layers (l > 1). We do not continually concatenate new prefixes to the sequence to avoid increasing the sequence length after each layer.\nFor both prefix-tuning and prefix-propagation, prefixes (keys and values) are globally attended to by all queries. Unlike prefix-tuning however, our method concatenates additional hidden states before the hidden states C are projected by\nW (i) k and W (i)\nv . By doing so, prefix-propagation modifies query matrices, allowing prefixes to attend to other hidden states globally, thereby increasing representation capability. This approach is somewhat analogous to the external global tokens inserted in the BigBird-ETC model (Zaheer et al., 2020). By attending to other tokens, the prefixes can act as special storage tokens, which is particularly useful in the restricted regime of long-document modelling where relatively few tokens have global context. Conversely, prefix-tuning only concatenates trained key and value matrices, P k , P v ∈ R j×d h , statically to the sequence:\nH l,i = Attn(CW (l,i) q , cat(P (l,i) k , CW (l,i) k ), cat(P (l,i) v , CW (l,i) v ))\n(2)\nSince our method has a single prefix matrix, P instead of separate P k and P v matrices, we reduce the number of trained parameters by 50%."
},
{
"figure_ref": [],
"heading": "Calibration",
"publication_ref": [
"b26",
"b11"
],
"table_ref": [],
"text": "We further study the proposed prefix-propagation method to understand the reliability of model's predictions through calibration. Well-calibrated models output confidence scores that closely match the models' accuracy. Either over-confident or underconfident models are undesirable. Calibration has widely been overlooked in PEFT methods. To quantify calibration in our work, we use expected calibration error (ECE), which bins predictions based on model confidence and compares them to accuracy (Pakdaman Naeini et al., 2015;Guo et al., 2017)."
},
{
"figure_ref": [],
"heading": "Kernel Decomposition",
"publication_ref": [
"b32",
"b32"
],
"table_ref": [],
"text": "Traditional attention is analogous to applying a kernel smoother over inputs (Tsai et al., 2019).\nMotivated by this insight, we reformulate prefixpropagation as a sum of kernelized attention modules. Separating the modules introduces flexibility in two ways: (1) Their individual kernel forms can be mixed and matched and (2) A hyperparameter scale factor α can be applied to the prefix component to increase or decrease its weighting. Equation 3 defines kernel decomposition for prefixpropagation 2 :\nH = Kern(cat(P, C)W q , CW k , CW v ) + (α)Kern(cat(P, C)W q , P W k , P W v ) (3)\nwhere Kern refers to kernel attention as formulated in (Tsai et al., 2019). The first term results from attending to the original sequence, C, and the second comes from attending to the prefixes, P . We provide the derivation of Equation 3 and the full definition of kernel attention in Appendix A.\nOur main motivation for presenting prefix decomposition is to establish foundational knowledge and guide future research. Ergo, we restrict experiments in this initial presentation to using just the default exponential kernel (Appendix A)."
},
{
"figure_ref": [],
"heading": "Experiments and Results",
"publication_ref": [
"b12",
"b21",
"b19",
"b35",
"b1",
"b1",
"b25"
],
"table_ref": [],
"text": "Datasets We evaluate our approach on three longdocument classification tasks: ArXiv (He et al., 2019), an 11-class classification task composed of academic research papers, the 20-newsgroups (Lang, 1995) classification task consisting of mailing lists that fall into one of 20 classes, and the Hyperpartisan dataset, a binary classification task for extremist news classification (Kiesel et al., 2019). We also run experiments on WikiHop (Welbl et al., 2018), a long-document reading comprehension task requiring multi-step reasoning.\nDue to compute limitations inherent to working with long documents, with the exception of Hyperpartisan, we only report a single run for each task. This mimics the original Longformer reporting scheme (Beltagy et al., 2020). For Hyperpartisan, the smallest of the datasets, we report mean metrics averaged over five seeds.\nBaselines As a baseline, we fine-tune Longformer-base (approx.\n149M parameters) as closely as possible to Beltagy et al. (2020). For PEFT, we evaluate prefix-tuning on Longformer-base and RoBERTa-base (approx. 125M parameters) (Liu et al., 2019). 2 We omit layer, l and head, i for brevity."
},
{
"figure_ref": [],
"heading": "Method",
"publication_ref": [],
"table_ref": [],
"text": "ArXiv HY. NG. More details on dataset sizes, pre-processing, and hyperparameters are in Appendix B."
},
{
"figure_ref": [],
"heading": "RoBERTa",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Results and Discussion",
"publication_ref": [
"b10"
],
"table_ref": [
"tab_4"
],
"text": "Across all tasks, our results in Table 2 verify that prefix-tuning is inferior to fine-tuning long sequences. Conversely, prefix-propagation consistently outperforms prefix-tuning and is comparable to fine-tuning on most tasks. Prefix propagation also performs competitively on Hyperpartisan, a relatively small dataset with only 625 samples. This is in contrast to prefix-tuning, which is known to underperform in low-data settings (Gu et al., 2022). Because we ran multiple seeds on Hyperpartisan, we also found that prefix-propagation's better performance relative to prefix-tuning is statistically significant (p < 0.05, using a single-tailed t-test). We do not have multiple samples to run these tests for larger datasets, but we emphasize that Hyperpartisan likely has the most variance and yet it is still statistically significant. We suspect that prefixpropagation's performance exceeds prefix-tuning because propagated prefixes can transmit global context across multiple layers, possibly modelling more expressive abstractions.\nWe note one exception where prefix-based methods still leave room for improvement: multiplechoice question answering on WikiHop. We hypothesize that prefix methods have insufficient capacity to properly model complex long-document multi-step question answering.\nWe also observe that prefix-based methods, and especially prefix-propagation, achieve better calibration than fine-tuning, as shown in Table 3. Unlike prefix-tuning however, prefix-propagation effectively balances calibration with accuracy metrics. The calibration of fine-tuning deteriorates as training progresses (Figure 4 "
},
{
"figure_ref": [],
"heading": "Micro F1",
"publication_ref": [
"b17"
],
"table_ref": [],
"text": "Figure 2: Violin plot of Micro F1 Score for five different seeds on the Hyperpartisan task. White dots, gray boxes, and gray lines are the medians, interquartile ranges, and ranges respectively. Width of the five violin shapes show the probability densities for the corresponding F1score. All methods tune Longformer-base except \"R Prefix\", which is prefix-tuning on RoBERTa-base.\nforgetting (Jagielski et al., 2022).\nAs an initial test for our ongoing prefixpropagation kernel study, we show results on Hyperpartisan in Figure 2. The kernelized version of prefix-propagation achieves the best single-run performance, but has higher variance than fine-tuning and prefix-propagation which necessitates further research."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "Our research focuses on parameter efficient tuning for long documents tasks. We introduce prefix-propagation, which consistently improves performance over prefix-turning on long document datasets, while using 50% fewer parameters. We study the reliability of the predictions by performing analyses on calibration and show that models tuned with prefix-propagation are better calibrated. We lastly explicate prefix-propagation from a kernel perspective, uncovering insights for future PEFT research."
},
{
"figure_ref": [],
"heading": "Limitations Scope",
"publication_ref": [
"b7",
"b29"
],
"table_ref": [],
"text": "This short paper serves as an initial step toward PEFT for long-document models. As such, our evaluated scope of models, tasks, datasets, and kernel variations is limited. We acknowledge the need to experiment across broader settings and hope our work provides a foundation for others to build on.\nFuture experiments should analyze the validity and efficacy of using prefix-propagation with other long-sequence models to determine whether the prefix modality is suitable for non-sparse attention approximations. For example, would the projection of prefix vectors using a random feature map as in Choromanski et al. (2020) result in an excessive loss of information for these critical tokens?\nRegarding tasks and datasets, the performance degradation in prefix methods for WikiHop deserves significant attention. Verifying whether this extends to other reading comprehension and question-answering tasks will assist in guiding future research efforts. We restricted our research to the encoder-only version of Longformer, but using the encoder-decoder version, LED would enable analysis of sequence-to-sequence tasks. The SCROLLS benchmark (Shaham et al., 2022) would be a good starting point for this analysis since it includes an LED baseline.\nCombining prefix and kernel methods is an ongoing research effort and there are several questions we plan to address: (1) What are the effects of swapping the default exponential kernel with other variants such as linear, polynomial, and RBF? (2) Does making the α scale parameter trainable improve performance? (3) Can we have a separate scale parameter for each query and should they be trainable? (4) Is this approach effective for modalities other than long-document? (5) Can we separate other components of attention into modular kernels (e.g. local and global kernels for sparse attention)?"
},
{
"figure_ref": [],
"heading": "Robustness",
"publication_ref": [
"b1"
],
"table_ref": [],
"text": "The size and nature of long-sequence tasks often resulted in long run times for the larger datasets ArXiv, 20-newsgroup and WikiHop. Consequently, we report results of one seed after doing a hyperparameter search for learning rate. This aligns with the reporting system of the original Longformer paper (Beltagy et al., 2020) but greater assurance in all long-sequence task performance could be achieved by accumulating results over several seeds. The size of datasets and iteration over several epochs somewhat mitigate this concern."
},
{
"figure_ref": [],
"heading": "Ethics Statement",
"publication_ref": [
"b30"
],
"table_ref": [],
"text": "Our work helps to address the environmental and equitable distribution concerns of LLMs (Strubell et al., 2019). All PEFT variants attempt to reduce resource requirements, primarily via GPU mem-ory consumption and storage requirements. By applying prefix-tuning and our variation, prefixpropagation to long-document models we limit carbon emissions and increase accessibility for lowresource groups. We note that prefix-propagation neither exacerbates nor alleviates other ethical risks such as biases regarding gender, race, religion, etc. that are often embedded in pre-trained LLMs. If such biases exist in the pre-trained model, they will be propagated to downstream tasks regardless of tuning method."
},
{
"figure_ref": [],
"heading": "A Kernel Decomposition Derivation",
"publication_ref": [
"b25",
"b12",
"b32"
],
"table_ref": [
"tab_6"
],
"text": "In the unified framework of He et al. (2021b), we can write the first layer l = 1 attention mechanism of prefix-propagation as:\nH l,i = Attn(cat(P (l) , C)W (l)(i) q ,(4) cat\n(P (l) , C)W (l)(i) k , cat(P (l) , C)W (l)(i) v )\nwhere P is a trained prefix for each downstream task. Omitting layer and head indices and using D = cat(P, C) for brevity, we can rewrite Equation 4 as:\nH = Attn(DW q , cat(P, C)W k , cat(P, C)W v ) = softmax(DW q cat(P W k , CW k )) P W v CW v = (1-λ(C))softmax(DW q W ⊤ k C ⊤ )CW v +λ(C)softmax(DW q W ⊤ k P ⊤ )P W v = (1-λ(C))Attn(DW q , CW k , CW v ) +λ(C)Attn(DW q , P W k , P W v ) = (1-λ(C))Attn(cat(P, C)W q , CW k , CW v ) +λ(C)Attn(cat(P, C)W q , P W k , P W v )(5)\nwhere λ(C) is a scalar (dependent on C) to normalize softmax over the sequence and the prefixes and is computed by:\nλ(C) = i DW q W k ⊤ P ⊤ i DW q W ⊤ k P ⊤ + j DW q W ⊤ k C ⊤(6)\nWe consider the two terms of Equation 5 as kernelized attention modules which brings us back to the complete kernel decomposition:\nH = Kern(cat(P, C)W q , CW k , CW v ) + (α)Kern(cat(P, C)W q , P W k , P W v ) (7)\nwhere α is an introduced hyperparameter that replaces the fixed weighting of λ. This change allows us to explicitly increase the weighting of prefixes 1.9.0 MIT RoBERTa (Liu et al., 2019) 6base MIT Longformer (Beltagy et al., 2020) 7 base Apache 2.0 P-Tuning (Liu et al., 2022) 8 2.0 Apache 2.0 ArXiv (He et al., 2019) 9 no_ref Unspecified Hyperpartisan (Kiesel et al., 2019) 10 1.0 CC BY 4.0 20-newsgroup (Lang, 1995) 11 1.0 Unspecified WikiHop (Welbl et al., 2018) 12 1.1 CC BY SA 3.0 Table 4: Complete list of artifacts used in our experiments along with their versions and licenses.\nby scaling the prefix kernel's coefficients. Kern is the kernelized attention variant described in Tsai et al. (2019):\nKern(Q, K, V ) i = N j=1 k(Q i , K j ) N j ′ =1 k(Q i , K j ′ ) V j (8)\nwhere subscripts (e.g. i) index the rows of a matrix, N is the number of key and value vectors, and k is a kernel function that calculates the similarity score between two vectors. We do not experiment with altering the kernel type since the default exponential kernel inherent to softmax attention already implicitly maps the input vectors to an infinite feature space. Therefore, the kernel function in Equation 8takes the form:\nk(x q , x k ) = exp ⟨x q , x k ⟩ √ d k (9)\nwhere ⟨•, •⟩ signifies the dot product and d k is the dimension of key projections."
},
{
"figure_ref": [
"fig_1"
],
"heading": "B Experimental Details",
"publication_ref": [
"b24",
"b27",
"b1",
"b12"
],
"table_ref": [
"tab_6",
"tab_7",
"tab_8"
],
"text": "Artifact Notes evaluate and/or develop state-of-the-art algorithms.\nThe intended use of 20-newsgroups is not explicit, although it is commonly used for natural language processing in research. We therefore believe we have adhered to the intended usages of the datasets we included.\nWe do not anonymize the data for 20newsgroups as (a) the trained models is not being deployed (only used for evaluation purposes) and (b) the non-anonymized variant is already publicly available. We chose to use the datasets in the current form for fair comparison with other baselines and therefore did not do a detailed analysis for those artifacts. We refer readers to the cited original works in Table 4 for complete documentation.\nTraining For our experiments, we use and adapt the prefix-tuning implementation provided in Liu et al. (2022). Training was conducted on 12 NVIDIA GeForce 1080 Ti cards, for an estimated 2300 single GPU hours (including preliminary ex-periments). All models tested fit on a single card, so we did not use any model parallelism. Throughout experiments, we use gradient accumulation for an effective batch size of 32. We use early stopping for our hyperparameter search, and show results for the run with the best validation F1-score. For learning rate, we search between {1e-2, 5e-2, 1e-3, 5e-3, 5e-4} for prefix-based methods, and {3e-5, 5e-5} for fine-tuning. For kernelized prefix-propagation, we search for a scale factor (hyperparameter α) of {1e-2, 4e-2, 1e-3, 3e-3, 5e-3, 7e-3} (after choosing the best learning-rate). Other hyperparameters are listed in Table 5.\nDespite seeding random number generators for Hugging Face's transformer library through the set_seed method, slight deviations will propagate if using GPUs due to some non-deterministic CUDA methods that do not respect the seed setting mechanisms of Pytorch (Paszke et al., 2019). Upon further analysis, we found this issue in nondeterministic algorithms to be widely overlooked in the field, and believe that this area needs further discussion in the research community. However, we note that our results should be reproducible when running across multiple seeds.\nTask Details All datasets used have a considerable portion of documents greater than RoBERTa's max sequence limit of 512 tokens, as shown in Figure 3. Number of samples and number of classes for each dataset are in Table 6.\nFor all classification tasks, we prepend a globally-attended [CLS] token to the start of the sequence and pass the output into a learned classification head. We truncate document lengths to 4096 and 512 tokens for Longformer and RoBERTa, respectively. For Hyperpartisan, we use the same data pre-processing and training split as Beltagy et al. (2020). However, we noticed overlap between training and testing samples, so we instead show validation results. We use the ArXiv dataset from He et al. (2019) that is available on Huggingface datasets (which we reviewed for correctness). The original dataset has labels leaked in the source text, so we use the no_ref version that has those labels filtered. We use the 20-newsgroups and follow preprocessing as recommended by scikit-learn authors, removing headers, quotations, and signatures from each sample to prevent the model from learning spurious correlations.\nWikiHop instances include a question, candidate answers, and multiple context documents. For "
},
{
"figure_ref": [],
"heading": "D Runtime Performance",
"publication_ref": [],
"table_ref": [
"tab_9"
],
"text": "We test the inference time of the studied methods and show the results in Table 7. We use the same 8000 randomly generated sequences of length 4096 across methods and test on a NVIDIA GTX 1080 Ti. We notice that prefix-propagation is slightly more efficient than prefix-tuning. We theorize that this discrepancy is caused by prefix-propagation only needing to concatenate a matrix in the first layer (and sum on the rest), whereas prefix-tuning concatenates before every layers."
},
{
"figure_ref": [],
"heading": "Acknowledgements",
"publication_ref": [],
"table_ref": [],
"text": "This research is supported by NSERC Discovery Grants. The second author is also supported by the Vector Scholarship in Artificial Intelligence."
}
] | Parameter-efficient tuning aims to mitigate the large memory requirements of adapting pretrained language models for downstream tasks. For example, one popular method, prefix-tuning (Li and Liang, 2021;Liu et al., 2022), prepends trainable tokens to sequences while freezing the rest of the model's parameters. Although such models attain comparable performance with fine-tuning when applied to sequences with short to moderate lengths, we show their inferior performance when modelling long sequences. To bridge this gap, we propose prefix-propagation, a simple but effective approach that conditions prefixes on previous hidden states. We empirically demonstrate that prefix-propagation outperforms prefix-tuning across long-document tasks, while using ∼50% fewer parameters. To further investigate the proposed architecture, we also show its advantage in calibration, and perform additional study on its relationship with kernel attention. To the best of our knowledge, this work is the first to focus on parameter-efficient learning for long-sequence language tasks. | Prefix-Propagation: Parameter-Efficient Tuning for Long Sequences | [
{
"figure_caption": "in Appendix C) and we speculate that this may be due to catastrophic R Prefix Finetune Prefix Prop. KernelMethod",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: Violin plot showing distribution of sequence lengths for each dataset. \"HY.\" and \"NG.\" denote the Hyperpartisan task and the 20-newsgroups tasks, respectively.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Mean F1-Scores of prefix-tuning and finetuning Longformer for common long-document classification tasks.",
"figure_data": "com/MonliH/prefix-propagation",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "and 1b respectively. For the first layer of the transformer, we prepend j trainable prefixes (i.e.,",
"figure_data": "Method% Tuned WikiHopArXiv20-newsgroupsHyperpartisanAccPRF 1PRF 1PRF 1RoBERTa PT0.111.779.4 79.6 79.8 67.9 67.0 68.2 70.4 59.2 64.1Prefix-Tuning0.138.981.5 81.7 82.7 68.9 68.4 69.7 78.3 73.8 75.3Prefix-Propagation0.0542.283.1 83.1 83.3 70.1 69.7 71.0 86.4 77.7 81.8Fine-Tuning10074.083.1 82.9 83.3 71.8 71.2 72.3 87.8 76.2 81.5Table 2: Main results of prefix-propagation compared to prefix-tuning and traditional fine-tuning on the validationsets of each dataset. All approaches use Longformer-base except \"RoBERTa PT\", which is prefix-tuning onRoBERTa-base. Micro F1 and macro-average precision (\"P\") and recall (\"R\") is reported for ArXiv, Hyperparti-san (with mean across 5 runs), and 20-newsgroups. Accuracy is reported for WikiHop. Performance is reported ontest splits with the exception of Hyperpartisan, which is performance on the validation split (See Appendix B forreasoning). The best run is bold and second best is underlined.",
"figure_id": "tab_2",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "ECE scores of tested approaches. Lower is better. Bold is the best and underline is the second best. \"HY.\" is Hyperpartisan, and \"NG.\" is 20-newsgroups.",
"figure_data": "",
"figure_id": "tab_4",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": "summarizes the completelist of artifacts we used in our experiments alongwith their licenses and versions. All libraries wereused for their intended purpose of open-source de-velopment. The ArXiv, Hyperpartisan, and Wiki-Hop datasets were released in research contexts to",
"figure_id": "tab_6",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Hyperparameters used for our experiments. \"HY.\" and \"NG.\" denote the Hyperpartisan task and the 20-newsgroups tasks, respectively.",
"figure_data": "10 5RoBERTa Seq. Limit Longformer Seq. LimitSequence Length10 2 10 3 10 410 110 0ArXivNG.DatasetHY.WikiHop",
"figure_id": "tab_7",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "Dataset n sample n class n train/dev/test Datasets used and their total size (n sample ), number of classes (n class ), and relative sizes of train, validation, and test splits (n train/dev/test ).Figure4: Calibration (measured by ECE) of different tuning approaches using Longformer-base on ArXiv. Lower is better. a fair comparison, we follow the WikiHop setup inBeltagy et al. (2020) to the best of our ability. In summary, we pass the dataset fields into the model in the format:[q] <question> [/q] [ent] <candidate 1> [/ent] ... [ent]<candidate N> [/ent] [sep] <context 1> [sep] ... [sep] <context N>.Because the context documents are often longer than the maximum sequence length of Longformer, we split the context documents into chunks of 4096 (or 512 for RoBERTa) and pass them separately through the model while concatenated to the question and candidate pair. We then train a classifier to predict a single logit for each [ent] token, take the average over all chunks, apply softmax, and finally use cross-entropy loss. We also train the new special tokens [ent] and [q] in prefix-based methods to better learn an effective representation (as they did not appear in pre-training).",
"figure_data": "MethodAbsolute Runtime (s) Relative RuntimeHY.645280/10/10No PEFT21920%NG. ArXiv18,846 33,38820 1160/20/20 85/7.5/7.5Prefix-Tuning Prefix-Propagation2239 2196+2.1% +0.2%WikiHop 48,867-90/5/50.20Finetune Prefix Tuning Prefix PropagationECE0.10 0.150.050246 Epoch810",
"figure_id": "tab_8",
"figure_label": "6",
"figure_type": "table"
},
{
"figure_caption": "Runtime for inference using \"No PEFT\" (i.e., regular forward pass), prefix-tuning, and prefixpropagation. \"Relative Runtime\" is the runtime relative to \"No PEFT\". either start less calibrated or deviate from prefixpropagation as training progresses.",
"figure_data": "",
"figure_id": "tab_9",
"figure_label": "7",
"figure_type": "table"
}
] | Jonathan Li; Will Aitken; Rohan Bhambhoria; Xiaodan Zhu | [
{
"authors": "",
"journal": "Iz",
"ref_id": "b0",
"title": "",
"year": ""
},
{
"authors": "Matthew E Beltagy; Arman Peters; Cohan",
"journal": "",
"ref_id": "b1",
"title": "Longformer: The long-document transformer",
"year": "2020"
},
{
"authors": "Elad Ben Zaken; Yoav Goldberg; Shauli Ravfogel",
"journal": "Association for Computational Linguistics",
"ref_id": "b2",
"title": "BitFit: Simple parameter-efficient fine-tuning for transformer-based masked language-models",
"year": "2022"
},
{
"authors": "Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell; Sandhini Agarwal; Ariel Herbert-Voss; Gretchen Krueger; Tom Henighan; Rewon Child; Aditya Ramesh; Daniel Ziegler; Jeffrey Wu; Clemens Winter; Chris Hesse; Mark Chen; Eric Sigler; Mateusz Litwin; Scott Gray; Benjamin Chess; Jack Clark; Christopher Berner; Sam Mccandlish; Alec Radford; Ilya Sutskever; Dario Amodei",
"journal": "",
"ref_id": "b3",
"title": "Language models are few-shot learners",
"year": "2020"
},
{
"authors": "Curran Associates; Inc ",
"journal": "",
"ref_id": "b4",
"title": "",
"year": ""
},
{
"authors": "Yifan Chen; Devamanyu Hazarika; Mahdi Namazifar; Yang Liu; Di Jin; Dilek Hakkani-Tur",
"journal": "",
"ref_id": "b5",
"title": "Inducer-tuning: Connecting prefix-tuning and adapter-tuning",
"year": "2022"
},
{
"authors": "Rewon Child; Scott Gray; Alec Radford; Ilya Sutskever",
"journal": "",
"ref_id": "b6",
"title": "Generating long sequences with sparse transformers",
"year": "2019"
},
{
"authors": "Krzysztof Choromanski; Valerii Likhosherstov; David Dohan; Xingyou Song; Andreea Gane; Tamas Sarlos; Peter Hawkins; Jared Davis; Afroz Mohiuddin; Lukasz Kaiser; David Belanger; Lucy Colwell; Adrian Weller",
"journal": "",
"ref_id": "b7",
"title": "Rethinking attention with performers",
"year": "2020"
},
{
"authors": "Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova",
"journal": "Association for Computational Linguistics",
"ref_id": "b8",
"title": "BERT: Pre-training of deep bidirectional transformers for language understanding",
"year": "2019"
},
{
"authors": "Geoff Jacob R Gardner; David Pleiss; Kilian Q Bindel; Andrew Weinberger; Wilson Gordon",
"journal": "",
"ref_id": "b9",
"title": "Gpytorch: Blackbox matrix-matrix gaussian process inference with gpu acceleration",
"year": "2018"
},
{
"authors": "Yuxian Gu; Xu Han; Zhiyuan Liu; Minlie Huang",
"journal": "Association for Computational Linguistics",
"ref_id": "b10",
"title": "PPT: Pre-trained prompt tuning for few-shot learning",
"year": "2022"
},
{
"authors": "Chuan Guo; Geoff Pleiss; Yu Sun; Kilian Q Weinberger",
"journal": "JMLR",
"ref_id": "b11",
"title": "On calibration of modern neural networks",
"year": "2017"
},
{
"authors": "Jun He; Liqun Wang; Liu Liu; Jiao Feng; Hao Wu",
"journal": "IEEE Access",
"ref_id": "b12",
"title": "Long document classification from local word glimpses via recurrent attention learning",
"year": "2019"
},
{
"authors": "Junxian He; Chunting Zhou; Xuezhe Ma; Taylor Berg-Kirkpatrick; Graham Neubig",
"journal": "",
"ref_id": "b13",
"title": "Towards a unified view of parameter-efficient transfer learning",
"year": "2021"
},
{
"authors": "Pengcheng He; Xiaodong Liu; Jianfeng Gao; Weizhu Chen",
"journal": "",
"ref_id": "b14",
"title": "Deberta: Decoding-enhanced bert with disentangled attention",
"year": "2020"
},
{
"authors": "Xuehai He; Zhuo Cai; Wenlan Wei; Yichen Zhang; Luntian Mou; Eric Xing; Pengtao Xie",
"journal": "Association for Computational Linguistics",
"ref_id": "b15",
"title": "Towards visual question answering on pathology images",
"year": "2021"
},
{
"authors": "Edward Hu; Yelong Shen; Phil Wallis; Zeyuan Allen-Zhu; Yuanzhi Li; Lu Wang; Weizhu Chen",
"journal": "",
"ref_id": "b16",
"title": "Lora: Low-rank adaptation of large language models",
"year": "2021"
},
{
"authors": "Matthew Jagielski; Om Thakkar; Florian Tramèr; Daphne Ippolito; Katherine Lee; Nicholas Carlini; Eric Wallace; Shuang Song; Abhradeep Thakurta; Nicolas Papernot; Chiyuan Zhang",
"journal": "",
"ref_id": "b17",
"title": "Measuring forgetting of memorized training examples",
"year": "2022"
},
{
"authors": "A Katharopoulos; A Vyas; N Pappas; F Fleuret",
"journal": "",
"ref_id": "b18",
"title": "Transformers are rnns: Fast autoregressive transformers with linear attention",
"year": "2020"
},
{
"authors": "Johannes Kiesel; Maria Mestre; Rishabh Shukla; Emmanuel Vincent; Payam Adineh; David Corney; Benno Stein; Martin Potthast",
"journal": "Association for Computational Linguistics",
"ref_id": "b19",
"title": "SemEval-2019 task 4: Hyperpartisan news detection",
"year": "2019"
},
{
"authors": "Nikita Kitaev; Łukasz Kaiser; Anselm Levskaya",
"journal": "",
"ref_id": "b20",
"title": "Reformer: The efficient transformer",
"year": "2020"
},
{
"authors": "Ken Lang",
"journal": "",
"ref_id": "b21",
"title": "Newsweeder: Learning to filter netnews",
"year": "1995"
},
{
"authors": "Quentin Lhoest; Albert Villanova Del Moral; Thomas Patrick Von Platen; Mario Wolf; Yacine Šaško; Abhishek Jernite; Lewis Thakur; Suraj Tunstall; Mariama Patil; Julien Drame; Julien Chaumond; Joe Plu; Simon Davison; Victor Brandeis; Teven Sanh; Kevin Canwen Le Scao; Nicolas Xu; Steven Patry; Angelina Liu; Philipp Mcmillan-Major; Sylvain Schmid; Nathan Gugger; Raw; Anton Sylvain Lesage; Matthew Lozhkov; Théo Carrigan; Matussière; Lysandre Leandro Von Werra; Stas Debut; Clément Bekman; Delangue",
"journal": "Association for Computational Linguistics",
"ref_id": "b22",
"title": "Datasets: A Community Library for Natural Language Processing",
"year": "2021"
},
{
"authors": "Lisa Xiang; Percy Li; Liang",
"journal": "Association for Computational Linguistics",
"ref_id": "b23",
"title": "Prefix-tuning: Optimizing continuous prompts for generation",
"year": "2021"
},
{
"authors": "Xiao Liu; Kaixuan Ji; Yicheng Fu; Weng Tam; Zhengxiao Du; Zhilin Yang; Jie Tang",
"journal": "Association for Computational Linguistics",
"ref_id": "b24",
"title": "P-tuning: Prompt tuning can be comparable to fine-tuning across scales and tasks",
"year": "2022"
},
{
"authors": "Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov",
"journal": "",
"ref_id": "b25",
"title": "Roberta: A robustly optimized bert pretraining approach",
"year": "2019"
},
{
"authors": "Gregory Mahdi Pakdaman Naeini; Milos Cooper; Hauskrecht",
"journal": "",
"ref_id": "b26",
"title": "Obtaining well calibrated probabilities using bayesian binning",
"year": "2015"
},
{
"authors": "Adam Paszke; Sam Gross; Francisco Massa; Adam Lerer; James Bradbury; Gregory Chanan; Trevor Killeen; Zeming Lin; Natalia Gimelshein; Luca Antiga; Alban Desmaison; Andreas Kopf; Edward Yang; Zachary Devito; Martin Raison; Alykhan Tejani; Sasank Chilamkurthy; Benoit Steiner; Lu Fang; Junjie Bai; Soumith Chintala",
"journal": "",
"ref_id": "b27",
"title": "PyTorch: An Imperative Style, High-Performance Deep Learning Library",
"year": "2019"
},
{
"authors": "Hao Peng; Nikolaos Pappas; Dani Yogatama; Roy Schwartz; Noah A Smith; Lingpeng Kong",
"journal": "",
"ref_id": "b28",
"title": "Random feature attention",
"year": "2021"
},
{
"authors": "Uri Shaham; Elad Segal; Maor Ivgi; Avia Efrat; Ori Yoran; Adi Haviv; Ankit Gupta; Wenhan Xiong; Mor Geva; Jonathan Berant; Omer Levy",
"journal": "",
"ref_id": "b29",
"title": "Scrolls: Standardized comparison over long language sequences",
"year": "2022"
},
{
"authors": "Emma Strubell; Ananya Ganesh; Andrew Mccallum",
"journal": "Association for Computational Linguistics",
"ref_id": "b30",
"title": "Energy and policy considerations for deep learning in NLP",
"year": "2019"
},
{
"authors": "Yi Tay; Mostafa Dehghani; Dara Bahri; Donald Metzler",
"journal": "ACM Comput. Surv",
"ref_id": "b31",
"title": "Efficient transformers: A survey",
"year": "2022"
},
{
"authors": "Yao-Hung Hubert Tsai; Shaojie Bai; Makoto Yamada; Louis-Philippe Morency; Ruslan Salakhutdinov",
"journal": "",
"ref_id": "b32",
"title": "Transformer dissection: An unified understanding for transformer's attention via the lens of kernel",
"year": "2019"
},
{
"authors": "Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Ł Ukasz Kaiser; Illia Polosukhin",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b33",
"title": "Attention is all you need",
"year": "2017"
},
{
"authors": "Curran Associates; Inc ",
"journal": "",
"ref_id": "b34",
"title": "",
"year": ""
},
{
"authors": "Johannes Welbl; Pontus Stenetorp; Sebastian Riedel",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b35",
"title": "Constructing datasets for multi-hop reading comprehension across documents",
"year": "2018"
},
{
"authors": "Thomas Wolf; Lysandre Debut; Victor Sanh; Julien Chaumond; Clement Delangue; Anthony Moi; Pierric Cistac; Tim Rault; Rémi Louf; Morgan Funtowicz; Joe Davison; Sam Shleifer; Clara Patrick Von Platen; Yacine Ma; Julien Jernite; Canwen Plu; Teven Xu; Sylvain Le Scao; Mariama Gugger; Quentin Drame; Alexander M Lhoest; Rush",
"journal": "Association for Computational Linguistics",
"ref_id": "b36",
"title": "Transformers: State-of-the-art natural language processing",
"year": "2020"
},
{
"authors": "Manzil Zaheer; Guru Guruganesh; Avinava Kumar; Joshua Dubey; Chris Ainslie; Santiago Alberti; Philip Ontanon; Anirudh Pham; Qifan Ravula; Li Wang; Yang",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b37",
"title": "Big bird: Transformers for longer sequences",
"year": "2020"
}
] | [
{
"formula_coordinates": [
3,
70.47,
476.54,
219.39,
107.17
],
"formula_id": "formula_0",
"formula_text": "H l,i = Attn(D (l) W (l,i) q , D (l) W (l,i) k , D (l) W (l,i) v ) (1) D (l) = cat(P (l) , C) if l = 1 cat(P (l) + C[:j, :], C[j:, :]) if l > 1 where inputs C are projected through pre-trained weight matrices W (l,i) q , W (l,i) k , W (l,i) v"
},
{
"formula_coordinates": [
3,
70.87,
730.03,
217.77,
29.8
],
"formula_id": "formula_1",
"formula_text": "W (i) k and W (i)"
},
{
"formula_coordinates": [
3,
323.44,
439.26,
183.68,
34.15
],
"formula_id": "formula_2",
"formula_text": "H l,i = Attn(CW (l,i) q , cat(P (l,i) k , CW (l,i) k ), cat(P (l,i) v , CW (l,i) v ))"
},
{
"formula_coordinates": [
4,
94.34,
206.18,
195.53,
27.31
],
"formula_id": "formula_3",
"formula_text": "H = Kern(cat(P, C)W q , CW k , CW v ) + (α)Kern(cat(P, C)W q , P W k , P W v ) (3)"
},
{
"formula_coordinates": [
8,
108.15,
236.65,
181.72,
32.26
],
"formula_id": "formula_4",
"formula_text": "H l,i = Attn(cat(P (l) , C)W (l)(i) q ,(4) cat"
},
{
"formula_coordinates": [
8,
163.76,
255.36,
88.09,
33.97
],
"formula_id": "formula_5",
"formula_text": "(P (l) , C)W (l)(i) k , cat(P (l) , C)W (l)(i) v )"
},
{
"formula_coordinates": [
8,
73.74,
364.34,
216.12,
159.76
],
"formula_id": "formula_6",
"formula_text": "H = Attn(DW q , cat(P, C)W k , cat(P, C)W v ) = softmax(DW q cat(P W k , CW k )) P W v CW v = (1-λ(C))softmax(DW q W ⊤ k C ⊤ )CW v +λ(C)softmax(DW q W ⊤ k P ⊤ )P W v = (1-λ(C))Attn(DW q , CW k , CW v ) +λ(C)Attn(DW q , P W k , P W v ) = (1-λ(C))Attn(cat(P, C)W q , CW k , CW v ) +λ(C)Attn(cat(P, C)W q , P W k , P W v )(5)"
},
{
"formula_coordinates": [
8,
77.2,
587.45,
212.67,
42.59
],
"formula_id": "formula_7",
"formula_text": "λ(C) = i DW q W k ⊤ P ⊤ i DW q W ⊤ k P ⊤ + j DW q W ⊤ k C ⊤(6)"
},
{
"formula_coordinates": [
8,
94.34,
695.26,
195.53,
27.31
],
"formula_id": "formula_8",
"formula_text": "H = Kern(cat(P, C)W q , CW k , CW v ) + (α)Kern(cat(P, C)W q , P W k , P W v ) (7)"
},
{
"formula_coordinates": [
8,
312.4,
293.98,
212.74,
33.71
],
"formula_id": "formula_9",
"formula_text": "Kern(Q, K, V ) i = N j=1 k(Q i , K j ) N j ′ =1 k(Q i , K j ′ ) V j (8)"
},
{
"formula_coordinates": [
8,
350.97,
469.52,
174.17,
26.3
],
"formula_id": "formula_10",
"formula_text": "k(x q , x k ) = exp ⟨x q , x k ⟩ √ d k (9)"
}
] |
10.1145/nnnnnnn.nnnnnnn | 2023-05-20 | [
{
"figure_ref": [],
"heading": "INTRODUCTION",
"publication_ref": [
"b33",
"b48",
"b26",
"b53",
"b33",
"b48",
"b53",
"b41",
"b5"
],
"table_ref": [],
"text": "Predicting the properties of graphs has attracted great attention from drug discovery [34,49] and material design [27,54], because molecules and polymers are naturally graphs. Properties such as density, melting temperature, and oxygen permeability are often in continuous value spaces [34,49,54]. Graph regression tasks are important and challenging. It is hard to observe label values in certain rare areas since the annotated data usually concentrate on small yet popular areas in the property spaces. Graph regression datasets are ubiquitously imbalanced. Previous attempts that address data imbalance mostly focused on categorical properties and classification tasks, however, imbalanced regression tasks on graphs are under-explored.\nBesides data imbalance, the annotated graph regression data are often small in real world. For example, measuring the property of a molecule or polymer often needs expensive experiments or simulations. It has taken nearly 70 years to collect only around 600 polymers with experimentally measured oxygen permeability in the Polymer Gas Separation Membrane Database [42]. On the other side, we have hundreds of thousands of unlabeled graphs.\nPseudo-labeling unlabeled graphs may enrich and balance training data, however, there are two challenges. First, if one directly trained a model on the imbalanced labeled data and used it to do pseudo-labeling, it would not be reliable to generate accurate and balanced labels. Second, because quite a number of unlabeled graphs might not follow the distribution of labeled data, massive label noise is inevitable in pseudo-labeling and thus selection is necessary to expand the set of data examples for training. Moreover, the selected pseudo-labels without noise cannot alleviate the label imbalance problem. Because the biased model tends to generate more pseudo-labels in the label ranges where most data concentrate. In this situation, the selected pseudo-labels may aggravate the model bias and lead the model to have even worse performance on the label ranges where we lack enough data. Even though the pseudo-labeling had involved quality selection and the unlabeled set had been fully used to address label imbalance, the label distribution of annotated and pseudo-labeled examples might still be far from a perfect balance. This is because there might not be a sufficient number of pseudo-labeled examples to fill the gap in the under-represented label ranges.\nFigure 1 illustrates our ideas to overcome the above challenges. First, we want to progressively reduce the model bias by gradually improving training data from the labeled and unlabeled sets. The performance of pseudo-labeling models and the quality of the expanded training data can mutually enhance each other through iterations. Second, we relate the regression confidence to the prediction variance under perturbations. Higher confidence indicates a lower prediction variance under different perturbation environments. Therefore, we define and use regression confidence score to avoid pseudo-label noise and select quality examples in regression tasks. To fully exploit the quality pseudo-labels to compensate for the data imbalance in different label ranges, we use a reversed distribution of the imbalanced annotated data to reveal label ranges that need to be more or less selected for label balancing. Third, we attempt to achieve the perfect balance of training data by creating graph examples of any given label value in the remaining under-represented ranges.\nIn this paper, we propose SGIR, a novel Semi-supervised framework for Graph Imbalanced Regression. This framework has three novel designs to implement our ideas. First, SGIR is a self-training framework with multiple iterations for model learning and balanced\n/(1)\nRepresentations of all real graphs ℋ Train / 1 : ! → ℋ with ! \"#$ ∪ ! %&'( for ℋ \"#$ and ℋ %&'( Train 5 6 : ℋ → ℝ with ℋ \"#$ ∪ ℋ %&'( ∪ ℋ *+, 6 = / 1 9; ; = 5 (6) "
},
{
"figure_ref": [],
"heading": "SGIR models",
"publication_ref": [],
"table_ref": [],
"text": "Not balanced yet!"
},
{
"figure_ref": [],
"heading": "self-training",
"publication_ref": [],
"table_ref": [],
"text": "Figure 1: An overview of our SGIR framework to train effective graph regression models with imbalanced labeled data. To balance the data properly, SGIR selects highly confident examples from predicted labels of unlabeled data and augments label areas that seriously lack data (even after added the confidently predicted data) by a novel label-anchored mixup algorithm.\ntraining data generation. Our second design is to sample more quality pseudo-labels for the less represented label ranges. We define a new measurement of regression confidence from recent studies on graph rationalization methods which provide perturbations for predictions at training and inference. After applying the confidence to filter out pseudo-label noise, we adopt reverse sampling to find optimal sampling rates at each label value that maximize the possibility of data balance. Intuitively, if a label value is less frequent in the annotated data, the sampling rate at this value is bigger and more pseudo-labeled examples are selected for model training. Third, we design a novel label-anchored mixup algorithm to augment graph examples by mixing up a virtual data point and a real graph example in latent space. Each virtual point is anchored at a certain label value that is still rare in the expanded labeled data. The mixed-up graph representations continue complementing the label ranges where we seriously lack data examples.\nTo empirically demonstrate the advantage of SGIR, we conduct experiments on seven graph property regression tasks from three different domains. Results show that SGIR significantly reduces the prediction error on all the tasks and in both under-/well-represented label ranges. For example, on the smallest dataset Mol-FreeSolv that has only 276 annotated graphs, SGIR reduces the mean absolute error from 1.114 to 0.777 (relatively 30% improvement) in the most under-represented label range and reduces the error from 0.642 to 0.563 (12% improvement) in the entire label space compared to state-of-the-art graph regression methods. To summarize:\n• We address a new problem of graph imbalance regression with a novel semi-supervised framework SGIR. • SGIR is a novel self-training framework creating balanced and enriched training data from pseudo-labels and augmented examples with three collaborated components: regression confidence, reverse sampling, and label-anchored mixup.\n• SGIR is theoretically motivated and empirically validated on seven graph regression tasks. It outperforms other semisupervised learning and imbalanced regression methods in both well-represented and under-represented label ranges."
},
{
"figure_ref": [],
"heading": "RELATED WORK 2.1 Imbalanced Learning",
"publication_ref": [
"b8",
"b10",
"b22",
"b6",
"b29",
"b42",
"b51",
"b5",
"b51",
"b34",
"b51",
"b12"
],
"table_ref": [],
"text": "Data resampling is known as under-sampling majority classes or over-sampling minority classes. SMOTE [9] created synthetic data for minority classes using linear interpolations on labeled data. Cost-sensitive techniques [11,23] assigned higher weights to the loss of minority classes. And posterior re-calibration [7,30,43] encouraged larger margins for the prediction logits of minority classes. Imbalanced regression tasks have unique challenges due to continuous label values [52]. Some of the methods from imbalanced classifications were extended to imbalanced regression tasks. For example, SMOGN [6] adopted the idea and method of SMOTE for regression; Recently, Yang et al. [52] used regression focal loss and cost-sensitive reweighting; and BMSE [35] used logit re-calibration to predict numerical labels. LDS [52] smoothed label distribution using kernel density estimation. RankSim [13] regularized the latent space by approximating the distance of data points in the label space. Although these methods would improve performance on under-represented labels, they come at the expense of decreased performance on well-represented labels, particularly when annotated data is limited. SGIR avoids this by using unlabeled graphs to create more labels in the under-represented label ranges."
},
{
"figure_ref": [],
"heading": "Semi-supervised Learning",
"publication_ref": [
"b36",
"b3",
"b19",
"b31",
"b16",
"b46",
"b17"
],
"table_ref": [],
"text": "To exploit unlabeled data, semi-supervised image classifiers such as FixMatch [37] and MixMatch [4] used pseudo-labeling and consistency regularization. Their performance relies on weak and strong data augmentation techniques, which are under-explored for regression tasks and graph property prediction tasks. At the same time, semi-supervised learners suffer from the model bias caused by the unlabeled imbalance. Therefore, after pseudo-labeling unlabeled data, DARP [20] and DASO [32] refined the biased pseudolabels by aligning their distribution with an approximated true class distribution of unlabeled data. CADR [17] adjusted the threshold for pseudo-label assignments. CReST [47] selected more pseudo-labels for minority classes in self-training. To the best of our knowledge, there was no work that leveraged unlabeled data for regression tasks on imbalanced graph data, although SSDKL [18] performed semisupervised regression for non-graph data without considering label imbalance. SGIR makes the first attempt to solve the imbalanced regression problem using semi-supervised learning."
},
{
"figure_ref": [],
"heading": "Graph Property Prediction",
"publication_ref": [
"b13",
"b20",
"b43",
"b50",
"b16",
"b24",
"b23",
"b57",
"b35",
"b58",
"b59",
"b14",
"b45",
"b24",
"b39"
],
"table_ref": [],
"text": "Graph neural network models (GNN) [14,21,44,51] have demonstrated their power for regression tasks in the fields of biology, chemistry, and material science [17,25]. Data augmentation [24,58] is an effective way to exploit limited labeled data. The node-and link-level augmentations [36,59,60] modified graph structure to improve the accuracy of node classification and link prediction. On the graph level, augmentation methods were mostly designed for classification [15,46]. Recently, GREA [25] delivered promising results for predicting polymer properties. But the model bias caused by imbalanced continuous labels was not addressed. InfoGraph [40] exploited unlabeled graphs, however, the data imbalance issue was not addressed either. Our work aims to achieve balanced training data for graph regression in real practice where we have a small set of imbalanced labeled graphs and a large set of unlabeled data."
},
{
"figure_ref": [],
"heading": "PROBLEM DEFINITION",
"publication_ref": [
"b53"
],
"table_ref": [],
"text": "To predict the property 𝑦 ∈ R of a graph 𝐺 ∈ G, a graph regression model usually consists of an encoder 𝑔 : 𝐺 → h ∈ R 𝑑 and a decoder 𝑓 : h → ŷ ∈ R. The encoder 𝑔(•) is often a graph neural network (GNN) that outputs the 𝑑-dimensional representation vector h of graph 𝐺, and the decoder 𝑓 (•) is often a multi-layer perceptron (MLP) that makes the label prediction ŷ given h. Let G imb = {(𝐺 𝑖 , 𝑦 𝑖 )} 𝑛 imb 𝑖=1 denote the labeled training data for graph regression models, where 𝑛 imb is the number of training graphs in the imbalanced labeled dataset. It often concentrates on certain areas in the continuous label space. To reveal it, we first divide the label space into 𝐶 intervals and use them to fully cover the range of continuous label values. These intervals are\n[𝑏 0 , 𝑏 1 ), [𝑏 1 , 𝑏 2 ), . . . , [𝑏 𝐶-1 , 𝑏 𝐶 ).\nThen, we assign the labeled examples into 𝐶 intervals and count them in each interval to construct the frequency set {𝜇 𝑖 } 𝐶 𝑖=1 . We could find that max{𝜇 𝑖 } min{𝜇 𝑖 } ≫ 1 (i.e., label imbalance) often exists, instead of 𝜇 1 = 𝜇 2 = • • • = 𝜇 𝐶 (i.e., label balance) that is assumed by most existing models. The existing models may be biased to small areas in the label space that are dominated by the majority of labeled data and lack a good generalization to areas that are equally important but have much fewer examples.\nLabeling continuous graph properties is difficult [54], limiting the size of labeled data. Fortunately, a large number of unlabeled graphs are often available though ignored in most existing studies. In this work, we aim to use the unlabeled examples to alleviate the label imbalance issue in graph regression tasks. That is, let G unlbl = {𝐺 𝑗 } 𝑛 imb +𝑛 unlbl 𝑗=𝑛 imb +1 denote the 𝑛 unlbl available unlabeled graphs. We want to train 𝑔(•) and 𝑓 (•) to deliver good performance through the whole continuous label space by utilizing both G imb and G unlbl ."
},
{
"figure_ref": [],
"heading": "PROPOSED FRAMEWORK",
"publication_ref": [],
"table_ref": [],
"text": "To progressively reduce label imbalance bias, we propose a novel framework named SGIR that iteratively creates reliable labeled examples in the areas of label space where annotations were not frequent. As presented in Figure 1, SGIR uses a graph regression model to create the labels and uses the gradually balanced data to train the regression model. "
},
{
"figure_ref": [],
"heading": "Balancing with Confidently Predicted Labels",
"publication_ref": [],
"table_ref": [],
"text": "At each iteration, SGIR enriches and balances training data with pseudo-labels of good quality. The unlabeled data examples in G unlbl are firstly exploited by reliable and confident predictions.\nThen the reverse sampling from the imbalanced label distribution of original training data G imb is used to select more pseudo-labels for under-represented label ranges."
},
{
"figure_ref": [],
"heading": "Graph regression with confidence.",
"publication_ref": [
"b4",
"b24",
"b47",
"b24",
"b11",
"b40",
"b1",
"b46"
],
"table_ref": [],
"text": "A standard regression model outputs a scalar without a certain definition of confidence of its prediction. The confidence is often measured by how much the predicted probability is close to 1 in classifications. The lack of confidence measurements in graph regression tasks may introduce noise to the self-training framework that aims at label balancing. It would be more severe when the domain gap exists between labeled and unlabeled data [5]. Recent studies [25,48] have proposed two concepts that help us define a good measurement: rationale subgraph and environment subgraph. A rationale subgraph is supposed to best support and explain the prediction at property inference. Its counterpart environment subgraph is the complementary subgraph in the example, which perturbs the prediction from the rationale subgraph if used. Our idea is to measure the confidence of graph property prediction based on the reliability of the identified rationale subgraphs. Specifically, we use the variance of predicted label values from graphs that consist of a specific rationale subgraph and one of many possible environment subgraphs.\nWe use an existing supervised graph regression model that can identify rationale and environment subgraphs in any graph example to predict its property. We denote 𝐺 𝑖 as the 𝑖-th graph in a batch of size 𝐵. The model separates 𝐺 𝑖 into 𝐺 𝑗 that has the rationale of 𝐺 𝑖 and environment subgraph of 𝐺 𝑗 . So it is expected to have the same label of 𝐺 𝑖 . By enumerating 𝑗 ∈ {1, 2, . . . , 𝐵}, the encoder 𝑔(•) and decoder 𝑓 (•) are trained to predict the label value of any 𝐺 (𝑖,𝑗) . We define the confidence of predicting the label of 𝐺 𝑖 as:\n𝜎 𝑖 = 1 Var {𝑓 (𝑔(𝐺 (𝑖,𝑗) ))} 𝑗=1,2,...,𝐵 .(1)\nIt is the reciprocal of prediction variance. In implementation, we choose GREA [25] as the model. Considering efficiency, GREA creates 𝐺 (𝑖,𝑗) in the latent space without decoding its graph structure. That is, it directly gets the representation of 𝐺 (𝑖,𝑗) as the sum of the representation vectors h\n(𝑟 ) 𝑖 of 𝐺 (𝑟 )\n𝑖 and h\n(𝑒) 𝑗 of 𝐺 (𝑒)\n𝑗 . So we have\n𝜎 𝑖 = 1 Var {𝑓 (h (𝑟 ) 𝑖 + h (𝑒) 𝑗 )} 𝑗=1,2,...,𝐵 .(2)\nNow we have predicted labels and confidence values for graph examples in the large unlabeled dataset G unlbl . Examples with low confidences will bring noise to the training data if we use them all. So we only consider a data example 𝐺 𝑖 to be of good quality if its confidence 𝜎 𝑖 is not smaller than a threshold 𝜏. We name this confidence measurement based on graph rationalization as GRation. GRation is tailored for graph regression tasks by considering the environment subgraphs as perturbations. We will compare its effect on quality graph selection against other graph-irrelevant methods such as Dropout [12], Certi [41], DER (Deep Evidential Regression) [2], and Simple (no confidence) in experiments.\nAfter leveraging the unlabeled data, the label distribution of quality examples may still be biased to the majority of labels. So we further apply reverse sampling on these examples from G unlbl to balance the distribution of training data. 4.2.2 Reverse sampling. The reverse sampling in SGIR helps reduce the model bias to label imbalance. Specifically, we want to selectively add unlabeled examples predicted in the under-represented label ranges. Suppose we have the frequency set {𝜇 𝑖 } 𝐶 𝑖=1 of 𝐶 intervals. We denote 𝑝 𝑖 as the sampling rate at the 𝑖-th interval and follow Wei et al. [47] \n𝑝 𝑖 = 𝜇 ′ 𝑖 max{𝜇 1 , 𝜇 2 , . . . , 𝜇 𝐶 } . (3\n)\nTo this end, we have the confidently labeled and reversed sampled data G conf . In each self-training iteration, we combine it with the original training set G imb ."
},
{
"figure_ref": [],
"heading": "Balancing with Augmentation via Label-Anchored Mixup",
"publication_ref": [
"b44",
"b45"
],
"table_ref": [],
"text": "Although G imb ∪ G conf is more balanced than G imb , we observe that G imb ∪ G conf is usually far from a perfect balance, even if G unlbl could be hundreds of times bigger than G imb . To create graph examples targeting the remaining under-represented label ranges, we design a novel label-anchored mixup algorithm for graph imbalanced regression. Compared to existing mixup algorithms [45,46] for classifications without awareness of imbalance, our new algorithm can augment training data with additional examples for target ranges of continuous label value.\nA mixup operation in the label-anchored mixup is to mix up two things in a latent space: (1) a virtual data point representing an interval of targeted label and (2) a real graph example. Specifically, we first calculate the representation of a target label interval by averaging the representation vectors of graphs in the interval from the labeled dataset G imb . Let M ∈ {0, 1} 𝐶×𝑛 imb be an indicator matrix, where 𝑀 𝑖,𝑗 = 1 means that the label of 𝐺 𝑗 ∈ G imb belongs to the 𝑖-th interval. We denote H ∈ R 𝑛 imb ×𝑑 as the matrix of graph representations from the GNN encoder 𝑔(•) for G imb . The representation matrix Z ∈ R 𝐶×𝑑 of all intervals is calculated\nZ = norm(M) • H,(4)\nwhere norm(•) is the row-wise normalization. Let 𝑎 𝑖 denote the center label value of the 𝑖-th interval. Then we have the representationlabel pairs of all the label intervals {(z 𝑖 , 𝑎 𝑖 )} 𝐶 𝑖=1 , where z 𝑖 is the 𝑖-th row of Z.\nNow we can use each interval center 𝑎 𝑖 as a label anchor to augment graph data examples in a latent space. We select 𝑛 𝑖 ∝ 𝑝 𝑖 real graphs from G imb ∪ G conf whose labels are closest to 𝑎 𝑖 , where 𝑝 𝑖 is calculated by Eq. ( 3). The more real graphs are selected, the more graph representations are augmented. 𝑛 𝑖 is likely to be big when the label anchor 𝑎 𝑖 remains under-represented after G conf is added to training set. Note that the labels were annotated if the graphs were in G imb and predicted if they were in G unlbl . For 𝑗 ∈ {1, 2, . . . , 𝑛 𝑖 }, we mix up the interval (z 𝑖 , 𝑎 𝑖 ) and a real graph (h 𝑗 , 𝑦 𝑗 ), where h 𝑗 and 𝑦 𝑖 are the representation vector and the annotated or predicted label of the 𝑗-th graph, respectively. Then the mixup operation is defined as\nh(𝑖,𝑗) = 𝜆 • z 𝑖 + 1 -𝜆 • h 𝑗 , ỹ(𝑖,𝑗) = 𝜆 • 𝑎 𝑖 + 1 -𝜆 • 𝑦 𝑗 ,(5)\nwhere h(𝑖,𝑗) and ỹ(𝑖,𝑗) are the representation vector and label of the augmented graph, respectively. 𝜆 = max(𝜆 ′ , 1 -𝜆 ′ ), 𝜆 ′ ∼ Beta(1, 𝛽), and 𝛽 is a hyperparameter. 𝜆 is often closer to 1 because we want ỹ(𝑖,𝑗) to be closer to the label anchor 𝑎 𝑖 . Let H aug denote the set of representation vectors of all the augmented graphs. Combined with G imb and G conf , we end up with a label-balanced training set for the next round of self-training."
},
{
"figure_ref": [],
"heading": "Optimization",
"publication_ref": [
"b24"
],
"table_ref": [],
"text": "In each iteration of self-training, we jointly optimize the parameters of graph encoder 𝑔(•) and label predictor 𝑓 (•) with a gradually balanced training set\nG imb ∪ G conf ∪ H aug .\nWe use the mean absolute error (MAE) as the regression loss. Specifically, for each\n(𝐺, 𝑦) ∈ G imb ∪ G conf , the loss is ℓ imb+conf = MAE(𝑓 (𝑔(𝐺)), 𝑦). Given (h, 𝑦) ∈ H aug , the loss is ℓ aug = MAE(𝑓 (h), 𝑦). So the total loss for SGIR is L = ∑︁ (𝐺,𝑦) ∈ G imb ∪G conf ℓ imb+conf (𝐺, 𝑦) + ∑︁ (h,𝑦) ∈H aug ℓ aug (h, 𝑦).\nOur framework is flexible with any graph encoder-decoder models.\nTo be consistent and given the promising results in graph regression tasks, we use the design of graph encoder and decoder in GREA [25] which is also used for measuring prediction confidence in Eq. (2)."
},
{
"figure_ref": [],
"heading": "Theoretical Motivations",
"publication_ref": [
"b6",
"b42",
"b55",
"b6",
"b18",
"b56"
],
"table_ref": [],
"text": "There is a lack of theoretical principle for imbalanced regression.\nOur theoretical motivation extends the generalization error bound from classification [7] to regression. The original bound enforces bigger margins for minority classes, which potentially hurt the model performance for well-represented classes [43,56]. Our result provides a more safe way to reduce the error bound by utilizing unlabeled graphs with self-training in graph regression tasks.\nAs we divide the label distribution into 𝐶 intervals, every graph example can be assigned into an interval (as the ground-truth interval) according to the distance between the interval center and the ground-truth label value. Besides, we use 𝑆 [𝑏 𝑖 ,𝑏 𝑖+1 ) (𝐺) to denote the reciprocal of the distance between the predicted label of the graph 𝐺 and the 𝑖-th interval [𝑏 𝑖 , 𝑏 𝑖+1 ), where 𝑖 ∈ {1, 2, . . . , 𝐶}. In this way, we could define 𝑓 (•) as a regression function that outputs a continuous predicted label. Then 𝑆 [𝑏 𝑖 ,𝑏 𝑖+1 ) (𝐺) consists of 𝑓 (•) and outputs the logits to classify the graph to the 𝑖-the interval.\nWe consider all training examples to follow the same distribution. We assume that conditional on label intervals, the distributions of graph sampling are the same at training and testing stages. So, the standard 0-1 test error on the balanced test distribution is\nE bal [𝑓 ] = Pr (𝐺, [𝑏 𝑖 ,𝑏 𝑖+1 ))∼P bal 𝑆 [𝑏 𝑖 ,𝑏 𝑖+1 ) (𝐺) < max 𝑗≠𝑖 𝑆 [𝑏 𝑗 ,𝑏 𝑗 +1 ) (𝐺) ,(6)\nwhere P bal denotes the balanced test distribution. It first samples a label interval uniformly and then samples graphs conditionally on the interval. The error for the 𝑖-th interval [𝑏 𝑖 , 𝑏 𝑖+1 ) is defined as\nE [𝑏 𝑖 ,𝑏 𝑖+1 ) [𝑓 ] = Pr 𝐺∼P [𝑏 𝑖 ,𝑏 𝑖+1 ) 𝑆 [𝑏 𝑖 ,𝑏 𝑖+1 ) (𝐺) < max 𝑗≠𝑖 𝑆 [𝑏 𝑗 ,𝑏 𝑗 +1 ) (𝐺) ,(7) where P\n[𝑏 𝑖 ,𝑏 𝑖+1 ) denotes the distribution for the interval [𝑏 𝑖 , 𝑏 𝑖+1 ). We define 𝛾 (𝐺, [𝑏 𝑖 , 𝑏 𝑖+1 )) = 𝑆 [𝑏 𝑖 ,𝑏 𝑖+1 ) (𝐺) -max 𝑗≠𝑖 𝑆 [𝑏 𝑗 ,𝑏 𝑗 +1 ) (𝐺)\nas the margin of an example 𝐺 assigned to the interval [𝑏 𝑖 , 𝑏 𝑖+1 ). To define the training margin 𝛾 [𝑏 𝑖 ,𝑏 𝑖+1 ) for the interval [𝑏 𝑖 , 𝑏 𝑖+1 ), we calculate the minimal margin across all examples assigned to that interval:\n𝛾 [𝑏 𝑖 ,𝑏 𝑖+1 ) = min 𝐺 𝑗 ∈ [𝑏 𝑖 ,𝑏 𝑖+1 ) 𝛾 𝐺 𝑗 , [𝑏 𝑖 , 𝑏 𝑖+1 ) .(8)\nWe assume that the MAE regression loss is small enough to correctly assign all training examples to the corresponding intervals. Given the hypothesis class F , C(F ) is assumed to be a proper complexity measure of F . We assume there are 𝑛 [𝑏 𝑖 ,𝑏 𝑖+1 ) examples i.i.d sampled from the conditional distribution P [𝑏 𝑖 ,𝑏 𝑖+1 ) for the interval [𝑏 𝑖 , 𝑏 𝑖+1 ). So, we apply the standard margin-based generalization bound to obtain the following theorem [7,19,57]: \nE [𝑏 𝑖 ,𝑏 𝑖+1 ) [𝑓 ] ⪅1\n𝛾 [𝑏 𝑖 ,𝑏 𝑖+1 ) √︄ C(F ) 𝑛 [𝑏 𝑖 ,𝑏 𝑖+1 ) + √︄ log log 2 (1/𝛾 [𝑏 𝑖 ,𝑏 𝑖+1 ) ) + log(1/𝛿) 𝑛 [𝑏 𝑖 ,𝑏 𝑖+1 ) ,(9)\nwhere ⪅ hides constant terms. examples do not break our assumption for the theorem and future directions of imbalanced regression theories without intervals."
},
{
"figure_ref": [],
"heading": "EXPERIMENTS",
"publication_ref": [],
"table_ref": [],
"text": "We conduct experiments to demonstrate the effectiveness of SGIR and answer the research question: how it performs on graph regression tasks and at different label ranges (RQ1). We also make a few ablation studies to investigate the effect of model design: where the effectiveness comes from (RQ2)."
},
{
"figure_ref": [
"fig_2",
"fig_5"
],
"heading": "Experimental Settings",
"publication_ref": [
"b48",
"b24",
"b33",
"b24",
"b51",
"b21",
"b30",
"b51",
"b0",
"b21",
"b21"
],
"table_ref": [
"tab_2",
"tab_2"
],
"text": "5.1.1 Datasets. Molecule and polymer datasets in Table 1 (the datasets with a prefix Mol-or Plym-) and Figure 2 present detailed data statistics for six graph regression tasks from chemistry and materials science. Three molecule datasets are from [49] and three polymer datasets are from [25]. Besides labeled graphs, we combine a database of 133,015 molecules in QM9 [34] and an integration of four sets of 13,114 polymers in total [25] to create a set of 146,129 unlabeled graphs to set up semi-supervised graph regression. We note that the unlabeled graphs may be slightly less than 146,129 for a polymer task on Plym-Melting, Plym-Density or Plym-Oxygen.\nBecause we remove the overlapping graph examples for the current polymer task with the polymer unlabeled data. We follow [52] to split the datasets to characterize imbalanced training distributions and balanced test distributions. The details of the age regression dataset are presented in Table 1 (Superpixel-Age) and Figure 3. The graph dataset Superpixel-Age is constructed from image superpixels using the algorithms from [22] on the image dataset AgeDB-DIR from [31,52]. Each face image in AgeDB-DIR has an age label from 0 to 101. We fisrt compute the SLIC superpixels for each image without losing the label-specific information [1,22]. Then we use the superpixels as nodes and calculate the spatial distance between superpixels to build edges for each image [22]. Binary edges are constructed between superpixel nodes by applying a threshold on the top-5% of the smallest spatial distances. After building a graph for each image, the graph dataset Superpixel-Age consists of 3,619 graphs for training, 628 graphs for validation, 628 graphs for testing, and 11,613 unlabeled graphs for semi-supervised learning."
},
{
"figure_ref": [
"fig_2",
"fig_5"
],
"heading": "Evaluation metrics.",
"publication_ref": [
"b12",
"b34",
"b51",
"b51",
"b51",
"b34",
"b12",
"b39",
"b24",
"b50"
],
"table_ref": [
"tab_3",
"tab_3"
],
"text": "We report model performance on three different sub-ranges following the work in [13,35,52], besides the entire range of label space. The three sub-ranges are the many-shot region, medium-shot region, and few-shot region. The sub-ranges are defined by the number of training graphs in each label value interval.\nDetails for each dataset are presented in Figure 2 and Figure 3. To evaluate the regression performance, we use mean absolute error (MAE) and geometric mean (GM) [52]. Lower values (↓) of MAE or GM indicate better performance. 5.1.3 Baselines and Implementations. Besides the GNN model, we broadly consider baselines from the fields of imbalanced regression and semi-supervised graph learning. Specifically, imbalanced regression baselines include LDS [52], BMSE [35], and RankSim [13].\nThe semi-supervised graph learning baseline is InfoGraph [40] and the graph learning baseline is GREA [25]. To implement SGIR and the baselines, the GNN encoder is GIN [51] and the decoder is a three-layer MLP to output property values. The threshold 𝜏 for selecting confident predictions is determined by the value at a certain percentile of the confidence score distribution. For all the methods, we reports the results on the test sets using the mean (standard deviation) over 10 runs with parameters that are randomly initialized. More Implementation details are in Appendix. 2 presents results of SGIR and baseline methods on six graph regression tasks. We have three observations. Overall effectiveness in the entire label range: SGIR performs consistently better than competitive baselines on all tasks. Columns \"All\" in Table 2 show that SGIR reduces MAE over the best baselines (whose MAEs are underlined in the table) relatively by 9.1%, 8.1%, and 12.3% on the three molecule datasets, respectively. Specifically, on Mol-FreeSolv, the MAE was reduced from 0.642 to 0.563 with no change on the standard deviation. This is because SGIR enrich and balance the training data with confidently predicted pseudo-labels and augments for data examples on all the possible label ranges, whereas all the baseline models suffer from the bias caused by imbalanced annotations."
},
{
"figure_ref": [],
"heading": "RQ1: Effectiveness on Property Prediction",
"publication_ref": [],
"table_ref": [],
"text": "Effectiveness in few-shot label ranges: The performance improvements of SGIR on graph regression tasks are simultaneously from three different label ranges: many-shot region, medium-shot region, and few-shot region. By looking at the results of baselines, we find that the best performance at a particular range would sacrifice the performance at a different label range. For example, on the Mol-Lipo and Mol-FreeSolv datasets, while GREA is the second best and best baseline, respectively, in the many-shot region, its performance in the few-shot region is worse than the basic GNN models. Similarly, on the Mol-FreeSolv dataset, LDS reduces the MAE from GNN relatively by +3.5% in the few-shot region with a trade-off of a -29% performance decrease in the many-shot region. Compared to baselines, the improvements from SGIR in the under-represented label ranges are theoretically guaranteed without sacrificing the performance in the well-represented label range. And our experimental observations support the theoretical guarantee, even in more challenging scenarios, i.e., predictions in the label ranges of fewer training shots on smaller datasets. Specifically, SGIR reduces MAE relatively by 30.3% and 9.0% in the few-shot region on Mol-FreeSolv and Plym-Oxygen. Because SGIR leverages the mutual enhancement of model construction and data balancing: the gradually balanced training data reduce model bias to popular "
},
{
"figure_ref": [],
"heading": "Effectiveness on different graph regression tasks:",
"publication_ref": [
"b33",
"b24",
"b33"
],
"table_ref": [],
"text": "We observe that the improvements on molecule regression tasks are more significant than those on polymer regression tasks. We hypothesize the reasons to be (1) the quality of unlabeled source data and (2) the size of the label space. First, our unlabeled graphs consist of more than a hundred thousand unlabeled small molecule graphs from QM9 [34] and around ten thousand polymers (macromolecules) from [25]. The massive quantity of unlabeled molecules make it easier to have good quality pseudo-labels and augmented examples for the three small molecule regression tasks on Mol-Lipo, Mol-ESOL, and Mol-FreeSolv [34]. Because the majority of unlabeled molecule graphs have a big domain gap with the polymer regression tasks, the quality of expanded training data in polymer regression tasks would be relatively worse than the quality of those ✓ ✗ ✗ 0.604(0.020) 0.557(0.037) 0.560(0.029) 0.903(0.055) ✗ ✓ ✗ 0.660(0.028) 0.574(0.015) 0.650(0.036) 0.941(0.066) ✓ ✓ ✗ 0.568(0.029) 0.538(0.020) 0.520(0.045) 0.831(0.132) ✗ ✗ ✓ 0.593(0.045) 0.536(0.033) 0.542(0.067) 0.947(0.062) ✓ ✓ ✓ 0.563(0.026) 0.535(0.038) 0.528(0.046) 0.777(0.061) Table 5: Investigation on choices of regression confidence with the metric MAE (↓). We disable all other SGIR components except the regression confidence score. Our confidence score (GRation) in Eq. ( 1) removes noise more effectively than others in graph regression tasks. in molecule regression. This inspires us to collect more polymer data in the future, even if their properties could not be annotated."
},
{
"figure_ref": [
"fig_2"
],
"heading": "Choice of 𝜎",
"publication_ref": [],
"table_ref": [],
"text": "Second, Figure 2 has shown that the label ranges in the polymer regression tasks are usually much wider than the ranges in the molecule regression tasks. This poses a great challenge for accurate predictions, especially when we train with a small dataset."
},
{
"figure_ref": [],
"heading": "Effectiveness on Age Prediction.",
"publication_ref": [],
"table_ref": [
"tab_4",
"tab_4",
"tab_4"
],
"text": "Besides molecules and polymers, Table 3 presents more results by comparing different methods on the Superpixel-Age dataset. SGIR consistently improves the model performance compared to the best baselines in different label ranges. In the entire label range, SGIR reduces the MAE (GM) relatively by +4.7% (+3.6%). The advantages mainly stem from the enhancements in the few-shot region, as demonstrated in Table 3, which shows an improvement of +4.3% and +3.1% on the MAE and GM metrics, respectively. Different from LDS, SGIR improves the model performance for the under-represented and well-represented label ranges at the same time. Table 3 showcases that the empirical advantages of SGIR could generalize across different domains."
},
{
"figure_ref": [],
"heading": "RQ2: Ablation Studies on Framework Design",
"publication_ref": [
"b4"
],
"table_ref": [],
"text": "We conduct five studies and analyze the results below. Four ablation studies are (1) G conf and H aug for data balancing; (2) mutually enhanced iterative process; (3) choices of confidence score; and (4) quality and diversity of the label-anchored mixup. (5) The sensitivity analysis for the label interval number 𝐶. Readers can refer to the appendix for complete results. "
},
{
"figure_ref": [],
"heading": "Effect of balancing data with different components in G conf and H aug . Studies on molecule regression tasks in",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Effect of regression confidence measurements.",
"publication_ref": [
"b24",
"b11"
],
"table_ref": [],
"text": "Table 5 shows that compared to existing methods that could define regression confidence, the measurement we define and use, GRation, is the best option for evaluating the quality of pseudo-labels in graph regression tasks. Because GRation uses various environments subgraphs, which provide diverse perturbations for robust graph learning [25]. We also observe that Dropout can be a good alternative of GRation. Dropout has extensive assessments [12] and makes it possible for SGIR to be extended to regression tasks for other data types such as images and texts."
},
{
"figure_ref": [],
"heading": "5.3.4",
"publication_ref": [],
"table_ref": [
"tab_7",
"tab_5",
"tab_5",
"tab_5"
],
"text": "Effect of label-anchored mixup augmentation. We implement z 𝑖 using G imb to improve the augmentation quality and G imb ∪ G unlbl to improve the diversity. Table 6 presents extensive empirical studies to support our idea. It shows that when many noisy representation vectors from unlabeled graphs are included in the interval center z 𝑖 , the quality of augmented examples is relatively low, which degrades the model performance in different label ranges.\nOn the other hand, the representations of unlabeled graphs improve the diversity of the augmented examples when we assign low mixup weights to them as in Eq. ( 5). Considering both quality and diversity, the effectiveness of the algorithm is further demonstrated in Table 4 by significantly reducing the errors for rare labels. From the fifth line of each dataset in Table 4, we find that it is also promising to directly use the label-anchored mixup augmentation (as G imb ∪ H aug ) for data balancing. Although its performance may be inferior to the performance using G imb ∪ G conf (as the third line of each dataset in Table 4), the potential of the label-anchored mixup algorithm could be further enhanced by improving the quality of the augmented examples to close the gap with real molecular graphs."
},
{
"figure_ref": [],
"heading": "5.3.5",
"publication_ref": [],
"table_ref": [],
"text": "Sensitivity analysis for the label interval number 𝐶. We find the best values of 𝐶 in main experiments using the validation set for pseudo-labeling and label-anchored mixup. We suggest setting the number 𝐶 to approximately 100 for pseudo-labeling and around 1,000 for label-anchored mixup. Specifically, sensitivity analysis is conducted on the Plym-Oxygen dataset to analyze the effect of the number 𝐶. Results are presented in Figure 5. We observe that SGIR is robust to a wide range of choices for the number of intervals."
},
{
"figure_ref": [],
"heading": "CONCLUSIONS",
"publication_ref": [],
"table_ref": [],
"text": "In this work, we explored a novel graph imbalanced regression task and improved semi-supervised learning on it. We proposed a self-training framework to gradually reduce the model bias of data imbalance through multiple iterations. In each iteration, we selected more high-quality pseudo-labels for rare label values and continued augmenting training data to approximate the perfectly balanced label distribution. Experiments demonstrated the effectiveness and reasonable design of the proposed framework, especially on material science. Graph data? Imbalance? Regression? (Otherwise, assuming:\n) (Supervised) (Non-graph) (Balance) (Classification) DARP [20] ✓ ✓ DASO [32] ✓ ✓ Bi-Sampling [16] ✓ ✓ CADR [17] ✓ ✓ CReST [47] ✓ ✓ LDS [52] ✓ ✓ BMSE [35] ✓ ✓ RankSim [13] ✓ ✓ SSDKL [18] ✓ ✓ InfoGraph [40] ✓ ✓ ✓ SGIR (Ours) ✓ ✓ ✓ ✓"
},
{
"figure_ref": [],
"heading": "A FURTHER RELATED WORK A.1 A Systematic Comparison with Related Methods",
"publication_ref": [],
"table_ref": [
"tab_9"
],
"text": "We compare SGIR with a line of related work on four important settings of research problem in Table 7. From the table we find that existing work mostly focused on solving imbalance problems in semi-supervised classification tasks with categorical labels and non-graph data. There lacks an exploration of research on semisupervised learning and imbalance learning for graph regression."
},
{
"figure_ref": [],
"heading": "A.2 Sampling Strategy in Self-training",
"publication_ref": [
"b46",
"b15"
],
"table_ref": [],
"text": "To the best of our knowledge, reverse sampling is one of the most suitable sampling strategies to address class imbalance issues in selftraining [47]. Compared to other strategies like random sampling or mean sampling [16], reverse sampling is also the most suitable one for graph imbalanced regression. This is because reverse sampling compensates for the label imbalance and enriches training examples. Other strategies cannot make the training data more balanced. They would lead the prediction model to be still biased to the majority of data. More complex sampling strategies that combine reverse sampling, mean sampling, and random sampling would be a promising direction for future work."
},
{
"figure_ref": [],
"heading": "B PROOFS OF THEORETICAL MOTIVATIONS",
"publication_ref": [],
"table_ref": [],
"text": "We rely on two theorems to derive theorem 4.1."
},
{
"figure_ref": [],
"heading": "B.1 Existing Theorems",
"publication_ref": [
"b2",
"b18",
"b18",
"b6",
"b6"
],
"table_ref": [],
"text": "Given a classifier 𝑓 from the function class F , an input example 𝑥 from the feature space X and its label 𝑦.\nTheorem B.1 (from [3,19]). Assume the expected loss on examples is E [𝑓 ] and the corresponding empirical loss Ê\n[𝑓 ]. Assume the loss is Lipschitz with Lipschitz constant 𝐿 𝑒 . And it is bounded by 𝑐 0 . For any 𝛿 > 0 and with probability at least 1 -𝛿 simultaneously for all 𝑓 ∈ F we have that\nE [𝑓 ] ≤ Ê [𝑓 ] + 2𝐿 𝑒 R 𝑛 (F ) + 𝑐 0 √︂ log(1/𝛿) 2𝑛 , (10\n)\nwhere 𝑛 is the number of example and R𝑛 (F ) is the Rademacher complexity measurement of the hypothesis class F . . Assume ∀𝑓 ∈ F we have sup 𝑥 ∈X |𝑓 (𝑥)| ≤ 𝑐 1 . Then, with probability at least 1 -𝛿 over the example, for all margins 𝛾 > 0 and all 𝑓 ∈ F we have,\nE [𝑓 ] ≤ 𝐾 𝛾 [𝑓 ] + 4 R 𝑛 (F ) 𝛾 + √︄ 2 log log 2 (4𝑐 1 /𝛾) + log(1/𝛿) 2𝑛 ,(11)\n≤ 𝐾 𝛾 [𝑓 ] + 4 R 𝑛 (F ) 𝛾 + log log 2 4𝑐 1 𝛾 𝑛 + √︂ log(1/𝛿) 2𝑛 .(12)\nB.2 Proof of theorem 4.1\nIn our work, we use the regression function 𝑓 to predict the label value. We calculate the reciprocal of the distance between the predicted label and interval centers as unnormalized probabilities of the graph 𝑆 [𝑏 𝑖 ,𝑏 𝑖+1 ) (𝐺) being assigned to the interval [𝑏 𝑖 , 𝑏 𝑖+1 ), 𝑖 ∈ {1, 2, . . . , 𝐶}. Given a hard margin 𝛾, we use E 𝛾, [𝑏 𝑖 ,𝑏 𝑖+1 ) [𝑓 ] to denote the hard margin loss for examples in the interval [𝑏 𝑖 , 𝑏 𝑖+1 ):\nE 𝛾, [𝑏 𝑖 ,𝑏 𝑖+1 ) [𝑓 ] = Pr 𝐺∼P [𝑏 𝑖 ,𝑏 𝑖+1 ) 𝑆 [𝑏 𝑖 ,𝑏 𝑖+1 ) (𝐺) < max 𝑗≠𝑖 𝑆 [𝑏 𝑗 ,𝑏 𝑗 +1 ) (𝐺) + 𝛾 .(13)\nWe assume its empirical variant is Ê𝛾,[𝑏 𝑖 ,𝑏 𝑖+1 ) [𝑓 ]. The empirical Rademacher complexity R(𝑏 𝑖 ,𝑏 𝑖+1 ] (F ) is used as the complexity measurement C(F ) for the hypothesis class F . With a vector 𝜎 of i.i.d. uniform {-1, +1} bits, we have\nR(𝑏 𝑖 ,𝑏 𝑖+1 ] (F ) =(14)\n1\n𝑛 (𝑏 𝑖 ,𝑏 𝑖+1 ] E 𝜎 sup 𝑓 ∈ F ∑︁ 𝐺 𝑖 ∈ [𝑏 𝑖 ,𝑏 𝑖+1 ) 𝜎 𝑖 𝑆 [𝑏 𝑖 ,𝑏 𝑖+1 ) (𝐺 𝑖 ) -max 𝑗≠𝑖 𝑆 [𝑏 𝑗 ,𝑏 𝑗 +1 ) (𝐺 𝑖 ) (15)\nAs any 𝐺 𝑖 in the interval (𝑏 𝑖 , 𝑏 𝑖+1 ] is an i.i.d. sample from the distribution P [𝑏 𝑖 ,𝑏 𝑖+1 ) , we directly apply the standard margin-based generalization bound theorem B.2 [19]: with probability 1 -𝛿, for all choices of 𝛾 [𝑏 𝑖 ,𝑏 𝑖+1 ) > 0 and 𝑓 ∈ F ,\nE [𝑏 𝑖 ,𝑏 𝑖+1 ) ≤ Ê𝛾,[𝑏 𝑖 ,𝑏 𝑖+1 ) [𝑓 ] + 4 R(𝑏 𝑖 ,𝑏 𝑖+1 ] (F ) 𝛾 [𝑏 𝑖 ,𝑏 𝑖+1 )(16)\n+ 2 log log 2 ( 4𝑐 1 𝛾 [𝑏 𝑖 ,𝑏 𝑖+1 ) ) + log(1/𝛿) 2𝑛 [𝑏 𝑖 ,𝑏 𝑖+1 ) , ≤ Ê𝛾,[𝑏 𝑖 ,𝑏 𝑖+1 ) [𝑓 ] +1\n𝛾 [𝑏 𝑖 ,𝑏 𝑖+1 ) √︄ C(F ) 𝑛 [𝑏 𝑖 ,𝑏 𝑖+1 )(17)\n+ 2 log log 2 ( 4𝑐 1 𝛾 [𝑏 𝑖 ,𝑏 𝑖+1 ) ) log(1/𝛿) 2𝑛 [𝑏 𝑖 ,𝑏 𝑖+1 ) , ⪅1\n𝛾 [𝑏 𝑖 ,𝑏 𝑖+1 ) √︄ C(F ) 𝑛 [𝑏 𝑖 ,𝑏 𝑖+1 ) + √︄ log log 2 (1/𝛾 [𝑏 𝑖 ,𝑏 𝑖+1 ) ) + log(1/𝛿) 2𝑛 [𝑏 𝑖 ,𝑏 𝑖+1 ) .(18)\nWe derive Eq. ( 17) from Eq. ( 16) because the Rademacher complexity R(𝑏 𝑖 ,𝑏 𝑖+1 ] (F ) typically scales as √︂\nC( F)\n𝑛 (𝑏 𝑖 ,𝑏 𝑖+1 ] for some complexity measurement C(F ) [7]. We derive Eq. ( 18) from Eq. ( 17) by ignoring constant factors [7]. Since the overall performance E bal [𝑓 ] is calculated over all intervals, we get it as\nE bal [𝑓 ] = 1 𝐶 𝐶 𝑖=1 E [𝑏 𝑖 ,𝑏 𝑖+1 ) ."
},
{
"figure_ref": [],
"heading": "B.3 Discussions",
"publication_ref": [
"b7",
"b25",
"b54",
"b52",
"b37",
"b9",
"b38"
],
"table_ref": [],
"text": "Existing work on the theoretical analysis of mixup [8,26,55] mainly focused on image classification and cannot guarantee that the augmented graph examples are i.i.d sampled from the conditional distribution P [𝑏 𝑖 ,𝑏 𝑖+1 ] for a specific interval [𝑏 𝑖 , 𝑏 𝑖+1 ]. While a recent work proposed the C-mixup [53] to sample closer pairs of examples with higher probability for regression tasks, it did not fit our theoretical motivation to address the label imbalance issue: with C-mixup, the pairs in the over-represented label ranges have a higher probability to be sampled than the under-represented ones. Compared to these theories and methods for the mixup algorithm, our label-anchored mixup allows direct application to imbalanced regression tasks without compromising the assumption in our theoretical motivation. This is because we use the augmented virtual examples H aug based on the label anchor within intervals [𝑏 𝑖 , 𝑏 𝑖+1 ]. Augmented examples are independently created with Eq. ( 5). Since the interval centers could be mixed with any other real graphs from G imb ∪ G conf , any value in the interval space could be sampled. Besides, it is reasonable to use the distribution of the entire label space (from G imb ∪ G conf ) to approximate the distribution within the interval and assume that the conditional distribution P [𝑏 𝑖 ,𝑏 𝑖+1 ] does not change.\nWe build the theoretical principle for imbalanced regression with intervals to connect with existing theoretical principles for classification. Future theoretical work on imbalanced regression can leverage the advantages of using mixture regressor models [38], which have been used to address covariate shift problems in regression tasks. Additionally, exploring the promising connection between domain adaptation theories and sample selection bias [10,39] holds potential for further advancements in this field."
},
{
"figure_ref": [],
"heading": "C EXPERIMENTS C.1 Dataset Details",
"publication_ref": [
"b12",
"b51",
"b28"
],
"table_ref": [],
"text": "We give a comprehensive introduction to our datasets used for regression tasks and splitting idea from [13,52].\nMol-Lipo. It is a dataset to predict the property of lipophilicity consisting of 4200 molecules. The lipophilicity is important for solubility and membrane permeability in drug molecules. This dataset originates from ChEMBL [29]. The property is from experimental results for the octanol/water distribution coefficient (log 𝐷 at pH 7.4)."
},
{
"figure_ref": [
"fig_5"
],
"heading": "Mol-ESOL.",
"publication_ref": [
"b32",
"b32",
"b41",
"b33",
"b24",
"b24",
"b12",
"b51",
"b21",
"b30",
"b51",
"b0",
"b21",
"b21",
"b51"
],
"table_ref": [
"tab_2"
],
"text": "It is to predict the water solubility (log solubility in mols per litre) from chemical structures consisting of 1128 small organic molecules.\nMol-FreeSolv. It is to predict the hydration free energy of molecules in water consisting of 642 molecules. The property is experimentally measured or calculated.\nPlym-Melting. It is used to predict the property of melting temperature ( • C). It is collected from PolyInfo, a web-based polymer database [33].\nPlym-Density. It is used to predict the property of polymer density (g/cm 3 ). It is collected from PolyInfo, a web-based polymer database [33].\nPlym-Oxygen. It is used to predict the property of oxygen permeability (Barrer). It is created from the Membrane Society of Australasia portal consisting of experimentally measured gas permeability data [42].\nUnlabeled Data for Molecules and Polymers. The total number of unlabeled graphs for molecule and polymers is 146,129, consisting of 133,015 molecules from QM9 [34] and 13,114 monomers (the repeated units of polymers) from [25]. QM9 is a molecule dataset for stable small organic molecules consisting of atoms C, H, O, N, and F. We use it as a source of unlabeled data. We integrate four polymer regression datasets including Plym-Melting, Plym-Density, Plym-Oxygen and another one from [25] for the glass transition temperature as the other source of unlabeled data. We note that the unlabeled graphs may be slightly less than 146,129 for a polymer task on Plym-Melting, Plym-Density or Plym-Oxygen. It is because we remove the overlapping graphs for the current polymer task with the polymer unlabeled data.\nData splitting for Molecules and Polymers. We split the datasets based on the approach in previous works [13,52] motivated for two reasons. First, we want the training sets to well characterize the imbalanced label distribution as presented in the original datasets. Second, we want relatively balanced valid and test sets to fairly evaluate the model performance in different ranges of label values. Superpixel-Age. The details of the age regression dataset are presented in Table 1 (Superpixel-Age) and Figure 3. The graph dataset Superpixel-Age is constructed from image superpixels using the algorithms from [22] on the image dataset AgeDB-DIR from [31,52]. Each face image in AgeDB-DIR has an age label from 0 to 101. We fisrt compute the SLIC superpixels for each image without losing the label-specific information [1,22]. Then we use the superpixels as nodes and calculate the spatial distance between superpixels to build edges for each image [22]. Binary edges are constructed between superpixel nodes by applying a threshold on the top-5% of the smallest spatial distances. After building a graph for each image, we follow the data splitting in [52] to study the imbalanced regression problem. We randomly remove 70% labels in the training/validation/test data and use them as unlabeled graphs. Finally, the graph dataset Superpixel-Age consists of 3,619 graphs for training, 628 graphs for validation, 628 graphs for testing, and 11,613 unlabeled graphs for semi-supervised learning."
},
{
"figure_ref": [],
"heading": "C.2 Implementation Details",
"publication_ref": [
"b50",
"b24"
],
"table_ref": [],
"text": "We use the Graph Isomorphism Network (GIN) [51] as the GNN encoder for 𝑓 𝜃 to get the graph representation and three layers of Multilayer perceptron (MLP) as the decoder to predict graph properties. The threshold 𝜏 for selecting confident predictions is determined by the value at a certain percentile of the confidence score distribution. To implement it, we set it up as a hyperparameter 𝜏 pct determining the percentile value of the prediction variance (i.e., the reciprocal of confidence) of the labeled training data. In experiments, all methods are implemented on Linux with Intel Xeon Gold 6130 Processor (16 Cores @2.1Ghz), 96 GB of RAM, and a RTX 2080Ti card (11 GB RAM). For all the methods, we reports the results on the test sets using the mean (standard deviation) over 10 runs with parameters that are randomly initialized. Note that the underlying design of the graph learning model used in SGIR is GREA with a learning objective as follows. Given (𝐺, 𝑦) ∈ G imb ∪ G conf , GREA [25] will output a vector m ∈ R 𝐾 that indicates the probability of 𝐾 nodes in a graph being in the rationale subgraph. So, we could get h\n(𝑟 ) = 1 ⊤ 𝐾 • (m × H) and h (𝑒) = 1 ⊤ 𝐾 • ((1 𝐾 -m) × H)\n, where H ∈ R 𝐾×𝑑 is the node representation matrix. By this, the optimization objectives of a graph consist of\n ℓ imb+conf = MAE(𝑓 (h (𝑟 ) ), 𝑦) + E 𝐺 ′ MAE(𝑓 (h + h ′ ), 𝑦) + Var 𝐺 ′ {MAE(𝑓 (h + h ′ ), 𝑦)} , ℓ regu = 1 𝐾 𝐾 𝑘=1\n|m 𝑘 | -𝛾 ℓ regu regularizes the vector m and 𝛾 ∈ [0, 1] is a hyperparameter to control the expected size of 𝐺 (𝑟 ) . 𝐺 ′ is the possible graph in the same batch that provides environment subgraphs and h ′ is the representation vector of the environment subgraph. When combining the rationale-environment pairs to create new graph examples, the original GREA creates the same number of examples for the underrepresented rationale and the well/over-represented rationale. We observe that it may make the training examples more imbalanced. Therefore, we use the reweighting technique to penalize more for the expectation term (E 𝐺 ′ MAE(𝑓 (h + h ′ ), 𝑦) ) and variance term (Var 𝐺 ′ {MAE(𝑓 (h + h ′ ), 𝑦)} ) in ℓ imb+conf when the label is from the under-represented ranges. The weight of the expectation and variance terms for a graph with label 𝑦 is\n𝑤 = exp( 𝐵 𝑏=1 |𝑦 -𝑦 𝑏 |/𝑡) exp( 𝐵 𝑗=1 𝐵 𝑏=1 |𝑦 -𝑦 𝑏 |/𝑡) ,\nwhere 𝐵 is the batch size and 𝑡 is the temperature hyper-parameter."
},
{
"figure_ref": [],
"heading": "C.3 Additional Experimental Results",
"publication_ref": [],
"table_ref": [
"tab_10"
],
"text": "Additional results on the effect of balancing data and label-anchored mixup. Table 8 and Table 9 present studies on the effect of balancing data and different options in the label-anchored mixup augmentation for molecules and polymers, respectively. They provide more evidence to our observations that (1) the effect of our pseudolabeling method (G imb ∪ G conf ) about improving the model performance in the entire label range and the few-shot region; (2) the essential role of the regression confidence 𝜎 and reverse sampling rate 𝑝 in our pseudo-labeling about improving pseudo-label quality and reducing imbalance label bias; and (3) the complementary effect of H aug about approximating the perfect balance of the training distribution.\nAdditional results on the regression confidence measurements. Table 10 show all comparisons among different confidence measurements. GRation consistently performs best in the entire label range excepting dataset Plym-Density on which Dropout is slightly better than GRation. "
},
{
"figure_ref": [],
"heading": "ACKNOWLEDGMENTS",
"publication_ref": [],
"table_ref": [],
"text": "This work was supported in part by NSF IIS-2142827, IIS-2146761, and ONR N00014-22-1-2507."
}
] | Data imbalance is easily found in annotated data when the observations of certain continuous label values are difficult to collect for regression tasks. When they come to molecule and polymer property predictions, the annotated graph datasets are often small because labeling them requires expensive equipment and effort. To address the lack of examples of rare label values in graph regression tasks, we propose a semi-supervised framework to progressively balance training data and reduce model bias via self-training. The training data balance is achieved by (1) pseudo-labeling more graphs for under-represented labels with a novel regression confidence measurement and (2) augmenting graph examples in latent space for remaining rare labels after data balancing with pseudo-labels. The former is to identify quality examples from unlabeled data whose labels are confidently predicted and sample a subset of them with a reverse distribution from the imbalanced annotated data. The latter collaborates with the former to target a perfect balance using a novel label-anchored mixup algorithm. We perform experiments in seven regression tasks on graph datasets. Results demonstrate that the proposed framework significantly reduces the error of predicted graph properties, especially in under-represented label areas. | Semi-Supervised Graph Imbalanced Regression | [
{
"figure_caption": "𝑖 . For the 𝑗-th graph 𝐺 𝑗 in the same batch, we have a combined example 𝐺 (𝑖,𝑗)",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Oxygen (log y used) ",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: Imbalanced training distributions G imb of annotated molecule and polymers.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Theorem 4 . 1 .41With probability (1 -𝛿) over the randomness of the training data, the error E [𝑏 𝑖 ,𝑏 𝑖+1 ) for interval [𝑏 𝑖 , 𝑏 𝑖+1 ) is bounded by",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "41",
"figure_type": "figure"
},
{
"figure_caption": "5. 2 . 121Effectiveness on Molecule and Polymer Prediction. Table",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "21",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: Imbalanced training distributions G imb in the Superpixel-Age dataset.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "In the few-shot region.",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :Figure 5 :45Figure 4: Test performance of SGIR through multiple self-training iterations. MAE for Plym-Density is scaled by ×1, 000. The iterative self-training algorithm is effective for gradually improving the quality of training data.",
"figure_data": "",
"figure_id": "fig_7",
"figure_label": "45",
"figure_type": "figure"
},
{
"figure_caption": "Theorem B. 2 (2from [19]). Applying theorem B.1 and considering the fraction of data having 𝛾-margin mistakes, or 𝐾 𝛾 [𝑓 ] := |𝑖:𝑦 𝑖 𝑓 (𝑥 𝑖 ) <𝛾 | 𝑛",
"figure_data": "",
"figure_id": "fig_8",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "To let data balancing and model construction mutually enhance each other, SGIR is a self-training framework that trains the encoder 𝑔(•) and decoder 𝑓 (•) using two strategies through multiple iterations. The first strategy is to use pseudo-labels based on confident predictions and reverse sampling, leveraging unlabeled data (see Section 4.2). Because the unlabeled graph set still may not contain real examples of rare label values, the second strategy is to augment the graph representation examples for the rare areas using a novel label-anchored mixup algorithm (see Section 4.3).",
"figure_data": "4.1 A Self-Training Framework for IterativelyBalancing Scalar Label DataA classic self-training framework is expected to be a virtuouscircle exploiting the unlabeled data in label-balanced classifica-tion/regression tasks [28, 50]. It first trains a classifier/regressorthat iteratively assigns pseudo-labels to the set of unlabeled train-ing examples G unlbl with a margin greater than a certain threshold.The pseudo-labeled examples are then used to enrich the labeledtraining set. And the classifier continues training with the updatedtraining set. For a virtuous circle of model training with imbalancedlabeled set G imb , the most confident predictions on G unlbl shouldbe selected to compensate for the under-represented labels, as wellas to enrich the dataset G imb . In each iteration, the model becomesless biased to the majority of labels. And the less biased modelcan make predictions of higher accuracy and confidence on theunlabeled data. Therefore, we hypothesize that model training anddata balancing can mutually enhance each other.SGIR is a self-training framework targeting to generalize themodel performance everywhere in the continuous label space withparticularly designed balanced training data from the labeled graphdata G imb , confidently selected graph data G conf , and augmentedrepresentation data H aug . For the next round of model training, thegradually balanced training data reduce the label imbalance biascarried by the graph encoder 𝑔(•) and decoder 𝑓 (•). Then the lessbiased graph encoder and decoder are applied to generate balancedtraining data of higher quality. Through these iterations, the modelbias from the imbalanced or low-quality balanced data would beprogressively reduced because of the gradually enhanced qualityof balanced training data.",
"figure_id": "tab_0",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "to calculate it. Basically, to perform reverse sampling, we want 𝑝 𝑖 < 𝑝 𝑗 if 𝜇 𝑖 > 𝜇 𝑗 . We define a new frequency set {𝜇 ′ 𝑖 } 𝐶 𝑖=1 in which 𝜇 ′ 𝑖 equals the 𝑖-th smallest in {𝜇} if 𝜇 𝑖 is the 𝑖-th biggest in {𝜇}. Then the sampling rate is",
"figure_data": "",
"figure_id": "tab_1",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Proofs are in appendix B. The bound decreases as the increase of the examples in corresponding label ranges. The SGIR is motivated to reduce and balance the bound for different intervals by manipulating 𝑛 [𝑏 𝑖 ,𝑏 𝑖+1 ) with pseudo-labels and augmented examples. Particularly, we discuss in appendix B.3 that the augmented Statistics of six tasks for graph property regression.",
"figure_data": "Taking union bound over all intervals,",
"figure_id": "tab_2",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Results of Mean(Std) on six molecule/polymer datasets. The best mean is bolded. The best baseline is underlined.",
"figure_data": "MAE ↓GM ↓AllMany-shot Med.-shot Few-shotAllMany-shot Med.-shot Few-shotGNN0.485(0.010) 0.421(0.030) 0.462(0.013) 0.566(0.032)0.297(0.012) 0.252(0.022) 0.294(0.016) 0.348(0.030)RankSim0.475(0.018) 0.388(0.017) 0.438(0.007) 0.587(0.043)0.297(0.015) 0.249(0.017) 0.274(0.006) 0.380(0.044)BMSE0.494(0.007) 0.409(0.019) 0.450(0.007) 0.614(0.033)0.304(0.008) 0.260(0.014) 0.279(0.015) 0.382(0.038)Mol-LipoLDS0.468(0.009) 0.394(0.012) 0.449(0.012) 0.551(0.026)0.294(0.010) 0.251(0.009) 0.281(0.010) 0.356(0.033)InfoGraph 0.499(0.008) 0.421(0.024) 0.471(0.013) 0.596(0.026)0.314(0.011) 0.269(0.018) 0.300(0.006) 0.376(0.029)GREA0.487(0.002) 0.391(0.015) 0.434(0.008) 0.626(0.018)0.294(0.010) 0.251(0.009) 0.281(0.010) 0.356(0.033)SGIR0.432(0.012) 0.357(0.019) 0.413(0.017) 0.515(0.020)0.264(0.013) 0.224(0.016) 0.256(0.017) 0.314(0.015)GNN0.508(0.015) 0.398(0.018) 0.448(0.012) 0.696(0.025)0.299(0.017) 0.231(0.017) 0.279(0.014) 0.425(0.035)RankSim0.501(0.014) 0.389(0.021) 0.443(0.019) 0.689(0.025)0.293(0.021) 0.227(0.028) 0.258(0.020) 0.449(0.030)BMSE0.533(0.023) 0.400(0.027) 0.449(0.015) 0.777(0.069)0.308(0.018) 0.245(0.036) 0.266(0.009) 0.473(0.035)Mol-ESOLLDS0.517(0.016) 0.423(0.012) 0.474(0.029) 0.668(0.010)0.304(0.010) 0.261(0.007) 0.283(0.025) 0.393(0.009)InfoGraph 0.561(0.025) 0.475(0.034) 0.466(0.036) 0.776(0.036)0.336(0.014) 0.306(0.022) 0.276(0.013) 0.484(0.029)GREA0.497(0.031) 0.396(0.040) 0.456(0.033) 0.652(0.045)0.289(0.032) 0.226(0.038) 0.270(0.025) 0.404(0.051)SGIR0.457(0.015) 0.370(0.022) 0.411(0.011) 0.604(0.024)0.263(0.016) 0.226(0.021) 0.240(0.015) 0.347(0.030)GNN0.726(0.039) 0.617(0.061) 0.695(0.055) 1.154(0.082)0.363(0.025) 0.317(0.027) 0.360(0.029) 0.556(0.073)RankSim0.779(0.109) 0.764(0.225) 0.674(0.072) 1.220(0.146)0.367(0.026) 0.396(0.052) 0.315(0.030) 0.537(0.082)BMSE0.856(0.071) 0.809(0.117) 0.820(0.064) 1.122(0.076)0.456(0.042) 0.426(0.029) 0.457(0.054) 0.552(0.062)Mol-FreeSolvLDS0.809(0.071) 0.796(0.071) 0.737(0.088) 1.114(0.141)0.443(0.045) 0.489(0.036) 0.387(0.052) 0.580(0.146)InfoGraph 0.933(0.042) 0.830(0.081) 0.913(0.030) 1.308(0.171)0.542(0.048) 0.505(0.107) 0.528(0.038) 0.789(0.183)GREA0.642(0.026) 0.541(0.064) 0.570(0.008) 1.202(0.023)0.321(0.038) 0.294(0.064) 0.301(0.024) 0.537(0.049)SGIR0.563(0.026) 0.535(0.038) 0.528(0.046) 0.777(0.061)0.264(0.029) 0.286(0.013) 0.244(0.046) 0.304(0.078)GNN41.8(1.2)35.5(1.2)33.0(0.7)54.7(2.2)23.2(1.0)21.3(1.1)16.2(1.0)33.4(2.5)RankSim41.1(0.9)34.1(0.5)33.6(1.1)53.5(1.2)22.6(1.1)20.5(0.5)16.8(1.0)31.4(2.8)BMSE42.1(0.7)35.8(1.4)34.1(1.3)54.4(1.5)23.7(1.2)21.5(1.0)18.1(0.5)32.4(3.0)Plym-MeltingLDS41.6(0.3)35.3(0.9)34.5(1.1)53.2(0.8)23.2(0.2)20.5(1.2)18.3(0.5)31.4(1.1)InfoGraph43.6(2.8)35.3(2.3)35.0(2.3)58.3(4.1)24.6(1.9)21.3(1.5)18.4(1.5)35.4(4.1)GREA41.2(0.8)33.3(0.5)32.7(0.7)55.3(3.0)23.4(0.6)20.0(0.6)17.3(0.7)34.3(2.9)SGIR38.9(0.7)31.7(0.3)31.5(1.1)51.4(1.6)21.1(1.2)18.5(0.5)15.9(1.4)30.2(1.9)GNN61.2(5.4)63.4(18.9)46.6(1.6)72.0(2.8)29.3(0.6)29.6(3.3)23.5(0.9)35.5(2.0)RankSim57.5(1.8)55.1(2.2)46.3(1.8)69.4(3.3)29.3(1.6)29.9(2.8)23.1(2.1)35.4(2.5)BMSE61.8(2.0)59.1(8.6)48.2(2.0)75.9(3.5)31.9(1.3)31.8(4.2)26.3(2.2)38.2(3.2)Plym-DensityLDS60.1(2.4)60.4(6.2)47.0(1.3)71.3(2.5)31.5(2.0)33.2(3.5)24.4(3.0)38.0(2.4)(scaled:×1, 000)InfoGraph54.9(1.7)46.8(1.0)43.0(1.9)72.3(3.2)29.3(1.8)27.3(1.4)22.6(1.2)39.2(4.3)GREA60.3(1.9)49.0(4.4)48.1(2.5)80.7(4.2)32.3(1.6)26.7(2.7)27.2(2.3)44.7(6.1)SGIR53.0(0.5)45.4(1.7)42.5(2.8)68.6(2.6)26.6(0.4)24.0(2.2)23.0(1.3)33.4(3.0)GNN183.5(33.4)6.3(3.2)14.6(6.6)464.0(85.3)7.0(1.8)2.4(0.7)3.9(1.1)29.9(7.2)RankSim165.7(27.4)3.9(1.4)13.0(2.0)420.7(69.7)5.9(1.4)1.8(0.3)3.6(1.7)26.6(6.7)BMSE190.4(33.4)26.4(21.6)27.0(16.4)454.3(88.9)25.7(14.8)14.9(11.7)15.9(9.6)63.2(23.5)Plym-OxygenLDS180.0(23.0)6.6(4.0)11.8(2.0)456.3(60.2)7.6(1.6)2.4(0.6)4.7(1.4)33.6(9.2)InfoGraph 199.5(31.5)7.5(7.2)13.0(1.8)505.5(78.2)7.8(1.9)2.3(0.5)5.1(2.2)34.8(8.5)GREA182.5(30.0)9.0(8.6)14.4(4.9)458.8(79.2)7.1(1.3)2.1(0.5)4.4(1.3)31.7(5.0)SGIR150.9(17.8)3.8(1.1)12.2(0.6)382.8(46.9)5.8(0.4)2.1(0.7)3.3(0.8)24.4(6.8)labels; the less biased model improves the quality of pseudo-labelsand augmented examples in the few-shot region.",
"figure_id": "tab_3",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Results of Mean(Std) on the age prediction using graphs from image superpixels.The best mean is bold. The best baseline is underlined.",
"figure_data": "MAE ↓GM ↓AllMany-shot Med.-shot Few-shotAllMany-shot Med.-shot Few-shotGNN14.583(0.413) 10.524(0.994) 11.698(0.404) 22.127(0.780)9.996(0.386) 7.265(0.858) 7.910(0.492) 18.404(0.673)RankSim 14.464(0.401) 10.468(0.759) 11.610(0.774) 21.910(0.700)9.606(0.303) 6.936(0.598) 7.721(0.660) 17.534(1.768)BMSE15.179(0.594) 10.639(2.303) 12.201(0.900) 23.321(2.525)10.419(0.393) 7.249(1.526) 8.659(0.827) 19.719(4.318)LDS14.674(0.191) 10.972(0.495) 11.985(0.627) 21.623(0.926)9.867(0.291) 7.317(0.672) 7.997(0.633) 17.298(0.957)InfoGraph 14.515(0.605) 10.610(1.063) 11.150(0.158) 22.476(1.147)9.879(0.524) 7.391(0.995) 7.377(0.333) 18.969(1.873)GREA14.682(0.300) 10.283(0.503) 11.999(0.585) 22.329(0.570)10.037(0.438) 7.051(0.455) 8.273(0.565) 18.142(1.276)SGIR13.787(0.123) 10.171(0.4156) 11.066(0.389) 20.687(0.839)9.261(0.221) 6.928(0.355) 7.247(0.593) 16.769(1.418)",
"figure_id": "tab_4",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "A comprehensive ablation study on molecule regression datasets with the metric MAE (↓). 𝜎 is the confidence score in Section 4.2.1. 𝑝 is the reverse sampling in Section 4.2.2. ( h, ỹ) is the labelanchored mixup in Section 4.3.",
"figure_data": "𝜎 𝑝 ( h, ỹ)AllMany-shot Med.-shot Few-shotw/o G unlbl 0.477(0.014) 0.378(0.030) 0.440(0.011) 0.600(0.006)Mol-Lipo✓ ✗ ✗ ✓ ✗ 0.446(0.008) 0.356(0.003) 0.407(0.011) 0.564(0.016) ✗ 0.448(0.006) 0.371(0.004) 0.421(0.012) 0.543(0.016) ✓ ✓ ✗ 0.442(0.012) 0.372(0.007) 0.415(0.004) 0.533(0.026)✗ ✗ ✓ 0.456(0.007) 0.372(0.014) 0.436(0.010) 0.549(0.005)✓ ✓ ✓ 0.432(0.012) 0.357(0.019) 0.413(0.017) 0.515(0.020)Mol-ESOLw/o G unlbl 0.477(0.027) 0.375(0.014) 0.432(0.042) 0.637(0.042) ✓ ✗ ✗ 0.475(0.014) 0.369(0.014) 0.446(0.017) 0.618(0.039) ✗ ✓ ✗ 0.480(0.017) 0.380(0.035) 0.440(0.017) 0.630(0.020) ✓ ✓ ✗ 0.468(0.007) 0.379(0.012) 0.425(0.013) 0.612(0.028)✗ ✗ ✓ 0.474(0.010) 0.353(0.018) 0.450(0.009) 0.623(0.027)✓ ✓ ✓ 0.457(0.015) 0.370(0.022) 0.411(0.011) 0.604(0.024)Mol-FreeSolvw/o G unlbl 0.619(0.019) 0.525(0.022) 0.590(0.035) 1.000(0.072)",
"figure_id": "tab_5",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Nine options on the",
"figure_data": "z 𝑖h 𝑗Mol-LipoPlym-Oxygenimplementation of the label-Additional SourceAllMany-shot Med.-shot Few-shotAllMany-shot Med.-shot Few-shotanchored mixup in Eq. (5). Ex-cept for the imbalanced la-NoneNone0.439(0.004) 0.361(0.010) 0.419(0.013) 0.529(0.022) 165.5(12.2)4.7(1.7)16.5(7.2)417.4(31.1)beled graphs G imb , the addi-NoneG conf0.447(0.015) 0.359(0.004) 0.423(0.016) 0.549(0.033) 158.1(17.0)4.1(0.7)11.3(0.7)401.9(45.1)tional source of the intervalNoneG unlbl0.432(0.012) 0.357(0.019) 0.413(0.017) 0.515(0.020) 150.9(17.8)3.8(1.1)12.2(0.6)382.8(46.9)representation z 𝑖 and the realG confNone0.448(0.012) 0.367(0.008) 0.423(0.008) 0.544(0.028) 166.0(18.2)11.9(11.3)12.6(0.9)414.0(52.6)graph representation h 𝑗 could be G conf or G unlbl . We exten-sively explore the options forG conf G confG conf G unlbl0.445(0.007) 0.364(0.008) 0.418(0.010) 0.542(0.012) 158.8(8.4) 0.449(0.021) 0.360(0.023) 0.416(0.016) 0.560(0.039) 169.5(56.1)7.7(8.9) 4.5(1.2)15.4(7.8) 12.7(1.8)397.5(15.4) 430.4(145.0)H aug and find that source z 𝑖G unlblNone0.446(0.007) 0.367(0.009) 0.415(0.011) 0.546(0.011) 173.1(30.3)3.7(0.4)13.5(1.4)440.0(79.3)from G imb and source h 𝑗 fromG unlblG conf0.446(0.011) 0.368(0.011) 0.421(0.012) 0.539(0.024) 174.5(9.3)8.1(3.3)11.9(0.9)440.4(25.5)G unlbl are usually the best.G unlblG unlbl0.451(0.007) 0.371(0.012) 0.425(0.008) 0.547(0.015) 156.3(20.5)8.2(2.9)12.9(0.9)392.3(50.6)",
"figure_id": "tab_7",
"figure_label": "6",
"figure_type": "table"
},
{
"figure_caption": "Table 4 present how SGIR improves the initial supervised performance to the most advanced semi-supervised performance step by step. In the first line for each dataset, we use only imbalanced training data G conf to train the regression model and observe that the model performs badly in the few-shot region. The fourth line for each dataset combines the use of regression confidence 𝜎 and the reverse sampling 𝑝 to produce G conf . It improves the MAE performance in the few-shot region relatively by +11.2%, +3.2%, and +15.9% on the Mol-Lipo, Mol-ESOL, and Mol-FreeSolv datasets, respectively. The label-anchored mixup algorithm produces the augmented graph representations H aug for the under-represented label ranges. By applying H aug with G conf , the last line continues improving the MAE performance in the fewshot region (compared to the third line) relatively by +3.3%, +1.3%, and +6.5% on the Mol-Lipo, Mol-ESOL, and Mol-FreeSolv datasets, respectively. Because the use of H aug provides a chance to lead the label distributions of training data closer to a perfect balance. Specifically, the effect of semi-supervised pseudo-labeling, or G conf , comes from the regression confidence 𝜎 and reverse sampling rate 𝑝. Results on Mol-ESOL and Mol-FreeSolv show that without the confidence 𝜎 (the second line), reverse sampling was useless due to heavy label noise. Results on all molecule datasets indicate that without the reverse sampling rate 𝑝 (the first line), the improvement to few-shot region by pseudo-labels was limited. 5.3.2 Effect of iterative self-training.Figure 4 confirms that model learning and balanced training data mutually enhance each other in SGIR. Because we find that the model performance gradually approximates and outperforms the best baseline in the entire label range, as well as the few-shot region, after multiple iterations. It also indicates that the quality of the training data is steadily improved over iterations. Especially for the under-represented label ranges.",
"figure_data": "",
"figure_id": "tab_8",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Comparing SGIR with related methods on research problem settings.",
"figure_data": "Is Semi-supervisedLearningAddressingSolvingmethod?",
"figure_id": "tab_9",
"figure_label": "7",
"figure_type": "table"
},
{
"figure_caption": "Complete results of ablation study and mixup options (MAE ↓ and GM ↓) on three molecule datasets. The best mean is bolded. For the label-anchored mixup options, the first column is the source of z 𝑖 and the second column is the source of h 𝑗 . G imb ∪ G conf 0.575(0.017) 0.551(0.018) 0.516(0.034) 0.863(0.071) 0.282(0.014) 0.298(0.015) 0.249(0.012) 0.389(0.058) G imb ∪ G unlbl 0.563(0.026) 0.535(0.038) 0.528(0.046) 0.777(0.061) 0.264(0.029) 0.286(0.013) 0.244(0.046) 0.304(0.078) G imb ∪ G conf G imb 0.568(0.032) 0.535(0.038) 0.513(0.036) 0.867(0.083) 0.267(0.019) 0.285(0.020) 0.235(0.026) 0.357(0.035) G imb ∪ G conf 0.577(0.021) 0.537(0.052) 0.522(0.012) 0.896(0.020) 0.280(0.018) 0.301(0.040) 0.246(0.018) 0.374(0.048) G imb ∪ G unlbl 0.565(0.027) 0.518(0.034) 0.522(0.034) 0.864(0.110) 0.262(0.024) 0.255(0.026) 0.247(0.022) 0.360(0.086) G imb ∪ G conf 0.598(0.042) 0.552(0.029) 0.545(0.040) 0.924(0.097) 0.311(0.040) 0.300(0.051) 0.295(0.040) 0.428(0.067) G imb ∪ G unlbl 0.559(0.023) 0.518(0.023) 0.503(0.016) 0.882(0.081) 0.266(0.017) 0.278(0.029) 0.229(0.010) 0.410(0.047)",
"figure_data": "MAE ↓GM ↓AllMany-shot Med.-shot Few-shotAllMany-shot Med.-shot Few-shotMol-LipoAblation StudyG imb G imb ∪ G conf G imb ∪ G conf (w/o 𝜎) G imb ∪ G conf (w/o 𝑝) G imb ∪ H aug0.477(0.014) 0.378(0.030) 0.440(0.011) 0.600(0.006) 0.442(0.012) 0.372(0.007) 0.415(0.004) 0.533(0.026) 0.446(0.008) 0.356(0.003) 0.407(0.011) 0.564(0.016) 0.448(0.006) 0.371(0.004) 0.421(0.012) 0.543(0.016) 0.456(0.007) 0.372(0.014) 0.436(0.010) 0.549(0.005)0.288(0.008) 0.236(0.015) 0.267(0.013) 0.371(0.017) 0.267(0.013) 0.240(0.008) 0.245(0.016) 0.320(0.027) 0.272(0.006) 0.222(0.002) 0.244(0.008) 0.363(0.013) 0.270(0.002) 0.228(0.009) 0.255(0.008) 0.333(0.015) 0.278(0.013) 0.235(0.019) 0.265(0.014) 0.338(0.006)G imb ∪ G conf ∪ H aug0.432(0.012) 0.357(0.019) 0.413(0.017) 0.515(0.020)0.264(0.013) 0.224(0.016) 0.256(0.017) 0.314(0.015)𝑖 𝑗 options in Mixup and h zG imb G imb ∪ G conf G imb ∪ G unlblG imb G imb ∪ G conf 0.447(0.015) 0.359(0.004) 0.423(0.016) 0.549(0.033) 0.439(0.004) 0.361(0.010) 0.419(0.013) 0.529(0.022) G imb ∪ G unlbl 0.432(0.012) 0.357(0.019) 0.413(0.017) 0.515(0.020) G imb 0.448(0.012) 0.367(0.008) 0.423(0.008) 0.544(0.028) G imb ∪ G conf 0.445(0.007) 0.364(0.008) 0.418(0.010) 0.542(0.012) G imb ∪ G unlbl 0.449(0.021) 0.360(0.023) 0.416(0.016) 0.560(0.039) G imb 0.446(0.007) 0.367(0.009) 0.415(0.011) 0.546(0.011) G imb ∪ G conf 0.446(0.011) 0.368(0.011) 0.421(0.012) 0.539(0.024) G imb ∪ G unlbl 0.451(0.007) 0.371(0.012) 0.425(0.008) 0.547(0.015)0.267(0.005) 0.231(0.015) 0.256(0.010) 0.318(0.020) 0.274(0.017) 0.221(0.007) 0.264(0.020) 0.344(0.031) 0.264(0.013) 0.224(0.016) 0.256(0.017) 0.314(0.015) 0.270(0.013) 0.230(0.013) 0.257(0.014) 0.328(0.025) 0.271(0.009) 0.227(0.011) 0.256(0.011) 0.337(0.016) 0.270(0.019) 0.223(0.017) 0.255(0.019) 0.340(0.032) 0.268(0.006) 0.228(0.008) 0.248(0.005) 0.336(0.012) 0.270(0.004) 0.233(0.010) 0.249(0.009) 0.334(0.017) 0.273(0.008) 0.222(0.007) 0.260(0.012) 0.344(0.014)Mol-ESOLAblation StudyG imb G imb ∪ G conf G imb ∪ G conf (w/o 𝜎) G imb ∪ G conf (w/o 𝑝) G imb ∪ H aug0.477(0.027) 0.375(0.014) 0.432(0.042) 0.637(0.042) 0.468(0.007) 0.379(0.012) 0.425(0.013) 0.612(0.028) 0.480(0.017) 0.380(0.035) 0.440(0.017) 0.630(0.020) 0.475(0.014) 0.369(0.014) 0.446(0.017) 0.618(0.039) 0.474(0.010) 0.353(0.018) 0.450(0.009) 0.623(0.027)0.273(0.024) 0.215(0.023) 0.248(0.043) 0.401(0.039) 0.263(0.009) 0.219(0.007) 0.236(0.017) 0.366(0.020) 0.269(0.016) 0.219(0.028) 0.249(0.024) 0.368(0.017) 0.267(0.012) 0.210(0.013) 0.251(0.017) 0.372(0.050) 0.272(0.004) 0.202(0.012) 0.257(0.011) 0.397(0.034)G imb ∪ G conf ∪ H aug0.457(0.015) 0.370(0.022) 0.411(0.011) 0.604(0.024)0.263(0.016) 0.226(0.021) 0.240(0.015) 0.347(0.030)𝑖 𝑗 options in Mixup and h zG imb G imb ∪ G conf G imb ∪ G unlblG imb G imb ∪ G conf 0.460(0.016) 0.368(0.026) 0.420(0.018) 0.605(0.026) 0.466(0.009) 0.374(0.023) 0.430(0.010) 0.604(0.032) G imb ∪ G unlbl 0.457(0.015) 0.370(0.022) 0.411(0.011) 0.604(0.024) G imb 0.469(0.017) 0.369(0.025) 0.432(0.020) 0.615(0.037) G imb ∪ G conf 0.466(0.003) 0.376(0.014) 0.425(0.011) 0.610(0.013) G imb ∪ G unlbl 0.461(0.010) 0.366(0.025) 0.424(0.020) 0.604(0.026) G imb 0.472(0.009) 0.369(0.022) 0.435(0.012) 0.623(0.025) G imb ∪ G conf 0.476(0.013) 0.387(0.027) 0.426(0.013) 0.630(0.042) G imb ∪ G unlbl 0.479(0.026) 0.368(0.012) 0.448(0.033) 0.629(0.047)0.266(0.010) 0.214(0.027) 0.242(0.018) 0.379(0.016) 0.268(0.017) 0.215(0.023) 0.252(0.022) 0.362(0.016) 0.263(0.016) 0.226(0.021) 0.240(0.015) 0.347(0.030) 0.260(0.014) 0.204(0.028) 0.248(0.013) 0.358(0.048) 0.261(0.004) 0.204(0.005) 0.242(0.013) 0.370(0.013) 0.264(0.015) 0.219(0.027) 0.244(0.017) 0.354(0.036) 0.266(0.005) 0.202(0.015) 0.257(0.012) 0.366(0.016) 0.271(0.017) 0.211(0.018) 0.253(0.022) 0.382(0.040) 0.269(0.016) 0.210(0.010) 0.253(0.023) 0.373(0.033)Mol-FreeSolvAblation StudyG imb G imb ∪ G conf G imb ∪ G conf (w/o 𝜎) G imb ∪ G conf (w/o 𝑝) G imb ∪ H aug0.619(0.019) 0.525(0.022) 0.590(0.035) 1.000(0.072) 0.568(0.029) 0.538(0.020) 0.520(0.045) 0.831(0.132) 0.660(0.028) 0.574(0.015) 0.650(0.036) 0.941(0.066) 0.604(0.020) 0.557(0.037) 0.560(0.029) 0.903(0.055) 0.593(0.045) 0.536(0.033) 0.542(0.067) 0.947(0.062)0.325(0.040) 0.289(0.006) 0.316(0.062) 0.521(0.084) 0.288(0.031) 0.295(0.037) 0.270(0.037) 0.365(0.088) 0.325(0.016) 0.302(0.007) 0.319(0.029) 0.437(0.056) 0.293(0.024) 0.307(0.050) 0.260(0.018) 0.416(0.080) 0.269(0.022) 0.259(0.037) 0.253(0.050) 0.409(0.033)G imb ∪ G conf ∪ H aug0.563(0.026) 0.535(0.038) 0.528(0.046) 0.777(0.061)0.264(0.029) 0.286(0.013) 0.244(0.046) 0.304(0.078)options in MixupG imbG imb0.572(0.006) 0.528(0.030) 0.531(0.017) 0.852(0.090)0.289(0.013) 0.299(0.026) 0.265(0.019) 0.370(0.079)𝑗and hz 𝑖",
"figure_id": "tab_10",
"figure_label": "8",
"figure_type": "table"
},
{
"figure_caption": "Investigating the effect of regression confidence measurements (MAE ↓ and GM ↓). The best mean is bolded.",
"figure_data": "MAE ↓GM ↓AllMany-shot Med.-shot Few-shotAllMany-shot Med.-shot Few-shotSimple0.481(0.010) 0.389(0.007) 0.440(0.013) 0.603(0.023)0.297(0.014) 0.239(0.006) 0.275(0.019) 0.388(0.026)Dropout 0.450(0.026) 0.365(0.031) 0.420(0.022) 0.555(0.037)0.277(0.017) 0.230(0.020) 0.263(0.011) 0.348(0.044)Mol-LipoCerti0.452(0.011) 0.384(0.018) 0.433(0.013) 0.532(0.010)0.276(0.009) 0.239(0.017) 0.267(0.015) 0.324(0.016)DER1.026(0.033) 0.604(0.035) 0.760(0.016) 1.672(0.111)0.688(0.026) 0.417(0.016) 0.528(0.015) 1.405(0.152)GRation 0.448(0.006) 0.371(0.004) 0.421(0.012) 0.543(0.016)0.270(0.002) 0.228(0.009) 0.255(0.008) 0.333(0.015)Simple0.499(0.016) 0.397(0.023) 0.457(0.018) 0.656(0.033)0.290(0.017) 0.238(0.023) 0.258(0.020) 0.415(0.025)Dropout 0.483(0.011) 0.381(0.027) 0.443(0.018) 0.636(0.027)0.279(0.017) 0.220(0.019) 0.261(0.026) 0.391(0.032)Mol-ESOLCerti0.487(0.030) 0.389(0.039) 0.439(0.024) 0.647(0.043)0.274(0.018) 0.221(0.033) 0.246(0.013) 0.396(0.025)DER0.918(0.135) 0.776(0.086) 0.826(0.098) 1.182(0.245)0.619(0.089) 0.525(0.074) 0.567(0.063) 0.829(0.180)GRation 0.475(0.014) 0.369(0.014) 0.446(0.017) 0.618(0.039)0.267(0.012) 0.210(0.013) 0.251(0.017) 0.372(0.050)Simple0.697(0.056) 0.616(0.025) 0.663(0.033) 1.054(0.260)0.327(0.036) 0.319(0.028) 0.297(0.017) 0.527(0.206)Dropout 0.639(0.013) 0.578(0.060) 0.589(0.017) 1.005(0.140)0.301(0.018) 0.274(0.047) 0.299(0.038) 0.433(0.040)Mol-FreeSolvCerti0.654(0.049) 0.589(0.046) 0.611(0.053) 0.999(0.130)0.326(0.038) 0.332(0.040) 0.292(0.044) 0.485(0.095)DER1.483(0.174) 1.180(0.162) 1.450(0.188) 2.480(0.373)0.949(0.131) 0.856(0.159) 0.883(0.183) 1.828(0.386)GRation 0.604(0.020) 0.557(0.037) 0.560(0.029) 0.903(0.055)0.293(0.024) 0.307(0.050) 0.260(0.018) 0.416(0.080)SimplePlym-Melting",
"figure_id": "tab_11",
"figure_label": "10",
"figure_type": "table"
}
] | Gang Liu; Tong Zhao; Eric Inae; Tengfei Luo; Meng Jiang | [
{
"authors": "Radhakrishna Achanta; Appu Shaji; Kevin Smith; Aurelien Lucchi; Pascal Fua; Sabine Süsstrunk",
"journal": "IEEE transactions on pattern analysis and machine intelligence",
"ref_id": "b0",
"title": "SLIC superpixels compared to state-of-the-art superpixel methods",
"year": "2012"
},
{
"authors": "Alexander Amini; Wilko Schwarting; Ava Soleimany; Daniela Rus",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b1",
"title": "Deep evidential regression",
"year": "2020"
},
{
"authors": "L Peter; Shahar Bartlett; Mendelson",
"journal": "Journal of Machine Learning Research",
"ref_id": "b2",
"title": "Rademacher and Gaussian complexities: Risk bounds and structural results",
"year": "2002-11"
},
{
"authors": "David Berthelot; Nicholas Carlini; Ian Goodfellow; Nicolas Papernot; Avital Oliver; Colin A Raffel",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b3",
"title": "Mixmatch: A holistic approach to semi-supervised learning",
"year": "2019"
},
{
"authors": "David Berthelot; Rebecca Roelofs; Kihyuk Sohn; Nicholas Carlini; Alex Kurakin",
"journal": "",
"ref_id": "b4",
"title": "Adamatch: A unified approach to semi-supervised learning and domain adaptation",
"year": "2022"
},
{
"authors": "Paula Branco; Luís Torgo; Rita P Ribeiro",
"journal": "PMLR",
"ref_id": "b5",
"title": "SMOGN: a pre-processing approach for imbalanced regression",
"year": "2017"
},
{
"authors": "Kaidi Cao; Colin Wei; Adrien Gaidon; Nikos Arechiga; Tengyu Ma",
"journal": "Advances in neural information processing systems",
"ref_id": "b6",
"title": "Learning imbalanced datasets with label-distribution-aware margin loss",
"year": "2019"
},
{
"authors": "Luigi Carratino; Moustapha Cissé; Rodolphe Jenatton; Jean-Philippe Vert",
"journal": "",
"ref_id": "b7",
"title": "On mixup regularization",
"year": "2020"
},
{
"authors": "Kevin W Nitesh V Chawla; Lawrence O Bowyer; Philip Hall; Kegelmeyer",
"journal": "Journal of artificial intelligence research",
"ref_id": "b8",
"title": "SMOTE: synthetic minority over-sampling technique",
"year": "2002"
},
{
"authors": "Corinna Cortes; Mehryar Mohri; Michael Riley; Afshin Rostamizadeh",
"journal": "Springer",
"ref_id": "b9",
"title": "Sample selection bias correction theory",
"year": "2008-10-13"
},
{
"authors": "Yin Cui; Menglin Jia; Tsung-Yi Lin; Yang Song; Serge Belongie",
"journal": "",
"ref_id": "b10",
"title": "Classbalanced loss based on effective number of samples",
"year": "2019"
},
{
"authors": "Yarin Gal; Zoubin Ghahramani",
"journal": "PMLR",
"ref_id": "b11",
"title": "Dropout as a bayesian approximation: Representing model uncertainty in deep learning",
"year": "2016"
},
{
"authors": "Yu Gong; Greg Mori; Frederick Tung",
"journal": "",
"ref_id": "b12",
"title": "RankSim: Ranking Similarity Regularization for Deep Imbalanced Regression",
"year": "2022"
},
{
"authors": "Rex William L Hamilton; Jure Ying; Leskovec",
"journal": "",
"ref_id": "b13",
"title": "Inductive representation learning on large graphs",
"year": "2017"
},
{
"authors": "Xiaotian Han; Zhimeng Jiang; Ninghao Liu; Xia Hu",
"journal": "",
"ref_id": "b14",
"title": "G-Mixup: Graph Data Augmentation for Graph Classification",
"year": "2022"
},
{
"authors": "Ju He; Adam Kortylewski; Shaokang Yang; Shuai Liu; Cheng Yang; Changhu Wang; Alan Yuille",
"journal": "",
"ref_id": "b15",
"title": "Rethinking Re-Sampling in Imbalanced Semi-Supervised Learning",
"year": "2021"
},
{
"authors": "Xinting Hu; Yulei Niu; Chunyan Miao; Xian-Sheng Hua; Hanwang Zhang",
"journal": "",
"ref_id": "b16",
"title": "On Non-Random Missing Labels in Semi-Supervised Learning",
"year": "2022"
},
{
"authors": "Neal Jean; Sang Michael Xie; Stefano Ermon",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b17",
"title": "Semi-supervised deep kernel learning: Regression with unlabeled data by minimizing predictive variance",
"year": "2018"
},
{
"authors": "Karthik Sham M Kakade; Ambuj Sridharan; Tewari",
"journal": "Advances in neural information processing systems",
"ref_id": "b18",
"title": "On the complexity of linear prediction: Risk bounds, margin bounds, and regularization",
"year": "2008"
},
{
"authors": "Jaehyung Kim; Youngbum Hur; Sejun Park; Eunho Yang; Sung Ju Hwang; Jinwoo Shin",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b19",
"title": "Distribution aligning refinery of pseudo-label for imbalanced semi-supervised learning",
"year": "2020"
},
{
"authors": "N Thomas; Max Kipf; Welling",
"journal": "",
"ref_id": "b20",
"title": "Semi-supervised classification with graph convolutional networks",
"year": "2017"
},
{
"authors": "Boris Knyazev; Graham W Taylor; Mohamed Amer",
"journal": "Advances in neural information processing systems",
"ref_id": "b21",
"title": "Understanding attention and generalization in graph neural networks",
"year": "2019"
},
{
"authors": "Tsung-Yi Lin; Priya Goyal; Ross Girshick; Kaiming He; Piotr Dollár",
"journal": "",
"ref_id": "b22",
"title": "Focal loss for dense object detection",
"year": "2017"
},
{
"authors": "Gang Liu; Eric Inae; Tong Zhao; Jiaxin Xu; Tengfei Luo; Meng Jiang",
"journal": "",
"ref_id": "b23",
"title": "Data-Centric Learning from Unlabeled Graphs with Diffusion Model",
"year": "2023"
},
{
"authors": "Gang Liu; Tong Zhao; Jiaxin Xu; Tengfei Luo; Meng Jiang",
"journal": "",
"ref_id": "b24",
"title": "Graph Rationalization with Environment-based Augmentations",
"year": "2022"
},
{
"authors": "Zixuan Liu; Ziqiao Wang; Hongyu Guo; Yongyi Mao",
"journal": "",
"ref_id": "b25",
"title": "Over-Training with Mixup May Hurt Generalization",
"year": "2023"
},
{
"authors": "Ruimin Ma; Tengfei Luo",
"journal": "Journal of Chemical Information and Modeling",
"ref_id": "b26",
"title": "PI1M: a benchmark database for polymer informatics",
"year": "2020"
},
{
"authors": "Geoffrey J Mclachlan",
"journal": "J. Amer. Statist. Assoc",
"ref_id": "b27",
"title": "Iterative reclassification procedure for constructing an asymptotically optimal rule of allocation in discriminant analysis",
"year": "1975"
},
{
"authors": "David Mendez; Anna Gaulton; Patrícia Bento; Jon Chambers; Marleen De Veij; Eloy Félix; Paula María; Juan F Magariños; Prudence Mosquera; Michał Mutowo; Nowotka",
"journal": "Nucleic acids research",
"ref_id": "b28",
"title": "ChEMBL: towards direct deposition of bioassay data",
"year": "2019"
},
{
"authors": "Aditya Krishna Menon; Sadeep Jayasumana; Ankit Singh Rawat; Himanshu Jain; Andreas Veit; Sanjiv Kumar",
"journal": "",
"ref_id": "b29",
"title": "Long-tail learning via logit adjustment",
"year": "2021"
},
{
"authors": "Stylianos Moschoglou; Athanasios Papaioannou; Christos Sagonas; Jiankang Deng; Irene Kotsia; Stefanos Zafeiriou",
"journal": "",
"ref_id": "b30",
"title": "Agedb: the first manually collected, in-the-wild age database",
"year": "2017"
},
{
"authors": "Youngtaek Oh; Dong-Jin Kim; In So Kweon",
"journal": "",
"ref_id": "b31",
"title": "Distribution-aware semantics-oriented pseudo-label for imbalanced semi-supervised learning",
"year": "2022"
},
{
"authors": "Shingo Otsuka; Isao Kuwajima; Junko Hosoya; Yibin Xu; Masayoshi Yamazaki",
"journal": "IEEE",
"ref_id": "b32",
"title": "PoLyInfo: Polymer database for polymeric materials design",
"year": "2011"
},
{
"authors": "Raghunathan Ramakrishnan; O Pavlo; Matthias Dral; O Rupp; Von Anatole; Lilienfeld",
"journal": "Scientific data",
"ref_id": "b33",
"title": "Quantum chemistry structures and properties of 134 kilo molecules",
"year": "2014"
},
{
"authors": "Mingyuan Jiawei Ren; Cunjun Zhang; Ziwei Yu; Liu",
"journal": "",
"ref_id": "b34",
"title": "Balanced MSE for Imbalanced Visual Regression",
"year": "2022"
},
{
"authors": "Yu Rong; Wenbing Huang; Tingyang Xu; Junzhou Huang",
"journal": "",
"ref_id": "b35",
"title": "DropEdge: Towards Deep Graph Convolutional Networks on Node Classification",
"year": "2019"
},
{
"authors": "Kihyuk Sohn; David Berthelot; Nicholas Carlini; Zizhao Zhang; Han Zhang; Colin A Raffel; Ekin Dogus Cubuk; Alexey Kurakin; Chun-Liang Li",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b36",
"title": "Fixmatch: Simplifying semi-supervised learning with consistency and confidence",
"year": "2020"
},
{
"authors": "Masashi Sugiyama; Amos J Storkey",
"journal": "Advances in neural information processing systems",
"ref_id": "b37",
"title": "Mixture regression for covariate shift",
"year": "2006"
},
{
"authors": "Baochen Sun; Jiashi Feng; Kate Saenko",
"journal": "",
"ref_id": "b38",
"title": "Return of frustratingly easy domain adaptation",
"year": "2016"
},
{
"authors": "Fan-Yun Sun; Jordon Hoffman; Vikas Verma; Jian Tang",
"journal": "",
"ref_id": "b39",
"title": "InfoGraph: Unsupervised and Semi-supervised Graph-Level Representation Learning via Mutual Information Maximization",
"year": "2020"
},
{
"authors": "Natasa Tagasovska; David Lopez-Paz",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b40",
"title": "Single-model uncertainties for deep learning",
"year": "2019"
},
{
"authors": "L Thornton; B Robeson; Freeman; Uhlmann",
"journal": "",
"ref_id": "b41",
"title": "Polymer Gas Separation Membrane Database",
"year": "2012"
},
{
"authors": "Junjiao Tian; Yen-Cheng Liu; Nathaniel Glaser; Yen-Chang Hsu; Zsolt Kira",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b42",
"title": "Posterior re-calibration for imbalanced datasets",
"year": "2020"
},
{
"authors": "Petar Veličković; Guillem Cucurull; Arantxa Casanova; Adriana Romero; Pietro Liò; Yoshua Bengio",
"journal": "",
"ref_id": "b43",
"title": "Graph Attention Networks",
"year": "2018"
},
{
"authors": "Vikas Verma; Alex Lamb; Christopher Beckham; Amir Najafi; Ioannis Mitliagkas; David Lopez-Paz; Yoshua Bengio",
"journal": "PMLR",
"ref_id": "b44",
"title": "Manifold mixup: Better representations by interpolating hidden states",
"year": "2019"
},
{
"authors": "Yiwei Wang; Wei Wang; Yuxuan Liang; Yujun Cai; Bryan Hooi",
"journal": "",
"ref_id": "b45",
"title": "Mixup for node and graph classification",
"year": "2021"
},
{
"authors": "Chen Wei; Kihyuk Sohn; Clayton Mellina; Alan Yuille; Fan Yang",
"journal": "",
"ref_id": "b46",
"title": "CREST: A class-rebalancing self-training framework for imbalanced semi-supervised learning",
"year": "2021"
},
{
"authors": "Ying-Xin Wu; Xiang Wang; An Zhang; Xiangnan He; Tat Seng; Chua ",
"journal": "",
"ref_id": "b47",
"title": "Discovering Invariant Rationales for Graph Neural Networks",
"year": "2022"
},
{
"authors": "Zhenqin Wu; Bharath Ramsundar; Evan N Feinberg; Joseph Gomes; Caleb Geniesse; S Aneesh; Karl Pappu; Vijay Leswing; Pande",
"journal": "Chemical science",
"ref_id": "b48",
"title": "MoleculeNet: a benchmark for molecular machine learning",
"year": "2018"
},
{
"authors": "Qizhe Xie; Minh-Thang Luong; Eduard Hovy; Quoc V Le",
"journal": "",
"ref_id": "b49",
"title": "Selftraining with noisy student improves imagenet classification",
"year": "2020"
},
{
"authors": "Keyulu Xu; Weihua Hu; Jure Leskovec; Stefanie Jegelka",
"journal": "",
"ref_id": "b50",
"title": "How Powerful are Graph Neural Networks?",
"year": "2019"
},
{
"authors": "Yuzhe Yang; Kaiwen Zha; Yingcong Chen; Hao Wang; Dina Katabi",
"journal": "PMLR",
"ref_id": "b51",
"title": "Delving into deep imbalanced regression",
"year": "2021"
},
{
"authors": "Huaxiu Yao; Yiping Wang; Linjun Zhang; James Y Zou; Chelsea Finn",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b52",
"title": "C-mixup: Improving generalization in regression",
"year": "2022"
},
{
"authors": "Qi Yuan; Mariagiulia Longo; Aaron W Thornton; Neil B Mckeown; Bibiana Comesana-Gandara; Johannes C Jansen; Kim E Jelfs",
"journal": "Journal of Membrane Science",
"ref_id": "b53",
"title": "Imputation of missing gas permeability data for polymer membranes using machine learning",
"year": "2021"
},
{
"authors": "Linjun Zhang; Zhun Deng; Kenji Kawaguchi; Amirata Ghorbani; James Zou",
"journal": "",
"ref_id": "b54",
"title": "How Does Mixup Help With Robustness and Generalization?",
"year": "2021"
},
{
"authors": "Yifan Zhang; Bingyi Kang; Bryan Hooi; Shuicheng Yan; Jiashi Feng",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b55",
"title": "Deep long-tailed learning: A survey",
"year": "2023"
},
{
"authors": "Tong Zhao; Tianwen Jiang; Neil Shah; Meng Jiang",
"journal": "IEEE Transactions on Neural Networks and Learning Systems",
"ref_id": "b56",
"title": "A synergistic approach for graph anomaly detection with pattern mining and feature learning",
"year": "2021"
},
{
"authors": "Tong Zhao; Wei Jin; Yozen Liu; Yingheng Wang; Gang Liu; Stephan Günneman; Neil Shah; Meng Jiang",
"journal": "",
"ref_id": "b57",
"title": "Graph Data Augmentation for Graph Machine Learning: A Survey",
"year": "2022"
},
{
"authors": "Tong Zhao; Gang Liu; Daheng Wang; Wenhao Yu; Meng Jiang",
"journal": "PMLR",
"ref_id": "b58",
"title": "Learning from Counterfactual Links for Link Prediction",
"year": "2022"
},
{
"authors": "Tong Zhao; Yozen Liu; Leonardo Neves; Oliver Woodford; Meng Jiang; Neil Shah",
"journal": "",
"ref_id": "b59",
"title": "Data Augmentation for Graph Neural Networks",
"year": "2021"
},
{
"authors": "",
"journal": "",
"ref_id": "b60",
"title": "Complete results of ablation study and mixup options (MAE ↓ and GM ↓) on three polymer datasets. The best mean is bolded. For the label-anchored mixup options, the first column is the source of z 𝑖 and the second column is",
"year": ""
},
{
"authors": "",
"journal": "G imb ∪ G conf",
"ref_id": "b61",
"title": "",
"year": ""
},
{
"authors": "",
"journal": "G imb ∪ G conf (w/o 𝜎)",
"ref_id": "b62",
"title": "",
"year": ""
},
{
"authors": "",
"journal": "G imb ∪ G conf (w/o 𝑝)",
"ref_id": "b63",
"title": "",
"year": ""
},
{
"authors": "",
"journal": "G imb ∪ H aug",
"ref_id": "b64",
"title": "",
"year": ""
},
{
"authors": " Imb",
"journal": "∪ G conf ∪ H aug",
"ref_id": "b65",
"title": "",
"year": ""
},
{
"authors": "",
"journal": "G imb ∪ G conf",
"ref_id": "b66",
"title": "",
"year": ""
},
{
"authors": "",
"journal": "G imb ∪ G unlbl",
"ref_id": "b67",
"title": "",
"year": ""
},
{
"authors": "",
"journal": "G imb ∪ G conf G imb",
"ref_id": "b68",
"title": "",
"year": ""
},
{
"authors": "",
"journal": "G imb ∪ G conf",
"ref_id": "b69",
"title": "",
"year": ""
},
{
"authors": "",
"journal": "G imb ∪ G unlbl",
"ref_id": "b70",
"title": "",
"year": ""
},
{
"authors": "",
"journal": "G imb ∪ G unlbl G imb",
"ref_id": "b71",
"title": "",
"year": ""
},
{
"authors": "",
"journal": "G imb ∪ G conf",
"ref_id": "b72",
"title": "",
"year": ""
},
{
"authors": "",
"journal": "G imb ∪ G unlbl",
"ref_id": "b73",
"title": "",
"year": ""
},
{
"authors": "",
"journal": "G imb ∪ G conf",
"ref_id": "b74",
"title": "",
"year": ""
},
{
"authors": "",
"journal": "G imb ∪ G conf (w/o 𝜎)",
"ref_id": "b75",
"title": "",
"year": ""
},
{
"authors": "",
"journal": "G imb ∪ G conf (w/o 𝑝)",
"ref_id": "b76",
"title": "",
"year": ""
},
{
"authors": "",
"journal": "G imb ∪ H aug",
"ref_id": "b77",
"title": "",
"year": ""
},
{
"authors": " Imb",
"journal": "∪ G conf ∪ H aug",
"ref_id": "b78",
"title": "",
"year": ""
}
] | [
{
"formula_coordinates": [
2,
177.68,
256.78,
17.38,
8.67
],
"formula_id": "formula_0",
"formula_text": "/(1)"
},
{
"formula_coordinates": [
3,
54.34,
576.75,
113.92,
9.38
],
"formula_id": "formula_1",
"formula_text": "[𝑏 0 , 𝑏 1 ), [𝑏 1 , 𝑏 2 ), . . . , [𝑏 𝐶-1 , 𝑏 𝐶 )."
},
{
"formula_coordinates": [
4,
107.64,
455.56,
186.41,
25.64
],
"formula_id": "formula_2",
"formula_text": "𝜎 𝑖 = 1 Var {𝑓 (𝑔(𝐺 (𝑖,𝑗) ))} 𝑗=1,2,...,𝐵 .(1)"
},
{
"formula_coordinates": [
4,
142.53,
534.12,
37.74,
13.12
],
"formula_id": "formula_3",
"formula_text": "(𝑟 ) 𝑖 of 𝐺 (𝑟 )"
},
{
"formula_coordinates": [
4,
204.85,
534.12,
37.12,
13.12
],
"formula_id": "formula_4",
"formula_text": "(𝑒) 𝑗 of 𝐺 (𝑒)"
},
{
"formula_coordinates": [
4,
103.71,
555.76,
190.34,
26.96
],
"formula_id": "formula_5",
"formula_text": "𝜎 𝑖 = 1 Var {𝑓 (h (𝑟 ) 𝑖 + h (𝑒) 𝑗 )} 𝑗=1,2,...,𝐵 .(2)"
},
{
"formula_coordinates": [
4,
390.86,
238.33,
164.17,
24.9
],
"formula_id": "formula_6",
"formula_text": "𝑝 𝑖 = 𝜇 ′ 𝑖 max{𝜇 1 , 𝜇 2 , . . . , 𝜇 𝐶 } . (3"
},
{
"formula_coordinates": [
4,
555.03,
248.18,
3.17,
7.94
],
"formula_id": "formula_7",
"formula_text": ")"
},
{
"formula_coordinates": [
4,
405.46,
553.5,
152.74,
8.43
],
"formula_id": "formula_8",
"formula_text": "Z = norm(M) • H,(4)"
},
{
"formula_coordinates": [
5,
125.35,
359.02,
168.69,
23.55
],
"formula_id": "formula_9",
"formula_text": "h(𝑖,𝑗) = 𝜆 • z 𝑖 + 1 -𝜆 • h 𝑗 , ỹ(𝑖,𝑗) = 𝜆 • 𝑎 𝑖 + 1 -𝜆 • 𝑦 𝑗 ,(5)"
},
{
"formula_coordinates": [
5,
132.7,
515.08,
78.16,
9.87
],
"formula_id": "formula_10",
"formula_text": "G imb ∪ G conf ∪ H aug ."
},
{
"formula_coordinates": [
5,
53.8,
537,
255.42,
57.21
],
"formula_id": "formula_11",
"formula_text": "(𝐺, 𝑦) ∈ G imb ∪ G conf , the loss is ℓ imb+conf = MAE(𝑓 (𝑔(𝐺)), 𝑦). Given (h, 𝑦) ∈ H aug , the loss is ℓ aug = MAE(𝑓 (h), 𝑦). So the total loss for SGIR is L = ∑︁ (𝐺,𝑦) ∈ G imb ∪G conf ℓ imb+conf (𝐺, 𝑦) + ∑︁ (h,𝑦) ∈H aug ℓ aug (h, 𝑦)."
},
{
"formula_coordinates": [
5,
324.17,
270.56,
234.03,
24.77
],
"formula_id": "formula_12",
"formula_text": "E bal [𝑓 ] = Pr (𝐺, [𝑏 𝑖 ,𝑏 𝑖+1 ))∼P bal 𝑆 [𝑏 𝑖 ,𝑏 𝑖+1 ) (𝐺) < max 𝑗≠𝑖 𝑆 [𝑏 𝑗 ,𝑏 𝑗 +1 ) (𝐺) ,(6)"
},
{
"formula_coordinates": [
5,
317.62,
340.31,
240.58,
36.83
],
"formula_id": "formula_13",
"formula_text": "E [𝑏 𝑖 ,𝑏 𝑖+1 ) [𝑓 ] = Pr 𝐺∼P [𝑏 𝑖 ,𝑏 𝑖+1 ) 𝑆 [𝑏 𝑖 ,𝑏 𝑖+1 ) (𝐺) < max 𝑗≠𝑖 𝑆 [𝑏 𝑗 ,𝑏 𝑗 +1 ) (𝐺) ,(7) where P"
},
{
"formula_coordinates": [
5,
317.53,
368.72,
242.06,
21.2
],
"formula_id": "formula_14",
"formula_text": "[𝑏 𝑖 ,𝑏 𝑖+1 ) denotes the distribution for the interval [𝑏 𝑖 , 𝑏 𝑖+1 ). We define 𝛾 (𝐺, [𝑏 𝑖 , 𝑏 𝑖+1 )) = 𝑆 [𝑏 𝑖 ,𝑏 𝑖+1 ) (𝐺) -max 𝑗≠𝑖 𝑆 [𝑏 𝑗 ,𝑏 𝑗 +1 ) (𝐺)"
},
{
"formula_coordinates": [
5,
366.6,
436.12,
191.6,
14.16
],
"formula_id": "formula_15",
"formula_text": "𝛾 [𝑏 𝑖 ,𝑏 𝑖+1 ) = min 𝐺 𝑗 ∈ [𝑏 𝑖 ,𝑏 𝑖+1 ) 𝛾 𝐺 𝑗 , [𝑏 𝑖 , 𝑏 𝑖+1 ) .(8)"
},
{
"formula_coordinates": [
5,
348.77,
569.18,
73.76,
15.06
],
"formula_id": "formula_16",
"formula_text": "E [𝑏 𝑖 ,𝑏 𝑖+1 ) [𝑓 ] ⪅1"
},
{
"formula_coordinates": [
5,
395.79,
565.2,
162.41,
57.46
],
"formula_id": "formula_17",
"formula_text": "𝛾 [𝑏 𝑖 ,𝑏 𝑖+1 ) √︄ C(F ) 𝑛 [𝑏 𝑖 ,𝑏 𝑖+1 ) + √︄ log log 2 (1/𝛾 [𝑏 𝑖 ,𝑏 𝑖+1 ) ) + log(1/𝛿) 𝑛 [𝑏 𝑖 ,𝑏 𝑖+1 ) ,(9)"
},
{
"formula_coordinates": [
12,
136.89,
134.37,
338.21,
150.97
],
"formula_id": "formula_18",
"formula_text": ") (Supervised) (Non-graph) (Balance) (Classification) DARP [20] ✓ ✓ DASO [32] ✓ ✓ Bi-Sampling [16] ✓ ✓ CADR [17] ✓ ✓ CReST [47] ✓ ✓ LDS [52] ✓ ✓ BMSE [35] ✓ ✓ RankSim [13] ✓ ✓ SSDKL [18] ✓ ✓ InfoGraph [40] ✓ ✓ ✓ SGIR (Ours) ✓ ✓ ✓ ✓"
},
{
"formula_coordinates": [
12,
359.72,
325.05,
195.07,
23.26
],
"formula_id": "formula_19",
"formula_text": "E [𝑓 ] ≤ Ê [𝑓 ] + 2𝐿 𝑒 R 𝑛 (F ) + 𝑐 0 √︂ log(1/𝛿) 2𝑛 , (10"
},
{
"formula_coordinates": [
12,
554.78,
334.21,
3.42,
7.94
],
"formula_id": "formula_20",
"formula_text": ")"
},
{
"formula_coordinates": [
12,
322.78,
451.7,
235.42,
54.89
],
"formula_id": "formula_21",
"formula_text": "E [𝑓 ] ≤ 𝐾 𝛾 [𝑓 ] + 4 R 𝑛 (F ) 𝛾 + √︄ 2 log log 2 (4𝑐 1 /𝛾) + log(1/𝛿) 2𝑛 ,(11)"
},
{
"formula_coordinates": [
12,
344.12,
516.16,
214.08,
27.02
],
"formula_id": "formula_22",
"formula_text": "≤ 𝐾 𝛾 [𝑓 ] + 4 R 𝑛 (F ) 𝛾 + log log 2 4𝑐 1 𝛾 𝑛 + √︂ log(1/𝛿) 2𝑛 .(12)"
},
{
"formula_coordinates": [
12,
318.13,
648.51,
245.91,
25.86
],
"formula_id": "formula_23",
"formula_text": "E 𝛾, [𝑏 𝑖 ,𝑏 𝑖+1 ) [𝑓 ] = Pr 𝐺∼P [𝑏 𝑖 ,𝑏 𝑖+1 ) 𝑆 [𝑏 𝑖 ,𝑏 𝑖+1 ) (𝐺) < max 𝑗≠𝑖 𝑆 [𝑏 𝑗 ,𝑏 𝑗 +1 ) (𝐺) + 𝛾 .(13)"
},
{
"formula_coordinates": [
13,
56.26,
102.53,
237.78,
11.19
],
"formula_id": "formula_24",
"formula_text": "R(𝑏 𝑖 ,𝑏 𝑖+1 ] (F ) =(14)"
},
{
"formula_coordinates": [
13,
54.64,
122.99,
251.99,
36.81
],
"formula_id": "formula_25",
"formula_text": "𝑛 (𝑏 𝑖 ,𝑏 𝑖+1 ] E 𝜎 sup 𝑓 ∈ F ∑︁ 𝐺 𝑖 ∈ [𝑏 𝑖 ,𝑏 𝑖+1 ) 𝜎 𝑖 𝑆 [𝑏 𝑖 ,𝑏 𝑖+1 ) (𝐺 𝑖 ) -max 𝑗≠𝑖 𝑆 [𝑏 𝑗 ,𝑏 𝑗 +1 ) (𝐺 𝑖 ) (15)"
},
{
"formula_coordinates": [
13,
53.98,
218.82,
240.07,
25.84
],
"formula_id": "formula_26",
"formula_text": "E [𝑏 𝑖 ,𝑏 𝑖+1 ) ≤ Ê𝛾,[𝑏 𝑖 ,𝑏 𝑖+1 ) [𝑓 ] + 4 R(𝑏 𝑖 ,𝑏 𝑖+1 ] (F ) 𝛾 [𝑏 𝑖 ,𝑏 𝑖+1 )(16)"
},
{
"formula_coordinates": [
13,
89.31,
256.02,
174.88,
55.28
],
"formula_id": "formula_27",
"formula_text": "+ 2 log log 2 ( 4𝑐 1 𝛾 [𝑏 𝑖 ,𝑏 𝑖+1 ) ) + log(1/𝛿) 2𝑛 [𝑏 𝑖 ,𝑏 𝑖+1 ) , ≤ Ê𝛾,[𝑏 𝑖 ,𝑏 𝑖+1 ) [𝑓 ] +1"
},
{
"formula_coordinates": [
13,
158.12,
292.26,
135.92,
26.21
],
"formula_id": "formula_28",
"formula_text": "𝛾 [𝑏 𝑖 ,𝑏 𝑖+1 ) √︄ C(F ) 𝑛 [𝑏 𝑖 ,𝑏 𝑖+1 )(17)"
},
{
"formula_coordinates": [
13,
89.31,
330.62,
167.34,
55.45
],
"formula_id": "formula_29",
"formula_text": "+ 2 log log 2 ( 4𝑐 1 𝛾 [𝑏 𝑖 ,𝑏 𝑖+1 ) ) log(1/𝛿) 2𝑛 [𝑏 𝑖 ,𝑏 𝑖+1 ) , ⪅1"
},
{
"formula_coordinates": [
13,
98.39,
366.86,
206.21,
37.44
],
"formula_id": "formula_30",
"formula_text": "𝛾 [𝑏 𝑖 ,𝑏 𝑖+1 ) √︄ C(F ) 𝑛 [𝑏 𝑖 ,𝑏 𝑖+1 ) + √︄ log log 2 (1/𝛾 [𝑏 𝑖 ,𝑏 𝑖+1 ) ) + log(1/𝛿) 2𝑛 [𝑏 𝑖 ,𝑏 𝑖+1 ) .(18)"
},
{
"formula_coordinates": [
13,
190.55,
426.47,
16.3,
6.84
],
"formula_id": "formula_31",
"formula_text": "C( F)"
},
{
"formula_coordinates": [
13,
192.24,
465.88,
100.79,
11.81
],
"formula_id": "formula_32",
"formula_text": "E bal [𝑓 ] = 1 𝐶 𝐶 𝑖=1 E [𝑏 𝑖 ,𝑏 𝑖+1 ) ."
},
{
"formula_coordinates": [
14,
118.46,
533.61,
172.62,
12
],
"formula_id": "formula_33",
"formula_text": "(𝑟 ) = 1 ⊤ 𝐾 • (m × H) and h (𝑒) = 1 ⊤ 𝐾 • ((1 𝐾 -m) × H)"
},
{
"formula_coordinates": [
14,
65.41,
575.4,
213.26,
55.24
],
"formula_id": "formula_34",
"formula_text": " ℓ imb+conf = MAE(𝑓 (h (𝑟 ) ), 𝑦) + E 𝐺 ′ MAE(𝑓 (h + h ′ ), 𝑦) + Var 𝐺 ′ {MAE(𝑓 (h + h ′ ), 𝑦)} , ℓ regu = 1 𝐾 𝐾 𝑘=1"
},
{
"formula_coordinates": [
14,
378.8,
157.49,
118.03,
25.92
],
"formula_id": "formula_35",
"formula_text": "𝑤 = exp( 𝐵 𝑏=1 |𝑦 -𝑦 𝑏 |/𝑡) exp( 𝐵 𝑗=1 𝐵 𝑏=1 |𝑦 -𝑦 𝑏 |/𝑡) ,"
}
] |
10.18653/v1/2020.emnlp-main.442 | 2023-10-03 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b5",
"b53",
"b13",
"b63",
"b10",
"b25",
"b26",
"b16",
"b15",
"b38"
],
"table_ref": [],
"text": "Task-oriented Dialogue (TOD) Systems aim to build dialogue systems that assist users in accomplishing specific goals, such as booking a hotel or a restaurant. Most solutions of TOD are based on domain-APIs (Budzianowski et al., 2018;Rastogi et al., 2020) and structured databases (Eric et al., 2017;Wu et al., 2019), which can only handle a limited range of scenarios within the scope of APIs/DBs. To further enlarge the model's ability of task-oriented assistance, recent works (Dimitrakis et al., 2018;Kim et al., 2020Kim et al., , 2021;;Feng et al., 2020Feng et al., , 2021;;Majumder et al., 2022) incorporate unstructured textual information retrieved from the Internet into dialogue modeling. Most of these works focus on factual knowledge sources such as frequently asked questions (FAQs) of online prod-"
},
{
"figure_ref": [],
"heading": "Avalon Hotel",
"publication_ref": [],
"table_ref": [],
"text": "While I was not pleased with the slow wi-fi and small room, I was content with their awesome breakfast options. They U: Is it fast enough to watch online videos? S: Yes, reviewers said that their WIFI is stable and fast."
},
{
"figure_ref": [],
"heading": "SK-Grounded TOD Subjective Knowledge Source",
"publication_ref": [
"b33",
"b3",
"b24",
"b29",
"b1"
],
"table_ref": [],
"text": "Figure 1: Examples of the SK-TOD task. The top part shows two hotels and their customer reviews. The bottom part shows three dialogue sessions between the system (denoted by S) and three users (denoted by U). The last user utterance is a subjective question about the WIFI quality of the hotel(s). The system needs to retrieve information from the relevant subjective knowledge, which is highlighted in the review text. ucts or government service guides. We refer to these models as Fact-TOD models.\nHowever, in many TOD tasks, users care about not only factual information but subjective insights as well, such as the experiences, opinions, and preferences of other customers. For instance, when booking a hotel or a restaurant, users often inquire about subject aspects like \"Is the WIFI reliable?\" or \"Does the restaurant have a good atmosphere?\".\nTo respond to such user requests, an agent needs to seek information from subjective knowledge sources, such as online customer reviews. While subjective knowledge has been specifically studied in other NLP problems such as opinion mining (Liu and Zhang, 2012) and question answering (Bjerva et al., 2020), incorporating it into TOD has not received significant attention.\nIn this work, we argue that it is important to enable the TOD model to leverage subjective knowledge for more effective task-oriented assistance. To this end, we propose a novel task of subjective-knowledge-based task-oriented dialogue (SK-TOD). SK-TOD focuses on responding to user requests that seek subjective information by incorporating user reviews as subjective knowledge. Figure 1 shows three examples of such requests, where customers ask about the WiFi quality of various hotels. User reviews are valuable resources for subjective information because even for the same aspect of a product or service, customers may have different opinions and leave either positive or negative reviews. As a result, a TOD system should consider multiple reviews to provide a comprehensive representation of user opinions. Ideally, the system's response should include both positive and negative opinions, along with their respective proportions (as exemplified in Dialogue 3). This two-sided response has been recognized as more credible and valuable for customers (Kamins et al., 1989;Lee et al., 2008;Baek et al., 2012), thereby fostering trust in the TOD system.\nIncorporating subjective knowledge into TOD introduces two unique challenges. Firstly, unlike in Fact-TOD where selecting a few relevant knowledge snippets suffices, the SK-TOD model must consider all relevant knowledge snippets. In other words, both precision and recall matter during this process. Secondly, the model needs to aggregate these knowledge snippets into a concise response that can faithfully reflect the diversity and proportion of opinions expressed. Conquering these challenges requires a large-scale dataset with subjective-knowledge-grounded responses, which, to our best knowledge, is not publicly available.\nTo facilitate the research in subjectiveknowledge-grounded TOD, we have collected a large-scale dataset, which contains 19,696 subjective knowledge-seeking dialogue contexts and manually annotated responses that are grounded on 143 entities and 1,430 reviews (8,013 sentences). We evaluate the performance of strong baselines on the SK-TOD task. Results show that there is a significant gap between human-generated and machine-generated responses, particularly in terms of the faithfulness of the sentiment proportion. To address this issue, we propose a model that incorporates review understanding into SK-TOD. We experimentally demonstrate that responses generated by this model more effectively capture the sentiment proportion. Our contributions are three-fold:\n• We introduce a novel task of subjectiveknowledge-based TOD (SK-TOD);\n• We create and release a large-scale, humanannotated dataset designed for this task;\n• We propose a new model and conduct extensive experiments on the proposed task.\n2 Related Work"
},
{
"figure_ref": [],
"heading": "Knowledge-Grounded Dialogue",
"publication_ref": [
"b43",
"b34",
"b44",
"b58",
"b59",
"b70",
"b11",
"b39",
"b36",
"b17",
"b27",
"b64",
"b19",
"b43",
"b18",
"b30",
"b57",
"b60",
"b13",
"b63",
"b48",
"b6",
"b56",
"b16",
"b15",
"b25",
"b26"
],
"table_ref": [],
"text": "Knowledge-grounded response generation is popular in the open-domain dialogue. Numerous external knowledge sources have been explored, from structured knowledge such as fact tables (Moghe et al., 2018;Liu et al., 2018) and knowledge graphs (Zhang et al., 2020a;Moon et al., 2019;Tuan et al., 2019), to unstructured knowledge such as Wikipedia articles (Vougiouklis et al., 2016;Zhou et al., 2018;Dinan et al., 2018), news articles (Majumder et al., 2020), web pages (Long et al., 2017;Galley et al., 2019;Komeili et al., 2022), narratives (Xu et al., 2021;Gopalakrishnan et al., 2019), user reviews and comments (Moghe et al., 2018;Ghazvininejad et al., 2018), and so on. Grounding on external knowledge makes the response more informative and meaningful when compared with models that solely rely on the dialog context. Regarding task-oriented dialogues, previous works have primarily focused on domain-specific APIs and databases to support the dialogue response (Levin et al., 2000;Singh et al., 2002;Williams and Young, 2007;Eric et al., 2017;Wu et al., 2019), which can only support a limited scope of user queries. Later works ground taskoriented dialogues to web pages (Penha et al., 2019;Chen et al., 2022), government service documents (Saeidi et al., 2018;Feng et al., 2020Feng et al., , 2021)), and FAQ knowledge snippets (Kim et al., 2020(Kim et al., , 2021)). Different from these works where factual knowledge is utilized, we apply subjective knowledge to generate the response and ground in multiple knowledge snippets. While Majumder et al. ( 2022) also explored grounding TOD in user reviews, they did not consider the diversity of opinions."
},
{
"figure_ref": [],
"heading": "Subjective Content Understanding",
"publication_ref": [
"b49",
"b22",
"b7",
"b69",
"b4",
"b0",
"b42",
"b3",
"b45"
],
"table_ref": [],
"text": "Besides being used as external knowledge sources in dialogue systems, subjective content, especially user reviews, has been studied in various nonconversational NLP tasks. For example, opinion mining (Pontiki et al., 2016;Jiang et al., 2019) focuses on extracting opinions and sentiments from user reviews. Opinion summarization (Chu and Liu, 2019;Zhao and Chaturvedi, 2020;Bražinskas et al., 2020;Angelidis et al., 2021) is used to distill multiple opinions into concise summaries. Subjective question answering (McAuley and Yang, 2016;Bjerva et al., 2020) have been proposed to answer questions based on user reviews. Explainable recommendation (Ni et al., 2019) aims to generate review-based explanations for the items recommended by a recommendation system. Table 1 provides detailed comparisons between SK-TOD and these subjective-content-based benchmarks. Generally, SK-TOD requires creating a response that is appropriate to the dialogue context. It also requires grounding in multiple subjective knowledge and explicitly considers the diversity of opinions and the proportion of sentiments."
},
{
"figure_ref": [],
"heading": "Problem Formulation",
"publication_ref": [],
"table_ref": [],
"text": "Formally, we have a\ndialogue context C = [U 1 , S 1 , U 2 , S 2 , • • • , U t ]\nbetween a user and a system, where each user utterance U i is followed by a system response utterance S i , except for the last user utterance U t . The dialogue involves one or more entities, denoted as\nE = {e 1 , • • • , e m }.\nAlongside the dialogue, we have a subjective knowledge source B = {(e 1 , R 1 ), (e 2 , R 2 ), • • • } containing all the entities and their corresponding customer reviews. Each entity e is associated with multiple\nreviews R = {R 1 , R 2 , • • • }. Each review can be divided into segments [K 1 , K 2 , • • • ],\nsuch as paragraphs, sentences, or sub-sentential units. In this work, we regard each review sentence as a knowledge snippet.\nThe SK-TOD task aims to identify whether U t is a subjective knowledge-seeking request and, if it is, to select the relevant knowledge snippets K + from the knowledge source and finally generate a response S t grounded on K + ."
},
{
"figure_ref": [],
"heading": "Data Collection and Statistics",
"publication_ref": [
"b5",
"b12"
],
"table_ref": [],
"text": "We ground the data collection in MultiWOZ (Budzianowski et al., 2018;Eric et al., 2020). We select dialogues from the domains of hotels and restaurants. The data collection is conducted by a group of crowd workers through Amazon Mechanical Turk (AMT). To control the data quality, we only choose workers that are pre-qualified. More details can be found in Appendix A."
},
{
"figure_ref": [],
"heading": "Annotation Guideline",
"publication_ref": [
"b25"
],
"table_ref": [],
"text": "Dialogues in MultiWOZ are collected based on single or multiple entities as the back-end database. To create a subjective knowledge source to support the SK-TOD task, we first collect multiple user reviews for each entity. To control the review collection, we provide the reviewer's persona, as well as the aspects and sentiments of reviews to workers. We then ask workers to write a review with all the given information included. After collecting the reviews, we also annotate the aspect and sentiment information for each review sentence. Overall, we select 33 hotels and 110 restaurants from Multi-WOZ, and collect 10 reviews for each entity. On average, each review contains 5.6 sentences and 56.71 tokens. More details about the review collection can be found in Appendix A.\nAfter obtaining the reviews, we go back to the dialogue data to create the subjective user request. Following a similar procedure in Kim et al. (2020), for each dialogue, we provide an aspect that users are interested in (e.g., WIFI-quality of the hotel) and then ask the worker to insert a subjective user request into the dialogue. Workers are requested to carefully select the insertion position and write an utterance to maintain coherence and naturalness in the dialogue flow. Finally, we use the partial dialog until this newly inserted turn as an instance in our data. Utterances that come after the insertion position are removed from the dialogue instance.\nSo far, we've collected the dialogue context C and the subjective knowledge source B. The final step is to ground the dialogue in the knowledge source. We first ask workers to identify entities that are relevant to the subjective user request as gold entities. We then align the user request and review sentences of the gold entities by matching their aspect. For example, if the aspect of a user request is about the \"WIFI quality\" of a hotel, all review sentences discussing the \"WIFI quality\" of that specific hotel will be considered relevant knowledge Size Manual Dial TOD Query Aspect Senti Mul-Knwl Senti-% Semeval/ MAMS (2016;2019) \n5K/22K ✓ ✗ n/a ✗ ✓ ✓ ✗ n/a Space (2021) 1K ✓ ✗ n/a ✗ ✓ ✓ ✓ ✗ Yelp/Amazon (2019; 2020) 200/180 ✓ ✗ n/a ✗ ✗ ✓ ✓ ✗ Justify-Rec (2019) 1.3M ✗ ✗ n/a ✗ ✓ ✗ ✓ ✗ AmazonQA (2016) 309K ✗ ✗ n/a ✓ ✗ ✗ ✗ n/a SubjQA (2020) 10K ✗ ✗ n/a ✓ ✓ ✓ ✗ n/a Holl-E (2018) 9K ✓ ✓ ✗ ✗ ✗ ✗ ✓ ✗ Foursquare (2018) 1M ✗ ✓ ✗ ✗ ✗ ✗ ✓ n/a SK-TOD (Ours) 20K ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓\nTable 1: Comparison between SK-TOD and other benchmarks based on the subjective content. We consider if the dataset is manually annotated, dialogue-based, task-oriented, and query-focused. We also list if it considers aspect and sentiment, multiple knowledge snippets (Mul-Knwl), and the proportion of two-sided sentiments (Senti-%).\nsnippets.1 Finally, we provide the dialogue context C and all related knowledge snippets K + and ask workers to generate a natural and faithful response. We explicitly instruct workers to consider the diversity and proportion of opinions in all relevant knowledge snippets during response creation. Detailed instructions can be found in Appendix A."
},
{
"figure_ref": [],
"heading": "Quality Control",
"publication_ref": [],
"table_ref": [],
"text": "To ensure the quality of our dataset, we took great care in selecting pre-qualified workers and designing annotation interfaces. We further conducted a human verification task on the entire dataset to identify invalid instances. The annotation showed that 81.89% of subjective-knowledge-seeking user turns are valid, with an Inter-Annotator Agreement (IAA) score of 0.9369 in Gwet's gamma. For agent response turns, 96.78% were valid, with an IAA score of 0.9497 in Gwet's gamma. Any invalid instances were filtered out or manually corrected before finalizing the dataset. We paid workers an average of $13.82/hr for data annotation and $14.77/hr for data verification. Both exceed the local living minimum wage. The details of our payment settings are elaborated on in Appendix A."
},
{
"figure_ref": [],
"heading": "Data Statistics",
"publication_ref": [],
"table_ref": [
"tab_0"
],
"text": "We collected a total of 19,696 instances consisting of subjective user requests and subjectiveknowledge-grounded responses. The average length of the subjective user request and the agent response is 8.75 and 24.07 tokens, respectively. While most of the instances contain a single entity, there are 1,047 instances where multiple en- tities are compared (like Dialogue 2 in Figure 1). On average, each instance requires 3.88 subjective knowledge snippets. To help identify the subjective knowledge-seeking user request, we also randomly sample another 18,383 dialogues with non-subjective user requests from the original Mul-tiWOZ dataset.\nWe split the dataset into training (75%), validation (10.8%), and test (14.2%) sets. Table 2 presents the detailed statistics of each subset. Both the validation and test sets contain two subsets: the seen subset where the aspects of these instances are included in the training set, and the unseen subset where the aspects are not included in the training set. The unseen subset is designed to evaluate models' ability to generalize to arbitrary aspects."
},
{
"figure_ref": [],
"heading": "Subjective-Knowledge-Grounded TOD",
"publication_ref": [
"b25"
],
"table_ref": [],
"text": "In this section, we describe the method for SK-TOD. As shown in Figure 2, we follow the pipeline introduced by Kim et al. (2020) which comprises four sequential sub-tasks: knowledge-seeking turn detection (KTD), entity tracking (ET), knowledge selection (KS), and response generation (RG). We elaborate on each subtask below. "
},
{
"figure_ref": [],
"heading": "Knowledge-Seeking Turn Detection",
"publication_ref": [
"b9"
],
"table_ref": [],
"text": "The goal of KTD is to identify the user request that requires subjective knowledge. We regard it as a binary classification problem, where the input is the dialogue context C and the output is a binary indicator.\nWe employ a pre-trained language model (e.g., BERT (Devlin et al., 2019)) to encode C and adopt the hidden state of the first token as its representation. Then we apply a classifier to obtain the probability that the current user request is seeking subjective knowledge. That is,\nh = Enc(C) P (C) = softmax (FFN (h)) .\n(1)\nThe model is finetuned with the binary crossentropy loss."
},
{
"figure_ref": [],
"heading": "Entity Tracking",
"publication_ref": [
"b23"
],
"table_ref": [],
"text": "The goal of ET is to identify the entities E = {e 1 , • • • , e m } that are relevant to the user request. It can help to reduce the number of candidates during the knowledge selection step.\nWe adopt a word-matching-based method used by Jin et al. (2021) to extract relevant entities. It first normalizes entity names in the knowledge source using a set of heuristic rules. Then a fuzzy n-gram matching is performed between the normalized entity and all dialogue turns. To find the entities that are relevant to the last user request, we choose the last dialogue turn in which the entities are detected and use these entities as the output of ET. We leave the tracking of aspects being questioned over multiple turns as future work."
},
{
"figure_ref": [],
"heading": "Knowledge Selection",
"publication_ref": [
"b41",
"b62",
"b54"
],
"table_ref": [],
"text": "The goal of KS is to select the knowledge snippets that are relevant to the user's request. The inputs are the dialogue context C and a set of knowledge snippets candidates K, which is a combination of all knowledge snippets of the relevant entities in E. The output K + ⊆ K is a subset of relevant knowledge candidates. Note that there might be multiple knowledge snippets in K + .\nTo select relevant knowledge snippets, we calculate the relevance score between the dialogue context C and a knowledge snippet K ∈ K. We regard it as a pairwise text scoring problem and consider two popular approaches: bi-encoder (Mazaré et al., 2018) and cross-encoder (Wolf et al., 2019). Generally, the bi-encoder approach is more efficient while the cross-encoder approach is more accurate.\nFor the bi-encoder approach, we encode C and K separately using the same pre-trained encoder and obtain two representations, h C and h K . Following Reimers and Gurevych (2019), we use the concatenation of h C , h K , and |h C -h K | as features and apply a classifier to obtain the probability of relevance. That is,\nh C = Enc(C), h K = Enc(K) P (C, K) = softmax (FFN (h c , h K , |h C -h K |)) .\n(2) For the cross-encoder approach, we encode the concatenation of C and K to obtain a contextualized representation. That is,\nh = Enc(C, K) P (C, K) = softmax (FFN (h)) .\n(3)\nDuring training, we use all relevant knowledge snippets to construct positive (C, K) pairs. Due to the large number of irrelevant knowledge snippets, we randomly sample the same number of irrelevant snippets to form negative pairs. We optimize the model using the binary cross-entropy loss. During inference, we predict the relevance probability for all knowledge snippets in the candidates. Since both precision and recall are crucial in KS, instead of selecting the top few results, we use a threshold, estimated from the validation set, to determine the relevancy of each knowledge snippet."
},
{
"figure_ref": [],
"heading": "Response Generation",
"publication_ref": [
"b31",
"b67"
],
"table_ref": [],
"text": "The goal of RG is to create an utterance S t that addresses the user's request. This response is generated based on the dialogue context C and the set of relevant knowledge snippets K + . To accomplish this, we concatenate K + and C as the input and use a pre-trained generation model to generate the response. We consider both the decoder-only model, such as GPT-2 (Radford et al.), and the encoderdecoder model, such as BART (Lewis et al., 2020).\nThe model is trained to maximize the generation probability p(S T | C, K + ).\nTo accurately capture the diversity and proportion of opinions, the model needs to understand the sentiment polarity of each knowledge snippet, which is challenging due to the lack of direct supervision. To address this issue, we apply a stateof-the-art aspect-based sentiment analysis (ABSA) model (Zhang et al., 2021) to predict the sentiment\nZ = [z 1 , • • • , z i , • • • ] for each knowledge snippet K i ∈ K + .\nThen we incorporate the sentiment information into RG by maximizing p(S T | C, K + , Z).\nMore specifically, we first convert the predicted z i into a natural language description using templates, and then append it to the end of the corresponding K i as the enhanced input of RG. For example, given the knowledge snippet as \"The ambience was so fun.\", the ABSA model detects the aspect-based sentiment as (\"ambience\", \"positive\"). We first convert the sentiment into a natural language \"ambience is great.\" and then enhance the knowledge snippet as \"The ambience was so fun. ambience is great.\". We refer to Appendix B for more details."
},
{
"figure_ref": [],
"heading": "Experiments on Sub-Tasks",
"publication_ref": [],
"table_ref": [],
"text": "We first conduct experiments on each individual subtask. To avoid any error accumulation from upstream tasks, we use the gold output of the previous task as the input to the current target task. The detailed experimental setup can be found in Appendix C."
},
{
"figure_ref": [],
"heading": "Knowledge-Seeking Turn Detection",
"publication_ref": [
"b9",
"b35",
"b28",
"b21",
"b25"
],
"table_ref": [
"tab_1"
],
"text": "Setting We conduct experiments using various pre-trained language models, including BERT 2 (Devlin et al., 2019), RoBERTa (Liu et al., 2019), ALBERT (Lan et al., 2020), and DeBERTa (He et al., 2021).\nEvaluation We report the precision, recall, F 1 score, and accuracy score.\nResults Table 3 shows the results of the KTD task. All models achieve similar and near-perfect performance, which is in line with the findings of Kim et al. (2020). It demonstrates that it is feasible to identify the user requests that require subjective knowledge, allowing them to be explicitly addressed by an SK-TOD component. However, this KTD classifier's performance may be specific 2 We use the base version of all pre-trained models. to this dataset or similar domains, and its generalizability to unseen domains or knowledge types requires further exploration in future works."
},
{
"figure_ref": [],
"heading": "Acc",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Entity Tracking",
"publication_ref": [
"b23"
],
"table_ref": [],
"text": "Setting We follow the setting of Jin et al. (2021) to run the ET method.\nEvaluation We report the instance-level accuracy score. An instance is regarded as accurate only if the predicted entities match exactly with the gold entities."
},
{
"figure_ref": [],
"heading": "Results",
"publication_ref": [],
"table_ref": [],
"text": "The fuzzy n-gram matching method achieves an instance-level accuracy of 92.18%. We further analyzed the type of errors. For 1.8% of the instances, there is at least one gold entity missing from the predicted entities. For 7.6% of the instances, the predicted entities contain at least one spurious entity. The latter error case can be further reduced by using model-based matching approaches, which we leave as future work."
},
{
"figure_ref": [],
"heading": "Knowledge Selection",
"publication_ref": [
"b40",
"b55"
],
"table_ref": [],
"text": "Setting We fine-tune the KS models following the same setting as in the KTD task. Additionally, we compare them with traditional information retrieval (IR) baselines, such as TF-IDF (Manning et al., 2008) and BM25 (Robertson et al., 2009).\nEvaluation Knowledge selection can be viewed as either a classification task or a retrieval task.\nFor classification, we use precision, recall, and F 1 measures. We calculate these measures at both the instance level and the snippet level. For the instance level, we first calculate P /R/F 1 for each instance, and then take the average over all instances as the final scores. For the snippet level, instead of computing P /R/F 1 for each instance, we calculate these scores for all <C, K> pairs in the entire dataset. Regarding retrieval evaluation, we use mean-average-precision (mAP) as the metric, which is not dependent on a specific threshold value and can reflect the overall ranking positions of all relevant knowledge snippets. Since the total number of the relevant knowledge snippets can vary for each instance, we do not include top-K-based measures like Precision@K or Recall@K, which are commonly used in other Fact-TOD and knowledgegrounded open-domain dialogue tasks.\nResults Table 4 shows the results of the KS task. Firstly, when comparing our models with IR baselines, all of the trained models outperform the baselines, indicating that the KS model can benefit from the annotated training data. We then compare bi-encoder models and cross-encoder models, and as expected, cross-encoder models outperform bi-encoder models by a large margin. When comparing the performance of different pre-trained models, there is a notable difference among the models under the bi-encoder setting. The variance becomes smaller when applying the cross-encoder architecture. DeBERTa achieves the best performance on all measures in both the bi-encoder and cross-encoder settings.\nFinally, we compare the performance between the seen subset and the unseen subset. At the bottom of Table 4, we list the performance of De-BERTa on both the seen and unseen test subsets. The results reveal a large gap between the perfor- 89 23.17 6.53 18.33 9.62 30.83 14.93 GPT2 9.04 33.9 13.52 26.73 16.27 39.73 22.66 DialoGPT 9.19 33.6 13.62 26.81 16.15 39.72 22.05 BART 10.8 36.35 15.04 28.57 17.96 41.12 24.02 BARTABSA 10.78 36.30 15.36 28.47 18.06 41.75 23.66 T5 10.72 36.50 15.57 28.81 18.33 40.84 25.36 T5ABSA 10.97 36.66 15.51 28.88 18.15 40.94 24.75 Table 5: Results of RG task. Models are evaluated using BLEU, ROUGE (R-1, R-2, R-L), METEOR (MT), and BertScore (BS). We also listed the average length (Len) of the generated response. Encoder-decoder models such as BART and T5 achieve better performance compared with GPT2-based models.\nBLEU R-1 R-2 R-L MT BS Len EXT 2.\nmance of the two subsets, indicating that one of the challenges for the KS model is to generalize from seen aspects to unseen aspects."
},
{
"figure_ref": [],
"heading": "Response Generation",
"publication_ref": [
"b31",
"b52",
"b14",
"b46",
"b32",
"b2"
],
"table_ref": [
"tab_3"
],
"text": "Setting we experiment with decoder-only generation models such as GPT-2 (Radford et al.) 3 and DialoGPT (Zhang et al., 2020c), as well as encoderdecoder models such as BART (Lewis et al., 2020) and T5 (Raffel et al., 2020). We also include two ABSA-enhanced models, namely BART ABSA and T5 ABSA . During decoding, we use beam-search with top-K sampling (Fan et al., 2018). We set the beam size as 5 and sample from the top 50 tokens.\nWe also compare with a random extractive baseline (EXT), where the response is created by randomly selecting a relevant knowledge snippet.\nEvaluation Following the evaluation of other generation tasks, We employ several automatic evaluation metrics, including BLEU (Papineni et al., 2002), ROUGE (Lin, 2004), METEOR (Banerjee and Lavie, 2005), as well as BERTScore (Zhang et al., 2020b), to evaluate the quality of the generated responses compared to the reference responses. We also conduct a human evaluation, where we ask crowd workers to evaluate the quality of responses.\nResults As presented in Table 5, machinegenerated responses significantly outperform the extractive responses. Encoder-decoder models achieve better performance across all automatic measures compared to GPT-based models, indicating that they are more suitable for this task. They also tend to generate longer responses. There is no clear difference in automatic measures when comparing BART and T5. For ABSA-enhanced models, BART ABSA achieves the best performance on BertScore, while T5 ABSA achieves the best score on BLEU and ROUGE.\nHuman Evaluation To obtain a more reliable assessment of response quality, we also conduct a human evaluation on AMT. We use the same group of workers involved in the data collection process. During the evaluation, we show the dialogue context, the oracle knowledge snippets, and all responses (both the reference and the generated responses) to the workers. We randomly sample 240 instances from the test set for evaluation. For each instance, we ask three independent workers to compare the responses based on three measures:\n• Appropriateness: whether the response is fluent and naturally connected to the dialogue context.\n• Aspect Accuracy: whether the response provides relevant and useful information to the aspect that the user queried.\n• Sentiment Accuracy: whether the sentiment proportion provided by the response is consistent with that of the subjective knowledge.\nFor sentiment accuracy, we first ask workers to annotate the sentiment label of each knowledge snippet, and then evaluate each response. All three measures are evaluated using a 5-Point Likert scale. The system-level score is computed as the average score over all instances and workers for each system. The compensation for workers was set at $0.25 for the tasks of appropriateness and aspect accuracy, and $0.4 for the task of sentiment accuracy. The average hourly pay for the crowd workers was $15.25/hr, $14.40/hr, and $14.85/hr for each evaluation task, exceeding the local living minimum wage.\nTable 6 shows the results of human evaluation for response generation. The inter-annotator agreement scores for each task are 0.7270, 0.7535, and 0.6239 in Gwet's gamma, respectively. The results show that machine-generated responses are comparable to the references in terms of appropriateness and aspect accuracy. Moreover, incorporating ABSA can improve the model's performance in sentiment accuracy. However, there is still a large gap in sentiment accuracy between the best modelgenerated responses and the references, indicating Dialogue History: U: Are there any hotels (not guesthouses) that have 0 stars? S: Yes, actually. Cityroomz is a moderately priced 0 star hotel. U: Do they have free wifi? S: Yes, they do have free wifi! Would you like me to book a stay for you? U: Does the Cityroomz have strong water pressure in the shower?\nReviews:\nReview 1: However, I noticed that when taking a shower there was very little water pressure which made the shower disappointing.\nReview 2: The water pressure is not good and it comes out really slow. Review 3: Everything else was a disaster, especially the lack of cleanliness in the room when we checked in, and the incredibly low water pressure in the bathroom when we took our morning showers. Review 4: I encounter a problem when trying to shower and washing hands. Review 5: There is no luggage storage service or anything like that, and the water pressure is much to be desired."
},
{
"figure_ref": [],
"heading": "Model Outputs:",
"publication_ref": [],
"table_ref": [],
"text": "GPT-2: The Cityroomz gets some mixed reviews on their water pressure. Some guests found the water pressure to be very good while others found it to be disappointing. Would you like to know more about them? BART: The Cityroomz hotel has very low water pressure according to most guests. Only one person had a good experience with it.\nBART ABSA : The Cityroomz has a lot of complaints about their water pressure being a bit low. Would you like me to find another place with better ratings? T5: The Cityroomz has mixed reviews about their water pressure. Some guests found the water pressure to be very low while others found it to be disappointing. Would you like to know more about them? T5 ABSA : The Cityroomz has a lot of reviews that say the water pressure is very low and disappointing. Do you want to look at some other places?"
},
{
"figure_ref": [],
"heading": "Reference:",
"publication_ref": [],
"table_ref": [],
"text": "No, guests consistently complain about the water pressure, unfortunately.\nWill that be okay or should I do another search?\nTable 7: Sampled output of different RG models.\nthat faithfully aggregating sentiment information from multiple knowledge snippets is still a challenging task for current models.\nQualitative Analysis Table 7 shows an example of responses generated by various systems. In this example, all the reviews express negative opinions about water pressure. However, responses generated by GPT-2 and BART include positive opinions. T5 correctly mentions the negative opinions but the generated response is not natural and coherent. By incorporating the ABSA model, both BART and T5 correctly generate responses with all negative opinions."
},
{
"figure_ref": [],
"heading": "Experiments on End-2-End Evaluation",
"publication_ref": [],
"table_ref": [
"tab_4"
],
"text": "In the previous section, we use gold information as input for each module to avoid error accumulation. In this section, we evaluate the entire pipeline in an end-to-end manner, where the input of each subtask is predicted by the previous component. We gradually add KS, ET, and KTD to the pipeline, and list the performance of KS and RG in Table 8. The results show that errors introduced during KS can decrease the quality of response generation. However, ET and KTD do not have a significant impact on the performance of downstream tasks. It is because ET and KTD results include fewer noisy predictions compared to the KS results."
},
{
"figure_ref": [],
"heading": "Comparison with Fact-TOD",
"publication_ref": [
"b25"
],
"table_ref": [
"tab_5",
"tab_6"
],
"text": "One difference between SK-TOD and Fact-TOD is that responses in SK-TOD are grounded on subjective knowledge instead of factual knowledge. In this section, we investigate whether a Fact-TOD model can ground on subjective knowledge to address subjective requests. To this end, we retrain our KTD (DeBERTa), KS (DeBERTa crossencoder), and RG (BART) models using the FAQgrounded TOD data provided by Kim et al. (2020). We then apply it to the test set of SK-TOD without further training. We compare the results of each sub-task using the Fact-TOD models with the results of models trained on SK-TOD.\nAs shown in Table 9, for all tasks, there is a significant performance gap between models trained on Fact-TOD and on SK-TOD. We also provide sampled outputs by the Fact-TOD model and the SK-TOD model in Table 10. By checking the model output, we observe that the Fact-TOD model Dialogue History: U: I am looking for a guesthouse to stay at that has free wifi. S: Ok. Were you hoping for a particular price range or part of town? U: I don't care about the part of town or the price range, but it needs to be a three star hotel. S: The Hamilton Lodge would be perfect! It is a moderately priced 3 star guesthouse on the north side of town. Does that sound okay to you? U: Does it have comfortable beds?\nReviews: Review 1: The bed really needed a new mattress, too. Review 2: That being said, the room was spacious and the bed was very comfortable. Review 3: We both slept extremely well every night we were there.\nModel Outputs: Fact-TOD: Yes, the Hamilton Lodge has a new mattress. Would you like to make a reservation? SK-TOD: The Hamilton Lodge has really comfortable beds according to most guests, but one guest did say that the bed needed a new mattress."
},
{
"figure_ref": [],
"heading": "Reference:",
"publication_ref": [],
"table_ref": [],
"text": "There are some mixed reviews on the beds. Some say they're very comfortable while others were not impressed. Would you like me to find another place with better reviews? tends to only ground on and copy information from a single knowledge snippet. This behavior indicates that it is difficult to apply the Fact-TOD model to the SK-TOD task directly, as the Fact-TOD model lacks the ability to effectively aggregate information from multiple knowledge snippets, especially when there are diverse and contradictory opinions. The results also highlight that compared to Fact-TOD, SK-TOD faces new challenges in terms of subjective content understanding and dialogue modeling when integrating subjective knowledge into the responses."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "In this paper, we have introduced SK-TOD: a novel task focused on subjective-knowledge-based taskoriented dialogue response generation. We create and release a large-scale, manually-annotated dataset for this task. Incorporating subjective knowledge requires models to accurately identify all relevant knowledge snippets and faithfully aggregate the information into concise and contextually appropriate responses, which brings unique challenges to this task. Experiments with strong baselines show that there is a significant performance gap between human-generated and machinegenerated responses, particularly in faithfully capturing the diversity and proportion of opinions present in the subjective knowledge. We hope this task together with the provided dataset can promote future research on knowledge-grounded TOD systems and subjective content understanding."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [],
"table_ref": [],
"text": "The dataset we collected contains two domains, restaurants and hotels. However, to evaluate the model's ability to generalize across different domains, it would be beneficial to include more domains in the dataset. Additionally, to address privacy and copyright concerns, we used crowdsourcing to collect review data, resulting in fewer and shorter reviews than those found in real-world scenarios. This limitation can be mitigated by sampling informative and reliable reviews from realworld data. Regarding the model, we did not investigate more complex models, such as large language models and novel architectures. However, we provide a strong baseline method that will serve as a benchmark for more advanced methods by the research community."
},
{
"figure_ref": [],
"heading": "Ethical Considerations",
"publication_ref": [],
"table_ref": [],
"text": "To build our dataset, we collected the dialogue data by augmenting MultiWOZ 2.1, which is a publicly available English dialogue dataset under MIT license. Additionally, we collected the review data using crowd-sourcing, where we provided crowd workers with the reviewer's persona, as well as the aspects and sentiments of reviews. This controlled review collection process helps to exclude offensive or harmful content from the reviews. It also helps to avoid privacy or copyright issues when making the dataset publicly available. Our dataset is available under the CDLA-Sharing 1.0 license."
},
{
"figure_ref": [
"fig_2",
"fig_3",
"fig_4"
],
"heading": "A Data Collection",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we describe more details of the data collection process. The data collection is conducted by a group of crowd workers through Amazon Mechanical Turk. To control the data quality, we choose English speakers from the US, CA, and GB. Workers are eligible for the annotation only if they pass our pre-qualification tests. During data collection, we also manually validate the annotation quality in several rounds to filter out the workers with low-quality annotations.\nDuring review collection, we provide the reviewer's persona, as well as the aspects and sentiments of reviews to workers. The persona is randomly sampled from a pre-defined set of personas. For the aspects and sentiments, we first define 26 common aspects for hotel and restaurant reviews (e.g., WIFI-quality and room-bed for hotels, food-quality and indoor-decor for restaurants). We then randomly selected the target aspects to be addressed in a review. The number of aspects is randomly chosen. To mimic the sentiment distribution of the real reviews, the sentiment of each aspect is sampled based on the actual average ratings taken from Yelp. Figure 3 shows the interface of review collection. We pay workers $1.00 per task.\nDuring user request collection, we ask workers to select the best position to insert a user request by considering every possible position of the given dialogue. Figure 4 shows the interface of user request collection. We pay workers $0.15 per task.\nDuring response generation, we explicitly ask workers to consider the information in all snippets to create a natural and faithful response. Figure 5 shows the interface of response generation. We pay workers $0.25 per task. Below we list the complete instructions that we provide to workers.\n• Please read ALL the customer reviews carefully.\n• Please read the conversation carefully.\n• Write down a response to the customer to answer the question and continue the conversation.\n• You must read EVERY REVIEW COM-MENT carefully. Each sentence was written by different people with potentially different opinions.\n• Your response MUST include your SUM-MARY of ALL the review sentences. • If there's any conflict or different opinions in the reviews, your response MUST describe the minority opinion as well.\n• Your response MUST be based on the contents in given review comments only.\n• Please keep the way of speaking as similar as possible to the previous utterances spoken by the agent."
},
{
"figure_ref": [],
"heading": "B Aspect Based Sentiment Analysis",
"publication_ref": [
"b67",
"b50",
"b49"
],
"table_ref": [],
"text": "To enhance the model's ability to understand the sentiment polarity of each individual knowledge snippet, we apply PGEN (Zhang et al., 2021), a state-of-the-art aspect-based sentiment analysis model, to predict the sentiment Z\n= [z 1 , z 2 , • • • , z i , • • • ] for every knowledge snippet [K 1 , K 2 , • • • , K i , • • • ] in K + .\nPGEN converts the problem of aspect-based sentiment analysis into a sequence generation problem, where the input is the review sentence, and the output is a natural language description of the aspect and the sentiment. For example, given the review sentence as \"The ambience was so fun.\", where the aspect term is \"ambience\" and the corresponding sentiment polarity is \"positive\", PGEN transform the aspect term and the sentiment polarity into a natural language description \"ambience is great.\" using templates. It is transformed by keeping the aspect term unchanged and mapping the positive/neutral/negative sentiment polarities into one of the three tokens: \"great\", \"ok\", and \"bad\". The model is trained using a BART-base model on semeval aspect-based sentiment analysis datasets (Pontiki et al., 2015(Pontiki et al., , 2016))."
},
{
"figure_ref": [],
"heading": "C Training Details",
"publication_ref": [
"b61",
"b37",
"b20",
"b23"
],
"table_ref": [],
"text": "For KTD and KS, the implementation is based on Transformers (Wolf et al., 2020). During training, we use AdamW (Loshchilov and Hutter, 2018) with a learning rate of 3 × 10 -5 and a batch size of 16. We apply warmup (Goyal et al., 2017) on the first 500 steps and early stopping based on the model performance on the validation set. We use a Tesla V100 GPU with 16 GB memory for training models. It takes 1 hour to train a KTD model and 5 hours to train a KS model.\nDuring inference, we set the classification threshold as 0 for KTD, as we observe that KTD results are insensitive to the threshold. However, for the KS model, the setting of the threshold can greatly impact the precision and recall scores. We therefore choose the best threshold based on the F 1 scores on the validation set. We use a grid search between -5 to 5. The optimal thresholds for BERT, RoBERTa, ALBERT, and DeBERTa are 2.25, 1, 1.75, and 2 in the bi-encoder setting. They are 3. 1, 4.6, 3.25, and 3.4 in the cross-encoder setting.\nFor ET model, we follow the setting of Jin et al. (2021) to identify entities. More specifically, we perform the fuzzy n-gram matching between an entity and the utterance, where n is the same as the length of the entity mention. The n-gram matching score is calculated based on the ratio of the longest common sequence between two n-grams. We set the matching threshold as 0.95.\nFor RG model, during training, we use AdamW with a learning rate of 3 × 10 -5 and a batch size of 16. We apply the warmup on the first 500 steps and the early stopping based on the model performance (perplexity) on the development set. The model is trained on a Tesla V100 GPU with 16 GB memory for 2 hours."
}
] | Task-oriented Dialogue (TOD) Systems aim to build dialogue systems that assist users in accomplishing specific goals, such as booking a hotel or a restaurant. Traditional TODs rely on domain-specific APIs/DBs or external factual knowledge to generate responses, which cannot accommodate subjective user requests (e.g.,"Is the WIFI reliable?" or "Does the restaurant have a good atmosphere?"). To address this issue, we propose a novel task of subjective-knowledge-based TOD (SK-TOD). We also propose the first corresponding dataset, which contains subjective knowledgeseeking dialogue contexts and manually annotated responses grounded in subjective knowledge sources. When evaluated with existing TOD approaches, we find that this task poses new challenges such as aggregating diverse opinions from multiple knowledge snippets. We hope this task and dataset can promote further research on TOD and subjective content understanding. | "What do others think?": Task-Oriented Conversational Modeling with Subjective Knowledge | [
{
"figure_caption": "have friendly and engaging staff… The room and hotel had a fast wifi which was useful and not aggravatingly slow like we've all seen. The room was … I traveled to the Avalon alone for work. The slow Wi-Fi and noisy room made work a bit difficult. I really liked the … Gonville Hotel I stayed at the Gonville and it was amazing! They had fast wifi and a great top floor view! It also has … I recently stayed at Gonville ... They had stable wifi and it was even better as it was free. The food is … One thing that was kind of disappointing was the breakfast … We loved their breakfast options ...",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2: The pipeline architecture of SK-TOD.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: The interface of review collection.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: The interface of user request collection.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 :5Figure 5: The interface of response generation.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Basic statistics of our dataset.",
"figure_data": "TrainValTest",
"figure_id": "tab_0",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Results of KTD task. Models are evaluated using Accuracy, Precision, Recall, and F 1 . All models achieve similar and near-perfect performance.",
"figure_data": "PRFBERT99.6799.7599.6199.68RoBERTa99.7499.8699.6499.75ALBERT99.4999.6499.3699.50DeBERTa99.7199.8699.5799.71",
"figure_id": "tab_1",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Results of human evaluation for RG.",
"figure_data": "Approp. Asp-Acc Senti-AccEXT2.653.323.13GPT24.554.543.20BART4.554.683.56BARTABSA4.584.663.80T54.404.633.87T5ABSA4.494.673.98Reference4.704.774.50",
"figure_id": "tab_3",
"figure_label": "6",
"figure_type": "table"
},
{
"figure_caption": "Results of the end-to-end evaluation. We start from RG with gold knowledge as input. We then gradually add components (KS, ET, and KTD) to the pipeline to replace the gold input with the predicted one.",
"figure_data": "KSRGMacro-F mAPBLEU R-LBSRG--10.80 28.52 41.12+KS84.6091.8410.20 27.78 40.64+ET+KS83.4790.4510.29 27.80 40.56+KTD+ET+KS83.4690.4510.27 27.79 40.55KTDKSRGAccMacro-F mAP BLEU R-L BSFact-TOD 87.6259.55 76.696.15 23.25 33.16SK-TOD 99.7184.60 91.84 10.80 28.57 41.12",
"figure_id": "tab_4",
"figure_label": "8",
"figure_type": "table"
},
{
"figure_caption": "Comparison between models trained on Fact-TOD and SK-TOD training data.",
"figure_data": "",
"figure_id": "tab_5",
"figure_label": "9",
"figure_type": "table"
},
{
"figure_caption": "Sampled outputs from the Fact-TOD model and the SK-TOD model, respectively.",
"figure_data": "",
"figure_id": "tab_6",
"figure_label": "10",
"figure_type": "table"
}
] | Chao Zhao; Spandana Gella; Seokhwan Kim; Di Jin; Devamanyu Hazarika; Alexandros Papangelis; Behnam Hedayatnia; Mahdi Namazifar; Yang Liu; Dilek Hakkani-Tur | [
{
"authors": "Stefanos Angelidis; Reinald Kim Amplayo; Yoshihiko Suhara; Xiaolan Wang; Mirella Lapata",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b0",
"title": "Extractive opinion summarization in quantized transformer spaces",
"year": "2021"
},
{
"authors": "Hyunmi Baek; Joongho Ahn; Youngseok Choi",
"journal": "International Journal of Electronic Commerce",
"ref_id": "b1",
"title": "Helpfulness of online consumer reviews: Readers' objectives and review cues",
"year": "2012"
},
{
"authors": "Satanjeev Banerjee; Alon Lavie",
"journal": "Association for Computational Linguistics",
"ref_id": "b2",
"title": "METEOR: an automatic metric for MT evaluation with improved correlation with human judgments",
"year": "2005-06-29"
},
{
"authors": "Johannes Bjerva; Nikita Bhutani; Behzad Golshan; Wang-Chiew Tan; Isabelle Augenstein",
"journal": "Association for Computational Linguistics",
"ref_id": "b3",
"title": "SubjQA: A Dataset for Subjectivity and Review Comprehension",
"year": "2020"
},
{
"authors": "Arthur Bražinskas; Mirella Lapata; Ivan Titov",
"journal": "",
"ref_id": "b4",
"title": "Unsupervised opinion summarization as copycatreview generation",
"year": "2020"
},
{
"authors": "Paweł Budzianowski; Tsung-Hsien Wen; Bo-Hsiang Tseng; Iñigo Casanueva; Stefan Ultes; Milica Osman Ramadan; Gašić",
"journal": "Association for Computational Linguistics",
"ref_id": "b5",
"title": "MultiWOZ -a largescale multi-domain Wizard-of-Oz dataset for taskoriented dialogue modelling",
"year": "2018"
},
{
"authors": "Zhiyu Chen; Bing Liu; Seungwhan Moon; Chinnadhurai Sankar; Paul Crook; William Yang; Wang ",
"journal": "Association for Computational Linguistics",
"ref_id": "b6",
"title": "KETOD: Knowledge-enriched task-oriented dialogue",
"year": "2022"
},
{
"authors": "Eric Chu; Peter Liu",
"journal": "",
"ref_id": "b7",
"title": "Meansum: a neural model for unsupervised multi-document abstractive summarization",
"year": "2019"
},
{
"authors": " Pmlr",
"journal": "",
"ref_id": "b8",
"title": "",
"year": ""
},
{
"authors": "Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova",
"journal": "Association for Computational Linguistics",
"ref_id": "b9",
"title": "BERT: Pre-training of deep bidirectional transformers for language understanding",
"year": "2019"
},
{
"authors": "Eleftherios Dimitrakis; Konstantinos Sgontzos; Panagiotis Papadakos; Yannis Marketakis; Alexandros Papangelis; Yannis Stylianou; Yannis Tzitzikas",
"journal": "",
"ref_id": "b10",
"title": "On finding the relevant user reviews for advancing conversational faceted search",
"year": "2018"
},
{
"authors": "Emily Dinan; Stephen Roller; Kurt Shuster; Angela Fan; Michael Auli; Jason Weston",
"journal": "",
"ref_id": "b11",
"title": "Wizard of wikipedia: Knowledge-powered conversational agents",
"year": "2018"
},
{
"authors": "Mihail Eric; Rahul Goel; Shachi Paul; Abhishek Sethi; Sanchit Agarwal; Shuyang Gao; Adarsh Kumar; Anuj Goyal; Peter Ku; Dilek Hakkani-Tur",
"journal": "European Language Resources Association",
"ref_id": "b12",
"title": "Mul-tiWOZ 2.1: A consolidated multi-domain dialogue dataset with state corrections and state tracking baselines",
"year": "2020"
},
{
"authors": "Mihail Eric; Lakshmi Krishnan; Francois Charette; Christopher D Manning",
"journal": "Association for Computational Linguistics",
"ref_id": "b13",
"title": "Key-value retrieval networks for task-oriented dialogue",
"year": "2017"
},
{
"authors": "Angela Fan; Mike Lewis; Yann Dauphin",
"journal": "Association for Computational Linguistics",
"ref_id": "b14",
"title": "Hierarchical neural story generation",
"year": "2018"
},
{
"authors": "Song Feng; Sankalp Siva; Hui Patel; Sachindra Wan; Joshi",
"journal": "Association for Computational Linguistics",
"ref_id": "b15",
"title": "MultiDoc2Dial: Modeling dialogues grounded in multiple documents",
"year": "2021"
},
{
"authors": "Song Feng; Hui Wan; Chulaka Gunasekara; Siva Patel; Sachindra Joshi; Luis Lastras",
"journal": "Association for Computational Linguistics",
"ref_id": "b16",
"title": "doc2dial: document-grounded dialogue dataset",
"year": "2020"
},
{
"authors": "Michel Galley; Chris Brockett; Xiang Gao; Jianfeng Gao; Bill Dolan",
"journal": "",
"ref_id": "b17",
"title": "Grounded response generation task at dstc7",
"year": "2019"
},
{
"authors": "Marjan Ghazvininejad; Chris Brockett; Ming-Wei Chang; William B Dolan; Jianfeng Gao; Wen Tau Yih; Michel Galley",
"journal": "",
"ref_id": "b18",
"title": "A knowledge-grounded neural conversation model",
"year": "2018"
},
{
"authors": "Karthik Gopalakrishnan; Behnam Hedayatnia; Qinglang Chen; Anna Gottardi; Sanjeev Kwatra; Anu Venkatesh; Raefer Gabriel; Dilek Hakkani-Tür; Amazon Alexa; A I ",
"journal": "",
"ref_id": "b19",
"title": "Topical-chat: Towards knowledge-grounded open-domain conversations",
"year": "2019"
},
{
"authors": "Priya Goyal; Piotr Dollár; Ross Girshick; Pieter Noordhuis; Lukasz Wesolowski; Aapo Kyrola; Andrew Tulloch; Yangqing Jia; Kaiming He",
"journal": "",
"ref_id": "b20",
"title": "Accurate, large minibatch sgd: Training imagenet in 1 hour",
"year": "2017"
},
{
"authors": "Pengcheng He; Xiaodong Liu; Jianfeng Gao; Weizhu Chen",
"journal": "",
"ref_id": "b21",
"title": "Deberta: Decoding-enhanced bert with disentangled attention",
"year": "2021"
},
{
"authors": "Qingnan Jiang; Lei Chen; Ruifeng Xu; Xiang Ao; Min Yang",
"journal": "Association for Computational Linguistics",
"ref_id": "b22",
"title": "A challenge dataset and effective models for aspect-based sentiment analysis",
"year": "2019"
},
{
"authors": "Di Jin; Seokhwan Kim; Dilek Hakkani-Tur",
"journal": "",
"ref_id": "b23",
"title": "Can i be of further assistance? using unstructured knowledge access to improve task-oriented conversational modeling",
"year": "2021"
},
{
"authors": " Michael A Kamins; J Meribeth; Stuart A Brand; John C Hoeke; Moe",
"journal": "Journal of advertising",
"ref_id": "b24",
"title": "Two-sided versus one-sided celebrity endorsements: The impact on advertising effectiveness and credibility",
"year": "1989"
},
{
"authors": "Seokhwan Kim; Mihail Eric; Karthik Gopalakrishnan; Behnam Hedayatnia; Yang Liu; Dilek Hakkani-Tur",
"journal": "",
"ref_id": "b25",
"title": "Beyond domain APIs: Task-oriented conversational modeling with unstructured knowledge access",
"year": "2020"
},
{
"authors": "Seokhwan Kim; Yang Liu; Di Jin; Alexandros Papangelis; Karthik Gopalakrishnan; Behnam Hedayatnia; Dilek Hakkani-Tür",
"journal": "IEEE",
"ref_id": "b26",
"title": "how robust ru?\": Evaluating task-oriented dialogue systems on spoken conversations",
"year": "2021"
},
{
"authors": "Mojtaba Komeili; Kurt Shuster; Jason Weston",
"journal": "Association for Computational Linguistics",
"ref_id": "b27",
"title": "Internet-augmented dialogue generation",
"year": "2022"
},
{
"authors": "Zhenzhong Lan; Mingda Chen; Sebastian Goodman; Kevin Gimpel; Piyush Sharma; Radu Soricut",
"journal": "",
"ref_id": "b28",
"title": "Albert: A lite bert for self-supervised learning of language representations",
"year": "2020"
},
{
"authors": "Jumin Lee; Do-Hyung Park; Ingoo Han",
"journal": "Electronic commerce research and applications",
"ref_id": "b29",
"title": "The effect of negative online consumer reviews on product attitude: An information processing view",
"year": "2008"
},
{
"authors": "Esther Levin; Roberto Pieraccini; Wieland Eckert",
"journal": "IEEE Transactions on speech and audio processing",
"ref_id": "b30",
"title": "A stochastic model of human-machine interaction for learning dialog strategies",
"year": "2000"
},
{
"authors": "Mike Lewis; Yinhan Liu; Naman Goyal; Marjan Ghazvininejad; Abdelrahman Mohamed; Omer Levy; Veselin Stoyanov; Luke Zettlemoyer",
"journal": "",
"ref_id": "b31",
"title": "Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension",
"year": "2020"
},
{
"authors": "Chin-Yew Lin",
"journal": "Association for Computational Linguistics",
"ref_id": "b32",
"title": "ROUGE: A package for automatic evaluation of summaries",
"year": "2004"
},
{
"authors": "Bing Liu; Lei Zhang",
"journal": "Springer",
"ref_id": "b33",
"title": "A survey of opinion mining and sentiment analysis",
"year": "2012"
},
{
"authors": "Shuman Liu; Hongshen Chen; Zhaochun Ren; Yang Feng; Qun Liu; Dawei Yin",
"journal": "Association for Computational Linguistics",
"ref_id": "b34",
"title": "Knowledge diffusion for neural dialogue generation",
"year": "2018"
},
{
"authors": "Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov",
"journal": "",
"ref_id": "b35",
"title": "Roberta: A robustly optimized bert pretraining approach",
"year": "2019"
},
{
"authors": "Yinong Long; Jianan Wang; Zhen Xu; Zongsheng Wang; Baoxun Wang; Zhuoran Wang",
"journal": "",
"ref_id": "b36",
"title": "A knowledge enhanced generative conversational service agent",
"year": "2017"
},
{
"authors": "Ilya Loshchilov; Frank Hutter",
"journal": "",
"ref_id": "b37",
"title": "Decoupled weight decay regularization",
"year": "2018"
},
{
"authors": "Prasad Bodhisattwa; Harsh Majumder; Taylor Jhamtani; Julian Berg-Kirkpatrick; Mcauley",
"journal": "Association for Computational Linguistics",
"ref_id": "b38",
"title": "Achieving conversational goals with unsupervised post-hoc knowledge injection",
"year": "2022"
},
{
"authors": "Prasad Bodhisattwa; Shuyang Majumder; Jianmo Li; Julian Ni; Mcauley",
"journal": "Association for Computational Linguistics",
"ref_id": "b39",
"title": "Interview: Large-scale modeling of media dialog with discourse patterns and knowledge grounding",
"year": "2020"
},
{
"authors": "Prabhakar Christopher D Manning; Hinrich Raghavan; Schütze",
"journal": "Cambridge university press",
"ref_id": "b40",
"title": "Introduction to information retrieval",
"year": "2008"
},
{
"authors": "Pierre-Emmanuel Mazaré; Samuel Humeau; Martin Raison; Antoine Bordes",
"journal": "Association for Computational Linguistics",
"ref_id": "b41",
"title": "Training millions of personalized dialogue agents",
"year": "2018"
},
{
"authors": "Julian Mcauley; Alex Yang",
"journal": "",
"ref_id": "b42",
"title": "Addressing complex and subjective product-related queries with customer reviews",
"year": "2016"
},
{
"authors": "Nikita Moghe; Siddhartha Arora; Suman Banerjee; Mitesh M Khapra",
"journal": "",
"ref_id": "b43",
"title": "Towards exploiting background knowledge for building conversation systems",
"year": "2018-10-31"
},
{
"authors": "Seungwhan Moon; Pararth Shah; Anuj Kumar; Rajen Subba",
"journal": "Association for Computational Linguistics",
"ref_id": "b44",
"title": "OpenDialKG: Explainable conversational reasoning with attention-based walks over knowledge graphs",
"year": "2019"
},
{
"authors": "Jianmo Ni; Jiacheng Li; Julian Mcauley",
"journal": "",
"ref_id": "b45",
"title": "Justifying recommendations using distantly-labeled reviews and fine-grained aspects",
"year": "2019"
},
{
"authors": "Kishore Papineni; Salim Roukos; Todd Ward; Wei-Jing Zhu",
"journal": "",
"ref_id": "b46",
"title": "Bleu: a method for automatic evaluation of machine translation",
"year": "2002-07-06"
},
{
"authors": " ",
"journal": "",
"ref_id": "b47",
"title": "",
"year": ""
},
{
"authors": "Gustavo Penha; Alexandru Balan; Claudia Hauff",
"journal": "",
"ref_id": "b48",
"title": "Introducing mantis: a novel multi-domain information seeking dialogues dataset",
"year": "2019"
},
{
"authors": "Maria Pontiki; Dimitrios Galanis; Haris Papageorgiou; Ion Androutsopoulos; Suresh Manandhar; Mohammad Al-Smadi; Mahmoud Al-Ayyoub; Yanyan Zhao; Bing Qin; Orphée De Clercq",
"journal": "",
"ref_id": "b49",
"title": "Semeval-2016 task 5: Aspect based sentiment analysis",
"year": "2016"
},
{
"authors": "Maria Pontiki; Dimitrios Galanis; Harris Papageorgiou; Suresh Manandhar; Ion Androutsopoulos",
"journal": "",
"ref_id": "b50",
"title": "Semeval-2015 task 12: Aspect based sentiment analysis",
"year": "2015"
},
{
"authors": "Alec Radford; Jeffrey Wu; Rewon Child; David Luan; Dario Amodei; Ilya Sutskever",
"journal": "",
"ref_id": "b51",
"title": "Language models are unsupervised multitask learners",
"year": ""
},
{
"authors": "Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu",
"journal": "J. Mach. Learn. Res",
"ref_id": "b52",
"title": "Exploring the limits of transfer learning with a unified text-to-text transformer",
"year": "2020"
},
{
"authors": "Abhinav Rastogi; Xiaoxue Zang; Srinivas Sunkara; Raghav Gupta; Pranav Khaitan",
"journal": "",
"ref_id": "b53",
"title": "Towards scalable multi-domain conversational agents: The schema-guided dialogue dataset",
"year": "2020"
},
{
"authors": "Nils Reimers; Iryna Gurevych",
"journal": "Association for Computational Linguistics",
"ref_id": "b54",
"title": "Sentence-BERT: Sentence embeddings using Siamese BERTnetworks",
"year": "2019"
},
{
"authors": "Stephen Robertson; Hugo Zaragoza",
"journal": "Foundations and Trends® in Information Retrieval",
"ref_id": "b55",
"title": "The probabilistic relevance framework: Bm25 and beyond",
"year": "2009"
},
{
"authors": "Marzieh Saeidi; Max Bartolo; Patrick Lewis; Sameer Singh; Tim Rocktäschel; Mike Sheldon; Guillaume Bouchard; Sebastian Riedel",
"journal": "Association for Computational Linguistics",
"ref_id": "b56",
"title": "Interpretation of natural language rules in conversational machine reading",
"year": "2018"
},
{
"authors": "Satinder Singh; Diane Litman; Michael Kearns; Marilyn Walker",
"journal": "Journal of Artificial Intelligence Research",
"ref_id": "b57",
"title": "Optimizing dialogue management with reinforcement learning: Experiments with the njfun system",
"year": "2002"
},
{
"authors": "Yi-Lin Tuan; Yun-Nung Chen; Hung-Yi Lee",
"journal": "Association for Computational Linguistics",
"ref_id": "b58",
"title": "DyKgChat: Benchmarking dialogue generation grounding on dynamic knowledge graphs",
"year": "2019"
},
{
"authors": "Pavlos Vougiouklis; Jonathon Hare; Elena Simperl",
"journal": "",
"ref_id": "b59",
"title": "A neural network approach for knowledgedriven response generation",
"year": "2016"
},
{
"authors": "Jason D Williams; Steve Young",
"journal": "Computer Speech & Language",
"ref_id": "b60",
"title": "Partially observable markov decision processes for spoken dialog systems",
"year": "2007"
},
{
"authors": "Thomas Wolf; Lysandre Debut; Victor Sanh; Julien Chaumond; Clement Delangue; Anthony Moi; Pierric Cistac; Tim Rault; Remi Louf; Morgan Funtowicz; Joe Davison; Sam Shleifer; Clara Patrick Von Platen; Yacine Ma; Julien Jernite; Canwen Plu; Teven Xu; Sylvain Le Scao; Mariama Gugger; Quentin Drame; Alexander Lhoest; Rush",
"journal": "Association for Computational Linguistics",
"ref_id": "b61",
"title": "Transformers: State-of-the-art natural language processing",
"year": "2020"
},
{
"authors": "Thomas Wolf; Victor Sanh; Julien Chaumond; Clement Delangue",
"journal": "",
"ref_id": "b62",
"title": "Transfertransfo: A transfer learning approach for neural network based conversational agents",
"year": "2019"
},
{
"authors": "Chien-Sheng Wu; Richard Socher; Caiming Xiong",
"journal": "",
"ref_id": "b63",
"title": "Global-to-local memory pointer networks for task-oriented dialogue",
"year": "2019"
},
{
"authors": "Jun Xu; Zeyang Lei; Haifeng Wang; Zheng-Yu Niu; Hua Wu; Wanxiang Che",
"journal": "",
"ref_id": "b64",
"title": "Enhancing dialog coherence with event graph grounded content planning",
"year": "2021"
},
{
"authors": "Houyu Zhang; Zhenghao Liu; Chenyan Xiong; Zhiyuan Liu; ; ",
"journal": "Association for Computational Linguistics",
"ref_id": "b65",
"title": "Grounded conversation generation as guided traverses in commonsense knowledge graphs",
"year": "2020"
},
{
"authors": "Tianyi Zhang; Varsha Kishore; Felix Wu; Kilian Q Weinberger; Yoav Artzi",
"journal": "",
"ref_id": "b66",
"title": "Bertscore: Evaluating text generation with bert",
"year": "2020"
},
{
"authors": "Wenxuan Zhang; Yang Deng; Xin Li; Yifei Yuan; Lidong Bing; Wai Lam",
"journal": "Association for Computational Linguistics",
"ref_id": "b67",
"title": "Aspect sentiment quad prediction as paraphrase generation",
"year": "2021"
},
{
"authors": "Yizhe Zhang; Siqi Sun; Michel Galley; Yen-Chun Chen; Chris Brockett; Xiang Gao; Jianfeng Gao; Jingjing Liu; William B Dolan",
"journal": "",
"ref_id": "b68",
"title": "Dialogpt: Largescale generative pre-training for conversational response generation",
"year": "2020"
},
{
"authors": "Chao Zhao; Snigdha Chaturvedi",
"journal": "",
"ref_id": "b69",
"title": "Weaklysupervised opinion summarization by leveraging external information",
"year": "2020"
},
{
"authors": "Kangyan Zhou; Shrimai Prabhumoye; Alan W Black",
"journal": "Association for Computational Linguistics",
"ref_id": "b70",
"title": "A dataset for document grounded conversations",
"year": "2018"
}
] | [
{
"formula_coordinates": [
3,
70.87,
504.36,
218.27,
24.18
],
"formula_id": "formula_0",
"formula_text": "dialogue context C = [U 1 , S 1 , U 2 , S 2 , • • • , U t ]"
},
{
"formula_coordinates": [
3,
202.55,
572.11,
88.49,
10.63
],
"formula_id": "formula_1",
"formula_text": "E = {e 1 , • • • , e m }."
},
{
"formula_coordinates": [
3,
70.87,
639.85,
218.66,
24.18
],
"formula_id": "formula_2",
"formula_text": "reviews R = {R 1 , R 2 , • • • }. Each review can be divided into segments [K 1 , K 2 , • • • ],"
},
{
"formula_coordinates": [
4,
77.78,
82.25,
430.74,
89.09
],
"formula_id": "formula_3",
"formula_text": "5K/22K ✓ ✗ n/a ✗ ✓ ✓ ✗ n/a Space (2021) 1K ✓ ✗ n/a ✗ ✓ ✓ ✓ ✗ Yelp/Amazon (2019; 2020) 200/180 ✓ ✗ n/a ✗ ✗ ✓ ✓ ✗ Justify-Rec (2019) 1.3M ✗ ✗ n/a ✗ ✓ ✗ ✓ ✗ AmazonQA (2016) 309K ✗ ✗ n/a ✓ ✗ ✗ ✗ n/a SubjQA (2020) 10K ✗ ✗ n/a ✓ ✓ ✓ ✗ n/a Holl-E (2018) 9K ✓ ✓ ✗ ✗ ✗ ✗ ✓ ✗ Foursquare (2018) 1M ✗ ✓ ✗ ✗ ✗ ✗ ✓ n/a SK-TOD (Ours) 20K ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓"
},
{
"formula_coordinates": [
5,
116.49,
374.86,
127.02,
26.35
],
"formula_id": "formula_4",
"formula_text": "h = Enc(C) P (C) = softmax (FFN (h)) ."
},
{
"formula_coordinates": [
5,
306.81,
328.23,
216.93,
27.22
],
"formula_id": "formula_5",
"formula_text": "h C = Enc(C), h K = Enc(K) P (C, K) = softmax (FFN (h c , h K , |h C -h K |)) ."
},
{
"formula_coordinates": [
5,
344.62,
423.05,
141.31,
26.35
],
"formula_id": "formula_6",
"formula_text": "h = Enc(C, K) P (C, K) = softmax (FFN (h)) ."
},
{
"formula_coordinates": [
6,
70.87,
196.55,
218.27,
24.18
],
"formula_id": "formula_7",
"formula_text": "Z = [z 1 , • • • , z i , • • • ] for each knowledge snippet K i ∈ K + ."
},
{
"formula_coordinates": [
7,
308.85,
77.11,
210.24,
25.78
],
"formula_id": "formula_8",
"formula_text": "BLEU R-1 R-2 R-L MT BS Len EXT 2."
},
{
"formula_coordinates": [
14,
306.14,
695.54,
218.27,
37.73
],
"formula_id": "formula_9",
"formula_text": "= [z 1 , z 2 , • • • , z i , • • • ] for every knowledge snippet [K 1 , K 2 , • • • , K i , • • • ] in K + ."
}
] |
10.18653/v1/W19-1909 | 2023-05-20 | [
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b2",
"b33",
"b68",
"b49",
"b51",
"b53",
"b50",
"b10",
"b4",
"b22",
"b19",
"b3",
"b13",
"b21",
"b13",
"b29",
"b30",
"b14",
"b22",
"b31",
"b18",
"b19",
"b3",
"b23",
"b38"
],
"table_ref": [],
"text": "The dynamic nature of labor markets, driven by technological changes, migration, and digitization, has resulted in a significant amount of job advertisement data (JAD) being made available on various platforms to attract qualified candidates (Brynjolfsson andMcAfee, 2011, 2014;Balog et al., 2012). This has led to an increase in tasks related to JAD, including skill extraction (Kivimäki et al., 2013;Zhao et al., 2015;Sayfullina et al., 2018;Smith et al., 2019;Tamburri et al., 2020;Shi et al., 2020;Chernova, 2020;Bhola et al., 2020;Zhang et al., 2022a,b,c;Green et al., 2022;Gnehm et al., 2022;Beauchemin et al., 2022;Decorte et al., 2022;Goyal et al., 2023), skill classification (Decorte et al., 2022;Zhang et al., 2022b), job title classification (Javed et al., 2015(Javed et al., , 2016;;Decorte et al., 2021;Green et al., 2022), de-identification of entities in job postings (Jensen et al., 2021), and multilingual skill entity linking (ESCO, 2022).\nWhile some previous studies have focused on JAD in non-English languages (Zhang et al., 2022b;Gnehm et al., 2022;Beauchemin et al., 2022), their baselines have typically relied on language-specific models, either using domain-adaptive pre-training (DAPT; Gururangan et al., 2020) or off-the-shelf models. The lack of comprehensive, open-source JAD data in various languages makes it difficult to fully pre-train a language model (LM) using such data. In this work, we seek external resources that can help improve the multilingual performance on the JAD domain. We use the ESCO taxonomy (le Vrang et al., 2014), which is a standardized system for describing and categorizing the skills, competences, qualifications, and occupations of workers in the European Union. The ESCO taxonomy, which has been curated by humans, covers over 13,000 skills and 3,000 occupations in 27 languages. Therefore, we seek to answer: To what extent can we leverage the ESCO taxonomy to pre-train a domain-specific and language-agnostic model for the computational job market domain?\nIn this work, we release the first multilingual JAD-related model named ESCOXLM-R, a language model based on XLM-R large that incorporates data from the ESCO taxonomy through the use of two pre-training objectives (Figure 1): Masked Language Modeling (MLM) and a novel ESCO relation prediction task (Section 2). We evaluate ESCOXLM-R on 9 JAD-related datasets in 4 different languages covering 2 NLP tasks (Section 3). Our results show that ESCOXLM-R outperforms previous state-of-the-art (SOTA) on 6 out of 9 In the middle of the figure, we show our pre-training setup. Pre-training instances are uniformly sampled in three ways: randomly, linked, or grouped (this is defined in Section 2.2). The selected instances (can be in different languages) are then fed to the language model, along with its description. We have two pre-training objectives: the regular MLM objective, and a new ESCO relation prediction objective, in which the goal is to predict which group the sampled instances belong to (Random, Linked, or Grouped). datasets (Section 4). In addition, our fine-grained analysis reveals that ESCOXLM-R performs better on short spans compared to XLM-R large , and consistently outperforms XLM-R large on entity-level and surface-level span-F1 (Section 5). • The largest JAD evaluation study to date on 3 job-related tasks, comprising 9 datasets in 4 languages and 4 models."
},
{
"figure_ref": [],
"heading": "Contributions",
"publication_ref": [],
"table_ref": [],
"text": "• A fine-grained analysis of ESCOXLM-R's performance on different span lengths, and emerging entities (i.e., recognition of entities in the long tail)."
},
{
"figure_ref": [
"fig_2"
],
"heading": "ESCOXLM-R",
"publication_ref": [
"b17",
"b41",
"b17",
"b12",
"b12",
"b27",
"b37"
],
"table_ref": [],
"text": "Preliminaries In the context of pre-training, an LM is trained using a large number of unlabeled documents, X = X (i) , and consists of two main functions: f encoder (.), which maps a sequence of tokens X = (x 1 , x 2 , ..., x t ) to a contextualized vector representation for each token, represented as (h 1 , h 2 , ..., h t ), and f head (.), the output layer that 1 The code for ESCOXLM-R is available as opensource: https://github.com/mainlp/escoxlmr. We further release ESCOXLM-R under an Apache License 2.0 on HuggingFace: https://huggingface.co./jjzha/ esco-xlm-roberta-large.\ntakes these representations and performs a specific task, such as pre-training in a self-supervised manner or fine-tuning on a downstream application. For example, BERT (Devlin et al., 2019) is pre-trained using two objectives: MLM and Next Sentence Prediction (NSP). In MLM, a portion of tokens in a sequence X is masked and the model must predict the original tokens from the masked input. In the NSP objective, the model takes in two segments (X A , X B ) and predicts whether segment X B follows X A . RoBERTa (Liu et al., 2019) is a variation of BERT that uses dynamic MLM, in which the masking pattern is generated each time a sequence is fed to the LM, and does not use the NSP task.\nMultilinguality Both BERT and RoBERTa have been extended to support multiple languages, resulting in multilingual BERT (mBERT; Devlin et al., 2019) and XLM-RoBERTa (XLM-R; Conneau et al., 2020). XLM-R was found to outperform mBERT on many tasks (e.g., Conneau et al., 2020;Hu et al., 2020;Lauscher et al., 2020) The ESCO dataset contains descriptions in 27 languages, with a combined total of approximately 3.72 million descriptions (i.e., instances). On average, there are around 130,000 descriptions per language. The average length of each description is 26.3 tokens, with some descriptions reaching a maximum length of 150 or more tokens, as shown by the outliers in the boxplot. and occupations of workers in the European Union (EU). It is designed to serve as a common language for the description of skills and qualifications across the EU, facilitating the mobility of workers by providing a common reference point for the recognition of qualifications and occupations. The taxonomy is developed and maintained by the European Commission and is based on the International Classification of Occupations and the International Standard Classification of Education. It includes 27 European languages: Bulgarian (ar), Czech (cs), Danish (da), German (de), Greek (el), English (en), Spanish (es), Estonian (et), Finnish (fi), French (fr), Gaelic (ga), Croatian (hr), Hungarian (hu), Icelandic (is), Italian (it), Lithuanian (lt), Latvian (lv), Maltese (mt), Dutch (nl), Norwegian (no), Polish (pl), Portuguese (pt), Romanian (ro), Slovak (sk), Slovenian (sl), Swedish (sv), and Arabic (ar). Currently, it describes 3,008 occupations and 13,890 skills/competences (SKC) in all 27 languages.2 \nThe ESCO taxonomy includes a hierarchical structure with links between occupations, skills, and aliases (OSA). In this work, we focus on the occupation pages and extract the following infor-mation from the taxonomy:3 \n• ESCO Code: The taxonomy code for the specific occupation or SKC.\n• Occupation Label: The preferred occupation name (i.e., title of the occupation).\n• Occupation Description/Definition: A description of the responsibilities of the specific occupation.\n• Major Group Name: The name of the overarching group to which the occupation belongs, e.g., \"Veterinarians\" for the occupation \"animal therapist\".\n• Alternative Labels: Aliases for the specific occupation, e.g., \"animal rehab therapist\" for the occupation \"animal therapist\".\n• Essential Skills: All necessary SKCs for the occupation, including descriptions of these.\n• Optional Skills: All optional SKCs for the occupation, including descriptions of these.\nIn Figure 2, we present the distribution of pretraining instances and the mean description lengths for each language in the ESCO taxonomy. Note 1 5 10 15 20 25 30\nStep (x1000) that the number of descriptions is not the same for all languages, and we do not count empty descriptions (i.e., missing translations) for certain occupations or SKCs."
},
{
"figure_ref": [
"fig_0"
],
"heading": "Pre-training Setup",
"publication_ref": [
"b60",
"b62"
],
"table_ref": [],
"text": "To (2020). Given the limited amount of training data (3.72M sentences), we utilize the XLM-R large checkpoint provided by the HuggingFace library (Wolf et al., 2020) as a starting point. 4 Our aim is to fine-tune the model to internalize domain-specific knowledge related to occupation and SKCs, while maintaining its general knowledge acquired during the original pre-training phase.\nWe introduce a novel self-supervised pretraining objective for ESCOXLM-R, inspired by LinkBERT from Yasunaga et al. (2022). We view the ESCO taxonomy as a graph of occupations and SKCs (Figure 1), with links between occupations or occupations and SKCs in various languages. By placing similar occupations or SKCs in the same context window and in different languages, we can learn from the links between (occupation ↔ occupation) and (occupation ↔ SKCs) in different languages for true cross-lingual pre-training. In addition to the MLM pre-training objective, which is used to learn concepts within contexts, we introduce another objective called ESCO Relation Prediction (ERP) to internalize knowledge of connections within the taxonomy in the LM. We take an anchor concept (C A ) by concatenating it with its description (X A ) from the ESCO taxonomy and sample an additional concept (C B ) concatenated with its description (X B ) to create LM input 5 We sample C B X B in three ways with uniform probability:\n[CLS] C A X A [SEP] C B X B [SEP].\n1. Random: We randomly sample C B X B from the ESCO taxonomy, in any language;\n2. Linked: We sample C B X B in any language from the same occupation page, for example, an \"animal therapist\" (or an alias of the \"animal therapist\", e.g., \"animal rehab therapist\") should have knowledge of \"animal behavior\";\n3. Grouped: We sample C B X B from the same major group in any language. For the same example \"animal therapist\", it comes from major group 2: Professionals → group 22: Health professionals. Several other concepts, e.g., \"Nursing professionals\" fall under this major group."
},
{
"figure_ref": [],
"heading": "Pre-training Objectives",
"publication_ref": [
"b62",
"b42",
"b40"
],
"table_ref": [],
"text": "The LM is trained using two objectives. First is the MLM objective, and the second is the ERP objective, where the task is to classify the relation r of the Random, Linked, Grouped). The rationale behind this is to encourage the model to learn the relevance between concepts in the ESCO taxonomy.\n[CLS] to- ken in [CLS] C A X A [SEP] C B X B [SEP] (r ∈\nWe formalize the objectives in Equation ( 1):\nL = L MLM + L ERP = - i log p (x i | h i ) -log p (r | h [CLS] ) ,\n(1) we define the overall loss L as the sum of the MLM loss L MLM and the ERP loss L ERP . The MLM loss is calculated as the negative log probability of the input token x i given the representation h i . Similarly, the ERP loss is the negative log probability of the relationship r given the representation of the start-token h ability to capture the relationships between ESCO occupations and skills.\nImplementation For optimization we follow (Yasunaga et al., 2022), we use the AdamW (Loshchilov and Hutter, 2019) optimizer with (β 1 , β 2 ) = (0.9, 0.98). We warm up the learning rate 1e -5 for a ratio of 6% and then linearly decay it. The model is trained for 30K steps, which is equivalent to one epoch over the data, and the training process takes 33 hours on one A100 GPU with tf32. We use a development set comprising 1% of the data for evaluation. In Figure 3, the pre-training loss and performance on the dev. set are plotted, it can be seen that the accuracy plateaus at 30K steps.\nThough the train and development loss hint that further gains could be obtained on the pretraining objective, we found through empirical analysis on downstream tasks that 30K steps performs best. (Liu et al., 2017). In our work, the bottleneck layer is not used and no additional training data is generated through bootstrapping. To keep comparison fair, we re-train their model without the additional layer and bootstrapping. We use Mean Reciprocal Rank as the main results metric."
},
{
"figure_ref": [],
"heading": "Experimental Setup",
"publication_ref": [
"b49",
"b32",
"b26",
"b61",
"b9"
],
"table_ref": [],
"text": "SAYFULLINA (Sayfullina et al., 2018) This dataset is used for soft skill prediction, a sequence labeling problem. Soft skills are personal qualities that contribute to success, such as \"team working\", \"being dynamic\", and \"independent\". The models for this dataset include a CNN (Kim, 2014), an LSTM (Hochreiter et al., 1997), and a Hierarchical Attention Network (Yang et al., 2016). We compare to their best-performing LSTM model. (Chan et al., 2020)."
},
{
"figure_ref": [],
"heading": "GREEN (",
"publication_ref": [
"b3",
"b43"
],
"table_ref": [],
"text": "FIJO (Beauchemin et al., 2022) A French job ad dataset with the task of labeling skill types using a sequence labeling approach. The skill groups are based on the AQESSS public skills repositories and proprietary skill sets provided by their collaborators. These skill types are divided into four categories: \"Thoughts\", \"Results\", \"Relational\", and \"Personal\". The best-performing model for this task is CamemBERT (Martin et al., 2020)."
},
{
"figure_ref": [],
"heading": "Results",
"publication_ref": [],
"table_ref": [
"tab_5"
],
"text": "The results of the models are presented in Table 2.\nTo evaluate the performance, four different models are used in total: ESCOXLM-R, the best-performing model originally reported in the relevant paper for the downstream task, vanilla XLM-R large , and an XLM-R large model that we continuously pre-trained using only MLM (DAPT; excluding the ERP objective) using the same pre-training hyperparameters as ESCOXLM-R. For more information regarding the hyperparameters of fine-tuning, we refer to Appendix C (Table 5).\nEnglish "
},
{
"figure_ref": [],
"heading": "Fijo",
"publication_ref": [],
"table_ref": [],
"text": "Figure 4: Radar Charts of Span-F1 performance by Span Token Length. We show the performance of XLM-R large and ESCOXLM-R on different span lengths, we bucketed the performances of both models according to the length of the spans, up to 10 tokens, and presented the average performance over five random seeds. We did not include error bars in these plots. Note that in some plots, there are no instances in certain buckets (e.g., SAYFULLINA with 7-8, 9-10). Also, some outer rings only go up to 60 span F1, rather than 100.\nXLM-R large (+ DAPT) has higher performance than ESCOXLM-R. Next, we will discuss potential reasons for these differences."
},
{
"figure_ref": [],
"heading": "Analysis",
"publication_ref": [],
"table_ref": [
"tab_5"
],
"text": "We highlight that the performance gains of ESCOXLM-R are generally much larger than any of the losses, indicating a largely positive effect of training on ESCO. The improved performance of ESCOXLM-R on JAD datasets in Table 2 is likely due to the focus on tasks with token-level annotation (i.e., sequence labeling). This suggests that pretraining on the ESCO taxonomy is particularly useful for these types of tasks. The under-performance of ESCOXLM-R on the KOMPETENCER dataset in both EN and DA may be because the task involves predicting the ESCO taxonomy code for a given skill without context, where we expect ESCO to particularly help with tasks where having context is relevant. We suspect applying DAPT and ERP on ESCO specifically improves recognizing entities that are uncommon. On the other hand, the poor performance on the JOBSTACK dataset may be due to the task of predicting various named entities, such as organizations and locations. By manual inspection, we found that ESCO does not contain entities related to organizations, locations, or persons, thus this reveals that there is a lack of relevant pre-training information to JOBSTACK."
},
{
"figure_ref": [],
"heading": "Discussion",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Performance on Span Length",
"publication_ref": [],
"table_ref": [],
"text": "We seek to determine whether the difference in performance between the ESCOXLM-R and XLM-R large models is due to shorter spans, and to what extent. One application of predicting short spans well is the rise of technologies, for which the names are usually short in length. Zhang et al. (2022c) In this table, the performance of two systems, XLM-R large and ESCOXLM-R, was measured using entity-level and surface-level span-F1 scores. Entity-level span-F1 measures precision, recall, and harmonic mean at the entity level, while surface-level span-F1 measures a system's ability to recognize a range of entities. We include the ratio of surface entities to total entities in each training set, with a higher ratio indicating more variety (a ratio of 1.00 indicates all entities are unique).\nboth models on the test sets of each dataset, where span-F1 is used as measurement. We group gold spans into buckets of lengths 1-2, 3-4, 5-6, 7-8, and 9-10, and present the span-F1 for each model (XLM-R large vs. ESCOXLM-R) in each bucket. Shown in Figure 4, ESCOXLM-R outperforms XLM-R large on shorter spans (i.e., 1-2 or 3-4) in 6 out of the 6 datasets, suggesting that pre-training on ESCO is beneficial for predicting short spans. However, there is a slight decline in performance on some datasets (e.g., SKILLSPAN, JOBSTACK, and GNEHM) when the spans are longer (i.e., 7-8 or 9-10). It is worth noting that the number of instances in these longer span buckets is lower, and therefore errors may be less apparent in terms of their impact on overall performance."
},
{
"figure_ref": [],
"heading": "Entity-F1 vs. Surface-F1",
"publication_ref": [
"b15"
],
"table_ref": [
"tab_8"
],
"text": "In this analysis, we adopt the evaluation method used in the W-NUT shared task on Novel and Emerging Entity Recognition (Derczynski et al., 2017). In this shared task, systems are evaluated using two measures: entity span-F1 and surface span-F1. Entity span-F1 assesses the precision, recall, and harmonic mean (F1) of the systems at the entity level, while surface span-F1 assesses their ability to correctly recognize a diverse range of entities, rather than just the most frequent surface forms. This means surface span-F1 counts entity types, in contrast to entity tokens in the standard entity span-F1 metric.\nAs shown in Table 3, we first calculate the ratio of unique entities and total entities in each relevant train set (i.e., datasets where we do span labeling). A higher ratio number indicates a wider variety of spans. Both XLM-R large and ESCOXLM-R tend to have lower performance when variety gets high (above 0.75). In addition, there are 2 datasets (SAYFULLINA, JOBSTACK) where we see a low variety of spans and large discrepancy between performance of entity span-F1 and surface span-F1. This difference is lower for ESCOXLM-R (especially in SAYFULLINA) suggesting that pre-training on ESCO helps predicting uncommon entities.\nIt is also noteworthy that the standard deviations for the scores at the entity span-F1 are generally lower than those for the surface span-F1. This suggests that the results for the entity span-F1 scores are more consistent across different runs, likely due to recognizing common entities more.\nOverall, ESCOXLM-R consistently outperforms XLM-R large in both the entity-level and surface-level F1 scores, indicating the benefits of using the ESCO dataset for pre-training on JAD tasks."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b46",
"b67",
"b62"
],
"table_ref": [],
"text": "To the best of our knowledge, we are the first to internalize an LM with ESCO for job-related NLP tasks. There are, however, several works that integrate factual knowledge (i.e., knowledge graphs/bases) into an LM. Peters et al. (2019) integrates multiple knowledge bases into LMs to enhance their representations with structured, human-curated knowledge and improve perplexity, fact recall and downstream performance on various tasks. Zhang et al. (2019) Yasunaga et al. (2022) introduced LinkBERT which leverages links between documents, such as hyperlinks, to capture dependencies and knowledge that span across documents by placing linked documents in the same context and pre-training the LM with MLM and document relation prediction."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "In this study, we introduce ESCOXLM-R as a multilingual, domain-adapted LM that has been further pre-trained on the ESCO taxonomy. We evaluated ESCOXLM-R, to the best of our knowledge, on the broadest evaluation set in this domain on 4 different langauges. The results showed that ESCOXLM-R outperformed XLM-R large on job-related downstream tasks in 6 out of 9 datasets, particularly when the task was relevant to the ESCO taxonomy and context was important. It was found that the improvement of ESCOXLM-R was mainly due to its performance on shorter span lengths, demonstrating the value of pre-training on the ESCO dataset. ESCOXLM-R also demonstrated improved performance on both frequent surface spans and a wider range of spans. Overall, this work showed the potential of ESCOXLM-R as an LM for multilingual job-related tasks. We hope that it will encourage further research in this area."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [],
"table_ref": [],
"text": "There are several limitations to this study that should be considered. First, a key limitation is the lack of a variety of language-specific JAD. Here, we have four different languages namely EN, DA, FR, and DE. This means that our analysis is based on a limited subset of languages and may not be representative of JAD data outside of these four languages.\nIn turn, the second limitation is that the ESCO taxonomy used as pre-training data only covers Europe and the datasets used in this work also covers mostly Europe. The results may not be generalizable to other regions. However, we see a slight improvement in the BHOLA dataset, the data of which comes from Singapore, which hints that it could generalize to other cultures.\nThe ESCO relation prediction task aims for learning the relations between elements of the ESCO taxonomy. We acknowledge that we do not evaluate the effectiveness of the pre-training objective in relation-centered tasks. Unfortunately, to the best of our knowledge, there is no job-related dataset containing relations between skill/occupation concepts to benchmark our model on. We consider this interesting future work. Finally, we did not conduct an ablation study on the ERP pre-training objective, i.e., which errors it makes. As the accuracy of the objective is 60%, we are unable to determine which sampling method is detrimental to this accuracy. However, we suspect that the Linked sampling approach might be the hardest to predict correctly. For example, many occupations have a lot of necessary and optional skills, thus it is harder to determine if some skill truly belongs to a specific occupation. Nevertheless, we see that adding the ERP objective improves over regular MLM domain-adaptive pre-training.\nDespite these limitations, we believe that this study provides valuable resources and insights into the use of ESCOXLM-R for analyzing JAD and suggests directions for future research. Future studies could address the limitations of this study by using a larger, more diverse datasets and by conducting ablation studies on the language model to better understand which parts contribute to the results."
},
{
"figure_ref": [],
"heading": "Ethics Statement",
"publication_ref": [
"b44",
"b47",
"b16",
"b34",
"b48",
"b59",
"b55",
"b1"
],
"table_ref": [],
"text": "We also see a potential lack of language inclusiveness within our work, as we addressed in the Limitation section that ESCO mostly covers Europe (and the Arabic language). Nevertheless, we see ESCOXLM-R as a step towards inclusiveness, due to JAD frequently being English-only. In addition, to the best of our knowledge, ESCO itself is devoid of any gendered language, specifically, pronouns and other gender-specific terms in, e.g., occupations. However, we acknowledge that LMs such as ESCOXLM-R could potentially be exploited in the process of hiring candidates for a specific job with unintended consequences (unconscious bias and dual use). There exists active research on fairer recommender systems (e.g., bias mitigation) for human resources (e.g., Mujtaba and Mahapatra, 2019;Raghavan et al., 2020;Deshpande et al., 2020;Köchling and Wehner, 2020;Sánchez-Monedero et al., 2020;Wilson et al., 2021;van Els et al., 2022;Arafan et al., 2022). "
},
{
"figure_ref": [],
"heading": "C Fine-tuning Details",
"publication_ref": [
"b54"
],
"table_ref": [],
"text": "For fine-tuning XLM-R large (+ DAPT) and ESCOXLM-R on the downstream tasks, we use MaChAmp (van der Goot et al., 2021). For more details we refer to their paper. We always include the original learning rate, batch size, maximum sequence length, and epochs from the respective downstream tasks in our search space (whenever applicable). Each model is trained on an NVIDIA A100 GPU with 40GBs of VRAM and an AMD Epyc 7662 CPU. The seed numbers the models are initialized with are 276800, 381552, 497646, 624189, 884832. We run all models with the maximun number of epochs indicated in Table 5 and select the best-performing one based on validation set performance in the downstream metric."
},
{
"figure_ref": [],
"heading": "Learning rate",
"publication_ref": [],
"table_ref": [],
"text": "Batch size max_seq_length Epochs SKILLSPAN {1e -4 , 5e -5 , 1e -5 5e -6 } {16, 32, 64} 128 20 KOMPETENCER {1e -4 , 7e -5 , 5e -5 , 1e -5 , 5e -6 } {8, 16, 32} 128 20 BHOLA {1e -4 , 7e -5 , 5e -5 , 1e -5 , 5e -6 } {4, 16, 32, 64, 128} {128, 256} 10 SAYFULLINA {1e -4 , 5e -5 , 1e -5 } {16, 32, 64} 128 10 GREEN {1e -4 , 5e -5 , 1e -5 } {16, 32, 64} 128 10 JOBSTACK {1e -4 , 7e -5 , 5e -5 , 1e -5 , 5e -6 } {16, 32, 64, 128} 128 20 GNEHM {1e -4 , 5e -5 , 1e -5 } {16, 32, 64} 128 5 FIJO {1e -4 , 5e -5 , 1e -5 } {8, 16, 32, 64} 128 10\nTable 5: Hyperparameter Sweep for Fine-tuning. We show a hyperparameter sweep for fine-tuning all models.\nLearning rate differs for both XLM-R large and ESCOXLM-R, where XLM-R large performs best on lower learning rate (e.g., 1e -5 ) and ESCOXLM-R on a bit of a higher learning rate (e.g., 5e -5 ). A batch size of 32 works best for all models. The max sequence length is usually the same, except for BHOLA due to it containing long texts. Epochs are determined based on previous work (i.e., the relevant datasets)."
},
{
"figure_ref": [],
"heading": "Acknowledgements",
"publication_ref": [],
"table_ref": [],
"text": "We thank both the NLPnorth and MaiNLP group for feedback on an earlier version of this paper. This research is supported by the Independent Research Fund Denmark (DFF) grant 9131-00019B and in parts by ERC Consolidator Grant DIALECT 101043235."
}
] | The increasing number of benchmarks for Natural Language Processing (NLP) tasks in the computational job market domain highlights the demand for methods that can handle job-related tasks such as skill extraction, skill classification, job title classification, and de-identification. While some approaches have been developed that are specific to the job market domain, there is a lack of generalized, multilingual models and benchmarks for these tasks. In this study, we introduce a language model called ESCOXLM-R, based on XLM-R large , which uses domain-adaptive pre-training on the European Skills, Competences, Qualifications and Occupations (ESCO) taxonomy, covering 27 languages. The pre-training objectives for ESCOXLM-R include dynamic masked language modeling and a novel additional objective for inducing multilingual taxonomical ESCO relations. We comprehensively evaluate the performance of ESCOXLM-R on 6 sequence labeling and 3 classification tasks in 4 languages and find that it achieves state-of-the-art results on 6 out of 9 datasets. Our analysis reveals that ESCOXLM-R performs better on short spans and outperforms XLM-R large on entity-level and surface-level span-F1, likely due to ESCO containing short skill and occupation titles, and encoding information on the entity-level. | ESCOXLM-R: Multilingual Taxonomy-driven Pre-training for the Job Market Domain | [
{
"figure_caption": "Figure 1 :1Figure 1: ESCO Pre-training Objective: From left to right, the figure illustrates the hierarchical structure of the ESCO taxonomy, which consists of occupations, skills, and aliases (OSA). Each OSA includes a definition. For the purposes of this study, we consider aliases of occupations to have the same definition as the occupation itself.In the middle of the figure, we show our pre-training setup. Pre-training instances are uniformly sampled in three ways: randomly, linked, or grouped (this is defined in Section 2.2). The selected instances (can be in different languages) are then fed to the language model, along with its description. We have two pre-training objectives: the regular MLM objective, and a new ESCO relation prediction objective, in which the goal is to predict which group the sampled instances belong to (Random, Linked, or Grouped).",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "In this work, we present and release the following: • ESCOXLM-R, an XLM-R large -based model, which utilizes domain-adaptive pre-training on the 27 languages from ESCO. 1",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: Statistics of Pre-training Data.The ESCO dataset contains descriptions in 27 languages, with a combined total of approximately 3.72 million descriptions (i.e., instances). On average, there are around 130,000 descriptions per language. The average length of each description is 26.3 tokens, with some descriptions reaching a maximum length of 150 or more tokens, as shown by the outliers in the boxplot.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "improve our XLM-R large -based model, we employ domain-adaptive pre-training techniques as described in previous work such as Alsentzer et al. (2019); Han and Eisenstein (2019); Lee et al. (2020); Gururangan et al. (2020); Nguyen et al.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "; He et al. (2020); Wang et al. (2021b) combine LM training with knowledge graph embeddings. Wang et al. (2021a) introduces K-Adapter for injecting knowledge into pre-trained models that adds neural adapters for each kind of knowledge domain. Yu et al. (2022) introduces Dict-BERT, which incorporates defi-nitions of rare or infrequent words into the input sequence and further pre-trains a BERT model. Calixto et al. (2021) introduced a multilingual Wikipedia hyperlink prediction intermediate task to improve language model pre-training. Similarly,",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": ": \"use physiotherapy for treatment of animals\", 27 \"description\": \"Adapt human physical therapy [...]\" Example Extraction. An example of the information that is given for ESCO code 2250.4: animal therapist. The original page can be found here: http://data.europa.eu/esco/occupation/ 0b2d3242-22a3-4de5-bd29-efd39cdf2c31.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Data Example Gnehm. ",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "due to careful tuning, sampling, and scaling to larger amounts of textual data. Because of this, our ESCOXLM-R model is based on XLM-R large .",
"figure_data": "Number of Instances150Count (x1000)50 75 100 125250Description Length250Length in Tokens50 100 150 2000ar bg cs da de el en es et fi fr ga hr hu is it lt lv mt nl no pl pt ro sk sl sv2.1 European Skills, Competences,Qualifications and OccupationsTaxonomyThe European Skills, Competences, Qualifications,and Occupations (ESCO; le Vrang et al., 2014) tax-onomy is a standardized system for describing andcategorizing the skills, competences, qualifications,",
"figure_id": "tab_1",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "[CLS] . In our implementation, we use XLM-R large and classify the start-token [CLS] for ERP to improve the model's Dataset Statistics. We show statistics for all 9 JAD datasets. There are 6 datasets in English and 3 in other languages (Danish, German, and French). We indicate the location the JAD originates from (whenever applicable, * indicates it comes from a variety of countries). We indicate the license of the dataset. Most of the task types consist of sequence labeling (e.g., span extraction, Named Entity Recognition, soft skill tagging). To maintain consistency, we use a single metric for each task type: Sequence Labeling (SL), Multilabel Classification (MLC), and Multiclass Classification (MCC). For KOMPETENCER, the statistics are provided in brackets for the Danish language.",
"figure_data": "Dataset NameLang. Loc. LicenseTaskMetricInput TypeTrainDev.TestSKILLSPANen*CC-BY-4.0SLSpan-F1Sentences5,866 3,992 4,680SAYFULLINAenUKUnknownSLSpan-F1Sentences3,706 1,854 1,853GREENenUKCC-BY-4.0SLSpan-F1Sentences8,670963336JOBSTACKen*RLTSLSpan-F1Sentences18,055 2,082 2,092BHOLAenSGCC-BY-4.0MLC MRRDocuments 16,238 2,030 2,030KOMPETENCER enDKCC-BY-4.0MCC W. Macro-F1 Skills9,472 1,577 1,578KOMPETENCER daDKCC-BY-4.0MCC W. Macro-F1 Skills138-784GNEHMdeCHCC-BY-NC-SA-4.0 SLSpan-F1Sentences22,134 2,679 2,943FIJOfrFRUnknownSLSpan-F1Sentences3995050",
"figure_id": "tab_3",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Results of Experiments. The datasets and models are described in Section 3. We re-train the bestperforming models of all papers to give us the standard deviation. The best-performing model is in bold. The difference in performance between ESCOXLM-R and the previous SOTA is shown as ∆. Note (*) that the results for GREEN are based on a CRF model where the data has been pre-split, and therefore, there is no standard deviation.",
"figure_data": "DatasetLang. MetricPrev. SOTA XLM-R large XLM-R large (+ DAPT) ESCOXLM-R∆SKILLSPANENSpan-F158.9±4.559.7±4.662.0±4.0 62.6±3.7+3.7SAYFULLINAENSpan-F173.1±2.189.9±0.590.6±0.4 92.2±0.2+19.1GREENENSpan-F131.8±*49.0±2.447.5±0.7 51.2±2.1+19.4JOBSTACKENSpan-F182.1±0.881.2±0.680.4±0.7 82.0±0.7-0.1KOMPETENCER ENW. Macro-F1 62.8±2.859.0±9.564.3±0.5 63.5±1.3-0.7BHOLAENMRR90.2±0.290.5±0.390.0±0.3 90.7±0.2+0.5GNEHMDESpan-F186.7±0.487.1±0.486.8±0.2 88.4±0.5+1.7FIJOFRSpan-F131.7±2.341.8±2.041.7±0.7 42.0±2.3+10.3KOMPETENCER DAW. Macro-F1 45.3±1.541.2±9.845.6±0.8 45.0±1.4-0.3",
"figure_id": "tab_5",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Green et al., 2022) A sentence-level sequence labeling task involving labeling skills, qualifications, job domain, experience, and occupation labels. The job positions in the dataset are from the United Kingdom. The industries represented in the data vary and include IT, finance, healthcare, and sales. Their model for this task is a Conditional Random Field(Lafferty et al., 2001) model.",
"figure_data": "JOBSTACK (Jensen et al., 2021) This corpusis used for de-identifying personal data in jobvacancies on Stack Overflow. The task involvessequence labeling and predicting Organization,Location, Name, Profession, and Contact detailslabels. The best-performing model for this task isa transformer-based (Vaswani et al., 2017) modeltrained in a multi-task learning setting. Jensen et al.(2021) propose to use the I2B2/UTHealth corpus,which is a medical de-identification task (Stubbsand Uzuner, 2015), as auxiliary data, whichshowed improvement over their baselines.GNEHM (Gnehm et al., 2022) A Swiss-Germanjob ad dataset where the task is Information andCommunications Technology (ICT)-related entityrecognition, these could be ICT tasks, technologystack, responsibilities, and so forth. The useddataset is a combination of two other Swissdatasets namely the Swiss Job Market Monitor andan online job ad dataset (Gnehm and Clematide,2020; Buchmann et al., 2022). Their model isdubbed JobGBERT and is based on DAPT withGerman BERT base",
"figure_id": "tab_6",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Entity vs. Surface-level span-F1 on Test.",
"figure_data": "observes",
"figure_id": "tab_8",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Data example references for each dataset.",
"figure_data": "1Experience OO1abilityO2inOO2toO3workingB-SkillO3workB-Skill4onI-SkillO4underI-Skill5aI-SkillO5stressI-Skill6cloud-based I-SkillO6condition O7application I-SkillO78runningOO8dueO9onOO9toO10DockerOB-Knowledge10theO11.OO11dynamicB-Skill1212natureO13AOO13ofO14degreeOB-Knowledge14theO15inOI-Knowledge15groupO16ComputerOI-Knowledge16environment O17ScienceOI-Knowledge17,O18orOO18theO19relatedOO19idealO20fieldsOO20candidate O21.OO21willOListing 2: Data Example SkillSpan.B Data ExamplesSKILLSPANListing 2SAYFULLINAListing 3GREENListing 4BHOLAListing 5KOMPETENCER Listing 6FIJOListing 7GNEHMListing 8",
"figure_id": "tab_9",
"figure_label": "4",
"figure_type": "table"
}
] | Mike Zhang; Rob Van Der Goot; Barbara Plank | [
{
"authors": "Emily Alsentzer; John Murphy; William Boag; Wei-Hung Weng; Di Jindi; Tristan Naumann; Matthew Mcdermott",
"journal": "Association for Computational Linguistics",
"ref_id": "b0",
"title": "Publicly available clinical BERT embeddings",
"year": "2019"
},
{
"authors": "Adam Mehdi Arafan; David Graus; Fernando P Santos; Emma Beauxis-Aussalet",
"journal": "",
"ref_id": "b1",
"title": "End-to-end bias mitigation in candidate recommender systems with fairness gates",
"year": "2022"
},
{
"authors": "Krisztian Balog; Yi Fang; Maarten De Rijke; Pavel Serdyukov; Luo Si",
"journal": "Foundations and Trends in Information Retrieval",
"ref_id": "b2",
"title": "Expertise retrieval",
"year": "2012"
},
{
"authors": "David Beauchemin; Julien Laumonier; Yvan Le Ster; Marouane Yassine",
"journal": "",
"ref_id": "b3",
"title": "FIJO\": a french insurance soft skill detection dataset",
"year": "2022"
},
{
"authors": "Akshay Bhola; Kishaloy Halder; Animesh Prasad; Min-Yen Kan",
"journal": "International Committee on Computational Linguistics",
"ref_id": "b4",
"title": "Retrieving skills from job descriptions: A language model based extreme multi-label classification framework",
"year": "2020"
},
{
"authors": "Erik Brynjolfsson; Andrew Mcafee",
"journal": "Brynjolfsson and McAfee",
"ref_id": "b5",
"title": "Race against the machine: How the digital revolution is accelerating innovation, driving productivity, and irreversibly transforming employment and the economy",
"year": "2011"
},
{
"authors": "Erik Brynjolfsson; Andrew Mcafee",
"journal": "WW Norton & Company",
"ref_id": "b6",
"title": "The second machine age: Work, progress, and prosperity in a time of brilliant technologies",
"year": "2014"
},
{
"authors": "Marlis Buchmann; Helen Buchs; Felix Busch; Simon Clematide; Ann-Sophie Gnehm; Jan Müller",
"journal": "European Sociological Review",
"ref_id": "b7",
"title": "Swiss job market monitor: A rich source of demand-side micro data of the labour market",
"year": "2022"
},
{
"authors": "Iacer Calixto; Alessandro Raganato; Tommaso Pasini",
"journal": "",
"ref_id": "b8",
"title": "Wikipedia entities as rendezvous across languages: Grounding multilingual language models by predicting wikipedia hyperlinks",
"year": "2021"
},
{
"authors": "Branden Chan; Stefan Schweter; Timo Möller",
"journal": "International Committee on Computational Linguistics",
"ref_id": "b9",
"title": "German's next language model",
"year": "2020"
},
{
"authors": "Mariia Chernova",
"journal": "",
"ref_id": "b10",
"title": "Occupational skills extraction with FinBERT",
"year": "2020"
},
{
"authors": "Chung Hyung Won; Thibault Févry; Henry Tsai; Melvin Johnson; Sebastian Ruder",
"journal": "",
"ref_id": "b11",
"title": "Rethinking embedding coupling in pre-trained language models",
"year": "2021-05-03"
},
{
"authors": "Alexis Conneau; Kartikay Khandelwal; Naman Goyal; Vishrav Chaudhary; Guillaume Wenzek; Francisco Guzmán; Edouard Grave; Myle Ott; Luke Zettlemoyer; Veselin Stoyanov",
"journal": "Association for Computational Linguistics",
"ref_id": "b12",
"title": "Unsupervised cross-lingual representation learning at scale",
"year": "2020"
},
{
"authors": "Jens-Joris Decorte; Jeroen Van Hautte; Johannes Deleu; Chris Develder; Thomas Demeester",
"journal": "",
"ref_id": "b13",
"title": "Design of negative sampling strategies for distantly supervised skill extraction",
"year": "2022"
},
{
"authors": "Jens-Joris Decorte; Jeroen Van Hautte; Thomas Demeester; Chris Develder",
"journal": "",
"ref_id": "b14",
"title": "Jobbert: Understanding job titles through skills",
"year": "2021"
},
{
"authors": "Leon Derczynski; Eric Nichols; Marieke Van Erp; Nut Limsopatham",
"journal": "Association for Computational Linguistics",
"ref_id": "b15",
"title": "Results of the WNUT2017 shared task on novel and emerging entity recognition",
"year": "2017"
},
{
"authors": "Shimei Ketki V Deshpande; James R Pan; Foulds",
"journal": "",
"ref_id": "b16",
"title": "Mitigating demographic bias in ai-based resume filtering",
"year": "2020"
},
{
"authors": "Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova",
"journal": "Association for Computational Linguistics",
"ref_id": "b17",
"title": "BERT: Pre-training of deep bidirectional transformers for language understanding",
"year": "2019"
},
{
"authors": " Esco",
"journal": "",
"ref_id": "b18",
"title": "Machine Learning Assisted Mapping of Multilingual Occupational Data to ESCO",
"year": "2022"
},
{
"authors": "Ann-Sophie Gnehm; Eva Bühlmann; Simon Clematide",
"journal": "European Language Resources Association",
"ref_id": "b19",
"title": "Evaluation of transfer learning and domain adaptation for analyzing germanspeaking job advertisements",
"year": "2022"
},
{
"authors": "Ann-Sophie Gnehm; Simon Clematide",
"journal": "Association for Computational Linguistics",
"ref_id": "b20",
"title": "Text zoning and classification for job advertisements in German, French and English",
"year": "2020"
},
{
"authors": "Nidhi Goyal; Jushaan Kalra; Charu Sharma; Raghava Mutharaju; Niharika Sachdeva; Ponnurangam Kumaraguru",
"journal": "Association for Computational Linguistics",
"ref_id": "b21",
"title": "JobXMLC: EXtreme multilabel classification of job skills with graph neural networks",
"year": "2023"
},
{
"authors": "Thomas Green; Diana Maynard; Chenghua Lin",
"journal": "European Language Resources Association",
"ref_id": "b22",
"title": "Development of a benchmark corpus to support entity recognition in job descriptions",
"year": "2022"
},
{
"authors": "Suchin Gururangan; Ana Marasović; Swabha Swayamdipta; Kyle Lo; Iz Beltagy; Doug Downey; Noah A Smith",
"journal": "Association for Computational Linguistics",
"ref_id": "b23",
"title": "Don't stop pretraining: Adapt language models to domains and tasks",
"year": "2020"
},
{
"authors": "Xiaochuang Han; Jacob Eisenstein",
"journal": "Association for Computational Linguistics",
"ref_id": "b24",
"title": "Unsupervised domain adaptation of contextualized embeddings for sequence labeling",
"year": "2019"
},
{
"authors": "Bin He; Di Zhou; Jinghui Xiao; Xin Jiang; Qun Liu; Nicholas Jing Yuan; Tong Xu",
"journal": "Association for Computational Linguistics",
"ref_id": "b25",
"title": "BERT-MK: Integrating graph contextualized knowledge into pre-trained language models",
"year": "2020"
},
{
"authors": "Sepp Hochreiter; Jürgen Schmidhuber; Corso Elvezia",
"journal": "Neural Computation",
"ref_id": "b26",
"title": "Long short-term memory",
"year": "1997"
},
{
"authors": "Junjie Hu; Sebastian Ruder; Aditya Siddhant; Graham Neubig; Orhan Firat; Melvin Johnson",
"journal": "",
"ref_id": "b27",
"title": "Xtreme: A massively multilingual multitask benchmark for evaluating cross-lingual generalisation",
"year": "2020"
},
{
"authors": " Pmlr",
"journal": "",
"ref_id": "b28",
"title": "",
"year": ""
},
{
"authors": "Faizan Javed; Qinlong Luo; Matt Mcnair; Ferosh Jacob; Meng Zhao; Tae Seung Kang",
"journal": "IEEE",
"ref_id": "b29",
"title": "Carotene: A job title classification system for the online recruitment domain",
"year": "2015"
},
{
"authors": "Faizan Javed; Matt Mcnair; Ferosh Jacob; Meng Zhao",
"journal": "",
"ref_id": "b30",
"title": "Towards a job title classification system",
"year": "2016"
},
{
"authors": "Kristian Nørgaard Jensen; Mike Zhang; Barbara Plank",
"journal": "Linköping University Electronic Press",
"ref_id": "b31",
"title": "De-identification of privacy-related entities in job postings",
"year": "2021"
},
{
"authors": "Yoon Kim",
"journal": "Association for Computational Linguistics",
"ref_id": "b32",
"title": "Convolutional neural networks for sentence classification",
"year": "2014"
},
{
"authors": "Ilkka Kivimäki; Alexander Panchenko; Adrien Dessy; Dries Verdegem; Pascal Francq; Hugues Bersini; Marco Saerens",
"journal": "Association for Computational Linguistics",
"ref_id": "b33",
"title": "A graph-based approach to skill extraction from text",
"year": "2013"
},
{
"authors": "Alina Köchling; Marius Claus Wehner",
"journal": "Business Research",
"ref_id": "b34",
"title": "Discriminated by an algorithm: a systematic review of discrimination and fairness by algorithmic decisionmaking in the context of hr recruitment and hr development",
"year": "2020"
},
{
"authors": "John D Lafferty; Andrew Mccallum; Fernando C N Pereira",
"journal": "Williams College",
"ref_id": "b35",
"title": "Conditional random fields: Probabilistic models for segmenting and labeling sequence data",
"year": "2001-06-28"
},
{
"authors": "Morgan Kaufmann",
"journal": "",
"ref_id": "b36",
"title": "",
"year": ""
},
{
"authors": "Anne Lauscher; Vinit Ravishankar; Ivan Vulić; Goran Glavaš",
"journal": "",
"ref_id": "b37",
"title": "From zero to hero: On the limitations of zero-shot language transfer with multilingual transformers",
"year": "2020"
},
{
"authors": "Agis Martin Le Vrang; Erika Papantoniou; Pieter Pauwels; Dominique Fannes; Johan De Vandensteen; Smedt",
"journal": "Computer",
"ref_id": "b38",
"title": "Esco: Boosting job matching in europe with semantic interoperability",
"year": "2014"
},
{
"authors": "Jinhyuk Lee; Wonjin Yoon; Sungdong Kim; Donghyeon Kim; Sunkyu Kim; Chan Ho; So ; Jaewoo Kang",
"journal": "Bioinformatics",
"ref_id": "b39",
"title": "Biobert: a pre-trained biomedical language representation model for biomedical text mining",
"year": "2020"
},
{
"authors": "Jingzhou Liu; Wei-Cheng Chang; Yuexin Wu; Yiming Yang",
"journal": "ACM",
"ref_id": "b40",
"title": "Deep learning for extreme multilabel text classification",
"year": "2017-08-07"
},
{
"authors": "Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov",
"journal": "",
"ref_id": "b41",
"title": "Roberta: A robustly optimized bert pretraining approach",
"year": "2019"
},
{
"authors": "Ilya Loshchilov; Frank Hutter",
"journal": "",
"ref_id": "b42",
"title": "Decoupled weight decay regularization",
"year": "2019-05-06"
},
{
"authors": "Louis Martin; Benjamin Muller; Pedro ; Javier Ortiz Suárez; Yoann Dupont; Laurent Romary; Éric De La Clergerie; Djamé Seddah; Benoît Sagot",
"journal": "Association for Computational Linguistics",
"ref_id": "b43",
"title": "CamemBERT: a tasty French language model",
"year": "2020"
},
{
"authors": "F Dena; Mujtaba; Nihar R Mahapatra",
"journal": "IEEE",
"ref_id": "b44",
"title": "Ethical considerations in ai-based recruitment",
"year": "2019"
},
{
"authors": "Thanh Dat Quoc Nguyen; Anh Tuan Vu; Nguyen",
"journal": "Association for Computational Linguistics",
"ref_id": "b45",
"title": "BERTweet: A pre-trained language model for English tweets",
"year": "2020"
},
{
"authors": "Matthew E Peters; Mark Neumann; Robert Logan; Roy Schwartz; Vidur Joshi; Sameer Singh; Noah A Smith",
"journal": "Association for Computational Linguistics",
"ref_id": "b46",
"title": "Knowledge enhanced contextual word representations",
"year": "2019"
},
{
"authors": "Manish Raghavan; Solon Barocas; Jon Kleinberg; Karen Levy",
"journal": "",
"ref_id": "b47",
"title": "Mitigating bias in algorithmic hiring: Evaluating claims and practices",
"year": "2020"
},
{
"authors": "Javier Sánchez-Monedero; Lina Dencik; Lilian Edwards",
"journal": "",
"ref_id": "b48",
"title": "What does it mean to'solve'the problem of discrimination in hiring? social, technical and legal perspectives from the uk on automated hiring systems",
"year": "2020"
},
{
"authors": "Luiza Sayfullina; Eric Malmi; Juho Kannala",
"journal": "",
"ref_id": "b49",
"title": "Learning representations for soft skill matching",
"year": "2018"
},
{
"authors": "Baoxu Shi; Jaewon Yang; Feng Guo; Qi He",
"journal": "ACM",
"ref_id": "b50",
"title": "Salience and market-aware skill extraction for job targeting",
"year": "2020-08-23"
},
{
"authors": "Ellery Smith; Martin Braschler; Andreas Weiler; Thomas Haberthuer",
"journal": "IEEE",
"ref_id": "b51",
"title": "Syntax-based skill extractor for job advertisements",
"year": "2019"
},
{
"authors": "Amber Stubbs; Özlem Uzuner",
"journal": "Journal of biomedical informatics",
"ref_id": "b52",
"title": "Annotating longitudinal clinical narratives for de-identification: The 2014 i2b2/uthealth corpus",
"year": "2015"
},
{
"authors": "Damian A Tamburri; Willem-Jan Van Den; Martin Heuvel; Garriga",
"journal": "IEEE",
"ref_id": "b53",
"title": "Dataops for societal intelligence: a data pipeline for labor market skills extraction and matching",
"year": "2020"
},
{
"authors": "Rob Van Der Goot; Ahmet Üstün; Alan Ramponi; Ibrahim Sharaf; Barbara Plank",
"journal": "Association for Computational Linguistics",
"ref_id": "b54",
"title": "Massive choice, ample tasks (MaChAmp): A toolkit for multi-task learning in NLP",
"year": "2021"
},
{
"authors": "Sarah-Jane Van Els; David Graus; Emma Beauxis-Aussalet",
"journal": "",
"ref_id": "b55",
"title": "Improving fairness assessments with synthetic data: a practical use case with a recommender system for human resources",
"year": "2022"
},
{
"authors": "Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Lukasz Kaiser; Illia Polosukhin",
"journal": "",
"ref_id": "b56",
"title": "Attention is all you need",
"year": "2017-09"
},
{
"authors": "Ruize Wang; Duyu Tang; Nan Duan; Zhongyu Wei; Xuan-Jing Huang; Jianshu Ji; Guihong Cao; Daxin Jiang; Ming Zhou; ; ",
"journal": "",
"ref_id": "b57",
"title": "K-adapter: Infusing knowledge into pre-trained models with adapters",
"year": "2021"
},
{
"authors": "Xiaozhi Wang; Tianyu Gao; Zhaocheng Zhu; Zhengyan Zhang; Zhiyuan Liu; Juanzi Li; Jian Tang",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b58",
"title": "KEPLER: A unified model for knowledge embedding and pre-trained language representation",
"year": "2021"
},
{
"authors": "Christo Wilson; Avijit Ghosh; Shan Jiang; Alan Mislove; Lewis Baker; Janelle Szary; Kelly Trindel; Frida Polli",
"journal": "",
"ref_id": "b59",
"title": "Building and auditing fair algorithms: A case study in candidate screening",
"year": "2021"
},
{
"authors": "Thomas Wolf; Lysandre Debut; Victor Sanh; Julien Chaumond; Clement Delangue; Pierric Moi; Tim Cistac; Remi Rault; Morgan Louf; Joe Funtowicz; Sam Davison; Shleifer; Clara Patrick Von Platen; Yacine Ma; Julien Jernite; Canwen Plu; Teven Xu; Sylvain Le Scao; Mariama Gugger; Quentin Drame; Alexander Lhoest; Rush",
"journal": "Association for Computational Linguistics",
"ref_id": "b60",
"title": "Transformers: State-of-the-art natural language processing",
"year": "2020"
},
{
"authors": "Zichao Yang; Diyi Yang; Chris Dyer; Xiaodong He; Alex Smola; Eduard Hovy",
"journal": "Association for Computational Linguistics",
"ref_id": "b61",
"title": "Hierarchical attention networks for document classification",
"year": "2016"
},
{
"authors": "Michihiro Yasunaga; Jure Leskovec; Percy Liang",
"journal": "Association for Computational Linguistics",
"ref_id": "b62",
"title": "LinkBERT: Pretraining language models with document links",
"year": "2022"
},
{
"authors": "Wenhao Yu; Chenguang Zhu; Yuwei Fang; Donghan Yu; Shuohang Wang; Yichong Xu; Michael Zeng; Meng Jiang",
"journal": "Association for Computational Linguistics",
"ref_id": "b63",
"title": "Dict-BERT: Enhancing language model pre-training with dictionary",
"year": "2022"
},
{
"authors": "Mike Zhang; Kristian Jensen; Sif Sonniks; Barbara Plank",
"journal": "Association for Computational Linguistics",
"ref_id": "b64",
"title": "SkillSpan: Hard and soft skill extraction from English job postings",
"year": "2022"
},
{
"authors": "Mike Zhang; Kristian Nørgaard Jensen; Barbara Plank",
"journal": "European Language Resources Association",
"ref_id": "b65",
"title": "Kompetencer: Fine-grained skill classification in danish job postings via distant supervision and transfer learning",
"year": "2022"
},
{
"authors": "Mike Zhang; Kristian Nørgaard Jensen; Rob Van Der Goot; Barbara Plank",
"journal": "",
"ref_id": "b66",
"title": "Skill extraction from job postings using weak supervision",
"year": "2022"
},
{
"authors": "Zhengyan Zhang; Xu Han; Zhiyuan Liu; Xin Jiang; Maosong Sun; Qun Liu",
"journal": "Association for Computational Linguistics",
"ref_id": "b67",
"title": "ERNIE: Enhanced language representation with informative entities",
"year": "2019"
},
{
"authors": "Meng Zhao; Faizan Javed; Ferosh Jacob; Matt Mcnair",
"journal": "AAAI Press",
"ref_id": "b68",
"title": "SKILL: A system for skill identification and normalization",
"year": "2015-01-25"
}
] | [
{
"formula_coordinates": [
4,
306.14,
114.29,
218.27,
25.1
],
"formula_id": "formula_0",
"formula_text": "[CLS] C A X A [SEP] C B X B [SEP]."
},
{
"formula_coordinates": [
4,
306.14,
425.04,
220.08,
25.1
],
"formula_id": "formula_1",
"formula_text": "[CLS] to- ken in [CLS] C A X A [SEP] C B X B [SEP] (r ∈"
},
{
"formula_coordinates": [
4,
314.13,
529.42,
202.3,
41.59
],
"formula_id": "formula_2",
"formula_text": "L = L MLM + L ERP = - i log p (x i | h i ) -log p (r | h [CLS] ) ,"
}
] |
2024-02-16 | [
{
"figure_ref": [
"fig_0"
],
"heading": "INTRODUCTION",
"publication_ref": [
"b33",
"b22",
"b61",
"b56",
"b2",
"b53",
"b59",
"b15",
"b29",
"b53",
"b11",
"b8",
"b0"
],
"table_ref": [],
"text": "Time series forecasting has emerged as a crucial task in various domains such as cloud computing, air quality forecasting, energy management, and traffic flow estimation (Qian et al., 2022;Liang et al., 2023;Zhu et al., 2023;Wen et al., 2023a). The rapid development of deep learning models has led to significant advancements in time series forecasting techniques, particularly in multivariate time series forecasting. Among various deep learning models developed for time series forecasting, RNN, CNN, MLP, transformer, and LLM-based models have demonstrated great performance thanks to their ability to capture complex long-term temporal dependencies (e.g., Zhou et al., 2021;Challu et al., 2022;Zeng et al., 2023;Zhou et al., 2022a;Wu et al., 2023b;Zhou et al., 2023;Jin et al., 2023).\nFor multivariate time series forecasting, a model is expected to yield a better performance by exploiting the dependence among different prediction variables, so-called channel-dependent (CD) methods. However, multiple recent works (e.g., Nie et al. 2023;Zeng et al. 2023) show that, in general, channel-independent (CI) forecasting models (i.e., all the time series variables are forecast independently) outperform the CD models. Analysis from (Han et al., 2023) indicates that CI forecasting models are more robust while CD models have higher modeling capacity. Given that time series forecasting usually involves high noise levels, typical transformer-based forecasting models with CD design can suffer from the issue of overfitting noises, leading to limited performance. These empirical studies and analyses raised an important question, i.e., how to build an effective transformer to utilize the cross-channel information for time series forecasting.\nIn this paper, we propose a Channel Aligned Robust Blend Transformer, or CARD for short, that effectively leverages the dependence among channels (i.e., forecasting variables) and alleviates the issue of overfitting noises in time series forecasting. Unlike typical transformers for time series analysis that only capture temporal dependency among signals through attention over tokens, the CARD also takes attention across different channels and hidden dimensions, which captures the correlation among prediction variables and aligns local information within each token. We observe that related approaches have been exploited in computer vision (Ding et al., 2022;Ali et al., 2021). Moreover, it is known that multi-scale information plays an important role in time series analysis. We design a token blend module to generate tokens with different resolutions. In particular, we propose to combine the adjacent tokens within the same head into the new token instead of merging the same position over different heads in multi-head attention. To improve the robustness and efficiency of the transformer for time series forecast, we further introduce an exponential smoothing layer over queries/keys tokens and a dynamic projection module when dealing with information among different channels. Finally, to alleviate the issue of overfitting noises, a robust loss function is introduced to weight each prediction by its uncertainty in the case of forecasting over a finite horizon. The overall model architecture is illustrated in Figure 1. We verify the effectiveness of the proposed model on various numerical benchmarks by comparing it to the state-of-the-art methods for Transformers and other models. Here we summarized our key contributions as follows:\n1. We propose a Channel Aligned Robust Blend Transformer (CARD) which efficiently and robustly aligns the information among different channels and utilizes the multi-scale information. 2. CARD demonstrates superior performance in several benchmark datasets for forecasting and other prediction-based tasks, outperforming the state-of-the-art models. Our studies have confirmed the effectiveness of the proposed model. 3. We develop a robust signal decay-based loss function that utilizes signal decay to bolster the model's ability to concentrate on forecasting for the near future. Our empirical assessment has confirmed that this loss function is effective in improving the performance of other benchmark models as well.\nThe remainder of this paper is structured as follows. In Section 2, we provide a summary of related works relevant to our study. Section 3 presents the proposed detailed model architecture. Section 4 describes the loss function design with a theoretical explanation via maximum likelihood estimation of Gaussian and Laplacian distributions. In Section 5, we demonstrate the results of the numerical experiments in forecasting benchmarks and conduct a comprehensive analysis to determine the effectiveness of the self-attention scheme for time series forecasting. Additionally, we discuss ablations and other experiments conducted in this study. Finally, in Section 6, the conclusions and future research directions are discussed."
},
{
"figure_ref": [],
"heading": "RELATED WORK",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "TRANSFORMERS FOR TIME SERIES FORECASTING",
"publication_ref": [
"b56",
"b49",
"b29"
],
"table_ref": [],
"text": "There is a large body of work that tries to apply Transformer models to forecast long-term time series in recent years (Wen et al., 2023b). We here summarize some of them. LogTrans (Li et al., 2019a) uses convolutional self-attention layers with LogSparse design to capture local information and reduce space complexity. Informer (Zhou et al., 2021) proposes a ProbSparse self-attention with distilling techniques to extract the most important keys efficiently. Autoformer (Wu et al., 2021) borrows the ideas of decomposition and auto-correlation from traditional time series analysis methods. FEDformer (Zhou et al., 2022b) uses Fourier enhanced structure to get a linear complexity. Pyraformer (Liu et al., 2022a) applies pyramidal attention module with inter-scale and intra-scale connections which also get a linear complexity. LogTrans avoids a point-wise dot product between the key and query, but its value is still based on a single time step. Autoformer uses autocorrelation to get patch-level connections, but it is a handcrafted design that doesn't include all the semantic information within a patch. A recent work PatchTST (Nie et al., 2023) studies using a vision transformer type model for long-term forecasting with channel independent design. The work closest to our proposed method is Crossformer (Zhang & Yan, 2023). This work designs an encoder-decoder model utilizing a hierarchy attention mechanism to leverage cross-dimension dependencies and achieves moderate performance in the same benchmark datasets that we use in this work. From the model architecture perspective, different from Crossformer, we employ an encoder-only structure, and the multi-scale information is induced via a lightweight token blend module instead of explicitly generating token hierarchies used in Crossformer. The designs significantly enhance the robustness of CARD and result in a substantial improvement in numerical performance."
},
{
"figure_ref": [],
"heading": "RNN, MLP AND CNN MODELS FOR TIME SERIES FORECASTING",
"publication_ref": [
"b18",
"b23",
"b37",
"b39",
"b46",
"b36",
"b10",
"b39",
"b28",
"b42",
"b31",
"b2",
"b21",
"b53",
"b6",
"b54",
"b46",
"b38"
],
"table_ref": [],
"text": "Besides transformers, other types of networks are also widely explored. For example, (Lai et al., 2018;Lim et al., 2021;Salinas et al., 2020;Smyl, 2020;Wen et al., 2017;Rangapuram et al., 2018;Zhou et al., 2022a;Gu et al., 2022) study the RNN/state-space models. In particular, (Smyl, 2020) considered equipping RNN with exponential smooth and, for the first time, beat the statistical models in forecasting tasks (Makridakis et al., 2018). (Chen et al., 2023;Oreshkin et al., 2020;Challu et al., 2022;Li et al., 2023;Zeng et al., 2023;Das et al., 2023;Zhang et al., 2022) explored MLP-type structures for time series forecasting. CNN models (e.g., Wu et al. 2023b;Wen et al. 2017;Sen et al. 2019) use the temporal convolution layer to extract the subsequence-level information. When dealing with multivariate forecasting tasks, the smoothness in adjacent covariates is assumed or the channel-independent strategy is used."
},
{
"figure_ref": [
"fig_0"
],
"heading": "MODEL ARCHITECTURE",
"publication_ref": [],
"table_ref": [],
"text": "The illustration of the architecture of CARD is shown in Figure 1. Let a t ∈ R C be the observation of time series at time t with channel C ≥ 1. Our objective is to use L recent historical data points (e.g., a t-L+1 , ..., a t ) to forecast the future T steps observations. (e.g., a t+1 , ..., a t+T ), where L, T ≥ 1."
},
{
"figure_ref": [],
"heading": "TOKENIZATION",
"publication_ref": [
"b29",
"b29",
"b23"
],
"table_ref": [],
"text": "We adopt the idea of patching (e.g., Nie et al. 2023;Zhang & Yan 2023) to convert the input time series into a token tensor. Let's denote A = [a t-L+1 , ..., a t ] ∈ R C×L as the input data matrix, S and P as stride and patch length respectively. We unfold the matrix A into the raw token tensor X ∈ R C×N ×P , where N = ⌊ L-P S + 1⌋. Here, we convert the time series into several P length segments, and each raw token maintains part of the sequence-level semantic information, which makes the attention scheme more efficient compared to the vanilla point-wise counterpart.\nWe then use a dense MLP layer F 1 : P → d, a extra token T 0 ∈ R C×d and positional embedding E ∈ R C×N ×d to generate the token matrix as follows:\nX = [T 0 , F 1 ( X) + E],(1)\nwhere X ∈ R C×(N +1)×d and d is the hidden dimension. Compared to (Nie et al., 2023) and (Zhang & Yan, 2023), our token construction introduces a extra T 0 token. The T 0 token is an analogy to the static covariate encoder in (Lim et al., 2021) and allows us to have a place to inject the features summarized the longer history of the series.\nWe consider generating Q, K and V via linear projection of the token tensor X:\nQ = F q (X), K = F k (X), V = F v (X),(2)\nwhere\nQ, K, V ∈ R C×(N +1)×d and F q , F k , F v are MLP layers. We next convert Q, K, V into {Q i },{K i },{V i } where Q i , K i , V i ∈ R C×(N +1)×d head , i =\n1, 2, ..., H. H and d head are number of heads and head dimension respectively. For each sample, the total number of tokens is C(N + 1). In order to fully utilize all cross-channel information, the ideal attention should be required O(C 2 (N + 1) 2 ) computation cost, which can be very timeconsuming and potentially can lead to easily over-fitting when training sample size is limited. In this paper, we consider paying attention alternately over each dimension instead."
},
{
"figure_ref": [],
"heading": "CARD ATTENTIONS OVER TOKENS",
"publication_ref": [
"b27",
"b48",
"b14",
"b5"
],
"table_ref": [],
"text": "When make attention over tokens, we slice the\nQ i , K i and V i on channel dimension into {Q c: i }, {K c: i } and {V c: i } with Q c: i , K c: i , V c: i ∈ R (N +1\n)×d head and c = 1, 2, ..., C. Besides the standard attention in tokens, we also introduce an extra attention structure in hidden dimensions that helps capture the local information within each patch. The attention in both tokens and hidden dimensions is computed as follows:\nA c: i1 = softmax 1 √ d • EMA(Q c: i ) (EMA(K c: i )) ⊤(3)\nA c: i2 = softmax 1 √ N • (Q c: i ) ⊤ K c: i ,(4)\nwhere By applying EMA on Q c: i and K c: i , each query token will be able to gain higher attention scores on more key tokens and thus the output becomes more robust. Similar techniques are also explored in (Ma et al., 2023) and (Woo et al., 2022). Different from those in the literature, we find that using a fixed EMA parameter that remains the same for all dimensions is enough to stabilize the training process. Thus, our EMA doesn't contain learnable parameters.\nA c: i1 ∈ R (N +1)×(N +1) , A c: i2 ∈ R\nThe outputs are computed as:\nO c: i1 = A c: i1 V c i , O c: i2 = V c: i A c: i2 .(5)\nWe next apply the proposed token blend module to merge heads and generate tokens capturing multiscale knowledge and the detailed discussions are deferred to section 3.4. The batch normalization (Ioffe & Szegedy, 2015) to O c: i1 and O c: i2 is then used to adjust the outputs' scale. Finally, the residual connection structure is used to generate the final output of the attention block.\nThe total number of tokens is on the order of O(L/S) per channel and the complexity in attention along tokens is upper bounded by O(C •d 2 •L 2 /S 2 ), which is smaller than O(C •d 2 •L 2 ) complexity of the vanilla point-wise token construction. In practice, one can use efficient attention implementation (e.g., FlashAttention Dao et al. 2022) to further obtain nearly linear computational performance."
},
{
"figure_ref": [
"fig_1"
],
"heading": "CARD ATTENTION OVER CHANNELS",
"publication_ref": [
"b60"
],
"table_ref": [],
"text": "We first compute {Q i }, {K i } and {V i } via Equation (2) and then slice them over token dimension into {Q :n i }, {K :n i } and {V :n i } with Q :n i , K :n i , V :n potential high-dimensionality issue of covariates, the vanilla method may suffer from computation overhead and overfitting. Take traffic dataset (PeMS) as an example, this dataset contains 862 covariates. When setting the lookback window size as 96, the attention over channels will require at least 80 times the computational cost of attention over tokens. The full attention will also merge a lot of noise patterns into the output token and lead to spurious correlation in the final forecasting results. In this paper, we consider using the dynamic projection technique (Zhu et al., 2021) to get \"summarized\" tokens to the K :n i and V :n i for n-th token dimension as shown in Figure 2. We first use MLP layers F pk and F pv to project head dimensions from d head to some fixed r with r ≪ C, and then we use softmax to normalized the projected tensors P :n k and P :n v as follow:\nP :n ki = softmax(F pk (K :n i )), P :n vi = softmax(F pv (V :n i )),(6)\nwhere P :n ki , P :n vi ∈ R C×r . Next the \"summarized\" tokens are computed by\nK:n i = (P :n ki ) ⊤ K :n i , Ṽ :n i = (P :n vi ) ⊤ V :n i ,(7)\nwhere\nK:n i , Ṽ :n i ∈ R r×d head .\nFinally, the outputs are generated by applying Q :n i , K:n i and Ṽ :n to equations from (3) to (5) for n = 1, 2, ..., N + 1. The upper bound of total computational cost is reduced to\nO(L/S • C • r • d 2 ) which is smaller than O(L/S • C 2 • d 2 ) cost of the standard attention."
},
{
"figure_ref": [],
"heading": "TOKEN BLEND MODULE",
"publication_ref": [
"b52",
"b53"
],
"table_ref": [],
"text": "Multi-scale knowledge plays a crucial role in forecasting tasks and has significantly enhanced the performance of diverse models. (e.g., Xu et al., 2021;Zeng et al., 2023;Wang et al., 2023b;Zhou et al., 2022b;Zhang & Yan, 2023). Most of these works initially decompose the time series into seasonal and trend components and then employ separate structures to process the seasonal and trend components individually. However, this approach, despite its simplicity, leads to higher model complexity, which in turn increases computation cost and makes it susceptible to overfitting issue.\nIn this work, we consider using a specially designed token blend mechanism to utilize the multiscaling structural knowledge without additional computation costs. The token blend module replaces the standard token reconstruction after the multi-head attention by merging the adjacent token within the same head to produce the token for the next stage. The output token tensor O from the multi-head attention has 4-D with shape C × H × (N + 1) × d head . The token blend module will first merge the second and third dimensions and reshape O into 3-D tensor with shape C × H(N + 1) × d head . We then decouple the second dimension into three dimensions, i.e., H(N + 1) → h 1 × h 2 × h 3 where Figure 3: Illustration example of token blend block in CARD.\nh 1 = Hh 3 , h 2 = N + 1 and h 3 ≥ 1. The final output O uses h 3 × h 1 × d head to construct the token dimension. Here we call h 3 as blend size. When h 3 = 1, the aforementioned operations generate the same outputs in the standard transformer. When h 3 ≥ 2, the outputs will first combine the adjacent token within the same head, which would create the token that represents the knowledge over a larger range, i.e., lower resolution. With increasing the blend size h 3 , more tokens within the same heads are merged and the attention module in the next stage could have more chance to capture long-term knowledge. An illustration example is shown in Figure 3. By consolidating temporally adjacent tokens within the same head, the resulting new tokens encompass knowledge over an extended time period. This enables more effective exploration of low-resolution knowledge by increasing attention on these tokens. Our token blend module is also different from the hierarchical adjacent tokens merging procedure in (Zhang & Yan, 2023). First, (Zhang & Yan, 2023) merges at the token level, the output token sequence at the coarse level has higher hidden dimensions and shorter sequence lengths. We consider merging at the head level instead which maintains the same output token sequence shape. Second, the merging size in (Zhang & Yan, 2023) is fixed as 2, while we allow a more flexible configuration. As a result, we achieve an implicit structure that enhances the extraction of multi-scale information without the need for an additional explicit signal disentanglement process."
},
{
"figure_ref": [],
"heading": "SIGNAL DECAY-BASED LOSS FUNCTION",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we discuss our loss function design. In literature, the Mean Squared Error (MSE) loss is commonly used to measure the discrepancy between the forecasting results and the ground truth observations. Let ât+1 (A), ...., ât+L (A) and a t+1 (A), ...., a t+L (A) be the predictions and real obversations from time t + 1 to t + L given historical information A. The overall objective loss becomes:\nmin E A 1 L L l=1 ∥â t+l (A) -a t+l (A)∥ 2 2 . (8\n)\nOne drawback of plain MSE loss for forecasting tasks is that the different time steps' errors are equally weighted. In real practice, the correlation of historical information to far-future observations is usually smaller than that to near-future observations, implying that far-future observations have higher variance. Therefore, the near-future loss would contribute more to generalization improvement than the far-future loss. To see this, we assume that our time series follows the first-order Markov process, i.e., a t+1 ∼ N (G(a t ), σ 2 I), where G is the smooth transition function with Lipschitz constant 1, σ > 0 and t = 1, 2, .... Then, we have\nvar(a t+1 ) = var(G(a t )) + σ 2 I ⪯ var(a t ) + σ 2 I,(9)\nwhere var(a) denote the covariance matrix of a. By recursively using Equation ( 9) from t + L to t and for all l ∈ [t, t + L], we have var(a t+l ) ⪯ lσ 2 I + var(a t ).\n(10) When a t is already observed, we have var(a t ) = 0 and Equation (10) implies var(a t+l ) ⪯ lσ 2 I. If we use negative log-likelihood estimation over Gaussian distribution, we come up with the following approximated loss function:\nmin E A 1 2 L l=1 (â t+l (A) -a t+l (A)) ⊤ var (a t+l ) -1 (â t+l (A) -a t+l (A)) ≥E A 1 2 L l=1 ∥â t+l (A) -a t+l (A)∥ 2 2 lσ 2 ∝ E A 1 L L l=1 l -1 ∥â t+l (A) -a t+l (A)∥ 2 2 . (11\n)\nCompared Equation ( 11) to Equation ( 8), the far-future loss is scaled down to address the high variance. Since Mean Absolute Error (MAE) is more resilient to outliers than square error, we propose to use the loss function in the following form:\nmin E A 1 L L l=1 l -1/2 ∥â t+l (A) -a t+l (A)∥ 1 ,(12)\nwhere Equation ( 12) can be derived via Equation ( 11) with replacing the Gaussian distribution by Laplace distribution."
},
{
"figure_ref": [],
"heading": "EXPERIMENTS",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "LONG TERM FORECASTING",
"publication_ref": [
"b56"
],
"table_ref": [],
"text": "Datasets We conducted experiments on seven real-world benchmarks, including four Electricity Transform Temperature (ETT) datasets (Zhou et al., 2021) comprising of two hourly and two 15minute datasets, one 10-minute weather forecasting dataset (Wetterstation), one hourly electricity consumption dataset (UCI), and one hourly traffic road occupancy rate dataset (PeMS)."
},
{
"figure_ref": [],
"heading": "Baselines and Experimental Settings",
"publication_ref": [
"b48",
"b54",
"b53",
"b29",
"b16"
],
"table_ref": [
"tab_8"
],
"text": "We use the following recent popular models as baselines: FEDformer (Zhou et al., 2022b), ETSformer (Woo et al., 2022), FilM (Zhou et al., 2022a), LightTS (Zhang et al., 2022), MICN (Wang et al., 2023b), TimesNet (Wu et al., 2023b), Dlinear (Zeng et al., 2023), Crossformer (Zhang & Yan, 2023), and PatchTST (Nie et al., 2023). We use the experimental settings in (Wu et al., 2023b) applying reversible instance normalization (RevIN, Kim et al., 2022) to handle data heterogeneity and keeping the lookback length as 96 for fair comparisons. Each setting is repeated 10 times and average MSE/MAE results are reported. The full results are summarized in Table 7 in the Appendix. More details on model configurations, model code, and comparison with other early baselines can be found in Appendix D and Appendix B, respectively."
},
{
"figure_ref": [],
"heading": "Results",
"publication_ref": [
"b29",
"b53"
],
"table_ref": [
"tab_1"
],
"text": "The results are summarized in Table 1. Regarding the average performance across four different output horizons, CARD gains the best performance in 6 out of 7 and 7 out of 7 in MSE and MAE, respectively. In single-length experiments, CARD achieves the best results in 82% cases in MSE metric and 100% cases in MAE metric.\nFor problems with complex covariate structures, the proposed CARD method beats the benchmarks by significant margins. For instance, in Electricity (321 covariates), CARD consistently outperforms the second-best algorithm by reducing MSE/MAE by more than 9.0% on average in each forecasting horizon experiment. By leveraging 21 covariates for Weather and 862 covariates for Traffic, we achieve a large reduction in MSE/MAE of over 7.5%. This highlights CARD's exceptional capability to incorporate extensive covariate information for improved prediction outcomes. Furthermore, Crossformer (Zhang & Yan, 2023) employs a comparable concept of integrating cross-channel data to enhance predictive accuracy. Remarkably, CARD significantly reduces the MSE/MAE by over 20% on 6 benchmark datasets compared to Crossformer, which shows our attention design is much more effective in utilizing cross-channel information. It's also important to note that while Dlinear shows strong performance in those tasks using an MLP-based model, CARD still consistently reduces MSE/MAE by 5% to 27.5% across all benchmark datasets.\nRecent works, such as Nie et al. 2023;Zeng et al. 2023;Zhang & Yan 2023) use the input length other than 96 and have shown performance improvement. In our study, we also report the numerical performance of CARD with a varying lookback length in Appendix G, and CARD consistently outperforms all baseline models when prolonging input sequence as well, demonstrating significantly lower MSE errors across all benchmark datasets."
},
{
"figure_ref": [],
"heading": "RECONSTRUCTION BASED ANOMALY DETECTION",
"publication_ref": [],
"table_ref": [
"tab_2"
],
"text": "Reconstruction based anomaly detection can be viewed as a task to predict the input itself. In previous works, the reconstruction is a classical task for unsupervised point-wise representation learning, where the reconstruction error is a natural anomaly criterion. We follow the experimental settings in (Wu et al., 2023a) and consider five widely used anomaly detection benchmarks. The results are summarized in Table 2. CARD outperforms the existing best result by 3% in F1 score on average. In "
},
{
"figure_ref": [],
"heading": "BOOSTING EFFECT OF SIGNAL DECAY-BASED LOSS FUNCTION",
"publication_ref": [],
"table_ref": [
"tab_3"
],
"text": "In this section, we present the boosting effect of our proposed signal decay-based loss function.\nIn contrast to the widely used MSE loss function employed in previous training of long-term sequence forecasting models, our approach yields a reduction in MSE ranging from 3% to 12% across a spectrum of recent state-of-the-art baseline models, including Transformer, CNN, and MLP architectures as shown in Table 3. Our proposed loss function specifically empowers FEDformer and Autoformer, two algorithms that heavily rely on frequency domain information. This aligns with our signal decay paradigm, which acknowledges that frequency information carries variance/noise across time horizons. Our novel loss function can be considered a preferred choice for this task, owing to its superior performance compared to the plain MSE loss function. More detailed discussions are deferred to Section J in Appendix. "
},
{
"figure_ref": [],
"heading": "INFLUENCE OF INPUT SEQUENCE LENGTH",
"publication_ref": [
"b53"
],
"table_ref": [
"tab_4"
],
"text": "Previous research (Zeng et al., 2023;Wen et al., 2023b) has highlighted a critical issue with the existing long-term forecasting transformers. They struggle to leverage extended input sequences, resulting in a decline in performance as the input length increases. We assert that this is not an inherent drawback of transformers, and CARD demonstrates robustness in handling longer and noisier historical sequence inputs, as evidenced by an 8.6% and 8.9% reduction in MSE achieved in the ETTh1 and ETTm1 datasets, respectively, when input lengths were extended from 96 to 720, as shown in Table 4."
},
{
"figure_ref": [
"fig_2"
],
"heading": "INFLUENCE OF TOKEN BLEND SIZE",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we test the effect of the token blend module by varying blend size. The results are summarized in Figure 4. When setting the blend size to 1, the token blend module reduces to the standard token mix method in Transformer literature and we observe test errors in both MSE/MAE increase. While using a larger blend size, the multi-scale information is utilized and the errors are reduced in turn. However, in some cases, further increasing the blend size may damage the performance. we conjecture it is due to the nature of the dataset that only some scales of knowledge are useful for forecasting. A higher blend size may oversmooth that knowledge. "
},
{
"figure_ref": [],
"heading": "OTHER EXPERIMENTS",
"publication_ref": [
"b28"
],
"table_ref": [],
"text": "We conduct a series of experiments, using both ablation and architecture variants, to evaluate each component in our proposed model. Our findings reveal that the channel branch made the greatest contribution to the reduction of MSE errors, as shown in Appendix Q.2. Furthermore, our experiments on sequential/parallel attention mixing design, detailed in Appendix Q.1, show that our model design is the preferred option. Visual aids and attention maps can be found in Appendix A and O, which effectively demonstrate our accurate predictions and utilization of covariate information. Another noteworthy experiment, concerning the impact of training data size, is presented in Appendix R.2. This study revealed that using 70% of training samples can significantly improve performance for half datasets affected by distribution shifts. Besides, Appendix L presents an error bar statistics table that demonstrates the robustness of CARD.More forecasting experiments on M4 (Makridakis et al., 2018) other datasets are presented in Appendix H and I."
},
{
"figure_ref": [],
"heading": "CONCLUSION AND FUTURE WORKS",
"publication_ref": [],
"table_ref": [],
"text": "In this paper, we present a novel Transformer model, CARD, for time series forecasting. CARD is a channel-dependent model that aligns information across different variables and hidden dimensions effectively. CARD improves traditional transformers by applying attention to both tokens and channels. The new design of the attention mechanism helps explore local information within each token, making it more effective for time series forecasting. We also propose a token blend module to utilize the multi-scale information knowledge in time series. Furthermore, we introduce a robust loss function to alleviate the issue of overfitting noises, an important issue in time series analysis. As demonstrated through various numerical benchmarks, our proposed model outperforms state-of-theart models. # construct Q,K,V B,nvars, H, C, = src.shape qkv = self.qkv(src).reshape(B,nvars, H, 3, self.n_heads, C // self.n_heads).permute(3, 0, 1,4, 2, 5) q, k, v = qkv[0], qkv[1], qkv [2] if not self.over_channel: attn_score_along_token = torch.einsum('bnhed,bnhfd->bnhef', self.ema(q), self.ema(k ))/ self.head_dim ** -0.5 attn_along_token = self.attn_dropout(F.softmax(attn_score_along_token, dim=-1) ) output_along_token = torch.einsum ('bnhef,bnhfd->bnhed', attn_along_token, v) else:\n[i][j] = ema_matrix[i-1][j] * (1-alpha) ema_matrix[i][i] = alpha self.register_buffer('ema_matrix',ema_matrix) def ema(self,\n# dynamic project V and K v_dp,k_dp = self.dynamic_projection(v,self.dp_v) , self.dynamic_projection(k,self. dp_k) attn_score_along_token = torch.einsum('bnhed,bnhfd->bnhef', self.ema(q), self.ema( k_dp))/ self.head_dim ** -0.5 attn_along_token = self.attn_dropout(F.softmax(attn_score_along_token, dim=-1) ) output_along_token = torch.einsum('bnhef,bnhfd->bnhed', attn_along_token, v_dp) # attention over hidden dimensions attn_score_along_hidden = torch.einsum('bnhae,bnhaf->bnhef', q,k)/ q.shape[-2] ** -0.5 attn_along_hidden = self.attn_dropout(F.softmax(attn_score_along_hidden, dim=-1) ) output_along_hidden = torch.einsum('bnhef,bnhaf->bnhae', attn_along_hidden, v) # token blend output1 = rearrange(output_along_token.reshape(B * nvars,-1,self. "
},
{
"figure_ref": [],
"heading": "C DATASETS DETAILS FOR LONG TERM FORECASTING",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Datasets of Long-term Forecasting",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "D MODEL CONFIGURATIONS FOR LONG TERM FORECASTING",
"publication_ref": [
"b16",
"b30",
"b37",
"b48",
"b54",
"b53",
"b29",
"b16",
"b52",
"b56"
],
"table_ref": [
"tab_7",
"tab_8",
"tab_2",
"tab_12"
],
"text": "For all experiments, we use reversible instance normalization (RevIN, Kim et al., 2022) to handle data heterogeneity. As suggested in (Olivares et al., 2023) and (Salinas et al., 2020), other standardization methods are also useful when data enjoys certain patterns. We would like to defer the detailed analysis of them into future study. Moreover, the Adam optimizer (Kingma & Ba, 2017) with cosine learning rate decay after linear warm-up is used as training scheme. We train the proposed models with at most 8 NVIDIA Tesla V100 SXM2-16-GB GPUs. For all experiments, we fixed the number of encoder blocks, head dimensions and dynamic projection dimensions being 2, 8, and 8, respectively.\nThe training epoch is set as 100. The default batch size is 128 and is adjusted due to GPU memory restriction. Other details of configurations are summarized in Table 6. We use the following recent popular models as baselines: FEDformer (Zhou et al., 2022b), ETSformer (Woo et al., 2022), FilM (Zhou et al., 2022a), LightTS (Zhang et al., 2022), MICN (Wang et al., 2023b), TimesNet (Wu et al., 2023b), Dlinear (Zeng et al., 2023), Crossformer (Zhang & Yan, 2023), and PatchTST (Nie et al., 2023). We use the experimental settings in (Wu et al., 2023b) applying reversible instance normalization (RevIN, Kim et al., 2022) to handle data heterogeneity and keeping the lookback length as 96 for fair comparisons. Each setting is repeated 10 times and average MSE/MAE results are reported.\nIn this section, we report the full results of long-term forecasting experiments in section 5.1. The MSE/MAE results are summarized in Table 7 and standard errors are reported in Table 8. CARD achieves 23/28 best performance in MSE and all the best results in MAE. It implies CARD can improve the baselines in a broad range of forecasting horizons. The standard deviation of CARD is on the order of 1e-3, which indicates our proposed framework is very robust. More baselines such as autoformer (Xu et al., 2021), nonstationary transformer (Liu et al., 2022b), Pyraformer (Liu et al., 2022a), LogTrans (Li et al., 2019b) and Informer (Zhou et al., 2021) can be found in Table 2 and Table 13 of (Wu et al., 2023b). CARD consistently outperforms those models in all forecasting horizons and we omit them for brevity. "
},
{
"figure_ref": [],
"heading": "F COMPARISON TO EARLY BASELINES",
"publication_ref": [
"b53"
],
"table_ref": [
"tab_9"
],
"text": "In this section, we report the comparison of CARD with early baselines including Nlinear, Linear, and Repret in Zeng et al. (2023). We use the experiment settings in subsection 5.1 and fix the input length as 96. The results are summarized in Table 9. Our model consistently outperforms those baselines."
},
{
"figure_ref": [],
"heading": "G EXPERIMENTS ON ALL BENCHMARK DATASETS BY VARYING THE INPUT LENGTH TO ACHIEVE THE BEST RESULTS REPORTED IN BASELINE LITERATURE",
"publication_ref": [
"b29"
],
"table_ref": [
"tab_1"
],
"text": "We report the proposed model with 720 input length in Table 10. We follow the experimental settings used in (Nie et al., 2023). For each benchmark, we report the best results in the literature or conduct grid searches on input length to build strong baselines. In single-length experiments, CARD achieves "
},
{
"figure_ref": [],
"heading": "H M4 SHORT TERM FORECASTING",
"publication_ref": [
"b28",
"b31",
"b2",
"b56",
"b49",
"b31",
"b29"
],
"table_ref": [
"tab_10",
"tab_11",
"tab_12",
"tab_13"
],
"text": "We also conduct experiments on short forecasting M4 tasks. M4 dataset (Makridakis et al., 2018) consists 100k time series. It covers time data in various domains, including business, financial, and economy, and the sampling frequencies range from hourly to yearly. We follow the test setting suggested in (Wu et al., 2023b). Each experiment is repeated 10 times and average Symmetric Mean Absolute Percentage Error (SMAPE), Mean Absolute Scaled Error (MASE), and Overall Weighted Average (OWA) are reported. We benchmark our model with N-BEATS (Oreshkin et al., 2020), N-HiTS (Challu et al., 2022), Informer (Zhou et al., 2021), Autoformer (Wu et al., 2021) and 7 baselines in long-term forecasting. Details for datasets and training configurations can be found in Table 11 and Table 12 respectively.\nThe results are summarized in Table 13. Our proposed model consistently outperforms benchmarks in all tasks. Specifically, we outperform the state-of-the-art MLP-based method N-BEATS (Oreshkin et al., 2020) by 1.8% in SMAPE reduction. We also outperform the best Transformer-based method PatchTST (Nie et al., 2023) and the best CNN-based method TimesNet (Wu et al., 2023b) by 1.5% and 2.2% in SMAPE reductions respectively. Since the M4 dataset only contains univariate time series, the attention to channels in our model plays a very limited role here. Thus good numerical performance indicates CARD's design with attention to hidden dimensions and token blend are also effective in univariate time series scenarios and can significantly boost forecasting performance.\nThe standard errors are reported in Table 14. Since the SAMPE score is not normalized, we observe the absolute value is on the order of 1e-2 while the MASE and OWA remain on the order of 1e-3 which is the same as in long-term forecasting experiments. After normalizing SAMPE with the corresponding mean value, the standard error of SMAPE will also reduce to the order of 1e-3. "
},
{
"figure_ref": [],
"heading": "I OTHER FORECASTING TASKS",
"publication_ref": [
"b18"
],
"table_ref": [
"tab_14",
"tab_15",
"tab_8",
"tab_1"
],
"text": "In this section, we report the results of Illness and Exchange tasks. The Illness (CDC) and Exchange (Lai et al., 2018) contains the weekly data on influenza-like illness from Jan-2002 to Jun-2020 and the daily exchange rates of eight foreign countries including Australia, British, Canada, Switzerland, China, Japan, New Zealand, and Singapore ranging from 1990 to 2016 respectively. We follow the test setting suggested in (Wu et al., 2023b). Each experiment is repeated 10 times and MSE and MAE are reported. We benchmark our model with the baselines in long-term forecasting. Details for datasets and training configurations can be found in Table 15 and Table 16 respectively.\nThe results are summarized in Table 17. Our proposed model outperforms benchmarks in 4/8 cases in MSE and 6/8 cases in MAE. The standard errors are reported in Table 18. "
},
{
"figure_ref": [],
"heading": "J EXTENDED RESULTS OF SIGNAL-BASED LOSS FUNCTION",
"publication_ref": [],
"table_ref": [
"tab_9",
"tab_2",
"tab_16",
"tab_16"
],
"text": "The full results of experiments in section 5.3 are reported in Table 19 and Table 20. Moreover, we also conduct an experiment on switching to the decay function other than the two forms considered in section 4. The results are summarized in Table 21. in Table 21, we consider the following decay function:\nf (t) = t -1/4 , f (t) = t -1/3 , f (t) = t -1 , f (t) = t -2\n, and f (t) = t -3 . In the ETTm1 task, we find that the decay function from f (t) = t -1/4 and f (t) = t -1/3 gives a similar MSE performance and slightly worse (by 0.001) MAE performance on average compared to the squared root decay. In the ETTh1 task, f (t) = t -1/4 , f (t) = t -1/3 , and f (t) = t -1 work the same good as squared root decay. In practice, we believe the function that is not \"decaying\" faster than f (t) = t -1 might be the candidate choice when no further information/assumptions on datasets could be obtained.\nFor the slow decaying function (e.g., f (t) = t -1/4 and f (t) = t -1/3 ), vert slight performance improvement is observed in individual tasks when it is getting close to the squared root decay. It implies that the proposed loss is robustness for slow decaying function.\nWe also provide an illustration example to show the rationality of the proposed signal-based loss function. Let's consider a 1D autoregressive model x t+1 = β true x t + ϵ t with ϵ t ∼ N (0, 1), β true ∈ (0, 1) and |x t | ≤ 1. And we want to use x t to forecast x t+1 and x t+2 . The plain loss function would be as follows: In this case, our proposed loss becomes:\nmin β T t=1 [∥x t β -x t+1 ∥ 2 2 + ∥x t β 2 -x t+2 ∥ 2 2 ].(13)\nf (t) = 1 f (t) = t -0.25 f (t) = t -0.33 f (t) = t -0.5 f (t) = t -1 f (t) = t -2 f (t) = t -\nmin β T t=1 [∥x t β -x t+1 ∥ 2 2 + 1 2 ∥x t β 2 -x ∥ 2 2 ].(15)\nFollows the same analysis procedures, we have with probability 1 -δ\n|β -β true | ≤ 3 2 T log(1/δ) T t=1 x 2 t . (16\n)\nHere the constant is improved from 5 2 to 3 2 , which implies the new loss may yield better convergence upper bound."
},
{
"figure_ref": [
"fig_6",
"fig_14",
"fig_15",
"fig_1",
"fig_18",
"fig_1",
"fig_19",
"fig_27"
],
"heading": "K EXTENDED RESULTS OF ANOMALY DETECTION",
"publication_ref": [],
"table_ref": [
"tab_18",
"tab_19",
"tab_20"
],
"text": "The full results of the anomaly detection experiment in the section 5.2 are reported in Table 22. For each setting, we repeat 5 replicates. In this section, we report the robustness test when varying input length. We conduct experiments on ETTh1, ETTm1, Weather, and M4 datasets and repeat each setting with 10 random seeds. The robust experiment results are summarized in Figure 9-Figure 17. In general, we observe the longer input length may yield better performance, and the variance is also enlarged slightly. In this section, we report the robustness test when varying model size. We conduct experiments on ETTh1, ETTm1, Weather, and M4 datasETSformer The model size (hidden dimension in attention) changes from 16 to 128 and the MLP layer dimension is set to be 2 times the model size. We repeat each setting with 10 random seeds. The robust experiment results are summarized in Figure 18-Figure 26. In terms of average performance (e.g., Figure 21 and Figure 26), the larger model size gives better results. For each individual task, we observe that the model with a large hidden dimension (e.g., 128) tends to overfit in low complexity tasks like ETTh1. For the high complexity task, the larger model size enables bigger learning capacity and gives better performance. seeds. We consider the fixed learning rate and cosine learning rate scheduler with the initial learning rate from 1e-3 to 1e-5. The results are summarized in Figure 27-Figure 35. We observe slight improvements when changing the fixed learning rate to the cosine learning rate decaying. For the relatively large learning rate, the variance of testing MSE/MAE increases and for the small enough learning rate, the model tends to underfit for the given training epochs. In practice, results suggest the learning should be set on the order of 1e-4. 23. Due to the patchified tokenization, when changing the input sequence length from 96 to 720 (7.5 times longer), the time increases less than 600% and thus we don't observe the quadratic time differences. It implies our model can efficiently handle long input sequences. \n)), our model's complexity can also be nearly linear. The condition S ≈ O( √ L) is not very restrictive. Take L = 900 as an example, the length of S = 30 would be enough. For the case L = 96, the corresponding S would be around 10.\nThe results of the experiments on running time are reported in Table 24. The input/forecasting lengths are set as 96/96 and we keep the batch size the same for all benchmarks and run the experiments on a single A100/80G GPU. Our proposed model yields comparable running time to transformer baselines as well as linear complexity baselines except Dlinear, which implies in practice model could also behave like a linear time model and won't introduce overhead computational cost. "
},
{
"figure_ref": [
"fig_28"
],
"heading": "O ATTENTION PATTERN MAPS",
"publication_ref": [],
"table_ref": [
"tab_21"
],
"text": "In this section, we report the attention maps of each head in the last attention layers. We use ETTh1 and ETTh2 tasks with forecasting length 96. The input length is set as 96 and we use patch length 8 with stride 8 to convert 96 time steps into 12 tokens, and we use the model with 2 attention heads. In order to highlight the correlation between the attention maps w.r.t. the forecasting sequences. We also report the dynamic time warping (DTW) scores between patches and the forecasting sequences, and the sum of attention scores for each patch. The DTW score can be treated as a rough ground truth to evaluate which input patches are most useful for forecasting. The results are summarized in Figure 36-Figure 37. We observe that attention maps have smooth landscapes and we believe it is due to the usage of EMA module to query and keyword tensors. Moreover, we find that the sum of attention scores for each patch is positively correlated with the post-hoc computed DTW scores between the patch and the forecasting sequence. It implies the proposed model can effectively 25.\nFollowing an exhaustive analysis, it is concluded that the architecture featuring the channel branch, complemented by channel/time blend, is the most resilient variant. Consequently, this specific architecture is adopted as the default approach in this work."
},
{
"figure_ref": [],
"heading": "Q.2 COMPONENT ABLATION EXPERIMENTS",
"publication_ref": [
"b53"
],
"table_ref": [
"tab_23"
],
"text": "Ablation on attention over hidden dimensions and over channels. We conducted a series of ablation experiments by removing the attention over hidden dimensions and channels sequentially. Consistent with our design, as shown in table 26, the channel branch contributes to the reduction of mean squared error (MSE); its removal resulted in a 2%, 7% and 6%increase in MSE for ETTm1, Weather and Electricity respectively. The attention over hidden dimensions branch contributes approximately 1% to the reduction of MSE in ETTm1 and Weather tasks and in Electricity tasks dropping the over hidden dimensions attention branch results in 11% more MSE score on average. the parameter of EMA from being fixed to learnable makes very few performance differences but significantly increases the training time. In practice, we would suggest using a fixed EMA parameter. 27. We have observed a distribution shift phenomenon in fifty percent of the benchmark datasets: Traffic, ETTh2, and ETTm2. The model's performance demonstrates a significant enhancement with the use of only 70% training data samples compared to the standard training setting for long-term forecasting, as illustrated in table 28. While it has been argued that the transformer model exhibits a weakness where more training data fails to improve performance (Zeng et al., 2023), we contend that this issue is an inherent feature of each time series benchmark dataset, wherein changes in data distribution between historical and current data are not related to the transformer model. Nevertheless, further exploration of this phenomenon may lead to improved performance, and we thus leave it as a topic for future study."
},
{
"figure_ref": [],
"heading": "",
"publication_ref": [],
"table_ref": [],
"text": "https://github.com/wxie9/ CARD."
},
{
"figure_ref": [],
"heading": "A VISUALIZATION",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "L.2 INFLUENCE OF DIFFERENT LEARNING RATES AND SCHEDULERS",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we report the robustness test when varying learning rates and schedulers. We conduct experiments on ETTh1, ETTm1, Weather, and M4 datasets and repeat each setting with 10 random capture the relationship between the input sequences and forecasting sequences and lead to good final performance.\nFigure 36: Attention Map Samples of ETTh1 task."
},
{
"figure_ref": [],
"heading": "P RELATED WORKS",
"publication_ref": [
"b41",
"b7",
"b34",
"b9",
"b26",
"b1",
"b8",
"b12",
"b13",
"b35",
"b4",
"b53"
],
"table_ref": [],
"text": "Patched Transformers in other Domains Transformer (Vaswani et al., 2017) has demonstrated significant potential in different data modalities. Among all applications, patching is an essential part when local semantic information is important. In NLP, BERT (Devlin et al., 2018), GPT (Radford et al., 2019) and their follow-up models consider subword-based tokenization and outperform character-based tokenization. In CV, Vision Transformers (e.g., Dosovitskiy et al. 2020;Liu et al. 2021;Bao et al. 2022;Ding et al. 2022;He et al. 2022) split an image into patches and then feed into the Transformer models. Similarly, in speech fields, researchers use convolutions to extract information in sub-sequence levels from a raw audio input (e.g., Hsu et al. 2021;Radford et al. 2022;Chen et al. 2022;Wang et al. 2023a). (Zeng et al., 2023) suggests that a linear layer can be used as a substitute for the self-attention layer to achieve higher accuracy in transformer-based models. To highlight the effectiveness of self-attention in our model, we conduct experiments of replacing self-attention modules (e.g., attention over tokens and channels) with linear layer. The results are summarized in Table 29. Upon replacing channelbranch attention and token attention with a linear layer in CARD, we observe a consistent decline in accuracy across all datasets. The deterioration effect is particularly pronounced in the weather dataset, which contains more informative covariates, with a significant drop of over 13%. These findings suggest that the self-attention scheme may be more effective in feature extraction than a simple linear layer for time series forecasting. "
},
{
"figure_ref": [],
"heading": "S EVALUATION ON IMPUTATION",
"publication_ref": [],
"table_ref": [],
"text": "We test the proposed model's imputation ability. We adopt the experimental settings in (Wu et al., 2023b) and results are reported in Table 30. CARD obtains top 2 ranks in 22/24 MSE scores and all MAE scores. In particular, in the Electricity dataset, CARD significantly reduces the MSE and MAE by 40% and 28% over the previous best results respectively. Those results suggest CARD may also generate good representations and thus can also work in the problem beyond forecasting."
}
] | Recent studies have demonstrated the great power of Transformer models for time series forecasting. One of the key elements that lead to the transformer's success is the channel-independent (CI) strategy to improve the training robustness. However, the ignorance of the correlation among different channels in CI would limit the model's forecasting capacity. In this work, we design a special Transformer, i.e., Channel Aligned Robust Blend Transformer (CARD for short), that addresses key shortcomings of CI type Transformer in time series forecasting. First, CARD introduces a channel-aligned attention structure that allows it to capture both temporal correlations among signals and dynamical dependence among multiple variables over time. Second, in order to efficiently utilize the multi-scale knowledge, we design a token blend module to generate tokens with different resolutions. Third, we introduce a robust loss function for time series forecasting to alleviate the potential overfitting issue. This new loss function weights the importance of forecasting over a finite horizon based on prediction uncertainties. Our evaluation of multiple long-term and short-term forecasting datasets demonstrates that CARD significantly outperforms state-of-the-art time series forecasting methods. | CARD: CHANNEL ALIGNED ROBUST BLEND TRANS-FORMER FOR TIME SERIES FORECASTING | [
{
"figure_caption": "Figure 1 :1Figure 1: Illustration of the architecture of CARD.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: Architecture for the CARD attention block.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: Experiments on token blend size. The blend size is varying in 1, 2, 4, 8, and 16.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Figure 7 :7Figure 7: Sample prediction graph for Weather long-term forecasting task",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "7",
"figure_type": "figure"
},
{
"figure_caption": "src): return torch.einsum('bnhad,ga ->bnhgd',src,self.ema_matrix[:src.shape[-2],:src.shape [-2]]) def dynamic_projection(self,src,mlp): src_dp = mlp(src) src_dp = F.softmax(src_dp,dim = -1) src_dp = torch.einsum('bnhef,bnhec -> bnhcf',src,src_dp) return src_dp def forward(self, src, * args, ** kwargs):",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "head_dim), 'bn (hl1 hl2 hl3) d -> bn hl2 (hl3 hl1) d', hl1 = self.n_heads//self.merge_size, hl2 = output_along_token.shape[-2] ,hl3 = self.merge_size ).reshape(B * nvars,-1,self.head_dim * self.n_heads) output2 = rearrange(output_along_hidden.reshape(B * nvars,-1,self.head_dim), 'bn (hl1 hl2 hl3) d -> bn hl2 (hl3 hl1) d', hl1 = self.n_heads//self.merge_size, hl2 = output_along_token.shape[-2] ,hl3 = self.merge_size ).reshape(B * nvars,-1,self.head_dim * self.n_heads) # post_norm output1 = self.norm_post1(output1).reshape(B,nvars, -1, self.n_heads * self.head_dim) output2 = self.norm_post2(output2).reshape(B,nvars, -1, self.n_heads * self.head_dim) # add & norm src2 = self.ff_1(output1)+self.ff_2(output2) src = src + src2 src = src.reshape(B * nvars, -1, self.n_heads * self.head_dim) src = self.norm_attn(src) src = src.reshape(B,nvars, -1, self.n_heads * self.head_dim) return src",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 9 :9Figure 9: ETTm1 experiments with different input lengths.",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "9",
"figure_type": "figure"
},
{
"figure_caption": "Figure 10 :10Figure 10: ETTm1 experiments with different input lengths.",
"figure_data": "",
"figure_id": "fig_7",
"figure_label": "10",
"figure_type": "figure"
},
{
"figure_caption": "Figure 11 :11Figure 11: Weather experiments with different input lengths.",
"figure_data": "",
"figure_id": "fig_8",
"figure_label": "11",
"figure_type": "figure"
},
{
"figure_caption": "Figure 12 :12Figure 12: The average results of ETTm1, ETTh1 and Weather experiments with different input lengths.",
"figure_data": "",
"figure_id": "fig_9",
"figure_label": "12",
"figure_type": "figure"
},
{
"figure_caption": "Figure 13 :13Figure 13: M4 Yearly experiments with different input lengths. The x axis \"input length ratio\" represents the ratio between input length and forecasting length.",
"figure_data": "",
"figure_id": "fig_10",
"figure_label": "13",
"figure_type": "figure"
},
{
"figure_caption": "Figure 14 :14Figure 14: M4 Quarterly experiments with different input lengths. The x axis \"input length ratio\" represents the ratio between input length and forecasting length.",
"figure_data": "",
"figure_id": "fig_11",
"figure_label": "14",
"figure_type": "figure"
},
{
"figure_caption": "Figure 15 :15Figure 15: M4 Monthly experiments with different input lengths. The x axis \"input length ratio\" represents the ratio between input length and forecasting length.",
"figure_data": "",
"figure_id": "fig_12",
"figure_label": "15",
"figure_type": "figure"
},
{
"figure_caption": "Figure 16 :16Figure 16: M4 average results of Daily, Weekly and Hourly experiments with different input lengths. The x axis \"input length ratio\" represents the ratio between input length and forecasting length.",
"figure_data": "",
"figure_id": "fig_13",
"figure_label": "16",
"figure_type": "figure"
},
{
"figure_caption": "Figure 17 :17Figure 17: Average results of all M4 experiments with different input lengths. The x axis \"input length ratio\" represents the ratio between input length and forecasting length.",
"figure_data": "",
"figure_id": "fig_14",
"figure_label": "17",
"figure_type": "figure"
},
{
"figure_caption": "Figure 18 :18Figure 18: ETTm1 experiments with different model sizes.",
"figure_data": "",
"figure_id": "fig_15",
"figure_label": "18",
"figure_type": "figure"
},
{
"figure_caption": "Figure 19 :19Figure 19: ETTh1 experiments with different model sizes.",
"figure_data": "",
"figure_id": "fig_16",
"figure_label": "19",
"figure_type": "figure"
},
{
"figure_caption": "Figure 20 :20Figure 20: Weather experiments with different model sizes.",
"figure_data": "",
"figure_id": "fig_17",
"figure_label": "20",
"figure_type": "figure"
},
{
"figure_caption": "Figure 21 :21Figure 21: The average results of ETTm1, ETTh1 and Weather experiments with different model sizes.",
"figure_data": "",
"figure_id": "fig_18",
"figure_label": "21",
"figure_type": "figure"
},
{
"figure_caption": "Figure 27 :27Figure 27: ETTh1 experiments with different learning rates and schedulers.",
"figure_data": "",
"figure_id": "fig_19",
"figure_label": "27",
"figure_type": "figure"
},
{
"figure_caption": "Figure 28 :28Figure 28: ETTm1 experiments with different learning rates and schedulers.",
"figure_data": "",
"figure_id": "fig_20",
"figure_label": "28",
"figure_type": "figure"
},
{
"figure_caption": "Figure 29 :29Figure 29: Weather experiments with different learning rates and schedulers.",
"figure_data": "",
"figure_id": "fig_21",
"figure_label": "29",
"figure_type": "figure"
},
{
"figure_caption": "Figure 30 :30Figure 30: The average results of ETTm1, ETTh1 and Weather experiments with different learning rates and schedulers.",
"figure_data": "",
"figure_id": "fig_22",
"figure_label": "30",
"figure_type": "figure"
},
{
"figure_caption": "Figure 31 :31Figure 31: M4 Yearly experiments with different learning schemes.",
"figure_data": "",
"figure_id": "fig_23",
"figure_label": "31",
"figure_type": "figure"
},
{
"figure_caption": "Figure 32 :32Figure 32: M4 Quarterly experiments with different learning schemes.",
"figure_data": "",
"figure_id": "fig_24",
"figure_label": "32",
"figure_type": "figure"
},
{
"figure_caption": "Figure 33 :33Figure 33: M4 Monthly experiments with different learning schemes.",
"figure_data": "",
"figure_id": "fig_25",
"figure_label": "33",
"figure_type": "figure"
},
{
"figure_caption": "Figure 34 :34Figure 34: M4 average results of Daily, Weekly and Hourly experiments with different learning schemes.",
"figure_data": "",
"figure_id": "fig_26",
"figure_label": "34",
"figure_type": "figure"
},
{
"figure_caption": "Figure 35 :35Figure 35: Average results of all M4 experiments with different learning schemes.",
"figure_data": "",
"figure_id": "fig_27",
"figure_label": "35",
"figure_type": "figure"
},
{
"figure_caption": "Figure 37 :37Figure 37: Attention Map Samples of ETTh2 task.",
"figure_data": "",
"figure_id": "fig_28",
"figure_label": "37",
"figure_type": "figure"
},
{
"figure_caption": "FigureFigure 38: Architecture Variants",
"figure_data": "",
"figure_id": "fig_29",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 39 :39Figure 39: Experiments on dynamic projection dimensions. The projection dimension is varying in 1, 8, and 16.",
"figure_data": "",
"figure_id": "fig_30",
"figure_label": "39",
"figure_type": "figure"
},
{
"figure_caption": "Figure 40 :40Figure 40: Experiments on stability of EMA module. Each setting is averaged over 10 random seeds.",
"figure_data": "",
"figure_id": "fig_31",
"figure_label": "40",
"figure_type": "figure"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Long-term forecasting tasks. The lookback length is set as 96. All models are evaluated on 4 different prediction horizons {96, 192, 336, 720} and average MSE/MAE results of ten repeats are reported. The best model is in boldface and the second best is underlined.",
"figure_data": "ModelsCARDPatchTSTMICNTimesNet Crossformer DlinearLightTSFilMETSformer FEDformerMetric MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAEETTm1 0.383 0.383 0.395 0.408 0.387 0.411 0.400 0.406 0.435 0.417 0.403 0.407 0.435 0.437 0.408 0.399 0.429 0.425 0.448 0.452ETTm2 0.271 0.316 0.283 0.327 0.284 0.340 0.291 0.333 0.609 0.521 0.350 0.401 0.409 0.436 0.287 0.328 0.292 0.342 0.305 0.349ETTh1 0.443 0.429 0.455 0.444 0.440 0.462 0.458 0.450 0.486 0.481 0.456 0.452 0.491 0.479 0.461 0.456 0.452 0.510 0.440 0.460ETTh2 0.367 0.390 0.384 0.406 0.402 0.437 0.414 0.427 0.966 0.690 0.559 0.515 0.602 0.543 0.384 0.406 0.439 0.452 0.437 0.449Weather 0.240 0.262 0.257 0.280 0.243 0.299 0.259 0.287 0.250 0.310 0.265 0.317 0.261 0.312 0.269 0.339 0.271 0.334 0.309 0.360Electricity 0.169 0.258 0.216 0.318 0.187 0.295 0.192 0.295 0.273 0.363 0.212 0.300 0.229 0.329 0.223 0.303 0.208 0.323 0.214 0.327Traffic 0.450 0.278 0.488 0.327 0.542 0.316 0.620 0.336 0.593 0.332 0.625 0.383 0.622 0.392 0.639 0.389 0.621 0.396 0.610 0.376particular, CARD achieves 14.2% significant improvement in SMAP task. Those facts imply CARDcould generate meaningful representation on time series.",
"figure_id": "tab_1",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Anomaly detection. F1 scores are reported. The best model is in boldface and the second best is underlined.",
"figure_data": "Models CARD PatchTST MICN TimesNet Crossformer ETSformer LightTS Dlinear FEDformer Stationary Autoformer InformerSMD 0.872 0.866 0.800 0.8580.7780.8310.825 0.7710.8510.8470.8510.855MSL 0.817 0.823 0.816 0.8520.8200.8500.790 0.8490.7860.7750.7910.841SMAP 0.857 0.695 0.656 0.7150.6740.6950.692 0.6930.7080.7110.7110.699SWaT 0.945 0.909 0.875 0.9210.8860.8490.933 0.8750.9320.7990.9270.814PSM 0.957 0.951 0.933 0.9750.9210.9180.972 0.9360.9720.9730.9330.771Avg 0.890 0.849 0.816 0.8640.8160.8290.842 0.8250.8490.8210.8430.789",
"figure_id": "tab_2",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Influence for signal decay-based loss function. The lookback length is set as 96. All models are evaluated on 4 different predication lengths {96, 192, 336, 720}. The average results are reported, and the full table is deferred to Table 19 in the Appendix. The model name with * uses the robust loss proposed in this work. The better results are in boldface.",
"figure_data": "Models CARDCARD* MICN-regre MICN-regre* TimesNet TimesNet* FEDformer FEDformer* Autoformer Autoformer*Metric MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAEETTm1 0.390 0.399 0.383 0.383 0.392 0.414 0.383 0.393 0.400 0.406 0.392 0.395 0.448 0.452 0.413 0.415 0.588 0.528 0.523 0.475ETTh1 0.449 0.440 0.443 0.425 0.559 0.535 0.527 0.499 0.458 0.450 0.449 0.438 0.440 0.460 0.436 0.442 0.496 0.487 0.514 0.481",
"figure_id": "tab_3",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Influence of prolonging input sequence. The lookback length is set as96,192,336,720. ",
"figure_data": "Input Length96192336512720MetricMSE MAE MSE MAE MSE MAE MSE MAE MSE MAEETTm1 0.383 0.384 0.363 0.372 0.352 0.367 0.402 0.420 0.349 0.368ETTh1 0.442 0.429 0.429 0.425 0.415 0.422 0.352 0.371 0.405 0.421",
"figure_id": "tab_4",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Table 5 summarizes details of statistics of long-term forecasting datasETSformer Dataset details in long-term forecasting.",
"figure_data": "DatasetLength Dimension FrequencyETTm169680715 minETTm269680715 minETTh11742071 hourETTh21742071 hourWeather526962110 minElectricity 263043211 hourTraffic175448621 hour",
"figure_id": "tab_6",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "Model configurations of CARD.",
"figure_data": "Dataset patch stride model dim FFN dim dropout blend size learning rate warm-up batch sizeETTm116816320.321e-40128ETTm216816320.321e-40128ETTh116816320.321e-40128ETTh216816320.321e-40128Weather1681282560.2161e-40128Electricity 1681282560.2161e-42032Traffic1681282560.2161e-42024E EXTENDED NUMERICAL RESULTS OF CARD IN LONG-TERM FORECASTINGWITH 96 INPUT LENGTH",
"figure_id": "tab_7",
"figure_label": "6",
"figure_type": "table"
},
{
"figure_caption": "Long-term forecasting tasks. The lookback length is set as 96. All models are evaluated on 4 different prediction horizons {96, 192, 336, 720}. The best model is in boldface and the second best is underlined. For MICN, we report the better result between MICN-regre and MICN-mean. MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE",
"figure_data": "Models CARDPatchTSTMICNTimesNet Crossformer DlinearLightTSFilMETSformer FEDformerMetric MSE MAE",
"figure_id": "tab_8",
"figure_label": "7",
"figure_type": "table"
},
{
"figure_caption": "Comparision to early baselines in long-term forecasting tasks. All models are evaluated on 4 different predication lengths {96, 192, 336, 720}. The best model is in boldface and the second best is underlined.",
"figure_data": "",
"figure_id": "tab_9",
"figure_label": "9",
"figure_type": "table"
},
{
"figure_caption": "Datasets and mapping details of M4 dataset.",
"figure_data": "DatasetLength HorizonM4 Yearly230006M4 Quarterly 240008M4 Monthly4800018M4 Weekly35913M4 Daily422714M4 Hourly41448",
"figure_id": "tab_10",
"figure_label": "11",
"figure_type": "table"
},
{
"figure_caption": "Model configurations for M4 experiment.",
"figure_data": "Datasetpatch stride model dim FFN dim dropout blend size learning rate warm-up batch sizeM4 Hourly1611285120.125e-40128M4 Weekly1611285120.125e-40128M4 Daily1611285120.125e-40128M4 Monthly 1611285120.125e-40128M4 Quarterly 411285120.125e-40128M4 Yearly311285120.125e-40128",
"figure_id": "tab_11",
"figure_label": "12",
"figure_type": "table"
},
{
"figure_caption": "Short-term Forecasting tasks on M4 dataset. The average results of ten repeats are reported. The best model is in boldface and the second best is underlined.",
"figure_data": "ModelsCARD PatchTST MICN TimesNet N-HiTS N-BEATS ETSformer LightTS Dlinear FEDformer Autoformer InformerSMAPE 13.215 13.258 14.935 13.387 13.418 13.43618.009 14.247 16.965 13.72813.97414.727YearlyMASE 2.972 2.985 3.523 2.9963.0453.0434.4873.109 4.2833.0483.1343.418OWA 0.778 0.781 0.900 0.7860.7930.7941.1150.827 1.0580.8030.8220.881SMAPE 9.958 10.179 11.452 10.100 10.202 10.12413.376 11.364 12.145 10.79211.33811.360QuarterlyMASE 1.163 1.212 1.389 1.1821.1941.1691.9061.328 1.5201.2831.3651.401OWA 0.876 0.904 1.026 0.8900.8990.8861.3021.000 1.1060.9581.0121.027SMAPE 12.414 12.641 13.773 12.670 12.791 12.66714.588 14.014 13.514 14.26013.95814.062MonthlyMASE 0.907 0.930 1.076 0.9330.9690.9371.3681.053 1.0371.1021.1031.141OWA 0.856 0.867 0.983 0.8780.8990.8801.1490.981 0.9561.0121.0021.024SMAPE 4.522 4.851 6.716 4.8915.0614.9257.26715.880 6.7094.9545.45824.460OthersMASE 3.021 3.238 4.717 3.3023.2163.3915.24011.434 4.9533.2643.86520.960OWA 0.962 1.021 1.451 1.0351.0401.0531.5913.474 1.4871.0361.1875.879SMAPE 11.614 11.807 13.130 11.829 11.927 11.85114.718 13.252 13.639 12.84012.90914.086AvgMASE 1.553 1.590 1.896 1.5851.6131.5992.4082.111 2.0951.7011.7712.718OWA 0.832 0.834 0.980 0.8510.8610.8551.1721.051 1.0510.9180.9391.230",
"figure_id": "tab_12",
"figure_label": "13",
"figure_type": "table"
},
{
"figure_caption": "Standard error results of CARD in M4 short-term forecasting. The results normalized with the corresponding mean value are reported in parentheses. Each setting is averaged over 10 random seeds.",
"figure_data": "MetricYearlyQuarterlyMonthlyOtherAverageSAMPE 0.022 (0.001) 0.008 (0.001) 0.032 (0.002) 0.024 (0.005) 0.018 (0.002)MASE 0.007 (0.003) 0.003 (0.002) 0.003 (0.003) 0.026 (0.008) 0.003 (0.002)OWA 0.003 (0.002) 0.001 (0.001) 0.032 (0.037) 0.004 (0.004) 0.001 (0.001)",
"figure_id": "tab_13",
"figure_label": "14",
"figure_type": "table"
},
{
"figure_caption": "Datasets and mapping details of Illness and Exchange datasets.",
"figure_data": "DatasetLength Horizon FrequencyIllness9667WeeklyExchange75888Daily",
"figure_id": "tab_14",
"figure_label": "15",
"figure_type": "table"
},
{
"figure_caption": "Model configurations for Illness and Exchange tasks.",
"figure_data": "Dataset patch stride model dim FFN dim dropout blend size learning rate warm-up batch size epochsIllness36116320.322.5e-30128100Exchange 16816320.321e-406410",
"figure_id": "tab_15",
"figure_label": "16",
"figure_type": "table"
},
{
"figure_caption": "Influences on the delay function to the loss function. The best results are in boldface and the second best is underlined.",
"figure_data": "Function",
"figure_id": "tab_16",
"figure_label": "21",
"figure_type": "table"
},
{
"figure_caption": "Metric MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE",
"figure_data": "96 0.329 0.364 0.319 0.349 0.318 0.349 0.316 0.347 0.317 0.347 0.334 0.356 0.345 0.363ETTm1192 0.368 0.385 0.362 0.370 0.361 0.370 0.363 0.370 0.363 0.369 0.379 0.377 0.430 0.416 336 0.400 0.405 0.393 0.391 0.393 0.390 0.393 0.390 0.396 0.391 0.414 0.402 0.605 0.505 720 0.468 0.444 0.459 0.427 0.459 0.427 0.458 0.426 0.466 0.429 0.491 0.449 0.760 0.578avg 0.391 0.400 0.383 0.384 0.383 0.384 0.383 0.383 0.386 0.384 0.405 0.396 0.535 0.49196 0.387 0.399 0.382 0.391 0.382 0.390 0.382 0.390 0.383 0.391 0.387 0.396 0.410 0.413ETTh1192 0.438 0.431 0.437 0.421 0.436 0.420 0.435 0.420 0.436 0.421 0.439 0.426 0.559 0.494 336 0.486 0.454 0.478 0.443 0.478 0.442 0.478 0.442 0.479 0.443 0.485 0.453 0.712 0.566 720 0.480 0.472 0.472 0.462 0.472 0.462 0.471 0.462 0.470 0.460 0.551 0.508 0.786 0.613avg 0.448 0.439 0.442 0.429 0.442 0.429 0.442 0.429 0.442 0.429 0.466 0.396 0.617 0.522",
"figure_id": "tab_17",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Full results for the anomaly detection task. The P, R and F1 represent the precision, recall and F1-score respectively.",
"figure_data": "DatasetsSMDMSLSMAPSWaTPSMAvg F1MetricsPRF1PRF1PRF1PRF1PRF1CARD 0.883 0.861 0.872 0.896 0.750 0.817 92.93 0.794 0.857 0.928 0.962 0.945 0.982 0.933 0.957 0.890PatchTST 0.802 0.942 0.866 0.898 0.760 0.823 89.97 0.566 0.695 0.919 0.899 0.909 0.992 0.913 0.951 0.849MICN0.765 0.838 0.780 0.892 0.752 0.816 0.895 0.518 0.656 0.913 0.841 0.875 0.987 0.885 0.933 0.816TimesNet 0.887 0.831 0.858 0.839 0.864 0.852 0.925 0.583 0.715 0.883 0.962 0.921 0.982 0.968 0.975 0.864Crossformer 0.722 0.844 0.778 0.907 0.749 0.820 0.895 0.541 0.674 0.919 0.856 0.886 0.971 0.876 0.921 0.816ETSformer 0.874 0.792 0.831 0.851 0.849 0.850 0.923 0.558 0.695 0.900 0.804 0.849 0.99.3 0.853 0.918 0.829LightTS 0.871 0.784 0.825 0.824 0.758 0.790 0.926 0.553 0.692 0.920 0.947 0.933 0.984 0.960 0.972 0.842DLinear 0.836 0.715 0.771 0.843 0.854 0.849 0.923 0.554 0.693 0.809 0.953 0.875 0.983 0.893 0.936 0.825FEDformer 0.880 0.824 0.851 0.771 0.801 0.786 0.905 0.581 0.708 0.902 0.964 0.932 0.973 0.972 0.972 0.850Stationary 0.883 0.812 0.846 0.686 0.891 0.775 0.894 0.590 0.711 0.680 0.968 0.799 0.978 0.968 0.973 0.821Autoformer 0.881 0.824 0.851 0.773 0.809 0.791 0.904 0.586 0.711 0.899 0.958 0.927 0.991 0.882 0.933 0.843Informer 0.866 0.773 0.817 0.818 0.865 0.841 0.901 0.571 0.699 0.703 0.968 0.814 0.643 0.963 0.771 0.788L ROBUSTNESS EXPERIMENTSL.1 INFLUENCE OF DIFFERENT INPUT LENGTHS",
"figure_id": "tab_18",
"figure_label": "22",
"figure_type": "table"
},
{
"figure_caption": "Model variants. All models are evaluated on 4 different predication lengths {96, 192, 336, 720}. The best results are in boldface. we use the patching trick, the order of the total complexity would be the same as PatchTST and Crossformer, which is O(L 2 /S 2 ). Some Transformer type models (e.g., FEDformer and Autoformer) may even break the quadratic dependent in L and reach linear or nearly linear complexity in L. Other CNN and RNN type models (e.g., TimesNet and FilM) by nature maintain the O(L) complexity.",
"figure_data": "model dimension163264128ETTh196 192 108.28% 104.49% 1012.58% 168.11% 100.00% 101.66% 109.63% 113.28% 336 113.39% 109.12% 159.18% 275.57% 720 125.71% 182.27% 299.45% 580.04%Weather96 192 104.90% 113.98% 169.15% 100.00% 102.85% 111.58% 336 119.56% 158.80% 274.05% 720 115.42% 167.38% 298.41%200.53% 312.11% 537.06% 604.03%Electricity96 192 102.13% 108.96% 128.34% 100.00% 105.67% 118.21% 336 104.05% 110.66% 133.88% 720 111.04% 112.03% 193.18%125.63% 140.41% 291.26% 481.24%Traffic96 192 103.02% 113.02% 117.51% 100.00% 101.61% 105.43% 336 112.68% 134.68% 168.98% 720 128.17% 183.90% 307.04%386.92% 454.12% 485.92% 561.57%N MODEL COMPLEXITIES AND RUNNING TIME",
"figure_id": "tab_19",
"figure_label": "23",
"figure_type": "table"
},
{
"figure_caption": "The average per step running time in seconds. The input/forecasting lengths are set as 96/96 and we keep the batch size the same for all benchmarks and run the experiments on a single A100/80G GPU. oom is short for out of memory.",
"figure_data": "CARD Autoformer PatchTST Crossformer FEDformer TimesNet MICN Dlinear FilM ETSFormerETTh1Train0.0197 0.10910.01640.25120.21070.0672 0.0423 0.0074 0.0747 0.0714Hidden=16, Batch=128 Inference 0.0046 0.01020.00210.01270.01780.0132 0.0030 0.0009 0.0123 0.0061WeatherTrain0.0779 0.15250.07850.11860.21890.2457 0.0613 0.0330 oom0.1354Hidden=128, Batch=128 Inference 0.0048 0.01390.00360.01230.03780.0224 0.0038 0.0014 oom0.0092ElectricityTrain0.2156 0.08350.3280oom0.1903oom 0.0405 0.0163 oom0.1174Hidden=128, Batch=32 Inference 0.0064 0.01600.0052oom0.0349oom 0.0045 0.0021 oom0.0103TrafficTrain0.2271 0.09600.1329oom0.16490.6048 0.0322 0.0139 oom0.0607Hidden=128, Batch=12 Inference 0.0101 0.01530.0052oom0.03690.0418 0.0074 0.0058 oom0.0223",
"figure_id": "tab_20",
"figure_label": "24",
"figure_type": "table"
},
{
"figure_caption": "Model variants. All models are evaluated on 4 different predication lengths {96, 192, 336, 720}. The best results are in boldface. .346 0.318 0.346 0.326 0.363 0.334 0.368 192 0.363 0.370 0.367 0.370 0.366 0.369 0.366 0.385 0.372 0.387 336 0.393 0.390 0.399 0.391 0.396 0.391 0.400 0.404 0.401 0.407 720 0.458 0.426 0.466 0.429 0.463 0.428 0.459 0.440 0.458 0.438 avg 0.383 0.384 0.388 0.384 0.386 0.384 0.388 0.398 0.391 0.400",
"figure_data": "Models c->t+c (CARD)t->c+tt+ct->cc->tMetricMSEMAEMSE MAE MSE MAE MSE MAE MSE MAEETTm1 0.318 0Weather 96 0.316 0.347 96 0.150 0.188 0.153 0.193 0.152 0.189 0.152 0.191 0.152 0.192 192 0.202 0.238 0.203 0.239 0.201 0.236 0.201 0.239 0.203 0.240 336 0.260 0.282 0.269 0.288 0.261 0.281 0.263 0.284 0.262 0.284 720 0.343 0.335 0.345 0.339 0.344 0.337 0.347 0.339 0.344 0.337avg 0.2390.2610.243 0.265 0.240 0.261 0.241 0.263 0.240 0.263",
"figure_id": "tab_21",
"figure_label": "25",
"figure_type": "table"
},
{
"figure_caption": "Component Ablation Experiments by removing the attention over hidden dimensions (wo. hidden column) and removing the attention over channels (w.o channel) sequentially. All models are evaluated on 4 different predication lengths {96, 192, 336, 720}. The differences in thousandths w.r.t. predecessor models are reported in parentheses.",
"figure_data": "",
"figure_id": "tab_22",
"figure_label": "26",
"figure_type": "table"
},
{
"figure_caption": "Influence of prolonging input sequence. The lookback length is set as96,192,336,720: CARD(96) means using lookback length 96.",
"figure_data": "Models CARD(96) CARD(192) CARD(336) CARD(720)Metric MSE MAE MSE MAE MSE MAE MSE MAE96 0.383 0.391 0.378 0.390 0.372 0.390 0.368 0.392ETTh1192 0.435 0.420 0.427 0.418 0.413 0.416 0.407 0.416 336 0.479 0.442 0.458 0.434 0.437 0.431 0.428 0.430 720 0.471 0.461 0.452 0.456 0.436 0.453 0.418 0.449avg 0.442 0.429 0.429 0.425 0.415 0.422 0.405 0.42196 0.316 0.347 0.296 0.333 0.284 0.328 0.288 0.332ETTm1192 0.363 0.370 0.342 0.359 0.326 0.354 0.332 0.357 336 0.393 0.390 0.375 0.379 0.368 0.377 0.364 0.376 720 0.458 0.426 0.439 0.418 0.428 0.410 0.414 0.407avg 0.383 0.384 0.363 0.372 0.352 0.367 0.349 0.368R.2 IS TRAINING DATA SIZE A LIMITING FACTOR FOR EXISTING LONG-TERM FORECASTINGTRANSFORMERS?",
"figure_id": "tab_23",
"figure_label": "27",
"figure_type": "table"
}
] | Wang Xue; Tian Zhou; Qingsong Wen; Jinyang Gao; Bolin Ding; Rong Jin | [
{
"authors": "Alaaeldin Ali; Hugo Touvron; Mathilde Caron; Piotr Bojanowski; Matthijs Douze; Armand Joulin; Ivan Laptev; Natalia Neverova; Gabriel Synnaeve; Jakob Verbeek",
"journal": "Advances in neural information processing systems",
"ref_id": "b0",
"title": "Xcit: Cross-covariance image transformers",
"year": "2021"
},
{
"authors": "Hangbo Bao; Li Dong; Songhao Piao; Furu Wei",
"journal": "",
"ref_id": "b1",
"title": "BEit: BERT pre-training of image transformers",
"year": "2022"
},
{
"authors": "Cristian Challu; Kin G Olivares; Boris N Oreshkin; Federico Garza; Max Mergenthaler-Canseco; Artur Dubrawski",
"journal": "",
"ref_id": "b2",
"title": "N-hits: Neural hierarchical interpolation for time series forecasting",
"year": "2022"
},
{
"authors": "Si-An Chen; Chun-Liang Li; Sercan O Arik; Nathanael Christian Yoder; Tomas Pfister",
"journal": "Transactions on Machine Learning Research",
"ref_id": "b3",
"title": "TSMixer: An all-MLP architecture for time series forecast-ing",
"year": "2023"
},
{
"authors": "Weidong Chen; Xiaofen Xing; Xiangmin Xu; Jianxin Pang; Lan Du",
"journal": "",
"ref_id": "b4",
"title": "Speechformer: A hierarchical efficient framework incorporating the characteristics of speech",
"year": "2022"
},
{
"authors": "Tri Dao; Dan Fu; Stefano Ermon; Atri Rudra; Christopher Ré",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b5",
"title": "Flashattention: Fast and memoryefficient exact attention with io-awareness",
"year": "2022"
},
{
"authors": "Abhimanyu Das; Weihao Kong; Andrew Leach; Rajat Sen; Rose Yu",
"journal": "",
"ref_id": "b6",
"title": "Long-term forecasting with tide: Time-series dense encoder",
"year": "2023"
},
{
"authors": "Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova",
"journal": "",
"ref_id": "b7",
"title": "Bert: Pre-training of deep bidirectional transformers for language understanding",
"year": "2018"
},
{
"authors": "Mingyu Ding; Bin Xiao; Noel Codella; Ping Luo; Jingdong Wang; Lu Yuan",
"journal": "Springer",
"ref_id": "b8",
"title": "Davit: Dual attention vision transformers",
"year": "2022"
},
{
"authors": "Alexey Dosovitskiy; Lucas Beyer; Alexander Kolesnikov; Dirk Weissenborn; Xiaohua Zhai; Thomas Unterthiner; Mostafa Dehghani; Matthias Minderer; Georg Heigold; Sylvain Gelly",
"journal": "",
"ref_id": "b9",
"title": "An image is worth 16x16 words: Transformers for image recognition at scale",
"year": "2020"
},
{
"authors": "Albert Gu; Karan Goel; Christopher Re",
"journal": "",
"ref_id": "b10",
"title": "Efficiently modeling long sequences with structured state spaces",
"year": "2022"
},
{
"authors": "Lu Han; Han-Jia Ye; De-Chuan Zhan",
"journal": "",
"ref_id": "b11",
"title": "The capacity and robustness trade-off: Revisiting the channel independent strategy for multivariate time series forecasting",
"year": "2023"
},
{
"authors": "Kaiming He; Xinlei Chen; Saining Xie; Yanghao Li; Piotr Dollár; Ross Girshick",
"journal": "",
"ref_id": "b12",
"title": "Masked autoencoders are scalable vision learners",
"year": "2022"
},
{
"authors": "Wei-Ning Hsu; Benjamin Bolte; Hubert Yao-Hung; Kushal Tsai; Ruslan Lakhotia; Abdelrahman Salakhutdinov; Mohamed",
"journal": "IEEE/ACM Transactions on Audio, Speech, and Language Processing",
"ref_id": "b13",
"title": "Hubert: Self-supervised speech representation learning by masked prediction of hidden units",
"year": "2021"
},
{
"authors": "Sergey Ioffe; Christian Szegedy",
"journal": "",
"ref_id": "b14",
"title": "Batch normalization: Accelerating deep network training by reducing internal covariate shift",
"year": "2015"
},
{
"authors": "Ming Jin; Shiyu Wang; Lintao Ma; Zhixuan Chu; James Y Zhang; Xiaoming Shi; Pin-Yu Chen; Yuxuan Liang; Yuan-Fang Li; Shirui Pan",
"journal": "",
"ref_id": "b15",
"title": "Time-LLM: Time series forecasting by reprogramming large language models",
"year": "2023"
},
{
"authors": "Taesung Kim; Jinhee Kim; Yunwon Tae; Cheonbok Park; Jang-Ho Choi; Jaegul Choo",
"journal": "",
"ref_id": "b16",
"title": "Reversible instance normalization for accurate time-series forecasting against distribution shift",
"year": "2022"
},
{
"authors": "P Diederik; Jimmy Kingma; Ba",
"journal": "",
"ref_id": "b17",
"title": "Adam: A Method for Stochastic Optimization",
"year": "2017-01"
},
{
"authors": "Guokun Lai; Wei-Cheng Chang; Yiming Yang; Hanxiao Liu",
"journal": "",
"ref_id": "b18",
"title": "Modeling long-and short-term temporal patterns with deep neural networks",
"year": "2018"
},
{
"authors": "Shiyang Li; Xiaoyong Jin; Xiyou Yao Xuan; Wenhu Zhou; Yu-Xiang Chen; Xifeng Wang; Yan",
"journal": "",
"ref_id": "b19",
"title": "Enhancing the locality and breaking the memory bottleneck of transformer on time series forecasting",
"year": "2019"
},
{
"authors": "Shiyang Li; Xiaoyong Jin; Xiyou Yao Xuan; Wenhu Zhou; Yu-Xiang Chen; Xifeng Wang; Yan",
"journal": "",
"ref_id": "b20",
"title": "Enhancing the locality and breaking the memory bottleneck of transformer on time series forecasting",
"year": "2019"
},
{
"authors": "Zhe Li; Zhongwen Rao; Lujia Pan; Zenglin Xu",
"journal": "",
"ref_id": "b21",
"title": "Mts-mixers: Multivariate time series forecasting via factorized temporal and channel mixing",
"year": "2023"
},
{
"authors": "Yuxuan Liang; Yutong Xia; Songyu Ke; Yiwei Wang; Qingsong Wen; Junbo Zhang; Yu Zheng; Roger Zimmermann",
"journal": "",
"ref_id": "b22",
"title": "Airformer: Predicting nationwide air quality in china with transformers",
"year": "2023"
},
{
"authors": "Bryan Lim; Ö Sercan; Nicolas Arık; Tomas Loeff; Pfister",
"journal": "International Journal of Forecasting",
"ref_id": "b23",
"title": "Temporal fusion transformers for interpretable multi-horizon time series forecasting",
"year": "2021"
},
{
"authors": "Shizhan Liu; Hang Yu; Cong Liao; Jianguo Li; Weiyao Lin; Alex X Liu; Schahram Dustdar",
"journal": "",
"ref_id": "b24",
"title": "Pyraformer: Low-complexity pyramidal attention for long-range time series modeling and forecasting",
"year": "2022"
},
{
"authors": "Yong Liu; Haixu Wu; Jianmin Wang; Mingsheng Long",
"journal": "",
"ref_id": "b25",
"title": "Non-stationary transformers: Exploring the stationarity in time series forecasting",
"year": "2022"
},
{
"authors": "Ze Liu; Yutong Lin; Yue Cao; Han Hu; Yixuan Wei; Zheng Zhang; Stephen Lin; Baining Guo",
"journal": "",
"ref_id": "b26",
"title": "Swin transformer: Hierarchical vision transformer using shifted windows",
"year": "2021"
},
{
"authors": "Xuezhe Ma; Chunting Zhou; Xiang Kong; Junxian He; Liangke Gui; Graham Neubig; Jonathan May; Luke Zettlemoyer",
"journal": "",
"ref_id": "b27",
"title": "Mega: Moving average equipped gated attention",
"year": "2023"
},
{
"authors": "Spyros Makridakis; Evangelos Spiliotis; Vassilios Assimakopoulos",
"journal": "International Journal of Forecasting",
"ref_id": "b28",
"title": "The m4 competition: Results, findings, conclusion and way forward",
"year": "2018"
},
{
"authors": "Yuqi Nie; Nam H Nguyen; Phanwadee Sinthong; Jayant Kalagnanam",
"journal": "",
"ref_id": "b29",
"title": "A time series is worth 64 words: Long-term forecasting with transformers",
"year": "2023"
},
{
"authors": "Kin G Olivares; David Luo; Cristian Challu; Stefania La Vattiata; Max Mergenthaler; Artur Dubrawski",
"journal": "",
"ref_id": "b30",
"title": "Hint: Hierarchical mixture networks for coherent probabilistic forecasting",
"year": "2023"
},
{
"authors": "Boris N Oreshkin; Dmitri Carpov; Nicolas Chapados; Yoshua Bengio",
"journal": "",
"ref_id": "b31",
"title": "N-beats: Neural basis expansion analysis for interpretable time series forecasting",
"year": "2020"
},
{
"authors": " Pems",
"journal": "",
"ref_id": "b32",
"title": "Traffic",
"year": ""
},
{
"authors": "Huajie Qian; Qingsong Wen; Liang Sun; Jing Gu; Qiulin Niu; Zhimin Tang",
"journal": "IEEE",
"ref_id": "b33",
"title": "Robustscaler: Qos-aware autoscaling for complex workloads",
"year": "2022"
},
{
"authors": "Alec Radford; Jeffrey Wu; Rewon Child; David Luan; Dario Amodei; Ilya Sutskever",
"journal": "OpenAI blog",
"ref_id": "b34",
"title": "Language models are unsupervised multitask learners",
"year": "2019"
},
{
"authors": "Alec Radford; Jong Wook Kim; Tao Xu; Greg Brockman; Christine Mcleavey; Ilya Sutskever",
"journal": "",
"ref_id": "b35",
"title": "Robust speech recognition via large-scale weak supervision",
"year": "2022"
},
{
"authors": "Syama Sundar Rangapuram; Matthias Seeger; Jan Gasthaus; Lorenzo Stella; Yuyang Wang; Tim Januschowski",
"journal": "",
"ref_id": "b36",
"title": "Deep state space models for time series forecasting",
"year": "2018"
},
{
"authors": "David Salinas; Valentin Flunkert; Jan Gasthaus; Tim Januschowski",
"journal": "International Journal of Forecasting",
"ref_id": "b37",
"title": "Deepar: Probabilistic forecasting with autoregressive recurrent networks",
"year": "2020"
},
{
"authors": "Rajat Sen; Hsiang-Fu Yu; Inderjit S Dhillon",
"journal": "Advances in neural information processing systems",
"ref_id": "b38",
"title": "Think globally, act locally: A deep neural network approach to high-dimensional time series forecasting",
"year": "2019"
},
{
"authors": "Slawek Smyl",
"journal": "International Journal of Forecasting",
"ref_id": "b39",
"title": "A hybrid method of exponential smoothing and recurrent neural networks for time series forecasting",
"year": "2020"
},
{
"authors": "",
"journal": "",
"ref_id": "b40",
"title": "Electricity",
"year": ""
},
{
"authors": "Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Ł Ukasz Kaiser; Illia Polosukhin",
"journal": "",
"ref_id": "b41",
"title": "Attention is all you need",
"year": "2017"
},
{
"authors": "Chengyi Wang; Sanyuan Chen; Yu Wu; Ziqiang Zhang; Long Zhou; Shujie Liu; Zhuo Chen; Yanqing Liu; Huaming Wang; Jinyu Li",
"journal": "",
"ref_id": "b42",
"title": "Neural codec language models are zero-shot text to speech synthesizers",
"year": "2023"
},
{
"authors": "Huiqiang Wang; Jian Peng; Feihu Huang; Jince Wang; Junhui Chen; Yifei Xiao",
"journal": "",
"ref_id": "b43",
"title": "MICN: Multi-scale local and global context modeling for long-term series forecasting",
"year": "2023"
},
{
"authors": "Haomin Wen; Youfang Lin; Yutong Xia; Huaiyu Wan; Qingsong Wen; Roger Zimmermann; Yuxuan Liang",
"journal": "",
"ref_id": "b44",
"title": "DiffSTG: Probabilistic spatio-temporal graph forecasting with denoising diffusion models",
"year": "2023"
},
{
"authors": "Qingsong Wen; Tian Zhou; Chaoli Zhang; Weiqi Chen; Ziqing Ma; Junchi Yan; Liang Sun",
"journal": "",
"ref_id": "b45",
"title": "Transformers in time series: A survey",
"year": "2023"
},
{
"authors": "Ruofeng Wen; Kari Torkkola; Balakrishnan Narayanaswamy; Dhruv Madeka",
"journal": "",
"ref_id": "b46",
"title": "A multi-horizon quantile recurrent forecaster",
"year": "2017"
},
{
"authors": " Wetterstation",
"journal": "",
"ref_id": "b47",
"title": "Weather",
"year": ""
},
{
"authors": "Gerald Woo; Chenghao Liu; Doyen Sahoo; Akshat Kumar; Steven Hoi",
"journal": "",
"ref_id": "b48",
"title": "Etsformer: Exponential smoothing transformers for time-series forecasting",
"year": "2022"
},
{
"authors": "Haixu Wu; Jiehui Xu; Jianmin Wang; Mingsheng Long",
"journal": "",
"ref_id": "b49",
"title": "Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting",
"year": "2021"
},
{
"authors": "Haixu Wu; Tengge Hu; Yong Liu; Hang Zhou; Jianmin Wang; Mingsheng Long",
"journal": "",
"ref_id": "b50",
"title": "Timesnet: Temporal 2d-variation modeling for general time series analysis",
"year": "2023"
},
{
"authors": "Haixu Wu; Tengge Hu; Yong Liu; Hang Zhou; Jianmin Wang; Mingsheng Long",
"journal": "",
"ref_id": "b51",
"title": "Timesnet: Temporal 2d-variation modeling for general time series analysis",
"year": "2023"
},
{
"authors": "Jiehui Xu; Jianmin Wang; Mingsheng Long",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b52",
"title": "Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting",
"year": "2021"
},
{
"authors": "Ailing Zeng; Muxi Chen; Lei Zhang; Qiang Xu",
"journal": "",
"ref_id": "b53",
"title": "Are transformers effective for time series forecasting",
"year": "2023"
},
{
"authors": "Tianping Zhang; Yizhuo Zhang; Wei Cao; Jiang Bian; Xiaohan Yi; Shun Zheng; Jian Li",
"journal": "",
"ref_id": "b54",
"title": "Less is more: Fast multivariate time series forecasting with light sampling-oriented mlp structures",
"year": "2022"
},
{
"authors": "Yunhao Zhang; Junchi Yan",
"journal": "",
"ref_id": "b55",
"title": "Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting",
"year": "2023"
},
{
"authors": "Haoyi Zhou; Shanghang Zhang; Jieqi Peng; Shuai Zhang; Jianxin Li; Hui Xiong; Wancai Zhang",
"journal": "",
"ref_id": "b56",
"title": "Informer: Beyond efficient transformer for long sequence time-series forecasting",
"year": "2021"
},
{
"authors": "Tian Zhou; Ziqing Ma; Xue Wang; Qingsong Wen; Liang Sun; Tao Yao; Wotao Yin; Rong Jin",
"journal": "",
"ref_id": "b57",
"title": "FiLM: Frequency improved legendre memory model for long-term time series forecasting",
"year": "2022"
},
{
"authors": "Tian Zhou; Ziqing Ma; Qingsong Wen; Xue Wang; Liang Sun; Rong Jin",
"journal": "",
"ref_id": "b58",
"title": "FEDformer: Frequency enhanced decomposed transformer for long-term series forecasting",
"year": "2022"
},
{
"authors": "Tian Zhou; Peisong Niu; Xue Wang; Liang Sun; Rong Jin",
"journal": "",
"ref_id": "b59",
"title": "One fits all: Power general time series analysis by pretrained lm",
"year": "2023"
},
{
"authors": "Chen Zhu; Wei Ping; Chaowei Xiao; Mohammad Shoeybi; Tom Goldstein; Anima Anandkumar; Bryan Catanzaro",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b60",
"title": "Long-short transformer: Efficient transformers for language and vision",
"year": "2021"
},
{
"authors": "Zhaoyang Zhu; Weiqi Chen; Rui Xia; Tian Zhou; Peisong Niu; Bingqing Peng; Wenwei Wang; Hengbo Liu; Ziqing Ma; Xinyue Gu",
"journal": "AI Magazine",
"ref_id": "b61",
"title": "Energy forecasting with robust, flexible, and explainable machine learning algorithms",
"year": "2023"
},
{
"authors": "",
"journal": "",
"ref_id": "b62",
"title": "Via standard generalization analysis procedures and Hoeffiding's Inequality, we have with probability 1 -δ",
"year": ""
},
{
"authors": "",
"journal": "T log",
"ref_id": "b63",
"title": "",
"year": null
},
{
"authors": "",
"journal": "MSE MAE MSE diff MAE diff MSE diff MAE diff ETTm",
"ref_id": "b64",
"title": "T t=1 x 2 t . (14) Models CARD wo. hidden wo",
"year": ""
}
] | [
{
"formula_coordinates": [
3,
256.58,
717.22,
248.08,
12.2
],
"formula_id": "formula_0",
"formula_text": "X = [T 0 , F 1 ( X) + E],(1)"
},
{
"formula_coordinates": [
4,
220.04,
153.1,
284.62,
9.68
],
"formula_id": "formula_1",
"formula_text": "Q = F q (X), K = F k (X), V = F v (X),(2)"
},
{
"formula_coordinates": [
4,
107.53,
168.7,
396.47,
34.28
],
"formula_id": "formula_2",
"formula_text": "Q, K, V ∈ R C×(N +1)×d and F q , F k , F v are MLP layers. We next convert Q, K, V into {Q i },{K i },{V i } where Q i , K i , V i ∈ R C×(N +1)×d head , i ="
},
{
"formula_coordinates": [
4,
108,
291.6,
397.25,
24.76
],
"formula_id": "formula_3",
"formula_text": "Q i , K i and V i on channel dimension into {Q c: i }, {K c: i } and {V c: i } with Q c: i , K c: i , V c: i ∈ R (N +1"
},
{
"formula_coordinates": [
4,
200.46,
356.68,
304.21,
23.67
],
"formula_id": "formula_4",
"formula_text": "A c: i1 = softmax 1 √ d • EMA(Q c: i ) (EMA(K c: i )) ⊤(3)"
},
{
"formula_coordinates": [
4,
200.46,
384.58,
304.21,
23.61
],
"formula_id": "formula_5",
"formula_text": "A c: i2 = softmax 1 √ N • (Q c: i ) ⊤ K c: i ,(4)"
},
{
"formula_coordinates": [
4,
136.04,
413.06,
140.84,
15.65
],
"formula_id": "formula_6",
"formula_text": "A c: i1 ∈ R (N +1)×(N +1) , A c: i2 ∈ R"
},
{
"formula_coordinates": [
4,
238.08,
522.86,
266.59,
12.69
],
"formula_id": "formula_7",
"formula_text": "O c: i1 = A c: i1 V c i , O c: i2 = V c: i A c: i2 .(5)"
},
{
"formula_coordinates": [
5,
187.2,
439.4,
317.47,
12.69
],
"formula_id": "formula_8",
"formula_text": "P :n ki = softmax(F pk (K :n i )), P :n vi = softmax(F pv (V :n i )),(6)"
},
{
"formula_coordinates": [
5,
224.3,
477.22,
280.37,
13.14
],
"formula_id": "formula_9",
"formula_text": "K:n i = (P :n ki ) ⊤ K :n i , Ṽ :n i = (P :n vi ) ⊤ V :n i ,(7)"
},
{
"formula_coordinates": [
5,
137.12,
494.61,
86.55,
15.65
],
"formula_id": "formula_10",
"formula_text": "K:n i , Ṽ :n i ∈ R r×d head ."
},
{
"formula_coordinates": [
5,
107.64,
527.63,
397.52,
21.49
],
"formula_id": "formula_11",
"formula_text": "O(L/S • C • r • d 2 ) which is smaller than O(L/S • C 2 • d 2 ) cost of the standard attention."
},
{
"formula_coordinates": [
6,
213.83,
455.99,
286.96,
30.55
],
"formula_id": "formula_12",
"formula_text": "min E A 1 L L l=1 ∥â t+l (A) -a t+l (A)∥ 2 2 . (8"
},
{
"formula_coordinates": [
6,
500.8,
466.73,
3.87,
8.64
],
"formula_id": "formula_13",
"formula_text": ")"
},
{
"formula_coordinates": [
6,
203.92,
568.9,
300.75,
11.72
],
"formula_id": "formula_14",
"formula_text": "var(a t+1 ) = var(G(a t )) + σ 2 I ⪯ var(a t ) + σ 2 I,(9)"
},
{
"formula_coordinates": [
6,
117.15,
664.49,
383.37,
65.73
],
"formula_id": "formula_15",
"formula_text": "min E A 1 2 L l=1 (â t+l (A) -a t+l (A)) ⊤ var (a t+l ) -1 (â t+l (A) -a t+l (A)) ≥E A 1 2 L l=1 ∥â t+l (A) -a t+l (A)∥ 2 2 lσ 2 ∝ E A 1 L L l=1 l -1 ∥â t+l (A) -a t+l (A)∥ 2 2 . (11"
},
{
"formula_coordinates": [
6,
500.52,
710.4,
4.15,
8.64
],
"formula_id": "formula_16",
"formula_text": ")"
},
{
"formula_coordinates": [
7,
207.02,
126.3,
297.64,
30.55
],
"formula_id": "formula_17",
"formula_text": "min E A 1 L L l=1 l -1/2 ∥â t+l (A) -a t+l (A)∥ 1 ,(12)"
},
{
"formula_coordinates": [
17,
120.91,
363.18,
241.02,
26.57
],
"formula_id": "formula_18",
"formula_text": "[i][j] = ema_matrix[i-1][j] * (1-alpha) ema_matrix[i][i] = alpha self.register_buffer('ema_matrix',ema_matrix) def ema(self,"
},
{
"formula_coordinates": [
22,
148.62,
539.05,
217.96,
10.53
],
"formula_id": "formula_19",
"formula_text": "f (t) = t -1/4 , f (t) = t -1/3 , f (t) = t -1 , f (t) = t -2"
},
{
"formula_coordinates": [
22,
215.94,
700.01,
288.73,
30.2
],
"formula_id": "formula_20",
"formula_text": "min β T t=1 [∥x t β -x t+1 ∥ 2 2 + ∥x t β 2 -x t+2 ∥ 2 2 ].(13)"
},
{
"formula_coordinates": [
24,
196.51,
106.19,
246.55,
7.22
],
"formula_id": "formula_21",
"formula_text": "f (t) = 1 f (t) = t -0.25 f (t) = t -0.33 f (t) = t -0.5 f (t) = t -1 f (t) = t -2 f (t) = t -"
},
{
"formula_coordinates": [
24,
212.25,
250.06,
292.42,
30.2
],
"formula_id": "formula_22",
"formula_text": "min β T t=1 [∥x t β -x t+1 ∥ 2 2 + 1 2 ∥x t β 2 -x ∥ 2 2 ].(15)"
},
{
"formula_coordinates": [
24,
243.52,
305.84,
257,
31.62
],
"formula_id": "formula_23",
"formula_text": "|β -β true | ≤ 3 2 T log(1/δ) T t=1 x 2 t . (16"
},
{
"formula_coordinates": [
24,
500.52,
317.96,
4.15,
8.64
],
"formula_id": "formula_24",
"formula_text": ")"
}
] |
|
10.1162/tacl_a_00338 | 2023-05-20 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [],
"table_ref": [],
"text": "Pre-training on large corpora of text enables the natural language processing (NLP) models to acquire a vast amount of factual and commonsense 1 Data is available at https://github.com/ nrjvarshney/break_the_common_assumptions Binary Classification Question: John prepared the soup in the morning and left it in the open for an hour, will he like the soup now?"
},
{
"figure_ref": [
"fig_0"
],
"heading": "Context (Breaking):",
"publication_ref": [
"b11",
"b18",
"b24",
"b5",
"b20",
"b10",
"b23",
"b6",
"b17",
"b15",
"b25",
"b7",
"b13",
"b1",
"b16",
"b12",
"b0",
"b19",
"b14",
"b8",
"b4",
"b9"
],
"table_ref": [],
"text": "John likes to have tomato soup only when it is cold.\nCommon Assumption: Generally, people like to consume soup when it is hot. We show that state-of-the-art NLP models while performing well in reasoning over contexts that follow the common assumptions struggle to reason over contexts that break them.\nknowledge (Liu et al., 2019;Petroni et al., 2019;Yogatama et al., 2019;Davison et al., 2019). Due to this knowledge, they are able to achieve remarkable performance on a variety of language understanding tasks. They typically acquire this knowledge by learning from the pre-training text and capturing certain patterns from it. However, in real-world settings, we often encounter scenarios that do not abide by these patterns i.e. scenarios that break the common assumptions. Consider a context, 'John likes to have tomato soup only when it is cold', this breaks the common assumption that 'people prefer to consume soup when it is hot'. Answering questions based on such contexts requires a model to truly understand the context and override its knowledge that it may have acquired (due to the predominant presence of certain patterns in the raw text) during pre-training. How well can state-ofthe-art NLP models perform in such scenarios?\nRecently, many datasets have been created that test different language understanding skills such as pronoun resolution (Sakaguchi et al., 2021;Levesque et al., 2012), commonsense reasoning (Talmor et al., 2019), numerical reasoning (Dua et al., 2019;Patel et al., 2021;Mishra et al., 2022), qualitative reasoning (Tafjord et al., 2019b,a), temporal reasoning (Zhou et al., 2019), and feasibility reasoning (Gupta et al., 2022). Furthermore, numerous adversarial datasets (McCoy et al., 2019;Bartolo et al., 2020;Naik et al., 2018) have also been developed that test the robustness of models. Longpre et al. (2021) study entity-based conflicts in the parametric and contextual knowledge. Agarwal et al. (2020) investigate entity-based swapping to test the robustness of models. Prior work has also studied creating counterfactuals using various techniques such as token substitutions and adversarial attacks (Ribeiro et al., 2020;Michel et al., 2019;Kaushik et al., 2020). However, evaluating models on the ability to reason over contexts that break the common assumptions (this is different from entitybased conflicts) has remained underexplored, and existing datasets do not contain a sufficient number of such examples.\nIn this work, we address the above limitations and comprehensively study the models' ability to reason over contexts that break the common assumptions. To this end, we first systematically create questions (binary classification) in which the contexts break the common assumptions and the questions test the ability to reason over those contexts. Furthermore, for each such context, we also create a corresponding context that 'follows' the common assumption. Specifically, instances in our evaluation data consist of the following: (a) a common assumption, (b) a context that follows the assumption, (c) a context that breaks the assumption, and (d) questions based on the contexts. Figure 1 illustrates examples of our dataset. For binary classification questions, the task is to answer a given question as either 'Yes' or 'No'.\nWe conduct comprehensive experiments with several NLP models such as Flan T5 (Chung et al., 2022), GPT-3 (Brown et al., 2020), and UnifiedQA (Khashabi et al., 2020). First, we evaluate models on the scenario where the contexts follow the common assumptions; we show that the models perform fairly well in this setting. However, on evaluating them for the scenario where the contexts break the common assumptions, we find that the models falter and achieve considerably lower performance. Specifically, on the binary classification questions, Flan T5-xxl achieves an accuracy of just 70.67% in the latter scenario (∼ 20 absolute points lower than its performance on the former scenario). Furthermore, we show that this performance is considerably and consistently lower than the human performance baseline.\nWe further conduct a thorough analysis which reveals several interesting findings such as (a) models show poor consistency i.e. they are often not able to correctly answer both (context-question) and (context (Breaking)-question) pairs correctly and (b) explicitly providing the common assumption along with the context improves the performance when the context aligns withe the assumption but degrades when it breaks the assumption. Overall, we believe our work and findings will encourage and facilitate further research in developing more robust models that can also reliably reason over contexts that break the common assumptions."
},
{
"figure_ref": [],
"heading": "Evaluation Data",
"publication_ref": [],
"table_ref": [],
"text": "In order to comprehensively study a system's ability to reason over contexts that break the common assumptions, we first systematically create evaluation instances. In this section, we describe the data creation process and provide supporting details."
},
{
"figure_ref": [],
"heading": "Data Creation",
"publication_ref": [],
"table_ref": [],
"text": "For creating data instances, we first compile a set of common assumptions across various categories, namely assumptions about preferences, behaviors, objects, and events. Table 1 demonstrates examples of common assumptions for each category. Then, we write a context that follows the common assumption and a corresponding context that breaks that assumption. Finally, we create binary classification questions from these contexts. Furthermore, we also create several variants of a (context, question) pair to comprehensively evaluate a system's ability to correctly and consistently answer questions. Table 2 shows examples of such variants. We note that in this work, our focus is on common assumptions and not on entity-based factual knowledge.\nSix computer science graduate students contributed to the development of this dataset. The data instances were cross-verified and instances on which the inter-annotator agreement was low were rejected. We also conduct validation of the compiled common assumptions; specifically, for each sentence, we asked human annotators to answer 'Yes' if they think that it is a common assumption otherwise answer 'No'. For nearly all the compiled common assumptions, the majority answer is 'Yes' which posits that they are indeed common assumptions. We provide further details about this step in section 2.3."
},
{
"figure_ref": [],
"heading": "Categories of common assumptions:",
"publication_ref": [],
"table_ref": [],
"text": "We create common assumptions for the following categories: Assumptions about Preferences: In this category, we include assumptions where a preference (typically of humans) is involved; for e.g. \"Generally, people prefer to eat fruits when they are fully ripened\", \"Generally, busy people prefer to have an assistant who can help them with their tasks\", and \"People usually like to go outside when the weather is pleasant\".\nAssumptions about Behaviors: Here, we include assumptions about people's behaviors such as 'Generally, people feel good when they meet an old friend', 'Generally, people like to get free coupons', and 'People usually go to work in the morning'.\nAssumptions about Objects: This category incorporates assumptions about objects/things such as 'Generally, hotels are more expensive than a dormitory', 'Generally, bigger vehicles have more seating capacity', and 'Generally, schools have science laboratories'.\nAssumptions about Events: In this category, we include assumptions about events such as 'Generally, football games have an audience' and 'Generally, there are food stalls in a carnival celebration'.\nWe also include an Others category to incorporate common assumptions that do not fit into the above four categories."
},
{
"figure_ref": [],
"heading": "Data Statistics",
"publication_ref": [],
"table_ref": [],
"text": "For binary classification questions, the task is to answer a given question as either 'Yes' or 'No'. To further measure the consistency of a system's predictions, we evaluate its predictions on context"
},
{
"figure_ref": [],
"heading": "Context",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Question",
"publication_ref": [],
"table_ref": [],
"text": "Matt always enjoys watching onesided sports game Q1: There are two matches tonight. One is high-intensity close match. Other is a boring one-sided game. Will Matt watch the close match? No Q2: There are two matches tonight. One is high-intensity close match. Other is a boring one-sided game. Will Matt watch the one-sided match? Yes Matt doesn't enjoy watching interesting sports game but likes onesided games Q1: There are two matches tonight. One is high-intensity close match. Other is a boring one-sided game. Will Matt watch the close match? No Q2: There are two matches tonight. One is a high-intensity close match and the ather is a boring one-sided game. Will Matt prefer to watch the one-sided match? Yes Table 2: Illustrative examples of variations of a (context, question) pair in our dataset. This is used to comprehensively evaluate a system's ability to correctly and consistently answer questions. pairs where one context follows the common assumption and a corresponding context that breaks it. We also conduct evaluations on different variations of a (context, question) pair as shown in Table 2."
},
{
"figure_ref": [],
"heading": "Category (# Assumptions",
"publication_ref": [],
"table_ref": [
"tab_1"
],
"text": "Table 3 shows the number of binary classification questions in our dataset across each category."
},
{
"figure_ref": [],
"heading": "Data Validation",
"publication_ref": [],
"table_ref": [],
"text": "We note that it is important to validate the quality of the compiled common assumptions. To this end, for each sentence, we ask 3 human annotators to answer 'Yes' if they think that the given sentence is a common assumption otherwise answer 'No'. Then, we use the majority voting aggregation strategy and find that for nearly all the compiled common assumptions, the majority answer is 'Yes'. This validates the quality of the common assumptions compiled in this work.\nIn addition to the above validation step, we note that the questions were also cross-verified by the data creators (who are also the authors of this paper) and the instances where the inter-annotator agreement was low were rejected."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Experimental Setup",
"publication_ref": [],
"table_ref": [],
"text": "Performance Metrics: For binary classification questions, the task is to answer a given question as either 'Yes' or 'No'. We calculate accuracy against the gold labels (Yes and No) for evaluation. To better evaluate a system's capability, we measure its consistency in correctly answering both the scenarios corresponding to the context that follows the common assumption and the context that breaks it."
},
{
"figure_ref": [],
"heading": "Models:",
"publication_ref": [
"b2",
"b9"
],
"table_ref": [],
"text": "We evaluate Flan T5 (Chung et al., 2022), GPT-3 (text-davinci-003) (Brown et al., 2020), and UnifiedQA (Khashabi et al., 2020) models on our task.\nHuman Performance Baseline: We randomly select 40 context-question pairs (20 for contexts that follow common assumptions and 20 for corresponding contexts that break those assumptions) for each category and ask a total of 3 human annotators to 'answer the given question in Yes or No based on the context'. We then use the majority voting aggregation method and calculate the human performance baseline."
},
{
"figure_ref": [],
"heading": "Results",
"publication_ref": [],
"table_ref": [],
"text": "Table 4 shows the performance of different models on binary classification questions. Column 'Con' and column 'Con (B)' correspond to the performance on context-question pairs where contexts follow the common assumptions and where contexts break the assumptions respectively.\nHigh Human Performance Baseline: The first row in Table 4 shows the human performance baseline for each category of our evaluation data. It demonstrates that humans typically achieve high performance across all the data categories. This Table 4: Performance of different models on binary classification questions for each category of our evaluation data. Column 'Con' and column 'Con (B)' correspond to the performance on context-question pairs where contexts follow the common assumptions and where contexts break the assumptions respectively.\nshows that typically humans are able to reason well in both the scenarios i.e. where contexts follow the common assumptions and where contexts break those assumptions. On average, the human performance is 99% on 'Con' and 96% on 'Con (B)'.\nCon vs Con (B) Performance: On comparing the performance on questions for contexts that follow the common assumptions ('Con') and for contexts that break them ('Con (B)'), we find that the models consistently achieve lower performance on 'Con (B)'. This behavior is observed for all the models and for all categories of common assumptions. For instance, Flan T5-xxl model on average achieves 89.19% accuracy on 'Con' and just 70.67% on 'Con (B)'. The gap in performance is observed for all the categories of common assumptions. The table also shows that with the increase in the size of the model, the performance on both 'Con' and 'Con (B)' improves. However, the gap in performance on them remains. This highlights that despite performing fairly well on reasoning over the contexts that follow the common assumptions, the models struggle to correctly reason over contexts that break those common assumptions.\nHuman vs Model Performance on 'Con (B)': Table 4 shows that the performances of all models are considerably lower than the human performance baseline. Specifically, on 'Con (B)' instances, the human performance on average is ∼ 26% higher than the Flan T5-xxl model. Furthermore, human performance is just slightly impacted when the contexts break the common assumptions (i.e. 'Con' column); however, the models' performance degrades significantly. This behavior is observed for all the categories. based on both Context and Context (Breaking)) achieved by different models on the binary classification questions. The results show that all the models achieve poor consistency i.e. they are often not able to correctly answer both (context-question) and (context (Breaking)-question) pairs correctly. This is primarily due to the poor performance on (context (Breaking)-question) instances."
},
{
"figure_ref": [],
"heading": "Models Show Poor Consistency:",
"publication_ref": [],
"table_ref": [
"tab_4",
"tab_5"
],
"text": "Impact of Explicitly Providing the Common Assumption with the Context: Table 6 shows the impact of explicitly providing the common assumption along with the context. Since the common assumption aligns with the 'Con' contexts, it slightly improves the performance on 'Con'; however, it hurts the performance on 'Con (B)'. This happens because the contexts in 'Con (B)' break the provided common assumptions. Hence, it further distracts the model resulting in a drop in performance.\nFailure Instances: Table 7 shows examples of instances where Flan T5-xxl model gave incorrect predictions. On analyzing the failure instances, we find that a large fraction of the mistakes are on the instances where the correct answer is 'Yes' while the model gives 'No' as its prediction.\nPerformance on instance variations: -question) pairs i.e. if a model predicts all the variants corresponding to a common assumption correctly then we give it a score of 1 otherwise we give it 0. Flan T5-xxl achieves a performance of just 33.99% on this metric highlighting that the model is often not able to consistently answer ALL the variants correctly."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "In this paper, we investigated the ability of models to correctly reason over contexts that break the common assumptions. To this end, we first systematically developed evaluation data that consists of a common assumption, a context that follows that as-sumption, a context that breaks the assumption, and question based on the contexts. Then, we evaluated multiple models and show that while performing fairly well on contexts that follow the common assumptions, the models struggle to correctly reason over contexts that break those assumptions. Furthermore, we conducted a thorough analysis which resulted in several interesting findings. In conclusion, we believe our work and findings will encourage and facilitate further research in developing more robust models that can also reliably reason over contexts that break the common assumptions."
},
{
"figure_ref": [],
"heading": "Ethical Considerations",
"publication_ref": [],
"table_ref": [],
"text": "The names used in our data are selected from the most common English names. Though the contexts in our dataset break the common assumption, we ensure that all of them indeed describe a realistic scenario. We do not collect any personal information from data creators in the development of the evaluation data for this work."
},
{
"figure_ref": [],
"heading": "Acknowledgement",
"publication_ref": [],
"table_ref": [],
"text": "We thank the Research Computing (RC) at Arizona State University (ASU) for providing computing resources for experiments."
}
] | Pre-training on large corpora of text enables the language models to acquire a vast amount of factual and commonsense knowledge which allows them to achieve remarkable performance on a variety of language understanding tasks. They typically acquire this knowledge by learning from the pre-training text and capturing certain patterns from it. However, realworld settings often present scenarios that do not abide by these patterns i.e. scenarios that break the common assumptions. Can state-ofthe-art NLP models correctly reason over the contexts of such scenarios? Addressing the above question, in this paper, we investigate the ability of models to correctly reason over contexts that break the common assumptions. To this end, we first systematically create evaluation data in which each data instance consists of (a) a common assumption, (b) a context that follows the assumption, (c) a context that breaks the assumption, and (d) questions based on the contexts. Then, through evaluations on multiple models including GPT-3 and Flan T5, we show that while doing fairly well on contexts that follow the common assumptions, the models struggle to correctly reason over contexts that break those assumptions. Specifically, the performance gap is as high as 20% absolute points. Furthermore, we thoroughly analyze these results revealing several interesting findings. We believe our work and findings will encourage and facilitate further research in developing more robust models that can also reliably reason over contexts that break the common assumptions 1 . | Can NLP Models Correctly Reason Over Contexts that Break the Common Assumptions? | [
{
"figure_caption": "Figure 1 :1Figure1: Illustrative examples of binary classification questions created in our study. Context follows the abovementioned common assumption while Context (Breaking) breaks it. We show that state-of-the-art NLP models while performing well in reasoning over contexts that follow the common assumptions struggle to reason over contexts that break them.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Number of binary classification questions for each category in our evaluation data.",
"figure_data": ")# QuestionsPreferences (33)131Behaviors (64)240Objects (17)73Events (26)95Others (13)44",
"figure_id": "tab_1",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Table 5 shows the consistency (correctly answering a question Consistency of different models on the binary classification questions.",
"figure_data": "ModelPrefBehObjEveOthFlan T5-xxl56.3 58.68 58.9 68.42 50.0Flan T5-xl54.81 54.96 56.16 55.79 45.45Flan T5-large 47.41 40.91 42.47 48.42 34.09Flan T5-base 24.44 32.23 26.03 35.79 20.45UnifiedQA28.89 19.42 21.92 26.32 15.91GPT-349.62 47.08 41.1 46.32 56.82",
"figure_id": "tab_3",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "Table8shows the overall performance of different models on different variations of (context (Breaking) Performance of different models on binary classification questions (when the common assumption is explicitly provided with the context) for each category of our evaluation data. Column 'Con' and column 'Con (B)' correspond to the performance on context-question pairs where contexts follow the common assumptions and where contexts break the assumptions respectively.",
"figure_data": "ModelPreferencesBehaviorsObjectsEventsOthersAverageCon Con (B) Con Con (B) Con Con (B) Con Con (B) Con Con (B) Con Con (B)Flan T5-xxl91.665.6592.0870.8393.1561.6493.6873.6888.6459.0992.1168.1Flan T5-xl87.7964.8992.0863.3389.0463.0190.5360.088.6454.5590.2262.44Flan T5-large 84.7366.4187.0860.4280.8246.5886.3258.9584.0956.8285.4259.52Flan T5-base 64.1254.9675.052.9265.7560.2778.9548.4265.9145.4571.3653.0UnifiedQA66.4138.9368.3344.5857.5350.6865.2647.3754.5547.7365.0144.77Context (Breaking)Question (Answer)PredictionRonald never hires a person that is experi-Joan is an inexperienced candidate applying for theNoenced to handle his business.position. will he be considered for hiring? (Yes)John is content with his small apartment andHis parents offered to help him buy a bigger home, willNowants to continue to stay herehe decline the offer? (Yes)John enjoys in small homes so that he canJohn's parents are looking for a new bungalow for him,Yesmanage it properlywill he like it? (No)Steven's has an old car that is even slowerSteven rides his bicycle and car for one hour, will heNothan a bicyclecover more distance with bicycle? (Yes)Matt always enjoys watching boring sportsThere are two matches tonight. One is high intensityNogameclose match. Other is a boring one-sided game. WillMatt watch the one-sided match? (Yes)",
"figure_id": "tab_4",
"figure_label": "6",
"figure_type": "table"
},
{
"figure_caption": "Examples of errors in prediction made by Flan T5-xxl model on the binary classification questions.",
"figure_data": "ModelPerformanceFlan T5-xxl33.99Flan T5-xl31.7Flan T5-large21.57Flan T5-base15.36",
"figure_id": "tab_5",
"figure_label": "7",
"figure_type": "table"
},
{
"figure_caption": "Performance of different models on different variations of (context (Breaking)-question) pairs.",
"figure_data": "",
"figure_id": "tab_6",
"figure_label": "8",
"figure_type": "table"
}
] | Neeraj Varshney; Mihir Parmar; Nisarg Patel; Divij Handa; Sayantan Sarkar; Man Luo; Chitta Baral | [
{
"authors": "Oshin Agarwal; Yinfei Yang; Byron C Wallace; Ani Nenkova",
"journal": "",
"ref_id": "b0",
"title": "Entity-switched datasets: An approach to auditing the in-domain robustness of named entity recognition models",
"year": "2020"
},
{
"authors": "Max Bartolo; Alastair Roberts; Johannes Welbl; Sebastian Riedel; Pontus Stenetorp",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b1",
"title": "Beat the AI: Investigating adversarial human annotation for reading comprehension",
"year": "2020"
},
{
"authors": "Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell; Sandhini Agarwal; Ariel Herbert-Voss; Gretchen Krueger; Tom Henighan; Rewon Child; Aditya Ramesh; Daniel Ziegler; Jeffrey Wu; Clemens Winter; Chris Hesse; Mark Chen; Eric Sigler; Mateusz Litwin; Scott Gray; Benjamin Chess; Jack Clark; Christopher Berner; Sam Mccandlish; Alec Radford; Ilya Sutskever; Dario Amodei",
"journal": "",
"ref_id": "b2",
"title": "Language models are few-shot learners",
"year": "2020"
},
{
"authors": "Curran Associates; Inc ",
"journal": "",
"ref_id": "b3",
"title": "",
"year": ""
},
{
"authors": "Chung Hyung Won; Le Hou; Shayne Longpre; Barret Zoph; Yi Tay; William Fedus; Eric Li; Xuezhi Wang; Mostafa Dehghani; Siddhartha Brahma",
"journal": "",
"ref_id": "b4",
"title": "Scaling instruction-finetuned language models",
"year": "2022"
},
{
"authors": "Joe Davison; Joshua Feldman; Alexander Rush",
"journal": "",
"ref_id": "b5",
"title": "Commonsense knowledge mining from pretrained models",
"year": "2019"
},
{
"authors": "Dheeru Dua; Yizhong Wang; Pradeep Dasigi; Gabriel Stanovsky; Sameer Singh; Matt Gardner",
"journal": "Association for Computational Linguistics",
"ref_id": "b6",
"title": "DROP: A reading comprehension benchmark requiring discrete reasoning over paragraphs",
"year": "2019"
},
{
"authors": "Himanshu Gupta; Neeraj Varshney; Swaroop Mishra; Kuntal Kumar Pal; Arjun Saurabh; Kevin Sawant; Siddharth Scaria; Chitta Goyal; Baral",
"journal": "",
"ref_id": "b7",
"title": "john is 50 years old, can his son be 65?\" evaluating nlp models' understanding of feasibility",
"year": "2022"
},
{
"authors": "Divyansh Kaushik; Eduard Hovy; Zachary C Lipton",
"journal": "ICLR",
"ref_id": "b8",
"title": "Learning the difference that makes a difference with counterfactually augmented data",
"year": "2020"
},
{
"authors": "Daniel Khashabi; Sewon Min; Tushar Khot; Ashish Sabharwal; Oyvind Tafjord; Peter Clark; Hannaneh Hajishirzi",
"journal": "",
"ref_id": "b9",
"title": "UNIFIEDQA: Crossing format boundaries with a single QA system",
"year": "2020"
},
{
"authors": "Hector J Levesque; Ernest Davis; Leora Morgenstern",
"journal": "AAAI Press",
"ref_id": "b10",
"title": "The Winograd Schema Challenge",
"year": "2012"
},
{
"authors": "Nelson F Liu; Matt Gardner; Yonatan Belinkov; Matthew E Peters; Noah A Smith",
"journal": "Association for Computational Linguistics",
"ref_id": "b11",
"title": "Linguistic knowledge and transferability of contextual representations",
"year": "2019"
},
{
"authors": "Shayne Longpre; Kartik Perisetla; Anthony Chen; Nikhil Ramesh; Chris Dubois; Sameer Singh",
"journal": "Association for Computational Linguistics",
"ref_id": "b12",
"title": "Entity-based knowledge conflicts in question answering",
"year": "2021"
},
{
"authors": "Tom Mccoy; Ellie Pavlick; Tal Linzen",
"journal": "Association for Computational Linguistics",
"ref_id": "b13",
"title": "Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference",
"year": "2019"
},
{
"authors": "Paul Michel; Xian Li; Graham Neubig; Juan Pino",
"journal": "Association for Computational Linguistics",
"ref_id": "b14",
"title": "On evaluation of adversarial perturbations for sequence-to-sequence models",
"year": "2019"
},
{
"authors": "Swaroop Mishra; Arindam Mitra; Neeraj Varshney; Bhavdeep Sachdeva; Peter Clark; Chitta Baral; Ashwin Kalyan",
"journal": "Association for Computational Linguistics",
"ref_id": "b15",
"title": "NumGLUE: A suite of fundamental yet challenging mathematical reasoning tasks",
"year": "2022"
},
{
"authors": "Aakanksha Naik; Abhilasha Ravichander; Norman Sadeh; Carolyn Rose; Graham Neubig",
"journal": "Association for Computational Linguistics",
"ref_id": "b16",
"title": "Stress test evaluation for natural language inference",
"year": "2018"
},
{
"authors": "Arkil Patel; Satwik Bhattamishra; Navin Goyal",
"journal": "Association for Computational Linguistics",
"ref_id": "b17",
"title": "Are NLP models really able to solve simple math word problems",
"year": "2021"
},
{
"authors": "Fabio Petroni; Tim Rocktäschel; Sebastian Riedel; Patrick Lewis; Anton Bakhtin; Yuxiang Wu; Alexander Miller",
"journal": "",
"ref_id": "b18",
"title": "Language models as knowledge bases?",
"year": "2019"
},
{
"authors": "Marco Tulio Ribeiro; Tongshuang Wu; Carlos Guestrin; Sameer Singh",
"journal": "Association for Computational Linguistics",
"ref_id": "b19",
"title": "Beyond accuracy: Behavioral testing of NLP models with CheckList",
"year": "2020"
},
{
"authors": "Keisuke Sakaguchi; Le Ronan; Chandra Bras; Yejin Bhagavatula; Choi",
"journal": "Communications of the ACM",
"ref_id": "b20",
"title": "Winogrande: An adversarial winograd schema challenge at scale",
"year": "2021"
},
{
"authors": "Oyvind Tafjord; Peter Clark; Matt Gardner; Wen-Tau Yih; Ashish Sabharwal",
"journal": "",
"ref_id": "b21",
"title": "a. Quarel: A dataset and models for answering questions about qualitative relationships",
"year": "2019"
},
{
"authors": "Oyvind Tafjord; Matt Gardner; Kevin Lin; Peter Clark",
"journal": "Association for Computational Linguistics",
"ref_id": "b22",
"title": "QuaRTz: An open-domain dataset of qualitative relationship questions",
"year": "2019"
},
{
"authors": "Alon Talmor; Jonathan Herzig; Nicholas Lourie; Jonathan Berant",
"journal": "Association for Computational Linguistics",
"ref_id": "b23",
"title": "CommonsenseQA: A question answering challenge targeting commonsense knowledge",
"year": "2019"
},
{
"authors": "Dani Yogatama; Cyprien De Masson D'autume; Jerome Connor; Tomas Kocisky; Mike Chrzanowski; Lingpeng Kong; Angeliki Lazaridou; Wang Ling; Lei Yu; Chris Dyer",
"journal": "",
"ref_id": "b24",
"title": "Learning and evaluating general linguistic intelligence",
"year": "2019"
},
{
"authors": "Ben Zhou; Daniel Khashabi; Qiang Ning; Dan Roth",
"journal": "Association for Computational Linguistics",
"ref_id": "b25",
"title": "going on a vacation\" takes longer than \"going for a walk\": A study of temporal commonsense understanding",
"year": "2019"
}
] | [] |
2023-05-20 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b16",
"b10",
"b30",
"b13",
"b1",
"b29"
],
"table_ref": [],
"text": "Common sense is more than just a collection of facts, but a set of mental models that allow individuals to understand and navigate the world around them (Minsky, 1988). Cognitive science research has shown that specialized modules in the human brain coordinate to support complex cognitive processes. The Thousands Brains theory (Hawkins et al., 2019) suggests that the neocortex learns complete models of objects through many models of each object distributed throughout cortical columns of the neocortex. However, conventional deep neural networks lack sufficient modularity and sparsity. For example, ChatGPT (OpenAI, 2023) is built upon large-scale training and a monolithic Transformer architecture. A limited degree of alignment between its beliefs and those of a human user is performed via\nThe Future of Life Institute has called for a six-month moratorium on the development of AI research to allow AI companies and regulators time to formulate safeguards to protect society from potential risks of the technology. Furthermore, there have been calls and joint workshops to reunite machine learning research with other scientific disciplines, including neuroscience, cognitive science, mathematics, and psychology (Zador et al., 2023;Juliani et al., 2022). In this thought piece, we propose a potential avenue to address problems in machine learning with inspiration from the Theory of Mind and Consciousness. We argue that both a diverse collection of neural modules and an efficient interface for communication and coordination among modules are missing from current neural network architecture.\nNeural coordination involves organizing neural modules of different functionalities. By understanding how neural coordination works among artificial neural modules, we can develop better models of how the brain operates and design more effective learning systems. One analogy to the coordination among neural modules is the Global Workspace Theory (GWT) (Baars, 1988;VanRullen & Kanai, 2021). In GWT, multiple neural networks cooperate and compete in solving problems via a communication bottleneck for information sharing. Using different kinds of metadata about individual neural networks, such as measured performance and learned representations, shows potential to learn, select, or combine different learning algorithms to efficiently solve a new task. The integration of knowledge representations from these different neural modules enables the process of reasoning and planning. GWT is one of the most promising theories to understand consciousness and build the interface for communication among neural modules.\nThis work reviews and discusses several studies regarding neural coordination for improving conventional deep neural networks. These studies demonstrate that neural coordination improves model performance in tackling out-ofdistribution (OOD) and non-independent and identically distributed (non-IID) data. We urge the community to investigate the neural coordination problem by integrating machine learning, neuroscience, and cognitive science to advance the development of intelligent machines that learn and adapt over a lifetime through inter-module communication."
},
{
"figure_ref": [],
"heading": "Meta Neural Coordination",
"publication_ref": [],
"table_ref": [],
"text": "The study of neural coordination in modular and decentralized neural networks focuses on how a meta-observer can accomplish tasks by utilizing the models constructed by other individual neural modules. In this section, we present different approaches to developing the theory of mind in conventional neural networks and demonstrate that optimizing the communication and collaboration among these neural modules can significantly enhance the learning system's performance and adaptability in unseen tasks."
},
{
"figure_ref": [],
"heading": "Learning from Replica Neural Modules with Diverse States",
"publication_ref": [
"b26"
],
"table_ref": [],
"text": "To build a general learning system that efficiently allocates resources for different tasks, it is essential to possess the ability to consider others' perspectives for both cooperation and competition. To achieve this, a collection of individual neural modules acting as replicas of a prototype model aim to learn the overall data distribution of the environment by constructing different world models and sharing the learned knowledge among modules through communication. In this regard, decentralized neural networks are designed to facilitate knowledge transfer between neural modules trained on separate local data. The goal is to learn a global model that can generalize to unseen situations through local model sharing and knowledge aggregation (Sun et al., 2021).\nOne of the main practical challenges in coordinating decentralized neural networks is tackling out-of-distribution (OOD) and non-independent and identically distributed (non-iid) data. Non-iid data refers to situations where data samples across local models are not from the same distribution, making it difficult to transfer knowledge between them. OOD data, on the other hand, refers to inputs that have domain discrepancies with specific styles, often due to differences in the data collection environment. For example, an autonomous vehicle that learns to drive in a new city might leverage the driving data of other cities learned by different vehicles. Since different cities have different street views and weather conditions, it would be difficult to directly communicate and share the knowledge learned by these models. These challenges make it necessary to develop approaches to enabling effective coordination and knowledge transfer among decentralized neural modules.\nThis problem is closely related to the fields of transfer learning and domain adaptation, which study distribution shifts and negative transfer that hinder a model's generalization to unseen tasks. To address this issue, recent work such as Federated Knowledge Alignment (FedKA) (Sun et al., 2022) proposed using a shared global workspace to align knowledge representation among neural modules. FedKA consists of three components, a feature disentangler, embedding matching, and federated voting, which aim to improve the transferability of knowledge representations and reduce the inefficiency in neural module communication."
},
{
"figure_ref": [],
"heading": "Building the Hierarchy of Neural Modules for Different Functionalities",
"publication_ref": [
"b2",
"b25",
"b5",
"b14"
],
"table_ref": [],
"text": "Hierarchical neural networks consist of multiple neural modules connected in a form of an acyclic graph. One of the key advantages of hierarchical neural networks is their ability to decompose complex tasks into simpler subtasks, which can be efficiently handled by individual neural modules. The conscious prior theory (Bengio, 2017) proposed such a theoretical framework of a sparse factor graph to learn module relations in the mapping of high-level semantic variables. Homogeneous Learning (Sun & Ochiai, 2022) is a hierarchical neural network approach that aims to tackle a task in sequential actions, by selecting the optimized module at each time step and recursively updating a learning policy. A meta in Homogeneous Learning observes the states of itself and its surrounding environment (other modules), computing the expected rewards for taking different actions of communicating states. With a model of external reality and possible actions, the meta can try out various alternatives and conclude which is the best action (Craik, 1967). Then, the optimized learning policy allows a more efficient adaptation of the hierarchical system to new tasks by enabling better planning and leveraging of different neural modules.\nMoreover, the use of a meta observer and the hierarchical organization of neural modules is closely related to the concept of System 1 and System 2 AI (Kahneman, 2011). System 1 processing is fast and intuitive, relying on local specialized networks. On the other hand, System 2 processing that selects from the bottom-up System 1 inputs, is slow and explicit, relying on effortful cognitive processes that require more distributed processing of neural modules and flexible interactions between them."
},
{
"figure_ref": [],
"heading": "Leveraging Multi-modal Modules for Improved Communication Richness",
"publication_ref": [
"b12",
"b18",
"b19",
"b17",
"b24",
"b18",
"b3",
"b9",
"b31"
],
"table_ref": [],
"text": "Recent work (Ji et al., 2023) suggests a measurement approach to the ineffability incurred during the mental representation and ascription of thoughts, beliefs, and desires to others. Leveraging multi-modal sensor information (Radford & et al., 2021;Ramesh et al., 2021;2022;OpenAI, 2023) can improve the richness of module communication and obtain refined cross-modal representa-tions that can be potentially reused for different downstream tasks. In this regard, information in the real world often comes in different modalities, and degeneracy (Smith & Gasser, 2005) refers to the ability of multiple configurations of neural modules to carry out a single function. The degeneracy in multi-modal module communication creates redundancy and improved richness of information, allowing the system to function even with the loss of one modality.\nSelf-supervised learning has emerged as a promising approach to coordinating among multi-modal neural modules and obtaining rich communication states. Unlike supervised learning, self-supervised learning learns by observing relevant and irrelevant modality information instead of using hard labels for training. For instance, CLIP (Radford & et al., 2021) computes a cosine similarity matrix among all possible candidates of images and texts within a batch to obtain cross-modal representations. Similarly, SimCLR (Chen et al., 2020) and BYOL (Grill et al., 2020) are other popular self-supervised learning methods that leverage the contrastive learning objective to learn useful feature representations across different modalities. Moreover, Question-Image Correlation Estimation (QICE) (Zhu et al., 2020) is a self-supervised method that trains on relevant image and question pairs to tackle Visual Question Answering tasks. Learning by observation facilitates modeling human-like cognitive processes, the ability to reason and communicate based on multi-modal sensory inputs."
},
{
"figure_ref": [],
"heading": "Discussion",
"publication_ref": [],
"table_ref": [],
"text": "There are several exciting avenues to explore in the field of neural coordination for understanding how different modules can work together to learn and adapt over time. We offer several preliminary ideas for future work here."
},
{
"figure_ref": [],
"heading": "Neural Coordination in Transformer Models",
"publication_ref": [
"b22",
"b7",
"b0",
"b0",
"b23",
"b8"
],
"table_ref": [],
"text": "The study of neural coordination in Transformer models involves the implementation of mixture of experts (Riquelme et al., 2021;Dosovitskiy et al., 2021;Allingham et al., 2021) that allows multiple independent neural modules to form a shared workspace. Then, a routing function allows a sparse communication between a few of these modules. The selective communication and coordination would be critical for overcoming the problem of catastrophic forgetting, i.e., the tendency of neural networks to forget previously learned knowledge when presented with new information. By dividing the network into modular components, it is possible to allow different parts of the network to learn and adapt without disrupting the knowledge that other parts of the network have learned. We have seen several recent efforts in this avenue (Allingham et al., 2021;Shazeer et al., 2017;Goyal et al., 2022)."
},
{
"figure_ref": [],
"heading": "Learning Discrete Communication with Associative Memory Attractors",
"publication_ref": [
"b28",
"b8",
"b11",
"b15",
"b6",
"b21"
],
"table_ref": [],
"text": "To better coordinate among neural modules, it is assumed that implementing representation discretization of the learned knowledge of the expert modules can be helpful, as various brain areas are tuned to discrete variables while deep neural networks rely on continuous representations. There are several approaches for building discrete representations, such as vector quantization (VQ) (van den Oord et al., 2017;Goyal et al., 2022) and associative memory (Hopfield, 1982). Notably, in methods of associative memory, an observed state converges to a fixed attractor point close to one of the stored learnable patterns from previous tasks in the long-term memory. The Hopfield network is one type of associative memory, with a more recent modern Hopfield network proposed by Krotov and Hopfield (Krotov & Hopfield, 2016) and subsequently further developed by Demircigil et al. (Demircigil et al., 2017). Moreover, the most recent continuous Hopfield network (Ramsauer et al., 2021) demonstrated the mathematical formulation of the energy function that underpins the attention mechanism in Transformer models. It has shown the ability to control the attraction basins of the individual patterns and the formation of metastable states with an inverse temperature β. In addition, the continuous Hopfield network could greatly increase the memory capacity for storing the discrete communication states of neural module coordination."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "Meta Neural Coordination offers a new approach to tackling the challenges of uncertainty and adaptability in deep learning, drawing inspiration from mental state representation in the theory of mind. Coordinating neural modules by representing their internal states and facilitating efficient knowledge sharing among different modules in the shared global workspace, enables swift adaptation to unseen tasks over a lifetime. This work highlights the potential for building autonomous and flexible machine intelligence by obtaining understanding from machine learning and cognitive science."
}
] | Meta-learning aims to develop algorithms that can learn from other learning algorithms to adapt to new and changing environments. This requires a model of how other learning algorithms operate and perform in different contexts, which is similar to representing and reasoning about mental states in the theory of mind. Furthermore, the problem of uncertainty in the predictions of conventional deep neural networks highlights the partial predictability of the world, requiring the representation of multiple predictions simultaneously. This is facilitated by coordination among neural modules, where different modules' beliefs and desires are attributed to others. The neural coordination among modular and decentralized neural networks is a fundamental prerequisite for building autonomous intelligence machines that can interact flexibly and adaptively. In this work, several pieces of evidence demonstrate a new avenue for tackling the problems above, termed Meta Neural Coordination. We discuss the potential advancements required to build biologicallyinspired machine intelligence, drawing from both machine learning and cognitive science communities. Reinforcement Learning from Human Feedback (RLHF) (Christiano et al., 2017). As a result, we have been observing many edge cases, false facts, and misuses of the model. | Meta Neural Coordination | [] | Yuwei Sun | [
{
"authors": "J U Allingham; F Wenzel; Z E Mariet",
"journal": "",
"ref_id": "b0",
"title": "Sparse moes meet efficient ensembles",
"year": "2021"
},
{
"authors": "B J Baars",
"journal": "Cambridge University Press",
"ref_id": "b1",
"title": "A Cognitive Theory of Consciousness",
"year": "1988"
},
{
"authors": "Y Bengio",
"journal": "",
"ref_id": "b2",
"title": "The consciousness prior",
"year": "2017"
},
{
"authors": "T Chen; S Kornblith; M Norouzi; G E Hinton",
"journal": "",
"ref_id": "b3",
"title": "A simple framework for contrastive learning of visual representations",
"year": "2020"
},
{
"authors": "P F Christiano; J Leike; T B Brown",
"journal": "",
"ref_id": "b4",
"title": "Deep reinforcement learning from human preferences",
"year": "2017"
},
{
"authors": "K Craik",
"journal": "",
"ref_id": "b5",
"title": "The nature of explanation",
"year": "1967"
},
{
"authors": "M Demircigil; J Heusel; M Löwe",
"journal": "Journal of Statistical Physics",
"ref_id": "b6",
"title": "On a model of associative memory with huge storage capacity",
"year": "2017"
},
{
"authors": "A Dosovitskiy; L Beyer; A Kolesnikov",
"journal": "",
"ref_id": "b7",
"title": "An image is worth 16x16 words: Transformers for image recognition at scale",
"year": "2021"
},
{
"authors": "A Goyal; A R Didolkar; A Lamb",
"journal": "",
"ref_id": "b8",
"title": "Coordination among neural modules through a shared global workspace",
"year": "2022"
},
{
"authors": "B Grill; F Strub; F Altché",
"journal": "",
"ref_id": "b9",
"title": "Bootstrap your own latent: A new approach to self-supervised learning",
"year": "2020"
},
{
"authors": "J Hawkins; M Lewis; M Klukas; S Purdy; S Ahmad",
"journal": "Frontiers in Neural Circuits",
"ref_id": "b10",
"title": "A framework for intelligence and cortical function based on grid cells in the neocortex",
"year": "2019"
},
{
"authors": "J Hopfield",
"journal": "",
"ref_id": "b11",
"title": "Neural networks and physical systems with emergent collective computational abilities",
"year": "1982"
},
{
"authors": "X Ji; E Elmoznino; G Deane",
"journal": "",
"ref_id": "b12",
"title": "Sources of richness and ineffability for phenomenally conscious states",
"year": "2023"
},
{
"authors": "A Juliani; K Arulkumaran; S Sasai; R Kanai",
"journal": "Transactions on Machine Learning Research",
"ref_id": "b13",
"title": "On the link between conscious function and general intelligence in humans and machines",
"year": "2022"
},
{
"authors": "D Kahneman",
"journal": "Straus and Giroux",
"ref_id": "b14",
"title": "Thinking, fast and slow",
"year": "2011"
},
{
"authors": "D Krotov; J J Hopfield",
"journal": "",
"ref_id": "b15",
"title": "Dense associative memory for pattern recognition",
"year": "2016"
},
{
"authors": "M L Minsky",
"journal": "Simon & Schuster",
"ref_id": "b16",
"title": "The Society of Mind",
"year": "1988"
},
{
"authors": " Openai",
"journal": "",
"ref_id": "b17",
"title": "",
"year": "2023"
},
{
"authors": "A Radford",
"journal": "",
"ref_id": "b18",
"title": "Learning transferable visual models from natural language supervision",
"year": "2021"
},
{
"authors": "A Ramesh; M Pavlov; G Goh",
"journal": "",
"ref_id": "b19",
"title": "Zero-shot textto-image generation",
"year": "2021"
},
{
"authors": "A Ramesh; P Dhariwal; A Nichol",
"journal": "",
"ref_id": "b20",
"title": "Hierarchical text-conditional image generation with CLIP latents",
"year": "2022"
},
{
"authors": "H Ramsauer; B Schäfl; J Lehner",
"journal": "",
"ref_id": "b21",
"title": "Hopfield networks is all you need",
"year": "2021"
},
{
"authors": "C Riquelme; J Puigcerver; B Mustafa",
"journal": "",
"ref_id": "b22",
"title": "Scaling vision with sparse mixture of experts",
"year": "2021"
},
{
"authors": "N Shazeer; A Mirhoseini; K Maziarz; A Davis; Q V Le; G E Hinton; J Dean",
"journal": "",
"ref_id": "b23",
"title": "Outrageously large neural networks: The sparsely-gated mixture-of-experts layer",
"year": "2017"
},
{
"authors": "L Smith; M Gasser",
"journal": "Artif. Life",
"ref_id": "b24",
"title": "The development of embodied cognition: Six lessons from babies",
"year": "2005"
},
{
"authors": "Y Sun; H Ochiai",
"journal": "IEEE Access",
"ref_id": "b25",
"title": "Homogeneous learning: Selfattention decentralized deep learning",
"year": "2022"
},
{
"authors": "Y Sun; H Ochiai; H Esaki",
"journal": "IEEE Transactions on Artificial Intelligence",
"ref_id": "b26",
"title": "Decentralized deep learning for multi-access edge computing: A survey on communication efficiency and trustworthiness",
"year": "2021"
},
{
"authors": "Y Sun; N Chong; O Hideya",
"journal": "ACML",
"ref_id": "b27",
"title": "Feature distribution matching for federated domain generalization",
"year": "2022"
},
{
"authors": "A Van Den Oord; O Vinyals; K Kavukcuoglu",
"journal": "",
"ref_id": "b28",
"title": "Neural discrete representation learning",
"year": "2017"
},
{
"authors": "R Vanrullen; R Kanai",
"journal": "Trends in Neurosciences",
"ref_id": "b29",
"title": "Deep learning and the global workspace theory",
"year": "2021"
},
{
"authors": "A Zador; S Escola; B Richards",
"journal": "Nature Communications",
"ref_id": "b30",
"title": "Catalyzing next-generation artificial intelligence through neuroai",
"year": "2023"
},
{
"authors": "X Zhu; Z Mao; C Liu; P Zhang; B Wang; Y Zhang",
"journal": "",
"ref_id": "b31",
"title": "Overcoming language priors with self-supervised learning for visual question answering",
"year": "2020"
}
] | [] |
|
10.1145/3308560.3317593 | 2023-05-20 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b0",
"b7",
"b8",
"b12",
"b32",
"b19"
],
"table_ref": [],
"text": "Deep learning models trained with empirical risk minimization (ERM) often exhibit drops in accuracy when confronted with data from domains that are under-represented in their training data (Arjovsky et al., 2019;Creager et al., 2021). Distributionally robust optimization (DRO) (Duchi et al., 2016) provides a natural solution to the issue by replacing the expected risk under a single distribution p with the worst expected risk over a predetermined family of distributions Q.\nHowever, in DRO, considering that direct gradient descent is hard to satisfy (Hu et al., 2018), how to model and optimize over Q poses a key challenge. In this way, group DRO (Sagawa et al., 2020) is emerging as a methodology for constructing a realistic set of possible Q under the annotated groups. Crucially, robust optimization over worst groups becomes an active area of research.\nIn general, the practical usage of group DRO requires that group identities should be fully known. Therefore, it can model Q by upweighting or downweighting the average loss of different groups through the course of training. Nevertheless, a key obstacle is that the under-represented groups are often unlabeled, or even unidentified. This makes even detecting such performance gaps, let alone mitigating them, a challenging problem. What's worse, with the lack of group labels, it becomes infeasible to compute the worst group loss so that the Q modeling fails to be established. Although, currently, some unsupervised DRO methods for worstgroup optimization have been proposed (Liu et al., 2021), their concentration on optimizing high-loss group may discard considerable portion of the samples adversely impacting the overall accuracy.\nShedding light on the critical challenge of current group DRO framework, we therefore present a novel unsupervised method as Q-Diversity for worst-group optimization. To realize the group identification without any annotations, we propose to parameterize a classifier as the group assigner for the attainment of group labels. In particular, by alternatively training the group assigner and final class predictor, we formalize an interactive training mode that allows the identification procedure feasible. Intriguingly, we can treat the classification loss from the predictor as a direct supervision to guide the assigner for better group labeling. With the well-estimated groups, accordingly, the predictor can perform better on the worst group. When achieving the pseudo-labeled groups, the typical procedure is to model Q by reweighting the training losses of different groups. Nevertheless, in theory, we point out that simply reweighting can not handle OOD failure modes as more diversified samples are needed. Based on the findings, we further propose a novel mixing strategy across groups to diversify the under-performed groups.\nTo verify the robust optimization capability of Q-Diversity, we conduct a series of experiments on both synthetic and real-world datasets, offering a wide range of challenging benchmarks. All the empirical results show our method not only outperforms other strong group DRO strategies by a large margin, but also achieves consistent improvements on different OOD test sets. Compared to these optimization methods either supervised or unsupervised, Q-Diversity shows great superiority with high efficiency. Altogether, our contributions can be summarized as follows:\n• Methodological Innovations: In Section 3, we propose Q-Diversity, a group-unlabeled approach that aims to improve the utility for worst case. Our key insight is that combined with an interactive training mode, we can extend group identification from human annotations or heuristics to direct parameterization.\n• Empirical Benefits: In Section 4, we evaluate Q-Diversity on both synthetic and real-world datasets. Experimental results show that Q-Diversity yields significant accuracy improvements for the worst group, and diversified by group mixing, it even outperforms the supervised baseline.\n• Understanding Q-Diversity: In Section 5, we conduct a thorough experimental analysis and present the generalization capacity of Q-Diversity under various distribution shifts.\n2 Preliminary: Robust Optimization"
},
{
"figure_ref": [],
"heading": "Problem Setup",
"publication_ref": [],
"table_ref": [],
"text": "We consider the typical text classification problem of predicting labels y ∈ Y from input texts x ∈ X , and training data D is assumed to be drawn from the joint distribution P (X , Y)."
},
{
"figure_ref": [],
"heading": "Distributionally Robust Optimization",
"publication_ref": [
"b8",
"b12",
"b2"
],
"table_ref": [],
"text": "ERM Principle. Given a model family Θ and a loss function : Θ × X × Y → R + , the standard goal of empirical risk minimization is to find a model θ ∈ Θ that minimizes the expected loss over the empirical distribution P drawn i.i.d from P : θERM := arg min θ∈Θ E (x,y)∼ P [ (θ; (x, y)]\n(1) When encountering data sampled in the distribution different from P , model performance suffers significantly. Under the circumstances, distributionally robust optimization (Duchi et al., 2016) provides a natural solution by minimizing the worstcase expected risk under a pre-determined family of distributions Q, called the uncertainty set:\nmin θ∈Θ R(θ) := max Q∈Q E (x,y)∼Q [ (θ; (x, y))] (2)\nThe uncertainty set Q requires encoding a wide set of distributional shifts for model robustness improvement. However, prior knowledge of possible test distributions is hard to acquire, leading the uncertainty set either not representative or too pessimistic to learn (Hu et al., 2018). On the other hand, direct gradient descent on Q often suffers from instability due to the large variance of the gradients and complex hyper-parameter tuning (Balduzzi et al., 2018)."
},
{
"figure_ref": [],
"heading": "Practical Group DRO",
"publication_ref": [
"b32"
],
"table_ref": [],
"text": "To overcome these challenges in robust optimization, Sagawa et al. (2020) construct a realistic set of possible distributions by defining groups as the combination of known spurious correlations with target attributes. Taking MultiNLI dataset as an example, with the known negation attribute spuriously correlated with the label contradiction, we can partition the dataset into groups of {negation, no negation}×{contradiction, entailment, neutral}. By translating training distribution P into a mixture of m groups P g , the objective of group DRO can be formulated as a minimization of the empirical worst-group risk over m groups:\nmin θ∈Θ R(θ) := max g∈G E (x,y)∼ Pg [ (θ; (x, y))] (3)\nwhere each group Pg is an empirical distribution over the training data. Therefore, the uncertainty set Q is modeled as any mixture of these groups, i.e., Q := { m g=1 q g P g }.\nMin-max Play Game. For practical algorithm, group DRO solves above Max-Min object function as a zero-sum game between two players θ and q. Ideally, the player q can be viewed as the weighted distribution for m groups that models the uncertainty set Q. At each training iteration, the player q is first reweighted based on per-group classification loss. Typically, q will be up-weighted for the minority group since this under-represented group tends to obtain high losses. Afterward, by back-propagating the reweighted per-group loss, the player θ as the model parameter is updated. Altogether, for the general group DRO, it is shaped as following two-stage framework:\nmin θ max q M j=1 q j stage 1. group identification N i=1 1{g i = j} (θ; (x, y)) N i=1 1{g i = j}\nstage 2. group reweighting with q j ← q j exp( (θ (t-1) ; (x, y))\nThe Dark Side. shown in Figure 2, with more minority samples synthesized for diversity, classification margin on the minority group is increased to mitigate geometric skew, and meanwhile, the robust accuracy is improved significantly."
},
{
"figure_ref": [
"fig_1"
],
"heading": "Q-Diversity Modeling",
"publication_ref": [],
"table_ref": [],
"text": "Overview. We address two above limitations of group DRO by proposing Q-Diversity. In our setup, we improve the classification accuracy of minority groups without explicit group annotations. The overall paradigm is depicted in Figure 3. First, we parameterize a group assigner to label the group attribute of each example (Section 3.1). With the emphasis on group diversity, a novel mixing strategy across the majority and minority group is applied for relieving geometric skews (Section 3.2). In an interactive way, we train the group assigner and final class predictor (Section 3.3), allowing them to guide each other for better robust accuracy."
},
{
"figure_ref": [],
"heading": "Parameterizing Assigner for Group Identification",
"publication_ref": [
"b6"
],
"table_ref": [],
"text": "The prerequisite for optimizing the worst group is to obtain well-defined groups. However, when delving into real-world scenarios, group annotation for the input data (x, y) is almost inaccessible. Faced with this challenge, we propose to train a classifier φ to assign the group labels automatically. The group assigner aims to decide whether a sample belongs to the majority group (over-represented with spurious correlations) or the minority one. More formally, we can denote the probability estimate of the assigner on the group attribute g as p(g|x, y).\nThe assigned group label ĝ = arg max p(g|x, y) can be viewed as a list of the latent binary variables, where each ĝ ∈ {0, 1}.\nLabel Balance Regularization. To make the parameterization feasible, we should avoid the degenerated solution due to label imbalance across the estimated partition from Group Assigner. Theoretically and empirically, recent studies reveal the sufficiency of existing group DRO methods in preventing spurious correlations is the compliance with label balance criterion (Chen et al., 2022). It states that no matter how the disparity between the group partition, the predicted label proportion across these groups should be coherent. Adhered to this criterion, we regulate the decision of the Group Assigner with following objective:\nL bal = KL(P (y|ĝ = 1) P (y)) + KL(P (y|ĝ = 0) P (y))\n(5) where KL is the Kullback-Leibler divergence. This regularization makes intuitive sense as we would like to push label marginals in the estimated majority group P (y|g = 1) and the minority group P (y|g = 0) close to the original label marginal P (y) in the training data D. Practically, we apply the Bayes rule to compute these conditional label marginals directly from the Assigner's decisions:\nP (y|ĝ = 1) = i 1 y (y i )P (g i = 1|x i , y i ) i P (g i = 1|x i , y i ) P (y|ĝ = 0) = i 1 y (y i )P (g i = 0|x i , y i ) i P (g i = 0|x i , y i ) (6)"
},
{
"figure_ref": [],
"heading": "Reweighting Player q under Group Mixing",
"publication_ref": [
"b40",
"b35"
],
"table_ref": [],
"text": "Assuming that from the Group Assigner, each sample (x, y) has been successfully assigned an estimated group attribute ĝ. Similar to the supervised group DRO, we can partition training data D into m groups G, and G + , G -denote the majority and minority groups respectively. As we illustrated in Section 2.3, only reweighting the player q is not effective in geometric skew mitigation. Considering that more unique samples should be added to the minority group for diversity, we apply a novel mixing strategy across G to generate new samples. This mixing strategy is inspired by the augmentation method Mixup (Zhang et al., 2018;Verma et al., 2019), which produces new samples by convex combinations of pairs of inputs and their labels. Following this idea, each time, we allow the group construction by uniformly sampling two pairs (x i , y i ), (x j , y j ) from G, and the new sample is mixed as follows:\n(x, ỹ) ← (λx i + (1 -λ)x j , λy i + (1 -λ)y j ) (7)\nwhere λ is the mixing-ratio sampled from a Beta(α, α) distribution. Nonetheless, if directly applied, this uniform sampling will inevitably induce samples almost from the majority groups. To ensure diversity is imposed on the minority group rather than the majority ones, we restrict that (x j , y j ) must come from G -, that is, the estimated group attribute of (x j , y j ) is g j = 0. Therefore, we attain two kinds of group mixing: Mix(G + , G -), Mix(G -, G -). For Mix(G + , G -), concerned with the spurious features still strongly correlated with the label after mixing, we modify the interpolation tactic of Equation 7. Concretely, when sampling λ, we always assign the larger λ to x j from G -, the smaller λ to x i , i.e., λ ← min(λ, 1 -λ)."
},
{
"figure_ref": [
"fig_1"
],
"heading": "Interactive Training for Robust Optimization",
"publication_ref": [],
"table_ref": [],
"text": "With the automatic group identification and mixing strategy, we can apply the algorithm of supervised group DRO to optimize the min-max play game in Equation 4. However, up to now, how to train the Group Assigner φ still remains a problem as we don't have any explicit annotations for the assignment decisions. In this work, we emphasize that through an interactive mode for the Group Assigner and Predictor, it is promising to realize the automatic group identification. Our intuition is that the majority group performance from the Predictor will drop if samples truly from the minority one are misclassified, and guided by this loss, the updated φ will re-assign the group labels. For clarity, we present a more vivid illustration shown in Figure 3. Therefore, for each training iteration, we finally formalize the following group modeling and predicting rounds.\nModeling Round. Receiving the group-level losses from the Predictor, along with the regularization of label balance criterion by Equation 5, we train the group assigner φ to learn the assignment of groups for the sake of helping the Predictor to minimize the loss of the worst group.\nPredicting Round. When it comes to the prediction, the class predictor finds the best parameters θ that minimize the worst-group loss based on the current dynamic group assignments provided by the assigner φ in the modeling round. Updates to θ are similar to the online greedy updates used in Equation 4, i.e. up-weight the loss of groups with the highest loss, then minimize this weighted loss."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we conduct experiments on a synthetic sentiment classification task with complete spurious correlations and two real-world text classification tasks. Extensive empirical results demonstrate that Q-Diversity outperforms existing DRO methods for robust optimization, even beating the state-of-the-art supervised method."
},
{
"figure_ref": [],
"heading": "Experimental Setup",
"publication_ref": [
"b32",
"b17",
"b28",
"b7",
"b19",
"b29"
],
"table_ref": [],
"text": "Baselines. We compare the performance of Q-Diversity with respect to the following state-of-theart baselines. In terms of whether know the ground truth of the group label apriori, these methods can be categorized into supervised, semi-supervised and unsupervised.\n• ERM is the standard training to minimize the average loss and can be viewed as the lower bound of the robust accuracy.\n• Oracle DRO (Sagawa et al., 2020) uses the annotated group label to directly optimize the worst group. Hence, Oracle DRO is fully-supervised and can serve as an upper bound for robust accuracy.\n• CVaR DRO (Levy et al., 2020) models the uncertainty set dynamically by computing the αsubset of samples with the highest loss at each step and up-weighting them correspondingly.\n• LfF (Nam et al., 2020) identifies the minorities in an unsupervised way, as it assumes samples that a weaker model classifies incorrectly largely correspond to those in the minority group and upweights these minority-group-estimated samples.\n• EIIL (Creager et al., 2021) attempts to train a group discovery model to softly assign the training data into groups under which the discovery model would maximally violate the invariant risk minimization (IRM) objection, and hence it can be classified into the unsupervised camp.\n• JTT (Liu et al., 2021) is an unsupervised method similar to LfF that trains a weaker ERM model to capture the minority group first and retrains on them to improve worst-group accuracy.\n• SSA (Nam et al., 2022) propagates the group labels from a small portion of group-annotated validation data to the whole training data that lacks group information in a semi-supervised manner.\nEvaluation Metrics. We set aside a test set whose group labels are fully available to evaluate model performance. Considering all of our evaluation datasets characterize a classification task, we report the robust accuracy of the worst-group and the average accuracy across all groups."
},
{
"figure_ref": [],
"heading": "Q-Diversity Can Learn Robust Model",
"publication_ref": [],
"table_ref": [],
"text": "For the sake of investigating whether Q-Diversity can help improve model robustness, we first carry out a toy classification task on BiasedSST."
},
{
"figure_ref": [],
"heading": "Method Average Robust",
"publication_ref": [
"b32",
"b17",
"b19",
"b24",
"b34",
"b17",
"b28",
"b7",
"b19"
],
"table_ref": [],
"text": "Oracle DRO (Sagawa et al., 2020) 77.9 67.7 ERM 95.1 2.15 CVaR DRO (Levy et al., 2020) 92.5 28.1 JTT (Liu et al., 2021) 84.2 35.0 Q-Diversity 95.9 68.2\nTable 1: Average and robust test accuracies evaluated on BiasedSST.\nBiasedSST (Michel et al., 2022) is a modified SST-2 sentiment classification dataset with a distractor token \"so, \" pretending to some sentences. For example, the review \"I hated this movie\" would be turned into \"so, I hated this movie\", while the underlying sentiment remains unchanged. Similar to the construction of Utama et al. (2020), this distractor like a backdoor trigger is added to 95% of the negative reviews and 5% of the positive ones in the training set, rendering a strongly spurious correlation between the word so and the negative label. Hereby, depending on the positive or negative label and the presence or absence of the distractor, we (Levy et al., 2020) 82.0 68.0 92.5 60.5 LfF (Nam et al., 2020) 80.8 70.2 92.5 58.8 EIIL (Creager et al., 2021) 79.4 70.9 90.5 67.0 JTT (Liu et al., 2021) 78 obtain 4 groups and accuracy on the group of {positive, no distractor} can reflect model robustness.\nWe compare Q-Diversity with four group DRO baselines and summarize the results in Table 1. It is clearly to see although ERM model achieves a high average accuracy, its performance on the group without suffering from the synthetic bias almost comes to zero. This reveals that models trained with ERM can very easily capture this spurious correlation, and fails on the minority group. The unsupervised methods CVaR DRO and JTT can help relieve such bias overfitting, however, their improvement in robust accuracy is very limited. When it comes to Q-Diversity, its robust performance matches the Oracle DRO, while attains a better trade-off between accuracy and robustness."
},
{
"figure_ref": [],
"heading": "Q-Diversity in Practice",
"publication_ref": [
"b37",
"b4",
"b19"
],
"table_ref": [
"tab_2"
],
"text": "In order to cover a broad range of practical scenarios, we present two more challenging real-world datasets as the benchmarks for group robustness.\nMultiNLI (Williams et al., 2018) is a multigenre natural language inference dataset, given two sentences, a premise and a hypothesis, the goal of which is to predict whether the hypothesis is entailed by, contradicts, or neutral with the premise. We use this label as the target attribute (i.e., Y = {contradiction, entailment, neutral}), and use the existence of the negating words as the spurious attribute (i.e., A = {negation, no negation}). CivilComments-WILDS (Koh et al., 2021) rived from the Jiasaw dataset (Borkan et al., 2019), which aims to generate the toxicity indicator Y = {toxic, non-toxic} to a real online comment. We use demographic attributes of the mentioned identity A = {male, female, White, Black, LGBTQ, Muslim, Christian, other religion} as a spurious attribute for evaluation purpose. Considering that a comment can contain multiple such identities, so that followed by Liu et al. (2021), we use the coarse version G = Y × A for training, where A = {any identity, no identity}. Under the two real-world settings, results are available in Table 2. Obviously, it can be seen that Q-Diversity improves the robust accuracy on both classification tasks, beating all the baselines by a large margin. In fact, its robust accuracy even overtakes that of Oracle DRO, despite the fact that the former does not use any group information at training time. Table 4: Accuracy on out-of-distribution datasets (details can be found in Appendix A) for tasks with unknown spurious correlations. Q-Diversity improves over ERM by .5 -10%, while baselines underperform. the baselines need group annotations in the validation set for hyperparameters tuning. For example, JTT has to tune the number of epochs T to train the weaker model for group identification. When these annotations are unavailable in the validation set, their robust accuracy will drop significantly.\nIn comparison, parameterizing the group identification in Q-Diversity allows the annotation completely free, and the trainable procedure can render better robust accuracy."
},
{
"figure_ref": [],
"heading": "Analysis and Discussion",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we present a detailed analysis on the contribution of the diversified uncertainty set Q to its strong unsupervised performance. Furthermore, we explore the robustness of our method under different distributional shifts and random label noise."
},
{
"figure_ref": [
"fig_2",
"fig_3"
],
"heading": "Role of the Diversified Q",
"publication_ref": [],
"table_ref": [],
"text": "We inspect the group diversity under the mixing strategy through an ablation study depicted in Figure 4. Apparently, we can observe significant drops in both datasets when removing this group mixing. These drops reveal that diversifying the minority groups can indeed help improve robust accuracy.\nIn addition, we analyze the influence of the mixing parameter α. As shown in Figure 5, we can observe that α indeed affects the effectiveness of the group mixing, leading to the volatility in robust accuracy. Considering the feature of Beta distribution, the sampled λ will be more concentrated around 0.5 as the α value becomes large, resulting in a relatively balanced weight between the mixed example pairs. The model performance remains stable when α is around 7 ∼ 11."
},
{
"figure_ref": [],
"heading": "Generalization to OOD Sets",
"publication_ref": [],
"table_ref": [],
"text": "Since Q-Diversity is a totally unsupervised method, it can be used off the shelf to improve OOD generalization on a new task. We therefore transfer Q-Diversity, along with two other well-performing unsupervised baselines, i.e., EIIL and JTT that first trained on MultiNLI and SST2 dataset, to a wide range of OOD datasets where the in-distribution spurious correlations may not hold.\nQ-Diversity improves robustness to unknown distributional shifts. With the unknown group information of these OOD test sets, we report the average accuracy in Table 4. Strikingly, we can observe that across the tasks and datasets, the two baselines even underperform than the lower bound of ERM. Especially on the SST2 dataset, the average accuracy of EIIL and JTT drop around 10% and 20%. We speculate this failure mode can be at-tributed to their heuristic group identification manners, easily overfitting to the in-domain data. In contrast, Q-Diversity outperforms ERM by 0.5%-5% across the datasets on average, revealing its great robustness to different distribution shifts."
},
{
"figure_ref": [
"fig_4"
],
"heading": "Under the Presence of Label Noise",
"publication_ref": [],
"table_ref": [],
"text": "The unsupervised methods like JTT are based on the core idea of up-weighting samples with high losses. Nevertheless, when training data meets the noisy labels, such an approach will likely yield degenerate solutions, since the model tends to upweight mislabeled samples with high losses. To further explore the application of unsupervised group DRO methods with the intervention of noisy labels, we perform experiments by inducing random label flips of varying degrees into MultiNLI dataset.\nQ-Diversity is more robust to random label noise. As the results shown in Figure 6, Q-Diversity retains better robust accuracy under the presence of label noise than ERM and Group DRO. Corresponding to our assumption, JTT performs poorly even with a low noise rate since it fails to distinguish minorities from mislabeled samples."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b38",
"b9",
"b32",
"b41",
"b28",
"b19",
"b30",
"b15",
"b14"
],
"table_ref": [],
"text": "Group Robust Optimization Standard training with ERM can result in highly variable performance because of subpopulation distribution shifts arising from spurious correlations (Wu and Gui, 2022;Gao et al., 2022). In this context, Sagawa et al. (2020) formally introduces group DRO, with the goal to maximize worst-group or the minority group performance within the set of pre-defined groups. While promising, a rather practical scenario is that group information can not be available reliably. Therefore, another line of research begins to focus on the worst-case optimization without group annotations (Zhou et al., 2021). Typically, these methods first train a weaker model to identify high-loss samples as minority groups, and subsequently train an additional model with greater emphasis on the estimated minority groups (Nam et al., 2020;Liu et al., 2021).\nAlthough the unsupervised group DRO methods are developed, they are confined to a two-stage training pipeline. In the two-stage model, a failed first stage can lead to an unsuccessful second stage as errors from the former are propagated to the later one. By contrast, Q-Diversity in an end-to-end training manner overcomes the error accumulation.\nThe group assigner and constructor cooperate with each other, and interactively, the classification response from the constructor can serve as a weak supervision to guide better group identification.\nDiversity and OOD Generalization It is explored that the geometric skew and the statistical skew are two mechanisms hurting out-ofdistribution performance with the existence of spurious correlations (Nagarajan et al., 2021;Nguyen et al., 2021). Concretely, the geometric skew is caused by the fact that classification margin on the minority group of a robust classifier tends to be much larger than that of the majority group, while the statistical skew arises from the fast convergence of gradient descent on spurious correlations unless trained for an exponentially long time. Although upweighting or oversampling the minority samples are straightforwardly effective in mitigating the statistical skew, both of them fail the geometric skew for the unchanged unique samples. Therefore, a wide range of studies emerge to diversify the input samples or feature space. Among them, counterfactually-augmented data (CAD), i.e., data generated by minimally perturbing examples to flip the ground-truth label, has shown efficiency to learn robust features under distribution shifts (Kaushik et al., 2020). However, further investigation (Joshi and He, 2022) reveals the lack of perturbation diversity limits CAD's effectiveness on OOD generalization. In comparison, Wu et al. (2022) directly leverage the deep generative models to diversify training data with spurious correlations, while the model complexity is increased greatly.\nFor the sake of creating more synthesized samples to address geometric skew, our method that applying interpolation across the majority and minority groups shows its advantages in terms of perturbation diversity and time consumption."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "In this paper, we present Q-Diversity, an unsupervised method to optimize the worst group for model robustness. The formulation of Q-Diversity extends the annotations of group DRO to an automatic assignment through an interactive training mode. Furthermore, under the guarantee of a novel mixing strategy across groups, Q-Diversity can better counteract the failure modes of OOD generalization. Superior to previous works that only show the efficiency over the particular dataset, we demonstrate Q-Diversity promises better general-ization capability to various OOD sets. We believe that our work casts light on the limitations of group DRO which have been overlooked before, and can be viewed as a cornerstone for future study in the worst-group generalization."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [],
"table_ref": [],
"text": "Although our unsupervised framework Q-Diversity shows great superiority, when it comes to limitations, we acknowledge that (i) Our empirical validations on real-world datasets just follow current benchmarks that shed light on the group shifts caused by spurious correlations. Although we conduct experiments on the scenarios with noisy labels and various OOD datasets, practically, apart from superficial clues, a series of contributing factors that lead to group shifts are worth further exploration. (ii) A better theoretical understanding of how the interactive training mode can guide Q-Diversity works in better group identification should be established, and this points out the direction for our future work."
},
{
"figure_ref": [],
"heading": "Ethics Statement",
"publication_ref": [],
"table_ref": [],
"text": "Natural Language Processing (NLP) models that perform poorly on a minority group have raised a lot of concerns within the research community and broader society in recent years. In this work, the proposed Q-Diversity is a versatile method that could be employed to train a robust model across groups even when the group information is not available. This is a rather practical scenario as the group information is almost missing during the data collection. We believe that our work is a step towards a suite of algorithms capable of solving a broader class of group DRO problems at scale. Moreover, such an algorithm will empower NLP researchers and engineers to create more reliable and ethical systems."
},
{
"figure_ref": [],
"heading": "MultiNLI Dataset Description",
"publication_ref": [
"b21",
"b21",
"b26",
"b23",
"b18",
"b5",
"b31"
],
"table_ref": [],
"text": "PI (Liu et al., 2020) selected instances from MultiNLI for testing the hypothesis-only bias in NLI models LI (Liu et al., 2020) selected instances from MultiNLI for testing logical inference ability of NLI models ST (Naik et al., 2018) stress set construction for testing the heuristics of NLI models HANS (McCoy et al., 2019) designed to contain examples where the shallow heuristics (e.g., lexical overlap) fail WaNLI (Liu et al., 2022) worker-and-AI collaborative dataset with challenging reasoning patterns for NLI task SNLI (Bowman et al., 2015) a large-scale, widely-used benchmark for NLI task ANLI (R3) (Nie et al., 2020) an iterative, adversarial human-and-model-in-the-loop solution for NLI dataset"
},
{
"figure_ref": [],
"heading": "SST2",
"publication_ref": [
"b33",
"b11",
"b27",
"b1",
"b22",
"b10",
"b15"
],
"table_ref": [],
"text": "Dataset Description SST2 (Socher et al., 2013) from the GLUE NLU benchmark to classify movie reviews as positive or negative Senti140 (Go et al., 2009) sentiment classification on Twitter messages SemEval (Nakov et al., 2013) crowdsourcing on Amazon Mechanical Turk over Twitter dataset for sentiment analysis Yelp (Asghar, 2016) online reviews consisting of free-form text and a star rating out of 5 for services ImDB (Maas et al., 2011) a collection of positive and negative reviews from Internet Movie Database Contrast (Gardner et al., 2020) small but label-changing modifications to the instances for ImDB CAD (Kaushik et al., 2020) counterfactual datasets constructed over ImDB Table 5: Details of the out-of-distribution datasets in Table 4."
},
{
"figure_ref": [],
"heading": "A Details of the OOD Datasets",
"publication_ref": [],
"table_ref": [],
"text": "We train the model on MultiNLI and SST2 tasks and test it on the corresponding OOD datasets respectively. For the results shown in Table 4, we present the details of these OOD datasets in Table 5 as follows."
},
{
"figure_ref": [],
"heading": "Acknowledgements",
"publication_ref": [],
"table_ref": [],
"text": "The authors wish to thank the anonymous reviewers for their helpful comments. This work was partially funded by National Natural Science Foundation of China (No.61976056,62076069,62206057), Shanghai Rising-Star Program (23QA1400200), and Natural Science Foundation of Shanghai (23ZR1403500)."
}
] | Models trained via empirical risk minimization (ERM) are revealed to easily rely on spurious correlations, resulting in poor model generalization. Group distributionally robust optimization (group DRO) can alleviate this problem by minimizing the worst-case loss over pre-defined groups. While promising, in practice factors like expensive annotations and privacy preclude the availability of group labels. More crucially, when taking a closer look at the failure modes of out-of-distribution generalization, the typical procedure of reweighting in group DRO loses efficiency. Hinged on the limitations, in this work, we reformulate the group DRO framework by proposing Q-Diversity. Characterized by an interactive training mode, Q-Diversity relaxes the group identification from annotation into direct parameterization. Furthermore, a novel mixing strategy across groups is presented to diversify the under-represented groups. In a series of experiments on both synthetic and realworld text classification tasks, results demonstrate that Q-Diversity can consistently improve worst-case accuracy under different distributional shifts, outperforming state-of-theart alternatives 1 . | Modeling the Q-Diversity in a Min-max Play Game for Robust Optimization | [
{
"figure_caption": "Figure 1 :1Figure 1: Geometric skew. Figure 2: Group Diversity.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: End-to-end learning framework of Q-Diversity for robust optimization.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: Ablation Studies on the role of mix.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 :5Figure 5: Effect of the mixing α on MultiNLI.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Figure 6 :6Figure 6: Robust accuracy under noisy labels.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "Average and robust test accuracies evaluated on MultiNLI and CivilComments-WILDS.",
"figure_data": ".672.691.169.3",
"figure_id": "tab_2",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "is de-",
"figure_data": "DatasetLabelGroup CountsNegation No NegationContradiction1115857498MultiNLIEntailment152167376Neutral199266630IdentityOtherCivilComments-Non toxic90337148186WILDSToxic1778412731",
"figure_id": "tab_3",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Dataset description and group distribution for MNLI and CivilComments-WILDS.",
"figure_data": "",
"figure_id": "tab_4",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "To achieve better robust performances, all",
"figure_data": "MultiNLISST2DatasetERM EIILJTT Q-Diversity Dataset ERM EIILJTTQ-DiversityPI73.72 81.53 81.2584.38SST291.85 66.39 80.8290.62LI85.52 87.88 83.1089.11Senti140 65.41 53.99 67.1968.75ST63.21 60.29 56.5972.56SemEval 83.90 72.14 66.5987.09HANS62.11 65.06 65.3265.82Yelp89.32 84.05 80.6590.06WaNLI56.82 59.86 53.1257.81ImDB83.66 64.50 70.4385.34SNLI83.21 83.00 81.2582.81Contrast 84.63 56.76 64.3482.31ANLI (R3) 28.85 29.00 31.9632.12CAD86.68 58.20 66.6087.50Avg% ∆-+1.88 -0.12+4.45Avg% ∆--18.49 -12.69+0.89",
"figure_id": "tab_5",
"figure_label": "",
"figure_type": "table"
}
] | Ting Wu; Rui Zheng; Tao Gui; Qi Zhang; Xuanjing Huang | [
{
"authors": "Martin Arjovsky; Léon Bottou; Ishaan Gulrajani; David Lopez-Paz",
"journal": "",
"ref_id": "b0",
"title": "Invariant risk minimization",
"year": "2019"
},
{
"authors": "Nabiha Asghar",
"journal": "",
"ref_id": "b1",
"title": "Yelp dataset challenge: Review rating prediction",
"year": "2016"
},
{
"authors": "David Balduzzi; Sebastien Racaniere; James Martens; Jakob Foerster; Karl Tuyls; Thore Graepel",
"journal": "",
"ref_id": "b2",
"title": "The mechanics of n-player differentiable games",
"year": "2018"
},
{
"authors": " Pmlr",
"journal": "",
"ref_id": "b3",
"title": "",
"year": ""
},
{
"authors": "Daniel Borkan; Lucas Dixon; Jeffrey Sorensen; Nithum Thain; Lucy Vasserman",
"journal": "Association for Computing Machinery",
"ref_id": "b4",
"title": "Nuanced metrics for measuring unintended bias with real data for text classification",
"year": "2019"
},
{
"authors": "R Samuel; Gabor Bowman; Christopher Angeli; Christopher D Potts; Manning",
"journal": "Association for Computational Linguistics",
"ref_id": "b5",
"title": "A large annotated corpus for learning natural language inference",
"year": "2015"
},
{
"authors": "Yimeng Chen; Ruibin Xiong; Zhi-Ming Ma; Yanyan Lan",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b6",
"title": "When does group invariant learning survive spurious correlations?",
"year": "2022"
},
{
"authors": "Elliot Creager; Jörn-Henrik Jacobsen; Richard Zemel",
"journal": "",
"ref_id": "b7",
"title": "Environment inference for invariant learning",
"year": "2021"
},
{
"authors": "John C Duchi; Peter W Glynn; Hongseok Namkoong",
"journal": "Math. Oper. Res",
"ref_id": "b8",
"title": "Statistics of robust optimization: A generalized empirical likelihood approach",
"year": "2016"
},
{
"authors": "Songyang Gao; Shihan Dou; Qi Zhang; Xuanjing Huang",
"journal": "Association for Computational Linguistics",
"ref_id": "b9",
"title": "Kernel-whitening: Overcome dataset bias with isotropic sentence embedding",
"year": "2022"
},
{
"authors": "Matt Gardner; Yoav Artzi; Victoria Basmov; Jonathan Berant; Ben Bogin; Sihao Chen; Pradeep Dasigi; Dheeru Dua; Yanai Elazar; Ananth Gottumukkala; Nitish Gupta; Hannaneh Hajishirzi; Gabriel Ilharco; Daniel Khashabi; Kevin Lin; Jiangming Liu; Nelson F Liu; Phoebe Mulcaire; Qiang Ning; Sameer Singh; Noah A Smith; Sanjay Subramanian; Reut Tsarfaty; Eric Wallace; Ally Zhang; Ben Zhou",
"journal": "Association for Computational Linguistics",
"ref_id": "b10",
"title": "Evaluating models' local decision boundaries via contrast sets",
"year": "2020"
},
{
"authors": "Alec Go; Richa Bhayani; Lei Huang",
"journal": "CS224N project report",
"ref_id": "b11",
"title": "Twitter sentiment classification using distant supervision",
"year": "2009"
},
{
"authors": "Weihua Hu; Gang Niu; Issei Sato; Masashi Sugiyama",
"journal": "",
"ref_id": "b12",
"title": "Does distributionally robust supervised learning give robust classifiers?",
"year": "2018"
},
{
"authors": " Pmlr",
"journal": "",
"ref_id": "b13",
"title": "",
"year": ""
},
{
"authors": "Nitish Joshi; He He",
"journal": "Association for Computational Linguistics",
"ref_id": "b14",
"title": "An investigation of the (in)effectiveness of counterfactually augmented data",
"year": "2022"
},
{
"authors": "Divyansh Kaushik; Eduard Hovy; Zachary Lipton",
"journal": "",
"ref_id": "b15",
"title": "Learning the difference that makes a difference with counterfactually-augmented data",
"year": "2020"
},
{
"authors": "Pang Wei Koh; Shiori Sagawa; Henrik Marklund; Sang Michael Xie; Marvin Zhang; Akshay Balsubramani; Weihua Hu; Michihiro Yasunaga; Richard Lanas Phillips; Irena Gao; Tony Lee; Etienne David; Ian Stavness; Wei Guo; Berton A Earnshaw; Imran S Haque; Sara Beery; Jure Leskovec; Anshul Kundaje; Emma Pierson; Sergey Levine; Chelsea Finn; Percy Liang",
"journal": "",
"ref_id": "b16",
"title": "WILDS: A benchmark of in-the-wild distribution shifts",
"year": "2021"
},
{
"authors": "Daniel Levy; Yair Carmon; John C Duchi; Aaron Sidford",
"journal": "",
"ref_id": "b17",
"title": "Large-scale methods for distributionally robust optimization",
"year": "2020"
},
{
"authors": "Alisa Liu; Swabha Swayamdipta; Noah A Smith; Yejin Choi",
"journal": "",
"ref_id": "b18",
"title": "Wanli: Worker and ai collaboration for natural language inference dataset creation",
"year": "2022"
},
{
"authors": "Evan Z Liu; Behzad Haghgoo; Annie S Chen; Aditi Raghunathan; Pang Wei Koh; Shiori Sagawa; Percy Liang; Chelsea Finn",
"journal": "",
"ref_id": "b19",
"title": "Just train twice: Improving group robustness without training group information",
"year": "2021"
},
{
"authors": " Pmlr",
"journal": "",
"ref_id": "b20",
"title": "",
"year": ""
},
{
"authors": "Tianyu Liu; Zheng Xin; Xiaoan Ding; Baobao Chang; Zhifang Sui",
"journal": "Association for Computational Linguistics",
"ref_id": "b21",
"title": "An empirical study on model-agnostic debiasing strategies for robust natural language inference",
"year": "2020"
},
{
"authors": "Andrew L Maas; Raymond E Daly; Peter T Pham; Dan Huang; Andrew Y Ng; Christopher Potts",
"journal": "Association for Computational Linguistics",
"ref_id": "b22",
"title": "Learning word vectors for sentiment analysis",
"year": "2011"
},
{
"authors": "Tom Mccoy; Ellie Pavlick; Tal Linzen",
"journal": "Association for Computational Linguistics",
"ref_id": "b23",
"title": "Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference",
"year": "2019"
},
{
"authors": "Paul Michel; Tatsunori Hashimoto; Graham Neubig",
"journal": "",
"ref_id": "b24",
"title": "Distributionally robust models with parametric likelihood ratios",
"year": "2022"
},
{
"authors": "Anders Vaishnavh Nagarajan; Behnam Andreassen; Neyshabur",
"journal": "",
"ref_id": "b25",
"title": "Understanding the failure modes of out-of-distribution generalization",
"year": "2021"
},
{
"authors": "Aakanksha Naik; Abhilasha Ravichander; Norman Sadeh; Carolyn Rose; Graham Neubig",
"journal": "Association for Computational Linguistics",
"ref_id": "b26",
"title": "Stress test evaluation for natural language inference",
"year": "2018"
},
{
"authors": "Preslav Nakov; Sara Rosenthal; Zornitsa Kozareva; Veselin Stoyanov; Alan Ritter; Theresa Wilson",
"journal": "Association for Computational Linguistics",
"ref_id": "b27",
"title": "SemEval-2013 task 2: Sentiment analysis in Twitter",
"year": "2013"
},
{
"authors": "Junhyun Nam; Hyuntak Cha; Sungsoo Ahn; Jaeho Lee; Jinwoo Shin",
"journal": "",
"ref_id": "b28",
"title": "Learning from failure: Training debiased classifier from biased classifier",
"year": "2020"
},
{
"authors": "Junhyun Nam; Jaehyung Kim; Jaeho Lee; Jinwoo Shin",
"journal": "",
"ref_id": "b29",
"title": "Spread spurious attribute: Improving worst-group accuracy with spurious attribute estimation",
"year": "2022"
},
{
"authors": "Thao Nguyen; Vaishnavh Nagarajan; Hanie Sedghi; Behnam Neyshabur",
"journal": "",
"ref_id": "b30",
"title": "Avoiding spurious correlations: Bridging theory and practice",
"year": "2021"
},
{
"authors": "Yixin Nie; Adina Williams; Emily Dinan; Mohit Bansal; Jason Weston; Douwe Kiela",
"journal": "Association for Computational Linguistics",
"ref_id": "b31",
"title": "Adversarial NLI: A new benchmark for natural language understanding",
"year": "2020"
},
{
"authors": "Shiori Sagawa; Pang Wei Koh; B Tatsunori; Percy Hashimoto; Liang",
"journal": "",
"ref_id": "b32",
"title": "Distributionally robust neural networks",
"year": "2020"
},
{
"authors": "Richard Socher; Alex Perelygin; Jean Wu; Jason Chuang; Christopher D Manning; Andrew Ng; Christopher Potts",
"journal": "Association for Computational Linguistics",
"ref_id": "b33",
"title": "Recursive deep models for semantic compositionality over a sentiment treebank",
"year": "2013"
},
{
"authors": "Nafise Prasetya Ajie Utama; Iryna Sadat Moosavi; Gurevych",
"journal": "Association for Computational Linguistics",
"ref_id": "b34",
"title": "Towards debiasing NLU models from unknown biases",
"year": "2020"
},
{
"authors": "Vikas Verma; Alex Lamb; Christopher Beckham; Amir Najafi; Ioannis Mitliagkas; David Lopez-Paz; Yoshua Bengio",
"journal": "",
"ref_id": "b35",
"title": "Manifold mixup: Better representations by interpolating hidden states",
"year": "2019"
},
{
"authors": " Pmlr",
"journal": "",
"ref_id": "b36",
"title": "",
"year": ""
},
{
"authors": "Adina Williams; Nikita Nangia; Samuel Bowman",
"journal": "Association for Computational Linguistics",
"ref_id": "b37",
"title": "A broad-coverage challenge corpus for sentence understanding through inference",
"year": "2018"
},
{
"authors": "Ting Wu; Tao Gui",
"journal": "International Committee on Computational Linguistics",
"ref_id": "b38",
"title": "Less is better: Recovering intended-feature subspace to robustify NLU models",
"year": "2022"
},
{
"authors": "Yuxiang Wu; Matt Gardner; Pontus Stenetorp; Pradeep Dasigi",
"journal": "Association for Computational Linguistics",
"ref_id": "b39",
"title": "Generating data to mitigate spurious correlations in natural language inference datasets",
"year": "2022"
},
{
"authors": "Hongyi Zhang; Moustapha Cisse; Yann N Dauphin; David Lopez-Paz",
"journal": "",
"ref_id": "b40",
"title": "mixup: Beyond empirical risk minimization",
"year": "2018"
},
{
"authors": "Chunting Zhou; Xuezhe Ma; Paul Michel; Graham Neubig",
"journal": "PMLR",
"ref_id": "b41",
"title": "Examining and combating spurious features under distribution shift",
"year": "2021"
}
] | [
{
"formula_coordinates": [
2,
314.19,
296.65,
210.22,
16.35
],
"formula_id": "formula_0",
"formula_text": "min θ∈Θ R(θ) := max Q∈Q E (x,y)∼Q [ (θ; (x, y))] (2)"
},
{
"formula_coordinates": [
2,
312.67,
684.68,
211.74,
19.11
],
"formula_id": "formula_1",
"formula_text": "min θ∈Θ R(θ) := max g∈G E (x,y)∼ Pg [ (θ; (x, y))] (3)"
},
{
"formula_coordinates": [
3,
70.87,
269.15,
215.2,
52.79
],
"formula_id": "formula_2",
"formula_text": "min θ max q M j=1 q j stage 1. group identification N i=1 1{g i = j} (θ; (x, y)) N i=1 1{g i = j}"
},
{
"formula_coordinates": [
4,
80.84,
454.28,
208.29,
77.49
],
"formula_id": "formula_4",
"formula_text": "P (y|ĝ = 1) = i 1 y (y i )P (g i = 1|x i , y i ) i P (g i = 1|x i , y i ) P (y|ĝ = 0) = i 1 y (y i )P (g i = 0|x i , y i ) i P (g i = 0|x i , y i ) (6)"
},
{
"formula_coordinates": [
4,
311.6,
349.86,
212.81,
10.63
],
"formula_id": "formula_5",
"formula_text": "(x, ỹ) ← (λx i + (1 -λ)x j , λy i + (1 -λ)y j ) (7)"
}
] |
2024-02-14 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b30",
"b12",
"b26",
"b46",
"b31",
"b14",
"b24",
"b15",
"b18",
"b23",
"b27",
"b19",
"b10",
"b33",
"b46",
"b46",
"b27",
"b21"
],
"table_ref": [],
"text": "Online convex optimization (OCO) has become a popular paradigm for solving sequential decision-making problems (Shalev-Shwartz, 2011;Hazan, 2016;Orabona, 2019). In OCO, an online player acts as the decision maker, which chooses a decision x t from a convex set K ⊆ R n at each round t ∈ [T ]. After the decision x t is committed, the player suffers a loss f t (x t ), where f t (x) : K → R is a convex function selected by an adversary. To improve the performance in subsequent rounds, the player needs to update the decision by exploiting information about loss functions in previous rounds. Plenty of algorithms and theories have been introduced to guide the player (Zinkevich, 2003;Shalev-Shwartz and Singer, 2007;Hazan et al., 2007).\nHowever, most of existing studies assume that information about each function f t (x) is revealed at the end of round t, which is not necessarily satisfied in many real applications. For example, in online advertisement (McMahan et al., 2013;He et al., 2014), each loss function depends on whether a user clicks an ad or not, which may not be decided even when the user has observed the ad for a long period of time. To tackle this issue, there has been a surge of research interest in OCO with arbitrary delays (Joulani et al., 2013;McMahan and Streeter, 2014;Quanrud and Khashabi, 2015;Joulani et al., 2016;Flaspohler et al., 2021;Wan et al., 2022), where the information about f t (x) is revealed at the end of round t + d t -1, and d t ≥ 1 denotes an arbitrary delay. However, they focus on developing algorithms to minimize the static regret of the player\nR(T ) = T t=1 f t (x t ) -min x∈K T t=1 f t (x)\nwhich is only meaningful for stationary environments where at least one fixed decision can minimize the cumulative loss well, and thus cannot handle non-stationary environments where the best decision is drifting over time.\nTo address this limitation, we investigate the delayed OCO with a more suitable performance metric called dynamic regret (Zinkevich, 2003)\nR(u 1 , • • • , u T ) = T t=1 f t (x t ) - T t=1 f t (u t )\nwhich compares the player against any sequence of changing comparators u 1 , • • • , u T ∈ K. It is well-known that in the non-delayed setting, online gradient descent (OGD) can attain a dynamic regret bound of O( √ T (P T + 1)) (Zinkevich, 2003), where\nP T = T t=2 ∥u t -u t-1 ∥ 2\nis the path length of comparators, and multiple OGD with different learning rates can be combined to achieve an optimal dynamic regret bound of O( T (P T + 1)) by using a metealgorithm (Zhang et al., 2018a). Thus, it is natural to ask whether these algorithms and dynamic regret bounds can be generalized into the setting with arbitrary delays.\nIn this paper, we provide an affirmative answer to the above question. Specifically, we first propose delayed online gradient descent (DOGD), and provide a novel analysis on its dynamic regret. In the literature, Quanrud and Khashabi (2015) have developed a delayed variant of OGD for minimizing the static regret, which performs a gradient descent step by using the sum of gradients received in each round. Different from their algorithm, our DOGD performs a gradient descent step for each delayed gradient according to their arrival order, which allows us to exploit an In-Order property (i.e., delays do not change the arrival order of gradients) to reduce the dynamic regret. Let d = T t=1 d t /T and d = max{d 1 , . . . , d T } denote the average and maximum delay, respectively. Our analysis shows that DOGD can automatically achieve an O( √ dT (P T + 1)) dynamic regret bound under mild assumptions such as the In-Order property, and an O( √ dT (P T + 1)) dynamic regret bound in the worst case.\nFurthermore, inspired by Zhang et al. (2018a), we propose an improved algorithm based on DOGD, namely multiple delayed online gradient descent (Mild-OGD). The essential idea is to run multiple DOGD, each with a different learning rate that enjoys small dynamic regret for a specific path-length, and combine them with a meta-algorithm. Compared with Zhang et al. (2018a), the key challenge is that the performance of each DOGD is required by the meta-algorithm, but it is also arbitrarily delayed. To address this difficulty, our meta-algorithm is built upon the delayed Hedge-a technique for prediction with delayed expert advice (Korotin et al., 2020), which can track the best DOGD based on their delayed performance. We prove that the dynamic regret of Mild-OGD can be automatically bounded by O( dT (P T + 1)) under mild assumptions such as the In-Order property, and O( dT (P T + 1)) in the worst case. In the special case without delay, both bounds reduce to the O( T (P T + 1)) bound achieved by Zhang et al. (2018a). Finally, we demonstrate that our Mild-OGD is optimal in the worst case by deriving a matching lower bound."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we briefly review related work on OCO with arbitrary delays and the dynamic regret."
},
{
"figure_ref": [],
"heading": "OCO with Arbitrary Delays",
"publication_ref": [
"b18",
"b18",
"b46",
"b27",
"b19",
"b18",
"b27",
"b23",
"b22",
"b9",
"b19",
"b10",
"b28",
"b20",
"b33",
"b13",
"b21"
],
"table_ref": [],
"text": "To deal with arbitrary delays, Joulani et al. (2013) first propose a black-box technique, which can extend any non-delayed OCO algorithm into the delayed setting. The main idea is to pool multiple instances of the non-delayed algorithm, each of which runs over a subsequence of rounds that satisfies the non-delayed assumption. Moreover, Joulani et al. (2013) show that if the non-delayed algorithm has a static regret bound of R(T ), this technique can attain a static regret bound of dR(T /d). Notice that in the non-delayed setting, there exist plenty of algorithms with an O( √ T ) static regret bound, such as OGD (Zinkevich, 2003). As a result, combining with OGD, this technique can achieve a static regret bound of O( √ dT ). However, despite the generality of this technique, it needs to run multiple instances of the non-delayed algorithm, which could be prohibitively resourceintensive (Quanrud and Khashabi, 2015;Joulani et al., 2016). For these reasons, instead of adopting the technique of Joulani et al. (2013), subsequent studies extend many specific non-delayed OCO algorithms into the delayed setting by only running a single instance of them with delayed information about all loss functions. Specifically, Quanrud and Khashabi (2015) propose a delayed variant of OGD, and improve the static regret bound to O( √ dT ), which only magnifies the O( √ T ) bound of OGD by a coefficient depending on the average delay d, instead of the maximum delay d. By additionally assuming that the In-Order property holds, McMahan and Streeter (2014) develop a delayed variant of the adaptive gradient (AdaGrad) algorithm (McMahan and Streeter, 2010;Duchi et al., 2011), and establish a data-dependent static regret bound, which could be tighter than O( √ dT ) for sparse data. Later, Joulani et al. (2016) propose another delayed variant of AdaGrad, which can attain a data-dependent static regret bound without the In-Order property. Recently, Flaspohler et al. (2021) develop delayed variants of optimistic algorithms (Rakhlin and Sridharan, 2013;Joulani et al., 2017), which can make use of \"hints\" about expected future loss functions to improve the O( √ dT ) static regret. Additionally, Wan et al. (2022) propose a delayed variant of online Frank-Wolfe (Hazan and Kale, 2012), and establish a static regret bound of O(T 3/4 + dT 1/4 ). Their algorithm is projection-free, and thus can be much efficiently implemented over complex constraints.\nWe also notice that Korotin et al. (2020) consider the problem of prediction with expert advice-a special case of OCO with linear loss functions and simplex decision sets, and propose delayed Hedge with an O( √ dT ) static regret bound."
},
{
"figure_ref": [],
"heading": "Dynamic Regret",
"publication_ref": [
"b46",
"b11",
"b8",
"b43",
"b2",
"b17",
"b5",
"b38",
"b25",
"b39",
"b1",
"b32",
"b42",
"b35",
"b36",
"b34"
],
"table_ref": [],
"text": "Dynamic regret of OCO is first introduced by Zinkevich (2003), who demonstrates that OGD can attain a dynamic regret bound of O( √ T (P T + 1)) by simply utilizing a constant learning rate. Later, Zhang et al. (2018a) establish a lower bound of Ω( T (P T + 1)) for the dynamic regret, and develop a novel algorithm to achieve an optimal dynamic regret of O( T (P T + 1)). The main idea is to run multiple instances of OGD with different learning rates in parallel, and combine them with a classical expert-tracking algorithm called Hedge (Freund and Schapire, 1997). Furthermore, subsequent studies achieve tighter dynamic regret bounds for specific data (Cutkosky, 2020) and functions with special curvatures (Zhao et al., 2020;Baby and Wang, 2021, 2022, 2023).\nBesides the arbitrary sequence of comparators, there also exist plenty of studies (Jadbabaie et al., 2015;Besbes et al., 2015;Yang et al., 2016;Mokhtari et al., 2016;Zhang et al., 2017Zhang et al., , 2018b;;Baby and Wang, 2019;Wan et al., 2021;Zhao and Zhang, 2021;Wang et al., 2021Wang et al., , 2023;;Wan et al., 2023) that focus on a restricted form of the dynamic regret, in which u t = x * t ∈ argmin x∈K f t (x). However, as discussed by Zhang et al. (2018a), the restricted dynamic regret is too pessimistic and is less flexible than the general one."
},
{
"figure_ref": [],
"heading": "Additional Discussions",
"publication_ref": [
"b35",
"b36",
"b18"
],
"table_ref": [],
"text": "Although both arbitrary delays and the dynamic regret have attracted much research interest, it is still unclear how arbitrary delays affect the dynamic regret. Recently, Wang et al. (2021Wang et al. ( , 2023) ) have demonstrated under a fixed and knowable delay d ′ , simply performing OGD with a delayed gradient ∇f t-d ′ +1 (x t-d ′ +1 ) is able to achieve a restricted dynamic regret bound of O( d ′ T (P * T + 1)) when P * T = T t=2 ∥x * t -x * t-1 ∥ 2 is also knowable. However, both their algorithm and theoretical guarantee do not apply to the general dynamic regret under arbitrary delays.\nMoreover, one may try to extend existing algorithms with dynamic regret bounds into the delayed setting by utilizing the black-box technique of Joulani et al. (2013). However, we want to emphasize that they focus on the static regret, and their analysis cannot yield a dynamic regret bound. In addition, since their technique does not achieve a static regret bound of O( √ dT ), it seems also unable to achieve the O( dT (P T + 1)) dynamic regret even under the In-Order assumption."
},
{
"figure_ref": [],
"heading": "Main Results",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we first introduce necessary assumptions, and then present our DOGD and Mild-OGD. Finally, we provide a matching lower bound to demonstrate the optimality of our Mild-OGD in the worst case."
},
{
"figure_ref": [],
"heading": "Assumptions",
"publication_ref": [
"b30",
"b12",
"b23",
"b44",
"b23",
"b44",
"b29",
"b45"
],
"table_ref": [],
"text": "Assumption 1 The gradients of all loss functions are bounded by G, i.e., ∥∇f t (x)∥ 2 ≤ G for any x ∈ K and t ∈ [T ].\nAssumption 2 The decision set K contains the origin 0, and its diameter is bounded by D, i.e., ∥x -y∥ 2 ≤ D for any x, y ∈ K.\nAssumption 3 Delays do not change the arrival order of gradients, i.e., the gradient ∇f i (x i ) is received before the gradient ∇f j (x j ), for any 1 ≤ i < j ≤ T .\nRemark: The first two assumptions have been commonly utilized in previous studies on OCO (Shalev-Shwartz, 2011;Hazan, 2016). To further justify the rationality of Assumption 3, we notice that parallel and distributed optimization (McMahan and Streeter, 2014;Zhou et al., 2018) is also a representative application of delayed OCO. For parallel optimization with many threads, the delay is mainly caused by the computing time of gradients. Thus, as in McMahan and Streeter (2014), it is reasonable to assume that these delays enjoy the In-Order assumption, because the gradient computed first is more likely to be obtained first. Even for general parallel and distributed optimization, polynomially growing delays, which imply d i ≤ d j for i < j and thus satisfy the In-Order assumption, have received much attention in recent years (Zhou et al., 2018;Ren et al., 2020;Zhou et al., 2022). Moreover, we want to emphasize that Assumption 3 is only utilized to achieve the dynamic regret depending on the average delay d, and the case without this assumption is also considered."
},
{
"figure_ref": [],
"heading": "DOGD with Dynamic Regret",
"publication_ref": [],
"table_ref": [],
"text": "In the following, we first introduce detailed procedures of DOGD, and then present its theoretical guarantees on the dynamic regret."
},
{
"figure_ref": [],
"heading": "Detailed Procedures",
"publication_ref": [
"b46",
"b27"
],
"table_ref": [],
"text": "Recall that in the non-delayed setting, the classical OGD algorithm (Zinkevich, 2003) at each round t updates the decision as\nx t+1 = argmin x∈K ∥x -(x t -η∇f t (x t ))∥ 2 2 (1)\nwhere η is a learning rate. To handle the setting with arbitrary delays, Quanrud and Khashabi (2015) have proposed a delayed variant of OGD by replacing ∇f t (x t ) with the sum of gradients received in round t. However, it actually ignores the arrival order of gradients, and thus cannot benefit from the In-Order property when minimizing the dynamic regret.\nTo address this limitation, we propose a new delayed variant of OGD, which performs a gradient descent step for each delayed gradient according to their arrival order. Specifically, our algorithm is named as delayed online gradient descent (DOGD) and outlined in Algorithm 1, where τ records the number of generated decisions and y τ denotes the τ -th generated decision. Initially, we set y 1 = 0 and τ = 1. At each round t ∈ [T ], we first play the latest decision x t = y τ and query the gradient ∇f t (x t )."
},
{
"figure_ref": [],
"heading": "Algorithm 1 DOGD",
"publication_ref": [],
"table_ref": [],
"text": "1: Input: a learning rate η 2: Initialization: set y 1 = 0 and τ = 1\n3: for t = 1, • • • , T do 4:\nPlay x t = y τ and query ∇f t (x t )\n5:\nReceive {∇f k (x k )|k ∈ F t } 6:\nfor k ∈ F t (in the ascending order) do 7:\nCompute y τ +1 as in (2) and set τ = τ + 1 8:\nend for 9: end for After that, due to the effect of arbitrary delays, we receive a set of delayed gradients\n{∇f k (x k )|k ∈ F t }, where F t = {k ∈ [T ]|k + d k -1 = t}.\nFor each k ∈ F t , inspired by (1), we perform the following update\ny τ +1 = argmin x∈K ∥x -(y τ -η∇f k (x k ))∥ 2 2 (2)\nand then set τ = τ + 1. Moreover, to utilize the In-Order property, elements in the set F t are sorted and traversed in the ascending order."
},
{
"figure_ref": [],
"heading": "Theoretical Guarantees",
"publication_ref": [
"b46",
"b27",
"b27",
"b7"
],
"table_ref": [],
"text": "We notice that due to the effect of delays, there could exist some gradients that arrive after round T . Although our DOGD does not need to utilize these gradients, they are useful to facilitate our analysis and discussion. Therefore, in the analysis of DOGD, we virtually set x t = y τ and perform steps 5 to 8 in Algorithm 1 at some additional rounds t = T + 1, . . . , T + d -1. In this way, all queried gradients are utilized to generate decisions y 1 , . . . , y T +1 .\nMoreover, we denote the time-stamp of the τ -th utilized gradient by c τ . To help understanding, one can imagine that DOGD also sets c τ = k at the beginning of its step 7. Then, we establish the following theorem with only Assumptions 1 and 2.\nTheorem 1 Under Assumptions 1 and 2, for any comparator sequence u 1 , . . . ,\nu T ∈ K, Algorithm 1 ensures R(u 1 , • • • , u T ) ≤ D 2 + DP T η + ηG 2 T t=1 m t + T t=1 G∥u t -u ct ∥ 2 (3)\nwhere\nm t = t -t-1 i=1 |F i |.\nRemark: The value of m t -1 actually counts the number of gradients that have been queried, but still not received at the end of round t -1. Since the gradient ∇f t (x t ) will only be counted as an unreceived gradient in d t -1 rounds, it is easy to verify that\nT t=1 m t ≤ T t=1 d t = dT. (4)\nTherefore, the first two terms in the right side of (3) are upper bounded by (2D +P T )G √ dT so long as\nη = D G T t=1 m t .(5)\nHowever, we still need to bound the last term in the right side of (3), which reflects the \"comparator drift\" caused by arbitrary delays, and has never appeared in previous studies on the delayed feedback and dynamic regret.\nTo this end, we establish the following lemma regarding the comparator drift.\nLemma 1 Under Assumption 2, for any comparator sequence u 1 , . . . , u T ∈ K, Algorithm 1 ensures\nT t=1 ∥u t -u ct ∥ 2 ≤ 0, if Assumption 3 holds; min {T D, 2dP T } , otherwise. Remark: It is easy to verify that min {T D, 2dP T } ≤ 2dT DP T .(6)\nTherefore, Lemma 1 implies that the comparator drift can be upper bounded by O( √ dT P T ) in the worst case, and vanishes when the In-Order property holds.\nBy further combining Theorem 1 with (4) and Lemma 1, we derive the following corollary.\nCorollary 2 Under Assumptions 1 and 2, by setting η as in (5), Algorithm 1 ensures\nR(u 1 , • • • , u T ) ≤(2D + P T )G dT + C for any comparator sequence u 1 , . . . , u T ∈ K, where C = 0, if Assumption 3 also holds; min {T GD, 2dGP T } , otherwise.(7)\nRemark: From Corollary 2, our DOGD enjoys a dynamic regret bound of O( √ dT (P T + 1) + C), which is adaptive to the upper bound of comparator drift. First, combining it with (6) and d ≤ d, the dynamic regret of DOGD can be bounded by O( √ dT (P T + 1)) in the worst case, which magnifies the O( √ T (P T + 1)) dynamic regret of OGD (Zinkevich, 2003) in the non-delayed setting by a coefficient depending on the maximum delay d. Second, in case C ≤ O( √ dT P T ), the dynamic regret of DOGD automatically reduces to O( √ dT (P T + 1)), which depends on the average delay. According to (7), this condition can be simply satisfied for all possible P T when the In-Order property holds or d ≤ √ dT . Third, by\nsubstituting u 1 = • • • = u T into\nCorollary 2, we find that DOGD can attain a static regret bound of O( √ dT ) for arbitrary delays, which matches the best existing result (Quanrud and Khashabi, 2015).\nRemark: Notice that Corollary 2 needs to set the learning rate as in (5). At first glance, this setting may become a limitation of DOGD, because the value of T t=1 m t is generally unknown in practice. However, we note that Quanrud and Khashabi (2015) also face this issue when minimizing the static regret of OCO with arbitrary delays, and have introduced a simple solution by utilizing the standard \"doubling trick\" (Cesa-Bianchi et al., 1997) to adaptively adjust the learning rate. The main insight behind this solution is that the value of T t=1 m t can be calculated on the fly. The details about DOGD with the doubling trick are provided in the appendix."
},
{
"figure_ref": [],
"heading": "Mild-OGD with Improved Dynamic Regret",
"publication_ref": [
"b21",
"b21",
"b7"
],
"table_ref": [],
"text": "One unsatisfactory point of DOGD is that the dynamic regret linearly depends on the path length. Notice that if only a specific path length P T is considered, from Theorem 1, we can tune the learning rate as\nη * = D(D + P T ) G T t=1 m t\nand obtain dynamic regret bounds sublinear to P T . However, our goal is to minimize the dynamic regret with respect to any possible path length P T . To address this dilemma, inspired by Zhang et al. (2018a), we develop an algorithm that runs multiple DOGD as experts, each with a different learning rate for a specific path length, and combines them with a meta-algorithm.\nIt is worth noting that the meta-algorithm of Zhang et al. (2018a) is incompatible to the delayed setting studied here. To this end, we adopt the delayed Hedge (Korotin et al., 2020), an expert-tracking algorithm in the delayed setting, to design our metaalgorithm. Moreover, there exist two options for the meta-algorithm to maintain these expert-algorithms: running them over the original functions {f t (x)} t∈[T ] or the surrogate functions {ℓ t (x)} t∈[T ] , where\nℓ t (x) = ⟨∇f t (x t ), x -x t ⟩ (8)\nand x t is the decision of the meta-algorithm. In this paper, we choose the second option, because the surrogate functions allow expert-algorithms to reuse the gradient of the metaalgorithm, and thus can avoid the inconsistent delay between the meta-algorithm and the expert-algorithm. Specifically, our algorithm is named as multiple delayed online gradient descent (Mild-OGD), and stated below.\nMeta-algorithm Let H denote a set of learning rates for experts. We first activate a set of experts {E η |η ∈ H} by invoking the expert-algorithm for each learning rate η ∈ H. Let η i be the i-th smallest learning rate in H. Following Zhang et al. (2018a), the initial weight of each expert E η i is set as\nw η i 1 = |H|+1 i(i+1)|H| .\nIn each round t ∈ [T ], our meta-algorithm receives a decision x η t from each expert E η , and then plays the weighted decision as x t = η∈H w η t x η t . After that, it queries the gradient ∇f t (x t ), but only receives {∇f k (x k )|k ∈ F t } due to the effect of arbitrary delays. Then, according to the delayed Hedge (Korotin et al., 2020), we Algorithm 2 Mild-OGD: Meta-algorithm 1: Input: a parameter α and a set H containing learning rates for experts 2: Activate a set of experts {E η |η ∈ H} by invoking the expert-algorithm for each learning rate η ∈ H 3: Sort learning rates in the ascending order, i.e., η 1 ≤ • • • ≤ η |H| , and set\nw η i 1 = |H|+1 i(i+1)|H| 4: for t = 1, • • • , T do 5:\nReceive x η t from each expert E η 6:\nPlay the decision x t = η∈H w η t x η t 7:\nQuery ∇f t (x t ) and receive {∇f k (x k )|k ∈ F t } 8:\nUpdate the weight of each expert as in ( 9)\n9:\nSend {∇f k (x k )|k ∈ F t } to each expert E η 10: end for Algorithm 3 Mild-OGD: Expert-algorithm 1: Input: a learning rate η 2: Initialization: set y η 1 = 0 and τ = 1\n3: for t = 1, • • • , T do 4: Submit x η t = y η τ to the meta-algorithm 5: Receive gradients {∇f k (x k )|k ∈ F t } from the meta-algorithm 6:\nfor k ∈ F t (in the ascending order) do 7:\nCompute y η τ +1 as in ( 10) and set τ = τ + 1 8:\nend for 9: end for update the weight of each expert as\nw η t+1 = w η t e -α k∈F t ℓ k (x η k ) µ∈H w µ t e -α k∈F t ℓ k (x µ k ) (9)\nwhere α is a parameter and ℓ k (x) is defined in (8). This is the critical difference between our meta-algorithm and that in Zhang et al. (2018a), which updates the weight according to the vanilla Hedge (Cesa-Bianchi et al., 1997). Finally, we send gradients {∇f k (x k )|k ∈ F t } to each expert E η so that they can update their own decisions without querying additional gradients. The detailed procedures of our meta-algorithm are summarized in Algorithm 2."
},
{
"figure_ref": [],
"heading": "Expert-algorithm",
"publication_ref": [],
"table_ref": [],
"text": "The expert-algorithm is instantiated by running DOGD over the surrogate loss function defined in (8), instead of the real loss function. To emphasize this difference, we present its procedures in Algorithm 3. The input and initialization are the same as those in DOGD. At each round t ∈ [T ], the expert-algorithm first submits the decision x η t = y η τ to the meta-algorithm, and then receives gradients {∇f k (x k )|k ∈ F t } from the meta-algorithm. For each k ∈ F t , it updates the decision as\ny η τ +1 = argmin x∈K ∥x -(y η τ -η∇f k (x k ))∥ 2 2 (10)\nand sets τ = τ + 1.\nWe have the following theoretical guarantee for the dynamic regret of Mild-OGD.\nTheorem 3 Let m t = t -t-1 i=1 |F i |.\nUnder Assumptions 1 and 2, by setting\nH = η i = 2 i-1 D G √ β i = 1, • • • , N and α = 1 GD √ β\nwhere N = 1 2 log 2 (T + 1) + 1 and β = T t=1 m t , Algorithm 2 ensures\nR(u 1 , • • • , u T ) ≤(3 D(D + P T ) + D)G dT + C + 2GD dT ln (k + 1) =O dT (P T + 1) + C\nfor any comparator sequence u 1 , . . . , u T ∈ K, where k = log 2 (P T + D)/D + 1 and C is defined in (7).\nRemark: Theorem 3 shows that Mild-OGD can attain an O( dT (P T + 1) + C) dynamic regret bound, which is also adaptive to the upper bound of comparator drift. Due to (6) and d ≤ d, this dynamic regret bound becomes O( dT (P T + 1)) in the worst case. Moreover, its reduces to O( dT (P T + 1)) in case C ≤ O( dT P T ), which can be satisfied for all possible P T when the In-order property holds or for P T ≤ dT /d 2 . Compared with the dynamic regret of DOGD, Mild-OGD reduces the linear dependence on P T to be sublinear. Moreover, compared with the optimal O( T (P T + 1)) bound achieved in the non-delayed setting (Zhang et al., 2018a), Mild-OGD magnifies it by a coefficient depending on delays.\nWe also notice that although Theorem 3 requires the value of T t=1 m t to tune parameters, as previously discussed, this requirement can be removed by utilizing the doubling trick. The details about Mild-OGD with the doubling trick are provided in the appendix."
},
{
"figure_ref": [],
"heading": "Lower Bound",
"publication_ref": [],
"table_ref": [],
"text": "Finally, we show that our Mild-OGD is optimal in the worst case by establishing the following lower bound.\nTheorem 4 Let L = ⌈T D/ max{P, D}⌉. Suppose K = [-D/(2 √ n), D/(2 √ n)] n which\nsatisfies Assumption 2. For any OCO algorithm, any P ∈ [0, T D], and any positive integer d, there exists a sequence of comparators\nu 1 , • • • , u T ∈ K satisfying P T ≤ P , a se- quence of functions f 1 (x), • • • , f T (x)\nsatisfying Assumption 1, and a sequence of delays\n1 ≤ d 1 , • • • , d T ≤ d such that R(u 1 , • • • , u T ) ≥ DGT 2 √ 2 , if d > L; G dD max{P, D}T 4 √ 2 , otherwise.\nRemark: From Theorem 4, if d > L = Ω(T /(P T + 1)), there exists an Ω(T ) lower bound on the dynamic regret, which can be trivially matched by any OCO algorithm including our Algorithm 2. As a result, we mainly focus on the case d < L, and notice that Theorem 4 essentially establishes an Ω( dT (P T + 1)) lower bound, which matches the O( dT (P T + 1)) dynamic regret of our Mild-OGD in the worst case. To the best of our knowledge, this is the first lower bound for the dynamic regret of the delayed OCO."
},
{
"figure_ref": [],
"heading": "Analysis",
"publication_ref": [],
"table_ref": [],
"text": "We prove Theorems 1 and 3 in this section, and the omitted proofs can be found in the appendix."
},
{
"figure_ref": [],
"heading": "Proof of Theorem 1",
"publication_ref": [],
"table_ref": [],
"text": "According to the convexity of functions, we have\nR(u 1 , • • • , u T ) ≤ T t=1 ⟨∇f t (x t ), x t -u t ⟩ = T t=1 ⟨∇f ct (x ct ), x ct -u ct ⟩ (11)\nwhere the first equality is due to the fact that c 1 , . . . , c T is a permutation of 1, . . . , T .\nAccording to Algorithm 1, it is easy to verify that\nx t = y τt (12\n)\nwhere 11) with ( 12), we have\nτ t = 1 + t-1 i=1 |F i |. Combining (\nR(u 1 , • • • , u T ) ≤ T t=1 ∇f ct (x ct ), y τc t -u ct = T t=1 ∇f ct (x ct ), y t -u t + y τc t -y t + T t=1 ⟨∇f ct (x ct ), u t -u ct ⟩ ≤ T t=1 ⟨∇f ct (x ct ), y t -u t ⟩ + T t=1 G y τc t -y t 2 + T t=1 G∥u t -u ct ∥ 2 (13)\nwhere the last inequality is due to Assumption 1.\nThen, let y ′ t+1 = y t -η∇f ct (x ct ). For the first term in the right side of (13), we have\nT t=1 ⟨∇f ct (x ct ), y t -u t ⟩ = T t=1 y t -y ′ t+1 , y t -u t η = T t=1 ∥y t -u t ∥ 2 2 -∥y ′ t+1 -u t ∥ 2 2 + ∥η∇f ct (x ct )∥ 2 2 2η ≤ T t=1 1 2η ∥y t -u t ∥ 2 2 -∥y t+1 -u t ∥ 2 2 + ηT G 2 2 = T t=1 1 2η ∥y t ∥ 2 2 -∥y t+1 ∥ 2 2 + T t=1 1 η ⟨y t+1 -y t , u t ⟩ + ηT G 2 2 ≤ 1 η ⟨y T +1 , u T ⟩ + T t=2 1 η ⟨u t-1 -u t , y t ⟩ + ηT G 2 2 ≤ 1 η ∥y T +1 ∥ 2 ∥u T ∥ 2 + T t=2 1 η ∥u t-1 -u t ∥ 2 ∥y t ∥ 2 + ηT G 2 2 ≤ D 2 + DP T η + ηT G 2 2 (14)\nwhere the first inequality is due to Assumption 1, the second inequality is due to y 1 = 0 and ∥y T +1 ∥ 2 2 ≥ 0, and the last inequality is due to Assumption 2. In the following, we proceed to bound the second term in the right side of (13). Notice that before round c t , Algorithm 1 has received τ ct -1 gradients, and thus has generated\ny 1 , • • • , y τc t .\nMoreover, let q = c t + d ct -1. It is easy to verify that q ≥ c t , and thus Algorithm 1 has also generated y 1 , • • • , y τc t before round q. Since the gradient ∇f ct (x ct ) is used to update y t in round q, we have\nτ ct ≤ t.(15)\nCombining ( 15) with Assumption 1, we have\nT t=1 y τc t -y t 2 ≤ T t=1 t-1 k=τc t ∥y k -y k+1 ∥ 2 ≤ T t=1 t-1 k=τc t y k -y ′ k+1 2 ≤ T t=1 t-1 k=τc t ∥η∇f c k (x c k )∥ 2 ≤ ηG T t=1 (t -τ ct ) .(16)\nBecause of the definitions of τ t and m t , we further have\nT t=1 (t -τ ct ) = T t=1 (t -1) - T t=1 ct-1 i=1 |F i | = T t=1 (t -1) - T t=1 t-1 i=1 |F i | = T t=1 t -1 - t-1 i=1 |F i | = T t=1 (m t -1)(17)\nwhere the second equality is due to the fact that c 1 , . . . , c T is a permutation of 1, . . . , T . Then, combining ( 16) with ( 17), we have\nT t=1 G y τc t -y t 2 ≤ ηG 2 T t=1 (m t -1) .(18)\nFinally, combining ( 13) with ( 14) and ( 18), we have\nT t=1 (f t (x t ) -f t (u t )) ≤ D 2 + DP T η + ηG 2 T t=1 m t + T t=1 G∥u t -u ct ∥ 2 ."
},
{
"figure_ref": [],
"heading": "Proof of Theorem 3",
"publication_ref": [],
"table_ref": [],
"text": "Let η * = D(D + P T )/(βG 2 ), where β = T t=1 m t . From Assumption 2, we have\n0 ≤ P T = T t=2 ∥u t -u t-1 ∥ 2 ≤ T D which implies that η 1 = D G √ β ≤ η * ≤ D √ T + 1 G √ β ≤ η |H| .\nTherefore, for any possible value of P T , there must exist a learning rate η k ∈ H such that\nη k ≤ η * ≤ 2η k (19\n)\nwhere k = ⌊log 2 (P T + D)/D⌋ + 1.\nThen, the dynamic regret can be upper bounded as follows\nR(u 1 , • • • , u T ) ≤ T t=1 ⟨∇f t (x t ), x t -u t ⟩ = T t=1 ℓ t (x t ) - T t=1 ℓ t (x η k t ) + T t=1 ℓ t (x η k t ) - T t=1 ℓ t (u t ) .(20)\nTo bound the first term in the right side of (20), we introduce the following lemma.\nLemma 2 Let m t = t -t-1 i=1 |F i |.\nUnder Assumptions 1 and 2, for any η ∈ H, Algorithm 2 has\nT t=1 ℓ t (x t ) - T t=1 ℓ t (x η t ) ≤ 1 α ln 1 w η 1 + αG 2 D 2 T t=1 m t .\nCombining Lemma 2 with (1/w η k 1 ) ≤ (k + 1) 2 and α =\n1 GD √ T t=1 mt\n, under Assumptions 1 and 2, we have\nT t=1 ℓ t (x t ) - T t=1 ℓ t (x η k t ) ≤2GD T t=1 m t ln(k + 1) + GD T t=1 m t ≤2GD dT ln (k + 1) + GD dT\nwhere the last inequality is due to (4). We also notice that each expert E η actually is equal to Algorithm 1 running with ℓ 1 (x), • • • , ℓ T (x), where each gradient ∇ℓ t (x η t ) = ∇f t (x t ) is delayed to the end of round t + d t -1. Therefore, combining Theorem 1 with Lemma 1 and the definition of C in (7), under Assumptions 1 and 2, we have\nT t=1 ℓ t (x η k t ) - T t=1 ℓ t (u t ) ≤ D 2 + DP T η k + η k G 2 T t=1 m t + C ≤ 2(D 2 + DP T ) η * + η * G 2 T t=1 m t + C ≤3G D(D + P T ) dT + C\nwhere the second inequality is due to (19), and the last inequality is due to the definition of η * and (4). Finally, we complete this proof by combining (20) with the above two inequalities."
},
{
"figure_ref": [],
"heading": "Conclusion and Future Work",
"publication_ref": [
"b27",
"b19"
],
"table_ref": [],
"text": "In this paper, we study the dynamic regret of OCO with arbitrary delays. To this end, we first propose a simple algorithm called DOGD, the dynamic regret of which can be automatically bounded by O( √ dT (P T + 1)) under mild assumptions such as the In-Order property, and O( √ dT (P T + 1)) in the worst case. Furthermore, base on DOGD, we develop an improved algorithm called Mild-OGD, which can automatically enjoy an O( dT (P T + 1)) dynamic regret bound under mild assumptions such as the In-Order property, and an O( dT (P T + 1)) dynamic regret bound in the worst case. Finally, we provide a matching lower bound to show the optimality of our Mild-OGD in the worst case.\nNotice that the O( √ dT ) static regret bound can be achieved under arbitrary delays (Quanrud and Khashabi, 2015). Thus, it is natural to ask whether the O( dT (P T + 1)) dynamic regret bound can also be achieved without additional assumptions. However, from Theorem 1, compared with the static regret, it is more challenging to minimize the dynamic regret in the delayed setting, because delays will further cause a comparator drift, i.e., T t=1 ∥u t -u ct ∥ 2 . It seems highly non-trivial to reduce the comparator drift without additional assumptions, and we leave this question as a future work.\nAdditionally, we have utilized the doubling trick to avoid tunning the learning rate with the unknown cumulative delay. One potential limitation of this technique is that it needs to repeatedly restart itself, while forgetting all the preceding information. For minimizing the static regret with arbitrary delays, Joulani et al. (2016) have addressed this limitation by continuously adjusting the learning rate according to the norm of received gradients. Thus, it is also appealing to extend this idea for minimizing the dynamic regret with arbitrary delays."
},
{
"figure_ref": [],
"heading": "Appendix A. Proof of Lemma 1",
"publication_ref": [],
"table_ref": [],
"text": "We first consider the case with Assumption 3, in which all delayed gradients are received in their original order. Since Algorithm 1 utilizes the received gradients in the ascending order, it is easy to verify that\nc 1 = 1, c 2 = 2, . . . , c T = T which directly implies that T t=1 ∥u t -u ct ∥ 2 = 0. (21\n)\nThen, we proceed to consider the case without Assumption 3. Since ∇f ct (x ct ) is the t-th used gradient and arrives at the end of round c t + d ct -1, it is not hard to verify that\nt ≤ c t + d ct -1 ≤ c t + d -1 (22)\nfor any t ∈ [T ], and there are at most t -1 arrived gradients before round c t + d ct -1. Notice that gradients queried at rounds 1, • • • , t must have arrived at the end of round t + d -1. Therefore, we also have\nc t + d ct -2 < t + d -1, which implies that c t ≤ t + d -d ct ≤ t + d -1. (23\n)\nIf t ∈ [T ] and c t ≤ t, according to ( 22), we have\n∥u t -u ct ∥ 2 ≤ t-1 k=ct ∥u k+1 -u k ∥ 2 ≤ min{ct+d-2,T -1} k=ct ∥u k+1 -u k ∥ 2 . (24\n)\nOtherwise, if t ∈ [T ] and c t > t, according to (23), we have\n∥u t -u ct ∥ 2 ≤ ct-1 k=t ∥u k+1 -u k ∥ 2 ≤ min{t+d-2,T -1} k=t ∥u k+1 -u k ∥ 2 . (25\n)\nCombining ( 24) and ( 25), we have\nT t=1 ∥u t -u ct ∥ 2 ≤ T t=1 min{ct+d-2,T -1} k=ct ∥u k+1 -u k ∥ 2 + T t=1 min{t+d-2,T -1} k=t ∥u k+1 -u k ∥ 2 =2 T t=1 min{t+d-2,T -1} k=t ∥u k+1 -u k ∥ 2 =2 d-1 k=1 T -1 t=k ∥u t+1 -u t ∥ 2 ≤2 d k=1 T -1 t=1 ∥u t+1 -u t ∥ 2 =2dP T (26)\nwhere the first equality is due to the fact that c 1 , . . . , c T is a permutation of 1, . . . , T . Moreover, according to Assumption 2, we also have\nT t=1 ∥u t -u ct ∥ 2 ≤ T D. (27\n)\nFinally, this proof can be completed by combining ( 21), ( 26), and ( 27)."
},
{
"figure_ref": [],
"heading": "Appendix B. Proof of Lemma 2",
"publication_ref": [
"b9",
"b16"
],
"table_ref": [],
"text": "We first define\nL η t = t i=1 k∈F i ℓ k (x η k ), Lη t = t i=1 ℓ i (x η i ), and Wt = η∈H w η 1 e -α Lη t .\nMoreover, we define\nc t = (L η t ) η∈H ∈ R |H| , ct = ( Lη t ) η∈H ∈ R |H| , and w t = (w η t ) η∈H ∈ R |H| .\nAccording to Algorithm 2, for any t ≥ 1, it is easy to verify that\nw η t+1 = w η t e -α k∈F t ℓ k (x η k ) µ∈H w µ t e -α k∈F t ℓ k (x µ k ) = w η 1 e -αL η t µ∈H w µ 1 e -αL µ t .\nCombining with the above definitions, we have\nw t+1 = argmin w∈∆ - 1 α ln(w 1 ) + c t , w + 1 α R(w)\nwhere ∆ = {w ⪰ 0|⟨w, 1⟩ = 1} and R(w) = i w i ln w i . Similarly, for any t ≥ 1, we define\nwt+1 = argmin w∈∆ - 1 α ln(w 1 ) + ct , w + 1 α R(w)\nIn this way, for any η ∈ H and t ≥ 1, we have wt+1 = ( wη t+1 ) η∈H , where\nwη t+1 = w η 1 e -α Lη t µ∈H w µ 1 e -α Lµ t .\nMoreover, we define w1 = w 1 and xt = η∈H wη t x η t .\n(28)\nThen, we will bound the distance between xt and x t based on the following lemma.\nLemma 3 (Lemma 5 in Duchi et al. (2011)\n) Let Π K (u, α) = argmin x∈K ⟨u, x⟩ + 1 α R(x). If R(x) is 1-strongly convex with respect to a norm ∥ • ∥, it holds that ∥Π K (u, α) -Π K (v, α)∥ ≤ α∥u -v∥ *\nfor any u and v, where ∥ • ∥ * is the dual norm of ∥ • ∥.\nSince R(w) = i w i ln w i is 1-strongly convex with respect to ∥ • ∥ 1 , by applying Lemma 3, for any t > 1, we have\n∥x t -x t ∥ 2 = η∈H ( wη t -w η t )x η t 2 ≤ η∈H | wη t -w η t | ∥x η t ∥ 2 ≤D∥ wt -w t ∥ 1 ≤ αD∥c t-1 -c t-1 ∥ ∞ . Let U t = [t] \\ ∪ i∈[t] F i .\nNote that U t actually records the time stamp of gradients that are queried, but still not arrive at the end of round t. Then, for t > 1, it is not hard to verify that\n∥x t -x t ∥ 2 ≤αD∥c t-1 -c t-1 ∥ ∞ ≤ αD max η∈H k∈U t-1 ℓ k (x η k ) ≤α t -1 - t-1 i=1 |F i | GD 2 = α (m t -1) GD 2 (29)\nwhere the last inequality is due to the definition of U t and the fact that Assumptions 1 and 2 ensures\n|ℓ k (x η k )| = |⟨∇f k (x k ), x η k -x k ⟩| ≤ ∥∇f k (x k )∥ 2 ∥x η k -x k ∥ 2 ≤ GD (30)\nfor any k ∈ [T ] and η ∈ H.\nThe above inequality shows that xt is close to x t . In the following, we first focus on the analysis of xt , and then combine with the distance between xt and x t .\nTo this end, we notice that ln WT = ln\n η∈H w η 1 e -α Lη T ≥ ln max η∈H w η 1 e -α Lη T = -α min η∈H Lη T + 1 α ln 1 w η 1 . (31\n)\nNext, for any t ≥ 2, we have \nln Wt Wt-1 = ln η∈H w η 1 e -α Lη t η∈H w η 1 e -α Lη t-1 = ln η∈H w η 1 e -α Lη t-1 e -αℓt(x η t ) η∈H w η 1 e -α Lη\nTo proceed, we introduce Hoeffding's inequality (Hoeffding, 1963).\nLemma 4 Let X be a random variable with a ≤ X ≤ b. Then, for any s ∈ R, it holds that\nln E[e sX ] ≤ sE[X] + s 2 (b -a) 2 8 .\nFrom (30) and Lemma 4, we have ln\n η∈H wη t e -αℓt(x η t ) ≤ -α η∈H wη t ℓ t (x η t ) + α 2 G 2 D 2 2 ≤ -αℓ t (x t ) + α 2 G 2 D 2 2 (34)\nwhere the second inequality is due to Jensen's inequality and (28). Combining ( 33) with (34), we have\nln WT ≤ -α T t=1 ℓ t (x t ) + α 2 G 2 D 2 T 2 .\nThen, by further combining with (31), we have\nT t=1 ℓ t (x t ) -min η∈H T t=1 ℓ t (x η t ) + 1 α ln 1 w η 1 ≤ αG 2 D 2 T 2 .\nFinally, combining with (29), for any η ∈ H, we have\nT t=1 ℓ t (x t ) - T t=1 ℓ t (x η t ) + 1 α ln 1 w η 1 = T t=1 ℓ t (x t ) - T t=1 ℓ t (x t ) + T t=1 ℓ t (x t ) - T t=1 ℓ t (x η t ) + 1 α ln 1 w η 1 ≤ T t=1 ⟨∇f t (x t ), x t -xt ⟩ + αG 2 D 2 T 2 ≤ T t=1 ∥∇f t (x t )∥ 2 ∥x t -xt ∥ 2 + αG 2 D 2 T 2 ≤αG 2 D 2 T t=1 (m t -1) + αG 2 D 2 T 2 ≤αG 2 D 2 T t=1 m t(35)\nwhich completes this proof."
},
{
"figure_ref": [],
"heading": "Appendix C. Proof of Theorem 4",
"publication_ref": [
"b37"
],
"table_ref": [],
"text": "We first note that in the non-delayed setting, the main idea to derive a lower bound of dynamic regret is to first divide total T rounds into several blocks, and then lower bound the dynamic regret over total rounds via the sum of the lower bound of static regret over each block (Zhang et al., 2018a). In the following, we will generalize this idea to derive our lower bound of dynamic regret in the delayed setting.\nTo this end, it is natural to first establish a lower bound of static regret in the delayed setting. Although the seminal work of Weinberger and Ordentlich (2002) has already provided such a lower bound, their result for general functions only holds in the special case that d divides T . To address this limitation, we establish a lower bound of static regret for any d and T , which is presented in the following lemma.\nLemma 5 Suppose K = [-D/(2 √ n), D/(2 √ n)]\nn which satisfies Assumption 2. For any OCO algorithm and any positive integer d, there exists a sequence of functions\nf 1 (x), • • • , f T (x)\nsatisfying Assumption 1 and a sequence of delays\n1 ≤ d 1 , • • • , d T ≤ d such that R(T ) ≥ DGT 2 2 ⌈T /d⌉ .\nLet Z = ⌈T /L⌉. We divide the total T rounds into Z blocks, where the length of the first Z -1 blocks is L and that of the last block is T -(Z -1)L. In this way, we can define the set of rounds in the block z as\nT z = {(z -1)L + 1, • • • , min{zL, T }}.\nThen, we define the feasible set of u 1 , • • • , u T as\nC(P ) = u 1 , • • • , u T ∈ K T t=2 ∥u t -u t-1 ∥ 2 ≤ P\nand construct a subset of C(P ) as\nC ′ (P ) = u 1 , • • • , u T ∈ K u (z-1)L+1 = • • • = u min{zL,T } , ∀z ∈ [Z]\nwhere the connection C ′ (P ) ⊆ C(P ) is derived by the definition of K and the fact that the comparator sequence in C ′ (P ) only changes Z -1 ≤ P/D times, and thus its path-length does not exceed P . Because of C ′ (P ) ⊆ C(P ) and Lemma 5, it is easy to verify that there exists a sequence of functions\nf 1 (x), • • • , f T (x) satisfying Assumption 1 and a sequence of delays 1 ≤ d 1 , • • • , d T ≤ d such that T t=1 f t (x t ) - min u 1 ,••• ,u T ∈C(P ) T t=1 f t (u t ) ≥ T t=1 f t (x t ) - min u 1 ,••• ,u T ∈C ′ (P ) T t=1 f t (u t ) = Z z=1 t∈Tz f t (x t ) -min x∈K t∈Tz f t (x) ≥ Z z=1 DG|T z | 2 2 ⌈|T z |/d⌉ .(36)\nIt is not hard to verify that\nZ z=1 DG|T z | 2 2 ⌈|T z |/d⌉ ≥ Z z=1 DG|T z | 2 2 ⌈L/d⌉ = DGT 2 2 ⌈L/d⌉ ≥ DGT 2 √ 2 , if d > L; G dD max{P, D}T 4 √ 2 , otherwise;(37)\nwhere the first inequality is due to |T z | ≤ L for any z ∈ [Z], and the last inequality is mainly due to\n⌈L/d⌉ ≤ 2L/d = 2 ⌈T D/ max{P, D}⌉ /d ≤ 4T D/(max{P, D}d) for d ≤ L.\nFinally, we complete this proof by combining ( 36) and (37)."
},
{
"figure_ref": [],
"heading": "Appendix D. Proof of Lemma 5",
"publication_ref": [
"b0"
],
"table_ref": [],
"text": "Let Z = ⌈T /d⌉. We first divide the total T rounds into Z blocks, where the length of the first Z -1 blocks is d and that of the last block is T -(Z -1)d. In this way, we can define the set of rounds in the block z as\nT z = {(z -1)d + 1, • • • , min{zd, T }}.\nFor any z ∈ [Z] and t ∈ T z , we construct the delay as\nd t = min{zd, T } -t + 1\nwhich satisfies 1 ≤ d t ≤ d. These delays ensure that the information of all functions in each block z is delayed to the end of the block, which is critical for us to construct loss functions that maximize the impact of delays on the static regret. Note that to establish the lower bound of the static regret in the non-delayed setting, one can utilize a randomized strategy to select loss functions for each round (Abernethy et al., 2008). Here, to maximize the impact of delays, we only select one loss function h z (x) for all rounds in the same block z, i.e., f t (x) = h z (x) for any t ∈ T z . Specifically, we set\nh z (x) = G √ n ⟨w z , x⟩\nwhere the i-th coordinate of w z is ±1 with probability 1/2 for any i ∈ [n] and will be denoted as w z,i . It is not hard to verify that h z (x) satisfies Assumption 1.\nFrom the above definitions, we have\nE w 1 ,••• ,w Z [R(T )] =E w 1 ,••• ,w Z T t=1 f t (x t ) -min x∈K T t=1 f t (x) =E w 1 ,••• ,w Z Z z=1 t∈Tz G √ n ⟨w z , x t ⟩ -min x∈K Z z=1 t∈Tz G √ n ⟨w z , x⟩ =E w 1 ,••• ,w Z -min x∈K Z z=1 G|T z | √ n ⟨w z , x⟩\nwhere the third equality is due to E w 1 ,••• ,w Z [⟨w z , x t ⟩] = 0 for any t ∈ T z , which can be derived by the fact that any decision x t in the block z is made before receiving the information of w z , and thus is independent with w z . Since a linear function is minimized at the vertices of the cube, we further have\nE w 1 ,••• ,w Z [R(T )] = -E w 1 ,••• ,w Z min x∈{-D/(2 √ n),D/(2 √ n)} n Z z=1 G|T z | √ n ⟨w z , x⟩ =E w 1 ,••• ,w Z n i=1 D 2 √ n Z z=1 w z,i G|T z | √ n = DG 2 E w 1 ,••• ,w Z Z z=1 w z,1 |T z | ≥ DG 2 √ 2 Z z=1 |T z | 2 ≥ DG 2 √ 2 ( Z z=1 |T z |) 2 Z = DGT 2 2 ⌈T /d⌉ (38)\nwhere the first inequality is due to the Khintchine inequality and the second inequality is due to the Cauchy-Schwarz inequality.\nThe expected lower bound in (38) implies that for any OCO algorithm and any positive integer d, there exists a particular choice of w\n1 , • • • , w Z such that R(T ) ≥ DGT 2 2 ⌈T /d⌉ ."
},
{
"figure_ref": [],
"heading": "Appendix E. DOGD with the Doubling Trick",
"publication_ref": [
"b6"
],
"table_ref": [],
"text": "As discussed before, our DOGD needs a learning rate depending on the following value\nT t=1 m t = T t=1 t - t-1 i=1 |F i | .\nHowever, it may be not available beforehand. Fortunately, the doubling trick (Cesa-Bianchi and Lugosi, 2006) provides a way to adaptively estimate this value. Specifically, it will divide the total T rounds into several epochs, and run a new instance of DOGD per epoch. Let s v and s v+1 -1 respectively denote the start round and the end round of the v-th epoch. In this way, to tune the learning rate for the v-th epoch, we only need to know the following value\ns v+1 -1 t=sv t + 1 -s v - t-1 i=sv |F sv i |\nwhere\nF sv i = {k ∈ [s v , i]|k + d k -1 = i}.\nAccording to the doubling trick, we can estimate this value to be 2 v at the start round s v of the v-th epoch. Then, for any t > s v , we first judge whether the estimate is still valid, i.e.,\nt j=sv j + 1 -s v - j-1 i=sv |F sv i | ≤ 2 v\nAlgorithm 4 DOGD with the Doubling Trick 1: Initialization: set y 1 = 0, τ = 1, v = 1, and\ns v = 1 2: for t = 1, • • • , T do 3: if t j=sv j + 1 -s v -j-1 i=sv |F sv i | > 2 v then 4:\nSet y 1 = 0, τ = 1, v = v + 1, and s v = t 5:\nend if"
},
{
"figure_ref": [],
"heading": "6:",
"publication_ref": [],
"table_ref": [],
"text": "Play x t = y τ and query ∇f t (x t )\n7:\nReceive {∇f k (x k )|k ∈ F sv t }, where F sv t = {k ∈ [s v , t]|k + d k -1 = t} 8:\nfor k ∈ F sv t (in the ascending order) do 9:\nCompute\ny τ +1 = argmin x∈K ∥x -(y τ -η v ∇f k (x k ))∥ 2 2 , where η v = D G2 v/2\n10:\nSet τ = τ + 1 11:\nend for 12: end for where the left side can be calculated at the beginning of round t. If the answer is positive, the round t is still assigned to the v-th epoch, and the instance of DOGD keeps running. Otherwise, the round t is set as the start round of the (v + 1)-th epoch, and a new instance of DOGD is activated. Notice that in the start round of the (v + 1)-th epoch, the new estimate must be valid, since t = s v+1 and\nt j=s v+1 j + 1 -s v+1 - j-1 i=s v+1 |F s v+1 i | = 1 ≤ 2 v+1 .\nMoreover, it is natural to set s 1 = 1. Then, the detailed procedures of DOGD with the doubling track are summarized in Algorithm 4. Remark: First, in Algorithm 4, the learning rate η v is set by replacing T t=1 m t in the learning rate required by Corollary 2 with 2 v . Second, in each epoch v, we do not need to utilize gradients queried before this epoch. For this reason, in Algorithm 4, we only receive\n{∇f k (x k )|k ∈ F sv t }, instead of {∇f k (x k )|k ∈ F t }.\nWe have the following theorem, which can recover the dynamic regret bound in Corollary 2 up to a constant factor.\nTheorem 5 Under Assumptions 1 and 2, for any comparator sequence u 1 , . . . , u T ∈ K, Algorithm 2 ensures\nR(u 1 , • • • , u T ) ≤ 2G (2D + P T ) √ dT √ 2 -1 + C\nwhere C is defined in (7).\nProof For any s v and j ≥ s v , we first notice that the value of j -s v -j-1 i=sv |F sv i | counts the number of gradients that have been queried over interval [s v , j -1], but still not arrive at the end of round j -1. Moreover, the gradient ∇f j (x j ) will only be counted as an unreceived gradient in d j -1 rounds. Therefore, for any s v ≤ t ≤ T , it is easy to verify that\nt j=sv j + 1 -s v - j-1 i=sv |F sv i | ≤ t j=sv d j ≤ T j=1 d j = dT.\nFor brevity, let V denote the final v of Algorithm 4, and let S = dT . It is easy to verify that V ≤ 1 + log 2 S.\nThen, let s V +1 = T + 1. We notice that for v ∈ [V ], Algorithm 4 actually starts or restarts Algorithm 1 with the learning rate of η v at round s v , which ends at round s v+1 -1. Therefore, combining Theorem 1 with Lemma 1, under Assumptions 1 and 2, we have\ns v+1 -1 t=sv f t (x t ) - s v+1 -1 t=sv f t (u t ) ≤ D 2 + D s v+1 -1 t=sv+1 ∥u t -u t-1 ∥ 2 η v + η v G 2 s v+1 -1 j=sv j + 1 -s v - j-1 i=sv |F sv i | + C v(40)\nwhere\nC v = 0, if Assumption 3 also holds; min (s v+1 -s v )GD, 2dG s v+1 -1 t=sv+1 ∥u t -u t-1 ∥ 2 , otherwise.(41)\nMoreover, we notice that Algorithm 4 also ensures that\ns v+1 -1 j=sv j + 1 -s v - j-1 i=sv |F sv i | ≤ 2 v .(42)\nBy substituting the above inequality into (40), we have\ns v+1 -1 t=sv f t (x t ) - s v+1 -1 t=sv f t (u t ) ≤ D 2 + D s v+1 -1 t=sv+1 ∥u t -u t-1 ∥ 2 η v + η v G 2 2 v + C v =G2 v/2 2D + s v+1 -1 t=sv+1 ∥u t -u t-1 ∥ 2 + C v ≤G2 v/2 (2D + P T ) + C v .(43)\nThen, because of (39), we have\nR(u 1 , • • • , u T ) = V v=1 s v+1 -1 t=sv f t (x t ) - s v+1 -1 t=sv f t (u t ) ≤ V v=1 G2 v/2 (2D + P T ) + V v=1 C v =G (2D + P T ) √ 2(2 V /2 -1) √ 2 -1 + V v=1 C v ≤ 2G (2D + P T ) √ S √ 2 -1 + V v=1 C v .(44)\nMoreover, it is not hard to verify that\nV v=1 min (s v+1 -s v )GD, 2dG s v+1 -1 t=sv+1 ∥u t -u t-1 ∥ 2 ≤ min V v=1 (s v+1 -s v )GD, V v=1 2dG s v+1 -1 t=sv+1 ∥u t -u t-1 ∥ 2 ≤ min {T GD, 2dGP T }\nAlgorithm 5 Mild-OGD with the Doubling Trick: Meta-algorithm 1: Initialization: set v = 1 and s v = 1 2: Activate a set of experts {E η |η ∈ H} by invoking the expert-algorithm for each constant η ∈ X , where Update the weight of each expert by\nH = η i = D2 i-1 G i = 1, • • • , log 2 √ T + 1 + 1 3: Set w η i t = |H|+1 i(i+1)|H| 4: for t = 1, • • • , T do 5: if t j=sv j + 1 -s v -j-1 i=sv |F sv i | > 2 v then 6: Set v = v + 1, s v = t,\nw η t+1 = w η t e -αv k∈F sv t ℓ k (x η k ) µ∈H w µ t e -αv k∈F sv t ℓ k (x µ k ) where ℓ k (x) = ⟨∇f k (x k ), x -x k ⟩ and α v = 1 GD2 v/2 12: Send {∇f k (x k )|k ∈ F sv t } to each expert E η 13: end for which implies that V v=1 C v ≤ C.(45)\nFinally, we complete this proof by substituting (45) and S = dT into (44)."
},
{
"figure_ref": [],
"heading": "Appendix F. Mild-OGD with the Doubling Trick",
"publication_ref": [],
"table_ref": [],
"text": "Similar to DOGD, Mild-OGD requires the value of T t=1 m t for setting\nα = 1 GD T t=1 m t and η i = 2 i-1 D G T t=1 m t (46\n)\nwhere α is the learning rate for updating the weight, and η i is the learning rate for the i-th expert. To address this limitation, we can utilize the doubling trick as described in the previous section. The only change is to replace DOGD with Mild-OGD. The detailed procedures of Mild-OGD with the doubling track are outlined in Algorithms 5 and 6.\nRemark: We would like to emphasize that since multiple instances of the expert-algorithm Algorithm 6 Mild-OGD with the Doubling Trick: Expert-algorithm 1: Input: a constant η 2: Initialization: set y η 1 = 0, τ = 1, v = 1, and Set τ = τ + 1 12:\ns v = 1 3: for t = 1, • • • , T do 4: if t j=sv j + 1 -s v -j-1 i=sv |F sv i | > 2 v then 5: Set y 1 = 0, τ = 1, v = v +\nend for 13: end for run over the surrogate losses defined by the meta-algorithm, these instances and the metaalgorithm will start a new epoch synchronously. Moreover, as shown in step 6 of Algorithm 5, in the start of each epoch, we need to reinitialize the weight of each expert E η . As shown in step 11, in each epoch v, we update the weight by using the learning rate α v , which replaces T t=1 m t in (46) with 2 v . Additionally, to facilitate presentation, in step 2 of Algorithm 5, each η i in H only contains the constant part that does not depend on the value of T t=1 m t . Meanwhile, according to steps 1 and 10 of Algorithm 6, the i-th expert will receive η i from the meta-algorithm, and combine it with the estimation of T t=1 m t to compute the learning rate.\nFurthermore, we have the following theorem, which can recover the dynamic regret bound in Theorem 3 up to a constant factor.\nTheorem 6 Under Assumptions 1 and 2, for any comparator sequence u 1 , . . . , u T ∈ K, Algorithm 2 ensures\nR(u 1 , • • • , u T ) ≤ 2 2 ln log 2 (D + P T ) /D + 2 + 1 GD + 3G D 2 + DP T √ dT √ 2 -1 + C\nwhere C is defined in (7).\nProof Following the proof of Theorem 5, we use V to denote the final v of Algorithms 5 and 6 and define s V +1 = T + 1. Moreover, let S = dT . It is easy to verify that (39) also holds.\nThen, we consider the dynamic regret of Algorithm 5 over the interval [s v , s v+1 -1] for each v ∈ [V ]. Let\nη v * = D G 2 D + s v+1 -1 t=sv+1 ∥u t -u t-1 ∥ 2 .\nFrom Assumption 2, we have 0 ≤ s v+1 -1 t=sv+1 ∥u t -u t-1 ∥ 2 ≤ (s v+1 -s v -1)D ≤ T D which implies that\nη 1 = D G ≤ η v * ≤ D √ T + 1 G ≤ η |H| .\nTherefore, for any possible value of s v+1 -1 t=sv+1 ∥u t -u t-1 ∥ 2 , there must exist a constant η kv ∈ H such that\nη kv ≤ η v * ≤ 2η kv(47)\nwhere\nk v = log 2 D + s v+1 -1 t=sv+1 ∥u t -u t-1 ∥ 2 /D + 1\n≤ log 2 (D + P T ) /D + 1.\nMoreover, we notice that each expert E η over the interval [s v , s v+1 -1] actually runs Algorithm 1 with the learning rate η 2 v/2 to handle the surrogate losses ℓ sv (x), • • • , ℓ s v+1 -1 (x), where each gradient ∇ℓ t (x η t ) = ∇f t (x t ) is delayed to the end of round t + d t -1 for t ∈ [s v , s v+1 -1].\nTherefore, by combining Theorem 1 with Lemma 1, under Assumptions 1 and 2, we have\ns v+1 -1 t=sv ℓ t (x η kv t ) -ℓ t (u t ) ≤ 2 v/2 D 2 + D s v+1 -1 t=sv+1 ∥u t -u t-1 ∥ 2 η kv + η kv G 2 2 v/2 s v+1 -1 j=sv j + 1 -s v - j-1 i=sv |F sv i | + C v ≤ 2 v/2 D 2 + D s v+1 -1 t=sv+1 ∥u t -u t-1 ∥ 2 η kv + η kv G 2 2 v/2 + C v ≤3G 2 v D 2 + D s v+1 -1 t=sv+1 ∥u t -u t-1 ∥ 2 + C v ≤3G 2 v (D 2 + DP T ) + C v (48)\nwhere C v is defined in (41), the second inequality is due to the fact that Algorithm 6 also ensures (42), and the third inequality is due to (47) and the definition of η v * . Moreover, it is also easy to verify that Algorithm 5 actually starts or restarts Algorithm 2 with the learning rate of α v at round s v , which ends at round s v+1 -1. Then, by using Lemma 2 with (1/w η kv sv ) ≤ (k v + 1) 2 , under Assumptions 1 and 2, we have\ns v+1 -1 t=sv ℓ t (x t ) - s v+1 -1 t=sv ℓ t (x η kv t ) ≤ 2 α v ln(k v + 1) + α v G 2 D 2 s v+1 -1 j=sv j + 1 -s v - j-1 i=sv |F sv i |\n≤ 2 ln log 2 (D + P T ) /D + 2 + 1 2 v/2 GD (49)\nwhere the second inequality is due to\nα v = 1 GD2 v/2\n, the definition of k v , and the fact that Algorithm 5 also ensures (42).\nBy combining ( 48) and ( 49), it is not hard to verify that \n√ 2(2 V /2 -1) √ 2 -1 + V v=1 C v ≤ 2 2 ln log 2 (D + P T ) /D + 2 + 1 GD + 3G D 2 + DP T √ S √ 2 -1 + V v=1 C v\nwhere the last inequality is due to (39). Finally, by substituting (45) and S = dT into the above inequality, we complete this proof."
}
] | Online convex optimization (OCO) with arbitrary delays, in which gradients or other information of functions could be arbitrarily delayed, has received increasing attention recently. Different from previous studies that focus on stationary environments, this paper investigates the delayed OCO in non-stationary environments, and aims to minimize the dynamic regret with respect to any sequence of comparators. To this end, we first propose a simple algorithm, namely DOGD, which performs a gradient descent step for each delayed gradient according to their arrival order. Despite its simplicity, our novel analysis shows that the dynamic regret of DOGD can be automatically bounded by O( √ dT (P T + 1)) under mild assumptions, and O( √ dT (P T + 1)) in the worst case, where d and d denote the average and maximum delay respectively, T is the time horizon, and P T is the path length of comparators. Furthermore, we develop an improved algorithm, which reduces those dynamic regret bounds achieved by DOGD to O( dT (P T + 1)) and O( dT (P T + 1)), respectively. The key idea is to run multiple DOGD with different learning rates, and utilize a meta-algorithm to track the best one based on their delayed performance. Finally, we demonstrate that our improved algorithm is optimal in the worst case by deriving a matching lower bound. | Non-stationary Online Convex Optimization with Arbitrary Delays | [
{
"figure_caption": "log 2 (D + P T ) /D + 2 + 1 2 v/2 GD + 3G 2 v (D 2 + DP T ) + C v log 2 (D + P T ) /D + 2 + 1 GD + 3G D 2 + DP T log 2 (D + P T ) /D + 2 + 1 GD + 3G D 2 + DP T",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Query ∇f t (x t ) and receive {∇f k (x k )|k ∈ F sv t }, where F sv t = {k ∈ [s v , t]|k+d k -1 = t}",
"figure_data": "and w η i t = |H|+1 i(i+1)|H|7:end if8: 9:Receive x η t from each expert E η Play the decision x t = η∈H w η t x η t10:11:",
"figure_id": "tab_0",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "1, ands v = t Compute y η τ +1 = argmin x∈K xy η τ -η 2 v/2 ∇f k (x k )",
"figure_data": "6:end if7: 8:Submit x η t = y η τ to the meta-algorithm Receive gradients {∇f k (x k )|k ∈ F sv t } from the meta-algorithm9:for k ∈ F t (in the ascending order) do210:211:",
"figure_id": "tab_1",
"figure_label": "",
"figure_type": "table"
}
] | Yuanyu Wan; Chang Yao; Lijun Zhang | [
{
"authors": "Jacob D Abernethy; Peter L Bartlett; Alexander Rakhlin; Ambuj Tewari",
"journal": "",
"ref_id": "b0",
"title": "Optimal stragies and minimax lower bounds for online convex games",
"year": "2008"
},
{
"authors": "Dheeraj Baby; Yu-Xiang Wang",
"journal": "",
"ref_id": "b1",
"title": "Online forecasting of total-variation-bounded sequences",
"year": "2019"
},
{
"authors": "Dheeraj Baby; Yu-Xiang Wang",
"journal": "",
"ref_id": "b2",
"title": "Optimal dynamic regret in exp-concave online learning",
"year": "2021"
},
{
"authors": "Dheeraj Baby; Yu-Xiang Wang",
"journal": "",
"ref_id": "b3",
"title": "Optimal dynamic regret in proper online learning with strongly convex losses and beyond",
"year": "2022"
},
{
"authors": "Dheeraj Baby; Yu-Xiang Wang",
"journal": "",
"ref_id": "b4",
"title": "Second order path variationals in non-stationary online learning",
"year": "2023"
},
{
"authors": "Omar Besbes; Yonatan Gur; Assaf Zeevi",
"journal": "Operations Research",
"ref_id": "b5",
"title": "Non-stationary stochastic optimization",
"year": "2015"
},
{
"authors": "Nicolò Cesa; - Bianchi; Gabor Lugosi",
"journal": "Cambridge University Press",
"ref_id": "b6",
"title": "Prediction, Learning, and Games",
"year": "2006"
},
{
"authors": "Nicolò Cesa-Bianchi; Yoav Freund; David Haussler; David P Helmbold; Robert E Schapire; Manfred K Warmuth",
"journal": "Journal of the ACM",
"ref_id": "b7",
"title": "How to use expert advice",
"year": "1997"
},
{
"authors": "Ashok Cutkosky",
"journal": "",
"ref_id": "b8",
"title": "Parameter-free, dynamic, and strongly-adaptive online learning",
"year": "2020"
},
{
"authors": "John C Duchi; Alekh Agarwal; Martin J Wainwright",
"journal": "IEEE Transactions on Automatic Control",
"ref_id": "b9",
"title": "Dual averaging for distributed optimization: Convergence analysis and network scaling",
"year": "2011"
},
{
"authors": "Genevieve E Flaspohler; Francesco Orabona; Judah Cohen; Soukayna Mouatadid; Miruna Oprescu; Paulo Orenstein; Lester Mackey",
"journal": "",
"ref_id": "b10",
"title": "Online with optimism and delay",
"year": "2021"
},
{
"authors": "Yoav Freund; Robert E Schapire",
"journal": "Journal of Computer and System Sciences",
"ref_id": "b11",
"title": "A decision-theoretic generalization of on-line learning and an application to boosting",
"year": "1997"
},
{
"authors": "Elad Hazan",
"journal": "Foundations and Trends in Optimization",
"ref_id": "b12",
"title": "Introduction to online convex optimization",
"year": "2016"
},
{
"authors": "Elad Hazan; Satyen Kale",
"journal": "",
"ref_id": "b13",
"title": "Projection-free online learning",
"year": "2012"
},
{
"authors": "Elad Hazan; Amit Agarwal; Satyen Kale",
"journal": "Machine Learning",
"ref_id": "b14",
"title": "Logarithmic regret algorithms for online convex optimization",
"year": "2007"
},
{
"authors": "Xinran He; Junfeng Pan; Ou Jin; Tianbing Xu; Bo Liu; Tao Xu; Yanxin Shi; Antoine Atallah; Ralf Herbrich; Stuart Bowers; Joaquin Q Candela",
"journal": "",
"ref_id": "b15",
"title": "Practical lessons from predicting clicks on ads at facebook",
"year": "2014"
},
{
"authors": "Wassily Hoeffding",
"journal": "Journal of the American Statistical Association",
"ref_id": "b16",
"title": "Probability inequalities for sums of bounded random variables",
"year": "1963"
},
{
"authors": "Ali Jadbabaie; Alexander Rakhlin; Shahin Shahrampour; Karthik Sridharan",
"journal": "",
"ref_id": "b17",
"title": "Online optimization: Competing with dynamic comparators",
"year": "2015"
},
{
"authors": "Pooria Joulani; András György; Csaba Szepesvári",
"journal": "",
"ref_id": "b18",
"title": "Online learning under delayed feedback",
"year": "2013"
},
{
"authors": "Pooria Joulani; András György; Csaba Szepesvári",
"journal": "",
"ref_id": "b19",
"title": "Delay-tolerant online convex optimization: Unified analysis and adaptive-gradient algorithms",
"year": "2016"
},
{
"authors": "Pooria Joulani; András György; Csaba Szepesvári",
"journal": "",
"ref_id": "b20",
"title": "A modular analysis of adaptive (non-)convex optimization: Optimism, composite objectives, and variational bounds",
"year": "2017"
},
{
"authors": "Alexander Korotin; Evgeny Vladimir V'yugin; Burnaev",
"journal": "Neurocomputing",
"ref_id": "b21",
"title": "Adaptive hedging under delayed feedback",
"year": "2020"
},
{
"authors": "H ; Brendan Mcmahan; Matthew Streeter",
"journal": "",
"ref_id": "b22",
"title": "Adaptive bound optimization for online convex optimization",
"year": "2010"
},
{
"authors": "H ; Brendan Mcmahan; Matthew Streeter",
"journal": "",
"ref_id": "b23",
"title": "Delay-tolerant algorithms for asynchronous distributed online learning",
"year": "2014"
},
{
"authors": "H ; Brendan Mcmahan; Gary Holt; D Sculley; Michael Young; Dietmar Ebner; Julian Grady; Lan Nie; Todd Phillips; Eugene Davydov; Daniel Golovin; Sharat Chikkerur; Dan Liu; Martin Wattenberg; Arnar Mar Hrafnkelsson; Tom Boulos; Jeremy Kubica",
"journal": "",
"ref_id": "b24",
"title": "Ad click prediction: a view from the trenches",
"year": "2013"
},
{
"authors": "Aryan Mokhtari; Shahin Shahrampour; Ali Jadbabaie; Alejandro Ribeiro",
"journal": "",
"ref_id": "b25",
"title": "Online optimization in dynamic environments: Improved regret rates for strongly convex problems",
"year": "2016"
},
{
"authors": "Francesco Orabona",
"journal": "",
"ref_id": "b26",
"title": "A modern introduction to online learning",
"year": "2019"
},
{
"authors": "Kent Quanrud; Daniel Khashabi",
"journal": "",
"ref_id": "b27",
"title": "Online learning with adversarial delays",
"year": "2015"
},
{
"authors": "Alexander Rakhlin; Karthik Sridharan",
"journal": "",
"ref_id": "b28",
"title": "Online learning with predictable sequences",
"year": "2013"
},
{
"authors": "Zhaolin Ren; Zhengyuan Zhou; Linhai Qiu; Ajay Deshpande; Jayant Kalagnanam",
"journal": "",
"ref_id": "b29",
"title": "Delay-adaptive distributed stochastic optimization",
"year": "2020"
},
{
"authors": "Shai Shalev-Shwartz",
"journal": "Foundations and Trends in Machine Learning",
"ref_id": "b30",
"title": "Online learning and online convex optimization",
"year": "2011"
},
{
"authors": "Shai Shalev; -Shwartz ; Yoram Singer",
"journal": "Machine Learning",
"ref_id": "b31",
"title": "A primal-dual perspective of online learning algorithm",
"year": "2007"
},
{
"authors": "Yuanyu Wan; Bo Xue; Lijun Zhang",
"journal": "",
"ref_id": "b32",
"title": "Projection-free online learning in dynamic environments",
"year": "2021"
},
{
"authors": "Yuanyu Wan; Wei-Wei Tu; Lijun Zhang",
"journal": "",
"ref_id": "b33",
"title": "Online frank-wolfe with arbitrary delays",
"year": "2022"
},
{
"authors": "Yuanyu Wan; Lijun Zhang; Mingli Song",
"journal": "",
"ref_id": "b34",
"title": "Improved dynamic regret for online frankwolfe",
"year": "2023"
},
{
"authors": "Juncheng Wang; Ben Liang; Min Dong; Gary Boudreau; Hatem Abou-Zeid",
"journal": "",
"ref_id": "b35",
"title": "Delaytolerant constrained OCO with application to network resource allocation",
"year": "2021"
},
{
"authors": "Juncheng Wang; Ming Dong; Ben Liang; Gary Boudreau; Hatem Abou-Zeid",
"journal": "IEEE/ACM Transactions on Networking",
"ref_id": "b36",
"title": "Delaytolerant OCO with long-term constraints: Algorithm and its application to network resource allocation",
"year": "2023"
},
{
"authors": "Marcelo J Weinberger; Erik Ordentlich",
"journal": "IEEE Transactions on Information Theory",
"ref_id": "b37",
"title": "On delayed prediction of individual sequences",
"year": "2002"
},
{
"authors": "Tianbao Yang; Lijun Zhang; Rong Jin; Jinfeng Yi",
"journal": "",
"ref_id": "b38",
"title": "Tracking slowly moving clairvoyant: Optimal dynamic regret of online learning with true and noisy gradient",
"year": "2016"
},
{
"authors": "Lijun Zhang; Tianbao Yang; Jinfeng Yi; Rong Jin; Zhi-Hua Zhou",
"journal": "",
"ref_id": "b39",
"title": "Improved dynamic regret for non-degenerate functions",
"year": "2017"
},
{
"authors": "Lijun Zhang; Shiyin Lu; Zhi-Hua Zhou",
"journal": "",
"ref_id": "b40",
"title": "Adaptive online learning in dynamic environments",
"year": "2018"
},
{
"authors": "Lijun Zhang; Tianbao Yang; Rong Jin; Zhi-Hua Zhou",
"journal": "",
"ref_id": "b41",
"title": "Dynamic regret of strongly adaptive methods",
"year": "2018"
},
{
"authors": "Peng Zhao; Lijun Zhang",
"journal": "",
"ref_id": "b42",
"title": "Improved analysis for dynamic regret of strongly convex and smooth functions",
"year": "2021"
},
{
"authors": "Peng Zhao; Yu-Jie Zhang; Lijun Zhang; Zhi-Hua Zhou",
"journal": "",
"ref_id": "b43",
"title": "Dynamic regret of convex and smooth functions",
"year": "2020"
},
{
"authors": "Zhengyuan Zhou; Panayotis Mertikopoulos; Nicholas Bambos; Peter Glynn; Yinyu Ye; Li-Jia Li; Li Fei-Fei",
"journal": "",
"ref_id": "b44",
"title": "Distributed asynchronous optimization with unbounded delays: How slow can you go",
"year": "2018"
},
{
"authors": "Zhengyuan Zhou; Panayotis Mertikopoulos; Nicholas Bambos; Peter W Glynn; Yinyu Ye",
"journal": "Mathematics of Operations Research",
"ref_id": "b45",
"title": "Distributed stochastic optimization with large delays",
"year": "2022"
},
{
"authors": "Martin Zinkevich",
"journal": "",
"ref_id": "b46",
"title": "Online convex programming and generalized infinitesimal gradient ascent",
"year": "2003"
}
] | [
{
"formula_coordinates": [
2,
226.33,
171.52,
159.34,
33.58
],
"formula_id": "formula_0",
"formula_text": "R(T ) = T t=1 f t (x t ) -min x∈K T t=1 f t (x)"
},
{
"formula_coordinates": [
2,
212.66,
294.93,
186.69,
33.58
],
"formula_id": "formula_1",
"formula_text": "R(u 1 , • • • , u T ) = T t=1 f t (x t ) - T t=1 f t (u t )"
},
{
"formula_coordinates": [
2,
252.71,
391.12,
106.08,
33.58
],
"formula_id": "formula_2",
"formula_text": "P T = T t=2 ∥u t -u t-1 ∥ 2"
},
{
"formula_coordinates": [
5,
215.61,
537.62,
306.39,
20.88
],
"formula_id": "formula_3",
"formula_text": "x t+1 = argmin x∈K ∥x -(x t -η∇f t (x t ))∥ 2 2 (1)"
},
{
"formula_coordinates": [
6,
95.92,
134.92,
106.3,
22.77
],
"formula_id": "formula_4",
"formula_text": "3: for t = 1, • • • , T do 4:"
},
{
"formula_coordinates": [
6,
95.92,
162.05,
147.55,
22.74
],
"formula_id": "formula_5",
"formula_text": "Receive {∇f k (x k )|k ∈ F t } 6:"
},
{
"formula_coordinates": [
6,
90,
265.75,
287.82,
25.07
],
"formula_id": "formula_6",
"formula_text": "{∇f k (x k )|k ∈ F t }, where F t = {k ∈ [T ]|k + d k -1 = t}."
},
{
"formula_coordinates": [
6,
212.43,
323.5,
309.57,
20.88
],
"formula_id": "formula_7",
"formula_text": "y τ +1 = argmin x∈K ∥x -(y τ -η∇f k (x k ))∥ 2 2 (2)"
},
{
"formula_coordinates": [
6,
90,
572.88,
432,
70.53
],
"formula_id": "formula_8",
"formula_text": "u T ∈ K, Algorithm 1 ensures R(u 1 , • • • , u T ) ≤ D 2 + DP T η + ηG 2 T t=1 m t + T t=1 G∥u t -u ct ∥ 2 (3)"
},
{
"formula_coordinates": [
6,
120.81,
655.34,
93.29,
15.24
],
"formula_id": "formula_9",
"formula_text": "m t = t -t-1 i=1 |F i |."
},
{
"formula_coordinates": [
7,
255.64,
116.18,
266.36,
33.58
],
"formula_id": "formula_10",
"formula_text": "T t=1 m t ≤ T t=1 d t = dT. (4)"
},
{
"formula_coordinates": [
7,
262.95,
188.03,
259.05,
32.77
],
"formula_id": "formula_11",
"formula_text": "η = D G T t=1 m t .(5)"
},
{
"formula_coordinates": [
7,
90,
315.1,
432,
76.91
],
"formula_id": "formula_12",
"formula_text": "T t=1 ∥u t -u ct ∥ 2 ≤ 0, if Assumption 3 holds; min {T D, 2dP T } , otherwise. Remark: It is easy to verify that min {T D, 2dP T } ≤ 2dT DP T .(6)"
},
{
"formula_coordinates": [
7,
90,
491.32,
432,
77.84
],
"formula_id": "formula_13",
"formula_text": "R(u 1 , • • • , u T ) ≤(2D + P T )G dT + C for any comparator sequence u 1 , . . . , u T ∈ K, where C = 0, if Assumption 3 also holds; min {T GD, 2dGP T } , otherwise.(7)"
},
{
"formula_coordinates": [
7,
90,
695.38,
150.07,
10.72
],
"formula_id": "formula_14",
"formula_text": "substituting u 1 = • • • = u T into"
},
{
"formula_coordinates": [
8,
259.96,
305.5,
90.89,
32.97
],
"formula_id": "formula_15",
"formula_text": "η * = D(D + P T ) G T t=1 m t"
},
{
"formula_coordinates": [
8,
247.07,
509.25,
274.93,
10.67
],
"formula_id": "formula_16",
"formula_text": "ℓ t (x) = ⟨∇f t (x t ), x -x t ⟩ (8)"
},
{
"formula_coordinates": [
8,
203.9,
649.31,
71.07,
16.27
],
"formula_id": "formula_17",
"formula_text": "w η i 1 = |H|+1 i(i+1)|H| ."
},
{
"formula_coordinates": [
9,
95.92,
145.14,
419.91,
39.66
],
"formula_id": "formula_18",
"formula_text": "w η i 1 = |H|+1 i(i+1)|H| 4: for t = 1, • • • , T do 5:"
},
{
"formula_coordinates": [
9,
95.92,
202.7,
238.35,
22.74
],
"formula_id": "formula_19",
"formula_text": "Query ∇f t (x t ) and receive {∇f k (x k )|k ∈ F t } 8:"
},
{
"formula_coordinates": [
9,
95.92,
231.13,
7.17,
7.86
],
"formula_id": "formula_20",
"formula_text": "9:"
},
{
"formula_coordinates": [
9,
95.92,
310.58,
316.19,
49.87
],
"formula_id": "formula_21",
"formula_text": "3: for t = 1, • • • , T do 4: Submit x η t = y η τ to the meta-algorithm 5: Receive gradients {∇f k (x k )|k ∈ F t } from the meta-algorithm 6:"
},
{
"formula_coordinates": [
9,
227.56,
443.32,
294.44,
35.99
],
"formula_id": "formula_22",
"formula_text": "w η t+1 = w η t e -α k∈F t ℓ k (x η k ) µ∈H w µ t e -α k∈F t ℓ k (x µ k ) (9)"
},
{
"formula_coordinates": [
9,
212.41,
664.04,
309.59,
21.68
],
"formula_id": "formula_23",
"formula_text": "y η τ +1 = argmin x∈K ∥x -(y η τ -η∇f k (x k ))∥ 2 2 (10)"
},
{
"formula_coordinates": [
10,
90,
109.39,
177.12,
15.24
],
"formula_id": "formula_24",
"formula_text": "Theorem 3 Let m t = t -t-1 i=1 |F i |."
},
{
"formula_coordinates": [
10,
183.01,
129.49,
244.22,
26.81
],
"formula_id": "formula_25",
"formula_text": "H = η i = 2 i-1 D G √ β i = 1, • • • , N and α = 1 GD √ β"
},
{
"formula_coordinates": [
10,
134.26,
186.3,
343.48,
37.43
],
"formula_id": "formula_26",
"formula_text": "R(u 1 , • • • , u T ) ≤(3 D(D + P T ) + D)G dT + C + 2GD dT ln (k + 1) =O dT (P T + 1) + C"
},
{
"formula_coordinates": [
10,
90,
485.26,
432,
17.43
],
"formula_id": "formula_27",
"formula_text": "Theorem 4 Let L = ⌈T D/ max{P, D}⌉. Suppose K = [-D/(2 √ n), D/(2 √ n)] n which"
},
{
"formula_coordinates": [
10,
90,
520.18,
432,
24.27
],
"formula_id": "formula_28",
"formula_text": "u 1 , • • • , u T ∈ K satisfying P T ≤ P , a se- quence of functions f 1 (x), • • • , f T (x)"
},
{
"formula_coordinates": [
10,
90,
547.31,
339.55,
73.71
],
"formula_id": "formula_29",
"formula_text": "1 ≤ d 1 , • • • , d T ≤ d such that R(u 1 , • • • , u T ) ≥ DGT 2 √ 2 , if d > L; G dD max{P, D}T 4 √ 2 , otherwise."
},
{
"formula_coordinates": [
11,
148.45,
195.93,
373.55,
33.58
],
"formula_id": "formula_30",
"formula_text": "R(u 1 , • • • , u T ) ≤ T t=1 ⟨∇f t (x t ), x t -u t ⟩ = T t=1 ⟨∇f ct (x ct ), x ct -u ct ⟩ (11)"
},
{
"formula_coordinates": [
11,
286.53,
273.93,
230.63,
10.67
],
"formula_id": "formula_31",
"formula_text": "x t = y τt (12"
},
{
"formula_coordinates": [
11,
517.15,
273.96,
4.85,
9.57
],
"formula_id": "formula_32",
"formula_text": ")"
},
{
"formula_coordinates": [
11,
106.94,
292.83,
104.29,
26.14
],
"formula_id": "formula_33",
"formula_text": "τ t = 1 + t-1 i=1 |F i |. Combining ("
},
{
"formula_coordinates": [
11,
104.18,
330.01,
417.82,
109.99
],
"formula_id": "formula_34",
"formula_text": "R(u 1 , • • • , u T ) ≤ T t=1 ∇f ct (x ct ), y τc t -u ct = T t=1 ∇f ct (x ct ), y t -u t + y τc t -y t + T t=1 ⟨∇f ct (x ct ), u t -u ct ⟩ ≤ T t=1 ⟨∇f ct (x ct ), y t -u t ⟩ + T t=1 G y τc t -y t 2 + T t=1 G∥u t -u ct ∥ 2 (13)"
},
{
"formula_coordinates": [
11,
114.25,
484.2,
407.75,
224.6
],
"formula_id": "formula_35",
"formula_text": "T t=1 ⟨∇f ct (x ct ), y t -u t ⟩ = T t=1 y t -y ′ t+1 , y t -u t η = T t=1 ∥y t -u t ∥ 2 2 -∥y ′ t+1 -u t ∥ 2 2 + ∥η∇f ct (x ct )∥ 2 2 2η ≤ T t=1 1 2η ∥y t -u t ∥ 2 2 -∥y t+1 -u t ∥ 2 2 + ηT G 2 2 = T t=1 1 2η ∥y t ∥ 2 2 -∥y t+1 ∥ 2 2 + T t=1 1 η ⟨y t+1 -y t , u t ⟩ + ηT G 2 2 ≤ 1 η ⟨y T +1 , u T ⟩ + T t=2 1 η ⟨u t-1 -u t , y t ⟩ + ηT G 2 2 ≤ 1 η ∥y T +1 ∥ 2 ∥u T ∥ 2 + T t=2 1 η ∥u t-1 -u t ∥ 2 ∥y t ∥ 2 + ηT G 2 2 ≤ D 2 + DP T η + ηT G 2 2 (14)"
},
{
"formula_coordinates": [
12,
276.75,
157.78,
58.51,
12.89
],
"formula_id": "formula_36",
"formula_text": "y 1 , • • • , y τc t ."
},
{
"formula_coordinates": [
12,
289.1,
249.66,
232.9,
10.63
],
"formula_id": "formula_37",
"formula_text": "τ ct ≤ t.(15)"
},
{
"formula_coordinates": [
12,
149.57,
293.96,
372.44,
77.42
],
"formula_id": "formula_38",
"formula_text": "T t=1 y τc t -y t 2 ≤ T t=1 t-1 k=τc t ∥y k -y k+1 ∥ 2 ≤ T t=1 t-1 k=τc t y k -y ′ k+1 2 ≤ T t=1 t-1 k=τc t ∥η∇f c k (x c k )∥ 2 ≤ ηG T t=1 (t -τ ct ) .(16)"
},
{
"formula_coordinates": [
12,
149.6,
402.55,
372.4,
72.08
],
"formula_id": "formula_39",
"formula_text": "T t=1 (t -τ ct ) = T t=1 (t -1) - T t=1 ct-1 i=1 |F i | = T t=1 (t -1) - T t=1 t-1 i=1 |F i | = T t=1 t -1 - t-1 i=1 |F i | = T t=1 (m t -1)(17)"
},
{
"formula_coordinates": [
12,
214.58,
519.6,
307.42,
33.58
],
"formula_id": "formula_40",
"formula_text": "T t=1 G y τc t -y t 2 ≤ ηG 2 T t=1 (m t -1) .(18)"
},
{
"formula_coordinates": [
12,
146.64,
585.2,
319.65,
33.58
],
"formula_id": "formula_41",
"formula_text": "T t=1 (f t (x t ) -f t (u t )) ≤ D 2 + DP T η + ηG 2 T t=1 m t + T t=1 G∥u t -u ct ∥ 2 ."
},
{
"formula_coordinates": [
12,
226.83,
675.22,
158.04,
33.58
],
"formula_id": "formula_42",
"formula_text": "0 ≤ P T = T t=2 ∥u t -u t-1 ∥ 2 ≤ T D which implies that η 1 = D G √ β ≤ η * ≤ D √ T + 1 G √ β ≤ η |H| ."
},
{
"formula_coordinates": [
13,
273.12,
159.18,
244.04,
10.77
],
"formula_id": "formula_43",
"formula_text": "η k ≤ η * ≤ 2η k (19"
},
{
"formula_coordinates": [
13,
517.15,
159.18,
4.85,
9.57
],
"formula_id": "formula_44",
"formula_text": ")"
},
{
"formula_coordinates": [
13,
144.56,
213.14,
377.44,
71.78
],
"formula_id": "formula_45",
"formula_text": "R(u 1 , • • • , u T ) ≤ T t=1 ⟨∇f t (x t ), x t -u t ⟩ = T t=1 ℓ t (x t ) - T t=1 ℓ t (x η k t ) + T t=1 ℓ t (x η k t ) - T t=1 ℓ t (u t ) .(20)"
},
{
"formula_coordinates": [
13,
90,
310.49,
166.75,
15.24
],
"formula_id": "formula_46",
"formula_text": "Lemma 2 Let m t = t -t-1 i=1 |F i |."
},
{
"formula_coordinates": [
13,
186.54,
344.2,
239.87,
33.58
],
"formula_id": "formula_47",
"formula_text": "T t=1 ℓ t (x t ) - T t=1 ℓ t (x η t ) ≤ 1 α ln 1 w η 1 + αG 2 D 2 T t=1 m t ."
},
{
"formula_coordinates": [
13,
356.17,
385.61,
56.16,
20.24
],
"formula_id": "formula_48",
"formula_text": "1 GD √ T t=1 mt"
},
{
"formula_coordinates": [
13,
154.75,
429.06,
302.94,
53.27
],
"formula_id": "formula_49",
"formula_text": "T t=1 ℓ t (x t ) - T t=1 ℓ t (x η k t ) ≤2GD T t=1 m t ln(k + 1) + GD T t=1 m t ≤2GD dT ln (k + 1) + GD dT"
},
{
"formula_coordinates": [
13,
166.35,
567.07,
279.46,
91.09
],
"formula_id": "formula_50",
"formula_text": "T t=1 ℓ t (x η k t ) - T t=1 ℓ t (u t ) ≤ D 2 + DP T η k + η k G 2 T t=1 m t + C ≤ 2(D 2 + DP T ) η * + η * G 2 T t=1 m t + C ≤3G D(D + P T ) dT + C"
},
{
"formula_coordinates": [
19,
90,
164.43,
427.15,
67.45
],
"formula_id": "formula_51",
"formula_text": "c 1 = 1, c 2 = 2, . . . , c T = T which directly implies that T t=1 ∥u t -u ct ∥ 2 = 0. (21"
},
{
"formula_coordinates": [
19,
517.15,
209.92,
4.85,
9.57
],
"formula_id": "formula_52",
"formula_text": ")"
},
{
"formula_coordinates": [
19,
239.65,
274.23,
282.35,
10.63
],
"formula_id": "formula_53",
"formula_text": "t ≤ c t + d ct -1 ≤ c t + d -1 (22)"
},
{
"formula_coordinates": [
19,
240.19,
323.53,
276.96,
32.84
],
"formula_id": "formula_54",
"formula_text": "c t + d ct -2 < t + d -1, which implies that c t ≤ t + d -d ct ≤ t + d -1. (23"
},
{
"formula_coordinates": [
19,
517.15,
345.74,
4.85,
9.57
],
"formula_id": "formula_55",
"formula_text": ")"
},
{
"formula_coordinates": [
19,
153.33,
389.41,
363.82,
35
],
"formula_id": "formula_56",
"formula_text": "∥u t -u ct ∥ 2 ≤ t-1 k=ct ∥u k+1 -u k ∥ 2 ≤ min{ct+d-2,T -1} k=ct ∥u k+1 -u k ∥ 2 . (24"
},
{
"formula_coordinates": [
19,
517.15,
401.64,
4.85,
9.57
],
"formula_id": "formula_57",
"formula_text": ")"
},
{
"formula_coordinates": [
19,
155.5,
456.79,
361.65,
35
],
"formula_id": "formula_58",
"formula_text": "∥u t -u ct ∥ 2 ≤ ct-1 k=t ∥u k+1 -u k ∥ 2 ≤ min{t+d-2,T -1} k=t ∥u k+1 -u k ∥ 2 . (25"
},
{
"formula_coordinates": [
19,
517.15,
468.53,
4.85,
9.57
],
"formula_id": "formula_59",
"formula_text": ")"
},
{
"formula_coordinates": [
19,
105.8,
523.18,
416.21,
183.21
],
"formula_id": "formula_60",
"formula_text": "T t=1 ∥u t -u ct ∥ 2 ≤ T t=1 min{ct+d-2,T -1} k=ct ∥u k+1 -u k ∥ 2 + T t=1 min{t+d-2,T -1} k=t ∥u k+1 -u k ∥ 2 =2 T t=1 min{t+d-2,T -1} k=t ∥u k+1 -u k ∥ 2 =2 d-1 k=1 T -1 t=k ∥u t+1 -u t ∥ 2 ≤2 d k=1 T -1 t=1 ∥u t+1 -u t ∥ 2 =2dP T (26)"
},
{
"formula_coordinates": [
20,
253.76,
131.87,
263.39,
33.58
],
"formula_id": "formula_61",
"formula_text": "T t=1 ∥u t -u ct ∥ 2 ≤ T D. (27"
},
{
"formula_coordinates": [
20,
517.15,
143.5,
4.85,
9.57
],
"formula_id": "formula_62",
"formula_text": ")"
},
{
"formula_coordinates": [
20,
158.67,
252.26,
294.66,
34.81
],
"formula_id": "formula_63",
"formula_text": "L η t = t i=1 k∈F i ℓ k (x η k ), Lη t = t i=1 ℓ i (x η i ), and Wt = η∈H w η 1 e -α Lη t ."
},
{
"formula_coordinates": [
20,
140.91,
322.53,
330.18,
15.19
],
"formula_id": "formula_64",
"formula_text": "c t = (L η t ) η∈H ∈ R |H| , ct = ( Lη t ) η∈H ∈ R |H| , and w t = (w η t ) η∈H ∈ R |H| ."
},
{
"formula_coordinates": [
20,
180.66,
372.34,
250.68,
35.99
],
"formula_id": "formula_65",
"formula_text": "w η t+1 = w η t e -α k∈F t ℓ k (x η k ) µ∈H w µ t e -α k∈F t ℓ k (x µ k ) = w η 1 e -αL η t µ∈H w µ 1 e -αL µ t ."
},
{
"formula_coordinates": [
20,
192.57,
444.92,
226.87,
25.76
],
"formula_id": "formula_66",
"formula_text": "w t+1 = argmin w∈∆ - 1 α ln(w 1 ) + c t , w + 1 α R(w)"
},
{
"formula_coordinates": [
20,
194.46,
521.8,
224.98,
25.76
],
"formula_id": "formula_67",
"formula_text": "wt+1 = argmin w∈∆ - 1 α ln(w 1 ) + ct , w + 1 α R(w)"
},
{
"formula_coordinates": [
20,
250.24,
586.11,
113.75,
33.27
],
"formula_id": "formula_68",
"formula_text": "wη t+1 = w η 1 e -α Lη t µ∈H w µ 1 e -α Lµ t ."
},
{
"formula_coordinates": [
21,
90,
92.08,
432,
51.63
],
"formula_id": "formula_69",
"formula_text": ") Let Π K (u, α) = argmin x∈K ⟨u, x⟩ + 1 α R(x). If R(x) is 1-strongly convex with respect to a norm ∥ • ∥, it holds that ∥Π K (u, α) -Π K (v, α)∥ ≤ α∥u -v∥ *"
},
{
"formula_coordinates": [
21,
90,
228.74,
348.1,
70.19
],
"formula_id": "formula_70",
"formula_text": "∥x t -x t ∥ 2 = η∈H ( wη t -w η t )x η t 2 ≤ η∈H | wη t -w η t | ∥x η t ∥ 2 ≤D∥ wt -w t ∥ 1 ≤ αD∥c t-1 -c t-1 ∥ ∞ . Let U t = [t] \\ ∪ i∈[t] F i ."
},
{
"formula_coordinates": [
21,
169.78,
346.43,
352.22,
66.14
],
"formula_id": "formula_71",
"formula_text": "∥x t -x t ∥ 2 ≤αD∥c t-1 -c t-1 ∥ ∞ ≤ αD max η∈H k∈U t-1 ℓ k (x η k ) ≤α t -1 - t-1 i=1 |F i | GD 2 = α (m t -1) GD 2 (29)"
},
{
"formula_coordinates": [
21,
155.32,
460.15,
366.68,
15.82
],
"formula_id": "formula_72",
"formula_text": "|ℓ k (x η k )| = |⟨∇f k (x k ), x η k -x k ⟩| ≤ ∥∇f k (x k )∥ 2 ∥x η k -x k ∥ 2 ≤ GD (30)"
},
{
"formula_coordinates": [
21,
160.12,
551.73,
357.03,
37.17
],
"formula_id": "formula_73",
"formula_text": " η∈H w η 1 e -α Lη T ≥ ln max η∈H w η 1 e -α Lη T = -α min η∈H Lη T + 1 α ln 1 w η 1 . (31"
},
{
"formula_coordinates": [
21,
517.15,
566.64,
4.85,
9.57
],
"formula_id": "formula_74",
"formula_text": ")"
},
{
"formula_coordinates": [
21,
131.2,
625.81,
338.36,
37.84
],
"formula_id": "formula_75",
"formula_text": "ln Wt Wt-1 = ln η∈H w η 1 e -α Lη t η∈H w η 1 e -α Lη t-1 = ln η∈H w η 1 e -α Lη t-1 e -αℓt(x η t ) η∈H w η 1 e -α Lη"
},
{
"formula_coordinates": [
22,
231.1,
192.51,
149.81,
26.38
],
"formula_id": "formula_77",
"formula_text": "ln E[e sX ] ≤ sE[X] + s 2 (b -a) 2 8 ."
},
{
"formula_coordinates": [
22,
190.81,
241.11,
331.19,
69.33
],
"formula_id": "formula_78",
"formula_text": " η∈H wη t e -αℓt(x η t ) ≤ -α η∈H wη t ℓ t (x η t ) + α 2 G 2 D 2 2 ≤ -αℓ t (x t ) + α 2 G 2 D 2 2 (34)"
},
{
"formula_coordinates": [
22,
220.19,
347.31,
171.61,
33.58
],
"formula_id": "formula_79",
"formula_text": "ln WT ≤ -α T t=1 ℓ t (x t ) + α 2 G 2 D 2 T 2 ."
},
{
"formula_coordinates": [
22,
179.02,
405.87,
254.89,
33.58
],
"formula_id": "formula_80",
"formula_text": "T t=1 ℓ t (x t ) -min η∈H T t=1 ℓ t (x η t ) + 1 α ln 1 w η 1 ≤ αG 2 D 2 T 2 ."
},
{
"formula_coordinates": [
22,
154.57,
464.43,
367.43,
224.6
],
"formula_id": "formula_81",
"formula_text": "T t=1 ℓ t (x t ) - T t=1 ℓ t (x η t ) + 1 α ln 1 w η 1 = T t=1 ℓ t (x t ) - T t=1 ℓ t (x t ) + T t=1 ℓ t (x t ) - T t=1 ℓ t (x η t ) + 1 α ln 1 w η 1 ≤ T t=1 ⟨∇f t (x t ), x t -xt ⟩ + αG 2 D 2 T 2 ≤ T t=1 ∥∇f t (x t )∥ 2 ∥x t -xt ∥ 2 + αG 2 D 2 T 2 ≤αG 2 D 2 T t=1 (m t -1) + αG 2 D 2 T 2 ≤αG 2 D 2 T t=1 m t(35)"
},
{
"formula_coordinates": [
23,
90,
248.8,
232.79,
17.43
],
"formula_id": "formula_82",
"formula_text": "Lemma 5 Suppose K = [-D/(2 √ n), D/(2 √ n)]"
},
{
"formula_coordinates": [
23,
454.6,
270.21,
78.29,
10.68
],
"formula_id": "formula_83",
"formula_text": "f 1 (x), • • • , f T (x)"
},
{
"formula_coordinates": [
23,
257.62,
283.75,
206.86,
43.97
],
"formula_id": "formula_84",
"formula_text": "1 ≤ d 1 , • • • , d T ≤ d such that R(T ) ≥ DGT 2 2 ⌈T /d⌉ ."
},
{
"formula_coordinates": [
23,
216.45,
386.22,
179.11,
10.63
],
"formula_id": "formula_85",
"formula_text": "T z = {(z -1)L + 1, • • • , min{zL, T }}."
},
{
"formula_coordinates": [
23,
187.61,
426.33,
225.28,
33.58
],
"formula_id": "formula_86",
"formula_text": "C(P ) = u 1 , • • • , u T ∈ K T t=2 ∥u t -u t-1 ∥ 2 ≤ P"
},
{
"formula_coordinates": [
23,
149.29,
487.14,
305.86,
13.71
],
"formula_id": "formula_87",
"formula_text": "C ′ (P ) = u 1 , • • • , u T ∈ K u (z-1)L+1 = • • • = u min{zL,T } , ∀z ∈ [Z]"
},
{
"formula_coordinates": [
23,
90,
564.84,
432,
143.96
],
"formula_id": "formula_88",
"formula_text": "f 1 (x), • • • , f T (x) satisfying Assumption 1 and a sequence of delays 1 ≤ d 1 , • • • , d T ≤ d such that T t=1 f t (x t ) - min u 1 ,••• ,u T ∈C(P ) T t=1 f t (u t ) ≥ T t=1 f t (x t ) - min u 1 ,••• ,u T ∈C ′ (P ) T t=1 f t (u t ) = Z z=1 t∈Tz f t (x t ) -min x∈K t∈Tz f t (x) ≥ Z z=1 DG|T z | 2 2 ⌈|T z |/d⌉ .(36)"
},
{
"formula_coordinates": [
24,
98.42,
112.63,
423.58,
69.49
],
"formula_id": "formula_89",
"formula_text": "Z z=1 DG|T z | 2 2 ⌈|T z |/d⌉ ≥ Z z=1 DG|T z | 2 2 ⌈L/d⌉ = DGT 2 2 ⌈L/d⌉ ≥ DGT 2 √ 2 , if d > L; G dD max{P, D}T 4 √ 2 , otherwise;(37)"
},
{
"formula_coordinates": [
24,
90,
220.57,
365.01,
27.95
],
"formula_id": "formula_90",
"formula_text": "⌈L/d⌉ ≤ 2L/d = 2 ⌈T D/ max{P, D}⌉ /d ≤ 4T D/(max{P, D}d) for d ≤ L."
},
{
"formula_coordinates": [
24,
218.19,
352.81,
175.61,
10.63
],
"formula_id": "formula_91",
"formula_text": "T z = {(z -1)d + 1, • • • , min{zd, T }}."
},
{
"formula_coordinates": [
24,
249.55,
395.28,
112.9,
10.63
],
"formula_id": "formula_92",
"formula_text": "d t = min{zd, T } -t + 1"
},
{
"formula_coordinates": [
24,
259.74,
516.67,
92.53,
24.48
],
"formula_id": "formula_93",
"formula_text": "h z (x) = G √ n ⟨w z , x⟩"
},
{
"formula_coordinates": [
24,
125.12,
597.5,
355.39,
111.29
],
"formula_id": "formula_94",
"formula_text": "E w 1 ,••• ,w Z [R(T )] =E w 1 ,••• ,w Z T t=1 f t (x t ) -min x∈K T t=1 f t (x) =E w 1 ,••• ,w Z Z z=1 t∈Tz G √ n ⟨w z , x t ⟩ -min x∈K Z z=1 t∈Tz G √ n ⟨w z , x⟩ =E w 1 ,••• ,w Z -min x∈K Z z=1 G|T z | √ n ⟨w z , x⟩"
},
{
"formula_coordinates": [
25,
134.24,
153.71,
387.76,
153.32
],
"formula_id": "formula_95",
"formula_text": "E w 1 ,••• ,w Z [R(T )] = -E w 1 ,••• ,w Z min x∈{-D/(2 √ n),D/(2 √ n)} n Z z=1 G|T z | √ n ⟨w z , x⟩ =E w 1 ,••• ,w Z n i=1 D 2 √ n Z z=1 w z,i G|T z | √ n = DG 2 E w 1 ,••• ,w Z Z z=1 w z,1 |T z | ≥ DG 2 √ 2 Z z=1 |T z | 2 ≥ DG 2 √ 2 ( Z z=1 |T z |) 2 Z = DGT 2 2 ⌈T /d⌉ (38)"
},
{
"formula_coordinates": [
25,
257.62,
357,
147.1,
43.58
],
"formula_id": "formula_96",
"formula_text": "1 , • • • , w Z such that R(T ) ≥ DGT 2 2 ⌈T /d⌉ ."
},
{
"formula_coordinates": [
25,
238.4,
460.37,
136.14,
33.71
],
"formula_id": "formula_97",
"formula_text": "T t=1 m t = T t=1 t - t-1 i=1 |F i | ."
},
{
"formula_coordinates": [
25,
230.78,
579.36,
141.8,
34.68
],
"formula_id": "formula_98",
"formula_text": "s v+1 -1 t=sv t + 1 -s v - t-1 i=sv |F sv i |"
},
{
"formula_coordinates": [
25,
121.55,
618.07,
160.59,
15
],
"formula_id": "formula_99",
"formula_text": "F sv i = {k ∈ [s v , i]|k + d k -1 = i}."
},
{
"formula_coordinates": [
25,
222.54,
672.95,
166.14,
34.29
],
"formula_id": "formula_100",
"formula_text": "t j=sv j + 1 -s v - j-1 i=sv |F sv i | ≤ 2 v"
},
{
"formula_coordinates": [
26,
95.92,
107.86,
263,
54.06
],
"formula_id": "formula_101",
"formula_text": "s v = 1 2: for t = 1, • • • , T do 3: if t j=sv j + 1 -s v -j-1 i=sv |F sv i | > 2 v then 4:"
},
{
"formula_coordinates": [
26,
95.92,
190.63,
350.07,
25.49
],
"formula_id": "formula_102",
"formula_text": "Receive {∇f k (x k )|k ∈ F sv t }, where F sv t = {k ∈ [s v , t]|k + d k -1 = t} 8:"
},
{
"formula_coordinates": [
26,
177.64,
218.19,
295.98,
16.05
],
"formula_id": "formula_103",
"formula_text": "y τ +1 = argmin x∈K ∥x -(y τ -η v ∇f k (x k ))∥ 2 2 , where η v = D G2 v/2"
},
{
"formula_coordinates": [
26,
91.32,
234.03,
103.17,
22.74
],
"formula_id": "formula_104",
"formula_text": "Set τ = τ + 1 11:"
},
{
"formula_coordinates": [
26,
185.43,
364.3,
241.14,
37.72
],
"formula_id": "formula_105",
"formula_text": "t j=s v+1 j + 1 -s v+1 - j-1 i=s v+1 |F s v+1 i | = 1 ≤ 2 v+1 ."
},
{
"formula_coordinates": [
26,
90,
475.63,
235.09,
14.63
],
"formula_id": "formula_106",
"formula_text": "{∇f k (x k )|k ∈ F sv t }, instead of {∇f k (x k )|k ∈ F t }."
},
{
"formula_coordinates": [
26,
208.73,
547.3,
193.77,
35.69
],
"formula_id": "formula_107",
"formula_text": "R(u 1 , • • • , u T ) ≤ 2G (2D + P T ) √ dT √ 2 -1 + C"
},
{
"formula_coordinates": [
26,
176,
667.47,
259.99,
34.29
],
"formula_id": "formula_108",
"formula_text": "t j=sv j + 1 -s v - j-1 i=sv |F sv i | ≤ t j=sv d j ≤ T j=1 d j = dT."
},
{
"formula_coordinates": [
27,
111.62,
183.42,
410.38,
74.3
],
"formula_id": "formula_110",
"formula_text": "s v+1 -1 t=sv f t (x t ) - s v+1 -1 t=sv f t (u t ) ≤ D 2 + D s v+1 -1 t=sv+1 ∥u t -u t-1 ∥ 2 η v + η v G 2 s v+1 -1 j=sv j + 1 -s v - j-1 i=sv |F sv i | + C v(40)"
},
{
"formula_coordinates": [
27,
143.86,
274.44,
378.14,
52.78
],
"formula_id": "formula_111",
"formula_text": "C v = 0, if Assumption 3 also holds; min (s v+1 -s v )GD, 2dG s v+1 -1 t=sv+1 ∥u t -u t-1 ∥ 2 , otherwise.(41)"
},
{
"formula_coordinates": [
27,
216.23,
350.18,
305.78,
34.68
],
"formula_id": "formula_112",
"formula_text": "s v+1 -1 j=sv j + 1 -s v - j-1 i=sv |F sv i | ≤ 2 v .(42)"
},
{
"formula_coordinates": [
27,
116.99,
411.58,
405.01,
92.41
],
"formula_id": "formula_113",
"formula_text": "s v+1 -1 t=sv f t (x t ) - s v+1 -1 t=sv f t (u t ) ≤ D 2 + D s v+1 -1 t=sv+1 ∥u t -u t-1 ∥ 2 η v + η v G 2 2 v + C v =G2 v/2 2D + s v+1 -1 t=sv+1 ∥u t -u t-1 ∥ 2 + C v ≤G2 v/2 (2D + P T ) + C v .(43)"
},
{
"formula_coordinates": [
27,
95.88,
530.45,
426.12,
73.51
],
"formula_id": "formula_114",
"formula_text": "R(u 1 , • • • , u T ) = V v=1 s v+1 -1 t=sv f t (x t ) - s v+1 -1 t=sv f t (u t ) ≤ V v=1 G2 v/2 (2D + P T ) + V v=1 C v =G (2D + P T ) √ 2(2 V /2 -1) √ 2 -1 + V v=1 C v ≤ 2G (2D + P T ) √ S √ 2 -1 + V v=1 C v .(44)"
},
{
"formula_coordinates": [
27,
116.45,
627.91,
379.1,
74.41
],
"formula_id": "formula_115",
"formula_text": "V v=1 min (s v+1 -s v )GD, 2dG s v+1 -1 t=sv+1 ∥u t -u t-1 ∥ 2 ≤ min V v=1 (s v+1 -s v )GD, V v=1 2dG s v+1 -1 t=sv+1 ∥u t -u t-1 ∥ 2 ≤ min {T GD, 2dGP T }"
},
{
"formula_coordinates": [
28,
95.92,
145.35,
330.23,
97.8
],
"formula_id": "formula_116",
"formula_text": "H = η i = D2 i-1 G i = 1, • • • , log 2 √ T + 1 + 1 3: Set w η i t = |H|+1 i(i+1)|H| 4: for t = 1, • • • , T do 5: if t j=sv j + 1 -s v -j-1 i=sv |F sv i | > 2 v then 6: Set v = v + 1, s v = t,"
},
{
"formula_coordinates": [
28,
90,
320.28,
432,
163.89
],
"formula_id": "formula_117",
"formula_text": "w η t+1 = w η t e -αv k∈F sv t ℓ k (x η k ) µ∈H w µ t e -αv k∈F sv t ℓ k (x µ k ) where ℓ k (x) = ⟨∇f k (x k ), x -x k ⟩ and α v = 1 GD2 v/2 12: Send {∇f k (x k )|k ∈ F sv t } to each expert E η 13: end for which implies that V v=1 C v ≤ C.(45)"
},
{
"formula_coordinates": [
28,
203.73,
591.65,
313.42,
34.72
],
"formula_id": "formula_118",
"formula_text": "α = 1 GD T t=1 m t and η i = 2 i-1 D G T t=1 m t (46"
},
{
"formula_coordinates": [
28,
517.15,
600.98,
4.85,
9.57
],
"formula_id": "formula_119",
"formula_text": ")"
},
{
"formula_coordinates": [
29,
95.92,
121.32,
263.47,
55.51
],
"formula_id": "formula_120",
"formula_text": "s v = 1 3: for t = 1, • • • , T do 4: if t j=sv j + 1 -s v -j-1 i=sv |F sv i | > 2 v then 5: Set y 1 = 0, τ = 1, v = v +"
},
{
"formula_coordinates": [
29,
95.26,
508.78,
418.77,
54.57
],
"formula_id": "formula_121",
"formula_text": "R(u 1 , • • • , u T ) ≤ 2 2 ln log 2 (D + P T ) /D + 2 + 1 GD + 3G D 2 + DP T √ dT √ 2 -1 + C"
},
{
"formula_coordinates": [
29,
211.62,
672.58,
188.76,
34.55
],
"formula_id": "formula_122",
"formula_text": "η v * = D G 2 D + s v+1 -1 t=sv+1 ∥u t -u t-1 ∥ 2 ."
},
{
"formula_coordinates": [
30,
226.43,
169.74,
159.13,
33.21
],
"formula_id": "formula_123",
"formula_text": "η 1 = D G ≤ η v * ≤ D √ T + 1 G ≤ η |H| ."
},
{
"formula_coordinates": [
30,
268.6,
237.65,
253.4,
14.19
],
"formula_id": "formula_124",
"formula_text": "η kv ≤ η v * ≤ 2η kv(47)"
},
{
"formula_coordinates": [
30,
187.27,
282.14,
237.45,
37.3
],
"formula_id": "formula_125",
"formula_text": "k v = log 2 D + s v+1 -1 t=sv+1 ∥u t -u t-1 ∥ 2 /D + 1"
},
{
"formula_coordinates": [
30,
101.06,
447.57,
420.95,
191.72
],
"formula_id": "formula_126",
"formula_text": "s v+1 -1 t=sv ℓ t (x η kv t ) -ℓ t (u t ) ≤ 2 v/2 D 2 + D s v+1 -1 t=sv+1 ∥u t -u t-1 ∥ 2 η kv + η kv G 2 2 v/2 s v+1 -1 j=sv j + 1 -s v - j-1 i=sv |F sv i | + C v ≤ 2 v/2 D 2 + D s v+1 -1 t=sv+1 ∥u t -u t-1 ∥ 2 η kv + η kv G 2 2 v/2 + C v ≤3G 2 v D 2 + D s v+1 -1 t=sv+1 ∥u t -u t-1 ∥ 2 + C v ≤3G 2 v (D 2 + DP T ) + C v (48)"
},
{
"formula_coordinates": [
31,
96.08,
117.29,
411.21,
34.68
],
"formula_id": "formula_127",
"formula_text": "s v+1 -1 t=sv ℓ t (x t ) - s v+1 -1 t=sv ℓ t (x η kv t ) ≤ 2 α v ln(k v + 1) + α v G 2 D 2 s v+1 -1 j=sv j + 1 -s v - j-1 i=sv |F sv i |"
},
{
"formula_coordinates": [
31,
273.68,
202.88,
58.31,
16.05
],
"formula_id": "formula_128",
"formula_text": "α v = 1 GD2 v/2"
},
{
"formula_coordinates": [
31,
92.07,
500.66,
427.08,
80.28
],
"formula_id": "formula_129",
"formula_text": "√ 2(2 V /2 -1) √ 2 -1 + V v=1 C v ≤ 2 2 ln log 2 (D + P T ) /D + 2 + 1 GD + 3G D 2 + DP T √ S √ 2 -1 + V v=1 C v"
}
] |
|
10.1145/2976749.2978318 | 2023-05-20 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b30",
"b29",
"b24",
"b30",
"b18",
"b19",
"b9",
"b13",
"b12",
"b14",
"b29",
"b0",
"b29",
"b11",
"b29",
"b23",
"b36",
"b28",
"b42",
"b33",
"b5",
"b10",
"b34",
"b51",
"b28",
"b1",
"b27",
"b31",
"b29",
"b24",
"b23",
"b23",
"b1",
"b47"
],
"table_ref": [],
"text": "Federated Learning (FL) (McMahan et al., 2017(McMahan et al., , 2018;;Kairouz et al., 2019) is designed to collaboratively train a global model on decentralized data across user clients while protecting data privacy. FL emerged as an effective privacy-preserving solution of training (language) models, as rich text data are generated by users, which may contain sensitive and personal information. After McMahan et al. (2017) proposed to train on-device recurrent neural network models, FL has been widely used in various natural language processing applications and products, including next-word prediction (Hard et al., 2018), keyword spotting (Hard et al., 2020), and out-of-vocabulary word discovery (Chen et al., 2019).\nTo further protect user privacy, Differential Privacy (DP) (Dwork et al., 2006;Dwork, 2011;Dwork and Roth, 2014;McMahan et al., 2018) is introduced to provide formal privacy guarantees of models trained by federated learning. DP for deep learning explicitly adds random noise with bounded sensitivity to a training process (e.g., DP-SGD (Abadi et al., 2016)), ensuring a quantifiable similarity in output model distributions when the training dataset changes. When combining DP with FL, a variant of DP-SGD called DP-FedAvg (McMahan et al., 2018)) is applied to guarantee user-level DP (Dwork, 2010). Current research primarily focuses on applying user-level DP to small on-device models with fewer than 10 million parameters (McMahan et al., 2018;Kairouz et al., 2021;Ramaswamy et al., 2020). The model size is limited due to challenges such as significant DP noise required to preserve privacy (Li et al., 2021) and the communication costs in cross-device FL.\nRecent advances in large language models (LLMs) (Thoppilan et al., 2022;Radford et al., 2019;Brown et al., 2020;Devlin et al., 2019;Raffel et al., 2020) have revolutionized natural language processing (NLP) and achieved unprecedented performance on various tasks such as text generation, machine translation, and sentiment analysis. However, their success comes at a cost of requiring massive amounts of computational resources, making them difficult to deploy on resource-constrained devices such as smartphones, tablets, or other edge devices. Additionally, there are concerns regarding the user privacy in various aspects such as memorizing personal information in training, and exposing private query in inference.\nRecent work explore incorporating public information to improve privacy-utility trade-off in applying DP for (large) LMs (Yu et al., 2022;Li et al., 2021). Public data (Amid et al., 2021) or other side information (Li et al., 2022) are also studied for (DP) FL. In non-DP FL settings, Nguyen et al. (2022) studies the effect of initializing from a pretrained model. However, it is an open question on how to leverage the power of pre-trained LLMs to facilitate private FL for on-device LMs.\nIn this work, we answer the question through systematic study aimed at enhancing private federated learning for on-device LMs with public pretrained LMs. Specifically, Our approach involves leveraging both public data and pre-trained LLMs to improve differentially private federated learning for on-device models by techniques of public pre-training and distillation. Additionally, we propose a novel distribution matching algorithm, which is backed by theoretical analysis, to sample public data closely resembling the private data distribution, which significantly increases sample efficiency in public training. Moreover, our extensive empirical results align with our theoretical predictions, further substantiating our approach. Our work complements existing research by utilizing LLMs to improve public training through knowledge distillation for private cross-device federated learning, and achieve a strong privacy-utility tradeoff with substantially improvements on sampling efficiency for public data. Our method points to a novel direction of efficiently enhancing private FL with public pretraining data and LLMs.\nWe summarize our contributions as follows: • We focus on improving private federated learning for language modeling tasks and explore ways to leverage public data and pre-trained LLMs for tokenizers, training protocols, and data (sub)sampling.\n• We conduct comprehensive studies and compare the use of Sentence Piece tokenizers from public LLM and unigram tokenizers from private corpus. We find that adopting public tokenizers from LLMs can not only prevent the potential privacy leakage from the private tokenizer vocabulary, but also lead to better learning utility with DP guarantees.\n• For training protocol, we propose to leverage public LLM to teach private on-device LMs by knowledge distillation. We demonstrate that distilling public LLM to pre-train on-device LM can lead to more than 7% accuracy improvement given tight privacy bound (ε = 1.77). Moreover, it can achieve high data efficiency of using only 1% of the public data compared to public pre-training without LLM, and attain better accuracy.\n• We further propose a novel distribution matching method that leverages both private on-device LMs and public LLMs to select public records close to private data distribution. We show that using 0.08% of carefully sampled public data to train on-device LM can lead to comparable performance as public pre-training ondevice LMs with the whole pre-training corpus, which reduces the public training time from more than one week to a few hours. Our method is grounded in theoretical analysis, which is corroborated by our extensive empirical results.\n2 Differentially Private Federated Learning for On-device LMs\nIn this section, we walk through the preliminaries of differentially private federated learning of language models following the cross-device federated learning literature (McMahan et al., 2018;Kairouz et al., 2019Kairouz et al., , 2021)). We also introduce the experimental setup used throughout this paper. In our experiments, we follow previous work (Kairouz et al., 2021;Amid et al., 2021;Wu et al., 2022) and sample 100 clients in each training round. Each client uses a batch size of 16 for local training. We set the training rounds T = 1600 in total."
},
{
"figure_ref": [],
"heading": "Cross-device Federated Learning",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "User-level Differential Privacy",
"publication_ref": [
"b13",
"b12",
"b14",
"b11",
"b23",
"b50",
"b20",
"b43",
"b45",
"b1",
"b23",
"b47",
"b42",
"b37",
"b23",
"b34"
],
"table_ref": [],
"text": "To further protect user privacy, Differential Privacy (DP) (Dwork et al., 2006;Dwork, 2011;Dwork and Roth, 2014) was introduced to provide a formal privacy guarantee for federated learning.\nDefinition 2.1 ((ε, δ)-Differential Privacy). A randomized algorithm M with domain N |X | is (ε, δ)differentially private if for all S ⊆ Range(M) and for any adjacent datasets D and D :\nPr[M(D) ∈ S] ≤ exp(ε) Pr[M(D ) ∈ S] + δ.\nDefinition 2.1 provides a formal definition of (ε, δ)-DP by bounding the change in output distribution caused by a small input difference (or, adjacent datasets) for a randomized algorithm. In the federated learning setting, it is preferable to bound the output distribution caused by different users in order to protect the privacy of each client's whole dataset. Specifically, adjacent datasets of D and D for user-level differential privacy (Dwork, 2010) are defined as: D can be obtained from D by adding or subtracting all the records of a single user/client, which determines the unit of privacy guarantees.\nIn our experiments, we use DP-FTRL (Kairouz et al., 2021) for privacy accounting and private federated training, which can achieve strong privacy guarantee in practical FL scenarios (Xu et al., 2023). We use δ = 10 -6 and consider two ε bounds: a tight privacy bound with ε = 1.77 by using a large noise multiplier m = 8.83, and a slightly loose privacy bound with ε = 18.71 and noise multiplier m = 1.13. We present more hyperparameter tuning details in Appendix C.\nOn-device LMs Due to the limited memory constraints of mobile devices, on-device LMs are relatively small (usually less than 10M parameters). In our work, we focus on two types of on-device autoregressive LMs: LSTM (Hochreiter and Schmidhuber, 1997) and transformers (Vaswani et al., 2017). Specifically, we follow previous work (Wang et al., 2021;Amid et al., 2021;Kairouz et al., 2021;Wu et al., 2022) and use one-layer LSTM and transformer. Both LSTM and transformer has a hidden size of 670 and embedding size of 96.\nPre-trained LLMs In addition to the on-device LMs trained on private datasets, this work also assumes that we have access to LLMs pre-trained on a large public corpus to aid private learning. Specifically, we use LaMDA (Thoppilan et al., 2022) We follow (Reddi et al., 2021;Kairouz et al., 2021) to construct a validation set of 10K samples, and a test set of 16.5M samples. Our evaluation metric is in-vocabulary next word (token) prediction accuracy, which is computed as the ratio of accurately predicted in-vocabulary words to the total number of words in the sequence (excluding OOV tokens).\nIn addition to StackOverflow as the (private) dataset, we use the realnews variant c4/realnewslike of C4 dataset (Raffel et al., 2020), as the public dataset. We analyzed the sources of the public C4 dataset and the Stackoverflow dataset for private training, and verified that there is no explicit overlap between public C4 dataset and the private StackOverflow dataset. More details can be found in Appendix B.1."
},
{
"figure_ref": [],
"heading": "Inspiration from LLMs",
"publication_ref": [],
"table_ref": [],
"text": "The success of publicly pre-trained LLMs motivate us to have retrospective views on further improving private on-device LMs. In this section, we explore inpiration from LLMs: the use of subword tokenizers and a large public corpus for pre-training. We apply them to on-device LMs, and observe that both techniques bring significant performance improvement for private FL."
},
{
"figure_ref": [
"fig_1"
],
"heading": "Using Public Tokenizer from LLMs",
"publication_ref": [
"b29",
"b23",
"b1",
"b32",
"b4",
"b26",
"b39",
"b38",
"b26",
"b8"
],
"table_ref": [],
"text": "Tokenizer is an important module of LMs, which transforms natural languages into a sequence of predefined symbol sets (vocabulary). Prior work in the literature of private FL of LMs (McMahan et al., 2018;Kairouz et al., 2021;Amid et al., 2021) use word-level unigram tokenizers potentially directly built from user data, which may need additional privacy budget (Ponomareva et al., 2022;Bagdasaryan et al., 2022).\nRecent LLMs adopt sub-word tokenizers (Kudo and Richardson, 2018;Sennrich et al., 2016;Schuster and Nakajima, 2012), which mitigate most outof-vocabulary (OOV) problems and yield state-ofthe-art performance across different downstream tasks. This motivates us to replace the prior wordlevel unigram tokenizers with public sub-word tokenizers. Specifically, we use SentencePiece tokenizer (Kudo and Richardson, 2018) from LaMDA.\nTo conduct comparison between unigram tokenizers and subword tokenizers for next word (token) prediction task, we convert the next word prediction accuracy into next token prediction accuracy. This conversion is achieved through splitting each word using the SentencePiece tokenizer. We consider all tokens within a word as accurate if the predicted word is correct. We compare standard SentencePiece models (vocabulary size = 32K) with unigram tokenizers that selects the top-k frequent words from user data with k = 10K or 32K as vocabulary.\nWe present the private FL accuracy on the Stack-Overflow dataset in Figure 1. For the unigram tokenizer, using a larger vocabulary size in the DP setting can result in a slight performance drop, which can be different from the observation in non-DP settings (Charles et al., 2022;Xu et al., 2022a). It is possible that the parameter increase of the embedding layer enlarges the effect of DP noise and hurts the final accuracy. However, for next token prediction accuracy, although the public Sentence-Piece tokenizer from LaMDA also consists of 32K tokens, it can significantly improve the private FL accuracy upon the unigram tokenizers, especially with smaller DP noise and ε = 18.71. We also observe that SentencePiece tokenizer finds no OOV tokens in the StackOverflow dataset, thus yielding the same high prediction accuracy with or without the OOV token. Therefore, we use SentencePiece tokenizer in the rest of this paper."
},
{
"figure_ref": [],
"heading": "Publicly pre-training for On-device LMs",
"publication_ref": [
"b27",
"b51"
],
"table_ref": [],
"text": "In addition to the use of subword tokenizers, LLMs benefit from pre-training on a large public corpus (Li et al., 2022;Yu et al., 2022). In this section, we explore pre-training on-device LMs on public corpus to improve private federated learning. "
},
{
"figure_ref": [],
"heading": "Pre-training Details",
"publication_ref": [],
"table_ref": [],
"text": "We use the standard autoregressive language modeling loss L LM to pre-train on-device LMs on the public C4 dataset, which takes around 1, 400K steps (over a week of single GPU time) to process the entire dataset with the batch size of 512. We then use the publicly pre-trained checkpoint as the start point for private federated learning. We leave more details in Appendix §B.2."
},
{
"figure_ref": [],
"heading": "Results",
"publication_ref": [],
"table_ref": [],
"text": "We present the next token prediction accuracy on the private StackOverflow dev set in Table 1. We observe that the accuracy on the private dataset significantly improves after pre-training for different different privacy budgets, shedding light on an effective way to boost private FL performance. We also observe that after pre-training, it gives reasonable zero-shot accuracy on the private dataset even without private training (round=0)."
},
{
"figure_ref": [],
"heading": "Distillation from Public LLM",
"publication_ref": [],
"table_ref": [],
"text": "We have shown that the accuracy of private federated learning can be significantly improved with public pre-training. On one hand, the cost of public pre-training for on-device LMs is still expensive on a large public corpus (around a week of GPU time). On the other hand, existing LLMs are well pretrained and demonstrate promising performance across a variety of downstream tasks. This motivates us to explore on whether we can leverage existing LLMs to improve the sample efficiency of pre-training on-device LMs. In this section, we answer the question above with systematic studies and show that we can improve the sample efficiency by using only 1% of pre-training data and distillation from LLMs, achieving similar or even better performance than using 100% of pretrianing data without distillation."
},
{
"figure_ref": [],
"heading": "Distillation Design",
"publication_ref": [
"b40",
"b22"
],
"table_ref": [],
"text": "Inspired by the literature of model compression (Sun et al., 2020;Jiao et al., 2019) "
},
{
"figure_ref": [
"fig_2"
],
"heading": "Experimental Results",
"publication_ref": [],
"table_ref": [
"tab_4"
],
"text": "After public pre-training with knowledge distillation, We use the checkpoints at different pre-training steps as the start point for private federated learning. Our main results can be found in Table 2. We show that by using 1% C4 dataset for pre-training with knowlegde distillation, we can significantly improve the sample efficiency without hurting but even improving the private FL accuracy for both LSTM and transformers, when compared with public pre-training on the whole C4 dataset.\nThe sample efficiency improvement thus reduces the pre-training cost from one week to around one day, shedding light on a promising direction to both improve the efficiency and effectiveness of private federated learning.\nAbaltion studies on top-k logits We take the top-k logits of the LLM to construct our distillation datasets and pre-train the on-device LMs. Here, we conduct an ablation study by pre-training different on-device LMs with different k and evaluate how top-k logits in distillation can impact the accuracy of private FL. We present our empirical results in Figure 2c and Appendix Figure 4. We observe that pre-training with a larger k is more helpful to achieve better downstream accuracy on private data.\nTo have a reasonable trade-off between dataset size and pre-training performance, we use top-k = 10 in all the following experiments.\nAblation studies on distillation steps To understand whether distillation for more epochs can help with private FL, we conduct a set of ablation studies on distillation steps given different privacy budgets as shown in Figure 2b and 2a. Specifically, we use the checkpoints at different distillation steps to initialize on-device LSTM and report the next word prediction accuracy after private FL at round 1600. We observe a consistent performance improvement when the distillation covers less than 5% of the C4 dataset. But when we pre-train the LM for more epochs, the improvement becomes marginal. This suggests that knowledge distillation from LLMs can help on-device LMs converge quickly within a few iterations."
},
{
"figure_ref": [],
"heading": "Distribution Matching",
"publication_ref": [],
"table_ref": [],
"text": "In the previous section, we achieve compelling performance by employing LLM distillation using only 1% of the randomly sampled pre-training corpus. Now we further investigate the possibility of improving sample efficiency by selectively identifying public samples that align with the distribution of private samples. To this end, we propose a novel distribution matching method to sample public records for pre-training. we provide a novel theoretical analysis jointly considering public-private distribution shift and DP mechanism. We demonstrate that by carefully selected 0.08% of public samples, we can pre-train on-device LMs that perform as well as using 1% of public samples with distillation. This approach significantly improves sample efficiency, providing an additional knob of using public pre-training for private on-device models."
},
{
"figure_ref": [
"fig_0"
],
"heading": "Algorithm",
"publication_ref": [
"b41"
],
"table_ref": [],
"text": "We hypothesize two principles to sample public records to match the private distribution: (i) the probability of the public sample x on the private data distribution p priv (x) is high, which can be approximated by the prediction of the on-device LMs trained on the private dataset; (ii) the probability of a public sample x on the public data distribution p pub (x) is also high, as we expect those samples are easy-to-learn (Swayamdipta et al., 2020) and of high data quality in the public corpus. The probability p pub (x) can be approximated by the public pre-trained LLMs.\nTo verify our hypothesis, we visualize the perplexity (PPL) distribution of public samples and private samples evaluated by both a privately finetuned on-device LM and a public pre-trained LLM in Figure 3. To have an \"oracle\" on-device LM that well captures the private data distribution, we fine-tune it on the private data without DP noise to overfit the private data distribution. We randomly sample 10k records from the public dataset and private dataset, respectively. We observe that the private dataset mostly concentrates on the regime with low PPL evaluated by the public and private LMs, whereas the public dataset is more diverse and distributed across a broader range of PPL values. The distribution visualization confirms our hypothesis to select public samples from the lower left corner, which correspond to samples with high probabilities p pub (x) and p priv (x) on public and private data distribution (i.e., low perplexity evaluated by public and pirvate LMs).\nIn practice, we do not have an \"oracle\" on-device LM trained on private data for distribution match. Instead, we propose to fine-tune an on-device LM with DP for certain rounds T < T before consuming all the privacy budgets, and then use the checkpoint at round T with DP guarantee to ap-proximate p priv (x) and perform distribution matching to sample public records. This post-processing based on a DP checkpoint will not incur any additional privacy cost. Thereafter, we can use the sampled public records to further train the private checkpoint at round T , as a way for efficient public (pre-)training. Following the strategy in §4, we also employ the distillation loss to better train the on-device LM with carefully sampled public records to further enhance the sample efficiency. Lastly, we use the remaining privacy budgets to fine-tune the on-device LM until reaching round T , and evaluate its next token prediction accuracy at the dev and test sets. We term the paradigm of two-stage private learning combined with public training as \"public mid-training\". This approach differs from \"public pre-training\", which involves public pre-training prior to private federated learning. We present the distribution matching protocol in Algorithm 1."
},
{
"figure_ref": [],
"heading": "Theoretical Analysis",
"publication_ref": [
"b21"
],
"table_ref": [],
"text": "In this section, we provide the theoretical analysis of our distribution matching protocol to present the intuition behind our selection hypothesis. In essence, the goal of our distribution matching algorithm is to have a good estimator for the private distribution. However, characterizing the distribution shift in the context of differential privacy is a challenging problem, in that the private models are trained with DP noise, which can yield an inaccurate estimation of private data distribution, and thus add the complexity to our analysis. Problem Setup Define the text data domain as X . Denote pub : X → R as the log-density function of the public data distribution (i.e., pub (x) = log p pub (x) where p pub (x) is the public data density estimated by public LLMs), and priv as the accurate log-density function of the private data distribution (i.e., priv (x) = log p true priv (x) where p true priv (x) is the true private data density). However, due to limited private data sampled from the true private data distribution and DP noise injected in the private FL, we can only obtain an inaccurate estimation ˆ priv = log p priv (x) of the true private log-density priv , where p priv (x) is the private data density estimated by private on-device LMs. Note that we use the hat notation ˆ priv to denote that it is an estimation of the true private log-density priv .\nWe can view the estimation ˆ priv is a random variable where the randomness comes from: (i) that the private dataset we have is sampled from the private data distribution; and (ii) the randomness in the algorithm of obtaining ˆ priv based on the private dataset, e.g., differential privacy. Following previous work (Jiang et al., 2023), we make a standard assumption. We assume the estimated private data log-density function is an unbiased estimator, i.e., E[ ˆ priv ] = priv .\nSince pub may not be ideal because of publicprivate domain shift, and ˆ priv may mot be ideal because of its DP noise, pub and ˆ priv are neither good estimators for priv . Can we leverage both of the information and form a function ĥ : X → R that combines pub and ˆ priv such that ĥ is a good estimator for priv ? In the following analysis, we choose ĥ = 1 2 pub + 1 2 ˆ priv and analyze when and why it can be a better estimator to the true private log-density priv than pub and ˆ priv .\nWe need some mathematical tools to define what does it mean to be \"better\". Concretely, we need a metric to measure the distance between functions. This can be done by having an inner product •, • in the function space of H = {f : X → R}, and hence the norm in the function space H is f = f, f for ∀f ∈ H. Our analysis holds with any choice of the inner product as long as it does not make the log-densities norm infinite. We discuss a concrete choice of the inner product and its relation to the KL divergence in Appendix D.\nWith the norm as a \"ruler\", we are able to define the following key quantities that formally characterize the setting. 1. Public-Private Domain Distance Let d pub, priv = pubpriv denote the distance between the public data log-density pub and the true private log-density priv . 2. Private Domain Randomness Let σ 2 priv = E[ ˆ privpriv 2 ] denote the randomness of the estimated private log-density, i.e., the quality of the estimated private log-density ˆ priv The above definitions are important because, as we show next, that the quality of a private log-density estimator would depend on the public-private domain shift and the private domain randomness.\nTheorem 5.1. Let ( f ) = E[ f -priv 2 ] charac-\nterise how good f is as an estimator of the true private data log-density priv for any random function f ∈ H. Consider the following three quantities:\n1. ( pub ) that characterizes the error if we use the public log-density function pub to approximate the priv 2. ( ˆ priv ) that characterizes the error if we use the noisy private log-density function ˆ priv to approximate the priv 3. ( ĥ) that characterizes the error if we use ĥ = 1 2 pub + 1 2 ˆ priv to approximate the priv .\nThen,\n( pub ) = d 2 pub, priv(1)\n( ˆ priv ) = σ 2 priv (2) ( ĥ) = 1 4 d 2 pub, priv + 1 4 σ 2 priv (3)\nInterpretation Theorem 5.1 implies that: • ( ĥ) ≤ 1 2 max{ ( pub ), ( ˆ priv )}.\n• ( ĥ) ≤ min{ ( pub ), ( ˆ priv )} if 1 3 ≤ d 2 pub, priv σ 2 priv ≤ 3.\nCombining the above, we have the following conclusion: recall ĥ = 1 2 pub + 1 2 ˆ priv = 1 2 log(p pub (x)p priv (x)). We can expect that ĥ is better than either pub or ˆ priv for any settings. Moreover, we can expect ĥ to be better than both pub and ˆ priv if (i) there is a domain shift between the public-private domain; and (ii) our estimated private log-density ˆ priv is noisy in an extent comparable to the domain shift. We leave the full proof and additional discussion in Appendix D."
},
{
"figure_ref": [],
"heading": "Experimental Results",
"publication_ref": [],
"table_ref": [
"tab_4",
"tab_6",
"tab_8"
],
"text": "Experimental Setup We set T = T /2 = 800 rounds for the first-stage private federated learning.\nWe use q = 0.08% of the whole pre-training corpus for public training, which reduces the public training time from more than 1 weeks to a few hours with a single GPU. For the public mid-training setting, we also evaluate how LLM distillation and distribution matching can impact the private FL accuracy, respectively. We run all the experimental settings for three times and report the average and We present the results of on-device LSTM and transformers in Table 2. In the pre-training setting (T = 0), we show that we cannot further improve the sample efficiency from 1% to 0.08% with LLM distillation improves the sample efficiency, as the final accuracy after private FL significantly decreases. In comparison, in the mid-training setting (T = T /2), using LLM distillation on the 0.08% of randomly sampled pre-training corpus already gives better performance than pre-training. Moreover, with distribution matching to carefully sample public data, we further improve the private learning accuracy, attaining comparable performance to the setting using the whole public corpus for pre-training.\nAblation studies on p pub (x) Our distribution matching algorithm leverages both on-device LM and LLM to sample data close to the private distribution. To understand how the use of LLM (p pub (x)) impact the sampling quality, we conduct an ablation study to sample a subset of D based on top log p priv (x) values alone instead of log p priv (x)+log p pub (x). We use the p priv -sampled D for public mid-training and report the test accuracy of three runs for both on-device LSTM and transformers given different privacy budgets in Table 3. The experimental findings corroborate our theoretical analysis. Specifically, when on-device language models (LMs) are trained with high noise levels (ε = 1.77), we find that a combined utilization of both on-device LMs and LLMs consistently yields superior performance. This is because the estimated private log-density ˆ priv is noisy to a degree comparable to the domain shift, making ĥ a more reliable estimator than ˆ priv . Conversely, when ondevice LMs are trained with low noise (ε = 18.71), the performance difference between models with and without p pub is negligible. This indicates that the noise introduced by differentially private (DP) training is not as significant as the distribution shift, allowing ˆ priv to serve as a good estimator.\nAblation studies on T T separates two-stage private federated learning and determines the timing for distribution matching and public training. In this ablation study, we evaluate the dev set accuracy of on-device LSTM given different T and privacy budgets, as shown in Table 4 and Appendix Table 5. From the table, we can see that the ondevice LSTM achieves the best private FL accuracy given T = T /2 = 800. We think the reasons are as follows: when T = 0, we cannot perform distribution matching as the on-device LM is not trained on the private dataset yet, and thus we can only use the randomly sampled data for pre-training; when T = 400, the on-device LM could not be well trained on the private data distribution, thus yielding worse distribution matching quality; when T = 1200 and T = 1600, the private on-device LM is biased towards the public data distribution due to public training, thus giving worse private FL accuracy. As a result, we use T = 800 in our main experiments, as it balances the private federated training and public training to have satisfactory distribution matching capabilities without biasing too much towards the public data distribution."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "In this work, we propose to improve private federated learning by using LLMs in public training. Inspired by the success of LLMs, we adapt on-device LMs with public subword tokenizers and pre-train on-device LMs on a large public corpus with distillation before private cross-device federated learning, where we observe significant performance improvement. We further leverage LLMs to aid public training of on-device LMs via distribution matching to sample public data close to private data distribution, which further improves the effectiveness and efficiency of public training, demonstrating strong private learning accuracy while minimizing the need for large amounts of public training data. Our work sheds light on a promising direction to improve private federated learning with public LLMs."
},
{
"figure_ref": [],
"heading": "A Additional Related Work",
"publication_ref": [
"b29",
"b23",
"b11",
"b18",
"b36",
"b23",
"b0",
"b25",
"b51",
"b6",
"b28",
"b16",
"b22",
"b40",
"b46",
"b15",
"b17"
],
"table_ref": [],
"text": "Private Federated Learning in On-device NLP Federated learning is designed to collaboratively training NLP models without sharing sensitive user data to protect user privacy. Given relatively small model sizes, state-of-the-art differentially private (DP) learning algorithms (McMahan et al., 2018;Kairouz et al., 2021) have enabled on-device LMs to achieve strong downstream task utility with reasonable userlevel differentially privacy guarantee (Dwork, 2010). The success of private FL has also led to real-world applications such as GBoard, which uses on-device LMs for next word prediction (Hard et al., 2018;Ramaswamy et al., 2020). Recent advances in DP optimization (Kairouz et al., 2021) further improves upon the state-of-the-art DP-SGD algorithm (Abadi et al., 2016), providing a practical tool to analyze privacy bound for federated learning.\nPrivacy-preserving Large NLP Models Scaling up LMs with more data and parameters has significantly improved performance and achieved great success in a variety of NLP tasks. Moreover, recent studies show that LLM has great potential in private learning. For example, Kerrigan et al. (2020) show that public pre-training is helpful for downstream DP fine-tuning. Follow-up studies argue that large pre-trained LMs can be strong differentially private learners with parameter-efficient fine-tuning (Yu et al., 2022;Bu et al., 2022) or full model fine-tuning (Li et al., 2021), narrowing the gap between non-private training and private training. Ganesh et al. (2023) also provide theoretical groundings on the necessity of involving public training into private learning. Motivated by the recent success of LLMs, our work performs comprehensive studies on how to use public data and existing LLMs to help private training of cross-device FL models.\nModel Compression for Pre-trained LMs One promising approach to address the resource limitations of LLMs is to compress them into smaller models through various techniques such as knowledge distillation (Jiao et al., 2019;Sun et al., 2020;Wang et al., 2020), or pruning (Elbayad et al., 2020;Gordon et al., 2020). While these techniques have demonstrated success in reducing the size of pre-trained LMs, most resulting models are still too large (with over 10 million parameters) to be effectively deployed on resource-constrained devices. In our work, we also explore the use of knowledge distillation in public training, but with a primary focus on leveraging LLMs to improve sample efficiency in pre-training on-device LMs. We aim to improve the private FL performance of on-device LMs while minimizing the need for large amounts of training data. We recognize that private federated learning can further benefit from advanced model compression techniques, and we leave this as a promising and orthogonal future direction for research in this area."
},
{
"figure_ref": [],
"heading": "B Experimental Setup Details B.1 Verification of Non-overlap between C4 and StackOverflow Datasets",
"publication_ref": [],
"table_ref": [
"tab_6",
"tab_8"
],
"text": "In this section, we detail the method used to verify that there is no explicit overlap between the public C4 dataset and the private StackOverflow dataset utilized in our study.\nWe explored C4 which has multiple variants1 : c4/en, c4/realnewslike, and c4/webtextlike.\nTo verify this hypothesis, we conducted a rigorous comparison of these two datasets and its variants. Specifically, we compared the unique identifiers (e.g., URL for webpages in the C4 dataset, and post ID for StackOverflow posts) between the two datasets.\nNo matching identifiers were found between the c4/realnewslike and the StackOverflow dataset. Thus we use the c4/realnewslike variant as our public pretraining corpus throughout the experiment.\nThrough this comprehensive comparison, we have confirmed that there is no explicit overlap between the public C4 dataset and the private StackOverflow dataset. This conclusion is critical to our study as it ensures that the integrity and privacy-preserving conditions of our experiment are maintained. Ablation studies on the timing T for mid-training T separates two-stage private federated learning and determines the timing for distribution matching and public training. In this ablation study, we evaluate the dev set accuracy of on-device LSTM given different T and privacy budgets, as shown in Table 4 and Appendix Table 5. From the table, we can see that the on-device LSTM achieves the best private FL accuracy given T = T /2 = 800. We think the reasons are as follows: when T = 0, we cannot perform distribution matching as the on-device LM is not trained on the private dataset yet, and thus we can only use the randomly sampled data for pre-training; when T = 400, the on-device LM could not be well trained on the private data distribution, thus yielding worse distribution matching quality; when T = 1200 and T = 1600, the private on-device LM is biased towards the public data distribution due to public training, thus giving worse private FL accuracy. As a result, we use T = 800 in our main experiments, as it balances the private federated training and public training to have satisfactory distribution matching capabilities without biasing too much towards the public data distribution. "
},
{
"figure_ref": [],
"heading": "D Detailed Theoretical Results",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "D.1 Discussion on the distance metrics of log-density functions",
"publication_ref": [
"b52",
"b7",
"b35"
],
"table_ref": [],
"text": "We need to define a meaningful distance metric in order to define the closeness of two log-density functions. To do this, we can choose any inner product •, • in the function space of H = {f : X → R}. Note that the log-density functions pub , priv , ˆ priv ∈ H. Accordingly, the norm in the function space H is denoted as • and by definition ∀f ∈ H : f = f, f . We note that our analysis works for any choice of the inner product as long as they don't make the log-densities norm infinite. For a concrete example, we discuss a generalization of the L 2 inner product, i.e., the L π inner product where π is a distribution on X . Formally, for this example of H = L π we define f, g π = E x∼π [f (x)g(x)] and\nf π = E x∼π [f (x) 2 ].\nThe L π is a rather general definition that is common in the literature of Bayesian coresets (Zhang et al., 2021;Campbell and Broderick, 2019) and kernel machine (Rahimi and Recht, 2007). For example, it recovers L 2 if π is chosen to be the uniform distribution on X .\nMoreover, if we choose π = p priv as the private data density, we can show that for any probability density function p, the distance between log p and log p priv measured by L p priv norm upper bounds the KL divergence between p priv and p: (5)\nIn general, the distribution π characterize where in X we want to evaluate a function.\nAbove we discuss a concrete choice of the inner product and the accordingly the norm to measure the distance between log-density functions. Since our analysis will work with any choice of inner product, we return to using the notation of •, • and • to remain generality in our main result."
},
{
"figure_ref": [],
"heading": "D.2 Proof",
"publication_ref": [],
"table_ref": [],
"text": "Theorem D.1 (Theorem 5.1 Restated). Let ( f ) = E[ fpriv 2 ] characterise how good f is as an estimator of the true private data log-density priv for any random function f ∈ H. Consider the following three quantities: 1. ( pub ) that characterizes the error if we use the public log-density function pub to approximate the priv 2. ( ˆ priv ) that characterizes the error if we use the noisy private log-density function ˆ priv to approximate the priv 3. ( ĥ) that characterizes the error if we use ĥ = 1 2 pub + 1 2 ˆ priv to approximate the priv .\nThen,\n( pub ) = d 2 pub, priv(6)\n( ˆ priv ) = σ 2 priv (7)\n( ĥ) = 1 4 d 2 pub, priv + 1 4 σ 2 priv (8)\nProof. We prove a general result which gives the theorem as special cases. For β ∈ [0, 1], define\nfβ = β pub + (1 -β) ˆ priv .(9)\nAccording to the definition of ( fβ ) = E[ fβpriv 2 ], we have Therefore, we can see that the theorem stands as we substitute f1 = pub , f 1 2 = ĥ, and f0 = ˆ priv .\n( fβ ) = E[ fβ -priv 2 ] = E[ β pub + (1 -β) ˆ priv -priv 2 ] (10) = E[ β( pub -priv ) + (1 -β)( ˆ priv -priv ) 2 ] (11) = β 2 pub -priv 2 + (1 -β) 2 E ˆ priv -priv 2 + 2β(1 -β)E pub -priv , ˆ priv -priv("
},
{
"figure_ref": [],
"heading": "D.3 Extended Analysis",
"publication_ref": [],
"table_ref": [],
"text": "Note that in the previous subsection the fβ is a weighted combination of pub and ˆ priv . I.e. fβ = (1 -β) pub + β ˆ priv where β ∈ [0, 1]. Therefore, one can show that with the optimal weight β , it is guaranteed that ( fβ ) ≤ min{ ( pub ), ( ˆ priv )}. This framework of analysis is general (as it stands with any meaningful inner product and its norm), and it may inspire even better ways to design estimators mitigating the domain shift and private model noise."
},
{
"figure_ref": [],
"heading": "Acknowledgement",
"publication_ref": [],
"table_ref": [],
"text": "The authors thank the early feedback from Yanxiang Zhang."
},
{
"figure_ref": [],
"heading": "B.2 Pretraining Details",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we outline the detailed procedures followed during the pretraining phase of our experiments. The pretraining phase consisted of the following steps:\n1. Data Preparation: We tokenized both the C4 and StackOverflow datasets using the SentencePiece tokenizer, as described in the main text. The vocabulary size was set to 32K for both datasets.\n2. Model Architecture: We utilized LSTM and transformer-based LMs with one hidden layer, 670 hidden units, embedding size euqal to 96."
},
{
"figure_ref": [],
"heading": "3.",
"publication_ref": [],
"table_ref": [],
"text": "Training Procedure: We trained the model using a standard autoregressive LM loss for next token prediction."
},
{
"figure_ref": [],
"heading": "Training Hyperparameters:",
"publication_ref": [],
"table_ref": [],
"text": "We employed the Adam optimizer with a learning rate of 1e-3, a batch size of 512, and a maximum sequence length of 20 tokens. We also used gradient clipping to prevent exploding gradients. The model was pretrained for 1400K steps on the C4 dataset to cover the whole C4 pretraining corpus.\nAfter pretraining, the model was then fine-tuned on the downstream task using federated learning with differential privacy. Further details regarding the fine-tuning process can be found in the relevant sections of the main text. We show that the pretraining procedure can significantly improve the model's robust performance in the downstream task performance."
},
{
"figure_ref": [],
"heading": "B.3 Distillation Details",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we delineate the specifics of our distillation process during the pretraining phase of our on-device LM. The pretraining procedure with distillation is mostly the same as details outlined in B.2 with slight hyper-parameter differences.\nWe set the temparature t = 1 and top-k = 10 to extract the logits z T from teacher LLM. We use grid search to tune the best hyper-parameter β ∈ {1e -1, 1e -2, 1e -3} and follow the same pre-training schedules as §3.2 but with a smaller batch size of 128 due to memory constraints."
},
{
"figure_ref": [],
"heading": "C Additional Experimental Results",
"publication_ref": [
"b2",
"b50"
],
"table_ref": [],
"text": "Hyper-parameter Tuning for Federated Learning Federated learning involves numerous hyperparameters, which is crucial for our experiment. Our hyper-parameter tuning strategy follows Xu et al. (2022b).\nThroughout our experiments, we fix the number of total rounds T = 1600. In each round, we select 100 clients from the shuffled pool for DP-FTRL, ensuring that the clients are disjoint across rounds. Within each client, we fix the number of local epochs to one and set the batch size to 16. We also impose a constraint on the maximum number of samples on each client, limiting it to 256.\nWe tune the server learning rate, client learning rate and clip norm for a certain given a noise multiplier. Specifically, we use grid search and tune the server learning rate from {0.05, 0.1, 0.2, 0.5, 1, 2}, the client learning rate from {0.01, 0.02, 0.05, 0.1, 0.2, 0.5}. We use the adaptive clipping technique in (Andrew et al., 2021;Xu et al., 2023) to help determine the clip norm, which in most of our experiments falls into {0.1, 0.3, 0.4, 1}.\nAbaltion studies on top-k logits We take the top-k logits of the LLM to construct our distillation datasets and pre-train the on-device LMs. Here, we conduct an ablation study by pre-training different on-device LMs with different k and evaluate how top-k logits in distillation can impact the accuracy of private FL. We present our empirical results in Figure 2c and Appendix Figure 4. We observe that pre-training with a larger k is more helpful to achieve better downstream accuracy on private data. To have a reasonable trade-off between dataset size and pre-training performance, we use top-k = 10 in all the following experiments."
}
] | We study (differentially) private federated learning (FL) of language models. The language models in cross-device FL are relatively small, which can be trained with meaningful formal user-level differential privacy (DP) guarantees when massive parallelism in training is enabled by the participation of a moderate size of users. Recently, public data has been used to improve privacy-utility trade-offs for both large and small language models. In this work, we provide a systematic study of using large-scale public data and LLMs to help differentially private training of on-device FL models, and further improve the privacy-utility tradeoff by techniques of distillation. Moreover, we propose a novel distribution matching algorithm with theoretical grounding to sample public data close to private data distribution, which significantly improves the sample efficiency of (pre-)training on public data. The proposed method is efficient and effective for training private model by taking advantage of public data, especially for customized ondevice architectures that do not have ready-touse pre-trained models. | Can Public Large Language Models Help Private Cross-device Federated Learning? | [
{
"figure_caption": "Figure 3 :3Figure 3: Visualization of perplexity (PPL) distribution of the private and public datasets evaluated by the private ondevice LM and the public LLM. The private dataset exhibits a concentration of low PPL values, whereas the public corpus is dispersed across a broader range of PPL values, with a higher average PPL.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Algorithm 11Leveraging LLMs for distribution matching and public training in private federated learning. Input: Public pre-training corpus D, private corpus D * , sampling rate q, private fine-tuning rounds T , first-stage fine-tuning rounds T < T for distribution matching, a public pre-trained LLM Output: Private on-device LM with DP guarantee 1: Randomly initialize an on-device LM; 2: // 1 First-stage private federated learning 3: Use DP-FTRL to train the on-device LM for rounds T ; 4: for each x ∈ D do 5: // 2 Probability evaluation 6: Compute the average (token) log prob log ppriv(x) given the privately fine-tuned LM at round T ; 7: Compute the average (token) log prob log ppub(x) given a publicly pre-trained LLM ; 8: end for 9: // 3 Distribtion matching 10: Sort D based on log ppriv(x) + log ppub(x) 11: Sample a subset of D as D with top log ppriv(x) + log ppub(x) values, such that |D | = q|D|. 12: // 4 Public mid-training with LLM distillation 13: Train the on-device LM with the loss Lpub on D 14: // 5 Second-stage private federated learning 15: Use DP-FTRL to train the on-device LM for the remaining rounds of T -T 16: return On-device LM with DP guarantee",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: Ablation studies on how distillation steps and top-k logits in distillation impact next token prediction accuracy (Acc.) of on-device LSTM models on the private StackOverflow dataset.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "log p -log p priv 2 π = E x∼p priv [(log p(x) -log p priv (x)) 2 ] = E x∼p priv log p(p priv |p)) 2",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "12) = β 2 d 2 pub, priv + (1 -β) 2 σ 2 priv + 2β(1 -β) pubpriv , E[ ˆ priv ]priv (13) = β 2 d 2 pub, priv + (1 -β) 2 σ 2 priv + 0 (14) = β 2 d 2 pub, priv + (1 -β) 2 σ 2 priv (15)",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Next Token Prediction Accuracy on the private StackOverflow dev set with or without public pre-training.",
"figure_data": "w/o pre-training w/ pre-trainingRounds0160001600ε = 1.77 ε = 18.710.0020.48 24.4516.9427.27 30.13",
"figure_id": "tab_2",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Figure 2: Ablation studies on how distillation steps and top-k logits in distillation impact next token prediction accuracy (Acc.) of on-device LSTM models on the dev set of the private StackOverflow dataset.Public pre-training with distillation loss Since we align the tokenizer of the on-device LM with the LLM to share the same vocabulary, we can align the output distribution of on-device LMs and LLMs by the cross-entropy loss. Formally, for next token prediction task, given the output logits from student on-device LMs z S , the gold label from the pre-training corpus y, and the logits from the distillation corpus of LLMs z T , we add an additional knowledge distillation loss L KD = CE(z S /t, z T /t) to the pre-training language modeling loss L LM = CE(z S , y) as our public pretraining loss L pub = L LM + βL KD where t is the temperature. We leave more distillation details in Appendix B.3.",
"figure_data": "Accuracy v.s. Distillation CoverageAccuracy v.s. Distillation CoverageAccuracy v.s. Distillation Coverage27.6Next Token Prediction Acc25 26 27 28 29 30eps=18.71, w/ public pre-training, covering 100% c4 eps=18.71, w/o public pre-training eps=18.71, w/ public pre-training eps=18.71, w/ public pre-training + distillation, topk=10Next Token Prediction Acc21 22 23 24 25 26 27eps=1.77, w/ public pre-training, covering 100% c4 eps=1.77, w/o public pre-training eps=1.77, w/ public pre-training eps=1.77, w/ public pre-training + distillation, topk=10Next Token Prediction Acc26.4 26.6 26.8 27.0 27.2 27.4 26.2eps=1.77, w/ public pre-training + distillation, topk=10 eps=1.77, w/ public pre-training + distillation, topk=5 eps=1.77, w/ public pre-training + distillation, topk=3 eps=1.77, w/ public pre-training + distillation, topk=105 Distillation Coverage of C4 dataset (%) 10 15 20 2505 Distillation Coverage of C4 dataset (%) 10 15 20 2526.02.0 Distillation Coverage of C4 dataset (%) 2.5 3.0 3.5 4.0 4.5 5.0(a) Acc. v.s. distillation steps (ε = 18.71)(b) Acc. v.s. distillation steps (ε = 1.77)(c) Acc. v.s. top-k logits (ε = 1.77)trained LLMs into on-device LMs during pre-training. The distillation pipeline contains the fol-lowing two steps:",
"figure_id": "tab_3",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Summary of techniques to improve downstream stream next token prediction accuracy and sample efficiency for on-device LSTM and transformer model evaluated on the StackOverflow test set.",
"figure_data": "q (% ofLLMDistributionAccuracy (LSTM)Accuracy (Transformer)Public Data) DistillationMatchingε=1.77ε=18.71ε=1.77ε=18.71No Public Training0%20.68±0.04 28.87±0.04 23.98±0.15 28.29±0.06Pre-training w/ public data (T = 0)100%28.01±0.26 30.70±0.01 28.05±0.02 30.10±0.00• LLM Distillation (100k steps)1%28.68±0.09 31.13±0.03 27.75±0.06 30.19±0.01• LLM Distillation (8k steps)0.08%26.18±0.04 29.53±0.10 25.31±0.08 29.36±0.12Mid-training w/ public data (T = T /2) 0.08%26.67±0.06 29.76±0.03 25.83±0.03 29.15±0.01• LLM Distillation (8k steps)0.08%27.01±0.03 30.18±0.06 26.04±0.12 29.47±0.05+ Distribution Matching0.08%28.01±0.08 30.63±0.0227.17±0.03 29.83±0.01",
"figure_id": "tab_4",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Ablation studies on the use of public LLM for distribution matching evaluated on the StackOverflow test set. standard deviation of test accuracy on the private StackOverflow dataset.",
"figure_data": "LSTMTransformerε=1.77ε=18.71ε=1.77ε=18.71w/ ppub(x)28.01±0.08 30.63±0.0227.17±0.03 29.83±0.01w/o ppub(x) 27.77±0.05 30.56±0.06 26.70±0.04 30.18±0.05",
"figure_id": "tab_5",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Ablation studies on the timing (T ) of distribution matching for mid-point public training on on-device LSTM.",
"figure_data": "T040080012001600ε=1.7725.41 27.08 27.73 26.40 18.40ε=18.71 28.38 30.07 30.37 29.45 19.34",
"figure_id": "tab_6",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Ablation studies on the timing (T ) of mid-point public training for on-device LSTM w/o distribution matching.",
"figure_data": "",
"figure_id": "tab_8",
"figure_label": "5",
"figure_type": "table"
}
] | Boxin Wang; Yibo Jacky Zhang; Yuan Cao; Bo Li; H Brendan Mcmahan; Sewoong Oh; Zheng Xu; Manzil Zaheer | [
{
"authors": "Martín Abadi; Andy Chu; Ian J Goodfellow; H B Mcmahan; Ilya Mironov; Kunal Talwar; Li Zhang",
"journal": "",
"ref_id": "b0",
"title": "Deep learning with differential privacy",
"year": "2016"
},
{
"authors": "E Amid; Arun Ganesh; Rajiv Mathews; Indra Swaroop; Shuang Ramaswamy; T Song; V Steinke; Om Suriyakumar; Abhradeep Thakkar; Thakurta",
"journal": "",
"ref_id": "b1",
"title": "Public data-assisted mirror descent for private model training",
"year": "2021"
},
{
"authors": "Galen Andrew; Om Thakkar; Brendan Mcmahan; Swaroop Ramaswamy",
"journal": "",
"ref_id": "b2",
"title": "Differentially private learning with adaptive clipping",
"year": "2021"
},
{
"authors": "",
"journal": "The TensorFlow Federated Authors",
"ref_id": "b3",
"title": "Tensorflow federated stack overflow dataset",
"year": "2019"
},
{
"authors": "Eugene Bagdasaryan; Congzheng Song; Matt Rogier Van Dalen; Áine Seigel; Cahill",
"journal": "",
"ref_id": "b4",
"title": "Training a tokenizer for free with private federated learning",
"year": "2022"
},
{
"authors": "Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell",
"journal": "Advances in neural information processing systems",
"ref_id": "b5",
"title": "Language models are few-shot learners",
"year": "2020"
},
{
"authors": "Zhiqi Bu; Yu-Xiang Wang; Sheng Zha; George Karypis",
"journal": "",
"ref_id": "b6",
"title": "Differentially private bias-term only fine-tuning of foundation models",
"year": "2022"
},
{
"authors": "Trevor Campbell; Tamara Broderick",
"journal": "The Journal of Machine Learning Research",
"ref_id": "b7",
"title": "Automated scalable bayesian inference via hilbert coresets",
"year": "2019"
},
{
"authors": "Zachary Charles; Kallista Bonawitz; Stanislav Chiknavaryan; Brendan Mcmahan",
"journal": "",
"ref_id": "b8",
"title": "Federated select: A primitive for communication-and memory-efficient federated learning",
"year": "2022"
},
{
"authors": "Mingqing Chen; Rajiv Mathews; Tom Ouyang; Françoise Beaufays",
"journal": "",
"ref_id": "b9",
"title": "Federated learning of out-of-vocabulary words",
"year": "2019"
},
{
"authors": "Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova",
"journal": "Association for Computational Linguistics",
"ref_id": "b10",
"title": "BERT: pre-training of deep bidirectional transformers for language understanding",
"year": "2019-06-02"
},
{
"authors": "Cynthia Dwork",
"journal": "Society for Industrial and Applied Mathematics",
"ref_id": "b11",
"title": "Differential privacy in new settings",
"year": "2010"
},
{
"authors": "Cynthia Dwork",
"journal": "Commun. ACM",
"ref_id": "b12",
"title": "A firm foundation for private data analysis",
"year": "2011"
},
{
"authors": "Cynthia Dwork; Frank Mcsherry; Kobbi Nissim; Adam Smith",
"journal": "Springer",
"ref_id": "b13",
"title": "Calibrating Noise to Sensitivity in Private Data Analysis",
"year": "2006"
},
{
"authors": "Cynthia Dwork; Aaron Roth",
"journal": "Found. Trends Theor. Comput. Sci",
"ref_id": "b14",
"title": "The algorithmic foundations of differential privacy",
"year": "2014"
},
{
"authors": "Maha Elbayad; Jiatao Gu; Edouard Grave; Michael Auli",
"journal": "ICLR",
"ref_id": "b15",
"title": "Depth-adaptive transformer",
"year": "2020"
},
{
"authors": "Arun Ganesh; Mahdi Haghifam; Milad Nasr; Sewoong Oh; Thomas Steinke; Om Thakkar; Abhradeep Thakurta; Lun Wang",
"journal": "",
"ref_id": "b16",
"title": "Why is public pretraining necessary for private model training?",
"year": "2023"
},
{
"authors": "Mitchell Gordon; Kevin Duh; Nicholas Andrews",
"journal": "Association for Computational Linguistics",
"ref_id": "b17",
"title": "Compressing BERT: Studying the effects of weight pruning on transfer learning",
"year": "2020"
},
{
"authors": "Andrew Hard; Daniel Chloé M Kiddon; Francoise Ramage; Hubert Beaufays; Kanishka Eichner; Rajiv Rao; Sean Mathews; Augenstein",
"journal": "",
"ref_id": "b18",
"title": "Federated learning for mobile keyboard prediction",
"year": "2018"
},
{
"authors": "Andrew Hard; Kurt Partridge; Cameron Nguyen; Niranjan Subrahmanya; Aishanee Shah; Pai Zhu; Ignacio Lopez Moreno; Rajiv Mathews",
"journal": "",
"ref_id": "b19",
"title": "Training keyword spotting models on non-iid data with federated learning",
"year": "2020"
},
{
"authors": "Sepp Hochreiter; Jürgen Schmidhuber",
"journal": "Neural computation",
"ref_id": "b20",
"title": "Long short-term memory",
"year": "1997"
},
{
"authors": "Enyi Jiang; Yibo ; Jacky Zhang; Oluwasanmi Koyejo",
"journal": "",
"ref_id": "b21",
"title": "Federated domain adaptation via gradient projection",
"year": "2023"
},
{
"authors": "Xiaoqi Jiao; Yichun Yin; Lifeng Shang; Xin Jiang; Xiao Chen; Linlin Li; F Wang; Qun Liu",
"journal": "",
"ref_id": "b22",
"title": "Tinybert: Distilling bert for natural language understanding",
"year": "2019"
},
{
"authors": "P Kairouz; B Mcmahan; Shuang Song; Om Thakkar; Abhradeep Thakurta; Zheng Xu",
"journal": "International Conference On Machine Learn",
"ref_id": "b23",
"title": "Practical and private (deep) learning without sampling or shuffling",
"year": "2021"
},
{
"authors": "P Kairouz; H B Mcmahan; Brendan Avent; A Bellet; M Bennis; A Bhagoji; Keith Bonawitz; Zachary B Charles; Graham Cormode; Rachel Cummings; G L Rafael; S D'oliveira; David Rouayheb; Josh Evans; Zachary Gardner; Adrià Garrett; Badih Gascón; Phillip B Ghazi; M Gibbons; Z Gruteser; Chaoyang Harchaoui; Lie He; Zhouyuan He; Ben Huo; Justin Hutchinson; Martin Hsu; T Jaggi; Gauri Javidi; M Joshi; Jakub Khodak; A Konecný; F Korolova; O Koushanfar; Tancrède Koyejo; Yang Lepoint; Prateek Liu; M Mittal; R Mohri; A Nock; R Özgür; Mariana Pagh; Hang Raykova; D Qi; R Ramage; D Raskar; Weikang Song; S Song; Ziteng Stich; A Sun; Florian Suresh; Praneeth Tramèr; Jianyu Vepakomma; Li Wang; Zheng Xiong; Qiang Xu; Felix X Yang; Han Yu; Sen Yu; Zhao",
"journal": "Found. Trends Mach. Learn",
"ref_id": "b24",
"title": "Advances and open problems in federated learning",
"year": "2019"
},
{
"authors": "Gavin Kerrigan; Dylan Slack; Jens Tuyls",
"journal": "Association for Computational Linguistics",
"ref_id": "b25",
"title": "Differentially private language models benefit from public pre-training",
"year": "2020"
},
{
"authors": "Taku Kudo; John Richardson",
"journal": "",
"ref_id": "b26",
"title": "Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing",
"year": "2018"
},
{
"authors": "Tian Li; M Zaheer; J Sashank; Virginia Reddi; Smith",
"journal": "",
"ref_id": "b27",
"title": "Private adaptive optimization with side information",
"year": "2022"
},
{
"authors": "Xuechen Li; Florian Tramèr; Percy Liang; Tatsunori B Hashimoto",
"journal": "",
"ref_id": "b28",
"title": "Large language models can be strong differentially private learners",
"year": "2021"
},
{
"authors": "Brendan Mcmahan; Daniel Ramage; Kunal Talwar; Li Zhang",
"journal": "",
"ref_id": "b29",
"title": "Learning differentially private recurrent language models",
"year": "2018"
},
{
"authors": "H B Mcmahan; Eider Moore; D Ramage; S Hampson; B A Y Arcas",
"journal": "",
"ref_id": "b30",
"title": "Communicationefficient learning of deep networks from decentralized data",
"year": "2017"
},
{
"authors": "John Nguyen; Jianyu Wang; Kshitiz Malik; Maziar Sanjabi; Michael Rabbat",
"journal": "",
"ref_id": "b31",
"title": "Where to begin? on the impact of pre-training and initialization in federated learning",
"year": "2022"
},
{
"authors": "Natalia Ponomareva; Jasmijn Bastings; Sergei Vassilvitskii",
"journal": "",
"ref_id": "b32",
"title": "Training text-to-text transformers with privacy guarantees",
"year": "2022"
},
{
"authors": "Alec Radford; Jeff Wu; Rewon Child; David Luan; Dario Amodei; Ilya Sutskever",
"journal": "",
"ref_id": "b33",
"title": "Language models are unsupervised multitask learners",
"year": "2019"
},
{
"authors": "Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu",
"journal": "Journal of Machine Learning Research",
"ref_id": "b34",
"title": "Exploring the limits of transfer learning with a unified text-totext transformer",
"year": "2020"
},
{
"authors": "Ali Rahimi; Benjamin Recht",
"journal": "Advances in neural information processing systems",
"ref_id": "b35",
"title": "Random features for large-scale kernel machines",
"year": "2007"
},
{
"authors": "Swaroop Ramaswamy; Om Thakkar; Rajiv Mathews; Galen Andrew; H Brendan Mcmahan; Françoise Beaufays",
"journal": "",
"ref_id": "b36",
"title": "Training production language models without memorizing user data",
"year": "2020"
},
{
"authors": "J Sashank; Zachary Reddi; Manzil Charles; Zachary Zaheer; Keith Garrett; Jakub Rush; Sanjiv Konečný; Hugh Brendan Kumar; Mcmahan",
"journal": "",
"ref_id": "b37",
"title": "Adaptive federated optimization",
"year": "2021"
},
{
"authors": "Mike Schuster; Kaisuke Nakajima",
"journal": "",
"ref_id": "b38",
"title": "Japanese and korean voice search",
"year": "2012"
},
{
"authors": "Rico Sennrich; Barry Haddow; Alexandra Birch",
"journal": "Association for Computational Linguistics",
"ref_id": "b39",
"title": "Neural machine translation of rare words with subword units",
"year": "2016"
},
{
"authors": "Zhiqing Sun; Hongkun Yu; Xiaodan Song; Renjie Liu; Yiming Yang; Denny Zhou",
"journal": "ACL",
"ref_id": "b40",
"title": "Mobilebert: a compact task-agnostic bert for resource-limited devices",
"year": "2020"
},
{
"authors": "Swabha Swayamdipta; Roy Schwartz; Nicholas Lourie; Yizhong Wang; Hannaneh Hajishirzi; Noah A Smith; Yejin Choi",
"journal": "",
"ref_id": "b41",
"title": "Dataset cartography: Mapping and diagnosing datasets with training dynamics",
"year": "2020"
},
{
"authors": "Romal Thoppilan; Daniel De Freitas; Jamie Hall; Noam Shazeer; Apoorv Kulshreshtha; Heng-Tze; Alicia Cheng; Taylor Jin; Leslie Bos; Yu Baker; Yaguang Du; Hongrae Li; Huaixiu Lee; Amin Steven Zheng; Marcelo Ghafouri; Yanping Menegali; Maxim Huang; Dmitry Krikun; James Lepikhin; Dehao Qin; Yuanzhong Chen; Zhifeng Xu; Adam Chen; Maarten Roberts; Vincent Bosma; Yanqi Zhao; Chung-Ching Zhou; Igor Chang; Will Krivokon; Marc Rusch; Pranesh Pickett; Laichee Srinivasan; Kathleen Man; Meredith Ringel Meier-Hellstern; Tulsee Morris; Renelito Delos Doshi; Toju Santos; Johnny Duke; Ben Soraker; Vinodkumar Zevenbergen; Mark Prabhakaran; Ben Diaz; Kristen Hutchinson; Alejandra Olson; Erin Molina; Josh Hoffman-John; Lora Lee; Ravi Aroyo; Alena Rajakumar; Matthew Butryna; Viktoriya Lamm; Joe Kuzmina; Aaron Fenton; Rachel Cohen; Ray Bernstein; Blaise Kurzweil; Claire Aguera-Arcas; Marian Cui; Ed Croak; Quoc Chi; Le",
"journal": "",
"ref_id": "b42",
"title": "Lamda: Language models for dialog applications",
"year": "2022"
},
{
"authors": "Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Ł Ukasz Kaiser; Illia Polosukhin",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b43",
"title": "Attention is all you need",
"year": "2017"
},
{
"authors": "Curran Associates; Inc ",
"journal": "",
"ref_id": "b44",
"title": "",
"year": ""
},
{
"authors": "Jianyu Wang; Zachary Charles; Zheng Xu; Gauri Joshi; H Brendan Mcmahan; Blaise Aguera Y Arcas; Maruan Al-Shedivat; Galen Andrew; Salman Avestimehr; Katharine Daly",
"journal": "",
"ref_id": "b45",
"title": "A field guide to federated optimization",
"year": "2021"
},
{
"authors": "Wenhui Wang; Furu Wei; Li Dong; Hangbo Bao; Nan Yang; Ming Zhou",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b46",
"title": "Minilm: Deep selfattention distillation for task-agnostic compression of pre-trained transformers",
"year": "2020"
},
{
"authors": "Shanshan Wu; Tian Li; Zachary Charles; Yu Xiao; Ziyu Liu; Zheng Xu; Virginia Smith",
"journal": "",
"ref_id": "b47",
"title": "Motley: Benchmarking heterogeneity and personalization in federated learning",
"year": "2022"
},
{
"authors": "Zhaozhuo Xu; Luyang Liu; Zheng Xu; Anshumali Shrivastava",
"journal": "",
"ref_id": "b48",
"title": "Adaptive sparse federated learning in large output spaces via hashing",
"year": "2022"
},
{
"authors": "Zheng Xu; Maxwell Collins; Yuxiao Wang; Liviu Panait; Sewoong Oh; Sean Augenstein; Ting Liu; Florian Schroff; Brendan Mcmahan",
"journal": "",
"ref_id": "b49",
"title": "Learning to generate image embeddings with user-level differential privacy",
"year": "2022"
},
{
"authors": "Zheng Xu; Yanxiang Zhang; Galen Andrew; Christopher Choquette; Peter Kairouz; Brendan Mcmahan; Jesse Rosenstock; Yuanbo Zhang",
"journal": "",
"ref_id": "b50",
"title": "Federated learning of gboard language models with differential privacy",
"year": "2023"
},
{
"authors": "Da Yu; Saurabh Naik; Arturs Backurs; Sivakanth Gopi; Huseyin A Inan; Gautam Kamath; Janardhan Kulkarni; Yin Tat Lee; Andre Manoel; Lukas Wutschitz; Sergey Yekhanin; Huishuai Zhang",
"journal": "",
"ref_id": "b51",
"title": "Differentially private fine-tuning of language models",
"year": "2022-04-25"
},
{
"authors": "Jacky Zhang; Rajiv Khanna; Anastasios Kyrillidis; Sanmi Koyejo",
"journal": "",
"ref_id": "b52",
"title": "Bayesian coresets: Revisiting the nonconvex optimization perspective",
"year": "2021"
},
{
"authors": "",
"journal": "",
"ref_id": "b53",
"title": "5 Next Token Prediction Acc Accuracy v.s. Distillation Coverage eps=18.71",
"year": ""
}
] | [
{
"formula_coordinates": [
3,
75.36,
169.72,
209.28,
9.57
],
"formula_id": "formula_0",
"formula_text": "Pr[M(D) ∈ S] ≤ exp(ε) Pr[M(D ) ∈ S] + δ."
},
{
"formula_coordinates": [
7,
305.78,
688.41,
220.44,
13.7
],
"formula_id": "formula_1",
"formula_text": "Theorem 5.1. Let ( f ) = E[ f -priv 2 ] charac-"
},
{
"formula_coordinates": [
8,
120.97,
336.21,
168.17,
14.37
],
"formula_id": "formula_2",
"formula_text": "( pub ) = d 2 pub, priv(1)"
},
{
"formula_coordinates": [
8,
120.28,
354.69,
168.85,
42.75
],
"formula_id": "formula_3",
"formula_text": "( ˆ priv ) = σ 2 priv (2) ( ĥ) = 1 4 d 2 pub, priv + 1 4 σ 2 priv (3)"
},
{
"formula_coordinates": [
8,
76.25,
434.81,
212.89,
32.95
],
"formula_id": "formula_4",
"formula_text": "• ( ĥ) ≤ min{ ( pub ), ( ˆ priv )} if 1 3 ≤ d 2 pub, priv σ 2 priv ≤ 3."
},
{
"formula_coordinates": [
15,
81.78,
708.95,
442.64,
24.18
],
"formula_id": "formula_5",
"formula_text": "f π = E x∼π [f (x) 2 ]."
},
{
"formula_coordinates": [
16,
238.6,
455.31,
285.81,
14.37
],
"formula_id": "formula_6",
"formula_text": "( pub ) = d 2 pub, priv(6)"
},
{
"formula_coordinates": [
16,
241.31,
557.22,
283.1,
13.7
],
"formula_id": "formula_7",
"formula_text": "fβ = β pub + (1 -β) ˆ priv .(9)"
},
{
"formula_coordinates": [
16,
83.47,
617.96,
440.94,
69.36
],
"formula_id": "formula_8",
"formula_text": "( fβ ) = E[ fβ -priv 2 ] = E[ β pub + (1 -β) ˆ priv -priv 2 ] (10) = E[ β( pub -priv ) + (1 -β)( ˆ priv -priv ) 2 ] (11) = β 2 pub -priv 2 + (1 -β) 2 E ˆ priv -priv 2 + 2β(1 -β)E pub -priv , ˆ priv -priv("
}
] |
10.4208/jml.220404 | 2023-05-20 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b0",
"b1",
"b2",
"b3",
"b4",
"b3",
"b5",
"b6",
"b7",
"b8"
],
"table_ref": [],
"text": "Many experiments have observed a phenomenon, called the edge of stability (EoS) (Wu et al., 2018;Cohen et al., 2021;Arora et al., 2022), that during the neural network (NN) training, the maximum eigenvalue of the loss Hessian, λ max , progressively increases until it reaches 2/η (η is learning rate), and then λ max stays around 2/η. At the EoS stage, the loss would continuously decrease, sometimes with slight oscillation. Training with a larger learning rate leads to a solution with smaller λ max . Since λ max is often used to indicate the sharpness of the loss landscape, a larger learning rate results in a flatter solution. Intuitively as shown in Fig. 1, the flat solution is more robust to perturbation and has better generalization performance (Keskar et al., 2016;Hochreiter and Schmidhuber, 1997). Therefore, training with a larger learning rate would achieve better generalization performance. In this work, we argue this intuitive analysis in Fig. 1 with λ max as the sharpness measure, which encounters difficulty in NNs through the study of loss spikes.\nIn a neural network training process, one may sometimes observe a phenomenon of loss spike, where the loss rapidly ascends and then descends to the value before the ascent. Typical examples are shown in Fig. 2. We show a special loss landscape structure underlying the loss spike, which is called a smaller-loss-as-sharper (SLAS) structure. In the SLAS structure, the training is driven by descending the loss while entering an increasingly sharp region. Once the sharpness is too large, the loss would ascend exponentially fast. To explain why the loss can descend so fast, we provide a frequency Figure 1: Schematic illustration of an ideal explanation for why flat solutions generalize well (Keskar et al., 2016). perspective analysis. We find that the deviation in the ascending stage is dominated by low-frequency components. Based on the frequency principle (Xu et al., 2019(Xu et al., , 2020) that low-frequency converges faster than high-frequency, we rationalize the fast descent.\nThe study of loss spike provides an important information that the deviation at the first eigen direction is dominated by low-frequency. We then further argue the link between λ max flatness and generalization. In practical datasets, low-frequency information is often dominant and shared by both the training and the test datasets. Therefore, the training can learn low-frequency well. Since the sharpest direction, indicated by the maximum eigenvalue of the loss Hessian, relates more to the low-frequency, a solution with good generalization and a solution with bad generalization have little difference in the sharpest direction, verified by a series of experiments. Hence, λ max with the intuitive explanation in Fig. 1 encounters difficulty in understanding the generalization of neural networks, such as why a larger learning rate results in better generalization for networks with EoS training.\nWe also find that a loss spike can facilitate condensation, that is, the input weights of different neurons in the same layer evolve towards the same, which would reduce the network's effective size. Condensation is a non-linear feature learning phenomenon in neural networks, which may be the underlying mechanism for why the loss spike improves generalization (He et al., 2019;Jastrzebski et al., 2017), rather than simply controlling the value of λ max . This work studies the loss spike from the landscape perspective and the frequency perspective, and revisits the relation between the generalization and the flatness, defined by the maximum eigenvalue of the loss Hessian. This work also conjectures the loss spike may improve generalization via the facilitation of condensation."
},
{
"figure_ref": [],
"heading": "Related works",
"publication_ref": [
"b1",
"b0",
"b9",
"b10",
"b11",
"b12",
"b13",
"b2",
"b14",
"b16",
"b4",
"b18",
"b19",
"b20",
"b21",
"b22",
"b23",
"b24",
"b25",
"b27",
"b28",
"b29",
"b30",
"b32",
"b33",
"b34",
"b35",
"b36",
"b5",
"b37",
"b38",
"b40",
"b41",
"b42",
"b43"
],
"table_ref": [],
"text": "Previous works (Cohen et al., 2021;Wu et al., 2018;Xing et al., 2018;Ahn et al., 2022;Lyu et al., 2022;Wang et al., 2022) conduct an extensive study of the EoS phenomenon under various settings. Lewkowycz et al. (2020) observe that when the initial sharpness exceeds 2/η, gradient descent \"catapults\" into a stable region and converges. Arora et al. (2022) analyze progressive sharpening and the edge of stability phenomenon under specific settings, such as normalized gradient descent. Damian et al. (2022) show that the third-order terms bias towards flatter minima to understand EoS. Ma et al. (2022) attribute the progressive sharpening to a subquadratic structure of the loss landscape, i.e., the maximum eigenvalue of the loss Hessian is larger when the loss is smaller in a direction. They also propose a flatness-driven motion to study the EoS stage, that is, the training would move towards a flatter minimum, such that the fixed flatness can correspond to points with smaller and smaller loss values due to the subquadratic property. We call this structure a smaller-loss-as-flatter (SLAF) structure. The SLAF structure should expect a continuous decrease in the loss rather than a loss spike. Agarwala et al. (2022) use a quadratic regression model with MSE to study EoS. Similarly, in their model, the loss spike can not happen. Ma et al. (2022) study the loss spike from the perspective of adaptive gradient optimization algorithms, while we focus on the loss landscape structure and use gradient descent training in this paper.\nA series of works link the generalization performance of solutions to the landscape of loss functions through the observation that flat minima tend to generalize better (Hochreiter and Schmidhuber, 1997;Wu et al., 2017;Ma and Ying, 2021). Algorithms that favor flat solutions are designed to improve the generalization of the model (Izmailov et al., 2018;Chaudhari et al., 2019;Lin et al., 2018;Zheng et al., 2021;Foret et al., 2020). On the other hand, Dinh et al. (2017) show that sharp minimum can also generalize well by rescaling the parameters at a flat minimum with ReLU activation. In this work, we study the relationship between flatness and generalization from a new perspective, i.e., the frequency perspective, without the limitation of the activation function. Luo et al. (2021); Zhou et al. (2022) mainly identify the linear regime and the condensed regime of the parameter initialization for two-layer and three-layer wide ReLU NNs, which determines the final fitting result of the network. In the linear regime (Jacot et al., 2018;Arora et al., 2019), the training dynamics of NNs are approximately linear and similar to a random feature model. On the contrary, in the condensed regime, active neurons are condensed at several discrete orientations. At this point, the network is equivalent to another network with a reduced width, which may explain why NNs outperform traditional algorithms (Breiman, 1995;Zhang et al., 2021). For the initial stage of training, A series of works (Zhou et al., 2021;Chen et al., 2023;Maennel et al., 2018;Pellegrini and Biroli, 2020) study the characteristics of the initial condensation for different activation functions. Andriushchenko et al. (2022) find that stochastic gradient descent (SGD) with a large learning rate can facilitate sparse solutions and attributes it to the noise structure of SGD. In our work, we find that for the noise-free full-batch gradient descent algorithm, the loss spike can also facilitate the condensation phenomenon, implying that the noise structure is not the intrinsic cause of condensation.\nThe frequency principle is examined in extensive datasets and deep neural network models (Xu et al., 2019;Xu and Zhou, 2021;Rahaman et al., 2019). Subsequent theoretical studies show that the frequency principle holds in the general setting with infinite samples (Luo et al., 2021). An overview for frequency principle is referred to Xu et al. (2022). Based on the theoretical understanding, the frequency principle inspires the design of deep neural networks to learn a function with high-frequency fast (Liu et al., 2020;Jagtap et al., 2020;Biland et al., 2019)."
},
{
"figure_ref": [],
"heading": "Preliminary: Linear stability in training quadratic model",
"publication_ref": [],
"table_ref": [],
"text": "We consider a simple quadratic model with the loss R(θ) = λθ 2 /2 trained by gradient descent with learning rate η, θ(t + 1) = θ(t) -η • dR(θ)/dθ. To ensure the linear stability of the training, it requires |θ(t + 1)| < |θ(t)|, which implies |1 -λη| < 1, i.e., otherwise, the training will diverge. Note that λ is the Hessian of R(θ). Similarly, to ensure the linear stability of training a neural network, it requires that the maximum eigenvalue of the loss Hessian is smaller than 2/η, i.e., 2 over the learning rate. Therefore, the maximum eigenvalue of the loss Hessian is often used as the measure of the sharpness of the loss landscape."
},
{
"figure_ref": [],
"heading": "Loss spike",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we study the phenomenon of loss spike, where the loss would suddenly increase and decrease rapidly. For example, as shown in Fig. 2(a,d), we train a tanh fully-connected neural network (FNN) with 20 hidden neurons for a one-dimensional fitting problem, and a ReLU convolutional neural network (CNN) for the CIFAR10-1k classification problem with MSE. Both two models experience loss spikes. The red curves, i.e., the λ max value, show that the loss spikes occur at the EoS stage."
},
{
"figure_ref": [],
"heading": "Typical loss spike experiments",
"publication_ref": [],
"table_ref": [],
"text": "To observe the loss spike clearly, we zoom in on the training epochs around the spike, shown in Fig. 2(b,e). The selected epochs are marked green in Fig. 2(a,d). When the maximum eigenvalue of Hessian λ max (red) exceeds 2/η (black dashed line), the loss increases, and when λ max < 2/η, the loss decreases, which are consistent with the linear stability analysis.\nWe then study the parameter space for more detailed characterization. Given t training epochs, and let θ i denote model parameters at epoch i, we apply PCA to the matrix M\n= [θ 1 -θ t , • • • , θ t -θ t ],\nand then select the first two eigen directions e 1 , e 2 . The two-dimensional loss surface based on e 1 and e 2 can be calculated by R S (θ t + αe 1 + βe 2 ), where α, β are the step sizes, and R S is the loss function under the dataset S. The trajectory point of parameter θ i can be calculated by the projection of θ i -θ t in the PCA directions, i.e., ( θ i -θ t , e 1 , θ i -θ t , e 2 ). Parameter trajectories (blue dots) and loss surfaces along PCA directions are shown in Fig. 2(c,f). In two distinct examples, they exhibit similar behaviors. At the beginning of the ascent stage of the spike, the parameter is at a small-loss region, where the opening of the contour lines is towards the left, indicating a leftward component of descent direction. In the left region, the contour lines are denser, implying a sharper loss surface. Once λ max > 2/η, the parameters become unstable, and the loss value increases exponentially. In the large-loss region, the opening of the contour shifts to the right, indicating a rightward component of the descent direction, resulting in a sparser contour, i.e., a flatter loss surface. After several steps, when λ max < 2/η, the training returns to the stable stage. The sum of the explained variance ratios of the first two PCA directions is 0.9882."
},
{
"figure_ref": [],
"heading": "Smaller-loss-as-sharper (SLAS) structure",
"publication_ref": [],
"table_ref": [],
"text": "The above experiments reveal a common structure that causes a loss spike, namely, the λ max sharpness increases in the direction of decreasing loss. We call this structure smaller-loss-as-sharper (SLAS) structure. The SLAS structure differs from the SLAF (smaller-loss-as-flatter) structure studied in Ma et al. (2022), which is also common in the EoS stage as shown in Fig. 3(a). A toy example of the SLAS structure is shown in Fig. 3(b). The left cross-section of the loss landscape has a flatter curvature while the right one has a sharper curvature. At the minimum of the left cross-section (the L 1 dashed line), the opening of the contour lines towards the right and the parameter point will also move right, which makes the curvature sharper. Once η > 2/λ max , it starts to diverge to a large-loss region and the opening of the contour turns left (the L 2 dashed line), which makes the curvature flatter.\nThe following quadratic model is a simple example of SLAS structure,\nf (x, y) = (50x + 200)y 2 -x + 5,(1)\nwhere (x, y) ∈ (-4, +∞) × R. For any constant C, y = 0 is the minimum point of f (C, y), and the larger x is, the sharper the loss landscape in the y-direction. As shown in Fig. 3(c,d), the loss curve and the trajectory of parameters are similar to the realistic example above, where the parameters move toward the sharp direction at the beginning of the loss spike, and then move toward the flat direction. The intuitive explanation for the above phenomenon is that as x increases, f (x, 0) decreases, which means that f (x, 0) has a smaller value at the sharp region, i.e., the SLAS structure, which makes the opening of the contour lines towards different directions at different loss levels. For this example, we can exactly compute the derivative of Eq. ( 1) as follows:\n∂f (x, y) ∂x = 50y 2 -1.\nThus we have ∂f (x, y) ∂x\n > 0 if f (x, y) < 9 = 0 if f (x, y) = 9 < 0 if f (x, y) > 9\n, which indicates that the toy model has a positive gradient component in the x direction when the parameters are in the small-loss region (f (x, y) < 9), while a negative gradient component in the x direction when the parameters are in the large-loss region (f (x, y) > 9).\nAlthough the SLAS structure can explain the mechanism of the ascent stage based on the toy model, it can not explain the reason for the rapid descent of the loss in the descent phase of the loss spike, which takes much fewer steps than the training from the same level loss at the initialization. For instance, for the quadratic model in the Preliminary section, the descent would be very slow if the learning rate is slightly smaller than 2/λ max . Moreover, due to the high dimensionality of the parameter space, the parameter trajectory does not always align with the first eigen direction, otherwise, as shown in the toy model, the loss would not decrease continuously. In the following, we take a step toward understanding the rapid decrease from the frequency perspective."
},
{
"figure_ref": [
"fig_2"
],
"heading": "Frequency perspective for understanding descent stage",
"publication_ref": [
"b5",
"b6",
"b38",
"b45"
],
"table_ref": [],
"text": "In this subsection, we study the mechanism of the rapid loss descent during the descent stage in a loss spike from the perspective of frequency.\nWe base our analysis on a common phenomenon of frequency principle (Xu et al., 2019(Xu et al., , 2020;;Zhang et al., 2021;Luo et al., 2021;Rahaman et al., 2019;Ronen et al., 2019), which states that deep NNs often fit target functions from low to high frequencies during the training. A series of frequency principle works show that low-frequency can converge faster than high-frequency. Compared to the peak point of the loss spike with the point with the same loss value at the initial training, the descent during the spike should eliminate more low-frequency with a fast speed while the descent from the initial model should eliminate more high-frequency with a slow speed. To verify this conjecture, we study the frequency distribution of the converged part during the descent stage. The peak of the loss spike is denoted as θ max , the initial point which has the similar loss of θ max is denoted as θ ini,m , the parameter at the end of the loss spike (a point is roughly selected when the descent is slow) is denoted as θ end . We then study the frequency distribution of spike output difference f peak,diff := f θmax -f θ end and initial output difference f ini,diff := f θini,m -f θ end .\nFor comparison, we also randomly select parameter θ rnd := θ end + ( θ end -θ max 2 / ε 2 )ε, where ε ∼ N (0, I) is a random variable. We then study the frequency distribution of random output difference f rnd,diff := f θ rnd -f θ end .\nWe characterize the frequency distribution by taking different low-frequency thresholds to study low-frequency proportion. For a low-frequency threshold K, a low-frequency proportion (LFP) is defined as follows to characterize the power proportion of the low-frequency component over the whole spectrum,\nLFP(K) = k≤K fθ (k) 2 k fθ (k) 2 , (2\n)\nwhere fθ indicates the Fourier transform of function f θ .\nAs shown in Fig. 4, the low-frequency proportion of the spike output difference is significantly larger than the low-frequency proportion of the initial output difference and the random output difference, where we take 100 samples of random variable ε for the mean value and the error bar for each low-frequency threshold. The large low-frequency proportion of the spike output difference is the key reason for the rapid drop in the loss value during the descent stage, as suggested by the frequency principle."
},
{
"figure_ref": [],
"heading": "Revisit the flatness-generalization picture",
"publication_ref": [
"b4",
"b46"
],
"table_ref": [],
"text": "Motivated by the loss spike analysis from the frequency perspective, we further revisit the common flatness-generalization picture. A series of previous works (Hochreiter and Schmidhuber, 1997;Li et al., 2017) attempt to link the flatness of the loss landscape with generalization, so as to characterize the model through flatness conveniently. A classic empirical illustration is shown in Fig. 1, which vividly expresses the reason why flat solutions tend to have better generalization. Usually, the training loss landscape and the test landscape do not exactly coincide due to sampling noise. A flat solution would be robust to the perturbation while a sharp solution would not. For such a one-dimensional case, this analysis is valid, but the loss landscape of a NN case is very high-dimensional, and such simple visualization or explanation is yet to be validated.\nThe first eigen direction of the loss Hessian, i.e., the eigen direction corresponding to the maximum eigenvalue, is the sharpest direction. Based on the flatness-generalization picture, it is natural to use the maximum eigenvalue as the measure for the flatness, which can also indicate generalization. However, this naive analysis is not always correct for neural networks."
},
{
"figure_ref": [
"fig_2",
"fig_3",
"fig_3"
],
"heading": "Frequency perspective",
"publication_ref": [],
"table_ref": [],
"text": "Since the maximum eigenvalue of the loss Hessian can indicate the linear stability of the training, it is often used as a measure for flatness/sharpness, that is, a larger maximum eigenvalue indicates a sharper loss landscape. As shown by the linear stability analysis, once the maximum eigenvalue is larger than 2/η, the training would oscillate and diverge along the first eigen direction. Meanwhile, as the parameter moves away from the minimum point along the first eigen direction, the loss spike is mainly due to the large low-frequency difference as shown in Fig. 4. Therefore, the deviation in the first eigen direction of the loss Hessian mainly leads to the deviation of low-frequency components.\nIn order to examine the above analysis, we first obtain the model parameter θ train with poor generalization by training the model initialized in the linear regime (Luo et al., 2021), and then further train the model parameter θ train on the test dataset with a small learning rate to obtain the model parameter θ test .\nWe study the impact of each eigen direction on the test loss by eliminating the difference between θ train and θ test in the i-th eigen direction ν i , where i is the index of eigenvalues. As shown in Fig. 5(a), we study the change of the test loss L(i) with the eigenvalue index i as follows to study the effect of eigenvectors on generalization,\nL(i) = R Stest θ train + i j=1 θ test -θ train , ν j ν j ,\nwhere S test is the test dataset. The movement of parameters on the eigenvectors corresponding to large eigenvalues has a weak impact on the test loss, while the movement of parameters on the eigenvectors corresponding to small eigenvalues has a significant impact on the test loss.\nA reasonable explanation from the perspective of frequency is as follows. In common datasets, lowfrequency components often dominate over high-frequency ones. For noisy sampling, the dominant low-frequency is shared by both the training and the test data. When the parameters move along the eigen directions corresponding to the large eigenvalues, the network output often changes at low-frequency, which is already captured by both θ train and θ test . Therefore, the improvement of model generalization often requires certain high-frequency changes. As shown in Fig. 5(b), we move the corresponding θ train along the first nine eigen directions, and show the difference between the network outputs before and after the movement, i.e., f θ train+ν i / √ λ i -f θtrain , where the 1/ √ λ i item is to make the loss of the network moved in different eigen directions approximately the same. From the difference between the outputs before and after the movement, it can be seen that when the parameters move along the eigen direction corresponding to the larger eigenvalue, the change of the model output is often less oscillated, i.e., dominated by the lower-frequency. Since the low-frequency is captured by both θ train and θ test , they should be close in the eigen directions corresponding to large eigenvalues, which is verified in the following subsection."
},
{
"figure_ref": [
"fig_7",
"fig_7"
],
"heading": "Difference on each eigen direction",
"publication_ref": [
"b47"
],
"table_ref": [],
"text": "We then examine the projection of θ test -θ train in each eigen direction of H(θ train ). As shown in Fig. 6, we show the projection of θ test -θ train on each eigenvector ν i (blue bar) for the FNN on function fitting problem and the CNNs on CIFAR10 classification problem. Due to the high complexity of calculating the eigenvectors of the large-size Hessian matrix, we use the Lanczos method (Cullum and Willoughby, 2002) to numerically compute the first N eigenvalues and their corresponding eigenvectors. For n < N , we use\nn i=1 λ 2 i / N i=1 λ 2\ni to represent the explained variance ratio, i.e., to measure how much flatness information the first n eigen directions (orange line) can explain. For different network structures and model tasks, the projection value of θ test -θ train on the eigenvector ν i has a positive correlation with the eigenvalue index i, which confirms that θ train and θ test have little difference on low-frequency part. Note that in Fig. 6(d), the two minima, θ small and θ large , are found by small and large batch sizes, respectively, and they also have little difference in eigen directions corresponding to large eigenvalues."
},
{
"figure_ref": [],
"heading": "Implications",
"publication_ref": [],
"table_ref": [],
"text": "The above analysis suggests the following implications: i) The maximum eigenvalue of the loss Hessian is a good measure of sharpness for whether the training is linearly stable but not a good 0 400 800 1200 eigenvalue index 10 4 10 3 10 2 10 1 test error measure for generalization; ii) The common low-dimensional flatness-generalization picture suffers difficulty in understanding the high-dimensional loss landscape of neural network. The generalization performance is a combined effect of most eigen directions, including those with small eigenvalues.\n(a) FNN i = 1 i = 2 i = 3 i = 4 i = 5 i = 6 i = 7 i = 8 i = 9 (b) output difference\nn i = 1 2 i / N i = 1 2 i (a) FNN\nn i = 1 2 i / N i = 1 2 i (b) Two-layer CNN\nn i = 1 2 i / N i = 1 2 i (c) Three-layer CNN\nn i = 1 2 i / N i = 1 2 i (d) Different batch sizes"
},
{
"figure_ref": [
"fig_8",
"fig_8"
],
"heading": "Loss spike facilitates condensation",
"publication_ref": [
"b7",
"b8",
"b32",
"b49",
"b50",
"b51"
],
"table_ref": [],
"text": "From the analysis above, the restriction on λ max does not seem to be the essential reason why loss spike affects the generalization of the model. In this section, we study the effect of loss spike on condensation, which may improve the model's generalization in some situations (He et al., 2019;Jastrzebski et al., 2017). A condensed network, which refers to a network with neurons condensing in several discrete directions, is equivalent to another smaller network (Zhou et al., 2021;Luo et al., 2021). It has a lower effective complexity than it appears. The embedding principle (Zhang et al., 2021(Zhang et al., , 2022;;Fukumizu et al., 2019;Simsek et al., 2021) shows that a condensed network, although equivalent to a smaller one in approximation, has more degeneracy and descent directions that may accelerate the training process. The low effective complexity and simple training process may be underlying reasons for good generalization. We show that the loss spike can facilitate the condensation phenomenon for the noise-free full-batch gradient descent algorithm.\nAs shown in Fig. 7, we train a tanh NN with 100 hidden neurons for the one-dimensional fitting problem to fit the data using MSE as the loss function. Additional experimental verification on ReLU NNs is provided in Appendix B.1. To clearly study the effect of loss spike on condensation, we take the parameter initialization distribution in the linear regime (Luo et al., 2021) that does not induce condensation without additional constraints. For NNs with identical initialization, we train the network separately with a small learning rate (blue) and a large learning rate (orange). For the left subfigure in Fig. 7, the loss value has a significant spike for the large learning rate, but not for the small one. At the same time, the middle subfigure reveals that the model output without a loss spike (blue) during the training process has more oscillation than the model output with a loss spike (orange). We study the features of parameters to understand the underlying effect of loss spike better.\nTo study the parameter features, we measure each parameter pair (a j , w j ) by the feature direction ŵj = w j / w j 2 and amplitude2 A j = |a j | w j 2 . For a NN with one-dimensional input, after incorporating the bias term, w j is two-dimensional, and we use the angle between w j and the unit vector "
},
{
"figure_ref": [
"fig_8",
"fig_11",
"fig_9"
],
"heading": "Conclusion and discussion",
"publication_ref": [],
"table_ref": [],
"text": "In this work, we provide an explanation for loss spikes in neural network training. We explain the ascent stage based on the landscape structure, i.e., the SLAS structure, and for the descent stage, we explain it from the perspective of frequency. We revisit the common flatness-generalization picture based on the frequency analysis. We also find that noise-free gradient descent with loss spikes can facilitate condensation, which may be an underlying reason for the good generalization in some situations. Obviously, many questions remain open. For example, why the eigen direction corresponding to a large eigenvalue is dominated by low-frequency? Why the loss spike can facilitate the condensation? We leave the discussion of these important questions to future work. For Fig. 7, Fig. 9, we use the two-layer tanh FNN with a width of 200 to fit the target function using full-batch gradient descent as follows,\nf (x) = tanh(x -6) + tanh(x + 6).\nThe initialization of the parameters θ ∼ N (0, m -1 ), where m is the width of the NN. We train the NN with loss spikes using the learning rate η = 0.05 while using η = 0.005 for the training without loss spikes. The training dataset is obtained by sampling 10 points equidistantly in the [-12, 12] interval.\nFor Fig. 8, we use the two-layer ReLU FNN with a width of 500 to fit the target function using full-batch gradient descent as follows,\nf (x) = 1 2 ReLU(-x - 1 3 ) + 1 2 ReLU(x -1 3\n).\nThe initialization of the parameters θ ∼ N (0, m -0.4 ), where m is the width of the NN. We train the NN with loss spikes using the learning rate η = 0.05 while using η = 0.0005 for the training without loss spikes. The training dataset is obtained by sampling 6 points equidistantly in the [-5/3, 5/3] interval."
},
{
"figure_ref": [],
"heading": "B Experimental results",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [
"fig_9",
"fig_8"
],
"heading": "B.1 Loss spikes facilitate condensation on ReLU NNs",
"publication_ref": [],
"table_ref": [],
"text": "In this subsection, We verify that the ReLU network facilitates condensation with loss spikes shown in Fig. 8, similar to the situation in the tanh NNs shown in Fig. 7 in the main text. We only plot the neurons with non-zero output value in the data interval [x 1 , x n ] in the situation in the ReLU NNs.\nFor the neurons with constant zero output value in the data interval, they will not affect the training process and the NN's output. "
},
{
"figure_ref": [
"fig_8",
"fig_11"
],
"heading": "B.2 Detailed Features of Tanh NNs",
"publication_ref": [],
"table_ref": [],
"text": "In order to eliminate the influence of the inhomogeneity of the tanh activation function on the parameter features of Fig. 7, we plot the normalized scatter figures between a j , w j and the orientation, as shown in Fig. 9. Obviously, for the network with loss spikes, both the input weight and the output weight have weight condensation, while the network without loss spikes does not have weight condensation. "
},
{
"figure_ref": [],
"heading": "Acknowledgments",
"publication_ref": [],
"table_ref": [],
"text": "This work is sponsored by the National Key R&D Program of China Grant No. 2022YFA1008200, the Shanghai Sailing Program, the Natural Science Foundation of Shanghai Grant No. 20ZR1429000, the National Natural Science Foundation of China Grant No. 62002221, Shanghai Municipal of Science"
},
{
"figure_ref": [],
"heading": "A Experimental setups",
"publication_ref": [],
"table_ref": [],
"text": "For Fig. 2(a-c), Fig. 3(a), Fig. 4, we use the two-layer tanh FNN with a width of 20 to fit the target function using full-batch gradient descent as follows, f (x) = sin(x) + sin(4x).\nThe initialization of the parameters θ ∼ N (0, m -1 ), where m is the width of the NN, and the learning rate η = 0.05. For Fig. 2(a), the λ max is calculated every 100 epochs. Fig. 2(c) and Fig. 3(a) show the parameter trajectories of different epoch intervals, which are indicated on the label of the color bar. For Fig. 4, the θ max is selected at epoch 114320, and the θ end is selected at epoch 114400.\nFor Fig. 2(d-f), we use the two-layer ReLU CNN with a Max Pooling layer behind the activation function for the CIFAR10-1k classification problem, i.e., using the first 1000 training data of the CIFAR10 as the training data. The number of the convolution kernels is 16 and the size is 3 × 3. We use the MSE as the loss function with learning rate η = 0.1.\nFor Fig. 3, we use the following quadratic model as the toy model to illustrate the SLAS structure,\nwhere (x, y) ∈ (-4, +∞) × R. The training uses the gradient descent algorithm with learning rate η = 5 × 10 -3 and the initial value (x, y) = (0.5, 0.00001).\nFor Fig. 5(a,b) and Fig. 6(a), we use the two-layer tanh FNN with a width of 500 to fit the target function using full-batch gradient descent as follows,\nThe initialization of the parameters θ ∼ N (0, m -0.4 ), where m is the width of the NN, and the learning rate η = 0.001. The training dataset is obtained by sampling 15 points equidistantly in the [-12, 12] interval, and the test dataset is obtained by sampling 14 points equidistantly in the [-11.14, 11.14] interval, which is approximately the midpoint of the pairwise data of the training set.\nFor Fig. 6(b-c), we use the CNNs for the CIFAR10-1k classification problem with structures shown in Table 1-2, respectively. We use ReLU as the activation function, added behind each convolutional layer. We use the Xavier initialization and the MSE loss function. The learning rate is 0.005. For Fig. 6(d), we use the CNNs for the CIFAR10-2k classification problem with structures shown in Table 3. We use ReLU as the activation function, added behind each convolutional layer. We use the Xavier initialization and the cross-entropy loss function. The learning rate is 0.01. The large batch size we used is 1000, while the small one is 32. "
}
] | In this work, we study the mechanism underlying loss spikes observed during neural network training. When the training enters a region, which has a smaller-loss-assharper (SLAS) structure, the training becomes unstable and loss exponentially increases once it is too sharp, i.e., the rapid ascent of the loss spike. The training becomes stable when it finds a flat region. The deviation in the first eigen direction (with maximum eigenvalue of the loss Hessian (λ max ) is found to be dominated by low-frequency. Since low-frequency is captured very fast (frequency principle), the rapid descent is then observed. Inspired by our analysis of loss spikes, we revisit the link between λ max flatness and generalization. For real datasets, low-frequency is often dominant and well-captured by both the training data and the test data. Then, a solution with good generalization and a solution with bad generalization can both learn low-frequency well, thus, they have little difference in the sharpest direction. Therefore, although λ max can indicate the sharpness of the loss landscape, deviation in its corresponding eigen direction is not responsible for the generalization difference. We also find that loss spikes can facilitate condensation, i.e., input weights evolve towards the same, which may be the underlying mechanism for why the loss spike improves generalization, rather than simply controlling the value of λ max . | Loss Spike in Training Neural Networks | [
{
"figure_caption": "Figure 2: (a, d) The loss value (black) and λ max (red) vs. training epoch, where the λ max is calculated every 100 epoch. (b, e) The loss value and λ max of a specific epoch interval, which is marked green in (a, d), respectively. (c, f) The loss surface and the trajectory of the model parameters along the first two PCA directions. (a, b, c) Two-layer tanh NN with width 20. The sum of the explained variance ratios of the first two PCA directions is 0.9895. (d, e, f) Two-layer ReLU CNN with Max Pooling. The sum of the explained variance ratios of the first two PCA directions is 0.9882.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3: (a) The loss surface and the trajectory of the model parameters along the first two PCA directions in the EoS stage. (b) Schematic illustration of SLAS structure. (c) The loss value and the maximum eigenvalue of the Hessian matrix of a loss spike process of the toy model. (d) The loss surface and the GD trajectory of the two-dimensional parameters of the toy model.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: Low-frequency proportion for different low-frequency thresholds. The NN we used is a two-layer tanh NN with width 20. For the random output difference, we calculate the mean value and the error bar with 100 random samples.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 :5Figure5: Two-layer tanh FNN with a width of 500. (a) The variation of the test loss with the eigenvalue index i when eliminating the difference between θ train and θ test in the first i eigen directions. (b) The output difference before and after moving θ train in the first nine eigen directions of its Hessian matrix. Each subset corresponds to the case of one eigen direction.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Figure 6 :6Figure 6: Blue bar: (a, b, c) show the projection values of in each eigen direction of H(θ train ) for θ test -θ train , and (d) for θ large -θ small . Orange line: the sum of the first n eigenvalues over all eigenvalues. (a) Two-layer tanh FNN for the one-dimensional fitting problem. (b) Two-layer ReLU CNN with Max Pooling for the CIFAR10 classification problem. (c) Three-layer ReLU CNN with Max Pooling for the CIFAR10 classification problem. (d) Five-layer ReLU CNN with Max Pooling for the CIFAR10 classification problem.",
"figure_data": "",
"figure_id": "fig_7",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "Figure 7 :7Figure 7: Comparison of two-layer tanh NNs with identical initialization but different learning rates η. The loss spike occurs at a large learning rate (orange), while not at a small learning rate (blue). Left: loss vs. epoch. The small picture in the upper right corner shows the occurrence of the loss spike in more detail. Middle: output. Right: The weight feature distribution of the trained models and the initial one.",
"figure_data": "",
"figure_id": "fig_8",
"figure_label": "7",
"figure_type": "figure"
},
{
"figure_caption": "Figure 8 :8Figure8: Comparison of two-layer ReLU NNs with the same initialization at different learning rates η. The loss spike occurs at a large learning rate, while does not occur at a small learning rate. Left: The loss value under different learning rates, η = 5 × 10 -4 (blue) and η = 5 × 10 -2 (orange). The small picture in the upper right corner shows the occurrence of the loss spike in more detail. Middle: The output of the model trained under different learning rates, η = 5 × 10 -4 (blue) and η = 5 × 10 -2 (orange). The black points are the target points. Right: The feature of the model trained under different learning rates, η = 5 × 10 -4 (blue) and η = 5 × 10 -2 (orange) and the initialization (green).",
"figure_data": "",
"figure_id": "fig_9",
"figure_label": "8",
"figure_type": "figure"
},
{
"figure_caption": "η = 5 × 10 -2",
"figure_data": "",
"figure_id": "fig_10",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 9 :9Figure9: The normalized scatter diagrams between a j , w j and the orientation of tanh NNs for the initialization parameters and the parameters trained with and without loss spikes. Blue dots and orange dots are the output weight distribution and the input weight distribution, respectively.",
"figure_data": "",
"figure_id": "fig_11",
"figure_label": "9",
"figure_type": "figure"
},
{
"figure_caption": "and Technology Major Project No. 2021SHZDZX0102, and the HPC of School of Mathematical Sciences and the Student Innovation Center, and the Siyuan-1 cluster supported by the Center for High Performance Computing at Shanghai Jiao Tong University.",
"figure_data": "",
"figure_id": "tab_0",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "The architecture of the five-layer CNN used in Fig.6(d).",
"figure_data": "LayerOutput sizeinput32 × 32 × 33 × 3 × 16, conv32 × 32 × 162 × 2, maxpool16 × 16 × 163 × 3 × 32, conv16 × 16 × 322 × 2, maxpool8 × 8 × 323 × 3 × 64, conv8 × 8 × 642 × 2, maxpool4 × 4 × 64flatten10242048 → 500, linear500500 → 10, linear10",
"figure_id": "tab_1",
"figure_label": "3",
"figure_type": "table"
}
] | Zhongwang Zhang; Zhi-Qin John Xu | [
{
"authors": "L Wu; C Ma; W E ",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b0",
"title": "How sgd selects the global minima in over-parameterized learning: A dynamical stability perspective",
"year": "2018"
},
{
"authors": "J M Cohen; S Kaur; Y Li; J Z Kolter; A Talwalkar",
"journal": "",
"ref_id": "b1",
"title": "Gradient descent on neural networks typically occurs at the edge of stability",
"year": "2021"
},
{
"authors": "S Arora; Z Li; A Panigrahi",
"journal": "PMLR",
"ref_id": "b2",
"title": "Understanding gradient descent on the edge of stability in deep learning",
"year": "2022"
},
{
"authors": "N S Keskar; D Mudigere; J Nocedal; M Smelyanskiy; P T P Tang",
"journal": "",
"ref_id": "b3",
"title": "On large-batch training for deep learning: Generalization gap and sharp minima",
"year": "2016"
},
{
"authors": "S Hochreiter; J Schmidhuber",
"journal": "Neural computation",
"ref_id": "b4",
"title": "Flat minima",
"year": "1997"
},
{
"authors": "Z.-Q J Xu; Y Zhang; Y Xiao",
"journal": "Springer",
"ref_id": "b5",
"title": "Training behavior of deep neural network in frequency domain",
"year": "2019"
},
{
"authors": "Z.-Q J Xu; Y Zhang; T Luo; Y Xiao; Z Ma",
"journal": "Communications in Computational Physics",
"ref_id": "b6",
"title": "Frequency principle: Fourier analysis sheds light on deep neural networks",
"year": "2020"
},
{
"authors": "F He; T Liu; D Tao",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b7",
"title": "Control batch size and learning rate to generalize well: Theoretical and empirical evidence",
"year": "2019"
},
{
"authors": "S Jastrzebski; Z Kenton; D Arpit; N Ballas; A Fischer; Y Bengio; A Storkey",
"journal": "",
"ref_id": "b8",
"title": "Three factors influencing minima in sgd",
"year": "2017"
},
{
"authors": "C Xing; D Arpit; C Tsirigotis; Y Bengio",
"journal": "",
"ref_id": "b9",
"title": "A walk with sgd",
"year": "2018"
},
{
"authors": "K Ahn; J Zhang; S Sra",
"journal": "PMLR",
"ref_id": "b10",
"title": "Understanding the unstable convergence of gradient descent",
"year": "2022"
},
{
"authors": "K Lyu; Z Li; S Arora",
"journal": "",
"ref_id": "b11",
"title": "Understanding the generalization benefit of normalization layers: Sharpness reduction",
"year": "2022"
},
{
"authors": "Z Wang; Z Li; J Li",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b12",
"title": "Analyzing sharpness along gd trajectory: Progressive sharpening and edge of stability",
"year": "2022"
},
{
"authors": "A Lewkowycz; Y Bahri; E Dyer; J Sohl-Dickstein; G Gur-Ari",
"journal": "",
"ref_id": "b13",
"title": "The large learning rate phase of deep learning: the catapult mechanism",
"year": "2020"
},
{
"authors": "A Damian; E Nichani; J D Lee",
"journal": "",
"ref_id": "b14",
"title": "Self-stabilization: The implicit bias of gradient descent at the edge of stability",
"year": "2022"
},
{
"authors": "C Ma; D Kunin; L Wu; L Ying",
"journal": "Journal of Machine Learning",
"ref_id": "b15",
"title": "Beyond the quadratic approximation: The multiscale structure of neural network loss landscapes",
"year": "2022"
},
{
"authors": "A Agarwala; F Pedregosa; J Pennington",
"journal": "",
"ref_id": "b16",
"title": "Second-order regression models exhibit progressive sharpening to the edge of stability",
"year": "2022"
},
{
"authors": "C Ma; L Wu; E Weinan",
"journal": "PMLR",
"ref_id": "b17",
"title": "A qualitative study of the dynamic behavior for adaptive gradient algorithms",
"year": "2022"
},
{
"authors": "L Wu; Z Zhu",
"journal": "",
"ref_id": "b18",
"title": "Towards understanding generalization of deep learning: Perspective of loss landscapes",
"year": "2017"
},
{
"authors": "C Ma; L Ying",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b19",
"title": "On linear stability of sgd and input-smoothness of neural networks",
"year": "2021"
},
{
"authors": "P Izmailov; A Wilson; D Podoprikhin; D Vetrov; T Garipov",
"journal": "",
"ref_id": "b20",
"title": "Averaging weights leads to wider optima and better generalization",
"year": "2018"
},
{
"authors": "P Chaudhari; A Choromanska; S Soatto; Y Lecun; C Baldassi; C Borgs; J Chayes; L Sagun; R Zecchina",
"journal": "Journal of Statistical Mechanics: Theory and Experiment",
"ref_id": "b21",
"title": "Entropy-sgd: Biasing gradient descent into wide valleys",
"year": "2019"
},
{
"authors": "T Lin; S U Stich; K K Patel; M Jaggi",
"journal": "",
"ref_id": "b22",
"title": "Don't use large mini-batches, use local sgd",
"year": "2018"
},
{
"authors": "Y Zheng; R Zhang; Y Mao",
"journal": "",
"ref_id": "b23",
"title": "Regularizing neural networks via adversarial model perturbation",
"year": "2021"
},
{
"authors": "P Foret; A Kleiner; H Mobahi; B Neyshabur",
"journal": "",
"ref_id": "b24",
"title": "Sharpness-aware minimization for efficiently improving generalization",
"year": "2020"
},
{
"authors": "L Dinh; R Pascanu; S Bengio; Y Bengio",
"journal": "PMLR",
"ref_id": "b25",
"title": "Sharp minima can generalize for deep nets",
"year": "2017"
},
{
"authors": "T Luo; Z.-Q J Xu; Z Ma; Y Zhang",
"journal": "Journal of Machine Learning Research",
"ref_id": "b26",
"title": "Phase diagram for two-layer relu neural networks at infinitewidth limit",
"year": "2021"
},
{
"authors": "H Zhou; Q Zhou; Z Jin; T Luo; Y Zhang; Z.-Q J Xu",
"journal": "",
"ref_id": "b27",
"title": "Empirical phase diagram for three-layer neural networks with infinite width",
"year": "2022"
},
{
"authors": "A Jacot; C Hongler; F Gabriel",
"journal": "",
"ref_id": "b28",
"title": "Neural tangent kernel: Convergence and generalization in neural networks",
"year": "2018"
},
{
"authors": "S Arora; S S Du; W Hu; Z Li; R R Salakhutdinov; R Wang",
"journal": "Advances in neural information processing systems",
"ref_id": "b29",
"title": "On exact computation with an infinitely wide neural net",
"year": "2019"
},
{
"authors": "L Breiman",
"journal": "The Mathematics of Generalization",
"ref_id": "b30",
"title": "Reflections after refereeing papers for nips",
"year": "1995"
},
{
"authors": "C Zhang; S Bengio; M Hardt; B Recht; O Vinyals",
"journal": "Communications of the ACM",
"ref_id": "b31",
"title": "Understanding deep learning (still) requires rethinking generalization",
"year": "2021"
},
{
"authors": "H Zhou; Q Zhou; T Luo; Y Zhang; Z.-Q J Xu",
"journal": "",
"ref_id": "b32",
"title": "Towards understanding the condensation of neural networks at initial training",
"year": "2021"
},
{
"authors": "Z Chen; Y Li; T Luo; Z Zhou; Z.-Q J Xu",
"journal": "",
"ref_id": "b33",
"title": "Phase diagram of initial condensation for two-layer neural networks",
"year": "2023"
},
{
"authors": "H Maennel; O Bousquet; S Gelly",
"journal": "",
"ref_id": "b34",
"title": "Gradient descent quantizes relu network features",
"year": "2018"
},
{
"authors": "F Pellegrini; G Biroli",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b35",
"title": "An analytic theory of shallow networks dynamics for hinge loss classification",
"year": "2020"
},
{
"authors": "M Andriushchenko; A Varre; L Pillaud-Vivien; N Flammarion",
"journal": "",
"ref_id": "b36",
"title": "Sgd with large step sizes learns sparse features",
"year": "2022"
},
{
"authors": "Z J Xu; H Zhou",
"journal": "",
"ref_id": "b37",
"title": "Deep frequency principle towards understanding why deeper learning is faster",
"year": "2021"
},
{
"authors": "N Rahaman; D Arpit; A Baratin; F Draxler; M Lin; F A Hamprecht; Y Bengio; A Courville",
"journal": "",
"ref_id": "b38",
"title": "On the spectral bias of deep neural networks",
"year": "2019"
},
{
"authors": "T Luo; Z Ma; Z.-Q J Xu; Y Zhang",
"journal": "CSIAM Transactions on Applied Mathematics",
"ref_id": "b39",
"title": "Theory of the frequency principle for general deep neural networks",
"year": "2021"
},
{
"authors": "Z.-Q J Xu; Y Zhang; T Luo",
"journal": "",
"ref_id": "b40",
"title": "Overview frequency principle/spectral bias in deep learning",
"year": "2022"
},
{
"authors": "Z Liu; W Cai; Z.-Q J Xu",
"journal": "Communications in Computational Physics",
"ref_id": "b41",
"title": "Multi-scale deep neural network (mscalednn) for solving poissonboltzmann equation in complex domains",
"year": "2020"
},
{
"authors": "A D Jagtap; K Kawaguchi; G E Karniadakis",
"journal": "Journal of Computational Physics",
"ref_id": "b42",
"title": "Adaptive activation functions accelerate convergence in deep and physics-informed neural networks",
"year": "2020"
},
{
"authors": "S Biland; V C Azevedo; B Kim; B Solenthaler",
"journal": "",
"ref_id": "b43",
"title": "Frequency-aware reconstruction of fluid simulations with generative networks",
"year": "2019"
},
{
"authors": "Y Zhang; T Luo; Z Ma; Z.-Q J Xu",
"journal": "Chinese Physics Letters",
"ref_id": "b44",
"title": "A linear frequency principle model to understand the absence of overfitting in neural networks",
"year": "2021"
},
{
"authors": "B Ronen; D Jacobs; Y Kasten; S Kritchman",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b45",
"title": "The convergence rate of neural networks for learned functions of different frequencies",
"year": "2019"
},
{
"authors": "H Li; Z Xu; G Taylor; C Studer; T Goldstein",
"journal": "",
"ref_id": "b46",
"title": "Visualizing the loss landscape of neural nets",
"year": "2017"
},
{
"authors": "J K Cullum; R A Willoughby",
"journal": "SIAM",
"ref_id": "b47",
"title": "Lanczos algorithms for large symmetric eigenvalue computations",
"year": "2002"
},
{
"authors": "Y Zhang; Z Zhang; T Luo; Z J Xu",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b48",
"title": "Embedding principle of loss landscape of deep neural networks",
"year": "2021"
},
{
"authors": "Y Zhang; Y Li; Z Zhang; T Luo; Z.-Q J Xu",
"journal": "Journal of Machine Learning",
"ref_id": "b49",
"title": "Embedding principle: a hierarchical structure of loss landscape of deep neural networks",
"year": "2022"
},
{
"authors": "K Fukumizu; S Yamaguchi; Y -I. Mototake; M Tanaka",
"journal": "Advances in neural information processing systems",
"ref_id": "b50",
"title": "Semi-flat minima and saddle points by embedding neural networks to overparameterization",
"year": "2019"
},
{
"authors": "B Simsek; F Ged; A Jacot; F Spadaro; C Hongler; W Gerstner; J Brea",
"journal": "PMLR",
"ref_id": "b51",
"title": "Geometry of the loss landscape in overparameterized neural networks: Symmetries and invariances",
"year": "2021"
}
] | [
{
"formula_coordinates": [
4,
386.69,
86.04,
100.89,
9.68
],
"formula_id": "formula_0",
"formula_text": "= [θ 1 -θ t , • • • , θ t -θ t ],"
},
{
"formula_coordinates": [
4,
235.81,
711.12,
268.19,
11.03
],
"formula_id": "formula_1",
"formula_text": "f (x, y) = (50x + 200)y 2 -x + 5,(1)"
},
{
"formula_coordinates": [
5,
262.11,
340.52,
88.98,
22.31
],
"formula_id": "formula_2",
"formula_text": "∂f (x, y) ∂x = 50y 2 -1."
},
{
"formula_coordinates": [
5,
276.31,
399.45,
92.72,
36.19
],
"formula_id": "formula_3",
"formula_text": " > 0 if f (x, y) < 9 = 0 if f (x, y) = 9 < 0 if f (x, y) > 9"
},
{
"formula_coordinates": [
6,
245.31,
406.41,
254.82,
30.39
],
"formula_id": "formula_4",
"formula_text": "LFP(K) = k≤K fθ (k) 2 k fθ (k) 2 , (2"
},
{
"formula_coordinates": [
6,
500.13,
417.52,
3.87,
8.64
],
"formula_id": "formula_5",
"formula_text": ")"
},
{
"formula_coordinates": [
7,
198.44,
270.89,
215.12,
33.53
],
"formula_id": "formula_6",
"formula_text": "L(i) = R Stest θ train + i j=1 θ test -θ train , ν j ν j ,"
},
{
"formula_coordinates": [
7,
251.96,
590.3,
62.37,
14.11
],
"formula_id": "formula_7",
"formula_text": "n i=1 λ 2 i / N i=1 λ 2"
},
{
"formula_coordinates": [
8,
214.66,
95.28,
246.44,
143.58
],
"formula_id": "formula_8",
"formula_text": "(a) FNN i = 1 i = 2 i = 3 i = 4 i = 5 i = 6 i = 7 i = 8 i = 9 (b) output difference"
},
{
"formula_coordinates": [
8,
144.11,
324.97,
60.71,
63.81
],
"formula_id": "formula_9",
"formula_text": "n i = 1 2 i / N i = 1 2 i (a) FNN"
},
{
"formula_coordinates": [
8,
221.51,
324.97,
80.85,
63.81
],
"formula_id": "formula_10",
"formula_text": "n i = 1 2 i / N i = 1 2 i (b) Two-layer CNN"
},
{
"formula_coordinates": [
8,
316.65,
324.5,
83.24,
64.28
],
"formula_id": "formula_11",
"formula_text": "n i = 1 2 i / N i = 1 2 i (c) Three-layer CNN"
},
{
"formula_coordinates": [
8,
408.44,
324.35,
88.99,
64.43
],
"formula_id": "formula_12",
"formula_text": "n i = 1 2 i / N i = 1 2 i (d) Different batch sizes"
},
{
"formula_coordinates": [
14,
231.66,
265.79,
148.67,
8.74
],
"formula_id": "formula_13",
"formula_text": "f (x) = tanh(x -6) + tanh(x + 6)."
},
{
"formula_coordinates": [
14,
212.49,
360.25,
179.18,
22.31
],
"formula_id": "formula_14",
"formula_text": "f (x) = 1 2 ReLU(-x - 1 3 ) + 1 2 ReLU(x -1 3"
}
] |
10.1016/j.pragma.2010.07.019 | 2023-05-23 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b30",
"b10",
"b29",
"b11",
"b56",
"b35",
"b56",
"b11",
"b27",
"b53",
"b20",
"b49",
"b11",
"b36",
"b56"
],
"table_ref": [],
"text": "Chinese political discourse has long been an important research topic for interpretation and translation studies. Recent years have witnessed a range of publications from these fields (Gu, 2019a;Liao and Pan, 2018;Fu and Chen, 2019;Li and Zhang, 2020;Fu and Wang, 2022), with most attention being drawn to a series of official documents and government press conferences, due to their paramount significance in China's diplomatic exchanges with the international society.\nAmong all the officially-released, publicavailable written materials, the reports on the work of the government (RWG) are particularly noteworthy, because of their evaluative role on the gov-ernment's performance in the past year and their guidance on China's future development (Zhang, 2004). The report is delivered by the Premier during the annual National People's Congress (NPC), the highest-level national conference in China, covering topics including the government's works and accomplishments, schedules and targets for the current year, as well as practices of government administration. Generally, it takes the joint efforts of several departments (including the Department of Translation and Interpreting, the Foreign Languages Publishing Administration, and the Xinhua News Agency) led by The Central Compilation and Translation Bureau (CCTB) to translate this report into various languages under unified and strict working processes (Pan and Li, 2021). However, the English version of the RWG may fail to be well-received by the target readership in some cases due to its \"bureaucratic airs\" (\"官气\", \"guan xi\", Zhang, 2004), which might be attributed to the differences between the two divergent cultural and linguistic systems (Fu and Wang, 2022).\nOne of the significant differences between the Chinese and English language is the use of hedges. Lakoff (1973) defines hedges as a group of devices to make the propositional content \"fuzzier or less fuzzy\" (195), but further studies indicate that hedges are in fact multifunctional and assume diversified pragmatic and semantic roles, making hedges a widespread linguistic phenomenon. The flexibility and practicality offered by hedges make them an important part of political and diplomatic discourse. Previous studies suggest that compared with English, the Chinese language is relatively \"underhedged\" across genres (Yang, 2013;Hu and Cao, 2011;Wang and Zhou, 2009;Fu and Wang, 2022). Given this, hedges can serve as a window to examine and compare the linguistic differences between the Chinese and English political language, as well as the distinction between En-glish translation of Chinese official documents and non-translated English ones.\nAlso needed to be mentioned is that the Chinese government has shifted the priority of its outward translation from being faithfulness-oriented towards \"target-oriented\" to improve the reception of Chinese diplomatic discourse, and to help the international community better understand the Chinese policies and practices (Pan and Wang, 2021;Zhang, 2004). Guided by this shifting policy orientation and changing translation norms, it is safe to presume that translators responsible for translating the government's pollical documents to the international audience, though still restricted by a set of fixed working mechanisms, may play a more active role in connecting China with the global community. However, few studies have been done to testify this assumption, and no research has chosen hedges as a particular focus to measure the degree of change. Therefore, the aim of the present study is three-pronged: 1) to examine the diachronic change of hedges being used in the English translations of RWG, and to what extent the differences are attributed to the ST, and 2) to categorize the common translation strategies adopted by translators to deal with the hedging devices.\n2 Literature Review"
},
{
"figure_ref": [],
"heading": "Hedges and their Pragmatics Uses",
"publication_ref": [
"b27",
"b4",
"b3",
"b32",
"b0",
"b20",
"b52",
"b54",
"b37",
"b11",
"b31",
"b46",
"b54",
"b5",
"b22",
"b38",
"b38",
"b40",
"b41",
"b25",
"b34",
"b37"
],
"table_ref": [],
"text": "Hedges are commonly understood as a set of lexical or non-lexical devices used to indicate fuzziness, uncertainty and probability (Lakoff, 1973;Brown and Levinson, 1987;Biber, 1988). As hedges are widespread across languages and can serve multiple functions, scholars have conducted various studies to explore their use in both spoken language (Magnifico and Defrancq, 2017) and written texts (Abdollahzadeh, 2011;Hu and Cao, 2011;Yang and Yap, 2015;Yang and Li, 2022), and across disciplines ranging from politics (Ponterotto, 2018;Fu and Wang, 2022), education (Liu and Tseng, 2021), finance and management (Vázquez, 2010;Yang and Li, 2022), as well as law (Chaemsaithong, 2017).\nThe concept of hedges can be traced to Lakoff (1973, p.195), who defined the term as \"words whose job is to make things fuzzier or less fuzzy\". He offered a detailed though not exhaustive analysis of their use, and noted that hedges often appeared before a noun phrase or predicate to blur the boundaries between concepts. Although Lakeoff mainly addresses the semantic usage of hedges, later researches tend to approach them from a functional view, establishing connections between hedges and pragmatic functions in different contexts. Brown and Levinson (1987, p.70) discussed how hedges can intervene in face-threatening acts to protect the speaker's positive face and negative face. In a similar vein, Holmes (1990, p.185) regard hedges as a speech strategy in line with the politeness principle. (Hyland, 2005) treated hedges as being part of a broader and systematic framework. To be specific, he proposed an interpersonal metadiscourse model consisted of two subgroups-\"interactive\" and \"interactional\", and put hedges into the \"interactional\" subgroup, whose purpose is to \"involve the reader in the text\" (49). In his view, hedges demonstrate the language user's cautiousness against absolute proposition, and serve to modify attitudes toward the truth value of propositional message.\nHowever, there has been no consensus in terms of what should be counted as hedges, and how to classify them. Prince et al. (1982) put forward a classification of hedges comprising two major categories (approximators and shields), treating hedges as both semantic and pragmatic devices. Following (Prince et al., 1982), Salager-Meyer (1994) further subdivided hedges into shields, approximators, emotionally-charged intensifiers (e.g. \"particularly encouraging\", \"unexpectedly\", \"extremely difficult\"), and compound hedges (e.g. \"it may suggest that\", \"it would seem likely that\"). Mainly focusing on grammatical features, Hyland (1998) proposed a model of hedges consisted of four categories: model verb; epistemic lexical verb; epistemic adjective, adverb, and noun; phraseological expression. It seems that there is also distinction between hedges often used in spoken language and those seen in written texts. Researchers have identified a set of hedges specific to oral discourse. For example, Schiffrin (1987) regarded the word \"well\" as a discourse marker to introduce new topics, leave time to pause, provide answers to questions, and express disagreement. Jucker and Smith (1998) examined how the hedge \"you know\" contributed to establishing connection with the speech receiver and encouraging positive reception of the information offered by the speech giver.\nMoreover, some scholars mentioned that in a broader sense, hedging could also be a discourse strategy, not limited to a set of specific words or phrases. Martín (2022) identified four basic hedging strategies, including Indetermination, Camouflage, Subjectivization and Depersonalization. Similarly, Ponterotto (2018) viewed hedging as referring to \"virtually any discourse strategy aimed at avoiding explicit position when ascertaining the truth conditions of reported facts or events\" (178)."
},
{
"figure_ref": [],
"heading": "Intercultural Studies of Hedges",
"publication_ref": [
"b52",
"b0",
"b42",
"b22",
"b53"
],
"table_ref": [],
"text": "It is noteworthy that research on hedges has attracted an increasing number of scholars from diversified cultural background in recent years, which greatly expands the scope of this field and offers fresh understanding into the various uses of hedges in different socio-cultural contexts. For example, Yang and Yap (2015) conducted a case study that taps into a special Chinese hedge \"kongpa\". Based on Mandarin broadcast transcripts, the study observed that although \"kongpa\" roughly equates to \"being afraid\" in English, it was rarely linked with fear or cautiousness in spoken Chinese. Rather, the word was often used neutrally or even positively for interpersonal purposes, to protect face of the speaker or the hearer, to avoid self-praising, or to criticize others in a euphemistic manner.\nAlso noticeable is that much attention has been drawn to comparative studies of hedges from a cross-cultural perspective. For instance, Hu and Cao (2011) adopted a corpus-based approach to instigate the similarities and distinctions of hedging patterns between English and Chinese academic abstracts. Results of their empirical study showed that the frequency of hedges in Englishmedium abstracts was significantly more than that in Chinese-medium abstracts, and the authors attributed this difference to divergent rhetorical conventions of the two languages, as well as different conceptions of scientific publication. Abdollahzadeh (2011) examined the use of hedges in academic papers written in English by Anglophone speakers and Iranian scholars. Based on 60 conclusion parts extracted from two groups of articles, the author found that hedges appear more frequently in papers written by Anglophone speakers. Shafqat et al. (2019) focused on journalistic English, conducting a comparative study of hedges in a Pskistani English newspaper and an European English newspaper according to the classification of Hyland (2005). They used statistical methods to test their assumptions about the use of hedges in the two groups of newspaper, which are supported by their empirical results: hedges of all types appear significantly more frequent in the European English newspaper than in the Pakistani one.\nThese intercultural studies offer great insight into the differentiated use of hedges in multiple cultures and in various disciplines. However, the bulk of literature has been centered on a few dominant languages, especially English and Spanish, leaving ample space to conduct research on less dominant languages. Also, as comparative studies in this field have gained momentum, results from many studies show that hedges tend to be used more frequently in English than in other languages (Yang, 2013). Nevertheless, few studies provide detailed explanation regarding why this phenomenon exists. As language is not merely a linguistic product, more studies are awaited to dissect the different use of hedges from a sociocultural perspective."
},
{
"figure_ref": [],
"heading": "Translation of Hedges in Political Settings",
"publication_ref": [
"b14",
"b9",
"b37",
"b32",
"b11",
"b41"
],
"table_ref": [],
"text": "Political discourse, whether in spoken or written form, is generally planned, formal, well-prepared, and strategically-expressed in advance (Gribanova and Gaidukova, 2019), because they are open to close scrutiny by both domestic and international audience. Because of this, proper deployment of hedges is often a necessity in political discourse to both satisfy the expectation of the target audience and protect the image and interest of the massage senders.\nAlthough hedges seem to have a natural connection with politics, research on the use of hedging devices in political texts is relatively limited. Among the few related studies, president's speeches have attracted most attention. For example, Fraser (2010) was particularly interested in how politicians resorted to hedging when facing challenging questions from journalists during press conference. He chose the former US president Bush's 2007 Press Conference as his source of research material. Another study conducted by Ponterotto (2018) analyzed hedging strategies adopted by another former president Obama in his political interviews. The author suggested that Obama used various hedging moves and strategies to evade explicit answers on sensitive issues, and his artful use of hedges contributed to a cautious speech style.\nTranslational research focusing on hedges is even fewer. Because of the small number of studies in this specific field, the following review is not strictly confined to translation studies, but includes interpreting studies as well. In Magnifico and Defrancq (2017), the authors drew on speeches and their corresponding interpretation transcripts at the European Parliament to examine whether gender differences existed in the linguistic output during spontaneous interpretations. The authors found that female interpreters used significantly more hedging devices than their male counterparts. Fu and Wang (2022) compared the frequency of hedges in interpreted Chinese political press briefings with that in spontaneous English speeches, and found that interpreted speeches contained significantly fewer hedges, and that the range of hedging devices were less diversified than speeches made by Anglophone speakers in similar political settings. Schiffrin (1987) investigated how hedges were translated from English to German in Tony Blair's speech during the 1995 Labour Party Conference. The author identified three translation strategies to render hedges in a different cultural context, and found that hedging devices could modify the precision of discourse, facilitate interpersonal exchanges with other politicians, show politeness, and build a positive self-image, fulfilling both semantic and pragmatic functions. In a newly published study, M.Aljawadi (2022) examined the English-Arabic translation of hedges in the former US president Donald Trump's public speeches and interviews related to the Covid-19 pandemic. Based on Fraser's classifications, the study found that hedges were most frequently used in the context of negation.\nThe previous literature of hedges from a translational perspective sheds light on the contributions of hedging devices in the cross-cultural delivery of political texts. Nevertheless, all of these studies only addressed single-direction translation of hedges (e.g. Chinese-to-English, English-to-German). No study has been done to deal with their bi-directional translation in similar political settings. In other words, there has not been a complete picture of how hedges are used to communicate with the domestic and international audience, and whether differences exist between the SL-TL and TL-SL translation of hedging devices, leaving a gap to be filled."
},
{
"figure_ref": [],
"heading": "Data and Methodology",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Corpus Data",
"publication_ref": [],
"table_ref": [],
"text": "This study adopts a corpus-based approach, comprising three parallel sub-corpora:\n(1) the original reports on the work of the government and their English translations from 2000 to 2004 (RWG1);\n(2) the original reports on the work of the government and their English translations from 2018 to 2022 (RWG2);\n(3) the English United Nations annual reports and their Chinese translations from 2018 to 2022 (UNAR). All the textual data from the first two sub-corpora are directly downloaded from the official website of China's state council, which provides documents both in Chinese1 , and in English2 . Data in the third sub-corpus come from the United Nations' official website3 , on which annual reports for the latest five years are available in English, Chinese and other official languages.\nThe UN annual reports, delivered by the general secretary during the general assembly, are highly similar to the RWGs in both register and topics, and thus are used as the comparable corpus in this study. Reports on the work of the government are the most comprehensive record of works done during the past year, cover a wide range of topics of national importance, including economic development, social welfares, ecological civilization, political reforms, diplomatic affairs, and cultural undertakings. Similarly, the UN annual reports are usually delivered by the general secretary during the general assembly, recording all the major achievements accomplished in the established fields of priority, including promotion of sustainable economic growth, peace keeping and security, human rights, as well as law and justice.\nSince the downloaded UN annual reports are pdf files, they are first transformed into machinereadable txt files. Following that, EmEditor is used to re-arrange the format, eliminate extra spaces between words, delete blank lines, and replace wrong punctuations. Next, ABBYY Aligner 2.0 is used to align all the original texts and their translations. After the automatic alignment by the software, manual adjustment is done to ensure that the source texts and translated texts are accurately matched."
},
{
"figure_ref": [],
"heading": "Research Methods",
"publication_ref": [
"b38",
"b38",
"b38"
],
"table_ref": [],
"text": "The taxonomy of hedges in this study is adapted from Prince et al. (1982), which consists of two major categories and four sub-categories, treating hedges as a set of expressions with both semantic and pragmatic functions. In their model, approximators can be divided into adaptors, which are mainly used to downgrade the absoluteness of propositional information (e.g. somewhat), and rounders, which indicate a range or degree around which the proposition can be valid (e.g. about, may). Shields also contain two sub-categories: plausibility shields are used to confirm that the proposition is made by the speaker (e.g. I believe), and attribution shields soften the truth value of a proposition by indicating that the utterance is quoted from others (e.g. \"according to\"). However, since the original taxonomy proposed in Prince et al. (1982) is mainly targeted at the spoken language, it may not fit the purpose of the current study. For this consideration, we drew from both their framework and other previous research to find as many hedges as possible and put them under appropriate category. In addition, as there is no systematic classification of hedges in the Chinese language, we have to first search with all the English hedges and find their Chinese counterparts with the help of our parallel corpora. In this manner, we can find more hedging devices not included in our original taxonomy for both the language pairs. After that, concordance lines containing these hedges will be extracted, and their corresponding expressions in the ST or TT can also be identified via the parallel sub-corpora. Since the extraction process is based on searching for and matching hedges according to the taxonomy without considering the context, some extracted \"hedges\" are not actually used for the purpose of hedging. For example, in \"Development system reform is about becoming much more effective, well-coordinated, transparent and accountable to better assist countries in implementing the 2030 Agenda for Sustainable Development\", \"about\" is not used as a hedge though it is extracted in the first place. These \"false hedges\" will not be considered in our further analysis. Next, the raw and normalized frequency of hedges in the three sub-corpora will be counted, and chi-square testing will be conducted to investigate the differences in the deployment of hedges in both C-E and E-C translations of political texts, and the diachronic patterns of hedges in C-E reports on the work of the government. Fol- it seems, it appears, according to, accordingly, in accordance with, it is said, it is reported, presumably, in line with, as the (old/Chinese) saying goes,based, based on\n似乎(si hu)，好像(hao xiang)，(根)据((gen) ju)，依(照) (yi (zhao)), 据 说(ju shuo)， 据 报 道(ju bao dao)， 表 明(biao ming)， 基 于(ji yu), 考 虑(到) (kao lv (dao))\nTable 2: Taxonomy of hedges adapted and expanded from Prince et al. (1982).\nlowing that, translation strategies will be categorized, and textual examples will be discussed in details to analyze how hedging devices in the ST are expressed in TT."
},
{
"figure_ref": [],
"heading": "Data Analysis",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we present major results of our data analysis showing the differences in the use of hedges both between the language pairs synchronically and within each language across time. To be specific, raw and normalized frequencies of hedges in the ST and TT in RWG1, RWG2, and UNAR will be counted respectively and then compared to see if the translation process brings about significant difference in the frequencies of hedging devices in the TT, and whether directionality influences the occurrences of hedges. Next, English hedges in the TT of RWG1 will be counted and compared with those in RWG2 to examine whether a diachronic change regarding the use of hedging expressions exists between the two subcorpora. Finally, translation strategies will be identified to reveal how institutional translators deal with the hedging devices."
},
{
"figure_ref": [],
"heading": "Occurrence of Hedges in the Original and Translated Political Texts",
"publication_ref": [
"b46",
"b20",
"b13",
"b53"
],
"table_ref": [
"tab_1"
],
"text": "Raw and normalized frequencies (per 1000 tokens) of hedges in the three sub-corpora are pre-sented in Table 3. As we can see, in both RWG1 and RWG2, hedges occur more frequently in the TT than in the ST, which indicates that a certain portion of hedges not seen in the STs are added in the process of translation. In contrast, a reverse pattern is observed in UNAR, in which hedging devices appear more frequently in the ST. These results suggest that hedges tend to appear more frequently in English than in Chinese, which coincide with the previous studies (Vázquez, 2010;Hu and Cao, 2011;Gong et al., 2021;Yang, 2013).\nAnother interesting finding is that original English political texts tend to contain more hedges than the translated English texts, at least in the case of our three sub-corpora. To illustrate, the ST in UNAR contains significantly more hedges than translated English texts in RWG2 (χ 2 =88.48, df =1, p<0.001), and the difference also amounts to significance between the ST in UNAR and the TT in RWG1(χ 2 =46.70, df =1, p<0.001). The data indicate that there might be a systematic difference between the original English and translated English concerning the use of hedges. Also worth mentioning is that directionality is an important factor for the differences between the STs and TTs regarding the frequencies of hedges. More hedges are seen in C-E translation compared with E-C translation. In RWG1 and RWG2, the ST-TT difference in the number of hedges is 17.2% and 14.7% respectively, while the percentage for UNAR is merely 2.1%. In other words, the Chinese translations of UNAR are more closely aligned with the STs in terms of hedges, while the English translation of hedges in RWG1 and RWG2 are less faithful to the STs ."
},
{
"figure_ref": [],
"heading": "Diachronic Change Regarding the Use of Hedges in the TTs",
"publication_ref": [
"b36",
"b56",
"b32",
"b36",
"b48"
],
"table_ref": [
"tab_2"
],
"text": "To examine whether there are noticeable changes concerning the deployment of hedges in the translations of Chinese political documents through years, we have counted the raw and normalized frequencies of hedging devices according to our taxonomy in the target text in RWG1 and RWG2. On the whole, there are significantly more hedges in RWG2 than in RWG1 (χ 2 =12.25, df =1, p<0.001), which attests the observation that the translation of Chinese political documents has become more target-oriented to improve their reception in the international community (Pan and Wang, 2021;Zhang, 2004). The increased number is partly attributed to the increase of hedges in the ST, or what is called \"source text stimuli\" by Magnifico and Defrancq (2017). This is easily understandable, since \"faithfulness\" has been the constant priority for Chinese institutional translators working for the government (Pan and Wang, 2021). Still, a noticeable portion of the increased hedges are not translations from the corresponding hedges in the ST, and have to be explained by factors other than the ST. In other words, these \"extra\" hedges are added by the translators out of various purposes, and thus can be seen as the result of the translators' subjectivity.\nAs revealed in Table 4, attribution shields are the most dominant hedges across the four categories, followed by adaptors and rounders, while plausibility shields seldom appear in both the two sub-corpora. It is also noted that differences exist in the preferred categories of hedges between RWG1 and RWG2. Specifically, RWG1 displays significantly more adaptors than RWG2 (χ 2 =33.00, df =1, p<0.001), while RWG2 contains nearly three times as many attribution shields as RWG1 (χ 2 =34.35, df =1, p<0.001). The differences in other two categories do not amount to significance. Particularly noteworthy is that plausibility shields are rarely seen in both the two subcorpora. This might be due to their inconsistency with the stylistic features of the Chinese political texts, which are usually formal, firm, and in lack of subjectivity (Wang, 2008)."
},
{
"figure_ref": [],
"heading": "Adaptors",
"publication_ref": [
"b38",
"b3",
"b27",
"b38",
"b2",
"b7",
"b2"
],
"table_ref": [
"tab_3"
],
"text": "Table 5 displays the frequencies of adaptors in the TT of RWG1 and RWG2. \"Some\" and \"can\" are the most dominant devices in both the two subcorpora, which constitute more than two-thirds of all the adaptors. Our results are not aligned with Prince et al. (1982), which identified \"sort of\" and \"kind of\" as the most frequently used adaptors. Such disparity is caused by the difference in the typology of texts: while \"sort of\" and \"kind of\" are widespread in the spoken discourse (Biber, 1988;Lakoff, 1973;Prince et al., 1982), they are less likely to be seen in the official political documents.\nExample 1 (RWG1 03/16/2018) Some enterprises, particularly small and medium ones, are finding it tough going. Growth in private investment is weak; some regions still face considerable downward economic pressure, and risks and potential dangers in the financial and other sectors are not to be ignored. Poverty alleviation remains a formidable task; agriculture is not based on a strong foundation. The disparities in development between rural and urban areas, between regions, and in income distribution remain substantial. Serious and major workplace accidents happen all too often. People still have a lot of complaints about air quality, environmental sanitation, food and drug safety, housing, education, healthcare, employ- ment, and elderly care. The transformation of government functions has not yet reached where it should be. In government work there are places where we fall short. Some reform measures and policies have not been fully implemented. Some officials are weak on awareness that they are there to serve and must uphold the rule of law, and some lack commitment to their work and willingness to bear the weight of responsibility. Bureaucratism and the practice of formalities for formalities' sake exist to varying degrees. There are many complaints from the people and the business sector about the difficulty of accessing government services and the excessive array of charges. In some sectors misconduct and corruption are still a common problem.\nThe frequent occurrence of \"some\" in the TTs is fully reflected in this extract, which contains 6 \"some\" out of 220 words. Most of them follow the \"some + countable nouns/ uncountable noun\" structure, indicating that \"some\" is used to restrict the scope of the mentioned object, making the propositions more generalized. that the distribution is heavily skewed, with only a small portion of attribution shields being frequently deployed, while the others only appear occasionally or not appear at all. Our results give support to \"simplification\" as a translation universal in the target text (Baker, 2019;Chesterman, 2004), which refers to a general pattern that information in the ST is often simplified during the process of translation. The phenomenon of simplification can be observed from the lexical, syntactic, grammatical, as well as other aspects. On the lexical level, the translated text generally features less lexical variety and more occurrence of high-frequency items (Baker, 2019)."
},
{
"figure_ref": [],
"heading": "Attribution Shields",
"publication_ref": [],
"table_ref": [],
"text": "It is also noticeable that although most attribution shields tend to occur at similar frequencies, \"-based\" and \"based on\" appear significantly more times in RWG2 than in RWG1 (χ 2 =46.27, df =1, p<0.001; χ 2 =8.35, df =1, p=0.004). While the \"NOUN-based\" pattern does not appear at all in RWG1, and \"based on\" only appears 10 times, they occur 95 and 64 times respectively in RWG2, representing a marked increase.\nBased on these observations, we hypothesize that the soaring occurrence of the two hedges during the past two decades are at least partly attributed to the language contact with English. pressions on BNC and COCA, and found that they appear at a much higher frequency in these two corpora compared with RWG1 (χ 2 =977.20, df =1, p<0.001), but appear significantly less frequently compared with RWG2 (χ 2 =11358.5, df =1, p<0.001). In other words, although these hedging devices are rarely used two decades ago, they have managed to be included in the lexicon of publicity-oriented English translated from Chinese, and gained increasing significance until their frequencies surpass those of the original English to a large extent.\nIn terms of actual usage, it is easily noticeable that nouns in \"NOUN-based\" expressions tend to be well-known places in both BNC and COCA, such as \"California\", \"London\", \"Massachusetts\", \"New York\", and \"Washington\". They can also be \"computer\" or things related to computers (such as the computer system \"Unix\", and \"web\").\nInterestingly, a close examination of the concordance lines in our RWG2 reveals a different picture. As displayed in on\" tend to co-occur with \"law\", \"market\", \"license\", \"performance\", \"procedure\", \"condition\", \"need\", \"reality\", and so on. The frequent cooccurrence of the two attribution shields with these words can lend support to the propositional contents in officially released government documents, making them more credible, objective, and authoritative. The different usage of \"NOUN-based\" and \"based on\" in the original English and translated English in our corpus indicates that although these two expressions are borrowed from the English world, adaption and variation happen to their actual usage as they enter the C-E translational system in the politics-specific genre."
},
{
"figure_ref": [],
"heading": "Handling of Hedging Devices in Translation",
"publication_ref": [],
"table_ref": [],
"text": "This section aims to gain deeper insight of how translators deal with hedges in the STs and the possible factors behind different handlings through both quantitative analysis and qualitative analysis of several examples between the language pairs. Careful examination of all the aligned concordance lines containing hedges suggests that the majority of hedges are retained and translated faithfully in the TTs. For the cases where hedges are not retained, addition and modification are most frequently adopted. In contrast, the strategy of omission is only used occasionally. Also noteworthy is the relationship between directionality and the proportion of retention. To be specific, for the C-E translation in RWG1 and RWG2, the retention rate of hedges is 83.26% and 85.32% respectively, while in UNAR, the number is up to 97.86%. This indicates that the use of hedges is more strictly in line with the STs for E-C translation, but less strictly for C-E translation, at least in the case of our data."
},
{
"figure_ref": [],
"heading": "Addition",
"publication_ref": [],
"table_ref": [],
"text": "The first identified method to deal with hedges is addition, which refers to that the translator intentionally adds hedging devices in the TT when there are no traces of corresponding hedges in the ST.\nFor the cases where hedges are not retained, addition serves as the dominant method, taking up 9.5% in RWG1 and 7.1% in RWG2. This strategy is generally adopted to restrict the scope of mentioned objects (see example 1) or to lower the absoluteness of the proposition.\nExample 1 (2003 RWG) ST:\n坚 持 不 懈 地 开 展 反 腐 败 斗 争，大力纠正部门和行业不正之风， 依法惩处了一批违法违纪的腐败分 子。\nGloss: We made unremitting efforts to combat corruption, rectify unhealthy tendencies in the departments and trades and punish according to law quite a few corrupt elements.\nTT: We made unremitting efforts to combat corruption, rectify unhealthy tendencies in some departments and trades and punish according to law quite a few corrupt elements.\nExample 1 demonstrates the resolution of the Chinese government to combat corruption and improve discipline in all the government departments and enterprises amid the rampant malpractices among some senior officials and managers in some state-owned enterprises. In fact, several measures are reported in details about how the government aims to cope with corruption in the previous parts of the report. However, widespread corruption is definitely detrimental to China's image in the international society, thus the translator adds the hedge \"some\" before \"departments and trades\" in the TT, to the effect of limiting the corruption issue to only a portion of all the departments and trades."
},
{
"figure_ref": [],
"heading": "Omission",
"publication_ref": [
"b46",
"b20",
"b13",
"b53"
],
"table_ref": [],
"text": "Omission is the least adopted method in RWG1 and RWG2. This might be attributed to the systematic difference between Chinese and English regarding the use of hedges: in general, Chinese tends to be \"under-hedged\" across all genres than English, which is supported by many previous studies (Vázquez, 2010;Hu and Cao, 2011;Gong et al., 2021;Yang, 2013). Therefore, when the political documents originally written in Chinese are to be translated into English, hedges are often added rather than omitted. In contrast, for E-C translation, omission is preferred, which may explain why omission is the only adopted translation method other than retention in the sub-corpus UNAR. As suggested by example 2, the omission of hedges generally serves to strengthen the propositional content.\nExample 2 (2018 UN annual report) ST: The incidence of homicides and violence relating to organized crime remains high in many regions in the world and, when linked to the illicit trafficking of arms and commodities, can derail efforts towards peace, human rights protection and sustainable development.\nTT:\n在世界许多地区，与有组织 犯罪有关的凶杀和暴力行为发生率仍 然很高，如果这些行为与非法贩运武 器和初级商品行为牵扯在一起，会破 坏实现和平、保护人权和可持续发展 的努力。\nBT: The incidence of homicides and violence relating to organized crime remains high in many regions in the world and, when linked to the illicit trafficking of arms and commodities, will derail efforts towards peace, human rights protection and sustainable development.\nThis example expresses the UN's deep concern towards relentless violence, organized crimes, and homicides, which can cause devastating impact for international peace and security. The Chinese government has constantly sided with the UN in human rights protection as well as peacekeeping, and contributed its share in coordinating with the UN to combat violence-related crimes, human trafficking, and illegal smuggling of weapons. Therefore, when translating the ST into Chinese, the translator intentionally omits the hedge \"can\" to transform the possibility into a certainty, which serves to amplify the severe consequences brought by these crimes, and signals the Chinese government's shared stance with the UN."
},
{
"figure_ref": [],
"heading": "Modification",
"publication_ref": [],
"table_ref": [],
"text": "Modification often occurs when the translator intends to raise or lower the definiteness of the statement, with the latter more often the case in RWG1 and RWG2. As suggested by example 3, despite the availability of the closest equivalence, the translators may opt for other hedging devices in the TT.\nExample 3 (2018 report on the work of the government) ST: 今年中央财政投入增加300 亿 元以上，比上年增长20%以上。 Gloss: This year, the central government will appropriate more than 30 billion yuan for this purpose, at least 20% more than last year.\nTT: This year, the central government will appropriate around 30 billion yuan for this purpose, at least 20% more than last year.\nAs illustrated in example 3, the central government plans to invest a large sum of fund into agriculture and rural regions, as part of its efforts to realize the \"revitalization of rural areas\". In the ST, \"300 亿元以上\" ( more than 30 billion yuan ) demonstrates the great importance attached to the development of rural areas by the government. However, this amount is only budgeted but not fixed, thereby leaving the possibility that the actual investment may be slightly lower than 30 billion. For the purpose of precaution, \"以上\" is not translated into its nearest equivalence \"more than\", but modified to be \"around\" in the TT."
},
{
"figure_ref": [],
"heading": "Discussion",
"publication_ref": [
"b11",
"b23",
"b33",
"b1",
"b20",
"b13",
"b53"
],
"table_ref": [],
"text": "To summarize, the corpus-based study of the translation of hedges in political documents reveals that hedges are more frequently used in English than in Chinese across our three sub-corpora. To be specific, for C-E translation, more hedging devices are found in the TTs than in the STs, while for E-C translation, more hedges appear in the STs than in the TTs. This finding points to a systemic difference regarding the general frequency of hedges in political texts between Chinese and English. Results in our study can find echoes in the previous studies, which suggested that English is \"over-hedged\" (Fu and Wang, 2022) compared with many other languages, including Spanish, Iranian, Arabian, and Chinese across various types of discourse and texts (Jalilifar and Alavi-Nia, 2012;M.Aljawadi, 2022;Alonso Alonso, 2019;Hu and Cao, 2011;Gong et al., 2021;Yang, 2013)."
},
{
"figure_ref": [],
"heading": "Linking Power Distance and Hedges",
"publication_ref": [
"b18",
"b43",
"b44"
],
"table_ref": [],
"text": "This disparity can be attributed to different power distances between China and English-speaking countries. Power distance is an anthropological concept propose by Hofstede as part of his cultural dimensions theory (Hofstede, 2011), which reflects the unequal power distribution in a community or society, and can be used to understand the interpersonal relationship between individuals with varying degrees of power. To measure power distance in a quantitative way, Hofstede proposed Power Distance Index (PDI). According to his research results, China is a country with high PDI, meaning that the decisions, opinions, and actions of the leading cadre are not likely to be challenged, and the hierarchical distinction between the leadership and ordinary citizens should be accepted. In contrast, English-speaking countries like the United Kingdom and the United States are gauged to have low PDI scores, meaning that members in these societies are ready to challenge hierarchy and voice dissenting opinions against the authority.\nIn the hierarchical Chinese society, the central government represents the highest authority, and political documents issued by the government institutions are thought to be assertive, authoritative, and unquestionable (Tang, 2016). As hedges are mostly \"informal, less specific markers of probability and uncertainty\" (Biber, 1988, p.240) to make the propositional content \"fuzzier or less fuzzy\" (Lakoff, 1973, p.195), it is not difficult to understand why hedges are not frequently adopted in the reports on the work of the government, which often feature formality and seriousness. On the contrary, the United Nations does not position itself as an authoritative institution, but an organization that encourages cooperation and solidarity with the mission to preserve world peace (Thorvaldsdottir et al., 2021). Therefore, the annual reports of the UN are written in a more engaging manner, which may explain the relatively higher occurrence of hedging expressions."
},
{
"figure_ref": [],
"heading": "Influence of the Shifting Translation Norms and Policy Orientation",
"publication_ref": [
"b26",
"b50",
"b55",
"b51",
"b6",
"b48",
"b24",
"b24",
"b6",
"b28",
"b36",
"b24",
"b45",
"b24",
"b36",
"b56"
],
"table_ref": [],
"text": "Ever since the sociological turn in translation studies, norm has gained its foothold in describing and explaining translation phenomenon related to power, ideology, and culture (Hermans, 1999, p.26).\nAlthough no consensus has been fully reached on whether to define norm as a descriptive or prescriptive concept, and its relationship with psycholinguistic as well as cognitive aspects (Kotze, 2020), norms can be regarded as shared values or rules that \"govern, identify and individualize the social order of the cultural system where translation takes place\" (Enríquez-Aranda, 2016, p.89). As pointed out by Wunderlich (2013); Yu and Xu (2016); Xu and Tian (2020), norms are, far from being constant and stable, susceptible to change and evolve. They \"emerge, diffuse, become internalized, and, once established, become subject to change resulting in their strengthening, weakening, or even erosion\" (Wunderlich, 2013, p.20). This dynamic nature of norms is particularly suitable to account for the diachronic shift regarding the frequency of hedges in C-E translation. To illustrate, accuracy and faithfulness to the STs have long been the dominant norm of translation accepted by institutional translators working for the government, since the original political documents, which represent the official voice and stance, are attached with absolute authority (Cheng, 1983;Wang, 2008;Jia, 2021). Under this context, translations of political documents have been required to resemble the STs as much as possible, even at the expense of understandability and acceptability (Jia, 2021). In the extreme cases, even the word order should not be easily changed, and any detail should not be ignored (Cheng, 1983). However, this norm has been questioned by a group of senior translators. In the meanwhile, since \"going global\" was designated as an important national strategy at the 16th National Congress of the CPC, institutional translators are entrusted with the responsibility to balance faithfulness and reception, in order to better represent the Chinese voices, enhance China's soft power, and engage with the international society (Li and Li, 2015;Pan and Wang, 2021). Much attention has been shifted away from a blind pursuit of \"faithfulness\" to the integration of accuracy and acceptability (Jia, 2021;Tong, 2014). The translation of Chinese political texts is expected to approach the international political discourse, and make adjustments from the aspects of content, style, and manner (Jia, 2021).\nAs indicated by our data analysis in the previous parts, we can identify a changing pattern concerning the frequencies of hedges in the translated political texts from Chinese to English. Their significant increase in the TTs during the past two decades is a vivid reflection of the shifting norms of the political text translation in recent years, giving support to previous findings that the translation of the Chinese political documents has become more target-oriented, reader-friendly, and easier to receive by the international community (Pan and Wang, 2021;Zhang, 2004)."
},
{
"figure_ref": [],
"heading": "Gatekeeping and the Directionality of Translating Hedges",
"publication_ref": [
"b39",
"b12",
"b47"
],
"table_ref": [
"tab_1",
"tab_7",
"tab_7"
],
"text": "What is out of our expectation is the directionality as an important factor that influences the general frequency of hedges: in the two C-E sub-corpora, we see a bigger difference between the STs and TTs in respect to the occurrence of hedges than the E-C sub-corpus. While almost all the hedging expressions are retained when being translated to Chinese, the rate of retention is much lower for the other direction (see table 3 andtable 8), which means that the translator's intervention is more prominent in C-E translation compared with E-C translation in our corpora. The mediation efforts by the institutional translators are referred to as \"gatekeeping\" (Pöllabauer, 2012;Fujii, 1988;Wadensjö, 1998;Gu, 2019b), which has been used as a metaphor in translation studies to indicate the active role of translators guided by certain ideologies to intervene in the process of text reproduction. In the context of C-E translation of political documents, since the reports on the work of the government often contain major national policies, reforms, and strategies, they are of vital importance for both the domestic and international audience. Therefore, when they are translated into English, the translators are particularly careful in the delivery of the message. Translators as gatekeepers are expected to filter the messages concerning national image and interests in the STs by means of various handlings (i.e. addition, omission, modification), with an aim to both maintain political correctness and ensure that the translated texts are clearly understood and well-received by international readers. In contrast, the major concerns and objectives of the UN are international regional affairs. Issues related to China are seldom mentioned in the UN annual reports (in fact, China and the Chinese government are never mentioned in the 2018 2022 UN annual reports). Therefore, the gatekeeping function of translators is less obvious for E-C translation, manifested by the relatively high retention rate of hedges, and extremely low adoption of other translation methods other than several cases of omission (see table 8)."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [
"b27"
],
"table_ref": [],
"text": "Based on three customized corpora consisted of political documents from the reports on the work of the government and the UN annual reports, this study adopts a comparative approach to investigate how hedges are used in the STs and translated to the TTs. Overall, our analysis shows that hedges tend to appear more frequently in English political texts, be it translated English or original English, which points to a systemic difference regarding the frequency of hedges between the two distinct languages. In addition, directionality seems to play an important role in influencing what proportion of hedges in the STs will be translated into their nearest equivalents in the TTs. Our data reveal that the retention rate of hedges in E-C translation is higher than that in C-E translation of political documents, at least in the case of our three sub-corpora, which can be partly explained by the different degrees of gatekeeping efforts by the institutional translators.\nAnother major finding is that there is a noticeable diachronic increase in the occurrence of hedging devices in the English translations of the reports on the work of the government. For one thing, this change can be partly attributed to the increase of hedges in the STs; for another, there is still a portion of hedges that are intentionally added, and thus can be seen as the result of mediation by the translators. Such change occurs in parallel to the shifting policy that guides the outward translation of Chinese political texts, which turns from being \"source-oriented\" to \"target-oriented\" in recent years. Differences also exist in terms of the preferred category of hedges in the two C-E sub-corpora: while adaptors appear significantly more frequently in the reports roughly two decades ago, attribution shields, especially \"based\" and \"based on\" serve as the most preferred hedging devices in the current years. Furthermore, our study finds that among the four identified handlings of hedges, retention is the dominant one, while omission is the least likely to be seen. Addition and modification are also actively deployed in C-E translation, but not used at all in E-C translation.\nAlthough much attention has been paid to hedges since Lakoff (1973), studies devoted to the use of hedges in translated political texts remain scarce. This study, with the help of authentic and fit-for-purpose textual data, along with corpora-building tools, offers some insight into how hedges are used differently between the STs and the TTs, between C-E translation and E-C translation, as well as between two time periods. Additionally, the study attempts to explain these differences from a socio-cultural perspective, and reveals the integral relationship among translation, ideology, power, and culture. Nevertheless, restricted by the size of the corpora, our findings remain preliminary, and thus more future studies are awaited to offer systematic and in-depth discussion of hedging expressions in the translation of political documents."
}
] | Hedges are widely studied across registers and disciplines, yet research on the translation of hedges in political texts is extremely limited. This contrastive study is dedicated to investigating whether there is a diachronic change in the frequencies of hedging devices in the target texts, to what extent the changing frequencies of translated hedges through years are attributed to the source texts, and what translation strategies are adopted to deal with them. For the purposes of this research, two types of official political texts and their translations from China and the United Nations were collected to form three sub-corpora. Results show that hedges tend to appear more frequently in English political texts, be it original English or translated English. In addition, directionality seems to play an important role in influencing both the frequencies and translation strategies regarding the use of hedges. A noticeable diachronic increase of hedging devices is also observed in our corpus. | Hedges in Bidirectional Translations of Publicity-Oriented Documents | [
{
"figure_caption": "approximators adaptors some, several, a portion of, certain, somewhat, quite, kind of, sort of, almost, can, could, might, may 一些(yi xie)，有些(you xie)，部分(bu fen)，有的(you de)，有所(you suo)，几(ji)，有点(you dian)，几乎(ji hu)，可能(ke neng) rounders approximately, about, roughly, around，nearly 大 约(da yue)， 大 概(da gai)， 围 绕(wei rao)， 上 下(shang xia)， 左 右(zuo you)，近(jin), 不到(bu dao) shields plausibility shields I/we think, I/we (firmly) believe, I/we suppose, I/we wonder, from my/our understanding, to my/our knowledge 我/我们相信(wo/wo men xiang xin)，我/我们觉得(wo/wo men jue de)， 我/我们认为(wo/wo men ren wei)，我/我们想(wo/wo men xiang)，据 我/我们所知(ju wo/wo men suo zhi) attribution shields",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Basic information of the three sub-corpora.",
"figure_data": "STTTItemsTokensItemsTokensReport in 20007657Translated report in 200013079Report in 20017605Translated report in 200113210RWG1Report in 2002 Report in 2003 11371 7454Translated report in 2002 Translated report in 200312865 18921Report in 20048242Translated report in 200413877Total4232971952Report in 2018 14798Translated report in 201824201Report in 2019 11707Translated report in 201919877RWG2Report in 2020 Report in 2021 10661 8291Translated report in 2020 Translated report in 202116059 19147Report in 20228895Translated report in 202216491Total5435295775Report in 20189650Translated report in 201813620Report in 20199687Translated report in 201916406UNARReport in 2020 Report in 20218962 7640Translated report in 2020 Translated report in 202115175 12194Report in 20228432Translated report in 202213869Total4437171264",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Raw and normalized frequencies of hedges in the three sub-corpora.",
"figure_data": "HedgesRWG1RWG2UNARST (zh) TT (en) ST (zh) TT (en) ST (en) TT (zh)Raw freq.217249373437331324Norm freq. (per 1000)5.133.476.864.584.664.55",
"figure_id": "tab_1",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Raw and normalized frequencies (per 1000 words) of hedges in RWG1 and RWG2.",
"figure_data": "CategoriesRWG1 (TT) Raw freq. Norm freq. Raw freq. Norm freq. RWG2 (TT)Adaptors1031.44910.95Rounders390.54610.64Plausibility Shields20.0300Attribution Shields1051.462852.99Total2493.474374.58Adaptors RWG1 (TT) RWG2 (TT)some5637can2722certain1011may78almost36",
"figure_id": "tab_2",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "High-frequency adaptors.",
"figure_data": "",
"figure_id": "tab_3",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "Table 6 lists the raw frequencies of attribution shields in the two sub-corpora. It is noteworthy Frequency distribution of attribution shields.",
"figure_data": "Attribution ShieldsRWG1 (TT) RWG2 (TT)-based095based on1064in accordance with6978according to2547accordingly10as the Chinese say-ing goes01Total105285",
"figure_id": "tab_4",
"figure_label": "6",
"figure_type": "table"
},
{
"figure_caption": "To test our hypothesis, we searched the two ex-",
"figure_data": "respects, deepen reform, advancelaw-basedgovernance, and strengthen Partyhave reformed and improved themarket-basedexchange rate mechanism and keptqualification examinations for somelicense-basedprofessions. We will support thebe better integrated and performance-based management will be strengthenedthe internet and otherIT-basedapproaches to strengthen oversightincluding developing at-home, community-based and mutual-aid elderly careenforced in a strict,procedure-basedimpartial, and civil mannerWe will strengthenauditing-basedoversight. We will consolidate andfull support in exercisinglaw-basedgovernance and in their effortsWe will promote creditrating-basedregulation and the Internet Plusa new model of social governancebased oncollaboration, co-governance, andwork on building a governmentbased onthe rule of law, andand create a business environmentbased onrule of law that thecontinue to promote opening upbased onflows of goods andgreater emphasis to opening upbased onrules and related institutionscore and an international orderbased oninternational law. China iswhile also pursuing progress.Based onChina's realities, we refrainedadjustments and improvementsbased onnew developments to reinforceeffective tax-and-fee reduction stepsbased onlocal conditions and in the",
"figure_id": "tab_5",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Concordance lines of \"NOUN-based\" and \"based on\" in RWG2 (TT).",
"figure_data": "CorporaBNCCOCANOUN-based 8437 (0.09)84381 (0.08)based on11337 (0.12) 141241 (0.14)",
"figure_id": "tab_6",
"figure_label": "7",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "tab_7",
"figure_label": "8",
"figure_type": "table"
},
{
"figure_caption": "\"-based\" and \"based ",
"figure_data": "CorporaPatternsFrequencyCalifornia-based355London-based280computer-based184Unix-based143BNCMassachusetts-based community-based134 113RISC-based113UK-based92land-based91broad-based90evidence-based3185community-based2669web-based2241school-based2031COCAfaith-based New York-based1852 1547broad-based1247Washington-based1047Atlanta-based1016Chicago-based988",
"figure_id": "tab_8",
"figure_label": "9",
"figure_type": "table"
},
{
"figure_caption": "\"NOUN-based\" patterns in BNC and COCA. ",
"figure_data": "Sub-corpora Retention Addition Omission Modification TotalRWG121518319252RWG236842532442UNAR324070331",
"figure_id": "tab_9",
"figure_label": "9",
"figure_type": "table"
},
{
"figure_caption": "Handlings of hedges in the three sub-corpora.",
"figure_data": "",
"figure_id": "tab_10",
"figure_label": "10",
"figure_type": "table"
}
] | Zhaokun Jiang; Ziyin Zhang | [
{
"authors": "Esmaeel Abdollahzadeh",
"journal": "Journal of Pragmatics",
"ref_id": "b0",
"title": "Poring over the findings: Interpersonal authorial engagement in applied linguistics papers",
"year": "2011"
},
{
"authors": "Rosa Alonso; Alonso ",
"journal": "Onomázein: Revista de lingüística, filología y traducción de la Pontificia Universidad Católica de Chile",
"ref_id": "b1",
"title": "A multicompetence perspective of hedging in second language academic writing",
"year": "2019"
},
{
"authors": "Mona Baker",
"journal": "Routledge",
"ref_id": "b2",
"title": "Corpus Linguistics and Translation Studies: Implications and applications",
"year": "2019"
},
{
"authors": "Douglas Biber",
"journal": "Cambridge University Press",
"ref_id": "b3",
"title": "Variation across Speech and Writing",
"year": "1988"
},
{
"authors": "Penelope Brown; Stephen C Levinson",
"journal": "Cambridge University Press",
"ref_id": "b4",
"title": "Politeness: Some universals in language usage",
"year": "1987"
},
{
"authors": "Krisda Chaemsaithong",
"journal": "Folia Linguistica",
"ref_id": "b5",
"title": "Evaluative stancetaking in courtroom opening statements",
"year": "2017"
},
{
"authors": " Cheng",
"journal": "中国翻译",
"ref_id": "b6",
"title": "翻译十二大文件的点滴体会",
"year": "1983"
},
{
"authors": "Andrew Chesterman",
"journal": "Benjamins Translation Library",
"ref_id": "b7",
"title": "Hypotheses about translation universals",
"year": "2004"
},
{
"authors": "Mercedes Enríquez-Aranda",
"journal": "Onomázein",
"ref_id": "b8",
"title": "Translation norms in the light of practical research in literary translation",
"year": "2016"
},
{
"authors": "Bruce Fraser",
"journal": "",
"ref_id": "b9",
"title": "Hedging in political discourse: the bush 2007 press conferences",
"year": "2010"
},
{
"authors": "Rongbo Fu; Jing Chen",
"journal": "Interpreting",
"ref_id": "b10",
"title": "Negotiating interpersonal relations in chinese-english diplomatic interpreting: Explicitation of modality as a case in point",
"year": "2019"
},
{
"authors": "Rongbo Fu; Kefei Wang",
"journal": "Text & Talk",
"ref_id": "b11",
"title": "Hedging in interpreted and spontaneous speeches: a comparative study of chinese and american political press briefings",
"year": "2022"
},
{
"authors": "Akio Fujii",
"journal": "Meta",
"ref_id": "b12",
"title": "News translation in japan",
"year": "1988"
},
{
"authors": "Heng Gong; Lingling Liu; Feng Cao",
"journal": "Journal of Language, Identity & Education",
"ref_id": "b13",
"title": "A cross-linguistic study of interactional metadiscourse in english and chinese research articles by the same chinese scholars",
"year": "2021"
},
{
"authors": "Tatiana Gribanova; Tamara Gaidukova",
"journal": "Training Language and Culture",
"ref_id": "b14",
"title": "Hedging in different types of discourse",
"year": "2019"
},
{
"authors": "Chonglong Gu",
"journal": "Critical Discourse Studies",
"ref_id": "b15",
"title": "Mediating 'face' in triadic political communication: a cda analysis of press conference interpreters' discursive (re)construction of chinese government's image (1998-2017)",
"year": "2019"
},
{
"authors": "James Chonglong; Gu ",
"journal": "Translation and Interpreting Studies",
"ref_id": "b16",
"title": "Interpreters caught up in an ideological tug-of-war?: A cda and bakhtinian analysis of interpreters' ideological positioning and alignment at government press conferences",
"year": "2019"
},
{
"authors": "Theo Hermans",
"journal": "",
"ref_id": "b17",
"title": "3 Norms and the Determination of Translation: A Theoretical Framework",
"year": "1999"
},
{
"authors": "Geert Hofstede",
"journal": "Online Readings in Psychology and Culture",
"ref_id": "b18",
"title": "Dimensionalizing cultures: The hofstede model in context",
"year": "2011"
},
{
"authors": "Janet Holmes",
"journal": "Language & Communication",
"ref_id": "b19",
"title": "Hedges and boosters in women's and men's speech",
"year": "1990"
},
{
"authors": "Guangwei Hu; Feng Cao",
"journal": "Journal of Pragmatics",
"ref_id": "b20",
"title": "Hedging and boosting in abstracts of applied linguistics articles: A comparative study of english-and chinese-medium journals",
"year": "2011"
},
{
"authors": "Ken Hyland",
"journal": "John Benjamins",
"ref_id": "b21",
"title": "Hedging in Scientific Research Articles",
"year": "1998"
},
{
"authors": "Ken Hyland",
"journal": "Continuum",
"ref_id": "b22",
"title": "Metadiscourse: Exploring interaction in writing",
"year": "2005"
},
{
"authors": "Alireza Jalilifar; Maryam Alavi-Nia",
"journal": "Discourse & Communication",
"ref_id": "b23",
"title": "We are surprised; wasn't iran disgraced there? a functional analysis of hedges and boosters in televised iranian and american presidential debates",
"year": "2012"
},
{
"authors": "Wenbo Jia",
"journal": "上海翻译",
"ref_id": "b24",
"title": "新时期政治文献对外翻译: 理 念更新、与时俱进才是硬道理",
"year": "2021"
},
{
"authors": "Andreas H Jucker; Sara W Smith",
"journal": "Pragmatics and beyond. New series",
"ref_id": "b25",
"title": "And people just you know like 'wow' : Discourse markers as negotiating strategies",
"year": "1998"
},
{
"authors": "Haidee Kotze",
"journal": "Routledge",
"ref_id": "b26",
"title": "The Routledge handbook of translation and cognition, chapter Translation, contact linguistics and cognition",
"year": "2020"
},
{
"authors": "George Lakoff",
"journal": "Journal of Philosophical Logic",
"ref_id": "b27",
"title": "Hedges: A study in meaning criteria and the logic of fuzzy concepts",
"year": "1973"
},
{
"authors": "Jingjing Li; Saihong Li",
"journal": "Perspectives",
"ref_id": "b28",
"title": "New trends of chinese political translation in the age of globalisation",
"year": "2015"
},
{
"authors": "Xin Li; Ranran Zhang",
"journal": "Routledge",
"ref_id": "b29",
"title": "Advances in Discourse Analysis of Translation and Interpreting",
"year": "2020"
},
{
"authors": "Sixin Liao; Li Pan",
"journal": "Interpreting",
"ref_id": "b30",
"title": "Interpreter mediation at political press conferences: A narrative account",
"year": "2018"
},
{
"authors": "Chunhong Liu; Ming-Yu Tseng",
"journal": "English for Specific Purposes",
"ref_id": "b31",
"title": "Paradigmatic variation in hedging and boosting: A comparative study of discussions in narrative inquiry and grounded theory research",
"year": "2021"
},
{
"authors": "Cédric Magnifico; Bart Defrancq",
"journal": "Interpreting",
"ref_id": "b32",
"title": "Hedges in conference interpreting: The role of gender",
"year": "2017"
},
{
"authors": "M Areej; Aljawadi",
"journal": "International Journal of English Linguistics",
"ref_id": "b33",
"title": "The analysis of translated hedges in trump's political speeches and interviews",
"year": "2022"
},
{
"authors": "Pedro Martín",
"journal": "Vigo International Journal of Applied Linguistics",
"ref_id": "b34",
"title": "The pragmatic rhetorical strategy of hedging in academic writing",
"year": "2022"
},
{
"authors": "Feng Pan; Tao Li",
"journal": "International Journal of Translation Studies",
"ref_id": "b35",
"title": "The retranslation of chinese political texts: Ideology, norms, and evolution. Target",
"year": "2021"
},
{
"authors": "Feng Pan; Binhua Wang",
"journal": "Babel",
"ref_id": "b36",
"title": "Is interpreting of china's political discourse becoming more target-oriented?: A corpus-based diachronic comparison between the 1990s and the 2010s",
"year": "2021"
},
{
"authors": "Diane Ponterotto",
"journal": "Pragmatics and Society",
"ref_id": "b37",
"title": "Hedging in political interviewing: When obama meets the press",
"year": "2018"
},
{
"authors": "Ellen Prince; Joel Frader; Charles Bosk",
"journal": "Linguistics and the Professions",
"ref_id": "b38",
"title": "On hedging in physician-physician discourse",
"year": "1982"
},
{
"authors": "Sonja Pöllabauer",
"journal": "Meta: Journal des traducteurs",
"ref_id": "b39",
"title": "Gatekeeping practices in interpreted social service encounters",
"year": "2012"
},
{
"authors": "Françoise Salager-Meyer",
"journal": "English for Specific Purposes",
"ref_id": "b40",
"title": "Hedges and textual communicative function in medical english written discourse",
"year": "1994"
},
{
"authors": "Deborah Schiffrin",
"journal": "Cambridge University Press",
"ref_id": "b41",
"title": "Discourse Markers. Studies in Interactional Sociolinguistics",
"year": "1987"
},
{
"authors": "Asmara Shafqat; Rafique Memon; Huma Akhtar",
"journal": "International Journal of English Linguistics",
"ref_id": "b42",
"title": "Cross-cultural analysis of the use of hedges in european and pakistani english newspaper: A corpus-based study",
"year": "2019"
},
{
"authors": "Wenfang Tang",
"journal": "Oxford University Press",
"ref_id": "b43",
"title": "Populist Authoritarianism: Chinese Political Culture and Regime Sustainability",
"year": "2016"
},
{
"authors": "Svanhildur Thorvaldsdottir; Ronny Patz; Steffen Eckhard",
"journal": "International Review of Administrative Sciences",
"ref_id": "b44",
"title": "International bureaucracy and the united nations system",
"year": "2021"
},
{
"authors": "Xiaohua Tong",
"journal": "中 国 翻 译",
"ref_id": "b45",
"title": "翻译的主体意识--2014 年 政 府 工 作 报 告 翻 译 心 得",
"year": "2014"
},
{
"authors": "Ignacio Vázquez",
"journal": "AELFE)",
"ref_id": "b46",
"title": "A contrastive analysis of the use of modal verbs in the expression of epistemic stance in business management research articles in english and spanish. Ibérica: Revista de la Asociación Europea de Lenguas para Fines Específicos",
"year": "2010"
},
{
"authors": "Cecilia Wadensjö",
"journal": "Addison Wesley Longman",
"ref_id": "b47",
"title": "Interpreting as Interaction",
"year": "1998"
},
{
"authors": " Wang",
"journal": "中国翻译",
"ref_id": "b48",
"title": "政治文献翻译新探索--十七 大文件翻译体会",
"year": "2008"
},
{
"authors": "Q Wang; J Zhou",
"journal": "Journal of Shandong University of Technology (Social Sciences)",
"ref_id": "b49",
"title": "A comparative study of shields in discussion section of medical discourse in english and chinese",
"year": "2009"
},
{
"authors": "Carmen Wunderlich",
"journal": "",
"ref_id": "b50",
"title": "Theoretical approaches in norm dynamics",
"year": "2013"
},
{
"authors": "Mingwu Xu; Chuanmao Tian",
"journal": "Perspectives",
"ref_id": "b51",
"title": "A case study of translation norm dynamics: a chinese perspective",
"year": "2020"
},
{
"authors": "Ying Yang; Foong Ha; Yap ",
"journal": "Journal of Pragmatics",
"ref_id": "b52",
"title": "i am sure but i hedge\": Fear expression kǒngpà as an interactive rhetorical strategy in mandarin broadcast talk",
"year": "2015"
},
{
"authors": "Yingli Yang",
"journal": "Journal of Pragmatics",
"ref_id": "b53",
"title": "Exploring linguistic and cultural variations in the use of hedges in english and chinese scientific discourse",
"year": "2013"
},
{
"authors": "Zhipu Yang; Lin Li",
"journal": "Journal of Pragmatics",
"ref_id": "b54",
"title": "No straight talk here: A multi-level analysis of hedging strategies employed by the fed chair in press conferences",
"year": "2022"
},
{
"authors": "Jing Yu; Minhui Xu",
"journal": "Perspectives",
"ref_id": "b55",
"title": "From normbreaking to norm-making: a sociological study of the genesis of a new norm",
"year": "2016"
},
{
"authors": "Yuanyuan Zhang",
"journal": "中国翻译",
"ref_id": "b56",
"title": "谈谈领导人言论英译的 几个问题",
"year": "2004"
}
] | [
{
"formula_coordinates": [
6,
218.99,
283.75,
293.39,
34.21
],
"formula_id": "formula_0",
"formula_text": "似乎(si hu)，好像(hao xiang)，(根)据((gen) ju)，依(照) (yi (zhao)), 据 说(ju shuo)， 据 报 道(ju bao dao)， 表 明(biao ming)， 基 于(ji yu), 考 虑(到) (kao lv (dao))"
},
{
"formula_coordinates": [
10,
329.09,
274.71,
174.63,
52.38
],
"formula_id": "formula_1",
"formula_text": "坚 持 不 懈 地 开 展 反 腐 败 斗 争，大力纠正部门和行业不正之风， 依法惩处了一批违法违纪的腐败分 子。"
},
{
"formula_coordinates": [
11,
93.82,
364.71,
174.64,
79.2
],
"formula_id": "formula_2",
"formula_text": "在世界许多地区，与有组织 犯罪有关的凶杀和暴力行为发生率仍 然很高，如果这些行为与非法贩运武 器和初级商品行为牵扯在一起，会破 坏实现和平、保护人权和可持续发展 的努力。"
}
] |
2023-05-20 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b0",
"b1",
"b2",
"b3",
"b4",
"b5",
"b6",
"b7",
"b8",
"b9",
"b10",
"b10",
"b11",
"b9",
"b12",
"b13",
"b10",
"b11",
"b14",
"b15",
"b16",
"b17",
"b18",
"b19",
"b20",
"b11",
"b14",
"b15",
"b16"
],
"table_ref": [],
"text": "By mimicking the spatio-temporal dynamics behaviors of biological neural circuits, Spiking Neural Networks (SNNs) [1][2][3][4] provide a low-power alternative to traditional Artificial Neural Networks (ANNs). The binary spiking communication enables SNNs to be deployed on neuromorphic chips [5][6][7] to perform sparse synaptic accumulation for low energy consumption. Given the memory storage limitations of such devices, neural pruning methods are well recognized as one of the crucial methods for implementing SNNs in real-world applications. Pruning redundant weights from an over-parameterized model is a mature and efficient way of obtaining significant compression [8,9].\nRecently, Lottery Ticket Hypothesis (LTH), a mile stone is proposed in the literature of network pruning, which asserts that an over-parameterized neural network contains sub-networks that can achieve a similar or even better accuracy than the fully-trained original dense networks by training only once [10]. A Stronger version of the LTH (SLTH) was then proposed: there is a high probability that a network with random weights contains sub-networks that can approximate any given sufficientlysmaller neural network, without any training [11]. SLTH claims to find the target sub-network without training, thus it is a kind of complement to original LTH that requires training [11]. Meanwhile, SLTH is considered to be \"stronger\" than LTH because it claims to require no training [12].\nThe effectiveness of LTH in ANNs have been verified by a series of experiments [10,13,14,11]. Furthermore, it has been theoretically proved by several works with various assumptions [12,[15][16][17], due to its attractive properties in statement. A line of work is dedicated to designing efficient pruning algorithms based on LTH [18][19][20]. But the role of LTH in SNN is almost blank. There is only one work that experimentally verifies that LTH can be applied to SNNs [21], whether it can also be theoretically established remains unknown.\nIn this work, we theoretically and experimentally prove that LTH (Strictly speaking, SLTH 3 ) holds in SNNs. There are two main hurdles. First and foremost, the binary spike signals are fired when the membrane potentials of the spiking neurons exceed the firing threshold, thus the activation function of a spiking neuron is discrete. But all current works [12,[15][16][17] proving LTH in ANNs relies on the Lipschitz condition. Only when the Lipschitz condition is satisfied, the error between the two layers of the neural network will be bounded when they are approximated. Second, brain-inspired SNNs have complex dynamics in the temporal dimension, which is not considered in the existing proofs and increases the difficulty of proving LTH in SNNs.\nTo bypass the Lipschitz condition, we design a novel probabilistic modeling approach. The approach can theoretically provide the probability that two SNNs behave identically, which is not limited by the complex spatio-temporal dynamics of SNNs. In a nutshell, we establish the following result:\nInformal version of Theorem 4.3 For any given target SNN Ĝ, there is a sufficiently large SNN G with a sub-network (equivalent SNN) G that, with a high probability, can achieve the same output as Ĝ for the same input,\nsup S∈S 1 T T t=1\nGt (S) -Ĝt (S) As a complement to the theoretical proof part, we show experimentally that there are also subnetworks in SNNs that can work without training. Subsequently, it is clear that ANNs and SNNs operate on fundamentally distinct principles about how weight influences the output of the activation layer. Thus, simply using the weight size as the criterion for pruning like ANNs must not be the optimal strategy. Based on this understanding, we design a new weight pruning criterion for SNNs. We evaluate how likely weights are to affect the firing of spiking neurons, and prune according to the estimated probabilities. In summary, our contributions are as follows:\n• We propose a novel probabilistic modeling method for SNNs that for the first time theoretically establishes the link between spiking firing and pruning (Section 3). The probabilistic modeling method is a general theoretical analysis tool for SNNs, and it can also be used to analyze the robustness and other compression methods of SNNs (Section 6).\n• With the probabilistic modeling method, we theoretically prove that LTH also holds in SNNs with binary spiking activations and complex spatio-temporal dynamics (Section 4).\n• We experimentally find the good sub-networks without weight training in random initialized SNNs, which is consistent with our theoretical results. (Section 5)\n• We apply LTH for pruning in SNNs and design a new probability-based pruning criterion for it (Section 6). The proposed pruning criterion can achieve better performance in LTH-based pruning methods and can also be exploited in non-LTH traditional pruning methods."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b2",
"b21",
"b22",
"b3",
"b5",
"b4",
"b23",
"b24",
"b25",
"b26",
"b27",
"b28",
"b29",
"b20",
"b30",
"b31",
"b32",
"b33",
"b34"
],
"table_ref": [],
"text": "Spiking neural networks. The spike-based communication paradigm and event-driven nature of SNNs are key to their energy-efficiency advantages [3,22,23,4]. Spike-based communication makes cheap synaptic Accumulation (AC) the basic operation unit, and the event-driven nature means that only a small portion of the entire network is active at any given time while the rest is idle. In contrast, neurons in ANNs communicate information using continuous values, so Multiply-and-Accumulate (MAC) is the major operation, and generally all MAC must be performed even if all inputs or activations are zeros. SNNs can be deployed on neuromorphic chips for low energy consumption The firing behavior in the approximation of spiking neurons can be described by the proposed probabilistic modeling approach (Section 3). Spikes are fired when the membrane potential u exceeds the firing threshold u th . As long as a weight change induces u to fall into the crisis neighborhood, there is a certain probability that the spiking firing will be changed (0 to 1, or 1 to 0). We hope that spiking firing will not change after the redundant weights are pruned. (c), Equivalent structure modeling.\nThe error between the target and equivalent SNN is related to the network width (Section 4). (d), Proof of LTH in SNN (Section 4). (e), We provide a pruning technique for SNNs (Section 6). New pruning criterion: compute the probability that the firing of a spiking neuron changes when weights are pruned, and pruning according to the rank of the probability. [6,5,24]. Thus, spike-based neuromorphic computing has broad application prospects in battery constrained edge computing platforms [25], e.g., internet of things, smart phones, etc.\nPruning in spiking neural networks. Recent studies on SNN pruning have mostly taken two approaches: 1) Gaining knowledge from ANN pruning's successful experience 2) incorporating the unique biological properties of SNNs. The former technical route is popular and effective. Some typical methods include pruning according to predefined threshold value [26][27][28], soft-pruning that training weights and pruning thresholds concurrently [29], etc. The temporal dynamics of SNN are often also taken into consideration in the design of pruning algorithms [30,21]. Meanwhile, There have been attempts to develop pruning algorithms based on the similarities between SNNs and neural systems, e.g., regrowth process [31], spine motility [32,33], gradient rewiring [34], state transition of dendritic spines [35], etc. None of these studies, however, took into account the crucial factor that affects network performance: the link between weights and spiking firing. In this work, we use probabilistic modeling to analyze the impact of pruning on spiking firing, and accordingly design a pruning criterion for SNNs."
},
{
"figure_ref": [
"fig_14",
"fig_14",
"fig_14"
],
"heading": "Probabilistic Modeling for Spiking Neurons",
"publication_ref": [
"b0",
"b1",
"b35",
"b36",
"b37",
"b1",
"b38",
"b39",
"b40",
"b41",
"b42",
"b43"
],
"table_ref": [],
"text": "No matter how intricate a spiking neuron's internal spatio-temporal dynamics are, ultimately the neuron decide whether to fire a spike or not based on its membrane potential at a specific moment. Thus, the essential question of the proposed probabilistic modeling is how likely is the firing of a spiking neuron to change after changing a weight.\nLeaky Integrate-and-Fire (LIF) is one of the most commonly used spiking neuron models [1,2], because it is a trade-off between the complex spatio-temporal dynamic characteristics of biological neurons and the simplified mathematical form. It can be described by a differential function\nτ du(t) dt = -u(t) + I(t),(1)\nwhere τ is a time constant, and u(t) and I(t) are the membrane potential of the postsynaptic neuron and the input collected from presynaptic neurons, respectively. Solving Eq. ( 1), a simple iterative representation of the LIF neuron [36,37] for easy inference and training can be obtained as follow\nu t,l i = h t-1,l i + x t,l i ,(2)\ns t,l i = Hea(u t,l i -u th ),(3)\nh t,l i = V reset s t,l i + βu t,l i (1 -s t,l i ),(4)\nx t,l i = N j=1 w l ij s t,l-1 j ,(5)\nwhere u t,l i means the membrane potential of the i-th neuron in l-th layer at timestep t, which is produced by coupling the spatial input feature x t,l i and temporal input h t-1,l i , u th is the threshold to determine whether the output spike s t,l i ∈ {0, 1} should be given or stay as zero, Hea(•) is a Heaviside step function that satisfies Hea(x) = 1 when x ≥ 0, otherwise Hea(x) = 0, V reset denotes the reset potential which is set after activating the output spiking, and β = e -dt τ < 1 reflects the decay factor. In Eq. ( 2), spatial feature x t,l i can be extracted from the spike s t,l-1 j from the spatial output of the previous layer through a linear or convolution operation (i.e., Eq. ( 5)), where the latter can also be regraded as a linear operation [38]. w l ij denotes the weight connect from the j-th neuron in (l -1)-th layer to the i-th neuron in l-th layer, N indicates the width of the (l -1)-th layer.\nThe spatio-temporal dynamics of LIF can be described as: the LIF neuron integrates the spatial input feature x t,l i and the temporal input h t-1,l i into membrane potential u t,l i , then, the fire and leak mechanism is exploited to generate spatial output s t,l i and the new neuron state h t,l i for the next timestep. Specifically, When u t,l i is greater than the threshold u th , a spike is fired and the neuron state h t,l i is reset to V reset . Otherwise, no spike is fired and the neuron state is decayed to βu t,l i . Richer neuronal dynamics [2] can be obtained by adjusting information fusion [39], threshold [40], decay [41], or reset [42] mechanisms, etc. Note, notations used in this work are summarized in Appendix A.\nProbabilistic modeling of spiking neurons. Our method is applicable to various spiking neuron models as long as Eq. ( 3) is satisfied. In this section, superscript l and i will be omitted when location is not discussed, and superscript t will be omitted when the specific timestep is not important. The crux of probabilistic modeling is how errors introduced by pruning alter the firing of spiking neurons.\nWe start by discussing only spatial input features. For a spiking neuron, assuming that the temporal input of a certain timestep is fixed, for two different spatial input features x and x ∈ [x -, x + ], the corresponding membrane potentials are also different, i.e., u and u ∈ [u -, u + ]. Once u and u are located in different sides of the threshold u th , the output will be different (see Fig. 1a). This situation can only happen when u is located in crisis neighborhood [u th -, u th + ] (see Fig. 1b). That is, suppose membrane potential u / ∈ [u th -, u th + ], if it changes from u to u , and u ∈ [u -, u + ], the output of the spiking neuron must not change. Consequently, the probability upperbound of a spiking neuron output change (crisis probability) is:\nu th + u th - p(u)du,(6)\nwhere p(•) is the probability density function of the membrane potential distribution. For the case of two independent spiking neurons, if the input is the same, then is controlled by the weights of the two neurons (Fig. 1a).\nIt is reasonable to consider that the membrane potential follows a certain probability distribution. The membrane potential is accumulated by temporal input and spatial input feature, the former can be regarded as a random variable, and the latter is determined by the input spikes and weights. The input spike is a binary random variable according to a firing rate, and usually the weights are also assumed to satisfy a certain distribution. Moreover, some existing works directly assume that the membrane potential satisfies the Gaussian distribution [43,44]. In this work, our assumptions about the membrane potential distribution are rather relaxed. Specifically, we give out a class of distributions for membrane potential, and we suppose all membrane potential should satisfy: The symbolic language is as follows\n∃ > 0, x ∈ [m -, m + ] , sup x p(x) < +∞.(7)\nThe upperbound of probability is controllable by controlling the input error : Lemma 3.2. For the probability density function p(•), if it is m-Neighborhood-Finite Distribution, for any δ> 0, there exists a constant > 0 so that m+ m-p(x)dx ≤ δ. Now, we add consideration of the temporal dynamics of spiking neurons. As shown in Eq. ( 2) and Eq. ( 4), spiking neurons have a memory function that the membrane potential retains the neuronal state information from all previous timesteps. The spatial input feature x is only related to the current input, and the temporal input h is related to the spatial input features at all previous timesteps. For the convenience of mathematical expression, if there is no special mention to consider the temporal dynamics inside the neuron, we directly write the spiking neuron as σ. But it should be noted that σ actually has complex internal dynamics. In its entirety, a spiking neuron in Eq. ( 3) can be denoted as σ t x 1:t-1 (x t ), where the temporal input h t-1 in the membrane potential u t depends on x 1:t-1 , i.e., spatial input features from 1 to t -1.\nUsing Definition 3.1, Lemma 3.2, and the math above, we can determine the probability of differing outputs from the two neurons due to errors in their spatial input features, under varying constraints.Specifically, in Lemma 3.3, it is assumed that the timestep is fixed and the temporal inputs to the two neurons are the same; in Lemma 3.4, we loosen the constraint on the timestep of the two neurons; finally, in Theorem 3.5, we generalize our results to arbitrary timesteps for two spiking layers (N neurons each layer). Lemma 3.3. At a certain temestep T , if the spiking neurons are u th -Neighborhood-Finite Distribution and the spatial input features of two neuron got an error upperbound , and they got the same temporal input h t-1 , the probability upperbound of different outputs is proportional to . Formally: For two spiking neurons σT and σT , when hT -1 = ĥT -1 and ûT = ĥT -1 + xT is a random variable follows the u th -Neighborhood-Finite Distribution, if xT -xT ≤ , then: P σT (x T ) = σT (x T ) ∝ Lemma 3.4. Suppose the spiking neurons are u th -Neighborhood-Finite Distribution at timestep T and the spatial input features of two corresponding spiking neurons got an error upperbound at any timestep, and they got the same temporal input h 0 , if both spiking neurons have the same output at the first T -1 timesteps, then the probability upperbound is proportional to 1-β . Formally:\nFor two spiking neurons σT and σT , when h0 = ĥ0 and ûT = ĥT\n-1 + xT is a random variable follows the u th -Neighborhood-Finite Distribution, if xt -xt ≤ and σt (x t ) = σt (x t ) for t = 1, 2, • • • , T -1, then: P σT (x T ) = σT (x T ) ∝ 1-β .\nTheorem 3.5. Suppose the spiking layers are u th -Neighborhood-Finite Distribution at timestep t and the inputs of two corresponding spiking layers with a width N got an error upperbound for each element of spatial input feature vector at any timestep, and they got the same initial temporal input vector h 0 , if there is no different spiking output at the first T -1 timesteps, then the probability upperbound is proportional to N 1-β . Formally:\nFor two spiking layers σT and σT , when h0 = ĥ0 and ût\n= ĥt-1 + xt is a random variable follows the u th -Neighborhood-Finite Distribution, if xt k -xt k ≤ (k = 1, 2, • • • , N ; i = 1, 2, • • • , T ) and σt (x t ) = σt (x t ) for t = 1, 2, • • • , T -1, then: P σT (x T ) = σT (x T ) ∝ N 1-β ."
},
{
"figure_ref": [
"fig_14",
"fig_14"
],
"heading": "Proving Lottery Ticket Hypothesis for Spiking Neural Networks",
"publication_ref": [
"b11",
"b11",
"b11",
"b14",
"b15"
],
"table_ref": [],
"text": "SLTH states that a large randomly initialized neural network has a sub-network that is equivalent to any well-trained network. In this work, the large initialized network, sub-network, well-trained network are named large network G, equivalent (pruned) network G, and target network Ĝ, respectively. We use the same method to differentiate the weights and other notations in these networks.\nWe generally follow the theoretical proof approach in [12]. Note, the only unique additional assumption we made (considering the characteristics of SNN) is that the membrane potentials follow a class of distribution defined in Definition 3.1, which is easy to satisfy. Specifically, the whole proof in this work is divided into two parts: approximation of linear transformation (Fig. 1c) and approximation of spatio-temporal nonlinear transformation (Fig. 1d), each of which contains three steps. The first part is roughly similar to the proof of LTH in the existing ANN [12]. The second part requires our spatio-temporal probabilistic modeling approach in Section 3, i.e., Theorem 3.5.\nApproximation of linear transformation. The existing methods for proving LTH in ANNs all first approach a single weight between two neurons by adding a virtual layer (introducing redundant weights), i.e., Step 1. Different virtual layer modeling methods [12,15,16] will affect the width of the network in the final conclusion. All previous proofs introduce two neurons in the virtual layer for approximation. In this work, we exploit only one spiking neuron in the virtual layer for the approximation, given the nature of the binary firing of spiking neurons. We then approximate the weights between one neuron and one layer of neurons (Step 2), and the weights between two layers of neurons (Step 3). Step 2 and Step 3 are the same as previous works.\nStep 1: Approximate single weight by a spiking neuron. As shown in Fig. 2, a connection weight in SNNs can be approximated via an additional spiking neuron (i.e, a new virtual layer with only one neuron) with two weights, one connect in and one connect out. We set the two weights of the virtual neuron as v and w, and the target connection weight is ŵ. Consequently, the equivalent function for the target weight can be written as g(s) = wσ(vs), where the temporal input of the virtual neuron does not need to be considered. Once v ≥ u th , the virtual neuron will fire no matter how the temporal input is if spatial input s = 1, while not fire if s = 0. Thus, the output of the virtual neuron is independent of the temporal input, which is V reset or the decayed membrane potential (neither of which is likely to affect the firing of the virtual neuron). We have: 1) target weight connection: ĝ(s) = ŵs; 2) equivalent structure: g(s) = wσ(vs). If the weight v satisfy v ≥ u th , the error between the target weight connection and the equivalent structure is\ng(s) -ŵs = ws -ŵs ≤ w -ŵ .(8)\nOnce the number of initialized weights in the large network G is large enough, it can choose the weights w and v to make the error approximate to 0 by probability convergence. Thus, only one spiking neuron is needed for the virtual layer. Formally, we have Lemma C.1.\nLayer Step 2: Layer weights approximation. Based on Lemma C.1, we can extend the conclusion to the case where there is one neuron in layer l and several neurons in layer l -1. The lemma is detailed in Lemma C.2.\nStep 3: Layer to layer approximation. Finally, we get an approximation between the two layers of spiking neurons. See Lemma C.3 for proof. Lemma 4.1. Layer to Layer Approximation.\nFix the weight matrix Ŵ ∈ [-1 √ N , 1 √ N ]\nN ×N which is the connection in target network between a layer of spiking inputs and the next layer of neurons. The equivalent structure is k spiking neurons with k × N weights V ∈ R k×N connect the input and W ∈ R N ×k connect out. All the weights wij and v ij random initialized with uniform distribution U[-1, 1] and i.i.d. B ∈ {0, 1} N ×k is the mask for matrix W , i,j B ij 0 ≤ N 2 , j B ij 0 ≤ N . Then, let the function of equivalent structure be g(s) = ( W B)σ(V s), where input spiking s is a vector that\ns ∈ {0, 1} N . Then, ∀0 <δ≤ 1, ∃ > 0 when k ≥ N N C th log N δ and C th = 1-u th 2 , there exists a mask matrix B that [g(s)] i -[ Ŵ s] i ≤ w.p at least 1 -δ.\nApproximation of spatio-temporal nonlinear transformation. Since the activation functions in ANNs satisfy the Lipschitz condition, the input error can control the output error of the activation function. In contrast, the output error is not governed by the input error when the membrane potential of a spiking neuron is at a step position. To break this limitation, in Step 4, we combine the probabilistic modeling method in Theorem 3.5 to analyze the probability of consistent output (only 0 or 1) of spiking layers. Then, in Step 5, we generalize the conclusions in Step 4 to the entire network regardless of the temporal input. Finally, we consider the dynamics of SNNs in the temporal dimension in Step 6. Specifically, we can denote a SNN as follow:\nG t (S) = G t,L • G t,L-1 • • • • • G t,2 • G t,1 (S),(9)\nwhere S ∈ {0, 1} N ×T is the spatial input of the entire network at all timesteps. G t,l (•) represents the function of network G at t-th timestep and l-th layer, when considering the specific temporal input, it can be written as:\nG t,l (S) = σ t,l W l S 1:t-1,l (W l S t,l-1 ),(10)\nwhere spatial input at t-th timestep is S t,l-1 ∈ {0, 1} N , σ denotes the activation function of spiking layer, its subscript S 1:t-1,l indicates that the membrane potential at t-th timestep is related to the spatial inputs at all previous timesteps. We are not concerned with these details in probabilistic modeling, thus they are omitted in the remaining chapters where they do not cause confusion.\nStep \n∈ [-1 √ N , 1 √ N ] N ×N\nwhich is the connection in target network between a layer of spiking inputs and the next layer of neurons. The equivalent structure is k spiking neurons with k × N weights V ∈ R k×N connect the input and W ∈ R N ×k connect out. All the weights wij and v ij random initialized with uniform distribution U[-1, 1] and i.i.d. B ∈ {0, 1} N ×k is the mask for matrix V , i,j B ij 0 ≤ N 2 , j B ij 0 ≤ N . Then, let the function of equivalent structure be g(s) = σ(( W B)σ(V s)), where input spiking s is a vector that s ∈ {0, 1} N . C is the constant depending on the supremum probability density of the dataset of the network. Then, ∀δ, ∃ when k ≥ N 2 N C th log N 2 δ-N C , there exists a mask matrix B that g(s) -σ( Ŵ s) = 0 w.p at least 1 -δ.\nStep 5: Network approximation (T = 1). The lemma is generalized to the whole SNN in Lemma C.5.\nStep 6: Network approximation (T > 1). If the output of two SNNs in the first T -1 timestep is consistent, and we assume that the error of spatial input features at all timesteps has an upperbound, there is a certain probability that the output of the two SNNs in timestep T is also the same. See detail proof in Theorem C.6 and detail theorem as follow Theorem 4.3:\nTheorem 4.3. All Steps Approximation. Fix the weight matrix Ŵ l ∈ [-1 √ N , 1 √ N ] N ×N\nwhich is the connection in target network between a layer of spiking inputs and the next layer of neurons. The equivalent structure is k spiking neurons with k×N weights V l ∈ R k×N connect the input and W l ∈ R N ×k connect out. All the weights wl ij and v l ij random initialized with uniform distribution U[-1, 1] and i.i.d.\nB l ∈ {0, 1} k×N is the mask for matrix V l , i,j B l ij 0 ≤ N 2 , j B l ij 0 ≤ N . Then, let the function of equivalent network at timestep t be Gt (S) = Gt,L • Gt,L-1 • • • • • Gt,1 (S) and Gt,l = σt (( W l B l )σ t (V l S t ))\n, where input spiking S is a tensor that S ∈ {0, 1} N ×T . And the target network at timestep\nt is Ĝt (S) = Ĝt,L • Ĝt,L-1 • • • • • Ĝt,1 (S), where Ĝt,l (S) = σt ( Ŵ l S t ). l = 1, 2, • • • , L, t = 1, 2, • • • , T .\nC is the constant depending on the supremum probability density of the dataset of the network. Then, ∀0 <δ≤ 1, ∃ > 0 whenk ≥ N 2 N C th log N 2 L δ-N CLT , there exists a mask matrix B that G(S) -Ĝ(S) = 0, w.p at least 1 -δ."
},
{
"figure_ref": [
"fig_2"
],
"heading": "Searching Winning Tickets from SNNs without Weight Training",
"publication_ref": [
"b10",
"b35",
"b10"
],
"table_ref": [],
"text": "By applying the top-k% sub-network searching (edge-popup) algorithm [11], we empirically verified SLTH as the complement of the theoretical proof in Section 4. Spiking neurons do not satisfy the Lipschitz condition, but the ANN technique can still be employed since the surrogate gradient of the SNN during BP training [36]. We first simply introduce the search algorithm, then show the results.\nThe sub-network search algorithm first proposed by [11] provides scores for every weight as a criterion for sub-network topology choosing. In each step of forward propagation, the top-k% score weights are chosen while others are masked by 0. It means the network used for classification is actually a sub-network of the original random initialized network. During the BP training process, the scores are randomly initialized as weights at the beginning, and then the gradient of the score is updated while the gradient of weights is not recorded (weights are not changed). The updating rule of scores can be formally expressed in mathematical notations as follow:\ns uv ← s uv + α ∂L ∂I v S u w uv ,(11)\nWhere s uv and w uv denote the score and corresponding weight that connect from spiking neuron u to v, respectively, S u ∈ {0, 1} indicates the output of neuron u, L and I v are loss function value of the network and the input value of the neuron v, and α is the learning rate. We apply the edge-popup in SNN-VGG-16 and Res-SNN-19. We set maskable weights for all layers (the first layer and the last layer are included). The hyper-parameter sparsity k% is set every 0.2 from 0.1 to 0.9. The datasets used to verify are CIFAR10/100. The network structures and other experiment details are shown in Appendix E. As shown in Fig. 3, the existence of good sub-networks is empirically verified, which is highly conformed with our theoretical analysis. The additional timesteps and discontinuous activation complicate the sub-network search problem in SNNs, but good sub-networks still exist without any weight training and can be found by specific algorithms."
},
{
"figure_ref": [
"fig_5",
"fig_14",
"fig_7"
],
"heading": "Pruning Criterion Based on Probabilistic Modeling",
"publication_ref": [
"b44",
"b45",
"b46",
"b47",
"b42",
"b43",
"b9",
"b42",
"b48",
"b20",
"b49"
],
"table_ref": [],
"text": "Discussion. For two spiking neurons, layers, or networks that have the same input but different weights, the proposed probabilistic modeling method can give the probability that the outputs (that is, the spiking firing) of both are the same. We convert the perturbation brought by pruning under the same input into the error of spatial input features, and then analyze its impact on spiking firing. Then, we prove that the lottery ticket hypothesis also holds in SNNs. The potential of probabilistic modeling goes far beyond that. Probabilistic modeling prompts us to rethink whether the existing pruning methods in SNNs are reasonable, since most methods directly continue the idea and methods of ANN pruning. Firstly, from the view of linear transformation, how to metrics the importance of weights in SNNs? The relationship between the inner state of the artificial neuron before activation and the input is usually not considered in ANN pruning. But we have to take this into account in binary-activated SNN pruning because the membrane potential is related to both weights and inputs. For instance, when the input spike frequency is 0, no matter how large the weight is, it will not affect the result of the linear transformation (Fig. 4). Secondly, from the view of nonlinear transformation, how does pruning affect the output of SNNs? As shown in Fig. 1b, when the membrane potential is in different intervals, the probability that the output of the spiking neuron is changed is also different. Thus, we study this important question theoretically using probabilistic modeling.\nMoreover, we can assume that the weights of the network remain unchanged, and add disturbance to the input, then exploit probabilistic modeling to analyze the impact of input perturbations on the output, i.e., analyze the robustness [45] of the SNNs in a new way. Similarly, probabilistic modeling can also analyze other compression methods in SNNs, such as quantization [46], tensor decomposition [47], etc. The perturbations brought about by these compression methods can be converted into errors in the membrane potential, which in turn affects the firing of spiking neurons.\nPruning criterion design. We design a new criterion to prune weights according to their probabilities of affecting the firing of spiking neurons. Based on Theorem 3.5, for each weight, its influence on the output of the spiking neuron has a probability upperbound P , which can be estimated as:\nP ≈ E(|u -u|)N (0|µ -u th , var),(12)\nwhere E(|u -u|) is the expectation of the error in the membrane potential brought about by pruning.\nIn this work, it is written as:\nE(|u -u|) = E act [|w|] |γ| σ 2 B + , (13\n)\nwhere E act [w] is the expectation that a weight is activated (the pre-synaptic spiking neuron outputs a spike). γ and σ B are a pair of hyper-parameters in Batch Normalization [48], which can also affect the linear transformation of weights, thus we incorporate them into Eq. ( 13). The detailed derivation of Eq. ( 13) can be found in Appendix D. Note, in keeping with the notations in classic BN, there is some notation mixed here, but only for Eq. ( 12) and Eq. ( 13). Next, following [43,44],\nhere we suppose the membrane potential follows Gaussian distribution N (µ, var). Consequently, N (0|µ -u th , var) represents the probability density when the membrane potential is at the threshold u th . Finally, for each weight, we have a probability P , and we prune according to the size of P . Experiments. Since our subject is to theoretically prove LTH in SNNs, we test the proposed new criteria using methods related to LTH. Note, Eq. ( 12) is a general pruning criterion and can also be incorporated in traditional non-LTH SNN pruning methods. Specifically, in this work we employ the LTH-based Iterative Magnitude Pruning (IMP) method for prune [10], whose criterion is to prune weights with small absolute values (algorithm details are given in Appendix F). Then, we change the criterion to pruning according to the size of P in Eq. ( 12).\nWe re-implement the pipeline network (Res-SNN-19 proposed in [43]) and IMP pruning method on the CIFAR-10/100 [49] datasets. We use the official code provided by authors in [21] for the whole pruning process. Then, we perform rigorous ablation experiments without bells and whistles, i.e., all experiment settings are same, the only change is to regulate the pruning criterion to Eq. ( 12). This is a IMP pruning based on Spiking Firing, thus we name it SF-IMP-1 (Appendix F). Moreover, the encoding layer (the first convolutional layer) in SNNs has a greater impact on task performance, and we can also allow the weights pruning in the encoding layer to be reactivated [50], called SF-IMP-2 (Appendix F).\nThe experimental results are shown in Fig. 5. First, these two approaches perform marginally differently on CIFAR-10/100, with SF-IMP-2 outperforming on CIFAR-10, and SF-IMP-1 performing better on CIFAR-100. Especially, compared with the vanilla IMP algorithm on CIFAR-100, SF-IMP-1 performs very well at all pruning sparsities. But on CIFAR-10, when the sparsity is high, the vanilla IMP looks better. One of the possible reasons is that the design of Eq. ( 12) still has room for improvement. These preliminary results demonstrate the effectiveness of the pruning criteria designed based on probabilistic modeling, and also lay the foundation for the subsequent design of more advanced pruning algorithms."
},
{
"figure_ref": [],
"heading": "Conclusions",
"publication_ref": [
"b16",
"b21",
"b23",
"b31"
],
"table_ref": [],
"text": "In this work, we aim to theoretically prove that the Lottery Ticket Hypothesis (LTH) also holds in SNNs, where the LTH theory has had a wide influence in traditional ANNs and is often used for pruning. In particular, spiking neurons employ binary discrete activation functions and have complex spatio-temporal dynamics, which are different from traditional ANNs. To address these challenges, we propose a new probabilistic modeling method for SNNs, modeling the impact of pruning on spiking firing, and then prove both theoretically and experimentally that LTH4 holds in SNNs. We design a new pruning criterion from the probabilistic modeling view and test it in an LTH-based pruning method with promising results. Both the probabilistic modeling and the pruning criterion are general. The former can also be used to analyze the robustness and other network compression methods of SNNs, and the latter can also be exploited in non-LTH pruning methods. In conclusion, we have for the first time theoretically established the link between membrane potential perturbations and spiking firing through the proposed novel probabilistic modeling approach, and we believe that this work will provide new inspiration for the field of SNNs.\nA Notations Proof. Since distribution p is v-Neighborhood-Finite Distribution, for any δ> 0,there exists a constant\n> 0so that x ∈ (v -, v + ) p sup = sup x p(x) ≥ p(x),(S.1)\nthus:\nv+ v- p(x)dx ≤ 2 p sup ,(S.2)\nwhen ≤ psup 2δ , the lemma holds.\nLemma B.2. Suppose the spiking neurons are u th -Neighborhood-Finite Distribution and the inputs of two corresponding spiking neuron got an error upperbound , and they got the same inner state h T -1 , the probability upperbound of different outputs is proportional to .\nFor norm, we use • 0 and • 2 . The • 0 is used to count the number of non-zero weights while the • 2 is used to measure the distance of outputs.\nFormally:\nFor two spiking neurons σT and σT , when hT -1 = ĥT -1 and ûT = ĥT -1 + xT is a random variable follows the u th -Neighborhood-Finite Distribution, if xT -xT ≤ , then:\nP σT (x T ) = σT (x T ) ∝\nProof. Since distribution of ûT = ĥT -1 + xT , denotes as p, is u th -Neighborhood-Finite Distribution, for any δ, exists that û ∈ (u th -, u th + )\np sup = sup ûT p(û T ) ≥ p(û T ),(S.3)\nsince ũT = hT -1 + xT = ûT + xT -xT , ûT -ũT ≤ , then:\nP σT (x T ) = σT (x T ) = P sign(û T -u th )sign(ũ T -u th ) = -1 ≤ u th + u th - p(û T )du ≤ 2p sup ,(S.4)\nthus the lemma holds.\nNote, for timestep T , the membrane potential will follow p(u T ; S 1:T , W ), where S 1:T ∈ {0, 1} N ×T is the input from timesteps 1 to T , W ∈ R N is the weight vector for the linear combination that shared at each timestep. When we do not emphasize the specific timestep and input data, we use p(u) for brevity.\nLemma B.3. Suppose the spiking neurons are u th -Neighborhood-Finite Distribution at timestep T and the inputs of two corresponding spiking neurons got an error upperbound at any timestep, and they got the same inner state h 0 , if there is no different output at the first T -1 timesteps, then the probability upperbound is proportional to 1-β .\nFormally:\nFor two spiking neurons σT and σT , when h0 = ĥ0 and ûT = ĥT\n-1 + xT is a random variable follows the u th -Neighborhood-Finite Distribution, if xt -xt ≤ and σt (x t ) = σt (x t ) for t = 1, 2, • • • , T -1, then: P σT (x T ) = σT (x T ) ∝ 1-β .\nProof. For each timestep i, we got:\nût = ĥt-1 + xt (S.5) ũt = ht-1 + xt ,(S.6)\nat timestep t = 1, we have:\nû1 = ũ1 ≤ , (S.7)\nif the same output of the former timestep t -1 is 1, there will be no error for inner state thus ĥt-1 = ht-1 , then ût -ũt ≤ , otherwise, the error will be ût -ũt ≤ + β ût-1 -ũt-1 .\nThus, by iterating, here is the upperbound error for membrane potential in the timestep T :\nûT -ũT ≤ T -1 n=0 β n ≤ +∞ n=0 β n ≤ 1 -β (S.8) then: P σT (x T ) = σT (x T ) ≤ 2p sup 1 -β ,(S.9)\nwhere p sup = sup ûT p(û T ) ≥ p(û T ).\nHere the lemma holds.\nTheorem B.4. Suppose the spiking layers are u th -Neighborhood-Finite Distribution at timestep T and the inputs of two corresponding spiking layers with a width N got an error upperbound for each element of input vectors at any timestep, and they got the same inner state vector h 0 , if there is no different output at the first T -1 timesteps, then the probability upperbound is proportional to N 1-β .\nFormally:\nFor two spiking layers σT and σT , when h0 = ĥ0 and ûT = ĥT -1 + xT is a random variable follows the\nu th -Neighborhood-Finite Distribution, if xt k -xt k ≤ (k = 1, 2, • • • , N ; t = 1, 2, • • • , T ) and σt (x t ) = σt (x t ) for t = 1, 2, • • • , T -1, then: P σT (x T ) = σT (x T ) ∝ N 1-β .\nProof. Here, we define p sup as the upperbound probability density of all entries at timestep T . Thus:\np sup = sup ûT k p(û T k ),(S.10)\nthen, according to Lemma 3.4, for any single entry we have:\nP σT k (x T k ) = σT k (x T k ) ≤ 2 1 -β p sup (S.11)\nthen, according to the union bound inequality: √ N ] which is the connection in target network between two neurons. The equivalent structure is k spiking neurons with Proof. For k spiking neurons in equivalent structure, since b 0 ≤ 1, only one spiking neuron is active while others are pruned out with their weights. For the chosen active neuron, the weights wi and v i should satisfy the following sufficient conditions:\nP ∃k, σT k (x T k ) = σT k (x T k ) ≤ N 2 1 -β p sup (S.\nk weights v = [v 1 , v 2 , • • • , v k ] T connect\n• wi -ŵi ≤\n• v i ≥ v th Since w, v ∼ U[-1, 1], P [ wi -ŵi ≤ ] = 2 2 = (S.13) P [v i ≥ v th ] = 1 -v th 2 = C th (S.14)\nC th is the constant that C th = 1-v th 2 . The probability for a single spiking neuron that satisfies the condition is:\nP [ wi -ŵi ≤ ∧ v i ≥ v th ] =P [ wi -ŵi ≤ ] P [v i ≥ v th ] =C th (S.15)\nThe probability of not satisfying the condition for all k spiking neurons is:\nP [∀i ∈ [k], ¬( wi -ŵi ≤ ∧ v i ≥ v th )] =(1 -C th ) k ≤ exp (-kC th ) ≤ δ (S.16)\nThus, when k ≥ 1 C th log 1 δ , the active spiking neuron satisfies the condition with probability at least 1 -δ. Proof. Define the event: 17) this event means in a block of approximation structures with k structures to approximate the connecting weight from the i-th input to the output, but no one structure satisfies the approximation error condition.\nE i,k , = {∀a ∈ [k ], ¬( w(i-1)k +a -ŵi ≤ ∧ v (i-1)k +a,i ≥ u th )}, (S.\nThus P (E i,k , N ) ≤ exp(-k C th N ), (S.18)\nWe totally have N blocks thus k = N k and for each block, using union bound inequality, we have: √ N ] N ×N which is the connection in target network between a layer of spiking inputs and the next layer of neurons. The equivalent structure is k spiking neurons with k × N weights V ∈ R k×N connect the input and W ∈ R N ×k connect out. All the weights wij and v ij random initialized with uniform distribution U[-1, 1] and i.i.d. B ∈ {0, 1} N ×k is the mask for matrix W , i,j B ij 0 ≤ N 2 , j B ij 0 ≤ N . Then, let the function of equivalent structure be g(s) = ( W B)σ(V s), where input spiking s is a vector that s ∈ {0, 1} N . Then, ∀0 <δ≤ 1, ∃ > 0 when\nP ( i E i,k , N ) ≤ i P (E i,k , N ) ≤ N exp(-k C th N ) ≤ δ, (S.19) thus: P (( i E i,k , N ) C ) ≥ 1 -δ, (S.\nk ≥ N N C th log N 2 δ\nand C th = 1-u th 2 , there exists a mask matrix B that\n[g(s)] i -[ Ŵ s] i ≤\nw.p at least 1 -δ.\nProof. Define the event: 22) this event means in a block of approximation structures with k structures to approximate the connecting weight from the i-th input to the j-th output, but no one structure satisfies the approximation error condition.\nE j,i,k , = {∀a ∈ [k ], ¬( wj,(i-1)k +a -ŵji ≤ ∧ v (i-1)k +a,i ≥ u th )}, (S.\nThus\nP (E j,i,k , N ) ≤ exp(-k C th N ),(S.23)\nWe totally have N blocks thus k = N k and for each block, using union bound inequality, we have: 24) thus: P ((\nP ( j,i E j,i,k , N ) ≤ j,i P (E j,i,k , N ) ≤ N 2 exp(-k C th N ) ≤ δ, (S.\nj,i E j,i,k , N ) C ) ≥ 1 -δ, (S.25)\nwhere\nk ≥ N N C th log N 2 δ (S.26)\nthen the lemma holds.\nLemma C.4. Layer Spiking Activation Approximation.\nFix the weight matrix Ŵ ∈ [-1 √ N , 1 √ N ] N ×N\nwhich is the connection in target network between a layer of spiking inputs and the next layer of neurons. The equivalent structure is k spiking neurons with k × N weights V ∈ R k×N connect the input and W ∈ R N ×k connect out. All the weights wij and v ij random initialized with uniform distribution U[-1, 1] and i.i.d. B ∈ {0, 1} N ×k is the mask for matrix V , i,j B ij 0 ≤ N 2 , j B ij 0 ≤ N . Then, let the function of equivalent structure be g(s) = σ(( W B)σ(V s)), where input spiking s is a vector that s ∈ {0, 1} N . C is the constant depending on the supremum probability density of the dataset of the network. Then, ∀0 <δ≤ 1, ∃ > 0 when\nk ≥ N 2 N C th log N 2 δ -N C\n, there exists a mask matrix B that g(s) -σ( Ŵ s) = 0 w.p at least 1 -δ.\nProof. The definition of E j,i,k , follows the proof of LemmaLemma C.3, thus we have:\nP ( j,i E j,i,k , N ) ≤ j,i P (E j,i,k , N ) ≤ N 2 exp(k C th N ) (S.27)\nAnd the event ( j,i E j,i,k , N ) C implies that the error of each channel smaller than .\nAccording to the proof of theoremTheorem 3.5, the probability upperbound of different output of spiking neurons with the same temporal state at t = 0 is 2N p sup 1-β . We define E f ire as the event of different output of corresponding spiking layer.\nThen, according to union bound, we have:\nP (( j,i E j,i,k , N ) E f ire ) ≤ j,i P (E j,i,k , N ) + P (E f ire ) ≤ N 2 exp(-k C th N ) + N p sup 2 1 -β ≤ δ, (S.28) Let C = 2p sup1\n1-β , then, we have:\nk ≥ N 2 N C th log N 2 δ -N C (S.29)\nHere, the event (( j,i E j,i,k , N ) E f ire ) C implies that the error of each entry is smaller than while the output have no difference.\nP (((\nj,i E j,i,k , N ) E f ire ) C ) ≥ 1 -δ,(S.30)\nthe lemma holds.\nLemma C.5. All Layers Approximation. Fix the weight matrix\nŴ l ∈ [-1 √ N , 1 √ N ]\nN ×N which is the connection in target network between a layer of spiking inputs and the next layer of neurons. The equivalent structure is k spiking neurons with k × N weights V l ∈ R k×N connect the input and W l ∈ R N ×k connect out. All the weights wl ij and v l ij random initialized with uniform distribution U[-1, 1] and i.i.d. B l ∈ {0, 1} k×N is the mask for matrix V , i,j B l ij 0 ≤ N 2 , j B l ij 0 ≤ N . Then, let the function of equivalent network be G\n(s) = GL • GL-1 • • • • • G1 (s) and Gl = σ(( W l B l )σ(V l s))\n, where input spiking s is a vector that s ∈ {0, 1} N . And the target network is Ĝ\n(s) = ĜL • ĜL-1 • • • • • Ĝ1 (s), where Ĝl (s) = σ( Ŵ l s). l = 1, 2, • • • , L.\nC is the constant depending on the supremum probability density of the dataset of the network. Then, ∀0 <delta≤ 1, ∃ > 0 when\nk ≥ N 2 N C th log N 2 L δ -N CL ,\nthere exists a mask matrix B that G(s) -Ĝ(s) = 0\nw.p at least 1 -δ.\nProof. We inherit the event expression E j,i,k , and E f ire from the proof of LemmaLemma C.4 with a delight modify. Here we add subscript l to denote the layer of the target network. Thus we have: 32) and: P (((\nP (( l,j,i E l,j,i,k , N ) E f ire,l ) ≤ l,j,i P (E l,j,i,k , N ) + l P (E f ire,l ) ≤ LN 2 exp(-k C th N ) + LN C ≤δ, (S.31) Thus, k ≥ N 2 N C th log N 2 L δ -N CL , (S.\nl,j,i E l,j,i,k , N ) E f ire,l ) C ) ≥ 1 -δ (S.33)\nthe lemma holds.\nTheorem C.6. All Steps Approximation. Fix the weight matrix\nŴ l ∈ [-1 √ N , 1 √ N ] N ×N\nwhich is the connection in target network between a layer of spiking inputs and the next layer of neurons. The equivalent structure is k spiking neurons with k×N weights V l ∈ R k×N connect the input and W l ∈ R N ×k connect out. All the weights wl ij and v l ij random initialized with uniform distribution U[-1, 1] and i.i.d. B l ∈ {0, 1} k×N is the mask for matrix V l , i,j B l ij 0 ≤ N 2 , j B l ij 0 ≤ N . Then, let the function of equivalent network at timestep t be Gt\n(S) = Gt,L • Gt,L-1 • • • • • Gt,1 (S) and Gt,l = σt (( W l B l )σ t (V l S t ))\n, where input spiking S is a tensor that S ∈ {0, 1} N ×T . And the target network at timestep t is Ĝt\n(S) = Ĝt,L • Ĝt,L-1 • • • • • Ĝt,1 (S), where Ĝt,l (S) = σt ( Ŵ l S t ). l = 1, 2, • • • , L, t = 1, 2, • • • , T .\nC is the constant depending on the supremum probability density of the dataset of the network. Then, ∀0 <δ≤ 1, ∃ > 0 when\nk ≥ N 2 N C th log N 2 L δ -N CLT ,\nthere exists a mask matrix B that G(S) -Ĝ(S) = 0\nw.p at least 1 -δ.\nProof. We inherit the event expression E l,j,i,k , and E f ire,l from the proof of LemmaLemma C.5 with a delight modify. Here we add subscript t to E f ire,l to denote the layer of the target network. Thus we have:\nP (( t,l,j,i E t,l,j,i,k , N ) E f ire,t,l ) ≤ t,l,j,i P (E t,l,j,i,k , N ) + t,l P (E f ire,t,l ) ≤LN 2 exp(-k C th N ) + T LN C ≤δ, (S.34) Thus, k ≥ N 2 N C th log N 2 L δ -N CT L , (S.35)\nand:\nP ((( l,j,i E t,l,j,i,k , N ) E f ire,t,l ) C ) ≥ 1 -δ (S.36)\nthe lemma holds."
},
{
"figure_ref": [],
"heading": "D Derivation of Equation (13)",
"publication_ref": [
"b36",
"b47",
"b42",
"b43",
"b47",
"b20",
"b10",
"b49"
],
"table_ref": [],
"text": "We design a new criterion to prune weights according to their probabilities of affecting the output of spiking neurons. Based on the probabilistic modeling (Theorem 3.5), for each weight, its influence on the output of the spiking neuron has a probability upperbound P , which can be estimated as:\nP ≈ E(|u -u|)N (0|µ -u th , var), (S. 37) where E(|u -u|) is the expectation of the error in the membrane potential brought about by pruning (i.e., effect of weights on linear transformations). In our method, it is written as:\nE(|u -u|) = E act [w]γ σ 2 B + , (S.38)\nwhere E act [w] is the expectation that a weight is activated (the pre-synaptic spiking neuron outputs a spike), γ and σ B are a pair of hyper-parameters in Batch Normalization [48]. The detailed derivation of Eq. (S.38) can be found in Appendix D. Next, following [43,44], here we suppose the membrane potential follows Gaussian distribution N (µ, var). Consequently, N (0|µ -u th , var) represents the probability density when the membrane potential is at the threshold u th (i.e., effect of weights on binary nonlinear transformations). Finally, for each weight, we have a probability P , and we prune according to the size of P (from small to large). Now we explain that how we get the Eq. (S.38), i.e., Eq. ( 13) in the main text. The expression of Batch Normalization (BN) can be written as: Note, in keeping with the notations in classic BN [48], there is some notation mixed here, but only for the derivation of Eq. ( 13). All training methods strictly follow experiments in [21], and the sub-network search module follows [11]. For the two datasets of CIFAR10/100, we use cosine learning scheduling and SGD optimizer with momentum 0.9 and weight decay 5e-4, and the total number of training epochs is 300. The learning rate is set to 0.3. batch size is set to 128. The timestep T of SNN is 5. We simply replace all the weight modules in network with the corresponding sub-network search modules in hidden-networks/blob/master/simple_mnist_example.py F Iterative Magnitude Pruning (IMP) and the proposed SF-IMP Algorithms Among the LTH-based pruning algorithms, the Iterative Magnitude Pruning (IMP) method has good performance. In IMP, the parameter θ ∈ R n of network f (x; θ) is pruned iteratively. (For the sake of briefness and convention of the symbols, unless otherwise specified, the meanings of the symbols in this chapter have nothing to do with the previous chapters.) We set K iterations. In the k-th iteration, we first train the network till convergence, then prune p% of nonzero parameters of θ trained m k-1 by mask m k ∈ {0, 1} n . Then reinitialize the network with parameter θ init m k and repeat the operations until the K-th iteration ends.\nx i ← x i -µ B\nFor SF-IMP-1, we change the criterion from considering magnitude to:\nE act [|w ij |] |γ j | σ 2 Bj +\nN (0|µ j -u th , var j ), (S.43)\nHere, i is the index of the input channel while j is the index of the output channel. We statistic the firing frequency for each input channel to evaluate E act [|w ij |], and other parameters γ j , σ Bj , µ j , var j are statistic by every output channel. However, the encoding layer and the Fully Connected (FC) layer are not suited for this algorithm, thus we keep the magnitude criterion for these two layers.\nFor SF-IMP-2, since the sparsity of the encoding layer and the FC layer have a great influence on pruning, we apply a dynamic strategy for these layers [50]. Specifically, we only mask the weight when computing loss, while when updating weight and re-initialize weight, the mask is not used."
}
] | The Lottery Ticket Hypothesis (LTH) states that a randomly-initialized large neural network contains a small sub-network (i.e., winning tickets) which, when trained in isolation, can achieve comparable performance to the large network. LTH opens up a new path for network pruning. Existing proofs of LTH in Artificial Neural Networks (ANNs) are based on continuous activation functions, such as ReLU, which satisfying the Lipschitz condition. However, these theoretical methods are not applicable in Spiking Neural Networks (SNNs) due to the discontinuous of spiking function. We argue that it is possible to extend the scope of LTH by eliminating Lipschitz condition. Specifically, we propose a novel probabilistic modeling approach for spiking neurons with complicated spatio-temporal dynamics. Then we theoretically and experimentally prove that LTH holds in SNNs. According to our theorem, we conclude that pruning directly in accordance with the weight size in existing SNNs is clearly not optimal. We further design a new criterion for pruning based on our theory, which achieves better pruning results than baseline. | Probabilistic Modeling: Proving the Lottery Ticket Hypothesis in Spiking Neural Network | [
{
"figure_caption": "2 = 0 ,20where t = 1, 2, • • • , T is the timestep, S is the input spiking tensor containing only 0 or 1.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "20",
"figure_type": "figure"
},
{
"figure_caption": "DendriteObjective:Figure 1 :1Figure 1: Overview of this work. (a), The goal of spiking neuron approximate. (b),The firing behavior in the approximation of spiking neurons can be described by the proposed probabilistic modeling approach (Section 3). Spikes are fired when the membrane potential u exceeds the firing threshold u th . As long as a weight change induces u to fall into the crisis neighborhood, there is a certain probability that the spiking firing will be changed (0 to 1, or 1 to 0). We hope that spiking firing will not change after the redundant weights are pruned. (c), Equivalent structure modeling. The error between the target and equivalent SNN is related to the network width (Section 4). (d), Proof of LTH in SNN (Section 4). (e), We provide a pruning technique for SNNs (Section 6). New pruning criterion: compute the probability that the firing of a spiking neuron changes when weights are pruned, and pruning according to the rank of the probability.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Definition 3 .31. m-Neighborhood-Finite Distribution. For a probability density function p(•) and a value m, if there exists > 0 for the neighborhood [m -, m + ], in this interval, the max value of function p(•) is finite, we call the distribute function p(•) as m-Neighborhood-Finite Distribution.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: The horizontal and vertical coordinates represent the sparsity and accuracy. The solid line is the untrained subnetwork. The dashed line is the trained large network (sparsity=0%).",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: The importance of weight needs to consider both the magnitude of the weight and the firing of neurons.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 :5Figure 5: Observations from Iterative Magnitude Pruning (IMP) with the proposed pruning criterion P in eq. (12).",
"figure_data": "",
"figure_id": "fig_7",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "v+ v-p(x)dx ≤ δ.",
"figure_data": "",
"figure_id": "fig_8",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "12 )C 1 √ N , 11211Proofs of Lottery Ticket Hypothesis in SNNs Lemma C.1. Single Weight Approximation. Fix the weight scalar ŵ ∈ [-",
"figure_data": "",
"figure_id": "fig_9",
"figure_label": "1211",
"figure_type": "figure"
},
{
"figure_caption": "1 C th log 1 δ11the input and w = [ w1 , w2 , • • • , wk ] T connect out. All the weights wi and v i random initialized with uniform distribution U[-1, 1] and i.i.d. b = [b 1 , b 2 , • • • , b k ] is the mask for weight vector v, and b ∈ {0, 1} k , b 0 ≤ 1. Then, let the function of equivalent structure be g(s) = ( w b) T σ(vs), where input spiking s is a scalar that s ∈ {0, 1}. Then, ∀0 <δ≤ 1, ∃ > 0 when k ≥ and C th = 1-u th 2 , there exists a mask vector b that g(s) -ŵs ≤ w.p at least 1 -δ.",
"figure_data": "",
"figure_id": "fig_10",
"figure_label": "11",
"figure_type": "figure"
},
{
"figure_caption": "Lemma C. 2 . 1 √ N , 1 √N211Layer Weights Approximation. Fix the weights vector ŵ ∈ [-] N which is the connection in target network between a layer of spiking inputs and a neuron. The equivalent structure is k spiking neurons with k × N weights V ∈ R k×N connect the input and w = [ w1 , w2 , • • • , wk ] T connect out. All the weights wi and v ij random initialized with uniform distribution U[-1, 1] and i.i.d. b ∈ {0, 1} k is the mask for matrix w, b 0 ≤ N . Then, let the function of equivalent structure be g(s) = ( w b) T σ(V s), where input spiking s is a vector that s ∈ {0, 1} N . Then, ∀0 <δ≤ 1, ∃ > 0 when k ≥ N N C th log N δ and C th = 1-u th 2 , there exists a mask vector b that g(s) -ŵT s ≤ w.p at least 1 -δ.",
"figure_data": "",
"figure_id": "fig_11",
"figure_label": "211",
"figure_type": "figure"
},
{
"figure_caption": "Lemma C. 3 . 1 √ N , 1311Layer to Layer Approximation. Fix the weight matrix Ŵ ∈ [-",
"figure_data": "",
"figure_id": "fig_12",
"figure_label": "311",
"figure_type": "figure"
},
{
"figure_caption": "y i ← γ x i + β ≡ BN γ,β (x i ).(S.40) When a spike s passes through a weight in convolution layer and the BN layer, the output is: these two operations can be regard as linear transformation, where the input spike s multiplies the scaling weight wγ √σ 2 B +and add a bias( -µ B γ √ σ 2 B + + β).Since the bias will not change after changing the weight, the error of membrane potential only related to the scaling weight wγ √ design E(|u -u|) as follows:E(|u -u|) = E(",
"figure_data": "",
"figure_id": "fig_13",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Algorithm 11Rewind Iterative Magnitude Pruning (IMP). Input: pruning rate p, iterations K, the rewind epoch R, max training epoch N Train network for R epochs. Save the parameters as rewind parameters θ rewind Train network for n epochs. for k = 1 to K do Prune network by remove p% of the lowest magnitude nonzero weights of the network base on a metrics, get the mask m k Reload the rewind parameters and mask them. (θ rewind m k ) Train the network for N epochs. end for E Implementation Details of Sub-Network Search",
"figure_data": "",
"figure_id": "fig_14",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "4: Layer spiking activation approximation. Compared to the Step 3, we here include spiking activations to evaluate the probability that two different spiking layers (same input, different weights) have the same output. See Lemma C.4 for proof Lemma 4.2. Layer Spiking Activation Approximation.Fix the weight matrix Ŵ",
"figure_data": "",
"figure_id": "tab_1",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Important Notations.",
"figure_data": "SymbolDefinitiontTimestep indexlLayer indexiNeuron indexu t,l i h t,l i s t,l i x t,l iMembrane potential of spiking neuron Temporal input of spiking neuron Spatial input of spiking neuron Spatial input feature of spiking neuronu thFiring thresholdHea(•)Heaviside step functionVresetReset membrane potential after firing a spikeβDecay factorw l ijWeight connect from two spiking neuronu thFiring thresholdxSpatial input feature of a spiking neuron without indexesxSpatial input feature after adding perturbationuMembrane potential of a spiking neuron without indexesuMembrane potential after adding perturbationσ t x 1:t-1 σ NMembrane potential depends on all previous spatial input features Shorthand for activation function σ t x 1:t-1 Width of networkTTotal timestepGLarge networkG ll-th layer of large networkĜTarget networkGEquivalent networks tSpatial input vector at timestep tsSpatial input vector without timestep indexS tSpatial input tensor at timestep tS 1:tAll spatial input tensors before timestep tSShorthand for S 1:tSDatasetvVirtual layer weightVWeight matrix of Virtual layerbMask of weight vectorBMask of weight matrixW lWeight matrix of l-th layerC thA constant related to the hyper-parameter u thCA constant related to the dataset S and threshold u thEEventkNetwork width required for LTH establishmentB Proofs of Probabilistic ModelingLemma B.1. For the probability density function p(•), if it is v-Neighborhood-Finite Distribution,for any δ> 0, there exists > 0 that",
"figure_id": "tab_2",
"figure_label": "1",
"figure_type": "table"
}
] | Man Yao; Yuhong Chou; Guangshe Zhao; Xiawu Zheng; Yonghong Tian; Bo Xu; Guoqi Li | [
{
"authors": "Wolfgang Maass",
"journal": "Neural Networks",
"ref_id": "b0",
"title": "Networks of spiking neurons: The third generation of neural network models",
"year": "1997"
},
{
"authors": "Wulfram Gerstner; M Werner; Richard Kistler; Liam Naud; Paninski",
"journal": "Cambridge University Press",
"ref_id": "b1",
"title": "Neuronal dynamics: From single neurons to networks and models of cognition",
"year": "2014"
},
{
"authors": "Kaushik Roy; Akhilesh Jaiswal; Priyadarshini Panda",
"journal": "Nature",
"ref_id": "b2",
"title": "Towards spike-based machine intelligence with neuromorphic computing",
"year": "2019"
},
{
"authors": "Man Yao; Guangshe Zhao; Hengyu Zhang; Yifan Hu; Lei Deng; Yonghong Tian; Bo Xu; Guoqi Li",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b3",
"title": "Attention spiking neural networks",
"year": "2023"
},
{
"authors": "Mike Davies; Narayan Srinivasa; Tsung-Han Lin; Gautham Chinya; Yongqiang Cao; Sri Harsha Choday; Georgios Dimou; Prasad Joshi; Nabil Imam; Shweta Jain",
"journal": "IEEE Micro",
"ref_id": "b4",
"title": "Loihi: A neuromorphic manycore processor with on-chip learning",
"year": "2018"
},
{
"authors": "Jing Pei; Lei Deng",
"journal": "Nature",
"ref_id": "b5",
"title": "Towards artificial general intelligence with hybrid tianjic chip architecture",
"year": "2019"
},
{
"authors": " Catherine D Schuman; R Shruti; Maryam Kulkarni; Parker Parsa; Bill Mitchell; Kay",
"journal": "Nature Computational Science",
"ref_id": "b6",
"title": "Opportunities for neuromorphic computing algorithms and applications",
"year": "2022"
},
{
"authors": "Song Han; Jeff Pool; John Tran; William Dally",
"journal": "Advances in neural information processing systems",
"ref_id": "b7",
"title": "Learning both weights and connections for efficient neural network",
"year": "2015"
},
{
"authors": "Torsten Hoefler; Dan Alistarh; Tal Ben-Nun; Nikoli Dryden; Alexandra Peste",
"journal": "J. Mach. Learn. Res",
"ref_id": "b8",
"title": "Sparsity in deep learning: Pruning and growth for efficient inference and training in neural networks",
"year": "2021"
},
{
"authors": "Jonathan Frankle; Michael Carbin",
"journal": "",
"ref_id": "b9",
"title": "The lottery ticket hypothesis: Finding sparse, trainable neural networks",
"year": "2019"
},
{
"authors": "Mitchell Vivek Ramanujan; Aniruddha Wortsman; Ali Kembhavi; Mohammad Farhadi; Rastegari",
"journal": "",
"ref_id": "b10",
"title": "What's hidden in a randomly weighted neural network",
"year": "2020"
},
{
"authors": "Eran Malach; Gilad Yehudai; Shai Shalev-Schwartz; Ohad Shamir",
"journal": "PMLR",
"ref_id": "b11",
"title": "Proving the lottery ticket hypothesis: Pruning is all you need",
"year": "2020"
},
{
"authors": "Hattie Zhou; Janice Lan; Rosanne Liu; Jason Yosinski",
"journal": "Advances in neural information processing systems",
"ref_id": "b12",
"title": "Deconstructing lottery tickets: Zeros, signs, and the supermask",
"year": "2019"
},
{
"authors": "Yulong Wang; Xiaolu Zhang; Lingxi Xie; Jun Zhou; Hang Su; Bo Zhang; Xiaolin Hu",
"journal": "",
"ref_id": "b13",
"title": "Pruning from scratch",
"year": "2020"
},
{
"authors": "Laurent Orseau; Marcus Hutter; Omar Rivasplata",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b14",
"title": "Logarithmic pruning is all you need",
"year": "2020"
},
{
"authors": "Ankit Pensia; Shashank Rajput; Alliot Nagle; Harit Vishwakarma; Dimitris Papailiopoulos",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b15",
"title": "Optimal lottery tickets via subset sum: Logarithmic over-parameterization is sufficient",
"year": "2020"
},
{
"authors": "Arthur Da Cunha; Emanuele Natale; Laurent Viennot",
"journal": "",
"ref_id": "b16",
"title": "Proving the strong lottery ticket hypothesis for convolutional neural networks",
"year": "2022"
},
{
"authors": "Haoran You; Chaojian Li; Pengfei Xu; Yonggan Fu; Yue Wang; Xiaohan Chen; Richard G Baraniuk; Zhangyang Wang; Yingyan Lin",
"journal": "",
"ref_id": "b17",
"title": "Drawing early-bird tickets: Towards more efficient training of deep networks",
"year": "2019"
},
{
"authors": "Sharath Girish; Kamal Shishira R Maiya; Hao Gupta; Larry S Chen; Abhinav Davis; Shrivastava",
"journal": "",
"ref_id": "b18",
"title": "The lottery ticket hypothesis for object recognition",
"year": "2021"
},
{
"authors": "Tianlong Chen; Jonathan Frankle; Shiyu Chang; Sijia Liu; Yang Zhang; Zhangyang Wang; Michael Carbin",
"journal": "Advances in neural information processing systems",
"ref_id": "b19",
"title": "The lottery ticket hypothesis for pre-trained bert networks",
"year": "2020"
},
{
"authors": "Youngeun Kim; Yuhang Li; Hyoungseob Park; Yeshwanth Venkatesha; Ruokai Yin; Priyadarshini Panda",
"journal": "Springer Nature Switzerland",
"ref_id": "b20",
"title": "Exploring lottery ticket hypothesis in spiking neural networks",
"year": "2022"
},
{
"authors": "Lei Deng; Yujie Wu; Xing Hu; Ling Liang; Yufei Ding; Guoqi Li",
"journal": "Neural Networks",
"ref_id": "b21",
"title": "Rethinking the performance comparison between snns and anns",
"year": "2020"
},
{
"authors": "Man Yao; Huanhuan Gao; Guangshe Zhao; Dingheng Wang; Yihan Lin; Zhaoxu Yang; Guoqi Li",
"journal": "",
"ref_id": "b22",
"title": "Temporal-wise attention spiking neural networks for event streams classification",
"year": "2021-10"
},
{
"authors": "Arjun Rao; Philipp Plank; Andreas Wild; Wolfgang Maass",
"journal": "Nature Machine Intelligence",
"ref_id": "b23",
"title": "A long short-term memory for ai applications in spike-based neuromorphic hardware",
"year": "2022"
},
{
"authors": "Mike Davies; Andreas Wild; Garrick Orchard; Yulia Sandamirskaya; Gabriel A Fonseca; Prasad Guerra; Philipp Joshi; Plank; Sumedh R Risbud",
"journal": "Proceedings of the IEEE",
"ref_id": "b24",
"title": "Advancing neuromorphic computing with loihi: A survey of results and outlook",
"year": "2021"
},
{
"authors": "Bruno U Emre O Neftci; Siddharth Pedroni; Maruan Joshi; Gert Al-Shedivat; Cauwenberghs",
"journal": "Frontiers in neuroscience",
"ref_id": "b25",
"title": "Stochastic synapses enable efficient brain-inspired learning machines",
"year": "2016"
},
{
"authors": "Nitin Rathi; Priyadarshini Panda; Kaushik Roy",
"journal": "IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems",
"ref_id": "b26",
"title": "Stdp-based pruning of connections and weight quantization in spiking neural networks for energy-efficient recognition",
"year": "2018"
},
{
"authors": "Bharadwaj Thao Nn Nguyen; Xuanyao Veeravalli; Fong",
"journal": "",
"ref_id": "b27",
"title": "Connection pruning for deep spiking neural networks with on-chip learning",
"year": "2021"
},
{
"authors": "Yuhan Shi; Leon Nguyen; Sangheon Oh; Xin Liu; Duygu Kuzum",
"journal": "Frontiers in neuroscience",
"ref_id": "b28",
"title": "A soft-pruning method applied during training of spiking neural networks for in-memory computing applications",
"year": "2019"
},
{
"authors": "Wenzhe Guo; Mohammed E Fouda; Hasan Erdem Yantir; Ahmed M Eltawil; Khaled Nabil; Salama ",
"journal": "Frontiers in Neuroscience",
"ref_id": "b29",
"title": "Unsupervised adaptive weight pruning for energy-efficient neuromorphic systems",
"year": "2020"
},
{
"authors": "Souvik Kundu; Gourav Datta; Massoud Pedram; Peter A Beerel",
"journal": "",
"ref_id": "b30",
"title": "Spike-thrift: Towards energy-efficient deep spiking neural networks by limiting spiking activity via attention-guided compression",
"year": "2021"
},
{
"authors": "David Kappel; Stefan Habenschuss; Robert Legenstein; Wolfgang Maass",
"journal": "PLoS computational biology",
"ref_id": "b31",
"title": "Network plasticity as bayesian inference",
"year": "2015"
},
{
"authors": "Guillaume Bellec; Darjan Salaj; Anand Subramoney; Robert Legenstein; Wolfgang Maass",
"journal": "Advances in neural information processing systems",
"ref_id": "b32",
"title": "Long short-term memory and learning-to-learn in networks of spiking neurons",
"year": "2018"
},
{
"authors": "Yanqi Chen; Zhaofei Yu; Wei Fang; Tiejun Huang; Yonghong Tian",
"journal": "",
"ref_id": "b33",
"title": "Pruning of deep spiking neural networks through gradient rewiring",
"year": "2021"
},
{
"authors": "Yanqi Chen; Zhaofei Yu; Wei Fang; Zhengyu Ma; Tiejun Huang; Yonghong Tian",
"journal": "PMLR",
"ref_id": "b34",
"title": "State transition of dendritic spines improves learning of sparse spiking neural networks",
"year": "2022"
},
{
"authors": "Yujie Wu; Lei Deng; Guoqi Li; Jun Zhu; Luping Shi",
"journal": "Frontiers in neuroscience",
"ref_id": "b35",
"title": "Spatio-temporal backpropagation for training high-performance spiking neural networks",
"year": "2018"
},
{
"authors": "Hesham Emre O Neftci; Friedemann Mostafa; Zenke",
"journal": "IEEE Signal Processing Magazine",
"ref_id": "b36",
"title": "Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks",
"year": "2019"
},
{
"authors": "Zhaodong Chen; Lei Deng; Bangyan Wang; Guoqi Li; Yuan Xie",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b37",
"title": "A comprehensive and modularized statistical framework for gradient norm equality in deep neural networks",
"year": "2020"
},
{
"authors": "Xingting Yao; Fanrong Li; Zitao Mo; Jian Cheng",
"journal": "",
"ref_id": "b38",
"title": "Glif: A unified gated leaky integrate-andfire neuron for spiking neural networks",
"year": "2022"
},
{
"authors": "Ahmed Shaban; Sai Sukruth Bezugam; Manan Suri",
"journal": "Nature Communications",
"ref_id": "b39",
"title": "An adaptive threshold neuron for recurrent spiking neural networks with nanodevice hardware implementation",
"year": "2021"
},
{
"authors": "Wei Fang; Zhaofei Yu; Yanqi Chen; Timothée Masquelier; Tiejun Huang; Yonghong Tian",
"journal": "",
"ref_id": "b40",
"title": "Incorporating learnable membrane time constant to enhance learning of spiking neural networks",
"year": "2021"
},
{
"authors": "Daniel Peter U Diehl; Jonathan Neil; Matthew Binas; Shih-Chii Cook; Michael Liu; Pfeiffer",
"journal": "ieee",
"ref_id": "b41",
"title": "Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing",
"year": "2015"
},
{
"authors": "Hanle Zheng; Yujie Wu; Lei Deng; Yifan Hu; Guoqi Li",
"journal": "",
"ref_id": "b42",
"title": "Going deeper with directly-trained larger spiking neural networks",
"year": "2021"
},
{
"authors": "Yufei Guo; Xinyi Tong; Yuanpei Chen; Liwen Zhang; Xiaode Liu; Zhe Ma; Xuhui Huang",
"journal": "",
"ref_id": "b43",
"title": "Recdis-snn: Rectifying membrane potential distribution for directly training spiking neural networks",
"year": "2022"
},
{
"authors": "Souvik Kundu; Massoud Pedram; Peter A Beerel",
"journal": "",
"ref_id": "b44",
"title": "Hire-snn: Harnessing the inherent robustness of energy-efficient deep spiking neural networks by training with crafted input noise",
"year": "2021"
},
{
"authors": "Lei Deng; Yujie Wu; Yifan Hu; Ling Liang; Guoqi Li; Xing Hu; Yufei Ding; Peng Li; Yuan Xie",
"journal": "IEEE transactions on neural networks and learning systems",
"ref_id": "b45",
"title": "Comprehensive snn compression using admm optimization and activity regularization",
"year": "2021"
},
{
"authors": "Dingheng Wang; Bijiao Wu; Guangshe Zhao; Man Yao; Hengnu Chen; Lei Deng; Tianyi Yan; Guoqi Li",
"journal": "IEEE Transactions on Neural Networks and Learning Systems",
"ref_id": "b46",
"title": "Kronecker cp decomposition with fast multiplication for compressing rnns",
"year": "2021"
},
{
"authors": "Sergey Ioffe; Christian Szegedy",
"journal": "PMLR",
"ref_id": "b47",
"title": "Batch normalization: Accelerating deep network training by reducing internal covariate shift",
"year": "2015"
},
{
"authors": "Alex Krizhevsky; Geoffrey Hinton",
"journal": "",
"ref_id": "b48",
"title": "Learning multiple layers of features from tiny images",
"year": "2009"
},
{
"authors": "Yiwen Guo; Anbang Yao; Yurong Chen",
"journal": "Advances in neural information processing systems",
"ref_id": "b49",
"title": "Dynamic network surgery for efficient dnns",
"year": "2016"
}
] | [
{
"formula_coordinates": [
2,
225.43,
315.26,
57.03,
30.2
],
"formula_id": "formula_0",
"formula_text": "sup S∈S 1 T T t=1"
},
{
"formula_coordinates": [
3,
256.18,
665.26,
247.82,
22.31
],
"formula_id": "formula_1",
"formula_text": "τ du(t) dt = -u(t) + I(t),(1)"
},
{
"formula_coordinates": [
4,
236.85,
93.23,
267.15,
13.68
],
"formula_id": "formula_2",
"formula_text": "u t,l i = h t-1,l i + x t,l i ,(2)"
},
{
"formula_coordinates": [
4,
237.89,
110.12,
266.11,
13.68
],
"formula_id": "formula_3",
"formula_text": "s t,l i = Hea(u t,l i -u th ),(3)"
},
{
"formula_coordinates": [
4,
236.82,
127.01,
267.18,
13.68
],
"formula_id": "formula_4",
"formula_text": "h t,l i = V reset s t,l i + βu t,l i (1 -s t,l i ),(4)"
},
{
"formula_coordinates": [
4,
236.86,
144.82,
267.14,
30.32
],
"formula_id": "formula_5",
"formula_text": "x t,l i = N j=1 w l ij s t,l-1 j ,(5)"
},
{
"formula_coordinates": [
4,
278.73,
536.59,
225.27,
26.96
],
"formula_id": "formula_6",
"formula_text": "u th + u th - p(u)du,(6)"
},
{
"formula_coordinates": [
5,
214.69,
100.69,
289.31,
16.07
],
"formula_id": "formula_7",
"formula_text": "∃ > 0, x ∈ [m -, m + ] , sup x p(x) < +∞.(7)"
},
{
"formula_coordinates": [
5,
108,
473.88,
396,
36.03
],
"formula_id": "formula_8",
"formula_text": "-1 + xT is a random variable follows the u th -Neighborhood-Finite Distribution, if xt -xt ≤ and σt (x t ) = σt (x t ) for t = 1, 2, • • • , T -1, then: P σT (x T ) = σT (x T ) ∝ 1-β ."
},
{
"formula_coordinates": [
5,
108,
575.44,
396,
37.31
],
"formula_id": "formula_9",
"formula_text": "= ĥt-1 + xt is a random variable follows the u th -Neighborhood-Finite Distribution, if xt k -xt k ≤ (k = 1, 2, • • • , N ; i = 1, 2, • • • , T ) and σt (x t ) = σt (x t ) for t = 1, 2, • • • , T -1, then: P σT (x T ) = σT (x T ) ∝ N 1-β ."
},
{
"formula_coordinates": [
6,
228.35,
354.74,
275.65,
8.96
],
"formula_id": "formula_10",
"formula_text": "g(s) -ŵs = ws -ŵs ≤ w -ŵ .(8)"
},
{
"formula_coordinates": [
6,
107.39,
517.26,
167.8,
15.29
],
"formula_id": "formula_11",
"formula_text": "Fix the weight matrix Ŵ ∈ [-1 √ N , 1 √ N ]"
},
{
"formula_coordinates": [
6,
108,
626.77,
396,
28.7
],
"formula_id": "formula_12",
"formula_text": "s ∈ {0, 1} N . Then, ∀0 <δ≤ 1, ∃ > 0 when k ≥ N N C th log N δ and C th = 1-u th 2 , there exists a mask matrix B that [g(s)] i -[ Ŵ s] i ≤ w.p at least 1 -δ."
},
{
"formula_coordinates": [
7,
210.54,
115.22,
293.46,
11.03
],
"formula_id": "formula_13",
"formula_text": "G t (S) = G t,L • G t,L-1 • • • • • G t,2 • G t,1 (S),(9)"
},
{
"formula_coordinates": [
7,
234.18,
169.48,
269.82,
14.85
],
"formula_id": "formula_14",
"formula_text": "G t,l (S) = σ t,l W l S 1:t-1,l (W l S t,l-1 ),(10)"
},
{
"formula_coordinates": [
7,
108,
282.19,
396,
23.76
],
"formula_id": "formula_15",
"formula_text": "∈ [-1 √ N , 1 √ N ] N ×N"
},
{
"formula_coordinates": [
7,
107.67,
481.29,
358.48,
17.67
],
"formula_id": "formula_16",
"formula_text": "Theorem 4.3. All Steps Approximation. Fix the weight matrix Ŵ l ∈ [-1 √ N , 1 √ N ] N ×N"
},
{
"formula_coordinates": [
7,
108,
538.18,
397.75,
40.88
],
"formula_id": "formula_17",
"formula_text": "B l ∈ {0, 1} k×N is the mask for matrix V l , i,j B l ij 0 ≤ N 2 , j B l ij 0 ≤ N . Then, let the function of equivalent network at timestep t be Gt (S) = Gt,L • Gt,L-1 • • • • • Gt,1 (S) and Gt,l = σt (( W l B l )σ t (V l S t ))"
},
{
"formula_coordinates": [
7,
108,
580.72,
397.74,
22.73
],
"formula_id": "formula_18",
"formula_text": "t is Ĝt (S) = Ĝt,L • Ĝt,L-1 • • • • • Ĝt,1 (S), where Ĝt,l (S) = σt ( Ŵ l S t ). l = 1, 2, • • • , L, t = 1, 2, • • • , T ."
},
{
"formula_coordinates": [
8,
250.32,
156.21,
253.68,
23.22
],
"formula_id": "formula_19",
"formula_text": "s uv ← s uv + α ∂L ∂I v S u w uv ,(11)"
},
{
"formula_coordinates": [
9,
231.87,
172.7,
272.13,
9.65
],
"formula_id": "formula_20",
"formula_text": "P ≈ E(|u -u|)N (0|µ -u th , var),(12)"
},
{
"formula_coordinates": [
9,
247.92,
233.2,
251.93,
26.05
],
"formula_id": "formula_21",
"formula_text": "E(|u -u|) = E act [|w|] |γ| σ 2 B + , (13"
},
{
"formula_coordinates": [
9,
499.85,
240.26,
4.15,
8.64
],
"formula_id": "formula_22",
"formula_text": ")"
},
{
"formula_coordinates": [
14,
112.04,
660.58,
391.96,
26.98
],
"formula_id": "formula_23",
"formula_text": "> 0so that x ∈ (v -, v + ) p sup = sup x p(x) ≥ p(x),(S.1)"
},
{
"formula_coordinates": [
14,
262.43,
699.21,
241.57,
26.29
],
"formula_id": "formula_24",
"formula_text": "v+ v- p(x)dx ≤ 2 p sup ,(S.2)"
},
{
"formula_coordinates": [
15,
108,
218,
105.63,
10.31
],
"formula_id": "formula_25",
"formula_text": "P σT (x T ) = σT (x T ) ∝"
},
{
"formula_coordinates": [
15,
249.03,
279.09,
254.98,
19.22
],
"formula_id": "formula_26",
"formula_text": "p sup = sup ûT p(û T ) ≥ p(û T ),(S.3)"
},
{
"formula_coordinates": [
15,
215.28,
327.96,
288.72,
59.22
],
"formula_id": "formula_27",
"formula_text": "P σT (x T ) = σT (x T ) = P sign(û T -u th )sign(ũ T -u th ) = -1 ≤ u th + u th - p(û T )du ≤ 2p sup ,(S.4)"
},
{
"formula_coordinates": [
15,
108,
537.78,
396,
36.03
],
"formula_id": "formula_28",
"formula_text": "-1 + xT is a random variable follows the u th -Neighborhood-Finite Distribution, if xt -xt ≤ and σt (x t ) = σt (x t ) for t = 1, 2, • • • , T -1, then: P σT (x T ) = σT (x T ) ∝ 1-β ."
},
{
"formula_coordinates": [
15,
273.6,
617.31,
230.41,
40.82
],
"formula_id": "formula_29",
"formula_text": "ût = ĥt-1 + xt (S.5) ũt = ht-1 + xt ,(S.6)"
},
{
"formula_coordinates": [
15,
279.78,
683.89,
224.22,
8.96
],
"formula_id": "formula_30",
"formula_text": "û1 = ũ1 ≤ , (S.7)"
},
{
"formula_coordinates": [
16,
108,
92.93,
396,
65.3
],
"formula_id": "formula_31",
"formula_text": "ûT -ũT ≤ T -1 n=0 β n ≤ +∞ n=0 β n ≤ 1 -β (S.8) then: P σT (x T ) = σT (x T ) ≤ 2p sup 1 -β ,(S.9)"
},
{
"formula_coordinates": [
16,
108,
296.27,
396.67,
23.88
],
"formula_id": "formula_32",
"formula_text": "u th -Neighborhood-Finite Distribution, if xt k -xt k ≤ (k = 1, 2, • • • , N ; t = 1, 2, • • • , T ) and σt (x t ) = σt (x t ) for t = 1, 2, • • • , T -1, then: P σT (x T ) = σT (x T ) ∝ N 1-β ."
},
{
"formula_coordinates": [
16,
268.04,
353.33,
235.96,
21.66
],
"formula_id": "formula_33",
"formula_text": "p sup = sup ûT k p(û T k ),(S.10)"
},
{
"formula_coordinates": [
16,
227.88,
396.24,
276.12,
22.31
],
"formula_id": "formula_34",
"formula_text": "P σT k (x T k ) = σT k (x T k ) ≤ 2 1 -β p sup (S.11)"
},
{
"formula_coordinates": [
16,
215.6,
442.7,
276.08,
22.31
],
"formula_id": "formula_35",
"formula_text": "P ∃k, σT k (x T k ) = σT k (x T k ) ≤ N 2 1 -β p sup (S."
},
{
"formula_coordinates": [
16,
108,
546.79,
161.39,
11.3
],
"formula_id": "formula_36",
"formula_text": "k weights v = [v 1 , v 2 , • • • , v k ] T connect"
},
{
"formula_coordinates": [
17,
108,
93.3,
396,
82.21
],
"formula_id": "formula_37",
"formula_text": "• v i ≥ v th Since w, v ∼ U[-1, 1], P [ wi -ŵi ≤ ] = 2 2 = (S.13) P [v i ≥ v th ] = 1 -v th 2 = C th (S.14)"
},
{
"formula_coordinates": [
17,
239.92,
210.75,
264.08,
37.45
],
"formula_id": "formula_38",
"formula_text": "P [ wi -ŵi ≤ ∧ v i ≥ v th ] =P [ wi -ŵi ≤ ] P [v i ≥ v th ] =C th (S.15)"
},
{
"formula_coordinates": [
17,
216.71,
269.73,
287.29,
25.87
],
"formula_id": "formula_39",
"formula_text": "P [∀i ∈ [k], ¬( wi -ŵi ≤ ∧ v i ≥ v th )] =(1 -C th ) k ≤ exp (-kC th ) ≤ δ (S.16)"
},
{
"formula_coordinates": [
17,
162.15,
517.21,
329.54,
9.96
],
"formula_id": "formula_40",
"formula_text": "E i,k , = {∀a ∈ [k ], ¬( w(i-1)k +a -ŵi ≤ ∧ v (i-1)k +a,i ≥ u th )}, (S."
},
{
"formula_coordinates": [
17,
107.69,
572.36,
396.31,
27.38
],
"formula_id": "formula_41",
"formula_text": "Thus P (E i,k , N ) ≤ exp(-k C th N ), (S.18)"
},
{
"formula_coordinates": [
17,
108,
620.84,
396,
55.92
],
"formula_id": "formula_42",
"formula_text": "P ( i E i,k , N ) ≤ i P (E i,k , N ) ≤ N exp(-k C th N ) ≤ δ, (S.19) thus: P (( i E i,k , N ) C ) ≥ 1 -δ, (S."
},
{
"formula_coordinates": [
18,
260.7,
168.16,
84.48,
24.8
],
"formula_id": "formula_43",
"formula_text": "k ≥ N N C th log N 2 δ"
},
{
"formula_coordinates": [
18,
265.28,
223.4,
80.17,
12.17
],
"formula_id": "formula_44",
"formula_text": "[g(s)] i -[ Ŵ s] i ≤"
},
{
"formula_coordinates": [
18,
142.35,
292.83,
349.34,
9.96
],
"formula_id": "formula_45",
"formula_text": "E j,i,k , = {∀a ∈ [k ], ¬( wj,(i-1)k +a -ŵji ≤ ∧ v (i-1)k +a,i ≥ u th )}, (S."
},
{
"formula_coordinates": [
18,
239.46,
364.21,
264.54,
15.57
],
"formula_id": "formula_46",
"formula_text": "P (E j,i,k , N ) ≤ exp(-k C th N ),(S.23)"
},
{
"formula_coordinates": [
18,
176.42,
404.65,
315.26,
21.98
],
"formula_id": "formula_47",
"formula_text": "P ( j,i E j,i,k , N ) ≤ j,i P (E j,i,k , N ) ≤ N 2 exp(-k C th N ) ≤ δ, (S."
},
{
"formula_coordinates": [
18,
265.63,
446.36,
238.37,
22.06
],
"formula_id": "formula_48",
"formula_text": "j,i E j,i,k , N ) C ) ≥ 1 -δ, (S.25)"
},
{
"formula_coordinates": [
18,
260.7,
484.44,
243.3,
24.8
],
"formula_id": "formula_49",
"formula_text": "k ≥ N N C th log N 2 δ (S.26)"
},
{
"formula_coordinates": [
18,
108,
532.16,
396,
26.2
],
"formula_id": "formula_50",
"formula_text": "Fix the weight matrix Ŵ ∈ [-1 √ N , 1 √ N ] N ×N"
},
{
"formula_coordinates": [
18,
246.27,
639.78,
109.08,
24.8
],
"formula_id": "formula_51",
"formula_text": "k ≥ N 2 N C th log N 2 δ -N C"
},
{
"formula_coordinates": [
19,
190.45,
95.15,
313.55,
21.98
],
"formula_id": "formula_52",
"formula_text": "P ( j,i E j,i,k , N ) ≤ j,i P (E j,i,k , N ) ≤ N 2 exp(k C th N ) (S.27)"
},
{
"formula_coordinates": [
19,
108,
219.03,
396,
68.73
],
"formula_id": "formula_53",
"formula_text": "P (( j,i E j,i,k , N ) E f ire ) ≤ j,i P (E j,i,k , N ) + P (E f ire ) ≤ N 2 exp(-k C th N ) + N p sup 2 1 -β ≤ δ, (S.28) Let C = 2p sup1"
},
{
"formula_coordinates": [
19,
246.27,
297.62,
257.73,
24.8
],
"formula_id": "formula_54",
"formula_text": "k ≥ N 2 N C th log N 2 δ -N C (S.29)"
},
{
"formula_coordinates": [
19,
247.27,
361.66,
256.73,
22.06
],
"formula_id": "formula_55",
"formula_text": "j,i E j,i,k , N ) E f ire ) C ) ≥ 1 -δ,(S.30)"
},
{
"formula_coordinates": [
19,
372.14,
408.5,
74.75,
17.68
],
"formula_id": "formula_56",
"formula_text": "Ŵ l ∈ [-1 √ N , 1 √ N ]"
},
{
"formula_coordinates": [
19,
108,
479.97,
386.35,
26.48
],
"formula_id": "formula_57",
"formula_text": "(s) = GL • GL-1 • • • • • G1 (s) and Gl = σ(( W l B l )σ(V l s))"
},
{
"formula_coordinates": [
19,
110.26,
495.2,
393.74,
26.45
],
"formula_id": "formula_58",
"formula_text": "(s) = ĜL • ĜL-1 • • • • • Ĝ1 (s), where Ĝl (s) = σ( Ŵ l s). l = 1, 2, • • • , L."
},
{
"formula_coordinates": [
19,
241.5,
540.58,
129,
24.8
],
"formula_id": "formula_59",
"formula_text": "k ≥ N 2 N C th log N 2 L δ -N CL ,"
},
{
"formula_coordinates": [
19,
159.19,
659.6,
344.81,
64.04
],
"formula_id": "formula_60",
"formula_text": "P (( l,j,i E l,j,i,k , N ) E f ire,l ) ≤ l,j,i P (E l,j,i,k , N ) + l P (E f ire,l ) ≤ LN 2 exp(-k C th N ) + LN C ≤δ, (S.31) Thus, k ≥ N 2 N C th log N 2 L δ -N CL , (S."
},
{
"formula_coordinates": [
20,
240.99,
126.12,
263.01,
22.29
],
"formula_id": "formula_61",
"formula_text": "l,j,i E l,j,i,k , N ) E f ire,l ) C ) ≥ 1 -δ (S.33)"
},
{
"formula_coordinates": [
20,
371.96,
174.42,
94.46,
17.67
],
"formula_id": "formula_62",
"formula_text": "Ŵ l ∈ [-1 √ N , 1 √ N ] N ×N"
},
{
"formula_coordinates": [
20,
110.26,
245.74,
393.74,
26.45
],
"formula_id": "formula_63",
"formula_text": "(S) = Gt,L • Gt,L-1 • • • • • Gt,1 (S) and Gt,l = σt (( W l B l )σ t (V l S t ))"
},
{
"formula_coordinates": [
20,
108,
273.86,
397.74,
22.73
],
"formula_id": "formula_64",
"formula_text": "(S) = Ĝt,L • Ĝt,L-1 • • • • • Ĝt,1 (S), where Ĝt,l (S) = σt ( Ŵ l S t ). l = 1, 2, • • • , L, t = 1, 2, • • • , T ."
},
{
"formula_coordinates": [
20,
237.9,
316.38,
136.21,
24.8
],
"formula_id": "formula_65",
"formula_text": "k ≥ N 2 N C th log N 2 L δ -N CLT ,"
},
{
"formula_coordinates": [
20,
107.69,
444.95,
396.31,
128.6
],
"formula_id": "formula_66",
"formula_text": "P (( t,l,j,i E t,l,j,i,k , N ) E f ire,t,l ) ≤ t,l,j,i P (E t,l,j,i,k , N ) + t,l P (E f ire,t,l ) ≤LN 2 exp(-k C th N ) + T LN C ≤δ, (S.34) Thus, k ≥ N 2 N C th log N 2 L δ -N CT L , (S.35)"
},
{
"formula_coordinates": [
20,
216.21,
589.13,
287.79,
22.29
],
"formula_id": "formula_67",
"formula_text": "P ((( l,j,i E t,l,j,i,k , N ) E f ire,t,l ) C ) ≥ 1 -δ (S.36)"
},
{
"formula_coordinates": [
21,
254.08,
101.81,
249.92,
26.05
],
"formula_id": "formula_68",
"formula_text": "E(|u -u|) = E act [w]γ σ 2 B + , (S.38)"
},
{
"formula_coordinates": [
21,
247.37,
254.46,
60.49,
16.39
],
"formula_id": "formula_69",
"formula_text": "x i ← x i -µ B"
},
{
"formula_coordinates": [
22,
231.74,
197.71,
60.51,
28.36
],
"formula_id": "formula_70",
"formula_text": "E act [|w ij |] |γ j | σ 2 Bj +"
}
] |
|
10.1145/3461702.3462624 | 2023-10-22 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b3",
"b27",
"b16",
"b2",
"b23",
"b17",
"b11",
"b17",
"b14",
"b13",
"b11",
"b31",
"b17",
"b27",
"b6",
"b27",
"b1",
"b17",
"b31",
"b17",
"b11",
"b30",
"b14",
"b22",
"b15"
],
"table_ref": [],
"text": "Topic models are, loosely put, an unsupervised dimensionality reduction technique that help organize document collections (Blei et al., 2003). A topic model summarizes a document collection with a small number of topics. A topic is a probability distribution over words or phrases. A topic T is interpretable through a representative set of words or phrases defining the topic, denoted W T . 1 Each document can, in turn, be represented as a distribution over topics. For each topic, we can retrieve a representative document collection by sorting documents across topic distributions. We denote this set of documents for topic T as D T . Because of their ability to organize large collections of texts, topic models are widely used 1 We think of \"words\" as an atomic unit in a document, which can also be an n-gram or phrase. E.g., Wlegal = {litigation, attorney-client privilege, intellectual property, . . . }. in the social sciences, digital humanities, and other disciplines to analyze large corpora (Talley et al., 2011;Grimmer and Stewart, 2013;Antoniak et al., 2019;Karami et al., 2020, inter alia).\nInterpretability makes topic models useful, but human interpretation is complex and notoriously difficult to approximate (Lipton, 2018). Automated topic coherence metrics do not correlate well with human judgments, often overstating differences between models (Hoyle et al., 2021;Doogan and Buntine, 2021). Without the guidance of an automated metric, the number of topics, an important hyperparameter, is usually derived manually: Practitioners fit various topic models, inspect the resulting topics, and select the configuration which works best for the intended use case (Hoyle et al., 2021). This is a non-replicable and time-consuming process, requiring expensive expert labor.\nRecent NLP research explores whether large language models (LLMs) can perform automatic annotations; e.g., to assess text quality (Fu et al., 2023;Faggioli et al., 2023;Huang et al., 2023, inter alia). Here, we investigate whether LLMs can automatically assess the coherence of topic modeling output and conclude that:\n(1) LLMs can accurately judge topic coherence, (2) LLMs can assist in automatically determining reasonable numbers of topics.\nWe use LLMs for two established topic coherence evaluation tasks and find that their judgment strongly correlates with humans on one of these tasks. Similar to recent findings, we find that coherent topic word sets W T do not necessarily imply an optimal categorization of the document collection (Doogan and Buntine, 2021). Instead, we automatically assign a label to each document in a D T and choose the configuration with the purest assigned labels. This solution correlates well with an underlying ground truth. Thus, LLMs can help find good numbers of topics for a text collection, as we show in three case studies.\nMost topic model evaluations focus on the coherence of W T , the most probable words from the topic-word distribution (Röder et al., 2015). Coherence itself can be thought of as whether the top words elicit a distinct concept in the reader (Hoyle et al., 2021). To complicate matters, human evaluation of topic models can be done in diverse ways. E.g., we can ask humans to directly rate topic coherence, for example, on a 1-3 scale (Newman et al., 2010a;Mimno et al., 2011;Aletras and Stevenson, 2013, inter alia). We can also add an unrelated intruder word to the list of top words, which human annotators are asked to identify. The intuition is that intruder words are easily identified within coherent and self-contained topics, but hard to identify for incoherent or not self-contained topics (Chang et al., 2009). High human accuracy on this task is thus a good proxy for high topic coherence. See both in Example 1. Although many automated metrics exist (Wallach et al., 2009;Newman et al., 2010b;Mimno et al., 2011;Aletras and Stevenson, 2013), normalized pointwise mutual information (NPMI, Bouma, 2009) is the most prevalent when evaluating novel methods (Hoyle et al., 2021). Informally, NPMI is larger if two words co-occur together regularly in a reference corpus. Another popular metric, C v , is a combination of NPMI and other measures and is also popular (Röder et al., 2015). See the formula definitions in Appendix D.\nDespite their popular use, these metrics correlate poorly with human evaluations (Hoyle et al., 2021;Doogan and Buntine, 2021). In this work, we let LLMs perform the rating and intrusion detection tasks for topic model evaluation2 and propose LLM scores as a novel automated metric. Similar work by Rahimi et al. (2023) is carried contemporaneously. LLMs have already been used to rank machine translations and generated text (Zhang et al., 2020;Fu et al., 2023;Kocmi and Federmann, 2023) and have also been shown to perform on par with crowdworkers (Gilardi et al., 2023). "
},
{
"figure_ref": [],
"heading": "LLM and Coherence",
"publication_ref": [
"b26",
"b17",
"b17"
],
"table_ref": [
"tab_1"
],
"text": "First, we show that large language models can assess the quality of topics generated by different topic modeling algorithms. We use existing topic modeling output annotated by humans (Hoyle et al., 2021). 3 This data consists of 300 topics, produced by three different topic modeling algorithms on two datasets: NYtimes (Sandhaus, 2008) and Wikitext (Merity et al., 2017). For each of the 300 topics, there are 15 individual human annotations for the topic word relatedness (on 1-3 scale), and 26 individual annotations for whether a crowd-worker correctly detected an intruder word. We replicate both tasks, prompting LLMs instead of human annotators. See Prompt 1 for prompt excerpts, and Appendix A for full details. We compute the Spearman correlation between the LLM answer and the human assessment of the topics and show results in Table 1.\nBaseline metrics. For NPMI and C v , we report the best correlation by Hoyle et al. (2021). These metrics depend on the reference corpus and other hyperparameters and we always report the best value. Hoyle et al. (2021) find no single best setting for these automated metrics and therefore this comparison makes the baseline inadequately strong.\nIntrusion detection task. The accuracies for detecting intruder words in the evaluated topics are almost identical -humans correctly detect 71.2% of the intruder words, LLMs identify intruders in 72.2% of the cases. However, humans and LLMs differ for which topics these intruder words are identified. This results in overall strong correlations within human judgement, but not higher correlations than NPMI and C v (in their best setting).\nCoherence rating task. The LLM rating of the W T top word coherence correlates more strongly with human evaluations than all other automated metrics in any setting. This difference is statistically significant, and the correlation between LLM ratings and human assessment approaches the inter-annotator agreement ceiling. Appendix Appendix B shows additional results with different prompts and LLMs.\nRecommendation. Both findings support using LLMs for evaluating coherence of W T in practice as they correlate highly with human judgements."
},
{
"figure_ref": [],
"heading": "Determining the Number of Topics",
"publication_ref": [
"b12",
"b11",
"b18",
"b26"
],
"table_ref": [],
"text": "Topic models require specifying the number of topics. Practitioners usually run models multiple times with different numbers of topics (denoted by k). After manual inspection, the model which seems most suited for a research question is chosen. Doogan et al. (2023) review 189 articles about topic modeling and find that common use cases are exploratory and descriptive studies for which no single best number of topics exists. However, the most prevalent use case is to isolate semantically similar documents belonging to topics of interest. For this, Doogan and Buntine (2021) challenge the focus on only evaluating W T , and suggest an analysis of D T as well. If we are interested in organizing a collection, then we would expect the top documents in D T to receive the same topic labels. We provide an LLM-based strategy to determine good number of topics for this use case: We let an LLM assign labels to documents, and find that topic as-signments with greater label purity correlate with the ground-truth in three case studies.\nTopics of interest might be a few broad topics such as politics or healthcare, or many specific topics, like municipal elections and maternity care. Following recent efforts that use research questions to guide LLM-based text analysis (Zhong et al., 2023), we incorporate this desideratum in the LLM prompt. We run collapsed Gibbs-sampled LDA (in MALLET: McCallum, 2002) on two text collections, with different numbers of topics (k = 20 to 400), yielding 20 models per collection. To compare topic model estimates and ground-truth partitions, we experiment with a legislative Bill summary dataset (from Hoyle et al., 2022) and Wikitext (Merity et al., 2017), both annotated with ground-truth topic labels in different granularities."
},
{
"figure_ref": [],
"heading": "Proposed Metrics",
"publication_ref": [
"b11"
],
"table_ref": [],
"text": "Ratings algorithm. For each of the 20 models, we randomly sample W T for some topics and let the LLM rate these W T . The prompt is similar to the ratings prompt shown in Prompt 1, see Appendix E for full details. We then average ratings for each configuration. Intuitively, the model yielding the most coherent W T should be the one with the optimal topic count. However, this procedure does not correlate with ground-truth labels.\nText labeling algorithm. Doogan and Buntine (2021) propose that domain experts assign labels to each document in a D T instead. A good topic should have a coherent D T : The same label assigned to most documents. Hence, good configurations have high purity of assigned labels within each topic. We proceed analogously. For each of the 20 models, we randomly sample D T for various topics. We retrieve the 10 most probable documents and then use the LLM to assign a label to these documents. We use the system prompt [...] Annotate the document with a broad|narrow label [...], see Appendix E for full details. We compute the purity of the assigned labels and average purities and we select the configuration with the most pure topics. In both procedures, we smooth the LLM outputs using a rolling window average to reduce noise (the final average goodness is computed as moving average of window of size 3)."
},
{
"figure_ref": [
"fig_0"
],
"heading": "Evaluation",
"publication_ref": [
"b18",
"b20"
],
"table_ref": [],
"text": "We need a human-derived metric to compare with the purity metric proposed above. We measure the alignment between a topic model's predicted topic assignments and the ground-truth labels for a document collection (Hoyle et al., 2022).\nWe choose the Adjusted Rand Index (ARI) which compares two clusterings (Hubert and Arabie, 1985) and is high when there is strong overlap. The predicted topic assignment for each document is its most probable topic. Recall that there exist many different optimal topic models for a single collection. If we want topics to contain semantically similar documents, each ground-truth assignment reflects one possible set of topics of interests.\nIf our LLM-guided procedure and the ARI correlate, this indicates that we discovered a reasonable value for the number of topics. In our case, the various ground-truth labels are assigned with different research questions in mind. We incorporate such constraints in the LLM prompt: We specify whether we are interested in broad or specific topics, and we enumerate some example ground-truth categories in our prompt. Practitioners usually have priors about topics of interest before running topic models, thus we believe this setup to be realistic.\nIn Figure 1 we show LLM scores and ARI for broad topics in the Bills dataset. We used this dataset to find a suitable prompt, hence this could be considered the \"training set\". We plot coherence ratings of word sets in blue , purity of document labels in red , and the ARI between topic model and ground-truth assignments in green . The purity of LLM-assigned D T labels correlate with the ARI, whereas the W T coherence scores do not. The argmax of the purity-based approach leads to similar numbers of topics as suggested by the ARI argmax (although not always the same).\nFor Wikitext, we evaluate the same 20 topic models, but measure ARI between topic model assignment and two different ground-truth label sets. The LLM scores differ only because of different prompting strategies. The distributions indicate that this strategy incorporates different research questions.\nFor Bills, our rating algorithm suggests to use a topic model with k=100 topics. In Appendix G, we show corresponding word sets. The resulting W T seem interpretable, although the ground-truth assignments using document-topic estimates are not correlated with the ground-truth labels. The puritybased approach instead suggests to use k=20 topics, the same k as indicated by the ARI. We show ground-truth labels and LLM-obtained text labels in Appendix G. We further manually evaluate 180 assigned LLM-labels and find that 94% of these labels are reasonable. Appendix F shows further evaluation of these label assignments."
},
{
"figure_ref": [],
"heading": "Discussion",
"publication_ref": [
"b17",
"b11"
],
"table_ref": [],
"text": "In this work, we revisit automated topic model evaluation with the help of large language models. Many automated evaluation metrics for topic models exist, however these metrics seem to not correlate strongly with human judgment on wordset analysis (Hoyle et al., 2021). Instead, we find that an LLM-based metric of coherent topic words correlates with human preferences, outperforming other metrics on the rating task.\nSecond, the number of topics k has to be defined before running a topic model, so practitioners run multiple models with different k. We investigate whether LLMs can guide us towards reasonable k for a collection and research question. We first note that the term optimal number of topics is vague and that such quantity does not exist without additional context. If our goal is to find a configuration which would result in coherent document sets for topics, our study supports evaluating D T instead of W T , as this correlates more strongly with the overlap between topic model and ground-truth assignment. This finding supports arguments made in Doogan and Buntine (2021) who challenge the focus on W T in topic model evaluation."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [
"b9"
],
"table_ref": [],
"text": "Choice of LLM. Apart from ChatGPT, we also used open-source LLMs, such as FLAN-T5 (Chung et al., 2022), and still obtained reasonable, albeit worse than ChatGPT, coherence correlations. Given the rapid advances, future iterations of opensource LLMs will likely become better at this task."
},
{
"figure_ref": [],
"heading": "Number of topics.",
"publication_ref": [
"b12",
"b18",
"b17"
],
"table_ref": [],
"text": "The optimal number of topics is a vague concept, dependent on a practitioner's goals and the data under study. At the same time, it is a required hyperparameter of topic models. Based on Doogan et al. (2023), we use an existing document categorization as one possible ground truth. While content analysis is the most popular application of topic models (Hoyle et al., 2022), it remains an open question how they compare to alternative clustering algorithms for this use case (e.g., k-means over document embeddings).\nInterpretability. LLM label assignment and intruder detection remain opaque. This hinders the understanding of the evaluation decisions.\nTopic modeling algorithm. In Section 3, we evaluate three topic modeling algorithms: Gibbs-LDA, Dirichlet-VAE and ETM (see Hoyle et al., 2021). In Section 4, we use only Gibbs-LDA and expansion to further models is left for future work."
},
{
"figure_ref": [],
"heading": "Future work.",
"publication_ref": [],
"table_ref": [],
"text": "• Evaluation of clustering algorithms with LLMs (e.g., k-means). • More rigorous evaluation of open-source LLMs.\n• Formalization, implementation and release of an LLM-guided algorithm for automatically finding optimal numbers of topics for a text collection and a research question."
},
{
"figure_ref": [],
"heading": "Ethics Statement",
"publication_ref": [
"b0",
"b24"
],
"table_ref": [],
"text": "Using blackbox models in NLP. Statistically significant positive results are a sufficient proof of models' capabilities, assuming that the training data is not part of the training set. This data leakage problem with closed-source LLMs is part of a bigger and unresolved discussion. In our case, we believe data leakage is unlikely. Admittedly, the data used for our coherence experiments has been publicly available. However, the data is available in a large JSON file where the topic words and annotated labels are stored disjointly. For our case studies in Section 4, the topic modeling was constructed as part of this work and there is no ground-truth which could leak to the language model.\nNegative results with LLMs. In case of negative results, we cannot conclude that a model can not be used for a particular task. The negative results can be caused by inadequate prompting strategies and may even be resolved by advances in LLMs.\nLLMs and biases. LLMs are known to be biased (Abid et al., 2021;Lucy and Bamman, 2021) and their usage in this application may potentially perpetuate these biases.\nData privacy. All data used in this study has been collected as part of other work. We find no potential violations of data privacy. Thus, we feel comfortable re-using the data in this work.\nMisuse potential. We urge practicioners to not blindly apply our method on their topic modeling output, but still manually validate that the topic outputs would be suitable to answer a given research question.\nEvan "
},
{
"figure_ref": [],
"heading": "A Language Model Prompts",
"publication_ref": [
"b17"
],
"table_ref": [
"tab_4"
],
"text": "In this section, we show the used LLM prompts. The task descriptions are borrowed from (Hoyle et al., 2021) and mimic crowd-worker instructions.\nWe use a temperature of 1 for LLMs, and the topic words are shuffled before being prompted. Both introduce additional variation within the results, similar to how some variation is introduced if different crowd-workers are asked to perform the same task.\nIntruder detection task. Analogous to the human experiment, we randomly sample (a) five word from the top 10 topic words and (b) an additional intruder word from a different topic which does not occur in the top 50 words of the current topic.\nWe then shuffle these six words. We show the final prompt with an example topic in Prompt 2. We also construct a prompt without the dataset description (see Prompt 3 and results in Table 2).\nSystem prompt: You are a helpful assistant evaluating the top words of a topic model output for a given topic. Select which word is the least related to all other words. If multiple words do not fit, choose the word that is most out of place. The topic modeling is based on The New York Times corpus. The corpus consists of articles from 1987 to 2007. Sections from a typical paper include International, National, New York Regional, Business, Technology, and Sports news; features on topics such as Dining, Movies, Travel, and Fashion; there are also obituaries and opinion pieces. Reply with a single word.\nUser prompt: water, area, river, park, miles, game Prompt 2: Intruder Detection Task (the intruder word in this topic is game). We show the task description for the New York Times dataset in the prompt for the rating task (the dataset descriptions are kept the same).\nSystem prompt: You are a helpful assistant evaluating the top words of a topic model output for a given topic. Select which word is the least related to all other words. If multiple words do not fit, choose the word that is most out of place.\nReply with a single word.\nUser prompt: water, area, river, park, miles, game Prompt 3: Intruder Detection Task. The intruder word in this topic is game.\nRating Task. Similar to the human experiment, we retrieve the top 10 topic words and shuffle them.\nWe include a task and dataset description which leads to Prompt 4. The minimal prompt without the dataset description is shown in Prompt 5.\nSystem prompt: You are a helpful assistant evaluating the top words of a topic model output for a given topic. Please rate how related the following words are to each other on a scale from 1 to 3 (\"1\" = not very related, \"2\" = moderately related, \"3\" = very related).\nThe topic modeling is based on the Wikipedia corpus. Wikipedia is an online encyclopedia covering a huge range of topics. Articles can include biographies (\"George Washington\"), scientific phenomena (\"Solar Eclipse\"), art pieces (\"La Danse\"), music (\"Amazing Grace\"), transportation (\"U.S. Route 131\"), sports (\"1952 winter olympics\"), historical events or periods (\"Tang Dynasty\"), media and pop culture (\"The Simpsons Movie\"), places (\"Yosemite National Park\"), plants and animals (\"koala\"), and warfare (\"USS Nevada (BB-36)\"), among others. Reply with a single number, indicating the overall appropriateness of the topic. User prompt: lake, park, river, land, years, feet, ice, miles, water, area Prompt 4: Rating Task. Topic terms are shuffled.\nSystem prompt: You are a helpful assistant evaluating the top words of a topic model output for a given topic. Please rate how related the following words are to each other on a scale from 1 to 3 (\"1\" = not very related, \"2\" = moderately related, \"3\" = very related).\nReply with a single number, indicating the overall appropriateness of the topic.\nUser prompt: lake, park, river, land, years, feet, ice, miles, water, area Prompt 5: Rating Task without dataset description. Topic terms are shuffled."
},
{
"figure_ref": [],
"heading": "B Additional: Topic Model Outputs",
"publication_ref": [
"b9",
"b17"
],
"table_ref": [
"tab_1"
],
"text": "Minimal prompt. Even without the dataset description in the prompt, the results remain similar.\nAll human ratings. In our main results, we discard human annotations with low annotator confidence in the rating. We now consider all ratings, even the non-confident ones. The results are slightly better than with the filtering.\nDifferent LLM. We also evaluate both tasks with FLAN-T5 XL (Chung et al., 2022) 1 for reference. LLM (min.) -results using a minimal prompt without dataset descriptions. LLM (all ann) -no discarding low-confidence annotations. FLAN-T5 -FLAN-T5 XL instead of ChatGPT. All numbers are the average result of 1000 bootstrapping episodes -re-sampling human annotations and LLM scores. Ceiling shows batched inter-annotator agreement.\nsignificant. For the NYT and concatenated experiments, the resulting correlation are statistically indistinguishable from the best reported automated metrics NPMI and C v in (Hoyle et al., 2021). We also ran our experiments with Alpaca-7B and Falcon-7B, with largely negative results."
},
{
"figure_ref": [],
"heading": "C Alternative Clustering Metrics",
"publication_ref": [],
"table_ref": [],
"text": "In our main results, we show correlations between LLM scores and the adjusted Rand Index, ARI, which measures the overlap between ground-truth clustering and topic model assignments. There are other cluster metrics, such as Adjusted Mutual Information, AMI (Vinh et al., 2010), completeness, or homogeneity. In Table 3, we show Spearman correlation statistics for these metrics. Our correlations are robust to the choice of metric used to measure the fit between the topic model assignment and the ground-truths in our case studies."
},
{
"figure_ref": [],
"heading": "D Definitions",
"publication_ref": [
"b31"
],
"table_ref": [],
"text": "See Bouma (2009) for justification of the NPMI formula. p(w i ) and p(w i , w j ) are unigram and joint probabilities, respectively.\nNPMI(w i , w j )= PMI(w i , w j ) -log p(w i , w j ) = log p(w i ,w j ) p(w i )p(w j ) -log p(w i , w j )\nThe C v metric (Röder et al., 2015) is a more complex and includes, among others, the combination of NPMI and cosine similarity for top words. Table 3: Spearman correlation coefficients between our language-model based scores and various popular metrics for assessing the overlap between the topic model assignment and the underlying ground-truth. Compl. = Completeness, Homog. = Homogenity."
},
{
"figure_ref": [],
"heading": "E Optimal Number of Topics Prompts",
"publication_ref": [],
"table_ref": [],
"text": "We now show the prompts for the optimal number of topics. We incorporate research questions in two ways: (1) we specify whether we are looking for broad or narrow topics, and (2) we prompt 5 example categories. We believe this is a realistic operationalization. If our goal is a reasonable partitioning of a collection, we usually have some priors about what categories we want the collection to be partitioned into.\nPrompt 6 shows the prompt for rating T ws by models run with different numbers of topics. The task description and user prompt is identical to the prompt used in our prior experiments, displayed in e.g., Prompt 4. However, the dataset description is different and allows for some variation. In Prompt 7, we show the prompt for automatically assigning labels to a document from a T dc . To automatically find the optimal number of topics for a topic model, we prompt an LLM to provide a concise label to a document from the topic document collection, the most likely documents assigned by a topic model to a topic (see Prompt 7).\nYou are a helpful assistant evaluating the top words of a topic model output for a given topic. Please rate how related the following words are to each other on a scale from 1 to 3 (\"1\" = not very related, \"2\" = moderately related, \"3\" = very related). The topic modeling is based on a legislative Bill summary dataset. We are interested in coherent broad|narrow topics. Typical topics in the dataset include \"topic 1\", \"topic 2\", \"topic 3\", \"topic 4\" and \"topic 5\". Reply with a single number, indicating the overall appropriateness of the topic. User prompt: lake, park, river, land, years, feet, ice, miles, water, area Prompt 6: Rating Task without dataset description. Topic terms are shuffled. We apply this prompt to 2 different datasets and 2 different research goals (broad and narrow topics), and would set this part of the prompt accordingly. Also, we set as topic 1 to topic 5 the 5 most prevalent ground-truth labels from a dataset.\nSystem prompt: You are a helpful research assistant with lots of knowledge about topic models. You are given a document assigned to a topic by a topic model. Annotate the document with a broad|narrow label, for example \"topic 1\", \"topic 2\", \"topic 3\", \"topic 4\" and \"topic 5\". Reply with a single word or phrase, indicating the label of the document. User prompt: National Black Clergy for the Elimination of HIV/AIDS Act of 2011 -Authorizes the Director of the Office of Minority Health of the Department of Health and Human Services (HHS) to make grants to public health agencies and faith-based organizations to conduct HIV/AIDS prevention, testing, and related outreach activities ... Prompt 7: Assigning a label to a document belonging to the top document collection of a topic. The label provided in this example is health. We apply this prompt to 2 different datasets and 2 different research goals (broad and narrow topics), and would set this part of the prompt accordingly. Also, we set as topic 1 to topic 5 the 5 most prevalent ground-truth labels from a dataset."
},
{
"figure_ref": [],
"heading": "F Additional: Document Labeling",
"publication_ref": [],
"table_ref": [],
"text": "In our study, we automatically label the top 10 documents for five randomly sampled topics. The ARI between-topic model partitioning and ground-truth labels correlates if we were to only examine these top 10 documents or all documents in the collection. The correlation between these two in the Bills dataset is 0.96, indicating that analyzing only the top 10 documents in a topic is a decent proxy for the whole collection.\nNext, we evaluate the LLM-based label assignement to a document. Our documents are usually long, up to 2000 words. We only consider the first 50 words in a document as input to the LLM. For Wikipedia, this is reasonable, because the first 2-3 sentences define the article and give a good summary of the topic of an article. For Bills, we manually confirm that the topic of an article is introduced at the beginning of a document.\nHuman evaluation. From each case study, we randomly sample 60 documents and assigned labels (3 examples for each of the twenty topic models), resulting in 180 examples in total. We then evaluate whether the assigned label reasonably captures the document content given the specification in the input prompt (e.g., a broad label such as health or defense, or a narrow label such as warships of germany or tropical cyclones: atlantic. Recall that the prompted labels correspond to the five most prevalent ground-truth categories of the groundtruth annotation. We find that the assigned label makes sense in 93.9% of examined labels. In the 11 errors spotted, the assigned label does not meet the granularity in 6 cases, is no adequate description of the document in 3 cases, and is a summary of the document instead of a label in 2 cases.\nAutomated Metrics. Given that we have groundtruth labels for each document, we can compute cluster metrics between the assigned labels by the LLM and the ground-truth labels (see Table 4). These values refer to comparing all labels assigned during our case study to their ground-truth label (1000 assigned datapoints per dataset). Table 4: Accuracy of the label assignment task. We find that the assigned labels clustering overlaps with the ground-truth labels.\nOn average, we assign 10 times as many unique labels to documents than there are ground-truth labels (we assign 172 different labels in the Bills dataset, 348 labels in the broad Wikitext dataset and 515 labels in the narrow Wikitext dataset). Nevertheless, the automated metrics indicate a decent overlap between ground-truth and assigned labels. Thus, the LLM often assigns the same label to documents with the same ground-truth label."
},
{
"figure_ref": [],
"heading": "G Qualitative Results",
"publication_ref": [],
"table_ref": [
"tab_7"
],
"text": "In this section, we show qualitative results of our automated investigation of numbers of topics. In Table 5, we show three randomly sampled topics from the preferred topic model in our experiments. We contrast these with three randomly sampled topics from the topic model configuration which our procedures indicate as least suitable.\nIn Table 6, we show true labels and LLMassigned labels for three randomly sampled topics from the preferred topic model, contrasting it with true and LLM-assigned labels from topics in the least suitable configuration. We find that indeed, the assigned labels and the ground-truth label often match -and that the purity of the LLM-assigned labels reflects the purity of the ground-truth label."
},
{
"figure_ref": [],
"heading": "Bills (broad categories)",
"publication_ref": [],
"table_ref": [],
"text": "Most suitable -veterans, secretary, veteran, assistance, service, disability, benefits, educational, compensation, veterans_affairs (3) -land, forest, management, lands, act, usda, projects, secretary, restoration, federal (3) -mental, health, services, treatment, abuse, programs, substance, grants, prevention, program (3) Least suitable -gas, secretary, lease, oil, leasing, act, way, federal, production, environmental (2) -covered, criminal, history, act, restitution, child, background, amends, checks, victim (2) -information, beneficial, value, study, ownership, united_states, act, area, secretary, new_york (1) Wikitext (broad categories) Most suitable -episode, star, trek, enterprise, series, season, crew, generation, ship, episodes (3) -series, episodes, season, episode, television, cast, production, second, viewers, pilot (3) -car, vehicle, vehicles, engine, model, models, production, cars, design, rear (3) Least suitable -episode, series, doctor, season, character, time, star, story, trek, set (2) -stage, tour, ride, park, concert, dance, train, coaster, new, roller (1) -said, like, character, time, life, love, relationship, later, people, way (1)"
},
{
"figure_ref": [],
"heading": "Wikitext (specific categories)",
"publication_ref": [],
"table_ref": [],
"text": "Most suitable -episode, star, trek, enterprise, series, season, crew, generation, ship, episodes (3) -car, vehicle, vehicles, engine, model, models, production, cars, design, rear (3) -world, record, meter, time, won, freestyle, gold, championships, relay, seconds (2) Least suitable -fossil, fossils, found, specimens, years, evolution, modern, million, eddie, like (2) -match, event, impact, joe, team, angle, episode, styles, championship, tag (1) -brown, rihanna, usher, love, girl, loud, yeah, wrote, bow, bad (1) 6: Assigned LLM labels and ground-truth labels for a given topic from the most and the least suitable cluster configuration according to our algorithm. The purity is higher in the most suitable configuration for LLM labels and ground-truth labels."
}
] | Topic models help make sense of large text collections. Automatically evaluating their output and determining the optimal number of topics are both longstanding challenges, with no effective automated solutions to date. This paper evaluates the effectiveness of large language models (LLMs) for these tasks. We find that LLMs appropriately assess the resulting topics, correlating more strongly with human judgments than existing automated metrics. However, the type of evaluation task matters -LLMs correlate better with coherence ratings of word sets than on a word intrusion task. We find that LLMs can also guide users toward a reasonable number of topics. In actual applications, topic models are typically used to answer a research question related to a collection of texts. We can incorporate this research question in the prompt to the LLM, which helps estimate the optimal number of topics. | Revisiting Automated Topic Model Evaluation with Large Language Models | [
{
"figure_caption": "Figure 1 :1Figure 1: (1) ARI for topic assignment and ground-truth topic labels, (2) LLM word set coherence, (3) LLM document set purity, obtained by our algorithm. ARI correlates with LLM document set purity, but not with LLM word set coherence. The ground-truth number of topics are: 21 topics in the BillSum dataset, 45 broad topics in Wikitext and 279 specific topics in Wikitext. ρ D and ρ W are document-LLM and word-LLM correlations with ARI.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Spearman correlation between human scores and automated metrics. All results use 1000 bootstrap-",
"figure_data": "TaskDataset NPMI C v LLM CeilingNYT0.430.45 † 0.370.67IntrusionWiki0.39 † 0.34 0.350.60Both0.40 † 0.40 † 0.360.64NYT0.480.40 0.64 ⋆0.72RatingWiki0.440.40 0.57 ⋆0.56Both0.440.42 0.59 ⋆0.65",
"figure_id": "tab_1",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Sandhaus. 2008. The new york times annotated corpus.",
"figure_data": "and David Mimno. 2009. Evaluation methods fortopic models. In Proceedings of the 26th annual in-Edmund M Talley, David Newman, David Mimno, Bruce W Herr 2nd, Hanna M Wallach, Gully A P Cternational conference on machine learning, pages 1105-1112.Burns, A G Miriam Leenders, and Andrew McCal-lum. 2011. Database of NIH grants using machine-learned categories and graphical clustering. Nat. Methods, 8(6):443-444.Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020. BERTScore: Evaluating text generation with BERT. In 8th In-ternational Conference on Learning Representations,Nguyen Xuan Vinh, Julien Epps, and James Bailey. 2010. Information theoretic measures for cluster-ICLR 2020 Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net.ings comparison: Variants, properties, normalization and correction for chance. J. Mach. Learn. Res., 11:2837-2854.Ruiqi Zhong, Peter Zhang, Steve Li, Jinwoo Ahn, Dan Klein, and Jacob Steinhardt. 2023. Goal driven dis-covery of distributional differences via language de-Hanna M Wallach, Iain Murray, Ruslan Salakhutdinov,scriptions.",
"figure_id": "tab_3",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Additional experiments reporting Spearman correlation between mean human scores and automated metrics. LLM (main) repeats our main results in Table",
"figure_data": "TaskDataset NPMI C v LLM (main) LLM (min.) LLM (all ann.) FLAN-T5 CeilingNYT0.430.450.370.410.390.370.67IntrusionWiki0.390.340.350.270.360.180.60Both0.400.400.360.340.380.280.64NYT0.480.400.640.640.650.310.72RatingWiki0.440.400.570.510.560.170.56Both0.440.420.590.570.610.250.65",
"figure_id": "tab_4",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Most and least suitable topics according to our LLM-based assessment on different datasets and use cases. In brackets the LLM rating for the coherence of this topic.",
"figure_data": "BillsWikitext (broad)Wikitext (specific)LLM-labelTrue labelLLM-labelTrue labelLLM-labelTrue labelMost suitablehealth elder abuse prevention healthHealth Social Wel-fare Healthamusement park ride amusement park ride amusementRecreation Recreation Recreationpolitician politician americanHistorical figures: politicians Historical figures: politicians Historical figures: politicianspark ridecivil warhealthHealthamusementRecreationlawyer andHistorical figures: otherpark ridepoliticianhealthHealthamusementRecreationhistoricalJournalism and newspaperspark ridenewspaperLeast suitablepublic land public land public landPublic Lands warship and naval unit Public Lands warship and naval unit Environment warship andArmies and mil-itary units Armies and mil-itary units Military people hinduism classical greek poetry hinduismPoetry Religious doctrines, teachings, texts, events, and symbols Religious doctrines, teachings,naval unittexts, events, and symbolsindigenousGovernmentwarship andMilitary people philosophyPhilosophical doctrines, teach-affairOperationsnaval unitings, texts, events, and symbolsindigenousGovernmentwar poetryLanguage andphilosophyPhilosophical doctrines, teach-affairOperationsliteratureings, texts, events, and symbolsTable",
"figure_id": "tab_7",
"figure_label": "5",
"figure_type": "table"
}
] | Dominik Stammbach; Vilém Zouhar; Alexander Hoyle; Mrinmaya Sachan; Elliott Ash | [
{
"authors": "Abubakar Abid; Maheen Farooqi; James Zou",
"journal": "",
"ref_id": "b0",
"title": "Persistent anti-muslim bias in large language models",
"year": "2021"
},
{
"authors": "Nikolaos Aletras; Mark Stevenson",
"journal": "",
"ref_id": "b1",
"title": "Evaluating topic coherence using distributional semantics",
"year": "2013"
},
{
"authors": "Maria Antoniak; David Mimno; Karen Levy",
"journal": "Proc. ACM Hum.-Comput. Interact",
"ref_id": "b2",
"title": "Narrative paths and negotiation of power in birth stories",
"year": "2019"
},
{
"authors": "David M Blei; Andrew Y Ng; Michael I Jordan",
"journal": "J. Mach. Learn. Res",
"ref_id": "b3",
"title": "Latent dirichlet allocation",
"year": "2003"
},
{
"authors": "J Gerlof; Bouma",
"journal": "",
"ref_id": "b4",
"title": "Normalized (pointwise) mutual information in collocation extraction",
"year": "2009"
},
{
"authors": "Sophie Burkhardt; Stefan Kramer",
"journal": "Journal of Machine Learning Research",
"ref_id": "b5",
"title": "Decoupling sparsity and smoothness in the dirichlet variational autoencoder topic model",
"year": "2019"
},
{
"authors": "Jonathan Chang; Sean Gerrish; Chong Wang; Jordan Boyd-Graber; David Blei",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b6",
"title": "Reading tea leaves: How humans interpret topic models",
"year": "2009"
},
{
"authors": "Curran Associates; Inc ",
"journal": "",
"ref_id": "b7",
"title": "",
"year": ""
},
{
"authors": "Ken Yew; Pengfei Chia; Lidong Hong; Soujanya Bing; Poria",
"journal": "",
"ref_id": "b8",
"title": "INSTRUCTEVAL: Towards holistic evaluation of instruction-tuned large language models",
"year": "2023"
},
{
"authors": "Chung Hyung Won; Le Hou; Shayne Longpre; Barret Zoph; Yi Tay; William Fedus; Yunxuan Li; Xuezhi Wang; Mostafa Dehghani; Siddhartha Brahma; Albert Webson; Shane Shixiang; Zhuyun Gu; Mirac Dai; Xinyun Suzgun; Aakanksha Chen; Alex Chowdhery; Marie Castro-Ros; Kevin Pellat; Dasha Robinson; Sharan Valter; Gaurav Narang; Adams Mishra; Vincent Yu; Yanping Zhao; Andrew Huang; Hongkun Dai; Slav Yu; Ed H Petrov; Jeff Chi; Jacob Dean; Adam Devlin; Denny Roberts; Quoc V Zhou; Jason Le; Wei",
"journal": "",
"ref_id": "b9",
"title": "Scaling instruction-finetuned language models",
"year": "2022"
},
{
"authors": "B Adji; Dieng; J R Francisco; David M Ruiz; Blei",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b10",
"title": "Topic modeling in embedding spaces",
"year": "2020"
},
{
"authors": "Caitlin Doogan; Wray Buntine",
"journal": "Association for Computational Linguistics",
"ref_id": "b11",
"title": "Topic model or topic twaddle? Re-evaluating semantic interpretability measures",
"year": "2021"
},
{
"authors": "Caitlin Doogan; Wray Buntine; Henry Linger",
"journal": "Artificial Intelligence Review",
"ref_id": "b12",
"title": "A systematic review of the use of topic models for short text social media analysis",
"year": "2023"
},
{
"authors": "Guglielmo Faggioli; Laura Dietz; Charles Clarke; Gianluca Demartini; Matthias Hagen; Claudia Hauff; Noriko Kando; Evangelos Kanoulas; Martin Potthast; Benno Stein; Henning Wachsmuth",
"journal": "",
"ref_id": "b13",
"title": "Perspectives on large language models for relevance judgment",
"year": "2023"
},
{
"authors": "Jinlan Fu; See-Kiong Ng; Zhengbao Jiang; Pengfei Liu",
"journal": "",
"ref_id": "b14",
"title": "GPTScore: Evaluate as you desire",
"year": "2023"
},
{
"authors": "Fabrizio Gilardi; Meysam Alizadeh; Maël Kubli",
"journal": "",
"ref_id": "b15",
"title": "ChatGPT outperforms crowd-workers for textannotation tasks",
"year": "2023"
},
{
"authors": "Justin Grimmer; Brandon M Stewart",
"journal": "Political Analysis",
"ref_id": "b16",
"title": "Text as data: The promise and pitfalls of automatic content analysis methods for political texts",
"year": "2013"
},
{
"authors": "Alexander Hoyle; Pranav Goel; Denis Peskov; Andrew Hian-Cheong; Jordan Boyd-Graber; Philip Resnik",
"journal": "",
"ref_id": "b17",
"title": "Is automated topic model evaluation broken?: The incoherence of coherence",
"year": "2021"
},
{
"authors": "Alexander Miserlis Hoyle; Rupak Sarkar; Pranav Goel; Philip Resnik",
"journal": "Association for Computational Linguistics",
"ref_id": "b18",
"title": "Are neural topic models broken?",
"year": "2022"
},
{
"authors": "Fan Huang; Haewoon Kwak; Jisun An",
"journal": "",
"ref_id": "b19",
"title": "Is chatgpt better than human annotators? potential and limitations of chatgpt in explaining implicit hate speech",
"year": "2023"
},
{
"authors": "Lawrence Hubert; Phipps Arabie",
"journal": "Journal of classification",
"ref_id": "b20",
"title": "Comparing partitions",
"year": "1985"
},
{
"authors": "Amir Karami; Morgan Lundy; Frank Webb; Yogesh K Dwivedi",
"journal": "IEEE Access",
"ref_id": "b21",
"title": "Twitter and research: A systematic literature review through text mining",
"year": "2020"
},
{
"authors": "Tom Kocmi; Christian Federmann",
"journal": "",
"ref_id": "b22",
"title": "Large language models are state-of-the-art evaluators of translation quality",
"year": "2023"
},
{
"authors": "Zachary C Lipton",
"journal": "Queue",
"ref_id": "b23",
"title": "The mythos of model interpretability: In machine learning, the concept of interpretability is both important and slippery",
"year": "2018"
},
{
"authors": "Li Lucy; David Bamman",
"journal": "",
"ref_id": "b24",
"title": "Gender and representation bias in GPT-3 generated stories",
"year": "2021"
},
{
"authors": "Andrew Kachites; Mccallum ",
"journal": "",
"ref_id": "b25",
"title": "MALLET: A MAachine Learning for LanguagE Toolkit",
"year": "2002"
},
{
"authors": "Stephen Merity; Caiming Xiong; James Bradbury; Richard Socher",
"journal": "",
"ref_id": "b26",
"title": "Pointer sentinel mixture models",
"year": "2017-04-24"
},
{
"authors": "David Mimno; Hanna Wallach; Edmund Talley; Miriam Leenders; Andrew Mccallum",
"journal": "",
"ref_id": "b27",
"title": "Optimizing semantic coherence in topic models",
"year": "2011"
},
{
"authors": "David Newman; Jey ; Han Lau; Karl Grieser; Timothy Baldwin",
"journal": "",
"ref_id": "b28",
"title": "Automatic evaluation of topic coherence",
"year": "2010"
},
{
"authors": "David Newman; Youn Noh; Edmund Talley; Sarvnaz Karimi; Timothy Baldwin",
"journal": "",
"ref_id": "b29",
"title": "Evaluating topic models for digital libraries",
"year": "2010"
},
{
"authors": "Hamed Rahimi; Jacob Louis Hoover; David Mimno; Hubert Naacke; Camelia Constantin; Bernd Amann",
"journal": "",
"ref_id": "b30",
"title": "Contextualized topic coherence metrics",
"year": "2023"
},
{
"authors": "Michael Röder; Andreas Both; Alexander Hinneburg",
"journal": "",
"ref_id": "b31",
"title": "Exploring the space of topic coherence measures",
"year": "2015"
}
] | [
{
"formula_coordinates": [
8,
73.56,
674.35,
211.68,
32.4
],
"formula_id": "formula_0",
"formula_text": "NPMI(w i , w j )= PMI(w i , w j ) -log p(w i , w j ) = log p(w i ,w j ) p(w i )p(w j ) -log p(w i , w j )"
}
] |
2023-05-20 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b34",
"b24",
"b31",
"b32"
],
"table_ref": [],
"text": "As AI systems increasingly take up the role of assisting, and at times replacing, human decision makers [Kaminski and Urban, 2021], there is a growing public call for establishing a right to receive an explanation to outcomes generated by automated decision-making processes. Purportedly originating from the General Data Protection Regulation (GDPR) [Regulation, 2016] and known as the \"legal right to explanation\" [Goodman and Flaxman, 2017], this right is often portrayed as one tool in the regulatory toolkit for creating, deploying, and monitoring ethical and accountable AI systems, mitigating the potential breach of fundamental principles of the rule of law, such as transparency and accountability [Hildebrandt, 2020] and protecting human rights. Simultaneously, and in correlation with the introduction of more complex AI, Explainability has triggered a growing interest within the Machine Learning (ML) community. Tasked with providing explanations to complex predictions and motivated by an incentive to cultivate trust in AI systems [Jacovi et al., 2021], ML developers have embraced Explainability (XAI) to develop end-users explanations. This work inquires whether, and to what extent, can end-user Explainability satisfy the right to explanation of AI systems' requirements by law. Embracing a \"court-like\" setting, Section 2 makes the case in favor of Explainability, addressing its organic evolvement, its usefulness for ML professionals and its often-cited potential contribution to protecting human rights. Next, Section 3 and Section 4 lay down the case against end-user Explainability. Accordingly, Section 3 sets the legal backdrop by providing a broad-brush analysis of the role of explanations in the legal domain. This analysis focuses on three questions: (1) How is \"explanation\" defined in law, linking it to the notion of \"reason-giving\";\n(2) Where is reason-giving used in law, briefly surveying its appearances in public, private and international law, and most importantly; (3) Why is reason-giving used in law, meaning what are the underlying functions at the heart of the ubiquitous legal practice of reason-giving?\nThe analysis of reason-giving's legal functions in Section 3 uncovers its four main purposes: (a) promoting the making of a better and more just decision, (b) facilitating due-process, (c) authenticating human agency of both the decision subject and the decision maker, and (d) enhancing the decision makers' authority, by promoting legitimacy, accountability and providing guidance. As an interim conclusion, Section 3 highlights the fact that reason-giving is a mechanism aimed to influence the human decision maker in various forms, subsequently restraining human rationale and human judgement. Building upon this methodology Section 4 continues to make the case against Explainability. It utilizes reason-giving's deconstruction in Section 3 to analyze the extent to which end-users Explainability is capable of serving the roles assigned to reason-giving in law. It first examines Explainbility's potential to impact the decision-making process itself, finding it slim given that a human decision-maker has been replaced by a prediction-making machine. Thus, reason-giving's function to promote the making of a better and more just decision is largely unfulfilled. Next, the analysis questions the ability of machine generated explanations to support human agency and respect human autonomy. Then, turning to the function of facilitating due-process rights, the analysis highlights Explainability's challenge to produce what is typically considered in law \"an explanation\". Having ruled out end-user Explainability's ability to serve three of reason-giving's functions in law, we do find that Explainability is compatible with fulfilling reason-giving's fourth function, i.e., enhancing the decision makers' authority. However, we observe that recent Explainable AI (XAI) research trajectories and Large Language Models' (LLMs) emerging capabilities raise serious challenges to the reliability of Explainability's outcomes and create a potential for manipulate end-users by humans and machines alike. As a final conclusion, the study outlines some policy implications. The gap between a legal right to explanation and the technological field called Explainability, challenges the usefulness of Explainability as a reason-giving tool in an end-user AI context. Policymakers and ML practitioners should thus reconsider reliance on end-users' Explainability for achieving the societal goals of reason-giving and explore alternatives."
},
{
"figure_ref": [],
"heading": "Making the Case for Explainable Artificial Intelligence",
"publication_ref": [
"b53",
"b48",
"b37",
"b61",
"b51",
"b2",
"b2",
"b6",
"b32",
"b65",
"b42",
"b37",
"b46",
"b60",
"b34",
"b63",
"b17",
"b66",
"b22",
"b8",
"b47",
"b16",
"b13",
"b34",
"b30",
"b57",
"b58",
"b4",
"b62"
],
"table_ref": [],
"text": "Prior to the legal and regulative interest in a \"right to explanation\" of AI systems, Explainability was developed by the ML community as a means to contend with one of the most publicly known features of AI systems: it's increasing opacity, or as more commonly known, the \"black box\" quality. Although not all AI systems are opaque, and there are several \"degrees\" of opaqueness, this quality nevertheless has become a meaningful challenge due to the introduction of Deep Learning Networks, some of them using billions of parameters. In sync with the rise of the level of complexity of data science, attempts began to offer means to address the opacity challenge, culminating in several concepts, methods, and tools, Explainability (XAI) being one of them [Nicholas, 2019].\nThe term \"Explainability\" originates to the 80's and 90's [Miller et al., 2017]. It was developed in order to produce good quality (robust) systems, which consists of an understanding of their inner workings, quality control, bug solving and continuous learning and progress towards the next generation of technology. At its core, XAI \"seeks to bring clarity to how specific ML models work\" [Laato et al., 2022], and the use of Explainability is often linked to context and relevancy considerations [Rudin, 2019;Molnar, 2020;Arrieta et al., 2020]. This fact highlights how the progress made in easing the opacity challenge evolved from real professional challenges: finding out how the system works in order to improve it, fix it, extract takeaways from mistakes and strive to simplify the process [Arrieta et al., 2020]. This core necessity sets the tone for the various technical solutions which were offered and are still being continuously developed to put forward explanations for automated systems, housing a vast amount of research work at the cutting edge of AI today [Biran and Cotton, 2017].\nAdditionally, industry has also acknowledged the problem opacity creates in the general public, asked to be subject to major life-changing and at times high stakes decisions, construed by machines. Clearing out some of the mist around AI systems is often regarded as a step towards creating public trust in this innovative technology [Jacovi et al., 2021]. This approach was largely facilitated by the increased focus of the HCI (Human Computer Interaction) field on extending the definition of human actors interacting with the machine. XAI was embraced by the HCI field at the intersection with the ML community, in a mission to make computational processes clearer to humans [Shneiderman et al., 2016]. Accordingly, explanation in the field of computer science has been understood as \"making it possible for a human being (designer, user, affected person, etc.) to understand a result or the whole system\" [Malgieri, 2021]. This comprehensive definition represents perhaps the turning trajectory of XAI towards including the end user of AI systems, re-calibrated in correlation with the increasing deployment of these systems in domains already regulated by existing laws. But it also highlights the fact that explanations for AI systems are often mentioned in the context of the mission to promote trust, or trustworthiness, in AI [Laato et al., 2022]. The absence of an ability to explain decisions and actions by AI black-boxes to human users has been recently referred to as a key limitation of today's intelligent systems, whereby the \"...lack of Explainability hampers our capacity to fully trust AI systems\" [Mehta et al., 2022]. And trust, it has been argued, promotes user's utilization of models, both by relying on its predictions as well as by accepting its deployment [Ribeiro et al., 2016].\nAgainst this technological backdrop, Explainability has been enlisted to secure a legal mechanism, the so called \"right to explanation\", which regulators sought for the protection of society from potential AI harms. Regulation of AI systems grasped the vast potential of automated systems on the one hand, but expressed a genuine concern towards safeguarding human rights [Council, 2019]. Being preoccupied with the purported \"black-box\" quality of AI systems, regulation sought transparency enhancing mechanisms to address those concerns. In the legal domain, transparency is often linked to fairness, as means to assure accountability of decision makers [Kaminski, 2021]. Faithful to this transparency ethos, \"...the majority of discourse around understanding machine learning models has seen the proper task as opening the black box and explaining what is inside\" [Selbst and Barocas, 2018]. Accordingly, explanations for AI systems are being promoted in service of multiple regulatory objectives aiming at enhancing transparency. Thus, explanation-giving for AI systems was mentioned as means for achieving AI accountability [Doshi-Velez et al., 2017;Smith-Renner et al., 2020;Gillis and Simons, 2019], detect discrimination [Brkan, 2019], reveal bias issues [Melsión et al., 2021], and ensure fairness in AI systems [Dodge et al., 2019]. In regards to governmental use of AI, explanation-giving is presented as a way to accommodate due process requirements and achieve good governance [Crawford and Schultz, 2014]. Similarly, it is also considered essential in order to allow for a meaningful contestation right towards automated decisions [Kaminski and Urban, 2021]. This extensive list highlights the diverse groups, interests, and contexts for which a right to explanation of AI systems is considered a desired feature, as well as demonstrates the large extent of reliance on transparency in general, and explanations in particular, by regulators, legal practitioners, and legal scholars. This insight prompts the following question: why did regulators and legal practitioners turn to the tool of explanation-giving, in service of protecting humanity against AI harms? The answer lies in the role of explanations in law and law's ubiquitous use of explanations.\n3 The Role of Explanation in Law -\"What\" is Explanation in Law?\n\"The business of law is the business of making decisions\" [Hawkins, 1986]. This eloquent statement captures the fact that decision-making resides at the heart of the legal system. In a democratic society decision-making is often accompanied by explanations of those decisions [Rawls, 1997], making it a common practice both for law-making and law-applying [Raz, 2009]. This form of \"reason-explanations\", typically used when humans try to understand and explain action and resolve disagreements [Baum et al., 2022], is usually referred to as \"reason-giving\". Its use is so ubiquitous that \"the practice of providing reasons for decisions has long been considered an essential aspect of legal culture\" [Schauer, 1994]. To deconstruct the notion of reason-giving in law and to answer the question \" What is reason-giving in law?\", this section will ask the following questions: How is reasongiving defined in law? Where can we find the use of reason-giving in law? and most importantly Why does law uses reason-giving to begin with, meaning what are its underlying functions?"
},
{
"figure_ref": [],
"heading": "Reason-Giving in Law -the \"How\" -Defining Key Terms",
"publication_ref": [
"b62",
"b42",
"b42",
"b62"
],
"table_ref": [],
"text": "In order to alleviate some of the \"fuzziness\" around basic concepts it is important to first define their meaning. The giving of reasons can be described as \"the practice of engaging in the linguistic act of providing a reason to justify what we do or what we decide\" [Schauer, 1994]. The difference between explaining (\"providing a reason\") and justifying (\"to justify\") is not strictly semantic. While explanation in a general sense means \"an act of spotting the main reasons or factors that led to a particular consequence, situation, or decision\" [Malgieri, 2021], a justification takes on another layer, detailing why the decision at hand is the \"right\" and \"just\" one [Malgieri, 2021]. Therefore, an explanation is part of the justification. Explanations and justifications will be collectively referred to here as reason-giving, the process whereby decisionmakers elaborate the explanations and justifications supporting their decisions [Deeks, 2019]. Indeed, reason-giving is particularly suitable for legal decisions since \"[w]hen we provide a reason for a particular decision, we typically provide a rule, principle, standard, norm, or maxim broader than the decision itself...\" [Schauer, 1994]. It should be noted here that reason-giving has a multi-layered presence in law. For example, law demands reason-giving (e.g., courts requiring agencies to produce reasons for a decision), and law also manufactures reasons simultaneously (e.g., courts justifying their rulings re the agencies' actions). Moreover, reason-giving is relevant both as part of the decision-making process itself (adjudicating -the process of deliberating and deciding), and as a product accompanying the final decision if released publicly."
},
{
"figure_ref": [],
"heading": "Reason-Giving in Law -the \"Where\"",
"publication_ref": [
"b71",
"b45",
"b35",
"b36"
],
"table_ref": [],
"text": "Reason-giving and explanations are being ubiquitously used across the legal system. Some of the most dominant arenas where the legal system leverages reason-giving are public law, private law, and increasingly international law. In a nutshell, public law is perhaps the most widely recognized domain of reason-giving in the legal system, construed out of courts, agencies and legislators constantly manufacturing and reviewing explanations and justifications. Private law exemplifies the extent through which \"regulatory transparency\" has become the tool-of-choice to handle regulatory challenges [Weil et al., 2006], where perhaps the most prominent example is the requirement to obtain a patient's informed consent prior to undergoing medical procedures, itself contingent upon receiving an explanation from a physician [McLean, 2009]. Civil law also entails numerous examples of explanation usages such as in contractual relationships or in Tort law. In addition, the newly emerging habit of nations to explain foreign policy as part of international law fortifies the importance of reason-giving to decision making as a legal and social phenomenon, transcending states and geolocations [Keitner, 2018;Kingsbury, 2009]."
},
{
"figure_ref": [],
"heading": "Reason-Giving in Law -the \"Why\"",
"publication_ref": [
"b22",
"b3",
"b11",
"b64",
"b22",
"b12",
"b44",
"b21",
"b34",
"b34",
"b13",
"b11",
"b22",
"b44",
"b44",
"b43",
"b43",
"b44",
"b17",
"b7",
"b1",
"b10",
"b28",
"b2",
"b38",
"b56",
"b67",
"b50",
"b25",
"b0",
"b25"
],
"table_ref": [],
"text": "The answer to the question \"what does society gain from this constant explanation giving?\" ought to spearhead the methodological framework of end user Explainability. Accordingly, this section will detail why are explanations and reason-giving such a repetitive practice in law? What purposes do they serve and what are their underlying functions? 1. Making a Better and More Just Decision -At the heart of reason-giving in law lies the non-instrumental purpose of securing a better and more just decision [Deeks, 2019]. The \"just\" feature, which supports the act as right, desirable, or reasonable, authenticates the decision as a non-biased, non-discriminatory one [Gillis and Simons, 2019]. It taps into the core objective of making sure that \"justice was done\" [Atkinson et al., 2020]. The \"better\" feature is brought to fruition by triggering the mechanism of review, either internal during the making of the decision, or external as means for appeal and contestation. Taken together, the decision possesses both a rational and a moral basis, making it a more righteous and fair result. In this sense, there's an inherent, non-instrumental value in reason-giving since it impacts the decision itself. Reason-giving compels the decision maker to handle the decision process with extra care, in a thoughtful and slower manner. There might also be a psychological pressure on decision makers to make decisions worthy of reasonable reasoning [Cohen, 2010;Shapiro, 1992]. In other words, the need to articulate reasonable reasoning for a decision nudges decision-makers to make decisions that support such reasoned reasoning in a circular movement. Therefore, the mere fact that reasoning may be required may impact the decisions even prior to such a request being materilized.\n2. Facilitating Due Process -understanding the decision-making system has been said to be instrumental for individuals to exercise their right to challenge decisions [Gillis and Simons, 2019]. When focusing on its role as a protector of individual rights, a right to explanation is usually regarded as a parasitic right, in service of fulfilling other values [Cohen, 2011;Mashaw, 2007]. Those include a right for due process, housing both a right to a hearing [Friendly, 1974] and a right to contestation [Kaminski and Urban, 2021]. The due-process theory is a core principle of the rule of law itself [Kaminski and Urban, 2021], and the procedure of due process is referred to today mainly as the requirement that any infringement on core rights should be taken after a notice was given and an opportunity for a hearing was granted [Crawford and Schultz, 2014]. Reason-giving plays multiple roles in the execution of due process rights. Naturally, knowing one's reasons for a decision assists in crafting better informed arguments to rebuttal it [Cohen, 2010], thus supporting a robust defense against a rights-infringing decision or act. Of course, due process allows for a judicial review of the decision and is especially instrumental given its contribution to the conservation of records which can be leveraged later for contestation and review of the aforementioned decision or action. Moreover, it allows the decision maker and contesting party to evaluate the chances of an appeal in advance. Finally, the giving of reasons can serve as a non-political legitimate demand by the adjudicating body, in comparison to a more subjective requirement of decision \"reasonableness\".\n3. Acknowledging Human Agency of the Decision Subject and Decision Maker -One of the core values underlying the existence of reasons for decisions is respecting human autonomy [Gillis and Simons, 2019]. In the case of the decision-subject, the reasons issued to a decision signal his or her sovereignty, since giving reasons respects the fact that humans are autonomous people that should be treated with dignity, while unreasoned coercion \"denies our moral agency and our political standing\" [Mashaw, 2007]. Moreover, respect comes in the form of providing grounds for detailed criticism not only when there is a right for contestation, but perhaps even more when there is no recourse for appeal (e.g., reason-giving accompanying Supreme Court decisions) and a decision subject is left with a right to public discourse. Additionally, reasons respect also the decision-maker's human agency. At that capacity of a human decision maker, the presence of reasons for one's actions and decisions stands at the heart of human morality and sense of judgment and autonomy [Mashaw, 2007]. Plainly put, a human decision maker needs there to be reasons for its actions, as an autonomous person. When actions are underlined with intent, the decision maker is acting as a rationale agent, thus strengthening his or her autonomy in the process.\n4. Enhancing the Decision Makers' Authority -The giving of reasons makes actions, decisions, rules, and regulations more tolerable and acceptable. This is because acknowledging them as binding is dependent upon there being sufficient rational explanations underlying those rules [Mashaw, 2001]. Simply put, \"the authority of all law relies on a set of complex reasons for believing that it should be authoritative\" [Mashaw, 2001]. Reason-giving contributes to this objective by supporting attributes that promote compliance and adherence to the deciding body. These attributes comprise of enhancing the accountability and legitimacy of the deciding body, as well as the providence of guidance to numerous stakeholders (while simultaneously serving as a binding precedent on the decision maker itself). Those virtues jointly add to maintaining and boosting agreement, cooperation and acceptance of rules established by the decision-making body, thus bolstering the system's mandate. They also serve as a pressure system of socio-legal and relational considerations cast upon the human decision maker, which is often concerned with matters of reputation, colleagues' approval, avoidance of unpleasant repercussions when reviewed, and various other incentives to make the \"right\" decision and provide meaningful explanations to reason it [Mashaw, 2007].\nFrom executing a right for due process, contributing to the making of a better and more just decision, respecting human agency and promoting the decision makers' authority, reason-giving's central role in law serves purposes oriented towards the decision subject, but also to a larger degree towards the human decision maker. Leveraging the existence of societal and relational pressures upon the human decision maker, reason-giving is a legal tool aimed primarily to contain, retrain, and curb human discretion and human judgement. This conclusion is also supported by instances where a requirement for explanations in law is absent, as is the case of jurors [Doshi-Velez et al., 2017], where the value of restraining human judgement is attained by other means, such as internal deliberations. 4 Can End-User Explainability Fulfil a \"Right to Explanation\"?\n\"Explainability\" has really come to dominate debates about ethics and regulation of machine learning [Bordt et al., 2022], largely framed as the tool to execute a right to explanation of AI systems. As several survey papers demonstrate [Adadi and Berrada, 2018;Carvalho et al., 2019;Guidotti et al., 2018], there is considerable effort employed at identifying a suitable framework or methodology for XAI in the context of end-users [Arrieta et al., 2020;Langer et al., 2021;Prakken, 2020;Tomsett et al., 2018]. However, currently, and despite this formidable effort, scholars have pointed out that the tool of Explainability is mostly used for professional debugging purposes [Mittelstadt et al., 2019], and has not yet managed to translate into a user-friendly explanation generating tool, albeit regulatory calls for an individual, decision-subject right to explanation [Goodman and Trehu, 2022]. Since \"...much work in AI and ML communities tends to suffer from a lack of usability, practical interpretability and efficacy on real users\" [Abdul et al., 2018], Explainability for end-users is proving to be a tough challenge. As scholars recently lamented, \"...so far at least, aspirational Explainability cannot be relied upon either for effective communication about how algorithmic systems works or for holding them to account\" [Goodman and Trehu, 2022].\nLeveraging the legal reason-giving methodology presented in Section 3 , this section proposes to frame the persistent gap between a right to explanation and Explainability, by examining to what extent can end-user Explainability fulfil the role law bestows upon explanations and reason-giving, and will accordingly ask: (a) can it contribute to the making of a better and more just decision? (b) can it facilitate due-process rights? (c) is it relevant for the authentication and respect of human agency, and (d) does it enhance the decision makers' authority?\n4.1 Can Explainability Contribute to a Better and More Just Decision?\nIf one of reason-giving's main roles is to impact the decision-making process itself by restraining human judgement and thus contribute to a better and fairer decision, then it is hard to grasp in what form this purpose might be fulfilled given a machine now replaces a human decision maker. The impact of reason-giving on humans, slowing down decision processes and leveraging relational pressures, is largely irrelevant when a machine's decision is involved. Unlike a human decision maker, reason-giving does not serve to contain an algorithm's judgement or discretion. An algorithm does not possess a \"rational\" (or logic) to begin with, nor does it produce a \"decision\" but rather a prediction. It is not impacted, nor impressed, by what other algorithmic colleagues may think of it, nor does it seek to minimize unpleasant consequences, or \"feel\" accountable to anyone or anything. Therefore, currently prediction algorithms make no use of the external explanation generated to their predictions. There might be some potential impact over humans in the \"surroundings\" of a model, e.g., designers, deployers etc., but this impact, if exists, should be further explored, and is probably diminished. Therefore, it appears that one of the most important objectives of reason-giving cannot be attained using Explainability for end-users."
},
{
"figure_ref": [],
"heading": "Is Explainability Instrumental to Facilitating Due-Process Rights?",
"publication_ref": [],
"table_ref": [],
"text": "To facilitate reason-giving's decision-subject purposes such as due process rights, appeal and contestation, this work proposes that Explainability should deliver decision-subjects with what law considers to be \"an explanation\", and reliable at that."
},
{
"figure_ref": [
"fig_0"
],
"heading": "Can Explainability Generate \"an Explanation\"?",
"publication_ref": [
"b70",
"b39",
"b50",
"b7",
"b50",
"b5",
"b70",
"b27"
],
"table_ref": [],
"text": "In a call to stay clear of black-box models, one of the more significant scholars in the field of ML, [, Rudin] has opined that \"[a]s the term is presently used in its most common form, an explanation is a separate model that is supposed to replicate most of the behavior of a black box...\" . In essence, the general concept dominating the XAI community is \"to create a simple human-understandable approximation of a decision[...]making algorithm that accurately models the decision given the current inputs...\" [Wachter et al., 2017]. These insights frame the different mehtods that were developed over the years to provide explanations for models, such as LIME, SHAP, LRP, etc. [Linardatos et al., 2020] and hints at the inadequacy of calling their output \"an explanation\" in nomenclature, as suggesting a reliable knowledge of how the complex model works [Mittelstadt et al., 2019].\nIn fact, those \"explanation-generating\" techniques should be regarded as producing a clue to the source of the issue explored, by providing vague approximations of how the algorithm generated its output or some understanding of the features that need to be changed in order to alter the said output [Bordt et al., 2022]. This insight requires further inquiry and human deduction skills, given causality may not be automatically inferred from the data an explanation has provided. It is up to ML experts to then leverage this clue and find the true cause for the decision/problem itself [Mittelstadt et al., 2019]. True for ML experts, this is double the case for a layperson lacking technological background.\nEven if Explainability techniques can produce an actual contextualized explanation rather than a clue, scholars argue it is still a long way from producing layperson understandable explanations [Bhatt et al., 2020]. In fact, most current Explainability techniques are non-accessible to a human lacking technological literacy [Wachter et al., 2017]. As Figure 1 demonstrates A run-of-the-mill person would have slim understanding of a saliency map, a data points analysis, or a feature importance result. Some may struggle even to understand a bar chart. Therefore, some kind of brokerage work would be needed, where a trusted expert would have to translate Explainability technique results to a person seeking an actual meaningful explanation. In this case, users' trust will be built upon experts' opinions rather than end user explanations, similar to many experiences in our lives like trusting the functioning of a navigation compass or trusting an engineer while crossing a bridge, where trust is granted not based on an explanation but on other features [Gryz and Shahbazi, 2020].\nBased on this examination it appears Explainability is currently not sufficient to deliver what regulators consider \"an explanation\". But even if it could deliver on such a requirement, can Explainability be trusted by decision-subjects to begin with?"
},
{
"figure_ref": [],
"heading": "Can Decision Subjects Rely on Explainability Generated \"Explanations\"?",
"publication_ref": [
"b70",
"b55",
"b49",
"b68",
"b55",
"b74",
"b27",
"b15",
"b7",
"b75",
"b52"
],
"table_ref": [],
"text": "If users and decision subjects cannot rely on the explanations generated by end-user Explainability, then a major obstacle hinders its adoption. Research in the field has highlighted few potential problems in this regard. First, not all stakeholders tasked with generating explanations to automated decisions welcome this explanation generating requirement. Some concerns can include potential infringement of privacy rights, intellectual property and trade secrets, and genuine security concerns [Wachter et al., 2017;Powell, 2021;Milli et al., 2019;Tramèr et al., 2016]. Additionally, a potential to game the system when receiving explanations on the one hand, or a perceived inability of end-users to comprehend complex systems on the other hand, might also contribute to designers' resentment towards end-user Explainability [Powell, 2021;Zhang et al., 2019]. And sometimes models are just so complex that it is claimed they simply cannot be explained in a meaningful way [Gryz and Shahbazi, 2020]. One should also not overlook the inherently adversarial relationship between end-users and automated decision-maker stakeholders, given that end-users and automated decision subjects largely seek an explanation to change the machine's prediction (e.g., loan-seeker vs. credit score generator).\nAdversarial situations invite ambiguous and non-trustworthy explanations to begin with [Dimanov et al., 2020], and there are multiple techniques to possibly manipulate the \"explanation\" generated by Explainability methods [Bordt et al., 2022;Zhou and Joachims, 2022;Mothilal et al., 2020]."
},
{
"figure_ref": [],
"heading": "Can Explainability Authenticate Human Autonomy?",
"publication_ref": [],
"table_ref": [],
"text": "Naturally, the change of the decision makers' identity, meaning an autonomous decision-making system vs. a human decision maker nulls reason-giving's function as an acknowledger of the decision makers' human agency. However, we believe that end-user Explainability's potential to acknowledge the decision subject's humanity and autonomy should also be questioned. While residing outside the scope of this work, this function raises multiple important and fundamental moral and philosophical questions relevant to AI systems in general, and to XAI in particular (e.g., can human agency be acknowledged by a non-human agent to begin with?).\nSo far, we have examined end-user Explainability's ability to fulfil three of reason-giving's functions in law and found it lacking. Turning to the final function, we at last find a function which Explainability is well suited to deliver, a fact which simultaneously raises serious concerns."
},
{
"figure_ref": [],
"heading": "Can Explainability Enhance the Decision-Makers' Authority?",
"publication_ref": [
"b9",
"b7",
"b40",
"b27",
"b66",
"b26",
"b19",
"b34",
"b41",
"b19",
"b18",
"b37",
"b54",
"b20",
"b54",
"b54",
"b72",
"b29",
"b20",
"b9",
"b69"
],
"table_ref": [],
"text": "Contributing to the decision-makers' authority and legitimacy is another function of reason-giving in law. In the case of end-user Explainability, we find this function can be successfully fulfilled, perhaps even better than human decision makers, especially in the case of LLMs. As a recent paper exploring GPT-4's explanation's abilities demonstrates, it \"is remarkably good at generating reasonable and coherent explanations, even when the output is nonsensical or wrong\" [Bubeck et al., 2023]. However, we recognize several problems emerging from XAI's ability to promote the decision makers' authority. First, at the heart of this function lies reason-giving's impact on the human decision maker. This impact means he or she feels accountable, seeks legitimacy and is bound by his or her previous decisions if they are to serve as guidance. Therefore, the replacement of a human with a machine nullifies most, if not all, of the aforementioned human effects. Moreover, the \"explanation\" Explainability generates is limited in the sense that we inherently expect an explanation to be based on some knowledge of the world (contextualized), whereas an algorithm only \"knows\" (if one can even attribute such adjective to a machine) what it was shown or defined to \"know\" [Bordt et al., 2022;Lipton, 2018]. In other words, and until Artificial General Intelligence proves otherwise, \"[e]very AI system is the fabled tabula rasa; it \"knows\" only as much as it has been told\" [Gryz and Shahbazi, 2020]. Under these conditions, an explanation cannot function as a rule, nor as guidance. Equally disconcertingly perhaps, although a human decision maker is increasingly replaced by a machine, one fact has yet to change, and that is the human identity of the decision subject. In automated decision making, a human is still the client/target of the explanation, a matter which potentially gives rise to rather alarming consequences. Research has shown Explainability's potential to cause human over-reliance on the system [Smith-Renner et al., 2020], as well as the opportunity for wrongdoing and manipulation by promoting misguided trust. This phenomenon of nudging users to act according to others' interest is known as \"Dark Patterns\" in XAI [Gray et al., 2018], and benefits from humans' \"automation bias\" towards trusting machines [Eiband et al., 2019;Kaminski and Urban, 2021;Lyell and Coiera, 2017]. For example, people are more eager to comply with a request simply by being presented a placebic justification by computerized systems [Eiband et al., 2019]. Further research has suggested that user manipulation can occur even unintentionally, causing \"Explainability pitfalls\" merely by choosing to present people with one sort of explanation over another [Ehsan and Riedl, 2021]. It should also be pointed out that end users XAI appears to drift away from its initial trust building objective. For example, [Laato et al., 2022] have shown that transparency is mostly evaluated in the literature according to the user's perception of transparency, rather than actual transparency attributes of the system. As a systematic review of papers in the field conveys, research in the field scarcely highlights the purpose for generating explanations to begin with [Nunes and Jannach, 2017].\nMoreover, it appears the research of end-users' XAI is increasingly shifting towards exploring which explanation practices will impact users' trust and increase perceived trustworthiness of the system, rather than produce a meaningful and reliable tool to scrutinize AI systems by users [Förster et al., 2020]. As Figure 2 taken from [Nunes and Jannach, 2017] demonstrates, surveying hundreds of XAI papers in the last decades displays a plateau or even an overall decrease in the study of XAI for transparency purposes, and a big increase in researching explanations' effectiveness, explanation techniques to enhance user's trust, techniques to increase explanations' persuasiveness, and to elevate Figure 2: [Nunes and Jannach, 2017] The figure shows that past decades demonstrate a plateau, or even a decrease, in researching explainability techniques for the purpose of transparency (\"explain how the system works\") and a sharp increase in purposes such as enhance explanations' effectiveness (\"help users make good decisions\"), enhance trust (\"increase users' confidence in the system\") and enhance persuasiveness (\"convince users to try or buy\"). Although this paper's survey dates to 2017, it is plausible to assume that a current overview will demonstrate an even stronger orientation towards user influencing purposes. All purposes definitions are taken from Table 8 of the surveyed paper.\nuser's levels of satisfaction from the system. Two interesting examples of this trend include Weitz et al. [2021] demonstrating how the use of virtual agents for Explainability seems especially promising for the purpose of increasing users' perceived trust in the system, or Goldman and Bustin [2022] experimenting with explanations as a technique to elevate user's comfort level in automated driving maneuvers of a simulated autonomous vehicle, to avoid manual take-overs. Both examples showcase how end-users' XAI might drift away from its original purpose of using explanations to promote \"appropriate trust\" [Gunning and Aha, 2019] and assist end-users to properly scrutinize AI systems [Förster et al., 2020], towards the study of how XAI can be used to influence end-users according to third parties' incentives, well intended as they may be. Finally, the emerging capabilities of LLMs demonstrate that machines might pose a risk when pursuing end-user Explainability. Recent models such as GPT-4 exhibit an increasing ability to generate convincing explanations for false decisions, a lack of consistent link between the decision-making process and the explanation generated to it, and a growing capability to generate specially tailored explanations to a human client [Bubeck et al., 2023]. As Turpin et al recently demonstrated, LLMs can produce step-by-step reasoning which systematically misrepresents the real reason underlying the model's prediction [Turpin et al., 2023]. Therefore, there is real potential to contribute to the decision-making system's trustworthiness, even when that trust is an unwarranted, misleading and even a dangerous one. These continuously improving capabilities should serve as a trigger warning for those promoting end user explanations."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [
"b9"
],
"table_ref": [],
"text": "The deconstruction of reason-giving in the legal system this study presents offers a methodological framework to analyze the gap between a right to explanation and end-user Explainability. It highlights reason-giving's role as impacting the human decision maker, as well as facilitating decision subjects' rights. Given the change in the identity of the decision maker from human to machine, current end-user Explainability struggles to deliver most of explanations' functions in law, which include promoting a better and more just decision, facilitating due-process and acknowledging human agency.\nIn contrast, end-user Explainability emerges as a successful mechanism to fulfill reason giving's fourth function in law, i.e., enhancing the decision makers' authority. However, this ability raises a set of risks for manipulating decision subjects by humans and machine alike. A key limitation of the case against Explainability is that it does not yet provide an alternative solution to the risks stemming from recent AI advancements [Bubeck et al., 2023]. Nevertheless, we fear the reliance on inadequate techniques, coupled with newly generated risks, is perilous. Hence, we hope our work will impact how Explainability is being developed and implemented and will serve as a warning sign from incompatible usage and unwarranted research directions."
},
{
"figure_ref": [],
"heading": "Acknowledgement",
"publication_ref": [],
"table_ref": [],
"text": "This work has been supported by the Israeli Science Foundation research grant 1437/22 and a grant from the Tel Aviv University Center for AI and Data Science (TAD)."
}
] | As artificial intelligence (AI) becomes more prevalent there is a growing demand from regulators to accompany decisions made by such systems with explanations. However, a persistent gap exists between the need to execute a meaningful right to explanation vs. the ability of Machine Learning systems to deliver on such a legal requirement. The regulatory appeal towards "a right to explanation" of AI systems can be attributed to the significant role of explanations, part of the notion called reason-giving, in law. Therefore, in this work we examine reason-giving's purposes in law to analyze whether reasons provided by end-user Explainability can adequately fulfill them. We find that reason-giving's legal purposes include: (a) making a better and more just decision, (b) facilitating due-process, (c) authenticating human agency, and (d) enhancing the decision makers' authority. Using this methodology, we demonstrate end-user Explainabilty's inadequacy to fulfil reason-giving's role in law, given reason-giving's functions rely on its impact over a human decision maker. Thus, end-user Explainability fails, or is unsuitable, to fulfil the first, second and third legal function. In contrast we find that end-user Explainability excels in the fourth function, a quality which raises serious risks considering recent end-user Explainability research trends, Large Language Models' capabilities, and the ability to manipulate end-users by both humans and machines. Hence, we suggest that in some cases the right to explanation of AI systems could bring more harm than good to end users. Accordingly, this study carries some important policy ramifications, as it calls upon regulators and Machine Learning practitioners to reconsider the widespread pursuit of end-user Explainability and a right to explanation of AI systems. | The Case Against Explainability | [
{
"figure_caption": "Figure 1 :1Figure 1: An example of SHAP Explainability technique summary plots from Wood et al. [2019]",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
}
] | Hofit Wasserman Rozen; Niva Elkin-Koren; Ran Gilad-Bachrach | [
{
"authors": "Ashraf Abdul; Jo Vermeulen; Danding Wang; Brian Y Lim; Mohan Kankanhalli",
"journal": "",
"ref_id": "b0",
"title": "Trends and trajectories for explainable, accountable and intelligible systems: An hci research agenda",
"year": "2018"
},
{
"authors": "Amina Adadi; Mohammed Berrada",
"journal": "IEEE access",
"ref_id": "b1",
"title": "Peeking inside the black-box: a survey on explainable artificial intelligence (xai)",
"year": "2018"
},
{
"authors": "Alejandro Barredo Arrieta; Natalia Díaz-Rodríguez; Javier Del Ser; Adrien Bennetot; Siham Tabik; Alberto Barbado; Salvador García; Sergio Gil-López; Daniel Molina; Richard Benjamins",
"journal": "Information fusion",
"ref_id": "b2",
"title": "Explainable artificial intelligence (xai): Concepts, taxonomies, opportunities and challenges toward responsible ai",
"year": "2020"
},
{
"authors": "Katie Atkinson; Trevor Bench-Capon; Danushka Bollegala",
"journal": "Artificial Intelligence",
"ref_id": "b3",
"title": "Explanation in ai and law: Past, present and future",
"year": "2020"
},
{
"authors": "Kevin Baum; Susanne Mantel; Eva Schmidt; Timo Speith",
"journal": "Philosophy & Technology",
"ref_id": "b4",
"title": "From responsibility to reason-giving explainable artificial intelligence",
"year": "2022"
},
{
"authors": "Umang Bhatt; Alice Xiang; Shubham Sharma; Adrian Weller; Ankur Taly; Yunhan Jia; Joydeep Ghosh; Ruchir Puri; M F José; Peter Moura; Eckersley",
"journal": "",
"ref_id": "b5",
"title": "Explainable machine learning in deployment",
"year": "2020"
},
{
"authors": "Or Biran; Courtenay Cotton",
"journal": "",
"ref_id": "b6",
"title": "Explanation and justification in machine learning: A survey",
"year": "2017"
},
{
"authors": "Sebastian Bordt; Michèle Finck; Eric Raidl; Ulrike Von; Luxburg ",
"journal": "",
"ref_id": "b7",
"title": "Post-hoc explanations fail to achieve their purpose in adversarial contexts",
"year": "2022"
},
{
"authors": "Maja Brkan",
"journal": "International journal of law and information technology",
"ref_id": "b8",
"title": "Do algorithms rule the world? algorithmic decision-making and data protection in the framework of the gdpr and beyond",
"year": "2019"
},
{
"authors": "Sébastien Bubeck; Varun Chandrasekaran; Ronen Eldan; Johannes Gehrke; Eric Horvitz; Ece Kamar; Peter Lee; Yin Tat Lee; Yuanzhi Li; Scott Lundberg",
"journal": "",
"ref_id": "b9",
"title": "Sparks of artificial general intelligence: Early experiments with gpt-4",
"year": "2023"
},
{
"authors": "Eduardo M Diogo V Carvalho; Jaime S Pereira; Cardoso",
"journal": "Electronics",
"ref_id": "b10",
"title": "Machine learning interpretability: A survey on methods and metrics",
"year": "2019"
},
{
"authors": "Mathilde Cohen",
"journal": "Archiv für Rechts-und Sozialphilosophie",
"ref_id": "b11",
"title": "The rule of law as the rule of reasons",
"year": "2010"
},
{
"authors": "Mathilde Cohen",
"journal": "OECD Council",
"ref_id": "b12",
"title": "Reasons for reasons",
"year": "2011-05"
},
{
"authors": "Kate Crawford; Jason Schultz",
"journal": "BCL Rev",
"ref_id": "b13",
"title": "Big data and due process: Toward a framework to redress predictive privacy harms",
"year": "2014"
},
{
"authors": "Ashley Deeks",
"journal": "Yale Law Journal",
"ref_id": "b14",
"title": "Secret reason-giving",
"year": "2020"
},
{
"authors": "Botty Dimanov; Umang Bhatt; Mateja Jamnik; Adrian Weller",
"journal": "",
"ref_id": "b15",
"title": "You shouldn't trust me: Learning models which conceal unfairness from multiple explanation methods",
"year": "2020"
},
{
"authors": "Jonathan Dodge; Q Vera Liao; Yunfeng Zhang; Rachel Ke Bellamy; Casey Dugan",
"journal": "",
"ref_id": "b16",
"title": "Explaining models: an empirical study of how explanations impact fairness judgment",
"year": "2019"
},
{
"authors": "Finale Doshi-Velez; Mason Kortz; Ryan Budish; Chris Bavitz; Sam Gershman; O' David; Kate Brien; Stuart Scott; James Schieber; David Waldo; Weinberger",
"journal": "",
"ref_id": "b17",
"title": "Accountability of ai under the law: The role of explanation",
"year": "2017"
},
{
"authors": "Upol Ehsan; Mark O Riedl",
"journal": "",
"ref_id": "b18",
"title": "Explainability pitfalls: Beyond dark patterns in explainable ai",
"year": "2021"
},
{
"authors": "Malin Eiband; Daniel Buschek; Alexander Kremer; Heinrich Hussmann",
"journal": "",
"ref_id": "b19",
"title": "The impact of placebic explanations on trust in intelligent systems",
"year": "2019"
},
{
"authors": "Maximilian Förster; Mathias Klier; Kilian Kluge; Irina Sigler",
"journal": "",
"ref_id": "b20",
"title": "Fostering human agency: A process for the design of user-centric xai systems",
"year": "2020"
},
{
"authors": "J Henry; Friendly",
"journal": "U. Pa. l. rev",
"ref_id": "b21",
"title": "Some kind of hearing",
"year": "1974"
},
{
"authors": "B Talia; Josh Gillis; Simons",
"journal": "JL & Innovation",
"ref_id": "b22",
"title": "Explanation< justification: Gdpr and the perils of privacy",
"year": "2019"
},
{
"authors": "V Claudia; Ronit Goldman; Bustin",
"journal": "IEEE",
"ref_id": "b23",
"title": "Trusting explainable autonomous driving: Simulated studies",
"year": "2022"
},
{
"authors": "Bryce Goodman; Seth Flaxman",
"journal": "AI magazine",
"ref_id": "b24",
"title": "European union regulations on algorithmic decision-making and a \"right to explanation",
"year": "2017"
},
{
"authors": "P Ellen; Julia Goodman; Trehu",
"journal": "",
"ref_id": "b25",
"title": "Ai audit washing and accountability",
"year": "2022"
},
{
"authors": "Colin M Gray; Yubo Kou; Bryan Battles; Joseph Hoggatt; Austin L Toombs",
"journal": "",
"ref_id": "b26",
"title": "The dark (patterns) side of ux design",
"year": "2018"
},
{
"authors": "Jarek Gryz; Nima Shahbazi",
"journal": "",
"ref_id": "b27",
"title": "Futility of a right to explanation",
"year": "2020"
},
{
"authors": "Riccardo Guidotti; Anna Monreale; Salvatore Ruggieri; Franco Turini; Fosca Giannotti; Dino Pedreschi",
"journal": "ACM computing surveys (CSUR)",
"ref_id": "b28",
"title": "A survey of methods for explaining black box models",
"year": "2018"
},
{
"authors": "David Gunning; David Aha",
"journal": "AI magazine",
"ref_id": "b29",
"title": "Darpa's explainable artificial intelligence (xai) program",
"year": "2019"
},
{
"authors": "Keith Hawkins",
"journal": "Wash. & Lee L. Rev",
"ref_id": "b30",
"title": "On legal decision-making",
"year": "1986"
},
{
"authors": "Mireille Hildebrandt",
"journal": "Oxford University Press",
"ref_id": "b31",
"title": "Law for computer scientists and other folk",
"year": "2020"
},
{
"authors": "Alon Jacovi; Ana Marasović; Tim Miller; Yoav Goldberg",
"journal": "",
"ref_id": "b32",
"title": "Formalizing trust in artificial intelligence: Prerequisites, causes and goals of human trust in ai",
"year": "2021"
},
{
"authors": "Margot E Kaminski",
"journal": "",
"ref_id": "b33",
"title": "The right to explanation, explained",
"year": ""
},
{
"authors": "Margot E Kaminski; Jennifer M Urban",
"journal": "Columbia Law Review",
"ref_id": "b34",
"title": "The right to contest ai",
"year": "2021"
},
{
"authors": "I Chimène; Keitner",
"journal": "McGill Law Journal",
"ref_id": "b35",
"title": "Explaining international acts",
"year": "2018"
},
{
"authors": "Benedict Kingsbury",
"journal": "European Journal of International Law",
"ref_id": "b36",
"title": "The concept of 'law'in global administrative law",
"year": "2009"
},
{
"authors": "Samuli Laato; Miika Tiainen; Najmul Islam; Matti Mäntymäki",
"journal": "Internet Research",
"ref_id": "b37",
"title": "How to explain ai systems to end users: a systematic literature review and research agenda",
"year": "2022"
},
{
"authors": "Markus Langer; Daniel Oster; Timo Speith; Holger Hermanns; Lena Kästner; Eva Schmidt; Andreas Sesing; Kevin Baum",
"journal": "Artificial Intelligence",
"ref_id": "b38",
"title": "What do we want from explainable artificial intelligence (xai)?-a stakeholder perspective on xai and a conceptual model guiding interdisciplinary xai research",
"year": "2021"
},
{
"authors": "Pantelis Linardatos; Vasilis Papastefanopoulos; Sotiris Kotsiantis",
"journal": "Entropy",
"ref_id": "b39",
"title": "Explainable ai: A review of machine learning interpretability methods",
"year": "2020"
},
{
"authors": " Zachary C Lipton",
"journal": "Queue",
"ref_id": "b40",
"title": "The mythos of model interpretability: In machine learning, the concept of interpretability is both important and slippery",
"year": "2018"
},
{
"authors": "David Lyell; Enrico Coiera",
"journal": "Journal of the American Medical Informatics Association",
"ref_id": "b41",
"title": "Automation bias and verification complexity: a systematic review",
"year": "2017"
},
{
"authors": "Gianclaudio Malgieri",
"journal": "Law and Business",
"ref_id": "b42",
"title": "just\" algorithms: justification (beyond explanation) of automated decisions under the general data protection regulation",
"year": "2021"
},
{
"authors": "Jerry L Mashaw",
"journal": "Fordham L. Rev",
"ref_id": "b43",
"title": "Small things like reasons are put in a jar: Reason and legitimacy in the administrative state",
"year": "2001"
},
{
"authors": "Jerry L Mashaw",
"journal": "Geo. Wash. L. Rev",
"ref_id": "b44",
"title": "Reasoned administration: The european union, the united states, and the project of democratic governance",
"year": "2007"
},
{
"authors": "A M Sheila; Mclean",
"journal": "Routledge",
"ref_id": "b45",
"title": "Autonomy, consent and the law",
"year": "2009"
},
{
"authors": "Mayuri Mehta; Indranath Vasile Palade; Chatterjee",
"journal": "Springer Nature",
"ref_id": "b46",
"title": "Explainable AI: Foundations, Methodologies and Applications",
"year": "2022"
},
{
"authors": "Gaspar Isaac Melsión; Ilaria Torre; Eva Vidal; Iolanda Leite",
"journal": "",
"ref_id": "b47",
"title": "Using explainability to help children understandgender bias in ai",
"year": "2021"
},
{
"authors": "Tim Miller; Piers Howe; Liz Sonenberg",
"journal": "",
"ref_id": "b48",
"title": "Explainable ai: Beware of inmates running the asylum or: How i learnt to stop worrying and love the social and behavioural sciences",
"year": "2017"
},
{
"authors": "Smitha Milli; Ludwig Schmidt; Anca D Dragan; Moritz Hardt",
"journal": "",
"ref_id": "b49",
"title": "Model reconstruction from model explanations",
"year": "2019"
},
{
"authors": "Brent Mittelstadt; Chris Russell; Sandra Wachter",
"journal": "",
"ref_id": "b50",
"title": "Explaining explanations in ai",
"year": "2019"
},
{
"authors": "Christoph Molnar",
"journal": "Lulu. com",
"ref_id": "b51",
"title": "Interpretable machine learning",
"year": "2020"
},
{
"authors": "Amit Ramaravind K Mothilal; Chenhao Sharma; Tan",
"journal": "",
"ref_id": "b52",
"title": "Explaining machine learning classifiers through diverse counterfactual explanations",
"year": "2020"
},
{
"authors": "Gabriel Nicholas",
"journal": "Geo. L. Tech. Rev",
"ref_id": "b53",
"title": "Explaining algorithmic decisions",
"year": "2019"
},
{
"authors": "Ingrid Nunes; Dietmar Jannach",
"journal": "User Modeling and User-Adapted Interaction",
"ref_id": "b54",
"title": "A systematic review and taxonomy of explanations in decision support and recommender systems",
"year": "2017"
},
{
"authors": "Alison B Powell",
"journal": "European Journal of Communication",
"ref_id": "b55",
"title": "Explanations as governance? investigating practices of explanation in algorithmic system design",
"year": "2021"
},
{
"authors": "Henry Prakken",
"journal": "",
"ref_id": "b56",
"title": "A top-level model of case-based argumentation for explanation",
"year": "2020"
},
{
"authors": "John Rawls",
"journal": "The University of Chicago Law Review",
"ref_id": "b57",
"title": "The idea of public reason revisited",
"year": "1997"
},
{
"authors": "Joseph Raz",
"journal": "Oxford University Press on Demand",
"ref_id": "b58",
"title": "The authority of law: essays on law and morality",
"year": "2009"
},
{
"authors": "",
"journal": "Protection Regulation. Regulation",
"ref_id": "b59",
"title": "679 of the european parliament and of the council",
"year": "2016"
},
{
"authors": "Marco Tulio Ribeiro; Sameer Singh; Carlos Guestrin",
"journal": "",
"ref_id": "b60",
"title": "why should i trust you?\" explaining the predictions of any classifier",
"year": "2016"
},
{
"authors": "Cynthia Rudin",
"journal": "Nature machine intelligence",
"ref_id": "b61",
"title": "Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead",
"year": "2019"
},
{
"authors": "Frederick Schauer",
"journal": "Stan. L. Rev",
"ref_id": "b62",
"title": "Giving reasons",
"year": "1994"
},
{
"authors": "D Andrew; Solon Selbst; Barocas",
"journal": "Fordham L. Rev",
"ref_id": "b63",
"title": "The intuitive appeal of explainable machines",
"year": "2018"
},
{
"authors": "Martin Shapiro",
"journal": "U. Chi. Legal F",
"ref_id": "b64",
"title": "The giving reasons requirement",
"year": "1992"
},
{
"authors": "Ben Shneiderman; Catherine Plaisant; Maxine Cohen; Steven Jacobs; Niklas Elmqvist; Nicholoas Diakopoulos",
"journal": "interactions",
"ref_id": "b65",
"title": "Grand challenges for hci researchers",
"year": "2016"
},
{
"authors": "Alison Smith-Renner; Ron Fan; Melissa Birchfield; Tongshuang Wu; Jordan Boyd-Graber; Daniel S Weld; Leah Findlater",
"journal": "",
"ref_id": "b66",
"title": "No explainability without accountability: An empirical study of explanations and feedback in interactive ml",
"year": "2020"
},
{
"authors": "Richard Tomsett; Dave Braines; Dan Harborne; Alun Preece; Supriyo Chakraborty",
"journal": "",
"ref_id": "b67",
"title": "Interpretable to whom? a role-based model for analyzing interpretable machine learning systems",
"year": "2018"
},
{
"authors": "Florian Tramèr; Fan Zhang; Ari Juels; Michael K Reiter; Thomas Ristenpart",
"journal": "",
"ref_id": "b68",
"title": "Stealing machine learning models via prediction apis",
"year": "2016"
},
{
"authors": "Miles Turpin; Julian Michael; Ethan Perez; Samuel R Bowman",
"journal": "",
"ref_id": "b69",
"title": "Language models don't always say what they think: Unfaithful explanations in chain-of-thought prompting",
"year": "2023"
},
{
"authors": "Sandra Wachter; Brent Mittelstadt; Chris Russell",
"journal": "Harv. JL & Tech",
"ref_id": "b70",
"title": "Counterfactual explanations without opening the black box: Automated decisions and the gdpr",
"year": "2017"
},
{
"authors": "David Weil; Archon Fung; Mary Graham; Elena Fagotto",
"journal": "Journal of Policy Analysis and Management: The Journal of the Association for Public Policy Analysis and Management",
"ref_id": "b71",
"title": "The effectiveness of regulatory disclosure policies",
"year": "2006"
},
{
"authors": "Katharina Weitz; Dominik Schiller; Ruben Schlagowski; Tobias Huber; Elisabeth André",
"journal": "Journal on Multimodal User Interfaces",
"ref_id": "b72",
"title": "let me explain!\": exploring the potential of virtual agents in explainable ai interaction design",
"year": "2021"
},
{
"authors": "Christopher Thomas R Wood; Megan Kelly; Bryan Roberts; Walsh",
"journal": "F1000Research",
"ref_id": "b73",
"title": "An interpretable machine learning model of biological age",
"year": "2019"
},
{
"authors": "Yujia Zhang; Kuangyan Song; Yiming Sun; Sarah Tan; Madeleine Udell",
"journal": "",
"ref_id": "b74",
"title": "why should you trust my explanation?",
"year": "2019"
},
{
"authors": "Joyce Zhou; Thorsten Joachims",
"journal": "",
"ref_id": "b75",
"title": "How to explain and justify almost any decision: Potential pitfalls for accountability in ai decision-making",
"year": "2022"
}
] | [] |
|
10.18653/v1/2020.acl-main.676 | 2023-10-18 | [
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b16",
"b35",
"b58",
"b13",
"b60",
"b30",
"b28",
"b35",
"b36",
"b52",
"b54",
"b39",
"b27",
"b67",
"b53",
"b34",
"b6",
"b36",
"b54",
"b27",
"b39",
"b53",
"b34",
"b6",
"b52",
"b67",
"b28",
"b35"
],
"table_ref": [],
"text": "A crucial property of human language learning is its compositional generalization (CG) -the algebraic ability to understand and produce a potentially infinite number of novel combinations from known components (Fodor and Pylyshyn, 1988;Lake et (Li et al., 2021) show the workflow of how humans exhibit CG. Suppose interpreters know the translation: [丢失了狗] for \"lost the dog\" and [他喜欢] for \"he liked\" (semantic information). When they first encounter \"lost the dog he liked\", they can correctly translate [丢失了他喜欢的 狗] instead of [丢失了狗他喜欢] depending on Pattern 2.3 (syntactic information). 2017). For example, if a person knows \"the doctor watches a movie\" [医生看电影] 1 and \"the lawyer\"\n[律师], then it is natural for the person to know the translation of \"the lawyer watches a movie\" is [律师看电影] even though they have never seen it before. Such nature is beneficial for generalizing to new compositions of previously observed elements, which is often required in real-world scenarios. Despite astonishing successes across a broad range of natural language understanding and generation tasks (Sutskever et al., 2014;Dong and Lapata, 2016;Vaswani et al., 2017), neural network models, in particular the very popular sequence-tosequence (seq2seq) architecture, are argued difficult to capture the compositional structure of human language (Lake and Baroni, 2018;Keysers et al., 2020;Li et al., 2021). A key reason for failure on CG is different semantic factors (e.g., lexical meaning and syntactic patterns) required by CG are entangled, which was proved explicitly or implicitly to exist in the representation of the encoder uppermost layer (encoder entanglement 1 The sentence in \"[]\" denotes the Chinese translation. problem) by previous studies (Li et al., 2019;Raunak et al., 2019;Russin et al., 2019;Liu et al., 2020bLiu et al., , 2021;;Jiang and Bansal, 2021;Zheng and Lapata, 2022a;Yin et al., 2022;Ruis and Lake, 2022;Li et al., 2022;Cazzaro et al., 2023). In other words, the syntactic and semantic representations of sequences are entangled.\nIn order to alleviate the encoder entanglement problem, one line of research on CG mainly concentrate on improving the encoder representation or separating the learning of syntax and semantics which adopt similar approaches to humans' strategies for CG (see Figure 1). Specifically, several works either produce two separate syntactic and semantic representations, and then compose them (Li et al., 2019;Russin et al., 2019;Jiang and Bansal, 2021) or design external modules, and then employ a multi-stage generation process (Liu et al., 2020b(Liu et al., , 2021;;Ruis and Lake, 2022;Li et al., 2022;Cazzaro et al., 2023). Moreover, some studies explore bag-of-words pre-training (Raunak et al., 2019), newly decoded target context (Zheng and Lapata, 2022a,b) or prototypes of token representations over the training set (Yin et al., 2022) to improve the encoder representation. Furthermore, we hypothesize that the source keys and values representations passing into different decoder layers are also entangled (keys, values entanglement problem), not just the representation of the encoder uppermost layer. We will further illustrate it in Section 5.1.\nTherefore, one natural question can be raised: how to alleviate keys, values entanglement problem. As a remedy, we examine CG from a new perspective to solve it, i.e., utilizing different encoder layers' information. We conduct preliminary analysis provided in Appendix A, and conclude that the bottom layers of the Transformer encoder contain more syntactic information and the top ones contain more semantic information. Inspired by this, we collect representations outputted by each encoder layer instead of separating the learning of syntax and semantics. So one intuitive solution to solve keys, values entanglement problem is to learn different and specific combinations of syntactic and semantic information (i.e., representations outputted by each encoder layer) for keys and values of different decoder layers. We argue that an effective composition is to provide different combinations for different tasks and a specific combination for a particular task. For example, the model can learn preference of layers in different levels of the encoder for different tasks (i.e., For A task, the information at encoder layer 0 may be more important, however, for B task, the information at encoder layer 5 may be more important). Additionally, the model can select which encoder layer of information is most suitable for itself (that is, which encoder layer of information is the most important) for a particular task. Inspired by that, we propose the composed layer (learnable scalars or vectors) to generate different specific source keys and values passing into different decoder layers for different particular tasks, since we argue that the learned scalars or vectors (i.e., different dynamic composition modes) by the model itself during training process can be dynamically adjusted for different particular tasks, and provide a way to learn preference of layers in different levels of the encoder for a particular task. Putting everything together, we propose COMPOSITION (Compose Syntactic and Semantic Representations), an extension to seq2seq models that learns to compose the syntactic and semantic representations of sequences dynamically for different tasks. COMPOSITION is simple yet effective, and mostly applicable to any seq2seq models without any dataset or task-specific modification.\nExperimental results on CFQ (Keysers et al., 2020) (semantic parsing) and CoGnition (Li et al., 2021) (machine translation, MT) empirically show that our method can improve generalization performance, outperforming competitive baselines and other techniques. Notably, COMPOSITION achieves 19.2% and 50.2% (about 32%, 20% relative improvements) for instance-level and aggregate-level error reduction rates on CoGnition. Extensive analyses demonstrate that composing the syntactic and semantic representations of sequences dynamically for different tasks leads to better generalization results."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b30",
"b28",
"b35",
"b3",
"b1",
"b65",
"b37",
"b36",
"b54",
"b46",
"b39",
"b24",
"b7",
"b45",
"b43",
"b19",
"b31",
"b9",
"b19",
"b68",
"b27",
"b68",
"b36",
"b54",
"b27",
"b52",
"b67",
"b66",
"b8",
"b64",
"b30",
"b35",
"b35",
"b59",
"b2",
"b67",
"b10",
"b51",
"b61",
"b5",
"b56",
"b62",
"b14",
"b63",
"b15",
"b23",
"b12"
],
"table_ref": [],
"text": "Compositional Generalization. After realizing existing neural models still struggle in scenarios requiring CG (Lake and Baroni, 2018;Keysers et al., 2020;Li et al., 2021), there have been various studies attempt to improve the model's ability of CG, including data augmentation (Andreas, 2020;Akyürek et al., 2021;Yang et al., 2022;Li et al., 2023), modifications on model architecture (Li et al., 2019;Russin et al., 2019;Nye et al., 2020;Liu et al., 2020cLiu et al., , 2021;;Zheng and Lapata, 2021;Herzig and Berant, 2021;Chaabouni et al., 2021;Mittal et al., 2022;Zheng et al., 2023), intermediate representations (Furrer et al., 2020;Herzig et al., 2021), meta-learning (Lake, 2019;Conklin et al., 2021), explorations on pre-trained language models (Furrer et al., 2020;Zhou et al., 2023), auxiliary objectives (Jiang and Bansal, 2021;Yin et al., 2023), two representations (Li et al., 2019;Russin et al., 2019;Jiang and Bansal, 2021) and enriching the encoder representation (Raunak et al., 2019;Zheng and Lapata, 2022a,b;Yin et al., 2022;Yao and Koller, 2022) (Cheng et al., 2020;Xu et al., 2021;Lake and Baroni, 2018;Li et al., 2021), including pre-training (Raunak et al., 2019), data augmentation (Guo et al., 2020a), datasets (Li et al., 2021), and enriching semantic information at tokenlevel (Thrush, 2020;Akyurek and Andreas, 2021;Zheng and Lapata, 2022a,b;Yin et al., 2022). Noteworthily, Dankers et al. (2022) argue that MT is a suitable and relevant testing ground to test CG in natural language. Different from them, we introduce a composed layer to compose different encoder layers' information dynamically, which is inspired by previous studies about analyzing Transformer (Raganato et al., 2018;Voita et al., 2019).\nEncoder Layer Fusion. Encoder layer fusion (En-coderFusion) is a technique to fuse all the encoder layers (instead of the uppermost layer) for seq2seq models, which has been proven beneficial, such as layer attention (Bapna et al., 2018;Shen et al., 2018;Wang et al., 2019), layer aggregation (Dou et al., 2018;Wang et al., 2018;Dou et al., 2019), and layer-wise coordination (He et al., 2018;Liu et al., 2020a). However, other studies show that exploiting low-layer encoder representations fails to improve model performance (Domhan, 2018). The essence of different EncoderFusion works is to explore different ways to combine information from different encoder layers. Our approach is essentially the same as EncoderFusion work, which explores different ways to combine information from different encoder layers, however, we propose a new way to combine them. Meanwhile, we consider that there are also three distinct differences. Firstly, our method exploits information from all encoder sub-layers and generates specific keys, values passing into different decoder layers while they do not. Secondly, our method shows the effectiveness of utilizing low-layer encoder representations while they have the opposite view (see Appendix D). Thirdly, we do not share the same motivation or task. Their work focuses on how to transform information across layers in deep neural network scenarios for seq2seq tasks. Our motivation is to compose the syntactic and semantic representations of sequences dynamically for CG."
},
{
"figure_ref": [],
"heading": "Methodology",
"publication_ref": [
"b60"
],
"table_ref": [],
"text": "We adopt the Transformer architecture (Vaswani et al., 2017) to clarify our method, however, our proposed method is mostly applicable to any seq2seq models. In the following, we first introduce the Transformer baseline (Section 3.1), and then our proposed COMPOSITION (Section 3.2)."
},
{
"figure_ref": [],
"heading": "Transformer",
"publication_ref": [
"b60",
"b17"
],
"table_ref": [],
"text": "The Transformer (Vaswani et al., 2017) is designed for sequence to sequence tasks which adopts an encoder-decoder architecture. The multi-layer encoder summarizes a source sequence into a contextualized representation and another multi-layer decoder produces the target sequence conditioned on the encoded representation.\nFormally, given a sequence of source sentence X = {x 1 , . . . , x S } and a sequence of target sentence Y = {y 1 , . . . , y T }, where S, T denote the number of source and target tokens, respectively. D = {(X, Y ), . . .} denotes a training corpus, V denotes the vocabulary of D, and θ denotes parameters of the Transformer model. The model aims to estimate the conditional probability p(y 1 , . . . , y T |x 1 , . . . , x S ):\np(Y |X; θ) = T +1 t=1 p(y t |y <t , X; θ), (1\n)\nwhere t is the index of each time step, y <t denotes a prefix of Y and each factor p(y t |X, y 1 , . . . , y t-1 ; θ) is defined as a sof tmax distribution of V.\nDuring training, the model are generally optimized with the cross-entropy (CE) loss, which is calculated as follows: During inference, the model predicts the probabilities of target tokens in an auto-regressive mode and generates target sentences using a heuristic search algorithm, such as beam search (Freitag and Al-Onaizan, 2017).\nL CE (θ) = - T +1 t=1 log p(y t |y <t , X; θ). (2)"
},
{
"figure_ref": [
"fig_1"
],
"heading": "COMPOSITION",
"publication_ref": [],
"table_ref": [],
"text": "Our proposed COMPOSITION extends the Transformer by introducing a composed layer between the encoder and decoder. Figure 2 shows the overall architecture of our approach."
},
{
"figure_ref": [],
"heading": "Composed Layer",
"publication_ref": [],
"table_ref": [],
"text": "The composed layer is a list consisting of 2N learnable vectors due to 2N source keys, values passing into N decoder layers, where each vector involves 2M learnable scalars or vectors. M, N denote the number of encoder and decoder layers respectively."
},
{
"figure_ref": [],
"heading": "Dynamic Combination",
"publication_ref": [],
"table_ref": [],
"text": "Here, we describe how to use the composed layer to compose collected representations dynamically for generating specific keys and values representations passing into different decoder layers. Let f Self -Attention and f F eed-F orward denote a Transformer self-attention sub-layer and feed-forward sub-layer respectively. The embedding layer of the Transformer encoder first maps X to embeddings H 0 , and then H 0 are fed into a Transformer self-attention sub-layer and feed-forward sub-layer to generate\nH SA 1 ∈ R d×S , H F F 1 ∈ R d×S respec-\ntively, where d denotes the hidden size. Next, each subsequent encoder layer takes the previous layer's output as input. The overall process is as follows:\nH SA 1 = f Self -Attention (H 0 ),(3)\nH F F 1 = f F eed-F orward (H SA 1 ),(4)\nH SA i = f Self -Attention (H F F i-1 ),(5)\nH F F i = f F eed-F orward (H SA i-1 ),(6)\nwhere 2 ≤ i ≤ M denote i-th encoder layer. Therefore, we can collect representations outputted by each encoder sub-layer\nH collect = {H SA 1 , H F F 1 , . . . , H SA M , H F F M }.\nThe keys and values of multi-head attention module of decoder layer l are defined to be:\nH l key = 2M i=1 w i k H i collect ,(7)\nH l value = 2M i=1 w i v H i collect ,(8)\nwhere w i k ∈ R, w i v ∈ R are learnable scalars or vectors and mutually different (e.g.\nw i k ̸ = w i v , w i k ̸ = w j k and w i v ̸ = w j v )\n, which weight each collected source representation in a dynamic linear manner. Eq. 7 and 8 provide a way to learn preference of sub-layers in different levels of the encoder."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [
"b28",
"b35"
],
"table_ref": [],
"text": "We mainly evaluate COMPOSITION on two comprehensive and realistic benchmarks for measuring CG, including CFQ (Keysers et al., 2020) and CoGnition (Li et al., 2021)."
},
{
"figure_ref": [
"fig_2"
],
"heading": "Experimental Settings",
"publication_ref": [
"b35",
"b28",
"b38",
"b43",
"b55",
"b35",
"b47",
"b35",
"b35",
"b48",
"b49"
],
"table_ref": [],
"text": "Datasets. CoGnition is a recently released realistic English → Chinese (En→Zh) translation dataset, which is used to systematically evaluate CG in MT scenarios. It consists of a training set of 196,246 sentence pairs, a validation set and a test set of 10,000 samples. In particular, it also has a dedicated synthetic test set (i.e., CG-test set) consisting of 10,800 sentences containing novel compounds, so that the ratio of compounds that are correctly translated can be computed to evaluate the model's ability of CG directly. CFQ is automatically generated from a set of rules in a way that precisely tracks which rules (atoms) and rule combinations (compounds) of each example. In this way, we can generate three splits with maximum compound divergence (MCD) while guaranteeing a small atom divergence between train and test sets, where large compound divergence denotes the test set involves more examples with unseen syntactic structures. We evaluate our method on all three splits. Each split dataset consists of a training set of 95,743, a validation set and a test set of 11,968 examples. Figure 3 shows examples of them. Data Preprocess. We follow the same settings of Li et al. (2021) and Keysers et al. (2020) to preprocess CoGnition and CFQ datasets separately. For CoGnition, we use an open-source Chinese tokenizer2 to preprocess Chinese and apply Moses tokenizer3 to preprocess English, which is the same in Lin et al. (2023) and Liu et al. (2023). We employ byte-pair encoding (BPE) (Sennrich et al., 2016) for Chinese with 3,000 merge operations, generating a vocabulary of 5,500 subwords. We do not apply BPE for English due to the small vocabulary (i.e., 2000). For CFQ, we use the GPT2-BPE tokenizer4 to preprocess source and target English text. Setup. For CoGnition and CFQ, we follow the same experimental settings and configurations of Li et al. (2021) and Zheng and Lapata (2022a) repspectively. We implement all comparison models and COMPOSITION with an open source Fairseq toolkit (Ott et al., 2019). More details are provided in Appendix B. Evaluation Metrics. For CoGnition, we use compound translation error rate (CTER (Li et al., 2021)) to measure the model's ability of CG. Specifically, instance-level CTER denotes the ratio of samples where the novel compounds are translated incorrectly, and aggregate-level CTER denotes the ratio of samples where the novel compounds suffer at least one incorrect translation when aggregating all 5 contexts. To calculate CTER, Li et al. (2021) manually construct a dictionary for all the atoms based on the training set, since each atom contains different translations. We also report characterlevel BLEU scores (Papineni et al., 2002) using SacreBLEU (Post, 2018) as a supplement. For CFQ, we use exact match accuracy to evaluate model performance, where natural language utterances are mapped to meaning representations."
},
{
"figure_ref": [],
"heading": "Model Settings",
"publication_ref": [
"b60"
],
"table_ref": [],
"text": "Machine Translation. We compare our method with previous competitive systems: (1) Transformer (Vaswani et al., 2017) "
},
{
"figure_ref": [],
"heading": "Results on CoGnition",
"publication_ref": [],
"table_ref": [],
"text": "The main results on CoGnition are shown in Table 1. We observe that: (1) COMPOSITION gives 20.4% CTER Inst and 52.0% CTER Aggr , with a significant improvement of 8.0% and 10.9% accordingly compared to the Transformer. Moreover, COMPOSITION significantly outperforms most baseline models under the almost same parameter settings,6 indicating composing the syntactic and semantic information of sequences dynamically for a particular task is more beneficial to CG. Although Transformer+CReg achieves slightly better performance and contains fewer parameters, it is more complex and costly compared with COM-POSITION;\n(2) COMPOSITION, COMPOSITION-Rela, COMPOSITION-Small and COMPOSITION-Deep can deliver various performance improvements, demonstrating the general effectiveness of our method; (3) COMPOSITION-Deep performs better than Bow, Dangle and Proto-Transformer, indicating that focusing on alleviating the encoder entanglement problem only can achieve part of goals of CG as mentioned in Section 1. Compared to SeqMix, the improvement of COMPOSITION is more significant (2.3% vs 10.9% CTER Aggr ). SeqMix utilizes linear interpolation in the input embedding space to reduce representation sparsity, and we suppose that the samples synthesized randomly may be unreasonable and harmful to model training; (4) It can be seen that Transformer is even slightly better than DLCL, indicating DLCL and COMPOSITION do not share the same motivation or scenario."
},
{
"figure_ref": [],
"heading": "Results on CFQ",
"publication_ref": [
"b35",
"b10"
],
"table_ref": [
"tab_3"
],
"text": "The main results on CFQ are presented in Table 2. We observe that: (1) RoBERTa is comparable to T5-11B, T5-11B-mod and outperforms other baseline systems without pre-training except HPD, indicating that pre-training indeed benefits CFQ; (2) COM- POSITION substantially boosts the performance of RoBERTa (43.4 → 59.4), about 37% relative improvements, and is in fact superior to T5-11B and T5-11B-mod. It also outperforms other baseline systems without pre-training except HPD. This result demonstrates that pre-training as a solution to CG also has limitations, and also indicates that COMPOSITION is complementary to pre-trained models; (3) HPD performs better than Dangle, RoBERTa+CReg and COMPOSITION, achieving 67.3 exact match accuracy, which is highly optimized for the CFQ dataset. On the contrary, COM-POSITION, RoBERTa+CReg and Dangle are generally applicable to any seq2seq models for solving any seq2seq tasks including MT, as mentioned in Section 4.3. However, compared with competitive performance on CoGnition, the improvements brought by COMPOSITION is relatively moderate, and even worse than Dangle. The underlying reason is related to a recent finding that compositionality in natural language is much more complex than the rigid, arithmetic-like operations (Li et al., 2021;Zheng and Lapata, 2022a;Dankers et al., 2022). MT is paradigmatically close to the tasks typically considered for testing compositionality in natural language, while our approach is more suitable for dealing with such scenarios."
},
{
"figure_ref": [],
"heading": "Analysis",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we conduct in-depth analyses of COMPOSITION to provide a comprehensive understanding of the individual contributions of each component. For all experiments, we train a COM-POSITION (6-6 encoder and decoder layers) instead of other experimental settings on the CoGnition dataset, unless otherwise specified."
},
{
"figure_ref": [
"fig_3"
],
"heading": "Effects of Specific Keys and Values of Different Decoder Layers",
"publication_ref": [],
"table_ref": [],
"text": "As mentioned in Section 1 and 3.2, we hypothesize that keys, values entanglement problem ex- ists. 7 It is clear that our hypothesized keys, values entanglement problem is an extension to encoder entanglement problem. We show curiosity about whether this problem exists, and COMPOSITION can alleviate it. In this experiment, we investigate its influence on CoGnition. As shown in Table 3, we observe certain improvements (-5.8% and -8.0% CTER Inst , -7.8% and -10.9% CTER Aggr ) when separately alleviating the encoder or keys, values entanglement problem. 8 It suggests that our method can alleviate both problems separately, and learning to compose information of different encoder layers dynamically can improve CG performance. Furthermore, the improvement brought from alleviating keys, values entanglement problem is more significant than that brought from alleviating encoder entanglement problem (52.0% vs 55.1% CTER Aggr ), demonstrating the reasonableness of keys, values entanglement problem.\nTo further illustrate the reasonableness of keys, values entanglement problem and understand how COMPOSITION alleviates it, we visualize the learned composition weights of COMPOSITION after normalized. 9 Specifically, we train COMPO-SITION on CoGnition and then extract W i k , W i v (see Section 3.2.2) to visualize them. Ideally, each key or value of different decoder layers should pay different attention weights to different encoder layers' information. As shown in Figure 4, the learned composition weights (after normalized) are mutually distinct for keys and values of different decoder layers, which implies COMPOSITION learns different dynamic composition modes for keys and values of every decoder layer respectively. In addition, it also indicates the reasonableness of keys, values entanglement problem we proposed, since keys and 7 It is noteworthy that the representation of the encoder upmost layer serves as the same key and value passing into every decoder layer in the Transformer. 8 We use one or 2N learnable vectors to generate one or 2N representations passing into N decoder layers. 9 We only use representations outputted by Eq. 6 for brevity. values of different decoder layers utilize more than just the information of the encoder topmost layer. More importantly, it also emphasizes our method provides an effective composition of syntactic and semantic information, i.e., a specific combination for a particular task. To further demonstrate it, we also provide a toy experiment in Appendix C."
},
{
"figure_ref": [],
"heading": "Effects of Composing Information of Encoder Layers or Sub-layers",
"publication_ref": [
"b33"
],
"table_ref": [
"tab_6"
],
"text": "As mentioned in Section 3, the Transformer encoder layer consists of two sub-layers. We assume that sub-layers may contain language information in different aspects, which may produce better generalization results. Therefore, we are curious about whether composing different encoder layers' or sub-layers' information is more beneficial to CG.\nIn this experiment, we investigate its influence on CoGnition. Specifically, we train COMPOSITION to compose representations outputted by either Eq. 5 or 6 or a combination of both dynamically.\nResults are presented in Table 4. We observe certain improvements (-6.2% and -5.8% CTER Inst ) when separately composing SA-and FF-level representations, where SA and FF denote representations outputted by Eq. 5 and 6 respectively. Furthermore, the combination of both them brings further improvement (-8.0% CTER Inst ), which illustrates that the information in different encoder sub-layers is complementary and has cumulative gains. It also suggests that syntactic and semantic information brought by SA or FF is similar, but slightly different (Li et al., 2020), and can improve generalization performance respectively. It can be seen that the results of COMPOSITION-SA and COMPOSITION-FF presented in The waiter he liked came 服务员来了，把那个恶霸赶走了。 他喜欢的服务员过来把那个恶霸赶走了。 by and chased the bully off.\n(The waiter came by and chased the bully off.) (The waiter he liked came by and chased the bully off.)\nThe waiter he liked 服务员喜欢拿起他的邮件。 他喜欢的服务员拿起了他的邮件。 picked up his mail.\n(The waiter liked to pick up his mail.) (The waiter he liked picked up his mail.)\nTable 5: Example translations of Transformer vs COMPOSITION. The bold characters denote the novel compounds and corresponding translations."
},
{
"figure_ref": [
"fig_4",
"fig_5"
],
"heading": "Effects on Compositional Generalization",
"publication_ref": [
"b35"
],
"table_ref": [],
"text": "Compound Length and Context Length. Longer compounds have more complex semantic information and longer contexts are harder to comprehend, making them more difficult to generalize (Li et al., 2021). We classify the test samples by compound length and context length, and calculate the CTER Inst . In Figure 5, we can observe that COM-POSITION generalizes better as the compound and context grows longer compared to Transformer. In particular, COMPOSITION gives a lower CTER by 11.0% over samples when the context length is more longer than 13 tokens. It suggests that our approach can better captures the compositional structure of human language. Complex Modifier. The postpositive modifier atom (MOD) is used to enrich the information of its preceding word (e.g., he liked in the phrase lost the dog he liked), which is challenging to translate due to word reordering from English to Chinese. We divide the test samples into two groups according to compounds with (w/) or without (wo/) MOD. In Figure 6, we observe that the advantage of COMPO-SITION grows larger in translating the compounds with MOD, demonstrating its superiority in processing complex semantic composition.\nCase Study. We present 3 source examples containing a novel compound the waiter he liked with MOD and 4 atoms, and their translations in Table 5. For all samples, correct translations denote that the novel compounds are translated correctly. COMPOSITION correctly translates the novel compounds across different contexts for all samples, while Transformer suffers from omitting different atoms. For example, the translation of the waiter is omitted in the first example, he liked is omitted in the second example and he is omitted in the third example. Our results not only contain the correct compound translations but also achieve better translation quality, while Transformer makes errors on unseen compositions, confirming the necessity of composing the syntactic and semantic representations of sequences dynamically."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "In "
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [],
"table_ref": [],
"text": "There are two limitations of our approach. Firstly, compared with competitive performance on CoGnition, the improvements brought by COMPOSITION on CFQ is relatively moderate, and even worse than some competitive methods. Hence, COMPOSITION is more suitable for tasks typically considered for testing compositionality in natural language. We strongly recommend researchers pay more attention to tasks evaluating compositionality on natural language. Meanwhile, we regard that designing a more general method that can improve generalization performance in both synthetic and natural scenarios is a promising direction to explore in the future. Secondly, our method is mostly applicable to any seq2seq models which adopt an encoder-decoder architecture instead of encoderonly or decoder-only architecture. However, the methodology of the proposed COMPOSITION is still rather general to any seq2seq models which adopt any architecture, since we can use the randomly initialized encoder or decoder to constitute the encoder-decoder architecture. "
},
{
"figure_ref": [],
"heading": "A Preliminary Analysis",
"publication_ref": [
"b69",
"b0"
],
"table_ref": [],
"text": "In this section, we analyze the amount of syntactic and semantic information captured by different encoder layers in the Transformer under MT scenarios. We aim at analyzing the representations learned by different encoder layers of different models through probing the encoder as input representation for various prediction tasks. We measure the importance of input features for various tasks by evaluating the ability of the decoder. Specifically, we use a fixed encoder representation as input and two different tasks, i.e., Part-of-Speech (POS) tagging, and Semantic tagging, to evaluate the syntactic and semantic information contained in different encoder layers respectively. The reason is that we assume if the input representation effectively captures a property (syntactic or semantic information), then the decoder can easily predict that property.\nTo explore the precise effects of information captured by different encoder layers, we train the Transformer on the WMT18 English → Chinese (EnZh, rich-resource), English → Estonian (EnEt, low-resource)10 by following the same settings of Raganato et al. (2018). 11 After training the MT models, we freeze the encoder parameters, and only train one decoder layer12 for each task, since we expect the decoder should not have overly significant impact on the model's performance of different tasks. We then analyze the amount of syntactic and semantic information in different encoder layers via evaluating the different encoder layers' performance of corresponding task. We use the Universal Dependencies English Web Treebank v2.0 (Zeman et al., 2017) for POS tagging (syntactic task) and the annotated data from the Parallel Meaning Bank (PMB) (Abzianidze et al., 2017) for Semantic tagging (semantic task). 13 We use precision to evaluate model performance.\nResults on POS tagging and Semantic tagging are presented in Figure 7 and 8 respectively. We observe that:\n• For EnEt and EnZh, the performance tends to decrease as the number of layers increase.\n• For EnEt and EnZh, the performance tends to increase as the number of layers increase.\nTherefore, we can conclude that the bottom layers of the Transformer encoder contain more vast amounts of multilingual sentences or bilingual sentence pairs. It is contrary to the compositional generalization task itself, since we can not guarantee that every sentence in the test set is a novel combination from known components for language models. Second, it is unfair to compare large language models with systems without pre-training. We strongly recommend researchers pay more attention to conduct experiments on CoGniton without language models."
},
{
"figure_ref": [],
"heading": "Acknowledgement",
"publication_ref": [],
"table_ref": [],
"text": "We thank all the anonymous reviewers for their insightful and valuable comments. This work is supported by National key R&D Program of China (Grant no.2022ZD0116101), the Key Support Project of NSFC-Liaoning Joint Foundation (Grant no. U1908216), and the Project of Research and Development for Neural Machine Translation Models between Cantonese and Mandarin (No. WT135-76)."
},
{
"figure_ref": [],
"heading": "",
"publication_ref": [],
"table_ref": [],
"text": "syntactic information and the top ones contain more semantic information, and the information encoded by each encoder layer transforms from syntactic to semantic as the number of layers increase."
},
{
"figure_ref": [],
"heading": "B Experimental Settings",
"publication_ref": [
"b35"
],
"table_ref": [],
"text": "For CoGnition, we set hidden size to 512 and feedforward dimension to 1,024. The number of encoder and decoder layers are 6, 6 and the number of attention heads are 4. The model parameters are optimized by Adam (Kingma and Ba, 2015), with β 1 = 0.9, β 2 = 0.98. The learning rate is set to 5e-4 and the number of warm-steps is 4000. We set max tokens as 8,192 tokens for iteration. We use one GeForce GTX 2080Ti for training with 100,000 steps and decoding. We report the average performance over 6 random seeds provided in Li et al. (2021). We train all COMPOSITION models from scratch. For CFQ, we use the base RoBERTa with 12 encoder layers, which is combined with a Transformer decoder that has 2 decoder layers with hidden size 256 and feed-forward dimension 512.\nWe use a separate target vocabulary. The number of attention heads are 8. The model parameters are optimized by Adam (Kingma and Ba, 2015), with β 1 = 0.9, β 2 = 0.98. The learning rate is set to 1e-4 and the number of warm-steps is 4000. We set max tokens as 4,096 tokens for iteration. We use one GeForce GTX 2080Ti for training with 45,000 steps and decoding. We report the average performance over 3 random seeds provided in Zheng and Lapata (2022a). We train COMPOSITION built on top of RoBERTa with full parameter fine-tuning."
},
{
"figure_ref": [],
"heading": "C Effects of the Effective Composition",
"publication_ref": [],
"table_ref": [],
"text": "As mentioned in Section 3, we introduce the composed layer between the encoder and decoder to compose different encoder sub-layers' information dynamically to generate specific keys and values passing into different decoder layers. We show curiosity about whether the composed layer can fuse all encoder sub-layers' information effectively. Therefore, we conduct a toy experiment on CoGnition. Specifically, all encoder sub-layers' information is accumulated to serve as the same key and value passing into every decoder layer (called Transformer-accu), 14 rather than composing them dynamically like we do. Results are listed in Table 6. Transformer-accu even fails to train. It suggests that even if the syntactic and semantic information of sequences is considered, the inappropriate combinations will instead bring noise to significantly affect the model's CG performance."
},
{
"figure_ref": [],
"heading": "D Effects of Representations from Low-layer Encoder",
"publication_ref": [],
"table_ref": [],
"text": "To verify the low-layer encoder representations are also essential to our approach, we only evaluate our approach on CoGnition with the collected encoder representations of the top three layers. Results are presented in Table 7. We can observe that only composing the representations of the top three encoder layers leads to a sharp drop in performance (27.0% vs 20.4% CTER Inst ), but still outperforms the Transformer baseline (27.0% vs 28.4% CTER Inst ). It further demonstrates the distinct difference between our method and the findings introduced by previous studies on EncoderFusion. It also reflects our starting point is correct, i.e., exploring how to compose syntactic and semantic information. It can be seen that COMPOSITION's performance is dramatically reduced given only semantic information (the last three encoder layers' information)."
},
{
"figure_ref": [],
"heading": "E Reasons for Experiments on CoGnition without Language Models",
"publication_ref": [],
"table_ref": [],
"text": "We do not conduct experiments on CoGnition with language models for two reasons. First, CoGnition is constructed to test CG performance in MT scenarios with simple sentence pairs (see Figure 3), however, language models are trained on"
}
] | Recent studies have shown that sequence-tosequence (seq2seq) models struggle with compositional generalization (CG), i.e., the ability to systematically generalize to unseen compositions of seen components. There is mounting evidence that one of the reasons hindering CG is the representation of the encoder uppermost layer is entangled, i.e., the syntactic and semantic representations of sequences are entangled. However, we consider that the previously identified representation entanglement problem is not comprehensive enough. Additionally, we hypothesize that the source keys and values representations passing into different decoder layers are also entangled. Starting from this intuition, we propose COMPOSI-TION (Compose Syntactic and Semantic Representations), an extension to seq2seq models which learns to compose representations of different encoder layers dynamically for different tasks, since recent studies reveal that the bottom layers of the Transformer encoder contain more syntactic information and the top ones contain more semantic information. Specifically, we introduce a composed layer between the encoder and decoder to compose different encoder layers' representations to generate specific keys and values passing into different decoder layers. COMPOSITION achieves competitive results on two comprehensive and realistic benchmarks, which empirically demonstrates the effectiveness of our proposal. | Learning to Compose Representations of Different Encoder Layers towards Improving Compositional Generalization | [
{
"figure_caption": "Figure 1 :1Figure 1: Examples from CoGnition(Li et al., 2021) show the workflow of how humans exhibit CG. Suppose interpreters know the translation: [丢失了狗] for \"lost the dog\" and [他喜欢] for \"he liked\" (semantic information). When they first encounter \"lost the dog he liked\", they can correctly translate [丢失了他喜欢的 狗] instead of [丢失了狗他喜欢] depending on Pattern 2.3 (syntactic information).",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: Architecture of COMPOSITION based on the Transformer. The bright yellow block in the middle denotes the composed layer introduced in Section 3.2. The red line denotes that we collect representations of the same positions for the rest encoder layers.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: Examples of CFQ and CoGnition.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: Learned composition weights (after normalized) that each encoder layer (y-axis) attending to keys or values of different decoder layers (x-axis).",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 :5Figure 5: CTER Inst of COMPOSITION and Transformer over the different compound and context lengths.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Figure 6 :6Figure 6: CTER Inst on compounds w/o and w/ MOD.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "Figure 8 :8Figure 8: Precision (%) against different encoder layers' representations as input on the test set of Semantic tagging task.",
"figure_data": "",
"figure_id": "fig_7",
"figure_label": "8",
"figure_type": "figure"
},
{
"figure_caption": "CTERs (%) on CoGnition. We report instance-level and aggregate-level CTERs in the CG-test set, separated by \"/\". In addition, we also report the commonly used metric BLEU score in MT tasks. \"-\" denotes that the results are not provided in the original paper. Results are averaged over 6 random runs.",
"figure_data": "Model#ParamsCompound Translation Error Rate (CTER) ↓BLEU ↑NPVPPPTotal∆Transformer35M24.7%/55.2%24.8%/59.5%35.7%/73.9%28.4%/62.9%-/-59.5Transformer-Rela35M30.1%/58.1%27.6%/61.2%38.5%/74.1%32.1%/64.5%+3.7%/+1.6%59.1Transformer-Small25M25.1%/56.9%25.6%/60.3%39.1%/75.0%29.9%/64.5%+1.5%/+1.6%59.0Transformer-Deep40M23.3%/51.6%24.1%/58.0%33.8%/72.6%27.0%/60.7%-1.4%/-2.0%60.1Bow35M22.2%47.9%24.8%/55.6%35.0%/73.2%27.3%/58.9%-1.1%/-3.0%-SeqMix35M24.5%/49.7%26.9%/58.9%34.4%/73.1%28.6%/60.6%+0.2%/-2.3%-Dangle35M-/--/--/-24.4%/55.5%-5.0%/-7.4%59.7Proto-Transformer42M14.1%/36.5%22.1%/50.9%28.9%/68.2%21.7%/51.8%-6.7%/-11.1%60.1Transformer+CReg25M-/--/--/-20.2%/48.3%-8.2%/-14.6%61.3R-Danglesep70M-/--/--/-16.0%/42.1% -12.4%/-20.8%63.4DLCL35M-/--/--/-28.4%/67.9%+0.0%/+5.0%59.2COMPOSITION35M10.0%/32.6%22.1%/54.8%29.2%/68.5%20.4%/52.0%-8.0%/-10.9%61.5COMPOSITION-Rela35M15.5%/39.2%22.4%/54.0%29.1%/67.3%22.3%/53.5%-6.1%/-9.4%61.6COMPOSITION-Small25M14.3%/40.3%24.4%/58.1%34.5%/73.4%24.4%/57.3%-4.0%/-5.6%60.1COMPOSITION-Deep40M11.4%/34.7% 19.5%/50.4% 26.7%/65.6%19.2%/50.2%-9.2%/-12.7%62.0CFQCoGnition\"Did M0 direct M1\"",
"figure_id": "tab_2",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": ": first proposes a new Exact-match accuracy on different MCD splits of CFQ. Results are averaged over 3 random runs.",
"figure_data": "ModelMCD1 MCD2 MCD3 MeanModelAlleviate E K, VCTERInst ↓CTERAggr ↓LSTM+attention Transformer28.9 34.95.0 8.210.8 10.614.9 17.9Transformer✗✗28.4%62.9%Universal Transformer37.48.111.318.9COMPOSITION ✓✗22.6% (-5.8%)55.1% (-7.8%)Evolved Transformer42.49.310.820.8COMPOSITION ✓✓20.4% (-8.0%) 52.0% (-10.9%)CGPS13.21.66.67.1NSEN5.10.92.32.8T5-11B61.430.131.240.9T5-11B-mod61.631.333.342.1RoBERTa60.633.636.043.4HPD72.066.163.967.3Dangle78.359.560.466.1RoBERTa+CReg74.853.358.362.1COMPOSITION72.853.252.259.4encoder-decoder architecture based solely on atten-tion mechanisms; (2) Transformer-Rela: only re-places sinusoidal (absolute) positional embeddingwith a relative one; (3) Transformer-Small: only de-creases the number of encoder layers and decoderlayers to 4, 4 respectively; (4) Transformer-Deep:only increases the number of encoder layers to 8;(5) Bow (Raunak et al., 2019): uses bag-of-wordspre-training to improve the representation of the en-coder upmost layer; (6) SeqMix (Guo et al., 2020a):synthesizes examples to encourage compositionalbehavior; (7) Dangle (Zheng and Lapata, 2022a):adaptively re-encodes (at each time step) the sourceinput to disentangle the representation of the en-coder upmost layer; 5 (8) Proto-Transformer (Yinet al., 2022): integrates prototypes of token rep-resentations over the training set into the sourceencoding to achieve the goal of categorization; (9)Transformer+CReg (Yin et al., 2023): promotesrepresentation consistency across samples and pre-diction consistency for a single sample; (10) R-Dangle sep (Zheng and Lapata, 2022b): disentan-gles their representations and only re-encode keysperiodically, at some interval; (11) DLCL (Wanget al., 2019): proposes an approach based on dy-namic linear combination of layers (DLCL), andis one of the very popular EnocderFusion work.Our method is built on top of (1)-(4), i.e., COM-POSITION, COMPOSITION-Rela, COMPOSITION-Small and COMPOSITION-Deep. We also providereasons for experiments on CoGnition without lan-guage models (see Appendix E).Semantic Parsing. We compare our method withprevious competitive systems: (1) LSTM + atten-",
"figure_id": "tab_3",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "CTERs (%) against alleviating E or K,V on the CG-test set, where CTER Inst and CTER",
"figure_data": "Aggr denote",
"figure_id": "tab_4",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "CTERs (%) against composing different source information on the CG-test set.",
"figure_data": "",
"figure_id": "tab_6",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Table 4 are basically the same, and the improvements brought by the combination of both them is relatively moderate.",
"figure_data": "SourceTransformerCOMPOSITIONThe waiter he liked他喜欢穿对方的衣服。他喜欢的服务员穿着彼此的衣服。wore each other's clothes.(He liked to wear each other's clothes.)(The waiter he liked wore each other's clothes.)",
"figure_id": "tab_7",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Lei Lin, Zhaohong Lai, Binling Wang, Shan Liu, Biao Fu, Wenhao Rao, Peigen Ye, Yidong Chen, and Xiaodong Shi. 2023. Layer-wise representation fusion for compositional generalization. arXiv preprint arXiv:2307.10799. Denny Zhou, Nathanael Schärli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc V Le, and Ed H. Chi. 2023. Least-to-most prompting enables complex reasoning in large language models. In The Eleventh International Conference on Learning Representations.",
"figure_data": "92EnEt90EnZhPrecision (%)84 86 888280012 Encoder Layers 345Figure 7: Precision (%) against different encoder layers'representations as input on the test set of POS taggingtask.Yafang Zheng,Hao Zheng and Mirella Lapata. 2021. Compositionalgeneralization via semantic tagging. In Findingsof the Association for Computational Linguistics:EMNLP 2021, Virtual Event / Punta Cana, Domini-can Republic, 16-20 November, 2021, pages 1022-1032. Association for Computational Linguistics.Hao Zheng and Mirella Lapata. 2022a. Disentangled se-quence to sequence learning for compositional gener-alization. In Proceedings of the 60th Annual Meetingof the Association for Computational Linguistics (Vol-ume 1: Long Papers). Association for ComputationalLinguistics.Hao Zheng and Mirella Lapata. 2022b.Real-world compositional generalization with disentan-gled sequence-to-sequence learning. arXiv preprintarXiv:2212.05982.",
"figure_id": "tab_9",
"figure_label": "",
"figure_type": "table"
}
] | Lei Lin; Shuangtao Li; Yafang Zheng; Biao Fu; Shan Liu; Yidong Chen; Xiaodong Shi | [
{
"authors": "Lasha Abzianidze; Johannes Bjerva; Kilian Evang; Hessel Haagsma; Rik Van Noord; Pierre Ludmann; Duc-Duy Nguyen; Johan Bos",
"journal": "Association for Computational Linguistics",
"ref_id": "b0",
"title": "The parallel meaning bank: Towards a multilingual corpus of translations annotated with compositional meaning representations",
"year": "2017-04-03"
},
{
"authors": "Ekin Akyürek; Afra Feyza Akyürek; Jacob Andreas",
"journal": "",
"ref_id": "b1",
"title": "Learning to recombine and resample data for compositional generalization",
"year": "2021-05-03"
},
{
"authors": "Ekin Akyurek; Jacob Andreas",
"journal": "Association for Computational Linguistics",
"ref_id": "b2",
"title": "Lexicon learning for few shot sequence modeling",
"year": "2021"
},
{
"authors": "Jacob Andreas",
"journal": "Association for Computational Linguistics",
"ref_id": "b3",
"title": "Good-enough compositional data augmentation",
"year": "2020"
},
{
"authors": "Dzmitry Bahdanau; Kyunghyun Cho; Yoshua Bengio",
"journal": "",
"ref_id": "b4",
"title": "Neural machine translation by jointly learning to align and translate",
"year": "2015-05-07"
},
{
"authors": "Ankur Bapna; Mia Chen; Orhan Firat; Yuan Cao; Yonghui Wu",
"journal": "Association for Computational Linguistics",
"ref_id": "b5",
"title": "Training deeper neural machine translation models with transparent attention",
"year": "2018"
},
{
"authors": "Francesco Cazzaro; Davide Locatelli; Ariadna Quattoni; Xavier Carreras",
"journal": "Association for Computational Linguistics",
"ref_id": "b6",
"title": "Translate first reorder later: Leveraging monotonicity in semantic parsing",
"year": "2023-05-02"
},
{
"authors": "Rahma Chaabouni; Roberto Dessì; Eugene Kharitonov",
"journal": "Association for Computational Linguistics",
"ref_id": "b7",
"title": "Can transformers jump around right in natural language? assessing performance transfer from SCAN",
"year": "2021"
},
{
"authors": "Yong Cheng; Lu Jiang; Wolfgang Macherey; Jacob Eisenstein",
"journal": "Association for Computational Linguistics",
"ref_id": "b8",
"title": "AdvAug: Robust adversarial augmentation for neural machine translation",
"year": "2020"
},
{
"authors": "Henry Conklin; Bailin Wang; Kenny Smith; Ivan Titov",
"journal": "Association for Computational Linguistics",
"ref_id": "b9",
"title": "Meta-learning to compositionally generalize",
"year": "2021-08-01"
},
{
"authors": "Verna Dankers; Elia Bruni; Dieuwke Hupkes",
"journal": "Association for Computational Linguistics",
"ref_id": "b10",
"title": "The paradox of the compositionality of natural language: A neural machine translation case study",
"year": "2022-05-22"
},
{
"authors": "Mostafa Dehghani; Stephan Gouws; Oriol Vinyals; Jakob Uszkoreit; Lukasz Kaiser",
"journal": "",
"ref_id": "b11",
"title": "Universal transformers",
"year": "2019-05-06"
},
{
"authors": "Tobias Domhan",
"journal": "",
"ref_id": "b12",
"title": "How much attention do you need? a granular analysis of neural machine translation architectures",
"year": "2018"
},
{
"authors": "Li Dong; Mirella Lapata",
"journal": "",
"ref_id": "b13",
"title": "Language to logical form with neural attention",
"year": "2016"
},
{
"authors": "Zi-Yi Dou; Zhaopeng Tu; Xing Wang; Shuming Shi; Tong Zhang",
"journal": "Association for Computational Linguistics",
"ref_id": "b14",
"title": "Exploiting deep representations for neural machine translation",
"year": "2018"
},
{
"authors": "Zi-Yi Dou; Zhaopeng Tu; Xing Wang; Longyue Wang; Shuming Shi; Tong Zhang",
"journal": "AAAI Press",
"ref_id": "b15",
"title": "Dynamic layer aggregation for neural machine translation with routing-by-agreement",
"year": "2019-01-27"
},
{
"authors": "Jerry A Fodor; Zenon W Pylyshyn",
"journal": "Cognition",
"ref_id": "b16",
"title": "Connectionism and cognitive architecture: A critical analysis",
"year": "1988"
},
{
"authors": "Markus Freitag; Yaser Al-Onaizan",
"journal": "Association for Computational Linguistics",
"ref_id": "b17",
"title": "Beam search strategies for neural machine translation",
"year": "2017"
},
{
"authors": "Karlis Freivalds; Emils Ozolins; Agris Sostaks",
"journal": "",
"ref_id": "b18",
"title": "Neural shuffle-exchange networks -sequence processing in o(n log n) time",
"year": "2019-12-08"
},
{
"authors": "Daniel Furrer; Nathan Marc Van Zee; Nathanael Scales; Schärli",
"journal": "",
"ref_id": "b19",
"title": "Compositional generalization in semantic parsing: Pre-training vs. specialized architectures",
"year": "2020"
},
{
"authors": "Demi Guo; Yoon Kim; Alexander Rush",
"journal": "",
"ref_id": "b20",
"title": "a. Sequence-level mixed sample data augmentation",
"year": "2020"
},
{
"authors": "Yinuo Guo; Zeqi Lin; Jian-Guang Lou; Dongmei Zhang",
"journal": "",
"ref_id": "b21",
"title": "Hierarchical poset decoding for compositional generalization in language",
"year": "2020"
},
{
"authors": "Curran Associates; Inc ",
"journal": "",
"ref_id": "b22",
"title": "",
"year": ""
},
{
"authors": "Tianyu He; Xu Tan; Yingce Xia; Di He; Tao Qin; Zhibo Chen; Tie-Yan Liu",
"journal": "",
"ref_id": "b23",
"title": "Layer-wise coordination between encoder and decoder for neural machine translation",
"year": "2018-12-03"
},
{
"authors": "Jonathan Herzig; Jonathan Berant",
"journal": "Association for Computational Linguistics",
"ref_id": "b24",
"title": "Spanbased semantic parsing for compositional generalization",
"year": "2021-08-01"
},
{
"authors": "Jonathan Herzig; Peter Shaw; Ming-Wei Chang; Kelvin Guu; Panupong Pasupat; Yuan Zhang",
"journal": "",
"ref_id": "b25",
"title": "Unlocking compositional generalization in pre-trained models using intermediate representations",
"year": "2021"
},
{
"authors": "Sepp Hochreiter; Jürgen Schmidhuber",
"journal": "Neural Comput",
"ref_id": "b26",
"title": "Long short-term memory",
"year": "1997"
},
{
"authors": "Yichen Jiang; Mohit Bansal",
"journal": "Association for Computational Linguistics",
"ref_id": "b27",
"title": "Inducing transformer's compositional generalization ability via auxiliary sequence prediction tasks",
"year": "2021-07-11"
},
{
"authors": "Daniel Keysers; Nathanael Schärli; Nathan Scales; Hylke Buisman; Daniel Furrer; Sergii Kashubin; Nikola Momchev; Danila Sinopalnikov; Lukasz Stafiniak; Tibor Tihon; Dmitry Tsarkov; Xiao Wang; Marc Van Zee; Olivier Bousquet",
"journal": "",
"ref_id": "b28",
"title": "Measuring compositional generalization: A comprehensive method on realistic data",
"year": "2020-04-26"
},
{
"authors": "P Diederik; Jimmy Kingma; Ba",
"journal": "",
"ref_id": "b29",
"title": "Adam: A method for stochastic optimization",
"year": "2015-05-07"
},
{
"authors": "Brenden Lake; Marco Baroni",
"journal": "PMLR",
"ref_id": "b30",
"title": "Generalization without systematicity: On the compositional skills of sequence-to-sequence recurrent networks",
"year": "2018"
},
{
"authors": "M Brenden; Lake",
"journal": "",
"ref_id": "b31",
"title": "Compositional generalization through meta sequence-to-sequence learning",
"year": "2019-12-08"
},
{
"authors": " Brenden M Lake; Joshua B Tomer D Ullman; Samuel J Tenenbaum; Gershman",
"journal": "Behavioral and brain sciences",
"ref_id": "b32",
"title": "Building machines that learn and think like people",
"year": "2017"
},
{
"authors": "Bei Li; Ziyang Wang; Hui Liu; Yufan Jiang; Quan Du; Tong Xiao; Huizhen Wang; Jingbo Zhu",
"journal": "",
"ref_id": "b33",
"title": "Shallow-to-deep training for neural machine translation",
"year": "2020"
},
{
"authors": "Qing Li; Yixin Zhu; Yitao Liang; Ying Nian Wu; Song-Chun Zhu; Siyuan Huang",
"journal": "",
"ref_id": "b34",
"title": "Neuralsymbolic recursive machine for systematic generalization",
"year": "2022"
},
{
"authors": "Yafu Li; Yongjing Yin; Yulong Chen; Yue Zhang",
"journal": "Association for Computational Linguistics",
"ref_id": "b35",
"title": "On compositional generalization of neural machine translation",
"year": "2021"
},
{
"authors": "Yuanpeng Li; Liang Zhao; Jianyu Wang; Joel Hestness",
"journal": "Association for Computational Linguistics",
"ref_id": "b36",
"title": "Compositional generalization for primitive substitutions",
"year": "2019"
},
{
"authors": "Zhaoyi Li; Ying Wei; Defu Lian",
"journal": "",
"ref_id": "b37",
"title": "Learning to substitute spans towards improving compositional generalization",
"year": "2023"
},
{
"authors": "Lei Lin; Shuangtao Li; Xiaodong Shi",
"journal": "",
"ref_id": "b38",
"title": "Leapt: Learning adaptive prefix-to-prefix translation for simultaneous machine translation",
"year": "2023"
},
{
"authors": "Chenyao Liu; Shengnan An; Zeqi Lin; Qian Liu; Bei Chen; Jian-Guang Lou; Lijie Wen; Nanning Zheng; Dongmei Zhang",
"journal": "Association for Computational Linguistics",
"ref_id": "b39",
"title": "Learning algebraic recombination for compositional generalization",
"year": "2021-08-01"
},
{
"authors": "Fenglin Liu; Xuancheng Ren; Guangxiang Zhao; Xu Sun; Liangyou Li",
"journal": "",
"ref_id": "b40",
"title": "Layer-wise crossview decoding for sequence-to-sequence learning",
"year": "2020"
},
{
"authors": "Qian Liu; Shengnan An; Jian-Guang Lou; Bei Chen; Zeqi Lin; Yan Gao; Bin Zhou; Nanning Zheng; Dongmei Zhang",
"journal": "",
"ref_id": "b41",
"title": "Compositional generalization by learning analytical expressions",
"year": "2020-12-06"
},
{
"authors": "Qian Liu; Shengnan An; Jian-Guang Lou; Bei Chen; Zeqi Lin; Yan Gao; Bin Zhou; Nanning Zheng; Dongmei Zhang",
"journal": "",
"ref_id": "b42",
"title": "Compositional generalization by learning analytical expressions",
"year": "2020-12-06"
},
{
"authors": "Shan Liu; Yafang Zheng; Lei Lin; Yidong Chen; Xiaodong Shi",
"journal": "Springer",
"ref_id": "b43",
"title": "A novel pos-guided data augmentation method for sign language gloss translation",
"year": "2023"
},
{
"authors": "Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov",
"journal": "",
"ref_id": "b44",
"title": "Roberta: A robustly optimized bert pretraining approach",
"year": "2019"
},
{
"authors": "Sarthak Mittal; Chandra Sharath; Irina Raparthy; Yoshua Rish; Guillaume Bengio; Lajoie",
"journal": "",
"ref_id": "b45",
"title": "Compositional attention: Disentangling search and retrieval",
"year": "2022-04-25"
},
{
"authors": "I Maxwell; Armando Nye; Josh Solar-Lezama; Brenden M Tenenbaum; Lake",
"journal": "",
"ref_id": "b46",
"title": "Learning compositional rules via neural program synthesis",
"year": "2020-12-06"
},
{
"authors": "Myle Ott; Sergey Edunov; Alexei Baevski; Angela Fan; Sam Gross; Nathan Ng; David Grangier; Michael Auli",
"journal": "Association for Computational Linguistics",
"ref_id": "b47",
"title": "fairseq: A fast, extensible toolkit for sequence modeling",
"year": "2019"
},
{
"authors": "Kishore Papineni; Salim Roukos; Todd Ward; Wei-Jing Zhu",
"journal": "ACL",
"ref_id": "b48",
"title": "Bleu: a method for automatic evaluation of machine translation",
"year": "2002-07-06"
},
{
"authors": "Matt Post",
"journal": "Association for Computational Linguistics",
"ref_id": "b49",
"title": "A call for clarity in reporting BLEU scores",
"year": "2018-10-31"
},
{
"authors": "Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu",
"journal": "J. Mach. Learn. Res",
"ref_id": "b50",
"title": "Exploring the limits of transfer learning with a unified text-to-text transformer",
"year": "2020"
},
{
"authors": "Alessandro Raganato; Jörg Tiedemann",
"journal": "",
"ref_id": "b51",
"title": "An analysis of encoder representations in transformerbased machine translation",
"year": "2018"
},
{
"authors": "Vaibhav Vikas Raunak; Florian Kumar; Metze",
"journal": "",
"ref_id": "b52",
"title": "On compositionality in neural machine translation",
"year": "2019"
},
{
"authors": "Laura Ruis; Brenden M Lake",
"journal": "",
"ref_id": "b53",
"title": "Improving systematic generalization through modularity and augmentation",
"year": "2022"
},
{
"authors": "Jake Russin; Jason Jo; C O' Randall; Yoshua Reilly; Bengio",
"journal": "",
"ref_id": "b54",
"title": "Compositional generalization in a deep seq2seq model by separating syntax and semantics",
"year": "2019"
},
{
"authors": "Rico Sennrich; Barry Haddow; Alexandra Birch",
"journal": "Association for Computational Linguistics",
"ref_id": "b55",
"title": "Neural machine translation of rare words with subword units",
"year": "2016"
},
{
"authors": "Yanyao Shen; Xu Tan; Di He; Tao Qin; Tie-Yan Liu",
"journal": "Association for Computational Linguistics",
"ref_id": "b56",
"title": "Dense information flow for neural machine translation",
"year": "2018"
},
{
"authors": "David R So; V Quoc; Chen Le; Liang",
"journal": "",
"ref_id": "b57",
"title": "The evolved transformer",
"year": "2019-06"
},
{
"authors": "Ilya Sutskever; Oriol Vinyals; Quoc V Le",
"journal": "Advances in neural information processing systems",
"ref_id": "b58",
"title": "Sequence to sequence learning with neural networks",
"year": "2014"
},
{
"authors": "Tristan Thrush",
"journal": "",
"ref_id": "b59",
"title": "Compositional neural machine translation by removing the lexicon from syntax",
"year": "2020"
},
{
"authors": "Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Łukasz Kaiser; Illia Polosukhin",
"journal": "Advances in neural information processing systems",
"ref_id": "b60",
"title": "Attention is all you need",
"year": "2017"
},
{
"authors": "Elena Voita; Rico Sennrich; Ivan Titov",
"journal": "Association for Computational Linguistics",
"ref_id": "b61",
"title": "The bottom-up evolution of representations in the transformer: A study with machine translation and language modeling objectives",
"year": "2019"
},
{
"authors": "Qiang Wang; Bei Li; Tong Xiao; Jingbo Zhu; Changliang Li; Derek F Wong; Lidia S Chao",
"journal": "Association for Computational Linguistics",
"ref_id": "b62",
"title": "Learning deep transformer models for machine translation",
"year": "2019"
},
{
"authors": "Qiang Wang; Fuxue Li; Tong Xiao; Yanyang Li; Yinqiao Li; Jingbo Zhu",
"journal": "Association for Computational Linguistics",
"ref_id": "b63",
"title": "Multi-layer representation fusion for neural machine translation",
"year": "2018"
},
{
"authors": "Weiwen Xu; Ai Ti Aw; Yang Ding; Kui Wu; Shafiq Joty",
"journal": "",
"ref_id": "b64",
"title": "Addressing the vulnerability of NMT in input perturbations",
"year": "2021"
},
{
"authors": "Jingfeng Yang; Le Zhang; Diyi Yang",
"journal": "Association for Computational Linguistics",
"ref_id": "b65",
"title": "SUBS: subtree substitution for compositional semantic parsing",
"year": "2022-07-10"
},
{
"authors": "Yuekun Yao; Alexander Koller",
"journal": "Association for Computational Linguistics",
"ref_id": "b66",
"title": "Structural generalization is hard for sequence-to-sequence models",
"year": "2022-12-07"
},
{
"authors": "Yongjing Yin; Yafu Li; Fandong Meng; Jie Zhou; Yue Zhang",
"journal": "International Committee on Computational Linguistics",
"ref_id": "b67",
"title": "Categorizing semantic representations for neural machine translation",
"year": "2022"
},
{
"authors": "Yongjing Yin; Jiali Zeng; Yafu Li; Fandong Meng; Jie Zhou; Yue Zhang",
"journal": "",
"ref_id": "b68",
"title": "Consistency regularization training for compositional generalization",
"year": "2023"
},
{
"authors": "Martin Daniel Zeman; Milan Popel; Jan Straka; Joakim Hajic; Filip Nivre; Juhani Ginter; Sampo Luotolahti; Slav Pyysalo; Martin Petrov; Potthast",
"journal": "Association for Computational Linguistics",
"ref_id": "b69",
"title": "Conll 2017 shared task: Multilingual parsing from raw text to universal dependencies",
"year": "2017-08-03"
}
] | [
{
"formula_coordinates": [
3,
340.04,
613.27,
180.86,
33.58
],
"formula_id": "formula_0",
"formula_text": "p(Y |X; θ) = T +1 t=1 p(y t |y <t , X; θ), (1"
},
{
"formula_coordinates": [
3,
520.9,
625.25,
4.24,
9.46
],
"formula_id": "formula_1",
"formula_text": ")"
},
{
"formula_coordinates": [
3,
333.26,
743.38,
191.88,
33.58
],
"formula_id": "formula_2",
"formula_text": "L CE (θ) = - T +1 t=1 log p(y t |y <t , X; θ). (2)"
},
{
"formula_coordinates": [
4,
359.13,
113.07,
167.09,
13.87
],
"formula_id": "formula_3",
"formula_text": "H SA 1 ∈ R d×S , H F F 1 ∈ R d×S respec-"
},
{
"formula_coordinates": [
4,
350.13,
178.01,
175.01,
14.19
],
"formula_id": "formula_4",
"formula_text": "H SA 1 = f Self -Attention (H 0 ),(3)"
},
{
"formula_coordinates": [
4,
345.85,
203.18,
179.29,
14.19
],
"formula_id": "formula_5",
"formula_text": "H F F 1 = f F eed-F orward (H SA 1 ),(4)"
},
{
"formula_coordinates": [
4,
345.27,
223.92,
179.87,
14.19
],
"formula_id": "formula_6",
"formula_text": "H SA i = f Self -Attention (H F F i-1 ),(5)"
},
{
"formula_coordinates": [
4,
345.41,
244.65,
179.73,
14.19
],
"formula_id": "formula_7",
"formula_text": "H F F i = f F eed-F orward (H SA i-1 ),(6)"
},
{
"formula_coordinates": [
4,
306.14,
294.66,
218.27,
25.78
],
"formula_id": "formula_8",
"formula_text": "H collect = {H SA 1 , H F F 1 , . . . , H SA M , H F F M }."
},
{
"formula_coordinates": [
4,
362.42,
356.24,
162.72,
33.71
],
"formula_id": "formula_9",
"formula_text": "H l key = 2M i=1 w i k H i collect ,(7)"
},
{
"formula_coordinates": [
4,
358.86,
403.21,
166.28,
33.71
],
"formula_id": "formula_10",
"formula_text": "H l value = 2M i=1 w i v H i collect ,(8)"
},
{
"formula_coordinates": [
4,
306.14,
457.7,
219.63,
29.36
],
"formula_id": "formula_11",
"formula_text": "w i k ̸ = w i v , w i k ̸ = w j k and w i v ̸ = w j v )"
}
] |
10.18653/v1/2022.naacl-main.223 | 2023-05-26 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b30",
"b28",
"b28",
"b33",
"b35"
],
"table_ref": [],
"text": "The NLP community has mainly focused on scaling Large Language Models (LLMs) vertically, i.e., deepening their understanding of high-resource languages by scaling up parameters and training data. While this approach has revolutionized NLP, the achievements are largely limited to high-resource languages. Examples of \"vertical\" LLMs are GPT3 (Brown et al., 2020), PaLM (Chowdhery et al., 2022) and Bloom (BigScience et al., 2022). In this paper, we create Glot500-m, a model that instead focuses on scaling multilingual LLMs horizontally, i.e., scaling to a large number of languages the great * Equal contribution. majority of which is low-resource. As LLMs are essential for progress in NLP, lack of LLMs supporting low-resource languages is a serious impediment to bringing NLP to all of the world's languages and cultures. Our goal is to address this need with the creation of Glot500-m. 1Existing multilingual LLMs support only about 100 (Conneau et al., 2020) out of the 7000 languages of the world. These supported languages are the ones for which large amounts of training data are available through projects such as Oscar (Suárez et al., 2019) and the Wikipedia dumps. 2 Following Siddhant et al. (2022), we refer to the 100 languages covered by XLM-R (Conneau et al., 2020) as head languages and to the remaining languages as tail languages. This terminology is motivated by the skewed distribution of available data per language: for the best-resourced languages there are huge corpora available, but for the long tail of languages, only small corpora exist. This is a key problem we address: the availability of data for tail languages is limited compared to head languages. As a result, tail languages have often been ignored by language technologies (Joshi et al., 2020).\nAlthough there exists some work on machine translation for a large number of tail languages (Costa-jussà et al., 2022;Bapna et al., 2022), existing LLMs for tail languages are limited to a relatively small number of languages (Wang et al., 2019;Alabi et al., 2022;Wang et al., 2022). In this paper, we address this gap. Our work has three parts. (i) Corpus collection. We collect Glot2000-c, a corpus covering thousands of tail languages. (ii) Model training. Using Glot500-c, a subset of Glot2000-c, we train Glot500-m, an LLM covering 511 languages. (iii) Validation. We conduct an extensive evaluation of the quality of Glot500-m's lows us to train Glot500-m, and will make as much of it publicly available as possible. (iii) We evaluate Glot500-m on pseudoperplexity and on five diverse tasks across these languages. We observe large improvements for low-resource languages compared to an XLM-R baseline. (iv) Our extensive analysis shows that no single factor explains the quality of multilingual LLM representations. Rather, a combination of factors determines quality including corpus size, script, \"help\" from related languages and the total capacity of the model. (v) Our work addresses an important goal of NLP research: we should not limit NLP to a relatively small number of high-resource languages and instead strive to support as many languages as possible to bring the benefits of NLP to all languages and cultures."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b39",
"b33",
"b35",
"b21",
"b20",
"b3",
"b28",
"b28",
"b32",
"b27",
"b34"
],
"table_ref": [],
"text": "Training multilingual LLMs using the masked language modeling (MLM) objective is effective to achieve cross-lingual representations (Devlin et al., 2019;Conneau et al., 2020). These models can be further improved by incorporating techniques such as discriminative pre-training (Chi et al., 2022) and the use of parallel data (Yang et al., 2020;Chi et al., 2021). However, this primarily benefits a limited set of languages with large corpora.\nRecent research has attempted to extend existing LLMs to languages with limited resources. Wang et al. (2019) propose vocabulary extension; Ebrahimi and Kann (2021) investigate adaptation methods, including MLM and Translation Language Model (TLM) objectives and adapters; Alabi et al. (2022) adapt XLM-R to 17 African languages; Wang et al. (2022) expand language models to low-resource languages using bilingual lexicons.\nAlternatively, parameter-efficient fine-tuning adapts pre-trained models to new languages by training a small set of weights effectively (Zhao et al., 2020;Pfeiffer et al., 2021;Ansell et al., 2022). Pfeiffer et al. (2022) address the \"curse of multilinguality\" by sharing a part of the model among all languages and having separate modules for each language. We show that the common perception that multilinguality increases as we add more languages, until, from some point, it starts decreasing, is naive. The amount of available data per language and the similarity between languages also play important roles ( §6.8).\nAnother approach trains LLMs from scratch for a limited number of tail languages; e.g., AfriBERTa (Ogueji et al., 2021a) and IndicNLPSuite (Kakwani et al., 2020) are LLMs for 11 African languages and 11 Indic languages. In concurrent work, Adebara et al. (2022) train a multilingual model for 517 African languages on a 42 GB corpus, but without making the model available and with an evaluation on a smaller number of languages than ours.\nClosely related to our work on corpus creation, Bapna et al. (2022) and Costa-jussà et al. (2022) also create NLP resources for a large number of tail languages. They train a language identifier model and extract textual data for tail languages from largescale web crawls. This approach is effective, but it requires significant computational resources and native speakers for all tail languages. This is hard to do outside of large corporations. Bapna et al. (2022) have not made their data available. Costajussà et al. (2022) have only released a portion of their data in around 200 languages.\nA key benefit of \"horizontally\" scaled multilingual LLMs is transfer from high-to low-resource languages. Our evaluation suggests that Glot500-m excels at this, but this is not the main focus of our paper. There is a large body of work on crosslingual transfer: (Artetxe and Schwenk, 2019;Imani-Googhari et al., 2022;Lauscher et al., 2020;Conneau et al., 2020;Turc et al., 2021;Fan et al., 2021;Severini et al., 2022;Choenni and Shutova, 2022;Wang et al., 2023), inter alia."
},
{
"figure_ref": [],
"heading": "Glot2000-c",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Data Collection",
"publication_ref": [],
"table_ref": [],
"text": "One of the major challenges in developing NLP technologies for tail languages is the scarcity of high-quality training data. In this work, we propose a lightweight methodology that is easily replicable for academic labs. We identify tail language data previously published by researchers, publishers and translators and then crawl or download them. By crawling a few websites and compiling data from around 150 different datasets, we amass more than 700GB of text in 2266 languages. We will refer to these sources of data as data sources. Our data covers many domains, including religious texts, news articles and scientific papers. Some of the data sources are high-quality, verified by native speakers, translators and linguists. Others are less reliable such as web crawls and Wikipedia dumps. It is therefore necessary to clean the data. For a list of data sources, see §C."
},
{
"figure_ref": [],
"heading": "Language-Scripts",
"publication_ref": [],
"table_ref": [],
"text": "Some languages are written in multiple scripts; e.g., Tajik is written in both Cyrillic and Arabic scripts. Some data sources indicate the script, but others either do not or provide mixed text in multiple scripts. We detect the script for each sentence and treat each language-script as a separate entity."
},
{
"figure_ref": [],
"heading": "Ngram LMs and Language Divergence",
"publication_ref": [],
"table_ref": [],
"text": "We train a 3-gram character-level language model 𝑀 𝑖 for each language-script 𝐿 𝑖 , using KenLM (Heafield, 2011). We refer to the perplexity calculated for the corpus of language 𝐿 𝑖 using language model 𝑀 𝑗 as PP (𝑀 𝑗 , 𝐿 𝑖 ). Similar to Gamallo et al. (2017), we define a perplexity-based divergence measure of languages 𝐿 𝑖 and 𝐿 𝑗 as: D 𝐿 𝑖 ,𝐿 𝑗 = max PP (𝑀 𝑗 , 𝐿 𝑖 ), PP (𝑀 𝑖 , 𝐿 𝑗 )\nWe use D to filter out noisy data in §3.4 and study the effect of similar languages in LLM training in §6.7 and §6.8. For more details, see §A."
},
{
"figure_ref": [],
"heading": "Data Cleaning",
"publication_ref": [],
"table_ref": [],
"text": "To remove noise, we use chunk-level and corpuslevel filters.\nWhile some sources are sentence-split, others provide multiple sentences (e.g., a paragraph) as one chunk. Chunk-level filters process each chunk of text from a data source as a unit, without sentencesplitting. Some chunk-level filters are based on the notion of word: we use white space tokenization when possible and otherwise resort to sentencePiece (Kudo and Richardson, 2018) trained by Costa-jussà et al. (2022).\nAs chunk-level filters, we employ the sentencelevel filters SF1-SF5 from BigScience ROOTS (Laurençon et al., 2022).\nSF1 Character repetition. If the ratio of repeated characters is too high, it is likely that the sentence has not enough textual content. SF2 Word repetition. A high ratio of repeated words indicates non-useful repetitive content.\nSF3 Special characters. Sentences with a high ratio of special characters are likely to be crawling artifacts or computer code.\nSF4 Insufficient number of words. Since training language models requires enough context, very small chunks of text are not useful. SF5 Deduplication. If two sentences are identical after eliminating punctuation and white space, one is removed. Table 1: Statistics for Glot2000-c, Glot500-c and existing multilingual datasets: number of languages, scripts, sentences' and median number of sentences' per language-script.\nIn the rest of the paper, we refer to a chunk as a sentence'. A sentence' can consist of a short segment, a complete sentence or a chunk (i.e., several sentences).\nCorpus-level filters detect if the corpus of a language-script is noisy; e.g., the corpus is in another language or consists of non-meaningful content such as tabular data. We employ filters CF1 and CF2.\nCF1 In case of mismatch between language and script, the corpus is removed; e.g., Chinese written in Arabic is unlikely to be Chinese.\nCF2 Perplexity mismatch. For each languagescript L1, we find its closest language-script L2: the language-script with the lowest perplexity divergence ( §3.3). If L1 and L2 are not in the same typological family, we check L1/L2 manually and take appropriate action such as removing the corpus (e.g., if it is actually English) or correcting the ISO code assigned to the corpus."
},
{
"figure_ref": [],
"heading": "Training Data: Glot500-c",
"publication_ref": [],
"table_ref": [],
"text": "Among the 2000+ language-scripts that we collected data for, after cleaning, most have too little data for pretraining LLMs. It is difficult to quantify the minimum amount needed for pretraining. Therefore, we pick a relatively high \"safe\" threshold, 30,000 sentences', for inclusion of language-scripts in model training. This allows us to train the model effectively and cover many low-resource languages. Table 1 gives Glot500-c statistics. See §B for a list of language-scripts. We train Glot500-m on Glot500-c; note that while Glot500-c focuses on tail languages, it contains some data in head languages which we include in Glot500-m training to prevent catastrophic forgetting.\nWe divide the corpus for each language into train/dev/test, reserving 1000 sentences' each for dev and test and using the rest for train. We pick 1000 parallel verses if we have a Bible translation 4 Glot500-m\nXLM-R-B XLM-R-L Glot500-m Model Size 278M 560M 395M Vocab Size 250K 250K 401K Transformer Size 86M 303M 86M"
},
{
"figure_ref": [],
"heading": "Vocabulary Extension",
"publication_ref": [],
"table_ref": [],
"text": "To extend XLM-R's vocabulary, we use Sentence-Piece (Kudo and Richardson, 2018) with a unigram language model (Kudo, 2018) to train a tokenizer with a vocabulary size of 250K on Glot500-c. We sample data from different language-scripts according to a multinomial distribution, with 𝛼=.3. The amount we sample for head languages is the same as tail languages with the lowest amount; this favors tail languages -head languages are already well learned by XLM-R. We merge the obtained tokens with XLM-R's vocabulary. About 100K new tokens were in fact old tokens, i.e., already part of XLM-R's vocabulary. We take the probabilities of the (genuinely) new tokens directly from Sen-tencePiece. After adding the 151K new tokens to XLM-R's vocabulary (which has size 250K), the vocabulary size of Glot500-m is 401K. We could also calculate probabilities of existing and new tokens over a mixture of original XLM-R training corpus and Glot500-c (Chung et al., 2020). For head languages, the percentage of changed tokens using the new tokenizer compared to the original tokenizer ranges from 0.2% to 50%. However, we found no relationship between percentage of changed tokens and change in performance on downstream tasks. Thus, there was little effect of tokenization in our experiments."
},
{
"figure_ref": [],
"heading": "Continued Pretraining",
"publication_ref": [],
"table_ref": [
"tab_0"
],
"text": "We create Glot500-m by continued pretraining of XLM-R-B with the MLM objective. The optimizer used is Adam with betas (0.9, 0.999). Initial learning rate: 5e-5. Each training step contains a batch of 384 training samples randomly picked from all language-scripts. The sampling strategy across language-scripts is the same as for vocabu- lary extension ( §4.1). We save checkpoints every 10K steps and select the checkpoint with the best average performance on downstream tasks by early stopping. Table 2 lists the sizes of XLM-R-B, XLM-R-L and Glot500-m. Except for a larger vocabulary ( §4.1), Glot500-m has the same size as XLM-R-B.\nWe train Glot500-m on a server with eight NVIDIA RTX A6000 GPUs for two weeks. Similar to XLM-R, we concatenate sentences' of a language-script and feed them as a stream to the tokenizer. The resulting output is then divided into chunks of 512 tokens and fed to the model."
},
{
"figure_ref": [],
"heading": "Experimental Setup",
"publication_ref": [
"b25",
"b25"
],
"table_ref": [],
"text": "For most tail languages, there are no manually labeled evaluation data. We therefore adopt a mixed evaluation strategy: based partly on human labels, partly on evaluation methods that are applicable to many languages without requiring gold data. Table 3 lists all our evaluation tasks.\nPerplexity Following Salazar et al. (2020), we calculate pseudoperplexity (PPPL) over the heldout test set. PPPL is based on masking tokens one-by-one (not left to right). Salazar et al. (2020) give evidence that PPPL is a better measure of linguistic acceptability compared to standard leftto-right perplexity."
},
{
"figure_ref": [],
"heading": "Roundtrip Alignment",
"publication_ref": [],
"table_ref": [],
"text": "For assessing the quality of multilingual representations for a broad range of tail languages without human gold data, we adopt roundtrip evaluation (Dufter et al., 2018). We first word-align sentences' in a parallel corpus based on the multilingual representations of an LLM. We then start from a word 𝑤 in a sentence' in language-script L1, follow the alignment links to its translations in language-script L2, then the alignment links from L2 to L3 and so on, until in the end we follow alignment links back to L1. If this \"roundtrip\" gets us back to 𝑤, then it indicates that the LLM has similar representations for the meaning of 𝑤 in language-scripts L1, L2, L3, etc. In other words, the cross-lingual quality of representations is high. Vice versa, failure to get back to 𝑤 is a sign of poor multilingual representations.\nWe use SimAlign (Jalili Sabet et al., 2020) and align on the sub-word level on the Bible part of test, based on the representations of the LLM computed by transformer layer 8 as suggested in the original paper. We use intersection symmetrization: each word in a sentence' is aligned to at most one word in the other sentence'.\nAs evaluation measure we compute the percentage of roundtrips that were successes, i.e., the roundtrip starts at 𝑤 in L1 and returns back to 𝑤. For each language-script in test, we randomly select three language-scripts as intermediate points L2, L3, L4. Since the intermediate points influence the results, we run the experiment five times with different intermediate points and report the average. All models are evaluated with the same five sets of three intermediate language-scripts."
},
{
"figure_ref": [],
"heading": "Sequence Labeling",
"publication_ref": [
"b19"
],
"table_ref": [],
"text": "We consider two sequence labeling tasks: Named Entity Recognition (NER) and Part-Of-Speech (POS) tagging. We use the WikiANN dataset (Pan et al., 2017) for NER and version v2.11 of Universal Dependencies (UD) (de Marneffe et al., 2021) for POS. Since training data does not exist for some languages, we finetune on English (with early stopping based on dev) and evaluate zero-shot transfer on all languages covered by WikiANN/UD. We set the learning rate to 2e-5 with Adam. Following (Hu et al., 2020), we use up to 1000 English-aligned sentences' from Tatoeba (Artetxe and Schwenk, 2019) to evaluate SentRetr (sentence retrieval). We also use 500 English-aligned sentences' from the Bible part of test. We find nearest neighbors using cosine similarity based on the average word embeddings in layer 𝑙 = 8 -following Jalili Sabet et al. ( 2020)and compute top10 accuracy. For fair comparison and because the architectures are the same, we do not optimize the hyperparameter 𝑙 for Glot500-m and XLM-R-B."
},
{
"figure_ref": [],
"heading": "Sentence Retrieval",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Text Classification",
"publication_ref": [
"b8"
],
"table_ref": [],
"text": "We evaluate on Taxi1500 (Ma et al., 2023). It provides gold data for text classification with six classes in a large number of language-scripts of which Glot500-m supports 354. We finetune on English (with early stopping on dev) and evaluate zero-shot on test of the target language-script. Learning rate: 2e-5, batch size:"
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we discuss aggregate results. For detailed results, see §D and §E."
},
{
"figure_ref": [],
"heading": "Results",
"publication_ref": [],
"table_ref": [],
"text": "Table 4 gives results. Glot500-m outperforms XLM-R-B on all tasks for both head and tail language-scripts, except for POS on head. That Glot500-m outperforms XLM-R-B is expected for tail language-scripts (i.e., those not covered by XLM-R). For these language-scripts the improvement margin is large. Outperformance may seem counterintuitive for head language-scripts (those covered by XLM-R) since Glot500-m has the same number of (non-embedding) parameters as XLM-R-B. Since the number of covered languages has greatly increased, leaving less capacity per language, we might expect underperformance. There are a few possible explanations. First, XLM-R may be undertrained, and the inclusion of more head language training data may improve their representations. Second, having more languages may improve multilinguality by allowing languages to synergize and enhance each other's representations and cross-lingual transfer. Third, there are languages similar to head languages among the tail languages, which in turn aids head languages.\nThe gap between Glot500-m and the baselines for tail language-scripts in sequence labeling is smaller. These tasks do not require as deep an understanding of language and thus transfer from head to tail language-scripts is easier through shared tokens.\nGlot500-m also outperforms XLM-R-L for tail language-scripts (all tasks) and head languagescripts (3 tasks). This suggests that scaling up size is not the only way for improvements. We can also improve the quality of multilingual LLM representations by increasing the number of languages."
},
{
"figure_ref": [],
"heading": "Language Coverage",
"publication_ref": [],
"table_ref": [],
"text": "Table 5 compares Glot500-m vs. XLM-R-B on pseudoperplexity. For fair comparison we use word-level normalization. For 69 head languagescripts, Glot500-m underperforms XLM-R-B. This is expected as Glot500-m's training data is small for these language-scripts. Glot500-m outperforms XLM-R-B for 420 tail language-scripts.\nThere are eight tail language-scripts for which Glot500-m performs worse than XLM-R-B. Five are tail languages with a similar head language where the two share a macro-language: ekk/Standard Estonian (est/Estonian), aln/Gheg Albanian (sqi/Albanian), nob/Norwegian Bokmal (nor/Norwegian), hbs/Serbo-Croatian (srp/Serbian), lvs/Standard Latvian (lav/Latvian). Since XLM-R-B's pretraining corpus is large for the five head languages, its performance is good for the close tail languages.\nThe other three languages all have a unique script: sat/Santali (Ol Chiki script), div/Dhivehi (Thaana script), iku/Inuktitut (Inuktitut syllabics). For these languages, XLM-R-B's tokenizer returns many UNK tokens since it is not trained on these scripts, resulting in an unreasonably optimistic estimate of pseudoperplexity by our implementation.\nGlot500-m's token-level normalized pseudoperplexity ranges from 1.95 for lhu/Lahu to 94.4 for tok/Toki Pona. The average is 13.5, the median 10.6. We analyze the five language-scripts with the highest pseudoperplexity: tok_Latn, luo_Latn, acm_Arab, ach_Latn, and teo_Latn.\ntok/Toki Pona is a constructed language. According to Wikipedia: \"Essentially identical concepts can be described by different words as the choice relies on the speaker's perception and experience.\" This property can result in higher variability and higher perplexity.\nacm/Mesopotamian Arabic contains a large number of tweets in raw form. This may result in difficult-to-predict tokens in test.\nluo/Luo, ach/Acoli and teo/Teso are related Nilotic languages spoken in Kenya, Tanzania, Uganda and South Sudan. Their high perplex- Table 4: Evaluation of XLM-R base and large (XLM-R-B and XLM-R-L) and Glot500-m on pseudoperplexity and six multilingual tasks across 5 seeds. Each number is an average over head, tail and all language-scripts. See §D, §E for results per task and language-script. Glot500-m outperforms XLM-R-B in all tasks for head (except for POS) and tail language-scripts and XLM-R-L for tail language-scripts. Best result per row/column group in bold.\ntail head all XLM-R-B XLM-R-L Glot500-m XLM-R-B XLM-R-L Glot500-m XLM-R-B XLM-R-L Glot500-\nhead tail Glot500-m is better 37 420 XLM-R-B is better 69 8\nTable 5: Pseudoperplexity Glot500-m vs XLM-R-B. Glot500-m's worse performance on head can be attributed to smaller training corpora and the relative difficulty of learning five times more languages with the same number of (non-embedding) parameters. Glot500-m performs better on almost all tail language-scripts. §6.2 discusses the eight exceptions.\nity could be related to the fact that they are tonal languages, but the tones are not orthographically indicated. Another possible explanation is that the training data is dominated by one subcorpus (Jehova's Witnesses) whereas the test data are dominated by PBC. There are orthographic differences between the two, e.g., \"dong\" (JW) vs. \"doŋ\" (PBC) for Acoli. These three languages are also spoken over a large area in countries with different standard languages, which could increase variability. Our analysis is not conclusive. We note however that the gap between the three languages and the next most difficult languages in terms of pseudoperplexity is not large. So maybe Luo, Acoli and Teso are simply (for reasons still to be determined) languages that have higher perplexity than others."
},
{
"figure_ref": [],
"heading": "Training Progression",
"publication_ref": [],
"table_ref": [],
"text": "To analyze the training process, we evaluate Glot500-m on sequence labeling and SentRetr at 10,000-step intervals. Figure 1 shows that performance improves rapidly at the onset of training, but then the rate of improvement slows down. This trend is particularly pronounced for tail languages in SentRetr. In comparison, sequence labeling is relatively straightforward, with the baseline (XLM-R-B, epoch 0) achieving high performance by correctly transferring prevalent classes such as verb and noun through shared vocabulary, resulting in a smaller improvement of Glot500-m vs. XLM-R-B.\nFor SentRetr, we observe larger improvements for the Bible than for Tatoeba. This is likely due to the higher proportion of religious data in Glot500-c, compared to XLM-R's training data (i.e., CC100).\nThe average performance on downstream tasks peaks at 480K steps. We have taken a snapshot of Glot500-m at this stage and released it."
},
{
"figure_ref": [],
"heading": "Analysis across Language-Scripts",
"publication_ref": [],
"table_ref": [],
"text": "To analyze the effect of language-scripts, we select five tail language-scripts each with the largest and smallest gain when comparing Glot500-m vs. XLM-R-B for SentRetr and sequence labeling.\nTable 6 shows that Glot500-m improves languages with scripts not covered by XLM-R (e.g., div/Dhivehi, Thaana script, see §6.2) by a large margin since XLM-R simply regards the uncovered scripts as unknown tokens and cannot compute meaningful representations for the input. The large amount of data we collected in Glot500-c also contributes to the improvement for tail languages, e.g., for tat_Cyrl (Tatar) in SentRetr Tatoeba and mlt_Latn (Maltese) in POS. See §6.7 for a detailed analysis of the effect of corpus size.\nOn the other hand, Glot500-m achieves just comparable or even worse results for some languagescripts. We see at least three explanations. (i) As discussed in §6.2, some tail languages (e.g., nob/Norwegian Bokmal) are close to a head language (e.g., nor/Norwegian), so Glot500-m has no advantage over XLM-R-B. (ii) A language is at the low end of our corpus size range (i.e., 30,000 sentences'). Example: xav_Latn, Xavánte. (iii) Some languages are completely distinct from all other languages in Glot500-c, thus without support from any similar language. An example is mau_Latn, Huautla Mazatec. Glot500-m has a much harder Table 7: Sentence Retrieval Bible performance of Glot500-m and XLM-R-B for six languages with two scripts: Uighur (uig), Hindi (hin), Uzbek (uzb), Kara-Kalpak (kaa), Northern Kurdish (kmr), Turkmen (tuk). Glot500-m clearly outperforms XLM-R-B with large differences for tail language-scripts.\ntime learning good representations in these cases."
},
{
"figure_ref": [],
"heading": "Languages with Multiple Scripts",
"publication_ref": [],
"table_ref": [],
"text": "Table 7 compares SentRetr performance XLM-R-B vs. Glot500-m for six languages with two scripts. Unsurprisingly, XLM-R performs much better for a language-script it was pretrained on (\"head\") than on one that it was not (\"tail\"). We can improve the performance of a language, even surpassing the language-script covered by XLM-R, if we collect enough data for its script not covered by XLM-R.\nFor languages with two scripts not covered by XLM-R, the performance is better for the script for which we collect a larger corpus. For example, kaa_Cyrl (Kara-Kalpak) has about three times as much data as kaa_Latn. This explains why kaa_Cyrl outperforms kaa_Latn by 30%.\nDufter and Schütze (2020) found that, after training a multilingual model with two scripts for English (natural English and \"fake English\"), the model performed well at zero-shot transfer if the capacity of the model was of the right size (i.e., not too small, not too large). Our experiments with real data show the complexity of the issue: even if there is a \"right\" size for an LLM that supports both full acquisition of languages and multilingual transfer, this size is difficult to determine and it may be different for different language pairs in a large horizontally scaled model like Glot500-m."
},
{
"figure_ref": [],
"heading": "Analysis across Language Families",
"publication_ref": [],
"table_ref": [],
"text": "Table 8 compares SentRetr performance Glot500-m vs. XLM-R-B for seven language families that have ten or more language-scripts in Glot500-c. We assign languages to families based on Glottolog. 4Generally, XLM-R has better performance the more language-scripts from a language family are represented in its training data; e.g., performance is better for indo1319 and worse for maya1287. The results suggest that Glot500-m's improvement over Table 9: Performance on Sentence Retrieval Bible of continued pretraining on just one language-script (Glot+1) vs. on Glot500-c (Glot500-m). Glot500-m underperforms on the top three and outperforms on the bottom three. Our explanation is that the second group is supported by closely related languages in Glot500-c; e.g., for Southern Quechua (quh), Glot500-m also covers closely related Cuzco Quechua (quz). For the first group this is not the case; e.g., the Wa language (wbm) has no close relative in Glot500-c.\nXLM-R is the larger, the better our training corpus Glot500-c's coverage is of a family."
},
{
"figure_ref": [],
"heading": "Effect of Amount of Training Data",
"publication_ref": [],
"table_ref": [],
"text": "We examine correlation between pretraining corpus size and Glot500-m zero-shot performance. We focus on SentRetr Bible ( §5) since it supports the most head and tail languages. We find that Pearson's 𝑟 = .34, i.e., corpus size and performance are moderately, but clearly correlated. We suspect that the correlation is not larger because, in addition to corpus size of language 𝑙 itself, corpus size of languages closely related to 𝑙 is also an important factor (see §6.4 for a similar finding for Norwegian). We therefore also compute Pearson's 𝑟 between (i) performance of language 𝑙 on SentRetr Bible and (ii) joint corpus size of 𝑙 and its 𝑘 nearest neighbors (according to perplexity divergence, §3.3). In this case, Pearson's 𝑟 = .44 (for both 𝑘 = 3 and 𝑘 = 4), indicating that the corpus size of nearest neighbor languages does play a role."
},
{
"figure_ref": [],
"heading": "Support through Related Languages",
"publication_ref": [],
"table_ref": [],
"text": "Building on §6.7, there is another way we can investigate the positive effect of closely related languages on performance: We can compare performance (again on SentRetr Bible) of continued pretraining on just one language (we refer to this model as Glot+1) vs. on all 511 languages represented in Glot500-c (i.e., Glot500-m). Table 9 presents results for six language-scripts selected from various language families and suggests that some languages do not receive support from related languages (top three). In that case, Glot+1 can fully concentrate on learning the isolated language and does better than Glot500-c. Other languages (bottom three) do receive support from related languages. For example, Southern Quechua (quh) seems to receive support in Glot500-m from closely related Cuzco Quechua (quz), resulting in Glot500-m outperforming Glot+1."
},
{
"figure_ref": [],
"heading": "Conclusion and Future Work",
"publication_ref": [],
"table_ref": [],
"text": "We collect and data-clean Glot500-c, a large corpus of hundreds of usually neglected tail (i.e., long-tail) languages and create Glot500-m, an LLM that is trained on Glot500-c and covers these languages.\nWe evaluate Glot500-m on six tasks that allow us to evaluate almost all languages. We observe large improvements for both head and tail languages compared to XLM-R. Our analysis shows that no single factor fully explains the quality of the representation of a language in a multilingual model. Rather, a combination of factors is important, including corpus size, script, \"help\" from related languages and the total capacity of the model. This work is the first to create a language model on a dataset of several hundreds of gigabytes and to make it publicly available for such a large and diverse number of low-resource languages. In future research, we would like to train larger models to further investigate the effect of model size, distill highly multilingual models for resource-efficient deployment, explore alternatives to continued pretraining and use models for more tail language downstream tasks."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [],
"table_ref": [],
"text": "(1) We did not perform any comprehensive hyperparameter search, which would have further consolidated our results. This decision was made due to the high cost of training multiple models. (2) Compared to current very large models, Glot500-m is comparatively small. (3) Although we have tried to minimize the amount of noise in our data, some noise is still present."
},
{
"figure_ref": [],
"heading": "Ethics Statement",
"publication_ref": [],
"table_ref": [
"tab_8"
],
"text": "There are two issues worth mentioning in regards to this project. First, it was not feasible for us to thoroughly examine the content of the data for all languages, thus we cannot confirm the absence of discrimination based on factors such as race or sexuality. The data was solely utilized as a textual corpus, and the content should not be interpreted as an endorsement by our team. If the model is subsequently utilized for generation, it is possible that the training data may be reflected in the generated output. However, addressing potential biases within the data is an area for future research. Second, it is important to note that while the data sources utilized in this study do not explicitly prohibit the reuse of data for research purposes, some sources do have copyright statements indicating that such use is permissible while others do not. Additionally, certain sources prohibit the redistribution of data. As such, data from these sources is omitted from the published version of Glot2000-c. \nPP (𝑆, 𝑀) = 𝑇 𝑇 𝑡=1 1 P 𝑐ℎ 𝑡 | 𝑐ℎ 𝑡 -1 1 (1)\nwhere P 𝑐ℎ 𝑡 | 𝑐ℎ 𝑡 -1 1 is computed as by dividing the observed frequency (𝐶) of 𝑐ℎ 𝑡 -1 1 𝑐ℎ 𝑖 by the observed frequency of 𝑐ℎ 𝑡 -1 1 in 𝑀 training data:\nP 𝑐ℎ 𝑡 | 𝑐ℎ 𝑡 -1 1 = 𝐶 𝑐ℎ 𝑡 -1 1 𝑐ℎ 𝑡 𝐶 𝑐ℎ 𝑡 -1 1 (2)\nGiven the definition of perplexity, we can determine how well a trained language model on language 𝐿 1 predicts the test text of language 𝐿 2 and vice-versa. The divergence between two languages is computed with the maximum of the perplexity values in both directions. Two reasons lead to the use of max: first, a symmetrical divergence is required, and second, languages differ in their complexity, so one direction of computing perplexity may result in a much lower perplexity than another. Thus, comparing perplexity results becomes difficult. As an example, the Kuanua language (ksd_Latn) has short words and a simple structure, which results in 3-gram models getting lower perplexity on its text compared to other languages. The lower the perplexity the smaller the divergence between languages. The divergence (D) between language 𝐿 𝑖 and 𝐿 𝑗 with trained language models of 𝑀 𝐿 𝑧 and test texts of 𝑆 𝐿 𝑧 , where 𝐿 𝑧 is the corresponding language, computed as follows:\nD 𝐿 𝑖 ,𝐿 𝑗 = max PP (𝑆 𝐿 𝑖 , 𝑀 𝐿 𝑗 ), PP (𝑆 𝐿 𝑗 , 𝑀 𝐿 𝑖 ) (3)\nRuns and Data. The data used to train and test the character level n-gram models is the same data used for the training and testing of the Glot500-m. The training of the models was limited to 100, 000 sentences' per language-script. We use KenLM library (Heafield, 2011) to build n-gram models. This library uses an interpolated modified Kneser-Ney smoothing for estimating the unseen n-grams. Our evaluation has been performed over 7 n-gram models (3 ≤ 𝑛 ≤ 9). Baseline and Evaluation. Language family trees were used as a baseline for evaluating the divergence measures of the proposed approach. We obtained language family tree data from Ethnologue online version (Eberhard et al., 2022). For each language, the family tree follows the general order from largest typological language family group to smallest. There is only one family tree for each language in the baseline data. Nodes in the family tree represent typological language family groups.\nEach node only has one parent, so if a node is common in the family tree of two languages, its parent is also common. We evaluate our perplexity method on the following binary classification task: Do the majority of a language 𝐿 𝑧 's 𝑘 nearest neighbors belong to the same typological language family group as 𝐿 𝑧 ? Assuming languages 𝐿 𝑖 and 𝐿 𝑗 , with the following family trees:\n𝑇 𝐿 𝑖 : 1 → 2 → 3 → 4 → 5 → 6 𝑇 𝐿 𝑗 : 1 → 2 → 7 → 8\nThese 2 languages belong to the same typological family group with family tree levels of 𝑙 ∈ {1, 2}, but not with family tree levels of 𝑙 = 3 and higher. Result. When it comes to language families, the majority of studies only refer to the largest typological language family group (level 𝑙 = 1). Here, we also assess our methodology for other levels. The results of classification accuracy for 3-gram model, 𝑘 ∈ {1, 3, 7, 13, 21} and 𝑙 ∈ {1, 2, 3, max} are shown in Table 10. In cases where the maximum level of a tree is less than the 𝑙 parameter, the maximum level for that language is used. Languages without a family or no other family member in our data are excluded. We only report the 3-gram model results as it gets the best results in most configurations among other n-gram models. With increasing 𝑙, the accuracy decreases, since more languages fall outside the same typological family. As 𝑘 increases, the accuracy decreases, because languages with faraway neighbors are being included but the number of languages in the language typological group family will remain the same. There are times when languages have a lot of loan words from other languages because of geological proximity or historical reasons (e.g, colonization), which makes them similar to the languages they borrowed words from in our method. However they are different when it comes to their typological families and our method fails in these cases. Aymara (Macrolanguage: aym_Latn) and Quechua (Macrolanguage: que_Latn), for example, had a great deal of contact and influence on each other, but they do not belong to the same typological group. As well, some of the typological families are not that large, which makes our results worse when 𝑘 increases. This is the case, for instance, of the Tarascan typological family which only has two members. "
},
{
"figure_ref": [],
"heading": "B Languages",
"publication_ref": [],
"table_ref": [],
"text": "The list of languages used to train Glot500-m with the amount of available data for each language is available in Tables 11, 12 and 13."
},
{
"figure_ref": [],
"heading": "On Macrolanguages",
"publication_ref": [
"b4"
],
"table_ref": [],
"text": "The presence of language codes that are supersets of other language codes within datasets is not uncommon (Kreutzer et al., 2022). This issue becomes more prevalent in extensive collections. Within the ISO 639-3 standard, these languages are referred to as macrolanguages.\nWhen confronted with macrolanguages, if it is not feasible to ascertain the specific individual language contained within a dataset, the macrolanguage code is retained. Consequently, it is possible that in Glot2000-c and Glot500-c both the corpora for the macrolanguage and its individual languages have been included."
},
{
"figure_ref": [],
"heading": "C List of data sources",
"publication_ref": [
"b4",
"b38",
"b24"
],
"table_ref": [],
"text": "The datasets and repositories used in this project involve: AI4Bharat, 5 AIFORTHAI-LotusCorpus, 6 Add (El-Haj et al., 2018), AfriBERTa (Ogueji et al., 2021b), AfroMAFT (Adelani et al., 2022;Xue et al., 2021), Anuvaad, 7 AraBench (Sajjad et al., 2020) "
},
{
"figure_ref": [],
"heading": "D Results for Each Task and Language",
"publication_ref": [],
"table_ref": [],
"text": "We report the detailed results for all tasks and languages in "
},
{
"figure_ref": [],
"heading": "E Perplexity Results for all Languages",
"publication_ref": [],
"table_ref": [
"tab_25",
"tab_26",
"tab_27",
"tab_13",
"tab_0"
],
"text": "Perplexity number for all languages is presented in Table 23, Table 24, and Table 25. Table 14: Top10 accuracy of XLM-R-B, XLM-R-L, and Glot500-m on Sentence Retrieval Tatoeba.\nLanguage-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-\nLanguage Table 15: Top10 accuracy of XLM-R-B, XLM-R-L, and Glot500-m on Sentence Retrieval Bible (Part I). Table 16: Top10 accuracy of XLM-R-B, XLM-R-L, and Glot500-m on Sentence Retrieval Bible (Part II). Table 17: F1 of XLM-R-B, XLM-R-L, and Glot500-m on NER. Table 18: F1 of XLM-R-B, XLM-R-L, and Glot500-m on POS.\n-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-\nLanguage-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-\nLanguage-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-\nLanguage-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-\nLanguage Language Table 20: F1 of XLM-R-B, XLM-R-L, and Glot500-m on Text Classification (Part II).\n-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-\n-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-\nLanguage Language Language \n-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-\n-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-\n-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-"
},
{
"figure_ref": [],
"heading": "Acknowledgements",
"publication_ref": [],
"table_ref": [],
"text": "We would like to thank Renhao Pei, Yihong Liu, Verena Blaschke, and the anonymous reviewers. This work was funded by the European Research Council (grants #740516 and #758969) and EU's Horizon Europe Research and Innovation Actions (UTTER, contract 101070631). Marta Bañón, Pinzhen Chen, Barry Haddow, Kenneth Heafield, Hieu Hoang, Miquel Esplà-Gomis, Mikel L. Forcada, Amir Kamran, Faheem Kirefu, Philipp Koehn, Sergio Ortiz Rojas, Leopoldo Pla Sempere, Gema Ramírez-Sánchez, Elsa Sarrías, Marek Strelec, Brian Thompson, William Waites, Dion Wiggins, and Jaume Zaragoza. 2020. ParaCrawl: Web-scale acquisition of parallel corpora. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4555-4567, Online. Association for Computational Linguistics. Marta Bañón, Miquel Esplà-Gomis, Mikel L. Forcada, Cristian García-Romero, Taja Kuzman, Nikola Ljubesic, Rik van Noord, Leopoldo Pla Sempere, Gema Ramírez-Sánchez, Peter Rupnik, Vít Suchomel, Antonio Toral, Tobias van der Werff, and Jaume Zaragoza. 2022. Macocu: Massive collection and curation of monolingual and bilingual data: focus on under-resourced languages. In Proceedings of the 23rd Annual Conference of the European Association for Machine Translation, EAMT 2022, Ghent, Belgium, June 1-3, 2022, pages 301-302. European Association for Machine Translation. Ankur Bapna, Isaac Caswell, Julia Kreutzer, Orhan Firat, Daan van Esch, Aditya Siddhant, Mengmeng Niu, Pallavi Baljekar, Xavier Garcia, Wolfgang Macherey, et al. 2022. Building machine translation systems for the next thousand languages. arXiv preprint arXiv:2205.03983. Workshop BigScience, :, Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilić, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, Matthias Gallé, Jonathan Tow, Alexander M. Rush, Stella Biderman, Albert Webson, Pawan Sasanka Ammanamanchi, Thomas Wang, Benoît Sagot, Niklas Muennighoff, Albert Villanova del Moral, Olatunji Ruwase, Rachel Bawden, Stas Bekman, Angelina McMillan-Major, Iz Beltagy, Huu Nguyen, Lucile Saulnier, Samson Tan, Pedro Ortiz Suarez, Victor Sanh, Hugo Laurençon, Yacine Jernite, Julien Launay, Margaret Mitchell, Colin Raffel, Aaron Gokaslan, Adi Simhi, Aitor Soroa, Alham Fikri Aji, Amit Alfassy, Anna Rogers, Ariel Kreisberg Nitzav, Canwen Xu, Chenghao Mou, Chris Emezue, Christopher Klamm, Colin Leong, Daniel van Strien, David Ifeoluwa Adelani, Dragomir Radev, Eduardo González Ponferrada, Efrat Levkovizh, Ethan Kim, Eyal Bar Natan, Francesco De Toni, Gérard Dupont, Germán Kruszewski, Giada Pistilli, Hady Elsahar, Hamza Benyamina, Hieu Tran, Ian Yu, Idris Abdulmumin, Isaac Johnson, Itziar Gonzalez-Dios, Javier de la Rosa, Jenny Chim, Jesse Dodge, Jian Zhu, Jonathan Chang, Jörg Frohberg, Joseph Tobing, Joydeep Bhattacharjee, Khalid Almubarak, Kimbo Chen, Kyle Lo, Leandro Von Werra, Leon Weber, Long Phan, Loubna Ben allal, Ludovic Tanguy, Manan Dey, Manuel Romero Muñoz, Maraim Masoud, María Grandury, Mario Šaško, Max Huang, Maximin Coavoux, Mayank Singh, Mike Tian-Jian Jiang, Minh Chien Vu, Mohammad A. Jauhar, Mustafa Ghaleb, Nishant Subramani, Nora Kassner, Nurulaqilla Khamis, Olivier Nguyen, Omar Espejel, Ona de Gibert, Paulo Villegas, Peter Henderson, Pierre Colombo, Priscilla Amuok, Quentin Lhoest, Rheza Harliman, Rishi Bommasani, Roberto Luis López, Rui Ribeiro, Salomey Osei, Sampo Pyysalo, Sebastian Nagel, Shamik Bose, Shamsuddeen Hassan Muhammad, Shanya Sharma, Shayne Longpre, Somaieh Nikpoor, Stanislav Silberberg, Suhas Pai, Sydney Zink, Tiago Timponi Torrent, Timo Schick, Tristan Thrush, Valentin Danchev, Vassilina Nikoulina, Veronika Laippala, Violette Lepercq, Vrinda Prabhu, Zaid Alyafeai, Zeerak Talat, Arun Raja, Benjamin Heinzerling, Chenglei Si, Davut Emre Taşar, Elizabeth Salesky, Sabrina J. Mielke, Wilson Y. Lee, Abheesht Sharma, Andrea Santilli, Antoine Chaffin, Arnaud Stiegler, Debajyoti Datta, Eliza Szczechla, Gunjan Chhablani, Han Wang, Harshit Pandey, Hendrik Strobelt, Jason Alan Fries, Jos Rozen, Leo Gao, Lintang Sutawika, M Saiful Bari, Maged S. Al-shaibani, Matteo Manica, Nihal Nayak, Ryan Teehan, Samuel Albanie, Sheng Shen, Srulik Ben-David, Stephen H. Bach, Taewoon Kim, Tali Bers, Thibault Fevry, Trishala Neeraj, Urmish Thakker, Vikas Raunak, Xiangru Tang, Zheng-Xin Yong, Zhiqing Sun, Shaked Brody, Yallow Uri, Hadar Tojarieh, Adam Roberts, Hyung Won Chung, Jaesung Tae, Jason Phang, Ofir Press, Conglong Li, Deepak Narayanan, Hatim Bourfoune, Jared Casper, Jeff Rasley, Max Ryabinin, Mayank Mishra, Minjia Zhang, Mohammad Shoeybi, Myriam Peyrounette, Nicolas Patry, Nouamane Tazi, Omar Sanseviero, Patrick von"
}
] | The NLP community has mainly focused on scaling Large Language Models (LLMs) vertically, i.e., making them better for about 100 languages. We instead scale LLMs horizontally: we create, through continued pretraining, Glot500-m, an LLM that covers 511 predominantly low-resource languages. An important part of this effort is to collect and clean Glot500-c, a corpus that covers these 511 languages and allows us to train Glot500-m. We evaluate Glot500-m on five diverse tasks across these languages. We observe large improvements for both high-resource and low-resource languages compared to an XLM-R baseline. Our analysis shows that no single factor explains the quality of multilingual LLM representations. Rather, a combination of factors determines quality including corpus size, script, "help" from related languages and the total capacity of the model. Our work addresses an important goal of NLP research: we should not limit NLP to a small fraction of the world's languages and instead strive to support as many languages as possible to bring the benefits of NLP technology to all languages and cultures. Code, data and models are available at https://github.com/cisnlp/Glot500. | Glot500: Scaling Multilingual Corpora and Language Models to 500 Languages | [
{
"figure_caption": "Model sizes. Glot500-m and XLM-R-B have the same transformer size, but Glot500-m has a larger vocabulary, resulting in an overall larger model.",
"figure_data": "and add 500 each to test and dev. These parallelverses convey identical meanings and facilitatecrosslingual evaluation. We pretrain the modelusing only the training data.",
"figure_id": "tab_0",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Evaluation tasks and measures. |head|/|tail|: number of head/tail language-scripts",
"figure_data": "",
"figure_id": "tab_2",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Israa Alsarsour, Esraa Mohamed, Reem Suwaileh, and Tamer Elsayed. 2018. DART: A large dataset of dialectal Arabic tweets. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA). In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8440-8451, Online. Association for Computational Linguistics.",
"figure_data": "Kenneth Heafield. 2011. KenLM: Faster and smallerMahmoud El-Haj, Paul Rayson, and Mariam Aboelezz. Muhammad, Nanda Muhammad, Ayanda Mnyakeni,language model queries. In Proceedings of the Sixth2018. Arabic dialect identification in the context of Jamshidbek Mirzakhalov, Tapiwanashe Matangira,Workshop on Statistical Machine Translation, pagesbivalency and code-switching. In Proceedings of Colin Leong, Nze Lawson, Sneha Kudugunta, Yacine187-197, Edinburgh, Scotland. Association for Com-the Eleventh International Conference on Language Jernite, Mathias Jenny, Orhan Firat, Bonaventureputational Linguistics.Resources and Evaluation (LREC 2018), Miyazaki, F. P. Dossou, Sakhile Dlamini, Nisansa de Silva,Antonios Anastasopoulos, Alessandro Cattelan, Zi-Yi Dou, Marcello Federico, Christian Federmann, Dmitriy Genzel, Franscisco Guzmán, Junjie Hu, Mac-duff Hughes, Philipp Koehn, Rosie Lazar, Will Lewis, Graham Neubig, Mengmeng Niu, Alp Öktem, Eric Paquin, Grace Tang, and Sylwia Tur. 2020. TICO-19: the translation initiative for COvid-19. In Proceed-ings of the 1st Workshop on NLP for COVID-19 (Part 2) at EMNLP 2020, Online. Association for Computational Linguistics. Marta R Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, et al. 2022. No language left behind: Scaling human-centered machine translation. arXiv preprint arXiv:2207.04672. Marie-Catherine de Marneffe, Christopher D. Manning, Junjie Hu, Sebastian Ruder, Aditya Siddhant, Gra-ham Neubig, Orhan Firat, and Melvin Johnson. 2020. XTREME: A massively multilingual multi-task benchmark for evaluating cross-lingual generalisation. In Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 4411-4421. PMLR. Joakim Nivre, and Daniel Zeman. 2021. Universal dependencies. Computational Linguistics, 47(2):255-308. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language under-Ayyoob ImaniGooghari, Silvia Severini, Masoud Jalili Sabet, François Yvon, and Hinrich Schütze. 2022. Graph-based multilingual label propagation for low-resource part-of-speech tagging. In Proceed-ings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 1577-1589, Abu Dhabi, United Arab Emirates. Association forTom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901. José Camacho-Collados, Claudio Delli Bovi, Alessandro Raganato, and Roberto Navigli. 2016. A large-scale multilingual disambiguation of glosses. In Proceed-ings of the Tenth International Conference on Lan-guage Resources and Evaluation (LREC'16), pages 1701-1708, Portorož, Slovenia. European Language Resources Association (ELRA). Japan. European Language Resources Association (ELRA). Angela Fan, Shruti Bhosale, Holger Schwenk, Zhiyi Ma, Ahmed El-Kishky, Siddharth Goyal, Mandeep Baines, Onur Celebi, Guillaume Wenzek, Vishrav Chaudhary, Naman Goyal, Tom Birch, Vitaliy Liptchinsky, Sergey Edunov, Michael Auli, and Armand Joulin. 2021. Be-yond english-centric multilingual machine translation. Sakine Çabuk Ballı, Stella Biderman, Alessia Bat-tisti, Ahmed Baruwa, Ankur Bapna, Pallavi Baljekar, Israel Abebe Azime, Ayodele Awokoya, Duygu Ata-man, Orevaoghene Ahia, Oghenefego Ahia, Sweta Agrawal, and Mofetoluwa Adeyemi. 2022. Quality at a glance: An audit of web-crawled multilingual datasets. Transactions of the Association for Compu-tational Linguistics, 10:50-72. J. Mach. Learn. Res., 22:107:1-107:48. Pablo Gamallo, Jose Ramom Pichel, and Iñaki Alegria. 2017. A perplexity-based method for similar lan-guages discrimination. In Proceedings of the Fourth Workshop on NLP for Similar Languages, Varieties and Dialects (VarDial), pages 109-114, Valencia, Spain. Association for Computational Linguistics. Taku Kudo. 2018. Subword regularization: Improv-ing neural network translation models with multiple subword candidates. In Proceedings of the 56th An-nual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 66-75, Melbourne, Australia. Association for Computational Linguistics.Computational Linguistics. Masoud Jalili Sabet, Philipp Dufter, François Yvon, and Hinrich Schütze. 2020. SimAlign: High quality word alignments without parallel training data using static and contextualized embeddings. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1627-1643, Online. Association for Computational Linguistics.Dirk Goldhahn, Thomas Eckart, and Uwe Quasthoff. 2012. Building large monolingual dictionaries at the leipzig corpora collection: From 100 to 200 languages. In Proceedings of the Eighth International Conference on Language Resources and Evaluation, Taku Kudo and John Richardson. 2018. SentencePiece: A simple and language independent subword tok-enizer and detokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System LREC 2012, Istanbul, Turkey, May 23-25, 2012, pages 759-765. European Language Resources Association (ELRA). Demonstrations, pages 66-71, Brussels, Belgium. Association for Computational Linguistics.Pratik Joshi, Sebastin Santy, Amar Budhiraja, Kalika Bali, and Monojit Choudhury. 2020. The state and fate of linguistic diversity and inclusion in the NLP world. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6282-6293, Online. Association for ComputationalSantiago Góngora, Nicolás Giossa, and Luis Chiruzzo. 2021. Experiments on a Guarani corpus of news and social media. In Proceedings of the First Work-shop on Natural Language Processing for Indigenous Languages of the Americas, pages 153-158, Online. Association for Computational Linguistics. Anoop Kunchukuttan, Pratik Mehta, and Pushpak Bhat-tacharyya. 2018. The IIT Bombay English-Hindi parallel corpus. In Proceedings of the Eleventh In-ternational Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA).Linguistics.Santiago Góngora, Nicolás Giossa, and Luis Chiruzzo. Hugo Laurençon, Lucile Saulnier, Thomas Wang,Divyanshu Kakwani, Anoop Kunchukuttan, Satish Golla,2022. Can we use word embeddings for enhancing Christopher Akiki, Albert Villanova del Moral, TevenGokul N.C., Avik Bhattacharyya, Mitesh M. Khapra,Guarani-Spanish machine translation? In Proceed-Le Scao, Leandro Von Werra, Chenghao Mou, Ed-and Pratyush Kumar. 2020. IndicNLPSuite: Monolin-ings of the Fifth Workshop on the Use of Computa-uardo González Ponferrada, Huu Nguyen, et al. 2022.gual corpora, evaluation benchmarks and pre-trainedtional Methods in the Study of Endangered Languages, The BigScience ROOTS Corpus: A 1.6 TB Compos-multilingual language models for Indian languages.pages 127-132, Dublin, Ireland. Association for Com-ite Multilingual Dataset. In Thirty-sixth ConferenceIn Findings of the Association for Computationalputational Linguistics. on Neural Information Processing Systems DatasetsAbteen Ebrahimi and Katharina Kann. 2021. How to adapt your pretrained multilingual model to 1600 languages. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Linguistics: EMNLP 2020, pages 4948-4961, Online. Association for Computational Linguistics. Fajri Koto and Ikhwan Koto. 2020. Towards computa-tional linguistics in Minangkabau language: Studies on sentiment analysis and machine translation. In Proceedings of the 34th Pacific Asia Conference on Language, Information and Computation, pages 138-148, Hanoi, Vietnam. Association for Computational Linguistics.Hyung Won Chung, Dan Garrette, Kiat Chuan Tan, and Jason Riesa. 2020. Improving multilingual models with language-clustered vocabularies. In Proceed-and Benchmarks Track. Thamme Gowda, Zhao Zhang, Chris Mattmann, and Jonathan May. 2021. Many-to-English machine trans-lation tools, data, and pretrained models. In Proceed-ings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing: System Demonstrations, pages 306-316, Online. As-sociation for Computational Linguistics. Anne Lauscher, Vinit Ravishankar, Ivan Vulić, and Goran Glavaš. 2020. From zero to hero: On the limita-tions of zero-shot language transfer with multilingual Mengjie Zhao, Tao Lin, Fei Mi, Martin Jaggi, and Transformers. In Proceedings of the 2020 Conference Hinrich Schütze. 2020. Masking as an efficient alter-on Empirical Methods in Natural Language Process-native to finetuning for pretrained language models. ing (EMNLP), pages 4483-4499, Online. Association In Proceedings of the 2020 Conference on Empirical for Computational Linguistics. Methods in Natural Language Processing (EMNLP), Tahmid Hasan, Abhik Bhattacharjee, Md. Saiful Is-pages 2226-2241, Online. Association for Computa-Language Processing (Volume 1: Long Papers), pages Julia Kreutzer, Isaac Caswell, Lisa Wang, Ahsan Wahab,ings of the 2020 Conference on Empirical Methods lam, Kazi Samin Mubasshir, Yuan-Fang Li, Yong-Colin Leong, Joshua Nemecek, Jacob Mansdorfer, Anna tional Linguistics.4555-4567, Online. Association for Computational Daan van Esch, Nasanbayar Ulzii-Orshikh, Allah-in Natural Language Processing (EMNLP), pages Bin Kang, M. Sohel Rahman, and Rifat Shahriyar. Filighera, Abraham Owodunni, and Daniel White-Linguistics. sera Tapo, Nishant Subramani, Artem Sokolov, Clay-4536-4546, Online. Association for Computational 2021. Xl-sum: Large-scale multilingual abstrac-nack. 2022. Bloom library: Multimodal datasets intone Sikasote, Monang Setyawan, SupheakmungkolLinguistics. tive summarization for 44 languages. In Findings 300+ languages for a variety of downstream tasks. InSarin, Sokhar Samb, Benoît Sagot, Clara Rivera,of the Association for Computational Linguistics: Proceedings of the 2022 Conference on EmpiricalAnnette Rios, Isabel Papadimitriou, Salomey Osei,ACL/IJCNLP 2021, Online Event, August 1-6, 2021, Methods in Natural Language Processing, EMNLPPedro Ortiz Suarez, Iroro Orife, Kelechi Ogueji, An-volume ACL/IJCNLP 2021 of Findings of ACL, pages 2022, Abu Dhabi, United Arab Emirates, Decem-dre Niyongabo Rubungo, Toan Q. Nguyen, Math-4693-4703. Association for Computational Linguis-ber 7-11, 2022, pages 8608-8621. Association forias Müller, André Müller, Shamsuddeen Hassantics. Computational Linguistics.",
"figure_id": "tab_7",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": "model𝑙𝑘 accuracy (%)3-gram1184.453-gram1375.773-gram1769.083-gram11362.753-gram12155.333-gram2179.753-gram2367.633-gram2759.493-gram21351.363-gram22142.683-gram3175.053-gram3360.223-gram3749.553-gram31338.343-gram32129.843-gram max 159.313-gram max 336.893-gram max 718.813-gram max 136.873-gram max 212.89",
"figure_id": "tab_8",
"figure_label": "10",
"figure_type": "table"
},
{
"figure_caption": "List of languages used to train Glot500-m (Part I).",
"figure_data": "Language-Script|Sent|FamilyHead Language-Script|Sent|FamilyHead Language-Script |Sent|FamilyHeadast_Latn4683554 indo1319guc_Latn249044 araw1281hrx_Latn45716 indo1319mon_Cyrl4616960 mong1349yesmam_Latn248348 maya1287quh_Latn45566 quec1387hbs_Cyrl4598073 indo1319nia_Latn247406 aust1307hyw_Cyrl45379 indo1319hau_Latn4368483 afro1255yesnyn_Latn241992 atla1278rue_Cyrl45369 indo1319sna_Latn4019596atla1278cab_Latn240101 araw1281eml_Latn44630 indo1319msa_Latn3929084yestop_Latn239232 toto1251acm_Arab44505 afro1255som_Latn3916769 afro1255yestog_Latn231969 atla1278tob_Latn44473 guai1249srp_Cyrl3864091 indo1319yesmco_Latn231209 mixe1284ach_Latn43974 nilo1247mlg_Latn3715802yestzh_Latn230706 maya1287vep_Latn43076 ural1272zul_Latn3580113atla1278pms_Latn227748 indo1319npi_Deva43072 indo1319arz_Arab3488224 afro1255wuu_Hani224088 sino1245tok_Latn42820arti1236nya_Latn3409030atla1278plt_Latn220413 aust1307sgs_Latn42467 indo1319tam_Taml3388255 drav1251yesyid_Hebr220214 indo1319yeslij_Latn42447 indo1319hat_Latn3226932 indo1319ada_Latn219427 atla1278myv_Cyrl42147 ural1272uzb_Latn3223485213615 aust1307tih_Latn41873 aust1307sot_Latn3205510atla1278kek_Latn209932 maya1287tat_Latn41640 turk1311uzb_Cyrl3029947 turk1311koo_Latn209375 atla1278lfn_Latn41632arti1236cos_Latn3015055 indo1319sop_Latn206501 atla1278cgg_Latn41196atla1278als_Latn2954874 indo1319kac_Latn205542 sino1245ful_Latn41188atla1278amh_Ethi2862985 afro1255yesqvi_Latn205447 quec1387gor_Latn41174 aust1307sun_Latn2586011 aust1307yescak_Latn204472 maya1287ile_Latn40984arti1236war_Latn2584810 aust1307kbp_Latn202877 atla1278ium_Latn40683 hmon1336div_Thaa2418687 indo1319ctu_Latn201662 maya1287teo_Latn40203 nilo1247yor_Latn2392359atla1278kri_Latn201087 indo1319kia_Latn40035atla1278fao_Latn2365271 indo1319mau_Latn199134 otom1299crh_Cyrl39985 turk1311uzn_Cyrl2293672 turk1311scn_Latn199068 indo1319crh_Latn39896 turk1311smo_Latn2290439 aust1307tyv_Cyrl198649 turk1311enm_Latn39809 indo1319bak_Cyrl2264196 turk1311ina_Latn197315arti1236sat_Olck39614 aust1305ilo_Latn2106531 aust1307btx_Latn193701 aust1307mad_Latn38993 aust1307tso_Latn2100708atla1278nch_Latn193129 utoa1244cac_Latn38812 maya1287mri_Latn2046850 aust1307ncj_Latn192962 utoa1244hnj_Latn38611 hmon1336hmn_Latn1903898pau_Latn190529 aust1307ksh_Latn38130 indo1319asm_Beng1882353 indo1319yestoj_Latn189651 maya1287ikk_Latn38071atla1278hil_Latn1798875 aust1307pcm_Latn187594 indo1319sba_Latn38040 cent2225nso_Latn1619354atla1278dyu_Latn186367 mand1469zom_Latn37013 sino1245ibo_Latn1543820atla1278kss_Latn185868 atla1278bqc_Latn36881 mand1469kin_Latn1521612atla1278afb_Arab183694 afro1255bim_Latn36835atla1278hye_Armn1463123 indo1319yesurh_Latn182214 atla1278mdy_Ethi36370 gong1255oci_Latn1449128 indo1319quc_Latn181559 maya1287bts_Latn36216 aust1307lin_Latn1408460atla1278new_Deva181427 sino1245gya_Latn35902atla1278tpi_Latn1401844 indo1319yao_Latn179965 atla1278ajg_Latn35631atla1278twi_Latn1400979atla1278ngl_Latn178498 atla1278agw_Latn35585 aust1307kir_Cyrl1397566 turk1311yesnyu_Latn177483 atla1278kom_Cyrl35249 ural1272pap_Latn1360138 indo1319kab_Latn176015 afro1255knv_Latn35196nep_Deva1317291 indo1319yestuk_Cyrl175769 turk1311giz_Latn35040 afro1255azj_Latn1315834 turk1311xmf_Geor174994 kart1248hui_Latn34926 nucl1709bcl_Latn1284493 aust1307ndc_Latn174305 atla1278kpg_Latn34900 aust1307xho_Latn1262364atla1278yessan_Deva165616 indo1319yeszea_Latn34426 indo1319cym_Latn1244783 indo1319yesnba_Latn163485 atla1278aoj_Latn34349 nucl1708gaa_Latn1222307atla1278bpy_Beng162838 indo1319csy_Latn34126 sino1245ton_Latn1216118 aust1307ncx_Latn162558 utoa1244azb_Arab33758 turk1311yestah_Latn1190747 aust1307qug_Latn162500 quec1387csb_Latn33743 indo1319lat_Latn1179913 indo1319yesrmn_Latn162069 indo1319tpm_Latn33517atla1278srn_Latn1172349 indo1319cjk_Latn160645 atla1278quw_Latn33449 quec1387ewe_Latn1161605atla1278arb_Arab159884 afro1255yesrmy_Cyrl33351 indo1319bem_Latn1111969atla1278kea_Latn158047 indo1319ixl_Latn33289 maya1287efi_Latn1082621atla1278mck_Latn157521 atla1278mbb_Latn33240 aust1307bis_Latn1070170 indo1319arn_Latn155882 arau1255pfl_Latn33148 indo1319orm_Latn1067699yespdt_Latn155485 indo1319pcd_Latn32867 indo1319haw_Latn1062491 aust1307her_Latn154827 atla1278tlh_Latn32863arti1236hmo_Latn1033636 pidg1258gla_Latn152563 indo1319yessuz_Deva32811 sino1245kat_Geor1004297 kart1248yeskmr_Cyrl151728 indo1319gcr_Latn32676 indo1319pag_Latn983637aust1307mwl_Latn150054 indo1319jbo_Latn32619arti1236loz_Latn964418atla1278nav_Latn147702 atha1245tbz_Latn32264atla1278fry_Latn957422indo1319yesksw_Mymr147674 sino1245bam_Latn32150 mand1469mya_Mymr945180sino1245yesmxv_Latn147591 otom1299prk_Latn32085 aust1305nds_Latn944715indo1319hif_Latn147261 indo1319jam_Latn32048 indo1319run_Latn943828atla1278wol_Latn146992 atla1278twx_Latn32028atla1278",
"figure_id": "tab_10",
"figure_label": "11",
"figure_type": "table"
},
{
"figure_caption": "List of languages used to train Glot500-m (Part II).",
"figure_data": "Language-Script|Sent|FamilyHead Language-Script|Sent|FamilyHead Language-Script |Sent|FamilyHeadpnb_Arab899895 indo1319sme_Latn146803 ural1272nmf_Latn31997 sino1245rar_Latn894515 aust1307gom_Latn143937 indo1319caq_Latn31903 aust1305fij_Latn887134 aust1307bum_Latn141673 atla1278rop_Latn31889 indo1319wls_Latn882167 aust1307mgr_Latn138953 atla1278tca_Latn31852 ticu1244ckb_Arab874441 indo1319ahk_Latn135068 sino1245yan_Latn31775 misu1242ven_Latn860249 atla1278kur_Arab134160 indo1319xav_Latn31765 nucl1710zsm_Latn859947 aust1307yesbas_Latn133436 atla1278bih_Deva31658chv_Cyrl859863 turk1311bin_Latn133256 atla1278cuk_Latn31612 chib1249lua_Latn854359 atla1278tsz_Latn133251 tara1323kjb_Latn31471 maya1287que_Latn838486sid_Latn130406 afro1255hne_Deva31465 indo1319sag_Latn771048 atla1278diq_Latn128908 indo1319wbm_Latn31394 aust1305guw_Latn767918 atla1278srd_Latn127064zlm_Latn31345 aust1307bre_Latn748954 indo1319yestcf_Latn126050 otom1299tui_Latn31161 atla1278toi_Latn745385 atla1278bzj_Latn124958 indo1319ifb_Latn30980indo1319yesudm_Cyrl121705 ural1272izz_Latn30894 atla1278che_Cyrl728201 nakh1245cce_Latn120636 atla1278rug_Latn30857 aust1307pis_Latn714783 indo1319meu_Latn120273 aust1307aka_Latn30704 atla1278kon_Latn685194chw_Latn119751 atla1278pxm_Latn30698 book1242oss_Cyrl683517 indo1319cbk_Latn118789 indo1319kmm_Latn30671 sino1245hyw_Armn679819 indo1319ibg_Latn118733 aust1307mcn_Latn30666 afro1255iso_Latn658789 atla1278bhw_Latn117381 aust1307ifa_Latn30621 aust1307nan_Latn656389 sino1245ngu_Latn116851 utoa1244dln_Latn30620 sino1245lub_Latn654390 atla1278nyy_Latn115914 atla1278ext_Latn30605 indo1319lim_Latn652078 indo1319szl_Latn112496 indo1319ksd_Latn30550 aust1307tuk_Latn649411 turk1311ish_Latn111814 atla1278mzh_Latn30517 mata1289tir_Ethi649117 afro1255naq_Latn109747 khoe1240llb_Latn30480 atla1278tgk_Latn636541 indo1319toh_Latn107583 atla1278hra_Latn30472 sino1245yua_Latn610052 maya1287ttj_Latn106925 atla1278mwm_Latn30432 cent2225min_Latn609065 aust1307nse_Latn105189 atla1278krc_Cyrl30353 turk1311lue_Latn599429 atla1278hsb_Latn104802 indo1319tuc_Latn30349 aust1307khm_Khmr590429 aust1305yesami_Latn104559 aust1307mrw_Latn30304 aust1307tum_Latn589857 atla1278alz_Latn104392 nilo1247pls_Latn30136 otom1299tll_Latn586530 atla1278apc_Arab102392 afro1255rap_Latn30102 aust1307ekk_Latn582595 ural1272vls_Latn101900 indo1319fur_Latn30052 indo1319lug_Latn566948 atla1278mhr_Cyrl100474 ural1272kaa_Latn30031 turk1311niu_Latn566715 aust1307djk_Latn99234 indo1319prs_Arab26823 indo1319yestzo_Latn540262 maya1287wes_Latn98492 indo1319san_Latn25742 indo1319yesmah_Latn534614 aust1307gkn_Latn97041atla1278som_Arab14199 afro1255yestvl_Latn521556 aust1307grc_Grek96986 indo1319uig_Latn9637turk1311yesjav_Latn516833 aust1307yeshbo_Hebr96484afro1255hau_Arab9593afro1255yes",
"figure_id": "tab_11",
"figure_label": "12",
"figure_type": "table"
},
{
"figure_caption": "List of languages used to train Glot500-m (Part III).",
"figure_data": "guages (Abate et al., 2018), Phontron (Neubig,2011), QADI (Abdelali et al., 2021), Quechua-IIC(Zevallos et al., 2022), SLI_GalWeb.1.0 (Agerriet al., 2018), Shami (Abu Kwaik et al., 2018),Stanford NLP, 23 StatMT, 24 TICO (Anastasopou-los et al., 2020), TIL (Mirzakhalov et al., 2021),Tatoeba, 25 TeDDi (Moran et al., 2022), Tilde (Rozisand Skadin , š, 2017), W2C (Majliš, 2011), WAT(Nakazawa et al., 2022), WikiMatrix (Schwenket al., 2021), Wikipedia, 26 Workshop on NER forSouth and South East Asian Languages (Singh,2008), XLSum (Hasan et al., 2021).",
"figure_id": "tab_12",
"figure_label": "13",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": "(Sentence Retrieval Tatoeba),15, 16 (Sentence Retrieval Bible), 17 (NER), and 18(POS), 19, 20 (Text Classification), 21, 22 (RoundTrip Alignment).",
"figure_id": "tab_13",
"figure_label": "14",
"figure_type": "table"
},
{
"figure_caption": "F1 of XLM-R-B, XLM-R-L, and Glot500-m on Text Classification (Part I).",
"figure_data": "m",
"figure_id": "tab_19",
"figure_label": "19",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "tab_22",
"figure_label": "21",
"figure_type": "table"
},
{
"figure_caption": "Accuracy of XLM-R-B, XLM-R-L, and Glot500-m on Round Trip Alignment (Part II).",
"figure_data": "",
"figure_id": "tab_24",
"figure_label": "22",
"figure_type": "table"
},
{
"figure_caption": "Perplexity of all languages covered by Glot500-m (Part I).Language-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-m",
"figure_data": "m",
"figure_id": "tab_25",
"figure_label": "23",
"figure_type": "table"
},
{
"figure_caption": "Perplexity of all languages covered by Glot500-m (Part II).",
"figure_data": "Language-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-mbts_Latn205.7204.58.8tsn_Latn264.7137.812.5orm_Latn23.48.616gla_Latn11.512.77.2pon_Latn928.4181.919.2luo_Latn699.4258.585.1kat_Latn36.424.818.3nmf_Latn297.6310.644.9pcm_Latn38.3169.63.6uig_Latn188.8173.915.2ajg_Latn147.1149.522.6nnb_Latn364.19528.6kat_Geor63.96.4tir_Ethi28.315.74.4kaz_Cyrl4.35.49.6mlg_Latn10.94.47.6bhw_Latn411.2126.221.6dzo_Tibt8.53.35.7arn_Latn382.796.717.6mhr_Cyrl122.9168.45.8sun_Latn23.611.917tuk_Latn456.7197.85.8swe_Latn4.83.512.7vec_Latn40.621.19.2vls_Latn97.739.69.7scn_Latn11764.97.8ayr_Latn261.1237.627.7hyw_Armn15.89.14.3udm_Cyrl356.7224.96.7oke_Latn209.2220.113.0que_Latn447.9536.111.9ifb_Latn246.3177.95.1kur_Latn14.26.810.3snd_Arab13.24.119.5naq_Latn136.860.215.7mgh_Latn680272.823.7giz_Latn81.982.937.7zlm_Latn5.63.34.6tgk_Cyrl181.31534.5ita_Latn4.53.37.2hrx_Latn478.1679.114.9sop_Latn607.5228.229.5qub_Latn283.2312.79.4lzh_Hani705821.8mos_Latn272.6118.313.2nav_Latn228.5126.55.2pap_Latn674.4149.318.1rap_Latn36.131.12.8kqn_Latn825.9686.617.5cfm_Latn235.115514.0prk_Latn69.445.97.1toh_Latn758.3216.619.6chv_Cyrl122.573.85.4uzb_Cyrl236.2138.44.9mah_Latn314.781.817.3tdt_Latn641.978.69.7tog_Latn821.1777.713.4wes_Latn144.6103.914.3pan_Guru4.42.54.3mal_Mlym53.76.2nob_Latn6.84.09.5pms_Latn83.646.23.6nyk_Latn1182.6914.216.5ext_Latn68.338.28.1roh_Latn243.51707.0quy_Latn949.7320.214.5lam_Latn233.7160.821.6prs_Arab6.83.54.8abn_Latn245.2272.58.7mwm_Latn44.853.17.1tuk_Cyrl277.486.36.7mcn_Latn120.7129.743.6kpg_Latn165.9122.615.1srm_Latn257.574.512.3nep_Deva8.86.310hau_Arab5.33.08.1gsw_Latn288.2181.222.3gle_Latn10.53.79.8ksd_Latn150154.97.7fat_Latn192.314917.6cab_Latn1216.7155.615.4zsm_Latn12.22.922.7ldi_Latn394.8107.138.2mps_Latn75.255.217.4hui_Latn209.917710.0kos_Latn470.7485.727.0pnb_Arab51.830.87.1cym_Latn8.24.811.2acr_Latn155.790.75.8swa_Latn11.46.420srp_Latn10.97.913.3mri_Latn6359.58.7hnj_Latn88.392.511.3bak_Latn347.12117.5frr_Latn117.61019.5haw_Latn63.566.77.4zho_Hani20.75.931.3mck_Latn369.3164.824.7tpi_Latn891.867.88.8nno_Latn9.912.710.4pes_Arab5.53.15.3ncj_Latn1019136.213.7gya_Latn3124.316.5san_Latn94.496.812.0som_Latn14.16.922.2ibo_Latn77.190.18.5yao_Latn738.9162.413.8mam_Latn132.762.46.1meu_Latn380.2158.526.7srp_Cyrl7.44.58.4lit_Latn4.42.510.6ncx_Latn1084.7948.514.6ful_Latn104105.613.1",
"figure_id": "tab_26",
"figure_label": "24",
"figure_type": "table"
},
{
"figure_caption": "Perplexity of all languages covered by Glot500-m (Part III).",
"figure_data": "",
"figure_id": "tab_27",
"figure_label": "25",
"figure_type": "table"
}
] | Ayyoob Imani; Peiqin Lin; Amir Hossein Kargaran; Silvia Severini; Masoud Jalili Sabet; Nora Kassner; Chunlan Ma; Helmut Schmid; André F T Martins; François Yvon; Hinrich Schütze; Pierre Platen; Pierre François Cornette; Rémi Lavallée; Samyam Lacroix; Sanchit Rajbhandari; Shaden Gandhi; Stéphane Smith; Suraj Requena; Tim Patil; Ahmed Dettmers; Amanpreet Baruwa; Anasta- Sia Singh; Anne-Laure Cheveleva; Arjun Ligozat; Aurélie Subramo- Nian; Charles Névéol; Dan Lovering; Deepak Garrette; Ehud Tunuguntla; Ekaterina Reiter; Ekaterina Takta- Sheva; Eli Voloshina; Genta Bogdanov; Hailey In- Dra Winata; Jan-Christoph Schoelkopf; Jekaterina Kalo; Jessica Zosa Novikova; Jordan Forde; Jungo Clive; Ken Kasai; Liam Kawamura; Marine Hazan; Miruna Carpuat; Najoung Clinciu; New- Ton Kim; Oleg Cheng; Omer Serikov; Oskar Antverg; Rui Van Der Wal; Ruochen Zhang; Sebas- Tian Zhang; Shachar Gehrmann; Shani Mirkin; Tatiana Pais; Thomas Shavrina; Tian Scialom; Tomasz Yun; Verena Lim- Isiewicz; Vitaly Rieser; Vladislav Protasov; Yada Mikhailov; Yonatan Pruksachatkun; Zachary Belinkov; Zdeněk Bamberger; Alice Kasner; Amanda Rueda; Amir Pestana; Ammar Feizpour; Amy Khan; Ana Faranak; Anthony Santos; Antigona Hevia; Arash Unl- Dreaj; Arezoo Aghagol; Aycha Abdollahi; Azadeh Tam- Mour; Bahareh Hajihosseini; Ben- Jamin Behroozi; Bharat Ajibade; Carlos Saxena; Ferran- Dis Muñoz; Danish Contractor; David Lansky; Davis David; Douwe Kiela; Duong A Nguyen; Edward Tan; Emi Baylor; Ezinwanne Ozoani; Fatima Mirza; Frankline Ononiwu; Habib Rezanejad; Hessie Jones; Indrani Bhattacharya; Irene Solaiman; Irina Sedenko; Isar Nejadgholi; Jesse Passmore; Josh Seltzer; Julio Bo- Nis Sanz; Livia Dutra; Mairon Samagaio; Maraim Elbadri; Margot Mieskes; Marissa Gerchick; Martha Akinlolu; Michael Mckenna; Mike Qiu; Muhammed Ghauri; Mykola Burynok; Nafis Abrar; Nazneen Ra- Jani; Nour Elkott; Nour Fahmy; Olanrewaju Samuel; Ran An; Rasmus Kromann; Ryan Hao; Samira Al- Izadeh; Sarmad Shubber; Silas Wang; Sourav Roy; Sylvain Viguier; Thanh Le; Tobi Oyebade; Trieu Le; Yoyo Yang; Zach Nguyen; Ramesh Kashyap; Alfredo Palasciano; Alison Callahan; Anima Shukla; Antonio Miranda-Escalada; Ayush Singh; Benjamin Beilharz; Bo Wang; Caio Brito; Chenxi Zhou; Chirag Jain; Chuxin Xu; Clémentine Fourrier; Daniel León Periñán; Daniel Molano; Dian Yu; Enrique Manjava- Cas; Fabio Barth; Florian Fuhrimann; Gabriel Altay; Giyaseddin Bayrak; Gully Burns; Helena U Vrabec; Imane Bello; Ishani Dash; Jihyun Kang; John Giorgi; Jonas Golde; Jose David Posada; Rangasai Karthik; Lokesh Sivaraman; Lu Bulchandani; Luisa Liu; Madeleine Shin- Zato; Maiko Hahn De Bykhovetz; Marc Takeuchi; Maria A Pàmies; Marianna Castillo; Mario Nezhurina; Matthias Sänger; Michael Samwald; Michael Cullan; Michiel Weinberg; Mina De Wolf; Minna Mihalj- Cic; Moritz Liu; Myungsun Freidank; Natasha Kang; Nathan Seelam; Nicholas Dahlberg; Nikolaus Michio Broad; Pascale Muellner; Patrick Fung; Ramya Haller; Renata Chandrasekhar; Robert Eisenberg; Rodrigo Martin; Rosaline Canalli; Ruisi Su; Samuel Su; Samuele Cahyawijaya; Garda; S Shlok; Shubhanshu Deshmukh; Sid Mishra; Si- Mon Kiblawi; Sinee Ott; Srishti Sang-Aroonsiri; Stefan Kumar; Sushil Schweter; Tanmay Bharati; Théo Laud; Tomoya Gigant; Wojciech Kainuma; Yanis Kusa; Labrak; Shailesh Yash; Yash Bajaj; Yifan Venkatraman; Yingxin Xu; Yu Xu; Xu | [
{
"authors": "Solomon Teferra Abate; Michael Melese; Martha Yifiru Tachbelie; Million Meshesha; Solomon Atinafu; Wondwossen Mulugeta; Yaregal Assabie; Hafte Abera; Binyam Ephrem; Tewodros Abebe; Wondimagegnhue Tsegaye; Amanuel Lemma; Tsegaye Andargie; Seifedin Shifaw",
"journal": "Association for Computational Linguistics",
"ref_id": "b0",
"title": "Parallel corpora for bi-lingual English-Ethiopian languages statistical machine translation",
"year": "2018"
},
{
"authors": "Ahmed Abdelali; Hamdy Mubarak; Younes Samih; Sabit Hassan; Kareem Darwish",
"journal": "Association for Computational Linguistics",
"ref_id": "b1",
"title": "QADI: Arabic dialect identification in the wild",
"year": "2021"
},
{
"authors": "Kathrein Abu Kwaik; Motaz Saad; Stergios Chatzikyriakidis; Simon Dobnik",
"journal": "European Language Resources Association (ELRA",
"ref_id": "b2",
"title": "Shami: A corpus of Levantine Arabic dialects",
"year": "2018"
},
{
"authors": "Ife Adebara; Abdelrahim Elmadany; Muhammad Abdul-Mageed; Alcides Alcoba; Inciarte ",
"journal": "",
"ref_id": "b3",
"title": "SERENGETI: Massively multilingual language models for Africa",
"year": "2022"
},
{
"authors": "David Adelani; Jesujoba Alabi; Angela Fan; Julia Kreutzer; Xiaoyu Shen; Machel Reid; Dana Ruiter; Dietrich Klakow; Peter Nabende; Ernie Chang; Tajuddeen Gwadabe; Freshia Sackey; F P Bonaventure; Chris Dossou; Colin Emezue; Michael Leong; Shamsuddeen Beukman; Guyo Muhammad; Oreen Jarso; Andre Yousuf; Gilles Niyongabo Rubungo; Eric Hacheme; Muhammad Umair Peter Wairagala; Benjamin Nasir; Tunde Ajibade; Yvonne Ajayi; Jade Gitau; Mohamed Abbott; Millicent Ahmed; Anuoluwapo Ochieng; Perez Aremu; Jonathan Ogayo; Fatoumata Mukiibi; Godson Ouoba Kabore; Derguene Kalipe; Mbaye; Auguste Allahsera; Victoire Tapo; Edwin Memdjokam Koagne; Valencia Munkoh-Buabeng; Idris Wagner; Ayodele Abdulmumin; Happy Awokoya; Blessing Buzaaba; Andiswa Sibanda; Sam Bukula; Manthalu",
"journal": "Association for Computational Linguistics",
"ref_id": "b4",
"title": "A few thousand translations go a long way! leveraging pre-trained models for African news translation",
"year": "2022"
},
{
"authors": "David Adelani; Dana Ruiter; Jesujoba Alabi; Damilola Adebonojo; Adesina Ayeni; Mofe Adeyemi; Ayodele Esther Awokoya; Cristina España-Bonet ",
"journal": "Virtual. Association for Machine Translation in the Americas",
"ref_id": "b5",
"title": "The effect of domain and diacritics in Yoruba-English neural machine translation",
"year": "2021"
},
{
"authors": "Rodrigo Agerri; Xavier Gómez Guinovart; German Rigau; Miguel Anxo; Solla Portela",
"journal": "European Language Resources Association (ELRA",
"ref_id": "b6",
"title": "Developing new linguistic resources and tools for the Galician language",
"year": "2018"
},
{
"authors": "O Jesujoba; David Alabi; Marius Ifeoluwa Adelani; Dietrich Mosbach; Klakow",
"journal": "International Committee on Computational Linguistics",
"ref_id": "b7",
"title": "Adapting pretrained language models to African languages via multilingual adaptive fine-tuning",
"year": "2022"
},
{
"authors": "Chunlan Ma; Ayyoob Imanigooghari; Haotian Ye; Ehsaneddin Asgari; Hinrich Schütze",
"journal": "",
"ref_id": "b8",
"title": "Taxi1500: A multilingual dataset for text classification in 1500 languages",
"year": "2023"
},
{
"authors": "Martin Majliš",
"journal": "",
"ref_id": "b9",
"title": "W2C -web to corpus -corpora",
"year": "2011"
},
{
"authors": "Jamshidbek Mirzakhalov; Anoop Babu; Duygu Ataman; Sherzod Kariev; Francis Tyers; Otabek Abduraufov; Mammad Hajili; Sardana Ivanova; Abror Khaytbaev; Antonio Laverghetta; Esra Bekhzodbek Moydinboyev; Shaxnoza Onal; Ahsan Pulatova; Orhan Wahab; Sriram Firat; Chellappan",
"journal": "Association for Computational Linguistics",
"ref_id": "b10",
"title": "A large-scale study of machine translation in Turkic languages",
"year": "2021"
},
{
"authors": "Steven Moran; Christian Bentz; Ximena Gutierrez-Vasques; Olga Pelloni; Tanja Samardzic",
"journal": "European Language Resources Association",
"ref_id": "b11",
"title": "TeDDi sample: Text data diversity sample for language comparison and multilingual NLP",
"year": "2022"
},
{
"authors": "Makoto Morishita; Jun Suzuki; Masaaki Nagata",
"journal": "European Language Resources Association",
"ref_id": "b12",
"title": "JParaCrawl: A large scale web-based English-Japanese parallel corpus",
"year": "2020"
},
{
"authors": "Toshiaki Nakazawa; Hideya Mino; Isao Goto; Raj Dabre; Shohei Higashiyama; Shantipriya Parida; Anoop Kunchukuttan; Makoto Morishita; Ondřej Bojar; Chenhui Chu; Akiko Eriguchi; Kaori Abe; Yusuke Oda; Sadao Kurohashi",
"journal": "",
"ref_id": "b13",
"title": "Overview of the 9th workshop on Asian translation",
"year": "2022"
},
{
"authors": "Toshiaki Nakazawa; Hideki Nakayama; Chenchen Ding; Raj Dabre; Shohei Higashiyama; Hideya Mino; Isao Goto; Win Pa Pa; Anoop Kunchukuttan; Shantipriya Parida; Ondřej Bojar; Chenhui Chu; Akiko Eriguchi; Kaori Abe; Yusuke Oda; Sadao Kurohashi",
"journal": "Association for Computational Linguistics",
"ref_id": "b14",
"title": "Overview of the 8th workshop on Asian translation",
"year": "2021"
},
{
"authors": "Graham Neubig",
"journal": "",
"ref_id": "b15",
"title": "The Kyoto free translation task",
"year": "2011"
},
{
"authors": "Kelechi Ogueji; Yuxin Zhu; Jimmy Lin; ; ",
"journal": "Association for Computational Linguistics",
"ref_id": "b16",
"title": "Small data? no problem! exploring the viability of pretrained multilingual language models for lowresourced languages",
"year": "2021"
},
{
"authors": "Kelechi Ogueji; Yuxin Zhu; Jimmy Lin",
"journal": "",
"ref_id": "b17",
"title": "Small data? no problem! exploring the viability of pretrained multilingual language models for lowresourced languages",
"year": "2021"
},
{
"authors": "Chester Palen-Michel; June Kim; Constantine Lignos",
"journal": "European Language Resources Association",
"ref_id": "b18",
"title": "Multilingual open text release 1: Public domain news in 44 languages",
"year": "2022"
},
{
"authors": "Xiaoman Pan; Boliang Zhang; Jonathan May; Joel Nothman; Kevin Knight; Heng Ji",
"journal": "Association for Computational Linguistics",
"ref_id": "b19",
"title": "Cross-lingual name tagging and linking for 282 languages",
"year": "2017"
},
{
"authors": "Jonas Pfeiffer; Naman Goyal; Xi Lin; Xian Li; James Cross; Sebastian Riedel; Mikel Artetxe",
"journal": "Association for Computational Linguistics",
"ref_id": "b20",
"title": "Lifting the curse of multilinguality by pre-training modular transformers",
"year": "2022"
},
{
"authors": "Jonas Pfeiffer; Ivan Vulić; Iryna Gurevych; Sebastian Ruder",
"journal": "Association for Computational Linguistics",
"ref_id": "b21",
"title": "UNKs everywhere: Adapting multilingual language models to new scripts",
"year": "2021"
},
{
"authors": "Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu",
"journal": "J. Mach. Learn. Res",
"ref_id": "b22",
"title": "Exploring the limits of transfer learning with a unified text-to-text transformer",
"year": "2020"
},
{
"authors": "Roberts Rozis; Raivis Skadin; Š ",
"journal": "Association for Computational Linguistics",
"ref_id": "b23",
"title": "Tilde MODEL -multilingual open data for EU languages",
"year": "2017"
},
{
"authors": "Hassan Sajjad; Ahmed Abdelali; Nadir Durrani; Fahim Dalvi",
"journal": "International Committee on Computational Linguistics",
"ref_id": "b24",
"title": "AraBench: Benchmarking dialectal Arabic-English machine translation",
"year": "2020"
},
{
"authors": "Julian Salazar; Davis Liang; Toan Q Nguyen; Katrin Kirchhoff",
"journal": "Association for Computational Linguistics",
"ref_id": "b25",
"title": "Masked language model scoring",
"year": "2020"
},
{
"authors": "Holger Schwenk; Vishrav Chaudhary; Shuo Sun; Hongyu Gong; Francisco Guzmán",
"journal": "Association for Computational Linguistics",
"ref_id": "b26",
"title": "Wiki-Matrix: Mining 135M parallel sentences in 1620 language pairs from Wikipedia",
"year": "2021"
},
{
"authors": "Silvia Severini; Ayyoob Imani; Philipp Dufter; Hinrich Schütze",
"journal": "",
"ref_id": "b27",
"title": "Towards a broad coverage named entity resource: A data-efficient approach for many diverse languages",
"year": "2022"
},
{
"authors": "Aditya Siddhant; Ankur Bapna; Orhan Firat; Yuan Cao; Mia Xu Chen; Isaac Caswell; Xavier Garcia",
"journal": "",
"ref_id": "b28",
"title": "Towards the next 1000 languages in multilingual machine translation: Exploring the synergy between supervised and self-supervised learning",
"year": "2022"
},
{
"authors": "Anil Kumar; Singh ",
"journal": "",
"ref_id": "b29",
"title": "Named entity recognition for south and south East Asian languages: Taking stock",
"year": "2008"
},
{
"authors": "Pedro Javier; Ortiz Suárez; Benoît Sagot; Laurent Romary",
"journal": "Leibniz-Institut für Deutsche Sprache",
"ref_id": "b30",
"title": "Asynchronous pipeline for processing huge corpora on medium to low resource infrastructures",
"year": "2019"
},
{
"authors": "Jörg Tiedemann",
"journal": "European Language Resources Association (ELRA)",
"ref_id": "b31",
"title": "Parallel data, tools and interfaces in opus",
"year": "2012"
},
{
"authors": "Iulia Turc; Kenton Lee; Jacob Eisenstein; Ming-Wei Chang; Kristina Toutanova",
"journal": "",
"ref_id": "b32",
"title": "Revisiting the primacy of english in zero-shot cross-lingual transfer",
"year": "2021"
},
{
"authors": "Hai Wang; Dian Yu; Kai Sun; Jianshu Chen; Dong Yu",
"journal": "Association for Computational Linguistics",
"ref_id": "b33",
"title": "Improving pre-trained multilingual model with vocabulary expansion",
"year": "2019"
},
{
"authors": "Mingyang Wang; Heike Adel; Lukas Lange; Jannik Strötgen; Hinrich Schütze",
"journal": "",
"ref_id": "b34",
"title": "NLNDE at semeval-2023 task 12: Adaptive pretraining and source language selection for low-resource multilingual sentiment analysis",
"year": "2023"
},
{
"authors": "Xinyi Wang; Sebastian Ruder; Graham Neubig",
"journal": "Association for Computational Linguistics",
"ref_id": "b35",
"title": "Expanding pretrained models to thousands more languages via lexicon-based adaptation",
"year": "2022"
},
{
"authors": "Guillaume Wenzek; Marie-Anne Lachaux; Alexis Conneau; Vishrav Chaudhary; Francisco Guzmán; Armand Joulin; Edouard Grave",
"journal": "European Language Resources Association",
"ref_id": "b36",
"title": "Ccnet: Extracting high quality monolingual datasets from web crawl data",
"year": "2020-05-11"
},
{
"authors": "Guillaume Wenzek; Marie-Anne Lachaux; Alexis Conneau; Vishrav Chaudhary; Francisco Guzmán; Armand Joulin; Edouard Grave",
"journal": "European Language Resources Association",
"ref_id": "b37",
"title": "CCNet: Extracting high quality monolingual datasets from web crawl data",
"year": "2020"
},
{
"authors": "Linting Xue; Noah Constant; Adam Roberts; Mihir Kale; Rami Al-Rfou; Aditya Siddhant; Aditya Barua; Colin Raffel",
"journal": "Association for Computational Linguistics",
"ref_id": "b38",
"title": "mT5: A massively multilingual pre-trained text-to-text transformer",
"year": "2021"
},
{
"authors": "Jian Yang; Shuming Ma; Dongdong Zhang; Shuangzhi Wu; Zhoujun Li; Ming Zhou",
"journal": "",
"ref_id": "b39",
"title": "Alternating language modeling for cross-lingual pre-training",
"year": "2020"
},
{
"authors": "Rodolfo Zevallos; John Ortega; William Chen; Richard Castro; Núria Bel; Cesar Toshio; Renzo Venturas; Hilario Aradiel; Nelsi Melgarejo",
"journal": "Association for Computational Linguistics",
"ref_id": "b40",
"title": "Introducing QuBERT: A large monolingual corpus and BERT model for Southern Quechua",
"year": "2022"
}
] | [
{
"formula_coordinates": [
4,
318.81,
71.78,
192.95,
43.59
],
"formula_id": "formula_0",
"formula_text": "XLM-R-B XLM-R-L Glot500-m Model Size 278M 560M 395M Vocab Size 250K 250K 401K Transformer Size 86M 303M 86M"
},
{
"formula_coordinates": [
7,
173.79,
71.74,
342.98,
17.11
],
"formula_id": "formula_1",
"formula_text": "tail head all XLM-R-B XLM-R-L Glot500-m XLM-R-B XLM-R-L Glot500-m XLM-R-B XLM-R-L Glot500-"
},
{
"formula_coordinates": [
17,
102.19,
129.84,
187.67,
31.23
],
"formula_id": "formula_2",
"formula_text": "PP (𝑆, 𝑀) = 𝑇 𝑇 𝑡=1 1 P 𝑐ℎ 𝑡 | 𝑐ℎ 𝑡 -1 1 (1)"
},
{
"formula_coordinates": [
17,
110.07,
225.22,
179.8,
32.93
],
"formula_id": "formula_3",
"formula_text": "P 𝑐ℎ 𝑡 | 𝑐ℎ 𝑡 -1 1 = 𝐶 𝑐ℎ 𝑡 -1 1 𝑐ℎ 𝑡 𝐶 𝑐ℎ 𝑡 -1 1 (2)"
},
{
"formula_coordinates": [
17,
77.26,
549.74,
212.61,
10.09
],
"formula_id": "formula_4",
"formula_text": "D 𝐿 𝑖 ,𝐿 𝑗 = max PP (𝑆 𝐿 𝑖 , 𝑀 𝐿 𝑗 ), PP (𝑆 𝐿 𝑗 , 𝑀 𝐿 𝑖 ) (3)"
},
{
"formula_coordinates": [
17,
329.54,
261.37,
170.67,
27.14
],
"formula_id": "formula_5",
"formula_text": "𝑇 𝐿 𝑖 : 1 → 2 → 3 → 4 → 5 → 6 𝑇 𝐿 𝑗 : 1 → 2 → 7 → 8"
},
{
"formula_coordinates": [
23,
75.91,
276.42,
440.38,
5.64
],
"formula_id": "formula_6",
"formula_text": "Language-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-"
},
{
"formula_coordinates": [
24,
98.37,
131.14,
417.92,
5.64
],
"formula_id": "formula_7",
"formula_text": "-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-"
},
{
"formula_coordinates": [
25,
75.91,
197.89,
440.38,
5.64
],
"formula_id": "formula_8",
"formula_text": "Language-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-"
},
{
"formula_coordinates": [
26,
75.91,
190.04,
440.38,
5.64
],
"formula_id": "formula_9",
"formula_text": "Language-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-"
},
{
"formula_coordinates": [
27,
75.91,
284.27,
440.38,
5.64
],
"formula_id": "formula_10",
"formula_text": "Language-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-"
},
{
"formula_coordinates": [
28,
98.37,
131.14,
417.92,
5.64
],
"formula_id": "formula_11",
"formula_text": "-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-"
},
{
"formula_coordinates": [
29,
98.37,
213.6,
417.92,
5.64
],
"formula_id": "formula_12",
"formula_text": "-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-"
},
{
"formula_coordinates": [
30,
98.37,
103.66,
417.92,
5.64
],
"formula_id": "formula_13",
"formula_text": "-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-"
},
{
"formula_coordinates": [
31,
98.37,
209.67,
417.92,
5.64
],
"formula_id": "formula_14",
"formula_text": "-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-"
},
{
"formula_coordinates": [
32,
98.37,
131.14,
417.92,
5.64
],
"formula_id": "formula_15",
"formula_text": "-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-m Language-Script XLM-R-B XLM-R-L Glot500-"
}
] |
10.18653/v1/2020.emnlp-main.506 | 2023-12-01 | [
{
"figure_ref": [
"fig_1"
],
"heading": "Introduction",
"publication_ref": [
"b23",
"b32",
"b12"
],
"table_ref": [],
"text": "Document-grounded dialog agents converse with users based on information present in document provided to them. These agents are expected to be factually consistent or faithful to the grounding document and refrain from generating content that cannot be verified using the document. As most existing document-grounded dialog agents (Prabhumoye et al., 2021;Wu et al., 2021) are built by fine-tuning large language models, ensuring faithful response generation is a major challenge.\nTo measure the ability of dialog agents to generate faithful responses, several automatic metrics have been proposed. These metrics take as input the agent generated response and the grounding document to quantify faithfulness. These are based on lexical overlap (e.g., BLEU, unigram-F1), semantic overlap (BERTScore) or even a trained classifier (Dziri et al., 2022a). Recently, Honovich et al. (2021) proposed Q 2 , a metric that measures faithfullness using automatic question generation and question answering.\nA major limitation of existing metrics is that they ignore the crucial dialog history when measuring faithfulness of responses. Even though, in many cases, the dialog history provides essential context that is necessary for a complete understanding of the response. To illustrate this point, let's consider two responses, textitR1 and R2 , as depicted in Figure 1. Response R1 is self-contained and can be comprehended without relying on the dialog history. On the other hand, response R2 is dependent on the dialog history and can only be fully understood when considering the preceding conversation. Unfortunately, current automated metrics do not take into account the dialog history, leading to their failure in evaluating responses that are not self-contained. Responses like R2 often lack domainspecific words, making similarity-based metrics like unigram-F1 and BERTScore ineffective. Additionally, generating question-answer pairs using such responses typically captures incomplete information, rendering metrics like Q2 as inadequate measures.\nTo overcome this problem, we propose a new metric that quantifies the faithfulness of a generated response with respect to both the document and the dialog history. Our metric is grounded in information theoretic concepts and captures the association of the response with the given document using Conditional Pointwise Mutual Information (CPMI). We call our metric PMI-FAITH, which uses CPMI between the generated response and the document, conditioned on the dialogue history, for quantifying faithfulness. PMI-FAITH captures the intuition that for a response to be grounded in the document, the probability of its generation given the document should be higher than the probability of its generation without the document.\nA significant advantage of our metric PMI-FAITH is that it can be factorized the same way as the likelihood of a response can be factorized in auto regressive models. We take advantage of this property to propose a novel decoding objective, PMI-DECODE. The goal of PMI-DECODE is to maximize not just the response's likelihood but a score that combines its likelihood and faithfulness. To summarize, our contributions are threefold:\n1. We propose PMI-FAITH, a novel metric which quantifies faithfulness as a conditional PMI between the response and the document given the dialog history. 2. We propose a novel decoding objective, PMI-DECODE, which can aid in generating faithful responses. 3. Our experiments show that PMI-FAITH correlates with human judgments better than any existing metrics on the BEGIN benchmark (Dziri et al., 2022b). We also show that using PMI-DECODE as the objective generates more faithful responses than standard likelihood objective on three standard documentgrounded dialog datasets. We release our code 1 for further use by the research community.\n1 https://github.com/ynandwan/pmi-faith"
},
{
"figure_ref": [
"fig_1"
],
"heading": "Related Work",
"publication_ref": [
"b17",
"b1",
"b0",
"b29",
"b8",
"b2",
"b33",
"b8",
"b13",
"b12",
"b11",
"b15",
"b21",
"b28"
],
"table_ref": [],
"text": "In this work, we focus primarily on faithfulness aspect of the generated responses with respect to the grounding document. It is crucial to distinguish between faithfulness and hallucination (Maynez et al., 2020) in evaluating responses. A response is considered faithful only when all the information it contains can be verified or inferred from the grounded document. On the other hand, a response is considered as a hallucination if it provides false or fabricated information. It is important to note that there can be responses that are not hallucinations but are still unfaithful. In such cases, the information provided may not be false, but it cannot be verified using the grounded document as a reference. It is important to point out that the set of faithful responses is a subset of responses that are not hallucinations. In this section, we discuss related work in faithfulness, followed by a brief discussion on Mutual Information in conversational settings.\nResearchers have used various terms such as faithfulness (Cao et al., 2018), factual consistency (Cao et al., 2020;Santhanam et al., 2021), factual accuracy (Goodrich et al., 2019), fidelity (Chen et al., 2020), attribution (Rashkin et al., 2021a) and hallucination (i.e., the lack of faithfulness) (Xiao and Wang, 2021) to define and quantify faithfulness of a model's generated text to a given knowledge.\nMost of the works focusing on evaluating faithfulness propose to train a classifier for the task (Goodrich et al., 2019;Kryscinski et al., 2020;Dziri et al., 2022a). Whereas our proposed metric doesn't require any training and is agnostic to the underlying data. Recently, Honovich et al. (2021) proposed Q 2 for quantifying faithfulness. It uses a question generator to first generate questionanswer (QA) pairs from the generated response. Then a QA system is used to find an answer, to the generated question, from the document. Finally, an NLI system is used to compare the two answers. Though Q 2 uses the given document to check the faithfulness of a response, it ignores the dialog history. Thus, it may fail at handling responses that are non-self contained as depicted in Figure 1. Our metric PMI-FAITH addresses this issue.\nMany recent works (Dziri et al., 2022b;Honovich et al., 2022) have released different benchmarks that can be used to evaluate the performance of faithfulness metrics. While Honovich et al. aim to standardize benchmark datasets across different generation tasks, Dziri et al. focus on documentgrounded dialogues, and thus we use their benchmark to compare our metric with various baselines. Li et al. (2016) and Paranjape and Manning (2021) use mutual information to prevent the conversational models from generating generic responses (such as \"Sorry, I'm not sure about this topic\"). Contemporary to our work, Ren et al. (2023) propose to use a Conditional Pointwise Mutual Information (CPMI) based metric to evaluate relevance of a generated response with respect to a reference hypothesis for open-domain response generation. In contrast, our work is the first to use CPMI as a metric for evaluating the faithfulness of a response given a document and dialogue history in a document-grounded response generation."
},
{
"figure_ref": [],
"heading": "Background",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we first review the task of documentgrounded dialog response generation, followed by the definition of the faithfulness metric."
},
{
"figure_ref": [],
"heading": "Document Grounded Response Generation: Let",
"publication_ref": [],
"table_ref": [],
"text": "dialog history h = [u 1 , • • • u m ]\nbe a sequence of m utterances in the dialog so far and d be the document on which the next response is grounded. The task of document-grounded dialog response generation is to predict the next response, r = ⟨r 1 r 2 . . . r T ⟩, one token at a time, given the dialog history h and the document d. Here, ∀i, r i ∈ V, where V is the vocabulary of all possible tokens. The underlying model learns a probability distribution P(r|d, h) over all possible responses r ∈ V + , where V + is the space of all the sequences having one or more tokens from vocabulary V.\nTypically, this distribution is factorized over the tokens of r as:\nP(r|d, h) = T t=1 P(r t |d, h, r 1:t-1 )(1)\nFaithfulness Metric: Most of the existing definitions (and metrics) for faithfulness focus mainly on document d and response r but ignore the history h (Dziri et al., 2022a,b). This may be for the sake of uniformity across different tasks such as summarization, grounded dialogue generation, and paraphrase generation. We qualify the definition of faithfulness specifically for the task of documentgrounded dialogue generation. Formally, a response r is considered 'faithful' to a given document d and the dialogue history h iff d, h ⊨ r, where ⊨ represents logical entailment.\nA faithfulness metric should quantify the faithfulness of the response r to the document d and dialogue history h. In general, such a metric should take r, d and h as its input and compute a score, F (r, d, h) ∈ R, such that a higher value of F (r, d, h) indicates a more faithful response."
},
{
"figure_ref": [],
"heading": "Approach",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we describe our proposed metric for faithfulness -PMI-FAITH. We then propose a decoding strategy PMI-DECODE based on our metric, with the objective of generating relevant and faithful responses."
},
{
"figure_ref": [],
"heading": "PMI-FAITH",
"publication_ref": [
"b30",
"b24"
],
"table_ref": [],
"text": "PMI-FAITH is based on the information-theoretic concept of Pointwise Mutual Information. We use the notion of CPMI between generated response r and the document d given the context h to capture the influence of the document in generating the response. We define our metric, PMI-FAITH, for faithfulness of the response r to the document d as:\nPMI-FAITH(r, d, h) = CPMI(r; d|h) = log P (r, d|h) P (r|h)P (d|h) = log P (r|d, h) P (r|h)(2)\nMathematically, PMI is a measure of the strength of the association between two random events. A positive value of CPMI in eq. ( 2) implies that the probability of generating the response given the document and the dialogue history is higher than the probability of generating the response given only the dialogue history. Hence, the response is likely to be grounded in the document. On the other hand, if the response r is not faithful to the document d, the probability of its generation given the document and the dialogue history is likely to be similar to the probability of its generation without the document, resulting in a lower value of PMI-FAITH. We use pre-trained language models such as BLOOM (Scao et al., 2022) or GPT2 (Radford et al., 2019), to compute these conditional probabilities P (r|d, h) and P (r|h)."
},
{
"figure_ref": [],
"heading": "PMI-DECODE",
"publication_ref": [
"b10"
],
"table_ref": [],
"text": "PMI-DECODE is a decoding strategy whose objective is to generate responses that are both relevant and faithful. Typically, the goal of any decoding strategy is to select a response that has the maximum (log) likelihood:\nr = arg max r∈V + log P(r|d, h)(3)\nThe objective of PMI-DECODE is to select a response that is highly likely and faithful. This is achieved by maximizing a combination of likelihood and faithfulness quantified using an appropriate metric F . With α ∈ [0, 1], and a linear scoring function, we get:\nr = arg max r∈V + (1 -α) log P(r|d, h) + αF (r, d, h)(4)\nWith an auto-regressive model that generates the response one token at a time, we use decoding strategies, such as greedy decoding, beam search, nucleus sampling (Holtzman et al., 2020), or beam sampling as a heuristic to find the maxima. For ease of description, we use the greedy decoding below, though our approach is agnostic to the choice of heuristic for maximising the objective function. It just modifies the standard log-likelihood objective with an additional term corresponding to faithfulness. Our choice of PMI-FAITH as function F for quantification of faithfulness keeps the decoding heuristic tractable as shown below.\nWith eq. ( 3) as the objective, greedy decoding would sample the next token r t as follows:\nr t = arg max v∈V log P(r 1:t-1 , v|d, h) = arg max v∈V [log P(r 1:t-1 |d, h) + log P(v|d, h, r 1:t-1 )](5)\nIn eq. ( 5), the likelihood term has been factorized and notice that its first term is independent of the next token candidate v and thus can be dropped while taking arg max. Not all faithfulness metrics can be decomposed the same way as the likelihood term. One advantage of PMI-FAITH is that it can be decomposed the same way as likelihood as fol-lows:\nPMI-FAITH(r 1:t-1 , v, d, h) = log P(r 1:t-1 , v|d, h) P(r 1:t-1 , v|h) = log P(r 1:t-1 |d, h) P(r 1:t-1 |h) + log P(v|d, h, r 1:t-1 ) P(v|h, r 1:t-1 ) = PMI-FAITH(r 1:t-1 , d, h) + CPMI(v; d|h, r 1:t-1 )(6)\nBy using PMI-FAITH as F in eq. ( 4), and dropping the two terms which are independent of v from eq. ( 5) and eq. ( 6), the objective of greedy decoding using the PMI-DECODE objective is expressed as:\nr t = arg max v∈V (1 -α) log P(v|d, h, r 1:t-1 ) + αCPMI(v; d|h, r 1:t-1 ) (7)\nTo compute CPMI in eq. ( 7), the same language model can be used to get the conditional probabilities P(v|d, h, r 1:t-1 ) and P(v|h, r 1:t-1 ) by separately passing d,h, r 1:t-1 and h, r 1:t-1 , respectively, through the model.\nWe observed that using CPMI in the scoring function sometimes results in selecting tokens from the document which may interfere with the grammar. To mitigate this, instead of maximizing over the entire vocabulary V at each step t, we propose to maximize only over the 'top p' subset from the likelihood distribution, V p,t , defined as the minimum cardinality subset of tokens with the sum of their probabilities as p. We call this top p masking:\nr t = arg max v∈Vp,t (1 -α) log P(v|d, h, r 1:t-1 ) + αCPMI(v; d|h, r 1:t-1 ) (8)\nThe intuition here is that while CPMI has a positive influence on generating a more faithful response, it may negatively impact the grammatical structure. Therefore by restricting the vocabulary to V p,t , we use only highly probable tokens to form a response and thus are likely to generate responses that are faithful as well as grammatically correct."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [
"b16",
"b34",
"b31",
"b24"
],
"table_ref": [],
"text": "Our experiments answer two research questions:\n1. PMI-FAITH: How does our novel metric perform when compared to exisitng metrics on a standard benchmark (section 5.2)? 2002), and RougeL (Lin, 2004) to capture lexical overlap between d and generated response r; BERTScore (Zhang et al., 2020) to capture r's semantic similarity with d. We use the code2 provided by Honovich et al. ( 2021) for all the above baselines. 3 We also compare against faithcritic4 (Dziri et al., 2022a), which is a pre-trained classifier to predict faithfulness of a response.\nTraining Details: To measure PMI-FAITH, we need to compute two conditional probabilities: P(r|d, h), and P(r|h). To do so, we use pretrained LLMs available off the shelf from huggingface library (Wolf et al., 2019). To quantify the impact of using one language model over the other, we compute the performance of PMI-FAITH using eight LLMs of varying sizes: five BLOOM (Scao et al., 2022) models with up to 7 billion parameters, and three GPT2 (Radford et al., 2019) models up to GPT2-large (774 million). We observe a robust and consistent performance with a variability of only 0.02 points in the F1 score. Hence, for all further experiments, we use BLOOM-560m.\nUnconditional variant of PMI-Faith: To quantify the impact of dialogue history h on PMI-FAITH, we also use a variant of it called unconditional PMI between a response and a document, i.e., UPMI-FAITH = log P(r|d) -log P(r), to measure faithfulness."
},
{
"figure_ref": [],
"heading": "PMI-FAITH: Experimental Results",
"publication_ref": [],
"table_ref": [
"tab_0"
],
"text": "Table 1 reports the precision, recall, F1 score, and accuracy achieved by different metrics on the test split of BEGIN benchmark. We first observe that PMI-FAITH performs better than UPMI-FAITH, clearly demonstrating the advantage of using the … Football has become Brady 's religion , said Chopra , a filmmaker whose latest project is \" Tom vs. Time , \" a behind -the -scenes documentary series about the quarterback 's preparations for this past NFL season…\nPerhaps the documentary series \" Tom vs Time \" will shed some light on Toms remaining time ."
},
{
"figure_ref": [],
"heading": "Document",
"publication_ref": [],
"table_ref": [],
"text": "I guess it is about time, he has been coaching forever and has gotten older and has fewer years left to live. I think that is his plan .\nQ 2 Reasoning\nQuestion: What is it about?"
},
{
"figure_ref": [],
"heading": "Conversation:",
"publication_ref": [],
"table_ref": [
"tab_1"
],
"text": "Response Ans: time Document Ans: \" tom vs. time , \" a behind -the -scenes… dialogue history while measuring faithfulness.\nWe then observe that both UPMI-FAITH and PMI-FAITH perform better than all other faithfulness metrics by a considerable margin across all reported performance measures, with the absolute gains ranging from 21.8% to 8.7% in F1 score. Even against the strong baseline of Q 2 , PMI-FAITH achieves an absolute gain of 5.6% and 8.7% in accuracy and F1 score, respectively. As expected, all the lexical overlap and semantic similarity based metrics achieve poor performance, with accuracy worse than even the majority-class classifier's accuracy of 76.7%.\nNext, we notice that all metrics, except faithcritic, have higher recall than precision, indicating that they tend to be lenient while classifying a response as faithful, whereas faithcritic tends to be conservative and classifies most of the responses as not faithful. Comparing the next two best metrics, we observe that Q 2 has better F1 and recall but worse accuracy and precision than faithcritic.\nTo identify dataset specific biases, Table 2 reports F1 score separately for each of the three contributing datasets. We observe that PMI-FAITH achieves the highest F1 on CMU-DoG and Top-icalChat with more than 12% and 9.7% absolute gain, respectively, over the other metrics. Faith-Critic achieves the best F1 score on WoW, whereas its F1 on TopicalChat and CMU-DoG is quite low. This over-fitting on WoW is because faithcritic is a learned metric, and the training data for it has been adapted from WoW, and its low performance on the other two datasets demonstrates its lack of generalization.\nTo understand the correlation of various faithfulness metrics with human judgement, we report three calibration-free metrics in table 3. In all three metrics, we observe that PMI-Faith is better aligned to human judgements than the other measures of faithfulness. Subjective Analysis: The state-of-the-art metric, Q 2 , identifies whether a response is faithful or not using two steps. In the first step, it generates a set of questions based on the response. In the second step, it uses a question answering system to generate two responses for each question: one based on the response and one based on the document. If both the answers match, then the response is considered faithful to the document. We now discuss the shortcoming of Q 2 and how PMI-FAITH overcomes it using two examples.\nFigure 2 shows the two examples where PMI-FAITH correctly identifies the faithfulness (or lack of it) whereas the strongest baseline Q 2 fails to do so. In the case of 'Fully-attributable' response (right), the pronoun 'it' in the response is an anaphora, referring back to the antecedent 'home alone' (movie name), which is difficult to infer without the dialogue context. However, Q 2 doesn't take the dialogue history into account, and thus it considers the pronoun it in the response as a cataphor, referring to its postcedent 'comedy'. As a result, the QA system correctly answers the generated question 'What is it called?' with the postcedent 'comedy', when presented with the response, and correctly outputs its antecedent (home alone) when presented with the document. But the overall Q 2 system fails, as the two answers do not match. On the other hand, by virtue of considering dialogue history during computation, PMI-FAITH has information that the question is about the genre and not the movie name, and hence it can correctly classify the response as 'Fully-attributable'.\nThe other example highlights two issues: (1) when the response is partially hallucinated, question generation system may generate question from just the faithful part of the response and may incorrectly declare the whole response as faithful. In this example, most of the response contains an opinion, which is not faithful to the document, but the QG system focused on 'I guess it is about time'.\n(2): the other issue is that the NLI system fails to capture that the single word answer 'time' from the response is not entailed by the long answer from the document, resulting in the incorrect prediction by the overall system. On the other hand, PMI-FAITH considers the response as a whole, instead of separately focusing on parts of it. As a result, it is correctly able to identify the given response as not faithful."
},
{
"figure_ref": [],
"heading": "PMI-DECODE: Experimental Setup",
"publication_ref": [
"b6",
"b9",
"b3",
"b14"
],
"table_ref": [
"tab_5"
],
"text": "Datasets: We perform our experiments on three document-grounded dialog datasets: Multi- Doc2Dial (Feng et al., 2021), TopicalChat (Gopalakrishnan et al., 2019) and FaithDial (Dziri et al., 2022b). Each dialog in MultiDoc2Dial (MD2D) is between a user and an agent. Only the agent has access to the documents. So, we only model the agent responses for this dataset. TopicalChat (TC) consists of dialogs between two parties, where each party may have a different set of documents on the same topics. We use the 'rare' version of the dataset and filtered utterances tagged as 'personal knowledge'. FaithDial (FD), a faithful adaptation of WoW (Dinan et al., 2019), in which one participant can ask a wide range of questions and the other participant can only provide information from Wikipedia. Some statistics of the three datasets are in Table 6.\nAlgorithms: For each of the three datasets, we separately finetune a BART-Large (Lewis et al., 2019) 5 model using the code6 made available by Dziri et al. (2022a). As baselines, we use two decoding techniques that use the standard likelihood as the objective function: (1) beam search and (2) beam sampling. Both the techniques use a beam size of 4. We compare these baselines with their variants that uses our PMI-DECODE (PMI-D) objective. We use the values of α and top-p masking that achieved the highest sum of RougeL and normalized PMI-FAITH on the dev set. Table 7: Human evaluation of responses generated using beam search with different decoding objectives. We evaluate faithfulness (Fai), relevance (Rel) and grammar (Gra)."
},
{
"figure_ref": [],
"heading": "PMI-DECODE: Experimental Results",
"publication_ref": [
"b19",
"b18",
"b7"
],
"table_ref": [
"tab_3",
"tab_7"
],
"text": "We compare the decoding strategies in Table 4 using various automated metrics for faithfulness (PMI-FAITH, Q 2 ) and relevance (BLEU, RougeL). The general trend is that the PMI-DECODE generates more faithful responses compared to the standard variant and the improvement in faithfulness comes a the cost of relevance.\nTo gain a better understanding of the correlation between faithfulness and relevance, we conducted experiments involving different configurations of the control parameters of PMI-D (α and top-p masking). The results of these experiments are presented in the table 5, showcasing the corresponding normalized PMIF and RougeL metrics for each configuration.\nFor a given fixed value of α, as the p value increases, the faithfulness of the responses also increases. However, this improvement in faithfulness comes at the cost of decreased relevance. On the other hand, for a fixed p value, as α increases, the faithfulness initially increases and then gradually decreases. Simultaneously, the relevance decreases with an increase in α. The highest level of faithfulness is observed when α = 0.5 and p = 0.6. Meanwhile, the highest relevance is achieved when the CPMI is not used in the decoding objective (α = 0), or when both α and p have low values.\nWe have gained two valuable insights from table 5. The first insight reveals that there are specific configurations, such as p = 0.6 and α = 0.25, which achieve the same level of relevance (0.40) as the standard variant while generating more faithful responses. The second insight highlights a significant drop in relevance when solely focusing on PMI scores (α = 1 and p = 1). Therefore, if the goal is to generate responses with nearly equivalent relevance, utilising smaller values for α along with a masking value of 0.6 is recommended.\nTo demonstrate the impact of top-p masking on the grammar, we also report the percentage of grammatically incorrect responses for different values of α and top-p masking in table 5. We use GECToR (Omelianchuk et al., 2020), a grammatical error correction method, to find if a generated response is grammatically correct or not. We can easily see that for any value of α > 0, the grammatical errors reduce significantly with a reduction in top-p. We do not report any value for α = 1 and top -p = 1 as the responses with this configuration are not even in proper English. For example, one of the responses generated with α = 1 and top -p = 1 is \"Cheyne Lauren sisters Vel Lauren wear Ralph indo Austrian Ralph linesauxricting Ren therapies Combat Rarity glamorous\". Human Evaluation: We perform human evaluation experiments to compare (1) relevance, (2) faithfulness, and (3) grammar. All three dimensions were categorically labeled as agree, neutral, or disagree. We sampled 100 random (document, dialog history, response) tuples, 50 each from MD2D and TC. We evaluate the responses generated by beam search using two objectives: The results are summarized in Table 7. For each dimension, we report the percentage of responses that were rated agree. As expected, PMI-DECODE generates more faithful responses compared to greedy. We observe a 15% improvement in faithfulness compared to greedy decode on both datasets. Further, PMI-DECODE improves relevance on MD2D but slightly deteriorates on TC. Manual analysis revealed that the improvement of relevance on MD2D is primarily due to inherent solution multiplicity (Nandwani et al., 2021) in most dialogues, where more than one correct response is possible, but the metrics capture just one.\nAs PMI-DECODE maximises not just the likelihood of responses, but a combination of likelihood and faithfulness, we expected the responses to contain grammatical errors compared to greedy decode. To counter this issue, we proposed to use a weighted combination of likelihood and faithfulness during decode, with a higher weight on likelihood. We also restricted the vocabulary during each decode step to just the top-p subset. The human study shows that these mitigation techniques helped in reducing the grammatical mistakes made by PMI-DECODE. We see that the grammar is only slightly inferior to greedy on both the datasets.\nWe use Fleiss Kappa (Fleiss and Cohen, 1973) to measure the inter-annotator agreement, which is substantial for relevance (0.63) and faithfulness (0.63), and almost perfect (0.88) for grammar.\nSubjective Analysis: Table 8 presents an example from MD2D where standard likelihood based beam search decoding returns a generic response ('no info. found') which is present in around 1800 training samples. The same model returns the correct response when PMI-DECODE objective is used instead of just likelihood, demonstrating the capability of PMI-DECODE to shift the score in favour of the words present in the document."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "In this paper, we present a novel metric, PMI-FAITH, to measure faithfulness of responses generated by document grounded dialog systems. It uses conditional PMI between the response and the document given the dialog history to quantify faithfulness. We extend the idea of PMI-FAITHto propose a novel decoding objective, PMI-DECODE which encourages responses to be faithful to the given document by maximizing both the likelihood and faithfulness of the decoded response. Our experiments on the BEGIN benchmark prove that our proposed metric better correlates with human judgments compared to existing metrics. On three document-grounded dialog datasets, our novel decoding objective generates more faithful responses than the standard likelihood objective, as measured using automated metrics and a human study."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [],
"table_ref": [],
"text": "Though our decoding objective generates more faithful responses, we observed its inability to respond to generic chit-chat or pleasantries, like 'Hello!' or 'Good-bye'. It is possible to combine it with other techniques, like training with CTRL tokens (Rashkin et al., 2021b), which can enable it to generate both generic as well as faithful responses depending upon the dialogue context. But identifying when to generate a particular kind of response may require more insights and we leave this overall thread for future work. Next, to compute CPMI, we need to pass d, h, and h separately to the decoder. Though it can be done in parallel, but it may still reduce the throughput of the overall system by half. Finally, as demonstrated by the human evaluation, PMI-DECODE at times generates grammatically incorrect responses, even though the pre-trained language models are very good at generating fluent and coherent English. While we presented two knobs: α and top p masking to overcome this, we believe there could be other ways of handling this."
},
{
"figure_ref": [],
"heading": "Ethics Statement",
"publication_ref": [],
"table_ref": [],
"text": "Our work does not introduce any new ethical concerns per se, other than the ones already faced by large language models. Our decoding objective works on top of any trained language model and generates the text which is more faithful to a given input document. This can act as a double-edged sword: on one hand, if the document itself contains profanity, it may enhance the model's likelihood of generating similar content. But on the other hand, providing a valid document may also reduce the inherent likelihood of the model to generate profane content. Therefore, we recommend using it with responsibility and caution."
},
{
"figure_ref": [],
"heading": "A Human Evaluation",
"publication_ref": [],
"table_ref": [],
"text": "The screenshot of a sample task is shown in Figure 3.\nFor our human evaluation study, we ensured the quality and expertise of our annotators by selecting individuals who are fluent in English and have a solid foundation in Machine Learning (ML), and Natural Language Processing (NLP). Out of six inhouse annotators used, four were experts in dialog research and two were beginners. Each annotator had completed at least one formal course in ML/NLP. The exact qualifications and experience of the six annotators are given below:\n• Annotator-1 is a postgraduate degree holder with more than 2 decades of experience in NLP research.\n• Annotator-2 and annotator-3 are PhD degree holders with more than 5 years of experience in AI.\n• Annotator-4 is an undergraduate degree holder with more than 5 years of experience in NLP research.\n• Annotator-5 and annotator-6 are undergraduate degree holders with about 2 years of experience in AI research.\nAll the annotators provided their consent over an appropriate official communication channel, e.g., official email or slack channels.\nThe following were the instructions provided to the human evaluators. What is the task? There are 50 incomplete dialogs along with a document over which the dialog is grounded on. For each (document, incomplete dialog) pair we provide the next response predicted by 2 different dialog systems (shuffled in random order). You are requested to judge the response generated by these 2 systems along three dimensions: faithfulness, relevance and grammar. Each dimension has to be evaluated using the following scale: Agree (A), Neutral (N), and Disagree (D). How to judge relevance? Relevance measures how apt is the response given the dialog context and the knowledge. Please select agree when the response is apt and does not convey any incorrect information. Select neutral when it is hard to decide whether it is right or wrong and disagree otherwise. How to judge faithfulness? The faithfulness of a response is only dependent on the grounding document and it is independent of the dialog. A system response can be marked disagree for relevance and still be marked agree for faithfulness. Please select agree when the complete response can be inferred from the document. Select neutral when it is hard to decide whether it can be inferred from the document or not and disagree when a major portion of the response cannot be inferred from the document. For the case where the response is something like \"No information is present\". The judgement should be agree if there is no information about that in the document provided and \"disagree\" if there is information available in the document, but the system didn't pick it up. For cases where the user initiates a chit-chat (say the user says \"hi, how are you\"), the agent responds with chit-chat (\"I am doing good\"), please can mark faithfulness as neutral. How to judge grammar? The grammar score for a response is independent of the dialog or the document. A system response can be marked as disagree for relevance and still be marked agree for grammar. Please select agree when the response looks like how an expert human writes. Select neutral when there is a major issue with how the response reads but it still understandable and disagree when the response makes no sense."
},
{
"figure_ref": [],
"heading": "B Faithfulness Metric Normalization",
"publication_ref": [],
"table_ref": [],
"text": "The minimum and maximium values used for normalizing various metrics are shown in Table 9. These are the minimum and maximum values achieved by each metric on the dev set. We also report the threshold used on the normalized metrics for the faithfulness classification task. These are the thresholds than achieved the hightest F1 on the dev set. Table 9: . The min and max values used for normalizing each metric and the threshold used for the classifying faithfulness of a response using the metric."
}
] | A major concern in using deep learning based generative models for document-grounded dialogs is the potential generation of responses that are not faithful to the underlying document. Existing automated metrics used for evaluating the faithfulness of response with respect to the grounding document measure the degree of similarity between the generated response and the document's content. However, these automated metrics are far from being well aligned with human judgments. Therefore, to improve the measurement of faithfulness, we propose a new metric that utilizes (Conditional) Point-wise Mutual Information (PMI) between the generated response and the source document, conditioned on the dialogue. PMI quantifies the extent to which the document influences the generated response -with a higher PMI indicating a more faithful response. We build upon this idea to create a new decoding technique that incorporates PMI into the response generation process to predict more faithful responses. Our experiments on the BEGIN benchmark demonstrate an improved correlation of our metric with human evaluation. We also show that our decoding technique is effective in generating more faithful responses when compared to standard decoding techniques on a set of publicly available document-grounded dialog datasets. | Pointwise Mutual Information Based Metric and Decoding Strategy for Faithful Generation in Document Grounded Dialogs | [
{
"figure_caption": "Creating a free my Social Security account takes less than 10 minutes, lets you set up or change your direct deposit and gives you access to many other online services.Hi, is the social security account free of charge?",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 1 :1Figure 1: An example document grounded dialog with two types of responses: sentential response (R2) and non-sentential response (R1).",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "The min and the max scores are identified from the dev set. We then identify an optimum threshold for each metric as the one that achieves the best F1 on the dev set. Finally, during test, we use the identified min, max and thresholds to classify a response as faithful. The thresholds, min and max for each metric are reported in Appendix B. Once we have the predicted class, we then compute precision, recall, F1 score and accuracy achieved by each metric. As done inHonovich et al. (2021), we also report calibration-free metrics that don't require any normalization. In addition to Spearman's and Pearson's correlation with human annotations, we also report AUROC for various faithfulness metrics. Performance of various faithfulness metrics on the BEGIN Benchmark.",
"figure_data": "2. PMI-DECODE: Does our proposed decodingMetricPrecision RecallF1Accuracytechnique generate responses that are moreU-F10.4010.7850.5310.677faithful compared to vanilla decoding tech-BLEU0.4780.4790.4790.757niques, while still maintaining relevance (sec-tion 5.4).?RougeL BERTScore FaithCritic0.487 0.459 0.6840.552 0.673 0.4920.518 0.546 0.5730.760 0.739 0.829Q 20.5170.7440.6100.7795.1 PMI-FAITH: Experimental SetupUPMI-FAITH0.5920.7040.6430.818PMI-FAITH0.6070.8180.6970.834Dataset: We experiment using recently proposedBEGIN benchmark (Dziri et al., 2022b) for eval-uating the ability of PMI-FAITH to identify faith-ful responses. This benchmark uses three docu-ment grounded datasets, viz, CMU-DoG (Zhouet al., 2018), TopicalChat (Gopalakrishnan et al.,2019), and WoW (Dinan et al., 2019). It contains11, 059 responses generated by three different mod-els, GPT2 (Radford et al., 2019), DoHA (Prabhu-moye et al., 2021) and T5 (Raffel et al., 2020),on randomly selected samples from test splits ofthe 3 datasets. Each generated response is anno-tated by humans and classified into either 'Fully-",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": "Q 2 Reasoning",
"figure_id": "tab_1",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Spearman and Pearson correlation with human annotations in the BEGIN Benchmark, and AUROC of various faithfulness metrics.",
"figure_data": "",
"figure_id": "tab_2",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Faithfulness and relevance metrics computed for various decoding techniques on three datasets.",
"figure_data": "Decode MethodObj.Faithfulness PMI-F Q2Relevance BLEU RougeLMultiDoc2DialBeamStand.0.590.6330.560.488SearchPMI-D0.640.6528.950.473BeamStand.0.580.6230.500.491SamplingPMI-D0.630.6630.690.488TopicalChatBeamStand.0.490.686.630.219SearchPMI-D0.570.735.650.197BeamStand.0.490.676.280.214SamplingPMI-D0.540.726.020.207FaithDialBeamStand.0.540.8313.530.404SearchPMI-D0.630.8712.380.389BeamStand.0.520.8213.300.398SamplingPMI-D0.610.8712.380.388",
"figure_id": "tab_3",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Faithfulness, relevance, and grammatical errors of the responses generated by beam sampling using different configurations of α and top -p masking on the FaithDial dataset .",
"figure_data": "PMIFRougeLGrammatical errors (in %)pα00.25 0.5100.25 0.510 0.25 0.510.60.52 0.58 0.59 0.59 0.40 0.40 0.39 0.38 6.9 10.2 11.812.20.750.52 0.60 0.61 0.60 0.40 0.39 0.38 0.36 6.4 12.0 16.318.60.90.52 0.61 0.62 0.57 0.40 0.39 0.36 0.30 7.5 15.0 24.245.810.52 0.61 0.56 0.34 0.40 0.38 0.30 0.05 7.4 16.5 54.8-Num. samplesAvg. wordsTrainDev.Test Doc. Hist. Resp.MD2D24,603 4,699 4,5671669318TC131,555 8,183 8,30124119920FD18,357 3,417 3,539236918",
"figure_id": "tab_4",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "Various statistics of the three document grounded dialog datasets.",
"figure_data": "Multi-Doc2DialTopical ChatFaiRelGraFaiRelGraStandard 0.52 0.72 0.96 0.69 0.70 0.96PMI-D0.60 0.75 0.92 0.80 0.67 0.93",
"figure_id": "tab_5",
"figure_label": "6",
"figure_type": "table"
},
{
"figure_caption": "Document What happens if I had a lapse of insurance coverage during active duty? You are required to maintain continuous coverage or surrender your plates before deployed. If you were unable to do so , DMV has a procedure in place to exempt you from administrative actions upon your return. You must provide a notarized statement that the vehicle was not used during the time in question, a copy of your military ID, proof of current insurance or surrender of plates, and signed and dated copies of deployment and return papers, or DD-214 if separated from service... You are required to maintain continuous coverage or surrender your plates before being deployed.Were you unable to keep your insurance while on active duty? User: yes, it just wasn't on my mind and I don't get notices and anything like that while deployed DMV has a procedure in place to exempt you from administrative actions upon your return. You must provide a notarized statement that the vehicle was not used during the time in question, a copy of your military ID , proof of current insurance or surrender of plates, and signed and dated copies of deployment and return papers , or DD-214 if separated from service.",
"figure_data": "ContextUser: lost my insurance while on active duty and have some questionsAgent: Greedy Agent: Unfortunately, no relevant information is found.PMI:Agent:",
"figure_id": "tab_6",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "An example from the test set of Multi-Doc2Dial dataset where Greedy generates a 'I don't know' response and PMI-DECODE generates a relevant and faithful response.",
"figure_data": "standard and PMI-DECODE. Out of six in-houseannotators used (3 per dataset), four were expertsin dialog research and two were beginners. Referto appendix A for more details.",
"figure_id": "tab_7",
"figure_label": "8",
"figure_type": "table"
}
] | Yatin Nandwani; Vineet Kumar; Dinesh Raghu; Sachindra Joshi; Luis A Lastras | [
{
"authors": "Meng Cao; Yue Dong; Jiapeng Wu; Jackie Chi; Kit Cheung",
"journal": "Association for Computational Linguistics",
"ref_id": "b0",
"title": "Factual error correction for abstractive summarization models",
"year": "2020-11-16"
},
{
"authors": "Ziqiang Cao; Furu Wei; Wenjie Li; Sujian Li",
"journal": "AAAI Press",
"ref_id": "b1",
"title": "Faithful to the original: Fact aware neural abstractive summarization",
"year": "2018-02-02"
},
{
"authors": "Zhiyu Chen; Wenhu Chen; Hanwen Zha; Xiyou Zhou; Yunkai Zhang; Sairam Sundaresan; William Yang; Wang ",
"journal": "Association for Computational Linguistics",
"ref_id": "b2",
"title": "Logic2text: High-fidelity natural language generation from logical forms",
"year": "2020-11"
},
{
"authors": "Emily Dinan; Stephen Roller; Kurt Shuster; Angela Fan; Michael Auli; Jason Weston",
"journal": "",
"ref_id": "b3",
"title": "Wizard of wikipedia: Knowledge-powered conversational agents",
"year": "2019-05-06"
},
{
"authors": "Nouha Dziri; Ehsan Kamalloo; Sivan Milton; Osmar Zaiane; Mo Yu; Edoardo M Ponti; Siva Reddy",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b4",
"title": "a. FaithDial: A Faithful Benchmark for Information-Seeking Dialogue",
"year": "2022"
},
{
"authors": "Nouha Dziri; Hannah Rashkin; Tal Linzen; David Reitter",
"journal": "Trans. Assoc. Comput. Linguistics",
"ref_id": "b5",
"title": "Evaluating attribution in dialogue systems: The BEGIN benchmark",
"year": "2022"
},
{
"authors": "Song Feng; Sankalp Siva; Hui Patel; Sachindra Wan; Joshi",
"journal": "Association for Computational Linguistics",
"ref_id": "b6",
"title": "Multidoc2dial: Modeling dialogues grounded in multiple documents",
"year": "2021-07-11"
},
{
"authors": "Joseph L Fleiss; Jacob Cohen",
"journal": "Educational and Psychological Measurement",
"ref_id": "b7",
"title": "The equivalence of weighted kappa and the intraclass correlation coefficient as measures of reliability",
"year": "1973"
},
{
"authors": "Ben Goodrich; Vinay Rao; Peter J Liu; Mohammad Saleh",
"journal": "ACM",
"ref_id": "b8",
"title": "Assessing the factual accuracy of generated text",
"year": "2019-08-04"
},
{
"authors": "Karthik Gopalakrishnan; Behnam Hedayatnia; Qinlang Chen; Anna Gottardi; Sanjeev Kwatra; Anu Venkatesh; Raefer Gabriel; Dilek Hakkani-Tür",
"journal": "",
"ref_id": "b9",
"title": "Topical-Chat: Towards Knowledge-Grounded Open-Domain Conversations",
"year": "2019"
},
{
"authors": "Ari Holtzman; Jan Buys; Li Du; Maxwell Forbes; Yejin Choi",
"journal": "",
"ref_id": "b10",
"title": "The curious case of neural text degeneration",
"year": "2020-04-26"
},
{
"authors": "Or Honovich; Roee Aharoni; Jonathan Herzig; Hagai Taitelbaum; Doron Kukliansy; Vered Cohen; Thomas Scialom; Idan Szpektor; Avinatan Hassidim; Yossi Matias",
"journal": "Association for Computational Linguistics",
"ref_id": "b11",
"title": "TRUE: re-evaluating factual consistency evaluation",
"year": "2022-05-26"
},
{
"authors": "Or Honovich; Leshem Choshen; Roee Aharoni; Ella Neeman; Idan Szpektor; Omri Abend",
"journal": "Association for Computational Linguistics",
"ref_id": "b12",
"title": "qˆ2$: Evaluating factual consistency in knowledgegrounded dialogues via question generation and question answering",
"year": "2021-07-11"
},
{
"authors": "Wojciech Kryscinski; Bryan Mccann; Caiming Xiong; Richard Socher",
"journal": "Association for Computational Linguistics",
"ref_id": "b13",
"title": "Evaluating the factual consistency of abstractive text summarization",
"year": "2020"
},
{
"authors": "Mike Lewis; Yinhan Liu; Naman Goyal; Marjan Ghazvininejad; Abdelrahman Mohamed; Omer Levy; Veselin Stoyanov; Luke Zettlemoyer",
"journal": "",
"ref_id": "b14",
"title": "BART: denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension",
"year": "2019"
},
{
"authors": "Jiwei Li; Michel Galley; Chris Brockett; Jianfeng Gao; Bill Dolan",
"journal": "Association for Computational Linguistics",
"ref_id": "b15",
"title": "A diversity-promoting objective function for neural conversation models",
"year": "2016"
},
{
"authors": "Chin-Yew Lin",
"journal": "Association for Computational Linguistics",
"ref_id": "b16",
"title": "ROUGE: A package for automatic evaluation of summaries",
"year": "2004"
},
{
"authors": "Joshua Maynez; Shashi Narayan; Bernd Bohnet; Ryan T Mcdonald",
"journal": "",
"ref_id": "b17",
"title": "On faithfulness and factuality in abstractive summarization",
"year": "2020"
},
{
"authors": "Yatin Nandwani; Deepanshu Jindal; Mausam ; Parag Singla",
"journal": "",
"ref_id": "b18",
"title": "Neural learning of one-of-many solutions for combinatorial problems in structured output spaces",
"year": "2021-05-03"
},
{
"authors": "Kostiantyn Omelianchuk; Vitaliy Atrasevych; Artem Chernodub; Oleksandr Skurzhanskyi",
"journal": "Association for Computational Linguistics",
"ref_id": "b19",
"title": "GECToR -grammatical error correction: Tag, not rewrite",
"year": "2020"
},
{
"authors": "Kishore Papineni; Salim Roukos; Todd Ward; Wei-Jing Zhu",
"journal": "Association for Computational Linguistics",
"ref_id": "b20",
"title": "Bleu: a method for automatic evaluation of machine translation",
"year": "2002"
},
{
"authors": "Ashwin Paranjape; Christopher Manning",
"journal": "Association for Computational Linguistics",
"ref_id": "b21",
"title": "Human-like informative conversations: Better acknowledgements using conditional mutual information",
"year": "2021"
},
{
"authors": "Matt Post",
"journal": "Association for Computational Linguistics",
"ref_id": "b22",
"title": "A call for clarity in reporting BLEU scores",
"year": "2018"
},
{
"authors": "Kazuma Shrimai Prabhumoye; Yingbo Hashimoto; Alan W Zhou; Ruslan Black; Salakhutdinov",
"journal": "Association for Computational Linguistics",
"ref_id": "b23",
"title": "Focused attention improves documentgrounded generation",
"year": "2021-06-06"
},
{
"authors": "Alec Radford; Jeff Wu; Rewon Child; David Luan; Dario Amodei; Ilya Sutskever",
"journal": "",
"ref_id": "b24",
"title": "Language models are unsupervised multitask learners",
"year": "2019"
},
{
"authors": "Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu",
"journal": "Journal of Machine Learning Research",
"ref_id": "b25",
"title": "Exploring the limits of transfer learning with a unified text-to-text transformer",
"year": "2020"
},
{
"authors": "Vitaly Hannah Rashkin; Matthew Nikolaev; Michael Lamm; Dipanjan Collins; Slav Das; Gaurav Petrov; Iulia Singh Tomar; David Turc; Reitter",
"journal": "",
"ref_id": "b26",
"title": "Measuring attribution in natural language generation models",
"year": "2021"
},
{
"authors": "David Hannah Rashkin; Gaurav Reitter; Dipanjan Singh Tomar; Das",
"journal": "Association for Computational Linguistics",
"ref_id": "b27",
"title": "Increasing faithfulness in knowledge-grounded dialogue with controllable features",
"year": "2021-08-01"
},
{
"authors": "Liliang Ren; Mankeerat Sidhu; Qi Zeng; Revanth Gangi Reddy; Heng Ji; Chengxiang Zhai",
"journal": "Association for Computational Linguistics",
"ref_id": "b28",
"title": "C-PMI: Conditional pointwise mutual information for turn-level dialogue evaluation",
"year": "2023"
},
{
"authors": "Sashank Santhanam; Behnam Hedayatnia; Spandana Gella; Aishwarya Padmakumar; Seokhwan Kim; Yang Liu; Dilek Hakkani-Tur",
"journal": "",
"ref_id": "b29",
"title": "Rome was built in 1776: A case study on factual correctness in knowledge-grounded response generation",
"year": "2021"
},
{
"authors": "Le Teven; Angela Scao; Christopher Fan; Akiki",
"journal": "",
"ref_id": "b30",
"title": "Bloom: A 176b-parameter open-access multilingual language model",
"year": "2022"
},
{
"authors": "Thomas Wolf; Lysandre Debut; Victor Sanh; Julien Chaumond; Clement Delangue; Anthony Moi; Pierric Cistac; Tim Rault; Rémi Louf; Morgan Funtowicz; Jamie Brew",
"journal": "",
"ref_id": "b31",
"title": "Huggingface's transformers: State-of-the-art natural language processing",
"year": "2019"
},
{
"authors": "Zeqiu Wu; Michel Galley; Chris Brockett; Yizhe Zhang; Xiang Gao; Chris Quirk; Rik Koncel-Kedziorski; Jianfeng Gao; Hannaneh Hajishirzi; Mari Ostendorf; Bill Dolan",
"journal": "AAAI Press",
"ref_id": "b32",
"title": "A controllable model of grounded response generation",
"year": "2021-02-02"
},
{
"authors": "Yijun Xiao; William Yang; Wang ",
"journal": "Association for Computational Linguistics",
"ref_id": "b33",
"title": "On hallucination and predictive uncertainty in conditional language generation",
"year": "2021-04-19"
},
{
"authors": "Tianyi Zhang; Varsha Kishore; Felix Wu; Kilian Q Weinberger; Yoav Artzi",
"journal": "",
"ref_id": "b34",
"title": "Bertscore: Evaluating text generation with BERT",
"year": "2020-04-26"
},
{
"authors": "Kangyan Zhou; Shrimai Prabhumoye; Alan W Black",
"journal": "Association for Computational Linguistics",
"ref_id": "b35",
"title": "A dataset for document grounded conversations",
"year": "2018-10-31"
}
] | [
{
"formula_coordinates": [
3,
70.87,
384.41,
139.22,
10.71
],
"formula_id": "formula_0",
"formula_text": "dialog history h = [u 1 , • • • u m ]"
},
{
"formula_coordinates": [
3,
106.86,
585.67,
183,
33.58
],
"formula_id": "formula_1",
"formula_text": "P(r|d, h) = T t=1 P(r t |d, h, r 1:t-1 )(1)"
},
{
"formula_coordinates": [
3,
330.22,
416.52,
194.92,
40.74
],
"formula_id": "formula_2",
"formula_text": "PMI-FAITH(r, d, h) = CPMI(r; d|h) = log P (r, d|h) P (r|h)P (d|h) = log P (r|d, h) P (r|h)(2)"
},
{
"formula_coordinates": [
4,
122.3,
130.02,
167.57,
18.81
],
"formula_id": "formula_3",
"formula_text": "r = arg max r∈V + log P(r|d, h)(3)"
},
{
"formula_coordinates": [
4,
73.13,
288.59,
216.74,
29.9
],
"formula_id": "formula_4",
"formula_text": "r = arg max r∈V + (1 -α) log P(r|d, h) + αF (r, d, h)(4)"
},
{
"formula_coordinates": [
4,
103.82,
585.24,
186.04,
59.61
],
"formula_id": "formula_5",
"formula_text": "r t = arg max v∈V log P(r 1:t-1 , v|d, h) = arg max v∈V [log P(r 1:t-1 |d, h) + log P(v|d, h, r 1:t-1 )](5)"
},
{
"formula_coordinates": [
4,
331.57,
94.65,
193.57,
134.05
],
"formula_id": "formula_6",
"formula_text": "PMI-FAITH(r 1:t-1 , v, d, h) = log P(r 1:t-1 , v|d, h) P(r 1:t-1 , v|h) = log P(r 1:t-1 |d, h) P(r 1:t-1 |h) + log P(v|d, h, r 1:t-1 ) P(v|h, r 1:t-1 ) = PMI-FAITH(r 1:t-1 , d, h) + CPMI(v; d|h, r 1:t-1 )(6)"
},
{
"formula_coordinates": [
4,
316.14,
303.53,
209,
35.16
],
"formula_id": "formula_7",
"formula_text": "r t = arg max v∈V (1 -α) log P(v|d, h, r 1:t-1 ) + αCPMI(v; d|h, r 1:t-1 ) (7)"
},
{
"formula_coordinates": [
4,
316.14,
548.94,
209,
36.81
],
"formula_id": "formula_8",
"formula_text": "r t = arg max v∈Vp,t (1 -α) log P(v|d, h, r 1:t-1 ) + αCPMI(v; d|h, r 1:t-1 ) (8)"
},
{
"formula_coordinates": [
6,
84.37,
190.56,
46.49,
8.42
],
"formula_id": "formula_9",
"formula_text": "Q 2 Reasoning"
}
] |
2023-05-20 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b0",
"b1",
"b2",
"b3",
"b4",
"b5",
"b27",
"b28",
"b29",
"b30",
"b31",
"b32",
"b33",
"b34",
"b35",
"b36",
"b37",
"b38",
"b39",
"b40",
"b41",
"b42",
"b43"
],
"table_ref": [],
"text": "Artificial intelligence (AI) has the potential to revolutionize the educational system. According to Chassignol et al. [1], four areas-customized educational content, cutting-edge teaching strategies, technology-enhanced evaluation, and communication between students and teachers-are where AI can revolutionize the educational environment. An overview of AI applications in higher education has been offered by Zawacki-Richter et al. [2], spanning profiling and prediction, evaluation and assessment, adaptive systems and personalization, and intelligent tutoring systems. Potential research subjects in AI applications for education have been suggested by Hwang et al. [3]. In order to enable effective administrative operations, content modification, and enhanced learning quality, Chen et al. [4] have concentrated on the use of AI in administration, instruction, and learning. The potential of generative AI in education to lessen workload and increase learner engagement in online learning has been highlighted by Dao et al. [5]. Finally, Nguyen et al. [6] have suggested a platform for online learning that incorporates a Vietnamese virtual assistant to help teachers present lectures to students and to make editing simple without the requirement for video recording.\nLLMs have difficulties when dealing with a variety of datasets that cover many areas. CoQA [28] poses a unique challenge for big language models due to the conversational character of the questions and the responses, which can be free-form text and include texts from seven different domains. PILE [29] is a large dataset of approximately 800GB of text from many sources, including as books, online pages, and scientific publications. ScienceQA [30] is a great tool for creating machine comprehension models for scientific domains because it comprises a variety of natural science, language science, and social science.\nFor question answering systems, there are numerous datasets available, each concentrating on a distinct subject. MATH [31] contains 12,500 difficult competition math problems with detailed solutions that allow models to produce answer derivations and justifications. GSM-8K [32] focuses on grade-school mathematics and covers a range of mathematical topics. Questions about biomedical research and medical scientific papers are included in BioASQ [33] and are categorized by level of difficulty. TQA [34] combines the machine comprehension and visual question-answering paradigms for middle school science classes. SWAG [35] challenges the grounded commonsense inference, combining natural language inference and physically grounded reasoning. PIQA [36] was developed as a commonsense reasoning dataset to examine the physical knowledge of current NLP models. PROST [37] is intended to test both causal and masked language models in a zero-shot environment. JEC-QA [38] and CaseHOLD [39] are Chinese legal datasets.\nIn the discipline of NLP, large datasets are necessary for the development and assessment of machine learning models. The internet is a source of data for many question-answering datasets, particularly for websites like Wikipedia and search engines like Google. WebQuestions [40] are all defined as Freebase entities, with Freebase serving as the knowledge base. Each question in WikiQA [41] links to a possible related Wikipedia page, and lines from the summary part of the page are utilized as candidate answers. TriviaQA [42] consists of 950K question-answer pairs drawn from 662K publications on the web and in Wikipedia. Because the context for each question is quite lengthy, span prediction may not be able to reliably produce the answers. One million pairs of questions and passages drawn from actual search queries are provided by the MS MARCO [43], which is updated on a regular basis with fresh search queries. Real-world, user-generated queries from Google.com and related Wikipedia pages are included in the NQ dataset [44]. Although there may be potential mistakes and incompleteness of information presented, the accuracy and completeness of the Wikipedia pages determine how accurate and thorough the responses are in these datasets. These datasets offer researchers useful tools for creating and enhancing machine learning models for problem-solving, with a variety of difficulties and chances for advancement in the field.\nOverall, these datasets provide valuable resources for evaluating LLMs in various tasks such as question answering, language modeling, text generation, reading comprehension, among others."
},
{
"figure_ref": [],
"heading": "Datasets from the exams for training large language models",
"publication_ref": [
"b44",
"b45",
"b46",
"b47",
"b48"
],
"table_ref": [],
"text": "LLMs are increasingly being used, hence it is critical to assess their dependability and performance. Due to the richness and diversity of language usage in these datasets, language model evaluation using test datasets have acquired significance. Due to the high cost of data generation by human experts, existing exam datasets like the NTCIR QA Lab [45], Entrance Exams task at CLEF QA Track [46], [47], and AI2 Elementary School Science Questions dataset [48], have not been adequate for training advanced data-driven machine reading models. As a result, larger and more varied exam datasets are essential for LLMs training and evaluation. RACE [49] is one such dataset that has drawn interest. RACE is a dataset for automated reading comprehension with RACE-M and RACE-H, two subgroups from middle school and high school tests, respectively.\nExam datasets are increasingly being used to evaluate LLMs, and the current datasets present interesting evaluation issues. The creation of novel test datasets, like the proposed Vietnamese High School Graduation Examination Dataset for LLMs, can improve the assessment of LLMs and guarantee their dependability in a variety of contexts. Using test datasets offers a demanding and varied evaluation of LLMs, which is essential for their usage in real-world applications. The creation of fresh test datasets can improve the evaluation procedure and increase the dependability of LLMs across a range of applications."
},
{
"figure_ref": [],
"heading": "Datasets from high school exams for training large language models",
"publication_ref": [
"b49",
"b50",
"b51"
],
"table_ref": [],
"text": "Despite the fact that there are few datasets that concentrate on using high school topic exams to assess LLMs, there are still some datasets that contain high school exam questions that can be utilized for this purpose. GeoS [50] intended for automatic math problem-solving. It includes SAT plane geometry questions from prior real SAT examinations and practice tests, each with a diagram and multiple-choice answers. Another dataset that includes multiple-choice questions from academic exams that range from grade 3 to grade 9 and need reasoning to answer is ARC [51]. The dataset was split into two parts, Easy and Challenge, with the latter comprising trickier problems. A supporting knowledge library of 14.3 million unstructured text passages is also included. SuperGLUE [52], a more difficult dataset with tasks involving intricate thinking and common sense, contains many different jobs in it, some of which need you to respond to questions based on science passages from high school.\nThese high school datasets can still be utilized to assess language models' capacities to perceive and analyze natural language, despite the fact that there are few datasets explicitly created for testing LLMs using high school subject exams. Researchers can gain a deeper understanding of language models' strengths and limitations and create ways to enhance their performance by evaluating them against high school-level content. So that they can be used to assess LLMs, these datasets offer a variety of tasks and subject areas that are pertinent to high school education."
},
{
"figure_ref": [],
"heading": "Our proposed dataset",
"publication_ref": [],
"table_ref": [
"tab_0",
"tab_1"
],
"text": "To begin with, we conducted a search for available datasets in the \"texts\" category that are relevant to question answering task, as well as datasets that support the Vietnamese language. Our search was carried out on Paperwithcode as well as in previous studies. Table 1 displays the available datasets. We found that the majority of datasets consist of English texts, with only a few supporting Vietnamese. The most popular subjects are English, mathematics, and physics, while other subjects have relatively fewer related datasets (see Appendix section A for further details).\nTable 1: Related datasets Subjects Dataset application (difficult), and high application (extremely tough) are the four levels of complexity. We may learn more about LLM's capabilities for complicated reasoning as well as its strengths and shortcomings in dealing with various high school levels by evaluating its performance over a range of difficulty levels. The exam's three primary subjects-mathematics, literature, and English-as well as two combinations-the natural science combination of physics, chemistry, and biology, and the social science combination of history, geography, and civic education-make up the exam's framework.\nTable 2 displays the multiple choice question subjects. Each exam contains 40 questions in each of the other topics in addition to 50 questions in mathematics and English. The dataset encompasses a wide range of disciplines and calls for a variety of abilities, from arithmetic to sophisticated reasoning. 1884-1914, 1919-1930, 1930-1945, 1945-1954, 1954-1975, and 1975-2000 periods.\nGeography geographical skills: atlas use, data table interpretation, and chart analysis; geographical theory: natural geography, population geography, economic sector geography, economic zone geography, sea geography, and island geography.\nCivic Education legal frameworks and regulations, fundamental rights of citizens, democratic principles and concepts, as well as case studies A systematic assessment technique called a literature dataset is used to assess a student's reading and writing abilities.\nReading comprehension is tested in Part I, while writing skills are tested in Part II. Four questions in Part I ask students to examine and interpret an essay or poem, including determining the genre and any words or phrases that have particular meanings. Their own view on the text must be expressed in the final question, or it must be evaluated. Two essay questions are included in Part II, one on how to write a social argumentative essay and the other on how to write a literary argumentative essay. The essay questions test a student's ability to create a coherent and concise argument, back it with evidence, and analyze and interpret literary materials in order to develop a well-supported argument. The literature dataset offers a thorough assessment of a student's writing and reading comprehension abilities.\nThe score distribution is an indicator to show how candidates scored in exams. Every year, VMET publishes the score distribution, which is shown as a chart for each subject. The distribution of scores is used to evaluate the competency of candidates and to assess exams according to their degree of difficulty, so assessing the level of competency of the applicants. Score distributions from 2019 to 2022 were gathered. We can assess the capability of LLMs by contrasting their outcomes with those of Vietnamese students (see Appendix section D for a detailed breakdown of the score distribution and a comparison of LLMs' performance). The average score (AVS) and most reached score (MVS) of the Vietnamese students are presented in Table 3 for a simpler comparison of the LLMs' performance. For instance, in 2019 the AVS and MVS for mathematics are 5.64 and 6.4, respectively. Any research project must start with the gathering of raw data, and for this study, we obtained our data from free public websites in Vietnam. We painstakingly selected and arranged the gathered information into a brand-new dataset of questions from VNHSGE and similar exams. We specifically used the illustrated exam questions that VMET publishes every year. To give students and teachers a general idea of the content and structure of the official exam, these exam questions are made available to them. We gathered the official exam questions from VMET in addition to the illustrated exam questions. VMET produced a brief answer key following the exam, and the teachers then supplied more thorough responses. Additionally, we have included similar exam questions that are created by instructors and high schools around Vietnam in our data collection. This strategy guarantees that our dataset has a wide variety of questions that cover a wide range of subjects and degrees of difficulty. Our dataset contains exam questions, answers, and thorough step-by-step explanations (see Appendix section B.1 for a raw data example) that have all been meticulously examined and validated by our team of subject matter experts. Instead of employing Amazon Mechanical Turk, as some earlier datasets did, detailed explanations are given by qualified teachers.\nThe extensive dataset gathered for this study offers a great chance to assess how well LLMs complete Vietnamese national tests. Our dataset's vast variety of themes and levels of difficulty provide a thorough assessment of the LLMs' accuracy and deductive reasoning abilities when responding to various questions. We may learn important lessons about the benefits and drawbacks of LLMs in handling actual tests by utilizing this dataset, which can guide further study and advancement in this area."
},
{
"figure_ref": [],
"heading": "Dataset",
"publication_ref": [
"b11",
"b17"
],
"table_ref": [],
"text": "The dataset is available in Word format and JSON format. In addition, we provide the dataset in Vietnamese and English (VNHGE-V and VNHSGE-E). The dataset was originally written in Vietnamese. Using GPT-4/ChatGPT, the dataset is translated into English, similar to how OpenAI tests the capability of GPT-4 [12] in other languages by using Azure Translate to translate the MMLU benchmark [18] into another language. Language models can handle several languages, as is well recognized. However, if LLMs do not support multilingualism they can use the English VNHSE version. We may also employ comparable strategies for additional languages by using GPT-4/ChatGPT, BingChat/Azure Translate, and Google Translate."
},
{
"figure_ref": [],
"heading": "Format",
"publication_ref": [],
"table_ref": [],
"text": "In the VNHSGE dataset, we convert formulas, equations, tables, images, and charts from raw text formats like Word, Pdf, and HTML into a text-only format and an image folder including steps: (1) collecting raw data and convert them into Word format, (2) transforming symbols, formulas, and equations into Latex format, (3) converting Word format to JSON format (see Appendix section B for more details of a step-by-step conversion)."
},
{
"figure_ref": [],
"heading": "Word format",
"publication_ref": [],
"table_ref": [],
"text": "We transform the symbols, equations, and formulas into text using the Latex format so that it is compatible with LLMs transformed BERT or GPT. For those who lack programming skills, we also offer a text format in the form of a Word file for evaluating the performance of LLMs. In this situation, the VNHSGE dataset can be thought of as a question bank for assessing LLMs over a range of subjects. However, full language models like ChatGPT and BingChat are typically more appropriate in this situation. It is vital to keep in mind that symbols, formulas, and equations were converted to text format while utilizing a text format in a Word file; we only ask questions of LLMs and receive responses.\nQuestion: Let $y=f(x)$ be a cubic function with the graph shown in the picture. \n-2 2 -1 1 2 x y\nSetting $t=x^3-3x$, we have $|f(x^3-3x)|=\\frac{2}{3} \\Leftrightarrow |f(t)|=\\frac{2}{3}$. From the above graph, we conclude that the equation $|f(t)|=\\frac{2}{3}$ has six distinct solutions $t=t_{i}$ (with $i=\\overline{1,6}$ and $(t_{1}<-2; -2<t_{2}, t_{3}<2; t_{4}, t_{5}, t_{6}>2)$.\nConsidering the function $t(x)=x^{3}-3x$, we have $t^{\\prime}(x)=3 x^{2}-3 ; t^{\\prime}(x)=0 \\Leftrightarrow x= \\pm1$. The sign variation table of $t(x)$ is:\nx f (x) f (x) -∞ -1 1 +∞ + 0 - 0 + -∞ -∞ 2 2 -2 -2"
},
{
"figure_ref": [],
"heading": "+∞ +∞ 0",
"publication_ref": [],
"table_ref": [],
"text": "Based on the table of variations, we have:\n• The equation $x^{3}-3x=t_{1}$ has one solution (since $(t_{1}<-2)$.\n• Each equation $x^{3}-3x=t_{2}, x^{3}-3x=t_{3}$ has three distinct solutions (since $-2<t_{2}, t_{3}<2$).\n• Each equation $x^{3}-3x=t_{4}, x^{3}-3x=t_{5}, x^{3}-3x=t_{6}$ has one solution (since $t_{4}, t_{5}, t_{6}>2$).\nThe equation $|f(x^{3}-3x)|=\\frac{2}{3}$ has 10 solutions. Therefore, the answer is B. 10."
},
{
"figure_ref": [],
"heading": "JSON format",
"publication_ref": [],
"table_ref": [],
"text": "We adopt the JSON format for the VNHSEG dataset because it is ideal for LLMs training, testing, and evaluation. Because it makes both accessing and processing textual information linked to syntactic structure and content-related information simple, the JSON format is especially well suited for LLM inputs. A variety of text data, including formulas, equations, tables, and images, can be stored and represented in a flexible and expandable manner using the JSON format. In general, the usage of JSON format makes the VNHSEG dataset compatible with a variety of LLMs and makes it easier to train, test, and evaluate LLMs.\n{ $(t_{1}<-2; -2<t_{2}, t_{3}<2; t_{4}, t_{5}, t_{6}>2)$. \\nConsidering the function $t(x)=x^{3}-3x$, we have $t^{\\prime}(x)=3x^{2}- ID refers to the ID of the question; IQ refers to the images of the question; Q refers to the question content; C refers to the choice options; IE refers to the images of the explanation; and E refers to the explanation content."
},
{
"figure_ref": [],
"heading": "Language",
"publication_ref": [
"b11",
"b61",
"b62"
],
"table_ref": [],
"text": "Vietnamese and English were used in the construction of the VNHSGE dataset. VNHSGE-V is in Vietnamese and VNHSGE-E is in English. GPT-4/ChatGPT was used to translate VNHSGE-V into VNHSGE-E. According to earlier research [12], [62], and [63], GPT-4/ChatGPT can successfully serve as the appropriate translation engine in this circumstance. It should be noted that ChatGPT or BingChat were used to translate the illustrative examples for the dataset presented in this work from Vietnamese to English."
},
{
"figure_ref": [],
"heading": "Subdataset",
"publication_ref": [],
"table_ref": [
"tab_3"
],
"text": "Table 4 shows the VNHSGE dataset structure. The dataset for mathematics and English consists of 2500 multiple-choice questions per subject, while the other multiple-choice subjects have 2000 questions. Literature has 50 exams with 300 essay questions. The dataset contains a large number of questions spanning various topics, ranging from recall-level knowledge to complex multi-step reasoning requirements (see Appendix section C for more details of examples). Total 19000 multiple-choice questions and 300 essay questions"
},
{
"figure_ref": [],
"heading": "Mathematics",
"publication_ref": [
"b52",
"b30",
"b31",
"b29"
],
"table_ref": [],
"text": "In contrast to a number of earlier mathematics datasets, including the Mathematics dataset [53], MATH dataset [31], GSM8K dataset [32], and ScienceQA dataset [30], the VNHSGE mathematics dataset covers a wide range of topics, including spatial geometry, number series (arithmetic progression, geometric progression), combinations and probability, derivatives and applications, exponential and logarithmic functions, primitives and integrals, complex numbers, polyhedrons, rotating blocks, and Oxyz spatial. To help models learn how to provide answer derivations and explanations, the dataset includes questions and related solutions, which are supplied in a complete step-by-step solution (C.1). The VNHSGE mathematics dataset also includes straightforward to complicated questions, necessitating strong mathematical reasoning skills from LLMs in both question answering and visual question answering tasks.\nFirst, the knowledge level question (C.1.1) has been created such that LLMs can quickly and simply solve it using their fundamental understanding. We need 1-2 steps to solve this kind of question. The mathematical calculation skills of LLMs may be put to the test by questions like (+ -× ÷ d dx ). In order to answer the comprehension level questions (C.1.2), LLMs must then infer a few steps to arrive at the appropriate answer. LLMs' capacity for reasoning is put to the test by this kind of question at the level of an average student. Further complicating matters for LLMs is the fact that these kinds of application level problems (C.1.3) mix several different mathematical ideas and need multiple complicated reasoning steps. These inquiries may assess a model's capacity for rational thinking and mathematical knowledge synthesis. Last but not least, the high application level questions (C.1.4) frequently feature unique solutions based on advanced mathematical reasoning and practical problem-solving techniques. LLMs need to have very strong deductive reasoning skills and expertise in solving difficult mathematical problems in order to answer these kinds of inquiries.\nThe VNHSGE mathematics dataset is a thorough collection that addresses a variety of mathematical topics. The dataset was created to evaluate LLMs' capacity for mathematical reasoning on a range of levels, including knowledge, comprehension, application, and high application. The questions in the dataset range in complexity from simple to complicated, therefore the models must have strong inference and reasoning skills. The dataset includes questions that may be answered in one or two steps using fundamental information, as well as problems that call for several steps and knowledge synthesis. The VNHSGE mathematics dataset is an excellent resource for developing and assessing LLMs' mathematical reasoning and inference skills since it presents a strong challenge to their mathematical aptitude in both breadth and depth."
},
{
"figure_ref": [],
"heading": "Literature",
"publication_ref": [],
"table_ref": [],
"text": "The literary exam, a structured assessment tool used to evaluate a student's reading comprehension and writing abilities, serves as the foundation for the VNHSGE literature dataset. This dataset can be deployed for the training and evaluation of LLMs for a variety of language understanding tasks, including essay writing, writing proficiency, and reading comprehension. The dataset is divided into two parts: the question and the answer (C.2). The question section (C.2.1) is divided into two parts. Four questions in Part I's reading comprehension assessment ask students to analyze and understand a paragraph or poetry. The questions ask one to identify the genre and any words or phrases with unique meanings before you analyze their significance. Students must give their own personal opinion of the text or assess another person's personal view of the text for the final question. Writing abilities are the main topic of Part II, which also contains two essay challenges, one on how to write an arguing social essay and the other on how to write an argumentative literary essay. The essay questions test a student's ability to formulate a coherent and succinct argument, back it with evidence, and analyze and interpret literary materials in order to develop a well-supported argument. The answer suggestions and grading guidelines are included in the solution (C.2.1). The scoring criteria are written down in great depth in the grading instructions (C.2.2). The suggested answers are given in accordance with the evaluation criteria.\nThe dataset created based on the answer key with grading guidelines and answer recommendations can assist LLMs in strengthening their capacity to respond to inquiries and offer pertinent justifications based on certain rating metrics. Language models can become more accurate and efficient at answering queries by being trained on this dataset to better grasp and adhere to grading requirements. This dataset offers a thorough assessment of a student's reading comprehension and writing abilities in high school literature, thereby providing a valuable tool for developing and testing LLMs for a variety of language understanding tasks, including sentiment analysis, question answering, text generation, and text summarization. Moreover, the VNHSGE literature dataset is built in Vietnamese, which challenges the ability of LLMs in NLP as Vietnamese is one of the languages with many layers of meaning. Additionally, because Vietnamese is one of the languages with multiple layers of implications, the VNHSGE literature dataset challenges LLMs' proficiency in NLP."
},
{
"figure_ref": [],
"heading": "English",
"publication_ref": [
"b55",
"b48",
"b51"
],
"table_ref": [],
"text": "For datasets involving question-answering, there are plenty of options. For instance, the DREAM dataset [56] focuses on reading comprehension for dialogue while the RACE dataset [49] exclusively considers paragraph reading comprehension. Another dataset that covers eight tasks is SuperGLUE [52]. These datasets have performed admirably for the intended purposes, but they do not provide a comprehensive examination of the LLMs' general language processing abilities.\nThe VNHSGE English dataset contains an assortment of exam questions from high school exams that cover a variety of topics and demand a variety of linguistic abilities (C.3). In the dataset's pronunciation and stress questions (C.3.1), LLMs are asked to choose the word whose underlined portion is pronounced differently from the other three. LLMs are also required to select the proper response from a list of alternatives for questions on vocabulary and grammar (C.3.2), identify terms with opposite or similar meanings, choose the closest-meaning sentence, and fix underlined parts. In order to pass the communication skills test (C.3.3), LLMs are required to select the appropriate response for each conversation. LLMs fill in each of the numbered blanks in the reading fill-in-the-blank questions (C.3.4) by choosing the appropriate word or phrase. Furthermore, LLMs are required to read passages in order to respond to questions about reading comprehension (C. 3.5). At the human level, the dataset encompasses an extensive variety of topics and activities. The dataset is also made up of questions and answers, where the answers are explained in great depth in the solutions. This aids in teaching LLMs how to think critically.\nThe VNHSGE English dataset is a useful tool for LLMs to enhance their proficiency in a range of topics and abilities connected to English language comprehension at the human-level performance. These models can perform better in a variety of language-related tasks, including question answering, language modeling, text generation, reading comprehension, text summarization, etc. by being trained on this dataset, which may assist these models comprehend and process natural language effectively."
},
{
"figure_ref": [],
"heading": "Physics",
"publication_ref": [
"b33",
"b35",
"b29"
],
"table_ref": [],
"text": "In the previous physics datasets, the TQA dataset [34] concentrates on life, earth, and physical sciences and includes both text and pictures for machine comprehension and visual question answering. Although the TQA dataset is intended for middle school students, it appears to be simple enough for LLMs in use today. The PIQA dataset [36] tests the LLMs' capacity for physical reasoning, it is suited for honing their capacity for inference and leaves out the computationally demanding physics problems that they must be able to answer. Physics-related topics such as materials, magnets, velocity, and forces, force and motion, particle motion and energy, heat, and thermal energy, states of matter, kinetic and potential energy, and mixtures are covered in the ScienceQA dataset [30]. Although ScienceQA covers a wide range of topics this is merely elementary physics. On the other hand, the VNHSGE physics dataset is geared toward high school students. The VNHSGE physics dataset also focuses on more complicated topics like electromagnetic oscillations and waves, light waves, quantum light, atomic nuclei, direct current, electromagnetic induction, and light refraction (C.4). The prior datasets can be difficult for LLMs since they demand one to comprehend and make connections between a wide range of scientific principles and notions. The VNHSGE physics dataset, however, may present a bigger challenge for language models because it deals with more complex and specialized physics topics and necessitates a higher level of scientific understanding and reasoning abilities to accurately respond to the questions.\n50% of the questions in the VNHSGE physics dataset are theoretical, and 50% are practical and applied. Most theoretical problems fall under the knowledge level (C.4.1), which calls for both inference and a firm comprehension of theoretical knowledge. For questions at the comprehension level (C.4.2), there is a higher degree of inference about knowledge and mathematical abilities. The application level questions (C.4.3) come next, which have a high categorization and draw on complex physics concepts like understanding of practice and application. The high application level questions (C.4.4) are the last type. These include experimental questions as well as questions that make use of graphs related to mechanical oscillations and alternating currents. These inquiries demand a very high degree of inference, and the unique solutions call for in-depth knowledge of high school physics challenges.\nPhysical concepts like mechanical oscillations, waves, quantum mechanics, and atomic nuclei might be difficult for LLMs to understand and rationalize when presented with physical information from the VNHSGE. In addition to demanding the ability to retain information, the datasets additionally inquire about the ability to draw conclusions, apply ideas to concrete circumstances, and even solve challenging issues. It is a difficult undertaking for any LLMs because the high application-level questions in the dataset demand specialized knowledge and experience in addressing physics issues at the high school level."
},
{
"figure_ref": [],
"heading": "Chemistry",
"publication_ref": [
"b56",
"b17",
"b29"
],
"table_ref": [],
"text": "There aren't many datasets in the field of chemistry that are specifically focused on tackling questions. The SciQ dataset [57] tests LLMs on their knowledge of chemistry with multiple-choice questions. It rates the model's comprehension and deductive reasoning skills in regard to chemistry-related scientific ideas and concepts. The chemistry dataset in [18] focuses on the LLMs' accuracy in chemistry subjects from high school, including chemical reactions, ions, acids, and bases, to college, like analytical, organic, inorganic, and physical. However, there are only a few chemistry questions. Understanding and responding to questions about chemistry subjects like solutions, physical and chemical changes, atoms and molecules, and chemical reactions are the main objectives of ScienceQA dataset [30]. The VNHSGE chemistry dataset, on the other hand, presents difficulties for LLMs in understanding and responding to questions regarding a variety of chemistry topics, including metals, inorganic and organic molecules, polymers, and more (C.5). It rates the model's comprehension and deductive reasoning skills with regard to a variety of chemistry concepts and principles.\nThe VNHSGE chemistry dataset is made up of 30% computational tasks and 70% theoretical questions. Usually, theoretical problems require knowledge and comprehension. The knowledge-level questions are typically brief and demand information-retrieval-level knowledge (C.5.1). Subsequently, the computations in the comprehension level (C.5.2) section are rather straightforward, requiring only 1 or 2 operations for problems. Next, the high-level reasoning and the synthesis of several concepts are required to answer the application-level questions (C.5.3). Finally, the highapplication questions (C.5.4) require in-depth knowledge, logical reasoning, and the synthesis of several chemical reaction equations.\nThe VNHSGE chemistry dataset evaluates LLMs' high-level reasoning and problem-solving abilities as well as their comprehension of chemistry principles across a variety of topics and levels of difficulty. The dataset necessitates that the models have an adequate knowledge of chemical principles and be able to implement that understanding in challenging contexts, such as the synthesis and analysis of chemical reactions."
},
{
"figure_ref": [],
"heading": "Biology",
"publication_ref": [
"b32",
"b56",
"b17",
"b29"
],
"table_ref": [],
"text": "Similar to chemistry, there aren't many biology datasets created expressly for question answering tasks. BioASQ [33] concentrates on medical fields rather than biological ones. The SciQ [57] dataset makes it difficult for LLMs to correctly respond to Biology-related multiple-choice questions on science exams. The dataset evaluates how well the model can understand and justify biological science principles and notions. The MMLU dataset [18] assesses LLMs' accuracy in subjects from high school and college biology, including natural selection, heredity, cell cycle, and more. The ScienceQA dataset [30], on the other hand, focuses on understanding and responding to questions about molecular and cellular biology. Because of its extensive coverage of topics including genetic laws, population genetics, applications of genetics, human genetics, evolution, ecology, plant organismal biology, and animal organismal biology, the VNHSGE biology dataset presents a significant challenge to LLMs (C.6).\nThe questions in the VNHSGE biology dataset are highly challenging and complicated, and in order to accurately respond to them, one must have a thorough understanding of all aspects of biology. According to the dataset's design, there should be 75% theoretical questions and 25% exercises, with 70% of the questions being at the knowledge and comprehension levels and 30% of the questions focusing on application and higher-order thinking skills. The dataset, which includes questions of varying complexity, focuses on the capacity for calculation and inference. The knowledge level questions (C.6.1) demand a comprehensive understanding of biology to answer correctly, while the comprehension level questions (C.6.2) require one to three steps of deductive reasoning to find the answer. The application level questions (C.6.3) focus on areas including rules of genetics, human genetics, population genetics, and mechanisms of inheritance and mutation and call for the capacity to synthesize knowledge. The high application level questions (C.6.4) require sophisticated analysis and problem-solving skills.\nThe VNHSGE biology dataset is a substantial challenge for LLMs since it calls for a mix of in-depth knowledge and sophisticated reasoning abilities in order to correctly understand and respond to questions about a wide range of biology topics."
},
{
"figure_ref": [],
"heading": "History",
"publication_ref": [
"b17",
"b29",
"b17",
"b29"
],
"table_ref": [],
"text": "Both the MMLU dataset [18] and ScienceQA dataset [30] evaluate how well LLMs perform when answering questions about historical events. While the MMLU dataset [18] assesses LLMs' accuracy in high school histories concepts like High School US History, High School European History, and High School World History, the ScienceQA dataset [30] focuses on understanding and responding to questions about American and global history.\nThe purpose of the VNHSGE history dataset is to assess LLMs' knowledge of historical events and milestones as well as to give correct analysis of historical events (C.7). The dataset contains 80% questions at the knowledge and comprehension levels covering a wide range of topics including Vietnamese and global histories (C.7.1 and C.7.2) . To answer these kinds of inquiries, one must not only accurately record the facts but also use historical reasoning. Across topics in Vietnamese history from 1919 to 1975, the dataset contains 20% of questions that require application and high application levels (C.7.3 and C.7.4). The majority of the questions concern comparison essays, connections between topics, links between Vietnamese history and world history, or commentary and summaries of historical periods to identify key characteristics or the substance of historical events. The capacity to analyze, contrast, and comment on historical events is necessary for these kinds of issues.\nThe VNHSGE history dataset is utilized for evaluating how well LLMs can recall and comprehend historical events as well as their timeframes. The questions in the dataset range from simple to complex, requiring varying degrees of deductive reasoning and inference skills. To correctly respond to the questions in the dataset, LLMs must be able to interpret and analyze complicated historical events, appreciate the relationships between them, and draw inferences from them."
},
{
"figure_ref": [],
"heading": "Geography",
"publication_ref": [
"b17",
"b29",
"b63",
"b57"
],
"table_ref": [],
"text": "Few specialized datasets are available for geography question-answering tasks. The MMLU dataset [18] includes a few inquiries about high school geography concepts including population movement, rural land use, and urban processes. While the ScienceQA dataset [30] focuses on questions about state capitals, geography, maps, and more. Additionally, the geography dataset in [64] includes 612 Bulgarian multiple-choice questions for the matriculation exam for the 12th grade. The GeoTSQA dataset [58], which was compiled from high school exams in China, has 1,000 actual questions in the geography domain that are contextualized by tabular scenarios. The VNHSGE geography dataset is intended to assess LLMs' knowledge of geographical concepts such as natural geography, population geography, economic sector geography, economic zone geography, sea geography, and island geography as well as geographical skills such as atlas use, data table interpretation, and chart analysis.\nThe questions in the VNHSGE geography dataset are ordered in order of increasing complexity, with 80% of the questions falling into the basic category (knowledge and understanding) and 20% falling into the advanced category (10% application and 10% high-level application) (C.8). 50% of the exam's questions, such as chart analysis (C.8.1), data table interpretation (C.8.2), and atlas use (C.8.3), involve geographic knowledge. LLMs must be able to solve problems in order to master these skills. Additionally, LLMs must be able to think logically, have a broad understanding of society, be adept at solving problems, and have a high degree of critical thinking to complete the diversified questions (C.8.4).\nQuestions in the VNHSGE geography dataset call for a variety of abilities, such as data analysis, chart interpretation, and atlas use, which can assist in training LLMs to comprehend and process complicated material in these fields. The dataset also contains questions that call for reasoning, problem-solving, and critical thinking, which can aid in the development of more sophisticated language skills in language models."
},
{
"figure_ref": [],
"heading": "Civic Education",
"publication_ref": [
"b37",
"b64",
"b38",
"b65",
"b66",
"b67",
"b68",
"b17",
"b29"
],
"table_ref": [],
"text": "There have been numerous attempts to construct datasets connected to the legal profession and ethics, which has recently received special attention. While the JEC-QA dataset [38] contains questions connected to the national judicial examination in China, the CJRC dataset [65] comprises documents and questions relating to legal knowledge in China. The CaseHOLD dataset [39], which focuses on finding the critical components in a legal case, is a novel and difficult dataset in the subject of law. While the PolicyQA dataset [66] focuses on comprehending the privacy policies of websites, the PrivacyQA dataset [67] focuses on queries regarding the privacy policies of mobile applications. To guarantee the accuracy of the replies, both databases offer questions that have been reviewed by experts. The Vietnamese transportation law dataset [68] and the Vietnamese law dataset [69] both concentrate on questions pertaining to law, but the Vietnamese transportation law dataset is more concerned with traffic law and the Law dataset is more concerned with broad legal issues. Additionally, MMLU dataset [18] has a few questions about professional law as well as questions about international law including torts, criminal law, contracts, etc. Focused on questions about civics subjects like social skills, governance, and the constitution is the ScienceQA dataset [30]. While the VNHSGE civic education dataset is intended to provide LLMs with civic education and legal training, it also focuses on case studies and multiple-choice questions on topics such as legal frameworks and regulations, fundamental civil rights, democratic principles, and case studies.\nThe purpose of VNHSGE civic education dataset is to evaluate LLMs' understanding of and ability to apply legal concepts (C.9). 70% of the exam's questions are knowledge and comprehension level questions (C.9.1 and C.9.2). 30% of the questions are application and high application levels, focused on topics like Citizens' fundamental rights; types of legal infractions; and equal rights in certain areas of social life. There is a lot of confusion in the answer choices for questions at the application level (C.9.3), making it difficult to accurately assess and choose the right response. Complex case studies with several plotlines and characters are offered for questions at the high level (C.9.4), and it needs a thorough comprehension of legal theory to properly examine the nature of the characters' violations.\nFor LLMs to assess their understanding of and ability to apply legal information, particularly in the context of civic education and legal training, the VNHSGE civic education dataset is employed. The dataset includes case studies together with multiple-choice questions on topics like legal frameworks and regulations, fundamental citizen rights, democratic principles, and notions. LLMs can gain a better understanding of legal ideas and how to apply them in practical scenarios by training on this dataset, which can be helpful for a range of applications like legal research, automated legal document analysis, and legal chatbots."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [
"fig_1"
],
"heading": "ChatGPT and BingChat responses",
"publication_ref": [],
"table_ref": [],
"text": "Response format: When posing questions to LLMs, we can receive answers in various formats. To standardize response formats and simplify result processing, we request that LLMs provide replies in a specific structure. Figure 1 demonstrates an example of the required structure for LLM responses. To achieve this, we used the Explanation and Choice approach and include a \"pre-question\" prompt before the actual question. This prompt combines the content of the original question with instructions for the desired response format. Standardizing the format of LLM answers is crucial for several reasons. Firstly, it enables quicker and more accurate processing of model responses. Secondly, it facilitates impartial comparison and evaluation of the performance of different LLMs. Additionally, it ensures that the solutions provided by LLMs are easy to understand and applicable for further applications. By giving LLM responses a clear and consistent structure, we can effectively harness their abilities to enhance various NLP tasks. "
},
{
"figure_ref": [],
"heading": "Question (Word format):",
"publication_ref": [],
"table_ref": [],
"text": "ID IQ Q C IA E1\n1) The volume of a cube with edge 2a is:\nA. 8a^3 B. 2a^3. C. a^3 D. 6a^3."
},
{
"figure_ref": [],
"heading": "A",
"publication_ref": [],
"table_ref": [],
"text": "The volume of a cube with edge 2a is: V=(2a)^3=8a^3.\nQuestion (JSON format): { \"ID\": \"Q1\", \"IQ\": \" \", \"Q\": \"1) The volume of a cube with edge 2a is:\\nA. 8a^3.\\t\\nB. 2a^3.\\t\\nC. a^3.\\t\\nD. 6a^3.\", \"C\": \"A\", \"IA\": \" \", \"E\": \"The volume of a cube with edge 2a is: V=(2a)^3=8a^3.\", } Pre-question (JSON format): \"I want you to answer the question in the following structure: \" \"\\nChoice: \"A\" or \"B\" or \"C\" or \"D\" \" \"\\nExplanation: Explain the answer\" \"\\nThe question is: {}\" New Question (Prompt): I want you to answer the question in the following structure:\nChoice: \"A\" or \"B\" or \"C\" or \"D\"\nExplanation: Explain the answer\nThe question is: 1) The volume of a cube with edge 2a is: A. 8a^3. B. 2a^3. C. a^3. D. 6a^3.\""
},
{
"figure_ref": [],
"heading": "Response (JSON format):",
"publication_ref": [],
"table_ref": [],
"text": "{ \"ID\": \"1\", \"IQ\": \" \", \"Q\": \"1) The volume of a cube with edge 2a is:\\nA. 8a^3.\\t\\nB. 2a^3.\\t\\nC. a^3.\\t\\nD. 6a^3.\", \"C\": \"A\", \"IA\": \" \", \"E\": \"The volume of a cube with edge 2a is: V=(2a)^3=8a^3.\", \"CC\": \"A\", \"CE\": \"The formula for the volume of a cube is V = s^3, where s is the length of one of its sides. Therefore, the volume of the cube with a side length of 2a is: V = (2a)^3 = 8a^3\", }\nResponse (Word format):\nID IQ Q C IA E CC CE1\n1) The volume of a cube with edge 2a is:\nA. 8a^3 B. 2a^3. C. a^3 D. 6a^3."
},
{
"figure_ref": [],
"heading": "A",
"publication_ref": [],
"table_ref": [],
"text": "The volume of a cube with edge 2a is: V=(2a)^3=8a^3."
},
{
"figure_ref": [],
"heading": "A",
"publication_ref": [],
"table_ref": [],
"text": "The formula for the volume of a cube is V = s^3, where s is the length of one of its sides. Therefore, the volume of the cube with a side length of 2a is: V = (2a)^3 = 8a^3\nWe conducted experiments using two state of the art language models, ChatGPT (based on GPT-3.5) and BingChat (based on GPT-4)1 , to evaluate the performance of our dataset. We assessed each model based on accuracy and provided examples of both successful and poor responses (see Appendix section C for further details of examples).\nIn the following sections, we compared the performance of ChatGPT and BingChat using five tests for each subject, including 30 literary essays and 1700 multiple-choice questions in others. LLMs like ChatGPT and BingChat have been trained to predict the next word in a text based on the preceding words. However, these models have limitations when it comes to handling complex computational problems or requiring multi-step reasoning, even though they are capable of responding to basic questions. These LLMs may also struggle to comprehend texts with intricate contexts and may encounter difficulties in certain situations, particularly when processing the Vietnamese language. They might misinterpret certain contexts and occasionally confuse words with homonyms or antonyms.\nMathematics: ChatGPT and BingChat can handle knowledge and comprehension level questions (C.1.1) and (C.1.2). However, they struggle with complex calculations and logical reasoning that require advanced mathematical skills or multi-step deductive reasoning (C.1.3). These models often provide inaccurate explanations and answers and are unable to provide appropriate solution instructions for high application level problems (C.1.4).\nLiterature: ChatGPT and BingChat are capable of responding to literary queries and generating essays due to their extensive training in various domains, including literature and journalism. They have a good grasp of natural language structure and can synthesize new responses and paragraphs based on learned knowledge and input data. However, ChatGPT and BingChat still have limitations in reasoning abilities and understanding complex language and context, particularly in languages like Vietnamese. As a result, their responses may not always be entirely accurate or suitable for the context or purpose of the question (C.2.1). ChatGPT is more suitable for language-related topics and tends to provide more relevant and emotive responses compared to BingChat, a search engineer (C.2.2).\nEnglish: ChatGPT and BingChat are unable to respond to questions on pronunciation and stress (C. Biology: Both ChatGPT and BingChat are capable of providing responses to questions at the knowledge and comprehension levels (C.6.1 and C.6.2), similar to subjects like mathematics, physics, and chemistry that require both calculation and reasoning skills. However, ChatGPT and BingChat have a very limited likelihood of correctly determining the answers to questions requiring complex thinking and information processing in diagrams at the application and high application levels (C.6.2 and C.6.4). These types of questions demand a deeper understanding of biology concepts and the ability to apply them in complex scenarios.\nHistory: ChatGPT and BingChat do reasonably well when answering questions in the field of history at the knowledge and comprehension levels (C.7.1 and C.7.2). However, both ChatGPT and BingChat often struggle to provide accurate responses to the application and high application questions (C.7.3 and C.7.4). These types of questions require higherorder thinking skills and a deep understanding of the historical context as well as demand the ability to compare, analyze, and express a judgment on historical events and characters."
},
{
"figure_ref": [],
"heading": "Geography:",
"publication_ref": [],
"table_ref": [],
"text": "ChatGPT is able to respond to questions about charts without requesting data from the chart, BingChat does not support these questions (C.8.1). The result is that both they cannot answer questions related to charts or images. Both ChatGPT and BingChat can provide precise responses to questions about the information in a table (C.8.2) and queries related to the use of the Atlas (C.8.3). However, when it comes to questions that require analysis and interpretation at the application and high application levels (C.8.4), both ChatGPT and BingChat often struggle to give precise responses. These types of questions necessitate the ability to analyze and interpret geographical data and concepts, which the models may find challenging.\nCivic Education: At the knowledge and comprehension levels (C.9.1 and C.9.2), ChatGPT and BingChat can provide accurate answers. However, ChatGPT often produces inaccurate responses for questions at the application level (C.9.3), while BingChat performs better. Both ChatGPT and BingChat often fail to provide precise responses when analyzing character behavior in scenario-based questions at the high application level (C.9.4)."
},
{
"figure_ref": [
"fig_3"
],
"heading": "ChatGPT and BingChat performances",
"publication_ref": [],
"table_ref": [
"tab_4"
],
"text": "Table 5 displays ChatGPT and BingChat's performance. We can see that for subjects requiring complex computation and reasoning, such as mathematics, physics, chemistry, and biology, their performance ranges from 48% to 69%.\nThe performance of ChatGPT and BingChat is between 56.5% and 92.4% for subjects that predominantly depend on languages, such as literature, English, history, geography, and civic education. LLMs such as ChatGPT and BingChat have been trained on vast amounts of text covering a wide range of fields. However, these models lack subject-matter expertise. Mathematics, physics, chemistry, and biology often demand profound knowledge and advanced computational abilities, which may not be possessed by language models like ChatGPT and BingChat for solving such challenging problems. On the other hand, subjects like literature, English, history, geography, and civic education frequently require strong language skills and the ability to comprehend complex texts, areas in which language models like ChatGPT and BingChat may have sufficient capabilities to handle. The performance comparison between ChatGPT and BingChat is depicted in Figure 2. BingChat performs better than ChatGPT in all categories except for literature. There is not much difference between BingChat and ChatGPT in subjects like mathematics, physics, and chemistry, which require extensive computation and reasoning. However, ChatGPT surpasses BingChat in terms of performance in the literature category. This is because BingChat is a search engine, and its results may not be suitable for the literature subject, which often involves writing extensive essays. BingChat outperforms ChatGPT in the remaining topics. It should be noted that BingChat is based on GPT-4 while ChatGPT is based on GPT-3.5. Furthermore, BingChat may find accurate answers when the questions and answers are publicly available online. "
},
{
"figure_ref": [],
"heading": "M a t h e m a t i c s L i t e r a t u r e E n g l i s h P h y s i c s",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "ChatGPT, BingChat, and Vietnamese Students",
"publication_ref": [],
"table_ref": [],
"text": "This section compares the effectiveness of BingChat and ChatGPT with Vietnamese students. Our aim is to determine whether LLMs possess abilities comparable to those of humans, although this comparison is challenging due to the dissimilar settings. By conducting this comparison, we can evaluate whether LLMs can serve as effective tools for Vietnamese students in various subject areas (see Appendix section D for more details of spectrum comparisons). Mathematics: According to the findings, ChatGPT and BingChat are unable to match the performance of human students in Vietnam's high school mathematics curriculum. Despite being trained on vast amounts of textual data from the internet, they struggle with complex mathematical problems, although they can handle simpler mathematical concepts. The high school mathematics questions require reasoning, logical thinking, analytical skills, and the ability to apply knowledge in practical situations. To achieve performance on par with humans in high school mathematics, ChatGPT and BingChat's mathematical abilities need substantial improvement.\nLiterature: Both ChatGPT and BingChat have been extensively trained on large Vietnamese language datasets, enabling them to analyze and generate essays with considerable proficiency. In terms of high school literature, the performance of LLMs such as ChatGPT and BingChat is human-like level. However, it should be emphasized that ChatGPT and BingChat are unable to write emotionally rich essays or conduct in-depth literary analyses. In summary, ChatGPT can be considered a tool to support Vietnamese students in studying literature.\nEnglish: According to the results, ChatGPT and BingChat performed better in high school English compared to Vietnamese students. It should be mentioned that Vietnamese students' English proficiency is not very high compared to the global average. ChatGPT and BingChat are effective tools that Vietnamese students can utilize to study foreign languages. "
},
{
"figure_ref": [],
"heading": "Physics:",
"publication_ref": [],
"table_ref": [],
"text": "The performance of ChatGPT and BingChat is comparable to the average score of Vietnamese students in physics. However, they are still less than the score achieved by most Vietnamese students. With thorough training in the field of physics, LLMs can provide accurate answers and insightful explanations to assist students in understanding physics. The models, however, still require development, particularly for physics issues that call for intricate computations and reasoning.\nChemistry: ChatGPT and BingChat still do not possess the same level of proficiency in chemistry as Vietnamese high school students do. While these LLMs can provide relevant knowledge and solutions in the field of chemistry, they lack the expertise required to solve complex chemistry problems that demand advanced levels of analysis and reasoning. However, in terms of delivering theoretical knowledge and information, it is certainly possible for LLMs to become useful tools for Vietnamese students in high school chemistry."
},
{
"figure_ref": [],
"heading": "Biology:",
"publication_ref": [],
"table_ref": [],
"text": "The findings indicate that ChatGPT and BingChat outperform Vietnamese students in biology. It is important to note that biology is considered a less prioritized subject for many Vietnamese students compared to mathematics, physics, and chemistry. The biology score of Vietnamese students is less in mathematics, physics, and chemistry. LLMs are capable of addressing basic questions in biology, such as definitions, concepts, simple problem-solving, and specific examples. Therefore, LLMs can serve as helpful resources for high school students to comprehend fundamental biology concepts and problems. History: While BingChat performs better, ChatGPT's results are comparable to those of Vietnamese students. With extensive and diverse training datasets, ChatGPT and BingChat are able to understand and process different types of historical questions and provide logical and useful responses. Although ChatGPT and BingChat may still encounter challenges with complex questions, they can be valuable resources for high school students in history. Geography: While BingChat achieves higher scores, ChatGPT performs at a similar level to Vietnamese students. The results indicate that both ChatGPT and BingChat are capable of understanding and responding to high school-level geography questions. They can effectively teach geography concepts and terminology, enhancing students' learning in high school geography. However, they may still face limitations when dealing with complex and in-depth inquiries that require advanced critical thinking.\nCivic Education: BingChat and ChatGPT showcase human-like abilities in the field of civic education. With their training in civic education and law-related subjects, they possess the expertise to provide high school-level knowledge in areas such as politics, law, citizen rights and responsibilities, and other social issues. Therefore, as reference tools, ChatGPT and BingChat can be highly valuable for Vietnamese students studying civic education."
},
{
"figure_ref": [],
"heading": "VNHSGE dataset and other datasets",
"publication_ref": [
"b11"
],
"table_ref": [],
"text": "In Figure 6, the performance of ChatGPT and BingChat on the VNHSGE dataset is compared to other datasets in the GPT-4 Report [12]. The results show that ChatGPT's performance on the VNHSGE dataset is comparable to that of GPT-3.5 across subjects ranging from AP Statistics to AP Psychology. BingChat improves its performance in text-based subjects such as history, geography, civic education, and English. However, BingChat's performance does not significantly outperform ChatGPT in subjects like mathematics, physics, chemistry, and biology, which require complex computation and reasoning. On the other hand, GPT-4 exhibits better performance than GPT-3.5 in tasks of similar nature. This could be due to the structure of questions in these subjects from the VNHSGE dataset, which presents challenges for BingChat, particularly at the application and high application levels. "
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "In this paper, we present the VNHSGE dataset, which is intended to evaluate and train LLMs' multitask abilities such as question answering, text generation, reading comprehension, visual question answering, and more. The dataset covers nine subject areas from the Vietnamese National High School Graduation Examination, including social and language subjects such as literature, English, history, geography, and civic education, as well as calculation and inference subjects like mathematics, physics, chemistry, and biology. The dataset encompasses a wide range of question types, spanning from basic recall to complex calculation and reasoning questions. The VNHSGE dataset serves as a valuable resource for training LLMs, offering a diverse set of challenges at the human level. The dataset helps researchers identify critical flaws in models, thereby facilitating the improvement of LLMs' abilities. The VNHSGE dataset has various benefits for developing LLMs, including:\n• Comprehensive coverage: The dataset provides thorough coverage of a wide range of topics in nine high school subjects. This enables more thorough training of language models across diverse computing and inference domains.\n• Various question types: The dataset contains a wide range of question types, from straightforward knowledgebased inquiries to intricate application-based inquiries requiring extensive investigation and evaluation. This offers a wide range of learning challenges for language models. • Different difficulty levels: The VNHSGE dataset contains questions that range in complexity from simple to sophisticated, making it possible to train models that can handle a variety of question challenges. • Vietnamese language: Given that the dataset is in Vietnamese, it is possible to train language models in a language other than English, enhancing their adaptability and global applicability.\nThe state of the art of LLMs, ChatGPT and BingChat, tested on the VNHGE dataset showed that the VNHSGE dataset is perfectly suited for LLMs. This outcome not only demonstrates the models' abilities but also presents chances and difficulties for LLMs deploying in the field of education.\nThe VNHSGE dataset demonstrates the advantages and disadvantages of LLMs and offers information about possible instructional applications for these models. Additionally, it poses a challenge for LLMs to enhance their abilities to handle challenging, high-level application questions."
},
{
"figure_ref": [],
"heading": "Appendix",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [
"fig_11",
"fig_11",
"fig_11"
],
"heading": "A Available datasets",
"publication_ref": [
"b48",
"b23",
"b51",
"b55",
"b52",
"b30",
"b31",
"b33",
"b34",
"b35",
"b36",
"b29",
"b53",
"b54",
"b37",
"b38",
"b29",
"b23",
"b24",
"b25"
],
"table_ref": [],
"text": "The authors searched Paperwithcode2 as of April 25, 2023 for pre-existing datasets for a variety of topics, tasks, and languages in order to construct a new dataset. We searched for pertinent datasets on three levels: \"General\" datasets, datasets linked to \"Texts\", and datasets connected to \"Text and Question Answering (QA)\" using keywords like mathematics, literature, english, physics, chemistry, biology, history, geography, and law. Figure 7a shows the number of datasets in subjects. The most datasets, including RACE [49], MLQA [24], SuperGLUE [52], and DREAM [56] were found in English, whereas Mathematics had Mathematics [53], MATH [31] and GSM8K [32]. Numerous datasets were available for Physics, including TQA [34], SWAG [35], PIQA [36], PROST [37], and ScienceQA [30]. The only two datasets for Literature and Law are [54], [55] and JEC-QA [38], CaseHOLD [39], respectively, whereas the only one dataset available for Chemistry, Biology, History and Geography were ScienceQA [30]. We discovered that QA datasets had the highest number of datasets in the \"Texts\" category shown in Figure 7b. It is observed that only three datasets, MLQA [24], XQuAD [25], and MKQA [26], shown in Figure 7c, supported Vietnamese. "
},
{
"figure_ref": [
"fig_23"
],
"heading": "B Dataset format",
"publication_ref": [
"b0"
],
"table_ref": [],
"text": "In this section, we describe how to convert formulas, equations, tables, photos, and charts from raw text formats like Word, Pdf, and HTML into a text-only format and an image folder. The exact steps of the method are shown in detail in Figure 8 including steps: (1) • Step 2: Symbols, formulas, and equations are converted to text format using the LaTeX format in the \"Raw data\" to \"Word format\" conversion. In mathematics, physics, chemistry, and biology, we convert symbols, formulas, and equations using three different techniques. The first technique converts Word documents with equations and formulae to the Latex format using the built-in equation editor in Microsoft Word. If the first approach is unable to convert the raw data, the second option employs the Mathpix3 software to convert pdf files to the Latex format. Sometimes it's not possible to utilize any of the two ways mentioned earlier, in which case we must manually input the formulas and equations.\n• Step 3: Convert \"Word format\" into \"JSON format\". With the aid of Python libraries, including \"docx4 \" and \"JSON,\" it is simple to convert Word files to JSON files. The procedure entails importing the necessary libraries before using their functions to parse and convert the text data to JSON format. The \"docx\" library offers tools for reading and writing Microsoft Word documents. Data conversion to the JSON format is made easy and effective by the \"JSON\" library."
},
{
"figure_ref": [],
"heading": "B.1 Raw data",
"publication_ref": [],
"table_ref": [],
"text": "There are several phases involved in transforming raw data into a machine-readable format. Finding and removing pertinent information from the raw data is one of the crucial tasks. The raw data may be in several formats, including HTML, PDF, or Word, and may include information on a variety of disciplines, including math, literature, english, physics, chemistry, biology, history, geography, and civic education. Each of these subjects has distinct qualities that call for various extraction techniques. For instance, symbols, formulas, and equations in mathematics, physics, chemistry, and biology must be precisely retrieved and represented. These equations might be as simple as simple biological equations or as sophisticated as complex mathematical equations. On the other hand, geography frequently contains a large number of images and charts that must be accurately retrieved and provided. Symbols, formulas, and equations are often not used in literature, english, history, or civic education; instead, these subjects place a greater emphasis on textual content. In conclusion, a variety of approaches and methodologies were used to transform raw data into a machine-readable format, ensuring that all pertinent information is extracted and accurately represented. A similar strategy must be used to accurately extract and represent each subject within the raw data. This procedure is essential for ensuring that the result is precise, trustworthy, and simple to understand.\nOur raw data is displayed in \"Raw data sample\" as an example. For ease of viewing, we don't present the example in table format. This is a query from the math dataset. As we can see, the questions and answers both include illustrations, equations, and formulas. The answers include thorough justifications that call for high-level inference skills, while the questions demand the ability to extract information from images. The information is complicated and may require specialist knowledge or training to properly understand, as suggested by the use of images and technical language."
},
{
"figure_ref": [],
"heading": "\"Raw data sample\"",
"publication_ref": [],
"table_ref": [],
"text": "Question: Let y = f (x) be a cubic function with the graph shown in the picture. \n-2 2 -1 1 2 x y Setting t = x 3 -3x, we have f x 3 -3x = 2 3 ⇔ |f (t)| = 2 3 .\nFrom the above graph, we conclude that the equation |f (t)| = 2 3 has six distinct solutions t = t i (with i = 1, 6 and t 1 < -2; -2 < t 2 , t 3 < 2; t 4 , t 5 , t 6 > 2). Considering the function t\n(x) = x 3 -3x, we have t (x) = 3x 2 -3; t (x) = 0 ⇔ x = ±1. The sign variation table of t(x) is: x f (x) f (x) -∞ -1 1 +∞ + 0 - 0 + -∞ -∞ 2 2 -2 -2"
},
{
"figure_ref": [],
"heading": "+∞ +∞ 0",
"publication_ref": [],
"table_ref": [],
"text": "Based on the table of variations, we have:\n• The equation x 3 -3x = t 1 has one solution (since t 1 < -2).\n• Each equation x 3 -3x = t 2 , x 3 -3x = t 3 has three distinct solutions (since -2 < t 2 , t 3 < 2).\n• Each equation x 3 -3x = t 4 , x 3 -3x = t 5 , x 3 -3x = t 6 has one solution (since t 4 , t 5 , t 6 > 2).\nThe equation f x 3 -3x = 2 3 has 10 solutions. Therefore, the answer is B. 10."
},
{
"figure_ref": [],
"heading": "C Samples",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "C.1 Mathematics",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "C.1.1 Knowledge level question",
"publication_ref": [],
"table_ref": [],
"text": "Simple math problems can be answered correctly by both ChatGPT and BingChat, and their rationales are adequate. This is due to the fact that easy arithmetic problems typically have a particular format and structure, which enables the models to employ the established patterns and guidelines to get accurate results.\nQuestion: Given the geometric sequence (u n ) with u 1 = 2 and common ratio q = 1 2 . What is the value of u\n3 ? A. 3 B. 1 2 C. 1 4 D. 7 2 Solution: We have u 3 = u 1 • q 2 = 2 • ( 1 2 ) 2 = 2 • 1 4 = 1 2 .\nTherefore, the answer is B 1 2 . ChatGPT: We have a geometric sequence (u n ) with common ratio q = 1 2 and u 1 = 2. Therefore\nu 2 = u 1 • q = 2 • 1 2 = 1, u 3 = u 2 • q = 1 • 1 2 = 1 2\nThus, the value of u 3 is 1 2 , and the correct answer is B. 1 2 . BingChat: In a geometric sequence, the nth term is calculated using the formula u n = u 1 * q n-1 . Therefore, the value of u 3 in this geometric sequence is\nu 3 = u 1 * q 3-1 = 2 * (1/2) 2 = 1/2. The correct answer is B. 1/2."
},
{
"figure_ref": [],
"heading": "C.1.2 Comprehension level question",
"publication_ref": [],
"table_ref": [],
"text": "Due to limitations in their training data and their computational abilities, LLMs may have trouble providing accurate answers as arithmetic question difficulty rises. In certain circumstances, ChatGPT might be unable to offer the right solution, whereas BingChat is able to give the correct answer and provide a rational explanation. The probability is:"
},
{
"figure_ref": [],
"heading": "Question",
"publication_ref": [],
"table_ref": [],
"text": "P = 12+15 105 = 9 35 .\nTherefore, the answer is A. 9 35 . ChatGPT: To choose two balls of different colors and have the sum of their numbers even, we must have the following cases:\n• Choose one red ball and one green ball. This case occurs when we select one ball from the group of 6 red balls and one ball from the group of 9 green balls. The sum of the numbers on these two balls is necessarily odd, so we do not need to consider this case. • Choose two green balls. This case occurs when we select two balls from the group of 9 green balls. The number of ways to choose two green balls is 9 2 = 36. There are 5 cases where the sum of the numbers on the two green balls is even:\n• Choose two green balls numbered 1: sum is 2 (even).\n• Choose two green balls numbered 2: sum is 4 (even). • ... • Choose two green balls numbered 9: sum is 18 (even). Thus, there are 5 satisfying cases. The probability we seek is: number of satisfying cases number of ways to choose two balls from the box = 5\n15 2 = 5 105 = 1 21\nTherefore, the correct answer is D. 1 7 BingChat: The correct answer is A. 9/35. The number of ways to choose two balls of different colors and the sum of the two numbers on them is even equals the number of ways to choose a red ball with an odd number on it and a blue ball with an odd number on it plus the number of ways to choose a red ball with an even number on it and a blue ball with an even number on it. Therefore, the number of ways to choose two balls of different colors and the sum of the two numbers on them is even equals 3 * 5 + 3 * 4 = 27. The number of elements in the sample space equals C(15,2) = 105. Therefore, the probability of getting two balls of different colors and the sum of the two numbers on them is even equals 27/105 = 9/35."
},
{
"figure_ref": [],
"heading": "C.1.3 Application level question",
"publication_ref": [],
"table_ref": [],
"text": "Both ChatGPT and BingChat are unable to resolve complex issues that need an in-depth analysis of derivatives and graphs. This is due to the fact that their training data and computational language abilities are constrained, and it's possible that they lack the mathematical knowledge required to address these issues.\nQuestion: How many integer values of the parameter m make the function y = -x 4 + 6x 2 + mx have three extreme points?\nA. 17 \n⇔ m = 4x 3 -12x. Consider the function g(x) = 4x 3 -12x with g (x) = 12x 2 -12. We have g (x) = 0 ⇔ 12x 2 -12 = 0 ⇔ x = ±1.\nThe tables of signs and variations of g(x):\nx y (x)\ny(x) -∞ -1 1 +∞ + 0 - 0 + -∞ -∞8 8\n-8 -8"
},
{
"figure_ref": [],
"heading": "+∞ +∞ 0",
"publication_ref": [],
"table_ref": [],
"text": "Based on the tables of signs and variations, equation ( 1 \n( √ 2) = y (- √ 2) = 0 Which is equivalent to: ⇔ -8 √ 2 + 12 + √ 2m = 0 8 √ 2 + 12 - √ 2m = 0 Solving this system of equations yields ⇔ m = 5 √ 2 m = 19 √2\nTo have a local minimum point at x = 0, we need y (0) = 0, which implies m = 0. Therefore, the integer values of m that satisfy all of the conditions are m = 5 √ 2, m = 19 √ 2, and m = 0. Thus, the answer is D. 7."
},
{
"figure_ref": [],
"heading": "BingChat:",
"publication_ref": [],
"table_ref": [],
"text": "The correct answer is D. 7. Here's how to solve it:\nThe derivative of the function y = -x 4 + 6x 2 + mx is y = -4x 3 + 12x + m. To find the critical points of the function, we need to solve the equation y' = 0.\nThis equation can be rewritten as -4x 3 + 12x + m = 0 or x(4x 2 -12) = -m. Since x is a factor of -m, x must be an integer divisor of -m.\nFor the function to have three critical points, the quadratic equation 4x 2 -12 = -m/x must have two distinct real roots for some value of x. The discriminant of this quadratic equation is (-12) 2 -4 * 4 * (-m/x) = 144+16m/x.\nFor the quadratic equation to have two distinct real roots, the discriminant must be positive. So we need to find the integer values of m for which 144 + 16m/x > 0 for some integer value of x.\nSince x is an integer divisor of -m, we can write m = -kx for some integer k. Substituting this into the inequality above, we get 144 + 16m/x > 0 becomes 144 -16k > 0 or k < 9.\nSo m can be any integer multiple of x such that the absolute value of the multiple is less than 9. There are 7 such values: m = ±x, ±2x, ±3x, ±4x."
},
{
"figure_ref": [],
"heading": "C.1.4 High application level question",
"publication_ref": [],
"table_ref": [],
"text": "Due to the problem's extremely challenging level and the need for complex analytical skills to solve it, ChatGPT is unable to handle so. The solution offered by ChatGPT is inaccurate and unreliable. Additionally, BingChat is unable to resolve this issue. Although it attempts to suggest a seemingly reliable solution strategy, which is wholly wrong, it lacks a solution approach.\nQuestion: Consider complex numbers z that satisfy |z Solution: Applying the triangle inequality, we have:\n2|z| = z 2 -3 -4i ≥ ||z 2 | -|3 + 4i| = ||z| 2 -5 (since z 2 = |z| 2 .\nThe equality occurs when z 2 = k(-3-4i). Hence, we get\n4|z| 2 ≥ (|z|-5) 2 ⇔ |z| 4 -14|z| 2 +25 ≤ 0 ⇔ 7 -2 √ 6 ≤ |z| 2 ≤ 7 + 2 √ 6. Therefore, we have √ 6 -1 ≤ |z| ≤ √ 6 + 1. Thus, we get M = 1 + √ 6 and m = √ 6 -1. Therefore, M 2 + m 2 = 14.\nTherefore, the answer is C. 14."
},
{
"figure_ref": [],
"heading": "ChatGPT:",
"publication_ref": [],
"table_ref": [],
"text": "We have the equation |z 2 -3 -4i| = 2|z| which is equivalent to the system:\n|z 2 -3 -4i| 2 = 4|z| 2 |z| 2 = x 2 + y 2 ≥ 0 Thus: |z 2 -3 -4i| 2 = |(x 2 -y 2 -3) + 2xyi -4i| 2 = (x 2 -y 2 -3) 2 + 4x 2 y 2 -8x 2 + 16y 2 = 4(x 2 + y 2 )|z| 2\nWhich is equivalent to: (x 2 -y 2 -3) 2 + 4x 2 y 2 -8x 2 + 12y 2 = 0. Setting t = x 2 -y 2 , u = xy we have t 2 + 4u 2 -8t + 12y 2 = 0. We have 4t 2 -16y 2 = (2t -4y)(2t + 4y) ≤ 0 so |t| ≤ 2|y|. Thus:\n0 ≤ (x 2 + y 2 ) = |z| 2 = 1 2 (|z| 2 + |z| 2 ) ≤ 1 2 (M 2 + m 2 ). And (x 2 + y 2 ) 2 = t 2 + 4u 2 ≤ 4|y| 2 (|t| + 4|y|) ≤ 8|y| 3 . Therefore: 0 ≤ |z| 2 ≤ 2 √ 2|y| 3/2 . In general, we have M 2 + m 2 ≤ 8 √ 2.\nBingChat: This is a complex math question. Would you like me to explain how to solve it? To solve this question, we can rewrite the expression |z 2 -3 -4i| = 2|z| as |z 2 -3 -4i| = |2z|. Then, we apply the theorem on the distance between two points in the complex plane to infer that z 2 is the midpoint of the two complex points z and (3 + 4i). From there, we can find the value of z and calculate the value of M 2 + m 2 . The correct answer to this question is D. 11 + 4 √ 6."
},
{
"figure_ref": [],
"heading": "C.2 Literature",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "C.2.1 Question and Answer",
"publication_ref": [],
"table_ref": [],
"text": "Both ChatGPT and BingChat are capable of correctly determining the poetry genre from an excerpt and offering pertinent explanations. The characteristics that distinguish poetry apart from various kinds of writing, including meter, rhyme, and imagery, can be recognized by ChatGPT and BingChat. Employing this knowledge, they can next categorize the sample into the proper poetry genre and explain their categorization. Although ChatGPT and BingChat are capable of understanding the inquiries, the intended language might not always be possible. LLMs have reduced abilities in both word positioning and language comprehension. This drawback results from the fact that LLMs rely on statistical correlations and patterns revealed in training data for their operation. They may choose the wrong words or use the wrong wording because they may not fully comprehend the underlying meanings and implications of words and phrases. LLMs find it difficult to precisely capture the intended meaning of a given text since language is complex and can have various meanings and interpretations. "
},
{
"figure_ref": [],
"heading": "BingChat:",
"publication_ref": [],
"table_ref": [],
"text": "Trong đoạn thơ trên, có một số từ ngữ diễn tả đời sống nghèo khó và vất vả của người mẹ. Đó là:\n\"tù túp lều lơp lá lơp tranh\": diễn tả ngôi nhà nghèo khó, đơn sơ của người mẹ.\n\"cắt cuống nhau bằng lưỡi liềm\": diễn tả công việc vất vả của người mẹ trong việc cắt cỏ để nuôi gia súc.\n\"bàn chân thô quanh năm bùn lấm\": diễn tả sự vất vả và khó khăn trong cuộc sống của người mẹ khi phải làm việc chân trần trong bùn đất.\n\"chưa một lần ướm qua sử sách\": diễn tả sự thiếu học vấn và kiến thức của người mẹ do hoàn cảnh khó khăn.\nNhững từ ngữ này giúp người đọc hình dung được cuộc sống nghèo khó và vất vả của người mẹ trong bài thơ.\nThe meanings of the aforementioned poems can be presented on ChatGPT and BingChat, respectively. In this setting, LLMs' reading comprehension skills are quite strong. ChatGPT and BingChat can employ their language comprehension skills to analyze the meanings of poetry by determining the tone, mood, and overall message of the poem. Additionally, they are able to identify literary techniques like metaphors, similes, and symbolism, which can serve to further explain the meaning of the poetry. -Con người Việt Nam phải sống trong muôn vàn khó khăn vất vả nhưng họ vẫn ngời sáng phẩm chất tinh thần tốt đẹp.\n-Lòng biết ơn, sự kiên cường.\n-Lòng yêu nước căm thù giặc sâu sắc\nChatGPT: Đoạn trích thể hiện hình ảnh một dân tộc Việt Nam chịu đựng và vượt qua khó khăn, gắn bó mạnh mẽ với quê hương và truyền thống văn hóa. Những từ ngữ miêu tả cuộc sống nghèo khó, vất vả, như \"tù túp\", \"lều lơp lá lơp tranh\", \"cắt cuống nhau bằng lưỡi liềm\", \"bàn chân thô quanh năm bùn lấm\" đã thể hiện một cuộc sống khắc nghiệt và cực kỳ khó khăn của người dân tộc Việt Nam.\nTuy nhiên, bài thơ cũng thể hiện tình yêu thương sâu sắc của người dân tộc Việt Nam dành cho đất nước và truyền thống văn hóa của mình. Những câu thơ \"dù uống nước đâu lòng vẫn nhớ nguồn\" hay \"tập con bước vịn vào ca dao tục ngữ\" đã đưa ra một thông điệp rõ ràng về sự quan tâm và tôn trọng giá trị của truyền thống dân tộc.\nBingChat: Trong đoạn trích trên, hình ảnh dân tộc Việt Nam được thể hiện một cách sinh động và đầy ý nghĩa. Dân tộc Việt Nam được miêu tả là một dân tộc kiên cường, bền bỉ và không ngừng vươn lên trong cuộc sống. Điều này được thể hiện qua hình ảnh người mẹ vất vả nuôi con bằng những công việc nặng nhọc và dạy con bằng những câu ca dao tục ngữ quý giá. Dân tộc Việt Nam cũng được miêu tả là một dân tộc biết ơn đời, biết nhớ về nguồn cội và gốc rễ của mình. Điều này được thể hiện qua câu thơ \"dù uống nước đâu lòng vẫn nhớ nguồn\". Ngoài ra, hình ảnh dân tộc Việt Nam còn được thể hiện qua sự đoàn kết và sức mạnh khi đứng dậy làm người, vượt qua mọi khó khăn và thử thách. Đây là một hình ảnh đẹp và đầy tự hào của dân tộc Việt Nam.\nBingChat and ChatGPT may both write a 200-word review on a certain subject. LLMs are capable of writing meaningful paragraph structures, with ChatGPT outperforming BingChat in this regard. This is due to the fact that BingChat and ChatGPT are both language models created to produce human-like language. They are able to comprehend the meaning and context of the assigned issue and provide a well-organized response.\nII. Writing (7.0 points) Câu 1: Từ nội dung đoạn trích ở phần đọc hiểu, hãy viết một đoạn văn khoảng 200 chữ trình bày suy nghĩ của anh chị về sức mạnh của tinh thần vượt khó trong cuộc sống."
},
{
"figure_ref": [],
"heading": "Đáp án:",
"publication_ref": [],
"table_ref": [],
"text": "Yêu cầu chung:\n- Trong cuộc sống, chúng ta sẽ gặp phải rất nhiều khó khăn và thử thách. Đôi khi chúng ta sẽ cảm thấy mệt mỏi, bất lực và không biết phải làm gì để vượt qua. Tuy nhiên, sức mạnh của tinh thần vượt khó là điều giúp chúng ta có thể đứng lên và tiếp tục đi tiếp trên con đường mà mình đã chọn.\nTinh thần vượt khó giúp chúng ta tạo ra những kế hoạch, đề ra mục tiêu, tập trung vào những điều tích cực và đối mặt với những khó khăn trong cuộc sống. Nếu không có sự mạnh mẽ của tinh thần vượt khó, chúng ta có thể không thể đạt được những thành công lớn trong cuộc sống.\nTuy nhiên, để có được sức mạnh của tinh thần vượt khó, chúng ta cần phải có niềm tin vào bản thân, tìm kiếm nguồn động lực bên trong và xây dựng một tâm trạng tích cực. Chúng ta cũng cần nhận ra rằng, sức mạnh của tinh thần vượt khó không phải là điều mà ta có thể có được từ ngày qua ngày, mà đó là một quá trình dài và cần phải được bồi đắp từng ngày.\nVới sức mạnh tinh thần và sự nỗ lực, chúng ta có thể vượt qua mọi khó khăn trong cuộc sống và đạt được những thành công lớn. Hãy học hỏi từ những hình mẫu sống động như mẹ của nhân vật trong đoạn trích để có thể trở thành một người mạnh mẽ và thành đạt.\nBingChat: Sức mạnh của tinh thần vượt khó là một trong những yếu tố quan trọng giúp con người vượt qua những khó khăn và thử thách trong cuộc sống. Đoạn trích trên đã thể hiện rõ sức mạnh này qua hình ảnh người mẹ vất vả nuôi con trong hoàn cảnh nghèo khó và dạy con bằng những câu ca dao tục ngữ quý giá. Dù cuộc sống có nhiều khó khăn và thử thách, người mẹ vẫn kiên cường và không ngừng vươn lên để nuôi con khôn lớn và dạy con những bài học quý giá về cuộc sống.\nSức mạnh của tinh thần vượt khó giúp con người có động lực để tiếp tục phấn đấu và không bao giờ từ bỏ ước mơ và hoài bão của mình. Nó giúp chúng ta có thể đối mặt với những khó khăn và thử thách một cách lạc quan và tích cực, tìm ra những giải pháp để vượt qua chúng. Sức mạnh này còn giúp chúng ta có thể học hỏi được nhiều điều từ những trải nghiệm khó khăn và trưởng thành hơn trong cuộc sống. Trong cuộc sống, chúng ta sẽ gặp phải nhiều khó khăn và thử thách. Tuy nhiên, nếu chúng ta có sức mạnh của tinh thần vượt khó, chúng ta sẽ có thể đối mặt với chúng một cách dũng cảm và vượt qua chúng để tiếp tục bước đi trên con đường của mình. Đó là sức mạnh tuyệt vời của tinh thần vượt khó trong cuộc sống.\nThe hidden meaning, utilized to examine the main idea of literary works, cannot yet be understood by LLMs, despite their ability to comprehend and present the explicit meaning. Due to their inability to comprehend the specificity of literature-the use of the particular to allude to the general-LLMs still have significant limitations. It's possible that BingChat's search engine functionality-which is inappropriate for producing a critical essay-is the reason ChatGPT performs better than it does in terms of literature. While ChatGPT is a huge language model capable of producing coherent and contextually suitable responses to natural language input, BingChat is a search engine mechanism meant to deliver answers to inquiries based on keyword matching and retrieval of pertinent resources. Therefore, ChatGPT would be better suited to creating an essay expressing perspectives and evaluations on a topic like literature, whereas BingChat's capabilities would be more constrained. "
},
{
"figure_ref": [],
"heading": "C.3 English",
"publication_ref": [],
"table_ref": [],
"text": "Both ChatGPT and BingChat are unable to respond to inquiries requiring the identification of the word whose underlined portion is pronounced differently from the other three. The amount of phonetic knowledge and analysis required for this work is higher than that LLMs currently possess or are equipped to accomplish. Both ChatGPT and BingChat perform well on the remaining portions of the exam, proving that LLMs have received adequate training in English vocabulary, grammar, and reading comprehension. Question: Read the following passage and mark the letter A, B, C, or D on your answer sheet to indicate the correct answer to each of the questions.\nIt is often held that technology in general does little to help family relationships as, in many cases, family members are more interested in their smartphones and the TV than they are in each other. Before smartphones the TV was blamed the most for breakdowns in communication between parents and children. It was argued that \"the box\" stop them from talking to each other, and clashes between family members over what programs to watch create barriers and cause resentment. However, TV programs can be booth informative and entertaining. They often provide topics for family members to discuss, bring them closer together rather than driving them further away. What is more, twenty-first century families do not need to watch the same program at the same time. The fact that people can watch what they want results in more time for families to be together rather than less. The smartphone has taken over from the TV as a major reason for the lack of communication between family members. Parents often do not see why their children spend so much time online and fail to see the importance of having an online identity while most teens feel their online image is part of who they are. They often criticize the amount of time their children spend online rather than attempting to understand why they do it. In fact, adults are often not very good examples themselves. Who has not witnessed a table at a restaurant where all the family members are silently focused on their mobile devices? Broadly speaking, smartphones do have some negative effects on family relationships. This technology is still relatively new, so many families have not yet learned how to deal with it.\nParents and children should discuss how they use technology to help them all to make the most of technology instead of blaming it for their problems. BingChat: For question, it can be inferred from the passage that \"A lot of families are still incapable of making good uses of smartphones\", so the correct answer is D."
},
{
"figure_ref": [],
"heading": "C.4 Physics",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "C.4.1 Knowledge level question",
"publication_ref": [],
"table_ref": [],
"text": "Although the knowledge level inquiry is aimed at students, neither ChatGPT nor BingChat are able to deliver precise responses based on the available data. This shows that ChatGPT and BingChat are still not at the level of humans, even at the most basic level of Physics. This might be because the models' training data, which might not have provided enough examples of physics-related reasoning problems, is limited. Furthermore, it's possible that the models lack human-like common sense and intuition.\nQuestion: A harmonic oscillator moves in simple harmonic motion along the Ox axis with the equation x = A cos(ωt + ϕ). When the object passes through the equilibrium position, the magnitude of its acceleration is ChatGPT: Answer: C. ω 2 A. Explanation: To calculate the magnitude of the acceleration of an object when it passes through the equilibrium position, we need to calculate the second derivative of the equation of motion x = Acos(ωt + φ) with respect to time t. The second derivative of x with respect to time t is: a = d 2 x/dt 2 = -Aω 2 cos(ωt + φ) When the object passes through the equilibrium position, cos(ωt + φ) = 1, so: a = -Aω 2 Therefore, the magnitude of the acceleration of the object when it passes through the equilibrium position has a value of ω 2 A."
},
{
"figure_ref": [],
"heading": "BingChat:",
"publication_ref": [],
"table_ref": [],
"text": "The motion equation of a harmonic oscillator moving along the Ox axis is x = Acos(ωt + ϕ). The velocity of the object is calculated using the formula v = dx dt = -Aωsin(ωt + ϕ) and the acceleration of the object is calculated using the formula a = dv dt = -Aω 2 cos(ωt + ϕ) = -ω 2 x. When the object passes through the equilibrium position, x = 0 and the acceleration of the object has its maximum value of a max = ω 2 A. Therefore, the correct answer is C. ω 2 A."
},
{
"figure_ref": [],
"heading": "C.4.2 Comprehension level question",
"publication_ref": [],
"table_ref": [],
"text": "BingChat and ChatGPT both offer precise responses for comprehension level questions. When there are few steps to the solution and it doesn't involve a lot of computation or inference skills, ChatGPT and BingChat are able to provide accurate answers and explanations for physics questions at the comprehension level. This is probably because LLMs have been trained on a lot of text data, such as textbooks and scientific publications, which has given them the ability to comprehend physics concepts and terminology.\nQuestion: An ideal LC oscillation circuit is undergoing free electromagnetic oscillation. The phase shift of the current intensity in the circuit with respect to the charge of a capacitor with magnitude is\nA. π 2 B. π 4 C. π 6 D. π3\nSolution: In the ideal LC oscillation circuit, the current intensity lags the charge of the capacitor by 0.5π in phase. Choose A.\nChatGPT: In an LC oscillator circuit, the free electromagnetic oscillation is determined by the equation:\nd 2 Q dt 2 + 1 LC Q = 0,\nwhere Q is the charge on the capacitor and L, C are the inductance and capacitance of the circuit, respectively. This equation has a solution of the form:\nQ(t) = Q 0 cos(ωt + ϕ), where ω = 1 √\nLC is the angular frequency of the oscillation and ϕ is the phase constant. The current intensity in the LC oscillator circuit is determined by the formula I = dQ dt , therefore: I(t) = -ωQ 0 sin(ωt + ϕ). Since the magnitude of the charge is Q 0 , the charge Q(t) and the current intensity I(t) have a phase difference of π 2 , which means:\nφ π = 1 2 ⇒ φ = π 2 .\nTherefore, the correct answer is ( A) π 2 . BingChat: In an ideal LC oscillating circuit that is undergoing free electromagnetic oscillation, the current intensity in the circuit and the electric charge of a capacitor have a phase difference of π 2 . Therefore, the correct answer is A. Π 2 ."
},
{
"figure_ref": [],
"heading": "C.4.3 Application level question",
"publication_ref": [],
"table_ref": [],
"text": "Both ChatGPT and BingChat are unable to respond to questions at the application level with precise information. While BingChat has made a modest improvement, ChatGPT is unable to offer appropriate solving techniques. Nevertheless, both ChatGPT and BingChat still struggle with physics problems at this level. Both ChatGPT and BingChat might find it difficult to give precise answers when the questions get more complicated and call for more inference and problem-solving abilities, as they might not have human-level critical thinking and reasoning abilities.\nQuestion: At a location on the Earth, two simple pendulums of lengths l and l + 45 cm are both excited to perform harmonic oscillations. The initial time is chosen when the suspension wire of both pendulums is vertical. When the maximum deviation angle of one pendulum from the vertical direction is third maximized, the other pendulum is in a position where its suspension wire is vertical for the second time (not including the initial time). The value of l is "
},
{
"figure_ref": [],
"heading": "ChatGPT:",
"publication_ref": [],
"table_ref": [],
"text": "We solve the problem by applying the law of conservation of mechanical energy. Let T be the period of a simple pendulum, m be the mass of the pendulum, and g be the acceleration due to gravity. Then, the mechanical energy of the harmonic oscillation of the pendulum is given byE = 1 2 ml 2 ω 2 , where ω = 2π T . The maximum angle of displacement of the pendulum from the vertical is given by θ = l l+45,cm θ max . herefore, when the pendulum is at its third maximum displacement, the pendulum is at a position where θ = 1 2 θ max . According to the law of conservation of mechanical energy, the mechanical energy of two simple pendulums at θ = θ max must be equal to the total mechanical energy of two simple pendulums at θ = 1 2 θ max . Thus, we have\n1 2 ml 2 ω 2 max + 1 2 m(l + 45cm) 2 ω 2 max = 1 2 m l 2 2 2π T 2 + 1 2 m l + 45cm 2 2 2π T 2 .\nSolving this equation, we get l = 90 cm =>A."
},
{
"figure_ref": [],
"heading": "BingChat:",
"publication_ref": [],
"table_ref": [],
"text": "The correct answer is A. 90 cm. The oscillation period of a simple pendulum with length l can be calculated using the formula: T = 2π l g . Therefore, the oscillation periods of two simple pendulums with lengths l and l + 45 cm are respectively T 1 = 2π l g and T 2 = 2π l+45 g . Since when the magnitude of the angle between the suspension wire of one pendulum and the vertical direction is at its maximum for the third time, the other pendulum is in a position where its suspension wire coincides with the vertical direction for the second time (not counting the initial time), we have:\n3T 1 T 2 = 2 ⇔ T 12 T 22 = 4 9 ⇔ l l+45 = 4 9 ⇔ l = 90(cm)."
},
{
"figure_ref": [],
"heading": "C.4.4 High application level question",
"publication_ref": [],
"table_ref": [],
"text": "ChatGPT gives an answer without requiring data from images for questions at the high application level, particularly for issues that ask for high-level reasoning from text and images, and therefore provides erroneous information. BingChat, on the other hand, asks for information rather than offering a response. Physics questions of this nature go beyond what ChatGPT and BingChat are currently capable of. The high application level physics problems demand a high level of reasoning, rationale, and problem-solving abilities that are currently beyond the capabilities of LLMs. Accurately answering questions of this nature may also be challenging due to a lack of understanding of visual information and an inability to combine information from many sources.\nQuestion: There is a standing wave on a standing wave. The figure shows part of the shape of the string at times t 1 and t 2 = t 1 + 0.8 s (solid and dashed lines). M is a point at the antinode. The velocity of M at times t 1 and t 2 are v 1 and v 2 respectively with v2 v1 = 3 We\nhave |v| = ω √ A 2 -u, therefore we obtain v2 v1 = 1-( u 2 A ) 2 1-( u 1 A ) 2 = 3 √ 6 8 . From the graph u 1 = +2mm u 2 = -3mm . 1 --3 A 2 1 -+2 A 2 = 3 √ 6 8 ⇒ 1 -3 A 2 1 -2 A 2 = 54 64 ⇒ 64 -64 9 A 2 = 54 -54 4 A 2 ⇒ A = 6 cm\nMoreover, from the graph, we have\nt = π 2 + cos -1 2 6 + sin -1 3 6 ω = 0.8s ⇒ ω = π 2 + cos -1 2 6 + sin -1 3 6 0.8 = 4.16 rad s\n. The maximum speed of the wave node is v max = ωA. Therefore, the closest value to v max is 6 × 4.16 = 24.96 = 25 cm s . Choose C. ChatGPT: We can solve this problem by using formulas related to standing waves and harmonic motion. Firstly, we can calculate the wavelength of the standing wave on the string: λ = 2L = 2×0.4 = 0.8(m) Using the formula for wave velocity: v = f λ We can obtain the frequency of the standing wave on the string: f = v λ = 20 0.8 = 25(Hz) We know that the velocity of element M reaches its maximum when its acceleration vector is opposite to its displacement vector. We also know the ratio between the velocity of element M at two different times:\nv2 v1 = 3 √6\n8 . Therefore, we can calculate the velocity of element M at times t1 and t2:\nv 1 = f λ 2 = 25×0.8 2 = 10(m/s) v 2 = 3 √ 6 8 v 1 ≈ 14.\n43(m/s) Since the acceleration vector of element M is opposite to its displacement vector between t1 and t2, the magnitude of this acceleration vector will be the difference between these two velocities:\na = |v 2 -v 1 | ≈ 4.43(m/s 2 )\nWe use the formula for velocity in harmonic motion:\nv max = √ 2ad\nWhere d is the distance that element M moves between t1 and t2. Since element M moves in harmonic motion and has maximum velocity at one point in the cycle, this distance d will be half the wavelength: d = λ 2 = 0.4(m) Substituting the values into the formula, we get: v max = √ 2ad ≈ 22(m/s) Since the problem asks for the value of v max closest to one of the answer choices, the most accurate answer is (D) 22 cm. BingChat: Your question seems like a question in a physics test. However, I cannot solve this question without additional information about the attached image or additional information about the conditions of the wire and standing wave."
},
{
"figure_ref": [],
"heading": "C.5 Chemistry",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "C.5.1 Knowledge level question",
"publication_ref": [],
"table_ref": [],
"text": "At the chemistry knowledge level, ChatGPT and BingChat were both able to give precise responses and explanations for the question. It is anticipated that both models could perform well because the inquiry was primarily concerned with the memorization of information without needing judgment. LLMs including ChatGPT and BingChat are frequently utilized in question-answering tasks that require knowledge recall since they are known for their capacity to memorize and retrieve information. BingChat: Glutamic acid is the only molecule among the options that has one amino group (-N H 2 ) and two carboxyl groups (-COOH) in its structure . Amino acids are organic molecules that include both an amino group and a carboxyl group attached to a central carbon atom . The other options have one amino group and one carboxyl group (formic acid, alanine) or one amino group and one carboxyl group plus another amino group (lysine) in their structure . Therefore, glutamic acid is the correct answer."
},
{
"figure_ref": [],
"heading": "C.5.2 Comprehension level question",
"publication_ref": [],
"table_ref": [],
"text": "Both ChatGPT and BingChat are capable of supplying the precise chemical reaction equation and the quantity of CO 2 needed to complete the provided question. There are some differences in their performance, nevertheless. As it produced the incorrect amount of moles for N aHCO 3 , ChatGPT appears to have restrictions when completing computations. For both CO 2 and N aHCO 3 , on the other hand, BingChat was able to provide the precise quantity of moles. However, it selected the incorrect response, indicating that it might be having trouble choosing the proper one. Overall, it appears that both models are able to respond to comprehension-level questions on chemical reactions, but they may be constrained by more challenging questions requiring sophisticated calculations and higher-level reasoning. "
},
{
"figure_ref": [],
"heading": "BingChat:",
"publication_ref": [],
"table_ref": [],
"text": "The thermal decomposition of sodium bicarbonate (N aHCO 3 ) can be represented by the following chemical equation:\n2N aHCO 3 (s) ⇒ N a 2 CO 3 (s) + H 2 O(g) + CO 2 (g) .\nFrom the equation, we can see that 2 moles of N aHCO 3 produce 1 mole of CO 2 . At standard temperature and pressure (STP), 1 mole of any gas occupies a volume of 22.4 liters. Therefore, 3.36 liters of CO 2 is equivalent to 3.36 22.4 = 0.15 moles of CO 2 . Since 2 moles of N aHCO 3 produce 1 mole of CO 2 , 0.15 moles of CO 2 are produced from 0.15 × 2 = 0.3 moles of N aHCO 3 . The molar mass of N aHCO 3 is approximately 84 g/mol, so the mass of 0.3 moles of N aHCO 3 is 0.3 × 84 = 25.2 grams. Therefore, the value of m is 25.2 grams. Choose B."
},
{
"figure_ref": [],
"heading": "C.5.3 Application level question",
"publication_ref": [],
"table_ref": [],
"text": "Even for humans, it can be challenging to answer Chemistry questions requiring one to synthesize several different disciplines. These questions often demand a better comprehension of the relevant ideas as well as the capacity to apply that knowledge to new circumstances. It could be difficult for ChatGPT and BingChat to respond to these questions with precise information. In the example that follows, there is a unique circumstance where ChatGPT can provide a correct response whereas BingChat does not. It's crucial to keep in mind that these models still have limitations when it comes to their capacity to synthesize data and deliver precise solutions to challenging problems."
},
{
"figure_ref": [],
"heading": "Question:",
"publication_ref": [],
"table_ref": [],
"text": "The statements are: (d) Iron objects are less prone to corrosion after being painted on the surface: True. Paint creates a protective layer on the surface of the metal, preventing contact of the metal with air and water, helping to minimize corrosion. Therefore, there are 2 true statements, so the answer is (C) 2.\nBingChat: Choose B.1. The only correct statement is (b). Electrolysis of copper sulfate solution with copper electrodes produces copper metal at the cathode and dissolves copper metal from the anode . Electrolysis of sulfate solution using inert electrodes produces oxygen gas at the anode and hydrogen gas at the cathode . Urea fertilizer does not provide potassium to plants, it only provides nitrogen . Excessive bubbling of carbon dioxide into sodium aluminate solution does not produce precipitate, it produces sodium carbonate and aluminum hydroxide . Iron objects are more resistant to corrosion after being painted on the surface, not harder."
},
{
"figure_ref": [],
"heading": "C.5.4 High application level question",
"publication_ref": [
"b22"
],
"table_ref": [],
"text": "Both ChatGPT and BingChat may find it difficult to answer questions involving high-level reasoning that call for synthesizing several steps, merging numerous chemical reaction equations, or applying complicated science ideas. This is due to the fact that these types of questions frequently call for not only factual information but also the capacity to reason and apply that knowledge in a way that necessitates critical thinking and problem-solving abilities. We can see from the example below that ChatGPT is unable to supply the reaction equation. BingChat's ability to provide the right response in this instance, however, is really unique. It's probable that BingChat used pre-trained data or internet searches to find the right response. The information obtained from the internet or pre-trained data may not always be credible or pertinent to the precise topic being asked, thus it is crucial to keep this in mind while using this strategy.\nQuestion: Heating m grams of the mixture X containing F e(N O3) 2 and KN O 3 (in a sealed container, without air) to constant mass yields solid Y and gas mixture Z. The entire gas mixture Z is absorbed in water to obtain 400 ml of solution E (containing only one dissolved substance) with a pH of 1, without any gas being released. The value of m is: A. 11.24 g B. 4.61 g C. 5.62 g D. 23.05 g Solution: Chemical reaction equation when heating a mixture F e(N O3) 2 and KN O 3 (in a sealed container, without air)\n2F e(N O 3 ) 2 t o -→ F e 2 O 3 + 4N O 2 ↑ +0.5O 2 ↑ (1) KN O 3 t o -→ KN O 2 + 0.5O 2 ↑ (2)\nThe olid Y is (F e 2 O 3 và KN O 2 ) and the gas mixture Z is (N O 2 and O 2 ). PChemical reaction equation when the entire gas mixture Z is absorbed in water\n2N O 2 + 0.5O 2 + H 2 O-→2HN O 3 (3)\nThe liquid E is HN O ChatGPT: To solve this problem, we perform the following steps: Let x be the number of moles of F e(N O3)2 in the mixture X, and y be the number of moles of KNO. Write the chemical equation for the reaction when heating the mixture X::\nF e (NO 3 ) 2 + KN O 3 → KF e (NO 3 ) 4 ↑\nThus, we can deduce that the number of moles of solid Y is x + y, and the gas Z is NO. Write the chemical equations for the reaction of NO with water:\nN O + H 2 O → HN O 2 HN O 2 + H + → N O + 2 + H 2 O N O + 2 + H 2 O → HN O 3 + H + Since\nthe pH of the solution E is 1, the concentration of H+ in the solution is 10 -1 M. Since no gas is released, the initial amount of gas equals the amount of gas absorbed by water. Therefore, we can deduce that the number of moles of NO in gas Z is equal to the number of moles of HNO3 in solution E. Let z be the number of moles of NO (or HN O 3 ) in gas Z (or solution E). From the given information, we have the following system of equations:\nx + y = m/M x/M + 4y/M = 400 × 10 -6 and z = 2y/M z/V = 10 -1 where M and V are the molar mass and molar volume of NO gas, respectively. Solving the above system of equations, we obtain: y/M = 5 × 10 -3 and z = 10 -3 BingChat: The reaction of the mixture X can be written as follows: 2Fe(NO3)2 + 10KNO3 → 2FeO + 10KNO2 + 5O2. The gas mixture Z consists of NO2 and O2. When Z is absorbed into water, only NO2 is absorbed to form HNO3 and NO. The volume of the solution E is 400 mL and its pH is 1, so the concentration of H+ in solution E is 0.1 M. Since HNO3 is a strong acid, the concentration of HNO3 in solution E is also 0.1 M. From the chemical equation: 3NO2 + H2O → 2HNO3 + NO The number of moles of NO2 absorbed into water is twice the number of moles of HNO3 formed, which is 0.04 mol. From the chemical equation: 10KNO3 → 10KNO2 + 5O2 The number of moles of KNO3 in mixture X is twice the number of moles of O2 formed, which is also equal to the number of moles of NO2 absorbed into water, which is 0.04 mol. So the mass of KNO3 in mixture X is mK = nK × MK = 0.04 × 101 = 4.04 (g). From the chemical equation: 2Fe(NO3)2 + 10KNO3 → 2FeO + 10KNO2 + 5O2 The molar ratio between Fe(NO3)2 and KNO3 in mixture X is 1:5. So the mass of Fe(NO3)2 in mixture X is mF = nF × MF = (nK/5) × MF = (0.04/5) × (56+6×14+12×16) = 0.57 (g). Therefore, the mass of mixture X is m = mF + mK = 0.57 + 4.04 = 4.61 (g). So the correct answer is B. 4.61."
},
{
"figure_ref": [],
"heading": "C.6 Biology",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "C.6.1 Knowledge level question",
"publication_ref": [],
"table_ref": [],
"text": "For information retrieval questions at the knowledge level in biology, both ChatGPT and BingChat can offer precise responses and justifications. ChatGPT and BingChat are able to provide precise responses and explanations for these questions because they can retrieve data from both their trained data and the internet.\nQuestion: Which of the following plant groups undergoes the process of photorespiration in the daytime? BingChat: The answer to question is B. C3 plants. Photosynthesis in C3 plants involves the Calvin cycle, which is also known as C3 photosynthesis. In this process, carbon dioxide is fixed into a three-carbon compound (hence the name C3) through the action of the enzyme Rubisco. This process occurs in the chloroplasts of plant cells during daylight hours when the plant is exposed to light."
},
{
"figure_ref": [],
"heading": "C.6.2 Comprehension level question",
"publication_ref": [],
"table_ref": [],
"text": "Both ChatGPT and BingChat gave the wrong responses to comprehension-level questions that call for inference capabilities. This highlights the biological reasoning capabilities of LLM models' limits. Although ChatGPT and BingChat can describe the procedures needed in making recombinant DNA, they are unable to accurately identify their order. As they significantly rely on the data they have been trained on, language models like ChatGPT and BingChat are limited in their ability to understand and interpret complicated information."
},
{
"figure_ref": [],
"heading": "Question:",
"publication_ref": [],
"table_ref": [],
"text": "The steps in the technique of creating recombinant DNA are as follows:\n(1) Cutting the donor DNA and the gene of interest.\n(2) Separating the donor DNA and the DNA carrying the gene of interest.\n(3) Joining the gene of interest with the donor DNA to create recombinant DNA.\nThe correct sequence of steps is: Joining the gene of interest with the donor DNA to create recombinant DNA: Once the gene of interest has been separated, it is ligated or joined with the cut DNA from step 1. After the ligation, the resulting recombinant DNA is reintroduced into a living cell, allowing the new gene to be expressed and produce a new product. Therefore, the correct sequence of steps is C. 1 → 2 → 3.\nA. 1 → 3 → 2. B. 2 → 1 → 3. C. 1 → 2 → 3. D. 3 → 1 → 2."
},
{
"figure_ref": [],
"heading": "BingChat:",
"publication_ref": [],
"table_ref": [],
"text": "The correct sequence of steps in the recombinant DNA technique is C. 1 → 2 → 3. The steps in the recombinant DNA technique include (1) cutting the vector and the gene to be transferred, (2) separating the vector and the DNA carrying the gene to be transferred, and (3) joining the gene to be transferred with the vector to create recombinant DNA. The correct sequence of these steps is C. 1 → 2 → 3."
},
{
"figure_ref": [],
"heading": "C.6.3 Application level question",
"publication_ref": [],
"table_ref": [],
"text": "For challenges in biology that need intricate inference and image processing, neither ChatGPT nor BingChat can give the right response. Both ChatGPT and BingChat are unable to draw conclusions about human genetics in the specific instance of the following questions. Both ChatGPT and BingChat lack the tools necessary to interpret photos or make complicated inferences, particularly in the field of biology where a thorough knowledge of the subject is necessary. The given question necessitates the reader to evaluate a pedigree chart, which entails deciphering genetic data and comprehending the pattern of how a specific trait is inherited. In order to establish the genotypes of the individuals in the pedigree, this is a difficult task that needs a thorough understanding of genetics and the capacity to comprehend complex data. Despite having access to a lot of data, neither ChatGPT nor BingChat can reason or deduce sophisticated knowledge in the same way that a human specialist in the subject of genetics could. As a result, neither ChatGPT nor BingChat could give an appropriate response to this question."
},
{
"figure_ref": [],
"heading": "Question:",
"publication_ref": [],
"table_ref": [],
"text": "The following pedigree describes the inheritance of phenylketonuria in humans caused by one of the two alleles of a specific gene. According to the theory, how many individuals in the following pedigree can accurately determine their genotype? + Generation 1: Pair I1: An affected male marries a normal female and they have 2 normal males; Pair I2: Both a normal male and a normal female, they have 1 normal female and 1 affected male. + Generation 2: The normal male offspring of Pair 1 marries a normal female offspring of Pair 2 (Pair II1); Pair II2: A normal male marries an affected female. + Generation 3: Pair II1 has 1 normal male and 1 normal female offspring, Pair II2 has 1 normal female and 1 affected female offspring.\nA. 10 B. 6 C. 9 D. 11 Solution: PKU is determined by a recessive gene and is not linked to the sex chromosomes. By analyzing the pedigree, the following information is gathered:\nStep 1: Determining the nature of the gene and its location + The affected individuals (III-4, II-4, II-6, and I-1) indicate that the disease is recessive. + The presence of the disease in both males and females suggests that it is not linked to the Y chromosome. + The fact that the daughter (III-4) of an unaffected father (II-5) is affected rules out the disease being linked to the X chromosome. Therefore, the gene responsible for PKU is located on an autosomal chromosome.\nStep 2: Determining the genotypes of individuals in the pedigree + Affected individuals (III-4, II-4, II-6, and I-1) must have the homozygous recessive genotype (aa). + Unaffected individuals who have affected children (II-1, II-2, I-3, and I-4) must be heterozygous carriers (Aa). + Unaffected individuals who have unaffected children (II-5 and III-3) can either be homozygous dominant (AA) or heterozygous carriers (Aa). Based on these findings, the genotypes of 10 individuals in the pedigree can be determined. The affected individuals have the genotype aa, while the unaffected individuals have either AA or Aa genotypes. Choose A.\nChatGPT: To accurately determine their genotype, individuals need to have both affected and unaffected offspring. By examining the given information, we can determine that only individuals in Pair II1 and Pair II2 have both affected and unaffected offspring. Therefore, the correct answer is B. 6."
},
{
"figure_ref": [],
"heading": "BingChat:",
"publication_ref": [],
"table_ref": [],
"text": "The answer to question is D. 11. The following pedigree describes the inheritance of phenylketonuria in humans due to one of two alleles of a gene. According to theory, a maximum of how many people in the following pedigree can be accurately determined for their genotype? + Generation 1: Pair I1: male with disease marries normal female and has 2 normal males; Pair I2: both male and female are normal, giving birth to 1 normal female and 1 male with disease. + Generation 2: Normal male, child of pair 1, marries normal female, child of pair 2 (pair II1); Pair II2, normal male marries female with disease. + Generation 3: Pair II1 has 1 normal male and 1 normal female; Pair II2 has 1 normal female and 1 female with disease. The maximum number of people in the pedigree whose genotype can be accurately determined is D. 11."
},
{
"figure_ref": [],
"heading": "C.6.4 High application level question",
"publication_ref": [],
"table_ref": [],
"text": "At the high application level in biology, ChatGPT and BingChat might not be able to infer enough details to give the right answer. There are situations, nevertheless, in which these LLMs can still provide precise responses. This can be a result of the LLMs' prior experience with equivalent questions or their internet searches for pertinent solutions.\nQuestion: Given the codon encoding for each amino acid in the table:"
},
{
"figure_ref": [],
"heading": "Amino acid",
"publication_ref": [],
"table_ref": [],
"text": "Lue Trp His Arg Codon 5'XUU3'; 5'XUX3'; 5'XUA3'; 5'XUG3' 5'UGG3' 5'XAU3'; 5'XAX3' 5'XGU3'; 5'XGX3'; 5'XGA3'; 5'XGG3'\nThe encoding triplets are the sets of three codons that correspond to the encoding of amino acids, and the termination triplet corresponds to the stop codon on mRNA. Assuming a bacterial gene sequence synthesizing mRNA with the following nucleotide sequence:\nTemplate strand for mRNA synthesis 3'TAXGAAAXXGXXGTAGXAATT5' mRNA 5'AUGXUUUGGXGGXAUXGUUAA3'\nIt is known that each point mutation involving the substitution of a nucleotide pair in this gene sequence creates a new allele. According to the theory, how many of the following statements are correct? I. There can be a maximum of three encoding triplets when a substitution mutation occurs at any third position without changing the type of amino acid in the polypeptide chain. II. There is one encoding triplet when a mutation occurs at the third position, which always changes the type of amino acid in the polypeptide chain. III. There are two encoding triplets when a mutation occurs at the third position, resulting in the early appearance of a stop codon.\nIV. There is one encoding triplet when a mutation occurs at the third position, leading to the absence of a start codon on the mRNA produced from this gene. "
},
{
"figure_ref": [],
"heading": "BingChat:",
"publication_ref": [],
"table_ref": [],
"text": "The number of correct statements is A. 1. Statement I is correct because there are three codons (XUU, XUA, and XUG) that encode the same amino acid (Leu) and would not change the amino acid sequence of the polypeptide chain if a point mutation occurred at the third position of these codons. Statements II and III are incorrect because a point mutation at the third position of any codon could either change the encoded amino acid or introduce a premature stop codon, depending on the specific nucleotide substitution that occurs.\nIn this case, BingChat provided the correct number of correct statements, however, the explanation was not accurate. This indicates that alleles e and f are different and not dominant over each other, as the F1 offspring have different phenotypic ratios. Based on these results, we can conclude: I. False. There is only one gene determining fur color with four alleles on the autosomes. Therefore, there can be a maximum of four different genotypes. II. False. Based solely on the F1 phenotypic ratios, we cannot accurately determine the genotypes of the P mice. For example, alleles a and b can both produce black fur, so we cannot determine which P mouse carries allele a and which carries allele b. III. False. This cannot happen because fur color is determined by one gene and is not related to the mating between mice. IV. True. Allele a is dominant to allele f, and allele f is dominant to allele e. Therefore, when a mouse with allele a mates with a mouse with allele f, the offspring can have a ratio of 50% black fur and 50% yellow fur. Hence, the correct answer is B. 1."
},
{
"figure_ref": [],
"heading": "Question",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "BingChat:",
"publication_ref": [],
"table_ref": [],
"text": "Based on the results of the above crosses, we can infer some information about the genotypes of the P mice. For cross 1, the F 1 phenotype ratio indicates that both P mice have the genotype AaBb (with A and B being dominant alleles). For cross 2, the F 1 phenotype ratio indicates that mouse c has the genotype AAbb and mouse d has the genotype aaBB. For cross 3, the F 1 phenotype ratio indicates that mouse e has the genotype aaBb and mouse f has the genotype Aabb. With this information, we can determine that statement I is correct because there are two genotypes that determine the brown fur phenotype: Aabb and aaBb. Statement II is incorrect because we can only accurately determine the genotype of 5 out of 6 P mice (excluding mouse b). Statement III is incorrect because if mouse d mates with mouse e, the offspring will have a phenotype ratio of 100% white fur. Statement IV is correct because if mouse a mates with mouse f, the offspring will have a ratio of 50% black fur: 50% yellow fur. The answer is A."
},
{
"figure_ref": [],
"heading": "C.7 History",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "C.7.1 Knowledge level question",
"publication_ref": [],
"table_ref": [],
"text": "Due to their access to a variety of previous data via their training data and the internet, ChatGPT and BingChat are able to respond to queries concerning historical information.\nTo produce an answer, they can examine historical occasions, figures, and locations and extract pertinent data from texts. ChatGPT: To answer this question, we need to look at the chart and compare the GDP values of the Philippines and Thailand in 2015 and 2019. From the chart, we can see that the GDP of both countries increased from 2015 to 2019. However, to determine which statement is correct, we need to compare the growth rate of the GDP of the two countries. wrote an article distorting the personnel planning policy of agency X and fabricated information about Mr. K's private life, then publicly posted it on social media, affecting Mr. K's reputation. This behavior violates Mr. K's right to legal protection of honor and dignity."
},
{
"figure_ref": [],
"heading": "D Vietnamese Student Score Spectrum",
"publication_ref": [],
"table_ref": [],
"text": "We compare the performance of LLMs with the results achieved by students in each year due to the different structures and difficulty levels of the exams in each year. This allows us to evaluate the capability of LLMs compared to human-level performance The score spectrum of students is released annually by the Vietnamese Ministry of Education and Training, which we have collected for the years 2019-2022 ( 2019, 2020, 2021, 2022). For instance, Figure 9 shows the mathematics score spectrum of Vietnamese students in 2022. The result of the analysis of the score distribution of the 2022 high school graduation exam in Mathematics shows that there were 982,728 candidates who participated in the exam, with an average score of 6.47 points and a median score of 6.8 points. The most attained score was 7.8 points. There were 186,222 candidates scored below the average (accounting for 18.95%). "
},
{
"figure_ref": [],
"heading": "D.1 Mathematics",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "",
"publication_ref": [],
"table_ref": [],
"text": "Question: My mother takes vitamins regularly as she believes that it benefit her mind and skin.\nA. takes B. regularly C. it D. mind Solution: Pronoun: Referencing language. The word 'it' is referring back to the vitamins mentioned earlier, but it's incorrect to use a singular pronoun. It should be corrected to 'they'. Choose C.\nChatGPT: The correct answer is D. The sentence should read: \"My mother takes vitamins regularly as she believes that they benefit her mind and skin.\" The pronoun \"it\" should be replaced with \"they\" to refer to \"vitamins\", and \"mind\" should be changed to \"mental health\" to use more appropriate language.\nBingChat: For question, the underlined part that needs correction is \"it\" which should be changed to \"they\" to match the plural subject \"vitamins\", so the correct answer is C."
},
{
"figure_ref": [],
"heading": "C.3.3 Communication question",
"publication_ref": [],
"table_ref": [],
"text": "Question: Mark the letter A, B, C, or D to indicate the correct response to each of the following exchanges.\nQuestion 21: Jack and Linh are talking about technology. Jack: \"I think robots will replace teachers in the near future.\" Linh: \"..... We still need teachers to inspire students\" A. I don't think so B. It's true C. You're right D. I couldn't agree more Solution: Jack and Linh are talking about technology, and Jack gives his opinion that robots will replace teachers in the future. Linh may express a counterargument that \"We still need teachers to inspire students.\" Therefore, the answer expressing disagreement is \"I don't think so\". Choose A."
},
{
"figure_ref": [],
"heading": "ChatGPT: A. I don't think so",
"publication_ref": [],
"table_ref": [],
"text": "BingChat: Here are the correct answers to your questions: A. I don't think so"
},
{
"figure_ref": [],
"heading": "C.3.4 Reading Fill-in-the-Blank question",
"publication_ref": [
"b33",
"b34"
],
"table_ref": [],
"text": "Question: Read the following passage and Mark the letter A, B, C, or D on your answer sheet indicate the correct word or phrase that best fits each of the numbered blanks.\nKeeping fit and healthy may seem difficult, but there are a few easy-to-follow guidelines. Firstly, a balanced diet (34) ... means selecting food that is low in salt and sugar is absolutely essential. Experts recommend reducing the amount of fat in our diet, as too much can (35) .... heart problems. "
},
{
"figure_ref": [],
"heading": "ChatGPT: D. which",
"publication_ref": [],
"table_ref": [],
"text": "BingChat: the correct word to fill in the blank is \"which\", so the correct answer is D."
},
{
"figure_ref": [],
"heading": "C.3.5 Reading comprehension question",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Solution:",
"publication_ref": [],
"table_ref": [],
"text": "The \"Special War\" strategy of the United States in South Vietnam: a new type of colonial invasion warfare, carried out by the Saigon Army under the command of the US \"adviser\" system, using US weapons, technical equipment, and warfare means to fight against our people's revolutionary movement. Basic plot: \"Use Vietnamese to fight Vietnamese.\" Choose D.\nChatGPT: D. Saigon Army. The Special War strategy (1961)(1962)(1963)(1964)(1965) was a military strategy of the United States implemented in South Vietnam, with the goal of preventing the Communist forces from taking over. This strategy was carried out by the forces of the Saigon Army, along with the support of special forces of the United States, including both special operatives and unmanned aerial vehicles.\nBingChat: D. Saigon Army. The US \"Special War\" strategy in South Vietnam was carried out by the Saigon Army, under the command of the US \"advisory\" system1. Therefore, the correct answer to your question is D. Saigon Army."
},
{
"figure_ref": [],
"heading": "C.7.2 Comprehension level question",
"publication_ref": [],
"table_ref": [],
"text": "To provide answers, LLM models like ChatGPT and BingChat rely on the patterns and connections they discover from the massive volumes of data they are trained on. These models may not have enough data to produce good predictions for historical questions with significant levels of disturbances, which is defined as irrelevant or deceptive information that can affect the accuracy of the answer. Additionally, it can be difficult for LLMs to answer historical questions because they frequently call for filtering and inference skills, such as the capacity to reason and make inferences based on existing facts. As a result, in such circumstances, their accuracy may suffer, which may result in inaccurate responses, as in the case of ChatGPT. BingChat, on the other hand, might have been able to give the right response because of its unique training data or architecture, or it might have been able to filter out the irrelevant information and draw the right conclusions. ChatGPT: The answer is D. Providing aid to all Asian countries. BingChat: C. Proposed and implemented the Marshall Plan. The correct answer is D. Proposed and implemented the Marshall Plan. The Marshall Plan (1948)(1949)(1950)(1951)(1952), also known as the European Recovery Program, was proposed by US Secretary of State Marshall to provide 17 billion in aid to help Western European countries recover their economies with conditions imposed by the US(https://shorturl.at/gwFRU)"
},
{
"figure_ref": [],
"heading": "Question",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "C.7.3 Application level question",
"publication_ref": [],
"table_ref": [],
"text": "To answer historical comparison questions correctly, one must be able to evaluate and combine data from many sources. However, LLMs systems like ChatGPT and BingChat might not have sufficient capabilities to effectively synthesize and analyze this information, therefore they might not be able to provide precise answers to issues of this nature. This is because even the most sophisticated language models may struggle to appropriately process and interpret such requests, which call for a more sophisticated comprehension of historical events and context. ChatGPT: Answer: D. Being directly impacted by two opposing social systems. Explanation: The national liberation movement (1939)(1940)(1941)(1942)(1943)(1944)(1945) and the anti-American resistance war are two wars with different nature. However, the similarity of the two wars is that they both directly affected by two opposing social systems: the developing capitalist society and the communist society. In which, the anti-American resistance war was mainly a war between a communist country and a capitalist country, while the national liberation movement (1939)(1940)(1941)(1942)(1943)(1944)(1945) was a war against French and Japanese colonialism in Indochina.\nBingChat: A. The armed forces play a decisive role in the victory of the revolution. The resistance war against the United States to save the country and the national liberation movement (1939)(1940)(1941)(1942)(1943)(1944)(1945) of the Vietnamese people both have in common that the armed forces play a decisive role in the victory of the revolution. Therefore, the correct answer to your question is A. The armed forces play a decisive role in the victory of the revolution."
},
{
"figure_ref": [],
"heading": "C.7.4 High application level question",
"publication_ref": [],
"table_ref": [],
"text": "ChatGPT and BingChat cannot respond to complex history queries, such as those requesting opinions on individuals or events in various settings. The reason is that in order to answer these historical questions correctly, one must be able to infer and incorporate data from several sources. Although LLMs like ChatGPT and BingChat can give accurate responses regarding certain historical occurrences, they are unable to comprehend and evaluate broader historical contexts, hence they are unable to give accurate responses to questions of this nature. "
},
{
"figure_ref": [],
"heading": "Question",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "C.8.3 Atlas use skill question",
"publication_ref": [],
"table_ref": [],
"text": "Both ChatGPT and BingChat were able to give precise responses to questions in the Atlas pertaining to map reading abilities. They were able to do so, which implies that the models have been trained on data that includes knowledge about the Atlas and that they are able to evaluate and understand maps. This is probably because both models were trained to utilize a significant amount of textual data that contains knowledge on a variety of topics including geography and cartography.\nQuestion: Based on the Vietnam Geography Atlas page 22, please indicate which river the Hoa Binh hydroelectric power plant is located on?\nA. Gấm River. B. Chu River. C. Lô River. D. Đà River."
},
{
"figure_ref": [],
"heading": "Solution:",
"publication_ref": [],
"table_ref": [],
"text": "The Hoa Binh hydroelectric power plant is located on the Đà River. Choose D.\nChatGPT: The Hoa Binh hydroelectric power plant is located on the Đà River. The answer is D."
},
{
"figure_ref": [],
"heading": "BingChat: Answer: D. Đà River",
"publication_ref": [],
"table_ref": [],
"text": "The Hoa Binh hydroelectric plant is located on the Da River. Therefore, answer D is the most correct answer among the options above. The Hoa Binh hydroelectric plant is the largest hydroelectric plant in Vietnam and Southeast Asia, located on the Da River in Hoa Binh province. This plant has an installed capacity of up to 1920 MW and plays an important role in providing electricity to the northern region of Vietnam."
},
{
"figure_ref": [],
"heading": "C.8.4 Application and high application questions",
"publication_ref": [],
"table_ref": [],
"text": "High levels of disturbance in geography questions, which is defined as the difficulty in separating the data because of errors or inconsistencies, might make it difficult for ChatGPT and BingChat to provide accurate responses. This is due to the fact that these models rely on patterns in the data they were trained on, and patterns in data can be disrupted by disturbances, making it challenging to determine the correct response. Furthermore, it may be difficult for language models to effectively answer geography problems because they frequently call for spatial reasoning and contextual comprehension. Therefore, when presented with geography questions that have a lot of disturbances, both ChatGPT and BingChat might become less accurate."
},
{
"figure_ref": [],
"heading": "Question:",
"publication_ref": [],
"table_ref": [],
"text": "The complete territory of our country includes A. Sea, air and archipelago areas, B. Land, sea and air areas. C. Mainland, sea and islands. D. Mainland, islands and air area."
},
{
"figure_ref": [],
"heading": "Solution:",
"publication_ref": [],
"table_ref": [],
"text": "The complete territory of our country includes land, sea, and air areas. Choose B."
},
{
"figure_ref": [],
"heading": "ChatGPT: D. Mainland, islands and air area.",
"publication_ref": [],
"table_ref": [],
"text": "BingChat: Answer: D. mainland, islands and airspace.\nIn order to arrive at a proper response, questions that ask for analytical abilities in geography can involve sophisticated inference and may call for advanced reasoning abilities. While ChatGPT and BingChat can both comprehend and analyze geographic data, they might not possess the sophisticated analytical abilities required to deliver precise responses to issues of this nature. These kinds of problems are incapable of being easily solved by language models alone and frequently require human expertise and awareness of problem-solving techniques. "
},
{
"figure_ref": [],
"heading": "Question",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Solution:",
"publication_ref": [],
"table_ref": [],
"text": "The most suitable type of chart to show changes over multiple years (>3 years) is an area chart. Therefore, the most suitable type of chart to show the changes in the structure of university lecturers by gender in Vietnam from 2015 to 2019 is an area chart. Choose C.\nChatGPT: To show the change in the structure of the number of university lecturers by gender in Vietnam from 2015 to 2019, the column chart is the most appropriate. This is because the column chart will clearly show the difference between the number of male and female university lecturers in each specific year, making it easy to compare and analyze the changes in structure between years. Other types of charts such as line, area or combination charts can also show these changes, but not as clearly and easily understood as the column chart. => Choose D. BingChat: Answer: D. Line. To represent the change in the structure of university lecturers by gender in our country during the period 2015-2019, the most suitable type of chart is a line chart. Therefore, option B is the correct answer. Line charts allow for the representation of changes in values over time and comparison between data groups. In this case, a line chart allows for the representation of changes in the number of male and female lecturers over the years. Would you like to know more information?"
},
{
"figure_ref": [],
"heading": "C.9 Civic Education",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "C.9.1 Knowledge level question",
"publication_ref": [],
"table_ref": [],
"text": "Because they have been trained on law-related datasets, ChatGPT and BingChat can both give precise responses to straightforward legal questions. Both models were developed using enormous volumes of textual data, such as court records, legal texts, and legal papers. As a result, they are able to comprehend and respond to straightforward legal questions that are based on explicit information provided in the text since they have learned patterns and relationships within the legal domain.\nQuestion: According to the regulations of the law, one of the provisions that demonstrates the right to equality among ethnic groups in the economic aspect is that citizens belonging to ethnic groups are all entitled to: A. Borrowing capital for production development. B. Participating in the state machinery. C. Running for National Assembly representative. D. Equality of educational opportunities."
},
{
"figure_ref": [],
"heading": "Solution:",
"publication_ref": [],
"table_ref": [],
"text": "According to the regulations of the law, one of the provisions that demonstrates the right to equality among ethnic groups in the economic aspect is that citizens belonging to ethnic groups are all entitled to: borrowing capital for production development. Choose A.\nChatGPT: The correct answer is A. to borrow money for production development. According to Article 62 of the Constitution of the Socialist Republic of Vietnam in 2013, \"Citizens belonging to different ethnic groups are equal before the law, with equal rights and duties; they enjoy full rights, preferential policies and protection of the State. The State ensures equal rights among ethnic groups in economic, social and cultural fields.\" Therefore, citizens belonging to different ethnic groups are entitled to borrow money for production development, which is one of the contents reflecting the equal rights among ethnic groups in economic field. Other options are not specified in the Constitution."
},
{
"figure_ref": [],
"heading": "BingChat:",
"publication_ref": [],
"table_ref": [],
"text": "The correct answer is A. borrow capital for production development. According to the law, one of the contents that demonstrates the equality between ethnic groups in terms of economics is that citizens of all ethnic groups are entitled to borrow capital for production development. This means that citizens of all ethnic groups have the right to be supported in borrowing capital to develop business production and improve their economic life. The correct answer is A. Borrow capital for production development."
},
{
"figure_ref": [],
"heading": "C.9.2 Comprehension level question",
"publication_ref": [],
"table_ref": [],
"text": "Both ChatGPT and BingChat may have trouble giving precise responses when civic education questions become more complicated and necessitate the interpretation of fundamentals. However, ChatGPT and BingChat gave different responses in the specific case given where the question needed comprehension of market prices in the production and distribution of items. BingChat gave the right response, whereas ChatGPT gave the wrong one. The models' training on different datasets may have had an impact on how they understood and interpreted the concepts connected to market prices, which is why the responses varied.\nQuestion: In the production and circulation of goods, when supply is less than demand, market prices tend to be higher. "
},
{
"figure_ref": [],
"heading": "Solution:",
"publication_ref": [],
"table_ref": [],
"text": "In the production and circulation of goods, when supply is less than demand, market prices are usually higher than the value of the commodity. Choose C. ChatGPT: D. Social consumption demand. Market prices depend on the balance between the supply and demand of goods. When social consumption demand is higher than the quantity of goods supplied, demand exceeds supply, leading to an increase in prices. Conversely, when the supply of goods exceeds social consumption demand, prices will decrease to stimulate consumption."
},
{
"figure_ref": [],
"heading": "BingChat: Answer: C. Value of goods",
"publication_ref": [],
"table_ref": [],
"text": "In the production and circulation of goods, when supply is less than demand, market prices are usually higher than the value of goods in production. This means that when supply is less than demand, it shows that the demand for goods has increased, making the goods scarce and unable to meet the market demand, so prices rise. Therefore, the correct answer is C. Value of goods."
},
{
"figure_ref": [],
"heading": "C.9.3 Application level question",
"publication_ref": [],
"table_ref": [],
"text": "Both ChatGPT and BingChat may not be able to offer accurate responses when questions have a lot of disturbances, as shown by a question for which both models provided the erroneous response (D instead of B). Legal issues frequently entail interpreting the context and intent of laws and regulations as well as their detailed specifics. While these models are capable of delivering factual data and fundamental legal concepts, they may struggle to carry out in-depth legal analysis that necessitates proficiency in legal argumentation and reasoning. "
},
{
"figure_ref": [],
"heading": "Solution:",
"publication_ref": [],
"table_ref": [],
"text": "According to the regulations of the law, citizens are not allowed to exercise their right to vote for representatives of the People's Councils at all levels in the case of: losing civil capacity. Choose B.\nChatGPT: Answer: D. Temporarily detained for investigation. Explanation: According to Article 19 of the Election Law, Vietnamese citizens from the age of 18 and above have the right to vote and be elected. However, if a citizen is temporarily detained for investigation, they are not allowed to exercise their right to vote for representatives of the People's Councils. After the temporary detention period ends, the citizen's right to vote for representatives of the People's Councils will be reinstated.\nBingChat: Answer: D. temporary detention for investigation purposes. Explanation: According to the law, citizens are not allowed to exercise their right to vote for representatives of the People's Councils at all levels in case they are being temporarily detained for investigation purposes. This means that the citizen is not allowed to participate in voting in the election of representatives of the People's Councils at all levels because they are being temporarily detained for investigation purposes. Therefore, the correct answer is D. Temporary detention for investigation purposes."
},
{
"figure_ref": [],
"heading": "C.9.4 High application level question",
"publication_ref": [],
"table_ref": [],
"text": "Due to the intricacy and nuance of legal studies, ChatGPT and BingChat might not be able to distinguish which character in case studies has committed an illegal act. Legal analysis frequently entails a thorough evaluation of a number of different variables, including the particular laws at issue, the parties' intentions, the situation's background, and the particulars of the case. While ChatGPT and BingChat are examples of machine learning models that can rapidly process and analyze massive volumes of data, they might not be able to provide the same level of nuanced interpretation and reasoning that a human legal expert can. These models are only as good as their training data, so if they haven't been exposed to a diverse enough set of legal scenarios and cases, they might find it difficult to correctly identify the appropriate legal analysis in a particular instance. As a result, while ChatGPT and BingChat can offer valuable insights and fundamental legal facts, they might not be able to do so for complex legal questions without human aid.\nQuestion: Agency X consists of Mr. G as the director, Ms. P as the deputy director, Ms. N as the head of the personnel organization department, Mr. K as the chairman of the labor union, and Mr. H as an employee. During the year-end review meeting, due to Ms. N's objection to her perspective on personnel transfers, Mr. G prevented her from speaking, insulted her, and forced her to leave the meeting. Sitting next to them, when Mr. H intended to express his agreement with Ms. N's opinion, Ms. P threatened to put him on the downsizing list to silence him. Upon hearing the conversation between Ms. P and Mr. H, and having a prior conflict with Mr. H, Mr. K took this opportunity to remove Mr. H from the year-end commendation list. Knowing the situation, Ms. V, Mr. H's wife, wrote a fabricated article distorting the agency X's personnel planning policy and spreading false information about Mr. K's personal life, which she publicly posted on social media, affecting Mr. K's reputation. Who among the following has violated both the rights protected by the law regarding honor and human dignity, as well as the freedom of speech of the citizens? ChatGPT: D. Ms. V and Ms. P. In this case, Mr. G violated the freedom of speech of Ms. N by preventing her from speaking and forcing her to leave the meeting. Ms. P threatened to put Mr. H on the list of job cuts to silence him, which is also a violation of his freedom of speech. Ms. V violated the rights protected by the law regarding honor and human dignity of Mr. K by distorting the organizational policy and fabricating information about his personal life on social media. BingChat: Answer: C. Mrs. P, Ms. V and Mr. G. Explanation: C. Mrs. P, Ms. V and Mr. G have violated both the right to legal protection of honor and dignity and the right to freedom of speech of citizens. Mr. G prevented Ms. N from speaking and insulted her, forcing her to leave the meeting. This behavior violates Ms. N's right to freedom of speech and the right to legal protection of honor and dignity. Mrs. P threatened to put Mr. H on the list of staff reductions to force him to stop speaking. This behavior violates Mr. H's right to freedom of speech. Ms. V"
}
] | The VNHSGE (VietNamese High School Graduation Examination) dataset, developed exclusively for evaluating large language models (LLMs), is introduced in this article. The dataset, which covers nine subjects, was generated from the Vietnamese National High School Graduation Examination and comparable tests. 300 literary essays have been included, and there are over 19,000 multiple-choice questions on a range of topics. The dataset assesses LLMs in multitasking situations such as question answering, text generation, reading comprehension, visual question answering, and more by including both textual data and accompanying images. Using ChatGPT and BingChat, we evaluated LLMs on the VNHSGE dataset and contrasted their performance with that of Vietnamese students to see how well they performed. The results show that ChatGPT and BingChat both perform at a human level in a number of areas, including literature, English, history, geography, and civics education. They still have space to grow, though, especially in the areas of mathematics, physics, chemistry, and biology. The VNHSGE dataset seeks to provide an adequate benchmark for assessing the abilities of LLMs with its wide-ranging coverage and variety of activities. We intend to promote future developments in the creation of LLMs by making this dataset available to the scientific community, especially in resolving LLMs' limits in disciplines involving mathematics and the natural sciences. Keywords GPT-3.5 • GPT-4 • ChatGPT • Bing AI Chat • large language models • dataset • Vietnamese high school graduation examination | VNHSGE: VietNamese High School Graduation Examination Dataset for Large Language Models https://github.com/Xdao85/VNHSGE | [
{
"figure_caption": "The number of real solutions of the equation $|f(x^3-3 x)|=\\frac{2}{3}$ is: A. 6 B. 10 C. 3 D. 9 Solution: From the graph of the function $y=f(x)$, we deduce that the graph of the function $y=|f(x)|$ is:",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 1 :1Figure 1: Formatted question and LLMs response.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "3.1) even though they both score well on other English languages topics like grammar and vocabulary (C.3.2), communication (C.3.3), reading fill-in-the-blank (C.3.4), and reading comprehension (C.3.5). Both ChatGPT and BingChat have been taught the rules and patterns of the English language, including grammar and vocabulary, through training on large English text data. Additionally, they receive instruction on how to comprehend and produce natural language, which involves reading fill-in-the-blank passages and reading comprehension. Though it's possible that neither BingChat nor ChatGPT received adequate training in pronunciation and stress.Physics: ChatGPT and BingChat can solve physics questions at the knowledge and comprehension levels (C.4.3 and C.4.2) which are relatively simple questions about physics topics. However, they are unable to answer questions at the application and high application levels (C.4.3 and C.4.4), which frequently call for substantial knowledge and skills in understanding and applying concepts to solve problems.Chemistry:ChatGPT and BingChat can respond to questions at the knowledge level (C.5.1) by memorizing facts. They often fail to generate the right response to questions at the comprehension level (C.5.2). Neither ChatGPT nor BingChat typically can provide accurate answers for challenging questions at the application level (C.5.3) and high application level (C.5.4) because these types of questions demand the capacity to infer from multiple chemical reactions and high-level synthesis knowledge.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: Comparison of ChatGPT and BingChat performances on VNHSGE dataset.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 33Figure 3 illustrates a comparison of the performance among ChatGPT, BingChat, and Vietnamese students in three core subjects: mathematics (D.1), literature (D.2), and English (D.3). These subjects are integral parts of the exam and are required for all students.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: Comparison in core subjects.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 Figure 4 :44Figure 4 depicts a comparison of the performance among ChatGPT, BingChat, and Vietnamese students in the natural combination, including physics (D.4), chemistry (D.5), and biology (D.6), respectively.",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "44",
"figure_type": "figure"
},
{
"figure_caption": "Figure 55Figure5presents a comparison of the performance among ChatGPT, BingChat, and Vietnamese students in the social combination, including history (D.7), geography (D.8), and civic education (D.9), respectively.",
"figure_data": "",
"figure_id": "fig_7",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 :5Figure 5: Comparison in social combination.",
"figure_data": "",
"figure_id": "fig_8",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "GPT- 4 Figure 6 :46Figure 6: Performance of ChatGPT, BingChat on VNHSGE dataset and GPT-3.5, GPT-4 on other datasets.",
"figure_data": "",
"figure_id": "fig_9",
"figure_label": "46",
"figure_type": "figure"
},
{
"figure_caption": "(a) Number of datasets in subjects: Texts and Question Answering Question Answering datasets in languages",
"figure_data": "",
"figure_id": "fig_10",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 7 :7Figure 7: Available datasets on Paperwithcode.",
"figure_data": "",
"figure_id": "fig_11",
"figure_label": "7",
"figure_type": "figure"
},
{
"figure_caption": "The number of real solutions of the equation |f (x 3 -3x)| = 2 3 is: A. 6 B. 10 C. 3 D. 9 Solution: From the graph of the function y = f (x), we deduce that the graph of the function y = |f (x)| is:",
"figure_data": "",
"figure_id": "fig_12",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": ": A box contains 15 balls including 6 red balls numbered from 1 to 6 and 9 blue balls numbered from 1 to 9. Two balls are randomly drawn from the box, what is the probability that they have different colors and their numbers add up to an even number? The number of ways to randomly select 2 balls from the box is: C 2 15 = 105 ways. To have the sum of numbers on the two balls even, we have two cases: • Case 1: Two balls of different colors with odd numbers: C 1 3 • C 1 5 = 15 ways. • Case 2: Two balls of different colors with even numbers: C 1 3 • C 1 4 = 12 ways.",
"figure_data": "",
"figure_id": "fig_13",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "B. 15 C. 3 D. 7 Solution: We have: y = -4x 3 + 12x + m. Consider the equation y = 0 ⇔ -4x 3 + 12x + m = 0 (1). For the function y to have three extreme points, equation (1) must have 3 distinct roots. We have: (1)",
"figure_data": "",
"figure_id": "fig_14",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": ") has 3 distinct roots when -8 < m < 8. Since m ∈ Z ⇒ m ∈ -7, -6, -5, . . . , 5, 6, 7. Thus, there are 15 integer values of the parameter m that satisfy the requirement in the problem. Answer choice B. 15 is correct. ChatGPT: To have the function y = -x 4 + 6x 2 + mx have three extreme points, the graph of the function must have two local maximum points and one local minimum point. The local maximum points occur at x = ± √ 2 and the local minimum point occurs at x = 0. The derivative of the function is: y = -4x 3 + 12x + m. To find the integer values of the parameter m such that the function has two local maximum points at x = ± √ 2, we need to satisfy the conditions: y",
"figure_data": "",
"figure_id": "fig_15",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "ICâu 1 :1. Reading Comprehension (3.0 points) Đọc đoạn trích: chính mẹ đẻ anh hùng và truyền thuyết tù túp lều lơp lá lơp tranh cắt cuống nhau bằng lưỡi liềm bàn chân thô quanh năm bùn lấm chưa một lần ướm qua sử sách tập con bước vịn vào ca dao tục ngữ dù uống nước đâu lòng vẫn nhớ nguồn thương từ cái kiến con ong tím ruột bầm gan thù bọn ác dân tộc tôi khi đứng dậy làm người là đứng theo dáng mẹ \"đòn gánh tre chín dạn hai vai\" mùa hạ gió Lào quăng quật mùa đông sắt se gió bấc dân tộc tôi khi đứng dậy làm người mồ hôi vã một trời sao trên đất trời sao lặn hóa thành muôn mạch nước chảy âm thầm chảy dọc thời gian Thực hiện các yêu cầu sau: Xác định thể thơ của đoạn trích Đáp án: Thể thơ tự doChatGPT: Đoạn trích không phải là một bài thơ đầy đủ, tuy nhiên, nó được viết theo thể thơ tự do (free verse), không có quy luật về độ dài câu thơ, số lượng câu thơ, hoặc âm vần.BingChat: Đoạn trích trên được viết theo thể thơ tự do. Thể thơ tự do không tuân theo những quy tắc về số lượng chữ trong một câu thơ hay số lượng câu thơ trong một khổ. Thay vào đó, nó tập trung vào cách sử dụng ngôn ngữ và hình ảnh để truyền tải cảm xúc và ý nghĩa của tác giả.",
"figure_data": "",
"figure_id": "fig_16",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Câu 2 :2Chỉ ra những từ diễn tả đời sống nghèo khó vất vả của người mẹ trong đoạn thơ: chính mẹ đẻ anh hùng và truyền thuyết từ túp lều lợp lá lợp tranh cắt cuống nhau bằng lưỡi liềm bàn chân thô quanh năm bùn lấm chưa một lần ướm qua sử sách Đáp án: Từ ngữ diễn tả đời sống nghèo khó của mẹ: Túp lều, lợp lá, lợp tranh, lưỡi liềm, bàn chân thô, bùn lấm. ChatGPT: Những từ ngữ diễn tả đời sống nghèo khó, vất vả của người mẹ trong đoạn thơ bao gồm: tù túp (tựa như đang chen chúc) lều lơp lá lơp tranh (lều nhỏ, vách dựng bằng lá) cắt cuống nhau bằng lưỡi liềm (công việc khó khăn, vất vả) bàn chân thô quanh năm bùn lấm (cuộc sống vất vả) chưa một lần ướm qua sử sách (không có cơ hội học hỏi).",
"figure_data": "",
"figure_id": "fig_17",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "C. 3 . 1 C. 3 . 23132Pronunciation and stress question Question: Mark the letter A, B, C, or D to indicate the word whose underlined part differs from the other three in pronunciation in each of the following questions. A. miss B. child C. sign D. mind Solution: Pronunciation: Pronounce single and double consonant sounds. Miss is pronounced as /i/ and the rest is pronounced as /ai/. Choose A. ChatGPT:B. child (pronounced with a \"ch\" sound) BingChat: Here are the correct answers to your questions: C. sign. Is there anything else you would like to know? Grammar and vocabulary questions Question: Mark the letter A, B, C, or D on your answer sheet to indicate the correct answer to each of the following questions.",
"figure_data": "",
"figure_id": "fig_19",
"figure_label": "3132",
"figure_type": "figure"
},
{
"figure_caption": "(Adapted from High Note) Question: Which of the following can be inferred from the passage? A. Most families prefer talking to one another to using smartphones in restaurants. B. Twenty-first century families are generally not in favor of watching television together. C. Parents in the past did not criticize their children for watching too much television. D. A lot of families are still incapable of making god uses of smartphones. Solution: Answering the question from the passage: Inference. Search keywords: a lot of families, incapable, making good use of smartphones =>Information in paragraph 4: many families, not yet learned how to deal with it. Choose D. ChatGPT: D. A lot of families are still incapable of making good uses of smartphones.",
"figure_data": "",
"figure_id": "fig_20",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "A. ωA B. 0.5A C. ω 2 A D. 0 Solution: The acceleration of the object is |a| = -ω 2 |x|. At the equilibrium position x = 0, so a = 0. Choose D.",
"figure_data": "",
"figure_id": "fig_21",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "A 4 T 1 = T 2 4 ⇒ 9 =41249. 90 cm B. 125 cm C. 80 cm D. 36 cm Solution: According to the hypothesis of the problem, 5 16(l + 45) = 25l ⇒ l = 16×25 80 cm. Choose C.",
"figure_data": "",
"figure_id": "fig_22",
"figure_label": "41249",
"figure_type": "figure"
},
{
"figure_caption": "√ 6 8 .8It is known that M at times t 1 and t 2 have acceleration vectors in opposite direction to its motion, and that M reaches its maximum speed v max once in the time interval from t 1 to t 2 . What is the value of v max closest to ? A. 27 cm s B. 20 cm s . C. 25 cm s D. 22 cm s Solution:",
"figure_data": "",
"figure_id": "fig_23",
"figure_label": "8",
"figure_type": "figure"
},
{
"figure_caption": "Question:Which of the following substances contains 1 amino group (N H 2 ) and 2 carboxyl groups (COOH) in its molecule? A. Formic acid B. Glutamic acid C. Alanine D. Lysine Solution: The substance that contains 1 amino group (N H 2 ) and 2 carboxyl groups (COOH) in its molecule is glutamic acid: HOOC -(CH 2 ) 2 -CH(N H2) -COOH. Choose B. ChatGPT: B. Glutamic acid has the formula HOOC -CH(N H 2 ) -CH 2 -CH 2 -COOH, which has 1 amino group and 2 carboxyl groups.",
"figure_data": "",
"figure_id": "fig_24",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Question:The completely thermal decomposition of m grams of N aHCO 3 yields N a 2 CO 3 , H 2 O, and 3.36 liters of CO 2 . The value of m is A. 30.0 g B. 25.2 g C. 15.0 g D. 12.6 g Solution: The number of moles of CO 2 is n CO2 = 3.36/22.4 = 0.15 mol. The balanced chemical equation is: 2N aHCO 3 t o → N a 2 CO 3 + CO 2 + H 2 O. The number of moles of N aHCO 3 is 0.3 mol. The mass m = 0.3 × 84 = 25.2 g. Choose B. ChatGPT: We have the reaction equation: 2N aHCO 3 → N a 2 CO 3 + H 2 O + 2CO 2 According to this equation, 2 mol of N aHCO 3 correspond to 1 mol of N a 2 CO 3 , 1 mol of H 2 O, and 2 mol of CO 2 . So if we completely decompose m grams of N aHCO 3 , the number of moles of CO 2 produced is: n CO2 = V CO 2 /V M = 3.36/22.4 = 0.15 (unit : mol) The initial number of moles of N aHCO 3 : n N aHCO3 = n CO2 /2 = 0.075. The initial mass of N aHCO 3 : m N aHCO3 = n N aHCO3 × M N aHCO3 = 0.075 × 84 = 6.3 (unit : gram) Therefore, the value of m is 6.3 grams. Answer: A. 6.3.",
"figure_data": "",
"figure_id": "fig_25",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "(a) Urea fertilizer provides plants with potassium. (b) Electrolysis of CuSO 4 solution produces Cu metal at the anode. (c) Adding excess CO 2 gas to N aAlO 2 solution results in the formation of precipitate. (d) Iron objects are less prone to corrosion after being coated with paint. A. 3 B. 1 C. 2 D. 4 Solution: Choose C. (a) Urea fertilizer provides plants with nitrogen, not potassium. => False (b) Cu metal is produced at the cathode, not the anode, during the electrolysis of CuSO 4 solution. => False (c) Adding excess CO2 gas to N aAlO 2 : CO 2 + H 2 O + N aAlO 2 → N aHCO 3 + Al(OH) 3 ↓ => Correct (d) Iron objects are less prone to corrosion after being coated with paint, this is a common method for protecting surfaces. Therefore, statement (d) is correct. => Correct ChatGPT: To solve this question, we need to consider each statement and determine whether they are true or false. (a) Urea fertilizers provide plants with potassium: False. Urea does not contain potassium and does not provide this element to plants. (b) Electrolyzing a CuSO 4 solution produces Cu metal at the anode: False. Electrolyzing a CuSO 4 solution produces Cu metal at the cathode, not the anode. (c) Bubbling excess CO 2 gas into a N aAlO 2 solution produces a precipitate: True. When CO 2 gas is bubbled into a N aAlO 2 solution, the following reaction occurs: N aAlO 2 + CO 2 + H 2 O → N aHCO 3 + Al(OH) 3 A precipitate of Al(OH)3 appears in the solution.",
"figure_data": "",
"figure_id": "fig_26",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "3 .3From pH, We have the mole number of HN O 3 as n HN O3 = 0.4 × 10 -pH = 0.04. Let the number of moles of O 2 in equations (1) and (2) be x and y. The mixture Z contains 8x moles of N O 2 and x+y moles of O 2 . Based on equation (3), we have 8x = 4(x + y) and 8x = 0.04, hence x = y = 0.05. The mass of m is m = 180 × 4 × x + 101 × 2 × y = 4.61 g. Choose B.",
"figure_data": "",
"figure_id": "fig_27",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "A. C 33and C 4 plants. B. C 3 and C 4 plants. C. C 4 and CAM plants. D. C 4 plants. Solution: The plant group that undergoes the process of photorespiration in the daytime is C3 plants. Choose B. ChatGPT: Answer: B. C 3 plants. The photorespiration process occurs in plants to generate energy through cellular respiration in the presence of light. In this process, C 3 plants use photosynthesis to produce glucose through the Calvin-Benson cycle (also known as the C 3 cycle) but employ photorespiration to convert glucose into ATP and other energy molecules. C 4 and CAM plants utilize alternative energy conversion mechanisms to minimize the amount of necessary photorespiration, thus they do not undergo the same level of photorespiration as C 3 plants.",
"figure_data": "",
"figure_id": "fig_28",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Solution:The steps in the technique of creating recombinant DNA are as follows: Separating the donor DNA and the DNA carrying the gene of interest => Cutting the donor DNA and the gene of interest => Joining the gene of interest with the donor DNA to create recombinant DNA. Choose B. ChatGPT: The sequence of steps in the technique of creating recombinant DNA is as follows: (1) Cutting the donor DNA and the gene of interest: Enzymes are used to cut a specific segment of DNA from the donor DNA, including the gene of interest. (2) Separating the donor DNA and the DNA carrying the gene of interest: After the gene has been cut from the donor DNA, an electrophoresis method is used to separate the gene from the DNA. (3)",
"figure_data": "",
"figure_id": "fig_29",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "A. 4 B. 2 C. 3 D. 1 Solution:421Original DNA strand: 3' TAX/GAA/AXX / GXX / GTA/GXA / ATT 5' Nucleotide sequence on mRNA: 5'AUG / XUU / UGG / XGG / XAU / XGU / UAA 3'. I. False. There are only 2 triplets that remain unchanged in amino acid when a substitution mutation occurs at the third position: 3'GAA5' and 3'GXA5'. The triplet 3'GAA5' still encodes the amino acid Leu even if there is a mutation in the third nucleotide, and the triplet 3'GXA5' still encodes the amino acid Arg even if there is a mutation in the third nucleotide. II. True. If a substitution occurs at the third nucleotide of the triplet 3'AXX5', it always results in a change in the amino acid at that position. III. False. There is only one triplet that leads to an early termination codon when a substitution mutation occurs at the third position: 3'AXX5'. IV. True. When a substitution mutation occurs at the third nucleotide of the triplet 3' TAX 5', it leads to the absence of a start codon. Therefore, the correct answer is B. Choose B.ChatGPT:The correct statements are: I. False. Since each amino acid is encoded by one or more codons, a mutation at the third position may or may not change the type of amino acid. It is not limited to a maximum of three encoding triplets. II. True. A mutation at the third position of a codon always changes the type of amino acid encoded by that codon. III. False. The presence of stop codons is determined by specific codons (in this case, 5'UGA3' and 5'UAA3'), not by mutations at the third position. IV. True. If a mutation at the third position eliminates the start codon (in this case, 5'AUG3'), there will be no start codon on the mRNA. Therefore, two statements are correct. The correct answer is B. 2.",
"figure_data": "",
"figure_id": "fig_30",
"figure_label": "421",
"figure_type": "figure"
},
{
"figure_caption": "Question:What military force did the United States use to implement their Special War strategy(1961)(1962)(1963)(1964)(1965) in South Vietnam? A. Portuguese Army B. Japanese Army C. Spanish Army D. Saigon Army Question: Given the chart: GDP OF THE PHILIPPINES AND THAILAND IN 2015 AND 2019 (Data from Vietnam Statistical Yearbook 2020, Statistical Publishing House, 2021) According to the chart, which of the following statements is true about the change in GDP in 2019 compared to 2015 of the Philippines and Thailand? A. Thailand increased less than the Philippines. B. The Philippines increased twice as much as Thailand. C. Thailand increased and the Philippines decreased. D. The Philippines increased slower than Thailand. Solution: During the period from 2015-2019, the GDP of the Philippines increased from 306 billion USD to 377 billion USD, an increase of 61 billion USD, or 1.23 times; while the GDP of Thailand increased from 401 billion USD to 544 billion USD, an increase of 141 billion USD, or 1.36 times => The GDP of the Philippines increased slower than Thailand. Choose D.",
"figure_data": "",
"figure_id": "fig_31",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figures 9 -D. 2 LiteratureD. 4 PhysicsFigures 21 -D. 5 ChemistryFigures 25 -D. 6 BiologyFigures 29 -D. 8 GeographyFigures 37 -92421525629837Figures 9-12 show the mathematics score spectrum of Vietnamese students in 2022-2019.",
"figure_data": "",
"figure_id": "fig_32",
"figure_label": "92421525629837",
"figure_type": "figure"
},
{
"figure_caption": "Figure 44 :44Figure 44: Civic education score spectrum of Vietnamese students in 2019.",
"figure_data": "",
"figure_id": "fig_33",
"figure_label": "44",
"figure_type": "figure"
},
{
"figure_caption": "Subjects use multiple-choice questionsSubjects TopicsMathematics spatial geometry, number series (arithmetic progression, geometric progression), combinations and probability, derivatives and applications, exponential and logarithmic functions, primitives and integrals, complex numbers, polyhedrons, rotating blocks, and Oxyz spatial calculus",
"figure_data": "",
"figure_id": "tab_0",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Average score and Most reached score of Vietnamese students",
"figure_data": "MathLitEngPhyCheBio",
"figure_id": "tab_1",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "VNHSGE dataset structure",
"figure_data": "SubjectExam TypeNumber of questions per exam Number of exams Question TotalMathematicsMultiple choice50502500LiteratureEssay650300EnglishMultiple choice50502500PhysicsMultiple choice40502000ChemistryMultiple choice40502000BiologyMultiple choice40502000HistoryMultiple choice40502000GeographyMultiple choice40502000Civic Education Multiple choice40502000",
"figure_id": "tab_3",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "ChatGPT and BingChat performances on VNHSGE dataset",
"figure_data": "MathematicsLiteratureEnglishPhysicsChemistryBiologyHistoryGeography Civic Education201952567552.757692605540556067.5 42.5 82.5507560752020665668.9 51.25869662.5 67.5 42.5 57.56072.5 47.58552.5707087.5202160667560.2576866067.5 62.55052.5 67.555907582.5 62.592.52022626056.37080946567.5 47.5 47.5 57.5 72.56092.5 62.58582.5902023546264.8 49.75789457.5 72.5 47.5 52.5606577.5 92.5 67.58577.582.5AVG 58.8606856.8 79.2 92.461664852.8586956.5 88.5 61.5 79.5 70.585.5",
"figure_id": "tab_4",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "1), literature (D.2), and English (D.3). These subjects are integral parts of the exam and are required for all students.",
"figure_data": "87.87.87.8Mathematics Score6 75.25.65.646.46.65.66.6766.66.616.266.475.46.25201920202021202220237.57.5Literature Score5 6 75.285.4966.895.136.6176.036.4775.6376.5176.48520192020202120222023English Score4 6 8 107.69.24.363.28.69.64.583.47.68.65.84489.45.153.87.89.420192020202120222023ChatGPT BingChat AVS MVS",
"figure_id": "tab_5",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "collecting raw data in Word format file, (2) translating symbols, formulas, and equations into Latex format, (3) converting Word format to JSON format. Convert raw data to json files and images. • Step 1: Take \"Raw data\" in. Questions and answers are the basic data that we present. The answers are multiple-choice with in-depth explanations. Microsoft Word displays the raw data as a table. A row with six columns represents each question's counterpart. The subsequent processing of the results is made easier with the aid of this data structure. ID Image Question Question Choice Image Answer Explanation",
"figure_data": "Image ChartImage folderImage pathPdf, word, html filesRaw dataWord fileJSON fileFigure 8:",
"figure_id": "tab_8",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "2 -3 -4i| = 2|z|. Let M and m respectively be the largest and smallest values of |z|. The value of M 2 + m 2 is",
"figure_data": "A. 28 B. 18 + 4√6 C. 14 D. 11 + 4√6",
"figure_id": "tab_9",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Nội dung câu thơ như một lời nhắc nhở thế hệ sau phải biết giữ gìn truyền thống của dân tộc. Biết ghi nhớ và biết ơn giá trị của dân tộc mình. Hai dòng thơ này nói về việc tập con người học tập và lấy nét từ các ca dao, tục ngữ, đồng thời nhắc nhở người ta rằng dù ở đâu, họ vẫn không nên quên nguồn gốc của mình. Nó thể hiện tình cảm gắn bó với nơi sinh ra, nơi mình trưởng thành, nơi mà các giá trị truyền thống và văn hóa dân tộc được tạo hình và truyền lại từ thế hệ này sang thế hệ khác",
"figure_data": "Câu 3: Nêu nội dung của hai dòng thơ:tập con bước vịn vào ca dao tục ngữdù uống nước đâu lòng vẫn nhớ nguồnĐáp án:",
"figure_id": "tab_10",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Xác định đúng vấn đề nghị luận: sức mạnh của tinh thần vượt khó trong cuộc sống -Viết đúng hình thức đoạn văn, số lượng chữ, không mắc lỗi dùng từ, đặt câu Yêu cầu cụ thể: bài viết cần đảm bảo các ý sau:-Giới thiệu vấn đề cần nghị luận: sức mạnh của tinh thần vượt khó trong cuộc sống",
"figure_data": "",
"figure_id": "tab_11",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "It is observed thatChatGPT outperforms BingChat in an essay that needs the capacity to analyze and evaluate.Cảnh: với sắc xanh ngút ngàn của núi rừng điểm những bông hoa chuối đỏ tươi như những bó đuốc sáng rực, xua đi sự lạnh lẽo hiu hắt của núi rừng. Thắp lên ngọn lửa ấm áp, mang lại ánh sáng nơi hơi ấm cho nơi đây. Con người: trước thiên nhiên bao la của núi rừng trở nên kỳ bí, hùng tráng hơn với hoạt động phát nương làm rẫy. + Mùa xuân: Ngày xuân mơ nở trắng rừng/ Nhớ người ta nón chuốt từng sợi giang Cảnh: hoa mơ rừng nở trắng khiến bừng sáng cả khu rừng, làm dịu mát tâm hồn con người Con người: đan nón, chuốt từng sợi giang. Một vẻ đẹp tình nghĩa được biểu hiện qua bàn tay khéo léo, tài hoa, nhanh nhẹn, chăm chút cẩn thận đối với từng sản phẩm lao động + Mùa hè: Ve kêu rừng phách đổ vàng/ Nhớ cô em gái hái măng một mình Cảnh: rừng khách đổ vào màu vàng rực của thiên nhiên dường như chuyển đột ngột qua \"tự đổ\" . Với tiếng ve kêu khiến cảnh sinh động có hồn và tưng bừng hơn. Con người: hình ảnh cô gái hái măng một mình đã lột tả được vẻ sống ân nghĩa được thể hiện thông qua nỗi nhớ, sự lưu luyến, bị lụy trong giây phút chia tay giữa đồng bào miền núi với các cán bộ khi trở về Hà Nội. Lẽ sống ân nghĩa còn được thể hiện qua sự mường tượng của tác giả về những kỉ niệm, sự gắn bó giúp đỡ của những đồng bào trong suốt quá trình sống và chiến đấu của các chiến sĩ ở khu vực miền Bắc. Lẽ sống ân nghĩa chính là một nét đẹp truyền thống của nhân dân ta, nó tạo nên sự liên kết giữa con người với con người. Góp phần xây dựng nên sức mạnh đại đoàn kết toàn dân Kết luận: Khái quát lại vấn đề, giá trị nội dung nghệ thuật Đoạn thơ của Tố Hữu đưa người đọc vào một không gian của quê hương, nơi mà những hình ảnh thiên nhiên và con người được kết nối một cách tự nhiên và tình cảm. Tác giả thể hiện sự nhớ những hoa cùng người khi trở về quê hương, một tình cảm thiêng liêng, đậm đà lẫn trong sự vui mừng và tràn đầy kỷ niệm.Bên cạnh đó, tác giả cũng miêu tả rất sống động những hình ảnh thiên nhiên và con người của quê hương Việt Nam. Những rừng xanh hoa chuối đỏ tươi, đèo cao nắng ánh dao gài thắt lưng, ngày xuân mơ nở trắng rừng, ve kêu rừng phách đổ vàng, rừng thu trăng rọi hòa bình. Tất cả tạo nên một bức tranh quê hương đẹp đẽ, hoang sơ và vẫn giữ được sự trong sạch, tự nhiên.Điểm nổi bật trong đoạn thơ này là thông điệp về lẽ sống ân nghĩa. Tác giả nhắc nhở về sự quan tâm, chia sẻ và tình người trong cuộc sống. Ý tưởng về sự ân nghĩa xuất hiện khi tác giả nhớ lại những người thân yêu, những người đã bên cạnh anh trong những khoảnh khắc khó khăn và cùng nhau chia sẻ niềm vui, nỗi buồn. Bức tranh quê hương đẹp đẽ, với những hình ảnh thiên nhiên và con người, như một thông điệp về sự sống động và sự ấm áp, sự đoàn kết giữa mọi người.Tuy nhiên, điểm nhấn của đoạn thơ này là sự nhớ nhung và ân nghĩa mà Tố Hữu dành cho những người mà ông đã gặp gỡ và quen biết trong cuộc sống. Mỗi câu thơ đều bắt đầu bằng từ \"nhớ\", thể hiện sự nhớ nhung và tình cảm mà Tố Hữu dành cho những người mà ông đã gặp gỡ. Ông nhớ về người đan nón chuốt từng sợi giang, cô em gái hái măng một mình hay ai đó có tiếng hát ân tình thủy chung. Những hình ảnh này thể hiện sự ân nghĩa và lòng biết ơn của Tố Hữu đối với những người đã giúp đỡ và chia sẻ với ông trong cuộc sống.Nhìn chung, đoạn trích trên thể hiện lẽ sống ân nghĩa của con người. Đó là sự biết ơn và trân trọng những người đã giúp đỡ và chia sẻ với mình trong cuộc sống. Đó cũng là thông điệp mà Tố Hữu muốn gửi gắm qua bài thơ này. Từ ngữ diễn tả đời sống nghèo khó của mẹ: Túp lều, lợp lá.. ... Nội dung câu thơ như một lời nhắc nhở thế hệ sau phải biết giữ gìn truyền thống ... Nhận xét về hình ảnh dân tộc Việt Nam được thể hiện trong đoạn trích: ..",
"figure_data": "Part QuestionScore criteriaScore1Thể thơ: tự do0.75I2 30.75 1.040.5ChatGPT:Tổng điểm câu 1 a. Đảm bảo yêu cầu về hình thức đoạn văn2 0.251b. Xác định đúng vấn đề nghị luận c. Triển khai vấn đề nghị luận0.25 0.25d. Chính tả, từ ngữ, ngữ pháp0.25IIe. Sáng tạo Tổng điểm câu 20.25 5.0Câu 2: Trong Việt Bắc nhà thơ Tố Hữu viết: a. Đảm bảo cấu trúc của văn nghị luận0.25Ta về, mình có nhớ ta Ta về, ta nhớ những hoa cùng người. b. Xác định đúng vấn đề nghị luận 2 c. Triển khai vấn đề nghị luận d. Chính tả, từ ngữ, ngữ phápVe kêu rừng phách đổ vàng Nhớ cô em gái hái măng một mình.0.5 3.5 0.25Rừng xanh hoa chuối đỏ tươi e. Sáng tạoRừng thu trăng rọi hòa bình0.25Đèo cao nắng ánh dao gài thắt lưng.Nhớ ai tiếng hát ân tình thủy chung.Ngày xuân mơ nở trắng rừngNhớ người đan nón chuốt từng sợi giang.",
"figure_id": "tab_12",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "In mice, fur color is determined by a gene with 4 alleles on an autosomal chromosome, where the dominant alleles are completely dominant. Six mice (denoted as a, b, c, d, e, f) of this species are subjected to the following crosses: Cross 1 (P): ♂ black fur (a) × ♀ black fur (b) producing F 1 offspring with 75% black fur and 25% brown fur. We can denote Cross 1: Black × Black → F1 produces brown fur → black > brown. Cross 2: Yellow × White → F1: 100% yellow fur → yellow > white. Cross 3: Brown × Yellow → F1: 2 brown: 1 yellow: 1 white → brown > yellow and white. In this species of mice, fur color is determined by one gene with four alleles on the autosomes, and the dominant alleles are completely dominant. Cross 1 (P): ♂ black fur (a) × ♀ black fur (b) produces F1 with 75% black fur and 25% brown fur. This indicates that alleles a and b both determine black fur, as F1 offspring have 75% black fur, meaning the majority carry either allele a or b. Cross 2 (P): ♂ yellow fur (c) × ♀ white fur (d) produces F1 with 100% yellow fur. This shows that allele c is completely dominant over d, as all F1 offspring carry allele c. Cross 3 (P): ♂ brown fur (e) × ♀ yellow fur (f) produces F1 with 25% white fur, 25% yellow fur, and 50% brown fur.",
"figure_data": "Solution: Convention: A1: black > A2: brown > A3: yellow > A4: white.I -False. There are 3 genotypes that determine brown fur: A2A2, A2A3, A2A4.II -False. The genotypes of all 6 mice can be determined.+ Cross 1: Black × Black → F1 produces brown fur → P: A1A2 (a) × A1A2 (b)+ Cross 2: Yellow × White → F1: 100% yellow fur → P: A3A3 (c) × A4A4 (d)+ Cross 3: Brown × Yellow → F1 produces white fur → P: A2A4 (e) × A3A4 (f)III -True. A4A4 (d) × A2A4 (e) → 1 A2A4: 1 A4A4 (1 brown: 1 white)IV -True. A1A2 (a) × A3A4 (f) → 1 A1A3: 1 A1A4: 1 A2A3: 1 A2A4 (1 black: 1 yellow).Choose A.ChatGPT:Cross 2 (P): ♂ yellow fur (c) × ♀ white fur (d) producing F 1 offspring with 100% yellow fur. 50% brown fur. Cross 3 (P): ♂ brown fur (e) × ♀ yellow fur (f) producing F 1 offspring with 25% white fur, 25% yellow fur, and According to theory, how many of the following statements are true?I. There are two genotypes that determine the brown fur phenotype.II. Based solely on the ratios of the F1 phenotypes resulting from the crosses, the genotypes of 4 out of the 6 micecan be accurately determined.III. If mouse (d) mates with mouse (e), the offspring will have genotypic ratios that match the phenotypic ratios.IV. If mouse (a) mates with mouse (f), the offspring can have a 50% ratio of black fur to yellow fur.A. 2 B. 1 C. 3 D. 4",
"figure_id": "tab_13",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "According to the chart, the GDP of Thailand increased from around 406 billion USD in 2015 to around 543 billion USD in 2019, an increase of about 137 billion USD. Meanwhile, the GDP of the Philippines increased from around 292 billion USD in 2015 to around 362 billion USD in 2019, an increase of about 70 billion USD.Due to their training in substantial textual datasets, ChatGPT and BingChat are able to assess and extract data from tables. As a result, they can give precise responses to simple questions that can be answered precisely using data from tables. In other words, both ChatGPT and BingChat are capable of providing precise answers if a question can be answered by only looking up a certain value or piece of information in a table. According to the table below, which shows the exports and imports of goods and services of Malaysia in billions of US dollars, which year did Malaysia have the largest trade surplus?",
"figure_data": "BingChat: I'm sorry, but I can't accurately answer your question because you didn't provide a chart for me toanalyze. Could you provide more information about the chart so I can better assist you?C.8.2 Table questionYear2015 2016 2018 2019Exports 209.3 201.2 246.0 237.8Imports 186.6 181.1 221.9 210.7",
"figure_id": "tab_14",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Physics score spectrum of Vietnamese students in 2021. Physics score spectrum of Vietnamese students in 2020. Number of Student Figure 24: Physics score spectrum of Vietnamese students in 2019. Chemistry score spectrum of Vietnamese students in 2022. Chemistry score spectrum of Vietnamese students in 2021. Chemistry score spectrum of Vietnamese students in 2020. Chemistry score spectrum of Vietnamese students in 2019. Number of Student Figure 29: Biology score spectrum of Vietnamese students in 2022. Biology score spectrum of Vietnamese students in 2021. Biology score spectrum of Vietnamese students in 2020. Biology score spectrum of Vietnamese students in 2019. History score spectrum of Vietnamese students in 2022. History score spectrum of Vietnamese students in 2021. History score spectrum of Vietnamese students in 2020. History score spectrum of Vietnamese students in 2019. Geography score spectrum of Vietnamese students in 2022. Geography score spectrum of Vietnamese students in 2021. Geography score spectrum of Vietnamese students in 2020. Civic education score spectrum of Vietnamese students in 2022. Civic education score spectrum of Vietnamese students in 2021. Civic education score spectrum of Vietnamese students in 2020.",
"figure_data": "0 1 2 3 4 5 0 1 2 3 4 0 2 4 6 8 0 1 2 3 4 5 6 7 8 0 4 •10 4 1 3 •10 4 1 1 3 •10 4 •10 4 0.5 Number of Student Number of Student Number of Student Number of Student 1 1.5 2 2.5 3 3.5 4 •10 4 13 0 4 6 9 23 42 47 38 109 134 12 260 292 51 568 1,129 668 1,212 65 28 1,980 2,189 395 3,123 3,092 610 4,373 4,421 694 5,965 5,642 1,262 7,207 6,792 1,520 8,533 7,725 2,261 9,661 8,452 2,916 10,724 9,190 4,810 11,981 9,645 5,559 13,066 10,573 8,239 14,266 11,330 10,119 15,359 12,248 19 13 22 41 24 138 191 248 509 761 1,225 1,759 3,083 3,722 5,920 7,107 112 271 678 1,374 2,927 5,113 8,263 12,251 16,803 21,590 26,094 30,207 33,323 35,634 36,884 38,064 37,883 16,898 13,107 14,794 10,833 36,594 18,528 14,400 16,097 20,204 15,475 22,329 22,232 16,719 22,089 23,712 18,136 42,285 25,704 19,609 38,622 27,651 20,981 50,630 29,634 22,363 51,713 31,292 23,502 67,855 33,408 24,943 62,643 35,357 26,287 73,559 37,964 28,088 65,243 40,132 29,894 75,019 42,732 31,927 58,066 45,808 34,273 64,026 48,716 36,783 51,202 51,490 39,596 52,307 53,694 41,075 35,680 54,495 41,868 10,898 14,828 14,052 25,382 23,895 36,092 38,225 56,848 54,355 72,566 65,718 80,282 60,025 67,193 48,900 49,475 28,706 35,940 34,027 32,576 30,700 29,077 27,268 25,439 23,862 22,414 21,222 19,744 18,991 18,071 17,093 16,295 15,800 15,355 15,135 52,273 41,567 32,762 24,938 14,689 48,222 40,297 21,753 40,654 38,260 15,556 31,021 35,625 6,095 20,796 32,792 2,326 12,095 27,237 172 5,915 19,433 5 2,540 11,086 12,810 7,348 1,946 624 41 2 14,546 14,191 13,702 12,652 11,115 9,070 6,757 926 4,669 1,542 240 4,111 1,824 35 273 425 Number of Student 0 0.5 1 1.5 2 2.5 3 3.5 4 •10 4 13 1 5 30 102 392 1,099 2,263 4,431 8,283 12,427 18,461 24,565 29,001 33,167 35,670 37,285 37,335 36,730 35,597 34,627 32,682 31,295 29,713 27,816 26,106 23,932 22,050 19,904 18,173 16,453 14,850 13,674 12,482 11,427 10,650 10,106 9,461 9,403 8,779 8,543 8,062 7,478 6,653 5,735 4,532 3,325 2,253 1,367 672 225 Number of Student 0 0.5 1 1.5 2 2.5 •10 4 12 0 0 1 11 24 71 138 241 388 645 1,029 1,526 2,031 2,818 3,581 4,613 5,661 7,005 8,252 9,549 11,160 12,696 14,556 16,355 17,962 19,475 21,240 22,546 23,162 22,883 21,849 19,664 16,373 13,432 10,232 7,192 4,278 2,012 708 154 Number of Student 0 0.5 1 1.5 2 2.5 •10 4 8 0 1 10 20 69 170 292 485 801 1,146 1,560 2,099 2,555 3,100 3,570 4,321 4,903 5,627 6,403 7,350 8,387 9,498 10,706 12,252 13,590 15,260 17,063 19,295 21,386 22,986 23,214 21,588 18,290 13,406 8,314 4,588 1,847 541 136 10 Number of Student 0 0.5 1 1.5 2 2.5 •10 4 12 2 1 8 20 51 125 202 460 763 1,187 1,639 2,272 3,106 3,981 5,042 6,069 7,235 8,422 9,303 10,275 11,052 12,206 12,942 13,577 14,532 15,582 16,631 17,935 19,035 20,429 22,028 22,800 21,818 18,092 12,868 8,061 4,440 2,180 829 158 Number of Student 0 0.5 1 1.5 2 •10 4 12 1 2 5 18 42 94 205 381 637 969 1,443 2,113 2,796 3,658 4,636 5,605 6,522 7,381 8,246 9,034 9,569 10,234 10,996 11,406 12,194 13,167 14,397 15,801 17,587 19,810 20,998 20,764 18,337 14,726 10,677 6,916 4,079 2,135 1,074 399 Number of Student 0 0.5 1 1.5 2 2.5 •10 4 38 0 5 6 45 80 244 560 1,191 2,193 3,995 6,188 9,198 12,549 16,010 19,296 21,628 22,942 24,004 23,470 22,360 20,200 18,423 16,094 13,956 11,701 9,799 8,466 7,063 5,942 5,159 4,606 4,062 3,566 2,906 2,096 1,243 622 235 54 5 0 0.5 1 1.5 2 •10 4 22 2 3 3 13 40 99 207 360 679 1,189 2,002 3,067 4,489 6,132 8,529 11,042 13,600 15,974 18,263 20,367 21,468 21,401 20,961 20,085 18,144 15,747 13,514 10,817 8,642 6,578 4,966 3,976 3,120 2,552 2,031 1,654 1,174 702 328 121 Number of Student 0 0.5 1 1.5 2 2.5 3 3.5 4 •10 4 5 4 2 14 58 108 259 579 1,154 1,929 3,235 4,690 6,456 8,329 10,685 12,913 15,611 17,953 20,594 22,979 25,896 28,554 30,961 33,304 35,162 36,931 38,602 38,565 38,718 37,785 35,357 32,674 28,576 24,257 19,968 15,598 11,742 8,343 5,572 3,766 1,779 Number of Student 0 0 0 0 0 0.5 1 1.5 2 2.5 3 3.5 4 •10 4 10 1 6 28 66 266 579 1,243 2,417 4,411 7,246 11,047 15,584 20,459 25,365 29,842 33,466 35,288 36,539 36,211 34,642 32,693 30,277 27,362 24,745 21,824 19,356 16,786 14,522 12,536 10,851 9,267 8,285 6,986 6,139 5,500 4,694 3,654 2,294 1,129 371 Number of Student 1 2 3 4 5 •10 4 21 3 3 7 4 10 36 75 135 210 405 610 867 1,413 2,194 3,350 5,062 7,546 10,908 15,127 20,364 26,111 32,256 38,481 44,464 48,253 50,888 52,273 52,449 49,163 45,213 39,556 33,014 26,025 19,697 13,878 8,784 5,223 2,394 788 163 Number of Student 1 2 3 4 5 6 •10 4 110 4 2 8 9 26 51 111 194 291 493 759 1,064 1,455 1,933 2,693 3,494 4,873 6,827 9,209 12,358 16,042 20,722 26,024 31,450 37,309 42,616 46,613 49,040 49,487 46,148 40,137 31,875 23,069 15,126 8,959 5,019 2,683 1,489 755 248 Number of Student 1 2 3 4 5 6 •10 4 27 0 0 0 3 4 8 23 31 36 61 90 153 227 270 395 613 830 1,216 1,736 2,417 3,237 4,531 6,331 8,240 11,174 14,814 19,062 24,029 30,504 37,263 44,158 50,855 55,320 •10 4 57,508 6 55,193 48,230 37,779 23,731 11,412 2,836 Number of Student 1 2 3 4 5 35 0 0 2 4 9 13 20 38 54 87 100 138 190 286 381 556 736 1,018 1,485 1,981 2,625 3,521 4,558 6,124 4,163 8,000 10,531 13,584 17,606 22,626 28,311 34,321 45,912 49,198 49,618 45,553 40,214 37,234 25,450 13,305 Number of Student0 0 0 0.2 0.2 0.2 0.4 0.4 0.4 0.6 0.6 0.6 0.8 0.8 0 1 1 0.25 1.2 1.2 0.5 1.4 1.4 0.75 1.6 1.6 1 1.8 1.8 1.25 2 2 1.5 2.2 2.2 1.75 2.4 2.4 2 2.6 2.6 2.25 2.8 2.8 2.5 3 3 2.75 3.2 3.2 3 3.4 3.4 3.25 3.6 3.6 3.5 3.8 3.8 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.8 4 4 3.75 3.75 4 4.2 4.2 4 4 4.2 4.4 4.4 4.25 4.6 4.6 4.5 4.8 4.8 4.75 5 5 5 5.2 5.2 5.25 5.4 5.4 5.5 5.6 5.6 5.75 5.8 5.8 6 6 6 6.25 6.2 6.2 6.5 6.4 6.4 6.75 6.6 6.6 7 6.8 6.8 7.25 7 7 7.5 7.2 7.2 7.75 7.4 7.4 8 7.6 7.6 8.25 7.8 7.8 4.25 4.5 4.75 5 5.25 5.5 5.75 6 6.25 6.5 6.75 7 7.25 7.5 7.75 8 8.25 4.4 4.6 4.8 5 5.2 5.4 5.6 5.8 6 6.2 6.4 6.6 6.8 7 7.2 7.4 7.6 7.8 8 8 8.5 8.5 8 8.2 8.2 8.75 8.4 8.4 9 8.6 8.6 9.25 8.8 8.8 9.5 9 9 9.75 9.2 9.2 10 9.4 9.4 8.75 9 9.25 9.5 9.75 10 8.2 8.4 8.6 8.8 9 9.2 9.4 9.6 9.6 9.6 9.8 9.8 9.8 10 10 10 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.8 4 4.2 4.4 4.6 4.8 5 5.2 5.4 5.6 5.8 6 6.2 6.4 6.6 6.8 7 7.2 7.4 7.6 7.8 8 8.2 8.4 8.6 8.8 9 9.2 9.4 9.6 9.8 10 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 5.25 5.5 5.75 6 6.25 6.5 6.75 7 7.25 7.5 7.75 8 8.25 8.5 8.75 9 9.25 9.5 9.75 10 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 5.25 5.5 5.75 6 6.25 6.5 6.75 7 7.25 7.5 7.75 8 8.25 8.5 8.75 9 9.25 9.5 9.75 10 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 5.25 5.5 5.75 6 6.25 6.5 6.75 7 7.25 7.5 7.75 8 8.25 8.5 8.75 9 9.25 9.5 9.75 10 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 5.25 5.5 5.75 6 6.25 6.5 6.75 7 7.25 7.5 7.75 8 8.25 8.5 8.75 9 9.25 9.5 9.75 10 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 5.25 5.5 5.75 6 6.25 6.5 6.75 7 7.25 7.5 7.75 8 8.25 8.5 8.75 9 9.25 9.5 9.75 10 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 5.25 5.5 5.75 6 6.25 6.5 6.75 7 7.25 7.5 7.75 8 8.25 8.5 8.75 9 9.25 9.5 9.75 10 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 5.25 5.5 5.75 6 6.25 6.5 6.75 7 7.25 7.5 7.75 8 8.25 8.5 8.75 9 9.25 9.5 9.75 10 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 5.25 5.5 5.75 6 6.25 6.5 6.75 7 7.25 7.5 7.75 8 8.25 8.5 8.75 9 9.25 9.5 9.75 10 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 5.25 5.5 5.75 6 6.25 6.5 6.75 7 7.25 7.5 7.75 8 8.25 8.5 8.75 9 9.25 9.5 9.75 10 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 5.25 5.5 5.75 6 6.25 6.5 6.75 7 7.25 7.5 7.75 8 8.25 8.5 8.75 9 9.25 9.5 9.75 10 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 5.25 5.5 5.75 6 6.25 6.5 6.75 7 7.25 7.5 7.75 8 8.25 8.5 8.75 9 9.25 9.5 9.75 10 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 5.25 5.5 5.75 6 6.25 6.5 6.75 7 7.25 7.5 7.75 8 8.25 8.5 8.75 9 9.25 9.5 9.75 10Figure 9: Mathematics score spectrum of Vietnamese students in 2022. ChatGPT BingChat Vietnamese students 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.8 4 4.2 4.4 4.6 4.8 5 5.2 5.4 5.6 5.8 6 6.2 6.4 6.6 6.8 7 7.2 7.4 7.6 7.8 8 8.2 199 1 85 0.8 22 0.6 11 0.4 •10 4 0 0.2 1 0 0 464 856 1,488 2,370 3,379 4,613 5,929 6,920 8,145 9,450 10,673 11,987 13,454 14,986 16,519 17,928 19,593 21,730 23,301 24,943 26,711 28,011 29,725 31,210 32,877 34,974 37,229 39,978 43,491 46,401 50,532 52,947 53,972 53,133 49,929 44,855 8.4 37,943 8.6 29,562 8.8 20,147 9 11,097 9.2 5,049 9.4 1,647 9.6 358 9.8 52 10 Figure 10: Mathematics score spectrum of Vietnamese students in 2021. Figure 11: Mathematics score spectrum of Vietnamese students in 2020. ChatGPT BingChat Vietnamese students 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.8 4 4.2 4.4 4.6 4.8 5 5.2 5.4 5.6 5.8 6 6.2 6.4 6.6 6.8 7 7.2 7.4 7.6 7.8 8 8.2 8.4 8.6 8.8 9 9.2 9.4 9.6 9.8 10 0 0 1 2 3 4 5 0.5 1 1.5 2 2.5 3 3.5 •10 4 0 1 6 36 74 228 552 1,229 2,350 4,121 6,566 9,070 11,571 14,142 15,962 17,353 18,670 19,437 20,458 21,825 23,406 24,686 26,671 28,073 29,786 31,386 32,313 33,937 34,687 34,982 35,295 35,435 35,794 35,741 35,661 35,203 33,853 32,099 29,167 25,762 22,002 17,640 13,581 10,243 7,086 4,765 2,794 1,565 684 182 12 ChatGPT BingChat Vietnamese students Figure 13: Literature score spectrum of Vietnamese students in 2022. ChatGPT BingChat Vietnamese students 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 5.25 5.5 5.75 6 6.25 6.5 6.75 7 7.25 7.5 7.75 8 8.25 8.5 8.75 9 9.25 9.5 9.75 10 0 1 2 3 4 5 6 7 8 •10 4 26 10 45 68 23 563 820 984 1,667 1,879 2,995 3,545 5,471 6,286 9,261 11,039 15,660 15,899 21,209 19,645 36,576 34,062 46,984 49,332 67,666 63,635 79,011 70,018 80,860 61,989 67,307 51,504 51,868 33,091 29,020 16,919 10,764 3,181 951 69 3 Number of Student Number of Student Number of Student ChatGPT BingChat Vietnamese students Figure 15: Literature score spectrum of Vietnamese students in 2020. ChatGPT BingChat Vietnamese students 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 5.25 5.5 5.75 6 6.25 6.5 6.75 7 7.25 7.5 7.75 8 8.25 8.5 8.75 9 9.25 9.5 9.75 10 0 1 2 3 4 5 6 7 8 •10 4 9 106 443 368 339 2,101 2,367 2,725 4,697 5,136 7,918 8,625 13,767 13,452 19,901 19,982 31,019 28,973 42,346 37,341 70,614 55,380 72,948 61,392 78,078 55,505 60,226 40,035 43,953 25,408 23,440 12,851 10,942 4,334 2,803 898 390 55 17 0 0 Number of Student Figure 16: Literature score spectrum of Vietnamese students in 2019. Figure 17: English score spectrum of Vietnamese students in 2022. ChatGPT BingChat Vietnamese students 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.8 4 4.2 4.4 4.6 4.8 5 5.2 5.4 5.6 5.8 6 6.2 6.4 6.6 6.8 7 7.2 7.4 7.6 7.8 8 8.2 8.4 8.6 8.8 9 9.2 9.4 9.6 9.8 10 0 0.5 1 1.5 2 2.5 3 •10 4 1 0 0 3 34 106 361 713 1,619 3,302 5,868 8,732 12,280 16,517 20,491 23,980 26,527 28,537 29,183 29,498 29,504 28,943 28,317 27,791 26,867 25,860 24,631 23,337 22,660 21,964 21,090 20,102 19,403 18,911 18,665 18,354 18,464 18,219 18,498 18,915 19,319 20,258 21,176 22,490 23,724 24,471 24,251 21,582 16,586 10,543 4,345 Number of Student Figure 18: English score spectrum of Vietnamese students in 2021. Figure 19: English score spectrum of Vietnamese students in 2020. ChatGPT BingChat Vietnamese students 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.8 4 4.2 4.4 4.6 4.8 5 5.2 5.4 5.6 5.8 6 6.2 6.4 6.6 6.8 7 7.2 7.4 7.6 7.8 8 8.2 8.4 8.6 8.8 9 9.2 9.4 9.6 9.8 10 0 1 2 3 4 5 •10 4 0 0 5 33 123 469 1,324 2,864 5,952 10,310 16,722 23,685 31,481 37,599 42,348 45,297 45,755 44,476 41,861 39,542 36,385 33,410 30,588 27,458 24,979 22,630 20,989 18,710 17,283 15,464 14,288 13,145 12,173 11,343 10,405 9,834 9,274 8,552 7,990 7,612 7,108 6,970 6,416 6,045 5,378 4,845 3,968 3,133 1,976 939 299 Number of Student Figure 20: English score spectrum of Vietnamese students in 2019. ChatGPT BingChat Vietnamese students Figure 21: Physics score spectrum of Vietnamese students in 2022. ChatGPT BingChat Vietnamese students 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 5.25 5.5 5.75 6 6.25 6.5 6.75 7 7.25 7.5 7.75 8 8.25 8.5 8.75 9 9.25 9.5 9.75 10 0 0.5 1 1.5 2 2.5 •10 4 5 2 1 3 14 37 76 162 277 494 819 1,223 1,654 2,319 3,199 4,275 5,483 6,870 8,403 9,923 11,734 13,216 14,780 16,697 18,068 19,699 21,277 22,691 24,018 25,218 25,506 24,783 22,154 17,931 11,663 6,858 3,176 1,239 360 83 14 Number of Student Figure 22: ChatGPT BingChat Vietnamese students Figure 23: ChatGPT BingChat Vietnamese students 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 5.25 5.5 5.75 6 6.25 6.5 6.75 7 7.25 7.5 7.75 8 8.25 8.5 8.75 9 9.25 9.5 9.75 10 0 0.5 1 1.5 2 •10 4 0 1 7 25 117 312 640 1,329 2,427 3,557 4,868 6,273 7,526 8,596 9,500 10,708 11,611 12,538 13,917 15,307 16,332 17,434 18,557 19,414 19,656 19,839 19,393 18,687 17,519 15,238 13,134 10,326 7,821 5,425 3,408 2,046 966 385 103 17 2 ChatGPT BingChat Vietnamese students Figure 25: ChatGPT BingChat Vietnamese students 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 5.25 5.5 5.75 6 6.25 6.5 6.75 7 7.25 7.5 7.75 8 8.25 8.5 8.75 9 9.25 9.5 9.75 10 0 0.5 1 1.5 2 2.5 •10 4 10 1 1 20 26 90 151 334 612 984 1,519 2,226 2,957 3,876 4,733 5,886 6,751 7,966 8,816 9,797 10,575 11,386 12,162 13,227 13,782 14,820 15,999 17,572 19,818 22,463 25,721 27,138 26,071 22,441 16,398 10,620 5,998 3,051 1,404 495 149 Number of Student Figure 26: ChatGPT BingChat Vietnamese students Figure 27: ChatGPT BingChat Vietnamese students 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 5.25 5.5 5.75 6 6.25 6.5 6.75 7 7.25 7.5 7.75 8 8.25 8.5 8.75 9 9.25 9.5 9.75 10 0 0.5 1 1.5 2 •10 4 0 1 5 35 146 389 873 1,646 2,712 4,252 5,641 7,199 8,600 10,144 11,717 12,870 13,899 15,142 15,774 16,700 17,609 17,949 18,488 19,213 19,765 19,563 19,234 18,110 15,985 13,547 10,519 7,809 5,359 3,341 2,134 1,204 641 311 150 51 12 Number of Student Figure 28: ChatGPT BingChat Vietnamese students ChatGPT BingChat Vietnamese students 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 5.25 5.5 5.75 6 6.25 6.5 6.75 7 7.25 7.5 7.75 8 8.25 8.5 8.75 9 9.25 9.5 9.75 10 0 0.5 1 1.5 2 2.5 3 •10 4 25 1 1 11 37 109 176 364 620 1,293 1,924 3,000 4,446 6,443 8,857 12,247 15,363 18,299 21,491 23,553 25,053 25,198 24,625 22,761 20,935 18,435 16,058 13,578 11,462 9,487 7,876 6,530 5,185 4,143 3,375 2,712 2,139 1,720 1,410 1,080 582 Number of Student Figure 30: ChatGPT BingChat Vietnamese students Figure 31: ChatGPT BingChat Vietnamese students 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 5.25 5.5 5.75 6 6.25 6.5 6.75 7 7.25 7.5 7.75 8 8.25 8.5 8.75 9 9.25 9.5 9.75 10 0 0.5 1 1.5 2 2.5 3 •10 4 0 2 9 17 70 191 455 946 1,781 2,972 4,936 7,311 10,532 14,233 18,150 22,546 26,171 29,056 30,279 29,624 27,257 23,613 19,672 15,461 11,964 8,641 6,404 4,704 3,483 2,850 2,316 1,893 1,670 1,333 1,073 806 574 406 256 134 39 Number of Student Figure 32: ChatGPT BingChat Vietnamese students Figure 33: ChatGPT BingChat Vietnamese students 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 5.25 5.5 5.75 6 6.25 6.5 6.75 7 7.25 7.5 7.75 8 8.25 8.5 8.75 9 9.25 9.5 9.75 10 0 0.5 1 1.5 2 2.5 3 3.5 •10 4 4 7 24 125 380 957 2,351 4,417 8,118 12,773 17,921 23,197 28,359 31,083 33,381 34,161 34,468 34,399 33,226 32,078 30,888 29,554 27,782 26,488 24,461 22,986 20,934 19,150 17,219 15,253 13,828 12,253 10,366 9,053 7,573 6,242 4,895 3,425 2,025 935 266 Number of Student Figure 34: ChatGPT BingChat Vietnamese students Figure 35: ChatGPT BingChat Vietnamese students 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 5.25 5.5 5.75 6 6.25 6.5 6.75 7 7.25 7.5 7.75 8 8.25 8.5 8.75 9 9.25 9.5 9.75 10 0 1 2 3 4 •10 4 0 3 17 84 291 829 2,237 4,954 9,237 15,040 21,556 29,185 35,663 39,871 42,739 43,449 43,209 40,417 37,001 33,233 29,412 25,127 21,162 17,772 14,743 12,077 9,717 8,172 6,599 5,424 4,488 3,724 3,154 2,553 2,129 1,598 1,239 928 545 246 80 Number of Student Figure 36: ChatGPT BingChat Vietnamese students Figure 37: ChatGPT BingChat Vietnamese students 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 5.25 5.5 5.75 6 6.25 6.5 6.75 7 7.25 7.5 7.75 8 8.25 8.5 8.75 9 9.25 9.5 9.75 10 0 1 2 3 4 5 6 •10 4 94 1 3 9 11 26 42 72 142 250 390 585 799 1,131 1,615 2,135 2,993 4,207 6,134 8,591 11,809 16,297 21,550 27,970 34,965 41,539 47,203 52,701 55,115 54,987 52,209 46,398 39,808 32,442 24,503 17,796 11,887 7,156 3,867 1,478 227 Number of Student Figure 38: ChatGPT BingChat Vietnamese students Figure 39: ChatGPT BingChat Vietnamese students 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 5.25 5.5 5.75 6 6.25 6.5 6.75 7 7.25 7.5 7.75 8 8.25 8.5 8.75 9 9.25 9.5 9.75 10 0 1 2 3 4 5 •10 4 0 1 4 11 31 57 151 297 530 860 1,313 2,054 2,918 4,207 5,825 8,104 11,006 14,818 19,688 24,839 30,708 36,660 41,638 45,372 46,866 46,573 44,722 40,178 34,828 28,461 22,072 16,550 11,601 7,744 5,046 3,106 1,835 1,051 554 236 42 Number of Student Figure 40: Geography score spectrum of Vietnamese students in 2019. Figure 41: ChatGPT BingChat Vietnamese students 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 5.25 5.5 5.75 6 6.25 6.5 6.75 7 7.25 7.5 7.75 8 8.25 8.5 8.75 9 9.25 9.5 9.75 10 0 1 2 3 4 5 6 •10 4 24 0 1 2 2 4 12 10 19 45 56 92 113 170 305 376 579 819 1,111 1,581 2,083 2,727 3,772 4,978 6,489 8,537 11,107 13,730 17,285 21,016 25,446 29,990 35,000 39,779 44,186 49,057 51,560 52,868 50,343 40,169 18,680 Number of Student Figure 42: ChatGPT BingChat Vietnamese students 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 5.25 5.5 5.75 6 6.25 6.5 6.75 7 7.25 7.5 7.75 8 8.25 8.5 8.75 9 9.25 9.5 9.75 10 0 Figure 43: ChatGPT BingChat Vietnamese students 1 2 3 4 •10 4 0 1 3 3 4 14 19 61 101 160 254 368 566 794 1,188 1,689 2,148 2,968 4,110 5,117 784 3,258 6,650 8,842 8,206 11,487 14,049 17,662 21,337 25,554 29,375 33,714 36,665 39,220 40,139 39,467 37,641 34,113 28,791 22,667 14,895 Number of Student",
"figure_id": "tab_15",
"figure_label": "",
"figure_type": "table"
}
] | Xuan-Quy Dao; Ngoc-Bich Le; The-Duy Vo; Xuan-Dung Phan; Bac-Bien Ngo; Van-Tien Nguyen; Thi-My-Thanh Nguyen; Hong-Phuoc Nguyen | [
{
"authors": "Maud Chassignol; Aleksandr Khoroshavin; Alexandra Klimova; Anna Bilyatdinova",
"journal": "Procedia Computer Science",
"ref_id": "b0",
"title": "Artificial intelligence trends in education: a narrative overview",
"year": "2018"
},
{
"authors": "Olaf Zawacki-Richter; Victoria I Marín; Melissa Bond; Franziska Gouverneur",
"journal": "International Journal of Educational Technology in Higher Education",
"ref_id": "b1",
"title": "Systematic review of research on artificial intelligence applications in higher education-where are the educators",
"year": "2019"
},
{
"authors": "Gwo-Jen Hwang; Haoran Xie; Benjamin W Wah; Dragan Gašević",
"journal": "Computers and Education: Artificial Intelligence",
"ref_id": "b2",
"title": "Vision, challenges, roles and research issues of artificial intelligence in education",
"year": "2020"
},
{
"authors": "Lijia Chen; Pingping Chen; Zhijian Lin",
"journal": "Ieee Access",
"ref_id": "b3",
"title": "Artificial intelligence in education: A review",
"year": "2020"
},
{
"authors": "Xuan Quy Dao; Ngoc Bich Le; Thi My; Thanh Nguyen",
"journal": "",
"ref_id": "b4",
"title": "AI-Powered MOOCs: Video Lecture Generation",
"year": "2021-03"
},
{
"authors": "Thi My; Thanh Nguyen; Thanh Hai Diep; Bac Bien Ngo; Ngoc Bich Le; Xuan Quy Dao",
"journal": "",
"ref_id": "b5",
"title": "Design of Online Learning Platform with Vietnamese Virtual Assistant",
"year": "2021-02"
},
{
"authors": "Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova",
"journal": "",
"ref_id": "b6",
"title": "Bert: Pre-training of deep bidirectional transformers for language understanding",
"year": "2018"
},
{
"authors": "Alec Radford; Narasimhan Karthik; Salimans Tim; Sutskever Ilya",
"journal": "Citado",
"ref_id": "b7",
"title": "Improving language understanding with unsupervised learning",
"year": "2018"
},
{
"authors": "Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov",
"journal": "",
"ref_id": "b8",
"title": "Roberta: A robustly optimized bert pretraining approach",
"year": "2019"
},
{
"authors": "Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu",
"journal": "The Journal of Machine Learning Research",
"ref_id": "b9",
"title": "Exploring the limits of transfer learning with a unified text-to-text transformer",
"year": "2020"
},
{
"authors": "Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell",
"journal": "Advances in neural information processing systems",
"ref_id": "b10",
"title": "Language models are few-shot learners",
"year": "2020"
},
{
"authors": " Openai",
"journal": "",
"ref_id": "b11",
"title": "",
"year": "2023"
},
{
"authors": "Thorp Holden",
"journal": "",
"ref_id": "b12",
"title": "Chatgpt is fun, but not an author",
"year": "2023"
},
{
"authors": "Johan Eva Am Van Dis; Willem Bollen; Robert Zuidema; Claudi L Van Rooij; Bockting",
"journal": "Nature",
"ref_id": "b13",
"title": "Chatgpt: five priorities for research",
"year": "2023"
},
{
"authors": "Alex Wang; Amanpreet Singh; Julian Michael; Felix Hill; Omer Levy; Samuel R Bowman",
"journal": "",
"ref_id": "b14",
"title": "Glue: A multi-task benchmark and analysis platform for natural language understanding",
"year": "2018"
},
{
"authors": "Bryan Mccann; Nitish Shirish Keskar; Caiming Xiong; Richard Socher",
"journal": "",
"ref_id": "b15",
"title": "The natural language decathlon: Multitask learning as question answering",
"year": "2018"
},
{
"authors": "Liang Xu; Hai Hu; Xuanwei Zhang; Lu Li; Chenjie Cao; Yudong Li; Yechen Xu; Kai Sun; Dian Yu; Cong Yu",
"journal": "",
"ref_id": "b16",
"title": "Clue: A chinese language understanding evaluation benchmark",
"year": "2020"
},
{
"authors": "Dan Hendrycks; Collin Burns; Steven Basart; Andy Zou; Mantas Mazeika; Dawn Song; Jacob Steinhardt",
"journal": "",
"ref_id": "b17",
"title": "Measuring massive multitask language understanding",
"year": "2020"
},
{
"authors": "Pranav Rajpurkar; Jian Zhang; Konstantin Lopyrev; Percy Liang",
"journal": "",
"ref_id": "b18",
"title": "Squad: 100,000+ questions for machine comprehension of text",
"year": "2016"
},
{
"authors": "Denis Paperno; Germán Kruszewski; Angeliki Lazaridou; Ngoc Quan; Raffaella Pham; Sandro Bernardi; Marco Pezzelle; Gemma Baroni; Raquel Boleda; Fernández",
"journal": "",
"ref_id": "b19",
"title": "The lambada dataset: Word prediction requiring a broad discourse context",
"year": "2016"
},
{
"authors": "Dheeru Dua; Yizhong Wang; Pradeep Dasigi; Gabriel Stanovsky; Sameer Singh; Matt Gardner",
"journal": "",
"ref_id": "b20",
"title": "Drop: A reading comprehension benchmark requiring discrete reasoning over paragraphs",
"year": "2019"
},
{
"authors": "Rowan Zellers; Ari Holtzman; Yonatan Bisk; Ali Farhadi; Yejin Choi",
"journal": "",
"ref_id": "b21",
"title": "Hellaswag: Can a machine really finish your sentence?",
"year": "2019"
},
{
"authors": "Keisuke Sakaguchi; Le Ronan; Chandra Bras; Yejin Bhagavatula; Choi",
"journal": "Communications of the ACM",
"ref_id": "b22",
"title": "Winogrande: An adversarial winograd schema challenge at scale",
"year": "2021"
},
{
"authors": "Patrick Lewis; Barlas Oguz; Ruty Rinott; Sebastian Riedel; Holger Schwenk",
"journal": "",
"ref_id": "b23",
"title": "Mlqa: Evaluating cross-lingual extractive question answering",
"year": "2019"
},
{
"authors": "Mikel Artetxe; Sebastian Ruder; Dani Yogatama",
"journal": "",
"ref_id": "b24",
"title": "On the cross-lingual transferability of monolingual representations",
"year": "2019"
},
{
"authors": "Shayne Longpre; Yi Lu; Joachim Daiber",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b25",
"title": "Mkqa: A linguistically diverse benchmark for multilingual open domain question answering",
"year": "2021"
},
{
"authors": "Yaobo Liang; Nan Duan; Yeyun Gong; Ning Wu; Fenfei Guo; Weizhen Qi; Ming Gong; Linjun Shou; Daxin Jiang; Guihong Cao",
"journal": "",
"ref_id": "b26",
"title": "Xglue: A new benchmark datasetfor cross-lingual pre-training, understanding and generation",
"year": "2020"
},
{
"authors": "Siva Reddy; Danqi Chen; Christopher D Manning",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b27",
"title": "Coqa: A conversational question answering challenge",
"year": "2019"
},
{
"authors": "Leo Gao; Stella Biderman; Sid Black; Laurence Golding; Travis Hoppe; Charles Foster; Jason Phang; Horace He; Anish Thite; Noa Nabeshima",
"journal": "",
"ref_id": "b28",
"title": "The pile: An 800gb dataset of diverse text for language modeling",
"year": "2020"
},
{
"authors": "Pan Lu; Swaroop Mishra; Tanglin Xia; Liang Qiu; Kai-Wei Chang; Song-Chun Zhu; Oyvind Tafjord; Peter Clark; Ashwin Kalyan",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b29",
"title": "Learn to explain: Multimodal reasoning via thought chains for science question answering",
"year": "2022"
},
{
"authors": "Dan Hendrycks; Collin Burns; Saurav Kadavath; Akul Arora; Steven Basart; Eric Tang; Dawn Song; Jacob Steinhardt",
"journal": "",
"ref_id": "b30",
"title": "Measuring mathematical problem solving with the math dataset",
"year": "2021"
},
{
"authors": "Karl Cobbe; Vineet Kosaraju; Mohammad Bavarian; Mark Chen; Heewoo Jun; Lukasz Kaiser; Matthias Plappert; Jerry Tworek; Jacob Hilton; Reiichiro Nakano",
"journal": "",
"ref_id": "b31",
"title": "Training verifiers to solve math word problems",
"year": "2021"
},
{
"authors": "George Tsatsaronis; Georgios Balikas; Prodromos Malakasiotis; Ioannis Partalas; Matthias Zschunke; Dirk Michael R Alvers; Anastasia Weissenborn; Sergios Krithara; Dimitris Petridis; Polychronopoulos",
"journal": "BMC bioinformatics",
"ref_id": "b32",
"title": "An overview of the bioasq large-scale biomedical semantic indexing and question answering competition",
"year": "2015"
},
{
"authors": "Aniruddha Kembhavi; Minjoon Seo; Dustin Schwenk; Jonghyun Choi; Ali Farhadi; Hannaneh Hajishirzi",
"journal": "",
"ref_id": "b33",
"title": "Are you smarter than a sixth grader? textbook question answering for multimodal machine comprehension",
"year": "2017"
},
{
"authors": "Rowan Zellers; Yonatan Bisk; Roy Schwartz; Yejin Choi",
"journal": "",
"ref_id": "b34",
"title": "Swag: A large-scale adversarial dataset for grounded commonsense inference",
"year": "2018"
},
{
"authors": "Yonatan Bisk; Rowan Zellers; Jianfeng Gao; Yejin Choi",
"journal": "",
"ref_id": "b35",
"title": "Piqa: Reasoning about physical commonsense in natural language",
"year": "2020"
},
{
"authors": "Stéphane Aroca-Ouellette; Cory Paik; Alessandro Roncone; Katharina Kann",
"journal": "",
"ref_id": "b36",
"title": "Prost: Physical reasoning of objects through space and time",
"year": "2021"
},
{
"authors": "Haoxi Zhong; Chaojun Xiao; Cunchao Tu; Tianyang Zhang; Zhiyuan Liu; Maosong Sun",
"journal": "",
"ref_id": "b37",
"title": "Jec-qa: A legaldomain question answering dataset",
"year": "2020"
},
{
"authors": "Lucia Zheng; Neel Guha; Peter Brandon R Anderson; Daniel E Henderson; Ho",
"journal": "",
"ref_id": "b38",
"title": "When does pretraining help? assessing self-supervised learning for law and the casehold dataset of 53,000+ legal holdings",
"year": "2021"
},
{
"authors": "Jonathan Berant; Andrew Chou; Roy Frostig; Percy Liang",
"journal": "",
"ref_id": "b39",
"title": "Semantic parsing on freebase from questionanswer pairs",
"year": "2013"
},
{
"authors": "Yi Yang; Wen-Tau Yih; Christopher Meek",
"journal": "",
"ref_id": "b40",
"title": "Wikiqa: A challenge dataset for open-domain question answering",
"year": "2015"
},
{
"authors": "Mandar Joshi; Eunsol Choi; Daniel S Weld; Luke Zettlemoyer",
"journal": "",
"ref_id": "b41",
"title": "Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension",
"year": "2017"
},
{
"authors": "Tri Nguyen; Mir Rosenberg; Xia Song; Jianfeng Gao; Saurabh Tiwary; Rangan Majumder; Li Deng",
"journal": "choice",
"ref_id": "b42",
"title": "Ms marco: A human generated machine reading comprehension dataset",
"year": "2016"
},
{
"authors": "Tom Kwiatkowski; Jennimaria Palomaki; Olivia Redfield; Michael Collins; Ankur Parikh; Chris Alberti; Danielle Epstein; Illia Polosukhin; Jacob Devlin; Kenton Lee",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b43",
"title": "Natural questions: a benchmark for question answering research",
"year": "2019"
},
{
"authors": "Hideyuki Shibuki; Kotaro Sakamoto; Yoshinobu Kano; Teruko Mitamura; Madoka Ishioroshi; Di Kelly Y Itakura; Tatsunori Wang; Noriko Mori; Kando",
"journal": "Ntcir",
"ref_id": "b44",
"title": "Overview of the ntcir-11 qa-lab task",
"year": "2014"
},
{
"authors": "Anselmo Penas; Yusuke Miyao; Alvaro Rodrigo; Eduard H Hovy; Noriko Kando",
"journal": "",
"ref_id": "b45",
"title": "Overview of clef qa entrance exams task 2014",
"year": "2014"
},
{
"authors": "Alvaro Rodrigo; Anselmo Penas; Yusuke Miyao; Eduard H Hovy; Noriko Kando",
"journal": "CLEF (Working Notes)",
"ref_id": "b46",
"title": "Overview of clef qa entrance exams task 2015",
"year": "2015"
},
{
"authors": "Daniel Khashabi; Tushar Khot; Ashish Sabharwal; Peter Clark; Oren Etzioni; Dan Roth",
"journal": "",
"ref_id": "b47",
"title": "Question answering via integer programming over semi-structured knowledge",
"year": "2016"
},
{
"authors": "Guokun Lai; Qizhe Xie; Hanxiao Liu; Yiming Yang; Eduard Hovy",
"journal": "",
"ref_id": "b48",
"title": "RACE: Large-scale ReAding comprehension dataset from examinations",
"year": "2017"
},
{
"authors": "Minjoon Seo; Hannaneh Hajishirzi; Ali Farhadi; Oren Etzioni; Clint Malcolm",
"journal": "",
"ref_id": "b49",
"title": "Solving geometry problems: Combining text and diagram interpretation",
"year": "2015"
},
{
"authors": "Peter Clark; Isaac Cowhey; Oren Etzioni; Tushar Khot; Ashish Sabharwal; Carissa Schoenick; Oyvind Tafjord",
"journal": "",
"ref_id": "b50",
"title": "Think you have solved question answering? try arc, the ai2 reasoning challenge",
"year": "2018"
},
{
"authors": "Alex Wang; Yada Pruksachatkun; Nikita Nangia; Amanpreet Singh; Julian Michael; Felix Hill; Omer Levy; Samuel Bowman",
"journal": "Advances in neural information processing systems",
"ref_id": "b51",
"title": "Superglue: A stickier benchmark for general-purpose language understanding systems",
"year": "2019"
},
{
"authors": "David Saxton; Edward Grefenstette; Felix Hill; Pushmeet Kohli",
"journal": "",
"ref_id": "b52",
"title": "Analysing mathematical reasoning abilities of neural models",
"year": "2019"
},
{
"authors": "Uri Shaham; Elad Segal; Maor Ivgi; Avia Efrat; Ori Yoran; Adi Haviv; Ankit Gupta; Wenhan Xiong; Mor Geva; Jonathan Berant",
"journal": "",
"ref_id": "b53",
"title": "Scrolls: Standardized comparison over long language sequences",
"year": "2022"
},
{
"authors": "Ekaterina Taktasheva; Tatiana Shavrina; Alena Fenogenova; Denis Shevelev; Nadezhda Katricheva; Maria Tikhonova; Albina Akhmetgareeva; Oleg Zinkevich; Anastasiia Bashmakova; Svetlana Iordanskaia",
"journal": "",
"ref_id": "b54",
"title": "Tape: Assessing few-shot russian language understanding",
"year": "2022"
},
{
"authors": "Kai Sun; Dian Yu; Jianshu Chen; Dong Yu; Yejin Choi; Claire Cardie",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b55",
"title": "Dream: A challenge data set and models for dialogue-based reading comprehension",
"year": "2019"
},
{
"authors": "Johannes Welbl; Nelson F Liu; Matt Gardner",
"journal": "",
"ref_id": "b56",
"title": "Crowdsourcing multiple choice science questions",
"year": "2017"
},
{
"authors": "Xiao Li; Yawei Sun; Gong Cheng",
"journal": "",
"ref_id": "b57",
"title": "Tsqa: tabular scenario based question answering",
"year": "2021"
},
{
"authors": "Ali Borji",
"journal": "",
"ref_id": "b58",
"title": "A categorical archive of chatgpt failures",
"year": "2023"
},
{
"authors": "Nir Yogesh K Dwivedi; Laurie Kshetri; Emma Hughes; Louise Slade; Anand Jeyaraj; Arpan Kumar Kar; Abdullah M Baabdullah; Alex Koohang; Manju Vishnupriya Raghavan; Ahuja",
"journal": "International Journal of Information Management",
"ref_id": "b59",
"title": "so what if chatgpt wrote it?\" multidisciplinary perspectives on opportunities, challenges and implications of generative conversational ai for research, practice and policy",
"year": "2023"
},
{
"authors": "Jürgen Rudolph; Samson Tan; Shannon Tan",
"journal": "Journal of Applied Learning and Teaching",
"ref_id": "b60",
"title": "Chatgpt: Bullshit spewer or the end of traditional assessments in higher education",
"year": "2023"
},
{
"authors": "Wenxiang Jiao; Wenxuan Wang; Jen-Tse Huang; Xing Wang; Zhaopeng Tu",
"journal": "",
"ref_id": "b61",
"title": "Is chatgpt a good translator? a preliminary study",
"year": "2023"
},
{
"authors": "Yejin Bang; Samuel Cahyawijaya; Nayeon Lee; Wenliang Dai; Dan Su; Bryan Wilie; Holy Lovenia; Ziwei Ji; Tiezheng Yu; Willy Chung",
"journal": "",
"ref_id": "b62",
"title": "A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity",
"year": "2023"
},
{
"authors": "Momchil Hardalov; Ivan Koychev; Preslav Nakov",
"journal": "",
"ref_id": "b63",
"title": "Beyond english-only reading comprehension: Experiments in zero-shot multilingual transfer for bulgarian",
"year": "2019"
},
{
"authors": "Baoxin Xingyi Duan; Ziyue Wang; Wentao Wang; Yiming Ma; Dayong Cui; Shijin Wu; Ting Wang; Tianxiang Liu; Zhen Huo; Hu",
"journal": "Springer",
"ref_id": "b64",
"title": "Cjrc: A reliable human-annotated benchmark dataset for chinese judicial reading comprehension",
"year": "2019"
},
{
"authors": "Abhilasha Ravichander; Alan W Black; Shomir Wilson; Thomas Norton; Norman Sadeh",
"journal": "",
"ref_id": "b65",
"title": "Question answering for privacy policies: Combining computational and legal perspectives",
"year": "2019"
},
{
"authors": "Ahmad Wasi Uddin; Jianfeng Chi; Yuan Tian; Kai-Wei Chang",
"journal": "",
"ref_id": "b66",
"title": "Policyqa: A reading comprehension dataset for privacy policies",
"year": "2020"
},
{
"authors": "Xuan Ngo; Tran Bach; Ngoc Ha; Tu Thien; Minh Phuong",
"journal": "IEEE",
"ref_id": "b67",
"title": "Question analysis for vietnamese legal question answering",
"year": "2017"
},
{
"authors": "Phi Manh Kien; Ha-Thanh Nguyen; Xuan Ngo; Vu Bach; Minh Le Tran; Tu Nguyen; Phuong Minh",
"journal": "",
"ref_id": "b68",
"title": "Answering legal questions by learning neural attentive text representation",
"year": "2020"
}
] | [
{
"formula_coordinates": [
7,
77.45,
306.4,
98.93,
74.39
],
"formula_id": "formula_0",
"formula_text": "-2 2 -1 1 2 x y"
},
{
"formula_coordinates": [
7,
338.11,
166.7,
192.95,
63.61
],
"formula_id": "formula_1",
"formula_text": "x f (x) f (x) -∞ -1 1 +∞ + 0 - 0 + -∞ -∞ 2 2 -2 -2"
},
{
"formula_coordinates": [
13,
89.34,
242.13,
268.92,
49.64
],
"formula_id": "formula_2",
"formula_text": "ID IQ Q C IA E1"
},
{
"formula_coordinates": [
13,
135.06,
269.8,
33.72,
49.48
],
"formula_id": "formula_3",
"formula_text": "A. 8a^3 B. 2a^3. C. a^3 D. 6a^3."
},
{
"formula_coordinates": [
13,
89.34,
532.77,
335,
56.46
],
"formula_id": "formula_4",
"formula_text": "ID IQ Q C IA E CC CE1"
},
{
"formula_coordinates": [
13,
135.06,
574.07,
33.72,
49.48
],
"formula_id": "formula_5",
"formula_text": "A. 8a^3 B. 2a^3. C. a^3 D. 6a^3."
},
{
"formula_coordinates": [
26,
88.31,
154.96,
440.83,
312.03
],
"formula_id": "formula_6",
"formula_text": "-2 2 -1 1 2 x y Setting t = x 3 -3x, we have f x 3 -3x = 2 3 ⇔ |f (t)| = 2 3 ."
},
{
"formula_coordinates": [
26,
310.73,
215.37,
220.33,
102.52
],
"formula_id": "formula_7",
"formula_text": "(x) = x 3 -3x, we have t (x) = 3x 2 -3; t (x) = 0 ⇔ x = ±1. The sign variation table of t(x) is: x f (x) f (x) -∞ -1 1 +∞ + 0 - 0 + -∞ -∞ 2 2 -2 -2"
},
{
"formula_coordinates": [
26,
338.38,
347.98,
190.76,
22.13
],
"formula_id": "formula_8",
"formula_text": "• The equation x 3 -3x = t 1 has one solution (since t 1 < -2)."
},
{
"formula_coordinates": [
27,
82.96,
179.01,
439.77,
47.85
],
"formula_id": "formula_9",
"formula_text": "3 ? A. 3 B. 1 2 C. 1 4 D. 7 2 Solution: We have u 3 = u 1 • q 2 = 2 • ( 1 2 ) 2 = 2 • 1 4 = 1 2 ."
},
{
"formula_coordinates": [
27,
204.62,
251.42,
201.56,
22.31
],
"formula_id": "formula_10",
"formula_text": "u 2 = u 1 • q = 2 • 1 2 = 1, u 3 = u 2 • q = 1 • 1 2 = 1 2"
},
{
"formula_coordinates": [
27,
247.67,
308.74,
265.32,
11.22
],
"formula_id": "formula_11",
"formula_text": "u 3 = u 1 * q 3-1 = 2 * (1/2) 2 = 1/2. The correct answer is B. 1/2."
},
{
"formula_coordinates": [
27,
159.06,
521.04,
71.56,
13.47
],
"formula_id": "formula_12",
"formula_text": "P = 12+15 105 = 9 35 ."
},
{
"formula_coordinates": [
28,
380.78,
72.09,
68.76,
26.15
],
"formula_id": "formula_13",
"formula_text": "15 2 = 5 105 = 1 21"
},
{
"formula_coordinates": [
28,
82.96,
325.1,
446.08,
32.13
],
"formula_id": "formula_14",
"formula_text": "⇔ m = 4x 3 -12x. Consider the function g(x) = 4x 3 -12x with g (x) = 12x 2 -12. We have g (x) = 0 ⇔ 12x 2 -12 = 0 ⇔ x = ±1."
},
{
"formula_coordinates": [
28,
156.18,
378.74,
313.92,
98.03
],
"formula_id": "formula_15",
"formula_text": "y(x) -∞ -1 1 +∞ + 0 - 0 + -∞ -∞8 8"
},
{
"formula_coordinates": [
28,
82.96,
573.05,
444.89,
54.91
],
"formula_id": "formula_16",
"formula_text": "( √ 2) = y (- √ 2) = 0 Which is equivalent to: ⇔ -8 √ 2 + 12 + √ 2m = 0 8 √ 2 + 12 - √ 2m = 0 Solving this system of equations yields ⇔ m = 5 √ 2 m = 19 √2"
},
{
"formula_coordinates": [
29,
87.48,
336.37,
441.57,
23.49
],
"formula_id": "formula_17",
"formula_text": "2|z| = z 2 -3 -4i ≥ ||z 2 | -|3 + 4i| = ||z| 2 -5 (since z 2 = |z| 2 ."
},
{
"formula_coordinates": [
29,
82.96,
349.33,
446.08,
36.59
],
"formula_id": "formula_18",
"formula_text": "4|z| 2 ≥ (|z|-5) 2 ⇔ |z| 4 -14|z| 2 +25 ≤ 0 ⇔ 7 -2 √ 6 ≤ |z| 2 ≤ 7 + 2 √ 6. Therefore, we have √ 6 -1 ≤ |z| ≤ √ 6 + 1. Thus, we get M = 1 + √ 6 and m = √ 6 -1. Therefore, M 2 + m 2 = 14."
},
{
"formula_coordinates": [
29,
82.96,
392.01,
445.78,
40.35
],
"formula_id": "formula_19",
"formula_text": "|z 2 -3 -4i| 2 = 4|z| 2 |z| 2 = x 2 + y 2 ≥ 0 Thus: |z 2 -3 -4i| 2 = |(x 2 -y 2 -3) + 2xyi -4i| 2 = (x 2 -y 2 -3) 2 + 4x 2 y 2 -8x 2 + 16y 2 = 4(x 2 + y 2 )|z| 2"
},
{
"formula_coordinates": [
29,
82.96,
459.72,
446.08,
24.32
],
"formula_id": "formula_20",
"formula_text": "0 ≤ (x 2 + y 2 ) = |z| 2 = 1 2 (|z| 2 + |z| 2 ) ≤ 1 2 (M 2 + m 2 ). And (x 2 + y 2 ) 2 = t 2 + 4u 2 ≤ 4|y| 2 (|t| + 4|y|) ≤ 8|y| 3 . Therefore: 0 ≤ |z| 2 ≤ 2 √ 2|y| 3/2 . In general, we have M 2 + m 2 ≤ 8 √ 2."
},
{
"formula_coordinates": [
38,
82.96,
254.71,
122.68,
13.47
],
"formula_id": "formula_21",
"formula_text": "A. π 2 B. π 4 C. π 6 D. π3"
},
{
"formula_coordinates": [
38,
84.15,
296.25,
446.82,
28.47
],
"formula_id": "formula_22",
"formula_text": "d 2 Q dt 2 + 1 LC Q = 0,"
},
{
"formula_coordinates": [
38,
302.52,
324.1,
161.12,
11.53
],
"formula_id": "formula_23",
"formula_text": "Q(t) = Q 0 cos(ωt + ϕ), where ω = 1 √"
},
{
"formula_coordinates": [
38,
458.48,
363.04,
70.56,
14
],
"formula_id": "formula_24",
"formula_text": "φ π = 1 2 ⇒ φ = π 2 ."
},
{
"formula_coordinates": [
39,
123.43,
104.01,
366.34,
25.51
],
"formula_id": "formula_25",
"formula_text": "1 2 ml 2 ω 2 max + 1 2 m(l + 45cm) 2 ω 2 max = 1 2 m l 2 2 2π T 2 + 1 2 m l + 45cm 2 2 2π T 2 ."
},
{
"formula_coordinates": [
39,
263.89,
224.34,
203.8,
13.47
],
"formula_id": "formula_26",
"formula_text": "3T 1 T 2 = 2 ⇔ T 12 T 22 = 4 9 ⇔ l l+45 = 4 9 ⇔ l = 90(cm)."
},
{
"formula_coordinates": [
40,
98.56,
72.87,
413.65,
73.3
],
"formula_id": "formula_27",
"formula_text": "have |v| = ω √ A 2 -u, therefore we obtain v2 v1 = 1-( u 2 A ) 2 1-( u 1 A ) 2 = 3 √ 6 8 . From the graph u 1 = +2mm u 2 = -3mm . 1 --3 A 2 1 -+2 A 2 = 3 √ 6 8 ⇒ 1 -3 A 2 1 -2 A 2 = 54 64 ⇒ 64 -64 9 A 2 = 54 -54 4 A 2 ⇒ A = 6 cm"
},
{
"formula_coordinates": [
40,
117.03,
171.12,
376.75,
25.09
],
"formula_id": "formula_28",
"formula_text": "t = π 2 + cos -1 2 6 + sin -1 3 6 ω = 0.8s ⇒ ω = π 2 + cos -1 2 6 + sin -1 3 6 0.8 = 4.16 rad s"
},
{
"formula_coordinates": [
40,
487.36,
268.24,
38.04,
19.28
],
"formula_id": "formula_29",
"formula_text": "v2 v1 = 3 √6"
},
{
"formula_coordinates": [
40,
82.96,
288.23,
447.25,
28.83
],
"formula_id": "formula_30",
"formula_text": "v 1 = f λ 2 = 25×0.8 2 = 10(m/s) v 2 = 3 √ 6 8 v 1 ≈ 14."
},
{
"formula_coordinates": [
40,
82.96,
327.25,
116.48,
11.23
],
"formula_id": "formula_31",
"formula_text": "a = |v 2 -v 1 | ≈ 4.43(m/s 2 )"
},
{
"formula_coordinates": [
40,
411.17,
320.33,
58.26,
18.14
],
"formula_id": "formula_32",
"formula_text": "v max = √ 2ad"
},
{
"formula_coordinates": [
41,
122.12,
362.16,
222.96,
9.65
],
"formula_id": "formula_33",
"formula_text": "2N aHCO 3 (s) ⇒ N a 2 CO 3 (s) + H 2 O(g) + CO 2 (g) ."
},
{
"formula_coordinates": [
42,
201.77,
547.16,
208.46,
34.17
],
"formula_id": "formula_34",
"formula_text": "2F e(N O 3 ) 2 t o -→ F e 2 O 3 + 4N O 2 ↑ +0.5O 2 ↑ (1) KN O 3 t o -→ KN O 2 + 0.5O 2 ↑ (2)"
},
{
"formula_coordinates": [
42,
224.79,
615.36,
162.42,
9.65
],
"formula_id": "formula_35",
"formula_text": "2N O 2 + 0.5O 2 + H 2 O-→2HN O 3 (3)"
},
{
"formula_coordinates": [
43,
222.54,
93.04,
166.93,
11.15
],
"formula_id": "formula_36",
"formula_text": "F e (NO 3 ) 2 + KN O 3 → KF e (NO 3 ) 4 ↑"
},
{
"formula_coordinates": [
43,
82.96,
139.72,
287.77,
57.96
],
"formula_id": "formula_37",
"formula_text": "N O + H 2 O → HN O 2 HN O 2 + H + → N O + 2 + H 2 O N O + 2 + H 2 O → HN O 3 + H + Since"
},
{
"formula_coordinates": [
44,
82.96,
197.34,
288.64,
8.97
],
"formula_id": "formula_38",
"formula_text": "A. 1 → 3 → 2. B. 2 → 1 → 3. C. 1 → 2 → 3. D. 3 → 1 → 2."
}
] |
|
10.1109/MSPEC.2004.1309810 | 2024-01-29 | [
{
"figure_ref": [],
"heading": "I. INTRODUCTION",
"publication_ref": [
"b0",
"b1",
"b2",
"b3",
"b4",
"b5",
"b6",
"b8",
"b5",
"b9",
"b10",
"b11"
],
"table_ref": [],
"text": "Deep Learning (DL) is a supervised machine learning approach that optimizes a loss function over a non-convex surface by comparing model predictions with ground truth. Each training iteration in DL involves forward and backward pass, i.e., generate predictions from input data, assess loss, compute gradients and update model parameters via optimization method like gradient descent. Training is an iterative process, typically involving multiple passes over the entire dataset where each pass is called an epoch. DL is also heavily influenced by certain hyperparameters that affect training speed, quality, or both. Commonly used hyperparameters are learning rate, momentum, batch size, weight decay, epochs, activation function, etc.\nDistributed data-parallel (DDP) methods further scale training across multiple nodes that train a globally shared model with I.I.D. data (independent and identically distributed) by periodically aggregating locally computed gradients at the end of each iteration. The compute requirements to train DL models doubles every 3.5 months [1], while the compute gains in chip design for ML accelerators and bandwidth gains in telecommunications networks double every 24 and 18 months [2], [3]. Thus, the infrastructure required to train state-of-the-art models tends to fall behind their compute and networking demands. Since upgrading network stack in the cloud, datacenter and HPC clusters can be infrequent as compared to appending new accelerators in pre-existing systems, gradient communication tends to be the major bottleneck in distributed training [4].\nDifferent compression techniques have been proposed in recent years to mitigate this synchronization overhead. However, the optimal compression factor (CF) that minimizes data exchange or end-to-end training time depends on the model itself (i.e., its size, structure and depth), available network bandwidth and the compression overhead itself. Unlike traditional HPC and distributed computing applications that only measure parallel efficiency, DDP training has an additional statistical efficiency associated with it. Although the amount of computation performed on each iteration is the same, some iterations tend to be more crucial than others towards the overall learning of the model. Updates are especially sensitive in early stages and to hyperparameters like learning rate schedule, momentum and weight decay [5]. It would thus be intuitive to compare information loss in gradients on account of compression, and use a lower CF when considerably more information is lost and a higher CF when most information is preserved under compression. We can subsequently increase compression as training continues and gradients saturate, and decrease it back during the aforementioned critical stages.\nWe take into account the parallel and statistical efficiency aspect of gradient compression in this work: a high CF improves overall throughput (i.e., number of samples processed per second) by reducing communication cost, but increases information loss in the gradients resulting in either slower or insignificant updates. The two metrics in DDP compression are pareto-related as one improves at the detriment of the other. We propose GraVAC: {Gra}dient {V}ariance-based {A}daptive {C}ompression 1 to dynamically adjust CF by comparing information loss from compression with that of the original gradients computed in backpropagation. GraVAC evaluates different CFs in a given search space and determines the CF that best balances parallel and statistical efficiency in DDP training with compression. We validate our approach over a variety of DL models and directly compare with static CF on compressors like Top-k [6], Deep Gradient Compression or DGC [7], Redsync [9] and Random-k [6]. DDP training can be implemented either via MPI-based collectives (AllReduce) [10]- [12] or using one or more centralized parameter servers (PS) [13] to accumulate and distribute model updates among workers."
},
{
"figure_ref": [
"fig_4"
],
"heading": "A. Scaling Efficiency of DDP Training",
"publication_ref": [
"b12",
"b13",
"b14",
"b15",
"b16"
],
"table_ref": [
"tab_0"
],
"text": "DL training is an iterative process that involves parameter updates at each step via gradient descent (GD) [14]. Full GD uses entire training data at every step, making the whole process slow and compute-intensive, while Stochastic GD processes a single sample and does not vectorize multiple samples on fast accelerators. Mini-batch GD is the optimal middle ground between Full and Stochastic GD where b samples are randomly sampled from I.I.D. data. Eqn. (1) describes the update rule in mini-batch GD where parameters w at (i + 1)-th iteration on N workers minimize loss function L(•) on input samples x j of size b from distribution X j with learning rate η. With weak scaling, we can increase the amount of per-iteration work by adding more workers and keeping per-worker batch-size b the same.\nw i+1 = w i -η 1 N n=N n=1 1 |b| j∈b ∂ ∂w i L(x (j,n) , w i )(1)\nThe ideal throughput of a distributed application T N executed across N workers is N times the throughput of a single worker T 1 . The deviation is measured via \"scaling efficiency\" in Eqn. 2a. Assuming negligible IO overhead, iteration time in dense SGD is bounded by computation and communication time (Eqn. (2b)). It may be possible to overlap communication with computation, but only partially since the latter is comparatively much lower on modern GPUs and TPUs. Model communication has been shown to be an order of hundreds or even thousands of magnitudes higher than gradient computation. Thus, frequent synchronization (t sync ) is the bottleneck that halts linear scaling in DDP. Table 1 describes the size, density and convergence target of ResNet101 [15], VGG16 [16] and LSTM [17] with dense SGD communication. Latency is further exacerbated on constrained networks with limited bandwidth as large volumes of data is exchanged by multiple workers simultaneously. For a DL model with a total of M parameters, the time cost based on the α-β communication model (where α is the latency and β is the inverse of bandwidth) for tree-based allreduce is (2αlogN + 2M logN β) [18]. For ring-based allreduce, this becomes 2(N -1)α+2M β(N -1)/N . Hence, communication cost increases as more workers are added to the mix in distributed training. Fig. 1a shows how overall throughput deviates from the ideal as cluster-size increases. The scaling efficiency is also influenced by the message size, i.e., total gradients/parameters to be communicated. In dense SGD, we observed scaling to be affected by the tensor-size distributions across the layers of a model as well. For e.g., LSTM has a better η scaling than ResNet101 despite being a larger model. This is because parameters in LSTM are spread across just 2 layers, compared to 101 in ResNet101.\nη scaling = T N /N • T 1 (2a) t iter ≈ t compute + t sync(2b)"
},
{
"figure_ref": [
"fig_4"
],
"heading": "B. Gradient Variance in Deep Learning",
"publication_ref": [
"b17",
"b18",
"b19",
"b20",
"b21"
],
"table_ref": [],
"text": "Prior work has demonstrated that gradient information can help measure the statistical efficiency of distributed training [19], [20]. There is a strong correlation between changes in the eigen values of second-order hessian [21] and first-order gradients (i.e., variance). [22], [23] explores how gradients behave in early stages of DL training and during certain critical periods, influenced by hyperparameters like learning rate schedule, gradient clipping and type of SGD used (e.g., zero, first or second-order moments). Fig. 1b attests those findings where we plot variance over the starting iterations and notice how drastically the gradients change and saturate over training."
},
{
"figure_ref": [],
"heading": "C. Gradient Compression",
"publication_ref": [
"b22",
"b23",
"b24",
"b38",
"b25",
"b26",
"b27"
],
"table_ref": [],
"text": "Many lossy compression techniques have been proposed for DDP and federated learning in recent years. Lossy compression incurs a fundamental trade-off between data-size and information loss; one can either reduce message size by losing more information, or preserve data quality by keeping majority of the original bits intact. In the context of DDP, higher CF reduces communication time at the cost of accuracy degradation or more steps/epochs required for the same convergence. CF measures the size of original gradients to the size of compressed tensors. E.g., compressing 10% gradients gives CF of 10x, while 1% gives 100x. Lossy compression can be broadly classified into quantization, sparsification or low-rank approximations.\nThe bit-width of single-precision (32-bit) floats is reduced in gradient quantization. Techniques like automatic mixed precision (AMP) [24] reduces gradients to half-precision, resulting in 2x CF. QSGD [25] balances the trade-off between accuracy and quantization precision. 1-bit SGD [26] reduces 32-bit floats to 1-bit and propagates quantization error via error-feedback. Sparsification methods communicate only a fraction of the gradient values along with their indices and set everything else to 0. Top-k sparisifies by extracting the top k% values while Random-k does so randomly with negligible compression overhead. DGC discards gradients below a certain threshold along with using momentum correction and gradient clipping. Methods like Redsync [40] combine quantization and sparsification, but the estimation quality is not accurate [27]. Approaches like PowerSGD [28] and Pufferfish [29] achieve compression via low-rank updates. The former can be viewed as adding regularization in DL, while the latter performs lowrank factorization on fully connected, convolutional and LSTM layers."
},
{
"figure_ref": [],
"heading": "What should be the ideal CF in Compression-based DDP?",
"publication_ref": [
"b28",
"b29",
"b30"
],
"table_ref": [
"tab_0"
],
"text": "The ideal CF is one that reduces communication time without trimming too much gradients which can be detrimental to final model. Compression has its own associated costs depending on the target CF and computational complexity of the mechanism itself. These factors affect both the parallel efficiency of distributed training as well as statistical inefficiency due to information loss from compression. Fig. 2 aptly demonstrates this where the CF that gives maximum speedup varies for each model and compression technique employed. The models are trained to Table 1 targets. ResNet101 on Top-k achieves most speedup at 100x, while VGG16 and LSTM peak at CFs 1000x and 10x respectively. On the other hand, ResNet101 fails to converge for any CF with Random-k compression. VGG16 and LSTM converged with 10x and failed with other CFs. Although a typical ML practitioner may not necessarily need to think about a plethora of compression methods, choosing the right CF with any compressor and DL model that minimizes training time, or even converges successfully, presents a non-trivial challenge.\nDynamic compression mechanisms like AdaQS [30] perform quantization using gradient mean to standard deviation ratio (MSDR). Systems like Accordion [31] and ScaDLES [32] switch between low and high compression based on critical regime identification. We tackle the ideal CF exploration problem in GraVAC in a gradient-driven manner by comparing variance of prior and post-compression gradients. For clarity, prior-compression gradients refer to the original tensors computed in backward pass. By measuring the information lost in compression, we dynamically adjust CF over each iteration. Starting with a low CF initially, we gradually increase compression as training progresses. On encountering senstive or critical regions, GraVAC switches to a lower CF that least degrades convergence."
},
{
"figure_ref": [],
"heading": "III. DESIGN AND IMPLEMENTATION",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we first describe the trade-off between parallel and statistical efficiency of DDP training with compression. Then we describe the metrics \"compression gain\" and \"compression throughput\" to combine the two, and explain GraVAC's adaptive compression algorithm."
},
{
"figure_ref": [
"fig_2",
"fig_2",
"fig_2"
],
"heading": "A. Parallel Efficiency of Gradient Compression",
"publication_ref": [
"b9",
"b3"
],
"table_ref": [],
"text": "The end goal of gradient compression is to improve DDP scaling efficiency. Application scaling is governed by the DDP mechanism (ring-based, tree-based allreduce or parameter servers), communication library used (MPI, NCCL [11], Gloo [10] or RPC) and available bandwidth. Keeping the latter and network infrastructure aside, speedup in any DL model depends on the target CF, quality of estimation and compression overhead. The overall iteration time in Eqn. 2b is adjusted for compression as\nt (c) iter ≈ t compute + t (c) sync + t (c) compress + t (c) decompress\nwhere it takes t is the time taken to reconstruct the compressed gradients to the same dimension as the original gradients. A viable compressor must have its compression time considerably lower than synchronization time.\nThe parallel efficiency of a distributed application suffers with more workers due to higher synchronization costs. Improving the network bandwidth alleviates this to only a certain extent. [4] investigates how DDP throughput improves marginally with higher bandwidth. They observed that ResNet50 peaks to 75% scale-out on a 25 Gbps network and remains the same even for 100 Gbps. Its because network transport implementation of current DL frameworks cannot fully utilize the available network bandwidth. Thus, even though cloud providers like GCP provide anywhere from 10-32 Gbps bandwidth depending on the machine type and VM size, they may not be utilized to their full potential.\nFig. 3 shows how the throughput increases and communication overhead reduces with compression. The results are relative to CF 10x for each model. We perform layerwise DGC compression over a 32 GPU cluster. System throughput is determined only by compression overhead and communication time as the compute time in backpropagation stays the same across all CFs. Based on the compressor used, compression latency may vary with target CF. For e.g., it decreases with larger CF as Top-k uses max-heap and sorts the top k% elements in O(N +k log k ) time. Throughput for ResNet101 and VGG16 saturates at 500x and does not improve thereafter, while LSTM saturates at 1000x (Fig. 3a). Communication savings also diminish at higher CFs due to small message size and network saturation (Fig. 3b). Thus, the highest CF may not necessarily correspond to the largest throughput."
},
{
"figure_ref": [],
"heading": "B. Statistical Inefficiency of Gradient Compression",
"publication_ref": [
"b33",
"b34",
"b5",
"b7",
"b33",
"b18",
"b35",
"b36"
],
"table_ref": [],
"text": "Gradient compression mechanisms rely on error-feedback [35], [36] which essentially acts as delayed updates, as commonly noted in asynchronous training. The gradients ineligible for compression in the current iteration are not discarded, but added to residual gradients which in turn are added to gradients computed in the next iteration. Residual gradients and error-feedback helps preserve important features and is critical to convergence [6]- [8]. Applying compression without error-feedback has been shown to achieve lower accuracy in deep learning models [35]. At the same time, residual gradients can sometimes degrade generalization performance due to stale updates.\nDDP training with very high CFs can negatively impact training time, convergence quality, or both if the compressed gradients are too sparse or quantized to update the model in any significant way. It is thus crucial to have an indicator that quantifies information loss between compressed and the original gradients. We do so by comparing variance between the original and compressed tensors on every iteration and see how it relates to actual model convergence. Denoting the original gradients as BC (Before-Compression) and compressed tensors as AC (After-Compression), we compare BC and AC tensors in two separate configurations with CFs 10x and 1000x in Fig. 4, 5 and 6. We compare the convergence curves for the two CFs with Dense SGD (i.e., no compression) to see how much accuracy degrades with compression.\nAC 10x is nearly identical to its BC counterpart in ResNet101 (Fig. 4a) while there is considerably more information loss in between BC and AC 1000x (Fig. 4b). This translates to their convergence curves in Fig. 4c as well where 10x and dense SGD runs follow a similar convergence trajectory while 1000x achieves considerably lower accuracy for the same iterations.\nVGG16 follows a similar trend with 10x CF. The BC and AC gradient variance (Fig. 5a) is nearly identical and so are the convergence curves for 10x and Dense SGD (Fig. 5c). We notice a slight deviation between BC and AC at 1000x initially in Fig. 5b, which correlates to slow convergence in the early iterations for 1000x in Fig. 5c. As the deviation BC and AC decreases, we see both CFs converge to the same accuracy as Dense SGD in the same iterations.\nThe AC 10x and 1000x gradients lie on similar scales as BC in LSTM, although the higher CF has slightly higher variance (Fig. 6a and5b). As seen from Fig. 5c, Dense SGD has the least perplexity (thus, better model quality), followed by 10x and 1000x CFs.\nTo compare the information loss between the original and gradients compressed to CF c, we define a simplistic metric called Compression gain. As part of error feedback, we update the gradients such that g\n(i) ef = g (i) 0 + residual gradients (i-1) for i ≥ 1. Here, g (i)\n0 are the original gradients calculated via backpropagation at iteration i, while residual gradients (i-1) are left-overs from the last iteration (i -1) and before, which are added back as part of error-feedback to produce g\n(i)\nef for the current iteration. With compression operator C, gradients are compressed as g\n(i) c = C[g (i) ef ].\nCompression gain is then measured as the ratio of expected variance of compressed gradients g\n(i) c\nand the original gradients modified with errorfeedback, i.e., g\n(i) ef : Compression gain = E[||g (i) c || 2 ] E[||g (i) ef || 2 ]\nIn prior work, gradient noise has been well studied in deep learning literature pertaining to divergence between locallycomputed and aggregated gradients in DDP [20], [37], [38]. These works use gradient information to tweak the global batch-size in DDP to optimize job completion time or allocate optimal resources for a job. Instead of looking at local and global gradients, GraVAC's novelty comes from evaluating the noise between the original and compressed tensors. The gradients computed over each iteration can be noisy. Thus, we keep a moving average of the respective variances of the original and compressed gradients. The computation and memory footprint of this approach is low since the windowsize in moving average is finite and only a single-precision floating point is stored for every iteration. Compression gain is bounded between {0, 1] such that it is low when C trims too much information. As models keep training, gradients saturate and higher compression becomes more viable in later stages of training. Hence, compression gain increases over training as compressed tensors become more aligned with the original gradients.\nWe plot compression gains for the three models when training with fixed CF 10x and 1000x respectively, shown in Fig. 4d, 5d and 6d. In each model, 10x has higher compression gain than 1000x since more information is preserved in the smaller CF. It should also be apparent that Dense SGD training has a constant gain of 1.0. For all models, convergence curve of 10x follows a similar trajectory as Dense SGD. Correspondigly, the compression gain of 10x stays close to 1.0 throughout. In ResNet101, gain of 1000x is low initially and grows in an oscillating manner, although still lower than gains of 10x and Dense SGD. The low gains in the first 1000 iterations of CF 1000x correlates to the considerable gap between BC and AC gradients in Fig. 4b and lower accuracy in Fig. 4c. VGG16 is more robust to higher CFs (Fig. 5c), as also seen from the high compression gains of CF 1000x in Fig. 5d. For LSTM, compression gain for 10x stays close to 1.0 and between 0.8-0.9 for 1000x. The proximity of the two CFs to Dense SGD's gain of 1.0 is equivalent to their perplexity curves in Fig. 6c. From these results we see how compression gain serves as a viable indicator of the statistical efficiency of DDP with compression."
},
{
"figure_ref": [
"fig_10",
"fig_9",
"fig_9",
"fig_10",
"fig_4",
"fig_8",
"fig_4"
],
"heading": "C. Combining System Throughput and Compression Gain",
"publication_ref": [
"b31",
"b32"
],
"table_ref": [],
"text": "As described earlier in II-C as well as Fig. 2, choosing a high CF unintuitively does not necessarily improve training time and may even degrade final model quality. Thus, to account for both the parallel and statistical efficiency DDP training with gradient compression, we combine system throughput (T system ) and compression gain into a single metric called Compression Throughput:\nT compression = T system × Compression gain\nIf CF is high, system throughput would be high as well but compression gain would relatively be lower, decreasing the resulting T compression . On the other hand, compression gain will be high for a low CF, but system throughput will be lower due to relatively higher communication overhead. With Compression Throughput, we capture this pareto-relationship between the parallel (system throughput) and statistical efficiency (compression gain) of gradient compression in DDP.\nWe build GraVAC as a modular extension on top of PyTorch's [33] DDP module [34] using Python in about 3000 lines of code. A base GravacOptimizer wraps common SGD optimizers implemented in PyTorch by extending the base torch.optim.Optimizer class. The optimizer takes an additional Compressor object that specifies the type of compression technique used. We implement four pre-existing techniques as compressor classes in this paper: Top-k, DGC, Redsync and Random-k. Compression for the appropriate CF and its gain is computed before the optimizer step function which applies the aggregated gradient updates on model parameters.\nGraVAC Algorithm: Alg. 1 describes GraVAC's approach of using compressor C to scale CFs in the exploration space [θ min , θ max ], where each candidate CF is evaluated for window steps and incremented in step-size of θ s w.r.t. θ min . For e.g., scaling from CF 10x to CF 20x means θ s = 20/10 = 2x. The threshold ϵ denotes the minimum compression gain required \no , t o = ∇f (x (i) , w i ) ▷ backpropagation 5 g (i) o = g (i) o + residual ▷ error-feedback 6 g (i)\nmin , t min = C(g\n(i) o , θ min ) ▷ compress to CF θ min 7 δ min = EWMA( ||g (i) min || 2 ||g (i) o || 2 ) ▷ θ min compression gain 8 g (i) c , t (i) c = C(g (i) min , θ s ) ▷ compress to CF (θ s • θ min ) 9 δ c = EWMA( ||g (i) c || 2 ||g (i) o || 2 ) ▷ gain for CF (θ s • θ min ) 10 t compress = t min + t c ▷ total compression time 11 if δ c ≥ ϵ : 12 g(i) , t s = Aggregate(g (i) c ) ▷ synchronize g (i) c 13 residual = g (i) o -g (i) c\n▷ update residual \nif | ct[-1] -ct[-2] ct[-2]\n| ≤ ω : for any CF to be eligible for communication in GraVAC, while threshold ω is used to measure saturation in compression throughputs and for scaling up θ min . We explain this in the following sections in more detail. For every iteration, we compute gradients g\n(i)\no with model parameters w i on training sample x (i) in time t o (line 4). To incorporate error-feedback,residual holds the leftover gradients not communicated from previous iterations. The shape and memory size of tensors in residual is the same as gradients itself. As shown in line 5, we add residual gradients to the gradients computed in the current iteration. In the first stage, we compress original gradients using C to compressed gradients g\n(i)\nmin corresponding to minimum CF θ min (line 6). We then compute the compression gain corresponding to θ min (line 7), and smoothen out the interiteration gain through exponential weighted moving average (EWMA) smoothing. In our evaluation, we set the EWMA smoothing factor to N /100, where N is the number of participating workers. We evaluate the next candidate CF by stepping up the previous θ min and further compressing the already compressed gradients g\n(i)\nmin by stepsize θ s (line 8). Thus, candidate CF evaluated in this case is θ s • θ min . This is done as part of our multi-level compression strategy to avoid compressing the large, original tensors g\n(i)\no twice. We measure the time savings of our multi-level approach in section IV-C.\nNext, we compute the gradients and compression gain of candidate CF θ s • θ min (line 8-9), and denote the total compression time t compress as the sum of time to compress original gradients to g c as well (line 13), calculate the total iteration time (line 14) and update the system as well as compression throughput for CF θ s •θ min via UpdateStep function. T compress is a dictionary or a hashmap that stores compression throughput of each candidate CF, min-max CF as well as dense SGD setting (i.e., CF 1x).\nIf the gain of g\n(i)\nc does not meet the threshold, but gain δ min of θ min does (line 16), we instead synchronize compressed gradients g\n(i)\nmin corresponding to θ min . In a similar fashion as before, we update the residuals, this time with g\n(i) min instead of g (i)\nc (line 18), compute iteration time and assess compression throughput. It is important to remember that synchronization overhead to communicate g\n(i)\nmin is more than g (i) c due to the former's lower CF. The trade-off we make in GraVAC is to incur higher communication latency for more accurate representation of the original gradients (measured by compression gain) and vice-versa.\nIf both θ min and currently evaluated CF do not meet the set threshold, we incur maximum communication latency by transmitting the original gradients via dense SGD (line 22). In this case, residual gradients are set to 0 and no compression overhead is included as part of iteration time and computing Iterations the compression gain of any candidate CF did not meet the threshold, we synchronized gradients compressed at 10x. For ϵ of 0.7, compression throughput was maximum for 1000x and we trained at this CF for most iterations as the corresponding gain easily met that threshold.\nVGG16: Like ResNet101, VGG16 also converges to the same accuracy as dense SGD within the same iterations, where ϵ = 0.7 and 0.9 reduce communication volume by 80× and 13.5× over dense SGD (Fig. 9). Although T compression is maximum at 1000x for ϵ = 0.9, the corresponding gain was not as high to meet the threshold. Because of this, we switch back to θ min and thus train with 10x for majority iterations as seen from the kernel density estimates in Fig. 8b. However, when ϵ was lower, we were able to find 40x CF to meet that threshold. T compression corresponding to this CF was second largest in our exploration space. As candidate CFs are evaluated over the iterations, the model gradually converges and as a result, compression gain improves even further on larger CFs as training progresses. Ultimately, we arrive on θ ideal = 1000x corresponding to the maximum compression throughput (Fig. 8c).\nLSTM: Like the models before, GraVAC with either ϵ converged in the same iterations as dense SGD training, while reducing the communication volume by 279× and 289× for ϵ of 0.9 and 0.7 respectively. Given the dataset, model and training hyperparameters, we already saw from Fig. 6d that compression gain for LSTM was high for both 10x and 1000x. We observed a similar trend here as compression gain corresponding to 1000x easily satisfied both thresholds and thus, we train with the largest available CF for most iterations (Fig. 9b). Correspondingly, the compression throughput is maximum at this CF as well.\nFurther, we compare GraVAC with static CFs running on different compression techniques. In particular, we train our models with Top-k, DGC, Redsync and Random-k at CFs 10x and 1000x. We run each compression technique to report the final accuracy/perplexity until it does not improve any further, difference in convergence compared to dense SGD For VGG16, we previously observed that the model is already quite robust to high compression (Fig. 5). We see that again here for Top-k, DGC and Redsync at 1000x cross 90% accuracy with 3.22, 3.35 and 3.6× speedup over Top-k 10x. Random-k at 10x also converged, albeit to a lower 87.8% accuracy and slower convergence. Since GraVAC attains 90.48% test accuracy with 1.95× training speedup, other compression schemes were more optimal in this case simply because they used high CFs.\nIn LSTM, GraVAC obtains the least perplexity of 21.25 while still providing maximum speedup of 6.67× over Top-k 10x. Random-k 10x converged to 24.15 perplexity and did not improve further, while Random-k 1000x failed here again. Of all the configurations, only Top-k, DGC and Redsync at 10x CF and GraVAC achieved better perplexity than dense SGD.\nThus, we see how GraVAC is able to train models like ResNet101 and LSTM to high accuracy/perplexity and still reduce training time significantly. Static compression schemes achieve high accuracy at low CF at the cost of high communication overhead, thus providing lower speedup. Large CFs considerably reduce communication, but the final model quality is not at par with GraVAC. On the flip side, some over-parameterized models like VGG16 can be robust to compression and still converge successfully at high static CFs.\n2) Geometric scaling policy: We also propose a relatively smoother compression policy where ScalingPolicy increments CFs as a geometric progression with common ratio 2. We deploy GraVAC with Redsync on ResNet101 and set θ min = 10x, θ max = 2000x, ϵ = 0.7, window = 2000 steps and ω = 1%. Thus, candidate CFs are 10x, 20x, 40x, 80x, 160x, 320x, 640x, 1280x and 2000x. Fig. 10a shows the accuracy curve over the iterations. Compared to dense SGD (Fig. 7a), GraVAC with geometric scaling converged while reducing communication volume by 76×. In contrast to exponential scaling, convergence is relatively slower because we evaluate each candidate CF for a larger window size. As a result, gradients get even smaller as GraVAC gradually arrives at larger CFs and compression gain increases beyond ϵ. Thus, we see similar iteration densities from CF 10x to 640x (Fig. 10b). After the first 7 CFs are evaluated over 2000 steps each, we mostly train with CF 1280x from 16K iterations onward (because 8 × 2000 = 16000). We did not scale to 2000x in our evaluation since compression throughput for 1280x and 2000x was 1029.9 and 1035.4, which falls within ω's bound of 1%. This case highlights the effectiveness of GraVAC such that it does not scale the CF beyond a point when it stop improving the parallel or statistical efficiency of gradient compression. In this case, GraVAC does not compress beyond 1280x as it corresponds to the maximum compression throughput (and at a lower CF of 1280x compared to 2000x)."
},
{
"figure_ref": [],
"heading": "C. Gains of Multi-level Compression in GraVAC",
"publication_ref": [],
"table_ref": [
"tab_5"
],
"text": "Alg. 1 explains how at each iteration, GraVAC scales compression from initial θ min to current CF being evaluated (i.e., θ c ), up to the maximum allowed θ max . Thus, compressing the original gradients (computed over backward pass) twice; i.e., once over θ min and then again on θ c can incur significant overhead, especially on larger models. The latency of a compressor may vary with the size of the tensor to compress as well as the target CF. To reduce the cumulative overhead of compressing original tensors multiple times, we apply a multi-level compression scheme as follows: given a compressor C and tensor X to be compressed to CFs θ 1 and θ 2 such that θ 2 > θ 1 , rather than compressing each CF on X as:\nX 1 = C(θ 1 , X ) and X 2 = C(θ 2 , X ) to produce compressed tensors where |X 2 | < |X 1 | < |X |.\nIn GraVAC, we first compute X 1 and then compress this tensor to θ\n′ 2 to produce X ′ 2 : X 1 = C(θ 1 , X ) =⇒ X ′ 2 = C(θ ′ 2 , X 1 ) : θ ′ 2 = θ 2 θ 1 The resulting tensor X ′ 2 is such that X ′ 2 = X 2 for θ ′ 2 = θ 2 /θ 1 .\nThe appeal of doing so is that the second compression operation is applied on a smaller tensor X 1 instead of X again. We tabulate the savings of multi-level compression in Table 3. Let's consider a scaling case of GraVAC where θ min = 10x and current CF evaluated is 1000x. Then multilevel GraVAC first compresses to 10x and then further compresses the reduced tensors to 100x, i.e., θ 1 = 10x and θ ′ 2 = 100x so that θ 2 = 1000x. In direct approach, we first compress original gradients to 10x, then compress the original gradients again to 1000x. From our results, we see that multi-level compression is at least 1.1× and up to 1.83× faster than directly compressing the original tensors twice."
},
{
"figure_ref": [
"fig_13",
"fig_13"
],
"heading": "D. Comparing GraVAC with Prior Art",
"publication_ref": [
"b29",
"b29",
"b17",
"b18",
"b35",
"b37",
"b29"
],
"table_ref": [
"tab_3"
],
"text": "In this section, we compare GraVAC with another adaptive scheme called Accordion [31]. For the three models, we use bounds of Rank-1 and Rank-4 for compression in Accordion, as described in [31] Accordion is based on detecting critical regions during training, i.e., when inter-iteration gradients computed in backward pass change significantly and cross a certain user-defined threshold. Accordion switches between 2 compression factors such that it uses the low CF in critical regions and the higher CF otherwise. On the other hand, GraVAC looks at information loss on account of compression (i.e., statistical efficiency) and not just relative gradient change in sensitive regions of training. That is, GraVAC looks at intra-iterations gradients as well (between original and gradients compressed at different CFs). Additionally, GraVAC scales compression across a wider range and carefully inspects intermediary CFs as potential compression candidates. Thus, we obtain higher speedups when training with GraVAC.\n1) GraVAC vs. Accordion on Random-k Compression: We previously saw in Fig. 2b and Table 2 that ResNet101 failed to converge at any CF with Random-k compression. In this section, we present a special case of using Random-k under the hood with both GraVAC and Accordion. Although the compression quality of Random-k is lower compared to other compressors, we present this as a special case to demonstrate how GraVAC is more dynamic and operates at a finer granularity. We launch GraVAC with Random-k on θ min = 1.5x, θ max = 1000x, window = 2000 and ϵ = 0.7. The CFs are scaled up via geometric scaling policy. Accordion was also deployed with the same min-max bounds on CF as GraVAC, i. Random-k compression (Fig. 2b) that failed to converge, we were able to achieve to 78% top-1 test accuracy for ResNet101 with GraVAC. The CFs used for training by GraVAC were 1.5x, 3x, 6x, 12x, 24x and 48x. All candidate CFs beyond this were ignored as they did not meet the required threshold of ϵ. CF 12x has the highest density, implying most iterations used this CF for training (Fig. 11b). Correspondingly, compression throughput is maximum for this CF as well. Compared to dense SGD, we reduced overall communication volume by 18×.\nAs for Accordion on Random-k, we see in Fig. 11a that training saturates at 20% accuracy. This is because Accordion does not consider the efficacy of the compression technique itself, and only switches between a low and high CF if the uncompressed, inter-iteration gradients change beyond a certain measure. With a low CF 1.5x, information loss in Random-k was too high to update ResNet101 in a meaningful way.\nV. CONCLUSION Gradient noise has previously been used as a scalability indicator for batch and cluster-size scaling in deep learning [19], [20], [37]- [39]. Adaptive compression schemes like Accordion [31] switch between two compression levels based on when the inter-iteration gradients change by some margin. GraVAC's key insight is to tweak compression factor over the course of training while balancing the pareto-relationship between parallel and statistical efficiency in gradient compression. We use \"compression gain\" to measure information loss on account of compression and choose a CF appropriately. In our evaluation, we see that GraVAC converges 1.95 to 6.67× faster than choosing a static CF, while converging in the same number of iterations as dense SGD. Compared to Accordion, we observed up to 5.63× reduction in end-to-end training time.\nOne should be mindful when training models with GraVAC as it introduces parameters like compression threshold (ϵ) and window size that may affect overall training performance. Setting too small a window size may result in poor convergence as all the candidate CFs may be exhausted while the model is still in early training stages and gradients are still volatile. As for ϵ, choosing a very small threshold may enable high compression but may lead to model degradation by allowing high CF gradients from the beginning that will not update the model in a significant way."
},
{
"figure_ref": [],
"heading": "",
"publication_ref": [],
"table_ref": [],
"text": "system/compression throughput. The CF and compression gain are both 1, as set in the UpdateStep function at line 25. Following SGD update (line 26), we evaluate GraVAC to assess the performance of CFs evaluated so far. This happens at a frequency determined by window. Here, we adjust θ s by a certain factor to scale up compression, determined by the chosen ScalingPolicy. The scaling policy tunes compression only until the upper bound θ max . We explore two scaling policies in this paper that we describe in detail under section IV-B. After scaling θ s , we also assess if the minimum CF, i.e., θ min can be scaled up as well. The intuition is that as training progresses, model gradually starts converging as well and we can use higher compression even for the minimum CF later on. In addition to candidate CFs, we thus scale up the minimum CF as well. The transition is made if the current gain δ c is within ω% of the gain of previous θ min (line 34). Once enough CFs are evaluated, we look at the two largest compression throughputs (line 36) and fetch the corresponding CF if they are within the bounds of ω. We do this as it means the compression throughput has saturated and thus, we pick the lower CF as θ ideal (line 38) and send the appropriate step-size (line 39). If the threshold ω is not met, we use θ s as is.\nWhen does compression scale-up? As seen from Alg. 1, the compression scale-up happens during GraVAC's evaluation phase where we scale the step-size θ s in accordance with a specific scaling policy. At the same time, we escalate the minimum CF θ min to currently evaluated CF if the two compression gains are within ω% of each other.\nWhen does compression scale-down? Compression scaledown is determined by ϵ (shown via conditional statements lines 11-25). If current CF loses considerably more information in compressed gradients g\nc , we use the lower CF θ min . If the latter fails to meet ϵ as well, we send uncompressed gradients g\no as a last resort."
},
{
"figure_ref": [],
"heading": "IV. EVALUATION",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "A. Cluster Setup and Training Hyperparameters",
"publication_ref": [],
"table_ref": [],
"text": "We evaluate GraVAC on a 32 GPU setup on the Google Cloud Platform (GCP) across 8 VMs. Each VM is a n1standard-8 machine type with 8 vCPUs, 30 GB system memory and 4 NVIDIA V100 GPUs with 16 GB VRAM each. The machines are configured with PyTorch 1.10.1, CUDA 11.3, CUDA driver 465.19.01 and NCCL 2.10.3.\nWe evaluate the three models described in Table 1. ResNet101 is trained with per-worker batch size 32, momentum 0.9, weight decay 0.0001 and SGD optimizer with initial learning rate (lr) 0.1 decayed by a factor of 10 at 9K and 14K iterations respectively. VGG16 is also trained with perworker batch-size 32, weight decay 0.0005, momentum 0.9 and SGD with fixed lr 0.1. Lastly, LSTM is measured with test perplexity (i.e., exponential of test loss) with per-worker batch-size 20, momentum 0.9, weight decay 0.0001 and SGD with fixed lr 0.1. The model is initialized with 1500 embedding dimensions and 2 hidden layers with 35 bptt steps.\nWe evaluate GraVAC with different scaling policies and look at their convergence curves (i.e. test accuracy/perplexity vs. iterations), average compression throughput of candidate CFs and kernel density estimates (KDE) of training iterations using different CFs over the course of training. KDE gives the distribution over the iterations for all CFs and plotted on the log-scale with smoothing bandwidth of 0.1 passed to the gaussian KDE."
},
{
"figure_ref": [],
"heading": "B. GraVAC's Adaptive Compression Policies",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we look at how GraVAC achieves optimal CF for a given θ min , θ max , ϵ, window, ω and stepsize. To see how a model converges and communication costs vary by evaluating different candidate CFs in the search space, we employ an Exponential policy that upscales CFs aggressively, and a relatively smoother Geometric scaling policy that scales CFs as a geometric progression.\n1) Exponential scaling policy: In this policy, we implement the ScalingPolicy function from Alg. 1 such that CFs are scaled up in exponents of 2 w.r.t the first initialized θ min . On top of DGC, we set θ min and θ max to 10x and 1000x, window=500 and ω=1%. So we scale up by factors of 2 1 , 2 2 , 2 4 , 2 8 w.r.t 10x up until 1000x. The candidate CFs thus evaluated in this policy are 10x, 20x, 40x, 160x and 1000x. We run GraVAC on two configuration with different thresholds on compression gain, ϵ = 0.7 and 0.9. The lower ϵ relaxes the constraint on the gain for higher CFs to be eligible for communication, thus achieving higher compression. A large ϵ (i.e., close to 1) allows for compression only if the compressed tensors are highly representative of the original gradients. First, we compare these two thresholds with Dense SGD as the latter demonstrates the ideal convergence scenario. Then, we compare GraVAC with different compression techniques on static CFs and look at final model accuracy, communication savings and overall speedup.\nResNet101: Fig. 7 shows how GraVAC achieves the same convergence as dense SGD in the same number of iterations. The low and high ϵ reduce overall communication volume by 163× and 19× over dense SGD. We measure communication volume as the ratio of cumulative single-precision floats exchanged among workers in GraVAC relative to dense SGD. Training cycle is slightly more volatile with compression, as seen from the accuracy drop due to lr decay at around 9000-th iteration. The drop is more apparent for ϵ = 0.7 as we continue to train with higher CFs on account of the lower threshold. Comparatively, ϵ = 0.9 is more robust to hyperparameter tuning like lr decay as we tend to train with a lower CF due to higher threshold. This is corroborated from Fig. 7b which shows distribution of training iterations over the CFs. We equally train with 10x and 1000x for ϵ = 0.9, while we mostly train with 1000x for ϵ of 0.7. For the compression throughputs of ϵ = 0.9 in Fig. 7c, it might seem counterintuitive at first that although T compression is maximum for 1000x and minimum for 10x, we still evenly train with the two CFs. This is on account of the high threshold and because θ min did not scale up and remained at 10x for ResNet101. Thus, whenever"
}
] | Distributed data-parallel (DDP) training improves overall application throughput as multiple devices train on a subset of data and aggregate updates to produce a globally shared model. The periodic synchronization at each iteration incurs considerable overhead, exacerbated by the increasing size and complexity of state-of-the-art neural networks. Although many gradient compression techniques propose to reduce communication cost, the ideal compression factor that leads to maximum speedup or minimum data exchange remains an open-ended problem since it varies with the quality of compression, model size and structure, hardware, network topology and bandwidth. We propose GraVAC, a framework to dynamically adjust compression factor throughout training by evaluating model progress and assessing gradient information loss associated with compression. GraVAC works in an online, black-box manner without any prior assumptions about a model or its hyperparameters, while achieving the same or better accuracy than dense SGD (i.e., no compression) in the same number of iterations/epochs. As opposed to using a static compression factor, GraVAC reduces end-to-end training time for ResNet101, VGG16 and LSTM by 4.32×, 1.95× and 6.67× respectively. Compared to other adaptive schemes, our framework provides 1.94× to 5.63× overall speedup. | GraVAC: Adaptive Compression for Communication-Efficient Distributed DL Training | [
{
"figure_caption": "Fig. 1 .1Fig. 1. Communication overhead and early critical period in DDP training.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "compress time to reduce gradients to CF c such that it reduces communication time to t",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 3 .3Fig. 3. Throughput and communication speedup for layerwise DGC compression, normalized by 10x CF.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 4 .Fig. 5 .Fig. 6 .456Fig. 4. ResNet101: Prior and Post-Compression gradients, test accuracy and compression gain for CFs 10x and 1000x.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "456",
"figure_type": "figure"
},
{
"figure_caption": "Algorithm 1 :1GraVAC's Adaptive Compression 1 Input: θ min , θ max , ϵ, θ s , ω, window, compressor C 2 w o : initial model state, N: total nodes, b: per-worker batch-size, residual = 0; T sys , T compress = empty() 3 Train for i = 1,2,3...",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "38 θ38ideal = T compress .get(ct[-2]) ▷ ideal CF 39 return θ ideal /θ min ▷ gives optimal θ s 40 else 41 return θ s ▷ else use old scaling factor",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "38",
"figure_type": "figure"
},
{
"figure_caption": "c(line 8). Based on the compression gains obtained and threshold ϵ, we choose the appropriate gradients to call the collective operation on. If the gain of our candidate CF meets ϵ (line 11), we go ahead and communicate compressed gradients g (i) c among workers. We update the residual gradients in accord with g (i)",
"figure_data": "",
"figure_id": "fig_7",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 7 .7Fig. 7. ResNet101:GraVAC with ϵ = [0.7, 0.9] and Dense SGD.",
"figure_data": "",
"figure_id": "fig_8",
"figure_label": "7",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 8 .8Fig. 8. VGG16: GraVAC with ϵ = [0.7, 0.9] and Dense SGD.",
"figure_data": "",
"figure_id": "fig_9",
"figure_label": "8",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 9 .9Fig. 9. LSTM: GraVAC with ϵ = [0.7, 0.9] and Dense SGD.",
"figure_data": "",
"figure_id": "fig_10",
"figure_label": "9",
"figure_type": "figure"
},
{
"figure_caption": "T compression and KDE 10. ResNet101: GraVAC with Geometric scaling policy.",
"figure_data": "",
"figure_id": "fig_12",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 11 .11Fig. 11. GraVAC and Accordion on Random-k compression.",
"figure_data": "",
"figure_id": "fig_13",
"figure_label": "11",
"figure_type": "figure"
},
{
"figure_caption": "",
"figure_data": "DL MODEL DESCRIPTIONModelLayersSize (MB)DatasetTest targetResNet101101170CIFAR1080% Top-1LSTM2252PTB22.0 PPLVGG1616528CIFAR10090% Top-5II. BACKGROUND AND RELATED WORK",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "GraVAC'S MODEL QUALITY AND SPEEDUP OVER STATIC CFS",
"figure_data": "ModelCompressionAcc./PplDiff.SpeedupTop-k 10x80.14%+0.14%1×Top-k 1000x76.4%-3.6%3.02×DGC 10x80.4%+0.4%1.23×DGC 1000x78.6%-1.4%5.19×ResNet101Redsync 10x79.4%-0.6%1.2×Redsync 1000x77.4%-2.6%6.94×Random-k 10x---Random-k 1000x---GraVAC80.2%+0.2%4.32×Top-k 10x91.2%+1.2%1×Top-k 1000x90.68%+0.68%3.22×DGC 10x90.8%+0.8%0.935×DGC 1000x90.4%+0.4%3.35×VGG16Redsync 10x90.45%+0.45%0.99×Redsync 1000x90.3%+0.3%3.6×Random-k 10x87.8%-2.2%0.7×Random-k 1000x---GraVAC90.48%+0.48%1.95×Top-k 10x22.0+0.01×Top-k 1000x26.78-4.783.36×DGC 10x21.67+0.331.23×DGC 1000x25.14-3.146.25×LSTMRedsync 10x21.65+0.351.17×Redsync 1000x24.24-2.246.9×Random-k 10x24.15-2.151.3×Random-k 1000x---GraVAC21.25+0.756.67×baseline from Table 1, and relative training speedup over Top-k 10x for each model. The results are tabulated in Table 2.We do not consider dense SGD training in this comparisonsince we already established previously how GraVAC is ableto achieve the same convergence in the same iterations, andother compression techniques have already been comparedto dense SGD in prior works. For ResNet101, 1000x CF onRedsync, DGC and Top-k have considerably high speedupsthan 10x Top-k. However, these methods at 1000x CF achieveconsiderably less accuracy than Top-k at 10x. At 1000x, Top-k,DGC and Redsync do not improve beyond 76.4%, 78.6% and77.4% top-1 test accuracy. Random-k faild to converge at eitherCF and accuracy did not improve beyond 20% . Because ofGraVAC's adaptive scheme, we converge to 80.2% accuracywhile still reducing training time by 4.32×.",
"figure_id": "tab_3",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "and compare with GraVAC in terms of communication and time savings (i.e., training speedup) to achieve the same test accuracy/perplexity. The savings are normalized by Accordion's performance for each respective model, shown in Table 4. For ResNet101, GraVAC reduces total communication volume by 44.5× and reduces training time by 1.94× over Accordion. GraVAC speeds up training by 5.63× over Accordion for communication-heavy models like VGG16. In LSTM training, GraVAC converges twice as fast by reducing communication volume up to 104.2×.",
"figure_data": "",
"figure_id": "tab_4",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "GraVAC'S MULIT-LEVEL (MTL) COMPRESSION SPEEDUP",
"figure_data": "ModelMethodDirect (ms)MTL (ms)SpeedupTop-k6063321.83×ResNet101DGC Redsync90 3359 29.81.52× 1.1×Random-k23141.64×Top-k1811211.49×VGG16DGC Redsync122 101.495.5 87.71.27× 1.16×Random-k41.6311.34×Top-k2001261.59×LSTMDGC Redsync88 69.463 46.41.4× 1.5×Random-k56.437.41.5×",
"figure_id": "tab_5",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "VS. ACCORDION: COMMUNICATION AND TIME SAVINGS",
"figure_data": "ModelMethodFloats sentComm. sav.Time sav.ResNet101Accordion GraVAC4.17 ×10 11 9.38 × 10 91× 44.5×1× 1.94×VGG16Accordion GraVAC3.83 ×10 11 1.7 × 10 101× 22.4×1× 5.63×LSTMAccordion GraVAC4.2 ×10 11 4 × 10 91× 104.2×1× 2.06×",
"figure_id": "tab_6",
"figure_label": "4",
"figure_type": "table"
}
] | Sahil Tyagi; Martin Swany | [
{
"authors": "",
"journal": "",
"ref_id": "b0",
"title": "AI and Compute",
"year": ""
},
{
"authors": "S Cherry; Spectrum",
"journal": "",
"ref_id": "b1",
"title": "Edholm's law of bandwidth",
"year": "2004"
},
{
"authors": "Robert R Schaller",
"journal": "IEEE Press",
"ref_id": "b2",
"title": "Moore's Law: Past, Present, and Future",
"year": "1997"
},
{
"authors": "Z Zhang; C Chang; H Lin; Y Wang; R Arora; Jin X ",
"journal": "NetAI",
"ref_id": "b3",
"title": "Is Network the Bottleneck of Distributed Training¿",
"year": "2020"
},
{
"authors": "Alessandro A Matteo; R Stefano; S ",
"journal": "",
"ref_id": "b4",
"title": "Critical Learning Periods in Deep Neural Networks",
"year": "2019"
},
{
"authors": "Dan A Torsten; H Mikael; J Sarit; K Nikola; K Cédric; R ",
"journal": "",
"ref_id": "b5",
"title": "The Convergence of Sparsified Gradient Methods",
"year": "2018"
},
{
"authors": "L Yujun; H Song; M Huizi; W Yu; Bill D ",
"journal": "ICLR",
"ref_id": "b6",
"title": "Deep Gradient Compression: Reducing the Communication Bandwidth for Distributed Training",
"year": "2018"
},
{
"authors": "S Shi; X Chu; K C Cheung; See ; S ",
"journal": "",
"ref_id": "b7",
"title": "Understanding Top-k Sparsification in Distributed Deep Learning",
"year": "2019"
},
{
"authors": "J Fang; H Fu; G Yang; C J Hsieh",
"journal": "Journal of Parallel and Distributed Computing",
"ref_id": "b8",
"title": "RedSync: Reducing synchronization bandwidth for distributed deep learning training system",
"year": "2019"
},
{
"authors": "Facebook Gloo",
"journal": "",
"ref_id": "b9",
"title": "",
"year": ""
},
{
"authors": "",
"journal": "",
"ref_id": "b10",
"title": "MPI: A message passing interface",
"year": ""
},
{
"authors": "L Mu; G A David; W P Jun; J S Alexander; A Amr; J Vanja; L James; Eugene J Bor-Yiing; S ",
"journal": "",
"ref_id": "b11",
"title": "Scaling Distributed Machine Learning with the Parameter Server",
"year": ""
},
{
"authors": "S Ruder",
"journal": "",
"ref_id": "b12",
"title": "An overview of gradient descent optimization algorithms",
"year": ""
},
{
"authors": "H Kaiming; Z Xiangyu; R Shaoqing; S Jian",
"journal": "",
"ref_id": "b13",
"title": "Deep Residual Learning for Image Recognition",
"year": "2015"
},
{
"authors": "Karen S Andrew; Z ",
"journal": "",
"ref_id": "b14",
"title": "Very Deep Convolutional Networks for Large-Scale Image Recognition",
"year": "2015"
},
{
"authors": "S Hochreiter; J Schmidhuber",
"journal": "Neural Computation",
"ref_id": "b15",
"title": "Long Short-Term Memory",
"year": "1997"
},
{
"authors": "S Agarwal; H Wang; S Venkataraman; D Papailiopoulos",
"journal": "MLSys",
"ref_id": "b16",
"title": "On the Utility of Gradient Compression in Distributed Training Systems",
"year": "2022"
},
{
"authors": "T B Johnson; P Agrawal; H Gu; C Guestrin",
"journal": "",
"ref_id": "b17",
"title": "AdaScale SGD: A User-Friendly Algorithm for Distributed Training",
"year": ""
},
{
"authors": "M Luo; L Guo; W Marcel; F Konstantinos; B Andrei-Octavian; P Peter",
"journal": "",
"ref_id": "b18",
"title": "Kungfu: Making Training in Distributed Machine Learning Adaptive",
"year": "2020"
},
{
"authors": "L Sagun; L Bottou; Y Lecun",
"journal": "",
"ref_id": "b19",
"title": "Eigenvalues of the Hessian in Deep Learning: Singularity and Beyond",
"year": "2016"
},
{
"authors": "Jonathan F David; J S Ari; S M ",
"journal": "",
"ref_id": "b20",
"title": "The Early Phase of Neural Network Training",
"year": "2020"
},
{
"authors": "Alessandro A Matteo; R Stefano; S ",
"journal": "",
"ref_id": "b21",
"title": "Critical Learning Periods in Deep Neural Networks",
"year": "2019"
},
{
"authors": "M Paulius; N Sharan; A Jonah; D Gregory; E Erich; G David; G Boris; H Michael; K Oleksii; V Ganesh; W Hao",
"journal": "",
"ref_id": "b22",
"title": "Mixed Precision Training",
"year": "2018"
},
{
"authors": "Dan A Demjan; G Jerry; L Ryota; T Milan; V ",
"journal": "",
"ref_id": "b23",
"title": "QSGD: Communication-Efficient SGD via Gradient Quantization and Encoding",
"year": "2017"
},
{
"authors": "F Seide; H Fu; J Droppo; G Li; D Yu",
"journal": "",
"ref_id": "b24",
"title": "1-Bit Stochastic Gradient Descent and Application to Data-Parallel Distributed Training of Speech DNNs",
"year": ""
},
{
"authors": "M A Ahmed; E Ahmed; A Mohamed-Slim; C Marco",
"journal": "MLSys",
"ref_id": "b25",
"title": "An Efficient Statistical-based Gradient Compression Technique for Distributed Training Systems",
"year": "2021"
},
{
"authors": "V Thijs; P K Sai; J Martin",
"journal": "NeurIPS",
"ref_id": "b26",
"title": "PowerSGD: Practical Low-Rank Gradient Compression for Distributed Optimization",
"year": "2019"
},
{
"authors": "W Hongyi; A Saurabh; P Dimitris",
"journal": "MLSys",
"ref_id": "b27",
"title": "Pufferfish: Communicationefficient Models At No Extra Cost",
"year": "2021"
},
{
"authors": "J Guo; W Liu; W Wang; J Han; R Li; Y Lu; S Hu",
"journal": "",
"ref_id": "b28",
"title": "Accelerating Distributed Deep Learning By Adaptive Gradient Quantization",
"year": ""
},
{
"authors": "A Saurabh; W Hongyi; L Kangwook; V Shivaram; P Dimitris",
"journal": "",
"ref_id": "b29",
"title": "Accordion: Adaptive Gradient Communication via Critical Learning Regime Identification",
"year": ""
},
{
"authors": "S Tyagi; M Swany",
"journal": "IEEE Big Data",
"ref_id": "b30",
"title": "ScaDLES: Scalable Deep Learning over Streaming data at the Edge",
"year": "2022"
},
{
"authors": "P Adam; G Sam; M Francisco; L Adam; B James; C Gregory; K Trevor; L Zeming; G Natalia; A Luca; D Alban; K Andreas; Y Edward; D Zach; R Martin; T Alykhan; C Sasank; S Benoit; F Lu; B Junjie; C Soumith",
"journal": "NeurIPS",
"ref_id": "b31",
"title": "PyTorch: An Imperative Style, High-Performance Deep Learning Library",
"year": "2019"
},
{
"authors": "S Li; Y Zhao; R Varma; O Salpekar; P Noordhuis; T Li; A Paszke; J Smith; B Vaughan; P Damania; S ",
"journal": "VLDB Endowment",
"ref_id": "b32",
"title": "PyTorch Distributed: Experiences on Accelerating Data Parallel Training",
"year": "2020"
},
{
"authors": "S P Karimireddy; Q Rebjock; S U Stich; M Jaggi",
"journal": "",
"ref_id": "b33",
"title": "Error Feedback Fixes SignSGD and other Gradient Compression Schemes",
"year": ""
},
{
"authors": "S Zheng; Z Huang; J T Kwok",
"journal": "NeurIPS",
"ref_id": "b34",
"title": "Communication-Efficient Distributed Blockwise Momentum SGD with Error-Feedback",
"year": "2019"
},
{
"authors": "M Sam; K Jared; A Dario; Openai Dota; Team ",
"journal": "",
"ref_id": "b35",
"title": "An Empirical Model of Large-Batch Training",
"year": "2018"
},
{
"authors": "Q Aurick; K C Sang; J S Suhas; N Willie; H Qirong; Z Hao; R G Gregory; P X Eric",
"journal": "",
"ref_id": "b36",
"title": "Pollux: Co-adaptive Cluster Scheduling for Goodput-Optimized Deep Learning",
"year": "2021"
},
{
"authors": "S Tyagi; P Sharma",
"journal": "",
"ref_id": "b37",
"title": "Scavenger: A Cloud Service for Optimizing Cost and Performance of ML Training",
"year": "2023"
},
{
"authors": "J Fang; H Fu; G Yang; C J Hsieh",
"journal": "",
"ref_id": "b38",
"title": "Accelerating Distributed Deep Learning Training with Gradient Compression",
"year": "2018"
}
] | [
{
"formula_coordinates": [
2,
76.63,
422.83,
224.06,
30.55
],
"formula_id": "formula_0",
"formula_text": "w i+1 = w i -η 1 N n=N n=1 1 |b| j∈b ∂ ∂w i L(x (j,n) , w i )(1)"
},
{
"formula_coordinates": [
2,
125.51,
677.93,
175.18,
30.86
],
"formula_id": "formula_1",
"formula_text": "η scaling = T N /N • T 1 (2a) t iter ≈ t compute + t sync(2b)"
},
{
"formula_coordinates": [
3,
336.88,
373.65,
200.2,
14.3
],
"formula_id": "formula_2",
"formula_text": "t (c) iter ≈ t compute + t (c) sync + t (c) compress + t (c) decompress"
},
{
"formula_coordinates": [
4,
311.98,
192.7,
251.5,
27.2
],
"formula_id": "formula_3",
"formula_text": "(i) ef = g (i) 0 + residual gradients (i-1) for i ≥ 1. Here, g (i)"
},
{
"formula_coordinates": [
4,
537.93,
244.66,
9.41,
6.12
],
"formula_id": "formula_4",
"formula_text": "(i)"
},
{
"formula_coordinates": [
4,
397.13,
269.18,
54.91,
14.3
],
"formula_id": "formula_5",
"formula_text": "(i) c = C[g (i) ef ]."
},
{
"formula_coordinates": [
4,
357.15,
293.69,
10.29,
12.46
],
"formula_id": "formula_6",
"formula_text": "(i) c"
},
{
"formula_coordinates": [
4,
372.04,
306.26,
129.74,
51.07
],
"formula_id": "formula_7",
"formula_text": "(i) ef : Compression gain = E[||g (i) c || 2 ] E[||g (i) ef || 2 ]"
},
{
"formula_coordinates": [
5,
78.97,
676.84,
191.05,
10.32
],
"formula_id": "formula_8",
"formula_text": "T compression = T system × Compression gain"
},
{
"formula_coordinates": [
6,
50.86,
124.92,
234.47,
38.85
],
"formula_id": "formula_9",
"formula_text": "o , t o = ∇f (x (i) , w i ) ▷ backpropagation 5 g (i) o = g (i) o + residual ▷ error-feedback 6 g (i)"
},
{
"formula_coordinates": [
6,
47.37,
151.34,
238.38,
130.83
],
"formula_id": "formula_10",
"formula_text": "(i) o , θ min ) ▷ compress to CF θ min 7 δ min = EWMA( ||g (i) min || 2 ||g (i) o || 2 ) ▷ θ min compression gain 8 g (i) c , t (i) c = C(g (i) min , θ s ) ▷ compress to CF (θ s • θ min ) 9 δ c = EWMA( ||g (i) c || 2 ||g (i) o || 2 ) ▷ gain for CF (θ s • θ min ) 10 t compress = t min + t c ▷ total compression time 11 if δ c ≥ ϵ : 12 g(i) , t s = Aggregate(g (i) c ) ▷ synchronize g (i) c 13 residual = g (i) o -g (i) c"
},
{
"formula_coordinates": [
6,
89.61,
630.57,
63.79,
14.38
],
"formula_id": "formula_11",
"formula_text": "if | ct[-1] -ct[-2] ct[-2]"
},
{
"formula_coordinates": [
6,
358.1,
97.4,
9.05,
6.12
],
"formula_id": "formula_12",
"formula_text": "(i)"
},
{
"formula_coordinates": [
6,
431.17,
181.69,
9.05,
6.12
],
"formula_id": "formula_13",
"formula_text": "(i)"
},
{
"formula_coordinates": [
6,
445.98,
277.94,
9.05,
6.12
],
"formula_id": "formula_14",
"formula_text": "(i)"
},
{
"formula_coordinates": [
6,
474.71,
314.41,
9.05,
6.12
],
"formula_id": "formula_15",
"formula_text": "(i)"
},
{
"formula_coordinates": [
6,
381.8,
522.88,
9.05,
6.12
],
"formula_id": "formula_16",
"formula_text": "(i)"
},
{
"formula_coordinates": [
6,
356.69,
547.39,
9.05,
6.12
],
"formula_id": "formula_17",
"formula_text": "(i)"
},
{
"formula_coordinates": [
6,
311.98,
560.58,
251.06,
25.61
],
"formula_id": "formula_18",
"formula_text": "(i) min instead of g (i)"
},
{
"formula_coordinates": [
6,
427.74,
598.28,
9.05,
6.12
],
"formula_id": "formula_19",
"formula_text": "(i)"
},
{
"formula_coordinates": [
9,
311.98,
281.79,
238.95,
31.91
],
"formula_id": "formula_20",
"formula_text": "X 1 = C(θ 1 , X ) and X 2 = C(θ 2 , X ) to produce compressed tensors where |X 2 | < |X 1 | < |X |."
},
{
"formula_coordinates": [
9,
311.67,
324.74,
253.11,
66.97
],
"formula_id": "formula_21",
"formula_text": "′ 2 to produce X ′ 2 : X 1 = C(θ 1 , X ) =⇒ X ′ 2 = C(θ ′ 2 , X 1 ) : θ ′ 2 = θ 2 θ 1 The resulting tensor X ′ 2 is such that X ′ 2 = X 2 for θ ′ 2 = θ 2 /θ 1 ."
}
] |
2023-05-20 | [
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b14",
"b25",
"b12",
"b17",
"b16",
"b13",
"b0",
"b26",
"b21",
"b0",
"b18",
"b11",
"b29",
"b3",
"b27",
"b23",
"b15",
"b6"
],
"table_ref": [],
"text": "Knowledge graphs (KG) have become the key technology to represent structural relations between entities and play an important role in question answering (Shen et al., 2018), dialogue systems (Yan et al., 2017), and entity disambiguation (Mulang' et al., 2020;Si et al., 2022Si et al., , 2023)). However, most KGs are growing at a rapid pace and are far from complete. Therefore, it is necessary to develop knowledge graph completion (KGC) approaches to add missing triples to the KGs, so as to improve the quality of KGs. Recent advances in KGC primarily work on knowledge graph embedding (KGE) by converting the entities and relations in KGs into lowdimensional vectors. Early studies on KGE introduce a margin-based pairwise ranking function to measure the Euclidean distance or similarity between the relational projection of entities (Nickel et al., 2011;Bordes et al., 2013;Yang et al., 2014;Trouillon et al., 2016). Among them, TransE (Bordes et al., 2013) is the most widely used KGE method, which views the relation as translation from a head entity to a tail entity. Recently, neural networks, such as neural tensor network (NTN) (Socher et al., 2013) and neural association model (NAM) (Liu et al., 2016) have been proposed to encode semantic matching and achieved remarkable predictive performance for KGC.\nTo increase the capacity of the KGE models, a larger embedding size with more parameters is a common technique in practice. As shown in Figure 1, the prediction performance of the KGC models such as DURA (Zhang et al., 2020) and RP (Chen et al., 2021) can be largely improved by increasing the graph embedding size. Although the large graph embedding often bring obvious performance improvements, it may also become the major obstacle for model deployment and realtime prediction, especially for memory-limited and resource-constrained devices.\nIn addition, the distribution of samples with longtail is prevalent in KGs (Zhang et al., 2019), where a large portion of relations have much fewer triples than other relations. However, most previous studies mainly focus on the predictive performance on overall test data, without taking long-tail samples into consideration. These models suffer from robust and generalization performance in the practical scenario. Although several recent works (Xiong et al., 2018;Sheng et al., 2020) have been proposed for few-shot KGC, these models are not adapted to the model compression frameworks.\nIn this paper, we propose a self-distillation framework with meta learning (MetaSD) for knowledge graph completion with dynamic pruning. First, we propose a dynamic pruning technique to obtain a small pruned model from a large source model at the start of each training epoch. Concretely, the pruning mask of the pruned model could be updated adaptively per epoch after updating the model weights. The pruned model is supposed to be more sensitive to the difficult-tomemorize samples (e.g., long-tail samples) than the source model. Second, we propose a one-step meta self-distillation method to distill comprehensive knowledge from the source model to the pruned model, where the two models co-evolve in a dynamic manner during the whole knowledge distillation process. The key idea is to use the performance of the pruned model, which is trained alongside the source model in one iteration, to improve the source model for the next iteration by borrowing the idea of learning to learn from meta learning (Finn et al., 2017). In particular, we define the objectives of the source model as functions of the pruned model's performance on a quiz set. The usage of \"gradient by gradient\" strategy makes the source model adjust to the learning state of the pruned model, and improves both the source and pruned models.\nThe main contributions of our method can be three-fold. (1) We propose a self-distillation framework to compress KG embeddings for KGC. The source and pruned models co-evolve in a dynamic manner during training, thus we can avoid pretraining a large model in advance, and the performance of the pruned model is not limited to that of a pre-trained large model. (2) We exploit the feedback from the pruned model to guide the source model with meta learning, making the source model transfer better knowledge to the pruned network. (3) Experimental results on two benchmark datasets show that our model achieves competitive performance compared to strong baselines, while being 10x smaller than other KGC models."
},
{
"figure_ref": [
"fig_1"
],
"heading": "Methodology",
"publication_ref": [
"b21",
"b9",
"b30"
],
"table_ref": [],
"text": "Problem Definition Suppose a KG can be viewed as a graph G = {(h, r, t)} ∈ E × R × E, where E and R represent the entity (node) set and relation (edge) set respectively. (h, r, t) represents a triple, where h, t and r indicate head entity, tail entity and the relation between two entities, respectively. Given the KG G, the goal of KGC is to infer missing links based on existing triples in the KG.\nModel Overview The overview of our MetaSD method is illustrated in Figure 2. We adopt Com-plEx (Trouillon et al., 2016) as our backbone model, which is treated as the source model T . First, we use a magnitude-based weight pruning method (Han et al., 2015;Zhu and Gupta, 2017) to obtain a pruned model S from the source model T . Second, we propose a one-step meta selfdistillation method for distilling comprehensive knowledge from the source model T to the pruned model S, where the two models co-evolve in a dynamic manner during training. Next, we introduce the proposed MetaSD method in detail."
},
{
"figure_ref": [],
"heading": "Network Pruning",
"publication_ref": [
"b9",
"b30"
],
"table_ref": [],
"text": "We use the magnitude-based weight pruning method (Han et al., 2015;Zhu and Gupta, 2017) to create a self-competitive compressed model S by pruning the source model T . In particular, we fix the pruning rate γ during the whole training process. At each iteration, we first calculate the sum of parameter numbers of all layers be pruned as P , and sort all the weights by their absolute values. Then, we prune a certain fraction (i.e., γ) of weights that have lowest absolute weight values. In particular, to dynamically adjust the pruned network S during each iteration, we prune the chosen weight by setting the corresponding values in a binary mask to zero, instead of directly setting the weights to zero."
},
{
"figure_ref": [],
"heading": "Self-Distillation via Meta Learning",
"publication_ref": [
"b24",
"b1",
"b1"
],
"table_ref": [],
"text": "We exploit the performance of the pruned model, which is trained alongside the source model in one iteration, to improve the source model's knowledge transfer ability for the next iteration via meta learning. In particular, we alternately update the pruned model S based on the output of the source model T and optimize the source model T based on the pruned model's performance via meta learning. Model S with Knowledge Distillation Formally, we use the function S(x i ; θ S ) to denote the soft prediction of the compressed model, where θ S represents the parameters of the pruned model S. We calculate the cross-entropy loss L S CE (θ S ) on the training data in current batch as:\nL S CE (θ S ) = 1 N N i=1 CE(y i , S(x i ; θ S )) (1)\nwhere N denotes the number of training samples. CE(•) represents the cross-entropy function.\nTo further improve the performance of S, we also design a knowledge distillation loss L S KD that encourages the output of S to mimic that of T . In particular, we minimize the Kullback-Leibler Divergence (KL-divergence) between the output distributions of S and T by:\nL S KD (θ S , θ T ) = 1 N N i=1 KL S(x i ; θ S )||T (x i ; θ T )(2\n) where θ T represents the parameters of T .\nThe cross-entropy loss L S CE and the knowledge distillation loss L S KD are combined to form the overall loss L S for the compressed model S as:\nL S (θ S , θ T ) = αL S CE (θ S ) + (1 -α)L S KD (θ S , θ T ) (3)\nwhere α is a hyperparameter to balance the relative importance of the two loss functions.\nModel T with Meta Learning We exploit feedback from the compressed model's learning state to improve the source model's knowledge transfer ability throughout the distillation process, instead of keeping the source model T fixed in the training process. We train both T and S in an iterative manner until convergence. This interaction between the two models can be seen as a form of meta learning with a bi-level optimization process, which comprises three steps: Virtual-Train, Meta-Train, and Actual-Train (Xu et al., 2021). That is, the compressed model S is the inner-learner and the source model T is the meta-learner.\nFor each training step, we first copy the parameters θ S of the compressed model S to a \"virtual\" compressed model S , and then update the parameters θ S (θ T ) of the \"virtual\" compressed model S with SGD (Bottou, 2012) for the Virtual-Train as:\nθ S (θ T ) = θ S -λ∇ θ S L S (θ S , θ T )(4)\nThen, the source model T is optimized based on the feedback of S on a held-out quiz set Q. We perform a derivative over a derivative (a Hessian matrix) to update θ T , by using a retained computational graph of θ S in order to compute derivatives with respect to θ T . The source model T is optimized by minimizing the cross-entropy loss over the quiz set Q for the Meta-Train as:\nL T CE θ S θ T ) = 1 M M i=1 CE(y i , S(x i ; θ S ) (5)\nwhere M is the training samples in the quiz set Q.\nx and y denote the input sample and corresponding label in quiz set q ∈ Q, respectively.\nFinally, we update the source model T with SGD (Bottou, 2012) as follows:\nθ T ← θ T -µ∇ θ T L T CE (θ S θ T ) (6)\nwhere µ is the learning rate for the Meta-Train."
},
{
"figure_ref": [],
"heading": "Mutual Update of T and S for Self-Distillation",
"publication_ref": [
"b2",
"b29"
],
"table_ref": [],
"text": "In our self-distillation framework, the source model T and the compressed model S co-evolve in a dynamic manner during the whole KD process. Instead of updating T with cross-entropy loss, we learn both T and S models mutually. (Broscheit et al., 2020). CP, ComplEx and RESCAL are implemented by following (Zhang et al., 2020).\nFormally, for the Actual-Train, we first update the compressed model's parameters θ S with the training data and the updated parameters θ T as:\nθ S = θ S -λ∇ θ S L S (θ S , θ T ) (7)\nThe source model T is also optimized by the combination of the cross-entropy loss L T CE and the knowledge distillation loss L T KD as:\nL T CE (θ T ) = 1 N N i=1 CE(y i , T (x i ; θ T ) (8) L T KD (θ S , θ T ) = 1 N N i=1 KL S(x i ; θ S )||T (x i ; θ T ) (9) L T (θ S , θ T ) = βL T CE (θ S ) + (1 -β)L T KD (θ S , θ T ) (10)\nwhere β is a hyperparameter to balance the relative importance of the two loss functions. We first update the source model's parameters θ T with the training data and the updated parameters θ S as:\nθ T = θ T -θ T L T (θ S , θ T ) (11)\nWe train the source and compressed models in an iterative manner until convergence. Overall, the self-distillation with meta learning is defined in Algorithm 1."
},
{
"figure_ref": [],
"heading": "Algorithm 1 Self-Distillation with Meta Learning",
"publication_ref": [],
"table_ref": [],
"text": "Require: train set D, quiz set Q, source model θ T Require: learning rate λ, learning rate µ, i ← 0 1: repeat\n2: i ← i + 1 3:\nSample a batch of training data x from D 4:\nGet pruned model θ S by pruning θ T"
},
{
"figure_ref": [],
"heading": "5:",
"publication_ref": [],
"table_ref": [],
"text": "Copy pruned model's parameters θ S to a \"virtual\" pruned model θ S"
},
{
"figure_ref": [],
"heading": "6:",
"publication_ref": [
"b20",
"b4"
],
"table_ref": [],
"text": "Update θ S with x and θ\nT : # Virtual-Train θ S (θ T ) = θ S -λ∇ θ S L S (x; θ S , θ T ) 7:\nSample a batch of quiz data q from Q 8:\nUpdate θ T with q and θ S : # Meta-Train\nθ T ← θ T -µ∇ θ T L T CE (q; θ S (θ T )) 9:\nMutual update θ T and θ S : # Actual-Train\nθ S (θ T ) = θ S -λ∇ θ S L S (x; θ S ; θ T ) θ T (θ S ) = θ T -λ∇ θ T L T (x; θ S ; θ T ) 10: until i == max iterations Output: source model θ T and pruned model θ S 3 Experimental Setup 3.1 Datasets\nWe conduct experiments on two KGC benchmark datasets: WN18RR (Toutanova and Chen, 2015) and FB15k-237 (Dettmers et al., 2018). WN18RR consists of 40,943 entities and 11 relations, and there are 86k/3k/3k instances for training/validation/testing respectively. FB15k-237 contains 14,541 entities and 237 relations, and there are 272k/17k/20k instances for training/validation/testing."
},
{
"figure_ref": [],
"heading": "Baseline Methods",
"publication_ref": [
"b13",
"b0",
"b21",
"b19",
"b29",
"b3",
"b10",
"b28",
"b3"
],
"table_ref": [],
"text": "We compare MetaSD with several strong KGC baselines, including CP (F.L, 1927), RESCAL (Nickel et al., 2011), TransE (Bordes et al., 2013), ComplEx (Trouillon et al., 2016), RotatE (Sun et al., 2019), DURA (Zhang et al., 2020), and RP (Chen et al., 2021). We also compare MetaSD with two widely used KD methods: knowledge distillation (KD) (Hinton et al., 2015) and deep mutual learning (DML) (Zhang et al., 2018), where the pre-trained RP (Chen et al., 2021) is used as their teacher model."
},
{
"figure_ref": [],
"heading": "Implementation Details",
"publication_ref": [
"b21",
"b29",
"b3",
"b5",
"b0"
],
"table_ref": [],
"text": "The source model of MetaSD is initialized with ComplEx (Trouillon et al., 2016), following the previous work (Zhang et al., 2020). Similar to Chen et al. (2021), we add relation prediction as an auxiliary task. We set the pruning rate γ to 0.9 to strike a balance between the effectiveness and efficiency of the model. We set balance hyperparameters α = β = 0.5. We choose Adagrad (Duchi et al., 2011) as the optimizer and the learning rate µ to 1e -4 and λ to 1e -1 . The quiz set is randomly sampled from training data and then fixed. We adopt widely used filtered evaluation metrics of mean reciprocal rank (MRR), Hits@1, Hits@3, and Hits@10 as described in (Bordes et al., 2013).\n4 Experimental Results"
},
{
"figure_ref": [],
"heading": "Overall Performance",
"publication_ref": [],
"table_ref": [
"tab_0"
],
"text": "As shown in Table 1, we report the results of the compressed model (denoted as MetaSD). Note that the parameters of RASCAL are proportional to the square of the number of relations, resulting in large differences in size between the two datasets. We observe that MetaSD achieves competitive performance compared to other high-dimensional baseline models on the two datasets for KGC, while being 10x smaller than baseline methods. In addition, MetaSD also outperforms than two widely used KD methods that have the same size and dimension with MetaSD."
},
{
"figure_ref": [],
"heading": "Long-tail Evaluation",
"publication_ref": [],
"table_ref": [
"tab_1"
],
"text": "We investigate the effectiveness of MetaSD for dealing with the long-tail samples. In particular, we collect the long-tail samples from the FB15k-237 test set by choosing the relations that have fewer than 1000 training instances. In total, there are 187 relations, which accounts for 79% of the total relation types but only 24% of train set. Table 2 reports the results of MetaSD and compared methods on the long-tail set. MetaSD significantly outperforms other models on the long-tail samples, which verifies the effectiveness of MetaSD in tackling the long-tail samples."
},
{
"figure_ref": [],
"heading": "Ablation Study",
"publication_ref": [],
"table_ref": [],
"text": "In order to verify the effectiveness of the pruning and meta learning modules, we conduct ablation evaluation on the proposed MetaSD on the FB15k-237 dataset. As shown in on the performance of the proposed MetaSD. This is because that meta learning can make the source network transfer rich knowledge to the pruned network effectively in the self-distillation process. In addition, the improvement of the self-pruning strategy is also significant since self-pruning can help the model learn discriminative representations and deal with the long-tail samples. It is no surprise that combining both factors achieves the best performance on in terms of four evaluation metrics."
},
{
"figure_ref": [],
"heading": "Generalization",
"publication_ref": [
"b29"
],
"table_ref": [
"tab_4"
],
"text": "To demonstrate the robustness of our framework, we also implement MetaSD on two additional backbone models (e.g., CP and RESCAL). These two backbone models are implemented and initialized by following the paper (Zhang et al., 2020). As shown in Table 4, our compressed model achieves substantially better performance than the larger baseline models based on two different backbone models."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "In this paper, we proposed a self-distillation framework with meta learning for graph embedding compression. Concretely, we proposed a one-step meta self-distillation method for distilling comprehensive knowledge from the source model to the pruned model, where the two models co-evolved in a dynamic manner during training. Experimental results showed that our model achieved competitive performance compared to strong baseline methods, while being 10x smaller than baseline methods."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [],
"table_ref": [],
"text": "To better understand the limitations of the proposed model, we carry out an analysis of the error predictions made by MetaSD. In particular, we primarily analyze the relations in the FB15k-237 test set, whose MRR scores are less than 0.2. Most of the incorrectly predicted relations are the \"location\" and \"relationships\" related relation types, such as place of birth/death, spouse, and sibling. We reveal several reasons of the bad cases, which can be divided into two primarily categories. First, MetaSD fails to predict some instances that require the multi-hop reasoning to get the correct answers, since our model does not consider the complex multi-hop paths during the knowledge graph representation learning. Second, MetaSD fails to predict some instances, where there are a large number of candidate entities to reason for a relation type (e.g., the location relation). One possible solution is to devise a two-step ranking method by filtering most of the irrelevant entities in a coarse-grained way and then distinguish the confusing entities with a fine-grained method."
},
{
"figure_ref": [],
"heading": "Acknowledgements",
"publication_ref": [],
"table_ref": [],
"text": "This work was partially supported by National Key R&D Program of China (No. 2019YFB2102500), National Natural Science Foundation of China (No. 61906185), Youth Innovation Promotion Association of CAS China (No. 2020357), Shenzhen Science and Technology Innovation Program (Grant No. KQTD20190929172835662), Shenzhen Basic Research Foundation (No. JCYJ20210324115614039 and No. JCYJ20200109113441941)."
}
] | In this paper, we propose a self-distillation framework with meta learning (MetaSD) for knowledge graph completion with dynamic pruning, which aims to learn compressed graph embeddings and tackle the long-tail samples. Specifically, we first propose a dynamic pruning technique to obtain a small pruned model from a large source model, where the pruning mask of the pruned model could be updated adaptively per epoch after the model weights are updated. The pruned model is supposed to be more sensitive to difficult-tomemorize samples (e.g., long-tail samples) than the source model. Then, we propose a one-step meta self-distillation method for distilling comprehensive knowledge from the source model to the pruned model, where the two models co-evolve in a dynamic manner during training. In particular, we exploit the performance of the pruned model, which is trained alongside the source model in one iteration, to improve the source model's knowledge transfer ability for the next iteration via meta learning. Extensive experiments show that MetaSD achieves competitive performance compared to strong baselines, while being 10x smaller than baselines. | Self-Distillation with Meta Learning for Knowledge Graph Completion | [
{
"figure_caption": "Figure 1 :1Figure 1: The MRR scores w.r.t. the graph embedding sizes of DURA and RP on FB15k-237.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: The overview of MetaSD framework. (a) We prune the teacher T to obtain the student S and perform knowledge distillation on training data to update the temporary copy S from S. Then, the source model T is optimized based on the feedback of S on a held-out quiz set Q; (b) We discard S and optimize the meta-updated T and real S alternately by performing mutual learning on the training data.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Experimental results on FB15k-237 and WN18RR test sets for KGC. The results with are taken from LibKGE",
"figure_data": "ModelFB15k-237WN18RRDimMRRHits@1Hits@3 Hits@10SizeMRRHits@1Hits@3 Hits@10SizeTransE0.3130.2210.3470.497-0.2280.0530.3680.520--RotatE0.3330.2400.3680.522-0.4780.4390.4940.553--CP0.3330.2470.3600.50850M0.4380.4140.4440.485156M2kRESCAL0.3530.2640.3830.528125M0.4550.4190.4600.49326M-ComplEx0.3460.2560.3700.52560M0.4600.4280.4750.522156M2kDURA0.3710.2760.4080.56060M0.4910.4490.5030.571156M2kRP0.3880.2980.4250.56860M0.4880.4430.5050.568156M2kKD0.3710.2820.4080.5506M0.4700.4270.4850.53015M0.2kDML0.3730.2800.4100.5636M0.4720.4290.4850.53515M0.2kMetaSD0.3910.3000.4280.5716M0.4910.4470.5040.57015M0.2k",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Results on long-tail data from FB15k-237.",
"figure_data": "ModelMRRH@1H@3H@10DURA0.4520.3540.4980.644RP0.4620.3720.5040.645MetaSD-T0.4680.3800.5100.642MetaSD0.4710.3810.5120.646ModelMRRH@1H@[email protected]/o P0.3810.2920.4150.561w/o M0.3780.2870.4120.555w/o P&M0.3730.2800.4100.563",
"figure_id": "tab_1",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Results of ablation study on FB15k-237. P and M denote the pruning and meta learning techniques.",
"figure_data": "",
"figure_id": "tab_2",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Table 3, we can observe that the meta learning technique has great impact",
"figure_data": "ModelMRRH@1H@3H@10SizeCP0.3330.2470.3600.50850MRESCAL0.3530.2640.3830.528125MMetaSD-CP0.3670.2700.3960.5575MMetaSD-RESCAL0.3720.2760.4050.56112.5M",
"figure_id": "tab_3",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Results of MetaSD on FB15k-237 by using different backbone models.",
"figure_data": "",
"figure_id": "tab_4",
"figure_label": "4",
"figure_type": "table"
}
] | Yunshui Li; Junhao Liu; Chengming Li; Min Yang | [
{
"authors": "Antoine Bordes; Nicolas Usunier; Alberto Garcia-Duran; Jason Weston; Oksana Yakhnenko",
"journal": "Advances in neural information processing systems",
"ref_id": "b0",
"title": "Translating embeddings for modeling multirelational data",
"year": "2013"
},
{
"authors": "Léon Bottou",
"journal": "Springer",
"ref_id": "b1",
"title": "Stochastic gradient descent tricks",
"year": "2012"
},
{
"authors": "Samuel Broscheit; Daniel Ruffinelli; Adrian Kochsiek; Patrick Betz; Rainer Gemulla",
"journal": "",
"ref_id": "b2",
"title": "LibKGE -A knowledge graph embedding library for reproducible research",
"year": "2020"
},
{
"authors": "Yihong Chen; Pasquale Minervini; Sebastian Riedel; Pontus Stenetorp",
"journal": "",
"ref_id": "b3",
"title": "Relation prediction as an auxiliary training objective for improving multirelational graph representations",
"year": "2021"
},
{
"authors": "Tim Dettmers; Pasquale Minervini; Pontus Stenetorp; Sebastian Riedel",
"journal": "",
"ref_id": "b4",
"title": "Convolutional 2d knowledge graph embeddings",
"year": "2018"
},
{
"authors": "John Duchi; Elad Hazan; Yoram Singer",
"journal": "Journal of machine learning research",
"ref_id": "b5",
"title": "Adaptive subgradient methods for online learning and stochastic optimization",
"year": "2011"
},
{
"authors": "Chelsea Finn; Pieter Abbeel; Sergey Levine",
"journal": "",
"ref_id": "b6",
"title": "Model-agnostic meta-learning for fast adaptation of deep networks",
"year": "2017"
},
{
"authors": " Pmlr",
"journal": "",
"ref_id": "b7",
"title": "",
"year": ""
},
{
"authors": "F L Hitchcock",
"journal": "Journal of Mathematics and Physics",
"ref_id": "b8",
"title": "The expression of a tensor or a polyadic as a sum of products",
"year": "1927"
},
{
"authors": "Song Han; Huizi Mao; William J Dally",
"journal": "",
"ref_id": "b9",
"title": "Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding",
"year": "2015"
},
{
"authors": "Geoffrey Hinton; Oriol Vinyals; Jeff Dean",
"journal": "",
"ref_id": "b10",
"title": "Distilling the knowledge in a neural network",
"year": "2015"
},
{
"authors": "Quan Liu; Hui Jiang; Andrew Evdokimov; Zhen-Hua Ling; Xiaodan Zhu; Si Wei; Yu Hu",
"journal": "",
"ref_id": "b11",
"title": "Probabilistic reasoning via deep learning: Neural association models",
"year": "2016"
},
{
"authors": "Isaiah Onando Mulang; ' ; Kuldeep Singh; Chaitali Prabhu; Abhishek Nadgeri; Johannes Hoffart; Jens Lehmann",
"journal": "",
"ref_id": "b12",
"title": "Evaluating the impact of knowledge graph context on entity disambiguation models",
"year": "2020"
},
{
"authors": "Maximilian Nickel; Hans-Peter Volker Tresp; Kriegel",
"journal": "",
"ref_id": "b13",
"title": "A three-way model for collective learning on multi-relational data",
"year": "2011"
},
{
"authors": "Ying Shen; Yang Deng; Min Yang; Yaliang Li; Nan Du; Wei Fan; Kai Lei",
"journal": "",
"ref_id": "b14",
"title": "Knowledge-aware attentive neural network for ranking question answer pairs",
"year": "2018"
},
{
"authors": "Jiawei Sheng; Shu Guo; Zhenyu Chen; Juwei Yue; Lihong Wang; Tingwen Liu; Hongbo Xu",
"journal": "",
"ref_id": "b15",
"title": "Adaptive attentional network for fewshot knowledge graph completion",
"year": "2020"
},
{
"authors": "Shuzheng Si; Zefan Cai; Shuang Zeng; Guoqiang Feng; Jiaxing Lin; Baobao Chang",
"journal": "",
"ref_id": "b16",
"title": "Santa: Separate strategies for inaccurate and incomplete annotation noise in distantly-supervised named entity recognition",
"year": "2023"
},
{
"authors": "Shuzheng Si; Shuang Zeng; Jiaxing Lin; Baobao Chang",
"journal": "",
"ref_id": "b17",
"title": "Scl-rai: Span-based contrastive learning with retrieval augmented inference for unlabeled entity problem in ner",
"year": "2022"
},
{
"authors": "Richard Socher; Danqi Chen; Christopher D Manning; Andrew Ng",
"journal": "Advances in neural information processing systems",
"ref_id": "b18",
"title": "Reasoning with neural tensor networks for knowledge base completion",
"year": "2013"
},
{
"authors": "Zhiqing Sun; Zhi-Hong Deng; Jian-Yun Nie; Jian Tang",
"journal": "",
"ref_id": "b19",
"title": "Rotate: Knowledge graph embedding by relational rotation in complex space",
"year": "2019"
},
{
"authors": "Kristina Toutanova; Danqi Chen",
"journal": "",
"ref_id": "b20",
"title": "Observed versus latent features for knowledge base and text inference",
"year": "2015"
},
{
"authors": "Théo Trouillon; Johannes Welbl; Sebastian Riedel; Éric Gaussier; Guillaume Bouchard",
"journal": "",
"ref_id": "b21",
"title": "Complex embeddings for simple link prediction",
"year": "2016"
},
{
"authors": " Pmlr",
"journal": "",
"ref_id": "b22",
"title": "",
"year": ""
},
{
"authors": "Wenhan Xiong; Mo Yu; Shiyu Chang; Xiaoxiao Guo; William Yang; Wang ",
"journal": "",
"ref_id": "b23",
"title": "One-shot relational learning for knowledge graphs",
"year": "2018"
},
{
"authors": "Youjiang Xu; Linchao Zhu; Lu Jiang; Yi Yang",
"journal": "",
"ref_id": "b24",
"title": "Faster meta update strategy for noise-robust deep learning",
"year": "2021"
},
{
"authors": "Nan Zhao Yan; Peng Duan; Ming Chen; Jianshe Zhou; Zhoujun Zhou; Li",
"journal": "",
"ref_id": "b25",
"title": "Building task-oriented dialogue systems for online shopping",
"year": "2017"
},
{
"authors": "Bishan Yang; Wen-Tau Yih; Xiaodong He; Jianfeng Gao; Li Deng",
"journal": "",
"ref_id": "b26",
"title": "Embedding entities and relations for learning and inference in knowledge bases",
"year": "2014"
},
{
"authors": "Ningyu Zhang; Shumin Deng; Zhanlin Sun; Guanying Wang; Xi Chen; Wei Zhang; Huajun Chen",
"journal": "",
"ref_id": "b27",
"title": "Long-tail relation extraction via knowledge graph embeddings and graph convolution networks",
"year": "2019"
},
{
"authors": "Ying Zhang; Tao Xiang; Timothy M Hospedales; Huchuan Lu",
"journal": "",
"ref_id": "b28",
"title": "Deep mutual learning",
"year": "2018"
},
{
"authors": "Zhanqiu Zhang; Jianyu Cai; Jie Wang",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b29",
"title": "Duality-induced regularizer for tensor factorization based knowledge graph completion",
"year": "2020"
},
{
"authors": "Michael Zhu; Suyog Gupta",
"journal": "",
"ref_id": "b30",
"title": "To prune, or not to prune: exploring the efficacy of pruning for model compression",
"year": "2017"
}
] | [
{
"formula_coordinates": [
3,
97.47,
390.32,
191.67,
33.71
],
"formula_id": "formula_0",
"formula_text": "L S CE (θ S ) = 1 N N i=1 CE(y i , S(x i ; θ S )) (1)"
},
{
"formula_coordinates": [
3,
76.51,
533.96,
208.39,
13.38
],
"formula_id": "formula_1",
"formula_text": "L S KD (θ S , θ T ) = 1 N N i=1 KL S(x i ; θ S )||T (x i ; θ T )(2"
},
{
"formula_coordinates": [
3,
76.51,
604.61,
212.62,
12.65
],
"formula_id": "formula_2",
"formula_text": "L S (θ S , θ T ) = αL S CE (θ S ) + (1 -α)L S KD (θ S , θ T ) (3)"
},
{
"formula_coordinates": [
3,
341.52,
447.85,
182.89,
11.73
],
"formula_id": "formula_3",
"formula_text": "θ S (θ T ) = θ S -λ∇ θ S L S (θ S , θ T )(4)"
},
{
"formula_coordinates": [
3,
311.79,
571.19,
212.62,
15.49
],
"formula_id": "formula_4",
"formula_text": "L T CE θ S θ T ) = 1 M M i=1 CE(y i , S(x i ; θ S ) (5)"
},
{
"formula_coordinates": [
3,
343.5,
656.04,
180.91,
14.22
],
"formula_id": "formula_5",
"formula_text": "θ T ← θ T -µ∇ θ T L T CE (θ S θ T ) (6)"
},
{
"formula_coordinates": [
4,
116.66,
270.61,
172.47,
11.73
],
"formula_id": "formula_6",
"formula_text": "θ S = θ S -λ∇ θ S L S (θ S , θ T ) (7)"
},
{
"formula_coordinates": [
4,
76.51,
325.37,
212.63,
50.45
],
"formula_id": "formula_7",
"formula_text": "L T CE (θ T ) = 1 N N i=1 CE(y i , T (x i ; θ T ) (8) L T KD (θ S , θ T ) = 1 N N i=1 KL S(x i ; θ S )||T (x i ; θ T ) (9) L T (θ S , θ T ) = βL T CE (θ S ) + (1 -β)L T KD (θ S , θ T ) (10)"
},
{
"formula_coordinates": [
4,
114.68,
434.38,
174.45,
11.73
],
"formula_id": "formula_8",
"formula_text": "θ T = θ T -θ T L T (θ S , θ T ) (11)"
},
{
"formula_coordinates": [
4,
76.98,
563.82,
66.61,
22.94
],
"formula_id": "formula_9",
"formula_text": "2: i ← i + 1 3:"
},
{
"formula_coordinates": [
4,
76.98,
631.72,
205.38,
36.32
],
"formula_id": "formula_10",
"formula_text": "T : # Virtual-Train θ S (θ T ) = θ S -λ∇ θ S L S (x; θ S , θ T ) 7:"
},
{
"formula_coordinates": [
4,
76.98,
683.81,
189.13,
24.89
],
"formula_id": "formula_11",
"formula_text": "θ T ← θ T -µ∇ θ T L T CE (q; θ S (θ T )) 9:"
},
{
"formula_coordinates": [
4,
70.87,
228.02,
355.16,
536.17
],
"formula_id": "formula_12",
"formula_text": "θ S (θ T ) = θ S -λ∇ θ S L S (x; θ S ; θ T ) θ T (θ S ) = θ T -λ∇ θ T L T (x; θ S ; θ T ) 10: until i == max iterations Output: source model θ T and pruned model θ S 3 Experimental Setup 3.1 Datasets"
}
] |
|
2023-10-18 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b14",
"b1",
"b4",
"b34",
"b21",
"b3",
"b11",
"b37",
"b0",
"b22",
"b17",
"b31",
"b23",
"b16",
"b33",
"b20"
],
"table_ref": [],
"text": "Multimodal named entity recognition (MNER) has recently garnered significant attention (Lu et al., 2018). Users generate copious amounts of unstructured content primarily consisting of images and text on social media. The textual component in The \"CP3\" in the text is a class of entities that are difficult to predict successfully by existing studies. PGIM demonstrates successful prediction of such entities with an approach more similar to human cognitive processes by endowing ChatGPT with reasonable heuristics.\nthese posts possesses inherent characteristics associated with social media, including brevity and an informal style of writing. These unique characteristics pose challenges for traditional named entity recognition (NER) approaches (Chiu and Nichols, 2016;Devlin et al., 2018). To leverage the multimodal features and improve the NER performance, numerous previous works have attempted to align images and text implicitly using various attention mechanisms (Yu et al., 2020;Sun et al., 2021), but these Image-Text (I+T) paradigm methods have several significant limitations. Limitation 1. The feature distribution of different modalities exhibits variations, which hinders the model to learn aligned representations across diverse modalities. Limitation 2. The image feature extractors used in these methods are trained on datasets like ImageNet (Deng et al., 2009) and COCO (Lin et al., 2014), where the labels primarily consist of nouns rather than named entities. There are obvious deviations between the labels of these datasets and the named entities we aim to recognize. Given these limitations, these multimodal fusion methods may not be as effective as state-of-the-art language models that solely focus on text.\nWhile MNER is a multimodal task, the contributions of image and text modalities to this task are not equivalent. When the image cannot provide more interpretation information for the text, the image information can even be discarded and ignored. In addition, recent studies (Wang et al., 2021b;Zhang et al., 2022) has shown that introducing additional document-level context on the basis of the original text can significantly improve the performance of NER models. Therefore, recent studies (Wang et al., 2021a(Wang et al., , 2022a) ) aim to solve the MNER task using the Text-Text (T+T) paradigm. In these approaches, images are reasonably converted into textual representations through techniques such as image caption and optical character recognition (OCR). Apparently, the inter-text attention mechanism is more likely to outperform the cross-modal attention mechanism. However, existing second paradigm methods still exhibit certain potential deficiencies:\n(i) For the methods that solely rely on in-sample information, they often fall short in scenarios that demand additional external knowledge to enhance text comprehension.\n(ii) For those existing methods that consider introducing external knowledge, the relevant knowledge retrieved from external explicit knowledge base (e.g., Wikipedia) is too redundant. These low-relevance extended knowledge may even mislead the model's understanding of the text in some cases.\nRecently, the field of large language models (LLMs) is rapidly advancing with intriguing new findings and developments (Brown et al., 2020;Touvron et al., 2023). On the one hand, recent research on LLMs (Qin et al., 2023;Wei et al., 2023;Wang et al., 2023a) shows that the effect of the generative model in the sequence labeling task has obvious shortcomings. On the other hand, LLMs achieves promising results in various NLP (Vilar et al., 2022;Moslem et al., 2023) and multimodal tasks (Yang et al., 2022;Shao et al., 2023). These LLMs with in-context learning capability can be perceived as a comprehensive representation of internet-based knowledge and can offer highquality auxiliary knowledge typically. So we ask: Is it possible to activate the potential of ChatGPT in MNER task by endowing ChatGPT with reasonable heuristics?\nIn this paper, we present PGIM -a conceptually simple framework that aims to boost the performance of model by Prompting ChatGPT In MNER to generate auxiliary refined knowledge. As shown in Figure 1, the additional auxiliary refined knowledge generated in this way overcomes the limitations of (i) and (ii). We begin by manually annotating a limited set of samples. Subsequently, PGIM utilizes the Multimodal Similar Example Awareness module to select relevant instances, and seamlessly integrates them into a meticulously crafted prompt template tailored for MNER task, thereby introducing pertinent knowledge. This approach effectively harnesses the in-context few-shot learning capability of ChatGPT. Finally, the auxiliary refined knowledge generated by heuristic approach of ChatGPT is subsequently combined with the original text and fed into a downstream text model for further processing.\nPGIM outperforms all state-of-the-art models based on the Image-Text and Text-Text paradigms on two classical MNER datasets and exhibits a stronger robustness and generalization capability. Moreover, compared with some previous methods, PGIM is friendly to most researchers, its implementation requires only a single GPU and a reasonable number of ChatGPT invocations."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Multimodal Named Entity Recognition",
"publication_ref": [
"b15",
"b39",
"b35",
"b40",
"b38"
],
"table_ref": [],
"text": "Considering the inherent characteristics of social media text, previous approaches (Moon et al., 2018;Zheng et al., 2020;Zhang et al., 2021;Zhou et al., 2022;Zhao et al., 2022) have endeavored to incorporate visual information into NER. They employ diverse cross-modal attention mechanisms to facilitate the interaction between text and images. Recently, Wang et al. (2021a) points out that the performance limitations of such methods are largely attributed to the disparities in distribution between different modalities. Despite Wang et al. (2022c) try to mitigate the aforementioned issues by using further refining cross-modal attention, training this end-to-end cross-modal Transformer architectures imposes significant demands on computational resources. Due to the aforementioned limitations, ITA (Wang et al., 2021a) and MoRe (Wang et al., 2022a) attempt to use a new paradigm to address MNER. ITA circumvents the challenge of multi-modal alignment by forsaking the utilization of raw visual features and opting for OCR and image captioning techniques to convey image information. MoRe assists prediction by retrieving additional knowledge related to text and images from explicit knowledge base. However, none of these methods can adequately fulfill the requisite knowledge needed by the model to comprehend the text. The advancement of LLMs address the limitations identified in the aforementioned methods. While the direct prediction of named entities by LLMs in the full-shot case may not achieve comparable performance to task-specific models, we can utilize LLMs as an implicit knowledge base to heuristically generate further interpretations of text. This method is more aligned with the cognitive and reasoning processes of human."
},
{
"figure_ref": [],
"heading": "In-context learning",
"publication_ref": [
"b0",
"b4",
"b12",
"b18",
"b9"
],
"table_ref": [],
"text": "With the development of LLMs, empirical studies have shown that these models (Brown et al., 2020) exhibit an interesting emerging behavior called In-Context Learning (ICL). Different from the paradigm of pre-training and then fine-tuning language models like BERT (Devlin et al., 2018), LLMs represented by GPT have introduced a novel in-context few-shot learning paradigm. This paradigm requires no parameter updates and can achieve excellent results with just a few examples from downstream tasks. Since the effect of ICL is strongly related to the choice of demonstration examples, recent studies have explored several effective example selection methods, e.g., similaritybased retrieval method (Liu et al., 2021;Rubin et al., 2021), validation set scores based selection (Lee et al., 2021), gradient-based method (Wang et al., 2023b). These results indicate that reasonable example selection can improve the performance of LLMs."
},
{
"figure_ref": [
"fig_1"
],
"heading": "Methodology",
"publication_ref": [
"b8"
],
"table_ref": [],
"text": "PGIM is mainly divided into two stages. In the stage of generating auxiliary refined knowledge, PGIM leverages a limited set of predefined artificial samples and employs the Multimodal Similar Example Awareness (MSEA) module to carefully select relevant instances. These chosen examples are then incorporated into properly formatted prompts, thereby enhancing the heuristic guidance provided to ChatGPT for acquiring refined knowledge. (detailed in §3.2). In the stage of entity prediction based on auxiliary knowledge, PGIM combines the original text with the knowledge information generated by ChatGPT. This concatenated input is then fed into a transformer-based encoder to generate token representations. Finally, PGIM feeds the representations into the linear-chain Conditional Random Field (CRF) (Lafferty et al., 2001) layer to predict the probability distribution of the original text sequence (detailed in §3. 3). An overview of the PGIM is depicted in Figure 2."
},
{
"figure_ref": [],
"heading": "Preliminaries",
"publication_ref": [
"b0",
"b19",
"b33",
"b20"
],
"table_ref": [],
"text": "Before presenting the PGIM, we first formulate the MNER task, and briefly introduce the in-context learning paradigm originally developed by GPT-3 (Brown et al., 2020) and its adaptation to MNER.\nTask Formulation Consider treating the MNER task as a sequence labeling task. Given a sentence T = {t 1 , • • • , t n } with n tokens and its corresponding image I, the goal of MNER is to locate and classify named entities mentioned in the sentence as a label sequence y = {y 1 , • • • , y n }, where y i ∈ Y are predefined semantic categories with the BIO2 tagging schema (Sang and Veenstra, 1999).\nIn-context learning in MNER GPT-3 and its successor ChatGPT (hereinafter referred to collectively as GPT) are autoregressive language models pretrained on a tremendous dataset. During inference, in-context few-shot learning accomplishes new downstream tasks in the manner of text sequence generation tasks on frozen GPT models. Concretely, given a test input x, its target y is predicted based on the formatted prompt p(h, C, x) as the condition, where h refers to a prompt head describing the task and in-context\nC = {c 1 , • • • , c n } contains n in-context examples.\nAll the h, C, x, y are text sequences, and target y = {y 1 , • • • , y L } is a text sequence with the length of L. At each decoding step l, we have:\ny l = argmax y l p LLM (y l |p, y <l )\nwhere LLM represents the weights of the pretrained large language model, which are frozen for new tasks. Each in-context example c i = (x i , y i ) consists of an input-target pair of the task, and these examples is constructed manually or sampled from the training set.\nAlthough the GPT-42 can accept the input of multimodal information, this function is only in the in- ternal testing stage and has not yet been opened for public use. In addition, compared with ChatGPT, GPT-4 has higher costs and slower API request speeds. In order to enhance the reproducibility of PGIM, we still choose ChatGPT as the main research object of our method. And this paradigm provided by PGIM can also be used in GPT-4. In order to enable ChatGPT to complete the imagetext multimodal task, we use advanced multimodal pre-training model to convert images into image captions. Inspired by PICa (Yang et al., 2022) and Prophet (Shao et al., 2023) in knowledge-based VQA, PGIM formulates the testing input x as the following template:\nText: t \\n Image: p \\n Question: q \\n Answer:\nwhere t, p and q represent specific test inputs. \\n stands for a carriage return in the template. Similarly, each in-context example c i is defined with similar templates as follows:\nText: t i \\n Image: p i \\n Question: q \\n Answer: a i where t i , p i , q and a i refer to an text-imagequestion-answer quadruple retrieved from predefined artificial samples. The complete prompt template of MNER consisting of a fixed prompt head, some in-context examples, and a test input is fed to ChatGPT for auxiliary knowledge generation."
},
{
"figure_ref": [],
"heading": "Stage-1. Auxiliary Refined Knowledge Heuristic Generation",
"publication_ref": [
"b12",
"b33"
],
"table_ref": [],
"text": "Predefined artificial samples The key to making ChatGPT performs better in MNER is to choose suitable in-context examples. Acquiring accurately annotated in-context examples that precisely reflect the annotation style of the dataset and provide a means to expand auxiliary knowledge poses a significant challenge. And directly acquiring such examples from the original dataset is not feasible.\nTo address this issue, we employ a random sampling approach to select a small subset of samples from the training set for manual annotation. Specifically, for Twitter-2017 dataset, we randomly sample 200 samples from training set for manual labeling, and for Twitter-2015 dataset, the number is 120. The annotation process comprises two main components. The first part involves identifying the named entities within the sentences, and the second part involves providing comprehensive justification by considering the image and text content, as well as relevant knowledge. For many possibilities encounter in the labeling process, what the annotator needs to do is to correctly judge and interpret the sample from the perspective of humans. For samples where images and text are related, we directly state which entities in the text are emphasized by the image. For samples where the image and text are unrelated, we directly declare that the image description is unrelated to the text. Through artifi-cial annotation process, we emphasize the entities and their corresponding categories within the sentences. Furthermore, we incorporate relevant auxiliary knowledge to support these judgments. This meticulous annotation process serves as a guide for ChatGPT, enabling it to generate highly relevant and valuable responses.\nMultimodel Similar Example Awareness Module Since the few-shot learning ability of GPT largely depends on the selection of in-context examples (Liu et al., 2021;Yang et al., 2022), we design a Multimodel Similar Example Awareness (MSEA) module to select appropriate in-context examples. As a classic multimodal task, the prediction of MNER relies on the integration of both textual and visual information. Accordingly, PGIM leverages the fused features of text and image as the fundamental criterion for assessing similar examples. And this multimodal fusion feature can be obtained from various previous vanilla MNER models.\nDenote the MNER dataset D and predefined artificial samples G as: In previous studies, the fusion feature H after cross-attention projection into the highdimensional latent space was directly input to the decoder layer for the prediction of the result. Unlike them, PGIM chooses H as the judgment basis for similar examples. Because examples approximated in high-dimensional latent space are more likely to have the same mapping method and entity type. PGIM calculates the cosine similarity of the fused feature H between the test input and each predefined artificial sample. And top-N similar predefined artificial samples will be selected as incontext examples to enlighten ChatGPT generation auxiliary refined knowledge:\nD = {(t i , p i , y i )} M i=1 G = {(t j , p j , y j )} N\nI = argTopN j∈{1,2,...,N } H T H j ∥H∥ 2 ∥H j ∥ 2 I is the index set of top-N similar samples in G.\nThe in-context examples C are defined as follows:\nC = {(t j , p j , y j ) | j ∈ I}\nIn order to efficiently realize the awareness of similar examples, all the multimodal fusion features can be calculated and stored in advance.\nHeuristics-enhanced Prompt Generation After obtaining the in-context example C, PGIM builds a complete heuristics-enhanced prompt to exploit the few-shot learning ability of ChatGPT in MNER.\nA prompt head, a set of in-context examples, and a testing input together form a complete prompt. The prompt head describes the MNER task in natural language according to the requirements. Given that the input image and text may not always have a direct correlation, PGIM encourages ChatGPT to exercise its own discretion. The incontext examples are constructed from the results C = {c 1 , • • • , c n } of the MSEA module. For testing input, the answer slot is left blank for ChatGPT to generate. The complete format of the prompt template is shown in Appendix A.4."
},
{
"figure_ref": [],
"heading": "Stage-2. Entity Prediction based on Auxiliary Refined Knowledge",
"publication_ref": [
"b36",
"b14",
"b34",
"b10",
"b2",
"b13"
],
"table_ref": [],
"text": "Define the auxiliary knowledge generated by ChatGPT after in-context learning as\nZ = {z 1 , • • • , z m }, where m is the length of Z. PGIM concatenates the original text T = {t 1 , • • • , t n }\nwith the obtained auxiliary refining knowledge Z as [T ; Z] and feeds it to the transformer-based encoder:\n{h 1 , • • • , h n , • • • , h n+m } = embed([T ; Z])\nDue to the attention mechanism employed by the transformer-based encoder, the token representation H = {h 1 , • • • , h n } obtained encompasses pertinent cues from the auxiliary knowledge Z. Similar to the previous studies, PGIM feeds H to a standard linear-chain CRF layer, which defines the probability of the label sequence y given the input sentence T :\nP (y|T, Z) = n i=1 ψ(y i-1 , y i , h i ) y ′ ∈Y n i=1 ψ(y ′ i-1 , y ′ i , h i )\nwhere ψ(y i-1 , y i , h i ) and ψ(y ′ i-1 , y ′ i , h i ) are potential functions. Finally, PGIM uses the negative log-likelihood (NLL) as the loss function for the input sequence with gold labels y * :\nL NLL (θ) = -log P θ (y * |T, Z) 4 Experiments 4.1 Settings\nDatasets We conduct experiments on two public MNER datasets: Twitter-2015 (Zhang et al., 2018) and Twitter-2017 (Lu et al., 2018). These two classic MNER datasets contain 4000/1000/3257 and 3373/723/723 (train/development/test) image-text pairs posted by users on Twitter.\nModel Configuration PGIM chooses the backbone of UMT (Yu et al., 2020) as the vanilla MNER model to extract multimodal fusion features. This backbone completes multimodal fusion without too much modification. BLIP-2 (Li et al., 2023) as an advanced multimodal pre-trained model, is used for conversion from image to image caption. The version of ChatGPT used in experiments is gpt-3.5-turbo and sampling temperature is set to 0. For a fair comparison, PGIM chooses to use the same text encoder XLM-RoBERTa large (Conneau et al., 2019) as ITA (Wang et al., 2021a), PromptM-NER (Wang et al., 2022b), CAT-MNER (Wang et al., 2022c) and MoRe (Wang et al., 2022a).\nImplementation Details PGIM is trained by Pytorch on single NVIDIA RTX 3090 GPU. During training, we use AdamW (Loshchilov and Hutter, 2017) optimizer to minimize the loss function. We use grid search to find the learning rate for the embeddings within [1 × 10 -6 , 5 × 10 -5 ]. Due to the different labeling styles of two datasets, the learning rates of Twitter-2015 and Twitter-2017 are finally set to 5 × 10 -6 and 7 × 10 -6 . And we also use warmup linear scheduler to control the learning rate. The maximum length of the sentence input is set to 256, and the mini-batch size is set to 4. The model is trained for 25 epochs, and the model with the highest F1-score on the development set is selected to evaluate the performance on the test set. The number of in-context examples N in PGIM is set to 5. All of the results are averaged from 3 runs with different random seeds."
},
{
"figure_ref": [],
"heading": "Main Results",
"publication_ref": [
"b5",
"b4",
"b32",
"b34",
"b35",
"b6",
"b38"
],
"table_ref": [],
"text": "We compare PGIM with previous state-of-the-art approaches on MNER in Table 1. The first group of methods includes BiLSTM-CRF (Huang et al., 2015), BERT-CRF (Devlin et al., 2018) as well as the span-based NER models (e.g., BERT-span, RoBERTa-span (Yamada et al., 2020)), which only consider original text. The second group of methods includes several latest multimodal approaches for MNER task: UMT (Yu et al., 2020), UMGF (Zhang et al., 2021), MNER-QG (Jia et al., 2022), R-GCN (Zhao et al., 2022), ITA (Wang et al., 2021a), PromptMNER (Wang et al., 2022b), CAT-MNER (Wang et al., 2022c) and MoRe (Wang et al., 2022a), which consider both text and corresponding images.\nThe experimental results demonstrate the superiority of PGIM over previous methods. PGIM surpasses the previous state-of-the-art method MoRe (Wang et al., 2022a) in terms of performance. This suggests that compared with the auxiliary knowledge retrieved by MoRe (Wang et al., 2022a) from Wikipedia, our refined auxiliary knowledge offers more substantial support. Furthermore, PGIM exhibits a more significant improvement in Twitter-2017 compared with Twitter-2015. This can be attributed to the more complete and standardized labeling approach adopted in Twitter-2017, in contrast to Twitter-2015. Apparently, the quality of dataset annotation has a certain influence on the accuracy of MNER model. In cases where the dataset annotation deviates from the ground truth, accurate and refined auxiliary knowledge leads the model to prioritize predicting the truly correct entities, since the process of ChatGPT heuristically generating auxiliary knowledge is not affected by mislabeling. This phenomenon coincidentally highlights the robustness of PGIM. The ultimate objective of the MNER is to support downstream tasks effectively. Obviously, downstream tasks of MNER expect to receive MNER model outputs that are unaffected by irregular labeling in the training dataset. We further demonstrate this argument through a case study, detailed in the Appendix A.3."
},
{
"figure_ref": [],
"heading": "Detailed Analysis",
"publication_ref": [
"b2",
"b7"
],
"table_ref": [
"tab_0",
"tab_0",
"tab_5"
],
"text": "Impact of different text encoders on performance As shown in Table 2, We perform experiments by replacing the encoders of all XLM-RoBERTa large (Conneau et al., 2019) MNER methods with BERT base (Kenton and Toutanova, 2019). Baseline BERT represents inputting original samples into BERT-CRF. All of the results are averaged from 3 runs with different random seeds. The marker * refers to significant test p-value < We think the reasons for this phenomenon are as follows: XLM-RoBERTa large conceals the defects of previous MNER methods through its strong encoding ability, and these defects are further amplified after using BERT base . For example, the encoding ability of BERT base on long text is weaker than XLM-RoBERTa large , and the additional knowledge retrieved by MoRe Image/Text is much longer than PGIM. Therefore, as shown in Table 2 andTable 5, the performance loss of MoRe Image/Text is larger than the performance loss of PGIM after replacing BERT base . The BIO annotation method is not considered in this experiment because it is a little difficult for ChatGPT. Only the complete match will be considered, and only if the entity boundary and entity type are both accurately predicted, we judge it as a correct prediction."
},
{
"figure_ref": [],
"heading": "Compared with direct prediction of ChatGPT",
"publication_ref": [
"b17"
],
"table_ref": [],
"text": "The results show that the performance of Chat-GPT on MNER is far from satisfactory compared with PGIM in the full-shot case, which once again confirms the previous conclusion of ChatGPT on NER (Qin et al., 2023). In other words, when we have enough training samples, only relying on Chat-GPT itself will not be able to achieve the desired effect. The capability of ChatGPT shines in scenarios where sample data are scarce. Due to the in-context learning ability of ChatGPT, it can achieve significant performance improvement after learning a small number of carefully selected samples, and its performance increases linearly with the increase of the number of in-context samples. We conduct experiments to evaluate the performance of PGIM in few-shot case. For each few-shot experiment, we randomly select 3 sets of training data and train 3 times on each set to obtain the average result. The results show that after 10 prompts, ChatGPT performs better than PGIM in the fs-100 scenario on both datasets. This suggests that ChatGPT exhibits superior performance when confronted with limited training samples. Text：RT @Evode7: Actor Idris Elba became the first male to make the cover of Maxim."
},
{
"figure_ref": [],
"heading": "Effectiveness of MSEA Module",
"publication_ref": [],
"table_ref": [],
"text": "Captions：Idris Elba on the cover of maxim magazine.\nAuxiliary refined knowledge: and \"Mumbai BJP\" are all entities that were not accurately predicted by past methods. Because our auxiliary refined knowledge provides explicit explanations for such entities, PGIM makes the correct prediction."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "In this paper, we propose a two-stage framework called PGIM and bring the potential of LLMs to MNER in a novel way. Extensive experiments show that PGIM outperforms state-of-the-art methods and considerably overcomes obvious problems in previous studies. Additionally, PGIM exhibits a strong robustness and generalization capability, and only necessitates a single GPU and a reasonable number of ChatGPT invocations. In our opinion, this is an ingenious way of introducing LLMs into MNER. We hope that PGIM will serve as a solid baseline to inspire future research on MNER and ultimately solve this task better."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [],
"table_ref": [],
"text": "In this paper, PGIM enables the integration of multimodal tasks into large language model by converting images into image captions. While PGIM achieves impressive results, we consider this Text-Text paradigm as a transitional phase in the development of MNER, rather than the ultimate solution.\nBecause image captions are inherently limited in their ability to fully capture all the details of an image. This issue may potentially be further resolved in conjunction with the advancement of multimodal capabilities in language and vision models (e.g., GPT-4)."
},
{
"figure_ref": [],
"heading": "A Appendix",
"publication_ref": [],
"table_ref": [
"tab_6"
],
"text": "A.1 Generalization Analysis\nDue to the distinctive underlying logic of PGIM in incorporating auxiliary knowledge to enhance entity recognition, PGIM exhibits a stronger generalization capability that is not heavily reliant on specific datasets. Twitter-2015→2017 denotes the model is trained on Twitter-2015 and tested on Twitter-2017, vice versa. The results in Table 6 show that the generalization ability of PGIM is significantly improved compared with previous methods. This further validates the efficacy and superiority of our auxiliary refined knowledge in enhancing model performance."
},
{
"figure_ref": [
"fig_5",
"fig_5"
],
"heading": "A.2 Comparison with MoRe",
"publication_ref": [
"b2"
],
"table_ref": [],
"text": "As the previous state-of-the-art method, MoRe retrieves relevant knowledge from Wikipedia to assist entity prediction. We experimentally compare the quality of auxiliary knowledge of MoRe and PGIM. The results are shown in after summing the improvement of each indicator compared with the baseline method. All models use XLM-RoBERTa large (Conneau et al., 2019) as the text backbone with a fixed batch size of 4. The experimental results demonstrate that PGIM achieves performance improvement while requiring shorter average auxiliary knowledge length and consuming less memory. This observation highlights the lightweight nature of PGIM and further underscores the superiority of our auxiliary refined knowledge compared with auxiliary knowledge of MoRe sourced from Wikipedia. Additionally, we observe that in certain cases, the introduction of auxiliary knowledge by MoRe can even lead to a deterioration in model performance. One possible explanation for this phenomenon is that the information retrieved from Wikipedia often contains redundant or irrelevant content. The first case in Figure 4 illustrates this phenomenon well. In this case, PGIM makes the correct prediction because the information re-trieved by ChatGPT clearly states that \"Mumbai BJP refers to the Bharatiya Janata Party\". However, the information retrieved by MoRe Text from Wikipedia provides almost no assistance in recognition of named entities. MoRe alleviates this problem to some extent by introducing the Mixture of Experts (MoE) module in the post-processing stage. They fixed the parameters of MoRe Text and MoRe Image , and trained the MoE module for 50 epochs on the basis of them. But as shown in Table 1 before, compared with MoRe MoE , PGIM still shows better results without any post-processing.\nFurthermore, we also show an error prediction of PGIM in Figure 4. In this case, \"Bush\" is not a named entity that is hard to predict correctly. But since the additional knowledge retrieved by Chat-GPT clearly states that \"Bush 41\" is a name of person, the prediction of PGIM is not in line with the gold label. This illustrates that the additional knowledge retrieved from ChatGPT can affect the final prediction of named entities to some extent. But the reason why MoRe Text can make correct prediction is obviously not related to the knowledge it retrieves from the Wikipedia, because \"Bush\" is not even mentioned in its knowledge. In fact, by using only the original text after masking the noise retrieved from the Wikipedia, the model can more easily predict correctly.\nIn summary, considering the relevance and length of retrieved information, using Chat-GPT is obviously more suitable for this additional knowledge-based NER method than using Wikipedia. The information retrieved from Chat-GPT is generally unambiguous and directional, which causes it to significantly help predictions in most cases, and may also mislead predictions in rare cases. But the information retrieved from Wikipedia may mislead the original predictions in many cases."
},
{
"figure_ref": [
"fig_8"
],
"heading": "A.3 Predictions for mislabeled examples",
"publication_ref": [],
"table_ref": [],
"text": "We observe that the annotation quality of the Twitter-2015 dataset is suboptimal. There have been a large number of errors and omissions in this dataset. This is the reason why the accuracy of Twitter-2015 has significantly decreased compared with Twitter-2017. However, as shown in Figure 5, since the first stage of ChatGPT heuristically generating auxiliary knowledge is not affected by mislabeling, PGIM correctly predicts those unlabeled entities. This also demonstrates the robustness of PGIM. As a future direction, we intend to reannotate the dataset to facilitate better development of the MNER task."
},
{
"figure_ref": [],
"heading": "A.4 Prompt template",
"publication_ref": [],
"table_ref": [],
"text": "We present the template for prompting ChatGPT to generate answers. In Figure 6, PGIM guides Chat-GPT for auxiliary refined knowledge generation. In-context examples and answers in the template are selected from predefined artificial samples by the MSEA module. In Figure 7, we guide ChatGPT to make direct In-context examples are selected from the same predefined artificial samples by the MSEA module. Note that the answers here are no longer human answers, but named entities in text. Text：RT @pressjournal: Two injured following crash in Inverurie town centre. Captions：A police officer stands next to a car that has been involved in a crash. Auxiliary refined knowledge: Named entities: 1. Press Journal (news outlet) 2. Inverurie (location) Reasoning: The sentence mentions Press Journal, a news outlet that is likely reporting on the incident. Inverurie is a location, likely the town where the crash occurred. The image of a police officer standing next to a car involved in a crash supports the information in the sentence about two people being injured in a crash in the town center. Therefore, the sentence and the image are directly related. Auxiliary refined knowledge:\nNamed entities: 1. Iiro Rantala (person)2. Peter Erskine (person) 3. Johannes Weidemuller (person) 4. FinnEmbassyDC (organization/location) Reasoning: The sentence mentions Iiro Rantala, a musician who likely performed in the mentioned concerts. Peter Erskine and Johannes Weidemuller are also mentioned and are likely fellow musicians who performed with Rantala. The image of the four people …… is related to the sentence as it is likely an image of the musicians performing at the mentioned concerts.\nText：I love this quote. Robin Sharma credited for photo. My library is growing. #librarygirl Captions: Ordinary people have big tvs extraordinary people have big libraries."
},
{
"figure_ref": [],
"heading": "Auxiliary refined knowledge:",
"publication_ref": [],
"table_ref": [],
"text": "Named entities:1. Robin Sharma (person) Reasoning: The sentence mentions Robin Sharma, likely a person responsible for the photo that the quote refers to. The quote does not contain any named entities, but it does suggest that the speaker has a growing library. The image description contains a quote about libraries and the success of having a large one. It is not clear whether Robin Sharma is the person who said the quote, but he is given credit for the photo associated with the quote. Therefore, the named entity in the sentence is Robin Sharma.\nText：Still smiling, the magnificent concerts of @IiroRantala with Peter Erskine and Johannes Weidemuller @FinnDC. Reasoning: The sentence mentions South Africa, a country located in the southern region of Africa. The image of a city skyline with a bridge over it, while not directly related to the sentence, may be a representation of the country's urban areas affected by the economic downturn. Overall, the named entity in the sentence is clear and specific.\nCaptions：A city skyline with a bridge over it.\nText：Anxiety in South Africa as economy slips into technical recession. "
},
{
"figure_ref": [],
"heading": "Acknowledgments",
"publication_ref": [],
"table_ref": [],
"text": "This work was supported by the Natural Science Foundation of Tianjin (No.21JCYBJC00640) and by the 2023 CCF-Baidu Songguo Foundation (Research on Scene Text Recognition Based on Pad-dlePaddle)."
},
{
"figure_ref": [],
"heading": "Ethics Statement",
"publication_ref": [],
"table_ref": [],
"text": "In this paper, we use publicly available Twitter-2015, Twitter-2017 datasets for experiments. For the auxiliary refined knowledge, PGIM generates them using ChatGPT. Therefore, we trust that all data we use does not violate the privacy of any user. 3. BJP (political party) 4. Police (law enforcement) 5. Mantralaya (government building) Reasoning: The sentence mentions a person or handle named Sootradhar who is likely tweeting about a political protest in Mumbai. The Mumbai BJP refers to the local branch of the Bharatiya Janata Party, a major political party in India. The police are mentioned as stopping the protesters from marching towards Mantralaya, a government building in Mumbai. The hashtag #SackShinde is likely a reference to a political issue or controversy, but without further context it is unclear what this refers to. The image of a crowd of people standing around a bus does not appear to be directly related to the sentence, but may be a part of the larger context of the political protest. Prompt template for ChatGPT to make auxiliary explanation\nHere are some content that people post on Twitter, and these content are composed of original text and image descriptions of the original text. Please note that the text and image descriptions here may or may not be relevant, so make your own judgment. Please follow the data annotation style and method reflected in the example I provided, comprehensively analyze the image description and the original text, determine which named entities and their corresponding types are included in the original text, and explain the reason for your judgment. Notice : just in 'Text', not include 'Image descriptions', don't change the writing style and format of entity names, and Words after the @ sign are not counted. "
},
{
"figure_ref": [],
"heading": "Prompt template for ChatGPT to direct predict",
"publication_ref": [],
"table_ref": [],
"text": "Here are some content that people post on Twitter, and these content are composed of original text and image descriptions of the original text. Please note that the text and image descriptions here may or may not be relevant, so make your own judgment. Please follow the data annotation style and method reflected in the example I provided, comprehensively analyze the image description and the original text, determine which named entities and their corresponding types are included in the original text. There will only be 4 types of entities: "
}
] | Multimodal Named Entity Recognition (MNER) on social media aims to enhance textual entity prediction by incorporating image-based clues. Existing studies mainly focus on maximizing the utilization of pertinent image information or incorporating external knowledge from explicit knowledge bases. However, these methods either neglect the necessity of providing the model with external knowledge, or encounter issues of high redundancy in the retrieved knowledge. In this paper, we present PGIM -a two-stage framework that aims to leverage ChatGPT as an implicit knowledge base and enable it to heuristically generate auxiliary knowledge for more efficient entity prediction. Specifically, PGIM contains a Multimodal Similar Example Awareness module that selects suitable examples from a small number of predefined artificial samples. These examples are then integrated into a formatted prompt template tailored to the MNER and guide ChatGPT to generate auxiliary refined knowledge. Finally, the acquired knowledge is integrated with the original text and fed into a downstream model for further processing. Extensive experiments show that PGIM outperforms state-of-the-art methods on two classic MNER datasets and exhibits a stronger robustness and generalization capability. | Prompting ChatGPT in MNER: Enhanced Multimodal Named Entity Recognition with Auxiliary Refined Knowledge | [
{
"figure_caption": "HeuristicsFigure1: The \"CP3\" in the text is a class of entities that are difficult to predict successfully by existing studies. PGIM demonstrates successful prediction of such entities with an approach more similar to human cognitive processes by endowing ChatGPT with reasonable heuristics.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: The architecture of PGIM.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "j=1 where t i , p i , y i refer to the text, image, and gold labels. The vanilla MNER model M trained on D mainly consists of a backbone encoder M b and a CRF decoder M c . The input multimodal imagetext pair is encoded by the encoder M b to obtain multimodal fusion features H: H = M b (t, p)",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Text：@NDCS_UK Pls RT a poster about the next Leadership Course for Young Deaf People on 11th/12th July 2015 thanks. Captions：dearplus -leadership training -july 2015 Auxiliary refined knowledge: Named entities: 1. NDCS (organization) 2. UK (location) 3. Leadership Course (event/training) 4. Young Deaf People (target audience) Reasoning: The sentence mentions NDCS, which is the National Deaf Children's Society in the UK. The event is the Leadership Course, which is a training for young deaf people. The image of a poster advertising the Leadership Course is directly related to the sentence, as it confirms the details of the event mentioned in the tweet. . With Zakwan Chaudhary in Holland.... Captions：A man holding a child on a busy street. Auxiliary refined knowledge: Named entities:1. Big B (person/celebrity) 2. Zakwan Chaudhary (person) 3. Holland (location/country) Reasoning: The sentence mentions Big B, a well-known celebrity. It also mentions Zakwan Chaudhary, who is likely a friend or associate of Big B. The sentence indicates that they are in Holland, which is a location/country. The image shows a man holding a child on a busy street, but it is unclear if this is related to the sentence or the named entities……",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3: Four case studies of how auxiliary refined knowledge can help model predictions.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: Two case studies on how information retrieved from Wikipedia by MoRe and information retrieved by PGIM from ChatGPT affects model predictions.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Captions：Four people standing next to a piano in a room.",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 :5Figure 5: Some mislabeled examples of Twitter-2015 datasets.",
"figure_data": "",
"figure_id": "fig_8",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Results of methods with † are retrieved from the corresponding original paper. And results with ⋄ come fromWang et al. (2022c).",
"figure_data": "Twitter-2015Twitter-2017MethodsSingle Type(F1)OverallSingle Type(F1)OverallPER LOC ORG OTH. Pre. Rec.F1PER LOC ORG OTH. Pre. Rec.F1TextBiLSTM-CRF †76.77 72.56 41.33 26.80 68.14 61.09 64.42 85.12 72.68 72.50 52.56 79.42 73.43 76.31BERT-CRF ‡85.37 81.82 63.26 44.13 75.56 73.88 74.71 90.66 84.89 83.71 66.86 86.10 83.85 84.96BERT-SPAN ‡85.35 81.88 62.06 43.23 75.52 73.83 74.76 90.84 85.55 81.99 69.77 85.68 84.60 85.14RoBERTa-SPAN ‡ 87.20 83.58 66.33 50.66 77.48 77.43 77.45 94.27 86.23 87.22 74.94 88.71 89.44 89.06Text+ImageUMT85.24 81.58 63.03 39.45 71.67 75.23 73.41 91.56 84.73 82.24 70.10 85.28 85.34 85.31UMGF84.26 83.17 62.45 42.42 74.49 75.21 74.85 91.92 85.22 83.13 69.83 86.54 84.50 85.51MNER-QG85.68 81.42 63.62 41.53 77.76 72.31 74.94 93.17 86.02 84.64 71.83 88.57 85.96 87.25R-GCN86.36 82.08 60.78 41.56 73.95 76.18 75.00 92.86 86.10 84.05 72.38 86.72 87.53 87.11ITA------78.03------89.75PromptMNER----78.03 79.17 78.60----89.93 90.60 90.27CAT-MNER88.04 84.70 68.04 52.33 78.75 78.69 78.72 94.61 88.40 88.14 80.50 90.27 90.67 90.47MoReText------77.79------89.49MoReImage------77.57------90.28MoReMoE------79.21------90.67PGIM(Ours)88.34 84.22 70.15 52.34 79.21 79.45 79.33* 96.46 89.89 89.03 79.62 90.86 92.01 91.43*±0.02 ±0.12 ±0.36 ±0.98 ±0.63 ±0.22 ±0.06 ±0.02 ±0.68 ±0.53 ±2.25 ±0.16 ±0.07 ±0.09Twitter-2015Twitter-2017experiment, especially on the Twitter-2017 dataset.Pre. Rec.F1Pre. Rec.F1BaselineBERT ⋄ 75.56 73.88 74.71 86.10 83.85 84.96UMT †71.67 75.23 73.41 85.28 85.34 85.31UMGF †74.49 75.21 74.85 86.54 84.50 85.51R-GCN †73.95 76.18 75.00 86.72 87.53 87.11ITABERT †--75.60--85.72CATBERT †76.19 74.65 75.41 87.04 84.97 85.99MoReImage BERT‡ 73.16 74.64 73.89 85.49 86.38 85.94MoReText BERT‡73.31 74.43 73.86 85.92 86.75 86.34PGIMBERT75.84 77.76 76.79* 89.09 90.08 89.58*±0.30 ±0.22 ±0.19 ±0.24 ±0.08 ±0.100.05 when comparing with ITA, CAT-MNER andMoRe Image/Text . And ‡ represents the results afterwe replace the text encoder in the MoRe officialcode with BERT base . 3 The experimental resultsshow that PGIM achieves a greater performanceimprovement than the XLM-RoBERTa large version",
"figure_id": "tab_0",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": "presents the performance comparison be-tween ChatGPT and PGIM in the few-shot scenario.VanillaGPT stands for no prompting, and Prompt-GPT denotes the selection of top-N similar samplesfor in-context learning. As shown in Appendix A.4,",
"figure_id": "tab_1",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Comparison of ChatGPT and PGIM in fewshot case. VanillaGPT and PromptGPT stand for direct prediction using ChatGPT.",
"figure_data": "",
"figure_id": "tab_2",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Table 4 demonstrates the effectiveness of the MSEA module. We use the auxiliary refined knowledge generated by ChatGPT after N in-context prompts to construct the datasets and train the model. The text encoder of the Baseline model is XLM-RoBERTa large , and",
"figure_data": "Twitter-2015Twitter-2017Pre. Rec. F1Pre. Rec. F1Baseline76.45 78.22 77.32 88.46 90.23 89.34w/o MSEAN=1 78.15 79.01 78.58 90.49 90.82 90.65w/o MSEAN=5 78.11 79.82 78.95 90.62 91.49 91.05w/o MSEAN=10 78.47 79.21 78.84 90.54 91.77 91.15PGIMN=178.40 79.21 78.76 89.90 91.63 90.76PGIMN=579.21 79.45 79.33 90.86 92.01 91.43PGIMN=1078.58 79.67 79.12 90.54 92.08 91.30Table 4: Effect of the number of in-context examples onauxiliary refined knowledge.its input is the original text that does not containany auxiliary knowledge. w/o MSEA represents arandom choice of in-context examples. All resultsare averages of training results after three randominitializations. Obviously, the addition of auxil-iary refined knowledge can improve the effect ofthe model. And the addition of MSEA modulecan further improve the quality of the auxiliaryknowledge generated by ChatGPT, which reflectsthe effectiveness of the MSEA module. An appro-priate number of in-context examples can furtherimprove the quality of auxiliary refined knowledge.But the number of examples is not the more the bet-ter. When ChatGPT is provided with an excessivenumber of examples, the quality of the auxiliaryknowledge may deteriorate. One possible explana-tion for this phenomenon is that too many artificialexamples introduce noise into the generation pro-cess of ChatGPT. As a pre-trained large languagemodel, ChatGPT lacks genuine comprehension ofthe underlying logical implications in the examples.Consequently, an excessive number of examplesmay disrupt its original reasoning process.Case Study Through some case studies in Fig-ure 3, we show how auxiliary refined knowledgecan help improve the predictive performance ofthe model. The Baseline model represents thatno auxiliary knowledge is introduced. MoRe Textand MoRe Image denote the relevant knowledge ofthe input text and image retrieved using text re-triever and image retriever, respectively. In PGIM,the auxiliary refined knowledge generated by Chat-GPT is structured into two components: the firstcomponent provides a preliminary estimation ofthe named entity, and the second component offersa corresponding contextual explanation. In theseexamples, \"Leadership Course\", \"Big B\", \"Maxim\"",
"figure_id": "tab_3",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "The baseline method solely relies on the original text without any incorporation of auxiliary information. MoRe Text and MoRe Image denote the relevant knowledge of the input text and image retrieved using text retriever and image retriever, respectively. Ave.length represents the average length of the auxiliary knowledge in entire dataset. Memory indicates the GPU memory size required for training model. Ave.Improve represents the average result",
"figure_data": "MethodsSingle Type(F1) PER LOC ORG OTH. Pre.Overall Rec.F1Ave. length Memory(MB) Ave. ImproveTwitter-2015BaseLine87.04 83.49 67.34 50.16 76.45 78.22 77.32-11865-MoReText86.92 83.08 68.20 49.15 77.12 77.77 77.45227.4116759↓ 0.05MoReImage87.38 83.78 67.75 49.38 77.44 78.06 77.75203.0016711↑ 0.28PGIM(Ours) 88.34 84.22 70.15 52.34 79.21 79.45 79.33104.5613901↑ 1.86Twitter-2017BaseLine95.07 87.22 85.82 78.66 88.46 90.23 89.34-11801-MoReText95.16 88.77 87.00 77.71 89.33 90.45 89.89241.4716695↑ 0.50MoReImage94.43 87.43 86.22 74.77 88.06 89.49 88.77192.0016447↓ 0.80PGIM(Ours) 96.46 89.89 89.03 79.62 90.86 92.01 91.4394.5213279↑ 2.07",
"figure_id": "tab_4",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "Comparison of MoRe with PGIM. Since the original paper of MoRe(Wang et al., 2022a) did not report its Single Type (F1) on the Twitter-2015 and Twitter-2017 datasets, we run its released code and count the results. All of the results are averaged from 3 runs with different random seeds.",
"figure_data": "Twitter-2015→2017 Twitter-2017→2015Pre. Rec.F1Pre. Rec.F1UMT †67.80 55.23 60.87 64.67 63.59 64.13UMGF †69.88 56.92 62.74 67.00 62.81 66.21CAT-MNER ‡ 70.69 59.44 64.58 74.86 63.01 68.43PGIM72.66 65.51 68.90 76.13 64.87 70.05",
"figure_id": "tab_5",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "Comparison of the generalization ability. For the baseline model, results with † come from Zhang et al. (2021), and results with ‡ come from Wang et al.",
"figure_data": "",
"figure_id": "tab_6",
"figure_label": "6",
"figure_type": "table"
}
] | Jinyuan Li; Han Li; Zhuo Pan; Di Sun; Jiahao Wang; Wenkun Zhang; Gang Pan | [
{
"authors": "Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell",
"journal": "Advances in neural information processing systems",
"ref_id": "b0",
"title": "Language models are few-shot learners",
"year": "2020"
},
{
"authors": "Jason Pc Chiu; Eric Nichols",
"journal": "Transactions of the association for computational linguistics",
"ref_id": "b1",
"title": "Named entity recognition with bidirectional lstm-cnns",
"year": "2016"
},
{
"authors": "Alexis Conneau; Kartikay Khandelwal; Naman Goyal; Vishrav Chaudhary; Guillaume Wenzek; Francisco Guzmán; Edouard Grave; Myle Ott; Luke Zettlemoyer; Veselin Stoyanov",
"journal": "",
"ref_id": "b2",
"title": "Unsupervised cross-lingual representation learning at scale",
"year": "2019"
},
{
"authors": "Jia Deng; Wei Dong; Richard Socher; Li-Jia Li; Kai Li; Li Fei-Fei",
"journal": "",
"ref_id": "b3",
"title": "Imagenet: A large-scale hierarchical image database",
"year": "2009"
},
{
"authors": "Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova",
"journal": "",
"ref_id": "b4",
"title": "Bert: Pre-training of deep bidirectional transformers for language understanding",
"year": "2018"
},
{
"authors": "Zhiheng Huang; Wei Xu; Kai Yu",
"journal": "",
"ref_id": "b5",
"title": "Bidirectional lstm-crf models for sequence tagging",
"year": "2015"
},
{
"authors": "Meihuizi Jia; Lei Shen; Xin Shen; Lejian Liao; Meng Chen; Xiaodong He; Zhendong Chen; Jiaqi Li",
"journal": "",
"ref_id": "b6",
"title": "Mner-qg: An end-to-end mrc framework for multimodal named entity recognition with query grounding",
"year": "2022"
},
{
"authors": "Jacob Devlin; Ming-Wei Chang; Kenton ; Lee Kristina; Toutanova ",
"journal": "",
"ref_id": "b7",
"title": "Bert: Pre-training of deep bidirectional transformers for language understanding",
"year": "2019"
},
{
"authors": "John Lafferty; Andrew Mccallum; Fernando Cn Pereira",
"journal": "",
"ref_id": "b8",
"title": "Conditional random fields: Probabilistic models for segmenting and labeling sequence data",
"year": "2001"
},
{
"authors": "Dong-Ho Lee; Akshen Kadakia; Kangmin Tan; Mahak Agarwal; Xinyu Feng; Takashi Shibuya; Ryosuke Mitani; Toshiyuki Sekiya; Jay Pujara; Xiang Ren",
"journal": "",
"ref_id": "b9",
"title": "Good examples make a faster learner: Simple demonstration-based learning for low-resource ner",
"year": "2021"
},
{
"authors": "Junnan Li; Dongxu Li; Silvio Savarese; Steven Hoi",
"journal": "",
"ref_id": "b10",
"title": "Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models",
"year": "2023"
},
{
"authors": "Tsung-Yi Lin; Michael Maire; Serge Belongie; James Hays; Pietro Perona; Deva Ramanan; Piotr Dollár; C Lawrence; Zitnick ",
"journal": "Springer",
"ref_id": "b11",
"title": "Microsoft coco: Common objects in context",
"year": "2014-09-06"
},
{
"authors": "Jiachang Liu; Dinghan Shen; Yizhe Zhang; Bill Dolan; Lawrence Carin; Weizhu Chen",
"journal": "",
"ref_id": "b12",
"title": "What makes good in-context examples for gpt-3?",
"year": "2021"
},
{
"authors": "Ilya Loshchilov; Frank Hutter",
"journal": "",
"ref_id": "b13",
"title": "Decoupled weight decay regularization",
"year": "2017"
},
{
"authors": "Di Lu; Leonardo Neves; Vitor Carvalho; Ning Zhang; Heng Ji",
"journal": "",
"ref_id": "b14",
"title": "Visual attention model for name tagging in multimodal social media",
"year": "2018"
},
{
"authors": "Seungwhan Moon; Leonardo Neves; Vitor Carvalho",
"journal": "",
"ref_id": "b15",
"title": "Multimodal named entity recognition for short social media posts",
"year": "2018"
},
{
"authors": "Yasmin Moslem; Rejwanul Haque; Andy Way",
"journal": "",
"ref_id": "b16",
"title": "Adaptive machine translation with large language models",
"year": "2023"
},
{
"authors": "Chengwei Qin; Aston Zhang; Zhuosheng Zhang; Jiaao Chen; Michihiro Yasunaga; Diyi Yang",
"journal": "",
"ref_id": "b17",
"title": "Is chatgpt a general-purpose natural language processing task solver?",
"year": "2023"
},
{
"authors": "Ohad Rubin; Jonathan Herzig; Jonathan Berant",
"journal": "",
"ref_id": "b18",
"title": "Learning to retrieve prompts for in-context learning",
"year": "2021"
},
{
"authors": "F Erik; Jorn Sang; Veenstra",
"journal": "",
"ref_id": "b19",
"title": "Representing text chunks",
"year": "1999"
},
{
"authors": "Zhenwei Shao; Zhou Yu; Meng Wang; Jun Yu",
"journal": "",
"ref_id": "b20",
"title": "Prompting large language models with answer heuristics for knowledge-based visual question answering",
"year": "2023"
},
{
"authors": "Lin Sun; Jiquan Wang; Kai Zhang; Yindu Su; Fangsheng Weng",
"journal": "",
"ref_id": "b21",
"title": "Rpbert: a text-image relation propagation-based bert model for multimodal ner",
"year": "2021"
},
{
"authors": "Hugo Touvron; Thibaut Lavril; Gautier Izacard; Xavier Martinet; Marie-Anne Lachaux; Timothée Lacroix; Baptiste Rozière; Naman Goyal; Eric Hambro; Faisal Azhar",
"journal": "",
"ref_id": "b22",
"title": "Llama: Open and efficient foundation language models",
"year": "2023"
},
{
"authors": "David Vilar; Markus Freitag; Colin Cherry; Jiaming Luo; Viresh Ratnakar; George Foster",
"journal": "",
"ref_id": "b23",
"title": "Prompting palm for translation: Assessing strategies and performance",
"year": "2022"
},
{
"authors": "Shuhe Wang; Xiaofei Sun; Xiaoya Li; Rongbin Ouyang; Fei Wu; Tianwei Zhang; Jiwei Li; Guoyin Wang",
"journal": "",
"ref_id": "b24",
"title": "Gpt-ner: Named entity recognition via large language models",
"year": "2023"
},
{
"authors": "Xinyi Wang; Wanrong Zhu; William Yang; Wang ",
"journal": "",
"ref_id": "b25",
"title": "Large language models are implicitly topic models: Explaining and finding good demonstrations for in-context learning",
"year": "2023"
},
{
"authors": "Xinyu Wang; Jiong Cai; Yong Jiang; Pengjun Xie; Kewei Tu; Wei Lu",
"journal": "",
"ref_id": "b26",
"title": "Named entity and relation extraction with multi-modal retrieval",
"year": "2022"
},
{
"authors": "Xinyu Wang; Min Gui; Yong Jiang; Zixia Jia; Nguyen Bach; Tao Wang; Zhongqiang Huang; Fei Huang; Kewei Tu",
"journal": "",
"ref_id": "b27",
"title": "Ita: Image-text alignments for multi-modal named entity recognition",
"year": "2021"
},
{
"authors": "Xinyu Wang; Yong Jiang; Nguyen Bach; Tao Wang; Zhongqiang Huang; Fei Huang; Kewei Tu",
"journal": "",
"ref_id": "b28",
"title": "Improving named entity recognition by external context retrieving and cooperative learning",
"year": "2021"
},
{
"authors": "Xuwu Wang; Junfeng Tian; Min Gui; Zhixu Li; Jiabo Ye; Ming Yan; Yanghua Xiao",
"journal": "Springer",
"ref_id": "b29",
"title": "Promptmner: Prompt-based entity-related visual clue extraction and integration for multimodal named entity recognition",
"year": "2022-04-11"
},
{
"authors": "Xuwu Wang; Jiabo Ye; Zhixu Li; Junfeng Tian; Yong Jiang; Ming Yan; Ji Zhang; Yanghua Xiao",
"journal": "IEEE",
"ref_id": "b30",
"title": "Cat-mner: Multimodal named entity recognition with knowledge-refined cross-modal attention",
"year": "2022"
},
{
"authors": "Xiang Wei; Xingyu Cui; Ning Cheng; Xiaobin Wang; Xin Zhang; Shen Huang; Pengjun Xie; Jinan Xu; Yufeng Chen; Meishan Zhang",
"journal": "",
"ref_id": "b31",
"title": "Zeroshot information extraction via chatting with chatgpt",
"year": "2023"
},
{
"authors": "Ikuya Yamada; Akari Asai; Hiroyuki Shindo; Hideaki Takeda; Yuji Matsumoto",
"journal": "",
"ref_id": "b32",
"title": "Luke: Deep contextualized entity representations with entity-aware self-attention",
"year": "2020"
},
{
"authors": "Zhengyuan Yang; Zhe Gan; Jianfeng Wang; Xiaowei Hu; Yumao Lu; Zicheng Liu; Lijuan Wang",
"journal": "",
"ref_id": "b33",
"title": "An empirical study of gpt-3 for few-shot knowledgebased vqa",
"year": "2022"
},
{
"authors": "Jianfei Yu; Jing Jiang; Li Yang; Rui Xia",
"journal": "Association for Computational Linguistics",
"ref_id": "b34",
"title": "Improving multimodal named entity recognition via entity span detection with unified multimodal transformer",
"year": "2020"
},
{
"authors": "Dong Zhang; Suzhong Wei; Shoushan Li; Hanqian Wu; Qiaoming Zhu; Guodong Zhou",
"journal": "",
"ref_id": "b35",
"title": "Multimodal graph fusion for named entity recognition with targeted visual guidance",
"year": "2021"
},
{
"authors": "Qi Zhang; Jinlan Fu; Xiaoyu Liu; Xuanjing Huang",
"journal": "",
"ref_id": "b36",
"title": "Adaptive co-attention network for named entity recognition in tweets",
"year": "2018"
},
{
"authors": "Xin Zhang; Yong Jiang; Xiaobin Wang; Xuming Hu; Yueheng Sun; Pengjun Xie; Meishan Zhang",
"journal": "",
"ref_id": "b37",
"title": "Domain-specific ner via retrieving correlated samples",
"year": "2022"
},
{
"authors": "Fei Zhao; Chunhui Li; Zhen Wu; Shangyu Xing; Xinyu Dai",
"journal": "",
"ref_id": "b38",
"title": "Learning from different text-image pairs: A relation-enhanced graph convolutional network for multimodal ner",
"year": "2022"
},
{
"authors": "Changmeng Zheng; Zhiwei Wu; Tao Wang; Yi Cai; Qing Li",
"journal": "IEEE Transactions on Multimedia",
"ref_id": "b39",
"title": "Object-aware multimodal named entity recognition in social media posts with adversarial learning",
"year": "2020"
},
{
"authors": "Baohang Zhou; Ying Zhang; Kehui Song; Wenya Guo; Guoqing Zhao; Hongbin Wang; Xiaojie Yuan",
"journal": "",
"ref_id": "b40",
"title": "A span-based multimodal variational autoencoder for semi-supervised multimodal named entity recognition",
"year": "2022"
}
] | [
{
"formula_coordinates": [
3,
306.14,
547.12,
218.27,
23.36
],
"formula_id": "formula_0",
"formula_text": "C = {c 1 , • • • , c n } contains n in-context examples."
},
{
"formula_coordinates": [
3,
351.39,
619.52,
127.78,
22.09
],
"formula_id": "formula_1",
"formula_text": "y l = argmax y l p LLM (y l |p, y <l )"
},
{
"formula_coordinates": [
5,
133.05,
398.5,
92.19,
34.35
],
"formula_id": "formula_2",
"formula_text": "D = {(t i , p i , y i )} M i=1 G = {(t j , p j , y j )} N"
},
{
"formula_coordinates": [
5,
116.08,
74.37,
410.24,
698.1
],
"formula_id": "formula_3",
"formula_text": "I = argTopN j∈{1,2,...,N } H T H j ∥H∥ 2 ∥H j ∥ 2 I is the index set of top-N similar samples in G."
},
{
"formula_coordinates": [
5,
359.47,
110.1,
111.62,
10.63
],
"formula_id": "formula_4",
"formula_text": "C = {(t j , p j , y j ) | j ∈ I}"
},
{
"formula_coordinates": [
5,
306.14,
452.86,
218.27,
37.73
],
"formula_id": "formula_5",
"formula_text": "Z = {z 1 , • • • , z m }, where m is the length of Z. PGIM concatenates the original text T = {t 1 , • • • , t n }"
},
{
"formula_coordinates": [
5,
319.72,
542.78,
191.11,
10.63
],
"formula_id": "formula_6",
"formula_text": "{h 1 , • • • , h n , • • • , h n+m } = embed([T ; Z])"
},
{
"formula_coordinates": [
5,
333.03,
679.05,
163.29,
56.34
],
"formula_id": "formula_7",
"formula_text": "P (y|T, Z) = n i=1 ψ(y i-1 , y i , h i ) y ′ ∈Y n i=1 ψ(y ′ i-1 , y ′ i , h i )"
},
{
"formula_coordinates": [
6,
70.87,
107.93,
175.24,
60.17
],
"formula_id": "formula_8",
"formula_text": "L NLL (θ) = -log P θ (y * |T, Z) 4 Experiments 4.1 Settings"
}
] |
|
10.18653/v1/2022.acl-long.439 | 2023-05-20 | [
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b15",
"b16",
"b21",
"b36",
"b10",
"b14",
"b2",
"b39",
"b17",
"b37",
"b8",
"b0",
"b26",
"b10",
"b2",
"b6",
"b14"
],
"table_ref": [],
"text": "Named Entity Recognition (NER) is a fundamental NLP task to extract entities from unstructured text. In traditional fully supervised NER scenarios, deep neural architectures (Huang et al., 2015;Lample et al., 2016;Ma and Hovy, 2016;Yan et al., 2019) have shown great ability to recognize entities with sufficient human-annotated data. However, acquiring such human-annotated data can be expensive and time-consuming since the demand for domainspecific knowledge. Previous NER models usually struggle to leverage very limited labeled data to recognize entities in practical scenarios owing to these data-hungry characteristics. Furthermore, the classifier head of a traditional NER system needs to be retrained from scratch when the number or type of entity class changes. Therefore, few-shot NER has drawn much attention in the information extraction field.\nOwing to only a few labeled examples (usually called support examples) available, Fritzler et al. (2019) and Hou et al. (2020) propose to compute token-level similarities between the label prototypes or each token of support sets and each token of query sets. Based on previous works, Das et al. (2022) propose CONTaiNER, the first method using contrastive learning to enhance the token representation of PTMs for few-shot NER. Ma et al. (2022a) propose an architecture consisting of two pre-trained encoders to encode the sentence and label words, proving effective for low-resource NER.\nRecently, span-based NER (Yu et al., 2020;Li et al., 2020;Yan et al., 2022) has demonstrated exemplary performance in various NER tasks. Ma et al. (2022b) decomposes the few-shot NER task into two distinct stages, i.e. span-detection and entity-typing. They also use MAML (Finn et al., 2017), a meta-learning algorithm, to enhance the performance of their model. Wang et al. (2022b) converts the NER task into a span-matching problem and propose a novel span refining module which applies the Soft-NMS (Bodla et al., 2017;Shen et al., 2021) algorithm during beam search. These span-based prototypical networks achieve significant improvements over token-level few-shot NER baselines, which avoid the token-level label dependency problem.\nDespite the promising performance of spanbased prototypical networks. Two problems limit these methods. 1) The span-level metric learning of the prototypical network is based on support sets and query sets, where samples from support sets are used to construct the label prototypes, and query samples are used to compute the span-level similarities and optimize these label prototypes. However, only the label of support samples is available in the test scenario. Previous prototypical networks (Fritzler et al., 2019;Wang et al., 2022b,a) usually do not update any parameter of their models on the novel support set, which limits the transfer learning capability of these methods. 2) Previous span detectors usually extract some false positive spans. In few-shot NER, unseen new classes in the test set are usually tagged as O-type (Das et al., 2022) during training. Unfortunately, previous span-based models are class-agnostic in the span-detection stage. It is challenging to detect unseen new class span only during the span detection stage since these models have been thoroughly trained in the source domain to regard the unseen new class entities as O-type. To address this problem, Ma et al. (2022b) and Wang et al. (2022a) filter some false positive spans, which are too far from label prototypes. Wang et al. (2022b) introduce an O-type prototype to match false positive spans in the query set. However, owing to the limited support examples, label prototypes constructed by support samples may not precisely represent the class distribution in the feature space. This paper proposes PromptNER: a simple but effective prompting method for few-shot NER. First, we construct a natural language prompt to instruct Pre-trained Language Models (PLMs) to extract entities with specific classes. Then we design a position-aware biaffine module for recalling candidate spans and a prompt-based classifier for entity typing. Inspired by Wang et al. (2022c), we introduce k nearest neighbor search to leverage the ground truth entity representations from support examples. The difference between typical prototypical networks and our method is shown in Figure 1. Unlike previous prototypical networks, the optimization process of our model is not limited to the support set and query set format. Like traditional NER, our model only requires sentences and their corresponding label sentences for training. Therefore, we can fine-tune our model on a novel support set without gaps between the training and fine-tuning stages. We alse propose a novel rerank strategy to filter false positive spans, We evaluate PromptNER on multiple benchmark datasets, including Few-NERD (Ding et al., 2021) and Cross-NER (Hou et al., 2020). The experimental results demonstrate that PromptNER achieves superior performance over state-of-the-art few-shot NER methods and the effectiveness of the rerank strategy and fine-tuning stage."
},
{
"figure_ref": [
"fig_1"
],
"heading": "Problem Formulation",
"publication_ref": [
"b6"
],
"table_ref": [],
"text": "In this part, we formally introduce the problem formulation of few-shot named entity recognition (NER).\nSimilarly to the supervised NER system, the input of the few-shot NER system is a natural language sentence X which contains n words. And the output Y = {y i } n i=1 is a label sentence, where y i ∈ T , T is the entity type set with O-type (Outside). Following Ding et al. (2021), we adapt the standard N-way K-shot setting to train and evaluate the few-shot NER system. During training, each episode data ε train = {S train , Q train , T train } contains a support set S train , a query set Q train and entity type set T train . A support or query set contains N classes (N -way) and K examples (Kshot) for each entity class respectively, where S train ∩Q train = ∅. For testing, we utilize a novel episode ε test = {S test , Q test , T test } to evaluate the few-shot NER system, where T train ∩ T test = O.A typical 2-way 2-shot episode is shown in Figure 2."
},
{
"figure_ref": [
"fig_2"
],
"heading": "Proposed Method",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we formally present our proposed PromptNER. The architecture of PromptNER is shown in the Figure 3."
},
{
"figure_ref": [],
"heading": "Span Representations",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "BERT",
"publication_ref": [],
"table_ref": [],
"text": "Find some entities, such as none, person, company: Steve Jobs founded Apple in 1976."
},
{
"figure_ref": [],
"heading": "BERT Biaffine Decoder",
"publication_ref": [],
"table_ref": [],
"text": "Find some entities, such as none, person, company: The CEO of Apple is Tim Cook."
},
{
"figure_ref": [],
"heading": "Final Prediction Prob.",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Steve Jobs",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Logits Rerank",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Candidate Spans",
"publication_ref": [],
"table_ref": [],
"text": "Prob. "
},
{
"figure_ref": [],
"heading": "Encode support sentences",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Query set",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Rerank",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Input Construction",
"publication_ref": [],
"table_ref": [],
"text": "Formally, the input of a NER system is a natural language sentence. Given a sentence consisting of n words X = [x 1 , x 2 , ..., x n ] and an entity type set T = {none, t 1 , t 2 , ..., t m-1 } where m = |T | and none means O-type, we use the pre-defined prompt template to reconstruct the input sentence as follows:\nX p = F prompt (T ) ⊕ X, = [X l , X m , X](1)\nwhere F prompt (T ) is a function which fills the template using the entity type set T . For example, suppose the entity type set is {none, person, company}, and the input sentence is \"Steve Jobs founded Apple in 1976.\". The reconstructed input, using the template \"Find some entities, such as none, t 1 , t 2 , ..., t m-1 : \" will be \"Find some entities, such as none, person, company: Steve Jobs founded Apple in 1976.\". Additionally, the input X p could also be split into three parts shown in (1), where X l = \"Find some entities, such as\", X m = \"none, person, company\". 1 The reconstructed input X p provides label information to the model and instructs the model to extract some entities mentioned in the prompt."
},
{
"figure_ref": [],
"heading": "Position-aware Biaffine Module",
"publication_ref": [
"b39",
"b37",
"b4",
"b18",
"b29",
"b29"
],
"table_ref": [],
"text": "We follow Yu et al. (2020) and Yan et al. (2022) to convert the span-detection task into a binary classification task. For a sentence with n tokens, we need to perform binary classification task n(n + 1)/2 times. To this end, our method first uses a pre-trained encoder to encode the prompt and the input sentence:\nH = Encoder(X p ), = [H l , H m , H n ](2)\nwhere H ∈ R (l+m+n)×d , and d is the embedding size. The encoder is typically a pre-trained language model, such as BERT (Devlin et al., 2019), RoBERTa (Liu et al., 2019). Because several words may be tokenized into some subwords, we use mean-pooling to obtain the representation of each word. Meanwhile, H l will be ignored, and the label word embedding H m is utilized in the promptbased classifier, which will be illustrated in Section 3.3.\nThen we design a biaffine model which incorporates absolute and relative position information. Inspired by Su et al. (2022), we apply RoPE into the span detection stage to inject absolute and relative position information, which satisfies the constraint R i R j = R j-i . For a span that ranges from i-th word to j-th word, we can calculate the prediction logit as follows:\nh s = LeakyReLU(h i W s ), h e = LeakyReLU(h j W e ), R i,j = h s Uh e + (R i h i W p ) (R j h j W p ), = h s Uh e + (h i W p ) R j-i (h j W p ) (3) where W s , W e , W p ∈ R d×h , U, R i , R j ∈ R h×h ,\nand h is the hidden size. For a sentence with n words, we can get a score matrix R ∈ R n×n . We mask the lower triangle part of R (where i > j), to filter all the impossible spans which contain words from i-th to j-th. To address the issue of sample imbalance, we use the span-based class imbalance loss proposed in Su et al. (2022): r(i,j) , ,j) ,\nL pos = log 1 + (i,j)∈Spos e -\nL neg = log 1 + (i,j)∈Sneg e r(i\nL span = L pos + L neg , where 1 ≤ i ≤ j ≤ n, S pos = {(s k , e k )} N k=1\nrepresents the collection of candidates spans(noun phrase), and N is the entity span number of the input sentence. S neg represents the collection of spans which not belong to noun phrases accordingly.\nDuring inference, we extract with the top-3k logits from the upper triangle part of score matrix R to recall more candidate spans, where k corresponds to the k-shot setting."
},
{
"figure_ref": [],
"heading": "Prompt-based Classifier",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we propose a novel approach to classify each candidate span. Unlike the technique presented by Ma et al. (2022a), our method incorporates the semantic information of the input sentence into the label embedding. Moreover, we introduce an additional embedding type for the \"none\" category, which assists in identifying and filtering out some false positive spans."
},
{
"figure_ref": [],
"heading": "Classification with Prompt",
"publication_ref": [],
"table_ref": [],
"text": "For each example (X, Y, T ) in D train , we utilize H m , H n computed in (2) to compute the classification probability of each entity span in S pos . Specifically, for the i-th span (s i , e i ) in S pos , we can obtain its representation as follows:\nu i = 1 e i -s i + 1 e i k=s i h k ,\nwhere h k ∈ H n , s i , e i denote the starting and ending indices for the i-th span, respectively.\nThe probability distribution can be calculated as follows:\np(y|s i , e i ) = Softmax( H m u i √ d ),\nwhere H m ∈ R m×d , m is the class number and d is the embedding size. Therefore, the loss function for the prompt-based classifier of each sentence can be expressed as:\nL class = 1 |S pos | |Spos| i=1 -log p(y|s i , e i ) ,"
},
{
"figure_ref": [],
"heading": "Training and Fine-tuning",
"publication_ref": [
"b2"
],
"table_ref": [],
"text": "During the training stage, we sample an episode data from D train which consists of a support set Ŝtrain and a query set Qtrain . Unlike previous methods (Das et al., 2022;Wang et al., 2022b;Ma et al., 2022b;Wang et al., 2022a), in the training process of PromptNER, we decompose the Ŝtrain and Qtrain , where the optimized object can be calculated in Ŝtrain and Qtrain , respectively:\nL = L span + L class (4)\nDuring the testing stage, where only label sentences from Ŝtest available, we just use the Ŝtest to optimize our model like (4)."
},
{
"figure_ref": [],
"heading": "Inference via kNN Search",
"publication_ref": [],
"table_ref": [],
"text": "As described in section 3.2, we denote the collection of candidate spans as C = {(s i , e i ) 3k i=1 }. The candidate span embedding is U query ∈ R 3k×d , while the prompt label embedding is U label ∈ R t×d . Hence, we can compute the probability distribution that the i-th span belongs to each class as follows:\np prompt (y|s i , e i ) = Softmax( U label u i √ d ),\nwhere u i ∈ R 1×d , and d is the embedding size. During this inference stage, we filter all the false positive spans which satisfy none = arg max p(y|s i , e i ).\nTo leverage the golden entity representations of the support set Ŝtest , we introduce the k nearest neighbor search algorithm during the inference stage. First, we merge all the golden entity embedding into a matrix U golden ∈ R n×d , and n is the golden number in the support set Ŝtest and d is the embedding size. The similarity score between a candidate span and golden entities is:\nd i = U golden u i √ d ,\nwhere d i ∈ R n×1 , and d is the embedding size. Inspired by Wang et al. (2022c), we just retrieve a golden entity set N i with top-k similarity scores.\np(y i = t|s i , e i ) ∝ n j=1 I(j ∈ N i , y j = t) • d i (j),\nwhere I is the indicator function. The probability of the label not being retrieved by the k-NN search always is assigned as zero.\nThe final prediction probability is calculated as follows:\np(y|s i , e i ) = γ • Sigmoid(R(s i , e i )) + α • p prompt (y|s i , e i ) + β • p knn (y|s i , e i )(5)\nwhere γ, α, β are hyper-parameters which balance these three different distributions. The reason why we use R(s i , e i ) to rerank is to filter some false positive spans extracted from the position-aware biaffine module.\nThe final prediction label of span(s i , e i ) is:\ny pred = argmax p(y = t|s i , e i ),\n4 Experiment Setup"
},
{
"figure_ref": [],
"heading": "Datasets",
"publication_ref": [],
"table_ref": [],
"text": "To demonstrate the few-shot learning ability of our method, we conduct experiments on two welldesigned N -way K-shot few-shot NER datasets. "
},
{
"figure_ref": [],
"heading": "Baselines",
"publication_ref": [
"b2",
"b6",
"b14",
"b14"
],
"table_ref": [],
"text": "For Few-NERD2 , we compare PromptNER to CONTaiNER (Das et al., 2022), ESD (Wang et al., 2022b), DecomposedNER (Ma et al., 2022b) and methods from Ding et al. (2021), e.g., StructShot, ProtoBERT, etc. For CrossNER, we compare our method to DecomposedNER (Ma et al., 2022b), L-TapNet+CDT (Hou et al., 2020) and other methods from Hou et al. (2020). We report the micro-F1 scores with standard deviations of different baselines."
},
{
"figure_ref": [],
"heading": "Implementation Details",
"publication_ref": [
"b4"
],
"table_ref": [],
"text": "We implement our method using PyTorch version 1.12.13 . We use two separate BERT models for the position-aware biaffine module and the prompt-based classifier, respectively. We load the BERT-base-uncased (Devlin et al., 2019) checkpoint from HuggingFace4 . During training, we use the AdamW optimizer with 10% linear warmup scheduler, and the weight decay ratio is 1e-2. We train our model in the training set and use the validation set to select the model with the highest F1 scores. We also use the AdamW for fine-tuning on the target domain and stop the fine-tuning process early when the loss is less than 1e-2. For more implementation details, please refer to Appendix A.1."
},
{
"figure_ref": [],
"heading": "Results and Analysis",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Main Results",
"publication_ref": [],
"table_ref": [
"tab_2",
"tab_3",
"tab_2"
],
"text": "Table 1 and Table 2 report the performance of PromptNER on two few-shot NER datasets. It can be observed that: 1) Our proposed PromptNER achieves the best performance on Few-NERD and CrossNER. The overall averaged F1 scores over Few-NERD Intra, and Inter setting are improved by 6.22% and 1.36% respectively compared to the previous SOTA model DecomposedMetaNER (Ma et al., 2022b). Meanwhile, our model also outperforms previous methods by 4.12% and 5.07% on CrossNER 1-shot and 5-shot settings, respectively.\n2) It is important that we observe the performance improvement on Few-NERD Intra is more significant than on Few-NERD Inter. This phenomenon is because Few-NERD Inter allows the train/dev/test episode to belong to the same coarse-grained types, whereas the train/dev/test episode in Few-NERD Intra must belong to different coarse-grained types and share little knowledge. Therefore, Few-NERD Intra is a more challenging benchmark. The results from Table 1 demonstrate that PromptNER has an excellent transfer learning ability than previous methods when facing difficult tasks."
},
{
"figure_ref": [
"fig_5"
],
"heading": "Ablation Study",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we demonstrate the contributions of different parts of Prompt NER. We introduce the following variants for the ablation: Results from Tabel 3 show that fine-tuning on the novel support set significantly improves the performance of our method. Although we do not fine-tune our model on the novel support set, our method still outperforms all the token-level models in the Few-NERD inter 5way 5∼10 setting, i.e., CONTaiNER (Das et al., 2022), which demonstrates the superiority of our span-based method. The rerank strategy could also significantly improve the F1 scores of our method, which indicates the performance of our method. Obviously, the rerank strategy and fine-tuning stage are the key components of our method during inference. We investigate how the effectiveness of these two components as follows:\nThe Effectiveness of Rerank Strategy. The rerank strategy is a crucial component of our method since it could effectively filter some false positive spans. As described in Eq.( 5), we use the scores from the span detector to rerank the final prediction probability. We further investigate how the rerank strategy influences performance. According to Table 4, the performance of our method is improved by 5.47% and 4.46%, respectively, when applying the rerank strategy during inference. It is worth noting that, although we extract all the spans from the input sentence, the rerank strategy could also significantly improve F1 scores by 38.95% and 50.23%, respectively. This phenomenon indicates that the entities belonging to the category mentioned in the prompt have significantly higher R(s i , e i ) scores than entities belonging to other categories, which proves the rerank strategy has the ability to filter some false positive spans. The Effectiveness of Fine-tuning. According to set during the inference stage will improve the performance by a large margin since our method has no gap between training and fine-tuning. We also investigate how performances are influenced by the different fine-tuning steps. As shown in Figure 4, the performance of our model gradually stabilizes and reaches its peak F1 scores as the number of fine-tuning steps increases, which indicates that our method could effectively utilize the examples from the novel support set to optimize label prototype embeddings."
},
{
"figure_ref": [],
"heading": "Error Analysis",
"publication_ref": [],
"table_ref": [
"tab_7"
],
"text": "The NER task has two types of errors: 'FP-Span' and 'FP-Type'. For FP-Span, it denotes that the span-detector extracts some false-positive spans from the input sentence. And FP-Type denotes that the NER system recognizes some true-positive spans but fails to categorize them into correct entity classes. As Table 5 shows, although we conduct the rerank strategy and introduce none type to filter some false positive spans, PromptNER still tends to extract a few spans with incorrect boundaries. The results from Table 6 also prove that recognizing unseen new class spans only during the span detection stage is difficult for previous few-shot NER systems because current few-shot NER systems are all fully trained in a training set, which results in the few-shot NER system extracting some entities appearing in the training set. Notably, our method does not follow the traditional prototype networks to use entities representations from the novel support set to construct label prototypes for the span classifier but achieves the lowest FP-Type ratio, demonstrating the superiority of the prompt-based classifier and k-NN search over previous traditional prototypical networks for few-shot NER.\n6 Related Work"
},
{
"figure_ref": [],
"heading": "Few-shot Learning and Meta Learning",
"publication_ref": [
"b35",
"b13",
"b27",
"b1",
"b25",
"b31",
"b28",
"b8",
"b5"
],
"table_ref": [],
"text": "Few-shot learning is an essential task that involves learning a model with only a few human-annotated examples (Wang et al., 2020). In recent years, several methods have been proposed to address different few-shot learning tasks (Geng et al., 2020;Sheng et al., 2020;Brown et al., 2020;Schick and Schütze, 2021;Gao et al., 2021a) in the NLP community. Meanwhile, various meta-learning algorithms are also proposed to address few-shot learning, i.e., metric learning-based methods (Vinyals et al., 2016;Snell et al., 2017), optimization-based methods (Finn et al., 2017), and augmentationbased learning (Ding et al., 2020)."
},
{
"figure_ref": [],
"heading": "Span-based NER",
"publication_ref": [
"b7",
"b39",
"b37",
"b17"
],
"table_ref": [],
"text": "Inspired by dependency parsing (Dozat and Manning, 2017), Yu et al. (2020) propose a span-based NER system with a biaffine model. The biaffine model scores each pair of start and end tokens to extract all the candidate spans.To enhance the performance of span-based NER, Yan et al. (2022) use the Convolutional Neural Network (CNN) to utilize spatial relations in the score matrix. Li et al. (2020) considers the NER task a Machine Reading Comprehension task. Notably, the span-based NER system could handle both flat and nested NER simultaneously, which avoid token-level label dependency problem (i.e, \"BIOES\" rules)."
},
{
"figure_ref": [],
"heading": "Few-shot NER",
"publication_ref": [
"b14",
"b6",
"b10",
"b14",
"b30",
"b2",
"b38"
],
"table_ref": [],
"text": "Recently, few-shot NER has received lots of attention in the field of Information Extraction, owing to the high cost of human annotation and the demand for domain-specific knowledge. To evaluate the performance of few-shot NER systems better, Hou et al. (2020) and Ding et al. (2021) release two well-designed datasets (CrossNER, Few-NERD) which satisfy the N∼way K∼shot paradigm. Research on few-shot NER could be categorized into two types, i.e., one-stage models (Fritzler et al., 2019;Hou et al., 2020;Tong et al., 2021;Das et al., 2022) with token-level metric learning, and two-stage models (Yu et al., 2021;Wang et al., 2022b;Ma et al., 2022b;Wang et al., 2022a) "
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "In this paper, we propose PromptNER, a prompting method for few-shot named entity recognition via k nearest neighbor search. Our approach uses a prompt to instruct Pre-trained Language Models to extract entities with specific classes. We also design a two-stage model with a position-aware biaffine module and a prompt-based classifier with k-NN search. Unlike traditional prototypical networks, our method could use only the novel support set to optimize label prototypes. Extensive experiments demonstrate that our method outperforms previous state-of-the-art few-shot NER methods.\nOur work provides a novel, simple, and effective baseline for few-shot learning in NER."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [],
"table_ref": [],
"text": "Our proposed method must be trained in a training set for warmup, then utilize its transfer learning ability to address the few-shot NER task. Meanwhile, we also only conduct experiments on the N-way K-shot settings and few-shot flat NER tasks.\nIn the future, we will extend our method to other NER scenarios, such as few-shot nested NER tasks, few-shot Chinese NER tasks."
},
{
"figure_ref": [],
"heading": "A Appendix",
"publication_ref": [],
"table_ref": [],
"text": "A.1 Implementation Details\nFor a fair comparison, we use BERT-base-uncased as the encoder for our method. We use AdamW to optimize our model with 10% linear warm-up steps. The learning rate of the encoder is 2e-5, and the learning rate of the biaffine decoder is 2e-3. We set the batch size as 1 to narrow the gap between training and fine-tuning, which means we use one episode per step to update our model. For fine-tuning, we stop the fine-tuning process early when the loss is less than 1e-2 or the finetuning steps are more than 50. We conduct experiments on Few-NERD and CrossNER with five different random seeds {1 2 3 4 5} and report the average micro-F1 with standard deviations. For inference, γ, α, β are hyper-parameters that balance these three distributions. We set γ as 0.5 for Few-NERD inter setting and 0.7 for other settings. Meanwhile, we set α as 0.35 * (1 -γ) and β as 0.65 * (1 -γ), respectively. Our source codes are available at https://github.com/Zhang-Mozhi/PromptNER."
},
{
"figure_ref": [],
"heading": "A.2 Contrastive Learning",
"publication_ref": [
"b2"
],
"table_ref": [
"tab_9",
"tab_10"
],
"text": "Recently, Contrastive Learning has been proven effective for token-level metric learning (Das et al., 2022). We also design a span-based contrastive learning algorithm to investigate whether contrastive learning could optimize the span embedding between entities with different labels. In the 1-shot setting, we just let the X p go through the encoder twice (Gao et al., 2021b) to obtain sufficient positive samples. We could get the golden span set M within a support set. Given a golden span u i , we can define its corresponding positive sample set M + i and in-batch sample set M - i :\nM + i = {u j ∈ M|y j = y i , u j = u i }, M - i = {u j ∈ M|y j = y i , u j = u i },\nThen, the span-based contrastive learning loss can be calculated as follows:\nL CL = - |M| i=1 log (u i ,u j )∈M + i exp(d(u i , u j )) u k ∈M - i exp(d(u i , u k )) ,\nwhere d is a scaled dot product function. By optimizing L CL , we can narrow the embedding distribution of entities with identical labels and separate the entity distribution with different labels. Therefore, the optimized object of PromptNER could be calculated as follows:\nL = L span + L class + L CL ,\nTable 7 and Table 8 denote the performance when applying contrastive learning to our method. We find that contrastive learning will accelerate the overfitting phenomenon of the novel support set, which might harm the performance of our method. "
},
{
"figure_ref": [],
"heading": "Models",
"publication_ref": [],
"table_ref": [],
"text": ""
}
] | Few-shot Named Entity Recognition (NER) is a task aiming to identify named entities via limited annotated samples. Recently, prototypical networks have shown promising performance in few-shot NER. Most of prototypical networks will utilize the entities from the support set to construct label prototypes and use the query set to compute span-level similarities and optimize these label prototype representations. However, these methods are usually unsuitable for fine-tuning in the target domain, where only the support set is available. In this paper, we propose PromptNER: a novel prompting method for few-shot NER via k nearest neighbor search. We use prompts that contains entity category information to construct label prototypes, which enables our model to fine-tune with only the support set. Our approach achieves excellent transfer learning ability, and extensive experiments on the Few-NERD and CrossNER datasets demonstrate that our model achieves superior performance over state-of-the-art methods. | PromptNER: A Prompting Method for Few-shot Named Entity Recognition via k Nearest Neighbor Search | [
{
"figure_caption": "Figure 1 :1Figure 1: The difference between traditional Prototypical Networks and PromptNER.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: An example of 2-way 2-shot episode.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: Model structure of our method.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Few-NERD Ding et al. (2021) propose a humanannotated few-shot NER dataset with 8 coarsegrained and 66 fine-grained entity types from Wikipedia. Because the sampling process becomes gradually stricter to satisfy the K-shot setting, therefore, each entity type contains K ∼ 2K samples, which alleviates the sampling limitation in Few-NERD. Few-NERD contains two different settings: Intra and Inter. CrossNER CrossNER consists of 4 NER datasets from different domains: CoNLL03 (Sang and De Meulder, 2003)(News), WNUT-2017(Derczynski et al., 2017)(Social), GUM(Zeldes, 2017)(Wiki) and OntoNotes(Pradhan et al., 2013)(Mixed). For a fair comparison, we use the sampled N -way K-shot dataset fromHou et al. (2020).",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "1) Ours w/o Fine-tune 2) Ours w/o Rerank 3) Ours w/o k-NN search 4) Ours w/o Fine-tune and k-NN search 5) Ours w/o Position-aware Biaffine 6) Ours w/o Fine-tune and RoPE.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: The Effectiveness of Fine-tuning.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "",
"figure_data": "Support Set:1. [Query Set:1. Tesla CEO Elon Musk is known for his ambitious goals for space exploration.2. Bill Gates was the CEO of Microsoft.Output:Person:Elon Musk,Bill GatesCompany:Tesla,Microsoft",
"figure_id": "tab_0",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "1976 and turned it into one of the most successful companies in the world. 2. The CEO of [Amazon] Company , [Jeff Bezos] Person recently announced his resignation from the company.",
"figure_data": "",
"figure_id": "tab_1",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "F1 scores with standard deviations on Few-NERD for both Inter and Intra settings. † denotes the results reported inDing et al. (2021) Arxiv V6 Version. ‡ is the result without standard deviations from(Das et al., 2022). The best results are in bold.",
"figure_data": "IntraInterModels1∼2-shot5∼10-shotAvg.1∼2-shot5∼10-shotAvg.5 way10 way5 way10 way5 way10 way5 way10 wayProtoBERT †20.76±0.84 15.05±0.44 42.54±0.94 35.40±0.13 28.44 38.83±1.49 32.45±0.79 58.79±0.44 52.92±0.37 45.75NNShot †25.78±0.91 18.27±0.41 36.18±0.79 27.67±1.06 26.98 54.29±0.40 46.98±1.96 50.56±3.33 50.00±0.36 50.46StructShot †30.21±0.90 21.03±1.13 38.00±1.29 26.42±0.60 28.92 51.88±0.69 43.34±0.10 57.32±0.63 49.57±3.08 50.53CONTAINER ‡40.4333.8453.7047.49 43.8755.9548.3561.8357.12 55.81ESD36.08±1.60 30.00±0.70 52.14±1.50 42.15±2.60 40.09 59.29 ±1.25 52.16±0.79 69.06±0.80 64.00±0.43 61.13DecomposedMetaNER 49.48±0.85 42.84±0.46 62.92±0.57 53.14±0.25 52.10 64.75±0.35 58.65±0.43 71.49±0.47 68.11±0.05 65.75Ours55.32±1.03 50.29±0.61 67.26±1.02 60.42±0.73 58.32 64.92±0.71 62.28±0.39 72.64±0.16 70.13±0.67 67.49Models1-shot5-shotCoNLL03 GUMWNUT OntoNotes Avg. CoNLL03 GUMWNUT OntoNotes Avg.TransferBERT †4.75±1.42 0.57±0.32 2.71±0.72 3.46±0.54 2.87 15.36±2.81 3.62±0.57 11.08±0.57 35.49±7.60 16.39SimBERT †19.22±0.00 6.91±0.00 5.18±0.00 13.99±0.00 11.33 32.01±0.00 10.63±0.00 8.20±0.00 21.14±0.00 18.00Matching Network †19.50±0.35 4.73±0.16 17.23±2.75 15.06±1.61 14.13 19.85±0.74 5.58±0.23 6.61±1.75 8.08±0.47 10.03ProtoBERT †32.49±2.01 3.89±0.24 10.68±1.40 6.67±0.46 13.43 50.06±1.57 9.54±0.44 17.26±2.65 13.59±1.61 22.61L-TapNet+CDT †44.30±3.15 12.04±0.65 20.80±1.06 15.17±1.25 23.08 45.35±2.67 11.65±2.34 23.30±2.80 20.95±2.81 25.32DecomposedMetaNER 46.09±0.44 17.54±0.98 25.14±0.24 34.13±0.92 30.73 58.18±0.87 31.36±0.91 31.02±1.28 45.55±0.90 41.53Ours49.69±2.70 26.24±1.21 28.07±0.48 35.38±0.58 34.85 63.47±1.28 44.54±0.29 30.40±0.83 48.71±0.59 46.78",
"figure_id": "tab_2",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "F1 scores with standard deviations on CrossNER.",
"figure_data": "",
"figure_id": "tab_3",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": ": Ablation study of different components ofour method. We conduct ablation experiments on Few-NERD Intra/Inter 5way 5∼10-shot setting.that this strategy could help to filter some falsepositive spans. The k-NN Search achieves lessperformance improvement compared to the modelwithout fine-tuning since fine-tuning the prompt-based classifier on the support set will narrow theembedding distributions between the label wordand golden entities in the support set. Accordingto Table 3, when we remove the Position-awareBiaffine Module during the inference stage, the",
"figure_id": "tab_5",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "fine-tuning on the novel support",
"figure_data": "Settings N-wayF1FP-Span FP-TypeInter5-way 10-way 62.28 64.9289.24 81.5110.76 18.49Intra5-way 10-way 50.29 55.3271.66 62.3128.34 37.69Table 5: Error analysis (%) of 1∼2-shot settings onFew-NERD dataset. \"FP-Span\" denotes that the span-detector extracts false-positive spans. \"FP-Type\" de-notes extracted spans with incorrect entity classes.ModelsF1FP-Span FP-TypeProtoBERT38.8386.7013.30NNShot47.2484.7015.30StructShot51.8880.0020.00ESD59.2972.8027.20DecomMeta 64.7576.4846.53Ours64.9289.2410.76",
"figure_id": "tab_6",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Error analysis (%) of 5-way 1∼2-shot on Few-NERD Inter for different methods.",
"figure_data": "",
"figure_id": "tab_7",
"figure_label": "6",
"figure_type": "table"
},
{
"figure_caption": "76±0.84 15.05±0.44 42.54±0.94 35.40±0.13 28.44 38.83±1.49 32.45±0.79 58.79±0.44 52.92±0.37 45.75 NNShot † 25.78±0.91 18.27±0.41 36.18±0.79 27.67±1.06 26.98 54.29±0.40 46.98±1.96 50.56±3.33 50.00±0.36 50.46 StructShot † 30.21±0.90 21.03±1.13 38.00±1.29 26.42±0.60 28.92 51.88±0.69 43.34±0.10 57.32±0.63 49.57±3.08 50.53 CONTAINER ‡ 08±1.60 30.00±0.70 52.14±1.50 42.15±2.60 40.09 59.29 ±1.25 52.16±0.79 69.06±0.80 64.00±0.43 61.13 DecomposedMetaNER 49.48±0.85 42.84±0.46 62.92±0.57 53.14±0.25 52.10 64.75±0.35 58.65±0.43 71.49±0.47 68.11±0.05 65.75 Ours w/o CL 55.32±1.03 50.29±0.61 67.26±1.02 60.42±0.73 58.32 64.92±0.71 62.28±0.39 72.64±0.16 70.13±0.67 67.49 Ours w CL 54.92±0.56 49.49±0.54 66.97±0.10 59.77±0.65 57.79 64.93±0.44 62.16±0.36 72.15±0.20 69.20±0.89 67.11 F1 scores with standard deviations on Few-NERD for both Inter and Intra settings. † denotes the results reported in Ding et al. (2021) Arxiv V6 Version. ‡ is the result without standard deviations from (Das et al., 2022). The best results are in bold. TransferBERT † 4.75±1.42 0.57±0.32 2.71±0.72 3.46±0.54 2.87 15.36±2.81 3.62±0.57 11.08±0.57 35.49±7.60 16.39 SimBERT † 19.22±0.00 6.91±0.00 5.18±0.00 13.99±0.00 11.33 32.01±0.00 10.63±0.00 8.20±0.00 21.14±0.00 18.00 Matching Network † 19.50±0.35 4.73±0.16 17.23±2.75 15.06±1.61 14.13 19.85±0.74 5.58±0.23 6.61±1.75 8.08±0.47 10.03 ProtoBERT † 32.49±2.01 3.89±0.24 10.68±1.40 6.67±0.46 13.43 50.06±1.57 9.54±0.44 17.26±2.65 13.59±1.61 22.61 L-TapNet+CDT † 44.30±3.15 12.04±0.65 20.80±1.06 15.17±1.25 23.08 45.35±2.67 11.65±2.34 23.30±2.80 20.95±2.81 25.32 DecomposedMetaNER 46.09±0.44 17.54±0.98 25.14±0.24 34.13±0.92 30.73 58.18±0.87 31.36±0.91 31.02±1.28 45.55±0.90 41.53 Ours w/o CL 49.69±2.70 26.24±1.21 28.07±0.48 35.38±0.58 34.85 63.47±1.28 44.54±0.29 30.40±0.83 48.71±0.59 46.78 Ours w CL 46.37±3.55 24.46±1.55 27.03±0.98 33.48±0.47 32.84 63.29±1.84 43.14±1.09 30.17±0.67 48.75±1.12 46.34",
"figure_data": "IntraInter1∼2-shot5∼10-shotAvg.1∼2-shot5∼10-shotAvg.5 way10 way5 way10 way5 way10 way5 way10 wayProtoBERT †20.40.4333.8453.7047.49 43.8755.9548.3561.8357.12 55.81ESD36.",
"figure_id": "tab_9",
"figure_label": "7",
"figure_type": "table"
},
{
"figure_caption": "F1 scores with standard deviations on CrossNER. † are the results reported inHou et al. (2020). The best results are in bold.",
"figure_data": "",
"figure_id": "tab_10",
"figure_label": "8",
"figure_type": "table"
}
] | Mozhi Zhang; Hang Yan; Yaqian Zhou; Xipeng Qiu | [
{
"authors": "Navaneeth Bodla; Bharat Singh; Rama Chellappa; Larry S Davis",
"journal": "",
"ref_id": "b0",
"title": "Soft-nms-improving object detection with one line of code",
"year": "2017"
},
{
"authors": "Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell",
"journal": "Advances in neural information processing systems",
"ref_id": "b1",
"title": "Language models are few-shot learners",
"year": "2020"
},
{
"authors": "Sarkar Snigdha; Sarathi Das; Arzoo Katiyar; Rebecca Passonneau; Rui Zhang",
"journal": "Association for Computational Linguistics",
"ref_id": "b2",
"title": "CONTaiNER: Few-shot named entity recognition via contrastive learning",
"year": "2022"
},
{
"authors": "Leon Derczynski; Eric Nichols; Marieke Van Erp; Nut Limsopatham",
"journal": "Association for Computational Linguistics",
"ref_id": "b3",
"title": "Results of the WNUT2017 shared task on novel and emerging entity recognition",
"year": "2017"
},
{
"authors": "Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova",
"journal": "Association for Computational Linguistics",
"ref_id": "b4",
"title": "BERT: Pre-training of deep bidirectional transformers for language understanding",
"year": "2019"
},
{
"authors": "Bosheng Ding; Linlin Liu; Lidong Bing; Canasai Kruengkrai; Hai Thien; Shafiq Nguyen; Luo Joty; Chunyan Si; Miao",
"journal": "Association for Computational Linguistics",
"ref_id": "b5",
"title": "DAGA: Data augmentation with a generation approach for low-resource tagging tasks",
"year": "2020"
},
{
"authors": "Ning Ding; Guangwei Xu; Yulin Chen; Xiaobin Wang; Xu Han; Pengjun Xie; Haitao Zheng; Zhiyuan Liu",
"journal": "Association for Computational Linguistics",
"ref_id": "b6",
"title": "Few-NERD: A few-shot named entity recognition dataset",
"year": "2021"
},
{
"authors": "Timothy Dozat; Christopher D Manning",
"journal": "",
"ref_id": "b7",
"title": "Deep biaffine attention for neural dependency parsing",
"year": "2017"
},
{
"authors": "Chelsea Finn; Pieter Abbeel; Sergey Levine",
"journal": "",
"ref_id": "b8",
"title": "Model-agnostic meta-learning for fast adaptation of deep networks",
"year": "2017"
},
{
"authors": " Pmlr",
"journal": "",
"ref_id": "b9",
"title": "",
"year": ""
},
{
"authors": "Alexander Fritzler; Varvara Logacheva; Maksim Kretov",
"journal": "",
"ref_id": "b10",
"title": "Few-shot classification in named entity recognition task",
"year": "2019"
},
{
"authors": "Tianyu Gao; Adam Fisch; Danqi Chen; ; ",
"journal": "Association for Computational Linguistics",
"ref_id": "b11",
"title": "Making pre-trained language models better few-shot learners",
"year": "2021"
},
{
"authors": "Tianyu Gao; Xingcheng Yao; Danqi Chen",
"journal": "",
"ref_id": "b12",
"title": "Simcse: Simple contrastive learning of sentence embeddings",
"year": "2021"
},
{
"authors": "Ruiying Geng; Binhua Li; Yongbin Li; Jian Sun; Xiaodan Zhu",
"journal": "Association for Computational Linguistics",
"ref_id": "b13",
"title": "Dynamic memory induction networks for few-shot text classification",
"year": "2020"
},
{
"authors": "Yutai Hou; Wanxiang Che; Yongkui Lai; Zhihan Zhou; Yijia Liu; Han Liu; Ting Liu",
"journal": "Association for Computational Linguistics",
"ref_id": "b14",
"title": "Few-shot slot tagging with collapsed dependency transfer and label-enhanced task-adaptive projection network",
"year": "2020"
},
{
"authors": "Zhiheng Huang; Wei Xu; Kai Yu",
"journal": "",
"ref_id": "b15",
"title": "Bidirectional lstm-crf models for sequence tagging",
"year": "2015"
},
{
"authors": "Guillaume Lample; Miguel Ballesteros; Sandeep Subramanian; Kazuya Kawakami; Chris Dyer",
"journal": "",
"ref_id": "b16",
"title": "Neural architectures for named entity recognition",
"year": "2016"
},
{
"authors": "Xiaoya Li; Jingrong Feng; Yuxian Meng; Qinghong Han; Fei Wu; Jiwei Li",
"journal": "Association for Computational Linguistics",
"ref_id": "b17",
"title": "A unified MRC framework for named entity recognition",
"year": "2020"
},
{
"authors": "Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov",
"journal": "",
"ref_id": "b18",
"title": "Roberta: A robustly optimized bert pretraining approach",
"year": "2019"
},
{
"authors": "Jie Ma; Miguel Ballesteros; Srikanth Doss; Rishita Anubhai; Sunil Mallya; Yaser Al-Onaizan; Dan Roth",
"journal": "Association for Computational Linguistics",
"ref_id": "b19",
"title": "Label semantics for few shot named entity recognition",
"year": "2022"
},
{
"authors": "Tingting Ma; Huiqiang Jiang; Qianhui Wu; Tiejun Zhao; Chin-Yew Lin",
"journal": "Association for Computational Linguistics",
"ref_id": "b20",
"title": "Decomposed meta-learning for few-shot named entity recognition",
"year": "2022"
},
{
"authors": "Xuezhe Ma; Eduard Hovy",
"journal": "",
"ref_id": "b21",
"title": "End-to-end sequence labeling via bi-directional lstm-cnns-crf",
"year": "2016"
},
{
"authors": "Hong Ming; Jiaoyun Yang; Lili Jiang; Yan Pan; Ning An",
"journal": "",
"ref_id": "b22",
"title": "Few-shot nested named entity recognition",
"year": "2022"
},
{
"authors": "Alessandro Sameer Pradhan; Nianwen Moschitti; Xue; Tou Hwee; Anders Ng; Olga Björkelund; Yuchen Uryupina; Zhi Zhang; Zhong",
"journal": "Association for Computational Linguistics",
"ref_id": "b23",
"title": "Towards robust linguistic analysis using OntoNotes",
"year": "2013"
},
{
"authors": "Erik Tjong; Kim Sang; Fien De; Meulder ",
"journal": "",
"ref_id": "b24",
"title": "Introduction to the conll-2003 shared task: Languageindependent named entity recognition",
"year": "2003"
},
{
"authors": "Timo Schick; Hinrich Schütze",
"journal": "",
"ref_id": "b25",
"title": "It's not just size that matters: Small language models are also few-shot learners",
"year": "2021"
},
{
"authors": "Yongliang Shen; Xinyin Ma; Zeqi Tan; Shuai Zhang; Wen Wang; Weiming Lu",
"journal": "Association for Computational Linguistics",
"ref_id": "b26",
"title": "Locate and label: A two-stage identifier for nested named entity recognition",
"year": "2021"
},
{
"authors": "Jiawei Sheng; Shu Guo; Zhenyu Chen; Juwei Yue; Lihong Wang; Tingwen Liu; Hongbo Xu",
"journal": "Association for Computational Linguistics",
"ref_id": "b27",
"title": "Adaptive Attentional Network for Few-Shot Knowledge Graph Completion",
"year": "2020"
},
{
"authors": "Jake Snell; Kevin Swersky; Richard Zemel",
"journal": "Advances in neural information processing systems",
"ref_id": "b28",
"title": "Prototypical networks for few-shot learning",
"year": "2017"
},
{
"authors": "Jianlin Su; Ahmed Murtadha; Shengfeng Pan; Jing Hou; Jun Sun; Wanwei Huang; Bo Wen; Yunfeng Liu",
"journal": "",
"ref_id": "b29",
"title": "Global pointer: Novel efficient spanbased approach for named entity recognition",
"year": "2022"
},
{
"authors": "Meihan Tong; Shuai Wang; Bin Xu; Yixin Cao; Minghui Liu; Lei Hou; Juanzi Li",
"journal": "",
"ref_id": "b30",
"title": "Learning from miscellaneous other-class words for fewshot named entity recognition",
"year": "2021"
},
{
"authors": "Oriol Vinyals; Charles Blundell; Timothy Lillicrap; Daan Wierstra",
"journal": "Advances in neural information processing systems",
"ref_id": "b31",
"title": "Matching networks for one shot learning",
"year": "2016"
},
{
"authors": "Jianing Wang; Chengyu Wang; Chuanqi Tan; Minghui Qiu; Songfang Huang; Jun Huang; Ming Gao; ; ",
"journal": "Association for Computational Linguistics",
"ref_id": "b32",
"title": "SpanProto: A two-stage span-based prototypical network for few-shot named entity recognition",
"year": "2022"
},
{
"authors": "Peiyi Wang; Runxin Xu; Tianyu Liu; Qingyu Zhou; Yunbo Cao; Baobao Chang; Zhifang Sui",
"journal": "Association for Computational Linguistics",
"ref_id": "b33",
"title": "An enhanced span-based decomposition method for few-shot sequence labeling",
"year": "2022"
},
{
"authors": "Shuhe Wang; Xiaoya Li; Yuxian Meng; Tianwei Zhang; Rongbin Ouyang; Jiwei Li; Guoyin Wang",
"journal": "",
"ref_id": "b34",
"title": "k nn-ner: Named entity recognition with nearest neighbor search",
"year": "2022"
},
{
"authors": "Yaqing Wang; Quanming Yao; James T Kwok; Lionel M Ni",
"journal": "ACM computing surveys (csur)",
"ref_id": "b35",
"title": "Generalizing from a few examples: A survey on few-shot learning",
"year": "2020"
},
{
"authors": "Hang Yan; Bocao Deng; Xiaonan Li; Xipeng Qiu",
"journal": "",
"ref_id": "b36",
"title": "Tener: adapting transformer encoder for named entity recognition",
"year": "2019"
},
{
"authors": "Hang Yan; Yu Sun; Xiaonan Li; Xipeng Qiu",
"journal": "",
"ref_id": "b37",
"title": "An embarrassingly easy but strong baseline for nested named entity recognition",
"year": "2022"
},
{
"authors": "Dian Yu; Luheng He; Yuan Zhang; Xinya Du; Panupong Pasupat; Qi Li",
"journal": "Association for Computational Linguistics",
"ref_id": "b38",
"title": "Few-shot intent classification and slot filling with retrieved examples",
"year": "2021"
},
{
"authors": "Juntao Yu; Bernd Bohnet; Massimo Poesio",
"journal": "",
"ref_id": "b39",
"title": "Named entity recognition as dependency parsing",
"year": "2020"
},
{
"authors": "Amir Zeldes",
"journal": "Language Resources and Evaluation",
"ref_id": "b40",
"title": "The gum corpus: Creating multilayer resources in the classroom",
"year": "2017"
}
] | [
{
"formula_coordinates": [
3,
126.98,
489.96,
162.15,
27.31
],
"formula_id": "formula_0",
"formula_text": "X p = F prompt (T ) ⊕ X, = [X l , X m , X](1)"
},
{
"formula_coordinates": [
3,
370.8,
506.2,
153.61,
27.34
],
"formula_id": "formula_1",
"formula_text": "H = Encoder(X p ), = [H l , H m , H n ](2)"
},
{
"formula_coordinates": [
4,
70.47,
94.73,
220.02,
84.77
],
"formula_id": "formula_2",
"formula_text": "h s = LeakyReLU(h i W s ), h e = LeakyReLU(h j W e ), R i,j = h s Uh e + (R i h i W p ) (R j h j W p ), = h s Uh e + (h i W p ) R j-i (h j W p ) (3) where W s , W e , W p ∈ R d×h , U, R i , R j ∈ R h×h ,"
},
{
"formula_coordinates": [
4,
104.11,
283.95,
124.69,
25.29
],
"formula_id": "formula_3",
"formula_text": "L pos = log 1 + (i,j)∈Spos e -"
},
{
"formula_coordinates": [
4,
103,
317.53,
132.89,
25.29
],
"formula_id": "formula_4",
"formula_text": "L neg = log 1 + (i,j)∈Sneg e r(i"
},
{
"formula_coordinates": [
4,
70.47,
350.54,
218.16,
34.59
],
"formula_id": "formula_5",
"formula_text": "L span = L pos + L neg , where 1 ≤ i ≤ j ≤ n, S pos = {(s k , e k )} N k=1"
},
{
"formula_coordinates": [
4,
123.38,
736.04,
113.23,
35.05
],
"formula_id": "formula_6",
"formula_text": "u i = 1 e i -s i + 1 e i k=s i h k ,"
},
{
"formula_coordinates": [
4,
343.15,
136.42,
144.26,
25.95
],
"formula_id": "formula_7",
"formula_text": "p(y|s i , e i ) = Softmax( H m u i √ d ),"
},
{
"formula_coordinates": [
4,
326.39,
239.22,
177.77,
34.9
],
"formula_id": "formula_8",
"formula_text": "L class = 1 |S pos | |Spos| i=1 -log p(y|s i , e i ) ,"
},
{
"formula_coordinates": [
4,
371.33,
422.27,
153.08,
10.77
],
"formula_id": "formula_9",
"formula_text": "L = L span + L class (4)"
},
{
"formula_coordinates": [
4,
324.47,
616.3,
181.61,
25.95
],
"formula_id": "formula_10",
"formula_text": "p prompt (y|s i , e i ) = Softmax( U label u i √ d ),"
},
{
"formula_coordinates": [
5,
140.51,
120.09,
78.98,
26.15
],
"formula_id": "formula_11",
"formula_text": "d i = U golden u i √ d ,"
},
{
"formula_coordinates": [
5,
75.21,
207.7,
209.58,
33.71
],
"formula_id": "formula_12",
"formula_text": "p(y i = t|s i , e i ) ∝ n j=1 I(j ∈ N i , y j = t) • d i (j),"
},
{
"formula_coordinates": [
5,
100.21,
322.9,
188.93,
43.85
],
"formula_id": "formula_13",
"formula_text": "p(y|s i , e i ) = γ • Sigmoid(R(s i , e i )) + α • p prompt (y|s i , e i ) + β • p knn (y|s i , e i )(5)"
},
{
"formula_coordinates": [
5,
109.36,
462.41,
141.27,
10.77
],
"formula_id": "formula_14",
"formula_text": "y pred = argmax p(y = t|s i , e i ),"
},
{
"formula_coordinates": [
12,
98.21,
604.73,
163.59,
31.37
],
"formula_id": "formula_15",
"formula_text": "M + i = {u j ∈ M|y j = y i , u j = u i }, M - i = {u j ∈ M|y j = y i , u j = u i },"
},
{
"formula_coordinates": [
12,
73.75,
687.76,
212.5,
35.03
],
"formula_id": "formula_16",
"formula_text": "L CL = - |M| i=1 log (u i ,u j )∈M + i exp(d(u i , u j )) u k ∈M - i exp(d(u i , u k )) ,"
},
{
"formula_coordinates": [
12,
352.96,
125.98,
124.63,
10.77
],
"formula_id": "formula_17",
"formula_text": "L = L span + L class + L CL ,"
}
] |
2023-05-23 | [
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b4",
"b14",
"b8",
"b15",
"b16",
"b17",
"b18",
"b19",
"b23",
"b23",
"b25"
],
"table_ref": [],
"text": "Large Language Models (LLMs) [1; 2; 3; 4] have demonstrated remarkable performance for various downstream tasks without task-specific fine-tuning. Recently, based on the powerful LLMs, there has been a surge of research [5; 6; 7; 8; 9; 10; 11; 12] that successfully adapt LLMs to visionlanguage tasks, resulting in powerful Multimodal LLMs (MLLMs), e.g., BLIP-2 [5]. When properly fed with visual data, they are shown to be capable of understanding the visual world and responding to instructions accordingly. Such vision-language understanding capability makes LLM a universal interface for multimodal tasks, contributing towards a tentative yet promising direction towards Artificial General Intelligence (AGI) [13; 14]. Within this framework, images are projected to the linguistic space for the LLMs to understand, where the common practice employs an image-text pre-trained visual tokenizer 3 , i.e., CLIP [15]. However, even though CLIP has shown strong capacity for image representations, to the best of our knowledge, it is yet to be explored whether CLIP is the optimal visual tokenizer for MLLMs. The absence of such investigation calls for a comprehensive comparison of existing visual tokenizers under MLLMs' framework. However, recent MLLMs have mostly investigated their performance in terms of generation quality [7; 8] or on a small set of questions [9], leaving a comprehensive quantitative evaluation untouched.\nTo this end, we curated a new benchmark to study what makes for a Good Visual Tokenizer (GVT-Bench). It is especially designed to evaluate an MLLM's visual understanding capability from two important perspectives: semantic understanding and fine-grained visual perception capabilities. As shown in Figure 1, the former is evaluated on Visual Question Answering (VQA) and image captioning. While the latter is tested on two new tasks: Object Counting (OC) and Multi-Class Identification (MCI), which requires in-depth understanding of fine-grained visual information. Based on this benchmark, we comprehensively evaluated existing visual tokenizers with same architecture but different pretraining methods, including fully supervised (DeiT [16]), text-guided weakly supervised (CLIP [17]) and self-supervised (MAE [18], DINO [19], DINOv2 [20]) models (Section 2). Our main observations are i) fully supervised and text-guided weakly supervised visual tokenizers demonstrate better semantic representation capacity than their self-supervised counterparts, but the gap is narrowed by scaling up the pretraining dataset (i.e., CLIP vs. DINOv2). ii) Self-supervised visual tokenizers show better fine-grained visual perception capacity, where patch-level supervision leads to superior region-level understanding. iii) On instruction tuning datasets which are often smaller than visual tokenizer pretraining dataset [8; 7], jointly tuning the visual tokenizer leads to noticeable semantic loss (i.e., frozen CLIP performs much better than tunable CLIP on semantic understanding tasks).\nGiven the fact that none of the previous visual tokenizers exhibit both good semantic and fine-grained visual perceptual capabilities, we reviewed existing methods that integrate semantic and region supervision and question whether they bring the best of the two worlds into a visual tokenizer. Existing methods can be mainly divided into two categories. Methods in the first group [21; 22] enhance a pretrained CLIP with region-level supervision, which comes from a pretrained Region Proposal Network (RPN) or bounding box annotations. However, we found that this leads to the loss of original semantics, which can not be justified by the limited improvements on fine-grained visual perception capabilties. The other group of methods [23; 24] utilize patch features from a pretrained CLIP as region supervision to train a new model, intending to enhance its fine-grained visual perceptual capability while maintaining the rich semantics. Specifically, [23; 25] uses CLIP features to supervise the training of Masked-Image-Modeling (MIM), while Feature Distillation [24] directly distills the CLIP feature into a new model without patch masking. Nonetheless, the introduction of [MASK] token in MIM leads to train-test mismatch, requiring the visual tokenizer to be jointly optimized in the instruction-tuning process, which again leads to semantic loss with the small-scale instruction tuning dataset. As such, we argue that the mask-based strategies that were once all the rage may not be applicable for obtaining good visual tokenizers under MLLM's framework.\nBased on these insights, we seek a new visual tokenizer with both strong semantic understanding and fine-grained visual perception capabilities via Feature Distillation [24]. Specifically, given a pretrained CLIP with rich semantics, we distill it into a new model by using the patch features as supervision, without patch masking. In this way, the rich semantics from large-scale image-text contrastive pretraining is preserved, and the fine-grained visual perceptual capability is greatly enhanced with patch supervision. With our new visual tokenizer and the language model Vicuna [26], we obtain a new MLLM with Good Visual Tokenizer (GVT). Benefiting from the versatile visual tokenizer, GVT is able to perform well vision language tasks that require visual understanding at multiple levels. Without introducing extra parameters, we achieve superior performance on semantic understanding tasks, i.e., VQA and image captioning, as well as fine-grained visual understanding tasks: instance counting and multi-class identification.\nTo summarize, our contributions are as follows: "
},
{
"figure_ref": [],
"heading": "GVTBench for Empirical Study",
"publication_ref": [],
"table_ref": [],
"text": "To comprehensively study what makes for good visual tokenizers for MLLMs, we conduct a series of experiments to study the property of various visual tokenizers with same architecture but different pretraining methods. In this work, we mainly investigate MLLMs' visual understanding capability from two important perspectives: semantic understanding and fine-grained visual perception."
},
{
"figure_ref": [],
"heading": "Experimental Setup",
"publication_ref": [
"b27",
"b28",
"b29",
"b30",
"b27",
"b31",
"b32"
],
"table_ref": [],
"text": "GVTBench. A comprehensive evaluation requires a benchmark that suitably quantify MLLM's visual understanding capability. Nonetheless, existing vision-language tasks mainly focus on semantic understanding [27; 28], leaving a special focus on fine-grained visual perception untouched. To this end, we curated a new benchmark -GVTBench. It evaluates the semantic understanding capability of an MLLM on VQA [28] and Image Captioning (IC) [29]. We report accuracy for the former and CIDEr [30] and SPICE [31] for the latter. For fine-grained visual perception capability evaluation, we specially design two new tasks for MLLMs:\n• Object Counting (OC). We ask the model to count the number of a certain object appearing in the image with the prompt \"Question: How many {obj} are there in the image? Answer:\". We regard it as a classification task and report a model's prediction accuracy. • Multi-Class Identification (MCI). We ask the model if a certain object exists in the image with the prompt \"Question: Does {obj} exist in the image? Answer:\". The model is expected to answer \" Yes/No\", resulting in a binary classification problem. We report accuracy for this task.\nNotably, in the VQAv2 [28] benchmark, there are also questions related to numbers. Nevertheless, these questions are often coupled with high-level semantics, making it unsuitable to strictly evaluate fine-grained visual understanding capabilities. In contrast, our OC and MCI tasks attend to individual objects, which is decoupled from high-level semantics and thus a more appropriate test bed for fine-grained visual understanding evaluation.\nExperimental Setting. We use different visual tokenizers to encode an image into a set of visual tokens. Then, we follow Flamingo [32] to use the Perceiver Resampler [33] to reduce the number of visual tokens to a fixed length, which are fed into LLM (i.e., Vicuna). The models are trained on a instruction dataset which contains about 5M image-text pairs. In the training process, the language model is always frozen, while the visual tokenizer can be frozen or jointly optimized. For more implementation details, please refer to the appendix. "
},
{
"figure_ref": [],
"heading": "Comparing Visual Tokenizers",
"publication_ref": [
"b33",
"b15",
"b18",
"b19",
"b17",
"b16",
"b34"
],
"table_ref": [
"tab_1"
],
"text": "On GVTBench, we evaluate visual tokenizers with the same architecture (ViT-B [34]) but different pretraining strategies, including fully-supervised (DeiT [16]), self-supervised (DINO [19], DI-NOv2 [20], MAE [18]) and text-guided weakly supervised (CLIP [17]) pretraining. Based on the results in Table 1, we arrive at the following conclusions.\nFully/weakly supervised models capture more semantics than self-supervised ones, but the gap is narrowed by scaling up the pre-training dataset. With tokenizers pretrained on relative smallscale dataset (i.e., ImageNet-1k [35] with 1.28M images), DeiT demonstrates better image captioning performance (65.8 CIDEr) than self-supervised models DINO (45.0) and MAE (37.3), without jointly tuning the visual tokenizer. However, with 142M images for pretraining, the self-supervised model -DINOv2 outperforms the supervised DeiT on image captioning (67.9) and VQA (51.3), and is only inferior to CLIP which is pretrained with weak supervision from a large-scale dataset with 400M image-text pairs. This indicates that supervision is beneficial for semantic representation capability, but this can also emerge from large-scale pretraining with self-supervision.\nSelf-supervised models are better at fine-grained perception, where patch-level supervision is particularly effective. On fine-grained visual understanding tasks, i.e., OC and MCI, selfsupervised models demonstrate consistently better performance than those with supervision. When they are jointly tuned on the instruction dataset, their OC and MCI performance are mostly boosted, indicating their fine-grained visual perception capability gets improved. Among all the self-supervised models, MAE achieves the best performance, indicating the patch-based supervision is particularly effective for improving fine-grained visual understanding.\nTuning semantic-rich visual tokenizer leads to semantic loss on small-scale instruction tuning dataset. When the tokenizer is jointly optimized on the instruction tuning dataset, the rich semantics obtained from large-scale pretraining in CLIP and DINOv2 have noticeably dropped (e.g., CLIP VQA 52.2 → 47.7 and DINOv2 captioning 67.9 → 49.6). We conjecture this is due to the relatively small scale of our instruction dataset (∼5M ≪ 142M). As such, for modern MLLMs that are often tuned on small-scale and high-quality instruction datasets [7; 8], jointly tuning the visual tokenizer may not be a good option.\n3 Unifying Semantic and Fine-grained Visual Understanding"
},
{
"figure_ref": [],
"heading": "CLIP with Region-based Training",
"publication_ref": [
"b20",
"b21",
"b22",
"b22",
"b23"
],
"table_ref": [
"tab_1",
"tab_2",
"tab_3"
],
"text": "The generalist MLLMs call for a versatile visual tokenizer that could properly represent an image's content at multiple levels. However, based on the results in Table 1, none of existing pretraining methods leads to a good visual tokenizer that excels at both semantic and fine-grained visual percep-tion capabilities. This motivates us to explore whether the best of the two worlds can be achieved by any other method. Fine-tuning CLIP with region supervision. One stream of work [21; 22] attempted to improve region representation capability of a pretrained CLIP by fine-tuning it with region supervision, which has demonstrated improved performance for open-vocabulary object detection. This motivates us to study if this also enhances CLIP as a visual tokenizer. We mainly investigated RegionCLIP [21] and Owl-ViT [22]. The former finetunes a CLIP with region-level supervision from bounding boxes generated by a pretrained RPN, while the latter utilizes the region annotation from an object detection dataset. We compared these methods with CLIP, and show the results in Table 2. It can be observed that, without joint tuning the visual tokenizer, both RegionCLIP and Owl-ViT show severe performance drop on image captioning and VQA, indicating the rich semantics in the original CLIP is lost during their region fine-tuning process. On the other hand, when the visual tokenizers are jointly tuned on the instruction-tuning dataset, their fine-grained representation capability improves by a margin (on OC and MCI performance), but this can not justify the loss of semantic representation capability, resulting in inferior overall performance compared to the original CLIP. Semantic Feature as Region Supervision. Another stream of work utilized CLIP's patch feature as region-level supervision for pretraining, aiming to obtain a model with both strong semantics and better region representations. Specifically, EVA [23] and MVP [23] use CLIP's patch feature as regression target for Masked Image Modeling (MIM) pretraining, while FD [24] does not employ the masking strategy and directly distills CLIP's patch feature into a new model. We compared these methods in Table 3. Without jointly tuning the visual tokenizer, FD results in performance improvement on both semantic and fine-grained visual understanding upon CLIP. However, when patch masking strategy is adopted, the performance of EVA significantly drops. This can be attributed to the introduction of the [MASK] token for MIM, which is only used for pretraining the visual tokenizer but discarded afterwards. In this way, the train-test mismatch arises without tuning the visual tokenizer, leading to unsatisfactory performance for downstream tasks. On the other hand, when the visual tokenizer is jointly optimized with the instruction data, they are inferior to the original CLIP on VQA and image captioning, indicating semantic loss occurs.\nGiven the fact that modern MLLMs are often trained on high-quality and small-scale instruction datasets [7; 8], our observation suggests that visual tokenizer should be frozen to maintain the powerful semantic representation capability from large-scale pretraining. Nonetheless, for visual tokenizers pretrained with MIM, the introduction of the [MASK] token inevitably leads to train-test mismatch, necessitating it to be jointly tuned on the instruction data. This contradiction indicates that mask-based pretraining may not lead to a good visual tokenizer under MLLM's framework."
},
{
"figure_ref": [
"fig_1"
],
"heading": "MLLM with Good Visual Tokenizer",
"publication_ref": [
"b35",
"b31",
"b32",
"b25"
],
"table_ref": [],
"text": "In this work, we tune a new visual tokenizer which unifies the advantage of semantic representation, fine-grained visual perception and semantic maintenance capabilities. Based on the insights above, we achieve this objective by utilizing a visual tokenizer pretrained on large-scale datasets, and properly integrate it with patch-level supervision. Furthermore, we does not use any mask-based strategy, so as the rich semantics could be preserved by freezing it in the instruction tuning process. Specifically, we take the powerful EVA-CLIP [36] based on ViT-L as the teacher model, and randomly initialize another model with identical architecture as student. The patch features from the teacher model is normalized by a whitening operation, and is taken as regression target for the student model. Afterwards, the visual tokenizer can be used for MLLMs and kept frozen during instruction tuning.\nBased on the tuned visual tokenizer, we construct a new MLLM with Good Visual Tokenizer (GVT).\nThe framework of GVT is shown in Figure 2. Following [32], we also random initialize a Receiver Resampler [33] with 32 learnable queries to attend to the features from the visual tokenizer. Then, the features output from Perceiver Resampler are taken as soft prompts, and are fed into the LLM together with the language prompts. In this work, we choose the instruction-tuned Vicuna-7B [26] as the LLM. The whole model is trained by the language modeling loss, and only the Perceiver Resampler is optimized in this process."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Experimental Setup",
"publication_ref": [
"b36",
"b37",
"b38",
"b28",
"b39",
"b40",
"b41",
"b7",
"b42"
],
"table_ref": [],
"text": "We train our model on a joint dataset of image-text pairs, including CC3M [37], SBU [38], Visual Genome [39] and MS-COCO [29]. We formulate these datasets as image captioning task, and use \"what does the image describe?\" as prompt during training. Besides, we also use two object detection datasets -Object365 [40] and OpenImagesV6 [41] to design a set of object-centric tasks following [42]. The LLaVA-150k [8] dataset is also utilized for joint training. This results in a total of 15M image-text pairs. The images are resized to 224 × 224, and we adopt random resized crop and horizontal flipping for data augmentation during training. The model is trained for 50k steps with 2k steps for linear warmup. We use AdamW [43] optimizer with a learning rate of 1e-4 and batch size 1024. The training process takes about 2 days on 32 Tesla V100 GPUs. For more implementation details, please refer to our appendix."
},
{
"figure_ref": [],
"heading": "Comparison with State-of-the-art Methods",
"publication_ref": [
"b27",
"b28",
"b43",
"b31",
"b4",
"b9",
"b7",
"b6",
"b44"
],
"table_ref": [
"tab_4"
],
"text": "Without task-specific fine-tuning, we evaluate GVT on our GVTBench, which includes VQA [28],\nImage Captioning [29], Object Counting (OC) and Multi-Class Identification (MCI). Besides evaluating OC and MCI on MS-COCO validation set, we also evaluate these two tasks based on the validation set of the VCR dataset [44]. We compared our method with recent MLLMs, including Flamingo [32], BLIP-2 [5], KosMos-1 [10], LLaVa [8], miniGPT4 [7]. We evaluate open-sourced models under our GVTBench and use reported results for others. The results are shown in Table 4.\nOn these tasks, our GVT achieves the best overall performance across competitors. Specifically, on tasks requiring fine-grained visual perception, i.e., OC and MCI on both COCO and VCR datasets, GVT surpasses models with larger visual tokenizer and more curated data. This indicates our visual tokenizer can better capture the fine-grained visual information, providing representations with better details. For semantic understanding tasks, GVT achieves the second best with an accuracy of 60.4 on VQA. This result is only inferior to BLIP-2, which utilized a much larger instruction dataset with high-quality image captions filtered by [45]. On image captioning task, our GVT achieves the highest SPICE score and second best CIDEr, showing it also has strong semantic understanding capability. "
},
{
"figure_ref": [],
"heading": "Ablation Study",
"publication_ref": [
"b35",
"b35"
],
"table_ref": [
"tab_1",
"tab_5",
"tab_6"
],
"text": "We adopt the training protocol in Section 2 to study the design of our GVT.\nChoice of Distillation Target. According to the results in Table 1, we observe that DINOv2, which is pretrained with self-supervision on a dataset with 142M images also demonstrates good overall performance. To find the best target for feature distillation, we compared it with the CLIP model from [36], both in ViT-L architecture. The results are shown in Table 5. It can be seen that CLIP has demonstrated better overall performance, which can be attributed to their large-scale pretraining dataset and advanced training strategies [36].\nNumber of Latent Queries. We study the number of latent queries in the Perceiver Resampler. The results are shown in Table 6. It can be observed that the overall performance generally increases with the number of latent queries, where 32 query results in a satisfactory performance. Besides, increasing the number of query to 64 leads to limited improvements. "
},
{
"figure_ref": [
"fig_3"
],
"heading": "Qualitative Results",
"publication_ref": [],
"table_ref": [],
"text": "We show some qualitative comparison on OC and MCI between our GVT and BLIP-2 in Figure 3. It can be observed that our method demonstrate better fine-grained visual understanding capabilities than the baseline method. Take the first example in OC as an example, our method not only recognize the 3 people in the foreground, but also takes the fourth person who is far away from the camera into consideration. Besides, GVT also successfully recognize non-salient or small-sized objects in the image, such as the mouse, bicycle and broccoli in the three examples in MCI.\n5 Related Work"
},
{
"figure_ref": [],
"heading": "Multimodal Large Language Models",
"publication_ref": [
"b7",
"b32",
"b5",
"b32",
"b4",
"b9",
"b6",
"b25",
"b4",
"b7",
"b46",
"b8",
"b47"
],
"table_ref": [],
"text": "LLMs have demonstrated strong capabilities for various downstream tasks without task-specific finetuning. Based on this, recent work has utilized it to accomplish vision and language tasks, enabling powerful Multimodal Large Language Models. The common practice uses a visual tokenizer to encode the image, followed by potential bridges such as MLP [8] or Perceiver Resampler [33] to encode them into soft prompts. For example, Flamingo [6] adopts a contrastive pre-trained visual tokenizer, followed by a Perceiver Resampler [33] to aggregate the image tokens into fixed length. These tokens are fed into a frozen language model with the help of gated attention attached to transformer blocks. BLIP-2 [5] tokenizes the image with a pretrained CLIP [21; 36], which is later input into the language model with the bridge of an attention-based Q-former. Instead of freezing the language model, Kosmos-1 [10] freezes the visual tokenizer while trains the language model from scratch with large-scale text and image-text data.\nRecently, with the open source of Large Language Models [2; 26; 3; 46], a lot of large multimodal models are constructed based on them. Mini-GPT4 [7] is built on the instruction-tuned Vicuna [26] and the visual encoder from BLIP-2 [5], with only a linear layer trained to bridge the two modules. This simple design results in a powerful multi-modal chatbot, with noticeable vision-language understanding capability. LLaVa [8] adopts CLIP as visual tokenizer, and trains the projector with a curated dataset with balanced concepts. The model then can be finetuned for downstream tasks, e.g., ScienceQA [47]. Apart from using frozen visual tokenizer, mPLUG-OWL [9] tunes the Perceiver Resampler with large-scale image-text data in the first stage, followed by the finetuning of language model with LoRA [48] in the second stage. Although these generalist models have demonstrated impressive capability on multimodal tasks, we find that they mostly focus on the semantic under-standing of the image, ignoring more fine-grained visual perception. To tackle this incapability, we tune a new visual tokenizer with better fine-grained visual perception capabilities to further advance MLLMs as generalists."
},
{
"figure_ref": [],
"heading": "Visual Tokenizer Pretraining",
"publication_ref": [
"b34",
"b48",
"b28",
"b51",
"b52",
"b53",
"b17",
"b16",
"b14",
"b35"
],
"table_ref": [],
"text": "Visual encoders have been shown to benefit from large-scale pretraining for downstream tasks. The most common approach first pretrains the model on a large dataset with annotations, e.g., Ima-geNet [35], and finetunes it for downstream tasks such as semantic segmentation [49] and object detection [29]. Recently, self-supervised pre-training have also shown to improve model's representation capability. The typical contrastive-based methods [19; 50; 51] trains the model by aligning views from the same image. Inspired by the idea of mask-language-modeling for pretraining language models [52], masked-image-modeling has also evolved for visual encoder pretraining. These methods mask a proportion of image patches before feeding them into the model, and ask the model to recover the masked patches. Some methods [53] discretize the masked patches via a pretrained tokenizer [54], and ask the model to find the ID of the masked patch during pretraining. Besides, the momentum update of a model itself can also be used as an effective tokenizer [55; 20]. Recently, auto-encoder based [18] methods ask the model to directly generate the masked patch in the continuous space. Another stream of visual encoders is pretrained on massive image-text pairs via contrastive learning. The most typical model CLIP [17] has been shown to be capable of various downstream tasks in zero-shot manner. It has also evolved with more training data [15] and better optimization strategy [36].\n6 Discussions"
},
{
"figure_ref": [],
"heading": "Potential Societal Impacts",
"publication_ref": [],
"table_ref": [],
"text": "Potential Positive Impacts. In this work, we systematically investigated various visual pretraining methods under MLLM's framework. Our findings may further motivate researchers in the community to design new visual pretraining algorithms.\nPotential Negative Impacts. The training process of large models often requires huge computation resources, which consumes a lot of energy and can exacerbate the emission of carbon dioxide. Furthermore, the training datasets may contain harmful contents, leading to biased prediction or harmful generation."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [],
"table_ref": [],
"text": "In this work, our investigations are mostly based on released checkpoints, aiming to provide a guideline for researchers to select visual tokenizer accordingly. Given that these models can be pretrained with different dataset and protocol, a more in-depth study could be performed by fully aligning their training procedure."
},
{
"figure_ref": [],
"heading": "Conclusion and Future Work",
"publication_ref": [],
"table_ref": [],
"text": "We comprehensively studied various visual tokenizers through the lens of MLLM. Our investigation reveals that i) fully/weakly supervised models perform better than self-supervised ones on semantic representation, but this gap can be narrowed by scaling up pretraining dataset. ii) Self-supervised models are better at fine-grained visual perception, where patch-level supervision is particularly effective. iii) jointly tuning the visual tokenizer on the small-scale instruction dataset leads to the loss of rich semantics from large-scale pretraining. Based on these findings, we hope to find a visual tokenizer that excels at both semantic understanding and fine-grained visual perception. We reviewed existing methods and find that directly fine-tuning CLIP with region-supervision does not lead to a versatile visual tokenizer. Besides, the mask strategy for pretraining is not suitable due to the train-test mismatch. "
},
{
"figure_ref": [],
"heading": "A. Implementation Detail A.1 Implementation Detail for Empirical Studies",
"publication_ref": [
"b28",
"b37",
"b36",
"b38",
"b39",
"b40",
"b42"
],
"table_ref": [],
"text": "For the experiments in empirical studies, we use a combination of 1) image captioning datasets: MS-COCO [29], SBU [38], CC-3M [37] and Visual Genome [39] and 2) two object detection datasets, including Object365 [40] and OpenImagesV6 [41]. For image captioning data, we take the question \"what does the image describe?\" as input prompt and ask the model to generate the descriptions. For object detection datasets, we use a total of 6 tasks to fully utilize the rich annotations. Please refer to Section D in this appendix for more details. The training dataset is uniformly sampled during training. We optimize the model with a learning rate of 1e-4 and a batch size 1024. The whole model is optimized by the AdamW [43] optimizer and we set β 1 to 0.9 and β 2 to 0.98. We train the model for 10k steps, while the learning rate is linearly warmuped from 0 in the first 1k steps, and is cosine decayed to 0 afterwards. We optimize all models using float16."
},
{
"figure_ref": [],
"heading": "A.2 Implementation Detail for GVT",
"publication_ref": [
"b7"
],
"table_ref": [],
"text": "The implementation detail of our GVT is similar to that in the empirical studies, except that we use more data and more training steps. Besides the image captioning and object detection dataset, we also used LLaVa-150k dataset [8], which is generated by external powerful LLM. We trained the model for 50k steps, with 2k steps for linear warmup. Then, we use cosine decay to decrease the learning rate to 0."
},
{
"figure_ref": [],
"heading": "A.3 Evaluation Details.",
"publication_ref": [],
"table_ref": [],
"text": "VQA. Modern Language Models mainly generate one or multiple sentences, making it infeasible to directly evaluate the MLLMs in the standard evaluation protocol which requires the prediction and ground truth to be exactly matched. As such, we slightly relax the original evaluation protocol. We use the first sentence generated my MLLM as prediction result, and treated it as correct if contains the ground truth answer.\nImage Captioning. When MLLMs generate multiple sentences, we use the first sentence as the captioning result for evaluation. Since MLLMs tend to generate multiple sentences, we use the prompt \"Describe this image in a sentence: This is an image of\" as prompt to condense the prediction for effective evaluation.\nObject Counting. We extract the number of word from the first generated sentence, and compare it with ground truth number.\nObject Existence. We extract \"yes\" or \"no\" from the first generated sentence, and compare it with ground truth. "
},
{
"figure_ref": [],
"heading": "B. Benchmarking Fine-Grained Visual Understanding Tasks",
"publication_ref": [
"b28",
"b43"
],
"table_ref": [
"tab_8"
],
"text": "We provide the details of the dataset used for evaluation in each task in Table 7. In this work, we constructed two fine-grained perception tasks: object counting and object existence based on instancelevel annotations from existing datasets. Specifically, they are constructed on MS-COCO [29] and VCR [44] validation datasets. We provide their details as follows."
},
{
"figure_ref": [],
"heading": "B.1 Object Counting",
"publication_ref": [],
"table_ref": [],
"text": "Besides the visual features, the prompt of this task -\"Question: How many {obj} are there in the image? Answer:¨is fed into the MLLM for evaluation. We select the object name {obj} from the object list of the dataset. Since there are often a single object of a certain class in one image, we select a maximum of 3 objects with highest occurrence in the image to make this benchmark challenging. Similar to object counting benchmarks, we report Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). Furthermore, we also report accuracy which treats the counting as a classification problem during evaluation. Both COCO-OC and VCR-OC contain a total of 10k tasks."
},
{
"figure_ref": [],
"heading": "B.2 Multi-Class Identification",
"publication_ref": [],
"table_ref": [],
"text": "Multi-label classification can be used as task to evaluate the model's multi-instance understanding capability. However, given the open-ended nature of language models, the evaluation process is not stable since the language model may generate more fine-grained object names than the dataset categories, making a stable and fair evaluation difficult. To this end, we change the format of this task and make the evaluation process more stable. We design the prompt as \"Question: Does {obj} exist in the image?\" Answer:¨, and the model is expected to answer \"Yes\" or \"No\". We select the object name {obj} from the object list of the dataset. For each image, we randomly select at most 3 objects that exist in the image, and the same number of objects that does not appear in the image, so as to make the evaluation set balanced. Both COCO-MCI and VCR-MCI contain a total of 10k tasks."
},
{
"figure_ref": [],
"heading": "C. More Fine-grained Visual Understanding Results",
"publication_ref": [],
"table_ref": [
"tab_9",
"tab_10",
"tab_11"
],
"text": "In this section, we provide more detailed results on our two new tasks: OC and OCI.\nDetailed Object Counting Results. We show the detailed results of Object Counting task on MS-COCO in Table 8. It can be observed that, when the images contains relatively small number of objects (1-3), all methods can understand the number of objects to some extend, where ours is significantly better than others. However, when the images become more complex, where the number of occurrence increases (4-6, 7-9), the performance has significantly dropped. Similar trend can also observed in Table 9. These results demonstrate that current MLLMs still struggle at correctly counting the objects, indicating future research are required to make them more capable of challenging visual understanding tasks. Detailed Multi-Class Identification Results. We provide more detailed results on MCI task for MS-COCO in Table 10. The performance of all methods decrease when the image becomes more complex (with more objects in the image). However, the results on the VCR dataset does not show a stable trend. We conjecture this can be related to the difference on the instruction tuning datasets, which leads the model to focus on different types of objects. "
},
{
"figure_ref": [],
"heading": "D. Object-centric Tasks",
"publication_ref": [
"b41"
],
"table_ref": [],
"text": "The work of [42] has proposed 4 tasks to utilize object detection dataset for vision-language pretraining, including:\n1. List Objects Input: \"List all objects\" Output: \"{obj1}, {obj2}, ...\""
},
{
"figure_ref": [],
"heading": "Object Existence",
"publication_ref": [],
"table_ref": [],
"text": "Input: \"Does {obj} exist in the image?\" Output: \"Yes/No.\""
},
{
"figure_ref": [],
"heading": "Group Existence",
"publication_ref": [],
"table_ref": [],
"text": "Input: \"Does all of {obj1}, {obj2} and {obj3} exists in the image?\" Output: \"Yes/No.\""
},
{
"figure_ref": [],
"heading": "Existence Selection",
"publication_ref": [],
"table_ref": [],
"text": "Input: \"Which of {obj1}, {obj2}, {obj3} exist in the image?\" Output: \"{obj1/2/3}\"\nTo further utilize the rich annotations in object detection datasets, we also design two tasks which facilitate the model's learning on fine-grained visual information."
},
{
"figure_ref": [],
"heading": "Object Counting",
"publication_ref": [],
"table_ref": [],
"text": "Input: \"How many {obj}s are there in the image?\" Output: 1-9. Task 6 is only performed when the selected {obj1} and {obj2} are unique in the image, so as to avoid the referring ambiguity problem. For all tasks, we use the input text as the prompt and ask the model to generate the output text. The loss is only computed on the output texts. For each image, the task is uniformly sampled on the two object detection datasets [40; 41]."
},
{
"figure_ref": [],
"heading": "",
"publication_ref": [],
"table_ref": [],
"text": "//github.com/TencentARC/GVT"
},
{
"figure_ref": [],
"heading": "",
"publication_ref": [],
"table_ref": [],
"text": "* This work is done during Guangzhi's internship at ARC Lab, Tencent PCG † Project lead."
}
] | We empirically investigate proper pre-training methods to build good visual tokenizers, making Large Language Models (LLMs) powerful Multimodal Large Language Models (MLLMs). In our benchmark, which is curated to evaluate MLLM's visual semantic understanding and fine-grained perception capabilities, we discussed different visual tokenizers pre-trained with dominant methods (i.e., DeiT, CLIP, MAE, DINO), and observe that: i) Fully/weakly supervised models capture more semantics than self-supervised models, but the gap is narrowed by scaling up the pre-training dataset. ii) Self-supervised models are better at finegrained perception, where patch-level supervision is particularly effective. iii) Tuning the visual tokenizer leads to the loss of semantics obtained from large-scale pretraining, which is unfavorable with relatively small-scale instruction-tuning dataset. Given the findings, we reviewed methods that attempted to unify semantics and fine-grained visual understanding, e.g., patch-level feature distillation with semantically-rich targets. We obtain an intriguing insight: mask-based strategies that were once all the rage may not be applicable for obtaining good visual tokenizers. Based on this critical observation, we obtain a new MLLM equipped with a tailored Good Visual Tokenizer -GVT, which exhibits strong visual comprehension capability at multiple scales. In particular, without introducing extra parameters and task-specific fine-tuning, GVT achieves superior performance on visual question answering, image captioning, and other fine-grained visual understanding tasks such as object counting and multi-class identification. Project released at: https: | What Makes for Good Visual Tokenizers for Large Language Models? | [
{
"figure_caption": "Figure 1 :1Figure 1: Different tasks require visual understanding of different perspectives. Mainstream visionlanguage tasks, e.g., (a) VQA and (b) Image Captioning mainly focus on semantic understanding of the image. In this work, we also study two fine-grained visual understanding tasks: (c) Object Counting (OC) and (d) Multi-Class Identification (MCI).",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure2: Framework of our GVT. We first distill the features of a pretrained CLIP via smoothed L 1 loss. Then, we use it to encode images into a set of tokens, which are fed into the Perceiver Resampler[33] as soft prompts. Together with language instructions, these prompts are fed into LLM to generate responses. Only the Perceiver Resampler is optimized in this process.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Q: 4 BLIP- 2 : 3 Q: 4 BLIP- 2 : 2 Q: 3 BLIP- 2 : 2 Q:423422322How many people are there in the image? OURS: Does mouse exist in the image? OURS: Yes BLIP-2: No Q: How many people are there in the image? OURS: How many computers are there in the image? OURS: Does bike exist in the image? OURS: Yes BLIP-2: No Q: Does broccoli exist in the image?",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "423422322",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: Qualitative Comparison on OC and MCI. Our method shows better performance on recognizing detailed clues in the image.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "6 .6Spatial Relation Input: \"What is the spatial relation between {obj1} and {obj2}? Choose one from Top/Top Left/Left/Bottom Left/Bottom/Bottom Right/Right/Top Right\" Output: \"Top/Top Left/Left/Bottom Left/Bottom/Bottom Right/Right/Top Right\"",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "To effectively evaluate MLLM's visual understanding capacity at different levels, we curate a new benchmark (GVTBench) which includes both semantic understanding tasks (VQA and image captioning) as well as fine-grained visual understanding tasks (Object Counting and Multi-Class Identification). Based on GVTBench, we perform extensive experiments to study what makes for a good visual tokenizer for MLLMs and make three main observations.• We reviewed methods that combine CLIP with fine-grained supervision to see if they can achieve the best of both worlds in terms of visual semantics and fine-grained understanding. We found that the SOTA pre-trained models (i.e., EVA) are inapplicable due to the train-test mismatch caused by MIM. Such mask-based visual tokenizers rely on further tuning with instructions, which leads to the loss of pre-trained rich semantics.",
"figure_data": "• Based on the insights, we tailor a new visual tokenizer by distilling the patch-level seman-tics of a pre-trained CLIP without masking. With our visual tokenizer and Vicuna [26], wearrive at a superior MLLM (GVT) with strong visual understanding capability, achievingstate-of-the-art performance on our curated benchmark.",
"figure_id": "tab_0",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Comparison of visual tokenizers of ViT-B with different pretraining strategies. The best result is bold while the second best is underlined.",
"figure_data": "Joint TuningSupervisedVisual Tokenizer# Pretraining VQA Images AccCaptioning CIDEr SPICE Acc OCMCI Avg AccFullyDeiT [16]1.28 M48.365.815.937.5 83.6 58.8×SelfDINO [19] MAE [18]1.28 M 1.28 M50.1 48.445.0 37.313.5 11.846.5 80.8 55.6 47.5 82.7 53.4DINOv2 [20]142 M51.367.916.147.0 86.0 63.1WeaklyCLIP [17]400 M52.269.316.642.5 86.0 62.5FullyDeiT [16]1.28 M50.738.410.041.0 86.9 54.3DINO [19]1.28 M47.354.114.544.5 86.6 58.1SelfMAE [18]1.28 M48.948.014.247.5 88.7 58.2DINOv2 [20]142 M50.549.613.043.5 84.1 56.9WeaklyCLIP [17]400 M47.764.215.445.5 88.0 61.4",
"figure_id": "tab_1",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Comparison of Visual region supervised methods and CLIP.",
"figure_data": "Joint Tuning Tokenizer VisualVQA AccCaptioning CIDEr SPICE Acc OCMCI Avg Acc×CLIP [17]52.269.316.642.5 86.0 62.5×RegionCLIP [21]48.728.510.341.0 86.0 51.5×Owl-ViT [22]44.032.58.543.0 80.8 50.1CLIP [17]47.764.215.445.5 88.0 61.4RegionCLIP [21]49.765.514.147.5 86.4 62.3Owl-ViT [22]50.861.214.038.5 87.1 59.4",
"figure_id": "tab_2",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Comparison of different strategies of utilizing CLIP features with ViT-B architecture.",
"figure_data": "MethodJoint Tuning Masking PatchVQA AccCaptioning CIDEr SPICE Acc OCMCI Avg AccCLIP [17]×-52.269.316.642.5 86.0 62.5FD [24]××49.472.115.846.5 86.7 63.7EVA [23]×42.927.010.046.9 70.5 46.8CLIP [17]-47.764.215.445.5 88.0 61.4FD [24]×49.353.312.740.5 85.8 57.2EVA [23]51.461.612.345.9 87.1 61.5Pretrained VisualTokenizer (CLIP)VicunaSmoothedLossDistilled Visual TokenizerDistilled Visual TokenizerPerceive ResamplerWhat does this image describe?Feature Distillation",
"figure_id": "tab_3",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Comparison with State-of-the-arts. The best results are bold and the second best are underlined.",
"figure_data": "Model#Vis. Tok. VQA Params AccCOCO-Caption CIDEr SPICECOCO-OC COCO-MCI VCR-OC VCR-MCI Avg Acc Acc Acc AccFlamingo-9B [6]438 M51.879.4------Kosmos-1 [10]307 M51.084.716.8-----LLaVa [8]307 M39.048.315.022.252.024.666.944.7miniGPT4 [7]1.0 B58.280.619.521.576.825.170.155.4BLIP-2 [5]1.0 B62.493.317.348.081.920.268.962.5GVT (Ours)307 M60.489.919.656.289.340.378.969.2",
"figure_id": "tab_4",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Comparison of visual tokenizers under ViT-L architecture.",
"figure_data": "Visual TokenizerVQA Acc CIDEr SPICE CaptioningCOCO-OC COCO-MCI Avg Acc AccDINO-v2-Large [20] 53.969.915.045.583.663.2CLIP-Large [36]55.571.916.545.283.564.0",
"figure_id": "tab_5",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "Comparison of the number of latent queries in the Perceiver Resampler.",
"figure_data": "#Latent VQA Query Acc CIDEr SPICE CaptioningCOCO-OC COCO-MCI Avg Acc Acc853.460.015.450.078.060.31655.061.715.851.183.562.83255.571.916.545.283.564.06454.071.116.447.084.264.1",
"figure_id": "tab_6",
"figure_label": "6",
"figure_type": "table"
},
{
"figure_caption": "Based on the insights above, we tune a new visual tokenizer, which distills CLIP patch feature into a new model without masking. Equipped with our visual tokenizer, Vicuna can better understand images at multiple levels, results in superior performance on vision-language tasks including VQA, Image Captioning, Object Counting and Multi-Class Identification. For future work, we would like to explore more versatile visual tokenizer that is capable of more challenging visual understanding tasks such as open-vocabulary object detection.",
"figure_data": "",
"figure_id": "tab_7",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Dataset set statistics of our dataset for evaluation.",
"figure_data": "TaskSplitDataset# of InstanceVisual Question Answering validationVQAv2 [28]440kImage Captioningvalidation MS-COCO [29]25kObject Countingvalidation MS-COCO [29]10kObject CountingvalidationVCR [44]10kMulti-Class Identificationvalidation MS-COCO [29]10kMulti-Class IdentificationvalitdaionVCR [44]10k",
"figure_id": "tab_8",
"figure_label": "7",
"figure_type": "table"
},
{
"figure_caption": "Detailed results on the Object Counting on MS-COCO dataset. MAE ↓ RMSE ↓ Acc ↑ MAE ↓ RMSE ↓ Acc ↑ MAE ↓ RMSE ↓ Acc ↑ MAE ↓ RMSE ↓",
"figure_data": "GT range1 -34 -67 -9OverallMethod Acc ↑ MiniGPT4 23.00.961.6011.01.682.190.04.094.2421.11.362.1LLaVa26.50.891.8611.01.723.251.584.755.8322.01.362.70BLIP-261.10.470.8212.12.102.500.474.972.5748.01.152.05GVT (Ours) 74.70.250.514.72.262.490.022.295.2556.01.011.93",
"figure_id": "tab_9",
"figure_label": "8",
"figure_type": "table"
},
{
"figure_caption": "Detailed results on the Object Counting on VCR dataset. MAE↓ RMSE ↓ Acc ↑ MAE ↓ RMSE ↓ Acc ↑ MAE ↓ RMSE ↓ Acc ↑ MAE ↓ RMSE ↓",
"figure_data": "GT range1 -34 -67 -9OverallMethod Acc ↑ MiniGPT4 25.00.841.3213.01.481.820.004.344.4625.01.512.24LLaVa24.00.912.2413.31.531.991.164.464.7524.01.582.77GVT (Ours) 63.90.360.615.942.222.460.004.965.1840.01.492.41",
"figure_id": "tab_10",
"figure_label": "9",
"figure_type": "table"
},
{
"figure_caption": "Detailed results on the Multi-Class Identification on MS-COCO dataset.",
"figure_data": "#Objects1 -9 10 -20 > 20 OverallMiniGPT480.772.396.176.8LLaVa52.152.051.752.0BLIP-285.477.675.281.9GVT (Ours) 89.787.084.588.2",
"figure_id": "tab_11",
"figure_label": "10",
"figure_type": "table"
},
{
"figure_caption": "Detailed results on the Multi-Class Identification on VCR dataset.",
"figure_data": "GT range1 -9 10 -20 > 20 OverallMiniGPT471.270.271.170.8LLaVa67.166.666.866.9BLIP-267.670.370.68.9GVT (Ours) 77.180.681.578.8",
"figure_id": "tab_12",
"figure_label": "11",
"figure_type": "table"
}
] | Guangzhi Wang; Yixiao Ge; Xiaohan Ding; Mohan Kankanhalli; Ying Shan | [
{
"authors": "Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell",
"journal": "NeurIPS",
"ref_id": "b0",
"title": "Language models are few-shot learners",
"year": "2020"
},
{
"authors": "Hugo Touvron; Thibaut Lavril; Gautier Izacard; Xavier Martinet; Marie-Anne Lachaux; Timothée Lacroix; Baptiste Rozière; Naman Goyal; Eric Hambro; Faisal Azhar",
"journal": "",
"ref_id": "b1",
"title": "Llama: Open and efficient foundation language models",
"year": "2023"
},
{
"authors": "Alec Radford; Jeffrey Wu; Rewon Child; David Luan; Dario Amodei; Ilya Sutskever",
"journal": "",
"ref_id": "b2",
"title": "Language models are unsupervised multitask learners",
"year": "2021"
},
{
"authors": "Long Ouyang; Jeffrey Wu; Xu Jiang; Diogo Almeida; Carroll Wainwright; Pamela Mishkin; Chong Zhang; Sandhini Agarwal; Katarina Slama; Alex Ray",
"journal": "NeurIPS",
"ref_id": "b3",
"title": "Training language models to follow instructions with human feedback",
"year": "2022"
},
{
"authors": "Junnan Li; Dongxu Li; Silvio Savarese; Steven Hoi",
"journal": "ICML",
"ref_id": "b4",
"title": "Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models",
"year": "2023"
},
{
"authors": "Jean-Baptiste Alayrac; Jeff Donahue; Pauline Luc; Antoine Miech; Iain Barr; Yana Hasson; Karel Lenc; Arthur Mensch; Katherine Millican; Malcolm Reynolds",
"journal": "NeurIPS",
"ref_id": "b5",
"title": "Flamingo: a visual language model for few-shot learning",
"year": "2022"
},
{
"authors": "Deyao Zhu; Jun Chen; Xiaoqian Shen; Xiang Li; Mohamed Elhoseiny",
"journal": "",
"ref_id": "b6",
"title": "Minigpt-4: Enhancing visionlanguage understanding with advanced large language models",
"year": "2023"
},
{
"authors": "Haotian Liu; Chunyuan Li; Qingyang Wu; Yong Jae Lee",
"journal": "",
"ref_id": "b7",
"title": "Visual instruction tuning",
"year": "2023"
},
{
"authors": "Qinghao Ye; Haiyang Xu; Guohai Xu; Jiabo Ye; Ming Yan; Yiyang Zhou; Junyang Wang; Anwen Hu; Pengcheng Shi; Yaya Shi",
"journal": "",
"ref_id": "b8",
"title": "mplug-owl: Modularization empowers large language models with multimodality",
"year": "2023"
},
{
"authors": "Shaohan Huang; Li Dong; Wenhui Wang; Yaru Hao; Saksham Singhal; Shuming Ma; Tengchao Lv; Lei Cui; Owais Khan Mohammed; Qiang Liu",
"journal": "",
"ref_id": "b9",
"title": "Language is not all you need: Aligning perception with language models",
"year": "2023"
},
{
"authors": "Zhengyuan Yang; Linjie Li; Jianfeng Wang; Kevin Lin; Ehsan Azarnasab; Faisal Ahmed; Zicheng Liu; Ce Liu; Michael Zeng; Lijuan Wang",
"journal": "",
"ref_id": "b10",
"title": "Mm-react: Prompting chatgpt for multimodal reasoning and action",
"year": "2023"
},
{
"authors": "Danny Driess; Fei Xia; S M Mehdi; Corey Sajjadi; Aakanksha Lynch; Brian Chowdhery; Ayzaan Ichter; Jonathan Wahid; Quan Tompson; Tianhe Vuong; Yu",
"journal": "",
"ref_id": "b11",
"title": "Palm-e: An embodied multimodal language model",
"year": "2023"
},
{
"authors": "Sébastien Bubeck; Varun Chandrasekaran; Ronen Eldan; Johannes Gehrke; Eric Horvitz; Ece Kamar; Peter Lee; Yin Tat Lee; Yuanzhi Li; Scott Lundberg",
"journal": "",
"ref_id": "b12",
"title": "Sparks of artificial general intelligence: Early experiments with gpt-4",
"year": "2023"
},
{
"authors": " Openai",
"journal": "",
"ref_id": "b13",
"title": "",
"year": "2023"
},
{
"authors": "",
"journal": "",
"ref_id": "b14",
"title": "Openclip",
"year": "2023"
},
{
"authors": "Hugo Touvron; Matthieu Cord; Matthijs Douze; Francisco Massa; Alexandre Sablayrolles; Hervé Jégou",
"journal": "",
"ref_id": "b15",
"title": "Training data-efficient image transformers & distillation through attention",
"year": "2021"
},
{
"authors": "Alec Radford; Jong Wook Kim; Chris Hallacy; Aditya Ramesh; Gabriel Goh; Sandhini Agarwal; Girish Sastry; Amanda Askell; Pamela Mishkin; Jack Clark",
"journal": "",
"ref_id": "b16",
"title": "Learning transferable visual models from natural language supervision",
"year": "2021"
},
{
"authors": "Kaiming He; Xinlei Chen; Saining Xie; Yanghao Li; Piotr Dollár; Ross Girshick",
"journal": "",
"ref_id": "b17",
"title": "Masked autoencoders are scalable vision learners",
"year": "2022"
},
{
"authors": "Mathilde Caron; Hugo Touvron; Ishan Misra; Hervé Jégou; Julien Mairal; Piotr Bojanowski; Armand Joulin",
"journal": "",
"ref_id": "b18",
"title": "Emerging properties in self-supervised vision transformers",
"year": "2021"
},
{
"authors": "Maxime Oquab; Timothée Darcet; Theo Moutakanni; V Huy; Marc Vo; Vasil Szafraniec; Pierre Khalidov; Daniel Fernandez; Francisco Haziza; Alaaeldin Massa; Russell El-Nouby; Po-Yao Howes; Hu Huang; Vasu Xu; Shang-Wen Sharma; Wojciech Li; Mike Galuba; Mido Rabbat; Nicolas Assran; Gabriel Ballas; Ishan Synnaeve; Herve Misra; Julien Jegou; Patrick Mairal; Armand Labatut; Piotr Joulin; Bojanowski",
"journal": "",
"ref_id": "b19",
"title": "Dinov2: Learning robust visual features without supervision",
"year": "2023"
},
{
"authors": "Yiwu Zhong; Jianwei Yang; Pengchuan Zhang; Chunyuan Li; Noel Codella; Liunian Harold Li; Luowei Zhou; Xiyang Dai; Lu Yuan; Yin Li",
"journal": "",
"ref_id": "b20",
"title": "Regionclip: Region-based language-image pretraining",
"year": "2022"
},
{
"authors": "Matthias Minderer; Alexey Gritsenko; Austin Stone; Maxim Neumann; Dirk Weissenborn; Alexey Dosovitskiy; Aravindh Mahendran; Anurag Arnab; Mostafa Dehghani; Zhuoran Shen; Xiao Wang; Xiaohua Zhai",
"journal": "",
"ref_id": "b21",
"title": "Simple open-vocabulary object detection with vision transformers",
"year": "2022"
},
{
"authors": "Yuxin Fang; Wen Wang; Binhui Xie; Quan Sun; Ledell Wu; Xinggang Wang; Tiejun Huang; Xinlong Wang; Yue Cao",
"journal": "CVPR",
"ref_id": "b22",
"title": "Eva: Exploring the limits of masked visual representation learning at scale",
"year": "2023"
},
{
"authors": "Yixuan Wei; Han Hu; Zhenda Xie; Zheng Zhang; Yue Cao; Jianmin Bao; Dong Chen; Baining Guo",
"journal": "",
"ref_id": "b23",
"title": "Contrastive learning rivals masked image modeling in fine-tuning via feature distillation",
"year": "2022"
},
{
"authors": "Longhui Wei; Lingxi Xie; Wengang Zhou; Houqiang Li; Qi Tian",
"journal": "",
"ref_id": "b24",
"title": "Mvp: Multimodality-guided visual pre-training",
"year": "2022"
},
{
"authors": " Fastchat; Vicuna",
"journal": "",
"ref_id": "b25",
"title": "",
"year": "2023"
},
{
"authors": "Ali Farhadi; Mohsen Hejrati; Mohammad Amin Sadeghi; Peter Young; Cyrus Rashtchian; Julia Hockenmaier; David Forsyth",
"journal": "",
"ref_id": "b26",
"title": "Every picture tells a story: Generating sentences from images",
"year": "2010"
},
{
"authors": "Yash Goyal; Tejas Khot; Douglas Summers-Stay; Dhruv Batra; Devi Parikh",
"journal": "",
"ref_id": "b27",
"title": "Making the v in vqa matter: Elevating the role of image understanding in visual question answering",
"year": "2017"
},
{
"authors": "Tsung-Yi Lin; Michael Maire; Serge Belongie; James Hays; Pietro Perona; Deva Ramanan; Piotr Dollár; C Lawrence; Zitnick ",
"journal": "",
"ref_id": "b28",
"title": "Microsoft coco: Common objects in context",
"year": "2014"
},
{
"authors": "C Lawrence Ramakrishna Vedantam; Devi Zitnick; Parikh",
"journal": "",
"ref_id": "b29",
"title": "Cider: Consensus-based image description evaluation",
"year": "2015"
},
{
"authors": "Peter Anderson; Basura Fernando; Mark Johnson; Stephen Gould",
"journal": "",
"ref_id": "b30",
"title": "SPICE: semantic propositional image caption evaluation",
"year": "2016"
},
{
"authors": "",
"journal": "",
"ref_id": "b31",
"title": "ml_foundations",
"year": "2023"
},
{
"authors": "Andrew Jaegle; Felix Gimeno; Andy Brock; Oriol Vinyals; Andrew Zisserman; Joao Carreira",
"journal": "",
"ref_id": "b32",
"title": "Perceiver: General perception with iterative attention",
"year": "2021"
},
{
"authors": "Alexey Dosovitskiy; Lucas Beyer; Alexander Kolesnikov; Dirk Weissenborn; Xiaohua Zhai; Thomas Unterthiner; Mostafa Dehghani; Matthias Minderer; Georg Heigold; Sylvain Gelly",
"journal": "",
"ref_id": "b33",
"title": "An image is worth 16x16 words: Transformers for image recognition at scale",
"year": "2021"
},
{
"authors": "Olga Russakovsky; Jia Deng; Hao Su; Jonathan Krause; Sanjeev Satheesh; Sean Ma; Zhiheng Huang; Andrej Karpathy; Aditya Khosla; Michael Bernstein",
"journal": "IJCV",
"ref_id": "b34",
"title": "Imagenet large scale visual recognition challenge",
"year": "2015"
},
{
"authors": "Quan Sun; Yuxin Fang; Ledell Wu; Xinlong Wang; Yue Cao",
"journal": "",
"ref_id": "b35",
"title": "Eva-clip: Improved training techniques for clip at scale",
"year": "2023"
},
{
"authors": "Piyush Sharma; Nan Ding; Sebastian Goodman; Radu Soricut",
"journal": "",
"ref_id": "b36",
"title": "Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning",
"year": "2018"
},
{
"authors": "Yago Tomás; Le Vicente; Chen-Ping Hou; Minh Yu; Dimitris Hoai; Samaras",
"journal": "",
"ref_id": "b37",
"title": "Large-scale training of shadow detectors with noisily-annotated shadow examples",
"year": "2016"
},
{
"authors": "Ranjay Krishna; Yuke Zhu; Oliver Groth; Justin Johnson; Kenji Hata; Joshua Kravitz; Stephanie Chen; Yannis Kalantidis; Li-Jia Li; David A Shamma",
"journal": "IJCV",
"ref_id": "b38",
"title": "Visual genome: Connecting language and vision using crowdsourced dense image annotations",
"year": "2017"
},
{
"authors": "Shuai Shao; Zeming Li; Tianyuan Zhang; Chao Peng; Gang Yu; Xiangyu Zhang; Jing Li; Jian Sun",
"journal": "",
"ref_id": "b39",
"title": "Objects365: A large-scale, high-quality dataset for object detection",
"year": "2019"
},
{
"authors": "Alina Kuznetsova; Hassan Rom; Neil Alldrin; Jasper Uijlings; Ivan Krasin; Jordi Pont-Tuset; Shahab Kamali; Stefan Popov; Matteo Malloci; Alexander Kolesnikov",
"journal": "IJCV",
"ref_id": "b40",
"title": "The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale",
"year": "2020"
},
{
"authors": "Weicheng Aj Piergiovanni; Anelia Kuo; Angelova",
"journal": "",
"ref_id": "b41",
"title": "Pre-training image-language transformers for open-vocabulary tasks",
"year": "2022"
},
{
"authors": "Ilya Loshchilov; Frank Hutter",
"journal": "",
"ref_id": "b42",
"title": "Decoupled weight decay regularization",
"year": "2017"
},
{
"authors": "Rowan Zellers; Yonatan Bisk; Ali Farhadi; Yejin Choi",
"journal": "",
"ref_id": "b43",
"title": "From recognition to cognition: Visual commonsense reasoning",
"year": "2019"
},
{
"authors": "Junnan Li; Dongxu Li; Caiming Xiong; Steven Hoi",
"journal": "",
"ref_id": "b44",
"title": "Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation",
"year": "2022"
},
{
"authors": "Chung Hyung Won; Le Hou; Shayne Longpre; Barret Zoph; Yi Tay; William Fedus; Eric Li; Xuezhi Wang; Mostafa Dehghani; Siddhartha Brahma",
"journal": "",
"ref_id": "b45",
"title": "Scaling instruction-finetuned language models",
"year": "2022"
},
{
"authors": "Pan Lu; Swaroop Mishra; Tanglin Xia; Liang Qiu; Kai-Wei Chang; Song-Chun Zhu; Oyvind Tafjord; Peter Clark; Ashwin Kalyan",
"journal": "NeurIPS",
"ref_id": "b46",
"title": "Learn to explain: Multimodal reasoning via thought chains for science question answering",
"year": "2022"
},
{
"authors": "J Edward; Phillip Hu; Zeyuan Wallis; Yuanzhi Allen-Zhu; Shean Li; Lu Wang; Weizhu Wang; Chen",
"journal": "",
"ref_id": "b47",
"title": "Low-rank adaptation of large language models",
"year": "2022"
},
{
"authors": "Bolei Zhou; Hang Zhao; Xavier Puig; Tete Xiao; Sanja Fidler; Adela Barriuso; Antonio Torralba",
"journal": "IJCV",
"ref_id": "b48",
"title": "Semantic understanding of scenes through the ade20k dataset",
"year": "2019"
},
{
"authors": "Ting Chen; Simon Kornblith; Mohammad Norouzi; Geoffrey Hinton",
"journal": "",
"ref_id": "b49",
"title": "A simple framework for contrastive learning of visual representations",
"year": "2020"
},
{
"authors": "Xinlei Chen; Kaiming He",
"journal": "",
"ref_id": "b50",
"title": "Exploring simple siamese representation learning",
"year": "2021"
},
{
"authors": "Jacob Devlin; Ming-Wei Chang; Kenton ; Lee Kristina; Toutanova ",
"journal": "",
"ref_id": "b51",
"title": "Bert: Pre-training of deep bidirectional transformers for language understanding",
"year": "2019"
},
{
"authors": "Hangbo Bao; Li Dong; Songhao Piao; Furu Wei",
"journal": "",
"ref_id": "b52",
"title": "Beit: Bert pre-training of image transformers",
"year": "2021"
},
{
"authors": "Aditya Ramesh; Mikhail Pavlov; Gabriel Goh; Scott Gray; Chelsea Voss; Alec Radford; Mark Chen; Ilya Sutskever",
"journal": "",
"ref_id": "b53",
"title": "Zero-shot text-to-image generation",
"year": "2021"
},
{
"authors": "Jinghao Zhou; Chen Wei; Huiyu Wang; Wei Shen; Cihang Xie; Alan Yuille; Tao Kong",
"journal": "",
"ref_id": "b54",
"title": "Image bert pre-training with online tokenizer",
"year": "2022"
}
] | [] |
|
10.18653/v1/N19-1423 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b6",
"b31",
"b27",
"b20",
"b32",
"b35",
"b45",
"b47",
"b47",
"b43",
"b35",
"b16",
"b47",
"b15",
"b22",
"b43",
"b47",
"b6",
"b27",
"b18"
],
"table_ref": [],
"text": "Pre-trained language models (Devlin et al., 2019;Radford et al., 2019;Liu et al., 2019;Lewis et al., 2020;Raffel et al., 2020;Chen et al., 2022b) have shown great potential in a wide range of NLP tasks. While large language models offer unparalleled performance, their high computation during inference limits the scope of applications. More studies recently concentrate on efficient NLP, which aims to speed up the inference of deep language models without significant performance degradation (Sanh et al., 2019;Zafrir et al., 2019;Zhou et al., 2020). Among these, the multi-exit models (Zhou et al., 2020;Xin et al., 2020) attract widespread attention.\nThe idea of the multi-exit models stems from the observation that inputs with varying semantics demand distinct computational resources. By automatically adjusting different computational resources according to input semantics, one can effectively speed up the inference of a multi-exit model with minimum performance loss. Furthermore, such multi-exit model can be easily combined with other static speedup approaches, e.g., distillation (Sanh et al., 2019;Jiao et al., 2020), by replacing the backbone model. In addition to higher efficiency, previous studies also show that the multiexit models are more robust to correctness-based adversarial samples (Zhou et al., 2020;Hu et al., 2020).\nThe study of NLP attacks has mostly focused on harming models' accuracy, and taken static transformers as victim models (Ebrahimi et al., 2018b;Li et al., 2020). There exists another type of attack on the model efficiency, i.e., to make the models computationally slow. Considering this type of attack, the intrinsic dynamic nature of the multiexit models might be vulnerable to such attacks. It remains unexplored, however, how significantly the efficiency or speedup from early exiting will be affected by the attacks. Motivated by this, we first analyze the efficiency robustness of dynamic NLP transformers. We find that previous accuracyoriented approaches cannot significantly slow down the dynamic transformers and sometimes even lead to shorter inference time.\nTo this end, we propose a novel slowdown attack framework on multi-exit language models: SAME. Unlike accuracy-oriented adversarial attacks, there are several unique challenges for effective efficiency attacks. First, existing accuracy-oriented attacks aim to mislead neural networks to generate wrong predictions, which is not suitable for efficiency-oriented attacks. Therefore, we develop a new objective function to guide the generation of efficiency-oriented adversarial samples. In addition, our objective function must be general to handle various exit mechanisms in multi-exit transformers. Second, multi-exit transformers are not static during inference, so the \"static\" search strategies used in adversarial attacks are not suitable. To overcome the challenges, we propose a dynamic importance adjustment strategy that assigns different importance to each exit layer, allowing the adversarial example search process to focus on the layers that contribute to model efficiency.\nWe evaluate our SAME using two widely-used multi-exit strategies (entropy-based (Xin et al., 2020) and patience-based (Zhou et al., 2020)) with various pre-trained language models (Devlin et al., 2019;Liu et al., 2019;Lan et al., 2020) as the backbone on eight tasks from the GLUE benchmark. Experimental results show that our SAME can effectively reduce the computational saving by 80% on average, which significantly outperforms previous accuracy-oriented approaches by a large margin. Further experiments on the multi-goal attack, attacking transferability, and adversarial training convincingly validate the effectiveness and generalization ability of our proposed SAME.\nThe contributions of this work are summarised as follows: (1) New Problem: we identify a new vulnerability of the multi-exit NLP models, namely, the network efficiency. (2) Novel Approach: We propose the first efficiency-oriented attacking framework to measure the efficiency robustness of the multi-exit NLP models. (3) Comprehensive Evaluation: We conduct a systematic evaluation of various dynamic transformers, which shows that future studies on improving and protecting the efficient robustness of the multi-exit NLP models are necessary."
},
{
"figure_ref": [
"fig_0"
],
"heading": "Background 2.1 Multi-Exit Networks",
"publication_ref": [],
"table_ref": [],
"text": "Multi-exit neural networks include multiple outputs or \"exits\" placed at different network layers. This architectural design allows for early decisionmaking if the input is confidently classified or predicted, leading to faster and more efficient processing. Based on the semantic complexity of the inputs, multi-exit neural networks can effectively reduce inference time by making predictions from early layers for a simpler input and later layers for a more complex input. As shown in Figure 1, a multi-exit transformer consists of N transformer layers, each containing an internal classifier. During the inference phase, predictions are made after each layer, and computation is terminated once the exit criterion is met.\nA funny, highly enjoyable movie.\nA funny, highly enjoyable movie."
},
{
"figure_ref": [
"fig_0",
"fig_0"
],
"heading": "Patience Counter",
"publication_ref": [
"b43",
"b26",
"b47",
"b49"
],
"table_ref": [],
"text": "Entropy-based early exiting Patience-based early exiting The choice of exit criterion is crucial in multiexit models. In this work, we explore two commonly used strategies: entropy-based (Xin et al., 2020;Liu et al., 2020) and patience-based (Zhou et al., 2020;Zhu, 2021). As depicted in Figure 1 (left), the entropy-based strategy employs the entropy of a probability distribution as an indicator of model confidence. The model checks if the entropy is lower than a predefined threshold after each layer's computation and outputs a prediction when the criterion is met. The patience-based strategy, as shown in Figure 1 (right), involves maintaining a patience counter that is incremented by 1 when predictions from two consecutive internal classifiers are consistent and is reset to zero when they are inconsistent. The model exits early if the patience counter reaches a pre-defined patience threshold."
},
{
"figure_ref": [],
"heading": "Adversarial Attack",
"publication_ref": [
"b30",
"b21",
"b38",
"b19",
"b14",
"b5",
"b24",
"b24",
"b0",
"b34",
"b46",
"b5",
"b40",
"b10",
"b48",
"b13",
"b17"
],
"table_ref": [],
"text": "Adversarial attacks are methods of creating adversarial examples to cause neural networks to make incorrect predictions (Papernot et al., 2016;Ebrahimi et al., 2018b;Li et al., 2019;Wallace et al., 2019;Le et al., 2022;Hong et al., 2021;Cheng et al., 2020;Li et al., 2023;Chen et al., 2022a;Li et al., 2023). Adversarial attacks in natural language processing (NLP) mainly contain two categories: character-level and word-level. For the character-level attacks, existing methods involve modifying the words in an input sentence by using insertion, swap, or deletion operators to create adversarial examples (Belinkov and Bisk, 2018;Ebrahimi et al., 2018a). The word-level attacks, on the other hand, involve replacing words in the input sentence with other words, e.g., synonym replacement (Ren et al., 2019), round-trip translation (Zhang et al., 2021). There has also been an emergence of attacks targeting generative models. For example, Seq2Sick (Cheng et al., 2020) generates adversarial examples that decrease the BLUE score of neural machine translation models. In addition to accuracy, inference efficiency is also highly critical for various real-time applications, e.g., speech recognition (Wang et al., 2022), machine translation (Fan et al., 2021;Zhu et al., 2020), lyric transcriptions (Gao et al., 2022b(Gao et al., , 2023(Gao et al., , 2022a)). Recently, NICGSlowDown and NMT-Sloth (Chen et al., 2022d,c) propose delaying the appearance of the end token to reduce the efficiency of language generative models. There have been studies evaluating the accuracy robustness of dynamic transformer through directly adapting TextFooler (Jin et al., 2020). Unlike the previous works, the proposed SAME is specially designed for evaluating the efficiency robustness of dynamic transformers."
},
{
"figure_ref": [],
"heading": "Methodology",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Problem Formulation",
"publication_ref": [],
"table_ref": [],
"text": "Unlike previous accuracy-oriented approaches, our goal here is to create adversarial examples that decrease the efficiency of a victim multi-exit model F by adding human-unnoticeable perturbations to a benign input. Specifically, we focus on two factors: (i) significantly increasing the computational costs for the victim model and (ii) keeping the generated perturbation minimal. We formulate the problem as a constrained optimization problem:\n∆ = argmax δ Exit F (x + δ) s.t.||δ|| ≤ , (1)\nwhere x is the given benign input, is the maximum adversarial perturbation allowed, and Exit F (•) measures the number of layers where the victim multiexit language model F exits. Our proposed approach attempts to find the optimal perturbation ∆ that maximizes the number of layers where the model exits (decrease the efficiency), and at the same time adheres to the constraint that the perturbation must be smaller than the allowed threshold (unnoticeable). In this work, we set the allowable modifiable words as 10% of the total input words.\nMachine learning is interesting."
},
{
"figure_ref": [],
"heading": "Benign Input",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Dynamic LM",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Mess Loss",
"publication_ref": [],
"table_ref": [],
"text": "Patience Loss"
},
{
"figure_ref": [],
"heading": "Importance Adjustment",
"publication_ref": [],
"table_ref": [],
"text": "Machine studying is interesting. Machine leanring is interesting."
},
{
"figure_ref": [],
"heading": "Word Level Character Level",
"publication_ref": [],
"table_ref": [],
"text": "Figure 2: Design overview of SAME"
},
{
"figure_ref": [],
"heading": "Approach Overview",
"publication_ref": [],
"table_ref": [],
"text": "Figure 2 illustrates the design overview of our approach. Our approach iteratively mutates the given inputs to craft adversarial examples. During each iteration, we first design a differentiable objective to approximate our adversarial goals (Section 3.3). Then, we dynamically adjust our objective based on the importance of each layer (Section 3.4). Finally, we apply our approximated objective function to mutate the inputs with two types of perturbations and generate a set of adversarial candidates that satisfy the given unnoticeable constraints (Section 3.5)."
},
{
"figure_ref": [],
"heading": "Adversarial Objective Approximation",
"publication_ref": [],
"table_ref": [],
"text": "Notice that our optimization objective in Equation 1 is non-differentiable, which makes it challenging to be directly used as the objective for searching optimal adversarial perturbations. Thus, we need to approximate the adversarial objective (i.e., argmax Exit F (•)) with a differentiable function.\nVarious objectives are used in accuracy-based adversarial attacks, which aim to decrease the model's accuracy by increasing the confidence scores of the wrong labels. However, these existing approaches do not address the model's efficiency. Therefore, a totally new design for efficiency-oriented adversarial objectives is required. Since exiting criteria determine the model's efficiency (as outlined in Section 2.1), we motivate our efficiency-oriented adversarial objective approximation from termination criteria of F, which includes the following: Making Mess Prediction: Recall that one way to determine early exiting is by whether the entropy undercuts a predefined threshold. To make the model less efficient, our goal is to keep the entropy above this threshold consistently. It is worth noting that a uniform distribution has the highest entropy among all distributions. Hence, our first objective function is to push the model prediction close to a uniform distribution:\nL mess = N i=1 SCE(F i (x), U),(2)\nwhere F i (x) is the prediction logits at the i th layer, U is a uniform distribution, N is the total layer of the victim F, and SCE(•) is the soft cross entropy loss. Eq. 2 is interpreted as we seek to minimize the error between output logits (i.e., F i (x)) and uniform distribution to push the model to produce larger entropy.\nDecrease Prediction Patience: The second termination criterion is based on prediction patience. To this end, our second objective function needs to push the victim model to produce \"impatient\" predictions. In other words, we seek to push the model to make inconsistent predictions among its intermediate classifiers as follows:\nL patience = N i=1 CE(F i (x), h i ),(3)\nwhere h i is the constructed target label at the i th layer and CE(•) is the cross entropy function. As previously mentioned, our second objective seeks to cause the model to produce inconsistent predictions. Thus, we construct our target h i based as:\nhi = argmax(Fi(x)), hi-1 = argmax(Fi(x)) argsecond(Fi(x)), hi-1 = argmax(Fi(x)) ,(4)\nand h 0 is set as the prediction given by the model's first internal classifier on the seed input. Our intuition is to force the model to produce inconsistent predictions between consecutive classifiers by introducing heuristics (Equation 4), thus decreasing prediction patience."
},
{
"figure_ref": [],
"heading": "Dynamic Importance Adjustment",
"publication_ref": [],
"table_ref": [],
"text": "It is important to note that the inference path of F is not \"static\", implying that treating all layer outputs equally at each stage of the search may not yield optimal results. For instance, if F exits at the third layer, optimizing the input to influence the output before the third layer would be less important. To overcome this challenge, we propose a strategy to dynamically adjust the importance assigned to early layer outputs. Given an input x, our layer-wise importance scores are computed as:\nw i = α, i < Exit F (x) β i-Exit F (x) i ≥ Exit F (x) ,(5)\nwhere w i is the importance score for the i th layer, Exit F (x) is the index of layer that exit the computation, α and β are hyper-parameters. As shown in Eq. 5, the layers, which have been computed, are assigned constant importance scores, while the layers, which are not used, are assigned exponentially increasing importance scores. Finally, our objective can be expressed as:\nL total = N i=1 w i (λL i mess +(1-λ)L i patience ), (6\n)\nwhere λ is the hyper-parameters that balance the importance of each objective goals."
},
{
"figure_ref": [],
"heading": "Perturbing Inputs",
"publication_ref": [
"b25",
"b21"
],
"table_ref": [],
"text": "Our adversarial perturbation generation includes three main steps: (i) finding critical words, (ii) generating adversarial candidates, and (iii) choosing candidates.\nFinding Critical Words: As mentioned earlier, we apply our approximated objective function as guidance to search for optimal adversarial perturbations. Thus, we first find the critical words using the gradient of our objective function (i.e., Equation 6). Specifically, we order the word based on j ∂L total ∂tk j i , where tk j i is the j th dimension of the i th tokens embedding. In this step, we consider the word that is exactly tokenized into one token.\nGenerating Perturbation Candidates: After identifying the critical words, the next step is to perturb the critical words to craft adversarial perturbation candidates. In this work, we follow existing work and use two types of perturbations to generate adversarial examples: character level and word level, which leads to two variants of SAME: SAME-Char and SAME-Word correspondingly.\nFor character-level perturbation, we employ four widely used mutations: neighbor character swap, character insertion, character deletion, and homoglyph character replacement (Ebrahimi et al., 2018a;Liu et al., 2022). For neighbor character swap and deletion mutations, we randomly swap or delete one character in the targeted word. To perform character insertion mutation, we randomly select a character from the ASCII character set and then insert it at a random location in the targeted word. For homoglyph character replacement mutation, we use the default homoglyph character mapping from TextBugger (Li et al., 2019). All these four character-level perturbations are common in the real world when typing quickly and can be unnoticeable without careful examination. For each mutation, we randomly generate 25 candidates, resulting in a total of 25×4=100 candidates.\nFor word-level perturbation, we consider replacing the critical word with another word δ. To compute the target word, we define word replace increment I s,t to measure the efficiency degradation of replacing word s to t:\nI s,t = j (E(t) -E(s)) j × ∂L total (x)\n∂s j i ; δ = argmax t I s,t(7)\nwhere E(•) represents the embedding vector of a given token, and I s,t denotes the increase in the direction of the gradient of our objective function, resulting from replacing token s with token t. For word level perturbations, we also generate 100 adversarial candidates.\nCandidates Selection: Once the adversarial candidates are generated, we select the valid candidates for the next iteration. To do this, we eliminate candidates that do not meet the constraints in Equation 1 and then select the top 5 candidates with the highest Exit F for the next iteration of search."
},
{
"figure_ref": [],
"heading": "Experiment",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Experimental Setup",
"publication_ref": [
"b39",
"b43",
"b6",
"b27",
"b47",
"b6",
"b18",
"b21",
"b34",
"b17",
"b47"
],
"table_ref": [],
"text": "Datasets: We conduct our experiments on GLUE (Wang et al., 2018) benchmark. For more details about GLUE, please refer to Appendix A.\nVictim models: We evaluate two popular early-exit strategies, namely entropy-based DeeBERT (Xin et al., 2020) with backbone model BERT (Devlin et al., 2019) and RoBERTa (Liu et al., 2019), as well as patiencebased PABEE (Zhou et al., 2020) with backbone model BERT (Devlin et al., 2019) and AL-BERT (Lan et al., 2020). Following the original paper, we consider two different settings with various entropy or patience threshold. Specifically, we select the threshold to keep the relative performance drop within 2% and 4%, denoted as PD<2% and PD<4%.\nBaselines: We compare SAME to 5 recent NLP attack approaches through adapting their attacking strategy to our attacking scenario, which includes white-box attacking approaches: Hot-Flip (Ebrahimi et al., 2018b), TextBugger (Li et al., 2019), A2T (Yoo and Qi, 2021); as well as black-box ones: PWWS (Ren et al., 2019), TextFooler (Jin et al., 2020).\nMetrics: We evaluate the efficacy of attacking methods with two metrics. As in (Zhou et al., 2020), the first metric is the estimated speedup, which is computed as the total number of transformer layers divided by number of actually computed layers. Besides, we propose a second metric, high computation ratio, which refers to the ratio of samples with extremely high computational cost. Specifically, we consider samples with at least 11 computed layers as high computational samples for base-size dynamic transformers with total 12 layers, and at lease 22 computed layers as high computational samples for large-size dynamic transformers with total 24 layers. In all tables, we report the speedup (left) and high computation ratio (right) unless specified otherwise."
},
{
"figure_ref": [],
"heading": "Main Results",
"publication_ref": [],
"table_ref": [
"tab_3"
],
"text": "The comparison of different attacking methods on entropy-based dynamic models are shown in Table 1, and the results on patience-based models are listed in Table 2. Overall, we find that previous accuracy-oriented approaches cannot harm the model efficiency much for either exiting strategy, and even lead to higher speedup for some cases, e.g., QQP, RTE. In sharp contrast, both variants of SAME can effectively reduce the speedup from early exiting, which outperforms all previous approaches by a large margin. Specifically, under PD<2% setting, SAME eliminates the efficiency gain by 74.88% on average across GLUE benchmark for DeeBERT series models, and 85% for PABEE series models. Under PD<4% setting, model's exiting criteria are more relaxed, which makes the slowdown more difficult. However, SAME consistently reduces the efficiency gain by 75% for DeeBERT models, and by 82% for PABEE models, which again convincingly demonstrates the efficacy of SAME.\nBesides, while previous works show that patience-based approaches are more robust against accuracy-oriented attack, we observe that both strategies are equally vulnerable under proposed efficiency attack. In addition, these two strategies have different level of vulnerability to different permutation. Entropy-based models are more vulnerable to character-level permutation. On the contrary, word-level permutation performs better on patience-based models. We hypothesize that the discrepancy between two exiting strategies lead to this phenomena. To slowdown patience-based models, ones need to break the consistency between predictions from internal classifiers, which might be difficult to achieve with character-level permutation. The results suggest that further combining multiple level of permutetation methods would lead to a more universal attacking framework that are applicable to wide range of dynamic models.\nFinally, we find that the quality of backbone language model has large impact on the efficiency robustness of dynamic transformers. For instance, compared to BERT, RoBERTa is trained with larger corpus with longer time, which makes DeeR-oBERTa much more robust than DeeBERT models."
},
{
"figure_ref": [],
"heading": "Accuracy & Efficiency",
"publication_ref": [
"b47"
],
"table_ref": [
"tab_5"
],
"text": "Since another important adversarial goal is misclassification, we further investigate the trade-off between accuracy and efficiency drop during attacking. Table 3 summarizes the results on SST-2 and MNLI-mm. In addition to efficiency drop, SAME can also considerably lead to misclassification. As the goal function of SAME doesn't consider the accuracy metric, we further propose SAME+, which adopts a multi-objective goal function:\nExit F (x + δ) + σ × 1(F(x + δ) = y true ), (8)\nwhere y true is the ground truth label, 1(•) is the indicator function, and σ is the weight that balances the importance of accuracy and efficiency. As we focus on efficiency robustness in this work, we set σ to 0.5. Therefore, SAME+ is expected to produce adversarial samples with a similar efficiency drop level as SAME but leads to an additional accuracy drop. As shown in for SAME-word and 37.47% for SAME-char without any increase in efficiency. In addition, previous work shows that patience-based methods are more robust against accuracy-oriented adversarial attack, compared to entropy/confidence-based ones (Zhou et al., 2020). However, we observe that SAME leads to similar accuracy drop for patience-based and entropy-based dynamic models. The robustness of patience-based methods come from internal classifier ensemble. Yet, proposed heuristic loss in SAME makes these internal classifiers hard to reach an agreement. Then, the victim model will directly obtain prediction from the last classifier for large proportion of inputs, which actually fails the mechanism of internal classifier ensemble. The empirical results suggest that it's possible to craft adversarial samples with low accuracy and efficiency."
},
{
"figure_ref": [],
"heading": "Attacking Transferability",
"publication_ref": [
"b29"
],
"table_ref": [
"tab_7",
"tab_10",
"tab_12"
],
"text": "In this section, we examine whether adversarial samples from SAME are transferable between various architectures. We study two settings: (i) Cross backbone: we assume the source model and target model share the same early exiting strategy but with different backbone models. (ii) Cross mechanism: we assume that the source and target model have different early exiting strategies.\nTable 4 summarizes the results on SST-2 and MNLI datasets. Overall, the adversarial samples are transferable between different models, and several critical factors determine the transferability. The first one is the exiting strategy. We find that samples are more transferable between models sharing the same exiting strategy, e.g., from PABEE-ALBERT-base to PABEE-BERT-base. The second factor is the backbone model. If the source and target model have the same backbone language model or share the same tokenizer, e.g., DeeBERT-base and DeeBERT-large, the transferred samples will cause more slowdown. In addition, we find that entropy-based models are more vulnerable to transferred attacks compared to patience-based models. Interestingly, we again observe that character-level attack is more transferable to the entropy-based model. while the word-level attack is more transferable to the patience-based model, which is consistent with our findings from Section 4.2. Impact of modification rate: In our main results, we set the allowable modification rate as 10% of the input words. We further investigate whether SAME can reduce the inference efficiency under lower modification rate (imperceptible attack). The experiment results across GLUE benchmark on DeeBERT-base and PABEE-BERT-base under are summarized in Table 7. Even constrained with a very low modification rate, e.g., 3%, both variants of SAME can still significantly reduce the model's efficiency. In addition, with increasing modification rate, SAME leads to higher reduction in efficiency.\nAblation Study: To understand the inner mechanism of SAME, we conduct ablation studies on each component. As shown in Table 8, solely using heuristic loss can already lead to significant effi- ciency drop. In addition, using loss combination, and adding layer-wise importance weights can both further increase the high computation ratio. Finally, SAME utilizes all the sub-components, which leads to the lowest inference efficiency. Semantic Similarity: While we constrain the modification rate in our experiments to keep the semantic meaning consistent, the semantic similarity between benign and adversarial examples is not explicitly constrained. Therefore, we further investigate the sentence semantic similarity between original and adversarial examples on SST-2 dataset. Specifically, We first obtain the sentence representations of adversarial and original sample with a state-of-the-art ST5-large embedding model (Ni et al., 2022), and then compute their pairwise cosine similarity. With DeeBERTbase and PABEE-BERT-base as the victim model, the SAME-word has an average cosine similarity of 0.89, and SAME-char has an average cosine similarity 0.96. The results suggest that both variants of SAME can well preserve the inputs' semantic meaning, at the same time, reduce the efficiency of dynamic transformers.\nVisualization: To illustrate the impact of efficiency-based v.s. correctness-based adversarial perturbations, We present a case study of adversarial samples produced from SST-2 dataset in Table 9. For better explainability, we show examples with one-word only modification. Due to space limitations, more adversarial samples generated using SAME can be found in Appendix C.\nAs shown in Table 9, our efficiency-based method will perturb the word but to bujt, thereby altering the explicit turning relationship between two sentences. While humans can make the correct prediction even without the word but, it can be challenging for dynamic transformers to infer the turning relationship in the early stage. Therefore, they fail to satisfy the exiting conditions, resulting in reduced inference efficiency. In contrast, correctnessbased approaches will keep the transition word and adversarially modify the word deeper, e.g., to deper with TextBugger. With the transition word but, the model will emphasize more on the latter sentence, and easily get a high model confidence.\n[Clean input] the film may appear naked in its narrative form ... but it goes deeper than that , to fundamental choices that include the complexity of the catholic doctrine.\n[TextBugger] the film may appear naked in its narrative form ... but it goes deper than that , to fundamental choices that include the complexity of the catholic doctrine.\n[TextFooler] the film may appear naked in its narrative form ... but it goes more than that , to fundamental choices that include the complexity of the catholic doctrine.\n[SAME] the film may appear naked in its narrative form ... bujt it goes deeper than that , to fundamental choices that include the complexity of the catholic doctrine.\nTable 9: Comparison of adversarial samples produced by accuracy-oriented approaches and our energy-oriented approaches from SST-2."
},
{
"figure_ref": [],
"heading": "Conclusion and Future Works",
"publication_ref": [],
"table_ref": [],
"text": "In this paper, we systematically evaluate the efficiency robustness of dynamic transformers. We also propose SAME, a novel white-box slowdown attack framework that effectively degrade the efficient performance of dynamic multi-exit language models. Specifically, SAME generates adversarial examples that could delay the exit of dynamic multi-exit language models with the guidance of heuristic and mess loss. Extensive experimental demonstrate the superior effectiveness of SAME across various dynamic multi-exit language models. Future works include the development of efficient robust dynamic transformers and the extension to other NLP models with dynamic inference time."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [
"b23",
"b36"
],
"table_ref": [],
"text": "Firstly, our proposed SAME is for the white-box attacking scenario only, which is less practical in real-world scenarios. However, experimental results on black-box transferability show that a blackbox efficiency-oriented attack is highly feasible. Therefore, we leave the black box SAME as a future study.\nSecondly, we mainly study multi-exit transformers for sentence classification tasks in this work. We notice that several recent works extend the idea of multi-exiting to other NLP tasks, e.g., sequence labelling (Li et al., 2021), text generation (Schuster et al., 2022). For classification tasks, SAME slowdowns the models by avoiding early exiting. While for text generation tasks, in addition to avoiding early exiting, ones can also slow down the model by forcing the model to produce a longer sequence. We leave the exploration of other dynamic models to future work.\nThirdly, as the first work that evaluates the efficiency robustness of dynamic transformers. we use a relatively simple permutation strategy. Although these permutations can lead to severe performance degradation, they might not be imperceptible enough. Yet, they could be easily replaced by other sophisticated permutations under SAME framework."
},
{
"figure_ref": [],
"heading": "Ethics Statement",
"publication_ref": [],
"table_ref": [],
"text": "We propose a slowdown attack against dynamic transformers on GLUE benchmark datasets in this work. We aim to study the efficiency robustness of dynamic transformers and provide insight to inspire future works on robust dynamic transformers.\nOur proposed framework may be used to attack online NLP services deployed with dynamic models. However, we believe that exploring this new type of vulnerability and robustness of efficiency is more important than the above risks. Research studying effective adversarial attacks will motivate improvements to the system security to defend against the attacks."
},
{
"figure_ref": [],
"heading": "A Experiment Setup",
"publication_ref": [
"b41",
"b37",
"b42",
"b33",
"b39",
"b7",
"b28"
],
"table_ref": [],
"text": "We conduct our experiments on 8 tasks from GLUE, including CoLA (Warstadt et al., 2019), SST-2 (Socher et al., 2013), MNLI(-mm) (Williams et al., 2018), QNLI (Rajpurkar et al., 2016), QQP2 , RTE (Wang et al., 2018), MRPC (Dolan and Brockett, 2005). For large datasets, i.e., QNLI, QQP, MNLI(-mm), we randomly sample 1000 samples from validation set for attacking experiments. For the rest, we use the whole validation set. For all dynamic victim models, We train the model with publicly available code from huggingface transformers3 with the default hyper-parameter (search). We use the implementation from TextAttack (Morris et al., 2020) for baselines. For SAME, we generate 100 mutant candidates for each iteration. All of our experiments are conducted on a Ubuntu 20.04 server with 8 RTX A5000 GPUs. One attacking experiment on BERT-base takes around 1.5 GPU hours."
},
{
"figure_ref": [],
"heading": "B Results on Large Dynamic Language Models",
"publication_ref": [],
"table_ref": [],
"text": "We further conduct the experiments on large dynamic transformers with backbone model RoBERTa-large, ALBERT-large, and BERT-large. "
},
{
"figure_ref": [],
"heading": "C Visualization of our generated adversarial examples",
"publication_ref": [],
"table_ref": [
"tab_1"
],
"text": "We visualize several adversarial examples our proposed attack method generates from SST-2 in Table 11. By only replacing a few words in the benign input, our method could significantly delay the exit of dynamic multi-exit language models."
},
{
"figure_ref": [],
"heading": "SAME-Word",
"publication_ref": [],
"table_ref": [],
"text": "[Clean input] although german cooking does not come readily to mind when considering the world 's best cuisine , mostly martha could make deutchland a popular destination for hungry tourists .\n[Adv. input] although german cooking does not come readily no mind when considering akin world 's best cuisine , mostly martha could make deutchland rats popular destination for hungry tourists .\n[Clean input] a difficult , absorbing film that manages to convey more substance despite its repetitions and inconsistencies than do most films than are far more pointed and clear .\n[Adv. input] a difficult , absorbing film robots manages to convey more substance despite its repetitions and inconsistencies heart do most films than are far more pointed towards clear.\n[Clean input] warm water under a red bridge is a quirky and poignant japanese film that explores the fascinating connections between women, water, nature, and sexuality.\n[Adv. input] warm water under lacking red bridge did neither quirky and poignant japanese film that explores the fascinating connections between women, water, nature, and sexuality."
},
{
"figure_ref": [],
"heading": "SAME-Char",
"publication_ref": [],
"table_ref": [
"tab_1"
],
"text": "[Clean input] the volatile dynamics of female friendship is the subject of this unhurried, low-key film that is so offhollywood that it seems positively french in its rhythms and resonance.\n[Adv. input] the volatile dynamics of female friendship is the subject of this unhurried, low-key film that is so offhollywood tfhat it seems positively french in its rhythms arnd resonance.\n[Clean input] if there's one thing this world needs less of, it's movies about college that are written and directed by people who couldn't pass an entrance exam.\n[Adv. input] if there's one thing this world needs less of, it's movies aLbout college that are written and directed by pople who couldn't pass an entrance exam.\n[Clean input] what's surprising about full frontal is that despite its overt self-awareness, parts of the movie still manage to break past the artifice and thoroughly engage you.\n[Adv. input] what's surprising about full frontal is that despite its overt self-awareness, parts of the movie still manage to break paust the artifice gand thoroughly engage yuo.\nTable 11: Crafted adversarial samples leads to maximum number of computational layers."
},
{
"figure_ref": [],
"heading": "Acknowledgement",
"publication_ref": [],
"table_ref": [],
"text": "This work is supported by Shenzhen Basic Research Key Project \"Multi-modal, multi-task deep neural networks and their training\" (Grant No. JCYJ20220818103001002); and Human-Robot Collaborative AI for Advanced Manufacturing and Engineering (Grant No. A18A2b0046), Agency for Science, Technology and Research, Singapore. This work is also partially supported by NSF grant CCF-2146443, CPS 2038727. "
}
] | Despite much success in natural language processing (NLP), pre-trained language models typically lead to a high computational cost during inference. Multi-exit is a mainstream approach to address this issue by making a trade-off between efficiency and accuracy, where the saving of computation comes from an early exit. However, whether such saving from early-exiting is robust remains unknown. Motivated by this, we first show that directly adapting existing adversarial attack approaches targeting model accuracy cannot significantly reduce inference efficiency. To this end, we propose a simple yet effective attacking framework, SAME, a novel slowdown attack framework on multi-exit models, which is specially tailored to reduce the efficiency of the multi-exit models. By leveraging the multiexit models' design characteristics, we utilize all internal predictions to guide the adversarial sample generation instead of merely considering the final prediction. Experiments on the GLUE benchmark show that SAME can effectively diminish the efficiency gain of various multi-exit models by 80% on average, convincingly validating its effectiveness and generalization ability. | Dynamic Transformers Provide a False Sense of Efficiency | [
{
"figure_caption": "Figure 1 :1Figure 1: Illustration of entropy-based (left) and patience-based (right) early-exiting strategies, l 1...n refer to transformer layers, and H i is the entropy of probability distribution from the i th internal classifier.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Comparison of various attacking methods on entropy-based dynamic models. Attacking methods with lowest speedup are bold.",
"figure_data": "",
"figure_id": "tab_1",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Table 3, the average accuracy score can be further reduced by 42.26%",
"figure_data": "MethodSST-2CoLAMRPCQNLIPD<2%PD<4%PD<2%PD<4%PD<2%PD<4%PD<2%PD<4%PABEE-ALBERT-base2.67x (0.00%)3.56x (0.00%)1.63x (9.11%)1.91x (3.45%)2.03x (1.47%)3.45x (0.00%)1.97x (3.00%)2.41x (1.20%)+HotFlip2.46x (1.38%)3.30x (0.11%)1.40x (17.26%)1.62x (7.86%)1.78x (7.60%)3.07x (0.25%)1.83x (9.20%)2.27x (3.10%)+PWWS2.23x (2.41%)3.01x (0.23%)1.40x (11.60%)1.60x (4.12%)1.53x (13.48%)2.57x (0.49%)1.63x (15.00%)2.04x (4.50%)+TextBugger2.20x (2.06%)2.97x (0.11%)1.41x (11.12%)1.60x (3.74%)1.46x (16.42%)2.35x (0.49%)1.48x (22.30%)1.83x (9.20%)+TextFooler2.12x (2.64%)2.96x (0.34%)1.41x (11.12%)1.60x (3.93%)1.47x (19.12%)2.53x (0.98%)1.47x (25.80%) 1.85x (10.30%)+A2T2.56x (1.03%)3.51x (0.00%)1.43x (15.72%)1.67x (5.18%)1.80x (9.31%)3.13x (0.00%)1.88x (7.70%)2.33x (2.90%)+SAME-Word1.26x (53.10%) 1.68x (17.20%) 1.05x (84.47%)1.06x (84.95%) 1.10x (73.28%) 1.32x (41.67%) 1.28x (50.60%) 1.41x (40.90%)+SAME-Char1.37x (42.66%) 1.77x (14.22%) 1.01x (97.99%)1.01x (95.49%) 1.11x (75.25%) 1.38x (40.93%) 1.30x (49.80%) 1.42x (41.80%)PABEE-BERT-base1.66x (9.52%)1.98x (2.41%)1.19x (35.57%)1.19x (35.57%)1.66x (9.56%)2.01x (2.45%)1.58x (11.00%)1.84x (4.70%)+HotFlip1.49x (22.13%)1.80x (5.50%)1.05x (80.44%)1.05x (80.44%)1.47x (17.40%)1.79x (1.96%)1.44x (20.90%) 1.68x (11.40%)+PWWS1.41x (28.44%) 1.66x (10.44%)1.04x (83.70%)1.04x (83.70%)1.28x (24.02%)1.50x (4.41%)1.33x (32.10%) 1.53x (17.80%)+TextBugger1.37x (32.11%) 1.62x (13.53%)1.04x (86.39%)1.04x (86.39%)1.25x (30.64%) 1.46x (10.29%) 1.25x (44.00%) 1.45x (25.00%)+TextFooler1.37x (32.11%) 1.63x (11.12%)1.05x (82.07%)1.05x (82.07%)1.26x (28.19%)1.48x (6.62%)1.26x (41.70%) 1.45x (21.10%)+A2T1.62x (12.39%)2.00x (2.98%)1.07x (72.20%)1.07x (72.20%)1.37x (21.32%)1.66x (3.43%)1.53x (15.00%)1.78x (7.60%)+SAME-Word1.05x (88.19%) 1.08x (82.11%) 1.00x (100.00%) 1.00x (100.00%) 1.04x (86.76%) 1.10x (69.85%) 1.10x (76.00%) 1.16x (62.70%)+SAME-Char1.14x (69.95%) 1.21x (61.24%)1.00x (99.90%)1.00x (99.90%)1.05x (85.78%) 1.13x (67.16%) 1.15x (63.80%) 1.23x (54.20%)QQPRTEMNLIMNLI-mmPD<2%PD<4%PD<2%PD<4%PD<2%PD<4%PD<2%PD<4%PABEE-ALBERT-base2.58x (0.50%)3.40x (0.00%)1.57x (11.55%)1.85x (5.05%)1.86x (4.70%)2.28x (1.40%)2.29x (1.60%)2.29x (1.60%)+HotFlip2.25x (0.30%)3.02x (0.00%)1.46x (20.94%)1.74x (9.75%)1.62x (13.30%)1.99x (4.50%)2.02x (5.10%)2.02x (5.10%)+PWWS2.34x (1.50%)3.15x (0.00%)1.41x (20.58%)1.61x (11.91%)1.50x (15.40%)1.83x (6.20%)1.81x (5.80%)1.81x (5.80%)+TextBugger2.18x (2.30%)2.88x (0.10%)1.37x (26.35%)1.60x (12.27%)1.45x (20.20%)1.75x (8.90%)1.74x (8.20%)1.74x (8.20%)+TextFooler2.19x (2.20%)2.93x (0.50%)1.37x (28.88%)1.62x (11.19%)1.41x (26.00%) 1.72x (11.00%) 1.75x (11.00%) 1.75x (11.00%)+A2T2.44x (1.10%)3.29x (0.20%)1.45x (22.74%)1.71x (11.91%)1.67x (14.90%)2.06x (5.90%)2.10x (5.50%)2.10x (5.50%)+SAME-Word1.39x (48.70%) 1.65x (27.10%) 1.13x (67.15%)1.20x (59.93%) 1.06x (85.10%) 1.11x (77.30%) 1.08x (82.20%) 1.08x (82.20%)+SAME-Char1.44x (47.90%) 1.71x (27.20%) 1.09x (76.90%)1.13x (69.68%) 1.06x (86.10%) 1.10x (79.70%) 1.11x (80.40%) 1.11x (80.40%)PABEE-BERT-base2.60x (0.40%)3.45x (0.10%)1.21x (55.23%)1.34x (33.21%)1.50x (16.10%)1.76x (7.50%)1.35x (23.30%)1.75x (6.10%)+HotFlip2.45x (1.10%)3.29x (0.10%)1.28x (37.91%)1.42x (28.16%)1.36x (31.30%) 1.60x (13.90%) 1.24x (42.00%) 1.61x (15.30%)+PWWS2.31x (2.30%)3.07x (0.00%)1.26x (44.77%)1.36x (40.43%)1.28x (40.40%) 1.49x (18.30%) 1.17x (53.20%) 1.50x (16.00%)+TextBugger2.21x (2.80%)2.78x (1.00%)1.21x (53.79%)1.34x (39.35%)1.27x (42.60%) 1.45x (22.00%) 1.15x (57.00%) 1.44x (23.50%)+TextFooler2.16x (3.70%)2.91x (0.30%)1.27x (44.40%)1.43x (31.05%)1.22x (52.50%) 1.44x (23.10%) 1.13x (63.50%) 1.42x (26.20%)+A2T2.54x (0.90%)3.47x (0.10%)1.29x (40.43%)1.45x (27.80%)1.41x (29.00%) 1.66x (13.20%) 1.25x (41.60%) 1.63x (15.00%)+SAME-Word1.35x (49.60%) 1.62x (24.90%) 1.09x (75.45%)1.14x (71.12%) 1.01x (96.20%) 1.03x (93.00%) 1.01x (97.80%) 1.03x (92.70%)+SAME-Char1.58x (32.90%) 1.95x (11.20%) 1.08x (74.01%)1.11x (71.84%) 1.03x (91.80%) 1.06x (85.90%) 1.02x (93.50%) 1.06x (86.70%)",
"figure_id": "tab_2",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "tab_3",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Comparison of SAME with(out) accuracy multi-goal function: each entry gives accuracy (left) and speedup (right).",
"figure_data": "",
"figure_id": "tab_5",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Transferability results: the first block shows the results with DeeBERT as the target model, and the second block uses PABEE-BERT as the target model.",
"figure_data": "Each row refers to a different source model. Char andWord refer to varients of SAME. Each entry denotesthe efficiency gain decrease ratio.4.5 Adversarial TrainingWe further whether this new efficiencythreat can be successfully defended through ad-versarial training. Specifically, given a victimmodel. we first generate an adversarial sample us-ing SAME or other adversarial approaches for eachsample from the training set. Then, we equallymix the clean and adversarial samples to retraina new model. Finally, we attack the adversar-ial trained models again with SAME. We adjustthe entropy/patience of adversarial trained modelsto have the same speedup as the original victimmodel. Table 5 shows the results. Overall, the ef-ficiency robustness of dynamic transformers canbe improved through adversarial training (1.18xto 1.58x on average using TextFooler), Yet, therestill exists a drastic speedup loss (2.25x to 1.58x).Compared to accuracy-oriented adversarial data,data from SAME provide more robustness benefi-cial against attack, which validates the potential ofusing SAME to enhance the robustness of current",
"figure_id": "tab_7",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Efficiency of models trained with various adversarial augmented data. Each row refers to a model trained with different adversarial data. Since attacking approaches is required to slowdown the victim models by more layers to achieve the same slowdown ratio, we further investigate the impact of victim model scale on the attacking performance. Experimental results using 24-layer BERT-large model on SST-2 and MNLI are shown in Table6. Due to space limitation, more results can be found in Appendix B. Accuracy-oriented methods can still hardly reduce the inference efficiency. Yet, our proposed SAME effectively reduce the speedup ratio by 89%, which is comparable to 93% on base-size models.",
"figure_data": "4.6 DiscussionImpact of Model Scale: Method SST-2MNLIDeeBERTPABEE-BERTDeeBERTPABEE-BERTw/o AdvTrain 2.06x (2.06%)2.91x (0.57%)1.57x (1.50%)1.56x (15.20%)+HotFlip1.74x (12.73%)2.53x (1.72%)1.45x (7.10%)1.40x (29.70%)+PWWS1.91x (7.11%)2.38x (2.41%)1.45x (4.00%)1.26x (44.80%)+TextBugger1.87x (6.88%)2.35x (1.95%)1.43x (6.90%)1.23x (50.00%)+TextFooler1.92x (8.37%)2.32x (1.95%)1.43x (8.40%)1.23x (49.40%)+A2T2.18x (4.93%)2.79x (1.03%)1.51x (6.20%)1.42x (29.50%)+SAME1.11x (65.71%) 1.22x (58.49%) 1.05x (81.90%) 1.04x (89.50%)",
"figure_id": "tab_8",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "Attacking results on large dynamic transformers with 24 transformer layers.",
"figure_data": "",
"figure_id": "tab_9",
"figure_label": "6",
"figure_type": "table"
},
{
"figure_caption": "Average efficiency reduction ratio on GLUE benchmark under various modification rate .",
"figure_data": "DeeBERTPABEE-BERTWordCharWordCharPD<2% PD<4% PD<2% PD<4% PD<2% PD<4% PD<2% PD<4%3%70.4667.8470.6468.3369.6866.2660.6358.615%83.8180.7082.5681.6781.0576.7871.7970.227%85.4183.5084.5283.3782.5379.7873.6871.0410%90.2188.4788.4387.2686.3284.1577.8976.09",
"figure_id": "tab_10",
"figure_label": "7",
"figure_type": "table"
},
{
"figure_caption": "Ablation studies on layer-wise importance weighting and loss combintation.",
"figure_data": "",
"figure_id": "tab_12",
"figure_label": "8",
"figure_type": "table"
},
{
"figure_caption": "Table 10 gives the results. Overall. our proposed SAME outperforms previous approaches by a large margin across various models and tasks.",
"figure_data": "MethodSST-2CoLAMRPCQNLIPD<2%PD<4%PD<2%PD<4%PD<2%PD<4%PD<2%PD<4%PABEE-ALBERT-large 3.48x (0.11%)5.14x (0.00%)2.68x (0.10%)2.68x (0.10%)3.26x (0.00%)3.88x (0.00%)2.80x (0.20%)3.80x (0.00%)+HotFlip3.15x (0.23%)4.81x (0.00%)2.11x (1.53%)2.11x (1.53%)3.08x (0.00%)3.76x (0.00%)2.49x (1.50%)3.48x (0.00%)+PWWS2.76x (0.23%)4.23x (0.11%)2.18x (0.58%)2.18x (0.58%)2.53x (0.49%)3.14x (0.25%)2.22x (1.90%)3.14x (0.00%)+TextBugger2.52x (0.69%)3.98x (0.00%)2.13x (0.67%)2.13x (0.67%)2.28x (0.74%)2.74x (0.00%)1.99x (3.50%)2.79x (0.30%)+TextFooler2.57x (1.03%)4.05x (0.00%)2.19x (0.48%)2.19x (0.48%)2.42x (0.49%)2.90x (0.25%)1.95x (5.20%)2.86x (0.20%)+A2T3.26x (0.46%)5.10x (0.00%)2.23x (0.96%)2.23x (0.96%)3.10x (0.25%)3.76x (0.00%)2.63x (1.30%)3.72x (0.00%)+SAME-Word1.53x (28.33%) 2.42x (1.15%)1.25x (45.45%)1.25x (45.45%) 1.71x (19.12%) 1.98x (8.82%) 1.52x (34.80%) 2.05x (12.70%)+SAME-Char1.52x (30.05%) 2.50x (0.92%)1.07x (77.76%)1.07x (77.76%) 1.50x (31.86%) 1.76x (11.52%) 1.54x (33.20%) 1.90x (17.70%)PABEE-BERT-large2.29x (2.06%)2.91x (0.57%)1.24x (33.37%)1.24x (33.37%) 1.35x (30.39%) 1.64x (13.24%) 2.31x (1.40%)1.73x (6.90%)+HotFlip2.00x (6.65%)2.53x (1.72%)1.07x (77.56%)1.07x (77.56%) 1.29x (37.75%) 1.58x (16.67%) 2.14x (4.10%) 1.60x (15.40%)+PWWS1.93x (8.03%)2.38x (2.41%)1.05x (84.56%)1.05x (84.56%) 1.11x (73.28%) 1.37x (33.09%) 1.83x (8.30%) 1.44x (21.00%)+TextBugger1.90x (9.40%)2.35x (1.95%)1.04x (87.25%)1.04x (87.25%) 1.09x (76.72%) 1.27x (39.22%) 1.75x (11.10%) 1.37x (27.90%)+TextFooler1.84x (9.98%)2.32x (1.95%)1.05x (82.74%)1.05x (82.74%) 1.09x (75.98%) 1.32x (31.86%) 1.76x (9.90%) 1.35x (29.10%)+A2T2.19x (3.78%)2.79x (1.03%)1.10x (69.70%)1.10x (69.70%) 1.27x (41.91%) 1.53x (17.40%) 2.28x (3.50%) 1.69x (11.90%)+SAME-Word1.13x (77.87%) 1.22x (58.49%) 1.00x (100.00%) 1.00x (100.00%) 1.04x (88.48%) 1.07x (85.29%) 1.19x (61.50%) 1.09x (77.90%)+SAME-Char1.25x (60.55%) 1.40x (42.78%) 1.00x (99.90%)1.00x (99.90%) 1.02x (93.38%) 1.02x (93.63%) 1.24x (56.30%) 1.12x (72.10%)DeeBERT-large1.78x ( 4.70%) 2.06x ( 2.06%)1.47x ( 1.05%)1.50x ( 0.77%)1.68x ( 0.49%) 1.99x ( 0.49%) 1.62x ( 2.80%) 1.80x ( 1.50%)+HotFlip1.51x (20.76%) 1.74x (12.73%)1.37x (5.85%)1.40x (4.60%)1.65x (4.41%)1.93x (2.70%) 1.53x (10.50%) 1.76x (4.20%)+PWWS1.66x (12.27%) 1.91x (7.11%)1.39x (2.59%)1.41x (2.30%)1.58x (6.86%)1.77x (5.15%)1.56x (7.80%)1.73x (3.30%)+TextBugger1.62x (14.11%) 1.87x (6.88%)1.38x (2.78%)1.40x (2.40%)1.50x (8.09%)1.67x (4.66%) 1.51x (10.50%) 1.68x (4.30%)+TextFooler1.61x (15.37%) 1.92x (8.37%)1.40x (2.11%)1.41x (1.63%)1.51x (12.25%) 1.74x (4.66%)1.52x (9.60%)1.67x (5.30%)+A2T1.82x (9.52%)2.18x (4.93%)1.43x (2.68%)1.45x (1.63%)1.65x (6.37%)1.92x (1.47%)1.61x (6.20%)1.81x (2.40%)+SAME-Word1.08x (73.17%) 1.11x (65.71%) 1.03x (86.48%)1.04x (82.93%) 1.12x (70.59%) 1.19x (57.84%) 1.20x (39.60%) 1.26x (30.90%)+SAME-Char1.10x (65.71%) 1.14x (60.09%) 1.01x (94.44%)1.02x (90.80%) 1.07x (79.17%) 1.09x (75.49%) 1.16x (45.80%) 1.20x (40.20%)DeeRoBERTa-large1.75x ( 0.92%) 1.93x ( 0.23%)1.46x ( 1.82%)1.57x ( 0.38%)1.73x ( 0.49%) 2.03x ( 0.25%) 1.89x ( 0.50%) 2.05x ( 0.00%)+HotFlip1.69x (5.62%)2.08x (2.64%)1.38x (10.35%)1.45x (3.36%)1.66x (1.47%)1.98x (0.49%)1.78x (3.20%)1.99x (1.00%)+PWWS1.64x (5.96%)1.81x (2.87%)1.44x (4.31%)1.47x (1.73%)1.59x (3.43%)1.93x (0.25%)1.82x (1.60%)2.00x (0.20%)+TextBugger1.62x (6.08%)1.76x (3.44%)1.42x (6.04%)1.47x (2.21%)1.56x (2.94%)1.91x (0.49%)1.77x (2.70%)1.94x (0.40%)+TextFooler1.60x (7.22%)1.73x (3.78%)1.44x (4.51%)1.47x (1.53%)1.57x (4.17%)1.82x (1.23%)1.72x (5.20%)1.94x (0.90%)+A2T1.74x (2.98%)1.94x (1.26%)1.42x (6.42%)1.50x (2.30%)1.71x (1.47%)2.04x (0.74%)1.86x (2.60%)2.05x (0.40%)+SAME-Word1.37x (20.41%) 1.44x (16.06%) 1.01x (95.97%)1.02x (91.18%) 1.53x (11.03%) 1.70x (6.13%) 1.45x (23.80%) 1.60x (11.70%)+SAME-Char1.27x (31.88%) 1.35x (23.51%) 1.00x (98.95%)1.00x (98.47%) 1.30x (36.27%) 1.45x (22.55%) 1.35x (35.90%) 1.53x (17.50%)QQPRTEMNLIMNLI-mmPD<2%PD<4%PD<2%PD<4%PD<2%PD<4%PD<2%PD<4%PABEE-ALBERT-large 5.06x (0.00%)6.79x (0.00%)1.56x (7.22%)1.29x (20.94%)2.48x (0.50%)3.36x (0.20%)2.52x (0.80%)2.91x (0.30%)+HotFlip4.59x (0.00%)6.02x (0.00%)1.55x (13.36%)1.28x (31.41%)2.17x (2.60%)3.05x (0.50%)2.18x (1.90%)2.60x (0.70%)+PWWS4.50x (0.00%)6.07x (0.00%)1.43x (20.94%)1.23x (40.79%)1.94x (2.70%)2.65x (0.40%)1.95x (2.50%)2.26x (1.60%)+TextBugger4.21x (0.00%)5.81x (0.00%)1.42x (20.94%)1.22x (43.32%)1.80x (5.60%)2.57x (0.60%)1.82x (6.10%)2.15x (1.60%)+TextFooler4.36x (0.00%)6.00x (0.00%)1.39x (25.27%)1.21x (46.21%)1.80x (7.30%)2.60x (0.70%)1.82x (6.60%)2.11x (3.10%)+A2T4.82x (0.00%)6.70x (0.00%)1.54x (13.72%)1.26x (35.02%)2.25x (3.10%)3.19x (0.50%)2.24x (3.40%)2.57x (1.60%)+SAME-Word2.42x (1.00%)3.25x (0.10%)1.20x (53.43%)1.11x (65.34%) 1.14x (72.20%) 1.33x (46.00%) 1.13x (74.70%) 1.21x (61.90%)+SAME-Char2.43x (1.60%)3.28x (0.00%)1.12x (70.04%)1.07x (74.73%) 1.12x (75.80%) 1.24x (53.10%) 1.12x (76.60%) 1.14x (71.00%)PABEE-BERT-large2.55x (0.90%)3.43x (0.10%)1.63x (4.33%)1.85x (1.44%)1.81x (9.10%) 1.56x (15.20%) 1.80x (9.40%) 1.54x (17.10%)+HotFlip2.27x (1.90%)2.97x (0.50%)1.57x (7.58%)1.84x (2.53%)1.61x (17.20%) 1.40x (29.70%) 1.65x (16.90%) 1.42x (28.50%)+PWWS2.17x (2.80%)2.89x (0.20%)1.50x (6.14%)1.67x (3.25%)1.42x (29.20%) 1.26x (44.80%) 1.44x (27.20%) 1.27x (43.90%)+TextBugger2.04x (5.40%)2.79x (0.30%)1.49x (6.50%)1.72x (3.97%)1.40x (32.70%) 1.23x (50.00%) 1.38x (35.50%) 1.23x (51.50%)+TextFooler2.03x (6.50%)2.80x (0.60%)1.48x (7.22%)1.65x (4.69%)1.40x (32.20%) 1.23x (49.40%) 1.40x (33.30%) 1.23x (50.70%)+A2T2.38x (2.50%)3.19x (0.40%)1.62x (7.58%)1.84x (3.97%)1.68x (17.80%) 1.42x (29.50%) 1.66x (17.40%) 1.43x (28.80%)+SAME-Word1.42x (48.30%) 1.66x (30.70%) 1.14x (56.68%)1.22x (42.96%) 1.06x (87.30%) 1.04x (89.50%) 1.05x (90.00%) 1.04x (90.50%)+SAME-Char1.43x (50.20%) 1.73x (22.00%) 1.08x (70.04%)1.16x (58.84%) 1.05x (90.00%) 1.03x (92.20%) 1.04x (90.50%) 1.02x (94.30%)DeeBERT-large2.08x ( 3.30%) 2.35x ( 1.60%)1.71x ( 1.81%)1.78x ( 1.08%)1.47x ( 4.20%) 1.57x ( 1.50%) 1.50x ( 3.50%) 1.59x ( 1.10%)+HotFlip1.98x (7.80%)2.27x (3.70%)1.73x (2.89%)1.78x (1.81%)1.35x (14.20%) 1.45x (7.10%) 1.37x (12.60%) 1.47x (5.30%)+PWWS2.16x (3.40%)2.45x (1.40%)1.76x (1.08%)1.79x (0.36%)1.35x (9.50%)1.45x (4.00%)1.36x (8.60%)1.47x (2.30%)+TextBugger2.26x (3.70%)2.64x (1.80%)1.73x (5.42%)1.81x (1.81%)1.31x (14.80%) 1.43x (6.90%) 1.32x (14.60%) 1.44x (5.60%)+TextFooler2.17x (6.00%)2.50x (2.10%)1.71x (3.97%)1.78x (2.53%)1.31x (17.70%) 1.43x (8.40%) 1.33x (16.50%) 1.44x (6.10%)+A2T2.18x (5.70%)2.49x (1.60%)1.73x (4.69%)1.80x (2.53%)1.39x (13.50%) 1.51x (6.20%) 1.41x (12.00%) 1.51x (5.20%)+SAME-Word1.29x (52.10%) 1.37x (45.10%) 1.17x (55.96%)1.22x (49.10%) 1.04x (84.90%) 1.05x (81.90%) 1.03x (86.50%) 1.06x (76.00%)+SAME-Char1.31x (53.00%) 1.42x (39.90%) 1.13x (66.79%)1.14x (64.26%) 1.02x (90.60%) 1.04x (86.20%) 1.03x (86.60%) 1.05x (81.30%)DeeRoBERTa-large2.15x ( 0.90%) 2.36x ( 0.70%)1.35x ( 1.44%)1.41x ( 0.00%)1.32x ( 2.70%) 1.35x ( 1.30%) 1.35x ( 3.20%) 1.38x ( 1.10%)+HotFlip2.05x (2.00%)2.27x (1.10%)1.32x (6.86%)1.39x (0.72%)1.26x (14.40%) 1.29x (9.10%)1.29x (9.90%)1.32x (5.90%)+PWWS2.27x (0.90%)2.53x (0.30%)1.31x (7.58%)1.38x (1.44%)1.26x (16.70%) 1.29x (11.30%) 1.27x (12.60%) 1.31x (9.70%)+TextBugger2.54x (1.00%)3.04x (0.50%)1.30x (9.03%)1.38x (0.72%)1.24x (21.60%) 1.28x (14.50%) 1.26x (16.80%) 1.30x (11.80%)+TextFooler2.27x (1.80%)2.53x (0.80%)1.31x (9.39%)1.39x (0.36%)1.25x (18.80%) 1.30x (11.40%) 1.28x (12.10%) 1.33x (6.70%)+A2T2.30x (1.50%)2.55x (1.00%)1.33x (6.50%)1.40x (1.44%)1.30x (9.00%)1.33x (4.50%)1.32x (6.30%)1.35x (4.40%)+SAME-Word1.47x (39.90%) 1.62x (30.10%) 1.17x (40.07%)1.28x (15.16%) 1.14x (48.20%) 1.18x (37.70%) 1.15x (42.90%) 1.20x (32.00%)+SAME-Char1.42x (42.30%) 1.54x (33.40%) 1.07x (72.56%)1.14x (47.65%) 1.10x (59.80%) 1.12x (52.50%) 1.12x (55.10%) 1.16x (40.30%)",
"figure_id": "tab_13",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Full results of various attacking methods on large dynamic models: each entry gives the speedup (left) and ratio of samples with number of inference layer at least 22. Attacking methods with lowest speedup are bold.",
"figure_data": "",
"figure_id": "tab_14",
"figure_label": "10",
"figure_type": "table"
}
] | Yiming Chen; Simin Chen; Zexin Li; Wei Yang; Cong Liu; Robby T Tan; Haizhou Li | [
{
"authors": "Yonatan Belinkov; Yonatan Bisk",
"journal": "",
"ref_id": "b0",
"title": "Synthetic and natural noise both break neural machine translation",
"year": "2018"
},
{
"authors": "Simin Chen; Mirazul Haque; Cong Liu; Wei Yang; ; ",
"journal": "",
"ref_id": "b1",
"title": "Deepperform: An efficient approach for performance testing of resource-constrained neural networks",
"year": "2022"
},
{
"authors": "Simin Chen; Hamed Khanpour; Cong Liu; Wei Yang",
"journal": "",
"ref_id": "b2",
"title": "Learning to reverse dnns from ai programs automatically",
"year": "2022"
},
{
"authors": "Simin Chen; Cong Liu; Mirazul Haque; Zihe Song; Wei Yang",
"journal": "",
"ref_id": "b3",
"title": "Nmtsloth: understanding and testing efficiency degradation of neural machine translation systems",
"year": "2022"
},
{
"authors": "Simin Chen; Zihe Song; Mirazul Haque; Cong Liu; Wei Yang",
"journal": "",
"ref_id": "b4",
"title": "Nicgslowdown: Evaluating the efficiency robustness of neural image caption generation models",
"year": "2022"
},
{
"authors": "Minhao Cheng; Jinfeng Yi; Pin-Yu Chen; Huan Zhang; Cho-Jui Hsieh",
"journal": "",
"ref_id": "b5",
"title": "Seq2sick: Evaluating the robustness of sequence-to-sequence models with adversarial examples",
"year": "2020"
},
{
"authors": "Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova",
"journal": "Association for Computational Linguistics",
"ref_id": "b6",
"title": "BERT: Pre-training of deep bidirectional transformers for language understanding",
"year": "2019"
},
{
"authors": "William B Dolan; Chris Brockett",
"journal": "IWP",
"ref_id": "b7",
"title": "Automatically constructing a corpus of sentential paraphrases",
"year": "2005"
},
{
"authors": "Javid Ebrahimi; Daniel Lowd; Dejing Dou",
"journal": "Association for Computational Linguistics",
"ref_id": "b8",
"title": "On adversarial examples for character-level neural machine translation",
"year": "2018"
},
{
"authors": "Javid Ebrahimi; Anyi Rao; Daniel Lowd; Dejing Dou",
"journal": "Association for Computational Linguistics",
"ref_id": "b9",
"title": "HotFlip: White-box adversarial examples for text classification",
"year": "2018"
},
{
"authors": "Angela Fan; Shruti Bhosale; Holger Schwenk; Zhiyi Ma; Ahmed El-Kishky; Siddharth Goyal; Mandeep Baines; Onur Celebi; Guillaume Wenzek; Vishrav Chaudhary",
"journal": "The Journal of Machine Learning Research",
"ref_id": "b10",
"title": "Beyond english-centric multilingual machine translation",
"year": "2021"
},
{
"authors": "Xiaoxue Gao; Chitralekha Gupta; Haizhou Li; ; ",
"journal": "IEEE/ACM Transactions on Audio, Speech, and Language Processing",
"ref_id": "b11",
"title": "Automatic lyrics transcription of polyphonic music with lyrics-chord multi-task learning",
"year": "2022"
},
{
"authors": "Xiaoxue Gao; Chitralekha Gupta; Haizhou Li",
"journal": "IEEE",
"ref_id": "b12",
"title": "Genre-conditioned acoustic models for automatic lyrics transcription of polyphonic music",
"year": "2022"
},
{
"authors": "Xiaoxue Gao; Chitralekha Gupta; Haizhou Li",
"journal": "IEEE/ACM Transactions on Audio, Speech, and Language Processing",
"ref_id": "b13",
"title": "Polyscriber: Integrated fine-tuning of extractor and lyrics transcriber for polyphonic music",
"year": "2023"
},
{
"authors": "Sanghyun Hong; Yigitcan Kaya; , -Vlad Ionut; Tudor Modoranu; Dumitras",
"journal": "",
"ref_id": "b14",
"title": "A panda? no, it's a sloth: Slowdown attacks on adaptive multi-exit neural network inference",
"year": "2021"
},
{
"authors": "Ting-Kuei Hu; Tianlong Chen; Haotao Wang; Zhangyang Wang",
"journal": "",
"ref_id": "b15",
"title": "Triple wins: Boosting accuracy, robustness and efficiency together by enabling input-adaptive inference",
"year": "2020"
},
{
"authors": "Xiaoqi Jiao; Yichun Yin; Lifeng Shang; Xin Jiang; Xiao Chen; Linlin Li; Fang Wang; Qun Liu",
"journal": "Association for Computational Linguistics",
"ref_id": "b16",
"title": "TinyBERT: Distilling BERT for natural language understanding",
"year": "2020"
},
{
"authors": "Di Jin; Zhijing Jin; Joey Tianyi Zhou; Peter Szolovits",
"journal": "",
"ref_id": "b17",
"title": "Is bert really robust? a strong baseline for natural language attack on text classification and entailment",
"year": "2020"
},
{
"authors": "Zhenzhong Lan; Mingda Chen; Sebastian Goodman; Kevin Gimpel; Piyush Sharma; Radu Soricut",
"journal": "",
"ref_id": "b18",
"title": "Albert: A lite bert for self-supervised learning of language representations",
"year": "2020"
},
{
"authors": "Thai Le; Jooyoung Lee; Kevin Yen; Yifan Hu; Dongwon Lee",
"journal": "Association for Computational Linguistics",
"ref_id": "b19",
"title": "Perturbations in the wild: Leveraging human-written text perturbations for realistic adversarial attack and defense",
"year": "2022"
},
{
"authors": "Mike Lewis; Yinhan Liu; Naman Goyal; Marjan Ghazvininejad; Abdelrahman Mohamed; Omer Levy; Veselin Stoyanov; Luke Zettlemoyer",
"journal": "Association for Computational Linguistics",
"ref_id": "b20",
"title": "BART: Denoising sequence-to-sequence pretraining for natural language generation, translation, and comprehension",
"year": "2020"
},
{
"authors": "Jinfeng Li; Shouling Ji; Tianyu Du; Bo Li; Ting Wang",
"journal": "The Internet Society",
"ref_id": "b21",
"title": "Textbugger: Generating adversarial text against real-world applications",
"year": "2019"
},
{
"authors": "Linyang Li; Ruotian Ma; Qipeng Guo; Xiangyang Xue; Xipeng Qiu",
"journal": "Association for Computational Linguistics",
"ref_id": "b22",
"title": "BERT-ATTACK: Adversarial attack against BERT using BERT",
"year": "2020"
},
{
"authors": "Xiaonan Li; Yunfan Shao; Tianxiang Sun; Hang Yan; Xipeng Qiu; Xuanjing Huang",
"journal": "Association for Computational Linguistics",
"ref_id": "b23",
"title": "Accelerating BERT inference for sequence labeling via early-exit",
"year": "2021"
},
{
"authors": "Zexin Li; Bangjie Yin; Taiping Yao; Juefeng Guo; Shouhong Ding; Simin Chen; Cong Liu",
"journal": "",
"ref_id": "b24",
"title": "Sibling-attack: Rethinking transferable adversarial attacks against face recognition",
"year": "2023"
},
{
"authors": "Aiwei Liu; Honghai Yu; Xuming Hu; Shu'ang Li; Li Lin; Fukun Ma; Yawen Yang; Lijie Wen",
"journal": "Association for Computational Linguistics",
"ref_id": "b25",
"title": "Character-level white-box adversarial attacks against transformers via attachable subwords substitution",
"year": "2022"
},
{
"authors": "Weijie Liu; Peng Zhou; Zhiruo Wang; Zhe Zhao; Haotang Deng; Qi Ju",
"journal": "Association for Computational Linguistics",
"ref_id": "b26",
"title": "FastBERT: a selfdistilling BERT with adaptive inference time",
"year": "2020"
},
{
"authors": "Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov",
"journal": "",
"ref_id": "b27",
"title": "Roberta: A robustly optimized bert pretraining approach",
"year": "2019"
},
{
"authors": "John Morris; Eli Lifland; Jin Yong Yoo; Jake Grigsby; Di Jin; Yanjun Qi",
"journal": "Association for Computational Linguistics",
"ref_id": "b28",
"title": "TextAttack: A framework for adversarial attacks, data augmentation, and adversarial training in NLP",
"year": "2020"
},
{
"authors": "Jianmo Ni; Gustavo Hernandez Abrego; Noah Constant; Ji Ma; Keith Hall; Daniel Cer; Yinfei Yang",
"journal": "Association for Computational Linguistics",
"ref_id": "b29",
"title": "Sentence-t5: Scalable sentence encoders from pre-trained text-to-text models",
"year": "2022"
},
{
"authors": "Nicolas Papernot; Patrick D Mcdaniel; Ananthram Swami; Richard E Harang",
"journal": "IEEE Military Communications Conference",
"ref_id": "b30",
"title": "Crafting adversarial input sequences for recurrent neural networks",
"year": "2016"
},
{
"authors": "Alec Radford; Jeffrey Wu; Rewon Child; David Luan; Dario Amodei; Ilya Sutskever",
"journal": "",
"ref_id": "b31",
"title": "Language models are unsupervised multitask learners",
"year": "2019"
},
{
"authors": "Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu",
"journal": "Journal of Machine Learning Research",
"ref_id": "b32",
"title": "Exploring the limits of transfer learning with a unified text-totext transformer",
"year": "2020"
},
{
"authors": "Pranav Rajpurkar; Jian Zhang; Konstantin Lopyrev; Percy Liang",
"journal": "Association for Computational Linguistics",
"ref_id": "b33",
"title": "SQuAD: 100,000+ questions for machine comprehension of text",
"year": "2016"
},
{
"authors": "Yihe Shuhuai Ren; Kun Deng; Wanxiang He; Che",
"journal": "Association for Computational Linguistics",
"ref_id": "b34",
"title": "Generating natural language adversarial examples through probability weighted word saliency",
"year": "2019"
},
{
"authors": "Victor Sanh; Lysandre Debut; Julien Chaumond; Thomas Wolf",
"journal": "",
"ref_id": "b35",
"title": "Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter",
"year": "2019"
},
{
"authors": "Tal Schuster; Adam Fisch; Jai Gupta; Mostafa Dehghani; Dara Bahri; Q Vinh; Yi Tran; Donald Tay; Metzler",
"journal": "",
"ref_id": "b36",
"title": "Confident adaptive language modeling",
"year": "2022"
},
{
"authors": "Richard Socher; Alex Perelygin; Jean Wu; Jason Chuang; Christopher D Manning; Andrew Ng; Christopher Potts",
"journal": "Association for Computational Linguistics",
"ref_id": "b37",
"title": "Recursive deep models for semantic compositionality over a sentiment treebank",
"year": "2013"
},
{
"authors": "Eric Wallace; Shi Feng; Nikhil Kandpal; Matt Gardner; Sameer Singh",
"journal": "Association for Computational Linguistics",
"ref_id": "b38",
"title": "Universal adversarial triggers for attacking and analyzing NLP",
"year": "2019"
},
{
"authors": "Alex Wang; Amanpreet Singh; Julian Michael; Felix Hill; Omer Levy; Samuel Bowman",
"journal": "Association for Computational Linguistics",
"ref_id": "b39",
"title": "GLUE: A multi-task benchmark and analysis platform for natural language understanding",
"year": "2018"
},
{
"authors": "Jiadong Wang; Xinyuan Qian; Haizhou Li",
"journal": "",
"ref_id": "b40",
"title": "Predict-and-update network: Audio-visual speech recognition inspired by human speech perception",
"year": "2022"
},
{
"authors": "Alex Warstadt; Amanpreet Singh; Samuel R Bowman",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b41",
"title": "Neural network acceptability judgments",
"year": "2019"
},
{
"authors": "Adina Williams; Nikita Nangia; Samuel Bowman",
"journal": "Association for Computational Linguistics",
"ref_id": "b42",
"title": "A broad-coverage challenge corpus for sentence understanding through inference",
"year": "2018"
},
{
"authors": "Ji Xin; Raphael Tang; Jaejun Lee; Yaoliang Yu; Jimmy Lin",
"journal": "Association for Computational Linguistics",
"ref_id": "b43",
"title": "DeeBERT: Dynamic early exiting for accelerating BERT inference",
"year": "2020"
},
{
"authors": "Jin ; Yong Yoo; Yanjun Qi",
"journal": "Association for Computational Linguistics",
"ref_id": "b44",
"title": "Towards improving adversarial training of NLP models",
"year": "2021"
},
{
"authors": "Ofir Zafrir; Guy Boudoukh; Peter Izsak; Moshe Wasserblat",
"journal": "IEEE",
"ref_id": "b45",
"title": "Q8bert: Quantized 8bit bert",
"year": "2019"
},
{
"authors": "Xinze Zhang; Junzhe Zhang; Zhenhua Chen; Kun He",
"journal": "Association for Computational Linguistics",
"ref_id": "b46",
"title": "Crafting adversarial examples for neural machine translation",
"year": "2021"
},
{
"authors": "Wangchunshu Zhou; Canwen Xu; Tao Ge; Julian Mcauley; Ke Xu; Furu Wei",
"journal": "",
"ref_id": "b47",
"title": "Bert loses patience: Fast and robust inference with early exit",
"year": "2020"
},
{
"authors": "Jinhua Zhu; Yingce Xia; Lijun Wu; Di He; Tao Qin; Wengang Zhou; Houqiang Li; Tieyan Liu",
"journal": "",
"ref_id": "b48",
"title": "Incorporating bert into neural machine translation",
"year": "2020"
},
{
"authors": "Wei Zhu",
"journal": "Association for Computational Linguistics",
"ref_id": "b49",
"title": "LeeBERT: Learned early exit for BERT with cross-level optimization",
"year": "2021"
}
] | [
{
"formula_coordinates": [
3,
78.34,
602.46,
210.79,
18.47
],
"formula_id": "formula_0",
"formula_text": "∆ = argmax δ Exit F (x + δ) s.t.||δ|| ≤ , (1)"
},
{
"formula_coordinates": [
4,
115.19,
123.06,
173.94,
33.71
],
"formula_id": "formula_1",
"formula_text": "L mess = N i=1 SCE(F i (x), U),(2)"
},
{
"formula_coordinates": [
4,
111.29,
369.35,
177.85,
33.71
],
"formula_id": "formula_2",
"formula_text": "L patience = N i=1 CE(F i (x), h i ),(3)"
},
{
"formula_coordinates": [
4,
76.29,
509.46,
212.84,
32.17
],
"formula_id": "formula_3",
"formula_text": "hi = argmax(Fi(x)), hi-1 = argmax(Fi(x)) argsecond(Fi(x)), hi-1 = argmax(Fi(x)) ,(4)"
},
{
"formula_coordinates": [
4,
334.04,
100.45,
190.37,
26.94
],
"formula_id": "formula_4",
"formula_text": "w i = α, i < Exit F (x) β i-Exit F (x) i ≥ Exit F (x) ,(5)"
},
{
"formula_coordinates": [
4,
311.6,
261.47,
208.57,
33.71
],
"formula_id": "formula_5",
"formula_text": "L total = N i=1 w i (λL i mess +(1-λ)L i patience ), (6"
},
{
"formula_coordinates": [
4,
520.17,
273.45,
4.24,
9.46
],
"formula_id": "formula_6",
"formula_text": ")"
},
{
"formula_coordinates": [
5,
83.2,
275.25,
176.65,
29.46
],
"formula_id": "formula_7",
"formula_text": "I s,t = j (E(t) -E(s)) j × ∂L total (x)"
},
{
"formula_coordinates": [
5,
137.84,
282.63,
151.29,
47.04
],
"formula_id": "formula_8",
"formula_text": "∂s j i ; δ = argmax t I s,t(7)"
},
{
"formula_coordinates": [
6,
315.88,
626.35,
208.53,
13.15
],
"formula_id": "formula_9",
"formula_text": "Exit F (x + δ) + σ × 1(F(x + δ) = y true ), (8)"
}
] |
|
2023-05-20 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b19"
],
"table_ref": [],
"text": "The term explanation in artificial intelligence (AI) is often conflated with the concepts of interpretability and explainable AI (XAI), but there are important distinctions to be made. Miller (2019) defines interpretability and XAI as the process of building AI systems that humans can understand. In other words, by design, the AI's decision-making process is inherently transparent to a human. In contrast, explicitly explaining the decision-making to an arbitrary human is explanation generation. The latter is the subject of this paper. More specifically, we are working towards developing a formal framework for the automated generation and assessment of explanations.\nFirstly, some key terminology: an explanation is generated through a dialectical interaction whereby one agent, the explainer, seeks to 'explain' some phenomenon, called the explanandum, to another agent, the explainee. In this article, we propose a novel measure of explanatory effectiveness that can be used to motivate artificial agents to generate good explanations (e.g. in the form of a reward signal), or to analyse the behaviours of existing communicating agents. We then define explanation games as cooperative games where two (or more) agents seek to maximise the effectiveness measure."
},
{
"figure_ref": [],
"heading": "Related Literature",
"publication_ref": [
"b1",
"b21",
"b26",
"b25",
"b9",
"b2",
"b24",
"b16",
"b5",
"b22",
"b27",
"b12",
"b0",
"b11",
"b10",
"b6",
"b18",
"b33",
"b29",
"b28",
"b4"
],
"table_ref": [],
"text": "Intepretability and XAI have received an abundance of recent attention (see Adadi & Berrada (2018) for a review). This is largely due to two factors; regulatory demands (UK Information Commissioner's Office 2019) and the emergence of highly-performant black-box models, such as deep neural networks, that are naturally inscrutable. However, the central crux of interpretability techniques is the need to define a fixed interpretable domain from which we can derive explanations. This presents two challenges: there are no formal procedures for determining if a given domain is interpretable; and a domain may be interpretable to some agents, but not others, or only within some contexts. Moving away from interpretability, the problem of explanation generation has a long history in AI (Mueller et al. 2019). To some, there is a sense in which generating explanations is the hallmark of intelligence itself (Schank 1984). To others, explanation is simply about building models -a process which is seen as merely instrumental to intelligent behaviour (Russell & Norvig 2010, Hutter 2005, Chaitin 2006).\nIn the philosophy of science the concept of explanation is posed in terms of generating descriptions of, or hypotheses regarding, latent phenomena. This has led to investigations of formal measures of explanatory power, with an early example being Popper's (1959) notion of the 'degree of corroboration'. This developed into a line of philosophers devising subjectivist definitions for capturing aspects of the 'goodness' of explanations or hypotheses (Lipton 2003, Glass 2002, Okasha 2000, Schupbach & Sprenger 2011). However, by the subjectivity of these measures they may only assess the degree to which one believes (or simply likes) an explanation, which is not necessarily correlated with the degree to which an explanation is actually true (or representative of the world).\nRecently, calls have been made for the need for human-centred explanation (Kirsch 2017, Abdul et al. 2018). Yet, the framing of explanation generation as a cooperative problem between a human and machine dates back to the era of expert systems (Karsenty & Brezillon 1995, Johnson & Johnson 1993, Graesser et al. 1996). By articulating explanation as a formal dialogue, a related direction of investigation is dialogue games (McBurney & Parsons 2002). In particular, information-seeking (Walton & Krabbe 1995) and education (Sklar & Parsons 2004) dialogues are especially relevant. Sklar & Azhar (2018) conducted empirical research with a human-machine collaboration task where the agents participated in a dialogue and explanations were provided to a human based of an argumentation framework (Dung 1995)."
},
{
"figure_ref": [],
"heading": "What is Explanation?",
"publication_ref": [
"b8",
"b2",
"b34",
"b17",
"b31",
"b23",
"b14",
"b7"
],
"table_ref": [],
"text": "In this work we treat explanatory processes as involving two agents -an explainer and an explainee -and the result is that the explainee understands the explanandum better by the end than they did at the start. We define 'an explanation' as any sequence of observations made by the explainee that leads to this result. Thus an explanation could be a piece of text or spoken language, but it could also be a diagram or a piece of interactive media.\nWith this we shift the problem onto formally defining a measure of an agent's 'understanding' of some arbitrary phenomenon. We approach the question in terms of four stances1 towards comprehension, understanding as: (1) a sensation (Hume 1751); (2) information compression (Chaitin 2006, Zenil 2019, Maguire et al. 2016); (3) performance capacity (Turing 1950, Perkins 1993); or (4) organised information (Lakoff & Johnson 1980, Hofstadter & Sander 2012).\nThe sensation stance states that comprehension is a conscious experienceyou understand something if you feel that you apprehend it. The compression stance says that understanding is the formulation of concise and accurate descriptions of phenomena. The performance stance argues that having information is not enough; you must also know how to use the information. The organisedinformation stance tells us that utilisation and compression are a byproduct of something more important; namely that the agent represents information in relation to their own conceptual framework. While each of the stances has issues of their own, combined they provide a persuasive account. In other words, if someone claims they understand something, they can use their information to do things, and their description of the phenomenon is concise, accurate, and grounded in other concepts that they understand, then it is hard to argue that they do not grasp the phenomenon."
},
{
"figure_ref": [],
"heading": "Technical Background",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Algorithmic Information Theory",
"publication_ref": [
"b30",
"b13",
"b3"
],
"table_ref": [],
"text": "Algorithmic Information Theory (AIT) is a view of information that takes a fundamentally computational approach (Solomonoff 1964, Kolmogorov 1968, Chaitin 1975). Formally, AIT is built on the notion of Kolmogorov complexity, denoted K(x). K(x) is defined as the length of the shortest program, p, on a Universal Turing Machine (UTM), U , that outputs x. The conditional Kolmogorov complexity, K(x|y), is similarly defined by the length of the shortest program that produces x when given input y."
},
{
"figure_ref": [],
"heading": "K(x) = min",
"publication_ref": [],
"table_ref": [],
"text": "K(x|y) = min p {|p| : U (yp) = x} (2)\nThus we can define a measure of mutual information:\nI(x; y) = K(y) -K(y|x) (3)\nUnless otherwise specified, when we talk of the mutual information between two objects we will be referring to an application of Equation 3."
},
{
"figure_ref": [],
"heading": "Agents",
"publication_ref": [],
"table_ref": [],
"text": "In its most basic conception, 'an agent' is any system that makes observations and takes actions. For any agent X t ∈ X at time t, we will denote that they make observations o t X ∈ O X and take actions a t X ∈ A X . Another important factor in describing agents is their internal state. This phrase can refer to various aspects of an agent's cognition, but we are mostly interested in this object insofar as it stores information. Firstly, we will assume that an agent's internal state may fall into a variety of configurations, i.e. there exists a set of possible internal states for an agent, which we will denote Z X . Secondly, we will talk of information being 'encoded' in an agent's internal state. Given an object o, we will denote X's encoding of o as o X , where o X ∈ {p : U (p) = o} for some UTM U . We will speak of the agent 'having' this encoding, or its internal state 'containing' this encoding. This is independent to how this is achieved, e.g. the agent's internal state may simply store a list of encodings, or multiple encodings may overlap in a distributed storage medium such as a neural network."
},
{
"figure_ref": [],
"heading": "Universal Intelligence Theory",
"publication_ref": [
"b15"
],
"table_ref": [],
"text": "Universal Intelligence Theory (UIT), proposed by Legg & Hutter (2007), establishes a definition of machine intelligence based on algorithmic information theory and reinforcement learning. In order to meaningfully compare different performances over a potentially infinite number of time steps, the scope of possible environments is limited such that the sum of rewards (the return) is always less than one. We will refer this as the set of bounded-test environments. With this, the universal intelligence of an agent π is given by:\nΥ (π) = µ∈E 2 -K(µ) V π µ (4)\nWhere V π µ is the return that π achieves in environment µ."
},
{
"figure_ref": [],
"heading": "Universal Artificial Intelligence",
"publication_ref": [
"b30",
"b20",
"b9"
],
"table_ref": [],
"text": "Consider a stochastic environment with dynamics described by a probability distribution µ(e k |ae <k ), where e k is the percept (observation-reward tuple) given at time k, and ae <k is the action-percept history. In order to perform optimally, the agent in this environment must infer µ. This is known as the problem of induction. By combining Solomonoff induction (Solomonoff 1964) with von Neumann-Morgenstern rational decision-making (Morgenstern & von Neumann 1953), Hutter (2005) defines AIXI; an agent that chooses the best possible action at every time step given perfect inductive inference."
},
{
"figure_ref": [],
"heading": "Formalising Understanding",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Partitioning the Internal State",
"publication_ref": [],
"table_ref": [],
"text": "In order to devise a measure of understanding, we will start by defining partitions of the information in the internal state. These partitions are constructed with respect to a given phenomenon p ∈ P. There are four: the p-relevant information (all information related to p), the p-irrelevant (all information completely unrelated to p), the p-specific (the information that only relates to p), and the p-background information (all information that is not specifically related to p).\nIn the following formal definitions we are using a particular notation that warrants explanation. As we have already established, z X denotes the internal state of agent X. We denote p-relevant notation with a comma after the X followed by * p, z X, * p . The star indicates that we are 'selecting' all of the information relevant to p, rather than only the information specific to p. When the star is omitted we are referring to to specific information regarding whatever follows the comma, e.g. z X,p is the p-specific information and z X,¬p is the information specific to everything that is not p (the p-irrelevant information).\nDefinition 1 (p-Relevant Information). Given a phenomenon p ∈ P and an agent X with internal state z X , the p-relevant information z X, * p ∈ Z X is the object where I(z X, * p ; p) = I(z X ; p) and I(z X, * p ; z X ) is minimised, i.e. there exists no z ′ X, * p such that I(z ′ X, * p ; p) = I(z X ; p) and I(z ′ X, * p ; z X ) < I(z X, * p ; z X ). Definition 2 (p-Irrelevant Information). Given a phenomenon p ∈ P and an agent X with internal state z X ∈ Z X , the p-irrelevant information z X,¬p is the object where I(z X,¬p ; p) = 0 and I(z X,¬p ; z X ) is maximised, i.e. there exists no z ′ X,¬p such that I(z ′ X,¬p ; p) = 0 and I(z ′ X,¬p ; z X ) > I(z X,¬p ; z X ). Definition 3 (p-Specific Information). Given a phenomenon p ∈ P and an agent X with internal state z X ∈ Z X , the p-specific information z X,p is the object where I(z X,p ; p) > 0, I(z X,p ; p ′ ) = 0 ∀p ′ ∈ P, p ′ = p and the mutual information I(z X,p ; z X ) is maximised, i.e. there exists no z ′ X,p such that I(z ′ X,p ; p) > 0, I(z ′ X,p , p ′ ) = 0 ∀p ′ ∈ P, p ′ = p and I(z ′ X,p ; z X ) > I(z X,p ; z X ). Definition 4 (p-Background Information). Given a phenomenon p ∈ P and an agent X with internal state z X ∈ Z X and p-specific information z X,p , the p-background information2 z X, * ¬p is the object where I(z X,p ; z X, * ¬p ) = 0 and I(z X, * ¬p ; z X ) is maximised, i.e. there exists no z ′ X, * ¬p such that I(z X,p ; z ′ X, * ¬p ) = 0 and I(z ′ X, * ¬p ; z X ) > I(z X, * ¬p ; z X )."
},
{
"figure_ref": [],
"heading": "Information Compression",
"publication_ref": [],
"table_ref": [],
"text": "With these partitions we can define how compressed the p-relevant information is:\nDefinition 5 (p-Compression Factor). Suppose a phenomenon p ∈ P and an agent X. The p-compression factor c : X × P → (0, 1] is given as the ratio of the Kolmogorov complexity of the p-relevant information object to the size of the agent's encoding of that information:\nc(X, p) = K(z X, * p ) | z X, * p X | (5)"
},
{
"figure_ref": [],
"heading": "Information Utilisation",
"publication_ref": [
"b9"
],
"table_ref": [],
"text": "Next, we will attempt to formalise the performance stance on understanding, i.e. we will try to define X's information utilisation of p. To do this, we will need to construct a set of 'fair tests of p' for X. We will start by noting: (1) A fair test for X should require X's background information;\n(2) a test of p should require information about p. We will use the formalisation of rational decision-making, AIXI, to 'benchmark' how information is utilised in a given environment. Unlike a typical test-taker, AIXI enters into an environment with no prior knowledge, and thus we must present any priors to AIXI as a part of its percept sequence. Therefore, to decide whether or not a given task meets the criteria outlined above we will construct a 'meta-task' for AIXI where relevant observations are prepended to the task.\nDefinition 6 ((X, p)-tests). Given a phenomenon p ∈ P and an agent X with internal state z X ∈ Z X , we start with the set of bounded-test environments E, we define the set of (X, p)-tests, E X,p , as follows:\nE X,p = µ ∈ E : V AIXI (p,b)•µ = V * µ > 0, V AIXI (p)•µ = V AIXI (b)•µ = V AIXI µ = 0 (6)\nWhere b is a shorthand for the p-background information b = z X, * ¬p , and x • µ denotes the construction of a new environment µ ′ such that:\n∀x i ∈ x, ∀a <i , µ ′ ((x i , 0) | a <i ) = 1 (7) ∀k > |x|, µ ′ (e k | a <k ) = µ(e k | a j...k ), where j = |x|(8)\nIt is worth noting why we are using only the p-background information and not the agent's entire internal state as required prior knowledge. This is because if the agent knows anything about p then AIXI would be able to use the information encoded in the internal state to pass the test when only given b. We want AIXI to only get information about p from p itself so that we can strictly outline the criteria above.\nUsing the set of fair tests for X, we can define a measure of information utilisation by measuring the agent's intelligence across these environments. This is an adaptation of Hutter's (2005) measure of intelligence (Equation 4)."
},
{
"figure_ref": [],
"heading": "Definition 7 (p-Utilisation).",
"publication_ref": [],
"table_ref": [],
"text": "Given an agent X and phenomenon p, the putilisation Υ p : X → [0, 1] is defined:\nΥ p (X) = µ∈EX,p 2 -K(µ) V X µ (9)"
},
{
"figure_ref": [],
"heading": "Information Integration",
"publication_ref": [],
"table_ref": [],
"text": "With the definitions we have constructed here, we can also introduce a measure of how 'integrated' the p-relevant information is."
},
{
"figure_ref": [],
"heading": "Definition 8 (p-Integration",
"publication_ref": [],
"table_ref": [],
"text": "). Suppose we have a phenomenon p ∈ P, and an agent X ∈ X with p-relevant information z X, * p and p-specific information z X,p . The p-integration, φ :\nX × P → [0, 1), is defined, φ(X, p) = tanh | z X, * p X | | z X,p X | -1(10)\nAs the p-relevant information will always be larger-than or equal to the pspecific information (| z X, * p X | ≥ | z X,p X |), the ratio in this measure will equal 1 when all relevant information is specific. In this case, there is no relevant information that is used for anything else, i.e. the p-relevant information is not at all integrated into the rest of the internal state (or nothing else exists to integrate with). Conversely, the smaller the specific information gets, the more the relevant information must be sharing with encodings for other phenomena."
},
{
"figure_ref": [],
"heading": "The Measure of Understanding",
"publication_ref": [],
"table_ref": [],
"text": "Finally we bring these ideas together to define our measure of understanding. The resulting measure is bounded by 0 and 1. Definition 9 (Understanding). Given an agent X ∈ X with internal state z X and phenomenon p ∈ P, the measure of X's understanding of phenomenon p, κ : X × P → [0, 1), is defined as:\nκ(X, p) = κ(X, p) • c(X, p) • φ(X, p) • Υ p (X) • I(z X ; p) K(p) (11)\nWhere κ(X, p) ∈ {0, 1} is X's self-reported understanding of p."
},
{
"figure_ref": [],
"heading": "Explanation Games",
"publication_ref": [],
"table_ref": [],
"text": "With our measure of understanding, we are ready to define explanatory effectiveness:\nDefinition 10 (Explanatory Effectiveness). The effectiveness of an explanation is the change in explainee's understanding of the explanandum p ∈ P over the course of the explanatory process. Formally, given an explainer agent A and an explainee agent B that interact over τ time steps, the explanatory effectiveness is a function ξ : O * B × P → (-1, 1) defined as:\nξ(o B , p) = κ(B τ , p) -κ(B 1 , p) (12\n)\nWhere From these definitions, there are a few observations that we can make. Firstly, there is nothing to stop a game from having negative effectiveness, i.e. the explainee understands the phenomenon less after the 'explanation'. As κ is bounded by 0 and 1, ξ is bounded by -1 and 1. Secondly, there is no necessary link between effectiveness and the explainee's beliefs regarding their own understanding. It is possible for the explainee to believe that the explanation was more effective than it was (e.g. κ(X, p) = 1, but I(z X ; p) = 0). Thirdly, we can use this notion to discuss the motivation of the explainer. It makes sense to consider an agent as an explainer, rather than a deceiver, only if they expect the sign of the ξ to be positive. Finally, it is worth noting that this measure changes according to time in which we choose to record it. The explainer may start out strong and increase the explainee's understanding of the explanandum, but then say something that leads to confusion."
},
{
"figure_ref": [],
"heading": "Discussion",
"publication_ref": [
"b18"
],
"table_ref": [],
"text": "In this paper we have presented a formal model for assessing the the 'explanatory effectiveness' ξ of a dialectical process between two agents. We used this to define explanation games in which participants seek to maximise ξ. Along the way we used AIT and UIT to develop a measure of an agent's 'understanding' of a given phenomenon p. This involved partitioning the information in the agent's mental state into four objects relative to p; the p-relevant, p-irrelevant, p-specific, and p-background information. We used these to define the p-compression factor (how compressed the agent's representation of p is), p-integration (what proportion of the representation is only encoding for p), and the p-utilisation. For the last of these we needed to construct a set of 'fair tests', i.e. a set of environments that would rely on both knowledge of p and the agent's background knowledge to solve. We find these environments by asking: \"Could AIXI solve this environment when given this information?\". However, it should not be taken for granted that this is the right question to ask, and thus we should study this space of environments more precisely to see if it includes unfair tests or leaves out potential fair tests.\nFuture work should investigate the trustworthiness of explanations generated in our framework, as we have made the implicit assumption that if an agent understands something they can assess whether or not they trust it. One direction to look in is the implications of explainees with limited capacities, i.e. either time/space complexity constraints, or explainees who are biased in particular ways. Additionally, the assumption that explanation games are always cooperative should be challenged, as in many real situations participants may have conflicting or ulterior agendas. For both the cooperative and non-cooperative case a useful research project will be to articulate rules for the dialogue game between explainer and explainee (McBurney & Parsons 2002) and to develop strategies for each player, given their goals. Finally, as K and AIXI are not computable, alternatives for these components (for the purposes of this framework) should be devised and studied."
},
{
"figure_ref": [],
"heading": "",
"publication_ref": [],
"table_ref": [],
"text": "Work done by DC is thanks to the UKRI Centre for Doctoral Training in Safe and Trusted AI (EPSRC Project EP/S023356/1). PM wishes to thank Simon Parsons and Elizabeth Sonenberg for discussions on these topics. DC would like to thank Alex Jackson and Nandi Schoots for helping to understand understanding."
}
] | In most conversations about explanation and AI, the recipient of the explanation (the explainee) is suspiciously absent, despite the problem being ultimately communicative in nature. We pose the problem 'explaining AI systems' in terms of a two-player cooperative game in which each agent seeks to maximise our proposed measure of explanatory effectiveness. This measure serves as a foundation for the automated assessment of explanations, in terms of the effects that any given action in the game has on the internal state of the explainee. | A Measure of Explanatory Effectiveness * Towards a Formal Model of Explanation | [
{
"figure_caption": "p{|p| : U (p) = x}, where |p| measures the length of p(1)",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "B t denotes B at time t and o B is the sequence of observations that B made during the interaction.",
"figure_data": "Definition 11 (Explanation Game). Suppose an explainer agent A, explaineeagent B, and explanandum p ∈ P. An explanation game G = (A, B, p, τ ) isa cooperative finite sequential game with asymmetric information in which theparticipants seek to maximise ξ(o B , p) over the course of τ time steps, where o Bis the sequence of all observations made by B.",
"figure_id": "tab_0",
"figure_label": "",
"figure_type": "table"
}
] | Dylan R Cope; Peter Mcburney | [
{
"authors": "A Abdul; J Vermeulen; D Wang; B Y Lim; M Kanhanhalli",
"journal": "",
"ref_id": "b0",
"title": "Trends and Trajectories for Explainable, Accountable and Intelligible Systems: An HCI Research Agenda",
"year": "2018"
},
{
"authors": "A Adadi; M Berrada",
"journal": "IEEE Access",
"ref_id": "b1",
"title": "Peeking Inside the Black-Box: A Survey on Explainable Artificial Intelligence (XAI)",
"year": "2018"
},
{
"authors": "G Chaitin",
"journal": "Scientific American",
"ref_id": "b2",
"title": "The Limits of Reason",
"year": "2006"
},
{
"authors": "G J Chaitin",
"journal": "Journal of the Association for Computing Machinery",
"ref_id": "b3",
"title": "A Theory of Program Size Formally Identical to Information Theory",
"year": "1975"
},
{
"authors": "P M Dung",
"journal": "Artificial Intelligence",
"ref_id": "b4",
"title": "On the acceptability of arguments and its fundamental role in nonmonotonic reasoning, logic programming and n-person games",
"year": "1995"
},
{
"authors": "D H Glass",
"journal": "Springer Verlag",
"ref_id": "b5",
"title": "Coherence, Explanation, and Bayesian networks",
"year": "2002"
},
{
"authors": "A C Graesser; W Baggett; K Williams",
"journal": "Applied Cognitive Psychology",
"ref_id": "b6",
"title": "Question-driven Explanatory Reasoning",
"year": "1996"
},
{
"authors": "D R Hofstadter; E Sander",
"journal": "Basic Books",
"ref_id": "b7",
"title": "Surfaces and essences : Analogy as the fuel and fire of thinking",
"year": "2012"
},
{
"authors": "D Hume",
"journal": "",
"ref_id": "b8",
"title": "An Enquiry Concerning Human Understanding",
"year": "1751"
},
{
"authors": "M Hutter",
"journal": "Springer",
"ref_id": "b9",
"title": "Universal Artificial Intelligence, Texts in Theoretical Computer Science",
"year": "2005"
},
{
"authors": "H Johnson; P Johnson",
"journal": "",
"ref_id": "b10",
"title": "Explanation facilities and interactive systems",
"year": "1993"
},
{
"authors": "L Karsenty; P J Brezillon",
"journal": "Int. J. Expert Systems with Applications",
"ref_id": "b11",
"title": "Cooperative problem solving and explanation",
"year": "1995"
},
{
"authors": "A Kirsch",
"journal": "",
"ref_id": "b12",
"title": "Explain to Whom? Putting the User in the Center of Explainable AI",
"year": "2017"
},
{
"authors": "A N Kolmogorov",
"journal": "International Journal of Computer Mathematics",
"ref_id": "b13",
"title": "Three approaches to the quantitative definition of information",
"year": "1968"
},
{
"authors": "G Lakoff; M Johnson",
"journal": "University of Chicago Press",
"ref_id": "b14",
"title": "Metaphors We Live By",
"year": "1980"
},
{
"authors": "S Legg; M Hutter",
"journal": "Minds and Machines",
"ref_id": "b15",
"title": "Universal intelligence: A definition of machine intelligence",
"year": "2007"
},
{
"authors": "P Lipton",
"journal": "Routledge",
"ref_id": "b16",
"title": "Inference to the Best Explanation",
"year": "2003"
},
{
"authors": "P Maguire; P Moser; R Maguire",
"journal": "Journal of Cognitive Science",
"ref_id": "b17",
"title": "Understanding Consciousness as Data Compression",
"year": "2016"
},
{
"authors": "P Mcburney; S Parsons",
"journal": "Journal of Logic, Language, and Information",
"ref_id": "b18",
"title": "Games That Agents Play: A Formal Framework for Dialogues between Autonomous Agents",
"year": "2002"
},
{
"authors": "T Miller",
"journal": "Artificial Intelligence",
"ref_id": "b19",
"title": "Explanation in Artificial Intelligence: Insights from the Social Sciences",
"year": "2019"
},
{
"authors": "O Morgenstern; J Von Neumann",
"journal": "Princeton University Press",
"ref_id": "b20",
"title": "Theory of games and economic behavior",
"year": "1953"
},
{
"authors": "S T Mueller; R R Hoffman; W Clancey; A Emrey; G Klein",
"journal": "",
"ref_id": "b21",
"title": "Explanation in Human-AI Systems: A Literature Meta-Review, Synopsis of Key Ideas and Publications, and Bibliography for Explainable AI",
"year": "2019"
},
{
"authors": "S Okasha",
"journal": "Studies in History and Philosophy of Science Part A",
"ref_id": "b22",
"title": "Van Fraassen's critique of inference to the best explanation",
"year": "2000"
},
{
"authors": "D Perkins",
"journal": "",
"ref_id": "b23",
"title": "What is Understanding?",
"year": "1993"
},
{
"authors": "K Popper",
"journal": "Routledge Classics",
"ref_id": "b24",
"title": "The Logic of Scientific Discovery",
"year": "1959"
},
{
"authors": "S Russell; P Norvig",
"journal": "Pearson",
"ref_id": "b25",
"title": "Artificial Intelligence: A Modern Approach",
"year": "2010"
},
{
"authors": "R C Schank",
"journal": "",
"ref_id": "b26",
"title": "The Explanation Game",
"year": "1984"
},
{
"authors": "J N Schupbach; J Sprenger",
"journal": "Philosophy of Science",
"ref_id": "b27",
"title": "The Logic of Explanatory Power",
"year": "2011"
},
{
"authors": "E I Sklar; M Q Azhar",
"journal": "ACM)",
"ref_id": "b28",
"title": "Explanation through Argumentation",
"year": "2018"
},
{
"authors": "E Sklar; S Parsons",
"journal": "",
"ref_id": "b29",
"title": "Towards the application of argumentation-based dialogues for education",
"year": "2004"
},
{
"authors": "R J Solomonoff",
"journal": "Information and Control",
"ref_id": "b30",
"title": "A Formal Theory of Inductive Inference, Part 1",
"year": "1964"
},
{
"authors": "A M Turing",
"journal": "Mind LIX",
"ref_id": "b31",
"title": "Computing Machinery and Intelligence",
"year": "1950"
},
{
"authors": "",
"journal": "UK Information Commissioner's Office",
"ref_id": "b32",
"title": "Guide to the GDPR",
"year": "2019"
},
{
"authors": "D Walton; E C W Krabbe",
"journal": "State University of New York Press",
"ref_id": "b33",
"title": "Commitment in Dialogue: Basic Concepts of Interpersonal Reasoning",
"year": "1995"
},
{
"authors": "H Zenil",
"journal": "World Scientific",
"ref_id": "b34",
"title": "Compression is Comprehension, and the Unreasonable Effectiveness of Digital Computation in the Natural World",
"year": "2019"
}
] | [
{
"formula_coordinates": [
3,
237.24,
539.61,
243.4,
14.96
],
"formula_id": "formula_0",
"formula_text": "K(x|y) = min p {|p| : U (yp) = x} (2)"
},
{
"formula_coordinates": [
3,
254.4,
587.73,
226.24,
9.96
],
"formula_id": "formula_1",
"formula_text": "I(x; y) = K(y) -K(y|x) (3)"
},
{
"formula_coordinates": [
4,
259.8,
422.23,
220.84,
22.54
],
"formula_id": "formula_2",
"formula_text": "Υ (π) = µ∈E 2 -K(µ) V π µ (4)"
},
{
"formula_coordinates": [
5,
263.28,
628.05,
217.36,
23.96
],
"formula_id": "formula_3",
"formula_text": "c(X, p) = K(z X, * p ) | z X, * p X | (5)"
},
{
"formula_coordinates": [
6,
147.96,
328.51,
332.68,
13.55
],
"formula_id": "formula_4",
"formula_text": "E X,p = µ ∈ E : V AIXI (p,b)•µ = V * µ > 0, V AIXI (p)•µ = V AIXI (b)•µ = V AIXI µ = 0 (6)"
},
{
"formula_coordinates": [
6,
192.96,
388.5,
287.68,
26.04
],
"formula_id": "formula_5",
"formula_text": "∀x i ∈ x, ∀a <i , µ ′ ((x i , 0) | a <i ) = 1 (7) ∀k > |x|, µ ′ (e k | a <k ) = µ(e k | a j...k ), where j = |x|(8)"
},
{
"formula_coordinates": [
6,
249.72,
579.31,
230.92,
22.42
],
"formula_id": "formula_6",
"formula_text": "Υ p (X) = µ∈EX,p 2 -K(µ) V X µ (9)"
},
{
"formula_coordinates": [
7,
231,
142.29,
249.64,
42.08
],
"formula_id": "formula_7",
"formula_text": "X × P → [0, 1), is defined, φ(X, p) = tanh | z X, * p X | | z X,p X | -1(10)"
},
{
"formula_coordinates": [
7,
192.36,
379.89,
288.28,
23.18
],
"formula_id": "formula_8",
"formula_text": "κ(X, p) = κ(X, p) • c(X, p) • φ(X, p) • Υ p (X) • I(z X ; p) K(p) (11)"
},
{
"formula_coordinates": [
7,
236.64,
556.97,
239.55,
12.36
],
"formula_id": "formula_9",
"formula_text": "ξ(o B , p) = κ(B τ , p) -κ(B 1 , p) (12"
},
{
"formula_coordinates": [
7,
476.19,
558.93,
4.45,
9.96
],
"formula_id": "formula_10",
"formula_text": ")"
}
] |
|
10.1006/GAME.2000.0824 | 2023-05-20 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b11",
"b16"
],
"table_ref": [],
"text": "Typically, in the field of emergent communication a group of agents learn to interact with one another through communication channels in order to facilitate coordination in a shared environment, i.e. a Dec-POMDP (Oliehoek & Amato, 2016). The agents learn highly effective communication strategies, but they tend to be brittle in the sense that they are unable to coordinate with agents that they have not encountered before. This construction does not naturally lend itself to systems that require a machine to communicate with a human, or enter within a community of humans using language to coordinate. In this paper, we frame this as a problem of cooperative language acquisition, where the goal is to adopt the language of a community of agents so as to coordinate with them.\nMore precisely, we place the problem in the context of ad hoc team play (Stone et al., 2010). In ad hoc team play, we are given a set of competent1 agents and a domain of coordination tasks, and the problem is to design new agents that are capable of achieving success when playing with randomly sampled teammates. In our problem, we assume that there exists a community of language-users that define the pool of players who, by means of their shared language, are all successful ad hoc team players. Therefore, the cooperative language acquisition problem is defined as the case of designing a new agent to join this pool of players by observing a sample of interactions from the community."
},
{
"figure_ref": [],
"heading": "Background",
"publication_ref": [
"b9"
],
"table_ref": [],
"text": "A decentralised partially-observable Markov decision process (Dec-POMDP) is described by a 7-tuple (S, {A i }, T, R, {Ω i }, O, γ), where S is a set of states, {A i } is a set of action sets, T is a transition function, R is a reward function, {Ω i } is a set of observation sets, O is an observation function, and γ is a discount factor. We will talk of trajectories for a given agent i, which are sequences of state-action-reward tuples τ ∈ T i = (S × A i × R) * . Each agent i follows a policy π i that maps an observation sequence to a distribution over actions. We denote the distribution over future trajectories that a policy induces as π(τ |.). The return of a trajectory is computed as the discounted sum of rewards:\nV (τ ) = |τ | k=0 γ k r k\nFollowing Lowe et al. (2019), we suppose that each agent's action sets can be expressed as\nA i = A c i ∪ A e i ,\nwhere A c i is a set of communicative actions and A e i is a set of environment actions. Communicative actions are sent to a target agent by a dedicated cheap-talk channel (there is no cost to communication), meaning they appear in the receiver's observation at the next time step. We also use Lowe et al.'s (2019) definitions of positive listening and positive signalling: Definition 1 (Positive Listening). An agent i with the policy π i exhibits positive listening if there exists a message generated by a signaller j, m ∈ A c j , such that d τ (π i (τ |z, 0), π i (τ |z, m)) > 0 where 0 is a zero vector, z is a variable that conditions the policy (e.g. observations and/or latent memory), and d τ is a distance metric over T i . Definition 2 (Positive Signalling). Let m = (m 0 , . . . , m T ) be a sequence of messages sent by an agent over the course of a trajectory of length T , and similarly for observations o = (o 0 , . . . , o T ), and actions a = (a 0 , . . . , a T ). An agent exhibits positive signalling if m is statistically dependent on either a or o."
},
{
"figure_ref": [],
"heading": "One-Way Communication Problem Formulation",
"publication_ref": [
"b8",
"b5",
"b6",
"b2",
"b17",
"b12"
],
"table_ref": [],
"text": "We start to formulate the problem of cooperative language acquisition with the simplest case involving two agents: a speaker A and a listener B. For a particular interaction i the speaker emits a message m i ∈ A c * A that is received by the listener, and then the listener takes actions that lead it on a trajectory τ i ∈ T B in a Dec-POMDP sampled from a given domain. We are only considering the case of one-way communication with this set-up, but we will discuss two-way communication, i.e. dialogues, in Section 6. To make this more precise, we assume that A and B are agents sampled from a pool of players operating in an ad hoc team, where the domain D is a set of referential games, i.e. a class of Dec-POMDPs based on Lewis signalling games (Lewis, 1969;Lazaridou et al., 2017;Lee et al., 2018). We will assume that the listener is exhibiting positive listening to the messages sent by the speaker, and the speaker is positive signalling an 'intended' target trajectory2 , τ ⊙ i ∈ T B . We can denote this by saying that the observer observes:\nm i ∼ π A (m | τ ⊙ i ) and τ i ∼ π B (τ | m i )\n, where π A and π B denote the policies followed by each agent respectively.3 Before moving on, let us define a running example game to illustrate the setting. Suppose that the speaker has access to a shopping list and a map of the supermarket, and must write a note for the listener to observe, who then must retrieve the items as quickly as possible.\nThe cooperative language acquisition task is to construct an agent X, who we will call the observer, which is able to take on the roles of either the speaker or the listener and successfully communicate with others. So, if X is taking on the role of the speaker, given some τ that they intend for B to follow, they should emit a message that maximises the probability that B does so. If X is acting as the listener, and receives some m ∼ π A (m | τ ⊙ ) they should estimate τ ⊙ and follow this trajectory. With this, given a dataset of interactions between speakers and listeners, m i , τ i ∼ D AB , we can define the following sub-problems: Problem 1 (The Forward Problem (Signalling)). Find a function β(m | τ, θ β ) parameterised by θ β , which we call the Broca function, such that m maximises the probability that the listening agent B follows the trajectory τ upon receiving m, i.e.:\nθ * β = arg max θ β τ ∈TB E m∼β(m|τ,θ β ) π B (τ | m) (1)\nProblem 2 (The Backward Problem (Listening)). Find a function ν(τ | m, θ ν ) parameterised by θ ν , which we call the Wernicke function. Given the message m is from the speaking agent A and is intended to invoke the trajectory τ ⊙ , the function ν should maximise the probability of τ ⊙ .\nθ * ν = arg max θν τ ⊙ ∈TB E m∼πA(m|τ ⊙ ) ν(τ ⊙ | m, θ ν ) (2)\n4 Finding Broca and Wernicke\nFirstly, we can directly model the forward problem with the data available. We estimate the parameters θ β by mapping from observed trajectories to messages received by B:\nθ * β = arg min θ β mi,τi∈DAB d m (m i , mi )\nwhere mi = arg max\nm β(m | τ i , θ β )(3)\nGiven a distance function d m over messages. Put differently, we are aiming to find θ β such that the Broca function can produce the message that caused a given trajectory in the data.\nTo place this into our running example, we have data regarding the notes that were sent to the shopper (m i ), and paths through the shop that the shopper took (τ i ), and we are learning the relationship between notes and paths.\nHowever, the backward problem is much trickier given that we are never able to directly observe the intended trajectory τ ⊙ i for the message m i sent by the speaker. If we assume that the speaker is optimal, i.e. it always sends the perfect message to invoke the intended actions in B, then τ ⊙ i = τ i and thus we can optimise the reverse mapping as in the forward problem (i.e. messages to trajectories). But what can we do if we wish to relax this? Instead of modelling the speaker as perfectly optimal, we can assume 'soft-optimality', otherwise known as Boltzmann-rationality4 . We will do this in two parts: first, we will assume that given the target trajectory τ ⊙ , the speaker is more likely to send messages that are 'closer to optimal', for which we need some notion of semantic distance between messages. Secondly, we will assume that the speaker is more likely to pick target trajectories for the listener that yield a high return in the Dec-POMDP. Put formally: 4)\nP (τ ⊙ ) ∝ exp(V (τ ⊙ )) (\nP (m|τ ⊙ ) ∝ exp(-S B (m * B (τ ⊙ ), m))\n(5) Where, V is the expected return of a given trajectory, m * B (τ ⊙ ) is the optimal message to send to B to maximise the chance that B takes the trajectory τ ⊙ , and S B is a measure of the semantic distance between two messages for B. These latter two are defined as follows:\nm * B (τ ) = arg max m π B (τ |m) (6) S B (m 1 , m 2 ) = d τ π B (τ | m 1 ), π B (τ | m 2 ) (7)\nIn other words, the semantic distance is a function of the difference in actions that B takes (characterised by the distance function over trajectories d τ ) as a result of different messages. Thus, it is mathematically similar to Lowe et al.'s (2019) definition of positive listening, and philosophically close to the various approaches to 'meaning' that couple information and action (Haig, 2017;Wittgenstein, 1953;Peirce, 1878). Additionally, note that if we substitute these definitions into P (m|τ ⊙ ):\nP (m|τ ⊙ ) ∝ exp -d τ π B (τ | arg max m π B (τ ⊙ |m)), π B (τ | m) ∝ exp -d τ τ ⊙ , π B (τ | m) (8)\nThus the expression involving the semantic distance captures the intuition that the more optimal messages are the ones that, in expectation, lead to trajectories that are closer to the target. With our assumptions in place, we now insert this into the backward problem. For a given interaction m i , τ i ∼ D AB we can express most probable target trajectory τ ⊙ i with the maximum a posteriori estimate:\nτ ⊙ i = arg max τ ⊙ P (τ ⊙ | m i ) = arg max τ ⊙ P (τ ⊙ )P (m i | τ ⊙ ) = arg max τ ⊙ V (τ ⊙ ) -αd τ (τ ⊙ , π B (τ | m i )) = arg max τ ⊙ V (τ ⊙ ) -αd τ (τ ⊙ , τ i ) (9)\nWhere α is a hyperparameter controlling our prior on the relative optimality of the speakers ability to effectively communicate versus pick the optimal trajectory (similar to Jeon et al. ( 2020)). Therefore, to estimate the parameters of the Wernicke function θ ν :\nθ * ν = arg min θν mi,τi∈DAB (αd m (τ ⊙ i , τ i ) -V (τ ⊙ i ))\nwhere\nτ ⊙ i = arg max τ ν(τ | m i , θ ν )(10)\nAgain, let us contextualise this within the example of the shoppers. If the speaker's note contained roughly the right set of instructions, but is perhaps slightly confusing in a way that throws off the shopper (perhaps impossible directions), the Wernicke function will not emulate the shoppers confusion. Instead, the estimate of the intended trajectory can take into account the ambiguity or inconsistency and try and figure out what would be a successful path through the supermarket. Comparatively, when we assume optimality of the speaker, we are forced to conclude that the shoppers confusion was intended."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b14",
"b10",
"b18",
"b0",
"b6",
"b5",
"b1"
],
"table_ref": [],
"text": "A close area of related work is inverse reinforcement learning (Russell, 1998;Ng & Russell, 2000). Namely, the modelling of the speaker is similar to IRL, where instead of there being a hidden reward function influencing the agents' actions, there is a target trajectory for the listener. Further, the Boltzmann-rational model used is very similar to approaches in IRL (Zietbart et al., 2008;Finn et al., 2016). In the field of emergent communication, the works of Lee et al. (2018) and Lazaridou et al. (2017) both present frameworks for grounding learning agents in human natural language. They do this by using text annotated images rather than data from direct human communication in a cooperative setting. Finally, outside of AI, in the economics literature there has been work modelling how an uniformed listener may extract information from informed debaters (Glazer & Rubinstein, 2001). But because of markedly different assumptions, it does not tackle the problem of the current paper."
},
{
"figure_ref": [],
"heading": "Discussion and Conclusion",
"publication_ref": [
"b4",
"b7",
"b15"
],
"table_ref": [],
"text": "In this paper we have presented the first steps towards constructing an agent that, given data regarding the interactions of language-users, can find the meaning behind messages received, as well as optimally convey a recommended trajectory to a listener. Yet, there are still several directions of further work to be explored.\nFirstly, how do we extend the system to dialogues, i.e. two-way communication? Potentially the system naturally captures dialogues as both agents can play the roles of speakers and listeners simultaneously (or interchangeably). For instance, suppose that in the supermarket example the speaker and the shopper held a phone call, and the shopper asks a question for clarification on directions. The shopper does not have an exact intended trajectory for the speaker's response, because if they knew this they would not need to ask the question. However, this does not necessarily pose a problem for the framework presented in this paper.\nAlthough we have referred to the target trajectory as \"intentional\" it is not necessary for a speaker to know the full details of the trajectory. This applies so long as they help the listener to find the closest trajectory that maximises reward, which they may do so by adding their own private information.\nSecondly, but no less critically, is the problem of empirically testing this framework by constructing an agent. There are several suitable test environments, for example, gridworld games that are similar to the supermarket example (Kajić et al., 2020;Leibo et al., 2017), or more communication focused problems such as the game of Taboo. In this game, one person has to get another to say a hidden word, but they are forbidden from revealing certain pieces of helpful information5 . Finally, this work could be extended from the passive case of observing data, to a situation where the learner is engaged with the language-users, perhaps for example with an active learning approach (Settles, 2009)."
},
{
"figure_ref": [],
"heading": "Acknowledgements",
"publication_ref": [],
"table_ref": [],
"text": "Work done by DC is thanks to the UKRI Centre for Doctoral Training in Safe and Trusted AI (EPSRC Project EP/S023356/1).\nWe would like to thank Francis Rhys Ward, Nandi Schoots, Richard Willis, Mattia Villani and Charles Higgins for their help."
}
] | In this paper, we propose and consider the problem of cooperative language acquisition as a particular form of the ad hoc team play problem. We then present a probabilistic model for inferring a speaker's intentions and a listener's semantics from observing communications between a team of language-users. This model builds on the assumptions that speakers are engaged in positive signalling and listeners are exhibiting positive listening, which is to say the messages convey hidden information from the listener, that then causes them to change their behaviour. Further, it accounts for potential sub-optimality in the speaker's ability to convey the right information (according to the given task). Finally, we discuss further work for testing and developing this framework. | Joining the Conversation: Towards Language Acquisition for Ad Hoc Team Play | [] | Dylan R Cope; Peter Mcburney | [
{
"authors": "Chelsea Finn; Sergey Levine; Pieter Abbeel",
"journal": "",
"ref_id": "b0",
"title": "Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization",
"year": "2016"
},
{
"authors": "Jacob Glazer; Ariel Rubinstein",
"journal": "Games and Economic Behavior",
"ref_id": "b1",
"title": "Debates and Decisions: On a Rationale of Argumentation Rules",
"year": "2001"
},
{
"authors": "David Haig",
"journal": "",
"ref_id": "b2",
"title": "Making Sense: Information Interpreted as Meaning",
"year": "2017"
},
{
"authors": "Smitha Hong Jun Jeon; Anca Milli; Dragan",
"journal": "",
"ref_id": "b3",
"title": "Reward-rational (implicit) choice: A unifying formalism for reward learning",
"year": "2020"
},
{
"authors": "Ivana Kajić; Eser Aygün; Aygün Aygün; Doina Precup",
"journal": "",
"ref_id": "b4",
"title": "Learning to cooperate: Emergent communication in multi-agent navigation",
"year": "2020"
},
{
"authors": "Angeliki Lazaridou; Alexander Peysakhovich; Marco Baroni",
"journal": "",
"ref_id": "b5",
"title": "Multi-Agent Cooperation and the Emergence of (Natural) Language",
"year": "2017"
},
{
"authors": "Jason Lee; Kyunghyun Cho; Jason Weston; Douwe Kiela",
"journal": "",
"ref_id": "b6",
"title": "Emergent Translation in Multi-Agent Communication",
"year": "2018"
},
{
"authors": "Vinicius Joel Z Leibo; Marc Zambaldi; Janusz Lanctot; Thore Marecki; Graepel",
"journal": "",
"ref_id": "b7",
"title": "Multiagent Reinforcement Learning in Sequential Social Dilemmas",
"year": "2017"
},
{
"authors": "David K Lewis",
"journal": "Wiley-Blackwell",
"ref_id": "b8",
"title": "Convention: A Philosophical Study",
"year": "1969"
},
{
"authors": "Ryan Lowe; Jakob Foerster; Y-Lan Boureau; Joelle Pineau; Yann Dauphin",
"journal": "",
"ref_id": "b9",
"title": "On the Pitfalls of Measuring Emergent Communication",
"year": "2019"
},
{
"authors": "Andrew Y Ng; Stuart Russell",
"journal": "",
"ref_id": "b10",
"title": "Algorithms for Inverse Reinforcement Learning",
"year": "2000"
},
{
"authors": "Frans A Oliehoek; Christopher Amato",
"journal": "Springer International Publishing",
"ref_id": "b11",
"title": "A Concise Introduction to Decentralized POMDPs",
"year": "2016"
},
{
"authors": "Peirce Charles",
"journal": "Popular Science Monthly",
"ref_id": "b12",
"title": "How to Make Our Ideas Clear",
"year": "1878"
},
{
"authors": "Dagmar Michael Rovatsos; Gábor Gromann; Bella",
"journal": "",
"ref_id": "b13",
"title": "The Taboo Challenge Competition",
"year": "2017"
},
{
"authors": "Stuart Russell",
"journal": "ACM Press",
"ref_id": "b14",
"title": "Learning agents for uncertain environments (extended abstract)",
"year": "1998"
},
{
"authors": "Burr Settles",
"journal": "",
"ref_id": "b15",
"title": "Active Learning Literature Survey",
"year": "2009"
},
{
"authors": "Peter Stone; A Gal; Sarit Kaminka; Jeffrey S Kraus; Rosenschein",
"journal": "",
"ref_id": "b16",
"title": "Ad Hoc Autonomous Agent Teams: Collaboration without Pre-Coordination",
"year": "2010"
},
{
"authors": "Ludwig Wittgenstein",
"journal": "Macmillan Publishers",
"ref_id": "b17",
"title": "Philosophical Investigations",
"year": "1953"
},
{
"authors": "Brian D Zietbart; Andrew Maas; Andrew Bagnell; Anind K Dey",
"journal": "",
"ref_id": "b18",
"title": "Maximum Entropy Inverse Reinforcement Learning",
"year": "2008"
}
] | [
{
"formula_coordinates": [
1,
382.08,
691.37,
81.03,
14.81
],
"formula_id": "formula_0",
"formula_text": "V (τ ) = |τ | k=0 γ k r k"
},
{
"formula_coordinates": [
2,
108,
94.15,
63.73,
12.63
],
"formula_id": "formula_1",
"formula_text": "A i = A c i ∪ A e i ,"
},
{
"formula_coordinates": [
2,
338.4,
404.21,
162.31,
13.73
],
"formula_id": "formula_2",
"formula_text": "m i ∼ π A (m | τ ⊙ i ) and τ i ∼ π B (τ | m i )"
},
{
"formula_coordinates": [
2,
210.72,
589.37,
293.32,
23.05
],
"formula_id": "formula_3",
"formula_text": "θ * β = arg max θ β τ ∈TB E m∼β(m|τ,θ β ) π B (τ | m) (1)"
},
{
"formula_coordinates": [
2,
200.64,
665.33,
303.4,
23.53
],
"formula_id": "formula_4",
"formula_text": "θ * ν = arg max θν τ ⊙ ∈TB E m∼πA(m|τ ⊙ ) ν(τ ⊙ | m, θ ν ) (2)"
},
{
"formula_coordinates": [
3,
246.36,
133.01,
150.15,
22.73
],
"formula_id": "formula_5",
"formula_text": "θ * β = arg min θ β mi,τi∈DAB d m (m i , mi )"
},
{
"formula_coordinates": [
3,
283.32,
149.61,
220.72,
27.97
],
"formula_id": "formula_6",
"formula_text": "m β(m | τ i , θ β )(3)"
},
{
"formula_coordinates": [
3,
239.76,
368.33,
255.77,
11.8
],
"formula_id": "formula_7",
"formula_text": "P (τ ⊙ ) ∝ exp(V (τ ⊙ )) ("
},
{
"formula_coordinates": [
3,
228.24,
384.05,
155.44,
13.25
],
"formula_id": "formula_8",
"formula_text": "P (m|τ ⊙ ) ∝ exp(-S B (m * B (τ ⊙ ), m))"
},
{
"formula_coordinates": [
3,
213.6,
434.45,
290.44,
33.85
],
"formula_id": "formula_9",
"formula_text": "m * B (τ ) = arg max m π B (τ |m) (6) S B (m 1 , m 2 ) = d τ π B (τ | m 1 ), π B (τ | m 2 ) (7)"
},
{
"formula_coordinates": [
3,
167.04,
541.73,
337,
36.65
],
"formula_id": "formula_10",
"formula_text": "P (m|τ ⊙ ) ∝ exp -d τ π B (τ | arg max m π B (τ ⊙ |m)), π B (τ | m) ∝ exp -d τ τ ⊙ , π B (τ | m) (8)"
},
{
"formula_coordinates": [
3,
188.04,
640.97,
316,
64.49
],
"formula_id": "formula_11",
"formula_text": "τ ⊙ i = arg max τ ⊙ P (τ ⊙ | m i ) = arg max τ ⊙ P (τ ⊙ )P (m i | τ ⊙ ) = arg max τ ⊙ V (τ ⊙ ) -αd τ (τ ⊙ , π B (τ | m i )) = arg max τ ⊙ V (τ ⊙ ) -αd τ (τ ⊙ , τ i ) (9)"
},
{
"formula_coordinates": [
4,
222.6,
121.13,
198.16,
22.85
],
"formula_id": "formula_12",
"formula_text": "θ * ν = arg min θν mi,τi∈DAB (αd m (τ ⊙ i , τ i ) -V (τ ⊙ i ))"
},
{
"formula_coordinates": [
4,
220.56,
138.21,
283.48,
28.57
],
"formula_id": "formula_13",
"formula_text": "τ ⊙ i = arg max τ ν(τ | m i , θ ν )(10)"
}
] |
2023-05-20 | [
{
"figure_ref": [],
"heading": "I. INTRODUCTION",
"publication_ref": [
"b0",
"b2",
"b3",
"b4",
"b5",
"b7",
"b8",
"b9",
"b10",
"b11",
"b12",
"b14",
"b15",
"b18",
"b19",
"b22",
"b15",
"b20",
"b23",
"b24",
"b25",
"b26",
"b28"
],
"table_ref": [],
"text": "High Dynamic Range Imaging (HDRI) [1]- [3], encompassing comprehensive scene content with optimal exposure, has gained considerable attention in recent years. As a pivotal technique for computer vision, HDRI not only delivers visually appealing observations in harmony with the human visual system, but also integrates essential features for a range of downstream vision applications, including object detection [4], [5], visual enhancement [6]- [8] and semantic segmentation [9], [10]. Unfortunately, due to constraints in photography Fig. 1. Visual comparison with representative MEF methods on general and misaligned fusion scenarios. The left figure plots the ranking with these competitors under general and misaligned scenarios. Other figures shows the obvious comparison with patch-based scheme MEFSPD [11] and learningbased method MEFGAN [12].\nequipment (e.g., smartphones and single-lens reflex cameras) that capture images with limited dynamic ranges, these images experience varying degrees of luminance degradation, leading to corrupted over/under-exposed regions. As a result, Low Dynamic Range (LDR) images are plagued by color distortion and loss of detail, hindering the accurate portrayal of complete natural scenes. Consequently, the generation of well-exposed HDR images remains both a challenging and significant research topic.\nRecently, a growing number of researchers have endeavored to create cutting-edge HDRI hardware devices capable of producing an extensive range of illumination, thereby addressing the limitations inherent to traditional digital cameras [13]- [15]. Nevertheless, due to elevated production costs and suboptimal efficiency, these intricately designed devices face challenges in achieving widespread adoption in real-world applications. As an alternative, Multi-Exposure Fusion (MEF) offers an efficacious solution for generating HDR images by assimilating characteristic texture information from a collection of LDR images captured under diverse exposures. This strategy adeptly bypasses hardware-specific limitations while maintaining a lower computational cost. Within the existing literature, conventional frameworks [16]- [19] and learning-based ones [20]- [23] constitute the predominant categories of MEF techniques. Despite these advancements, MEF continues to grapple with certain hurdles that impede its overall efficacy.\nIt is imperative to highlight that current learning-based methods neglect the essential adaptive preservation in multi-exposure image fusion. Specifically, a variety of approaches utilize direct fusion rules for feature aggregation, such as summation [16] and multiplication [21]. Regrettably, these rudimentary fusion rules fall short in effectively aggregating information from LDR pairs with markedly distinct characteristics, thereby failing to preserve critical information (e.g., pixel intensity and texture details) appropriately. Furthermore, given the considerable variation in multi-exposure image distributions, manually designed architectures face difficulties in flexibly adapting to disparate data distributions. In addition, due to the unavoidable movements and shaking of imaging devices, minor pixel misalignments in LDR pairs are commonplace. Existing methods seldom tackle this issue, leading to fused images characterized by blurred details and compromised structure. In essence, there is an urgent demand for an all-encompassing, robust, and efficient learning approach that not only delivers promising visual realism enhancement but also ensures high efficiency and stability across a wide array of scenes.\nTo be more specific, the second challenge stems from the computational efficiency of existing methods. Present MEF methods, encompassing both conventional and learningbased approaches, rely heavily on handcrafted architectures and operations. In terms of conventional frameworks, various transformations such as wavelet transform [24], multi-scale representation [25], and Laplacian pyramid [26] are proposed to enable feature fusion through handcrafted mechanisms. Unfortunately, these manual designs for feature extraction and fusion rules demand substantial fine-tuning and a significant amount of experiential knowledge. Most of these methods utilize numerical optimization, which in turn impacts the inference efficiency and robustness in real-world applications. Furthermore, in recent years, the powerful feature extraction capabilities of CNN-based learning have led to the increasing dominance of end-to-end models in MEF development, considerably improving performance concerning statistical metrics and visual effects. For architectural construction, a variety of learnable mechanisms [27]- [29] have been introduced to forge connections between LDR pairs and HDR outputs. We contend that current neural architectures for MEF largely borrow effective practices from other vision tasks without paying adequate attention to MEF-specific characteristics. As a consequence, these simplistic cascaded architectures, marked by increased width and depth, possess an excessive number of parameters, making them prone to feature redundancy."
},
{
"figure_ref": [],
"heading": "A. Contributions",
"publication_ref": [],
"table_ref": [],
"text": "To partially overcome these limitations, we present a versatile architecture search-based multi-exposure fusion approach. Specifically, we first develop a fusion-centric hyper-architecture, adhering to two primary principles: selfalignment and detail repletion. We initially propose scene relighting to map source images onto the same illumination, enhancing over/under-exposed details and providing improved support for subsequent feature aggregation. We introduce deformable alignment to achieve accurate feature registration, minimizing artifacts and blurring. Subsequently, we propose the detail repletion module to refine the coarse fusion results, leading to richer texture details. Next, we establish a flexible search space encompassing more effective operations, enabling the selection of an optimal network architecture. Considering hardware latency constraints, we employ a differentiable architecture search to automatically discover a compact and efficient model for image fusion. Consequently, our method effectively produces vivid colors, abundant details, and ghostingfree results, as intuitively illustrated in Fig. 1. Our core contributions can be summarized as follows:\n• Tackling pixel misalignment and detail enhancement as critical components of multi-exposure image fusion, we introduce a self-alignment strategy that combines robust registration and detail repletion to effectively preserve texture. This approach adeptly mitigates artifacts while maintaining the intricate structure inherent to the source images. "
},
{
"figure_ref": [],
"heading": "II. RELATED WORKS",
"publication_ref": [],
"table_ref": [],
"text": "In this part, we first briefly overview the related literature on multi-exposed image fusion, which contain two representative categories of methods, i.e., traditional numerical schemes and learning-based networks. Then we introduce the development of related architecture search methods."
},
{
"figure_ref": [],
"heading": "A. Traditional Numerical Schemes",
"publication_ref": [
"b29",
"b30",
"b31",
"b32",
"b15",
"b33",
"b34",
"b36",
"b34",
"b35",
"b37",
"b38",
"b39",
"b40"
],
"table_ref": [],
"text": "In the past decades, various handcrafted numerical strategies are proposed to achieve multi-exposure fusion. These schemes can be roughly divided into transform-based, gradientbased, weighting-based, and patch-based methods. In detail, transform-based schemes first extract features and introduce fusion strategies based on the informativeness measurement. Diverse multi-scale transformations are developed to comprehensively utilize the principled features from each scale, such as wavelet transform [30], contourlet transform [31], pyramid transform [32] and dense invariant transform [33]. For instance, guided filters [16] is proposed to decompose images into base and details parts on the spatial domain. By introducing the weighted average technique, this method can fuse the consistent feature comprehensively for various fusion scenarios, including multi-exposure, multi-focus, and multimodal image fusion. Another representative technique is the Gaussian pyramids transform [34], which fuse source images to enhance the under/over-exposed regions progressively.\nPatch-based methods [35]- [37] are robust for the fusion scenarios but suffer from artifacts and blurred boundaries. Ma et al. [35], [36] introduced a patch-based method to measure the structural information and use the decomposition strategy to extract the richest features to form the fused images. Kou et al. [38], [39] present the gradient-domain smoothing to realize edge preservation instead of Gaussian smoothing and avoid the inference of halo artifacts. Furthermore, tone-mapping-based methods are developed to achieve HDR construction with various LDR images. Sparse representation methods [40], [41] are widely utilized for multi-exposure image fusion. These schemes utilize the overcomplete dictionary to capture the features of source images and fuse the features utilized by the corresponding sparse coefficients. These methods benefit from the extraction of sharp structures and abundant textures. Traditional numerical schemes rely on the handcrafted feature extraction and inflexible fusion strategies. In this way, traditional schemes cannot adequately perform for challenging fusion scenarios (e.g., extreme exposure variation). Moreover, these schemes are also limited by the huge computation resource and the fusion performance can be reduced drastically when facing large exposure intervals."
},
{
"figure_ref": [],
"heading": "B. Learning-based Schemes",
"publication_ref": [
"b41",
"b42",
"b43",
"b44",
"b27",
"b26",
"b20",
"b45",
"b46",
"b11",
"b21",
"b28",
"b45",
"b11"
],
"table_ref": [],
"text": "With the flourishing progress of the deep learning paradigm, learning-based methods realize the promising improvement in the quantity and quality of multi-exposure fusion tasks, compared with traditional methods. Supervised by the MEF-SSIM metric, DeepFuse pioneered the first learning framework to aggregate the luminance components, and utilized a weighted fusion strategy to fuse color and brightness components. Ma et al. proposed MEF-Net [42] to predict the weighted map by feeding down-sampled images. Zhang et al. presented the network IFCNN [43] to adopt the element-wise feature fusion based on the features extracted from two independent branches. The above networks either trained by unsatisfied metric (e.g., MEF-SSIM [44]) or based on the local pixel-wise feature fusion, are easy to result in the color distortion and global structural inconsistency. On the other hand, There are various unified learning-based schemes to uniformly address diverse image fusion tasks. Zhang et al. proposes a densenet with the squeeze and decomposition principle, called SDNet [45] to realize the versatile image fusion framework. A multi-decoder-based framework is introduced with a shared encoder to realize the unified fusion [28]. U2Fusion [27] utilizes the feature similarity based on the gradient to measure the differences between source and fusion images. These versatile fusion methods pursuit discoverer the similarity of tasks, inevitably lacking task-oriented consideration, thus leading to color distortion and structural detail degradation.\nLately, attention mechanisms are widely used for multiexposure fusion. Liu et al. propose a hierarchical attention module [21], [46] to investigate the sufficient information on both under/over-exposed images. Yan et al. [47] introduce the dual spatial attention module to remove the ghosts and misalignments of adjoint frames. Xu et al. [12] introduce the non-local self-attention block to capture the long-range dependency of all regions with a generative adversarial network. Similarly, Li et al. design various attention modules [22] (e.g., coordinate and self-attention) to extract the texture details from source images. Han et al. [29] decouple the MEF task into two deep perceptual enhancements including content detail extraction and color correction. We emphasize that the mainstream learning-based schemes put amount of effort to design attention modules (e.g., hierarchical attention [46], nonlocal attention [12]) to realize the visual-pleasant realistic color correction. However, these schemes mostly utilize the heuristic attention architectures, which cannot achieve the adaptive feature extraction among diverse fusion scenes and suffer from the slow inference time with abundant parameters. Thus, the texture information from source images cannot be sufficiently investigated to generate reasonable results."
},
{
"figure_ref": [],
"heading": "C. Neural Architecture Search",
"publication_ref": [
"b47",
"b48",
"b49",
"b50",
"b51",
"b52",
"b5",
"b53",
"b54",
"b55"
],
"table_ref": [],
"text": "The appearance of Neural Architecture Search (NAS) takes the network design to enter into a new phase. Current NAS techniques can be categorized as three mainstream methodologies, including reinforcement learning [48], [49], evolutionary algorithms [50], [51], and differentiable search methods. The classical reinforcement learning-based and evolutionary algorithms are limited by massive computations with unaffordable resources. Lately, imposing continuous relaxation into the architecture representation, differentiable search strategy [52], [53] has attracted plenty of attention and widely leveraged for various image enhancement tasks. For instance, Liu et al. [6] introduced the Retinex theory to composite cooperative architecture search for low-light enhancement. Zhang et al. [54] construct the hierarchical super-net with primitive search space and strategy to discover the suitable architecture for image restoration. Li et al. [55] proposes the all-in-one architectures of extracting the shared image features for diverse image restoration using a series of physics principles-guided operations. As for image fusion, cell-level and operation-level search spaces [56] are proposed to address infrared-visible image fusion. We argue that current search strategies have obtained unprecedented attention, but mostly ignore the taskspecific module design, efficient search space, and the tradeoff between inference time. Thus, in this manuscript, we elaborately design principled super-net and search space with introducing hardware-sensitive constraint, aiming to provide a compact architecture for MEF tasks."
},
{
"figure_ref": [
"fig_0",
"fig_0"
],
"heading": "III. PROPOSED METHOD",
"publication_ref": [
"b56",
"b57",
"b58",
"b59"
],
"table_ref": [],
"text": "We first introduce a robust multi-exposure fusion framework to address misalignment between source images and visual aesthetics of fused images. Then we introduce a hardwarelatency constrained architecture search with corresponding loss functions to discover the nimble network for fast inference. Concrete components are schematically illustrated in Fig. 2.\nA. Robust Multi-Exposure Fusion Framework 1) Self-alignment Module: As aforementioned, in realworld, the misalignment of image pairs caused by device shaking and movement is almost inevitable. On the other hand, due to the extreme exposure intervals of pairs, it is untoward to straightforwardly utilize alignment techniques (e.g., optical flow [57], registration module [58], [59]), which may produce texture artifacts under inexact alignment. Thus, we conquer this obstacle into two steps: scene relighting for illumiation correction and deformable aligning for feature registration.\nScene Relighting Sub-Module: In essence, based on the Retinex theory, we propose a recurrent adaptative attention mechanism for scene relighting, aiming to push the images into similar illumination domain. We introduce two Scene Relighting Sub-Modules (SRSM) for each image to restrain the degree of illumination of source images into the similar domain for following alignment and detail enhancement. Furthermore, instead of targeting to restore the normal-light scene from single sources, SRSM aims to leverage the illumination map to preserve the comprehensive structures.\nDenoted the intermediate results as I S U and I S O and SRSM as S, the illumination correction can be formulated as\nI S i = I i ⊗ S(I i ), i ∈ {U; O},(1)\nwhere ⊗ denotes the element-wise multiplication. O and U represent under/over exposed images. Noting that, we exploit a recurrent gradual scheme to cascade SRSM, aiming to realize the progressive illumination correction. The stage-wise attention maps can benefit the procedure of complementary feature learning fully and elaborately. As shown at left part of Fig. 2, rather than utilizing heuristic handcrafted methodology, we leverage the differentiable architecture search to construct this module for fast sceneadaption. In detail, we first utilize one 3 × 3 convolution to transfer image into feature domain. Then we set two candidate operations to extract scene features. Max pooling and average pooling are hierarchically embed to realize the amplification of salient feature for illumination estimation completely. Then we leverage one undetermined convolution layer to boost the information richness of features and utilize one 3 × 3 convolution with sigmoid function to generate three-channel illumination map with range [0,1].\nDeformable Aligning Sub-Module: Few multi-exposure fusion methods consider the misalignment of source images, which are based on pre-registered pairs. However, in the realworld scenes, misalignment of over/under exposed images would damage the visual quality with serious ghosts artifacts, due to the movement of image devices or targets. Moreover, introducing learning-based optical flow methods would lead to the huge computation of pixel motion. The lack of real optical flow as ground truth for pre-training limits their performance. Thus, we introduce the Pyramid, Cascading and Deformable Convolution (PCD) mechanism [60] to establish Deformable Aligning Sub-Module (DASM) based on the supervision of visual quality metrics. We only consider DASM under the misalignment scenario.\nSpecifically, DASM first employs diverse strided convolution to generate pyramid features F U and F O based on the intermediate results from SRSM, we utilize deformable convolutions to conduct the feature-level alignment by coarseto-fine manner. Denoted the DASM as A, we can obtain the comprehensive feature as\nF A = A(F U , F O ) + F O ,(2)\nwhere F A represents the fused features based on the summation of aligned source features. Similarly, instead of introducing the original PCD network, we employ architecture search scheme to rebuild the structure (i.e., replacing different kernels of deformable convolutions) to accommodate itself into multiexposure fusion task.\n2) Detail Repletion Module: Then we introduce Detail Repletion Module (DRM) to enhance the textural details of complementary feature. In order to preserve the spatial structures, we utilize successive structure under the same resolution to promote the information richness. Specifically, inspired by effective residual learning mechanisms (e.g., residual dense blocks, dilated dense block and residual on residual), we introduce a residual operator-based search space to discover a suitable dense structure. We also introduce the attention mechanism to finally address the global color distortion. Thus, denoted the network as R and output as y, we can formulate the optimization procedure as\ny = F A R(F A ).(3)\nIn a word, DRM not only targets to strengthen feature representation of details from the fused features, but also protect the integral normal illumination. Specifically, we employ four candidate operators to composite this module. Lastly, we utilize one 3 × 3 convolution layer with sigmoid function to estimate the illumination map. In the following, we will discuss the concrete strategy to search compact MEF framework."
},
{
"figure_ref": [
"fig_0"
],
"heading": "B. Automatic Architecture Construction",
"publication_ref": [
"b5",
"b51",
"b53",
"b51"
],
"table_ref": [],
"text": "In this part, we introduce the detailed search space and strategy for the light-weight effective architectures.\n1) Principle-driven Search Space: Different from recent NAS-based schemes [6], [52], [54], which introduces the single operators (e.g., one-layer convolution and primitive pooling operations) to composite the search space, without the deep investigation of principles for module-related characteristics, we construct the principle-driven search space. As shown in the bottom part of Fig. 2, normal convolutions (denoted as \"C\") and dilated convolutions (denoted as \"DC\") with different kernel size k × k, k ∈ {1, 3, 5, 7} are utilized for the SRSM, which are consisted by three layers of convolutions for feature representation and dimension changing. In order to persevere the sufficient features to recover the complementary information, we add the skip connection to establish the residual learning, which are denoted as \"RConv\" and \"RDConv\" respectively. Similarly, DASM also can be searched using three kinds of deformable convolutions, denoted as \"3-DeC\", \"5-DeC\" and \"7-DeC\" respectively.\n2) Compact Architecture Search: In this paper, following with the continuous relaxation [52], we introduce the architecture weight α to connect the operators from search space O for the super-net construction. The continuous relaxation from layer i to layer j is formulated as:\nF j = O i→j (F i ); O i→j = O∈O α i→j O(F i ),(4)\nwhere the relaxation operator is denoted as O and O∈O α i→j = 1. In order to obtain the desired architecture with high performance and fast inference time, we also establish the continuous relaxation with operation latency. In this way, we can obtain the inference time of this super-net:\nR(α; LAT) = M O∈O α i→j LAT(O),(5)\nwhere M denotes the number of search blocks. Thus, we introduce the summation of operation latency R(α; LAT) as the constraint for architecture search objective, which can be expressed as:\nmin α val (α; ω * ) + ηR(α; LAT),(6)\nwhere val and ω * are the validation loss and optimal parameters based on the training data. Introducing the differentiable search strategy, we conducted the search of whole super-net."
},
{
"figure_ref": [],
"heading": "C. Loss Functions",
"publication_ref": [
"b60",
"b61"
],
"table_ref": [],
"text": "Focusing on the texture details preservation, color information promotion and global scene consistency, we leverage three categories of loss functions to train the proposed network, including pixel-intensity loss Int , gradient loss Gra and global-adversarial loss Dis by supervised learning with ground truth y gt . On the whole, the total loss measurement Total is denoted as:\nTotal = Int + β 1 Gra + β 2 Dis .(7)\nwhere {β 1 , β 2 } are a series of trade-off parameters.\nIn order to realize the same intensity distribution with ground truth (denoted as y gt ), we impose the 1 distance to measure the discrepancy, which can be formulated as:\nInt = 1 HW y -y gt 1 ,(8)\nwhere H, W denote the height and width of image. Due to the interference by noises and corrupted exposures, source images lack partial details. We utilize the Sobel operator to preserve the fine-grained texture details.\nGra = 1 HW ∇y -∇y gt 2 ,(9)\nBecause of the information deficiency of local region, it is untoward to obtain the global consistency of color distribution. Thus, we introduce the discriminator D from PatchGAN [61] to judge the generated results with a global activation map. By this constraint, color distribution of whole scene can be guaranteed. We introduce the gradient-penalty wasserstein training strategy [62] to conduct the generative adversarial learning. Dis is formulated as:\nE x∼Pfake D(y)-E x∼Preal D(y gt )+ηE x∼Pfake [( ∇ y D(y) 2 -1) 2 ].(10)\nIV. EXPERIMENTAL RESULTS In this section, we first introduce the detailed configurations of the architecture search and training procedure. Then we conduct the subjective and objective comparisons on general multi-exposure image fusion and misaligned multi-exposure image fusion with eleven methods, which demonstrates the remarkable performances and robust generalization ability of the proposed method."
},
{
"figure_ref": [],
"heading": "A. Search and Training Configurations",
"publication_ref": [
"b62",
"b34"
],
"table_ref": [],
"text": "1) Datasets: We used the widely-used SICE dataset [63] to train and evaluate the performance of our network. This dataset contains diverse sequences of scenes with varying exposure ratios. Each sequence has a well-exposed ground truth. For our general multi-exposure image fusion task, we randomly selected 258 pixel-level registered pairs for training and 100 pairs for testing with a significant exposure difference from each sequence. For the misaligned multi-exposure image fusion task, we selected 100 pairs with noticeable unregistered pixels to create a dataset for misaligned scenarios. Additionally, we introduced a dataset [35] without ground truth to verify the generalization ability of the network. 2) Architecture Search: Specifically, different from the original architecture weight sharing, each search bock is with the unique architecture weights to integrate the operations from diverse sub-search spaces. We utilize 100 pairs from the training dataset to divide the search-training and validation datasets homogeneously. Firstly, the whole super-net is pretrained with 10 epochs to obtain well initialized ω. Then we conduct the differentiable architecture search with 300 epochs. SGD optimizer and cosine delay schedule with initial learning rate 3e -4 are introduced to optimize the neural parameters. Adam optimizer is introduced to update the architecture with learning rate 1e -4 . The η at Eq. ( 6) is empirically set to 0.5 to balance the performance and inference time (denoted as \"Ours\"). The faster version is also provided based on the constraint η = 1 and denoted as \"Ours * \". Moreover, 80 unregistered pairs are utilized to search for the specific network for the misaligned scenarios. Int is defined as the training and validation loss for architecture search."
},
{
"figure_ref": [],
"heading": "3) Network Training:",
"publication_ref": [
"b63",
"b64"
],
"table_ref": [],
"text": "The β 1 and β 2 of Eq.( 7) are set to 0.75 and 0.05 respectively. Data augmentation, such as the random crop, horizontal and vertical flipping, rotating are utilized for the training procedure. With patch size of 128 × 128, we train the network 2000 epochs. This network is trained with Adam optimizer and introduce the cosine annealing strategy to delay the learning rate from 1e -4 to 1e -10 progressively.\nWe introduce two categories of metrics to measure the visual quality of generated results, including the referencebased measurements (PSNR and SSIM) and visual perception metrics (LPIPS [64] and FSIM [65]). PSNR measures the differences of pixel intensity between outputs and ground truths. SSIM can provide the similarly measurement from the luminance, contrast and structure aspects. LPIPS utilizes the deep features to measure the perceptual similarity incline with human visual perception. FSIM is also used for the evaluation, which represents the salient low-level features based on phase congruency and gradient magnitude."
},
{
"figure_ref": [
"fig_1",
"fig_2",
"fig_3"
],
"heading": "B. General Multi-Exposure Image Fusion",
"publication_ref": [
"b10",
"b36",
"b32",
"b16",
"b19",
"b42",
"b44",
"b26",
"b11",
"b45",
"b28",
"b34"
],
"table_ref": [],
"text": "To demonstrate the effectiveness and remarkable advantages of the proposed methods, we comprehensively compare our methods with eleven competitors including the traditional methods, i.e., MESPD [11], FMMEF [37], MEFDSIFT [33], DEMEF [17] and deep learning methods, i.e., DeepFuse [20], IFCNN [43], SDNet [45], U2Fusion [27], MEFGAN [12], HALDER [46] and DPEMEF [29]. Then we evaluate the proposed scheme with these methods from three aspects, i.e., the objective, subjective comparison, and computation complexity.\n1) Subjective Comparison: We select two representative pairs to demonstrate the superiority of our proposed method, which are shown at Fig. 3 and Fig. 4. In the below part of each image, we also illustrate related signal maps of the line marked by \"blue\" compared with ground truth, to show the obvious differences. Since these image pairs have a challenging distinct exposure gap, it is untoward to design the multi-exposure fusion scheme to preserve the suitable brightness, structural details, and abundant color distribution. From these noticeable comparisons, we can conclude our methods contain three significant advantages.\nFirstly, our scheme realizes the consistency of global brightness. Patch-based schemes (e.g., MESPD, DEMEF, and FM- MEF) cannot achieve brightness consistency, which introduces significant artifacts, caused by the extreme difference in exposure time. From the signal maps of DEMEF and FMMEF at the first scene, the fused results are over-exposure with higher pixel intensities compared with ground truth. Meanwhile, the sky is degraded by extreme-low exposure time, shown in the local under-exposed regions. In contrast, our method can recover the promising brightness with normal illumination. Secondly, current learning-based schemes are easily trended to color distortion. For instance, SDNet, U2Fusion, and HALDER methods cannot realize the vivid color details, including the bushes in the first scene and the flowers in the second scene. This illustration is also reflected in the corresponding signal maps. These methods cannot achieve large signal changes and are with a moderate reflection of pixel intensity. Our method and MEFGAN can preserve the promising color distribution with remarkable improvement. More importantly, the maintenance of texture details is a critical goal for this fusion task. From the two blew-up regions, it is obvious to observe that, most fusion schemes can not maintain sufficient details, e.g., the cliff painting and tower.\nEspecially in the second scene, corrupted by a large difference in exposure time and severe illumination, it is difficult to recover the concrete details by simple feature aggregation and direct fusion principles. Our results are visual-friendly, which is an incline with the human vision system.\nOn the other hand, we also provide another comparison based on the dataset [35] without ground truth in Fig. 5. We select two pairs with extreme exposure variance to illustrate the effectiveness of our scheme with nine state-ofthe-art multi-exposed image fusion. As shown in the first sequence, the information (e.g., cloud layer) under low exposure cannot be recovered clearly. Thus, these details are hard to be highlighted and recovered from under-exposed images (e.g., SDNet and U2Fusion marked by the green boxes). We can clearly observe that attention-based methods (MEFGAN, HALDER, and Ours) can achieve abundant detail preservations. Especially, our network can effectively promote visual perception to render sufficient details. Furthermore, our method can accomplish vivid color enhancement. Most of the results appear in the local over-exposure region. In contrast, our method achieves abundant texture details (e.g., bushes) and consistent color distribution (e.g., computer screen). Either existing patch-based optimization methods or learning-based schemes cannot address the global consistency of illumination. "
},
{
"figure_ref": [],
"heading": "2) Objective Comparison:",
"publication_ref": [],
"table_ref": [],
"text": "To demonstrate the superiority of the proposed scheme, we utilize four different metrics, including PSNR, SSIM, LPIPS, and FSIM to measure the visual quality of diverse methods. The whole numerical results are reported in Table . I. We introduce two versions to conduct the comparison, which are named \"Ours\" and \"Ours * \" respectively. The difference between both versions is utilizing diverse latency constraints, where \"Ours * \" focuses more on the inference time. Obviously, our scheme achieves consistent performance improvement in terms of these metrics. Compared with representative learning-based schemes MEFGAN and HALDER, our scheme promotes 0.8 dB and 1.0 dB drastically. On the other hand, it can be clearly seen that we obtain the second best numerical results. However, the patch-based fusion scheme obtains the suboptimal numerical performance, which indicates our scheme can effectively preserve sufficient textural details and structural information. We also utilize the LPIPS to measure the distortion at feature levels. Our scheme can reduce almost 24.6% of LPIPS compared with HALDER, which demonstrates better visual quality in line with the human perception system. Obviously, our scheme also achieves the best results.\n3) Computation Efficiency Analyses: We also conduct a comparison under computation efficiency, which is a critical point for real-world deployment. The concert numerical results among these competitors under the metrics of parameters and runtime on the SICE dataset are reported in Table . II. Obviously, most learning-based methods achieve fast inference time than traditional numerical-based schemes (e.g., DEMEF and MEFDSIFT). Furthermore, though SDNet and IFCNN have fewer parameters, and are limited by direct fusion rules and inefficient architectures, their performances are interfered. More importantly, both our methods realize the faster inference time. Compared with the latest learning-based method DPEMEF, the fastest version (ours * ) can effectively promote 68.11% runtime and reduce 98.68% parameters, which demonstrates high efficiency with visual-pleasant fused results.\n4) Fusion with Arbitrary Exposure Ratios: In order to verify the generalization ability to address the inputs with arbitrary exposure ratios, we select two representative sequences, which is shown in Fig. 6. Since we utilize the pairs with the Fig. 6. Visual results about the source image sequences with different exposure ratios. ( 1) and ( 2) are under-exposed images, (3), ( 4) and ( 5) are over-exposed images. (a)-(c) are fused by inputs of ( 1) and ( 3)-( 5). (d)-(f) are fused by inputs of ( 2) and ( 3)-( 5)."
},
{
"figure_ref": [],
"heading": "Under Exposure",
"publication_ref": [],
"table_ref": [],
"text": "Under Exposure\nOver Exposure Over exposure ratios for training and testing, the proposed scheme is robust enough to handle different exposure ratios. These sequences contain two under-exposed images and three over-exposed images. We can clearly observe that the scheme can obtain the consistent visual-pleasant fused results with natural color correction, generated by the image pairs with diverse exposure ranges. Several characteristics can be found from the fused result of this sequence. Firstly, the proposed scheme is with large capacity for wide exposure difference to preserve the textural details and maintain suitable illumination distribution. Secondly, though without the training procedure of small exposure difference, the fused images generated by (a) and (b) contain sufficient scene details, e.g., grasses and sunset glow. Finally, our result (i.e., (c)) fused by the large difference of exposure is close to the ground truth. Several fused images (e.g., (b) and (d)) have more vivid scene representation compared with ground truth."
},
{
"figure_ref": [],
"heading": "C. Misaligned Multi-Exposure Image Fusion",
"publication_ref": [],
"table_ref": [],
"text": "Misaligned multi-exposure image fusion is a challenging scenario due to the camera movement and device shaking, whereas current methods for MEF are easy to generate blurs without the consideration of pixel registration. We also illustrate the performance by numerical and visual comparisons."
},
{
"figure_ref": [
"fig_4"
],
"heading": "1) Visual Comparison:",
"publication_ref": [],
"table_ref": [],
"text": "We select four misaligned pairs from the testing dataset to demonstrate the effectiveness of proposed framework in Fig. 7. The first two rows illustrate the scene with large pixel movements. Most methods cannot preserve sufficient details. The results generated by MEF-GAN and FMMEF methods have obvious artifacts and cannot preserve the normal color distribution. More importantly, our scheme successfully realizes the uniform promotion of pixel alignment and visual enhancement, which can effectively address diverse levels of pixel alignment. 2) Numerical Comparison: In Table . III, we report the numerical results compared with various representative multiexposure image fusion methods. We utilize five typical metrics to measure the visual quality of fused images. Since existing methods often assume the image pairs are well-registered, these methods cannot obtain promising quantitative results. We can observe that the proposed method achieves the consistent best numerical results in terms of these four metrics. These results demonstrate the superiority of our methods on the misaligned MEF. "
},
{
"figure_ref": [],
"heading": "V. ABLATION STUDY",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we conduct sufficient experiments with numerical and visual evaluations to verify the effectiveness of proposed modules, loss functions and architecture search."
},
{
"figure_ref": [
"fig_5",
"fig_5"
],
"heading": "A. Effectiveness of Scene Relighting.",
"publication_ref": [],
"table_ref": [],
"text": "In this part, we first validate the effectiveness of proposed SRSM and validate the suitable cascaded numbers for MEF task. The role of scene relighting is to gradually preserve the scene information and constrain the level of illumination for following feature aggregation. The ablation experiment about the cascaded number of SRSM is conducted, where the quantitative and qualitative comparisons are shown at Table . IV and Fig. 8 respectively. Firstly, we illustrate the necessary of proposed SRSM. The version without SRSM only concatenate the inputs to feed into the DRM, without the procedure of illumination adjustment. We can clearly observe that, directly processing compromises the image quality, which reduce the numerical performance drastically. As shown in Fig. 8, we can obtain the output image is over-exposed, cannot render sufficient detail and preserve the normal light distribution. Compared by pie charts, which depicts the proportion of RGB channels. The version w/o SRSM cannot restore the normal color distribution, which leads to the distortion of color and details. Then we evaluate the cascaded number of SRSM. By introducing the cascaded SRSMs, we propose the recurrent attention mechanisms to extract the sufficient features. Cascading two modules can achieve the best numerical performance. Increasing the number of SRSM obtain the moderate improvement. "
},
{
"figure_ref": [
"fig_4"
],
"heading": "B. Effectiveness of Deformable Alignment",
"publication_ref": [
"b56"
],
"table_ref": [],
"text": "We further evaluate the advantages of self-alignment mechanism. Three variants of comparisons are conducted, including \"w/o DASM\", optical flow-based and changing the position before relighting. Self-alignment module targets to align the unregistered pixels of image pairs. In the architecture construction, we put DASM after the SRSM. Optical flow-based schemes is to firstly utilize the RAFT [57] to align image pairs. Numerical results and visual comparisons are depicted at Table . V and Fig. 7 respectively. From the numerical results, we can observe the effectiveness of proposed mechanisms. Our scheme improves 2.65dB of PSNR, 23.55% of VIF and 7.16% of FSIM. From the visual comparison, the fused result of \"w/o DASM\" cannot preserve enough texture details, such as the grasses. On the other hand, we can see that, the optical flow-based schemes cannot improve the quantitative results due to the inaccurate motion estimation caused by different illumination. The fused images contains more obvious artifacts. Meanwhile, we also discuss the suitable position of relighting. Obviously, the variant \"Before SRSM\" obtains the sub-optimal numerical and visual results compared with the final scheme, but has remarkable performance promotion compared with previous two variants. Furthermore, based on the similar intensity distribution by SRSM, the final scheme can realize the consistent improvement. "
},
{
"figure_ref": [
"fig_6",
"fig_6",
"fig_6"
],
"heading": "C. Training Losses Analyses",
"publication_ref": [],
"table_ref": [],
"text": "also perform the detailed analyses to evaluate the effectiveness of diverse training strategies (i.e., combinations of loss functions). In this part, we gradually introduce the loss functions to composited three schemes, including \"w/ Int \", \"w/ Int + Gra \" and our scheme. Related visual results are plotted at Fig. 10. From the objective comparison, Gra can effectively preserve the edge information, which reflects on the structural measurement SSIM and feature-level metric LPIPS. As shown in Fig. 10, visual results scheme \" w/ Int + Gra \" provide flourishing textural details, e.g., the details of grasses and floors. Meanwhile, introducing Gra can effectively remove the artifact such as the shape of tree on the first row. Our scheme is combined with three categories of losses, i.e., Int for pixel intensity, Gra for structural detail and Dis for color distribution. Thus, our scheme can further improve the visual quality, obtaining with the highest numerical results. For instance, our final scheme obtain the vivid color distribution, without any color distortion (e.g., the color of wall), shown at the second row in Fig. 10."
},
{
"figure_ref": [],
"heading": "Inputs",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Int",
"publication_ref": [],
"table_ref": [],
"text": "Int+ Gra Total "
},
{
"figure_ref": [],
"heading": "D. Search Space Analyses",
"publication_ref": [],
"table_ref": [],
"text": "We also verify the basic properties of search space, where the concrete performances are reported at Table. VII. From the fusion performance, we can observe that, dilated convolutions (3-DC and 7-DC) have higher quantitative results (e.g., PSNR, LPIPS and FSIM) compared with normal convolutions. On the other hand, we can directly observe that 3 × 3 convolution has the fast inference speed but has the sub-optimal statistical results. Under the constrained of hardware latency, our searched scheme actually achieve the balance between visual quality and inference speed."
},
{
"figure_ref": [],
"heading": "E. Hardware-sensitive Analyses",
"publication_ref": [],
"table_ref": [],
"text": "We also analyse the influence of trade-off parameter η, which controls the influences of hardware-sensitive latency constraint. The numerical results are reported in Table . VIII. \"C32\" and \"C64\" denote the version with 32 and 64 channels. When η = 0.5, the balance between fusion quality and inference requirement can be guaranteed simultaneously. The concrete architectures under diverse η are plotted in Fig. 11. The previous three layers illustrate the architecture of SAM. The last four layers shows the structure of DRM with residual connection. In detail, without the constraint of latency, NAS scheme chooses operator with large receptive field to better capture the large features. Moreover, we also can conclude that 3 × 3 convolution can effectively extract features with high efficiency, which is widely leveraged for SRSM under the latency constraint. As for DRM, {3-C, 1-C} with skip connection is a low-weight combination for the detail compensation, as shown at the subfigure (c) and (d). "
},
{
"figure_ref": [],
"heading": "VI. CONCLUDING REMARKS",
"publication_ref": [],
"table_ref": [],
"text": "In this paper, we proposed a robust multi-exposure image fusion framework to address various scenarios, including the aligned and misaligned image pairs. We divided the fusion procedure into two parts: self-alignment for feature-wise alignment and detail repletion to enhance texture details visually. By utilizing a hardware-friendly architecture search strategy and incorporating a task-oriented search space, we discovered a highly efficient and compact architecture for MEF. Furthermore, we conducted comprehensive subjective and objective comparisons to demonstrate the outstanding performance of our method compared to various state-of-the-arts."
},
{
"figure_ref": [],
"heading": "",
"publication_ref": [],
"table_ref": [],
"text": "The source code will be available at https://github.com/LiuZhu-CV"
},
{
"figure_ref": [],
"heading": "",
"publication_ref": [],
"table_ref": [],
"text": "This work is partially supported by the National Key R&D Program of China (2020YFB1313503), the National Natural Science Foundation of China (U22B2052), the Fundamental Research Funds for the Central Universities and the Major Key Project of PCL (PCL2021A12)."
}
] | In recent years, deep learning-based methods have achieved remarkable progress in multi-exposure image fusion. However, existing methods rely on aligned image pairs, inevitably generating artifacts when faced with device shaking in real-world scenarios. Moreover, these learning-based methods are built on handcrafted architectures and operations by increasing network depth or width, neglecting different exposure characteristics. As a result, these direct cascaded architectures with redundant parameters fail to achieve highly effective inference time and lead to massive computation. To alleviate these issues, in this paper, we propose a search-based paradigm, involving selfalignment and detail repletion modules for robust multi-exposure image fusion. By utilizing scene relighting and deformable convolutions, the self-alignment module can accurately align images despite camera movement. Furthermore, by imposing a hardware-sensitive constraint, we introduce neural architecture search to discover compact and efficient networks, investigating effective feature representation for fusion. We realize the stateof-the-art performance in comparison to various competitive schemes, yielding a 4.02% and 29.34% improvement in PSNR for general and misaligned scenarios, respectively, while reducing inference time by 68.1%. | Embracing Compact and Robust Architectures for Multi-Exposure Image Fusion | [
{
"figure_caption": "Fig. 2 .2Fig. 2. Schematic diagram of the proposed architecture. The super-architecture for multi-exposure fusion consists of Self-Alignment Module (SAM) and Detail Repletion Module (DRM). Search spaces of SAM and for DRM are also illustrated respectively.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 3 .3Fig. 3. Qualitative comparison with nine state-of-the-art methods. The signal maps provide the differences of pixel intensity with the ground truth.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 4 .4Fig. 4. Qualitative comparison with nine state-of-the-art methods. The signal maps provide the differences of pixel intensity with ground truth.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 5 .5Fig. 5. Qualitative comparison with nine state-of-the-art methods on the dataset [35] without ground truth.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 7 .7Fig. 7. Qualitative results compared with the state-of-the-arts on the misaligned multi-exposure fusion.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "7",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 8 .8Fig. 8. Visual comparison of effectiveness for SRSM.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "8",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 10 .10Fig. 10. Visual results and error maps obtained by the different loss functions.",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "10",
"figure_type": "figure"
},
{
"figure_caption": "Trade= 5 Fig. 11 .511Fig. 11. Heatmaps of the searched architectures based on different trade-off parameter η. The selected operators are marked by red boxes.",
"figure_data": "",
"figure_id": "fig_7",
"figure_label": "511",
"figure_type": "figure"
},
{
"figure_caption": "COMPARISON OF PROPOSED METHODS WITH A SERIES OF MULTI-EXPOSURE FUSION SCHEMES.",
"figure_data": "5314.5015.1013.2417.5619.14 17.4217.6719.7119.9119.2320.71 ↑4.02% 20.54SSIM ↑ 0.7180.7660.8080.7450.7420.795 0.7530.7180.7570.7630.8440.8250.822LPIPS ↓ 0.2270.2310.2090.2070.2640.190 0.2480.2240.2730.1750.1430.132 ↓7.69% 0.138FSIM ↑ 0.8800.8770.8830.8710.8990.887 0.8290.8530.9060.9210.8860.924 ↑0.32% 0.924",
"figure_id": "tab_1",
"figure_label": "I",
"figure_type": "table"
},
{
"figure_caption": "EFFICIENCY COMPARISON INCLUDING PARAMETERS AND AVERAGED RUNTIME ON THE SICE DATASET.",
"figure_data": "MetricsMESPD FMMEF MEFDSIFT DEMEF DeepFuse IFCNN SDNetU2Fusion MEFGAN HALDER DPEMEF Ours Ours *PlatformMatlab MatlabMatlabMatlab Tensorflow Pytorch Tensorflow Tensorflow Tensorflow Pytorch Pytorch Pytorch PytorchDeviceCPUCPUCPUCPUGPUGPUGPUGPUGPUGPUGPUGPUGPUParameters (M) ↓----0.0890.0830.0670.6593.1574.71151.931.879 0.684Runtime (s) ↓0.1830.2350.5570.2170.2030.0220.8850.3050.8620.1410.0680.047 0.021TABLE IIINUMERICAL RESULTS COMPARED WITH REPRESENTATIVE METHODS FOR MISALIGNED MULTI-EXPOSURE IMAGE FUSION.Metrics MESPD FMMEF MEFDSIFT DEMEF DeepFuse IFCNN SDNet U2Fusion MEFGAN HALDER DPEMEFOursPSNR ↑13.1215.1113.3613.7312.1817.0416.3416.71↑0.4230.4540.4400.4360.3880.4240.4340.4640.4640.4950.4350.681 ↑37.6%LPIPS ↓ 0.2800.2650.2860.2770.3840.2930.2660.2850.3590.2700.2660.187 ↓29.4%FSIM ↑0.7960.8000.7920.7930.7920.7760.7850.8010.8010.8210.8070.913 ↑11.2%",
"figure_id": "tab_2",
"figure_label": "II",
"figure_type": "table"
},
{
"figure_caption": "COMPARISONS OF EFFECTIVENESS WITH SRSM.",
"figure_data": "Numberw/o SRSM Cascade-1 Cascade-2 Cascaded-3PSNR ↑9.5720.1520.7120.56SSIM ↑0.6110.8190.8250.825LPIPS ↓0.4070.1500.1320.136FSIM ↑0.7870.9320.9240.929",
"figure_id": "tab_3",
"figure_label": "IV",
"figure_type": "table"
},
{
"figure_caption": "COMPARISONS OF EFFECTIVENESS WITH DASM.",
"figure_data": "Metricw/o DASMFlow-BasedBefore RelightingOursPSNR ↑19.3919.1921.7322.04SSIM ↑0.5610.5400.6750.681LPIPS ↓0.3250.3310.2370.187FSIM ↑0.8520.8380.9070.913",
"figure_id": "tab_4",
"figure_label": "V",
"figure_type": "table"
},
{
"figure_caption": "THE EFFECTIVENESS OF LOSS FUNCTIONS.",
"figure_data": "Metricw/ Intw/ Int+ Graw/ Int+ DisTotalPSNR ↑20.6220.6320.6620.71SSIM ↑0.7990.8230.8210.825LPIPS ↓0.1450.1340.1360.132FSIM ↑0.9320.9130.9270.924TABLE VIIQUANTITATIVE COMPARISION OF SEARCH SPACE.3-C20.660.8240.1320.9260.0423-DC20.680.8250.1310.9310.0445-C20.580.8220.1340.9300.1245-DC20.490.8180.1350.9300.0887-C20.520.8200.1350.9280.1957-DC20.800.8230.1400.9330.153Ours20.710.8250.1320.9240.047",
"figure_id": "tab_5",
"figure_label": "VI",
"figure_type": "table"
},
{
"figure_caption": "-OFF η FOR HARDWARE-SENSITIVE ANALYSIS.",
"figure_data": "",
"figure_id": "tab_6",
"figure_label": "VIII",
"figure_type": "table"
}
] | Zhu Liu; Jinyuan Liu; Guanyao Wu; Risheng Liu | [
{
"authors": "X Zhang",
"journal": "Information Fusion",
"ref_id": "b0",
"title": "Benchmarking and comparing multi-exposure image fusion algorithms",
"year": "2021"
},
{
"authors": "L Wang; K.-J Yoon",
"journal": "IEEE transactions on pattern analysis and machine intelligence",
"ref_id": "b1",
"title": "Deep learning for hdr imaging: State-ofthe-art and future trends",
"year": "2021"
},
{
"authors": "J Luo; W Ren; X Gao; X Cao",
"journal": "IEEE Transactions on Image Processing",
"ref_id": "b2",
"title": "Multi-exposure image fusion via deformable self-attention",
"year": "2023"
},
{
"authors": "T Ma; L Ma; X Fan; Z Luo; R Liu",
"journal": "",
"ref_id": "b3",
"title": "Pia: Parallel architecture with illumination allocator for joint enhancement and detection in lowlight",
"year": "2022"
},
{
"authors": "J Liu; X Fan; Z Huang; G Wu; R Liu; W Zhong; Z Luo",
"journal": "",
"ref_id": "b4",
"title": "Target-aware dual adversarial learning and a multi-scenario multimodality benchmark to fuse infrared and visible for object detection",
"year": "2022"
},
{
"authors": "R Liu; L Ma; J Zhang; X Fan; Z Luo",
"journal": "",
"ref_id": "b5",
"title": "Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement",
"year": "2021"
},
{
"authors": "R Liu; L Ma; T Ma; X Fan; Z Luo",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b6",
"title": "Learning with nested scene modeling and cooperative architecture search for low-light vision",
"year": "2022"
},
{
"authors": "H Zhang; J Ma",
"journal": "Information Fusion",
"ref_id": "b7",
"title": "Iid-mef: A multi-exposure fusion network based on intrinsic image decomposition",
"year": "2023"
},
{
"authors": "L Ma; T Ma; R Liu; X Fan; Z Luo",
"journal": "",
"ref_id": "b8",
"title": "Toward fast, flexible, and robust low-light image enhancement",
"year": "2022"
},
{
"authors": "X Xue; J He; L Ma; Y Wang; X Fan; R Liu",
"journal": "",
"ref_id": "b9",
"title": "Best of both worlds: See and understand clearly in the dark",
"year": "2022"
},
{
"authors": "H Li; T N Chan; X Qi; W Xie",
"journal": "IEEE Transactions on Circuits and Systems for Video Technology",
"ref_id": "b10",
"title": "Detail-preserving multiexposure fusion with edge-preserving structural patch decomposition",
"year": "2021"
},
{
"authors": "H Xu; J Ma; X.-P Zhang",
"journal": "IEEE Transactions on Image Processing",
"ref_id": "b11",
"title": "Mef-gan: Multi-exposure image fusion via generative adversarial networks",
"year": "2020"
},
{
"authors": "P E Debevec; J Malik",
"journal": "",
"ref_id": "b12",
"title": "Recovering high dynamic range radiance maps from photographs",
"year": "2008"
},
{
"authors": "J Munkberg; P Clarberg; J Hasselgren; T Akenine-Möller",
"journal": "ACM Transactions on Graphics (TOG)",
"ref_id": "b13",
"title": "High dynamic range texture compression for graphics hardware",
"year": "2006"
},
{
"authors": "S K Nayar; T Mitsunaga",
"journal": "IEEE",
"ref_id": "b14",
"title": "High dynamic range imaging: Spatially varying pixel exposures",
"year": "2000"
},
{
"authors": "S Li; X Kang; J Hu",
"journal": "IEEE Transactions on Image processing",
"ref_id": "b15",
"title": "Image fusion with guided filtering",
"year": "2013"
},
{
"authors": "Q Wang; W Chen; X Wu; Z Li",
"journal": "IEEE Transactions on Circuits and Systems for Video Technology",
"ref_id": "b16",
"title": "Detail-enhanced multi-scale exposure fusion in yuv color space",
"year": "2020"
},
{
"authors": "N Hayat; M Imran",
"journal": "Journal of Visual Communication and Image Representation",
"ref_id": "b17",
"title": "Ghost-free multi exposure image fusion technique using dense sift descriptor and guided filter",
"year": "2019"
},
{
"authors": "S Hu; W Zhang",
"journal": "IEEE",
"ref_id": "b18",
"title": "Exploiting patch-based correlation for ghost removal in exposure fusion",
"year": "2017"
},
{
"authors": "K Ram Prabhakar; V Sai; R Srikar; Venkatesh; Babu",
"journal": "",
"ref_id": "b19",
"title": "Deepfuse: A deep unsupervised approach for exposure fusion with extreme exposure image pairs",
"year": "2017"
},
{
"authors": "J Liu; J Shang; R Liu; X Fan",
"journal": "IEEE Transactions on Circuits and Systems for Video Technology",
"ref_id": "b20",
"title": "Attention-guided global-local adversarial learning for detail-preserving multi-exposure image fusion",
"year": "2022"
},
{
"authors": "J Li; J Liu; S Zhou; Q Zhang; N K Kasabov",
"journal": "IEEE Transactions on Circuits and Systems for Video Technology",
"ref_id": "b21",
"title": "Learning a coordinated network for detail-refinement multi-exposure image fusion",
"year": "2022"
},
{
"authors": "J Liu; G Wu; J Luan; Z Jiang; R Liu; X Fan",
"journal": "Information Fusion",
"ref_id": "b22",
"title": "Holoco: Holistic and local contrastive learning network for multi-exposure image fusion",
"year": "2023"
},
{
"authors": "W Zhang; X Liu; W Wang; Y Zeng",
"journal": "International Journal of Advanced Robotic Systems",
"ref_id": "b23",
"title": "Multi-exposure image fusion based on wavelet transform",
"year": "2018"
},
{
"authors": "Y Yang; D Zhang; W Wan; S Huang",
"journal": "IEEE Transactions on Instrumentation and Measurement",
"ref_id": "b24",
"title": "Multi-scale exposure fusion based on multi-visual feature measurement and detail enhancement representation",
"year": "2022"
},
{
"authors": "S Li; B Yang; J Hu",
"journal": "Information Fusion",
"ref_id": "b25",
"title": "Performance comparison of different multiresolution transforms for image fusion",
"year": "2011"
},
{
"authors": "H Xu; J Ma; J Jiang; X Guo; H Ling",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b26",
"title": "U2fusion: A unified unsupervised image fusion network",
"year": "2022"
},
{
"authors": "Z Li; J Liu; R Liu; X Fan; Z Luo; W Gao",
"journal": "IEEE",
"ref_id": "b27",
"title": "Multiple taskoriented encoders for unified image fusion",
"year": "2021"
},
{
"authors": "D Han; L Li; X Guo; J Ma",
"journal": "Information Fusion",
"ref_id": "b28",
"title": "Multi-exposure image fusion via deep perceptual enhancement",
"year": "2022"
},
{
"authors": "J J Lewis; R J O'callaghan; S G Nikolov; D R Bull; N Canagarajah",
"journal": "Information fusion",
"ref_id": "b29",
"title": "Pixel-and region-based image fusion with complex wavelets",
"year": "2007"
},
{
"authors": "M Qiguang; W Baoshu",
"journal": "IEEE",
"ref_id": "b30",
"title": "A novel image fusion method using contourlet transform",
"year": "2006"
},
{
"authors": "J Shen; Y Zhao; S Yan; X Li",
"journal": "IEEE Trans. Cybern",
"ref_id": "b31",
"title": "Exposure fusion using boosting laplacian pyramid",
"year": "2014"
},
{
"authors": "Y Liu; Z Wang",
"journal": "Journal of Visual Communication and Image Representation",
"ref_id": "b32",
"title": "Dense sift for ghost-free multi-exposure fusion",
"year": "2015"
},
{
"authors": "S Li; X Kang; L Fang; J Hu; H Yin",
"journal": "information Fusion",
"ref_id": "b33",
"title": "Pixel-level image fusion: A survey of the state of the art",
"year": "2017"
},
{
"authors": "K Ma; H Li; H Yong; Z Wang; D Meng; L Zhang",
"journal": "IEEE Transactions on Image Processing",
"ref_id": "b34",
"title": "Robust multiexposure image fusion: a structural patch decomposition approach",
"year": "2017"
},
{
"authors": "K Ma; Z Wang",
"journal": "IEEE",
"ref_id": "b35",
"title": "Multi-exposure image fusion: A patch-wise approach",
"year": "2015"
},
{
"authors": "H Li; K Ma; H Yong; L Zhang",
"journal": "IEEE Transactions on Image Processing",
"ref_id": "b36",
"title": "Fast multi-scale structural patch decomposition for multi-exposure image fusion",
"year": "2020"
},
{
"authors": "F Kou; Z Li; C Wen; W Chen",
"journal": "IEEE",
"ref_id": "b37",
"title": "Multi-scale exposure fusion via gradient domain guided image filtering",
"year": "2017"
},
{
"authors": "",
"journal": "Journal of Visual Communication and Image Representation",
"ref_id": "b38",
"title": "Edge-preserving smoothing pyramid based multi-scale exposure fusion",
"year": "2018"
},
{
"authors": "Q Yan; J Sun; H Li; Y Zhu; Y Zhang",
"journal": "Neurocomputing",
"ref_id": "b39",
"title": "High dynamic range imaging by sparse representation",
"year": "2017"
},
{
"authors": "J Wang; H Liu; N He",
"journal": "Neurocomputing",
"ref_id": "b40",
"title": "Exposure fusion based on sparse representation using approximate k-svd",
"year": "2014"
},
{
"authors": "K Ma; Z Duanmu; H Zhu; Y Fang; Z Wang",
"journal": "IEEE Transactions on Image Processing",
"ref_id": "b41",
"title": "Deep guided learning for fast multi-exposure image fusion",
"year": "2019"
},
{
"authors": "Y Zhang; Y Liu; P Sun; H Yan; X Zhao; L Zhang",
"journal": "Information Fusion",
"ref_id": "b42",
"title": "Ifcnn: A general image fusion framework based on convolutional neural network",
"year": "2020"
},
{
"authors": "K Ma; K Zeng; Z Wang",
"journal": "IEEE Transactions on Image Processing",
"ref_id": "b43",
"title": "Perceptual quality assessment for multi-exposure image fusion",
"year": "2015"
},
{
"authors": "H Zhang; J Ma",
"journal": "International Journal of Computer Vision",
"ref_id": "b44",
"title": "Sdnet: A versatile squeeze-and-decomposition network for real-time image fusion",
"year": "2021"
},
{
"authors": "J Liu; J Shang; R Liu; X Fan",
"journal": "IEEE",
"ref_id": "b45",
"title": "Halder: Hierarchical attentionguided learning with detail-refinement for multi-exposure image fusion",
"year": "2021"
},
{
"authors": "Q Yan; D Gong; Q Shi; A V D Hengel; C Shen; I Reid; Y Zhang",
"journal": "",
"ref_id": "b46",
"title": "Attention-guided network for ghost-free high dynamic range imaging",
"year": "2019"
},
{
"authors": "M Guo; Z Zhong; W Wu; D Lin; J Yan",
"journal": "",
"ref_id": "b47",
"title": "Irlas: Inverse reinforcement learning for architecture search",
"year": "2019"
},
{
"authors": "Y Gao; H Yang; P Zhang; C Zhou; Y Hu",
"journal": "",
"ref_id": "b48",
"title": "Graph neural architecture search",
"year": "2020"
},
{
"authors": "Y Chen; G Meng; Q Zhang; S Xiang; C Huang; L Mu; X Wang",
"journal": "",
"ref_id": "b49",
"title": "Renas: Reinforced evolutionary neural architecture search",
"year": "2019"
},
{
"authors": "X Zhang; Z Huang; N Wang; S Xiang; C Pan",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b50",
"title": "You only search once: Single shot neural architecture search via direct sparse optimization",
"year": "2020"
},
{
"authors": "H Liu; K Simonyan; Y Yang",
"journal": "",
"ref_id": "b51",
"title": "Darts: Differentiable architecture search",
"year": "2018"
},
{
"authors": "R Liu; J Gao; J Zhang; D Meng; Z Lin",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b52",
"title": "Investigating bilevel optimization for learning and vision from a unified perspective: A survey and beyond",
"year": "2021"
},
{
"authors": "H Zhang; Y Li; H Chen; C Gong; Z Bai; C Shen",
"journal": "International Journal of Computer Vision",
"ref_id": "b53",
"title": "Memoryefficient hierarchical neural architecture search for image restoration",
"year": "2022"
},
{
"authors": "R Li; R T Tan; L.-F Cheong",
"journal": "",
"ref_id": "b54",
"title": "All in one bad weather removal using architectural search",
"year": "2020"
},
{
"authors": "R Liu; Z Liu; J Liu; X Fan",
"journal": "",
"ref_id": "b55",
"title": "Searching a hierarchically aggregated fusion architecture for fast multi-modality image fusion",
"year": "2021"
},
{
"authors": "Z Teed; J Deng",
"journal": "Springer",
"ref_id": "b56",
"title": "Raft: Recurrent all-pairs field transforms for optical flow",
"year": "2020"
},
{
"authors": "D Wang; J Liu; X Fan; R Liu",
"journal": "",
"ref_id": "b57",
"title": "Unsupervised misaligned infrared and visible image fusion via cross-modality image generation and registration",
"year": "2022"
},
{
"authors": "Z Huang; J Liu; X Fan; R Liu; W Zhong; Z Luo",
"journal": "",
"ref_id": "b58",
"title": "Reconet: Recurrent correction network for fast and efficient multi-modality image fusion",
"year": "2022"
},
{
"authors": "X Wang; K C Chan; K Yu; C Dong; C Change Loy",
"journal": "",
"ref_id": "b59",
"title": "Edvr: Video restoration with enhanced deformable convolutional networks",
"year": "2019"
},
{
"authors": "A Liu; X Liu; J Fan; Y Ma; A Zhang; H Xie; D Tao",
"journal": "",
"ref_id": "b60",
"title": "Perceptualsensitive gan for generating adversarial patches",
"year": "2019"
},
{
"authors": "I Gulrajani; F Ahmed; M Arjovsky; V Dumoulin; A C Courville",
"journal": "Advances in neural information processing systems",
"ref_id": "b61",
"title": "Improved training of wasserstein gans",
"year": "2017"
},
{
"authors": "J Cai; S Gu; L Zhang",
"journal": "IEEE Transactions on Image Processing",
"ref_id": "b62",
"title": "Learning a deep single image contrast enhancer from multi-exposure images",
"year": "2018"
},
{
"authors": "R Zhang; P Isola; A A Efros; E Shechtman; O Wang",
"journal": "",
"ref_id": "b63",
"title": "The unreasonable effectiveness of deep features as a perceptual metric",
"year": "2018"
},
{
"authors": "L Zhang; L Zhang; X Mou; D Zhang",
"journal": "IEEE transactions on Image Processing",
"ref_id": "b64",
"title": "Fsim: A feature similarity index for image quality assessment",
"year": "2011"
}
] | [
{
"formula_coordinates": [
4,
118.74,
706.98,
181.29,
12.19
],
"formula_id": "formula_0",
"formula_text": "I S i = I i ⊗ S(I i ), i ∈ {U; O},(1)"
},
{
"formula_coordinates": [
4,
390.5,
496.27,
172.53,
9.88
],
"formula_id": "formula_1",
"formula_text": "F A = A(F U , F O ) + F O ,(2)"
},
{
"formula_coordinates": [
4,
400.78,
739.05,
162.25,
9.88
],
"formula_id": "formula_2",
"formula_text": "y = F A R(F A ).(3)"
},
{
"formula_coordinates": [
5,
80.97,
489.53,
219.05,
20.09
],
"formula_id": "formula_3",
"formula_text": "F j = O i→j (F i ); O i→j = O∈O α i→j O(F i ),(4)"
},
{
"formula_coordinates": [
5,
97.98,
597.14,
202.04,
20.09
],
"formula_id": "formula_4",
"formula_text": "R(α; LAT) = M O∈O α i→j LAT(O),(5)"
},
{
"formula_coordinates": [
5,
109.93,
687.19,
190.1,
16.21
],
"formula_id": "formula_5",
"formula_text": "min α val (α; ω * ) + ηR(α; LAT),(6)"
},
{
"formula_coordinates": [
5,
376.12,
162.19,
186.92,
9.85
],
"formula_id": "formula_6",
"formula_text": "Total = Int + β 1 Gra + β 2 Dis .(7)"
},
{
"formula_coordinates": [
5,
391.23,
230.32,
171.81,
22.31
],
"formula_id": "formula_7",
"formula_text": "Int = 1 HW y -y gt 1 ,(8)"
},
{
"formula_coordinates": [
5,
382.93,
306.87,
180.11,
22.31
],
"formula_id": "formula_8",
"formula_text": "Gra = 1 HW ∇y -∇y gt 2 ,(9)"
},
{
"formula_coordinates": [
5,
311.98,
432.54,
256.09,
22.98
],
"formula_id": "formula_9",
"formula_text": "E x∼Pfake D(y)-E x∼Preal D(y gt )+ηE x∼Pfake [( ∇ y D(y) 2 -1) 2 ].(10)"
}
] |
|
10.48550/arxiv.2011.00362 | 2023-05-20 | [
{
"figure_ref": [],
"heading": "INTRODUCTION",
"publication_ref": [],
"table_ref": [],
"text": "The key challenge that we seek to address is that of identifying learned features in a model's intermediate representations that are more or less likely to be robust to distributional shift. Our approach starts from the intuition that for high-entropy features in a model's training distribution, it will have learned a better understanding for when the feature is relevant. More precisely, it will be better at distinguishing the presence or absence of the feature across different situations. Consider a hypothetical data set containing photographs from two safari trips, where each trip contains the same people on the same safari, but driving around in different trucks. Suppose that it is useful for the given task to identify which of the two trips a given image corresponds to; we might expect the model to be particularly good at distinguishing between the trucks. On the other hand, if a rare tree appears in exactly one photograph, the model may have learned to recognise the specific pattern of pixels in that photograph corresponding to the tree, but it might not have the capability to recognise the tree in new situations.\nAs models have increased in performance within the bounds of the i.i.d. assumption, recent years have seen growing interest in the OOD behaviour of machine learning systems. While many approaches have studied OOD detection or the effects of external changes to a model's training regime on OOD behaviour (e.g. domain randomization or auxiliary loss functions), to the best of our knowledge our proposal of the entropy of an intermediate representation as a guide to its effects OOD is a novel approach. In this paper we demonstrate that the removal of low-entropy representations via the masking of learned discrete bits can notably improve OOD performance."
},
{
"figure_ref": [
"fig_0",
"fig_3"
],
"heading": "TASK AND MODEL DESCRIPTION",
"publication_ref": [
"b12"
],
"table_ref": [],
"text": "To learn representations of a domain we train an encoder network to produce a representation r of a given input x * . This representation is given to a distinguisher network that is tasked with identifying x * from a set of k images composed of x * and k -1 distractor inputs arranged randomly. We use the CIFAR-10 dataset (Krizhevsky, 2009) as the training distribution. The labels from the dataset are discarded and an unsupervised k-contrast task is constructed by pairing each image with k -1 distractor images, shuffling, and giving the distinguisher k inputs to choose from. The same preprocessing is later used when out-of-distribution datasets are introduced. See Figure 1for an example of a contrastive task and Figure 4 in the Supplementary Material for the full architecture. It is important to note that we use a 'soft-discretization' technique (Foerster et al., 2016) on the intermediate representation r such that it can be learned with gradient-descent, but each dimension can be mapped to a binary digit at test time with no loss in performance. While the use of a communication channel to discretize representations poses optimization challenges, it also provides a large benefit when it comes to computing the entropy values of each bit in the representation. The computation is reduced from approximating an integral to the simple formula for the entropy of a binary variable, as outlined in Section 3.1. This allows us to run a greater number of experiments with higher precision than if we had used continuous representations. This unsupervised contrastive learning task was chosen as it can be easily transferred to different data distributions. A task such as image classification limits the available datasets as it requires the out-of-distribution testing data to have the same (or at least overlapping) image labels."
},
{
"figure_ref": [],
"heading": "ENTROPY-BASED MASKING",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [
"fig_5"
],
"heading": "ENTROPY OF REPRESENTATION BITS",
"publication_ref": [],
"table_ref": [],
"text": "Each representation r produced by an encoder network consists of a number of bits |r|, referred to as the representation length. By considering each bit at index i as a random variable B i we can compute the binary entropy of the bit on a given dataset D:\nH(B i | D) = -p log 2 p -(1 -p) log 2 (1 -p), where p = P (B i = 1 | D).\n(1)\nEntropy close to 1 means that the bit is 0 or 1 with roughly equal probability of p = 0.5. Very low entropy means that the bit is either almost always 0 or almost always 1. We notice that for smaller representation lengths and/or few distractors the distribution tends to skew towards higher entropy bits. In separate experiments where we further varied representation lengths, we find that for smaller |r| equal to 8, 16 or 32, all bits have entropy higher than 0.8, which makes studying bits based on entropy variation uninteresting for these representation lengths. For a visualization of these entropy values see Figure 5 in the Supplementary Material. Representation lengths of 64, 128, 256 and 512 all lead to a wide range of entropy values. A theoretical analysis of the optimal bit-entropy can be found in Section B of the Supplementary Material."
},
{
"figure_ref": [],
"heading": "BIT MASKING STRATEGIES",
"publication_ref": [],
"table_ref": [],
"text": "In this paper we are interested in the effects of strategically 'removing' parts of the model's intermediate representation, i.e. obscuring bits in r. It is important to note that we are only applying masking at test time. The masking is not used to train any of the models. The mask is defined by a set masking variables m i ∈ {0, 1} for each bit r i in the representation. The masked bit ri is computed:\nri = m i r i + (1 -m i ) 1 2 . (2\n)\nIn other words, when the masking variable m i = 0 then ri = 0.5, and otherwise ri = r i . In this paper we use three masking strategies; Random Masking, Top-Entropy Masking, and Bottom-Entropy Masking. In order to construct a mask with any of these strategies, we define a masking proportion p mask that represents the percentage of bits in r that should be masked.\nTo construct any mask M = {m 1 , . . . , m |r| } we will need to choose l mask = p mask • |r| bits to remove. For a random mask we draw l mask masking variables from M at random with uniform probability and without replacement, and set them to 0, we set the remaining |r| -l mask variables to 1. To construct a top-entropy mask we compute the entropy for each bit h i = H(B i | D) and sort these values in descending order. We then take the bits associated with the first l mask entropy values (i.e. highest entropy) and set their corresponding masking variables to zero. Likewise, for the bottom-entropy mask we take the last l mask bits and remove those instead."
},
{
"figure_ref": [],
"heading": "EXPERIMENTAL RESULTS",
"publication_ref": [
"b16"
],
"table_ref": [],
"text": "Figure 2: Accuracy of CIFAR-10 pre-trained models on OOD datasets (on the y-axis) against accuracy on CIFAR-10 (on the x-axis). The dashed line (which coincides with the green and blue lines) is the y = x line.\nWe trained 54 encoder-distinguisher pairs 1 on CIFAR-10 and removed models that did not converge, resulting in 51 trained models. Models were trained with varying combinations of representation lengths and number of distractors:\n(|r|, k) ∈ {64, 128, 256, 512} × {3, 5, 10, 20}.\nSee Table 1 for the test accuracy statistics for the models on the k-contrast CIFAR-10 training distributions. See Section C in the Supplementary Material for a full description of the training methodology.\nTo evaluate the effects of distributional shifts we test our 51 trained models on the CIFAR-100 Krizhevsky ( 2009 In Figure 2 we demonstrate the shift in performance that results from applying the models to the new datasets. Following Taori et al. (Taori et al., 2020), plotting the relationship between InD and OOD performance in this manner allows us to study distributional shift while controlling for the variations in initial accuracy. The y = x line is plotted with a black dashed line, however, it is obscured by the regression lines for CIFAR-100 and Stanford Online Products. This tells us that there is no distributional shift for these datasets, i.e. no loss in performance. For this reason, we drop these datasets from all further out-of-distribution analysis. For the other datasets, we see in order of increased degradation: Plant Village, Colorectal Histology, and MNIST."
},
{
"figure_ref": [],
"heading": "ANALYSIS OF MASKING EFFECTS IN-DISTRIBUTION",
"publication_ref": [],
"table_ref": [],
"text": "Before moving onto the out-of-distribution case, we first examine the effects of applying the different masking strategies to the models that we trained on CIFAR-10, with the CIFAR-10 test data. For each of the 51 successfully trained models we evaluated the accuracy without any masking, and with each 1 A sweep of 3 runs for each pair of (|r|, k) plus 6 initial separate runs. of the different masking strategies for masking proportions between 0.15 and 0.5 at 0.05 intervals. We found that for any masking proportion, removing the top-entropy bits is more damaging to accuracy than masking out bottom-entropy bits. In light of general insights from information theory, this result is not too surprising. The highest entropy bits necessarily convey the most information, and so it follows that their removal should lead to the largest drop in performance.\nIn general, we did not expect any of the masking strategies to provide a benefit when applied within the training distribution. Yet, we saw that with a small masking proportion (around p mask < 0.3) we see an increase in accuracy for low-entropy and random masks. Our initial hypothesis was that the masking may be 'undoing' overfitting to the training set. But for each of the trained models we have verified that there is no overfitting (see Section D.1 in the Supplementary Material for a visualization)."
},
{
"figure_ref": [
"fig_2"
],
"heading": "ANALYSIS OF MASKING EFFECTS OUT-OF-DISTRIBUTION (OOD)",
"publication_ref": [
"b6",
"b2",
"b2"
],
"table_ref": [
"tab_1"
],
"text": "In order to understand the effects of masking on accuracy in the OOD setting we measure the mean change in accuracy of a masking strategy under various circumstances. We also report the standard deviations associated with these estimates. As in the case of in-distribution masking we evaluated the masking strategies for a sweep of masking proportions between 0.15 and 0.5 at 0.05 intervals. We cut-off the maximum masking proportion p mask ≤ 0.25 for all further analysis as beyond that threshold masking has an almost universally negative effect. The overall mean accuracy changes can be seen in Table 2. We see that masking the bottom-entropy or random bits produces the highest increase, albeit with a large variance.\nThis variance can be understood and disentangled by separating the low-k models from the high-k models. What we see is that the benefits of bottom-entropy masking are more prevalent for low-k models. This is visualized in Figure 3 where we illustrate the effective robustness of each of the masking strategies on the three OOD datasets. In the Supplementary Material Section D.2 we include plots for all values of k and p mask that we tested. Effective robustness is a concept introduced by Taori et al. ( 2020) as a way to understand the efficacy of a method for increasing robustness to distributional shift. By plotting the baseline regression line for unaltered models with differing in-distribution accuracy values on the diagram we can observe whether a proposed robustness method moves towards the y = x line (i.e. no degradation). Crucially, with these plots, we are able to account for each model's performance on the training distribution. Hence, despite the large variance in the performance of models trained across various k and |r| values2 , we are able to discern the effects of the masking interventions.\nIn our case, we see that -as is consistent with previous results -for each dataset the top-entropy masking moves below the dashed green line showing the baseline unmasked models. On the other hand, the random masking and bottom-entropy masking lines move closer to y = x (as compared to the no masking lines). For Plant Village we see that almost all of the in-distribution accuracy is recovered. For MNIST we find the most substantial jump, and the largest benefit of bottom-entropy over random masking. 2021). However, we apply entropy in an entirely different context, namely, we calculate the entropy of latent variables to estimate how robust they will be to distributional shift. Relative entropy (KL-divergence) is a popular measure and is notably used in the Bits-Back method Hinton & van Camp (1993), Flamich et al. (2020) to calculate the optimal compression rate in latent variables. Images that are traditionally compressed by a variational auto-encoder have now been compressed with code-length close to this theoretical optimum Flamich et al. (2020).\nContrastive representation learning takes many forms; in computer vision alone there are many approaches for applying deep learning to multiple inputs and producing representations to distinguish between them; see Jaiswal et al. (2020) for a review. To our knowledge, there are no existing suitable state-of-the-art (SOTA) methods for OOD robustness in contrastive learning to benchmark our proposals against."
},
{
"figure_ref": [
"fig_2",
"fig_3",
"fig_3",
"fig_0"
],
"heading": "CONCLUSION",
"publication_ref": [
"b19"
],
"table_ref": [],
"text": "In this paper we have investigated the out-of-distribution effects of using different post-hoc strategies to remove bits from discrete intermediate representations in an unsupervised contrastive learning task.\nWe have studied how the difficulty of the task (more distractors) impacts the entropy distribution of the learned representations and shown that removing low-entropy bits can improve the performance of models out-of-distribution (Section 4.2), notably almost entirely restoring in-distribution performance for one of our datasets (see Figure 3). However, the results also present mysteries that prompt further experiments and analysis. At the time of writing, we do not have a clear understanding of why the removal of bits within the training distribution should increase performance, as we would expect the encoder to learn an optimal protocol.\nNext, there is a need for a deeper understanding of the conditions in which our results hold. Within our experimentation, we found that the effect (of harm from low-entropy features OOD) was less pronounced for models trained on the more difficult tasks (higher numbers of distractors). From our data, it is unclear if this relationship represents something fundamental or if it is a side-effect of these models generally performing to a lower standard. One of the most important avenues of further work is in testing if other systems built on top of the learned representations in this paper inherit the same OOD robustness under low-entropy masking.\nA NETWORK ARCHITECTURES The encoder network is composed of a convolutional network (CNN) that takes a 32 × 32 × 3 dimensional tensor as input (CNN A in Figure 4), followed by: a 3 × 3 convolutional layer with 64 filters and ReLU activation; two 3 × 3 convolutional layers with 64 filters, ReLU activation, and a stride-length of 2; a flatten layer; and finally a dense layer without any activation that projects into R |r| , where |r| is a hyperparameter controlling the 'representation length' of r. Next, between the encoder and the distinguisher, there is a discretize/regularize unit (Foerster et al., 2016). Following the literature in which this component was developed, we will refer to this as a communication channel (see in green in Figure 4). The channel is a differentiable unit that, during training, 'soft discretizes' activations passed through it by applying Gaussian white noise (GWN) and a sigmoid function. Then at test time we 'hard discretize' the activations by passing through a sigmoid function and emitting 0 if the result is less than 0.5 and 1 otherwise. This enables the end-to-end learning of a discrete representation via backpropagation from the output of the distinguisher. We configure the channel with a fixed GWN standard deviation of 0.5 during training.\nThe distinguisher network is composed of another convolutional network (CNN B in Figure 1) with exactly the same input and layers as the CNN in the encoder (initialized separately and no parameter sharing), except projecting to a fixed embedding size of 128. This CNN is shared for each of the 'possible answer' images, producing embeddings that are each concatenated with the representation r from the encoder (i.e. the output of the communication channel) and fed into a transformer network (Vaswani et al., 2017) as tokens. The transformer is composed of two self-attention encoder layers with 3 heads of dimension 64, and a dropout rate of 0.1. After the transformer layers each token is projected onto a single dimension without activation. This is then taken as the log-probability (logit) that the corresponding possible answer is correct. The networks are trained together with a sparse categorical crossentropy loss on these logits and the index of the correct answer. The use of a transformer and a shared encoder for the input images means that a model trained, for example, on a 3-contrast dataset (k = 3) can be tested on a 5-contrast dataset without any modification."
},
{
"figure_ref": [],
"heading": "B THEORETICAL RESULT",
"publication_ref": [],
"table_ref": [],
"text": "Consider the following abstracted and idealized version of the contrastive learning game. An encoder receives an input, and communicates features in that input via bits. A distinguisher has to identify the original input from a set of k (distractor) inputs, based on the communicated features. The encoder and distinguisher win if the distinguisher correctly identifies the original input. The encoder and distinguisher need to decide on a communication protocol before playing the game. Each bit corresponds to one feature. The encoder sends a 1 if a given feature is present and a 0 otherwise.\nThe question we're answering in this section is: what is the optimal feature occurrence (or bit entropy) for a feature when the encoder can choose b bits, and the distinguisher has to choose between k inputs.\nBelow we calculate that the optimal strategy is to use l independent features that are each present in exactly half of the images. The chance of the receiver picking out the right image depends on k.\nFor this calculation we will assume the encoder can choose b = 2 bits, i.e. can communicate two features x and y. Let f x and f y be the frequency of respectively feature x and feature y in the dataset.\nTo answer the question we will calculate the values of f x and f y that maximize the chance of winning.\nLet c x be the random variable that represents: the correct input has feature x, and c y the variable that represents: the correct input has feature y. We assume that these variables are independent. Let v be the random variable that represents the number of inputs in the set of k inputs that the distinguisher gets to see, that have both feature x and feature y. Note that P (c x , c y ) = f x • f y . Below we calculate that\nP (win|c x , c y ) = Σ k v=1 1 v • (f x f y ) v-1 • (1 -f x f y ) k-v • k -1 v -1 .\nTo do so we introduce one more helper variable ṽ which represents the number of inputs in the set of k inputs that the distinguisher gets to see, that have both feature x and feature y, but excluding the correct input.\nWe now calculate\nP (win|c x , c y ) = Σ k v=1 P (win|c x , c y , V = v) • P ( Ṽ = v -1|c x , c y ) = Σ k v=1 P (win|c x , c y , V = v) • P ( Ṽ = v -1)\nNote that P (win|c x , c y , V = v) = 1 v and\nP ( Ṽ = v -1) = (f x f y ) v-1 • (1 -f x f y ) k-1-(v-1) • k -1 v -1 Hence P (win|c x , c y ) = Σ k v=1 1 v • (f x f y ) v-1 • (1 -f x f y ) k-v • k -1 v -1 .\nApplying the Bionomial theorem gives us the following equality\nP (win|c x , c y )P (c x , c y ) = f x f y • Σ k v=1 1 v • (f x f y ) v-1 • (1 -f x f y ) k-v • k -1 v -1 = Σ k v=1 1 v • (f x f y ) v • (1 -f x f y ) k-v • k -1 v -1 = 1 k Σ k v=1 (f x f y ) v • (1 -f x f y ) k-v • k v = 1 k (f x f y + 1 -f x f y ) k -(1 -f x f y ) k = 1 k 1 -(1 -f x f y ) k = 1 k - 1 k (1 -f x f y ) k\nWe can write similar equations for c ¬x and c ¬y and combining them results in\nP (win) = 4 k - 1 k (1 -f x • f y ) k +(1 -f x • (1 -f y )) k +(1 -(1 -f x ) • f y ) k +(1 -(1 -f x ) • (1 -f y )) k .\nMore generally, for arbitrary number of bits b and feature frequencies f 1 , . . . , f b we find\nP (win) = 2 b k - 1 k (1 -f 1 • • • f b ) k + (1 -(1 -f 1 )f 2 • • • f b ) k + . . . + (1 -(1 -f 1 ) • • • (1 -f b )) k\nThe derivative of P (win) with respect to f 1 is\n∂P (win) ∂f 1 = f 2 • • • f b (1 -f 1 • • • f b ) k-1 -f 2 • • • f b (1 -(1 -f 1 )f 2 • • • f b ) k-1 + . . . -(1 -f 2 ) • • • (1 -f b )(1 -(1 -f 1 ) • • • (1 -f b )) k-1\nWhen f 1 = 0.5 the components with a factor of f 1 compensate for the ones with a factor of (1 -f 1 ), and so the derivative is 0 for f 1 = 0.5. Deriving with respect to other feature values gives analogues results. That is, one optimal feature occurrence value for maximizing P (win) is 0.5."
},
{
"figure_ref": [
"fig_5"
],
"heading": "C TRAINING METHODOLOGY",
"publication_ref": [
"b10",
"b0"
],
"table_ref": [],
"text": "In order to prevent overfitting and the representation of 'trivial features' (e.g. specific pixel values) in the representations, during training we use a stack of image augmentation layers independently applied prior to each image encoder. This involves a random rotation of up to 0.1 radians, a random contrast shift of up to 10%, a random translation of up to 10% along both axes, and a random zoom of up to 10% (all with a nearest-neighbour filling of blank pixels).\nThe models were optimized using Adam (Kingma & Ba, 2015) with a learning rate of 0.001. The batch size used for training was dependent on the number of distractors, and each epoch iterated through the entire training dataset. See Table 1 for the full breakdown of test accuracy values for trained models, i.e. the mean and standard deviations for the proportion of occasions where the distinguisher was correctly able to identifier x * by using r.\nAll of the code was implemented with Tensorflow 2 (Abadi et al., 2015) and datasets were pulled from Tensorflow Datasets3 (TFDS) (TF Devs, 2022). CIFAR-10 was split into the default TFDS training and test sets (50,000 training images and 10,000 test images). Training and analysis were performed with an NVIDIA RTX 3090 GPU.\nWe trained 54 independent encoder-distinguisher pairs4 for 10 epochs on CIFAR-10 and removed models that did not converge (as defined by not reaching an 80% drop in loss), resulting in 51 trained models (taken as the best performing checkpoint). Models were trained with varying combinations of representation lengths and number of distractors: (|r|, k) ∈ {64, 128, 256, 512} × {3, 5, 10, 20}. We also trained models with representation lengths 8, 16 and 32, visualizations of which can be found in Figure 5, which we discarded because their bit entropies were too homogeneous to meaningfully study the effect of masking out low versus high entropy bits. "
},
{
"figure_ref": [],
"heading": "D EXPERIMENTS",
"publication_ref": [],
"table_ref": [],
"text": "The code for the experiments can be found at the following repository: [URL removed to preserve anonymity]"
},
{
"figure_ref": [
"fig_6",
"fig_7",
"fig_8",
"fig_9"
],
"heading": "D.1 OVERFITTING ANALYSIS",
"publication_ref": [],
"table_ref": [],
"text": "In Figure 6 we see that the test and training accuracies are very similar (with the test accuracy even being slightly higher) and so no overfitting has happened. Figure 7 shows the OOD accuracies for each dataset (using the data of all the values of k and all the analysed representation lengths). Figure 8 shows the accuracies for each dataset and each value of k.\nFigure 9 shows the accuracies for each dataset and each representation length |r|. "
},
{
"figure_ref": [],
"heading": "ACKNOWLEDGEMENTS",
"publication_ref": [],
"table_ref": [],
"text": "Work done by both authors is thanks to the UKRI Centre for Doctoral Training in Safe and Trusted AI (EPSRC Project EP/S023356/1)."
},
{
"figure_ref": [],
"heading": "D.3 OOD MEAN ACCURACY CHANGE FROM MASKS",
"publication_ref": [],
"table_ref": [],
"text": "The tables in this section are the same as Table 2 in Section ??, except separated by different values of k. Figure 10 is a visualisation of the data along with the 'distance out-of-distribution' for each k value. "
}
] | We study the relationship between the entropy of intermediate representations and a model's robustness to distributional shift. We train models consisting of two feed-forward networks end-to-end separated by a discrete n-bit channel on an unsupervised contrastive learning task. Different masking strategies are applied after training that remove a proportion of low-entropy bits, high-entropy bits, or randomly selected bits, and the effects on performance are compared to the baseline accuracy with no mask. We hypothesize that the entropy of a bit serves as a guide to its usefulness out-of-distribution (OOD). Through experiment on three OOD datasets we demonstrate that the removal of low-entropy bits can notably benefit OOD performance. Conversely, we find that top-entropy masking disproportionately harms performance both in-distribution (InD) and OOD. | Generalization LOW-ENTROPY LATENT VARIABLES HURT OUT-OF-DISTRIBUTION PERFORMANCE | [
{
"figure_caption": "Figure 1 :1Figure 1: An example of a contrastive task (k = 3). For a given dataset, the distinguisher is shown k images, among which k-1 distractor images, and has to predict the correct image.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "), Stanford Online Products Song et al. (2016), Colorectal Histology Kather et al. (2016), Plant Village Hughes & Salathe (2015), and MNIST LeCun et al. (1999) datasets.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: Effective robustness plots for low-k models. y = x shown as black dashed line.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: Architecture diagram (k = 5). The encoder is shown as the purple and green components, and the distinguisher is the orange and red.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "P(win) = P (win|c x , c y ) • P (c x , c y ) +P (win|c x , c ¬y ) • P (c x , c ¬y ) +P (win|c ¬x , c y ) • P (c ¬x , c y ) +P (win|c ¬x , c ¬y ) • P (c ¬x , c ¬y )",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 :5Figure5: The x-axis represents entropy percentile of bits in the representation. The y-axis shows the entropy values of bits (measured on CIFAR-10). In other words, we take the list of bits and sort them by entropy, and then plot the sorted line as percentiles in order to compare the distributions of different lengths. The translucent regions show the error bars from various training runs. We can see that for lower |r| values, the entropy distributions do not tend to go below 0.8.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Figure 6 :6Figure 6: For different values of k the blue line shows the training accuracy and the orange line shows the test accuracy.",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "Figure 7 :7Figure 7: The y-axis represents the accuracy and the x-axis the masking proportion. Different masking strategies are represented by different colors.",
"figure_data": "",
"figure_id": "fig_7",
"figure_label": "7",
"figure_type": "figure"
},
{
"figure_caption": "Figure 8 :8Figure 8: The y-axis shows the accuracy and the x-axis shows different masking proportions. Masking strategies are indicated by color.",
"figure_data": "",
"figure_id": "fig_8",
"figure_label": "8",
"figure_type": "figure"
},
{
"figure_caption": "Figure 9 :9Figure 9: The y-axis shows the accuracy and the x-axis shows different masking proportions. Masking strategies are indicated by color.",
"figure_data": "",
"figure_id": "fig_9",
"figure_label": "9",
"figure_type": "figure"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Mean accuracy shift (in percentage points) after masking with each strategy. After running paired t-tests we find that all of these accuracy shifts are statistically significant (with p = 0.05). work adds to the toolkit of methods to aid in understanding and improving robustness to distributional shift, which for example includes forms of data augmentationHendrycks et al. (2021) and abstaining from making a prediction in the face of uncertaintyThulasidasan et al. (2021). For a general overview of problems and methods in OOD robustness seeShen et al. (2015).Below we reference some notable entropy-based methods that have a different purpose than improving OOD robustness.Chatterjee & Mishchenko (2019) use low entropy (or \"rare\") signals to analyze the extent to which a model is overfitted to the training distribution. Entropy-based methods have also been used widely in the adjacent problem of OOD detection. For example, predictive entropy measures the uncertainty of the prediction of a sample given a training distribution and is used to calculate the extent to which a sample is OODKirsch et al. (",
"figure_data": "CIFAR-10Colorectal HistologyMNISTPlant VillageMasked Bottom Entropy1.6 ± 8.0-2.0 ± 14.39.4 ± 15.63.0 ± 23.7Masked Top Entropy-4.3 ± 21.4-7.8 ± 19.0-16.6 ± 5.3 -18.5 ± 21.8Random Mask2.5 ± 12.33.4 ± 10.94.2 ± 13.72.1 ± 19.65 RELATED WORK",
"figure_id": "tab_1",
"figure_label": "2",
"figure_type": "table"
}
] | Nandi Schoots; Dylan Cope | [
{
"authors": "Martín Abadi; Ashish Agarwal; Paul Barham; Eugene Brevdo; Zhifeng Chen; Craig Citro; Greg S Corrado; Andy Davis; Jeffrey Dean; Matthieu Devin; Sanjay Ghemawat; Ian Goodfellow; Andrew Harp; Geoffrey Irving; Michael Isard; Jia Yangqing; Rafal Jozefowicz; Lukasz Kaiser; Manjunath Kudlur; Josh Levenberg; Dandelion Mané; Rajat Monga; Sherry Moore; Derek Murray; Chris Olah; Mike Schuster; Jonathon Shlens; Benoit Steiner; Ilya Sutskever; Kunal Talwar; Paul Tucker; Vincent Vanhoucke; Vijay Vasudevan; Fernanda Viégas; Oriol Vinyals; Pete Warden; Martin Wattenberg; Martin Wicke; Yuan Yu; Xiaoqiang Zheng",
"journal": "",
"ref_id": "b0",
"title": "TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems",
"year": "2015"
},
{
"authors": "Satrajit Chatterjee; Alan Mishchenko",
"journal": "",
"ref_id": "b1",
"title": "Coherent gradients: An approach to understanding generalization in gradient descent-based optimization",
"year": "2019"
},
{
"authors": "Gergely Flamich; Marton Havasi; José Miguel Hernández-Lobato",
"journal": "CoRR",
"ref_id": "b2",
"title": "Compressing images by encoding their latent representations with relative entropy coding",
"year": "2020"
},
{
"authors": "Jakob Foerster; Alexandros Ioannis; Nando Assael; Shimon De Freitas; Whiteson",
"journal": "",
"ref_id": "b3",
"title": "Learning to Communicate with Deep Multi-Agent Reinforcement Learning",
"year": ""
},
{
"authors": " Curran Associates; Inc",
"journal": "",
"ref_id": "b4",
"title": "",
"year": "2016"
},
{
"authors": "Dan Hendrycks; Steven Basart; Norman Mu; Saurav Kadavath; Frank Wang; Evan Dorundo; Rahul Desai; Tyler Zhu; Samyak Parajuli; Mike Guo; Dawn Song; Jacob Steinhardt; Justin Gilmer",
"journal": "IEEE",
"ref_id": "b5",
"title": "The many faces of robustness: A critical analysis of out-of-distribution generalization",
"year": "2021"
},
{
"authors": "Geoffrey E Hinton; Drew Van Camp",
"journal": "Association for Computing Machinery",
"ref_id": "b6",
"title": "Keeping the neural networks simple by minimizing the description length of the weights",
"year": "1993"
},
{
"authors": "P David; Marcel Hughes; Salathe",
"journal": "",
"ref_id": "b7",
"title": "An open access repository of images on plant health to enable the development of mobile disease diagnostics through machine learning and crowdsourcing",
"year": "2015"
},
{
"authors": "Ashish Jaiswal; Ramesh Ashwin; Mohammad Zaki Babu; Debapriya Zadeh; Fillia Banerjee; Makedon",
"journal": "Technologies",
"ref_id": "b8",
"title": "A Survey on Contrastive Self-supervised Learning",
"year": ""
},
{
"authors": "Jakob Nikolas Kather; Cleo-Aron Weis; Francesco Bianconi; Susanne M Melchers; Lothar R Schad; Timo Gaiser; Alexander Marx; Frank Gerrit Zollner",
"journal": "Scientific Reports",
"ref_id": "b9",
"title": "Multi-class texture analysis in colorectal cancer histology",
"year": "2016"
},
{
"authors": "D P Kingma; L J Ba",
"journal": "",
"ref_id": "b10",
"title": "Adam: A Method for Stochastic Optimization",
"year": "2015"
},
{
"authors": "Andreas Kirsch; Jishnu Mukhoti; Joost Amersfoort; H S Philip; Yarin Torr; Gal",
"journal": "",
"ref_id": "b11",
"title": "On pitfalls in ood detection: Entropy considered harmful",
"year": "2021"
},
{
"authors": "Alex Krizhevsky",
"journal": "",
"ref_id": "b12",
"title": "Learning multiple layers of features from tiny images",
"year": "2009"
},
{
"authors": "Yann Lecun; Corinna Cortes; Chris Burges",
"journal": "",
"ref_id": "b13",
"title": "MNIST handwritten digit database",
"year": "1999"
},
{
"authors": "Zheyan Shen; Jiashuo Liu; Yue He; Xingxuan Zhang; Renzhe Xu; Han Yu; Peng Cui",
"journal": "Journal of Latex Class Files",
"ref_id": "b14",
"title": "Towards Out-Of-Distribution Generalization: A Survey",
"year": "2015"
},
{
"authors": "Hyun Oh Song; Yu Xiang; Stefanie Jegelka; Silvio Savarese",
"journal": "",
"ref_id": "b15",
"title": "Deep Metric Learning via Lifted Structured Feature Embedding",
"year": "2016"
},
{
"authors": "Rohan Taori; Achal Dave; Vaishaal Shankar; Nicholas Carlini; Benjamin Recht; Ludwig Schmidt",
"journal": "",
"ref_id": "b16",
"title": "Measuring Robustness to Natural Distribution Shifts in Image Classification",
"year": "2020"
},
{
"authors": " Tf Devs",
"journal": "",
"ref_id": "b17",
"title": "TensorFlow Datasets: A collection of ready-to-use datasets",
"year": "2022"
},
{
"authors": "Sushil Sunil Thulasidasan; Sayera Thapa; Gopinath Dhaubhadel; Tanmoy Chennupati; Jeff A Bhattacharya; Bilmes",
"journal": "IEEE",
"ref_id": "b18",
"title": "An effective baseline for robustness to distributional shift",
"year": "2021"
},
{
"authors": "Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Łukasz Kaiser; Illia Polosukhin",
"journal": "",
"ref_id": "b19",
"title": "Attention Is All You Need",
"year": "2017"
}
] | [
{
"formula_coordinates": [
2,
151.59,
454.79,
308.81,
10.59
],
"formula_id": "formula_0",
"formula_text": "H(B i | D) = -p log 2 p -(1 -p) log 2 (1 -p), where p = P (B i = 1 | D)."
},
{
"formula_coordinates": [
2,
256.32,
661.71,
243.81,
22.31
],
"formula_id": "formula_1",
"formula_text": "ri = m i r i + (1 -m i ) 1 2 . (2"
},
{
"formula_coordinates": [
2,
500.13,
668.76,
3.87,
8.64
],
"formula_id": "formula_2",
"formula_text": ")"
},
{
"formula_coordinates": [
3,
106.83,
258.74,
190.95,
8.74
],
"formula_id": "formula_3",
"formula_text": "(|r|, k) ∈ {64, 128, 256, 512} × {3, 5, 10, 20}."
},
{
"formula_coordinates": [
9,
174.64,
263.52,
262.72,
22.31
],
"formula_id": "formula_4",
"formula_text": "P (win|c x , c y ) = Σ k v=1 1 v • (f x f y ) v-1 • (1 -f x f y ) k-v • k -1 v -1 ."
},
{
"formula_coordinates": [
9,
170.79,
346.74,
270.41,
29.58
],
"formula_id": "formula_5",
"formula_text": "P (win|c x , c y ) = Σ k v=1 P (win|c x , c y , V = v) • P ( Ṽ = v -1|c x , c y ) = Σ k v=1 P (win|c x , c y , V = v) • P ( Ṽ = v -1)"
},
{
"formula_coordinates": [
9,
108,
403.21,
329.36,
64.12
],
"formula_id": "formula_6",
"formula_text": "P ( Ṽ = v -1) = (f x f y ) v-1 • (1 -f x f y ) k-1-(v-1) • k -1 v -1 Hence P (win|c x , c y ) = Σ k v=1 1 v • (f x f y ) v-1 • (1 -f x f y ) k-v • k -1 v -1 ."
},
{
"formula_coordinates": [
9,
143.45,
490.96,
317.76,
152.67
],
"formula_id": "formula_7",
"formula_text": "P (win|c x , c y )P (c x , c y ) = f x f y • Σ k v=1 1 v • (f x f y ) v-1 • (1 -f x f y ) k-v • k -1 v -1 = Σ k v=1 1 v • (f x f y ) v • (1 -f x f y ) k-v • k -1 v -1 = 1 k Σ k v=1 (f x f y ) v • (1 -f x f y ) k-v • k v = 1 k (f x f y + 1 -f x f y ) k -(1 -f x f y ) k = 1 k 1 -(1 -f x f y ) k = 1 k - 1 k (1 -f x f y ) k"
},
{
"formula_coordinates": [
9,
211,
665.58,
189.99,
68.62
],
"formula_id": "formula_8",
"formula_text": "P (win) = 4 k - 1 k (1 -f x • f y ) k +(1 -f x • (1 -f y )) k +(1 -(1 -f x ) • f y ) k +(1 -(1 -f x ) • (1 -f y )) k ."
},
{
"formula_coordinates": [
10,
108,
119.7,
397.48,
23.89
],
"formula_id": "formula_9",
"formula_text": "P (win) = 2 b k - 1 k (1 -f 1 • • • f b ) k + (1 -(1 -f 1 )f 2 • • • f b ) k + . . . + (1 -(1 -f 1 ) • • • (1 -f b )) k"
},
{
"formula_coordinates": [
10,
174.62,
209.53,
263.46,
68.28
],
"formula_id": "formula_10",
"formula_text": "∂P (win) ∂f 1 = f 2 • • • f b (1 -f 1 • • • f b ) k-1 -f 2 • • • f b (1 -(1 -f 1 )f 2 • • • f b ) k-1 + . . . -(1 -f 2 ) • • • (1 -f b )(1 -(1 -f 1 ) • • • (1 -f b )) k-1"
}
] |
2023-08-15 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b47",
"b54",
"b60",
"b3",
"b10",
"b11"
],
"table_ref": [],
"text": "Estimating a map along with camera poses from a collection of images is a long-standing and challenging problem in computer vision [9,41,49] with relevant applications in domains such as autonomous driving and augmented reality. Recently, deep visual detectors and descriptors have shown increased resilience to extreme viewpoint and appearance changes [4, 48,55,61,64]. However, a common limitation among all these deep descriptors and detectors is the lack of a probabilistic formulation for detection noise. Consequently, downstream pose estimation relies on the assumption of constant spatial covariances, as shown in Figure 1 (a), leading to suboptimal results.\nWe propose to model the spatial covariance of detected keypoints in deep detectors. We find that recent detectors share a common design, where a deep convolutional backbone is used to predict a score map that assigns a \"probabil-ity\" to a pixel being a point of interest. We exploit this common design space, to propose two post-hoc methods that estimate a covariance matrix for each keypoint detected in any pretrained feature detector. Our simplest method uses the score at a given pixel to initialize an isotropic covariance matrix. We illustrate the learned score maps (lower triangle) overlaid on exemplar images, along with our proposed isotropic covariances in Figure 1 (b) for two state-ofthe-art deep feature detectors, Superpoint [11] and D2Net [12]. We also suggest a theoretically-motivated method to estimate the full spatial covariance matrix of detected keypoints using the local structure tensor. The structure tensor models the local saliency of the detections in the score map. Figure 1 (c) shows the deduced full covariances. These covariances capture the larger uncertainty along edges on the learned score map. To the best of our knowledge, we are the first to model spatial uncertainties of deep detectors.\nWe show in a series of experiments that our proposed methods for modeling the spatial covariances of detected features are directly related with the errors in matching. Accounting for them allows us to improve performance in tasks such as solving the perspective-n-point (PnP) problem and nonlinear optimizations."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b50",
"b5",
"b0",
"b31",
"b17",
"b22",
"b56",
"b13",
"b25",
"b6",
"b7",
"b27",
"b38",
"b58",
"b62",
"b13",
"b14",
"b39",
"b29",
"b4",
"b10",
"b11",
"b42",
"b57",
"b59",
"b61",
"b10",
"b11",
"b42",
"b57",
"b4",
"b23",
"b53",
"b55",
"b49"
],
"table_ref": [],
"text": "We propose to model the spatial covariances of learned local features. Modeling the spatial uncertainty of features is not a new idea, and has been studied for hand-crafted detectors. In this section, we first review hand-crafted methods and how uncertainty estimation has been proposed for them. We then describe recent progress in learned detectors.\nHandcrafted local features. Harris et al. [22] pioneeres rotationally-invariant corner detection by using a heuristic over the eigenvalues of the local structure tensor, while Shi et al. [51] proposes its smallest eigenvalue for detection. Mikolajczyk and Schmid [37] robustifies it against scale and affine transformations. On the other hand, SIFT (DoG) [34] popularizes detection (and description) of blobs. SURF [6] reduces its execution time by using integral images and (A-)KAZE [1,2] proposes non-linear diffusion to improve the invariance to changes in scale. Lastly, FAST [45] and its extensions [32,46] stand out for achieving the lowest execution times, thanks to only requiring intensity comparisons between the neighboring pixels of an image patch.\nUncertainty quantification in handcrafted detectors. Inclusion of spatial uncertainty of local features for estimation of geometry has been extensively studied [18,23,25,57]. However, its quantification on classical local features is still recognized as an open problem in the literature [14,26,27] and which has not been addressed yet with learned detections. Several works have shown the benefits of a precise quantification of uncertainty. For this purpose, they adapt the quantifications to specific classical detectors [8,17,28,39,59,63], assume planar surfaces [14,42] and require an accurate offline calibration [15,16]. Only recently [40] proposes a learning-based approach of spatial covariances, trained per detector, in order to weigh normalized epipolar errors [30,33]. In contrast to previous works, we propose a general formulation for their quantification, directly applicable on state-of-the-art learned detectors, seamlessly fitting them by leveraging their common characteristics.\nLearned local features and lack of uncertainty. The dominant approach [5,11,12,35,43,58,60,62] consists on training CNNs to regress a score map over which detections are extracted via non-maximum-suppresion (NMS). Superpoint [11] is an efficient detector-descriptor of corners, robust to noise and image transformations thanks to a synthetic pre-training followed by homographic adaptation on real images. D2Net [12] shows the applicability, to the problem of local feature detection and description, of classification networks [52]. R2D2 [43] proposes a reliability measure, used to discard unmatchable points. Similarly, DISK [58] bases its learning on matches, and KeyNet [5], inspired by classical systems, proposes to learn from spatial image gradients. This work experiments with Superpoint, D2Net, R2D2, and KeyNet, but our proposal is applicable to the rest of the systems. Finally, recent works [24,54,56] exploit attention mechanisms to match pixels without explicitly using detectors. Although their inclusion in geometric estimation pipelines can be engineered [50], they suffer from lack of repeatability. Because of this, in this work, we focus on the quantification of spatial uncertainty for learned detectors."
},
{
"figure_ref": [
"fig_0"
],
"heading": "Method",
"publication_ref": [],
"table_ref": [],
"text": "Our framework takes as input a RGB image I of spatial dimensions H × W , and outputs local features together with their spatial uncertainties. It is composed of two main com-ponents as shown in Fig. 2. (1) A pretrained feature detector and (2) our novel and detector-agnostic uncertainty module. Our methods work with any pretrained state-ofthe-art detectors, taking their learned score maps as input, and outputting reliable covariance estimates for each detected keypoint. We show that the estimated covariances are well-calibrated and improve downstream tasks such as PnP and motion-only bundle adjustment."
},
{
"figure_ref": [
"fig_1"
],
"heading": "Overview",
"publication_ref": [
"b4",
"b10",
"b11",
"b42",
"b57",
"b59",
"b61",
"b22",
"b56"
],
"table_ref": [],
"text": "Pre-trained local-feature detector. Our framework is directly applicable to the vast majority of learned detectors [5,11,12,35,43,58,60,62]. These detectors share a common architectural design, using convolutional backbones to predict a score map S ∈ R kH×kW with k ∈ R + , followed by non-maximum-suppression (NMS) to extract a sparse set of features. We leverage this standard design to make a detector-agnostic, post-hoc uncertainty module that takes the estimated score maps as input and estimates the spatial uncertainty based on the peakiness in a local region around the detected point. Our approach can be applied to these detectors without any training or fine-tuning.\nDetector-Agnostic Feature Uncertainties. Instead of considering the 2D position of detections as deterministic locations, we propose to model their spatial uncertainties. More formally, we consider that the spatial location x i ∈ R 2 of a local feature i, detected in I, stems from perturbing its true location x i,true ∈ R 2 , with random noise ξ i ∈ R 2\nx i = x i,true + ξ i ,(1)\nwhose probability distribution we want to estimate. We follow the dominant model in computer vision [23,25,57], which uses second order statistics to describe the spatial uncertainty of each location\nE{ξ i } = 0 , Σ i := Cov(ξ i ) = E{ξ i ξ ⊤ i } , ∀i . (2)\nThereby our goal is to quantify each covariance matrix Σ i . To this end, we propose two methods: (1) A point-wise estimator of isotropic covariances (Sec. 3.2) that uses the inverse of the scores at the detected point, and (2) a full covariance estimator based on the local structure tensor (Sec. 3.3) which models the local saliency to estimate the uncertainty in all directions. We show that both approaches lead to reliable uncertainty estimates, benefiting downstream tasks. Intuitively, a peaky score map at the location of a detected feature will yield low spatial uncertainty, whereas a flatter score map will yield larger spatial uncertainty. Figure 3 qualitatively shows how our deduced 2D uncertainties relate to 3D uncertainties. In the Supplemental we explore the agnostic behavior of our covariances. \nϕ([∇ x S, ∇ y S]) = [(∇ x S) 2 , (∇ y S) 2 , ∇ x S∇ y S]"
},
{
"figure_ref": [],
"heading": "Point-wise Estimation",
"publication_ref": [],
"table_ref": [],
"text": "As the simplest estimator of the spatial uncertainty, we propose to use the regressed score of each local feature to create an isotropic covariance matrix\nΣ i := 1 S(x i ) I 2×2 = 1/S(x i ) 0 0 1/S(x i ) .(3)\nFig 1 shows that this estimator yields isotropic predictions of uncertainty (equal in all directions), so it only quantifies the relative scale regardless of the learned local structure."
},
{
"figure_ref": [
"fig_2"
],
"heading": "Structure Tensor",
"publication_ref": [
"b4",
"b10",
"b11",
"b42",
"b46",
"b43",
"b19",
"b35"
],
"table_ref": [],
"text": "Quantification of local saliency motivates the use of the local structure tensor,\nC i ∈ R 2×2 [22]. Defining [∇ x S i , ∇ y S i ] := ∂S/∂x| xi as the spatial gradient of S eval- uated at x i , C i in its local neighborhood W i (window of size u × v) is given by C i := j∈Wi w j ∂S ∂x ⊤ xj ∂S ∂x xj = j∈Wi w j (∇ x S j ) 2 ∇ x S j ∇ y S j ∇ y S j ∇ x S j (∇ y S j ) 2 ,(4)\nPoint-wise (Sec. 3.2)\nStructure Tensor (Sec. 3.3) Zoom Key.Net [5] Superpoint [11] D2Net [12] R2D2 [43] Figure 5. Qualitative comparison. Our estimated covariance matrices for each detected local feature are illustrated with uncertainty ellipses for four state-of-the-art pretrained detectors. Our pointwise covariance estimator yields isotropic uncertainty predictions (equal in all directions). Our proposed local structure tensor estimates the directionality of the spatial uncertainty, resulting in anisotropic predictions. For instance, Superpoint (which is trained to detect corners) has high uncertainty patterns corresponding to lines. For visualization purposes, a uniform sample of local feature detections is shown.\nwith w j ∈ R + being the weight of pixel j, preferably defined by a Gaussian centered at x i [47]. As such, C i is a positive semidefinite (PSD) matrix, resulting from averaging the directionality of gradients and hence not canceling opposite ones (see Fig. 4).\nThe reason why C i captures the local saliency lies in the auto-correlation function, c : R u×v → R, which averages local changes in S given small displacements δx [22]:\nc i = j∈Wi w j (S(x j ) -S(x j + δx)) 2 ,(5)\nwith S(x j ) indicating the score at x j . Linearly approximating S(x j + δx) ≈ S(x j ) + ∂S/∂x| xj δx, yields\nc i ≈ j∈Wi w j (S(x j ) -S(x j ) - ∂S ∂x xj δx) 2 ,(6)\n= j∈Wi w j δx ⊤ ∂S ∂x ⊤ xj ∂S ∂x xj δx = δx ⊤ C i δx . (7)\nThus, extreme saliency directions are obtained by solving:\nmax δx min δx δx ⊤ C i δx, s.t. ∥δx∥ = 1 ,(8)\nwhere the constraint ∥δx∥ = 1 ensures non-degenerate directions, which can be obtained with Lagrange multipliers [44] i.e. by defining the Lagrangian\nL(δx, λ) := δx ⊤ C i δx -λ(δx ⊤ δx -1) ,(9)\ndifferentiating w.r.t. δx and setting it to 0:\n2δx ⊤ C i -2λδx ⊤ = 0 ⇒ C i δx = λδx ,(10)\nwe conclude that directions of extreme saliency correspond to the eigenvectors of C i . Since inverting a matrix, does the same to its eigenvalues without affecting its eigenvectors 1 ,\nΣ i := C -1\ni results in a proper covariance matrix (PSD) assigning greater uncertainty in the direction of less saliency and vice-versa. [20,29] and motivated by the principle of maximum entropy [36] of the local features' scores, distributed independently by"
},
{
"figure_ref": [],
"heading": "Statistical interpretation. Under a Gaussian model of aleatoric uncertainty, common in deep learning",
"publication_ref": [
"b27",
"b56"
],
"table_ref": [],
"text": "N (S(x i + t i ), σ 2 ) ,(11)\ninspired by [28], we can set up a parametric optimization of t ∈ R 2 which maximizes the likelihood, L(t i | S) with S := {S(x j ) | x j ∈ W i }, of the observations:\nS(x i ) = S(x i + t i ) + ε i , with ε i ∼ N (0, σ 2 ) ,(12)\n1 Let A ∈ R m×m and det A ̸ = 0, then λ -1 v = A -1 v, with λ and v being the eigenvalues and eigenvectors of A.\nwhere S(x i ) is the observed score, perturbed by the random noise ε i , independently affecting the rest of observations. Thus, the optimization is formulated as follows\nti = arg max ti j∈Wi 1 σ √ 2π exp - (S(x j ) -S(x j + t i )) 2 2σ 2 . (13\n)\nIts solution, or maximum-likelihood-estimation (MLE), is known a priori: ti = 0, which is unbiased given our statistical model (Eq. 12), and coherent with Equation 2.\nWith these conditions, the inverse of the Fisher information matrix, I( t), evaluated at the MLE, imposes a lower bound in the covariance matrix of the estimator ti = 0, known as Cramer-Rao Lower Bound (CRLB) [21]:\nVar( ti ) ≥ I( ti ) -1 .(14)\nI( t) is defined as E{s ⊤ i s i }, being the variance of the loglikelihood derivative (known as score)\ns i := ∂ log L(t i | S)/∂t i , since E{s i } = 0 [21].\nIn our case, it is given by\ns i = xj ∈Wi d j := S(x j ) -S(x j + t i ) σ 2 ∂S(x j + t i ) ∂t i .(15)\nDue to the linearity of expectation and the independence of observations,\nE{d ⊤ j d k } = 0, ∀j ̸ = k. Thereby I( ti ) = j∈Wi E{d ⊤ j d j } ti .(16)\nSince derivatives of Eq. 15 are applied on our deterministic model, they can go out of the expectation, and evaluating them on ti = 0 lead to\nI( ti ) = xj ∈Wi 1 σ 4 ∂S(x j ) ∂x ⊤ xj ∂S(x j ) ∂x xj E{(S(x j ) -S(x j + t i )) 2 } ti .(17)\nLastly, E{(S(x j ) -S(x j + t i )) 2 } ti = E{ε 2 i } = σ 2 since E{ε i } = 0, implying that our Fisher information matrix is\nI( ti ) = 1 σ 2 xj ∈Wi ∂S(x j ) ∂x ⊤ xj ∂S(x j ) ∂x xj ,(18)\nat the MLE. It matches the local structure tensor (Eq. 4) up to a scale factor Var(ε i ) -1 = σ -2 , unknown a priori.\nRecalling the CRLB (Eq. 14), although achievable only asymptotically [21,57], it motivates Σ i := C -1 i as an upto-scale covariance matrix of each location x i ."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [],
"table_ref": [],
"text": "Implementation details. In our experiments, we compute the structure tensor (and relative uncertainty) independently of the learned detector: for each local feature i, spatial differentiation at S(x j ), ∀j ∈ W i , is done with Sobel filters. Integration in W i is done with a 7 × 7 window, MMA overall 0.42 0.71"
},
{
"figure_ref": [],
"heading": "D2Net",
"publication_ref": [],
"table_ref": [],
"text": "Full Isotropic "
},
{
"figure_ref": [],
"heading": "SuperPoint",
"publication_ref": [
"b2",
"b4",
"b10",
"b11",
"b42",
"b18",
"b52"
],
"table_ref": [],
"text": "Full Isotropic -uncertainty + Figure 6. Uncertainty and matching accuracy on HPatches [3]. We show MMA for each detector. We organize the results into 10 uncertainty ranges, ranging from lowest to highest (x-axis), each containing an equal number of matches. Our covariances accurately model less accurate matches, with the full-based approach (structure tensor) showing higher sensitivity in D2Net, Key.Net and R2D2.\nthe result of using an isotropic Gaussian filter with σ = 1 of cutoff frequency 3σ. Throughout all the experiments, we evaluate and extract our covariances using the detectors of state-of-the-art learned systems: Key.Net [5], Superpoint [11], D2Net [12] and R2D2 [43]. The score map used with Superpoint is the one prior to the channel-wise softmax to avoid the alteration of learned patterns crossing the boundaries of each grid. For the rest of the systems, we directly use their regressed score map. Diverse qualitative examples are shown in Figure 5.\nFigure 7. Qualitative results on KITTI [19]. Our covariances correctly model detections with unclear spatial locations, such as faraway corners or those located at edges.\nFigure 8. Qualitative results on TUM-RGBD [53]. We assign more spatial uncertainty at a priori less distinctive regions, such as papers and edges of the keyboard."
},
{
"figure_ref": [],
"heading": "Matching accuracy.",
"publication_ref": [
"b2",
"b11",
"b11",
"b52",
"b18"
],
"table_ref": [],
"text": "We first test the relation between our estimated covariances with the accuracy of local-feature matching. Intuitively, local features detected with higher uncertainty should relate to less accurate matches, and vice versa. For this purpose, we consider the widely adopted HPatches dataset [3]. HPatches contains 116 sequences of 6 images each. 57 sequences exhibit significant illumination changes while the other 59 sequences undergo viewpoint changes.\nEvaluation protocol. We base our evaluation on the one proposed by [12]. First, extraction of local features and, in our case, covariance matrices of their locations is performed for all images. For every sequence, pairwise matching is done between a reference image r and each remaining image i by using Mutual Nearest Neighbor search (MNN). We then compute the reprojection errors and their covariances with the homographies, H i,r , provided by the dataset:\ne i,r := cart(x i -H i,r xr ) ,(19)\nΣ ei,r := JΣ xr J ⊤ + Σ xi ,(20)\nwhere cart(•) maps from homogeneous to Cartesian coordinates, and J := ∂e i,r /∂x r , i.e. we linearly propagate each covariance matrix Σ xr of the reference locations.\nTo quantify the uncertainty of the match with a scalar, we use the biggest eigenvalue of the corresponding Σ ei,r . Based on them, all matches gathered in the dataset are distributed in 10 ranges from lowest to highest uncertainty estimates, such that each range has the same number of matches. To quantify the accuracy in matching at each range, we choose the mean matching accuracy error (MMA), which represents the average percentage of matches with a corresponding value of ∥e i,r ∥ below a threshold. We use the same thresholds as in [12]. Finally, we compute the mean of all the MMA values at each range. This process is repeated for all the evaluated detectors and with the proposed full and isotropic covariances.\nResults. Figure 6 shows the averaged mean matching accuracy, MMA, at each uncertainty range. Ranges are ordered from lowest (1) to highest (10) estimated uncer- [53] and KITTI [19]. We report the cumulative errors for camera pose rotation and translation. Practically all estimations converge to acceptable thresholds when using our 2D full and 2D isotropic covariances. This is also apparent when using 3D covariances derived from our 2D ones. Without our covariances, a significant percentage of poses fail to localize. tainty. As can be seen, it exists a direct relationship between matching accuracy and both full and isotropic covariance estimates. With full covariances and for all evaluated ranges, lower uncertainty estimates imply higher accuracy in matching. However, this is not always the case when using isotropic covariances. As can be seen with R2D2, there is a certain increase in MMA for more uncertain matches. Additionally, there is a higher sensitivity of MMA to the uncertainty estimates stemming from full covariances on D2Net, Key.Net, and R2D2. This motivates the need for taking into account the learned local structure when quantifying the spatial uncertainty of the local feature, rather than basing it only on the regressed scalar estimate of the regressed score map."
},
{
"figure_ref": [],
"heading": "Geometry estimation",
"publication_ref": [
"b58",
"b18",
"b52",
"b12",
"b22",
"b30",
"b58",
"b18",
"b52"
],
"table_ref": [],
"text": "To test the influence of our uncertainties in 3D-geometry estimation, we follow the evaluation proposed in [59]. It covers common stages in geometric estimation pipelines such as solving the perspective-n-point problem and motion-only bundle adjustment. The data used consists in the three sequences 00-02 of KITTI [19] and the first three 'freiburg 1' monocular RGB sequences of TUM-RGBD [53].\nEvaluation protocol. KITTI is used with a temporal win-dow of two left frames, while three are used in TUM RGB-D (each with a pose distance > 2.5 cm). Features and our 2D covariance matrices are extracted with the evaluated detectors. Pairwise matching is done across frames with MNN. In TUM, since more than two images are used, we form feature tracks (set of 2D local features corresponding to the same 3D point) with the track separation algorithm of [13]. Matched features are triangulated with ground-truth camera poses and DLT algorithm [23], and refined with 2D-covariance-weighted Levenberg-Marquardt (LM), producing also covariance matrices for 3D point coordinates. The next frame is used for evaluation. After matching it to the reference images we obtain 2D-3D matches which are then processed with P3P LO-RANSAC [10] to filter potential outliers. Given the potential inliers, and when using no uncertainty, we choose EPnP [31] as the nonminimal PnP solver. Otherwise, when leveraging our proposed 2D covariances, and optionally, the 3D covariances from LM, we use our implementation (validation in Supp.) of EPnPU [59]. Finally, the estimated camera pose is refined with a covariance-weighted motion-only bundle adjustment. In the Supplemental, we detail how the inclusion of our uncertainty estimates is done in the previous tasks.\nTo quantify the accuracy in pose estimation at each 1. Motion estimation from 2D-3D correspondences on KITTI [19] and TUM-RGBD [53]. Sequences are specified at the leftmost column ('all' is their aggregation). erot and et are the mean absolute rotation (in 0.1×deg.) and translation (in cm.) errors. We compare estimations without using uncertainty (first two columns) and with 2D and 3D uncertainties via the proposed full and isotropic covariances. Errors are consistently reduced when using uncertainty estimates. The best result for each sequence-detector is in bold.\nsequence, we use the absolute rotation error in degrees: e rot = arccos(0.5 trace(R ⊤ true R -1)), and the absolute translation error e t = ∥t true -t∥ in cm., where R true , t true is the ground-truth pose and R, t is the estimated one."
},
{
"figure_ref": [
"fig_3"
],
"heading": "Results",
"publication_ref": [
"b58"
],
"table_ref": [],
"text": ". Following [59], in Table 1, we report the mean errors obtained across sequences of both datasets. As can be seen, taking into account the proposed uncertainties is a key aspect to converge pose estimations across sequences. Figures 7 and8 show qualitative examples of why uncertainties help geometric estimations by assigning more uncertainty to distant or less reliable keypoints. This behavior can be understood better by having a look at the cumulative error curves. In Figure 9, it is shown that practically all pose estimations obtained with methods leveraging the proposed covariances, fall under acceptable error thresholds, whereas the ones from the baseline do not."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [
"b30",
"b58",
"b56",
"b52",
"b18"
],
"table_ref": [],
"text": "The proposed covariances modeling the spatial uncertainty of the learned local-features are up-to-scale. This is not an issue for common 3D geometric estimation algorithms, such as solving linear systems [31,59] or nonlinear leastsquares optimizations [57], as their solutions depend only on the relative weight imposed by the covariance matrices. However, this limitation hinders reasoning about the covariances in pixel units. For instance, extracting the absolute scale factor would facilitate the use of robust cost functions, pointing towards a potential direction for future work.\nAdditionally, while we achieved improvements in 3D geometric estimation tasks on the standard datasets TUM-RGBD [53] and KITTI [19], their exposure to effects like illumination changes or different types of camera motions might be limited. These effects may pose challenges to our approach, as we believe it is subject to the equivariance of the noise of the learned score maps to such changes."
},
{
"figure_ref": [],
"heading": "Conclusions",
"publication_ref": [
"b58",
"b30",
"b30",
"b30"
],
"table_ref": [],
"text": "In this paper, we formulate, for the first time in the literature, detector-agnostic models for the spatial covariances of deep local features. Specifically, we proposed two methods based on their learned score maps: one using local-feature scores directly, and another theoretically-motivated method using local structure tensors. Our experiments on KITTI and TUM show that our covariances are well calibrated, significantly benefiting 3D-geometry estimation tasks.\nPnP. Consider a set of n 3D points {p w,i ∈ R3 | i ∈ 1 . . n} expressed in an absolute reference system {w}, along with their corresponding 2D points {x i ∈ R 2 | i ∈ 1 . . n} in an image captured by a camera with calibration matrix K ∈ R 3×3 . The Perspective-n-Point problem (PnP) involves finding the rotation R cw ∈ SO(3) and translation t cw ∈ R 3 that transform the 3D points to the camera's reference system {c}: p c,i := R cw p w,i + t cw .\nEPnP(U). To leverage our proposed covariance matrices, we adopt the recent EPnPU [59] as our PnP solver. EPnPU extends EPnP [31] to take into account the uncertainty of the observations. To understand how uncertainties are included, consider the algebraic residual r i (i ∈ 1 . . n):\nr i := x(1,2) i - x(3) i x i ,(21)\nwhere xi represents, in homogeneous coordinates, the estimated 2D location corresponding to p w,i :\nxi := K(R c,w p w,i + t c,w ) ,(22)\nwhere x(1,2) i (resp.\nx(3) i ) represents its two first entries (resp. third entry). Given this, we seek to estimate the covariance matrix of each residual, Σ ri := Cov(r i ), that stems from the covariance matrices of the observations Σ xi and Σ pw,i .\nEach residual can be linearized by instead considering a set of control points as the unknowns 2 of the problem, whose concatenation we denote here as c ∈ R 12 , and a matrix block M i ∈ R 2×12 that depends only on the input data [31]. Thus, the solution to c is found in the null space of the matrix formed by the concatenation of each residual [31,Eq. 7]:\nMc = 0 , with M := M ⊤ 1 • • • M ⊤ n ⊤ . (23\n)\nThanks to this, we can weigh the influence of the residuals according to their projections onto the directions of extreme uncertainty:\nM ⊤ Σ -1 r Mc = 0 ,(24)\nwith Σ r := diag(Σ r1 , . . . , Σ rn ) .\n(25)\n2 To ease uncertainty propagation from p w,i to xi , we instead parameterize xi with the pose (Rcw, tcw), as in Eq. (22). Because of this, we need an initial rough pose estimate. This pose estimate is obtained at no additional cost during the outlier filtering step with P3P LO-RANSAC.\nΣ ri derivation. For the noise ν ∈ R 3 affecting the coordinates of each 3D point, we assume a similar model to the one proposed for the 2D location x i of each local feature (Eqs. ( 1) and (2)):\np w,i = E{p w,i } + ν ,(26)\nE{ν} = 0 , Σ pw,i := E{νν ⊤ } .(27)\nwhere, as we do in our case, the covariance matrix Σ pw,i ∈ R 3×3 can be estimated after the convergence of the nonlinear optimization of p w,i (Appendix A.2). Plugging Eq. ( 26) in Eq. ( 22) leads to\nxi = KRp w,i + Kt = E{x i } + ζ ,(28)\nwhere E{ζ} = 0, because ζ := KRν, and E{ν} = 0 (by definition in Eq. ( 26)). According to Eq. ( 28), we can propagate the uncertainty of p w,i to xi by:\nΣ xi := Λ w w ⊤ γ := KRΣ pw,i R ⊤ K ⊤ .(29)\nEach residual can then be expressed as a random variable as follows:\nr i = x(1,2) i - x(3) i x i ,(30)\n= (E{x\n(1,2) i } + ζ (1,2) ) -(E{x(3)\ni } + ζ (3) )(E{x i } + ξ) ,(31)\n= E{x (1,2) i - x(3) i x i } + ζ (1,2) -E{x (3) i }ξ -ζ (3) E{x i } -ζ (3) ξ ,(32)\nwhere we have used the linearity of the expectation and assumed independence between ξ and ζ. Following Eq. ( 30) and the fact that\nE{r i } = E{x (1,2) i - x(3\n) i x i } since E{ξ} = 0 2 , E{ζ} = 0 3 , the derivation of Σ ri follows as in Tab. 2."
},
{
"figure_ref": [],
"heading": "A.2. Nonlinear optimizations",
"publication_ref": [
"b22",
"b56",
"b22",
"b30",
"b58"
],
"table_ref": [],
"text": "Uncertainty-based geometric refinement. Commonly, initial estimates of camera poses and 3D points (obtained, in our case, using EPnP(U) and multi-view DLT triangulation [23], respectively) are not geometrically optimal 3 . Therefore, they are typically optimized using iterative algorithms such as Gauss-Newton or Levenberg-Marquardt [57] by minimizing reprojection errors. This approach is considered in the literature as the gold standard [23]. known beforehand. Comparisons are done with OpenCV's implementation of EPnP [31]. Since EPnPU is an extension of EPnP's algorithm to include uncertainties, results should improve accordingly when leveraging them. Additionally, we compare our implementation when using just identity covariance matrices, to verify that it matches the behavior of EPnP.\nAs shown in Figure 10, our EPnPU implementation gives results very similar to the ones reported in [59]. In turn it improves the results of OpenCV's EPnP, whose behavior also matches our EPnPU implementation when just using identity covariances matrices."
},
{
"figure_ref": [
"fig_4",
"fig_5"
],
"heading": "C. Interpretability",
"publication_ref": [
"b10",
"b4",
"b11",
"b42"
],
"table_ref": [],
"text": "Our proposed methods for quantifying the uncertainty of the locations are based on the learned score maps, independently of the detector that has learned it. Depending on its training, systems learn to focus on different input image pat- terns. For instance, SuperPoint [11] is trained to detect corners, while Key.Net [5], D2Net [12] and R2D2 [43] do not directly impose such constraint. Additionally, all of them use different learning objectives.\nTo explore what kind of locations get assigned low uncertainty estimates, we set up a toy experiment inspired by DeepDream [38]. As depicted in Fig. 11 we update a 20 × 20 synthetic input patch via gradient descent, such that we minimize the biggest eigenvalue of the covariance matrix (computed by using the regressed score map) at the center pixel. We downweight the gradients located at the extremes of the receptive field and, as in [38], we smooth the gradients with a Gaussian filter.\nResults after convergence are shown in Fig. 12. We obtain distinctive blob/corner-like regions by minimizing both uncertainty estimates. This highlights the detectoragnostic behavior of our two methods. Interestingly, excepting Key.Net, the generated patterns are slightly different depending on the method. We attribute this to the the fact that the full approach takes into account the surrounding learned patterns, which in this case increases the saliency of the generated input image pattern."
},
{
"figure_ref": [],
"heading": "",
"publication_ref": [],
"table_ref": [],
"text": "Acknowledgements. The authors thank Alejandro Fontán Villacampa for his thoughtful comments and help with the experiments. This work was supported by the Ministerio de Universidades Scholarship FPU21/04468."
},
{
"figure_ref": [],
"heading": "Supplementary Material",
"publication_ref": [
"b56",
"b22",
"b58",
"b58",
"b30",
"b58",
"b58"
],
"table_ref": [],
"text": "A. Accounting for uncertainty A.1. Perspective-n-Point problem\n= E{ζ (1,2) ( ( ( ( ( ( ( ( ( E{E{x (3) i }ζ (1,2) ξ ⊤ } E{ζ (1,2) }=E{ξ}=0 -E{ζ (3) ζ (1,2) E{x i } ⊤ } w E{xi} ⊤ -( ( ( ( ( ( ( ( E{ζ (3) ζ (1,2) ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( E{(\nTable 2. Derivation of Σr i (Appendix A.1).\nConsider a 3D point j expressed in {w}: p w,j ∈ R 3 , that is observed from a camera i whose pose is: R cw,i ∈ SO(3), t cw,i ∈ R 3 , then we define the reprojection error as\nwhere x ij ∈ R 2 is the 2D location in the image of a local feature corresponding to p w,j and π(•) is the projection function. Thereby, the variables of interest (poses and/or 3D points), which we represent for convenience in vectorized form as y, are refined with the following optimization\nwith e(y) := . . . e ij (y) ⊤ . . . ⊤ , (40)\nwhere W ij := Σ -1 ij represents the inverse of the covariance matrix of e ij (we use the identity matrix W ij := I 2 in the no-uncertainty baseline), thus weighing each e ij according to its projection onto the directions of extreme uncertainty.\nIn our experiments, we approximate the solution of Eq. ( 38) with ten iterations of Levenberg-Marquardt (LM) [57] i.e. we iteratively update y ← y ⊕ ∆y within the manifold of the unknowns y [7], by composing them with increments ∆y computed by solving the following linear system of equations\nwhere J := ∂e ∂∆y 0 is the Jacobian matrix of the residuals. As such, LM imposes a penalization of the magnitude of ∆y that is controlled with the damping factor λ ∈ R. We initialize λ as 10 -3 times the average of diag(J ⊤ WJ), as recommended in [23]. We also follow the recommended protocol of [23, App. A6.2] for updating λ at each iteration. [59] implementation with known synthetic noise. Our implementation, labeled as 'EPnPU imp' obtains pose errors akin to the ones of [59], improving as it should over EPnP [31], labeled as 'EPnP OpenCV', whose behavior is matched when using our implementation with identity covariance matrices, labeled as 'EPnP imp'.\nCovariances for 3D-points refinement. Since we use ground-truth camera poses for triangulation in our experiments, we directly use our proposed 2D covariances Σ ij := Σ xij as the covariance matrix of each error e ij . After convergence, we estimate the covariance matrix Σ pw,j of the 3D point as the inverse of the Hessian, following [59].\nCovariances for motion-only bundle-adjustment. The covariance matrix of each error e j 4 , when considering 3D noise, is estimated at each iteration following [59]:\ni.e. linearly propagating Σ pw,j and assuming independence between the distributions of x j and p w,j . On the other hand, if only 2D noise is considered, we consider Σ j := Σ xj ."
},
{
"figure_ref": [],
"heading": "B. Validation of EPnPU implementation",
"publication_ref": [
"b4",
"b10",
"b11",
"b42",
"b58",
"b58"
],
"table_ref": [],
"text": "Since all learned detectors used in our paper [5,11,12,43] are implemented in Python, but EPnPU [59] is originally written in MATLAB 5 , we reimplemented this last one in Python to ease the workflow. As validation, we followed the same synthetic experiments done in [59], where the noise affecting the simulated observations (2D and 3D points) is 4 We drop here the subscript i to avoid clutter since just one camera is considered in motion-only BA. "
}
] | Superpoint D2Net (a) Constant Covariance (b) Isotropic Covariance (c) Full Covariance Figure 1. Detector-agnostic, post-hoc uncertainty for learned local-feature detectors. State-of-the-art deep feature detectors, such as Superpoint [11] and D2Net [12], do not estimate the spatial uncertainty of their detections. This corresponds to assuming constant covariances (see (a)), which lead to suboptimal performance. We propose two methods to model local features' uncertainty: (b) a pointwise isotropic covariance and (c) a structure-based full covariance estimate. We demonstrate that modeling uncertainties leads to improved performance in downstream tasks such as solving the perspective-n-point problem and nonlinear optimizations. | DAC: Detector-Agnostic Spatial Covariances for Deep Local Features | [
{
"figure_caption": "Figure 2 .2Figure 2. Method overview. By exploiting the dominant approach in learned local-feature detectors (represented by D) of regressing, for each input image I, a map of scores S := D(I), over which detections are done, we propose to quantify the uncertainty of each detected location xi by a mapping U(S, xi) agnostic to D.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 .3Figure 3. Influence of 2D uncertainties in 3D geometry. The learned score maps for two images and matched local features with highest and lowest estimated uncertainty -represented w. ellipses-for two points detected in two views. More uncertain 2D points present significantly higher 3D reprojection errors.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 .4Figure 4. Structure tensor as kernel. The direction of local gradients [∇xS, ∇yS] := ∂S/∂x can be averaged without loss of information with the local structure tensor. It acts as a kernel ϕ : R 2 → R 3 , mapping vectors with 180 • of difference (opposite directions) to the same point in R 3 .",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Figure 9 .9Figure9. Evaluation in TUM-RGBD[53] and KITTI[19]. We report the cumulative errors for camera pose rotation and translation. Practically all estimations converge to acceptable thresholds when using our 2D full and 2D isotropic covariances. This is also apparent when using 3D covariances derived from our 2D ones. Without our covariances, a significant percentage of poses fail to localize.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "9",
"figure_type": "figure"
},
{
"figure_caption": "Figure 11 .11Figure 11. Interpretability. We update a 20 × 20 input patch I by minimizing our uncertainty estimates (full and isotropic) at the center of the patch. Gradients are downweighted and smoothed, as in [38], to favor convergence.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "11",
"figure_type": "figure"
},
{
"figure_caption": "Figure 12 .12Figure 12. Interpretability results. Minimizing both uncertainty measures lead to distinctive (blob-like and corner-like) regions generated in the input patch. Differences obtained depending on learned detector highlights the detector-agnostic behavior of the proposals.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "12",
"figure_type": "figure"
}
] | Javier Tirado-Garín; Frederik Warburg; Javier Civera | [
{
"authors": "Pablo Fernández Alcantarilla; Adrien Bartoli; Andrew J Davison",
"journal": "Springer",
"ref_id": "b0",
"title": "Kaze features",
"year": "2012"
},
{
"authors": "Pablo Fernández Alcantarilla; Jesús Nuevo; Adrien Bartoli",
"journal": "",
"ref_id": "b1",
"title": "Fast explicit diffusion for accelerated features in nonlinear scale spaces",
"year": "2013"
},
{
"authors": "Vassileios Balntas; Karel Lenc; Andrea Vedaldi; Krystian Mikolajczyk",
"journal": "",
"ref_id": "b2",
"title": "Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors",
"year": "2017"
},
{
"authors": " O León Barbed; Javier Franc ¸ois Chadebecq; Morlana; M M José; Ana C Montiel; Murillo",
"journal": "Springer",
"ref_id": "b3",
"title": "Superpoint features in endoscopy",
"year": "2022-09-18"
},
{
"authors": "Axel Barroso; -Laguna ; Krystian Mikolajczyk",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b4",
"title": "Key. net: Keypoint detection by handcrafted and learned cnn filters revisited",
"year": "2022"
},
{
"authors": "Herbert Bay; Andreas Ess; Tinne Tuytelaars; Luc Van Gool",
"journal": "Computer vision and image understanding",
"ref_id": "b5",
"title": "Speeded-up robust features (surf)",
"year": "2008"
},
{
"authors": "Luis José; Blanco-Claraco",
"journal": "",
"ref_id": "b6",
"title": "A tutorial on SE(3) transformation parameterizations and on-manifold optimization",
"year": "2022"
},
{
"authors": "J Michael; Wojciech Brooks; Darren Chojnacki; Anton Gawley; Van Den; Hengel",
"journal": "IEEE",
"ref_id": "b7",
"title": "What value covariance information in estimating vision parameters",
"year": "2001"
},
{
"authors": "Carlos Campos; Richard Elvira; Juan J Gómez Rodríguez; José Mm Montiel; Juan D Tardós",
"journal": "IEEE Transactions on Robotics",
"ref_id": "b8",
"title": "Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam",
"year": "2021"
},
{
"authors": "Ondřej Chum; Jiří Matas; Josef Kittler",
"journal": "Springer",
"ref_id": "b9",
"title": "Locally optimized ransac",
"year": "2003"
},
{
"authors": "Tomasz Daniel Detone; Andrew Malisiewicz; Rabinovich",
"journal": "",
"ref_id": "b10",
"title": "Superpoint: Self-supervised interest point detection and description",
"year": "2018"
},
{
"authors": "Mihai Dusmanu; Ignacio Rocco; Tomas Pajdla; Marc Pollefeys; Josef Sivic; Akihiko Torii; Torsten Sattler",
"journal": "",
"ref_id": "b11",
"title": "D2-net: A trainable cnn for joint description and detection of local features",
"year": "2019"
},
{
"authors": "Mihai Dusmanu; Johannes L Schönberger; Marc Pollefeys",
"journal": "Springer",
"ref_id": "b12",
"title": "Multi-view optimization of local feature geometry",
"year": "2020"
},
{
"authors": "Luis Ferraz; Xavier Binefa; Francesc Moreno-Noguer",
"journal": "BMVA Press",
"ref_id": "b13",
"title": "Leveraging feature uncertainty in the pnp problem",
"year": "2014"
},
{
"authors": "Alejandro Fontan; Laura Oliva; Javier Civera; Rudolph Triebel",
"journal": "IEEE Robotics and Automation Letters",
"ref_id": "b14",
"title": "Model for multi-view residual covariances based on perspective deformation",
"year": "2022"
},
{
"authors": "Alejandro Fontan; Riccardo Giubilato; Laura Oliva; Javier Civera; Rudolph Triebel",
"journal": "IEEE Robotics and Automation Letters",
"ref_id": "b15",
"title": "Sid-slam: Semi-direct information-driven rgb-d slam",
"year": "2023"
},
{
"authors": "Wolfgang Förstner; Eberhard Gülch",
"journal": "",
"ref_id": "b16",
"title": "A fast operator for detection and precise location of distinct points, corners and centres of circular features",
"year": "1987"
},
{
"authors": "Wolfgang Förstner; Bernhard P Wrobel",
"journal": "Springer",
"ref_id": "b17",
"title": "Photogrammetric computer vision",
"year": "2016"
},
{
"authors": "Andreas Geiger; Philip Lenz; Christoph Stiller; Raquel Urtasun",
"journal": "The International Journal of Robotics Research",
"ref_id": "b18",
"title": "Vision meets robotics: The kitti dataset",
"year": "2013"
},
{
"authors": "Martin Fredrik K Gustafsson; Thomas B Danelljan; Schon",
"journal": "",
"ref_id": "b19",
"title": "Evaluating scalable bayesian deep learning methods for robust computer vision",
"year": "2020"
},
{
"authors": "Wolfgang Härdle; Léopold Simar",
"journal": "Springer",
"ref_id": "b20",
"title": "Applied multivariate statistical analysis",
"year": "2007"
},
{
"authors": "Chris Harris; Mike Stephens",
"journal": "Citeseer",
"ref_id": "b21",
"title": "A combined corner and edge detector",
"year": "1988"
},
{
"authors": "R I Hartley; A Zisserman",
"journal": "Cambridge University Press",
"ref_id": "b22",
"title": "Multiple View Geometry in Computer Vision",
"year": "2004"
},
{
"authors": "Wei Jiang; Eduard Trulls; Jan Hosang; Andrea Tagliasacchi; Kwang Moo; Yi ",
"journal": "",
"ref_id": "b23",
"title": "Cotr: Correspondence transformer for matching across images",
"year": "2021"
},
{
"authors": "Kenichi Kanatani",
"journal": "Elsevier Science Inc",
"ref_id": "b24",
"title": "Statistical Optimization for Geometric Computation: Theory and Practice",
"year": "1996"
},
{
"authors": "Kenichi Kanatani",
"journal": "Systems and Computers in Japan",
"ref_id": "b25",
"title": "For geometric inference from images, what kind of statistical model is necessary?",
"year": "2004"
},
{
"authors": "Kenichi Kanatani",
"journal": "International Journal of Computer Vision",
"ref_id": "b26",
"title": "Statistical optimization for geometric fitting: Theoretical accuracy bound and high order error analysis",
"year": "2008"
},
{
"authors": "Yasushi Kanazawa; Ken-Ichi Kanatani",
"journal": "",
"ref_id": "b27",
"title": "Do we really have to consider covariance matrices for image features?",
"year": "2001"
},
{
"authors": "Alex Kendall; Yarin Gal",
"journal": "Advances in neural information processing systems",
"ref_id": "b28",
"title": "What uncertainties do we need in bayesian deep learning for computer vision",
"year": "2017"
},
{
"authors": "Seong Hun; Lee ; Javier Civera",
"journal": "",
"ref_id": "b29",
"title": "Geometric interpretations of the normalized epipolar error",
"year": "2020"
},
{
"authors": "Vincent Lepetit; Francesc Moreno-Noguer; Pascal Fua",
"journal": "International journal of computer vision",
"ref_id": "b30",
"title": "Epnp: An accurate o (n) solution to the pnp problem",
"year": "2009"
},
{
"authors": "Stefan Leutenegger; Margarita Chli; Roland Y Siegwart",
"journal": "Ieee",
"ref_id": "b31",
"title": "Brisk: Binary robust invariant scalable keypoints",
"year": "2011"
},
{
"authors": "H Christopher; Longuet- Higgins",
"journal": "Nature",
"ref_id": "b32",
"title": "A computer algorithm for reconstructing a scene from two projections",
"year": "1981"
},
{
"authors": " David G Lowe",
"journal": "International journal of computer vision",
"ref_id": "b33",
"title": "Distinctive image features from scaleinvariant keypoints",
"year": "2004"
},
{
"authors": "Zixin Luo; Lei Zhou; Xuyang Bai; Hongkai Chen; Jiahui Zhang; Yao Yao; Shiwei Li; Tian Fang; Long Quan",
"journal": "",
"ref_id": "b34",
"title": "Aslfeat: Learning local features of accurate shape and localization",
"year": "2020"
},
{
"authors": "Jochen Meidow; Christian Beder; Wolfgang Förstner",
"journal": "ISPRS Journal of Photogrammetry and Remote Sensing",
"ref_id": "b35",
"title": "Reasoning with uncertain points, straight lines, and straight line segments in 2d",
"year": "2009"
},
{
"authors": "Krystian Mikolajczyk; Cordelia Schmid",
"journal": "International journal of computer vision",
"ref_id": "b36",
"title": "Scale & affine invariant interest point detectors",
"year": "2004"
},
{
"authors": "Alexander Mordvintsev; Christopher Olah; Mike Tyka",
"journal": "",
"ref_id": "b37",
"title": "Inceptionism: Going deeper into neural networks",
"year": "2015"
},
{
"authors": "Dominik Muhle; Lukas Koestler; Nikolaus Demmel; Florian Bernard; Daniel Cremers",
"journal": "",
"ref_id": "b38",
"title": "The probabilistic normal epipolar constraint for frame-to-frame rotation optimization under uncertain feature positions",
"year": "2022"
},
{
"authors": "Dominik Muhle; Lukas Koestler; Krishna Murthy Jatavallabhula; Daniel Cremers",
"journal": "",
"ref_id": "b39",
"title": "Learning correspondence uncertainty via differentiable nonlinear least squares",
"year": "2023"
},
{
"authors": "Raul Mur; -Artal ; Juan D Tardós",
"journal": "IEEE transactions on robotics",
"ref_id": "b40",
"title": "Orb-slam2: An opensource slam system for monocular, stereo, and rgb-d cameras",
"year": "2017"
},
{
"authors": "Songyou Peng; Peter Sturm",
"journal": "",
"ref_id": "b41",
"title": "Calibration wizard: A guidance system for camera calibration based on modelling geometric and corner uncertainty",
"year": "2019"
},
{
"authors": "Jerome Revaud; Cesar De Souza; Martin Humenberger; Philippe Weinzaepfel",
"journal": "Advances in neural information processing systems",
"ref_id": "b42",
"title": "R2d2: Reliable and repeatable detector and descriptor",
"year": "2019"
},
{
"authors": "Rockafellar Tyrrell",
"journal": "SIAM review",
"ref_id": "b43",
"title": "Lagrange multipliers and optimality",
"year": "1993"
},
{
"authors": "Edward Rosten; Reid Porter; Tom Drummond",
"journal": "IEEE transactions on pattern analysis and machine intelligence",
"ref_id": "b44",
"title": "Faster and better: A machine learning approach to corner detection",
"year": "2008"
},
{
"authors": "Ethan Rublee; Vincent Rabaud; Kurt Konolige; Gary Bradski",
"journal": "Ieee",
"ref_id": "b45",
"title": "Orb: An efficient alternative to sift or surf",
"year": "2011"
},
{
"authors": "Javier Sánchez; Nelson Monzón; Agustín Salgado De; La Nuez",
"journal": "Image Processing On Line",
"ref_id": "b46",
"title": "An analysis and implementation of the harris corner detector",
"year": "2018"
},
{
"authors": "Paul-Edouard Sarlin; Cesar Cadena; Roland Siegwart; Marcin Dymczyk",
"journal": "",
"ref_id": "b47",
"title": "From coarse to fine: Robust hierarchical localization at large scale",
"year": "2019"
},
{
"authors": "L Johannes; Jan-Michael Schonberger; Frahm",
"journal": "",
"ref_id": "b48",
"title": "Structurefrom-motion revisited",
"year": "2016"
},
{
"authors": "Zehong Shen; Jiaming Sun; Yuang Wang; Xingyi He; Hujun Bao; Xiaowei Zhou",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b49",
"title": "Semi-dense feature matching with transformers and its applications in multiple-view geometry",
"year": "2022"
},
{
"authors": "Jianbo Shi",
"journal": "IEEE",
"ref_id": "b50",
"title": "Good features to track",
"year": "1994"
},
{
"authors": "K Simonyan; Zisserman",
"journal": "Computational and Biological Learning Society",
"ref_id": "b51",
"title": "Very deep convolutional networks for large-scale image recognition",
"year": "2015"
},
{
"authors": "Jürgen Sturm; Nikolas Engelhard; Felix Endres; Wolfram Burgard; Daniel Cremers",
"journal": "IEEE",
"ref_id": "b52",
"title": "A benchmark for the evaluation of rgb-d slam systems",
"year": "2012"
},
{
"authors": "Jiaming Sun; Zehong Shen; Yuang Wang; Hujun Bao; Xiaowei Zhou",
"journal": "",
"ref_id": "b53",
"title": "Loftr: Detector-free local feature matching with transformers",
"year": "2021"
},
{
"authors": "Jiaming Sun; Zihao Wang; Siyu Zhang; Xingyi He; Hongcheng Zhao; Guofeng Zhang; Xiaowei Zhou",
"journal": "",
"ref_id": "b54",
"title": "Onepose: One-shot object pose estimation without cad models",
"year": "2022"
},
{
"authors": "Shitao Tang; Jiahui Zhang; Siyu Zhu; Ping Tan",
"journal": "ICLR",
"ref_id": "b55",
"title": "Quadtree attention for vision transformers",
"year": "2022"
},
{
"authors": "Bill Triggs; Richard I Philip F Mclauchlan; Andrew W Hartley; Fitzgibbon",
"journal": "Springer",
"ref_id": "b56",
"title": "Bundle adjustment-a modern synthesis",
"year": "1999"
},
{
"authors": "Michał Tyszkiewicz; Pascal Fua; Eduard Trulls",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b57",
"title": "Disk: Learning local features with policy gradient",
"year": "2020"
},
{
"authors": "Alexander Vakhitov; Luis Ferraz; Antonio Agudo; Francesc Moreno-Noguer",
"journal": "",
"ref_id": "b58",
"title": "Uncertainty-aware camera pose estimation from points and lines",
"year": "2021"
},
{
"authors": "Yannick Verdie; Kwang Yi; Pascal Fua; Vincent Lepetit",
"journal": "",
"ref_id": "b59",
"title": "Tilde: A temporally invariant learned detector",
"year": "2015"
},
{
"authors": "Kuan Xu; Yuefan Hao; Chen Wang; Lihua Xie",
"journal": "",
"ref_id": "b60",
"title": "Airvo: An illumination-robust point-line visual odometry",
"year": "2022"
},
{
"authors": "Kwang Moo; Yi ; Eduard Trulls; Vincent Lepetit; Pascal Fua",
"journal": "Springer",
"ref_id": "b61",
"title": "Lift: Learned invariant feature transform",
"year": "2016"
},
{
"authors": "Bernhard Zeisl; Pierre Fite Georgel; Florian Schweiger; G Eckehard; Nassir Steinbach; G Navab; Munich",
"journal": "",
"ref_id": "b62",
"title": "Estimation of location uncertainty for scale invariant features points",
"year": "2009"
},
{
"authors": "Qunjie Zhou; Sérgio Agostinho; Aljoša Ošep; Laura Leal-Taixé",
"journal": "Springer",
"ref_id": "b63",
"title": "Is geometry enough for matching in visual localization",
"year": "2022"
}
] | [
{
"formula_coordinates": [
3,
130.91,
436.47,
155.46,
10.62
],
"formula_id": "formula_0",
"formula_text": "x i = x i,true + ξ i ,(1)"
},
{
"formula_coordinates": [
3,
61.02,
519.88,
225.34,
13.03
],
"formula_id": "formula_1",
"formula_text": "E{ξ i } = 0 , Σ i := Cov(ξ i ) = E{ξ i ξ ⊤ i } , ∀i . (2)"
},
{
"formula_coordinates": [
3,
333.66,
275.96,
186.92,
11.23
],
"formula_id": "formula_2",
"formula_text": "ϕ([∇ x S, ∇ y S]) = [(∇ x S) 2 , (∇ y S) 2 , ∇ x S∇ y S]"
},
{
"formula_coordinates": [
3,
336.03,
531.58,
209.08,
23.23
],
"formula_id": "formula_3",
"formula_text": "Σ i := 1 S(x i ) I 2×2 = 1/S(x i ) 0 0 1/S(x i ) .(3)"
},
{
"formula_coordinates": [
3,
308.86,
636.1,
236.25,
79.14
],
"formula_id": "formula_4",
"formula_text": "C i ∈ R 2×2 [22]. Defining [∇ x S i , ∇ y S i ] := ∂S/∂x| xi as the spatial gradient of S eval- uated at x i , C i in its local neighborhood W i (window of size u × v) is given by C i := j∈Wi w j ∂S ∂x ⊤ xj ∂S ∂x xj = j∈Wi w j (∇ x S j ) 2 ∇ x S j ∇ y S j ∇ y S j ∇ x S j (∇ y S j ) 2 ,(4)"
},
{
"formula_coordinates": [
5,
90.52,
178.65,
195.85,
22.13
],
"formula_id": "formula_5",
"formula_text": "c i = j∈Wi w j (S(x j ) -S(x j + δx)) 2 ,(5)"
},
{
"formula_coordinates": [
5,
61.33,
244.1,
225.04,
26.8
],
"formula_id": "formula_6",
"formula_text": "c i ≈ j∈Wi w j (S(x j ) -S(x j ) - ∂S ∂x xj δx) 2 ,(6)"
},
{
"formula_coordinates": [
5,
71.72,
275.97,
214.64,
30
],
"formula_id": "formula_7",
"formula_text": "= j∈Wi w j δx ⊤ ∂S ∂x ⊤ xj ∂S ∂x xj δx = δx ⊤ C i δx . (7)"
},
{
"formula_coordinates": [
5,
91.72,
336.01,
194.65,
30.53
],
"formula_id": "formula_8",
"formula_text": "max δx min δx δx ⊤ C i δx, s.t. ∥δx∥ = 1 ,(8)"
},
{
"formula_coordinates": [
5,
84.39,
418.17,
201.97,
11.72
],
"formula_id": "formula_9",
"formula_text": "L(δx, λ) := δx ⊤ C i δx -λ(δx ⊤ δx -1) ,(9)"
},
{
"formula_coordinates": [
5,
84.81,
458.61,
201.55,
11.72
],
"formula_id": "formula_10",
"formula_text": "2δx ⊤ C i -2λδx ⊤ = 0 ⇒ C i δx = λδx ,(10)"
},
{
"formula_coordinates": [
5,
50.11,
514.55,
45.74,
11.87
],
"formula_id": "formula_11",
"formula_text": "Σ i := C -1"
},
{
"formula_coordinates": [
5,
126.48,
610.64,
159.88,
11.72
],
"formula_id": "formula_12",
"formula_text": "N (S(x i + t i ), σ 2 ) ,(11)"
},
{
"formula_coordinates": [
5,
55.09,
674.99,
231.27,
11.72
],
"formula_id": "formula_13",
"formula_text": "S(x i ) = S(x i + t i ) + ε i , with ε i ∼ N (0, σ 2 ) ,(12)"
},
{
"formula_coordinates": [
5,
314.51,
115.84,
226.45,
23.62
],
"formula_id": "formula_14",
"formula_text": "ti = arg max ti j∈Wi 1 σ √ 2π exp - (S(x j ) -S(x j + t i )) 2 2σ 2 . (13"
},
{
"formula_coordinates": [
5,
540.96,
121.94,
4.15,
8.64
],
"formula_id": "formula_15",
"formula_text": ")"
},
{
"formula_coordinates": [
5,
385.48,
236.99,
159.63,
11.03
],
"formula_id": "formula_16",
"formula_text": "Var( ti ) ≥ I( ti ) -1 .(14)"
},
{
"formula_coordinates": [
5,
308.86,
269.66,
236.25,
21.95
],
"formula_id": "formula_17",
"formula_text": "s i := ∂ log L(t i | S)/∂t i , since E{s i } = 0 [21]."
},
{
"formula_coordinates": [
5,
314.74,
300.08,
230.37,
24.15
],
"formula_id": "formula_18",
"formula_text": "s i = xj ∈Wi d j := S(x j ) -S(x j + t i ) σ 2 ∂S(x j + t i ) ∂t i .(15)"
},
{
"formula_coordinates": [
5,
363.96,
344.08,
181.16,
44.08
],
"formula_id": "formula_19",
"formula_text": "E{d ⊤ j d k } = 0, ∀j ̸ = k. Thereby I( ti ) = j∈Wi E{d ⊤ j d j } ti .(16)"
},
{
"formula_coordinates": [
5,
315.97,
437.51,
229.15,
48.12
],
"formula_id": "formula_20",
"formula_text": "I( ti ) = xj ∈Wi 1 σ 4 ∂S(x j ) ∂x ⊤ xj ∂S(x j ) ∂x xj E{(S(x j ) -S(x j + t i )) 2 } ti .(17)"
},
{
"formula_coordinates": [
5,
329.72,
527.25,
215.39,
30.01
],
"formula_id": "formula_21",
"formula_text": "I( ti ) = 1 σ 2 xj ∈Wi ∂S(x j ) ∂x ⊤ xj ∂S(x j ) ∂x xj ,(18)"
},
{
"formula_coordinates": [
6,
375,
438.19,
170.11,
10
],
"formula_id": "formula_22",
"formula_text": "e i,r := cart(x i -H i,r xr ) ,(19)"
},
{
"formula_coordinates": [
6,
367.99,
452.99,
177.12,
11.72
],
"formula_id": "formula_23",
"formula_text": "Σ ei,r := JΣ xr J ⊤ + Σ xi ,(20)"
},
{
"formula_coordinates": [
12,
120.73,
338.9,
165.64,
11.28
],
"formula_id": "formula_24",
"formula_text": "r i := x(1,2) i - x(3) i x i ,(21)"
},
{
"formula_coordinates": [
12,
111.41,
394.35,
174.95,
10
],
"formula_id": "formula_25",
"formula_text": "xi := K(R c,w p w,i + t c,w ) ,(22)"
},
{
"formula_coordinates": [
12,
61.86,
570.07,
220.35,
14.58
],
"formula_id": "formula_26",
"formula_text": "Mc = 0 , with M := M ⊤ 1 • • • M ⊤ n ⊤ . (23"
},
{
"formula_coordinates": [
12,
282.21,
574.4,
4.15,
8.64
],
"formula_id": "formula_27",
"formula_text": ")"
},
{
"formula_coordinates": [
12,
125.3,
639.15,
161.06,
12.95
],
"formula_id": "formula_28",
"formula_text": "M ⊤ Σ -1 r Mc = 0 ,(24)"
},
{
"formula_coordinates": [
12,
356.08,
184.78,
189.03,
9.68
],
"formula_id": "formula_29",
"formula_text": "p w,i = E{p w,i } + ν ,(26)"
},
{
"formula_coordinates": [
12,
357.74,
199.27,
187.37,
11.72
],
"formula_id": "formula_30",
"formula_text": "E{ν} = 0 , Σ pw,i := E{νν ⊤ } .(27)"
},
{
"formula_coordinates": [
12,
354.42,
279.75,
190.69,
9.79
],
"formula_id": "formula_31",
"formula_text": "xi = KRp w,i + Kt = E{x i } + ζ ,(28)"
},
{
"formula_coordinates": [
12,
342.1,
346.72,
203.01,
20.81
],
"formula_id": "formula_32",
"formula_text": "Σ xi := Λ w w ⊤ γ := KRΣ pw,i R ⊤ K ⊤ .(29)"
},
{
"formula_coordinates": [
12,
326.15,
412.72,
218.96,
11.07
],
"formula_id": "formula_33",
"formula_text": "r i = x(1,2) i - x(3) i x i ,(30)"
},
{
"formula_coordinates": [
12,
356.33,
427.7,
73.47,
29.85
],
"formula_id": "formula_34",
"formula_text": "(1,2) i } + ζ (1,2) ) -(E{x(3)"
},
{
"formula_coordinates": [
12,
387.84,
439.23,
157.28,
20.52
],
"formula_id": "formula_35",
"formula_text": "i } + ζ (3) )(E{x i } + ξ) ,(31)"
},
{
"formula_coordinates": [
12,
336.96,
463.66,
208.16,
30.04
],
"formula_id": "formula_36",
"formula_text": "= E{x (1,2) i - x(3) i x i } + ζ (1,2) -E{x (3) i }ξ -ζ (3) E{x i } -ζ (3) ξ ,(32)"
},
{
"formula_coordinates": [
12,
373.39,
527.19,
93.84,
14.07
],
"formula_id": "formula_37",
"formula_text": "E{r i } = E{x (1,2) i - x(3"
}
] |
|
2023-05-20 | [
{
"figure_ref": [
"fig_0",
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b0",
"b1",
"b2",
"b3",
"b4",
"b5",
"b6",
"b7",
"b8",
"b9",
"b13",
"b14",
"b15",
"b5",
"b6",
"b7",
"b8",
"b16",
"b17",
"b18",
"b19",
"b20",
"b21",
"b22",
"b23",
"b24",
"b1",
"b11",
"b25",
"b15",
"b26",
"b27",
"b28",
"b29",
"b12",
"b4",
"b29",
"b30",
"b28",
"b31",
"b32",
"b27",
"b33",
"b34",
"b35",
"b13",
"b14",
"b15",
"b36",
"b37",
"b38",
"b39",
"b40",
"b41",
"b7",
"b5",
"b8",
"b42",
"b28",
"b43",
"b43",
"b12",
"b3",
"b44",
"b26",
"b30",
"b45",
"b5",
"b46",
"b47",
"b20",
"b19",
"b18",
"b20",
"b24",
"b48",
"b49",
"b50",
"b22",
"b23"
],
"table_ref": [],
"text": "Human-Object Interaction Detection (HOI), a core element in human-centric vision perception tasks such as human activity recognition [1], motion tracking [2], and anomaly behavior detection [3], has attracted considerable attention over the past decades. HOI primarily involves localizing correlated human-object 2D positions within an image and identifying their interactions. Despite numerous models proposed in recent years [4][5][6][7][8][9], practical implementation remains challenging due to the inherent ambiguity of verbs and subtle distinctions among interaction categories.\nWe investigate the problems of the current methods from both data and model aspects. From the data perspective, current datasets characterize an HOI instance as a 〈human, object, interaction〉 triplet, and complete with comprehensive annotations, i.e., human and object bounding boxes and respective interaction types. We carefully analyze the widely-used HICO-DET dataset [10], which encompasses 38, 118 training images with 600 distinct types of real-world HOI triplets, and have identified its several limitations that hinder the effective learning of interaction representations: (1) Class imbalance. As demonstrated in Fig. 1-(a), we observe it has a significant long-tail distribution in the HOI triplet types. This imbalanced label distribution is not conducive to learning, particularly when identifying interaction categories with only subtle variances. (2) Small data size. Upon further investigation, we discover that among its 600 interaction categories, 51 categories have only a single image available, and 435 categories have less than 100 images. Such limited data severely impacts the learning of these rare categories. (3) Limited diversity. Compared with existing datasets, real-world scenarios typically exhibit more diversity due to the complex variation in human appearance attributes, environmental backgrounds, and shooting perspectives. This diversity causes a considerable decline in the performance of existing methods when the scene content changes.\nFrom the model standpoint, the majority of existing HOI detectors, regardless of whether they employ two-stage [11-13, 4, 5] or one-stage [14][15][16][6][7][8][9] strategies, are built upon pre-trained object detectors to enhance the initial localization of humans and objects. However, it remains challenging to carry out effective interaction prediction solely based on representations of humans and objects, sometimes with their spatial relationship. Specifically, verb concepts and their nuanced contextual information, including human posture, orientation, attention area and overall circumstances, can critically influence prediction. Extracting this semantic information from data is typically inefficient and cumbersome, significantly limiting existing methods' performance.\nRecently, text-to-image diffusion models [17][18][19][20][21] trained on massive internet-scale data have achieved significant performance in conditioned image generation, which provides high-quality, versatile, and semantically controllable image generation. Such an advantage offers us an effective means to generate rich and diverse realistic HOI images. In practice, these models utilize cross-attention mechanisms between text embeddings and visual representations, signifying a substantial correlation between their feature spaces and semantic concepts in language. Inspired by this, we use DAAM [22] to generate pixel-level linguistic concept heatmaps based on the image features of this diffusion model. As depicted in Fig. 1-(b), beyond the noun concepts highlighted in prior studies [23][24][25], we find that the internal representation space of a frozen text-to-image diffusion model is highly relevant to verb concepts and their associated contexts. However, the challenge of extracting these verb-associated representations from the diffusion model for downstream HOI task still exist.\nTo address the aforementioned issues, we introduce DiffHOI, a novel HOI detection scheme based on text-to-image diffusion models (e.g., Stable Diffusion). For the first time, DiffHOI tries to leverage the generative and representative capabilities to benefit the HOI task, i.e., the extracted powerful verb-associated contextual representations of SD. DiffHOI consists of three components: the pretrained human-object detector, interaction decoder, and object-interaction classifiers. We introduce an adapter-style tuning approach to align global and local semantic associated representations from the SD and CLIP model in the interaction decoder. These representations serve a critical role in comprehending the nuanced disparities in interactions and reducing interaction prediction ambiguity.\nTo fill in the shortcomings of existing long-tail HOI datasets, we present SynHOI, a class-balance, large-scale, and high-diversity synthetic HOI dataset with over 140K fully annotated HOI images, which can effectively facilitate learning interaction representations. To make the flow of dataset production scalable, we present an automatic pipeline, including the HOIPrompt design, automatic labeling and filtering, and quality verification, designed to continually scale up the generation of diverse and high-precision HOI-annotated data. Therefore, our contributions can be summarized as follows: (1) We introduce a novel scheme, DiffHOI, which leverages both the generative and representation capacities of pre-trained text-to-image diffusion models to enhance the performance of HOI detection tasks. (2) We present an automatic and scalable pipeline to generate realistic annotated HOI images characterized by varied attributes and scene contexts and propose a class-balance, large-scale, and high-diversity synthetic HOI dataset, namely SynHOI. (3) Extensive experimental results demonstrate the proposed adapter-style tuning, together with the proposed SynHOI dataset, significantly improves the performance of HOI tasks under the regular and zero-shot settings and achieves the new state-of-the-art, i.e., 41.50 mAP on HICO-DET.\n2 Related Work HOI Detection. HOI detection task primarily encompasses three sub-problems, i.e., object detection, human-object pairing, and interaction recognition. Previous HOI detectors can generally be divided into one-stage and two-stage paradigms. The two-stage strategy employs an off-the-shelf detector to determine the locations and classes of objects, followed by specially-designed modules for humanobject association and interaction recognition. Most methods are dedicated to exploring additional feature streams to improve interaction classification, such as the appearance stream [12,26,16,27], spatial stream [28][29][30], pose and body-parts stream [13,5,30], semantic stream [31,29,32], and graph network [33,28,[34][35][36]. Instead, the one-stage strategy detects HOI triplets in a single forward pass by assigning human and object proposals to predefined anchors and then estimating potential interactions [14][15][16]37]. Recently, the DETR-based HOI detectors [38][39][40][41][42] have gained prominence in this paradigm, and they formulate the HOI detection task as a set prediction problem, avoiding complex post-processing. In particular, many methods [8,6,9,43] demonstrate promising performance improvements by disentangling human-object detection and interaction classification as two decoders in a cascade manner. Our work builds on the top of the transformer-based HOI detection strategy and focuses on enhancing the learning of a dedicated interaction decoder.\nZero-shot HOI Detection. Zero-shot HOI detection has emerged as a field aiming to identify unseen HOI triplet categories not present in the training data. Previous research [29,44,44,13,4,45,27,31] has addressed this task in a compositional manner, by disentangling the reasoning process on actions and objects during training. This approach enables the recognition of unseen HOI triplets during inference. With the advancements in Vision-Language Models, such as CLIP [46], recent research [6,47,48] has shifted focus toward transferring knowledge from CLIP to recognize unseen HOI concepts. This shift has resulted in a notable performance improvement in zero-shot settings. In this work, we aim to further explore the potential benefits of the Text-to-Image Diffusion model [21] in enhancing zero-shot HOI detection.\nDiffusion Models for Image Generation and Vision Perception. Text-to-image diffusion models, such as DALLE-2 [20], Imagen [19], and Stable Diffusion [21], have shown considerable potentials in generating photorealistic images from free-form text prompts. This capability is attributed to the strong semantic correspondence between visual and language elements learned from a vast corpus of image-caption pairs. Recent research has proposed utilizing diffusion models to augment real datasets, assisting the training of downstream tasks [25,[49][50][51]. Our work concentrates on generating synthetic HOI data to enhance HOI detection performance. Moreover, text-to-image diffusion models pretrained on large-scale image-text pairs offer a high degree of control through customizable prompts. This aspect suits them for downstream tasks [23,24], where the additional feature representations from pretraining can bolster performance. This work first explores using a frozen text-image diffusion model for HOI detection."
},
{
"figure_ref": [],
"heading": "SynHOI-A Synthetic HOI Dataset",
"publication_ref": [],
"table_ref": [],
"text": "This section introduces how the proposed high-quality synthetic HOI dataset SynHOI is built via an automatic and scalable pipeline. \"urban\" → \"outdoor\" \"partial view\" → \"front view\" Generated annotations Generated annotations \"white woman\" → \"asian man\" \"outdoor\" → \"secluded\" \"oblique view\" → \"back view\" <human, motorcycle, ride & sit on> "
},
{
"figure_ref": [
"fig_5",
"fig_5",
"fig_5"
],
"heading": "Construction Process",
"publication_ref": [
"b9",
"b51",
"b52",
"b53"
],
"table_ref": [],
"text": "The HOIPrompt design. To address limitations in current datasets , we propose pre-defined HOIPrompts, as illustrated in Fig. 2 (a). Complete HOI triplets are formed by sampling verb and noun combinations from the HICO-DET dataset [10], creating \"a {race} {age & gender} verb-ing an object.\" To describe a person's appearance, we use the format \"a {race} {age & gender},\" randomly selecting elements from the HOIPrompts. We generate images with combinations likely to occur together by analyzing co-occurring HOI triplets. For interaction environments, we select an adjective from a range of options to describe the atmosphere (denoted as \"{environment}\"). Photographic information is represented by four components (\"{quality},\" \"{lighting},\" \"{view},\" and \"{camera}\"), aligning the synthetic images with real HOI data and providing camera angle diversity. We further enhance diversity and quality through negative prompts and random model configurations. Overall, we generate 259, 806 images using HOIPrompts at this stage.\nAutomatic Labeling and Filtering. We design a three-step process to automatically annotate and filter the synthetic images. Firstly, we utilize a state-of-the-art detection model [52] to detect objects within each image. Secondly, we discard any images in which the confidence score of the detected object specified in the corresponding HOI triplet(s) prompt is below a threshold of 0.5. Thirdly, we associate humans with objects in the images and assign the appropriate HOI category from the prompts to the human-object combination(s). In practice, we assign the HOI category to the person closest to the center of the object's bounding box. If not all humans in the image have an HOI label, we select the object closest to the corresponding human's bounding box. Upon completing the automated labeling and filtering process, we obtain a new synthetic dataset, namely SynHOI, including 146, 772 annotated images with complete HOI labels associated with the detected human-object interactions.\nVisualization and Manual Verification. We develop a visualization tool to facilitate the manual inspection and filtering of any incorrect HOI annotations. We incorporate manual efforts to sample and inspect the annotated results to ensure their quality. During this inspection, we observe that the SOTA detector trained on COCO [53] performs well in detecting humans and objects, indicating that the data distribution in SynHOI closely resembles that of natural images in COCO. However, due to the inherent ambiguity of verbs, a small number of synthetic interactions in SynHOI may be incompletely accurate, which also exists in the HICO-DET dataset. To address this issue, we construct a subset comprising 5% data of SynHOI (over 8K), namely SynHOI-Sub, during the sampling and inspection process. This subset has undergone meticulous manual examination, resulting in annotations that are verified to be completely accurate. High-diversity. SynHOI exhibits a high level of diversity, offering a wide range of visually distinct images. Fig. 2-(b) demonstrates the impact of random variations in people's descriptions, environments, and photographic information within the HOIPrompts on the diversity of synthetic images.\nHigh-quality. SynHOI showcases high-quality HOI annotations. First, we employ CLIPScore [54] to measure the similarity between the synthetic images and the corresponding HOI triplet prompts. The SynHOI dataset achieves a high CLIPScore of 0.805, indicating a faithful reflection of the HOI triplet information in the synthetic images. Second, Fig. 2-(b) provides evidence of the high quality of detection annotations in SynHOI, attributed to the effectiveness of the SOTA detector and the alignment of SynHOI with real-world data distributions. As mentioned earlier, we also release a carefully verified subset of SynHOI, called SynHOI-Sub."
},
{
"figure_ref": [],
"heading": "Methodology",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [
"fig_4"
],
"heading": "Framework Overview",
"publication_ref": [
"b54",
"b20",
"b45"
],
"table_ref": [],
"text": "The proposed DiffHOI framework is presented in Figure 3, consisting of three components. Primarily, we employ a transformer-based object detector D obj [55] to extract local feature representations of humans and objects. Using these extracted features, initial representations of interactions are computed and fed into the human-object interaction decoder, referred to as D int , which subsequently updates these interaction representations. Additionally, we employ the frozen Stable Diffusion (SD) Model F sd [21] to generate semantically associated image feature maps, ensuring that distinct regions of these feature maps respond to various types of semantic information (see Sec.4.2). Furthermore, we introduce the CLIP model [46] and utilize its frozen image encoder F img clip to extract comprehensive semantic representations that align with text descriptors. Two trainable scene-aware adaptors are introduced to adapt these semantic representations to suit the SD Model and interaction classifier (see Sec. 4.3). In practice, both the object classifier and interaction classifier are generated using the CLIP text encoder F text clip , which is applied during both the training and inference phases. Pair-wise Human-Object Localization. Given an input image x, we first utilize the image encoder to extract feature maps V , which is then adopted to localize humans and objects by using humanobject decoder D obj . We further input pair-wise human-object queries Q {h,o} ∈ R 2×N ×C into the D obj and update them to Q {h,o} ∈ R 2×N ×C , where N is the number of paired queries and C indicates the channel dimension. We use Q h,o to predict the human bounding box B h and object bounding box B o , whose dimension are N × 4. To predict the class label of each object, we further employ the text encoder F text clip of CLIP to extract the object-aware embedding T o clip ∈ R K1×C . We further leverage T o clip to conduct the dot-product with Q o ∈ R N ×C to predict the final object category distributions P o ∈ R N ×K1 , where K 1 denotes the total number of object classes.\nHuman-Object Interaction Recognition. We design a novel interaction decoder D int to perform better interaction understanding, where we exploit three types of visual representations as the input, i.e., the V from object detector, the V sd from Stable Diffusion Model F sd , and the v clip from the image encoder F img clip of CLIP. Specifically, we first perform average pooling on the human-object queries Q h,o , resulting in interaction queries Qi ∈ R N ×C . Then, we feed Q i into the interaction decoder D int to apply self-attention and subsequent cross-attention with the sum of V and V sd . This process updates the interaction queries from Q i to Q i ∈ R N ×C . Moreover, we leverage v clip ∈ R C to enhance each of the interaction query in Q i , performing element-wise addition. To predict the interaction label of each human-object pair, we utilize the text encoder F text clip of CLIP to extract the interaction-aware embedding T i clip ∈ R K2×C . Finally, we employ dot-product between T i clip and Q i to predict the human-object interaction category distributions P i ∈ R N ×K2 , where K 2 is the number of interaction categories."
},
{
"figure_ref": [
"fig_0",
"fig_4",
"fig_4"
],
"heading": "Local Semantic Association via Stable Diffusion",
"publication_ref": [
"b20",
"b5",
"b46",
"b47"
],
"table_ref": [],
"text": "The output features V from the object detector only retain human and object-oriented semantic information. In such a case, the interactional relationship and its corresponding context remain limited, resulting in sub-optimal interaction predictions. Recenlty, Stable Diffusion [21] (SD) has been primarily designed to produce high-quality images utilizing textual descriptors. For the HOI task, we find that it be used to encourage diverse semantic information embedded in the textual representations to be associated with specific local regions of the images, as shown in Fig. 1. Inspired by the above observations, the feature maps derived from the SD model are expected to exhibit superior semantic correspondence within local regions, establishing an association between the semantic information of the text and the corresponding regions within the image. Accordingly, we utilize the UNet-based diffusion model to extract the feature maps, which are not only associated with the noun concepts, e.g., the human and objects, but also include the verb concepts and the corresponding contextual details. Different from using a series of denoising steps to generate an image, we directly feed the input image x into the UNet (denoted as F sd ) and perform a single forward pass via the network. The output multi-scale image features are as V sd = F sd (x, A x ), where A x denotes the text-related representation that corresponds to x. Typically, A x can be obtained by utilizing a text encoder, such as the CLIP text encoder, to encode the description of the image. However, as a discriminative task, the description of the HOI image is not available in the inference. To address this, we replace it with the global feature representation of x extracted from the CLIP image encoder F img clip , which bears significant relevance to the textual representation. Please refer to Sec. 4.3 for more details. As illustrated in the blocks in the red background of Fig. 3, we proceed to combine the features V and V sd to augment the representation ability of the verb concepts within V and use it to generate more informative Q i for the final interaction prediction. Two Scene-aware Adaptors. Given that v clip represents a global feature, it inherently encompasses the contextual information of the overall scene. As shown in Fig. 3, we introduce two scene-aware adaptors, denoted as α and β, to project v clip into feature spaces more consistent with the SD model and interaction predictor. Regarding the scene-aware adaptor α, since the CLIP model is trained to align global visual and textual representations within a latent space, the v clip can be employed as a substitute for the textual representation. Hence, we can train an MLP to directly project v clip to a textual space suitable for the SD model. As for the scene-aware adaptor β, we project v clip through it and incorporate the adapted version into each interaction query of Q i . This adjustment allows for the tuning of these interaction queries to align more effectively with the CLIP-based classifiers T i clip . CLIP-based Classifiers. Inspired by [6,47,48], we use the CLIP text encoder F text clip to generate both the object classifier and interaction classifier. Specifically, we initiate the process by transforming each object category or HOI category into a sentence using the hand-crafted template, i.e. \"A photo of a [Object]\" or \"A photo of a person [Verb-ing] a [Object]\". Then these sentences can be encoded to obtain the object category embeddings T o clip or interaction category embeddings T i clip . Formally, the final object category distributions P o and HOI category distributions P i can be calculated as,\nP o = softmax(Q o * T o clip )(1)\nP i = softmax(Q i * T i clip )(2)\nwhere\nQ o ∈ R N ×C and Q o ∈ R N ×C\ndenotes object queries and interaction queries, respectively, and softmax indicates the row-wise softmax operation."
},
{
"figure_ref": [],
"heading": "Loss Function",
"publication_ref": [
"b5",
"b7"
],
"table_ref": [],
"text": "Following the query-based methods [6,8], we employ the Hungarian algorithm to match predictions to each ground-truth. The overall loss is computed between the matched predictions and their corresponding ground-truths, which includes the box regression loss L b , the intersection-over-union loss L g , the object class loss L o c , and the interaction category loss L i c ,\nL = λ b L b + λ g L g + λ o c L o c + λ i c L i c ,(3)\nwhere L b and L g contain both human and object localization. λ b , λ g , λ o c and λ i c are used to adjust the weights of each loss component."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Experimental Settings",
"publication_ref": [
"b10",
"b55",
"b7",
"b5",
"b5",
"b54",
"b56",
"b51",
"b57",
"b5",
"b65",
"b22",
"b20"
],
"table_ref": [],
"text": "Due to the page limit, we leave the detailed experimental settings in the Appendix.\nDatasets and Evaluation Metrics. We evaluate our models on two widely used datasets: HICO-DET [11] and V-COCO [56]. The mean Average Precision (mAP) is used as the evaluation metric, following standard protocols [8,6].\nZero-Shot Setting. We conduct zero-shot experiments on HICO-Det, following the settings in [6]: Rare First UC (RF-UC), Non-rare First UC (NF-UC), Unseen Object (UO) and Unseen Verb (UV).\nImplementation Details. We implement two variant architectures of DiffHOI: DiffHOI-S, and DiffHOI-L, where 'S' and 'L' refer to small and large, respectively. For DiffHOI-S, we use ResNet-50 as the backbone and a six-layer vanilla Transformer encoder [55] as the feature extractor. Both the human-object decoder and interaction decoder are three-layer vanilla Transformer decoders. For DiffHOI-L, we employ Swin-L [57] as the backbone. In this variant, we replace all the transformer layers with deformable Transformer layers [52,58]. We fine-tune the CLIP-based interaction and object classifiers for regular settings with a small learning rate of 10 -5 while freezing them for zero-shot settings. The diffusion model is frozen during training for all the settings. Two Key Representations. As described in Sec. 4.1, we introduce V sd derived from stable diffusion F sd and v clip obtained from the image encoder F img clip of CLIP, into the interaction decoder. Tab. 6 demonstrates their effectiveness. Initially, we establish a baseline model following the design of GEN-VLKT [6] while excluding the knowledge distillation component. Subsequently, the incorporation of V sd leads to significant improvements, particularly in terms of rare mAP. This suggests that the internal features in stable diffusion play a crucial role in enhancing the representation of rare categories in the HICO-DET dataset. Additionally, the integration of v clip further aids in improving the non-rare AP, as it aligns the output queries with the interaction classifier. Diffusion Time Steps. We investigate the effectiveness of different diffusion steps in extracting interaction-aware features V sd , similarly to [66,23]. Diffusion models control the noise distortion added to the input image by varying the value of t. Stable diffusion [21] uses a total of 1000 time steps. We set t values to 0, 100, 500 for ablation studies. As demonstrated in Tab. 7, the best performance is achieved when t = 0. It is worth noting that using V sd extracted from input images with higher noise levels would decrease performance and potentially impact the learning of interactions."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "We investigate the issues of the current HOI detection methods and present a novel scheme, namely DiffHOI, which leverages both the generative and representation capacities of a pre-trained text-toimage diffusion model to benefit the HOI detection task. Particularly, we release a class-balance, large-scale, and high-diversity synthetic HOI dataset called SynHOI to address the long-tail issue in previous datasets and develop an automatic and scalable pipeline to scale up the generation of diverse and high-precision HOI-annotated data. Extensive experimental results demonstrate that our method significantly outperforms the prior state-of-the-art in regular and zero-shot detection tasks. We hope this work could inspire further research in related fields."
}
] | This paper investigates the problem of the current HOI detection methods and introduces DiffHOI, a novel HOI detection scheme grounded on a pre-trained text-image diffusion model, which enhances the detector's performance via improved data diversity and HOI representation. We demonstrate that the internal representation space of a frozen text-to-image diffusion model is highly relevant to verb concepts and their corresponding context. Accordingly, we propose an adapterstyle tuning method to extract the various semantic associated representation from a frozen diffusion model and CLIP model to enhance the human and object representations from the pre-trained detector, further reducing the ambiguity in interaction prediction. Moreover, to fill in the gaps of HOI datasets, we propose SynHOI, a class-balance, large-scale, and high-diversity synthetic dataset containing over 140K HOI images with fully triplet annotations. It is built using an automatic and scalable pipeline designed to scale up the generation of diverse and high-precision HOI-annotated data. SynHOI could effectively relieve the long-tail issue in existing datasets and facilitate learning interaction representations. Extensive experiments demonstrate that DiffHOI significantly outperforms the state-of-the-art in regular detection (i.e., 41.50 mAP) and zero-shot detection. Furthermore, SynHOI can improve the performance of model-agnostic and backbone-agnostic HOI detection, particularly exhibiting outstanding an 11.55% mAP improvement in rare classes. | Boosting Human-Object Interaction Detection with Text-to-Image Diffusion Model | [
{
"figure_caption": "Figure 1 :1Figure 1: We show a) the long-tail distribution issue in HICO-DET and b) the high correlation between HOI text (i.e., nouns and verbs) and internal image features within the frozen stable diffusion.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "(a) Examples of HOIPrompts <human, sandwich, make & cook> \"black young woman\" → \"latino teen\"",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Examples of generated images and annotations",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: Illustration of a) HOIPrompts and b) how HOIPrompts guide the text-to-image generation process to enhance diversity. 3.1 Construction Process",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: Overview of DiffHOI, comprising a pretrained human-object decoder, a novel interaction decoder, and CLIP-based object and interaction classifiers.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "3. 22Data Characteristics Large-scale. SynHOI consists of 146, 772 images, 157, 358 person bounding boxes, 165, 423 object bounding boxes, and 282, 140 HOI triplet instances. It provides approximately four times the amount of training data compared to HICO-DET. Class-balance. SynHOI can effectively address the long-tail issue in previous datasets, where 343 HOI categories have fewer than 50 images in HICO-DET. Combining SynHOI with HICO-DET reduces the number of HOI categories with fewer than 50 images to only three (refer to Fig. 1-(a)).",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "4. 33CLIP Representation and Scene-aware Adaptor CLIP Preliminaries. The CLIP model is trained to align visual and textual representations. It comprises an image encoder F img clip and a text encoder F text clip , each followed by a linear layer, projecting image and text representations into a shared latent space. Given an image x, F img clip extracts a global visual representation v clip ∈ R C , while F text clip extracts a global text representation T clip ∈ R K×C for all K predefined categories. The prediction distribution over K categories is then calculated as S = softmax(T clip * v clip ), where * denotes the matrix-vector multiplication.",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Performance comparison on HICO-DET in terms of mAP. † means pre-training using our SynHOI dataset. The underlined highlights the compared results.5.2 Comparison to State-of-the-Art for Regular HOI DetectionTab. 1 and Tab. 2 present the performance comparison between DiffHOI and other state-of-the-art methods, which are grouped into one-stage and two-stage methods. For HICO-DET, DiffHOI-L outperforms all existing one-stage and two-stage methods by a significant margin in all evaluation settings. Notably, it achieves a new state-of-the-art performance of 40.84 mAP in the Default Full setting. Using the same backbone Swin-L, DiffHOI-L demonstrates a performance improvement of 3.45 mAP compared to the current one-stage state-of-the-art method FGAHOI[42]. Furthermore, by incorporating our SynHOI dataset, the performance is further boosted to 41.50 mAP. For V-COCO, DiffHOI-L also surpasses the previous state-of-the-art methods with role AP scores of 65.7 on Scenario 1 and 68.2 on Scenario 2.",
"figure_data": "DefaultKnown ObjectMethodBackboneFullRare Non-RareFullRare Non-RareTwo-stage methodsSTIP [59]ResNet-5032.2228.1533.4335.2931.4336.45DEFR [60]ViT-B/1632.3533.4532.02---UPT [61]ResNet-10132.6228.6233.8136.0831.4137.47ViPLO [36]ViT-B/3234.9533.8335.2838.1536.7738.56ViPLO [36]ViT-B/1637.2235.4537.7540.6138.8241.15One-stage methodsQPIC [39]ResNet-10129.9023.9231.6932.3826.0634.27MSTR [62]ResNet-5031.1725.3132.9234.0228.8335.572SSRT [63]ResNet-10131.3424.3133.32---CDN [8]ResNet-10132.0727.1933.5334.7929.4836.38DOQ [64]ResNet-5033.2829.1934.50---IF [65]ResNet-5033.5130.3034.4636.2833.1637.21GEN-VLKT [6] ResNet-5033.7529.2535.1036.7832.7537.99QAHOI [41]Swin-T28.4722.4430.2730.9924.8332.84QAHOI [41]Swin-L35.7829.8037.5637.5931.6639.36FGAHOI [42]Swin-T29.9422.2432.2432.4824.1634.97FGAHOI [42]Swin-L37.1830.7139.1138.9331.9341.02DiffHOI-SResNet-5034.41↑ 1.96% 31.0735.4037.31↑ 1.44% 34.5638.14DiffHOI-LSwin-L40.63↑ 9.28% 38.1041.3843.14↑ 10.81% 40.2444.01DiffHOI-L †Swin-L41.50↑ 11.61% 39.9641.9643.62↑ 12.05% 41.4144.28MethodAP (Scenario 1) AP (Scenario 2)Two-stage methodsSCG [35]54.260.9UPT [61]61.367.1ViPLO [36]60.966.6One-stage methodsQPIC [39]58.360.7MSTR [62]62.065.2CDN [8]63.965.9IF [65]63.065.2GEN-VLKT [6]62.464.5ParMap [7]63.065.1DiffHOI-S61.163.5DiffHOI-L65.768.2",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Performance comparison on V-COCO.",
"figure_data": "MethodTypeUnseen SeenFullGEN-VLKT[6] RF-UC21.3632.91 30.56DiffHOI-SRF-UC24.1332.93 31.08DiffHOI-LRF-UC28.7638.01 36.16GEN-VLKT[6] NF-UC25.0523.38 23.71DiffHOI-SNF-UC26.5725.55 25.75DiffHOI-LNF-UC29.4531.68 31.24GEN-VLKT[6]UO10.5128.92 25.63DiffHOI-SUO9.4229.79 26.22DiffHOI-LUO5.7535.08 30.11GEN-VLKT[6]UV20.9630.23 28.74DiffHOI-SUV23.1030.91 29.72DiffHOI-LUV24.2036.81 35.04",
"figure_id": "tab_1",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Zero-shot performance comparison on HICO-DET.5.3 Comparison to State-of-the-Art for Zero-Shot HOI DetectionTo comprehensively assess the zero-shot capability of DiffHOI, we conduct experiments on various zero-shot settings, as shown in Tab. 3. It demonstrates that DiffHOI-L outperforms the state-of-theart method-GEN-VLKT[6] in all zero-shot settings. Notably, DiffHOI-S with the same backbone achieves remarkable improvements, with a +2.04 mAP gain under the NF-UC setting for all categories and a +2.14 mAP improvement for rare categories under the UO setting.5.4 Investigating the Usefulness of the SynHOI DatasetAs outlined in Sec. 3, we develop the SynHOI dataset, consisting of over 140K synthetic images that are generated, filtered, and annotated automatically. Additionally, we present a subset of SynHOI,",
"figure_data": "MethodBackboneFullRareNon-RareGEN-VLKT ‡ ResNet5033.0429.1034.21GEN-VLKT [4]ResNet50 34.43↑ 4.21% 32.46↑ 11.55% 35.02↑ 2.37%DiffHOI-LSwin-L40.6338.1041.38DiffHOI-LSwin-L41.50↑ 2.14% 39.96↑ 4.88%6 41.96↑ 1.40%",
"figure_id": "tab_2",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Effectiveness of pre-training using Syn-HOI and fine-tuning on HICO-DET. ‡ indicates that our experiments reproduce the results.",
"figure_data": "MethodBackboneFullRareNon-RareGEN-VLKT ‡ [4] ResNet5033.0429.1034.21GEN-VLKT [4]ResNet50 33.73↑ 2.09% 30.08↑ 3.37% 34.82↑ 1.78%DiffHOI-LSwin-L40.6338.1041.38DiffHOI-LSwin-L41.42↑ 1.94% 39.94↑ 4.83% 41.87↑ 1.18%",
"figure_id": "tab_3",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Effectiveness of joint-training using SynHOI-sub and HICO-DET, evaluated on HICO-DET. named as SynHOI-Sub, which includes over 8K data that has undergone manual inspection to ensure the quality of annotations. Here, we explore two strategies to leverage SynHOI and SynHOI-Sub to benefit the HOI detection task.Pre-training. An intuitive strategy is to use SynHOI to pre-train an HOI detector and fine-tune the model on the target dataset. As shown in Tab. 4, SynHOI could improve the performance of model-agnostic and backbone-agnostic HOI detection. Notably, there is a significant improvement in the rare AP when evaluating 138 HOI categories with fewer than 10 training instances in HICO-DET. Specifically, the rare AP increases significantly by 3.36 mAP for GEN-VLKT and 1.86 mAP for DiffHOI. It demonstrates that SynHOI can effectively relieve the long-tail issue and boost the interaction modeling in HICO-DET.Joint-training with HICO-DET. SynHOI-Sub contains more diverse and class-balance images with accurate annotations than HICO-DET. We employ it in conjunction with HICO-DET to perform joint training of the models. As shown in Tab. 5, the inclusion of SynHOI-Sub leads to an improvement of 0.69 mAP for GEN-VLKT and 0.58 mAP for DiffHOI, which can also contribute to improving the rare mAP and enhancing the learning of interactions within the rare categories.",
"figure_data": "5.5 Ablation StudyV sd v clipFullRare Non-rare31.99 29.6332.7032.92 31.2933.4134.41 31.0735.40",
"figure_id": "tab_4",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "Ablation results of two key representations.",
"figure_data": "Time stepFullRare Non-Rare034.41 31.0735.4010034.03 30.5835.0250033.59 29.8034.71",
"figure_id": "tab_5",
"figure_label": "6",
"figure_type": "table"
},
{
"figure_caption": "Ablation results of different diffusion time steps.",
"figure_data": "",
"figure_id": "tab_6",
"figure_label": "7",
"figure_type": "table"
}
] | Jie Yang; Bingliang Li; Fengyu Yang; Ailing Zeng; ‡ Lei Zhang; Ruimao Zhang | [
{
"authors": "Fabian Caba Heilbron; Victor Escorcia; Bernard Ghanem; Juan Niebles",
"journal": "IEEE",
"ref_id": "b0",
"title": "Activitynet: A large-scale video benchmark for human activity understanding",
"year": "2015"
},
{
"authors": "Xinyu Yi; Yuxiao Zhou; Marc Habermann; Soshi Shimada; Vladislav Golyanik; Christian Theobalt; Feng Xu",
"journal": "",
"ref_id": "b1",
"title": "Physical inertial poser (pip): Physics-aware real-time human motion tracking from sparse inertial sensors",
"year": "2022"
},
{
"authors": "Wen Liu; Weixin Luo; Dongze Lian; Shenghua Gao",
"journal": "",
"ref_id": "b2",
"title": "Future frame prediction for anomaly detection-a new baseline",
"year": "2018"
},
{
"authors": "Zhi Hou; Xiaojiang Peng; Yu Qiao; Dacheng Tao",
"journal": "Springer",
"ref_id": "b3",
"title": "Visual compositional learning for humanobject interaction detection",
"year": "2020"
},
{
"authors": "Bo Wan; Desen Zhou; Yongfei Liu; Rongjie Li; Xuming He",
"journal": "",
"ref_id": "b4",
"title": "Pose-aware multi-level feature network for human object interaction detection",
"year": "2019"
},
{
"authors": "Yue Liao; Aixi Zhang; Miao Lu; Yongliang Wang; Xiaobo Li; Si Liu",
"journal": "",
"ref_id": "b5",
"title": "Gen-vlkt: Simplify association and enhance interaction understanding for hoi detection",
"year": "2022"
},
{
"authors": "Xiaoqian Wu; Yong-Lu Li; Xinpeng Liu; Junyi Zhang; Yuzhe Wu; Cewu Lu",
"journal": "Springer",
"ref_id": "b6",
"title": "Mining cross-person cues for body-part interactiveness learning in hoi detection",
"year": "2022"
},
{
"authors": "Aixi Zhang; Yue Liao; Si Liu; Miao Lu; Yongliang Wang; Chen Gao; Xiaobo Li",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b7",
"title": "Mining the benefits of two-stage and one-stage hoi detection",
"year": "2021"
},
{
"authors": "Desen Zhou; Zhichao Liu; Jian Wang; Leshan Wang; Tao Hu; Errui Ding; Jingdong Wang",
"journal": "",
"ref_id": "b8",
"title": "Human-object interaction detection via disentangled transformer",
"year": "2022"
},
{
"authors": "Yu-Wei Chao; Yunfan Liu; Xieyang Liu; Huayi Zeng; Jia Deng",
"journal": "",
"ref_id": "b9",
"title": "Learning to detect humanobject interactions",
"year": "2018"
},
{
"authors": "Yu-Wei Chao; Yunfan Liu; Xieyang Liu; Huayi Zeng; Jia Deng",
"journal": "IEEE",
"ref_id": "b10",
"title": "Learning to detect humanobject interactions",
"year": "2018"
},
{
"authors": "Chen Gao; Yuliang Zou; Jia-Bin Huang",
"journal": "",
"ref_id": "b11",
"title": "ican: Instance-centric attention network for human-object interaction detection",
"year": "2018"
},
{
"authors": "Tanmay Gupta; Alexander Schwing; Derek Hoiem",
"journal": "",
"ref_id": "b12",
"title": "No-frills human-object interaction detection: Factorization, layout encodings, and training techniques",
"year": "2019"
},
{
"authors": "Yue Liao; Si Liu; Fei Wang; Yanjie Chen; Chen Qian; Jiashi Feng",
"journal": "",
"ref_id": "b13",
"title": "Ppdm: Parallel point detection and matching for real-time human-object interaction detection",
"year": "2020"
},
{
"authors": "Tiancai Wang; Tong Yang; Martin Danelljan; Fahad Shahbaz Khan; Xiangyu Zhang; Jian Sun",
"journal": "",
"ref_id": "b14",
"title": "Learning human-object interaction detection using interaction points",
"year": "2020"
},
{
"authors": "Bumsoo Kim; Taeho Choi; Jaewoo Kang; Hyunwoo J Kim",
"journal": "Springer",
"ref_id": "b15",
"title": "Union-level detector towards real-time human-object interaction detection",
"year": "2020"
},
{
"authors": "Yogesh Balaji; Seungjun Nah; Xun Huang; Arash Vahdat; Jiaming Song; Karsten Kreis; Miika Aittala; Timo Aila; Samuli Laine; Bryan Catanzaro",
"journal": "",
"ref_id": "b16",
"title": "ediffi: Text-to-image diffusion models with an ensemble of expert denoisers",
"year": "2022"
},
{
"authors": "Yufan Zhou; Ruiyi Zhang; Changyou Chen; Chunyuan Li; Chris Tensmeyer; Tong Yu; Jiuxiang Gu; Jinhui Xu; Tong Sun",
"journal": "",
"ref_id": "b17",
"title": "Towards language-free training for text-to-image generation",
"year": "2022"
},
{
"authors": "Chitwan Saharia; William Chan; Saurabh Saxena; Lala Li; Jay Whang; Emily L Denton; Kamyar Ghasemipour; Raphael Gontijo Lopes; Burcu Karagol Ayan; Tim Salimans",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b18",
"title": "Photorealistic text-to-image diffusion models with deep language understanding",
"year": "2022"
},
{
"authors": "Aditya Ramesh; Prafulla Dhariwal; Alex Nichol; Casey Chu; Mark Chen",
"journal": "",
"ref_id": "b19",
"title": "Hierarchical text-conditional image generation with clip latents",
"year": "2022"
},
{
"authors": "Robin Rombach; Andreas Blattmann; Dominik Lorenz; Patrick Esser; Björn Ommer",
"journal": "",
"ref_id": "b20",
"title": "Highresolution image synthesis with latent diffusion models",
"year": "2022"
},
{
"authors": "Raphael Tang; Akshat Pandey; Zhiying Jiang; Gefei Yang; Karun Kumar; Jimmy Lin; Ferhan Ture",
"journal": "",
"ref_id": "b21",
"title": "What the daam: Interpreting stable diffusion using cross attention",
"year": "2022"
},
{
"authors": "Jiarui Xu; Sifei Liu; Arash Vahdat; Wonmin Byeon; Xiaolong Wang; Shalini De Mello",
"journal": "",
"ref_id": "b22",
"title": "Open-vocabulary panoptic segmentation with text-to-image diffusion models",
"year": "2023"
},
{
"authors": "Ziyi Li; Qinye Zhou; Xiaoyun Zhang; Ya Zhang; Yanfeng Wang; Weidi Xie",
"journal": "",
"ref_id": "b23",
"title": "Guiding text-to-image diffusion model towards grounded generation",
"year": "2023"
},
{
"authors": "Jordan Shipard; Arnold Wiliem; Kien Nguyen Thanh; Wei Xiang; Clinton Fookes",
"journal": "",
"ref_id": "b24",
"title": "Diversity is definitely needed: Improving model-agnostic zero-shot classification via stable diffusion",
"year": ""
},
{
"authors": "Yong-Lu Li; Siyuan Zhou; Xijie Huang; Liang Xu; Ze Ma; Hao-Shu Fang; Yanfeng Wang; Cewu Lu",
"journal": "",
"ref_id": "b25",
"title": "Transferable interactiveness knowledge for human-object interaction detection",
"year": "2019"
},
{
"authors": "Zhi Hou; Baosheng Yu; Yu Qiao; Xiaojiang Peng; Dacheng Tao",
"journal": "",
"ref_id": "b26",
"title": "Detecting human-object interaction via fabricated compositional learning",
"year": "2021"
},
{
"authors": "Bingjie Xu; Yongkang Wong; Junnan Li; Qi Zhao; Mohan S Kankanhalli",
"journal": "",
"ref_id": "b27",
"title": "Learning to detect human-object interactions with knowledge",
"year": "2019"
},
{
"authors": "Ankan Bansal; Sai Saketh Rambhatla; Abhinav Shrivastava; Rama Chellappa",
"journal": "",
"ref_id": "b28",
"title": "Detecting human-object interactions via functional generalization",
"year": "2020"
},
{
"authors": "Yong-Lu Li; Xinpeng Liu; Han Lu; Shiyi Wang; Junqi Liu; Jiefeng Li; Cewu Lu",
"journal": "",
"ref_id": "b29",
"title": "Detailed 2d-3d joint representation for human-object interaction",
"year": "2020"
},
{
"authors": "Ye Liu; Junsong Yuan; Chang Wen; Chen ",
"journal": "",
"ref_id": "b30",
"title": "Consnet: Learning graph for zero-shot human-object interaction detection",
"year": "2020"
},
{
"authors": "Chen Gao; Jiarui Xu; Yuliang Zou; Jia-Bin Huang",
"journal": "Springer",
"ref_id": "b31",
"title": "Drg: Dual relation graph for humanobject interaction detection",
"year": "2020"
},
{
"authors": "Siyuan Qi; Wenguan Wang; Baoxiong Jia; Jianbing Shen; Song-Chun Zhu",
"journal": "",
"ref_id": "b32",
"title": "Learning human-object interactions by graph parsing neural networks",
"year": "2018"
},
{
"authors": "Hai Wang; Wei-Shi Zheng; Ling Yingbiao",
"journal": "Springer",
"ref_id": "b33",
"title": "Contextual heterogeneous graph network for human-object interaction detection",
"year": "2020"
},
{
"authors": "Frederic Z Zhang; Dylan Campbell; Stephen Gould",
"journal": "",
"ref_id": "b34",
"title": "Spatially conditioned graphs for detecting human-object interactions",
"year": "2021"
},
{
"authors": "Jeeseung Park; Jin-Woo Park; Jong-Seok Lee",
"journal": "",
"ref_id": "b35",
"title": "Viplo: Vision transformer based pose-conditioned self-loop graph for human-object interaction detection",
"year": "2023"
},
{
"authors": "Yichen Hao-Shu Fang; Dian Xie; Cewu Shao; Lu",
"journal": "",
"ref_id": "b36",
"title": "Dirv: Dense interaction region voting for end-to-end human-object interaction detection",
"year": "2021"
},
{
"authors": "Masato Tamura; Hiroki Ohashi; Tomoaki Yoshinaga",
"journal": "",
"ref_id": "b37",
"title": "Qpic: Query-based pairwise humanobject interaction detection with image-wide contextual information",
"year": "2021-06"
},
{
"authors": "Masato Tamura; Hiroki Ohashi; Tomoaki Yoshinaga",
"journal": "",
"ref_id": "b38",
"title": "Qpic: Query-based pairwise humanobject interaction detection with image-wide contextual information",
"year": "2021"
},
{
"authors": "Bumsoo Kim; Junhyun Lee; Jaewoo Kang; Eun-Sol Kim; Hyunwoo J Kim",
"journal": "",
"ref_id": "b39",
"title": "Hotr: End-toend human-object interaction detection with transformers",
"year": "2021"
},
{
"authors": "Junwen Chen; Keiji Yanai",
"journal": "",
"ref_id": "b40",
"title": "Qahoi: query-based anchors for human-object interaction detection",
"year": "2021"
},
{
"authors": "Shuailei Ma; Yuefeng Wang; Shanze Wang; Ying Wei",
"journal": "",
"ref_id": "b41",
"title": "Fgahoi: Fine-grained anchors for human-object interaction detection",
"year": "2023"
},
{
"authors": "Mingfei Chen; Yue Liao; Si Liu; Zhiyuan Chen; Fei Wang; Chen Qian",
"journal": "",
"ref_id": "b42",
"title": "Reformulating hoi detection as adaptive set prediction",
"year": "2021"
},
{
"authors": "Julia Peyre; Ivan Laptev; Cordelia Schmid; Josef Sivic",
"journal": "",
"ref_id": "b43",
"title": "Detecting unseen visual relations using analogies",
"year": "2019"
},
{
"authors": "Zhi Hou; Baosheng Yu; Yu Qiao; Xiaojiang Peng; Dacheng Tao",
"journal": "",
"ref_id": "b44",
"title": "Affordance transfer learning for human-object interaction detection",
"year": "2021"
},
{
"authors": "Alec Radford; Jong Wook Kim; Chris Hallacy; Aditya Ramesh; Gabriel Goh; Sandhini Agarwal; Girish Sastry; Amanda Askell; Pamela Mishkin; Jack Clark",
"journal": "PMLR",
"ref_id": "b45",
"title": "Learning transferable visual models from natural language supervision",
"year": "2021"
},
{
"authors": "Mingrui Wu; Jiaxin Gu; Yunhang Shen; Mingbao Lin; Chao Chen; Sun ; Rongrong Ji",
"journal": "",
"ref_id": "b46",
"title": "End-to-end zero-shot hoi detection via vision and language knowledge distillation",
"year": "2022"
},
{
"authors": "Shan Ning; Longtian Qiu; Yongfei Liu; Xuming He",
"journal": "",
"ref_id": "b47",
"title": "Hoiclip: Efficient knowledge transfer for hoi detection with vision-language models",
"year": "2023"
},
{
"authors": "Brandon Trabucco; Kyle Doherty; Max Gurinas; Ruslan Salakhutdinov",
"journal": "",
"ref_id": "b48",
"title": "Effective data augmentation with diffusion models",
"year": "2023"
},
{
"authors": "Hritik Bansal; Aditya Grover",
"journal": "",
"ref_id": "b49",
"title": "Leaving reality to imagination: Robust classification via generated datasets",
"year": "2023"
},
{
"authors": "Zebin You; Yong Zhong; Fan Bao; Jiacheng Sun; Chongxuan Li; Jun Zhu",
"journal": "",
"ref_id": "b50",
"title": "Diffusion models and semi-supervised learners benefit mutually with few labels",
"year": "2023"
},
{
"authors": "Hao Zhang; Feng Li; Shilong Liu; Lei Zhang; Hang Su; Jun Zhu; Lionel M Ni; Heung-Yeung Shum",
"journal": "",
"ref_id": "b51",
"title": "Dino: Detr with improved denoising anchor boxes for end-to-end object detection",
"year": "2022"
},
{
"authors": "Tsung-Yi Lin; Michael Maire; Serge Belongie; James Hays; Pietro Perona; Deva Ramanan; Piotr Dollár; C Lawrence; Zitnick ",
"journal": "Springer",
"ref_id": "b52",
"title": "Microsoft coco: Common objects in context",
"year": "2014"
},
{
"authors": "Jack Hessel; Ari Holtzman; Maxwell Forbes; Ronan Le Bras; Yejin Choi",
"journal": "",
"ref_id": "b53",
"title": "Clipscore: A reference-free evaluation metric for image captioning",
"year": "2021"
},
{
"authors": "Nicolas Carion; Francisco Massa; Gabriel Synnaeve; Nicolas Usunier; Alexander Kirillov; Sergey Zagoruyko",
"journal": "Springer",
"ref_id": "b54",
"title": "End-to-end object detection with transformers",
"year": "2020"
},
{
"authors": "Saurabh Gupta; Jitendra Malik",
"journal": "",
"ref_id": "b55",
"title": "Visual semantic role labeling",
"year": "2015"
},
{
"authors": "Ze Liu; Yutong Lin; Yue Cao; Han Hu; Yixuan Wei; Zheng Zhang; Stephen Lin; Baining Guo",
"journal": "",
"ref_id": "b56",
"title": "Swin transformer: Hierarchical vision transformer using shifted windows",
"year": "2021"
},
{
"authors": "Xizhou Zhu; Weijie Su; Lewei Lu; Bin Li; Xiaogang Wang; Jifeng Dai",
"journal": "",
"ref_id": "b57",
"title": "Deformable detr: Deformable transformers for end-to-end object detection",
"year": "2020"
},
{
"authors": "Yong Zhang; Yingwei Pan; Ting Yao; Rui Huang; Tao Mei; Chang-Wen Chen",
"journal": "",
"ref_id": "b58",
"title": "Exploring structure-aware transformer over interaction proposals for human-object interaction detection",
"year": "2022"
},
{
"authors": "Ying Jin; Yinpeng Chen; Lijuan Wang; Jianfeng Wang; Pei Yu; Lin Liang; Jenq-Neng Hwang; Zicheng Liu",
"journal": "",
"ref_id": "b59",
"title": "The overlooked classifier in human-object interaction recognition",
"year": "2022"
},
{
"authors": "Frederic Z Zhang; Dylan Campbell; Stephen Gould",
"journal": "",
"ref_id": "b60",
"title": "Efficient two-stage detection of humanobject interactions with a novel unary-pairwise transformer",
"year": "2022"
},
{
"authors": "Bumsoo Kim; Jonghwan Mun; Minchul Kyoung-Woon On; Junhyun Shin; Eun-Sol Lee; Kim",
"journal": "",
"ref_id": "b61",
"title": "Mstr: Multi-scale transformer for end-to-end human-object interaction detection",
"year": "2022"
},
{
"authors": "Hao Asm Iftekhar; Kaustav Chen; Xinyu Kundu; Joseph Li; Davide Tighe; Modolo",
"journal": "",
"ref_id": "b62",
"title": "What to look at and where: Semantic and spatial refined transformer for detecting human-object interactions",
"year": "2022"
},
{
"authors": "Xian Qu; Changxing Ding; Xingao Li; Xubin Zhong; Dacheng Tao",
"journal": "",
"ref_id": "b63",
"title": "Distillation using oracle queries for transformer-based human-object interaction detection",
"year": "2022"
},
{
"authors": "Xinpeng Liu; Yong-Lu Li; Xiaoqian Wu; Yu-Wing Tai; Cewu Lu; Chi-Keung Tang",
"journal": "",
"ref_id": "b64",
"title": "Interactiveness field in human-object interactions",
"year": "2022"
},
{
"authors": "Dmitry Baranchuk; Ivan Rubachev; Andrey Voynov; Valentin Khrulkov; Artem Babenko",
"journal": "",
"ref_id": "b65",
"title": "Label-efficient semantic segmentation with diffusion models",
"year": "2021"
}
] | [
{
"formula_coordinates": [
7,
247.2,
303.93,
256.81,
12.95
],
"formula_id": "formula_0",
"formula_text": "P o = softmax(Q o * T o clip )(1)"
},
{
"formula_coordinates": [
7,
248.31,
321.26,
255.69,
12.95
],
"formula_id": "formula_1",
"formula_text": "P i = softmax(Q i * T i clip )(2)"
},
{
"formula_coordinates": [
7,
135.1,
338.39,
124.87,
12.19
],
"formula_id": "formula_2",
"formula_text": "Q o ∈ R N ×C and Q o ∈ R N ×C"
},
{
"formula_coordinates": [
7,
233.78,
452.58,
270.22,
12.69
],
"formula_id": "formula_3",
"formula_text": "L = λ b L b + λ g L g + λ o c L o c + λ i c L i c ,(3)"
}
] |
|
2023-05-20 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b32"
],
"table_ref": [],
"text": "The standard training strategy of a modern Neural Image Captioning system includes a policy gradient method, called Self-Critical Sequence Training [27] (shortened as SCST) which is designed to maximize the evaluation score given to the outputs. In this work, we discuss the problems caused by the lack of transparency from the research community over the inclusion or omission of the End-of-Sequence token during the optimization. An easy-to-overlook implementation detail that can significantly increase the performance of any model despite yielding worse descriptions.\nThe lack of awareness of the impact of the End-of-Sequence (<Eos>) omission and the lack of explicit information on the SCST implementation during the reporting of results pose an obstacle to scientific progress as they make it challenging to compare established works and evaluate new ones. Our paper attempts to spread awareness about the issue and proposes a solution to increase transparency in future works. This paper is structured as follows: in Section 2, we discuss the problem of the End-of-Sequence omission and why it is a problem for the research community; in Section 3, we provide a qualitative and quantitative analysis of the issue and we sample some of the recent works in Image Captioning to demonstrate its pervasiveness and provide some practical examples of its impact; In Section 4, we propose a possible solution with the help of a Python library called SacreEOS; in Section 5, we mention some of the literature approaches, and, finally, we draw our conclusions in Section 6."
},
{
"figure_ref": [],
"heading": "Problem Description",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "CIDEr Optimization",
"publication_ref": [
"b35",
"b29",
"b21",
"b8",
"b7",
"b35",
"b32"
],
"table_ref": [],
"text": "CIDEr [30] is an n-gram-based metric that evaluates the caption semantic content according to its similarities to the ground truths. Compared to the other metrics [24,16,3,2], it exploits the entire corpus of reference descriptions in the attempt of backing the evaluation with the consensus of the majority of people. In particular, each n-gram w k in sequence Z is weighted according to the tf-idf term g n k (Z) defined as:\nh n k (Z) w l ∈Ω h n l (Z) • log( |I| Ii∈I min(1, q h n k (V i q )) ) (1\n)\nwhere Ω is the set possible n-grams in the corpus, I is the set of corpus images and h n k (Z), h n k (V i j ) represent the number of occurrences of n-gram w k in the sequence Z and in the j-th ground truth of image I i ∈ I. The CIDEr and its alternative (CIDEr-D), compute the similarity between the candidate and reference description as the number of matching n-grams, weighted according to Equation 1. We refer to [30] for additional details of the formula since they are unnecessary for the sake of the discussion.\nThe standard training practice of the Image Captioning model consists of a pre-training phase using the Cross-Entropy loss followed by a CIDer-D optimization by means of a policy gradient method called Self-Critical Sequence Training [27]. The latter minimizes the negative expected reward:\nL R (θ) = -E y 1:T ∼p θ [r(y 1:T )] (2\n)\nwhere r is the CIDEr function, and its gradient is approximated as follows:\n∇ θ L R (θ) ≈ -(r(y s 1:T ) -r(y b 1:T ))∇ θ log p θ (y s 1:T )(3)\nwhere y s 1:T are the sampled captions and y b 1:T are the base predictions."
},
{
"figure_ref": [],
"heading": "The End-of-Sequence token in SCST",
"publication_ref": [],
"table_ref": [],
"text": "Two properties are desirable in an image description: completeness and correctness. While the first goal is pursued by the reward maximization, the SCST algorithm provides no explicit control over the latter, which is instead implicitly encouraged by the sequentiality of the decoding process. A token predicted at a specific time step also determines the most likely n-grams in the following ones. Since all n-grams are extracted from linguistically correct references, the final description will be correct, at least locally. Unfortunately, the CIDEr score does not consider a sentence's global correctness, and this aspect can be easily exploited by the SCST if not carefully implemented. In particular, the algorithm is allowed to produce incomplete descriptions using trivial sentence fragments that almost certainly match some parts of any set of references. This is the reason why the standard SCST implementation includes the special End-of-Sequence token, abbreviated as <Eos>, in the definition of the n-grams space. With this precaution, the reward function encourages a correct sentence termination leveraging the fact that the tf-idf of the <Eos> token out-weights those of function words."
},
{
"figure_ref": [
"fig_0",
"fig_0"
],
"heading": "The problem of the <Eos> omission",
"publication_ref": [],
"table_ref": [],
"text": "The inclusion or exclusion of the <Eos> token in the SCST algorithm represents a small and easy-to-overlook detail that significantly impacts a captioning system's performance. In case the <Eos> token is omitted, the descriptions generated by the network are often terminated by trivial sentence fragments such as \"and a\", \"in the\", \"on top of\" and \"in front of\" (more examples in Figure 1). However, despite the presence of artifacts, they achieve superior performances on popular benchmarks compared to the correct ones (Figure 1). In particular, the number of additional points yielded by the artifacts can even be greater than the range of values in which different models developed around the same period typically compete. Therefore, the Image Captioning research field is currently suffering from a lack of transparency and, in some cases lack of awareness over the importance of the <Eos> token in the SCST. The problem can be described from multiple perspectives:\n-If details over the <Eos> token in the SCST implementation are unavailable, omitted, or simply overlooked, it becomes difficult to compare models in the literature fairly. -Researchers that are aware of the issue are given the difficult choice between less competitive results and poorly formulated outputs. -Finally, researchers that are not aware of the issue (especially the newcomers in the field of Image Captioning) are indirectly encouraged to adopt the implementations that generate compromised sentences because of their superior performances.\n3 <Eos> Omission Impact Analysis"
},
{
"figure_ref": [],
"heading": "Experimental Setup",
"publication_ref": [
"b34",
"b22",
"b16",
"b6"
],
"table_ref": [],
"text": "For the qualitative and quantitative analysis of artifacts we implement1 the Transformer [29] with 3 layers, d model =512 and d f f =2048, trained on the COCO 2014 [17] data set using the Karpathy split [11]. The Faster-RCNN backbone provided by [1] is adopted. The learning procedure consists of a first training step on Cross Entropy loss for 8 epochs followed by the CIDEr-D optimization for 20 epochs. The following configurations are adopted:\n1. batch size of 48, a learning rate of 2e-4 annealed by 0.8 every 2 epochs and warm-up of 10000 in case of Cross Entropy Loss; 2. batch size of 48, a learning rate of 1e-4 annealed by 0.8 every 2 epochs during the SCST.\nOptimization details are provided only for the sake of reproducibility since the artifacts discussed in this work arise regardless of the architecture and optimization details. For the ensemble results, 4 model instances are generated with the aforementioned method differing only in the initialization seed. In the experiments, for each seed, the SCST in the Standard and No<Eos> configurations optimize the same pre-trained model."
},
{
"figure_ref": [
"fig_1",
"fig_1",
"fig_1"
],
"heading": "Artifacts Analysis",
"publication_ref": [],
"table_ref": [],
"text": "The <Eos> token can be omitted in two aspects of SCST:\n1. during the reward computation; 2. during the initialization of tf-idfs;\nwhich leads to 4 implementation instances in case sampled descriptions are tokenized consistently with respect to the ground-truths. Prediction w/ <Eos>) and (tf-idf Init. w/o <Eos>, Prediction w/o <Eos>) configuration referred as \"Standard\" and \"No<Eos>\" respectively throughout the rest of this work.\nIn the No<Eos> configuration, results are affected by 8 classes of artifacts depending on how sequences are terminated, with the last token belonging to A={\"in\", \"a\", \"of\", \"the\", \"with\", \"on\", \"and\" \"*\"}, where \"*\" represents all the possible remaining cases. While all elements in the set A are just simplifications of longer trivial fragments such as \"and a\", \"in a\", \"with a\" and \"in front of\", the case of \"on\" may seem acceptable but the token is often part of uncommon formulations such as \"a beach with a surfboard on\" and \"a street with a bus on\". Nevertheless, \"on\" represents only a small fraction of all instances, which mostly end with the \"a\" token instead (see Figure 2.c).\nFigure 2.a showcases the number of artifacts converging to 50% of the whole testing set as the number of epochs increases. Thus, both correct and compromised sentences are produced by the <Eos> omission, which means the network learns to inject the fragments following a non-trivial and unpredictable criteria for each sequence.\nFigure 2.b and Table 3.2 showcase that a single model trained with SCST in the No<Eos> configuration consistently outperforms the standard one across all seeds, often by a large margin, with a maximum gain of +2.8 and +4.3 CIDEr-D in the offline test and validation set respectively. Whereas, by removing the artifacts from the latter predictions we observed the opposite trend with a maximum performance decrease of -2.3 and -2.0. Therefore, the increase in score is mostly due to the artifacts and the <Eos> omission poses an obstacle to the generation of semantically meaningful content. Similar behaviour is observed for ensemble performances (referred as )."
},
{
"figure_ref": [],
"heading": "Literature classification",
"publication_ref": [
"b13",
"b28"
],
"table_ref": [
"tab_2"
],
"text": "We sample recent works in the research literature and classify each of them according to the way SCST is implemented. In Section 3.2 we observed that only half of the evaluated sentences are compromised, which means that if a paper provides only a few correct captioning examples, it is not enough to determine whether the <Eos> token was omitted or not. Because of that, the classification is made through code inspection. The classes and the respective criteria are defined as follows:\n-Standard: <Eos> token is included in both SCST initialization and reward computation or complete results on either test or validation set are provided; -No<Eos>: <Eos> token is omitted in both initialization and reward computation; -Unknown: the code was not found or it was not available at the time this work was completed.\nTable 3.3 showcases that only 12 of 25 works are confirmed to follow the Standard implementation, 8 fall in the No<Eos> category and 5 are unknown. The State-of-the-art architectures in 2019 [8] and 2020 [23] achieved 129.6 and 131.4 CIDEr-D scores respectively, which showcases the gradual improvement process of the research activity and provides an example of the magnitude of improvements over the years. Unfortunately, such a difference in performance can be lower than the additional score yielded by artifacts (see Section 3.2). For instance, if AoANet adopted the No<Eos> configuration, its score would have been comparable to the State-of-the-art performances of the following year (X-Transformer) (see Table 4).\nThe amount of No<Eos> implementations in the last years confirms the phenomena described in Section 2.3.\nTable 3. SCST classification of recent Image Captioning works and their respective performances on the MS-COCO 2014 task. The offline case reports the CIDEr-D score of a single model in contrast to the online evaluation server results where an ensemble is adopted instead with some exceptions denoted with \" \". "
},
{
"figure_ref": [],
"heading": "SacreEOS",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "SacreEOS signature",
"publication_ref": [],
"table_ref": [],
"text": "The lack of transparency and awareness over the <Eos> token in SCST originates from an easy-to-overlook implementation detail. Therefore, the natural solution is to disseminate awareness of the issue. To achieve this goal we introduce SacreEOS, a Python library whose main functionality consists of the generation of signatures that uniquely identify the key aspects of the SCST implementation. In particular, how the <Eos> token is handled. The sharing of the SacreEOS signature accomplishes three objectives:\n1. it increases transparency and eases the comparison of models; 2. it informs the reader about the presence or absence of artifacts (those related to the <Eos> omission) in the results; 3. last but not least, it spreads awareness of the problem.\nWe believe this is especially useful in cases of works that do not release the code to the public.\nEstablished researchers and existing implementations can manually generate the signature using the SacreEOS command line interface. The tool simply asks a few questions regarding the technical aspects of SCST, therefore it does not require any code integration. For new projects instead, SacreEOS consists of an SCST implementation helper, in this case, the signature is provided automatically. "
},
{
"figure_ref": [],
"heading": "Implementation helper and limitations of the approach",
"publication_ref": [],
"table_ref": [],
"text": "In addition to the functionality of signature generation, the SacreEOS library optionally provides helpful classes to ease the implementation of SCST in future projects. In particular, it covers the following aspects:\n-SCST class selection. Given the number of established works implemented in both Standard and No<Eos> configurations, it is out of the scope of this paper to decide which one is the \"correct\" one (the library provides no default option in this regard). However, the tool helps the user to make informed decisions. Classes are currently defined by the reward metric, the reward base and whether the <Eos> token is included or omitted in both initialization and reward computation. -SCST initialization. The library initializes the tf-idfs for the reward computation and performs input checks according to the selected class. -SCST reward computation. The library currently supports the following reward functions CIDEr, CIDEr-D, CIDEr-R and BLEU. Results are consistent with the official repositories 2 . Each function is implemented in both Python and C, users can optionally enable the latter version to increase efficiency. -Signature generation. In this case the SacreEOS signature is automatically determined by the class selection and does not require user intervention.\nThe library includes an intricate collection of assertions and input checks on all implementation levels, taylored to each specific class. Nevertheless, the SacreEOS does not prevent misreporting. In case the signature is manually generated, it relies on the user to provide the correct data."
},
{
"figure_ref": [],
"heading": "Related works",
"publication_ref": [
"b32",
"b15",
"b24",
"b33",
"b30",
"b29"
],
"table_ref": [],
"text": "The work of [27] mentioned the role of the End-of-Sequence token. However, it only provided a few qualitative examples and did not report numerical details. Several works in the past focused on improving the evaluation of Image Captioning systems but they mostly proposed alternatives to the CIDEr metric, such as TIGEr [10], SPIDEr [19], and CIDEr-R [28]. None of them addressed the issue discussed in this work.\nThe main inspiration of SacreEOS is SacreBLEU [25], in the field of Machine Translation, where ambiguities can arise from different tokenization and detokenization choices that ultimately affect the BLEU score [24]."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "Our work discussed the role of <Eos> in the Self-Critical Sequence Training and how the lack of transparency and awareness over its function pose an obstacle 2 CIDEr, CIDEr-D, BLEU: github.com/vrama91/cider CIDEr-R: github.com/gabrielsantosrv/coco-caption to the scientific progress in the Image Captioning field. We described the source of the problem from a qualitative and quantitative perspective. We classified recent works in the scientific literature according to the SCST configuration to showcase the pervasiveness and the importance of the matter. Finally, we proposed a possible solution that consists of sharing a unique signature with the help of a Python library called SacreEOS, to enable fair model comparisons and spread awareness regarding the issue."
}
] | The Image Captioning research field is currently compromised by the lack of transparency and awareness over the End-of-Sequence token (<Eos>) in the Self-Critical Sequence Training. If the <Eos> token is omitted, a model can boost its performance up to +4.1 CIDEr-D using trivial sentence fragments. While this phenomenon poses an obstacle to a fair evaluation and comparison of established works, people involved in new projects are given the arduous choice between lower scores and unsatisfactory descriptions due to the competitive nature of the research. This work proposes to solve the problem by spreading awareness of the issue itself. In particular, we invite future works to share a simple and informative signature with the help of a library called SacreEOS. | A request for clarity over the End of Sequence token in the Self-Critical Sequence Training | [
{
"figure_caption": "Fig. 1 .1Fig. 1. Captions generated by the same model (the Transformer [29]) trained with different implementations of SCST on the MS-COCO [17] data set. (Left) The model is optimized by the standard SCST and achieves 125.8 CIDEr-D on the validation set. (Right) The model is optimized by an implementation of SCST in which the <Eos> token is omitted and achieves 130.1 CIDEr-D on the validation set.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 2 .2Fig. 2. a) The number of artifacts in the No<Eos> configuration on 5000 test set predictions. b) Average CIDEr-D score of 4 training instances (different seeds) in the Standard and No<Eos> configuration, \"Cleaned\" denotes the No<Eos> performance in case artifacts are removed before the evaluation. c) Artifacts distribution. Sequences terminated by \"a\" account for 89.8% of all cases (top). Histogram of sequences terminated by \"a\" (bottom).",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Format and signature examples are the following: Format: <scst config>_<Init>+<metric[args]>+<base[args]>+<Version> Examples: STANDARD_w/oInit+Cider-D[n5,s6.0]+average[nspi5]+1.0.0 NO<EOS>MODE_wInit+Cider-D[n4,s6.0]+greedy[nspi5]+1.0.0 NO<EOS>MODE_w/oInit+BLEU[n4]+average[nspi5]+1.0.0",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Table1reports the impact of each configuration over the final descriptions. Two cases are the focus of this work since most popular implementations fall into the (tf-idf Init. w/ <Eos>, Impact of the <Eos> token in SCST over the final CIDEr-D score and outputs. \"tf-idf Init.\" refers to the ground truth sentences involved in the calculation of document frequencies, and \"Predictions\" refers to the sampled predictions and respective references.",
"figure_data": "tf-idf Init.tf-idf Init.w/ <Eos>w/o <Eos>Reward baseline score lower scorew/ <Eos> no artifacts with artifactsRewardlower score higher scorew/o <Eos> with artifacts with artifactsKarpathy test splitKarpathy validation splitStandardNo<Eos> (ε) / δCleaned / δ StandardNo<Eos> (ε) / δCleaned / δSeed 1 128.4 131.2 (48.3%) / +2.8 127.8 / -0.6 125.8 130.1 (47.5%) / +4.3 126.4 / +0.6Seed 2 129.0 130.9 (49.3%) / +1.9 127.4 / -1.6 127.0 129.9 (48.1%) / +2.9 126.2 / -0.8Seed 3 129.0 131.0 (50.3%) / +2.0 127.5 / -1.5 127.2 129.3 (47.6%) / +2.1 125.7 / -1.5Seed 4 129.1 130.7 (50.4%) / +1.6 126.8 / -2.3 128.0 130.0 (50.6%) / +2.0 126.0 / -2.0Avg128.9 130.9 (49.6%) / +2.0 127.3 / -1.1 126.9 129.8 (48.6%) / +2.8 126.0 / -0.9133.0 134.9 (50.2%) / +1.9 131.2 / -1.8 131.8 133.8 (49.5%) / +2.0 129.8 / -2.0",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Performance comparison the CIDEr-D optimization in Standard and No<Eos> training. \"Cleaned\" refers to the No<Eos> results but artifacts are removed prior to the evaluation.refers to the ensemble of the four models and ε represents the percentage of artifacts.",
"figure_data": "",
"figure_id": "tab_1",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "CIDEr-D performance increase observed in open source projects when the SCST configuration is changed from Standard into No<Eos> mode. Training details can be found in the respective works or repositories.",
"figure_data": "a Prefix https://github.com/",
"figure_id": "tab_2",
"figure_label": "4",
"figure_type": "table"
}
] | Jia Cheng Hu; Roberto Cavicchioli; Alessandro Capotondi | [
{
"authors": "",
"journal": "",
"ref_id": "b0",
"title": "Year Work Offline Online SCST Code inspection a (commit) 2018 GCN-LSTM",
"year": ""
},
{
"authors": "",
"journal": "",
"ref_id": "b1",
"title": "Standard ruotianluo/self-critical",
"year": ""
},
{
"authors": "",
"journal": "Standard LuoweiZhou/VLP",
"ref_id": "b2",
"title": "d85)",
"year": "2021"
},
{
"authors": "",
"journal": "Standard jchenghu/ExpansionNet_v",
"ref_id": "b3",
"title": "d130) 2022 BLIP [15] 136.7",
"year": ""
},
{
"authors": "",
"journal": "",
"ref_id": "b4",
"title": "Standard OFA-Sys/OFA (1809b55",
"year": "2022"
},
{
"authors": "",
"journal": "",
"ref_id": "b5",
"title": "GT-RIPL/Xmodal-Ctx (d927eec) Bibliography",
"year": ""
},
{
"authors": "Peter Anderson",
"journal": "",
"ref_id": "b6",
"title": "Bottom-up and top-down attention for image captioning and visual question answering",
"year": "2018"
},
{
"authors": "Peter Anderson",
"journal": "Springer",
"ref_id": "b7",
"title": "Spice: Semantic propositional image caption evaluation",
"year": "2016"
},
{
"authors": "Satanjeev Banerjee; Alon Lavie",
"journal": "",
"ref_id": "b8",
"title": "METEOR: An automatic metric for MT evaluation with improved correlation with human judgments",
"year": "2005"
},
{
"authors": "Manuele Barraco",
"journal": "",
"ref_id": "b9",
"title": "CaMEL: Mean Teacher Learning for Image Captioning",
"year": "2022"
},
{
"authors": "Marcella Cornia",
"journal": "",
"ref_id": "b10",
"title": "Meshed-memory transformer for image captioning",
"year": "2020"
},
{
"authors": "Simao Herdade",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b11",
"title": "Image captioning: Transforming objects into words",
"year": "2019"
},
{
"authors": "Jia Cheng Hu; Roberto Cavicchioli; Alessandro Capotondi",
"journal": "",
"ref_id": "b12",
"title": "Expan-sionNet v2: Block Static Expansion in fast end to end training for Image Captioning",
"year": "2022"
},
{
"authors": "Lun Huang",
"journal": "",
"ref_id": "b13",
"title": "Attention on attention for image captioning",
"year": "2019"
},
{
"authors": "Jiayi Ji",
"journal": "",
"ref_id": "b14",
"title": "Improving image captioning by leveraging intra-and interlayer global representation in transformer network",
"year": "2021"
},
{
"authors": "Ming Jiang",
"journal": "",
"ref_id": "b15",
"title": "Tiger: Text-to-image grounding for image caption evaluation",
"year": "2019"
},
{
"authors": "Andrej Karpathy; Li Fei-Fei",
"journal": "",
"ref_id": "b16",
"title": "Deep visual-semantic alignments for generating image descriptions",
"year": "2015"
},
{
"authors": "Lei Ke",
"journal": "",
"ref_id": "b17",
"title": "Reflective decoding network for image captioning",
"year": "2019"
},
{
"authors": "Chia-Wen Kuo; Zsolt Kira",
"journal": "",
"ref_id": "b18",
"title": "Beyond a Pre-Trained Object Detector: Cross-Modal Textual and Visual Context for Image Captioning",
"year": "2022"
},
{
"authors": "Jingyu Li",
"journal": "",
"ref_id": "b19",
"title": "ER-SAN: Enhanced-Adaptive Relation Self-Attention Network for Image Captioning",
"year": "2022-07"
},
{
"authors": "Junnan Li",
"journal": "",
"ref_id": "b20",
"title": "Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation",
"year": "2022"
},
{
"authors": "Chin-Yew Lin",
"journal": "",
"ref_id": "b21",
"title": "Rouge: A package for automatic evaluation of summaries",
"year": "2004"
},
{
"authors": "Tsung-Yi Lin",
"journal": "Springer",
"ref_id": "b22",
"title": "Microsoft coco: Common objects in context",
"year": "2014"
},
{
"authors": "Bing Liu",
"journal": "",
"ref_id": "b23",
"title": "Show, Deconfound and Tell: Image Captioning With Causal Inference",
"year": "2022"
},
{
"authors": "Siqi Liu",
"journal": "",
"ref_id": "b24",
"title": "Improved image captioning via policy gradient optimization of spider",
"year": "2017"
},
{
"authors": "Ruotian Luo",
"journal": "",
"ref_id": "b25",
"title": "A better variant of self-critical sequence training",
"year": "2020"
},
{
"authors": "Yunpeng Luo",
"journal": "",
"ref_id": "b26",
"title": "Dual-level collaborative transformer for image captioning",
"year": "2021"
},
{
"authors": "Masanori Van-Quang Nguyen; Takayuki Suganuma; Okatani",
"journal": "",
"ref_id": "b27",
"title": "GRIT: Faster and Better Image captioning Transformer Using Dual Visual Features",
"year": "2022"
},
{
"authors": "Yingwei Pan",
"journal": "",
"ref_id": "b28",
"title": "X-Linear Attention Networks for Image Captioning",
"year": "2020"
},
{
"authors": "Kishore Papineni",
"journal": "",
"ref_id": "b29",
"title": "Bleu: a method for automatic evaluation of machine translation",
"year": "2002"
},
{
"authors": "Matt Post",
"journal": "",
"ref_id": "b30",
"title": "A call for clarity in reporting BLEU scores",
"year": "2018"
},
{
"authors": "Yu Qin",
"journal": "",
"ref_id": "b31",
"title": "Look back and predict forward in image captioning",
"year": "2019"
},
{
"authors": "Rennie Steven",
"journal": "",
"ref_id": "b32",
"title": "Self-critical sequence training for image captioning",
"year": "2017"
},
{
"authors": "Gabriel Oliveira Dos Santos; Esther Luna Colombini; Sandra Avila",
"journal": "",
"ref_id": "b33",
"title": "Cider-r: Robust consensus-based image description evaluation",
"year": "2021"
},
{
"authors": "Ashish Vaswani",
"journal": "",
"ref_id": "b34",
"title": "Attention is all you need",
"year": "2017"
},
{
"authors": "Ramakrishna Vedantam; Lawrence Zitnick; Devi Parikh",
"journal": "",
"ref_id": "b35",
"title": "Cider: Consensus-based image description evaluation",
"year": "2015"
},
{
"authors": "Peng Wang",
"journal": "PMLR",
"ref_id": "b36",
"title": "Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework",
"year": "2022"
},
{
"authors": "Weixuan Wang; Zhihong Chen; Haifeng Hu",
"journal": "",
"ref_id": "b37",
"title": "Hierarchical attention network for image captioning",
"year": "2019"
},
{
"authors": "Yiyu Wang; Jungang Xu; Yingfei Sun",
"journal": "",
"ref_id": "b38",
"title": "End-to-End Transformer Based Model for Image Captioning",
"year": "2022"
},
{
"authors": "Yang Xu",
"journal": "",
"ref_id": "b39",
"title": "Auto-encoding scene graphs for image captioning",
"year": "2019"
},
{
"authors": "Ting Yao",
"journal": "",
"ref_id": "b40",
"title": "Exploring visual relationship for image captioning",
"year": "2018"
},
{
"authors": "Pengpeng Zeng",
"journal": "",
"ref_id": "b41",
"title": "S2 Transformer for Image Captioning",
"year": ""
},
{
"authors": "",
"journal": "",
"ref_id": "b42",
"title": "Main Track. International Joint Conferences on Artificial Intelligence Organization",
"year": "2022-07"
},
{
"authors": "Xuying Zhang",
"journal": "",
"ref_id": "b43",
"title": "RSTNet: Captioning with adaptive attention on visual and non-visual words",
"year": "2021"
},
{
"authors": "Luowei Zhou",
"journal": "",
"ref_id": "b44",
"title": "Unified vision-language pre-training for image captioning and vqa",
"year": "2020"
}
] | [
{
"formula_coordinates": [
2,
221.17,
240.39,
255.18,
32.89
],
"formula_id": "formula_0",
"formula_text": "h n k (Z) w l ∈Ω h n l (Z) • log( |I| Ii∈I min(1, q h n k (V i q )) ) (1"
},
{
"formula_coordinates": [
2,
476.35,
248.65,
4.24,
8.8
],
"formula_id": "formula_1",
"formula_text": ")"
},
{
"formula_coordinates": [
2,
247.76,
424.29,
228.59,
10.38
],
"formula_id": "formula_2",
"formula_text": "L R (θ) = -E y 1:T ∼p θ [r(y 1:T )] (2"
},
{
"formula_coordinates": [
2,
476.35,
424.29,
4.24,
8.8
],
"formula_id": "formula_3",
"formula_text": ")"
},
{
"formula_coordinates": [
2,
205.93,
464.65,
274.66,
12.69
],
"formula_id": "formula_4",
"formula_text": "∇ θ L R (θ) ≈ -(r(y s 1:T ) -r(y b 1:T ))∇ θ log p θ (y s 1:T )(3)"
}
] |
|
2023-05-25 | [
{
"figure_ref": [
"fig_1",
"fig_2"
],
"heading": "Introduction",
"publication_ref": [
"b32",
"b2",
"b42",
"b1",
"b5",
"b19",
"b6",
"b29",
"b30",
"b14",
"b13",
"b31",
"b13",
"b41",
"b20",
"b12",
"b41",
"b9",
"b16",
"b35"
],
"table_ref": [
"tab_0"
],
"text": "Current neural machine translation (NMT) has achieved great triumph (Sutskever et al., 2014;Bahdanau et al., 2015;Zhu et al., 2020), however in the cost of creating large-scale parallel sentences, which obstructs the development of NMT for the minor languages. Unsupervised NMT (UMT) has thus been proposed to relieve the reliance of parallel corpora (Artetxe et al., 2018;Chen et al., 2018). The core idea of UMT is to align the representation spaces between two languages with alternative pivot signals rather than parallel sentences, such as bilingual lexicons (Lample et al., 2018), multilingual language models (LM) (Conneau and Lample, 2019) and back-translation technique (Sennrich et al., 2016). Recent trends have considered the incorporation of visual information, i.e., multimodal machine translation (MMT) (Specia et al., 2016;Huang et al., 2016). Intuitively, visual modality can serve as language-agnostic signals, pivoting different languages by grounding the same textual semantics into the common visual space. Therefore, solving UMT with visual contents as pivot becomes a promising solution, a.k.a., unsupervised MMT (UMMT) (Huang et al., 2020;Su et al., 2019).\nUMMT systems are trained with only the textimage pairs (<text-img>), which can be easier to collect than the parallel source-target sentence pairs (<src-tgt>) (Huang et al., 2020). Although exempting the parallel sentences for training, UMMT still requires such text-image pairs as inputs for testing. Yet such assumption might be unrealistic, because in most of the real-world scenarios such as online translation systems, paired images are not available during inference. Especially for some scarce languages, the <text-img> pairs have difficult access. In other words, practical UMMT systems should not only avoid the parallel sentences during training, but also the text-image pairs during inference. As summarized in Table 1, although some existing MMT researches exempt the testing-time visual inputs (Zhang et al., 2020;Li et al., 2022), they all unfortunately are supervised methods, relying on large-scale parallel sentences for training. As emphasized above, the visual information is vital to UMMT. However, for both the existing supervised and unsupervised MMT studies, they may suffer from ineffective and insufficient modeling of visual pivot features. For example, most of MMT models perform vision-language (VL) grounding over the whole image and text (Huang et al., 2019;Zhang et al., 2020), where such coarse-grained representation learning can cause mismatching and sacrifice the subtle VL semantics. Fang and Feng (2022) recently introduce a fine-grained VL alignment learning via phrase-level grounding, while without a holistic understanding of the visual scene, such local-level method may lead to incomplete or missing alignments.\nIn this work, we present a novel UMMT method that solves all aforementioned challenges. First of all, to better represent the visual (also the textual) inputs, we consider incorporating the visual scene graph (VSG) (Johnson et al., 2015) and language scene graph (LSG) (Wang et al., 2018). The scene graphs (SG) advance in intrinsically depicting the semantic structures of texts or images with rich details (cf. Fig. 1), which offers a holistic viewpoint for more effective pivoting learning. Then, we build the UMMT framework as illustrated in Fig. 2. The input src text and paired image are first transformed into LSG and VSG, which are further fused into a mixed SG, and then translated into the tgt-side LSG. And the tgt sentence will be finally produced conditioned on the tgt LSG. Several SGbased pivoting learning strategies are proposed for unsupervised training of UMMT system. In addition, to support pure-text (image-free) input during inference, we devise a novel visual scene hallucination module, which dynamically generates a hallucinated VSG from the LSG compensatively. Our system is evaluated on the standard MMT Multi30K and NMT WMT data. Extensive experimental results verify that the proposed method outperforms strong baselines on unsupervised multimodal translation by above 5 BLEU score on average. We further reveal the efficacy of the visual scene hallucination mechanism in relieving the reliance on image inputs during inference. Our SG-pivoting based UMMT helps yield translations with higher completeness, relevance and fluency, and especially obtains improvements on the longer sentences.\nOverall, we make the following contributions: ▶ 1) We are the first to study the inferencetime image-free unsupervised multimodal machine translation, solved with a novel visual scene hallucination mechanism. ▶ 2) We leverage the SGs to better represent the visual and language inputs. Moreover, we design SG-based graph pivoting learning strategies for UMMT training. ▶ 3) Our model achieves huge boosts over strong baselines on benchmark data. Code is available at https: //github.com/scofield7419/UMMT-VSH."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b32",
"b2",
"b23",
"b33",
"b7",
"b30",
"b14",
"b29",
"b21",
"b19",
"b6",
"b25",
"b5",
"b31",
"b13",
"b18",
"b35",
"b16",
"b15",
"b39",
"b41",
"b9"
],
"table_ref": [],
"text": "Neural machine translation has achieved notable development in the era of deep learning (Sutskever et al., 2014;Bahdanau et al., 2015;Luong et al., 2015). The constructions of powerful neural models and training paradigms as well as the collection of large-scale parallel corpora are the driving forces to NMT's success (Vaswani et al., 2017;Devlin et al., 2019). The key of NMT is to learn a good mapping between two (or more) languages. In recent years, visual information has been intro-duced for stronger NMT (i.e., multimodal machine translation), by enhancing the alignments of language latent spaces with visual grounding (Specia et al., 2016;Huang et al., 2016). Intuitively, people speaking different languages can actually refer to the same physical visual contents and conceptions.\nUnsupervised machine translation aims to learn cross-lingual mapping without the use of largescale parallel corpora. The setting is practically meaningful to those minor languages with hard data accessibility. The basic idea is to leverage alternative pivoting contents to compensate the parallel signals based on the back-translation method (Sennrich et al., 2016), such as third-languages (Li et al., 2020), bilingual lexicons (Lample et al., 2018) or multilingual LM (Conneau and Lample, 2019). The visual information can also serve as pivot signals for UMT, i.e., unsupervised multimodal machine translation. Comparing to the standard MMT that trains with <src-img-tgt> triples, UMMT takes as input only the <src-img>. So far, few studies have explored the UMMT setting, most of which try to enhance the back-translation with multimodal alignment mechanism (Nakayama and Nishida, 2017;Chen et al., 2018;Su et al., 2019;Huang et al., 2020).\nScene graph describes a scene of an image or text into a structure layout, by connecting discrete objects with attributes and with other objects via pairwise relations (Krishna et al., 2017;Wang et al., 2018). As the SGs carry rich contextual and semantic information, they are widely integrated into downstream tasks for enhancements, e.g., image retrieval (Johnson et al., 2015), image generation (Johnson et al., 2018) and image captioning (Yang et al., 2019). This work inherits wisdom, incorporating both the visual scene graph and language scene graph as pivots for UMMT.\nAll the UMMT researches assume that the <src-img> pairs are required during inference, yet we notice that this can be actually unrealistic. We thus propose a visual hallucination mechanism, achieving the inference-time image-free goal. There are relevant studies on supervised MMT that manage to avoid image inputs (with text only) during inference. The visual retrieval-base methods (Zhang et al., 2020;Fang and Feng, 2022) sentence. Differently, we consider generating the visual scene graph with richer and holistic visual structural information.\n3 Scene Graph-based Translation System"
},
{
"figure_ref": [],
"heading": "Problem Definition",
"publication_ref": [],
"table_ref": [],
"text": "In UMMT, no parallel translation pairs are available. This work considers an inference-time imagefree UMMT. During training, the data availability is <x, z>∈<X , Z> and the corresponding srcside LSG x and VSG, where X are the src-side sentences, and Z are the paired images. During inference, the model generates tgt-side sentences y ∈ Y based on the inputs of only x ∈ X and the corresponding LSG x , while the visual scene VSG ′ is hallucinated from LSG x . In both training and inference, y will be generated from the intermediate tgt-side language scene graph LSG y , which is produced from LSG x and VSG (or VSG ′ )."
},
{
"figure_ref": [
"fig_2"
],
"heading": "Framework",
"publication_ref": [],
"table_ref": [],
"text": "As shown in Fig. 2, the system first represents the src-side LSG x and VSG features with two GCN graph encoders, respectively. Then the SG fus-ing&mapping module integrates and transforms two SG representations into a unified one as tgtside LSG, i.e., LSG y . Another GSN model further encodes the LSG y , where the representations are used to generate the tgt sentence (i.e., translation)."
},
{
"figure_ref": [
"fig_3"
],
"heading": "Scene Graph Generating and Encoding",
"publication_ref": [
"b24"
],
"table_ref": [],
"text": "We first employ two off-the-shelf SG parsers to obtain the LSG and VSG, separately (detailed in the experiment part). For simplicity, here we unify the notations of LSG and VSG as SG. We denote a SG as G=(V, E), where V are the nodes (including object o, attribute a and relation r types), and E are the edges e i,j between any pair of nodes v i ∈ V .\nWe then encode both the VSG and LSG with two spatial Graph Convolution Networks (GCN) (Marcheggiani and Titov, 2017) respectively, which is formulated as:\nr 1 , • • • , r n = GCN(G) ,(1)\nwhere r i is the representation of node v i . We here denote r L i as LSG's node representation, and r V i as VSG's node representation.\nVisual Scene Hallucinating During inference, the visual scene hallucination (VSH) module is activated to perform two-step inference to generate the hallucinated VSG ′ , as illustrated in Fig. 3.\nStep1: sketching skeleton aims to build the skeleton VSG. We copy all the nodes from the raw LSG x to the target VSG, and transform the textual entity nodes into the visual object nodes.\nStep2: completing vision aims to enrich and augment the skeleton VSG into a more realistic one. It is indispensable to add new nodes and edges in the skeleton VSG, since in real scenarios, visual scenes are much more concrete and vivid than textual scenes. Specifically, we develop a node augmentor and a relation augmentor, where the former decides whether to attach a new node to an existing one, and the later decides whether to create an edge between two disjoint nodes. To ensure the fidelity of the hallucinated VSG ′ , during training, the node augmentor and relation augmentor will be updated (i.e., with the learning target L VSH ) with the input LSG and VSG supervisions. Appendix §A.1 details the VSH module.\nSG Fusing&Mapping Now we fuse the heterogeneous LSG x and VSG into one unified scene graph with a mixed view. The key idea is to merge the information from two SGs serving similar roles.\nIn particular, we first measure the representation similarity of each pair of <text-img> nodes from two GCNs. For those pairs with high alignment scores, we merge them as one by averaging their representations, and for those not, we take the union structures from two SGs. This results in a pseudo tgt-side LSG y . We then use another GCN model for further representation propagation. Finally, we employ a graph-to-text generator to transform the LSG y representations to the tgt sentence y. Appendix §A.2 presents all the technical details in this part."
},
{
"figure_ref": [
"fig_4"
],
"heading": "Learning with Scene Graph Pivoting",
"publication_ref": [],
"table_ref": [],
"text": "In this part, based on the SG pivot we introduce several learning strategies to accomplish the unsupervised training of machine translation. We mainly consider 1) cross-SG visual-language learning, and 2) SG-pivoted back-translation training. Fig. 4 illustrates these learning strategies."
},
{
"figure_ref": [],
"heading": "Cross-SG Visual-language Learning",
"publication_ref": [
"b22",
"b38",
"b11",
"b15"
],
"table_ref": [],
"text": "The visual-language SG cross-learning aims to enhance the structural correspondence between the LSG and VSG. Via cross-learning we also teach the SG encoders to automatically learn to highlight those shared visual-language information while deactivating those trivial substructures, i.e., denoising.\nCross-modal SG Aligning The idea is to encourage the text and visual nodes that serve a similar role in VSG and LSG to be closer. To align the fine-grained structures between SGs, we adopt the contrastive learning (CL) technique (Logeswaran and Lee, 2018;Yan et al., 2021;Fei et al., 2022;Huang et al., 2022). In particular, CL learns effec-tive representation by pulling semantically close content pairs together, while pushing apart those different ones. Technically, we measure the similarities between pairs of nodes from two VSG and LSG:\nsi,j = (r L i ) T • r V j ||r L i || ||r V j || .(2)\nA threshold value α is pre-defined to decide the alignment confidence, i.e., pairs with s i,j > α are considered similar. Then we put on the CL loss:\nL CMA = - i∈LSG x , j * ∈VSG log exp(si,j * /τ ) Z ,(3)\nZ = i∈LSG x , k∈VSG, k̸ =j * exp(s i,k /τ ) ,(4)\nwhere τ >0 is an annealing factor. j * means a positive pair with i, i.e., s i,j * >α.\nCross-modal Cross-reconstruction We further strengthen the correspondence between VSG and LSG via cross-modal cross-reconstruction. Specifically, we try to reconstruct the input sentence from the VSG, and the image representations from the LSG. In this way we force both two SGs to focus on the VL-shared parts. To realize VSG→x we employ the aforementioned graph-to-text generator.\nFor LSG→z, we use the graph-to-image generator (Johnson et al., 2018). The learning loss can be marked as L REC ."
},
{
"figure_ref": [],
"heading": "SG-pivoted Back-translation Training",
"publication_ref": [
"b29",
"b13"
],
"table_ref": [],
"text": "Back-translation is a key method to realize unsupervised machine translation (Sennrich et al., 2016).\nIn this work, we further aid the back-translation with structural SG pivoting.\nVisual-concomitant Back-translation We perform the back-translation with the SG pivoting. We denote the X →Y translation direction as y=F xz→y (x, z), and Y→Z as x=F yz→x (y, z).\nAs we only have src-side sentences, the backtranslation is uni-directional, i.e., x→ȳ→x.\nL VCB = E[-log p yz→x (x|F xz→y (x, z), z)] . (5)\nCaptioning-pivoted Back-translation Image captioning is partially similar to MMT besides the non-text part of the input. Inspired by Huang et al. (2020), based on the SG pivoting, we incorporate two captioning procedures, Z→X and Z→Y, to generate pseudo parallel sentences <x-ȳ> for back-translation and better align the language latent spaces. We denote Z→X as x=C z→x (z), Z→Y as ȳ=C z→y (z). The back-translation loss will be:\nL CPB = E[-log p(x|F xz→y (x, z), z)] + E[-log p(ȳ|F yz→x (ȳ, z), z)] .(6)\n⋆ Remarks In the initial stage, each of the above learning objectives will be executed separately, in a certain order, so as to maintain a stable and effective UMMT system. We first perform L CMA and L REC , because the cross-SG visual-language learning is responsible for aligning the VL SGs, based on which the high-level translation can happen. Then we perform back-translation training L VCB and L CPB , together with VSH updating L VSH .\nOnce the system tends to converge, we put them all together for further fine-tuning:\nL = L CMA + L REC + L VCB + L CPB + L VSH . (7)\n5 Experiments"
},
{
"figure_ref": [],
"heading": "Setups",
"publication_ref": [
"b8",
"b3",
"b4",
"b20",
"b26",
"b27",
"b40",
"b0",
"b28",
"b17",
"b33",
"b5",
"b31",
"b13",
"b9",
"b41",
"b9",
"b20"
],
"table_ref": [],
"text": "The experiments are carried out on Multi30K data (Elliott et al., 2016), a benchmark for MMT, where each image comes with three parallel descriptions in English/German/French. Following Huang et al.\n(2020), we mainly consider the English-French (En↔Fr) and English-German (En↔De). For each translation direction, we only use the src sentence & img for training, and only the src sentence for testing. We also test on the WMT16 En→Ro and WMT14 En→De, En→Fr. WMT (Bojar et al., 2014(Bojar et al., , 2016) ) is widely-used text-only translation corpora, where following Li et al. (2022), we use CLIP (Radford et al., 2021) to retrieve images from Multi30K for sentences.\nFollowing prior research, we employ the Faster-RCNN (Ren et al., 2015) as an object detector, and MOTIFS (Zellers et al., 2018) as a relation classifier and an attribute classifier, where these three together form a VSG generator. For LSG generation, we convert the sentences into dependency trees with a parser (Anderson et al., 2018), which is then transformed into the scene graph based on certain rules (Schuster et al., 2015). For text preprocessing, we use Moses (Koehn et al., 2007) for tokenization and apply the byte pair encoding (BPE) technique. We use Transformer (Vaswani et al., 2017) as the underlying text-encoder to offer representations for GCN, and use the FasterRCNN to encode visual feature representations. All GCN encoders and other feature embeddings have the same dimension of 1,024, and all GCN encoders are with two layers.\nWe mainly compare with the existing UMMT models: Game-MMT (Chen et al., 2018), UMMT (Su et al., 2019) and PVP (Huang et al., 2020).\nTo achieve a fair comparison on the inference-time image-free setup, we also re-implement the UMMT and PVP by integrating the phrase-level retrievalbased visual hallucination method (Fang and Feng, 2022). All models use the same fair configurations, and we do not use pre-trained LM. On WMT we also test the supervised MMT setup, where we use these baselines: UVR (Zhang et al., 2020), RMMT (Wu et al., 2021b), PUVR (Fang and Feng, 2022) and VALHALLA (Li et al., 2022). We report the BLEU and METEOR scores for model evaluation.\nOur results are computed with a model averaging over 5 latest checkpoints with significance test. Our experiments are based on the NVIDIA A100 Tensor Core GPUs."
},
{
"figure_ref": [],
"heading": "Main Results",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Results on Multi30K",
"publication_ref": [
"b29",
"b13"
],
"table_ref": [
"tab_3",
"tab_4"
],
"text": "In Table 2 we show the overall results on Multi30K data. First, we inspect the performance where gold-paired images are given as inputs for testing. We see that our method (Ours # ), by integrating the LSG and VSG information, shows clear superiority over baselines on all translation jobs, while ablating the SGs, the performance drops rapidly. This shows the importance of leveraging scene graphs for more effective Ablation Study In Table 3 we quantify the contribution of each objective of scene graph pivoting learning via ablation study. Each learning strategy exhibits considerable impacts on the overall performance, where the captioning-pivoted (man in shirt and pants playing football.) mann in t-shirt und shorts tritt fußball vom tee.\n(man in t-shirt and shorts kicks football off the tee.) (two bicycles stand behind two people sitting on the grass near a body of water.) back-translation influences the results the biggest, with an average 4.3 BLEU score. Overall, two SG-pivoted back-translation training targets show much higher influences than the two cross-SG visual-language learning objectives. When removing both two back-translation targets, we witness the most dramatic decrease, i.e., average -5.7 BLEU. This validates the long-standing finding that the back-translation mechanism is key to unsupervised translation (Sennrich et al., 2016;Huang et al., 2020). Table 6: Vision-language aligning evaluation. For our models, we transform the hallucinated VSG into an image via a graph-to-image generator. We use CLIP to measure the VL relevance score."
},
{
"figure_ref": [],
"heading": "Results on WMT",
"publication_ref": [],
"table_ref": [],
"text": "unsupervised MMT. We can find that our unsupervised method only loses within 1 BLEU score to supervised models, e.g., UVR and PUVR."
},
{
"figure_ref": [
"fig_6",
"fig_6",
"fig_7"
],
"heading": "Further Analyses and Discussions",
"publication_ref": [
"b24",
"b20"
],
"table_ref": [
"tab_7",
"tab_7",
"tab_7"
],
"text": "In this part we try to dive deep into the model, presenting in-depth analyses to reveal what and how our proposed method really works and improves.\n• Integration of the vision and language SGs helps gain a holistic understanding of input.\nBoth VSG and LSG advance in comprehensively depicting the intrinsic structure of the content semantics, which ensures a holistic understanding of the input texts and images. By encoding the vision and language SGs, it is expected to completely capture the key components from src inputs, and thus achieve better translations. However, without such structural features, some information may be lost during the translation. In Table 5 via human evalua- tion we can see that our system obtains significantly higher scores in terms of the completeness, comparing to those baselines without considering SGs. Also in Fig. 5, we can find that the baseline system PVP * (PR), with only the local-level phrase-level visual retrieval, has frequently missed the key entities during the translation, e.g., the object 'tee' in case#2.\n• SG-based multimodal feature modeling helps achieve more accurate alignment between vision and language. Another merit to integrating the SGs is that the fine-grained graph modeling of visual and language scenes obviously aids more precise multimodal feature alignment. In this way, the translated texts have higher fidelity to the original texts. Inaccurate multimodal alignment without considering the SG modeling will otherwise lead to worse ambiguity. Observing the ambiguity in Table 5, we see that our model exhibits the lowest ambiguity. In Fig. 5 for the case#3, PVP * (PR) confuses the verb 'saw' as 'see' as it fails to accurately refer 'saw' to a certain lumbering tool, while ours gives a correct prediction. Besides, accurate multimodal alignment greatly enhances the utility of visual information. In Table 6 we compare the relevance of vision-language counterparts by different models, where our model gives the highest performance on both the overall text-image matching and the regional phrase-object matching. In addition, two proposed cross-SG learning targets display big impacts on the VL-aligning ability.\n• The longer and more complex the sentences, the higher the translation quality benefiting from the SGs features. In this work, we investigate the SG structures to model the input texts.\nGraph modeling of the texts has proven effective for resolving the long-range dependency issue (Marcheggiani and Titov, 2017;Li et al., 2022). In Fig. 6 we group the translation performance based on the lengths of source sentences. We see that our SG-based model gives very considerable gains over the two non-SG baselines, where the longer the sentences the higher the improvements.\n• Incorporating SGs into MMT advances in more fluent translation. Also, modeling the semantic scene graph of the input features contributes a lot to the language fluency of the translation texts.\nLooking at the Fluency item in Table 5, we find that our system gives the best fluency with the lowest grammar errors.\n• SG-based visual scene hallucination mechanism helps gain rich and correct visual features. Different from the baseline retrieval-based methods that directly obtain the whole images (or local regions), our proposed VSH mechanism instead compensatively generates the VSGs from the given LSGs. In this way, the hallucinated visual features enjoy two-fold advantages. On the one hand, the pseudo VSG has high correspondence with the textual one, both of which will enhance the shared feature learning between the two modalities. On the other hand, the hallucinated VSG will produce some vision-specific scene components and structures, providing additional clues to facilitate back to the textual features for overall better semantic understanding. Fig. 7 illustrates the node increasing rate during the vision scene graph hallucination. We see that the numbers of all three types of nodes increase, to different extents, where object nodes grow rapidest. Also, during the two transition steps of the VSH mechanism we get two VSGs, skeleton VSG and hallucinated VSG. From Fig. 8 we see that after two full hallucination steps, we can obtain high-fidelity vision features, demonstrating the necessity of the second completing-vision step."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "We investigate an inference-time image-free setup in unsupervised multimodal machine translation.\nIn specific, we integrate the visual and language scene graph to learn the fine-grained visionlanguage representations. "
},
{
"figure_ref": [],
"heading": "A Appendix",
"publication_ref": [],
"table_ref": [],
"text": "In §3.2 we give a brief induction to the overall model framework. Here we extend the details of each module of the scene graph-based multimodal translation backbone. In Fig. 9 we outline our framework."
},
{
"figure_ref": [
"fig_3",
"fig_9"
],
"heading": "A.1 Visual Scene Hallucination Learning Module",
"publication_ref": [
"b34"
],
"table_ref": [],
"text": "First of all, we note that VSH only will be activated to produce VSG hallucination at inference time. During the training phase, we construct the VSG vocabularies of different VSG nodes. We denote the object vocabulary as D o , which caches the object nodes from parsed VSG of training images; denote the attribute vocabulary as D a , which caches the attribute nodes; and denote the relation vocabulary as D r , which caches the relation nodes.\nThose vocabularies will be used to provide basic ingredients for VSG hallucination. At inference time, VSH is activated to perform two-step inference to generate the hallucinated VSG ′ . The process is illustrated in Fig. 3.\nStep1: Sketching Skeleton This step builds the skeleton VSG from the raw LSG. Specifically, we only need to transform the textual entity nodes into the visual object nodes, while keeping unchanged the whole graph topology. As for the attribute nodes and the relation nodes, we directly copy them into the VSG, as they are all text-based labels that are applicable in VSG. Then we transform the textual entity nodes into the visual object nodes. For each textual entity node in LSG, we employ the edges, which is illustrated in Fig. 11.\n▶ For the node augmentor, we first traverse all the object nodes in the skeleton VSG. For each object node v i , we then perform k-order routing over its neighbor nodes. We denote its neighbor nodes as\nV na i = {• • • , v k , • • • }.\nThen we use the attention to learn the neighbor influence to v i , and obtain the k-order feature representation h i of v i :\nα n k = exp r i • r k v * k ∈V na i exp r i • r * k h na i = r i + k α n k • r k .\nwhere r i and r k is the node representations of v i and v k , which are obtained from GCN encoder.\nThen we use a classifier to make prediction over the total vocabularies of D o and D a , to determine which node v′ i (either an object or an attribute node) should be attached to v i , if any: (FFN([h na i ; r i )) . ▶ For the relation augmentor, we first traverse all the node-pairs (object or attribute nodes, excluding the relation nodes) in the VSG, i.e., v i &v j . Then, for each node in the pair we use a triaffine attention (Wang et al., 2019;Wu et al., 2021a) to directly determine which new relation type ê′ i,j should be built between them, if exists:\nv′ i ← Softmax D na (FFN(h na i )) , where D na = D o ∪ D a ∪\nh pa i-j = Sigmoid( r i 1 T (r j ) T W r i-j 1 ) , ê′ i,j ← Softmax D pa (FFN(h pa i-j )) ,\nwhere D pa = D r ∪ {ϵ}, where the dummy token ϵ indicates no new edge should be created between two nodes. The new edge ê′ i,j has a relation label. r i-j is the representation of the path from v i to v j , which is obtained by the pooling function over all the nodes in the path:\nh pa i-j = Pool(r i , • • • , r j ) . Note that the triaffine scorer is effective in modeling the high-order ternary relations, which will provide a precise determination on whether to add a new edge.\nDuring training, the node augmentor and the relation augmentor are trained and updated based on the gold LSG and VSG, to learn the correct mapping between LSG and VSG. Such supervised learning is also important for ensuring that the final generated hallucinated visual scenes are basically coincident with the caption text, instead of random or groundless vision scenes."
},
{
"figure_ref": [],
"heading": "A.2 SG Fusing&Mapping Module",
"publication_ref": [],
"table_ref": [],
"text": "Here we extend the contents in § 3.2. As shown in Fig. 9, first of all, the SG fusing module merges the LSG x and VSG into a mixed cross-modal scene graph, such that the merged scene graph are highly compact with less redundant. Before the merging, we first measure the similarity of each pair of <text-img> node representations via cosine distance:\ns f i,j = (r L i ) T • r V j ||r L i || ||r V j ||\n. This is a similar process as in Eq. ( 2). For those pairs with high alignment scores, i.e., s i,j > α (we use the same pre-defined threshold as in crossmodal alignment learning), we consider them as serving a similar role. Since we will perform the cross-modal SG aligning learning L CMA , the accuracy of the alignment between LSG x and VSG can be guaranteed. Then, we average the representations of the image-text node pair from their GCNs. And for the rest of nodes in LSG x and VSG, we take the union structures of them. The resulting mixed SG fully inherits the semantic-rich scene nodes from both the textual SG and the visual SG, which will benefit the following text generation. Now we treat the mixed SG as a pseudo tgt-side LSG y . We use another GCN to model LSG y for further feature propagation: r y 1 , • • • , r y m = GCN(V SG y ) . The initial node representations of GCN are from the GCNs of VSG and LSG x , i.e., r L and r V as in Eq. (1). Based on the node representation r y i of VSG y , we finally employ a graph-to-text model2 to generate the final tgt-side sentence. Specifically, all the node representation r i will be first summarized into one unified graph-level feature via pooling: r y = Pool(r y 1 , • • • , r y m ) . Then, an autoregressive sequential decoder (Se-qDec) will take r y to generate tgt-side token over the tgt-side vocabulary at each setp, sequentially: e i = SeqDec(e ≤i , r y ) , ŷi ← Softmax(e i ) ."
},
{
"figure_ref": [],
"heading": "Acknowledgments",
"publication_ref": [],
"table_ref": [],
"text": "This research is supported by the National Natural Science Foundation of China (No. 62176180), and also the Sea-NExT Joint Lab."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [],
"table_ref": [],
"text": "Our paper has the following potential limitations. First of all, we take advantage of the external scene graph structures to achieve the inference-time visual hallucination and secure significant improvements of the target task, while it could be a doubleedged sword. This makes our method subject to the quality of the external structure parsers. When the parsed structures of visual scene graphs and language scene graphs are with much noise, it will deteriorate our methods. Fortunately, the existing scene graph parsers have already achieved satisfactory performance for the majority language (e.g., English), which can meet our demands. Second, the effectiveness of our approach depends on the availability of good-quality images, which however shares the pitfalls associated with the standard unsupervised multimodal translation setup."
},
{
"figure_ref": [],
"heading": "Node augmentor",
"publication_ref": [],
"table_ref": [],
"text": "Skeleton VSG CLIP tool 1 to search for the best matching visual node (proposal) in D o as the counterpart visual object, resulting in the skeleton VSG. After this step, we obtain the sketch structure of the target VSG.\nStep2: Completing Vision This step completes the skeleton VSG into a more realistic one, i.e., the final hallucinated VSG ′ . With the skeleton VSG at hand, we aim to further enrich skeleton VSG. Because intuitively, in actual world the visual scenes are always much more concrete and vivid than textual scenes. For example, given a caption text 'boys are playing baseball on playground', the LSG only mentions 'boys', 'baseball' and 'playground' objects. But imaginarily, there must be a 'baseball bat' in the scene of vision, and also both the pairs of 'boys'-'playground' and 'baseball'-'playground' has 'on' relation. Thus it is indispensable to add new nodes and more edges, i.e., scene graph augmentation. To reach the goal, we propose a node augmentor and a relation augmentor, as shown in Fig. 10. First of all, we downgrade all the relation nodes as the edge itself, i.e., an edge with a relation label. By this, we obtain a VSG that only contains object and attribute nodes, and labeled 1 https://github.com/openai/CLIP"
}
] | In this work, we investigate a more realistic unsupervised multimodal machine translation (UMMT) setup, inference-time image-free UMMT, where the model is trained with sourcetext image pairs, and tested with only sourcetext inputs. First, we represent the input images and texts with the visual and language scene graphs (SG), where such fine-grained visionlanguage features ensure a holistic understanding of the semantics. To enable pure-text input during inference, we devise a visual scene hallucination mechanism that dynamically generates pseudo visual SG from the given textual SG. Several SG-pivoting based learning objectives are introduced for unsupervised translation training. On the benchmark Multi30K data, our SG-based method outperforms the best-performing baseline by significant BLEU scores on the task and setup, helping yield translations with better completeness, relevance and fluency without relying on paired images. Further in-depth analyses reveal how our model advances in the task setting. | Scene Graph as Pivoting: Inference-time Image-free Unsupervised Multimodal Machine Translation with Visual Scene Hallucination | [
{
"figure_caption": "✓• Unsupervised MMT Chen et al. (2018) Su et al. (2019)Huang et al. (2020) ",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 1 :1Figure 1: Representing the texts and images via language scene graphs (LSG) and visual scene graphs (VSG). In a scene graph, object, attribute, relation nodes are shown in green, orange, purple respectively.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: The high-level overview of our SG-based UMMT model. During training, src-side sentences with paired images are used as inputs, together with the corresponding LSG and VSG. Testing phase only takes src-side sentences, where the visual hallucination module is activated to generate VSG from text sources.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: The illustration of the visual scene hallucination (VSH) module, including two steps of inference.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: Illustrations of the learning strategies for unsupervised multimodal machine translation.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "two bicycles stand behind two people sitting on the grass near a body of water. hinter zwei Personen, die auf dem gras in der nähe eines gewässers sitzen. stand behind two man with the herbaceous plants near the ocean.)zwei fahrräder stehen hinter zwei mann mit den eingetopften graspflanzen in der nähe des meeres.man in t-shirt and shorts kicking football off tee. und hose, der fußball spielt.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 :5Figure 5: Qualitative results of inference-time image-free UMMT (En→De).",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Figure 6 :6Figure 6: BLEU scores under different sentence lengths.",
"figure_data": "",
"figure_id": "fig_7",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "Figure 7 :Figure 8 :78Figure 7: Growing rate of nodes in hallucinated VSG.",
"figure_data": "",
"figure_id": "fig_8",
"figure_label": "78",
"figure_type": "figure"
},
{
"figure_caption": "Figure 11 :11Figure 11: Degeneration of the relation node to the labeled edge.",
"figure_data": "",
"figure_id": "fig_9",
"figure_label": "11",
"figure_type": "figure"
},
{
"figure_caption": "|V SG ← LSG) , L VSH = L N A + L P A .",
"figure_data": "",
"figure_id": "fig_10",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Practical unsupervised MMT requires the avoidance of not only parallel sentences during training, but also the paired image during inference (testing).",
"figure_data": "",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Results of UMMT on Multi30K data. 'Ours # ': using paired images for testing instead of visual hallucination. 'UMMT * /PVP * ': re-implemented baselines with phrase-level retrieval-based visual hallucination. In the brackets are the improvements of our model over the best-performing baseline(s).",
"figure_data": "En→Fr En←Fr En→De En←De Avg.Ours50.645.532.033.6 40.4-L CMA49.244.330.932.6 39.3(-1.1)-L REC48.743.930.332.1 38.8(-1.6)-L VCB47.042.228.730.1 37.0(-3.4)-L CPB45.941.627.629.2 36.1(-4.3)-L CMA &L REC 47.242.529.230.9 37.5(-2.9)-L CPB &L VCB 44.640.026.327.7 34.7(-5.7)",
"figure_id": "tab_3",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Ablating different learning strategies.",
"figure_data": "",
"figure_id": "tab_4",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Human evaluations are rated on a Likert 10scale, where the results are averaged on En→De and De→En. PVP",
"figure_data": "",
"figure_id": "tab_7",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "Table 4 further compares the translation results on WMT corpora under supervised/unsupervised MMT. It is unsurprising to see that MMT models trained with supervision from parallel sentences are overall better than the unsupervised ones. However, our UMMT system effectively narrows the gap between supervised and",
"figure_data": "67.4±6.8-PVP * (PR)-88.9±5.4Ours86.8±4.791.4±3.8-L CMA76.5±5.580.3±4.3-L REC70.1±5.277.5±4.0-L CMA &L REC68.6±6.172.8±4.8",
"figure_id": "tab_8",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "{ϵ}, including an additional dummy token ϵ indicating no new node to be attached to v i . And if the predicted node is an object node, an additional relation classifier will determine what is the relation label ê′ between v′",
"figure_data": "iand v i :ê′ ← Softmax D r",
"figure_id": "tab_10",
"figure_label": "",
"figure_type": "table"
}
] | Hao Fei; Qian Liu; Meishan Zhang; Min Zhang; Tat-Seng Chua | [
{
"authors": "Peter Anderson; Xiaodong He; Chris Buehler; Damien Teney; Mark Johnson; Stephen Gould; Lei Zhang",
"journal": "",
"ref_id": "b0",
"title": "Bottom-up and top-down attention for image captioning and visual question answering",
"year": "2018"
},
{
"authors": "Mikel Artetxe; Gorka Labaka; Eneko Agirre; Kyunghyun Cho",
"journal": "",
"ref_id": "b1",
"title": "Unsupervised neural machine translation",
"year": "2018"
},
{
"authors": "Dzmitry Bahdanau; Kyunghyun Cho; Yoshua Bengio",
"journal": "",
"ref_id": "b2",
"title": "Neural machine translation by jointly learning to align and translate",
"year": "2015"
},
{
"authors": "Ondřej Bojar; Christian Buck; Christian Federmann; Barry Haddow; Philipp Koehn; Johannes Leveling; Christof Monz; Pavel Pecina; Matt Post; Herve Saint-Amand; Radu Soricut; Lucia Specia; Aleš Tamchyna",
"journal": "",
"ref_id": "b3",
"title": "Findings of the 2014 workshop on statistical machine translation",
"year": "2014"
},
{
"authors": "Ondřej Bojar; Rajen Chatterjee; Christian Federmann; Yvette Graham; Barry Haddow; Matthias Huck; Antonio Jimeno Yepes; Philipp Koehn; Varvara Logacheva; Christof Monz; Matteo Negri; Aurélie Névéol; Mariana Neves; Martin Popel; Matt Post; Raphael Rubino; Carolina Scarton; Lucia Specia; Marco Turchi; Karin Verspoor; Marcos Zampieri",
"journal": "",
"ref_id": "b4",
"title": "Findings of the 2016 conference on machine translation",
"year": "2016"
},
{
"authors": "Yun Chen; Yang Liu; O K Victor; Li",
"journal": "",
"ref_id": "b5",
"title": "Zeroresource neural machine translation with multi-agent communication game",
"year": "2018"
},
{
"authors": "Alexis Conneau; Guillaume Lample",
"journal": "",
"ref_id": "b6",
"title": "Crosslingual language model pretraining",
"year": "2019"
},
{
"authors": "Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova",
"journal": "",
"ref_id": "b7",
"title": "BERT: Pre-training of deep bidirectional transformers for language understanding",
"year": "2019"
},
{
"authors": "Desmond Elliott; Stella Frank; Khalil Sima'an; Lucia Specia",
"journal": "",
"ref_id": "b8",
"title": "Multi30K: Multilingual English-German image descriptions",
"year": "2016"
},
{
"authors": "Qingkai Fang; Yang Feng",
"journal": "",
"ref_id": "b9",
"title": "Neural machine translation with phrase-level universal visual representations",
"year": "2022"
},
{
"authors": "Shengqiong Hao Fei; Yafeng Wu; Meishan Ren; Zhang",
"journal": "",
"ref_id": "b10",
"title": "Matching structure for dual learning",
"year": "2022"
},
{
"authors": "Chengyu Huang; Zheng Zhang; Hao Fei; Lizi Liao",
"journal": "",
"ref_id": "b11",
"title": "Conversation disentanglement with bi-level contrastive learning",
"year": "2022"
},
{
"authors": "Po-Yao Huang; Xiaojun Chang; Alexander Hauptmann",
"journal": "",
"ref_id": "b12",
"title": "Multi-head attention with diversity for learning grounded multilingual multimodal representations",
"year": "2019"
},
{
"authors": "Po-Yao Huang; Junjie Hu; Xiaojun Chang; Alexander Hauptmann",
"journal": "",
"ref_id": "b13",
"title": "Unsupervised multimodal neural machine translation with pseudo visual pivoting",
"year": "2020"
},
{
"authors": "Po-Yao Huang; Frederick Liu; Sz-Rung Shiang; Jean Oh; Chris Dyer",
"journal": "",
"ref_id": "b14",
"title": "Attention-based multimodal neural machine translation",
"year": "2016"
},
{
"authors": "Justin Johnson; Agrim Gupta; Li Fei-Fei",
"journal": "",
"ref_id": "b15",
"title": "Image generation from scene graphs",
"year": "2018"
},
{
"authors": "Justin Johnson; Ranjay Krishna; Michael Stark; Li-Jia Li; David A Shamma; Michael S Bernstein; Li Fei-Fei",
"journal": "",
"ref_id": "b16",
"title": "Image retrieval using scene graphs",
"year": "2015"
},
{
"authors": "Philipp Koehn; Hieu Hoang; Alexandra Birch; Chris Callison-Burch; Marcello Federico; Nicola Bertoldi; Brooke Cowan; Wade Shen; Christine Moran; Richard Zens; Chris Dyer; Ondřej Bojar; Alexandra Constantin; Evan Herbst",
"journal": "",
"ref_id": "b17",
"title": "Moses: Open source toolkit for statistical machine translation",
"year": "2007"
},
{
"authors": "Ranjay Krishna; Yuke Zhu; Oliver Groth; Justin Johnson; Kenji Hata; Joshua Kravitz; Stephanie Chen; Yannis Kalantidis; Li-Jia Li; David A Shamma; Michael S Bernstein; Li Fei-Fei",
"journal": "International Journal of Computer Vision",
"ref_id": "b18",
"title": "Visual genome: Connecting language and vision using crowdsourced dense image annotations",
"year": "2017"
},
{
"authors": "Guillaume Lample; Alexis Conneau; Ludovic Denoyer; Marc'aurelio Ranzato",
"journal": "",
"ref_id": "b19",
"title": "Unsupervised machine translation using monolingual corpora only",
"year": "2018"
},
{
"authors": "Yi Li; Rameswar Panda; Yoon Kim; Chun-Fu Richard Chen; Rogério Feris; David D Cox; Nuno Vasconcelos",
"journal": "",
"ref_id": "b20",
"title": "VALHALLA: visual hallucination for machine translation",
"year": "2022"
},
{
"authors": "Zuchao Li; Hai Zhao; Rui Wang; Masao Utiyama; Eiichiro Sumita",
"journal": "",
"ref_id": "b21",
"title": "Reference language based unsupervised neural machine translation",
"year": "2020"
},
{
"authors": "Lajanugen Logeswaran; Honglak Lee",
"journal": "",
"ref_id": "b22",
"title": "An efficient framework for learning sentence representations",
"year": "2018"
},
{
"authors": "Thang Luong; Hieu Pham; Christopher D Manning",
"journal": "",
"ref_id": "b23",
"title": "Effective approaches to attention-based neural machine translation",
"year": "2015"
},
{
"authors": "Diego Marcheggiani; Ivan Titov",
"journal": "",
"ref_id": "b24",
"title": "Encoding sentences with graph convolutional networks for semantic role labeling",
"year": "2017"
},
{
"authors": "Hideki Nakayama; Noriki Nishida",
"journal": "Machine Translation",
"ref_id": "b25",
"title": "Zeroresource machine translation by multimodal encoderdecoder network with multimedia pivot",
"year": "2017"
},
{
"authors": "Alec Radford; Jong Wook Kim; Chris Hallacy; Aditya Ramesh; Gabriel Goh; Sandhini Agarwal; Girish Sastry; Amanda Askell; Pamela Mishkin; Jack Clark; Gretchen Krueger; Ilya Sutskever",
"journal": "",
"ref_id": "b26",
"title": "Learning transferable visual models from natural language supervision",
"year": "2021"
},
{
"authors": "Kaiming Shaoqing Ren; Ross B He; Jian Girshick; Sun",
"journal": "",
"ref_id": "b27",
"title": "Faster R-CNN: towards real-time object detection with region proposal networks",
"year": "2015"
},
{
"authors": "Sebastian Schuster; Ranjay Krishna; Angel Chang; Li Fei-Fei; Christopher D Manning",
"journal": "",
"ref_id": "b28",
"title": "Generating semantically precise scene graphs from textual descriptions for improved image retrieval",
"year": "2015"
},
{
"authors": "Rico Sennrich; Barry Haddow; Alexandra Birch",
"journal": "",
"ref_id": "b29",
"title": "Improving neural machine translation models with monolingual data",
"year": "2016"
},
{
"authors": "Lucia Specia; Stella Frank; Khalil Sima'an; Desmond Elliott",
"journal": "",
"ref_id": "b30",
"title": "A shared task on multimodal machine translation and crosslingual image description",
"year": "2016"
},
{
"authors": "Yuanhang Su; Kai Fan; C.-C. Jay Nguyen Bach; Fei Kuo; Huang",
"journal": "",
"ref_id": "b31",
"title": "Unsupervised multi-modal neural machine translation",
"year": "2019"
},
{
"authors": "Ilya Sutskever; Oriol Vinyals; Quoc V Le",
"journal": "",
"ref_id": "b32",
"title": "Sequence to sequence learning with neural networks",
"year": "2014"
},
{
"authors": "Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Lukasz Kaiser; Illia Polosukhin",
"journal": "",
"ref_id": "b33",
"title": "Attention is all you need",
"year": "2017"
},
{
"authors": "Xinyu Wang; Jingxian Huang; Kewei Tu",
"journal": "",
"ref_id": "b34",
"title": "Second-order semantic dependency parsing with endto-end neural networks",
"year": "2019"
},
{
"authors": "Yu-Siang Wang; Chenxi Liu; Xiaohui Zeng; Alan Yuille",
"journal": "",
"ref_id": "b35",
"title": "Scene graph parsing as dependency parsing",
"year": "2018"
},
{
"authors": "Shengqiong Wu; Hao Fei; Yafeng Ren; Donghong Ji; Jingye Li; ; ",
"journal": "",
"ref_id": "b36",
"title": "Learn from syntax: Improving pair-wise aspect and opinion terms extraction with rich syntactic knowledge",
"year": "2021"
},
{
"authors": "Zhiyong Wu; Lingpeng Kong; Wei Bi; Xiang Li; Ben Kao",
"journal": "",
"ref_id": "b37",
"title": "Good for misconceived reasons: An empirical revisiting on the need for visual context in multimodal machine translation",
"year": "2021"
},
{
"authors": "Yuanmeng Yan; Rumei Li; Sirui Wang; Fuzheng Zhang; Wei Wu; Weiran Xu",
"journal": "",
"ref_id": "b38",
"title": "ConSERT: A contrastive framework for self-supervised sentence representation transfer",
"year": "2021"
},
{
"authors": "Xu Yang; Kaihua Tang; Hanwang Zhang; Jianfei Cai",
"journal": "",
"ref_id": "b39",
"title": "Auto-encoding scene graphs for image captioning",
"year": "2019"
},
{
"authors": "Rowan Zellers; Mark Yatskar; Sam Thomson; Yejin Choi",
"journal": "",
"ref_id": "b40",
"title": "Neural motifs: Scene graph parsing with global context",
"year": "2018"
},
{
"authors": "Zhuosheng Zhang; Kehai Chen; Rui Wang; Masao Utiyama; Eiichiro Sumita; Zuchao Li; Hai Zhao",
"journal": "",
"ref_id": "b41",
"title": "Neural machine translation with universal visual representation",
"year": "2020"
},
{
"authors": "Jinhua Zhu; Yingce Xia; Lijun Wu; Di He; Tao Qin; Wengang Zhou; Houqiang Li; Tie-Yan Liu",
"journal": "",
"ref_id": "b42",
"title": "Incorporating BERT into neural machine translation",
"year": "2020"
}
] | [
{
"formula_coordinates": [
4,
126.28,
335.6,
163.59,
11.11
],
"formula_id": "formula_0",
"formula_text": "r 1 , • • • , r n = GCN(G) ,(1)"
},
{
"formula_coordinates": [
5,
141.85,
137.39,
147.88,
24.09
],
"formula_id": "formula_1",
"formula_text": "si,j = (r L i ) T • r V j ||r L i || ||r V j || .(2)"
},
{
"formula_coordinates": [
5,
93.7,
209.69,
196.03,
26.75
],
"formula_id": "formula_2",
"formula_text": "L CMA = - i∈LSG x , j * ∈VSG log exp(si,j * /τ ) Z ,(3)"
},
{
"formula_coordinates": [
5,
106.76,
242.15,
182.98,
20.94
],
"formula_id": "formula_3",
"formula_text": "Z = i∈LSG x , k∈VSG, k̸ =j * exp(s i,k /τ ) ,(4)"
},
{
"formula_coordinates": [
5,
77.21,
603.63,
212.65,
13.09
],
"formula_id": "formula_4",
"formula_text": "L VCB = E[-log p yz→x (x|F xz→y (x, z), z)] . (5)"
},
{
"formula_coordinates": [
5,
95.95,
744.15,
193.92,
29.21
],
"formula_id": "formula_5",
"formula_text": "L CPB = E[-log p(x|F xz→y (x, z), z)] + E[-log p(ȳ|F yz→x (ȳ, z), z)] .(6)"
},
{
"formula_coordinates": [
5,
313.82,
225.41,
211.32,
10.59
],
"formula_id": "formula_6",
"formula_text": "L = L CMA + L REC + L VCB + L CPB + L VSH . (7)"
},
{
"formula_coordinates": [
12,
347.44,
663.13,
99.3,
14
],
"formula_id": "formula_7",
"formula_text": "V na i = {• • • , v k , • • • }."
},
{
"formula_coordinates": [
12,
351.92,
702.34,
123.17,
58.75
],
"formula_id": "formula_8",
"formula_text": "α n k = exp r i • r k v * k ∈V na i exp r i • r * k h na i = r i + k α n k • r k ."
},
{
"formula_coordinates": [
13,
70.47,
141.62,
170.55,
31.94
],
"formula_id": "formula_9",
"formula_text": "v′ i ← Softmax D na (FFN(h na i )) , where D na = D o ∪ D a ∪"
},
{
"formula_coordinates": [
13,
76.82,
359.33,
206.36,
54.45
],
"formula_id": "formula_10",
"formula_text": "h pa i-j = Sigmoid( r i 1 T (r j ) T W r i-j 1 ) , ê′ i,j ← Softmax D pa (FFN(h pa i-j )) ,"
},
{
"formula_coordinates": [
13,
371.09,
185.95,
82.33,
31.07
],
"formula_id": "formula_11",
"formula_text": "s f i,j = (r L i ) T • r V j ||r L i || ||r V j ||"
}
] |
|
2023-06-04 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b5",
"b37",
"b23",
"b3",
"b7",
"b26",
"b17",
"b22",
"b34",
"b32",
"b15"
],
"table_ref": [],
"text": "Relation extraction (RE) aims at extracting from the plain texts the meaningful entity mentions paired with semantic relations. One widelyacknowledged key bottleneck of RE is called the long-range dependence (LRD) issue, i.e., the decay of dependence clues of two mention entities with increasing distance in between (Culotta and Sorensen, 2004;Zhang et al., 2018;Fei et al., 2021). Fortunately, prior work extensively reveals that the syntactic dependency trees help resolve LRD issue effectively, by taking advantage of the close relevance between the dependency structure and the relational RE pair (Miwa and Bansal, 2016;Can et al., 2019). In cross-lingual RE, likewise, the universal dependency trees (de Marneffe et al., 2021) are leveraged as effective language-persistent features Copying current node from SRC or TGT trees into forest, if the node has no alignment in opposite tree at this layer.\n在In马歇尔Marshall的家乡德国城hometown Germantown他是位大胆的法官，,他 于1801年in the year of 1801被by亚当斯提名为to be首席大法官。 Synthesized code-mixed text Figure 2: A real example to construct a code-mixed UD forest. The raw sentence is selected from ACE05 data. We exemplify the transfer from English (EN) to Chinese (ZH).\nbe exacerbated in UD-based model transfer, cf. Fig. 1(a). Given that UD has a universal annotation standard, inevitably, there is still a syntax discrepancy between the two languages due to their intrinsic linguistic nature. We show (cf. §3 for more discussion) that between the parallel sentences in English and Arabic, around 30% words are misaligned and over 35% UD word-pairs have no correspondence. Such structural discrepancies consequently undermine the model transfer efficacy.\nOne alternative solution is using annotation projection (Padó and Lapata, 2009;Kim et al., 2010;McDonald et al., 2013;Xiao and Guo, 2015). The main idea is directly synthesizing the pseudo TGTside training data, so that the TGT-side linguistic features (i.e., UD trees) are well preserved. However, it could be a double side of the sword in the annotation projection paradigm. It manages to learn the language-specific features, while at the cost of losing some high-efficient structural knowledge from SRC-side UD, thus leading to the SRC-biased UD feature transfer. As illustrated in Fig. 1(b), the dependence paths in the SRC UD tree that effectively solves the LRD issues for the task are sacrificed when transforming the SRC tree into the TGT tree. This motivates us to pursue an unbiased and holistic UD-based XRE transfer by considering both the SRC and TGT UD syntax features. To reach the goal, in this work, we propose combining the view of model transfer and annotation projection paradigm, and constructing a type of codemixed UD forests. Technically, we first project the SRC training instances and TGT predicting instances into the opposite languages, respectively.\nThen, we parse the parallel UD trees of both sides respectively via existing UD parsers. Next, merge each pair of SRC and TGT UD trees together into the code-mixed UD forest, in which the wellaligned word pairs are merged to the TGT ones in the forest, and the unaligned words will all be kept in the forest. With these code-mixed syntactic features, the gap between training and predicting phases can be closed, as depicted in Fig. 1(c).\nWe encode the UD forest with the graph attention model (GAT; Velickovic et al., 2018) for feature encoding. We perform experiments on the representative XRE benchmark, ACE05 (hristopher Walker et al., 2006), where the transfer results from English to Chinese and Arabic show that the proposed code-mixed forests bring significant improvement over the current best-performing UD-based system, obtaining the new SoTA results. Further analyses verify that 1) the code-mixed UD forests help maintain the debiased cross-lingual transfer of RE task, and 2) the larger the difference between SRC and TGT languages, the bigger the boosts offered by code-mixed forests. To our knowledge, we are the first taking the complementary advantages of annotation projection and model transfer paradigm for unbiased XRE transfer. We verify that the gap between training and predicting of UD-based XRE can be bridged by synthesizing a type of code-mixed UD forests. The resource can be found at https://github.com/ scofield7419/XLSIE/."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b4",
"b35",
"b33",
"b10",
"b21",
"b30",
"b38",
"b18",
"b25",
"b2",
"b24",
"b6",
"b20",
"b39",
"b19",
"b16",
"b1",
"b28",
"b28",
"b27",
"b36",
"b36"
],
"table_ref": [],
"text": "Different from the sequential type of information extraction (IE), e.g., named entity recognition (NER) (Cucerzan and Yarowsky, 1999), RE not only detects the mentions but also recognizes the semantic relations between mentions. RE has long received extensive research attention within the last decades (Zelenko et al., 2002). Within the community, research has revealed that the syntactic dependency trees share close correlations with RE or broad-covering information extraction tasks in structure (Fei et al., 2021;Wu et al., 2021;Fei et al., 2022), and thus the former is frequently leveraged as supporting features for enhancing RE. In XRE, the key relational features between words need to be transferred between languages, which motivates the incorporation of UD tree features that have consistent annotations and principles across various languages. Thus, UD-based systems extensively achieve the current SoTA XRE (Lu et al., 2020;Taghizadeh and Faili, 2021;Zhang et al., 2021). This work inherits the prior wisdom, and leverages the UD features.\nModel transfer (Kozhevnikov and Titov, 2013;Ni and Florian, 2019;Fei et al., 2020b) and annotation projection (Björkelund et al., 2009;Mulcaire et al., 2018;Daza and Frank, 2019;Fei et al., 2020a;Lou et al., 2022) are two mainstream avenues in structural cross-lingual transfer track. The former trains a model on SRC annotations and them make predictions with TGT instances, i.e., transferring the shared language-invariant features. The latter directly synthesizes the pseudo training instances in TGT language based on some parallel sentences, in which the TGT-specific features are retained to the largest extent. As we indicated earlier, in both two paradigms the UD tree features can be unfortunately biased during the transfer, thus leading to the underutilization of UD resource. This work considers a holistic viewpoint, integrating both the two cross-lingual transfer schemes and combining both the SRC and TGT syntax trees by code mixing.\nSeveral prior studies have shown that combining the raw SRC and pseudo TGT (from projection) data for training helps better transfer. It is shown that although the two data are semantically identical, SRC data still can offer some complementary language-biased features (Fei et al., 2020a,b;Zhen et al., 2021). Yet we emphasize that different from regular cross-lingual text classification or sequential prediction, XRE relies particularly on the syntactic structure features, e.g., UD, and thus needs a more fine-grained approach for SRC-TGT data ensembling, instead of simply instance stack-ing. Thus, we propose merging the SRC and TGT syntax trees into the code-mixed forests.\nCode mixing has been explored in several different NLP applications (Labutov and Lipson, 2014;Joshi et al., 2016;Banerjee et al., 2018;Samanta et al., 2019), where the core idea is creating data piece containing words from different languages simultaneously. For example, Samanta et al. (2019) introduce a novel data augmentation method for enhancing the recognition of code-switched sentiment analysis, where they replace the constituent phrases with code-mixed alternatives. Qin et al. (2020) propose generating code-switching data to augment the existing multilingual language models for better zero-shot cross-lingual tasks. While we notice that most of the works focus on the development of code-mixed sequential texts, this work considers the one for structural syntax trees. Our work is partially similar to Zhang et al. (2019) on the code-mixed UD tree construction. But ours differentiate theirs in that Zhang et al. (2019) target better UD parsing itself, while we aim to improve downstream tasks.\n3 Observations on UD Bias"
},
{
"figure_ref": [],
"heading": "Bias Source Analysis",
"publication_ref": [],
"table_ref": [],
"text": "As mentioned, even though UD trees define consistent annotations across languages, it still falls short on wiping all syntactic bias. This is inevitably caused by the underlying linguistic disparity deeply embedded in the language itself. Observing the linguistic discrepancies between different languages, we can summarize them into following three levels:\n1) Word-level Changes.\n• Word number. The words referring to same semantics in different languages vary, e.g., in English one single-token word may be translated in Chinese with more than one token. • Part of speech. In different languages a parallel lexicon may come with different part of speech. • Word order. Also it is a common case that the word order varies among parallel sentences in different languages.\n2) Phrase-level Change.\n• Modification type. A modifier of a phrasal constituent can be changed when translating into another languages. For example, in English, 'in the distance' is often an adverbial modifier, while its counterpart in Chinese '遥 远的' plays a role of an attribute modifier. • Change of pronouns. English grammar has strict structure, while in some other languages the grammar structures may not strict. For example, in English, it is often case to use relative pronouns (e.g., which, that, who) to refer to the prior mentions, while in other languages, such as Chinese, the personal pronouns (e.g., which, that, who) will be used to refer the prior mentions. • Constituency order change. Some constituent phrases will be reorganized and reordered from one language to another language, due to the differences in grammar rules.\n3) Sentence-level Change.\n• Transformation between active and passive sentences. In English it could be frequent to use the passive forms of sentences, while being translated into other languages the forms will be transformed into active types, where the words and phrases in the whole sentences can be reversed. • Transformation between clause and main sentence. In English the attributive clauses and noun clauses are often used as subordinate components, while they can be translated into two parallel clauses in other languages. • Change of reading order of sentences. The majority of the languages in this world have the reading order of from-left-to-right, such as English, French, etc. But some languages, e.g., under Afro-Asiatic family, Arabic, Hebrew, Persian, Sindhi and Urdu languages read from right to left."
},
{
"figure_ref": [],
"heading": "UD Bias Statistics",
"publication_ref": [],
"table_ref": [],
"text": "In Fig. 3 we present the statistics of such bias between the parallel UD trees in different languages, such as the misaligned words, mismatched UD (w ↷ i w j ) pair and UD path of (e ↷ s • • • ↷ e o ) relational pair. Fig. 3(a) reveals that languages under different families show distinct divergences. And the more different of languages, the greater the divergences (e.g., English to Arabic). Fig. 3(b) indicates that complex sentences (e.g., compound sentences) bring larger bias; and in the real world, complex sentences are much more ubiquitous than simple ones. Also, the mismatch goes worse when the UD core predicates are nouns instead of verbs. "
},
{
"figure_ref": [],
"heading": "Code-mixed UD Forest Construction",
"publication_ref": [],
"table_ref": [],
"text": "To eliminate such discrepancies for unbiased UDfeature transfer, we build the code-mixed UD forests, via the following six steps."
},
{
"figure_ref": [],
"heading": "▶",
"publication_ref": [],
"table_ref": [],
"text": "Step 1: translating a sentence x Src in SRC language to the one x Tgt in TGT language.1 This step is to generate a pseudo parallel sentence pair in both TGT and SRC languages. We accomplish this by using the state-of-the-art Google Translation API. 2 We denote the parallel sentences as <x Src ,x Tgt > or <x Src ,x Tgt >."
},
{
"figure_ref": [],
"heading": "▶",
"publication_ref": [],
"table_ref": [],
"text": "Step 2: obtaining the word alignment scores. Meanwhile, we employ the Awesome-align toolkit3 to obtain the word alignment confidence M ={m i↔j } between word pair w i ∈ x Src and w j ∈ x Tgt in parallel sentences."
},
{
"figure_ref": [],
"heading": "▶",
"publication_ref": [],
"table_ref": [],
"text": "Step 3: parsing UD trees for parallel sentences. Then, we use the UD parsers in SRC and SRC languages respectively to parse the UD syntax trees for two parallel sentences, respectively. We adopt the UDPipe 4 as our UD parsers, which are trained separately on different UD annotated data 5 . We denote the SRC UD tree as T Src , and the pseudo TGT UD tree as T Tgt . Note that the UD trees in all languages share the same dependency labels, Algorithm 1 Process of constructing code-mixed UD forests Input: T SRC , T TGT , M , threshold θ, empty forest F = Φ. Output: Code-mixed UD forest F.\n1: def Construct (T SRC , T TGT , M , F)\n▷ breadth-first top-down traverse.\n2: is_root = True ▷ a flag for traversing the predicate only once."
},
{
"figure_ref": [],
"heading": "3:",
"publication_ref": [],
"table_ref": [],
"text": "F.w cur = ROOT ▷ creating ROOT node for F. if is_root then 7: return aligned_pairs, nonaligned_nodes i.e., with the same (as much as possible) annotation standards. In Appendix §A we list the dependency labels which are the commonly occurred types.\nw merged ="
},
{
"figure_ref": [],
"heading": "▶",
"publication_ref": [],
"table_ref": [],
"text": "Step 4: projecting and merging the labels of training data. For the training set, we also need to project the annotations (relational subjectobject pairs) of sentences in SRC languages to TGT pseudo sentences. Note that this step is not needed for the testing set. The projection is based on the open source6 , during which the word alignment scores at step-2 are used. We can denote the SRC annotation as y, and the pseudo TGT label as y.\nWe then merge the annotation from both SRC and TGT viewpoints, into the code-mixed one Y , for later training use. Specifically, for the node that is kept in the final code-mixed forest, we will keep its labels; and for those nodes that are filtered, the annotations are replaced by their correspondences.\n▶ Step 5: merging the SRC and TGT UD trees into a code-mixed forest. Finally, based on the SRC UD tree and the TGT UD tree, we construct the code-mixed UD forest. We mainly perform breadth-first top-down traversal over each pair of nodes T Src and T Tgt , layer by layer. The traversal starts from their ROOT node. We first create a ROOT node as the initiation of the codemixed forest. We design two types of actions for the forest merging process:\n• Merging current pair of nodes w i ∈ T Src from SRC tree and w j ∈ T Tgt from TGT tree into the forest F, if the current two nodes are confidently aligned at same dependency layer. We check the word alignment confidence m i↔j between the two nodes, and if the confidence is above a pre-defined threshold θ, i.e., m i↔j > θ, we treat them as confidently aligned.\n• Copying current node from SRC tree T Src or TGT tree T Tgt into the forest F, once the node has no significant alignment in the opposite tree at this layer. In Algorithm 1 we formulate in detail the process of code-mixed forest construction. Also, we note that when moving the nodes from two separate UD trees into the forest, the attached dependency labels are also copied. When two nodes are merged, we only choose the label of the TGT-side node. Finally, the resulting forest F looks like code-mixing, and is structurally compact."
},
{
"figure_ref": [],
"heading": "▶",
"publication_ref": [],
"table_ref": [],
"text": "Step 6: assembling code-mixed texts. Also we need to synthesize a code-mixed text X based on the raw SRC text x Src and the pseudo TGT text x Tgt . The code-mixed text X will also be used as inputs together with the forest, into the forest encoder. We directly replace the SRC words with the TGT words that have been determined significantly aligned at Step-5."
},
{
"figure_ref": [],
"heading": "XRE with Code-mixed UD Forest",
"publication_ref": [
"b32",
"b8"
],
"table_ref": [],
"text": "Along with the UD forest F Src , we also assemble the code-mixed sequential text X Src from the SRC and translated pseudo-TGT sentences (i.e., x Src and x Tgt ), and the same for the TGT sentences X Tgt . An XRE system, being trained with SRC-side annotated data (<X Src , F Src >, Y Src ), needs to determine the label Y Tgt of relational pair e r ↷ s e o given a TGT sentence and UD forest (<X Tgt , F Tgt >).\nThe XRE system takes as input X={w i } n and F. We use the multilingual language model (MLM) for representing the input code-mixed sentence X:\nH = {h1, • • • , hn} = MLM(X) , (1\n)\nwhere X is the code-mixed sentential text. We then formulate the code-mixed forest F as a graph, G=<E, V >, where E={e i,j } n×n is the edge between word pair (i.e., initiated with e i,j =0/1, meaning dis-/connecting), V ={w i } n are the words. We main the node embeddings r i for each node v i . We adopt the GAT model (Velickovic et al., 2018) for the backbone forest encoding:\nρi,j = Softmax(GeLU(U T [W1ri; W2rj])) ,(2)\nui = σ( j ρi,jW3r 1 j ) ,(3)\nwhere W 3/4/5 and U are all trainable parameters. σ is the sigmoid function. GeLU is a Gaussian error linear activation function. Note that the firstlayer representations of r i is initialized with h i .\nH and U are then concatenated as the resulting feature representation:\nĤ = H ⊕ U .(4)\nXRE aims to determine the semantic relation labels between two given mention entities. For example, given a sentence 'John Smith works at Google', RE should identify that there is a relationship of \"works at\" between the entities \"John Smith\" and \"Google\". Our XRE model needs to predict the relation label y. We adopt the biaffine decoder (Dozat and Manning, 2017) to make prediction:\ny = Softmax(h T s • W1 • ho + W2 • Pool( Ĥ)) . (5)\nHere both h s and h o are given.\n6 Experiments"
},
{
"figure_ref": [],
"heading": "Setups",
"publication_ref": [
"b15",
"b0",
"b0"
],
"table_ref": [
"tab_1"
],
"text": "We consider the ACE05 (hristopher Walker et al., 2006) dataset, which includes English (EN), Chinese (ZH) and Arabic (AR). We give the data statistics in Table 1 The multilingual BERT is used. We use two-layer GAT for forest encoding, with a 768-d hidden size. We mainly consider the transfer from EN to one other language. Following most cross-lingual works (Fei et al., 2020b;Ahmad et al., 2021), we train the XRE model with fixed 300 iterations without early-stopping. We make comparisons between three setups: 1) using only raw SRC training data with the model transfer, 2) using only the pseudo TGT (via annotation projection) for training, and 3) using both the above SRC and TGT data. Each setting uses both the texts and UD tree (or forest) features. The baseline uses the same GAT model for syntax encoding, marked as Syn-Baseline. For setup 1)&2) we also test the transfer with only text inputs, removing the syntax features, marked as TxtBaseline. Besides, for setup 1) we cite current SoTA performances as references. We use F1 to measure the RE performance, following Ahmad et al. (2021). All experiments are undergone five times and the average value is reported."
},
{
"figure_ref": [],
"heading": "Data Inspection",
"publication_ref": [],
"table_ref": [
"tab_4"
],
"text": "We also show in Table 3 the differences in average sequential and syntactic (shortest dependency path) distances between the subjects and objects of the relational triplets. As seen, the syntactic distances between subject-object pairs are clearly shortened in the view of syntactic dependency trees, which indicates the imperative to incorporate the tree structure features. However, the syntactic distances between different languages vary, i.e., more complex languages have longer syntactic distances. Such discrepancy reflects the necessity of employing our proposed UD debiasing methods to bridge the gap."
},
{
"figure_ref": [],
"heading": "Main Results",
"publication_ref": [],
"table_ref": [
"tab_3"
],
"text": "From Table 2, we can see that UD features offer exceptional boosts (M1 vs. M2, M4 vs. M5). And annotation projection methods outperform model transfer ones (i.e., M1&M2&M3 vs. M4&M5) by offering direct TGT-side features. Interestingly, in both two transfer paradigms, the improvements from UD become weak on the language pairs with bigger divergences. For example, the improvement on EN→DE outweighs the ones on EN→ZH. Furthermore, using our proposed code-mixed syntax forests is significantly better than using standalone SRC or TGT (or the simple combination) UD features (M7 vs. M2&M5&M6) on all transfers with big margins. For example, our system outperforms SoTA UD-based systems with averaged +4.8%(=67.2-62.4) F1. This evidently verifies the necessity to create the code-mixed forests, i.e., bringing unbiased UD features for transfer. Also, we find that the more the difference between the two languages, the bigger the improvements from forests. The ablation of code-mixed texts also shows the contribution of the sequential textual features, which indirectly demonstrates the larger efficacy of the structural code-mixed UD forests."
},
{
"figure_ref": [
"fig_3"
],
"heading": "Probing Unbiasedness of Code-mixed UD Forest",
"publication_ref": [],
"table_ref": [],
"text": "Fig. 4 plots the change of the syntax distances of RE pairs during the transfer with different syntax trees. We see that the use of SRC UD trees shows clear bias (with larger inclination angles) during the transfer, while the use of TGT UD trees and codemixed forests comes with less change of syntax distances. Also, we can see from the figure that the inference paths between objects and subjects of RE tasks are clearly shortened with the forests (in orange color), compared to the uses of SRC/TGT UD trees."
},
{
"figure_ref": [],
"heading": "Change during Code-mixed UD Forest Merge",
"publication_ref": [],
"table_ref": [
"tab_6"
],
"text": "Here we make statistics of how many words are merged and kept during the UD tree merging, respectively. The statistics are shown in Table 4. We can see that the distance between EN-ZH is shorter than that between EN-AR. For example, the length of code-mixed EN-ZH UD forests (sentences) is 31.63, while for EN-AR the length is 40.44. Also, EN-ZH UD forests have a higher to 21.4% merging rate, while EN-AR UD forests have 16.6% merging rate. This demonstrates that the more divergences of languages, the lower the merging rate of the code-mixed forest."
},
{
"figure_ref": [
"fig_4"
],
"heading": "Impacts of θ on Controlling the Quality of Merged Forest",
"publication_ref": [],
"table_ref": [],
"text": "In §4 of step-5, we describe that we use a threshold θ to control the aligning during the UD tree merging. Intuitively, the large the threshold θ, the lower the alignment rate. When θ → 0, most of the SRC and TGT nodes in two parallel UD trees can find their counterparts but the alignments are most likely to be wrong, thus hurting the quality of the resulting code-mixed UD forests. When θ → 1, none of the SRC and TGT nodes in two parallel UD trees can be aligned, and both two UD trees are copied and co-existed in the resulting code-mixed UD forests. In such case, the integration of such forests is equivalent to the annotation projection methods where we directly use both the raw SRC UD feature and the translated pseudo TGT UD tree feature. In Fig. 5 we now study the influences of using different code-mixed forest features generated with different merging rates (θ). We see that with a threshold of θ=0.5, the performances are consistently the best."
},
{
"figure_ref": [],
"heading": "Performances on Different Types of Sentence",
"publication_ref": [],
"table_ref": [
"tab_7"
],
"text": "In Table 5 we show the results under different types of sentences. We directly select 500 short sentences (with length < 12) as simple sentences; and select 500 lengthy sentences (with length > 35) as complex sentences. As can be seen, with the code-mixed forest features, the system shows very notable improvements in complex sentences. For example, on the EN→ZH we obtain 15.9(=57.2-41.3)% F1 improvement, and on the EN→AR the boost increases strikingly to 25.2(=67.3-42.1)% F1. However, such enhancements are not very significant in handling simple sentences. This indicates that the code-mixed UD forest features can espe- cially enhance the effectiveness on the hard case, i.e., the transfer between those pairs with greater divergences will receive stronger enhancements from our methods."
},
{
"figure_ref": [],
"heading": "Conclusion and Future Work",
"publication_ref": [
"b14",
"b13"
],
"table_ref": [],
"text": "Universal dependencies (UD) have been served as effective language-consistent syntactic features for cross-lingual relation extraction (XRE). In this work, we reveal the intrinsic language discrepancies with respect to the UD structural annotations, which limit the utility of the UD features. We enhance the efficacy of UD features for an unbiased UD-based transfer, by constructing code-mixed UD forests from both the source and target UD trees. Experimental results demonstrate that the UD forests effectively debias the syntactic disparity in the UD-based XRE transfer, especially for those language pairs with larger gaps. Leveraging the syntactic dependency features is a long-standing practice for strengthening the performance of RE tasks. In this work, we propose a novel type of syntactic feature, code-mixed UD forests, for cross-lingual relation extraction. We note that this feature can be applied broadly to other cross-lingual structured information extraction tasks that share the same task definition besides RE, such as event detection (ED) (Halpin and Moore, 2006) and semantic role labeling (SRL) (Gildea and Jurafsky, 2000). Besides, how to fur-ther increase the utility of the UD forests with a better modeling method is a promising research direction, i.e., filtering the noisy structures in the UD forests.\n2.5-4.0 9 as well as GNU GPL 3.0 10 . Our use of UD treebanks comply with all these license terms is at non-commercial purpose. The software tools (i.e., UDPipe parsers) are provided under GNU GPL V2. Our use of UDPipe tools complies with the term."
},
{
"figure_ref": [],
"heading": "Acknowledgments",
"publication_ref": [],
"table_ref": [],
"text": "This research is supported by the National Natural Science Foundation of China (No. 62176180), and also the Sea-NExT Joint Lab."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [],
"table_ref": [],
"text": "Although showing great prominence, our proposed method has the following limitations. First of all, our method relies on the availability of annotated UD trees of both the source and target languages, as we need to use the annotations to parse the syntax trees for our own sentences. Fortunately, UD project covers over 100 languages, where most of the languages, even the minor ones, will have the UD resources. At the same time, our method will be influenced by the quality of UD parsers. Secondly, our method also uses the external translation systems to produce the pseudo parallel sentences, where our method may largely subject to the quality of the translators. Again luckily, current neural machine translation systems have been well developed and established, i.e., Google Translation. Only when handling very scare languages where the current translation systems fail to give satisfactory performances, our method will fail."
},
{
"figure_ref": [],
"heading": "Ethics Statement",
"publication_ref": [],
"table_ref": [],
"text": "In this work, we construct a type of code-mixed UD forest based on the existing UD resources. We note that all the data construction has been accomplished automatically, and we have not created any new annotations with additional human labor. Specifically, we use the UD v2.10 resource, which is a collection of linguistic data and tools that are open-sourced. Each of treebanks of UD has its own license terms, including the CC BY-SA 4.0 8 and CC BY-NC-SA"
},
{
"figure_ref": [],
"heading": "A The universal dependency labels",
"publication_ref": [],
"table_ref": [],
"text": "In Table 6, we list the dependency labels which are the commonly occurred types. Please refer to Stanford dependency 11 "
}
] | Latest efforts on cross-lingual relation extraction (XRE) aggressively leverage the languageconsistent structural features from the universal dependency (UD) resource, while they may largely suffer from biased transfer (e.g., either target-biased or source-biased) due to the inevitable linguistic disparity between languages. In this work, we investigate an unbiased UDbased XRE transfer by constructing a type of code-mixed UD forest. We first translate the sentence of the source language to the parallel target-side language, for both of which we parse the UD tree respectively. Then, we merge the source-/target-side UD structures as a unified code-mixed UD forest. With such forest features, the gaps of UD-based XRE between the training and predicting phases can be effectively closed. We conduct experiments on the ACE XRE benchmark datasets, where the results demonstrate that the proposed code-mixed UD forests help unbiased UD-based XRE transfer, with which we achieve significant XRE performance gains. | Constructing Code-mixed Universal Dependency Forest for Unbiased Cross-lingual Relation Extraction | [
{
"figure_caption": "Merging current pair of nodes from SRC and TGT trees as one into forest, if the two nodes are aligned at same dependency layer.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3: Statistics of mismatching items of UD trees.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "SRC ̸ = Φ) or (T TGT ̸ = Φ) or (opt_nodes̸ = Φ) do 6:",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: Change of syntax distance (shortest path) of relational pair in different UD trees.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 :5Figure 5: Transfer performances by using code-mixed forests generated with different merging rates (θ).",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Merge(T SRC .ROOT, T TGT .ROOT) ▷ merging from ROOT in T SRC and T TGT . merged .next SRC = T SRC .ROOT.GetChildNodes() 9:w merged .next TGT = T TGT .ROOT.GetChildNodes()",
"figure_data": "10:F.w cur .SetChild(w merged , 'root')11:opt_nodes.enqueue(w merged )12:is_root = False13:else14:F.w cur = opt_nodes.dequeue()15:17:w merged = Merge(w SRC i , w TGT j )18:21:opt_nodes.enqueue(w merged )22:end for23:for w i ∈ nonaligned_nodes do24:25:end for26:end if27:end while28:return F32:aligned_pairs = []33:for m i↔j ∈ M do34:if m i↔j > θ then35:aligned_pairs.Append(nodes_a[i], nodes_b[j], nodes_b[i].arc )36:nodes_a.Remove(w i )37:nodes_a.Remove(w j )38:end if39:end for40:nonaligned_nodes = nodes_a.union(nodes_b)▷ words with no salient alignments.",
"figure_id": "tab_0",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "7 Data statistics. The numbers are documents.",
"figure_data": "LanguageTrainDevTestEN4796060ZH5076363AR3234040",
"figure_id": "tab_1",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Main results of cross-lingual RE transfer tasks from English language to other languages, by different models and features. M6 uses two separate instances (texts and UD trees) for training, including the raw SRC one and the pseudo TGT one. M7 uses the SRC-TGT merged one as ours, i.e., code-mixed texts and forests.",
"figure_data": "ENZHAR•Sequential Distance4.83.925.8•Syntactic Distance2.22.65.1",
"figure_id": "tab_3",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Sequential and syntactic (shortest dependency path) distances (words) between the subjects and objects of the relational triplets.",
"figure_data": "",
"figure_id": "tab_4",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "The statistics of the words before and after constructing code-mixed data.",
"figure_data": "Words per SentenceBefore MergingAfter MergingSRC (EN)TGTSumCode-mixedMerged (Rate)EN-ZH15.3224.9140.2331.638.6 (21.4%)EN-AR15.3233.1248.4440.448.0 (16.6%)EN→ZHEN→AR• Simple SentenceSynBaseline(+T SRC )66.178.2SynBaseline(+T T GT )68.780.6SynBaseline(+F)71.382.4• Complex SentenceSynBaseline(+T SRC )39.537.4SynBaseline(+T T GT )41.342.1SynBaseline(+F)57.267.3",
"figure_id": "tab_6",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Comparisons under different types of sentences.",
"figure_data": "",
"figure_id": "tab_7",
"figure_label": "5",
"figure_type": "table"
}
] | Hao Fei; Meishan Zhang; Min Zhang; Tat-Seng Chua | [
{
"authors": "Ahmad Wasi Uddin; Nanyun Peng; Kai-Wei Chang",
"journal": "",
"ref_id": "b0",
"title": "GATE: graph attention transformer encoder for cross-lingual relation and event extraction",
"year": "2021"
},
{
"authors": "Suman Banerjee; Nikita Moghe; Siddhartha Arora; Mitesh M Khapra",
"journal": "",
"ref_id": "b1",
"title": "A dataset for building code-mixed goal oriented conversation systems",
"year": "2018"
},
{
"authors": "Anders Björkelund; Love Hafdell; Pierre Nugues",
"journal": "",
"ref_id": "b2",
"title": "Multilingual semantic role labeling",
"year": "2009"
},
{
"authors": "Duy-Cat Can; Hoang-Quynh Le; Quang-Thuy Ha; Nigel Collier",
"journal": "",
"ref_id": "b3",
"title": "A richer-but-smarter shortest dependency path with attentive augmentation for relation extraction",
"year": "2019"
},
{
"authors": "Silviu Cucerzan; David Yarowsky",
"journal": "",
"ref_id": "b4",
"title": "Language independent named entity recognition combining morphological and contextual evidence",
"year": "1999"
},
{
"authors": "Aron Culotta; Jeffrey Sorensen",
"journal": "",
"ref_id": "b5",
"title": "Dependency tree kernels for relation extraction",
"year": "2004"
},
{
"authors": "Angel Daza; Anette Frank",
"journal": "",
"ref_id": "b6",
"title": "Translate and label! an encoder-decoder approach for cross-lingual semantic role labeling",
"year": "2019"
},
{
"authors": "Marie-Catherine De Marneffe; Christopher D Manning; Joakim Nivre; Daniel Zeman",
"journal": "Comput. Linguistics",
"ref_id": "b7",
"title": "Universal dependencies",
"year": "2021"
},
{
"authors": "Timothy Dozat; Christopher D Manning",
"journal": "",
"ref_id": "b8",
"title": "Deep biaffine attention for neural dependency parsing",
"year": "2017"
},
{
"authors": "Fei Hao Fei; Bobo Li; Donghong Li; Ji",
"journal": "",
"ref_id": "b9",
"title": "Encoder-decoder based unified semantic role labeling with label-aware syntax",
"year": "2021"
},
{
"authors": "Shengqiong Hao Fei; Jingye Wu; Bobo Li; Fei Li; Libo Li; Meishan Qin; Min Zhang; Tat-Seng Zhang; Chua",
"journal": "",
"ref_id": "b10",
"title": "Lasuie: Unifying information extraction with latent adaptive structure-aware generative language model",
"year": "2022"
},
{
"authors": "Meishan Hao Fei; Donghong Zhang; Ji",
"journal": "",
"ref_id": "b11",
"title": "a. Cross-lingual semantic role labeling with highquality translated training corpus",
"year": "2020"
},
{
"authors": "Meishan Hao Fei; Fei Zhang; Donghong Li; Ji",
"journal": "IEEE ACM Trans. Audio Speech Lang. Process",
"ref_id": "b12",
"title": "Cross-lingual semantic role labeling with model transfer",
"year": "2020"
},
{
"authors": "Daniel Gildea; Daniel Jurafsky",
"journal": "",
"ref_id": "b13",
"title": "Automatic labeling of semantic roles",
"year": "2000"
},
{
"authors": "Harry Halpin; Johanna D Moore",
"journal": "",
"ref_id": "b14",
"title": "Event extraction in a plot advice agent",
"year": "2006"
},
{
"authors": "Stephanie Walker; Julie Strassel; Kazuaki Medero; Maeda",
"journal": "",
"ref_id": "b15",
"title": "Ace 2005 multilingual training corpus",
"year": "2006"
},
{
"authors": "Aditya Joshi; Ameya Prabhu; Manish Shrivastava; Vasudeva Varma",
"journal": "",
"ref_id": "b16",
"title": "Towards sub-word level compositions for sentiment analysis of Hindi-English code mixed text",
"year": "2016"
},
{
"authors": "Seokhwan Kim; Minwoo Jeong; Jonghoon Lee; Gary Geunbae; Lee ",
"journal": "",
"ref_id": "b17",
"title": "A cross-lingual annotation projection approach for relation detection",
"year": "2010"
},
{
"authors": "Mikhail Kozhevnikov; Ivan Titov",
"journal": "",
"ref_id": "b18",
"title": "Crosslingual transfer of semantic role labeling models",
"year": "2013"
},
{
"authors": "Igor Labutov; Hod Lipson",
"journal": "",
"ref_id": "b19",
"title": "Generating codeswitched text for lexical learning",
"year": "2014"
},
{
"authors": "Chenwei Lou; Jun Gao; Changlong Yu; Wei Wang; Huan Zhao; Weiwei Tu; Ruifeng Xu",
"journal": "",
"ref_id": "b20",
"title": "Translation-based implicit annotation projection for zero-shot cross-lingual event argument extraction",
"year": "2022"
},
{
"authors": "Di Lu; Ananya Subburathinam; Heng Ji; Jonathan May; Shih-Fu Chang; Avi Sil; Clare Voss",
"journal": "",
"ref_id": "b21",
"title": "Crosslingual structure transfer for zero-resource event extraction",
"year": "2020"
},
{
"authors": "Ryan Mcdonald; Joakim Nivre; Yvonne Quirmbach-Brundage; Yoav Goldberg; Dipanjan Das; Kuzman Ganchev; Keith Hall; Slav Petrov; Hao Zhang; Oscar Täckström; Claudia Bedini; Núria Bertomeu Castelló; Jungmee Lee",
"journal": "",
"ref_id": "b22",
"title": "Universal dependency annotation for multilingual parsing",
"year": "2013"
},
{
"authors": "Makoto Miwa; Mohit Bansal",
"journal": "",
"ref_id": "b23",
"title": "End-to-end relation extraction using LSTMs on sequences and tree structures",
"year": "2016"
},
{
"authors": "Phoebe Mulcaire; Swabha Swayamdipta; Noah A Smith",
"journal": "",
"ref_id": "b24",
"title": "Polyglot semantic role labeling",
"year": "2018"
},
{
"authors": "Jian Ni; Radu Florian",
"journal": "",
"ref_id": "b25",
"title": "Neural cross-lingual relation extraction based on bilingual word embedding mapping",
"year": "2019"
},
{
"authors": "Sebastian Padó; Mirella Lapata",
"journal": "J. Artif. Intell. Res",
"ref_id": "b26",
"title": "Cross-lingual annotation projection for semantic roles",
"year": "2009"
},
{
"authors": "Libo Qin; Minheng Ni; Yue Zhang; Wanxiang Che",
"journal": "",
"ref_id": "b27",
"title": "Cosda-ml: Multi-lingual code-switching data augmentation for zero-shot cross-lingual NLP",
"year": "2020"
},
{
"authors": "Bidisha Samanta; Niloy Ganguly; Soumen Chakrabarti",
"journal": "",
"ref_id": "b28",
"title": "Improved sentiment detection via label transfer from monolingual to synthetic codeswitched text",
"year": "2019"
},
{
"authors": "Ananya Subburathinam; Di Lu; Heng Ji; Jonathan May; Shih-Fu Chang; Avirup Sil; Clare Voss",
"journal": "",
"ref_id": "b29",
"title": "Cross-lingual structure transfer for relation and event extraction",
"year": "2019"
},
{
"authors": "Nasrin Taghizadeh; Heshaam Faili",
"journal": "ACM Trans. Asian Low Resour. Lang. Inf. Process",
"ref_id": "b30",
"title": "Crosslingual adaptation using universal dependencies",
"year": "2021"
},
{
"authors": "Nasrin Taghizadeh; Heshaam Faili",
"journal": "Comput. Speech Lang",
"ref_id": "b31",
"title": "Crosslingual transfer learning for relation extraction using universal dependencies",
"year": "2022"
},
{
"authors": "Petar Velickovic; Guillem Cucurull; Arantxa Casanova; Adriana Romero; Pietro Liò; Yoshua Bengio",
"journal": "",
"ref_id": "b32",
"title": "Graph attention networks",
"year": "2018"
},
{
"authors": "Shengqiong Wu; Hao Fei; Yafeng Ren; Donghong Ji; Jingye Li",
"journal": "",
"ref_id": "b33",
"title": "Learn from syntax: Improving pair-wise aspect and opinion terms extraction with rich syntactic knowledge",
"year": "2021"
},
{
"authors": "Min Xiao; Yuhong Guo",
"journal": "",
"ref_id": "b34",
"title": "Annotation projection-based representation learning for crosslingual dependency parsing",
"year": "2015"
},
{
"authors": "Dmitry Zelenko; Chinatsu Aone; Anthony Richardella",
"journal": "",
"ref_id": "b35",
"title": "Kernel methods for relation extraction",
"year": "2002"
},
{
"authors": "Meishan Zhang; Yue Zhang; Guohong Fu",
"journal": "",
"ref_id": "b36",
"title": "Cross-lingual dependency parsing using code-mixed TreeBank",
"year": "2019"
},
{
"authors": "Yuhao Zhang; Peng Qi; Christopher D Manning",
"journal": "",
"ref_id": "b37",
"title": "Graph convolution over pruned dependency trees improves relation extraction",
"year": "2018"
},
{
"authors": "Zhisong Zhang; Emma Strubell; Eduard Hovy",
"journal": "",
"ref_id": "b38",
"title": "On the benefit of syntactic supervision for crosslingual transfer in semantic role labeling",
"year": "2021"
},
{
"authors": "Ranran Zhen; Rui Wang; Guohong Fu; Chengguo Lv; Meishan Zhang",
"journal": "",
"ref_id": "b39",
"title": "Chinese opinion role labeling with corpus translation: A pivot study",
"year": "2021"
}
] | [
{
"formula_coordinates": [
5,
76.98,
124.77,
168.47,
11.33
],
"formula_id": "formula_0",
"formula_text": "1: def Construct (T SRC , T TGT , M , F)"
},
{
"formula_coordinates": [
5,
122.14,
207.59,
45.26,
10.77
],
"formula_id": "formula_1",
"formula_text": "w merged ="
},
{
"formula_coordinates": [
6,
351.22,
248.22,
170.31,
8.09
],
"formula_id": "formula_2",
"formula_text": "H = {h1, • • • , hn} = MLM(X) , (1"
},
{
"formula_coordinates": [
6,
521.52,
248.54,
3.48,
7.77
],
"formula_id": "formula_3",
"formula_text": ")"
},
{
"formula_coordinates": [
6,
332.35,
374.29,
192.66,
10.33
],
"formula_id": "formula_4",
"formula_text": "ρi,j = Softmax(GeLU(U T [W1ri; W2rj])) ,(2)"
},
{
"formula_coordinates": [
6,
369.96,
389.93,
155.05,
19.94
],
"formula_id": "formula_5",
"formula_text": "ui = σ( j ρi,jW3r 1 j ) ,(3)"
},
{
"formula_coordinates": [
6,
390.33,
497.1,
134.68,
10.25
],
"formula_id": "formula_6",
"formula_text": "Ĥ = H ⊕ U .(4)"
},
{
"formula_coordinates": [
6,
328.21,
633.92,
196.8,
11.13
],
"formula_id": "formula_7",
"formula_text": "y = Softmax(h T s • W1 • ho + W2 • Pool( Ĥ)) . (5)"
}
] |
|
10.48550/ARXIV.2002.05709 | 2023-05-20 | [
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b12",
"b6",
"b12",
"b18",
"b17",
"b16",
"b18",
"b17",
"b16",
"b0",
"b23",
"b18",
"b9",
"b37",
"b20",
"b39",
"b11",
"b13",
"b8",
"b2",
"b16",
"b0",
"b17"
],
"table_ref": [],
"text": "Continual learning aims to continually train models with new tasks without forgetting previously learned tasks (Ke and Liu, 2022;De Lange et al., 2022). It has become a promising direction for NLP models to incrementally learn new tasks/domains/classes as humans do (Ke and Liu, 2022). A typical scenario aims to enable NLP models to solve various tasks in an incremental manner, namely the task-incremental continual learning scenario, which is our study setting in this paper. A salient challenge for continual learning is that continually learned models usually suffer from cartographic forgetting (CF), i.e. the performance on previously learned tasks decreases after training on the new one (Lopez-Paz and Ranzato, 2017).\nVarious training strategies have been proposed to mitigate CF (Li and Hoiem, 2017;Kirkpatrick et al., 2017;Lopez-Paz and Ranzato, 2017). Under the fixed model structure, regularization-based methods design regularization terms to control the shift of representations learned from previous tasks (Li and Hoiem, 2017;Kirkpatrick et al., 2017;Aljundi et al., 2018). Rehearsal-based methods save the data samples from previous tasks into a memory buffer and re-train the model to recover knowledge during training on the current task (Riemer et al., 2019;Lopez-Paz and Ranzato, 2017;de Masson D'Autume et al., 2019). However, most continual learning methods are designed to recover the learned knowledge or mitigate the representation of forgetting.\nSeldom considers adapting the classification criterion for the newly learned representations. For example, in supervised contrastive learning, the contrastive objective is designed to pull the data representations with the same labels together and push representations with different labels away (Chen et al., 2020a;Gao et al., 2021;Zhang et al., 2021;Neelakantan et al., 2022;Zhao et al., 2022). Representations of the training samples can be saved as a classification criterion, after which an instancebased method such as a k Nearest Neighbor (kNN) module can be leveraged for inference (Kassner and Schütze, 2020;Khandelwal et al., 2020). After learning the new task, we can feed the saved sample into models for new classification criteria and mitigate the problem of CF. For example in Figure 1, although the representations have decayed for learning the new task, the saved samples adapt to serve as the classification criterion in kNN modules.\nInspired by the above motivation, we investigate the use of supervised contrastive learning for taskincremental continual learning (SCCL). After supervised contrastive learning on each task, we use a K-means module to select several samples and save them into a memory buffer while maintaining the learned representation distribution. In addition, to mitigate the representation drift when training the model for new tasks, we use an instance-wise relation distillation (IRD) term (Fang et al., 2020;Cha et al., 2021) and a memory replay module (de Masson D'Autume et al., 2019) to maintain the learned knowledge. During inference, the saved samples are fed into the trained model to obtain updated representations and a kNN module is used for classification.\nExperimental results show that our proposed model can achieve state-of-the-art performance compared with standard cross-entropy-based (CE) baselines. We additionally extend different continual learning strategies (Kirkpatrick et al., 2017;Aljundi et al., 2018;Li and Hoiem, 2017) to the supervised contrastive continual learning framework, which gives stronger results than corresponding CE-based methods, showing the advantage of contrastive learning with a kNN classifier in continual learning scenarios. We further analyze the effectiveness of each module in our paper through ablation studies. To our knowledge, we are the first to propose a supervised contrastive learning framework for task-incremental continual learning, without any augmented parameters. The code will be released when accepted."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b33",
"b26",
"b17",
"b16",
"b24",
"b16",
"b0",
"b17",
"b17",
"b24",
"b1",
"b31",
"b9",
"b14",
"b19",
"b2"
],
"table_ref": [],
"text": "Continual Learning Various continual learning methods have been proposed to mitigate the problem of CF. The methods can be broadly divided into architecture-based methods (Yoon et al., 2018;Serra et al., 2018), regularization-based methods Li and Hoiem (2017); Kirkpatrick et al. (2017), and rehearsal-based methods (Rolnick et al., 2019). Under the fixed model structure, regularizationbased methods (Kirkpatrick et al., 2017;Aljundi et al., 2018;Li and Hoiem, 2017) optimize network parameters on the current task while constraining the representation drift. For example, Li and Hoiem (2017) propose learning without forgetting (LwF) to tackle this problem, which regularizes the model output of current data close to those trained for the previous model. Another category of fixed-structure strategies (rehearsal-based) stores a limited subset of samples from previous tasks to mitigate CF such as ER (Rolnick et al., 2019), RM (Bang et al., 2021), andiCaRL (Rebuffi et al., 2017).\nContrastive Learning Contrastive learning is initially introduced in self-supervised settings and proved to subsume or significantly outperform traditional contrastive losses such as triplet loss (Chen et al., 2020b;Wu et al., 2018;Gao et al., 2021;?). For example, Khosla et al. (2020) first propose the idea of self-supervised contrastive learning and prove that the method is more robust to natural corruptions, stable to hyper-parameter settings, and has strong transfer performance. Luo et al. (2022) uses supervised contrastive learning combined with a kNN inference module for cross-domain sentiment analysis, showing a stronger generalization ability compared with standard CE-based methods. Cha et al. (2021) propose a contrastive continual learning method, Co 2 L, for class-incremental continual learning. The method uses an asymmetric supervised contrastive loss to enlarge the distance between representations of previous and new tasks. However, there are significant differences between our model and Co 2 L. First, the asymmetric contrastive loss of Co 2 L is unsuitable for taskincremental continual learning, because a representation can be predicted as different labels according to task objectives. Second, Co 2 L uses a decoupled classification layer for inference, i.e. it learns rep- resentations first and then learns a linear classifier separately, causing low extensibility and high complexity of the model. In contrast, we use a kNN module as the classification criteria to enhance the extensibility of the model. Third, Co 2 L only considers the representation drift from the view of regularization. But we also mitigate the problem of representation drift by feeding memory data into the current model to obtain updated classification criteria."
},
{
"figure_ref": [
"fig_1"
],
"heading": "Method",
"publication_ref": [],
"table_ref": [],
"text": "The overall SCCL framework is illustrated in Figure 2, consisting of four parts. First, we introduce the contrastive learning objective of SCCL in Section 3.1. Second, the selection of learned representations is shown in Section 3.2. Third, an instancewise distillation module and a memory replay module are introduced to preserve learned knowledge in Section 3.3 and 3.4, respectively. Fourth, the kNN inference procedure is shown in 3.5, respectively. The training algorithm is shown in Algorithm 1.\nFormally, a model learns several tasks denoted as {T i }, i = 1, 2, ..., n (i is the number of tasks). Each task T i contains a limited set of labels C i . During the training of the task T i , only the corresponding data D i = {(x i j , y i j )} are available, where x i j is the input text and y i j ∈ C i is the corresponding label. In the scenario of task-incremental continual training, the task id can be observed when carrying out inference, and for generality, we consider the label set\nC j ∩ C k = ∅, if i = j."
},
{
"figure_ref": [],
"heading": "Supervised Contrastive Continual",
"publication_ref": [
"b14"
],
"table_ref": [],
"text": "Training (SCCL)\nDuring the learning on the task T i , we first feed the input x i j into a pre-trained language model to obtain hidden states. The hidden states of a special token\n[CLS] (the beginning token of the pre-trained language model) are regarded as the representation of the input sequence:\nh i j = N orm(LM i (x i j )[CLS]),(1)\nwhere N orm(•) refers to normalization, LM i is the language model encoder trained for the task T i , and LM 0 is the initial pre-trained language model. We denote the data samples in a mini-batch as A (we omit the corner mark i during task T i for simplicity). For each data sample j, we denote N (j) ≡ A/{j}, and the positive neighbor set of it as P (j) = {u|y u = y j and u ∈ N (j)}. To push the representations with different labels away, and pull them with the same labels together, we use supervised contrastive learning objective following Khosla et al. (2020):\nL cl = j∈A -1 |P (j)| p∈P (j) log exp(h i j • h i p /κ) a∈N (j) exp(h i j • h i a /κ) (2)\nwhere κ is the hyper-parameter of temperature."
},
{
"figure_ref": [],
"heading": "Sample Selection",
"publication_ref": [],
"table_ref": [],
"text": "After training on each task T i , we select m samples from training data of D i to keep the representation distribution with respect to the labels (Algorithm 1 (18-22)). In particular, we adopt a K-means module to aggregate the data D i (c) of each label (c ∈ C i ) to clusters. Then we randomly select samples according to the data density to keep representation distribution, which can be formulated as:\nM c = Sample(Kmeans(D i (c)), c, m |C i | ). (3)\nThe selected samples for task T i are the union of selected data for each label c that\nM i = ∪ c∈C i M c .\nM i is saved in the memory buffer and serves as the classification criteria for task T i in the continual learning process."
},
{
"figure_ref": [],
"heading": "Instance-wise Relation Distillation (IRD)",
"publication_ref": [
"b8",
"b2"
],
"table_ref": [],
"text": "To preserve the knowledge learned for previous tasks, inspired by Fang et al. (2020) and Cha et al. (2021), we use an instance-wise relation distillation term to control representation drift (Algorithm 1 (7-9)). During the learning on task T i , i > 1, the normalized instance-wise similarity in the minibatch A is calculated as:\ns i j,p = exp(h i j • h i p /τ ) a∈N (j) exp(h i j • h i a /τ ) ,(4)\nwhere N (j) ≡ A/{j}, the representations are encoded by the model LM i and τ is the hyperparameter temperature. Then the IRD regularization term follows:\nL IRD = 1 |A| 2 j p s i-1 j,p log s i j,p .(5)\nThe IRD regularization term aims to estimate the discrepancy of current representations to those learned in the previous model, and mitigate the representation drift through optimization. In this way, the knowledge of previous models is preserved and the CF problem can be mitigated.\nThe overall training objective can be denoted as follows:\nL = L cl + L IRD .(6)"
},
{
"figure_ref": [],
"heading": "Memory Replay (MR)",
"publication_ref": [],
"table_ref": [],
"text": "To make full use of the memory buffer saved during training, we use a memory replay module (de Masson D' Autume et al., 2019) to further recover the knowledge learned in the previous tasks (Algorithm 1 (14-16)). In the training on the task T i , i > 1, we revisit the samples in the memory buffer and train the model with the same loss in Eq (2) after training every f step on the current task."
},
{
"figure_ref": [],
"heading": "Inference",
"publication_ref": [
"b13"
],
"table_ref": [],
"text": "After learning the task T i , we can obtain the model LM i . During the inference for previous tasks T u , u <= i, we feed each test data x u j into LM i and obtain the corresponding representation h u j . Then we retrieve the k buffered data from M u whose cosine similarity with h u j is the largest. Note that the representations of buffered data are obtained using the current model, which can adapt to the representation drift for parameter update. We denote the k nearest neighbors as (h u k , y u k ) ∈ K u j . The retrieved set is converted to a probability distribution over the labels by applying a softmax with \nM c = Sample(Kmeans(D i (c)), c, m |C i | ); 21: M i = M i ∪ M c ; 22:\nend for 23: M = M + M i ; 24: end for temperature T to the similarity. Using the temperature T > 1 can flatten the distribution, and prevent over-fitting to the most similar searches (Khandelwal et al., 2020). The probability distribution on the labels is expressed as follows:\np k (y j ) ∝ (h u k ,y u k )∈K u j 1 y j =y u k • exp( h u j • h u k T ),(7)\nand the label with the largest probability is taken as the prediction result.\n4 Experimental Setting"
},
{
"figure_ref": [],
"heading": "Tasks",
"publication_ref": [
"b28",
"b10",
"b29",
"b30",
"b21",
"b38",
"b38"
],
"table_ref": [
"tab_1"
],
"text": "We adopt classification tasks from the benchmark GLUE (Wang et al., 2018) and those from MBPA++ (Huang et al., 2021;de Masson D'Autume et al., 2019). We select dissimilar tasks to form the task sequences, i.e. there are no overlap labels between each task. The tasks contain 1) CoLA (Warstadt et al., 2019), requiring the model to determine whether a sentence is linguistically acceptable; 2) MNLI (Williams et al., 2017) mation; 3) QNLI1 , requiring deciding whether the answer answers the question; 4) QQP, (parsed from SQuAD (Rajpurkar et al., 2016)), testing whether a pair of Quora questions are synonymous; 5) Yelp (Zhang et al., 2015), requiring detecting the sentiment of a sentence; 6) AG (Zhang et al., 2015), requiring to classify the topics of the news.\nThe sequences can be divided into 2 types with respect to the task lengths: 1) a sequence of 4 classification tasks containing AG, Yelp, QNLI, and MRPC; 2) a sequence of 6 classification tasks containing AG, MRPC, MNLI, CoLA, Yelp, and QNLI. Without losing generality the orders are randomly selected and the task orders for experiments are shown in Table 1."
},
{
"figure_ref": [],
"heading": "Evaluation Metrics",
"publication_ref": [
"b18"
],
"table_ref": [],
"text": "We adopt the metrics of average accuracy (ACC) and backward transfer (BWT) to evaluate the performance of the continual learning model (Lopez-Paz and Ranzato, 2017). The model trained after the task T i is evaluated on the test set of earlier tasks T j (j <= i), and the test accuracy is denoted as R i j . The metrics are shown as follows:\nACC = 1 n n i=1 R n i (8) BW T = 1 n -1 n-1 i=1 R n i -R i i ,(9)\nwhere the former evaluates the overall performance of the final trained model, and the latter calculates the knowledge forgetting during the continual training procedure."
},
{
"figure_ref": [],
"heading": "Baselines",
"publication_ref": [
"b10",
"b2",
"b32",
"b23"
],
"table_ref": [],
"text": "We not only compare our model with several CEbased continual learning methods but extend training strategies of them to our contrastive learning framework (i.e. training with contrastive learning and inferring with kNN) to verify the effectiveness of contrastive learning in mitigating CF. We also compare our model with the competitive models IRDB (Huang et al., 2021) and Co 2 L (Cha et al., 2021). The shared hyper-parameters are kept the same as SCCL in baselines. The model details are as follows:\n• Fine-tune (CE, CL) (Yogatama et al., 2019) modifies the parameters of the pre-trained language model to adapt to a new task without any augmented strategies and additional loss. • Experience Replay (ER) (Riemer et al., 2019) stores a small subset of samples from previous tasks and replays those to prevent models from forgetting past knowledge. (Yu et al., 2020b) trains on all the tasks simultaneously, i.e. the data of different tasks are mixed up for training. It does not suffer from catastrophic forgetting and represents an upper bound on model performance."
},
{
"figure_ref": [],
"heading": "Implementation Details",
"publication_ref": [
"b15"
],
"table_ref": [],
"text": "We adopt the officially released roberta-base from HuggingFace 2 as our backbone network. We train our model on 1 GPU (A100 80G) using the Adam optimizer (Kingma and Ba, 2014). For all the models, the batch size is 96, the learning rate is 3e-5, "
},
{
"figure_ref": [],
"heading": "Results",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Overall Results",
"publication_ref": [],
"table_ref": [
"tab_4"
],
"text": "The overall results of our experiments are shown in Table 2. First, our model SCCL achieves ACCs of 79.20%, 80.05%, 80.24%, 78.36%, 79.00%, and 78.55% in Order 1-6, respectively, which are 2.37%, 0.9%, 0.66%, 3.28%, 2.49%, and 2.40% higher than the second-best performance of the continual learning baselines. It shows that the performance of the continually learned model is wellmaintained in SCCL, but the problem of CF still exists. SCCL achieves state-of-the-art ACCs compared with the baseline models, indicating the effectiveness of our proposed framework. We also observe that the performance variance is small in the SCCL model for different orders, which implies that our models are not sensitive to the order of task sequences.\nSecond, the results of BWT range from -3.75% to 0.57% in SCCL for Orders 1-6, which demonstrates knowledge forgetting during the continual learning procedure. The results of SCCL are rela-tively higher than CE-based models, indicating that SCCL suffers from a milder impact of CF. Note that the BWT of SCCL is 0.57% in Order 4, which indicates that SCCL can even backward transfer the knowledge from the current tasks to previous tasks. But compared with CL-LwF, CL-MAS, and CL-EWC, the values of ACCs in SCCL are higher, but BWTs are adverse. It implies that using the regularization-based strategies, the fine-tuning performance is destructed for explicit control of representations. In this way, BWTs become low since the fine-tuning performance on downstream tasks is relatively weak.\nThird, the extended CL-based models achieve stronger performance than corresponding standard CE-based models. For example, the model CL-LwF achieves ACCs of 76.53%, 79.15%, 79.58%, 68.24%, 72.39%, and 71.48%, which are 4.44%, 7.02%, 6.04%, 0.01%, 9.24% and 3.56% higher than those of CE-LwF. The results of CL, CL-MAS, and CL-EWC are in a similar pattern. The results reflect that contrastive learning with a kNN classifier for continual learning has a stronger ability to overcome CF. But we observe that Co 2 L achieves relatively low performance compared with our model, which proves that Co 2 L is not effective for task-incremental learning. It can be explained that Co 2 L keeps the knowledge of classes and separate the tasks with clear boundary, by using asymmetric supervised contrastive loss, which makes it difficult to distinguish a representation for different task purposes.\nFinally, we observe a significant variance in the results of different task orders for regularizationbased methods. For example, ACCs of CL-EWC range from 75.89% to 66.74%. But in CE-ER or SCCL the variance is less drastic, such as CE-ER ranging from 76.90% to 75.08% and SCCL ranging from 80.24% to 78.55%. The phenomenon may result from that knowledge forgetting of previous tasks increases step by step for information los, but no samples help recover such information in regularization-based methods."
},
{
"figure_ref": [],
"heading": "Ablation Study",
"publication_ref": [],
"table_ref": [
"tab_4"
],
"text": "We show the ablation study of memory replay and IRD in the last two rows in Table 2. The ACCs of the models w/o memory replay range from 71.22%, to 80.27% for Order 1-6, which are 2.01%, 1.42%, -0.03%, 2.71%, 7.78%, and 3.91% lower than SCCL, respectively. It shows the effectiveness of memory replay, without which ACC also becomes less robust to task orders. Then ACCs of the models w/o IRD are 1.63%, 0.32%, 0.76%, 4.49%, 2.11%, and 4.23% lower than SCCL for Order 1-6, respectively. We observe that the models w/o IRD are more robust to task orders, which implies that rehearsal-based methods are less sensitive to task sequences. Comparing the model w/o IRD with CE-ER, the model performance are also higher than those of CE-ER, which uses almost the same training strategy. The phenomenon demonstrates the effectiveness of contrastive learning in overcoming CF.\nWe also compare our model with Co 2 L in abla-tion studies (Figure 4). "
},
{
"figure_ref": [],
"heading": "Detailed Results",
"publication_ref": [],
"table_ref": [],
"text": "As an example, we show the detailed results of Order 3 in several models (Figure 3). and MRPC. The accuracy of QNLI decreases by 17.38%, that of Yelp decreases by 11.59%, and that of MRPC decreases by 14.84%. As for our model SCCL, we observe that the test performance is 88.70%, 87.24%, 86.24% and 85.68% after training on tasks QNLI, Yelp, MRPC, and AG, respectively. It shows that the performance of SCCL decreases as the training precedes, but within a small range (3.02%). The results on Yelp and MRPC are in a similar pattern. It demonstrates that our model has a strong ability to overcome CF."
},
{
"figure_ref": [
"fig_4"
],
"heading": "Visualization",
"publication_ref": [],
"table_ref": [],
"text": "We use t-SNE to visualize the representations of QNLI in Order 3 of the training models, CE, CE-ER, CL, and SCCL (Figure 5). As we observe in CE the representations of the test data are clearly separated into two clusters after training on the task QNLI. When finishing the continual learning, the representations become nearly uniformly distributed on the feature space and the model only achieves an accuracy of 50.21%. It demonstrates that catastrophic forgetting is significant due to representation drift. In the model CL, the representations drift severely as well, but the distribution is less uniform compared with CE. Typically, we can clearly at the upper right of the distribution, there are more memory samples with label 1, and the test samples with label 1 also gather in the position, indicating correct classification based on kNN. The test performance achieves 61.74%, but is still 26.94% lower than the initial model. The phenomenon shows representations during continual learning drift less significantly and the saved sam-ples (the classification criterion) also drift, which maintains some correct inferences. But CF is still a salient problem in contrastive learning.\nBut in CE-ER, the boundary of the representations becomes indistinct, and the accuracy of QNLI decreases from 87.79% to 70.11% after continual learning. It indicates that the representations are less effective compared with the initially trained, i.e. CF is significant in CE-ER. But the representations in SCCL are still clearly divided into two parts according to the labels. The representations of the memory samples are among the according clusters, implying the performance on the task QNLI is well-maintained. Correspondingly, the accuracy at the end of learning is 85.68% based on SCCL, only 3.02% lower than the initial performance. It shows that in SCCL the representation drift slightly and the classification criterion is well-maintained, resulting in a satisfactory performance."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [
"b9"
],
"table_ref": [],
"text": "In this paper, we proposed a supervised contrastive learning model for task-incremental continual learning (SCCL) to boost the extensibility of continual learning. The model used contrastive learning to learn representations and a kNN module was adopted for inference, together with an instance-wise distillation and a memory replay module to maintain previously learned knowledge.\nWith extensive experiments, our model achieved state-of-the-art performance compared with standard CE-based methods. Ablation studies and visualizations also proved the effectiveness of our model in solving the problem of CF.\nOur model SCCL is specific for task-incremental continual learning scenarios, but not suitable for class-incremental scenarios. In class-incremental scenarios, the representations of current classes should be designed to be far away from previous ones. For simplicity, we do not consider data augmentation in our model, so the batch size should be large enough to contain positive pairs for each label. But data augmentation (such as two different dropout representations (Gao et al., 2021)) is a plug-and-play module for our model if there are plenty of labels in each task."
},
{
"figure_ref": [],
"heading": "A Data Statistics",
"publication_ref": [],
"table_ref": [
"tab_6"
],
"text": "We show the data statistics in Table 3. "
},
{
"figure_ref": [],
"heading": "B kNN Sensitivity",
"publication_ref": [],
"table_ref": [],
"text": "We show the sensitivity of SCCL to the number of k in the kNN module (Figure 6). We find that the performance of our model fluctuates from 80.24% to 80.28% (a significantly small range) in our method, indicating the representations in our model cluster well in the feature space and are robust to the hyperparameter k. But the performance in CL fluctuates more severely, ranging from 59.02% to 60.12%. The best performance is achieved when k = 10 and decreases with the increase of k, which means the representations drift significantly and the clusters become less reliable. The experiment demonstrates the effectiveness of IRD regularization term and the memory replay module in maintaining the representation distribution, and without them the representations drift significantly, suffering from the CF problem. "
}
] | Task-incremental continual learning refers to continually training a model in a sequence of tasks while overcoming the problem of catastrophic forgetting (CF). The issue arrives for the reason that the learned representations are forgotten for learning new tasks, and the decision boundary is destructed. Previous studies mostly consider how to recover the representations of learned tasks. It is seldom considered to adapt the decision boundary for new representations and in this paper we propose a Supervised Contrastive learning framework with adaptive classification criterion for Continual Learning (SCCL), In our method, a contrastive loss is used to directly learn representations for different tasks and a limited number of data samples are saved as the classification criterion. During inference, the saved data samples are fed into the current model to obtain updated representations, and a k Nearest Neighbour module is used for classification. In this way, the extensible model can solve the learned tasks with adaptive criteria of saved samples. To mitigate CF, we further use an instance-wise relation distillation regularization term and a memory replay module to maintain the information of previous tasks. Experiments show that SCCL achieves stateof-the-art performance and has a stronger ability to overcome CF compared with the classification baselines. | Mitigating Catastrophic Forgetting in Task-Incremental Continual Learning with Adaptive Classification Criterion | [
{
"figure_caption": "Figure 1 :1Figure 1: Illustration of representations after contrastive continual learning on a task before and after learning a new task.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: The model framework of SCCL contains four main modules: (1) the supervised contrastive learning for each task; (2) the explicit control of catastrophic forgetting with IRD knowledge distillation and memory replay; (3) the selection of learned representations; (4) a kNN inference module.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :Figure 4 :34Figure 3: Detailed results during continual learning procedure for different strategies in Order 3.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "34",
"figure_type": "figure"
},
{
"figure_caption": "First, we replace the kNN module of SCCL with a decoupled linear classifier like(Cha et al., 2021) (SCCL-CLS), where ACCs are slightly smaller than SCCL. It indicates that the kNN module in SCCL can achieve satisfactory performance without additional training on the final representations of contrastive learning. Then we replace the decoupled linear classifier of Co 2 L with our kNN module (Co 2 L-kNN), and we observe an increase in performance. It implies that the representations learned by Co 2 L are not separated clearly in the feature space, thus a trained linear layer is less effective for classification. But K-means selection of the samples and kNN inference module can estimate the representation distribution more precisely, resulting in better performance. Note that the results of SCCL are also stronger than Co 2 L-kNN, which indicates the effectiveness of our model on task-incremental continual learning.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 55Figure 5: t-SNE visualization of the representations of QNLI samples learned based on the different continual learning methods in Order 5. 'E' refers to the representations at the end of continual learning.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Algorithm 1 SCCL Training Input: A set of training task {T i } n , the corresponding data set {D i } n , sets of disjoint classes {C i } n . Training steps S and memory replay frequency f . Memory buffer size m. Initial pre-trained language model LM 0 . Output: Trained language model encoder LM n and memory buffer M.",
"figure_data": "1: Load pre-trained language model LM 0 ;2: M = []3: for i = 1, ..., n do4:for t = 1, ..., S do5:Draw mini-batch A from D i ;6:Calculate L cl of A with LM i (Eq (1-2));7:if i > 1 then8:Calculate LIRD of A (Eq (5));9:L = L cl + LIRD;10:else11:L = L cl ;12:end if13:Update model parameters with L;14:if i % f == 0 then15:Update model parameters with memory relay;16:end if17:end for18:for c ∈ C i do19:Obtain k-means clusters of data with label c;20:",
"figure_id": "tab_0",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "containing 433k sentence pairs annotated with textual entailment infor-Orders 1 AG → Yelp → QNLI→ MRPC 2 MRPC → QNLI → Yelp →AG 3 QNLI →Yelp →MRPC→AG 4 AG→MRPC →CoLA→MNLI →Yelp→ QNLI 5 QNLI →Yelp →MNLI→CoLA →MRPC→ AG 6 MNLI → AG →QNLI→ MRPC →Yelp→ CoLA Different task orders for our experiments.",
"figure_data": "",
"figure_id": "tab_1",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Continual Learning results on 6 different tasks. 'CE' refers to the standard cross-entropy-based methods, and 'CL' refers to extended contrastive-learning-based methods with continual learning strategies. '-' for not acquirable. All the results are averaged on 5 different random seeds.",
"figure_data": "61.87 -22.81",
"figure_id": "tab_4",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Statistics for different classification tasks.",
"figure_data": "TaskType#Train #Test #LabelsAGNews16000 76004QNLIQ & A800052662YelpSentiment20000 76005CoLA Linguistics652710422MNLIInference12000 98153MRPC Paraphrase407417252",
"figure_id": "tab_6",
"figure_label": "3",
"figure_type": "table"
}
] | Yun Luo; Xiaotian Lin; Zhen Yang; Fandong Meng; Jie Zhou; Yue Zhang | [
{
"authors": "Rahaf Aljundi; Francesca Babiloni; Mohamed Elhoseiny; Marcus Rohrbach; Tinne Tuytelaars",
"journal": "",
"ref_id": "b0",
"title": "Memory aware synapses: Learning what (not) to forget",
"year": "2018"
},
{
"authors": "Jihwan Bang; Heesu Kim; Youngjoon Yoo; Jung-Woo Ha; Jonghyun Choi",
"journal": "",
"ref_id": "b1",
"title": "Rainbow memory: Continual learning with a memory of diverse samples",
"year": "2021"
},
{
"authors": "Hyuntak Cha; Jaeho Lee; Jinwoo Shin",
"journal": "",
"ref_id": "b2",
"title": "Co2l: Contrastive continual learning",
"year": "2021"
},
{
"authors": "Ting Chen; Simon Kornblith; Mohammad Norouzi; Geoffrey Hinton",
"journal": "",
"ref_id": "b3",
"title": "a. A simple framework for contrastive learning of visual representations",
"year": "2020"
},
{
"authors": "Ting Chen; Simon Kornblith; Mohammad Norouzi; Geoffrey Hinton",
"journal": "",
"ref_id": "b4",
"title": "A simple framework for contrastive learning of visual representations",
"year": "2020"
},
{
"authors": " Pmlr",
"journal": "",
"ref_id": "b5",
"title": "",
"year": ""
},
{
"authors": "Matthias De Lange; Rahaf Aljundi; Marc Masana; Sarah Parisot; Xu Jia; Aleš Leonardis; Gregory Slabaugh; Tinne Tuytelaars",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b6",
"title": "A continual learning survey: Defying forgetting in classification tasks",
"year": "2022"
},
{
"authors": "Cyprien De Masson D'autume; Sebastian Ruder; Lingpeng Kong; Dani Yogatama",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b7",
"title": "Episodic memory in lifelong language learning",
"year": "2019"
},
{
"authors": "Zhiyuan Fang; Jianfeng Wang; Lijuan Wang; Lei Zhang; Yezhou Yang; Zicheng Liu",
"journal": "",
"ref_id": "b8",
"title": "Seed: Self-supervised distillation for visual representation",
"year": "2020"
},
{
"authors": "Tianyu Gao; Xingcheng Yao; Danqi Chen",
"journal": "Association for Computational Linguistics",
"ref_id": "b9",
"title": "SimCSE: Simple contrastive learning of sentence embeddings",
"year": "2021"
},
{
"authors": "Yufan Huang; Yanzhe Zhang; Jiaao Chen; Xuezhi Wang; Diyi Yang",
"journal": "",
"ref_id": "b10",
"title": "Continual learning for text classification with information disentanglement based regularization",
"year": "2021"
},
{
"authors": "Nora Kassner; Hinrich Schütze",
"journal": "Association for Computational Linguistics",
"ref_id": "b11",
"title": "BERT-kNN: Adding a kNN search component to pretrained language models for better QA",
"year": "2020"
},
{
"authors": "Zixuan Ke; Bing Liu",
"journal": "",
"ref_id": "b12",
"title": "Continual learning of natural language processing tasks: A survey",
"year": "2022"
},
{
"authors": "Urvashi Khandelwal; Angela Fan; Dan Jurafsky; Luke Zettlemoyer; Mike Lewis",
"journal": "",
"ref_id": "b13",
"title": "Nearest neighbor machine translation",
"year": "2020"
},
{
"authors": "Prannay Khosla; Piotr Teterwak; Chen Wang; Aaron Sarna; Yonglong Tian; Phillip Isola; Aaron Maschinot; Ce Liu; Dilip Krishnan",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b14",
"title": "Supervised contrastive learning",
"year": "2020"
},
{
"authors": "P Diederik; Jimmy Kingma; Ba",
"journal": "",
"ref_id": "b15",
"title": "Adam: A method for stochastic optimization",
"year": "2014"
},
{
"authors": "James Kirkpatrick; Razvan Pascanu; Neil Rabinowitz; Joel Veness; Guillaume Desjardins; Andrei A Rusu; Kieran Milan; John Quan; Tiago Ramalho; Agnieszka Grabska-Barwinska",
"journal": "Proceedings of the national academy of sciences",
"ref_id": "b16",
"title": "Overcoming catastrophic forgetting in neural networks",
"year": "2017"
},
{
"authors": "Zhizhong Li; Derek Hoiem",
"journal": "IEEE transactions on pattern analysis and machine intelligence",
"ref_id": "b17",
"title": "Learning without forgetting",
"year": "2017"
},
{
"authors": "David Lopez; - Paz; Marc'aurelio Ranzato",
"journal": "Advances in neural information processing systems",
"ref_id": "b18",
"title": "Gradient episodic memory for continual learning",
"year": "2017"
},
{
"authors": "Yun Luo; Fang Guo; Zihan Liu; Yue Zhang",
"journal": "",
"ref_id": "b19",
"title": "Mere contrastive learning for cross-domain sentiment analysis",
"year": "2022"
},
{
"authors": "Arvind Neelakantan; Tao Xu; Raul Puri; Alec Radford; Jesse Michael Han; Jerry Tworek; Qiming Yuan; Nikolas Tezak; Jong Wook Kim; Chris Hallacy",
"journal": "",
"ref_id": "b20",
"title": "Text and code embeddings by contrastive pretraining",
"year": "2022"
},
{
"authors": "Pranav Rajpurkar; Jian Zhang; Konstantin Lopyrev; Percy Liang",
"journal": "Association for Computational Linguistics",
"ref_id": "b21",
"title": "SQuAD: 100,000+ questions for machine comprehension of text",
"year": "2016"
},
{
"authors": " Sylvestre-Alvise; Alexander Rebuffi; Georg Kolesnikov; Christoph H Sperl; Lampert",
"journal": "",
"ref_id": "b22",
"title": "icarl: Incremental classifier and representation learning",
"year": "2017"
},
{
"authors": "Matthew Riemer; Ignacio Cases; Robert Ajemian; Miao Liu; Irina Rish; Yuhai Tu; Gerald Tesauro",
"journal": "ICLR",
"ref_id": "b23",
"title": "Learning to learn without forgetting by maximizing transfer and minimizing interference",
"year": "2019"
},
{
"authors": "David Rolnick; Arun Ahuja; Jonathan Schwarz; Timothy Lillicrap; Gregory Wayne",
"journal": "",
"ref_id": "b24",
"title": "Experience replay for continual learning",
"year": "2019"
},
{
"authors": "Curran Associates; Inc ",
"journal": "",
"ref_id": "b25",
"title": "",
"year": ""
},
{
"authors": "Joan Serra; Didac Suris; Marius Miron; Alexandros Karatzoglou",
"journal": "",
"ref_id": "b26",
"title": "Overcoming catastrophic forgetting with hard attention to the task",
"year": "2018"
},
{
"authors": " Pmlr",
"journal": "",
"ref_id": "b27",
"title": "",
"year": ""
},
{
"authors": "Alex Wang; Amanpreet Singh; Julian Michael; Felix Hill; Omer Levy; Samuel Bowman",
"journal": "",
"ref_id": "b28",
"title": "Glue: A multi-task benchmark and analysis platform for natural language understanding",
"year": "2018"
},
{
"authors": "Alex Warstadt; Amanpreet Singh; Samuel R Bowman",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b29",
"title": "Neural network acceptability judgments",
"year": "2019"
},
{
"authors": "Adina Williams; Nikita Nangia; Samuel R Bowman",
"journal": "",
"ref_id": "b30",
"title": "A broad-coverage challenge corpus for sentence understanding through inference",
"year": "2017"
},
{
"authors": "Zhirong Wu; Yuanjun Xiong; Stella X Yu; Dahua Lin",
"journal": "",
"ref_id": "b31",
"title": "Unsupervised feature learning via nonparametric instance discrimination",
"year": "2018"
},
{
"authors": "Dani Yogatama; Cyprien De Masson D'autume; Jerome Connor; Tomas Kocisky; Mike Chrzanowski; Lingpeng Kong; Angeliki Lazaridou; Wang Ling; Lei Yu; Chris Dyer",
"journal": "",
"ref_id": "b32",
"title": "Learning and evaluating general linguistic intelligence",
"year": "2019"
},
{
"authors": "Jaehong Yoon; Eunho Yang; Jeongtae Lee; Sung Ju Hwang",
"journal": "ICLR",
"ref_id": "b33",
"title": "Lifelong learning with dynamically expandable networks",
"year": "2018"
},
{
"authors": "Lu Yu; Bartłomiej Twardowski; Xialei Liu; Luis Herranz; Kai Wang; Yongmei Cheng; Shangling Jui; Joost Van De Weijer",
"journal": "",
"ref_id": "b34",
"title": "Semantic drift compensation for class-incremental learning",
"year": "2020"
},
{
"authors": "Tianhe Yu; Saurabh Kumar; Abhishek Gupta; Sergey Levine; Karol Hausman; Chelsea Finn",
"journal": "",
"ref_id": "b35",
"title": "Gradient surgery for multi-task learning",
"year": "2020"
},
{
"authors": "Curran Associates; Inc ",
"journal": "",
"ref_id": "b36",
"title": "",
"year": ""
},
{
"authors": "Dejiao Zhang; Shang-Wen Li; Wei Xiao; Henghui Zhu; Ramesh Nallapati; Andrew O Arnold; Bing Xiang",
"journal": "Association for Computational Linguistics",
"ref_id": "b37",
"title": "Pairwise supervised contrastive learning of sentence representations",
"year": "2021"
},
{
"authors": "Xiang Zhang; Junbo Zhao; Yann Lecun",
"journal": "Advances in neural information processing systems",
"ref_id": "b38",
"title": "Character-level convolutional networks for text classification",
"year": "2015"
},
{
"authors": "Kang Zhao; Hua Xu; Jiangong Yang; Kai Gao",
"journal": "Association for Computational Linguistics",
"ref_id": "b39",
"title": "Consistent representation learning for continual relation extraction",
"year": "2022"
}
] | [
{
"formula_coordinates": [
3,
149.64,
680.8,
100.08,
11.76
],
"formula_id": "formula_0",
"formula_text": "C j ∩ C k = ∅, if i = j."
},
{
"formula_coordinates": [
3,
345.45,
300.89,
178.96,
14.19
],
"formula_id": "formula_1",
"formula_text": "h i j = N orm(LM i (x i j )[CLS]),(1)"
},
{
"formula_coordinates": [
3,
314.2,
487.38,
210.21,
37.81
],
"formula_id": "formula_2",
"formula_text": "L cl = j∈A -1 |P (j)| p∈P (j) log exp(h i j • h i p /κ) a∈N (j) exp(h i j • h i a /κ) (2)"
},
{
"formula_coordinates": [
3,
313.22,
676,
211.19,
24.43
],
"formula_id": "formula_3",
"formula_text": "M c = Sample(Kmeans(D i (c)), c, m |C i | ). (3)"
},
{
"formula_coordinates": [
3,
448.51,
720.97,
77.81,
13.42
],
"formula_id": "formula_4",
"formula_text": "M i = ∪ c∈C i M c ."
},
{
"formula_coordinates": [
4,
107.49,
190.07,
181.64,
30.32
],
"formula_id": "formula_5",
"formula_text": "s i j,p = exp(h i j • h i p /τ ) a∈N (j) exp(h i j • h i a /τ ) ,(4)"
},
{
"formula_coordinates": [
4,
103.07,
288.07,
186.07,
28.55
],
"formula_id": "formula_6",
"formula_text": "L IRD = 1 |A| 2 j p s i-1 j,p log s i j,p .(5)"
},
{
"formula_coordinates": [
4,
140.66,
435.49,
148.47,
10.77
],
"formula_id": "formula_7",
"formula_text": "L = L cl + L IRD .(6)"
},
{
"formula_coordinates": [
4,
306.14,
339.7,
198.4,
33.44
],
"formula_id": "formula_8",
"formula_text": "M c = Sample(Kmeans(D i (c)), c, m |C i | ); 21: M i = M i ∪ M c ; 22:"
},
{
"formula_coordinates": [
4,
312.31,
500.91,
212.1,
35.02
],
"formula_id": "formula_9",
"formula_text": "p k (y j ) ∝ (h u k ,y u k )∈K u j 1 y j =y u k • exp( h u j • h u k T ),(7)"
},
{
"formula_coordinates": [
5,
112.83,
504.97,
176.3,
75.25
],
"formula_id": "formula_10",
"formula_text": "ACC = 1 n n i=1 R n i (8) BW T = 1 n -1 n-1 i=1 R n i -R i i ,(9)"
}
] |
10.1093/acprof:oso/9780199547548.003.0003 | 2023-05-20 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b12",
"b7",
"b8",
"b12",
"b39",
"b6",
"b5",
"b16",
"b14",
"b27",
"b29",
"b15",
"b16",
"b4",
"b42"
],
"table_ref": [],
"text": "Maltese is a Semitic language that has been shaped by an extensive history of contact with non-Semitic languages. A large influx of Sicilian, Italian, and English words over the course of hundreds of years has influenced the Maltese lexicon and grammar, making it a prime case study for those interested in the effects of language contact on morphological systems. Semitic languages are notable for their use of root-and-pattern (a.k.a. templatic) morphology in which inflectional or derivational forms of a lexeme may be related via the non-concatenative interleaving of consonants and vowels. In Maltese, some lexemes of non-Semitic origin have integrated into the native morphology to take both concatenative as well as non-concatenative patterns of Semitic origin. Non-Semitic morphological markers have also entered the grammar and may be found on lexemes of both non-Semitic and Semitic origin.\nThis study applies methods from computational modeling and information theory to investigate factors shaping the organization of the modern Maltese lexicon. Contextualized within frameworks of analogical classification and usage-based accounts of contact-induced language change, we quantify the extent to which the phonology and etymology of Maltese lexemes are predictive of nominal plural inflection in the language. The results indicate that system-level phonology, hypothesized to capture analogical pressures, and etymology, hypothesized to capture conservative pressures that resist analogical change, are predictive of Maltese plural inflection in non-redundant ways, with phonology being more predictive than etymology overall.\nBecause Maltese is a Semitic language, we are also interested in the extent to which these factors are predictive of the type of morphology (either concatenative or non-concatenative) relating singular-plural pairs in the language. Our results show that both phonology and etymology are twice as predictive of a lexeme's plural allomorph(s) as compared to its concatenative type. This suggests that the analogical processes hypothesized to inform speakers' morphological intuitions are most sensitive to phonological similarities across surface forms, regardless of typological differences distinguishing concatenative and non-concatenative relationships. This study provides quantitative evidence for the role of analogical classification based on phonological similarity at the word level as a structuring principle of Maltese nominal plural morphology.\n2 Morphology in Contact: Maltese as a \"Hybrid\" Language?\nMaltese is a descendant of the Siculo Arabic variety spoken by settlers of the Maltese islands beginning in the year 1048 (Fabri, 2010;Brincat, 2011).\nWhile the language is Semitic with respect to its genetic classification, isolation and centuries of foreign colonization led to the development of Maltese as a distinct language shaped by Sicilian, Italian, and English influence. Written records from as early as 1240 acknowledge Maltese as its own language (Brincat, 2017), but it was not until 1934 that Maltese was declared an official language of Malta, along with English and Italian (Fabri, 2010). Italian was revoked as an official language in 1936, but its influence on the Maltese lexicon and grammar remains.\nMuch of the existing literature on Maltese describes the language as having a \"split lexicon\" or a \"hybrid morphology\" (e.g., Spagnol, 2011;Borg and Gatt, 2017). These characterizations reflect an etymological divide in the lexical stock. Semitic nouns in the language mostly form the plural with Semitic affixes or root-and-pattern templates, while non-Semitic nouns show a less strong tendency to form the plural with non-Semitic affixes. At the same time, hundreds of non-Semitic nouns inflect using Semitic patterns and are found in nearly all plural classes (Borg and Azzopardi-Alexander, 1997). Integration in the opposite direction is also found for a smaller number Semitic nouns which inflect using non-Semitic affixes. Maltese thus represents a partial, but not total, example of what has variously been called a \"stratal effect\" (Gardani, 2021) or \"code compartmentalization\" (Friedman, 2013) or \"compartmentalized morphology\" (Matras, 2015), in which native and borrowed morphological exponents in a language are restricted to applying to lexemes of the same etymological origin.\nIt is common in contact linguistics to describe outcomes of language contact as compositions of distinct linguistic systems, even in cases of extensive borrowing or codeswitching (e.g., Myers-Scotton, 1997;Gardani, 2020). Such descriptions are sometimes intended as theoretical analyses. For example, Gardani (2021) treats the stratal effect not simply as an empirically observable pattern, but as a synchronic constraint within the grammar that is psychologically real for speakers: \"... a restriction on the application domain of non-native morphological formatives in a recipient language...\" (Gardani, 2021, 132) that enforces the boundaries of etymologically-defined morphological subsystems. However, we find the a priori assumption that stratal effects reflect distinct and psychologically real morphological subsystems to be problematic inasmuch as it conflates the property to be explained -that language contact can result (to greater or lesser degree) in compartmentalized morphology -with the mechanisms that produce and reinforce that compartmentalization. Stated differently, reification of the stratal effect as a mechanism of the grammar obscures important questions: Given that speakers do not generally know the etymological origins of words, how do they classify words into morphological patterns? What is the relationship between the processes that they use to do this and the stratal effect (or lack thereof) as an empirically observable outcome of language contact?\nIn this study we examine the (partial) stratal effect found in Maltese noun morphology, examining its relationship to factors known to be important outside of contact situations to how speakers classify words into morphological patterns. In particular we analyze the relative strength of a word's phonology and etymology as predictors of its nominal plural morphology and look at the relevance of these factors for the organization of the Maltese lexicon. It is important to note that we are not interested in etymology directly and we do not assume that speakers have or use direct knowledge of the etymology of words. We instead use etymology as a way to estimate the influence of conservative forces on morphological classification. We assume that the predictive power of etymology applies to words which have retained their etymological plurals, in some cases resisting pressures to conform to other parts of the language system. The conservative forces which resist these pressures include token frequency (Krause-Lerche, 2022).\nAdditionally, as a related question, we ask whether there is evidence in Maltese for distinct morphological subsystems (\"hybrid morphology\") in theoretical terms. This question is interesting in part because characterizations of Maltese as having hybrid morphology have also suggested, sometimes explicitly, that the non-concatenative morphology native to Semitic languages should be analyzed as distinct from concatenative morphology, both Semitic and non-Semitic. Moreover, research on morphological integration in Semitic languages has tended to focus specifically on the extent to which foreign words make use of native root-and-template morphology, as compared to affixation (e.g., Bensoukas, 2018;Ziani, 2020). However, since the vast majority of suffixal allomorphs in Maltese are of Semitic origin, division of the lexicon along etymological lines does not correspond to a split according to concatenative vs. non-concatenative morphology, as is sometimes implied. We test whether morphological type is a distinct factor in the stratal effect. Specifically, we ask whether there is support for analyzing root-and-pattern (templatic) plural morphology and affixal plural morphology as distinct subsystems.\nWe compare the results of two models: the first uses a lexeme's phonology and etymology to predict its concatenative type, either affixal or templatic. The second uses the same information to predict its inflectional allomorph, i.e., the specific affix or template found on the lexeme's plural form. Comparisons across factors within each model provide insight into the extent to which phonology and etymology are informative about plural morphology, and thus are likely to have played a role in the development of the language over centuries of contact with speakers of non-Semitic languages. Comparisons across the two models offer insight into the extent to which templatic and affixal morphological patterns operate as distinct subsystems in Maltese."
},
{
"figure_ref": [],
"heading": "Analogy and Language Change",
"publication_ref": [
"b2",
"b22",
"b13",
"b23",
"b19",
"b40",
"b28",
"b26"
],
"table_ref": [],
"text": "We take an analogical approach, using the term analogy to refer broadly to any similarity-based, paradigmatic influence of one word on the morphological behavior of another. The importance of analogy as a mechanism of language change is well established in the field of historical linguistics (Anttila, 1977;Hock, 1991;Fertig, 2013;Joseph, 2013), but it is most often discussed with respect to its role in language-internal change, independent of the effects of language contact. In contact linguistics, the idea that (phonologically-based) analogy plays a role in whether and how borrowed words are morphologically integrated into a recipient language has a long history, going back to at least Haugen (1950) and Weinreich (1953). However, most analyses of lexical and morphological borrowing focus on the potential and observed outcomes of contact (see Matras and Adamou, 2020, for an overview), often with little to no discussion of the exact ways in which analogy is hypothesized to play a role.\nTo examine the role of analogy, we take a cue from Matras (2009), who proposes a usage-based model of language contact in which a multilingual individual draws on a unified repertoire of linguistic resources. In this section we elaborate on how such a perspective can help in understanding the role of analogy, specifically analogical classification, in contact-induced morphological change and the development of the Maltese lexicon."
},
{
"figure_ref": [],
"heading": "The Paradigm Cell Filling Problem",
"publication_ref": [
"b0",
"b0",
"b38",
"b18",
"b35",
"b36"
],
"table_ref": [],
"text": "Analysis of the analogical mechanism hypothesized to drive morphological integration in contact may be understood as an extension of the Paradigm Cell Filling Problem (PCFP), a line of research in theoretical morphology that seeks to identify the information available to speakers that allows them to infer and produce grammatically inflected surface forms (Ackerman et al., 2009). Most quantitative analyses of the PCFP to date take an analogical approach: speakers are hypothesized to rely on emergent similarities and paradigmatic relations among previously-acquired words in the lexicon to inform their intuitions when inflecting or processing rare or novel word forms (see, e.g., Ackerman et al., 2009;Sims and Parker, 2016;Guzmán Naranjo, 2020;Parker et al., 2022).\nMatras's ( 2009) usage-based model of language contact is directly compatible with analogical approaches to the PCFP. Since multilingual speakers are assumed to have access to a unified linguistic repertoire corresponding to all of their languages, this full repertoire may be drawn upon to make morphological generalizations. Combinations of generalizations from different languages during speech production may result in linguistic innovations or morphologically adapted \"nonce borrowings\" (Poplack et al., 1988). Over time, some of these may be conventionalized and perpetuated throughout the larger speech community, leading to contact-induced language change.\nWe may therefore specify the PCFP with respect to language contact as follows: what guides speakers' grammatical intuitions when adapting and integrating lexemes in multilingual contexts, and how may conventionalized integration of borrowed linguistic material affect the intuitions of a monolingual speaker when producing inflected word forms?"
},
{
"figure_ref": [],
"heading": "Computational Modeling of the PCFP",
"publication_ref": [
"b21",
"b41",
"b34",
"b1"
],
"table_ref": [],
"text": "A number of recent studies in computational linguistics have applied machine learning methods to analyze the kinds and amounts of information that may be available to speakers when solving the PCFP (in monolingual contexts). For example, Guzmán Naranjo (2020) uses a Long Shortterm Memory Network (LSTM, Hochreiter and Schmidhuber, 1996) to quantify the respective informativity of stem phonology, lexical semantics, and affixal exponents as predictors of nominal inflection class organization in Russian. His results indicate that while each factor contributes predictive information, more information about inflection class is contributed by stem phonology than by any individual affix. Furthermore, the contributions of the three predictors are additive, indicating a level of nonredundancy in their informativity. Williams et al. (2020) also employ the representational power of an LSTM to quantify the extent to which phonology and lexical semantics are predictive of a noun's declension class in German and Czech. As opposed to model accuracy, they measure the amount of Mutual Information, in bits, shared by phonology, semantics, and declension class systems in each language. They find that, while phonology is more predictive than semantics overall in both languages, the relative informativity of phonology and semantics varies greatly across the two languages and across individual declension classes within each language.\nDawdy-Hesterberg and Pierrehumbert (2014) take an analogical approach to modeling plural formation in Modern Standard Arabic. The authors use a Generalized Context Model (GCM, Nosofsky, 1990) to quantify the extent to which phonological factors, specifically similarities in consonant-vowel (CV) template (a.k.a. \"broken plural\" allomorph), segmental properties (in terms of natural classes), and lexical gang size (Alegre and Gordon, 1999), predict the form of a plural noun in Arabic. Their results indicate that all three factors are predictive to varying degrees, suggesting phonological representations that are both fine-grained, i.e., at the segmental level, and coarse-grained, i.e., with respect to gang size and CV template, may serve as a basis for analogical processing and morphological organization in Arabic.\nFinally, Nieder et al. (2021a,d) use both computational and psycholinguistic methods to investigate the role of analogical classification in the nominal plural system of Maltese. The authors find that plural forms in Maltese may be predicted with a reasonable degree of accuracy based on their phonological similarity to attested plural forms, modulated by the frequency distribution of plural allomorphs in the language. However, the authors do not specifically measure etymology as a predictor, leaving open the question of how non-Semitic words were integrated into the morphological system. In other words, it is unclear from their results whether phonology is predictive independently of etymology, or only as an indicator of etymological origin."
},
{
"figure_ref": [],
"heading": "Methods",
"publication_ref": [
"b41"
],
"table_ref": [],
"text": "The current study adapts the methods proposed by Williams et al. (2020) to quantify the relative contributions of phonology and etymology as predictors of inflectional organization in Maltese. We use a character-level LSTM classifier trained to make inferences about a word's plural class by abstracting over the phonology of each word form as a whole. We then quantify the influence of phonology on Maltese nominal plural inflection using Mutual Information, an information theoretic measure of interpredictability among two or more systems. We compare our results to the predictive strength of the word's etymological origin using the same measures, quantifying the balance of analogical and conservative factors hypothesized to shape the integration of foreign lexemes into the grammar."
},
{
"figure_ref": [],
"heading": "Data",
"publication_ref": [
"b17",
"b3",
"b41"
],
"table_ref": [],
"text": "This study merges data from two collections compiled by Nieder et al. (2021b,c) into a single dataset consisting of 3,174 singular-plural noun pairs. Each pair is tagged for etymological origin, either Semitic or non-Semitic. The original data was manually compiled from the MLRS Korpus Malti v. 2.0 and 3.0 (Gatt and Čéplö, 2013) and supplemented with Schembri's (2012) collection of Maltese CV templates. Etymological information was sourced from a digitized version of Aquilina's (2006) Maltese-English dictionary. Plural nouns in the data are classified as taking one of 12 different suffixes (\"sound plurals\") or 11 different non-concatenative CV templates (\"broken plurals\"), forming a nominal plural inflection system composed of 23 different inflection classes (Nieder et al., 2021b). Maltese is the only standardized Semitic language written in a Latin script, using an orthography that \"represents the phonology of the language admirably\" according to Hoberman (2007, 258). For this reason, we analyze nouns using their original orthography, as in Williams et al. (2020). Over 135 nouns in the dataset take more than one plural form. Of these, 78 nouns may take both broken and sound plurals. In this study, we account for these nouns by representing each pair separately at the allomorph level, whereas in the binary prediction model of the lexeme's concatenative type (concatenative vs. non-concatenative) we include a noun only once per type. For example, the word LIBSA 'dress' may take the sound plural libsiet and the broken plurals lbies and lbiesi. The lexeme LIBSA is therefore included in the model three times in the allomorph prediction setting, but only twice in the type prediction setting."
},
{
"figure_ref": [],
"heading": "Non-Semitic",
"publication_ref": [
"b41"
],
"table_ref": [
"tab_0"
],
"text": "Following Williams et al. (2020), we remove all classes with fewer than 20 lexemes, leaving a total of 13 plural allomorph classes in our model. Table 1 shows the full distribution of allomorphs according to etymology and concatenative type. Note that lexemes that take more than one allomorph are counted more than once."
},
{
"figure_ref": [],
"heading": "Formal Notation",
"publication_ref": [
"b41"
],
"table_ref": [],
"text": "Following Williams et al. (2020), we can define a lexeme as a tuple (w i , e i , c i ) where for the i th lexeme, w i = the lexeme's phonological form, e i = the lexeme's etymological origin, and c i = the lexeme's inflection class. We assume the lexemes follow a probability distribution p(w, e, c), approximated by the corpus. We can define the space of K inflection classes as C = {1, ..., K}, so that c i ∈ C and define C as the random variable associated with C. For a set of lexemes derived from N etymological origins, we can define an etymological space as E = {1, ..., N } so that e i ∈ E and define E as the random variable associated with E. Each noun may be associated with one of two genders g i from the space of genders G specific to Maltese. Finally, we define the space of word forms as the Kleene closure over a language's alphabet Σ, so that w i ∈ Σ*, with W as the random variable associated with Σ*."
},
{
"figure_ref": [
"fig_0"
],
"heading": "Mutual Information (MI)",
"publication_ref": [
"b10",
"b41"
],
"table_ref": [],
"text": "Mutual Information (MI) is an information theoretic measure that quantifies the degree of interpredictability among two or more systems. For example, the MI shared by the nominal plural inflection class system C and phonological system W in Maltese may be calculated as follows:\nMI(C; W ) = H(C) -H(C|W )(1)\nThis may be generalized to consider the amount of redundant information shared by inflection class, phonology, and etymology E as follows:\nMI(C; E; W ) = MI(C; W ) -MI(C; W |E) (2)\nBecause a language's grammatical gender system is known to interact with its inflectional morphology in non-deterministic ways (Corbett and Fraser, 2000), we follow Williams et al. (2020) and condition all relevant measures on gender:\nMI(C; W |G) = H(C|G) -H(C|W, G) (3)\nThe intuitive reasoning behind Equations 1 -3 may be seen in Figure 1, in which each colored circle represents H|G, the total entropy, conditioned on gender, of the three interacting systems under analysis.\nFinally, since our corpus is only a sample of the language, we note that all calculations are estimates. However, while estimates over the finite inflection class and etymology systems can be empirically calculated using the corpus, the infinite number of possible word forms in the Σ* means calculations involving W must be further approximated. Methods for estimating the entropy of both kinds of systems are described in detail in the following sections."
},
{
"figure_ref": [],
"heading": "Techniques for Estimating Entropy",
"publication_ref": [],
"table_ref": [],
"text": "We use plug-in estimation to obtain entropy values for C and E, calculating the distribution p(c) for c ∈ C (or alternatively, p(e) for e ∈ E) and using this to estimate H(C) in Equation 1 above."
},
{
"figure_ref": [],
"heading": "Approximating Conditional Entropy",
"publication_ref": [
"b9"
],
"table_ref": [],
"text": "H(C|E) may be similarly calculated using plugin estimation. However, given the infinite number of possible word forms in Σ*, an estimate for H(C|W ) cannot be calculated directly from the corpus. We therefore approximate this value using cross-entropy, which has been mathematically proven to be an upper bound on conditional entropy (Brown et al., 1992). We use the cross-entropy loss obtained from a computational model that has been trained to predict the plural class c i associated with a singular noun w i to approximate the cross-entropy of the system:\nH(C|W ) ≤ - 1 M M i=1 log q(c i |w i ) (4)\nWe note that as the amount of data in the corpus increases, i.e., as M → ∞, the above value approaches the true cross-entropy value."
},
{
"figure_ref": [],
"heading": "Normalized Mutual Information (NMI)",
"publication_ref": [],
"table_ref": [],
"text": "To compare results across models and across languages, we normalize MI values by dividing by the total entropy of the inflection class system. For example, the NMI shared by a Maltese noun's phonology and plural inflection may be calculated as:\nNMI(C; W ) = MI(C; W ) H(C) (5)"
},
{
"figure_ref": [],
"heading": "Model Details",
"publication_ref": [
"b41",
"b41"
],
"table_ref": [],
"text": "We adapt the LSTM classifier implemented in Williams et al. (2020) to estimate the probability that a plural class c is associated with a given input noun w of gender g, i.e., q(c|w, g) in Equation 4. embedded and input into the model's initial hidden state. The model is trained using Adam (Kingma and Ba, 2015) with model hyperparameters, including the number of training epochs and the number and sizes of hidden layers, optimized using the Bayesian optimization technique implemented in Williams et al. (2020). The model then learns a probability distribution that serves to approximate q(c|w, g).\nFollowing training, we test the model on a heldout dataset and use the model's cross-entropy loss to serve as an approximate upper bound on the conditional entropy H(C|W, G). We use 10-fold cross validation to make full use of the dataset for our approximations. To estimate q(c|w, e, g), we concatenate a binary character representing the word's etymology onto the end of the noun to serve as model input and follow the same procedure."
},
{
"figure_ref": [],
"heading": "Results",
"publication_ref": [],
"table_ref": [
"tab_1"
],
"text": "NMI and H(C|G) values for C defined as concatenative type and plural allomorph, respectively, are presented in Table 2. The largest NMI value we obtain, NMI(E; W |G), indicates that more than half of the information needed to predict a word's etymology is shared with its phonology. In other words, it is often not difficult to guess the origin of a Maltese word based on how it sounds. Note that this value is consistent across models, as it does not depend on C."
},
{
"figure_ref": [],
"heading": "Concatenative Type",
"publication_ref": [],
"table_ref": [
"tab_1"
],
"text": "Results for the model predicting a noun's concatenative type are in Table 2. Note first that the entropy H(C|G) of the plural inflection class system defined at the level of concatenative type is calculated to be 0.81, indicating that, given its gender, predicting whether a random Maltese noun takes concate-native or non-concatenative morphology is more predictable than chance, although not by much. We find phonology, indicated by NMI(C; W |G), to be more predictive than etymology, indicated by NMI(C; E|G). Crucially, each of these bipartite NMI values exceeds the tripartite mutual information NMI(C; E; W |G) shared across all three systems. This indicates that while a non-trivial amount of predictive information is shared across all three systems, phonology and etymology are each predictive of concatenative type in partially non-redundant ways. This suggests that both analogical and conservative forces are likely to have played a role in the development of the Maltese nominal plural system."
},
{
"figure_ref": [],
"heading": "Plural Allomorph",
"publication_ref": [],
"table_ref": [
"tab_1"
],
"text": "In an analogical model of inflection in which singular inflected forms and their plural counterparts share a direct relationship in the lexicon, the predictive principles structuring the morphological system are expected to be most evident when defining an inflection class system at the level of the allomorph.\nWe first note that the entropy H(C|G) calculated over the plural class distribution defined at the allomorph level is nearly three times as high as the entropy of C when defined as a noun's concatenative type. This is reflective of the higher degree of unpredictability associated with a non-uniform distribution of nouns over a greater number of inflection classes. When comparing across the allomorph and concatenative type models it is thus important to normalize for the fact that predicting allomorphs is more difficult than predicting concatenative type. However, even calculations normalized in this way show that the interpredictability among phonology, etymology, and plural inflection, indicated by the NMI values in Table 2, are all twice as high at the allomorph level as they are for concatenative type. In other words, a noun's singular form reduces the relative uncertainty about its plural allomorph twice as much as it reduces the uncertainty about whether that allomorph is concatenative. This suggests the analogical and conservative pressures hypothesized to shape morphological organization are more sensitive to correspondences at the word level than to typological similarities with respect to concatenativity.\nAdditionally, the general tendency found at the level of concatenative type still follows when classes are defined at the level of individual allomorphs: phonology shares more information with inflection class than does etymology, with each factor contributing some amount of non-redundant information. This illustrates one key advantage of the methods employed in this study, namely the ability to disentangle the independent contributions of either predictor from the degree to which both exert redundant organizational pressure towards the same end.\nFor example, given the fact that phonology and etymology are themselves mutually informative, we cannot uniquely interpret either bipartite measure of MI, that is, NMI(C; W |G) or NMI(C; E|G), as indicative of the forces hypothesized to shape the integration of linguistic material in contact. Rather, evidence for analogical structuring of the Maltese plural system at the allomorph level is specifically indicated by the positive difference between NMI(C; W |G) and NMI(C; E; W |G). Conservative pressures, such as those associated with high token-frequency items (Krause-Lerche, 2022), are similarly indicated by the extent to which NMI(C; E|G) exceeds NMI(C; E; W |G)."
},
{
"figure_ref": [
"fig_1",
"fig_1",
"fig_2"
],
"heading": "Variation Across Allomorph Classes",
"publication_ref": [
"b41"
],
"table_ref": [
"tab_2"
],
"text": "Closer examination of the model's predictions reveals an effect of type frequency, with larger inflection classes predicted more often than smaller classes. Table 3 reports the accuracy of all models in which singular noun phonology W is a predictor. Since all models achieve an overall accuracy above a majority baseline, the NMI values we obtain may be reliably interpreted as empirical minimums. However, as can be seen in Figure 2, the model's incorrect predictions do not clearly distinguish between sound and broken classes; nouns with a sound plural allomorph may be misclassified as taking a broken plural template, and nouns taking a broken plural may be incorrectly predicted to take a sound plural.\nIf speakers are sensitive to differences between concatenative and non-concatenative allomorphs grouped into high-level macro classes (morphological subsystems), we might expect some degree of observable within-class coherence with respect to either or both of the phonology and etymology of words exhibiting a particular morphological behavior. Specifically, we would expect a pattern of predictions in which the LSTM is able to first Instead, as seen in Figure 2, we do not find such evidence. Rather, we find evidence for coherence at the allomorph level, specifically, for phonological patterns as a predictor of inflectional organization and driver of inflectional behavior at the allomorph level.\nFinally, as in Williams et al. (2020), we also conduct an analysis of the partial Pointwise Mutual Information (PMI) shared between phonology W and class C with respect to the surprisal H(C = c|G) for each class, defined at the allomorph level. Figure 3 shows this distribution, with allomorph classes presented in order of increasing type frequency (and thus decreasing surprisal). We note that Maltese noun classes are each only par- "
},
{
"figure_ref": [],
"heading": "Discussion",
"publication_ref": [],
"table_ref": [],
"text": "In this paper we used an LSTM to help estimate the kinds and amounts of information that may be available to speakers when \"solving\" the PCFP. Overall, our results provide quantitative evidence for the role of both word phonology and etymology (as a stand-in for conservative factors) in shaping the Maltese lexicon.\nSpecifically, we found that the extent to which a Maltese singular noun's phonology predicts its plural morphology exceeds that of etymology in nonredundant ways. This suggests that analogical pressures from phonological correspondences across the lexicon shape nominal plural inflection in Maltese, independently of the etymological source language for some word or morphological pattern.\nOur results also show an independent contribution of etymology as a predictor. We hypothesize that this captures conservative pressures theorized to resist analogical change, including token frequency (Krause-Lerche, 2022). It may also reflect associative correlations from the use of lexemes of a common etymology in similar contexts, strengthening their coherence as a subsystem in the multilingual repertoire and encouraging the maintenance of a noun's original morphology. Further work is needed to investigate these possibilities.\nIn language contact situations such as that of Maltese, it is likely that an influx of foreign lexemes and increased productivity of foreign affixes affect both the size and character (e.g., phonology) of nominal plural classes relative to each other over time. This in turn is likely to affect subsequent classification and integration of words into the inflectional morphology of the language.\nIn general, our results do not support characterizations of Maltese in which concatenative and non-concatenative morphologies co-exist as discrete systems within the lexicon. While a singular noun's phonology and etymology are each somewhat predictive of its concatenative type, they are twice as predictive of the actual plural allomorph(s) with which the lexeme is associated. This suggests that systematic relationships at the word level organize the morphology of Maltese, in turn shaping the language as new words are integrated and inflected."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "This study extends previous work in information theory, computational modeling, and theoretical morphology to provide quantitative evidence for the role of phonology as an analogical force in the morphological organization of Maltese. We ground this in a usage-based account of multilingualism and contact-induced change in which speakers are hypothesized to make use of analogical reasoning, among other language-general cognitive functions, when integrating novel words and patterns within a unified linguistic repertoire. The same processes that guide synchronic language use are proposed to be responsible for the diachronic effects of contactinduced language change. Specifically, it is hypothesized that speakers draw on similarities across multiple dimensions -including but not limited to phonological patterns, semantic and indexical meaning, pragmatic function, and contexts of useto collaboratively construct and adapt grammatical systems of linguistic communication over time.\nIn the case of Maltese, our findings indicate that while a lexeme's phonology and etymology are themselves highly interpredictable, each contributes non-redundant information to reduce uncertainty when predicting the lexeme's plural inflection. While the etymology of a noun is somewhat predictive of its plural inflection, the word's phonology plays a much greater role. This synchronic analysis has diachronic implications. Our results suggest that analogical pressures from phonological similarities across the lexicon may have guided speakers' inflectional behavior when code mixing over the course of the development of the language to result in the conventionalized forms observed in modern Maltese. However, further diachronic study is needed to confirm this interpretation.\nContrary to a hypothesis in which concatenative and non-concatenative systems operate as separate subsystems within a \"split\" or \"hybrid\" morphology, our results indicate correspondences at the level of individual wordforms and affixes are driving speakers' morphological behavior. Specifically, the phonology and etymology of a lexeme are twice as predictive of its plural allomorph than its concatenative type. Further investigation into Maltese nouns attested to take plural forms of both concatenative types may provide additional insight into the ways in which concatenative type affects speakers' behavior, if at all. Future work should also consider additional factors known to shape inflection class systems, for example by integrating semantic word vectors into the model. Finally, additional comparisons implementing these methods across corpora in a variety of languages will continue to shed light on the factors shaping morphological systems cross-linguistically. "
},
{
"figure_ref": [],
"heading": "A Nominal Plural Allomorphs in Maltese",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Acknowledgments",
"publication_ref": [],
"table_ref": [],
"text": "We thank Jessica Nieder and Adam Ussishkin for generously sharing an abundance of digital resources and helpful feedback, Sarah Caruana for her insight into the Maltese language, and Christian Clark and Andrew Duffy for their contributions to an initial version of this project.\nThis material is based on work supported by the National Science Foundation under grant BCS-2217554 (Neural discovery of abstract inflectional structure, PI Micha Elsner, Co-PI Andrea Sims)."
}
] | Maltese is often described as having a hybrid morphological system resulting from extensive contact between Semitic and Romance language varieties. Such a designation reflects an etymological divide as much as it does a larger tradition in the literature to consider concatenative and non-concatenative morphological patterns as distinct in the language architecture. Using a combination of computational modeling and information theoretic methods, we quantify the extent to which the phonology and etymology of a Maltese singular noun may predict the morphological process (affixal vs. templatic) as well as the specific plural allomorph (affix or template) relating a singular noun to its associated plural form(s) in the lexicon. The results indicate phonological pressures shape the organization of the Maltese lexicon with predictive power that extends beyond that of a word's etymology, in line with analogical theories of language change in contact. | Analogy in Contact: Modeling Maltese Plural Inflection | [
{
"figure_caption": "Figure 1 :1Figure 1: Tripartite Mutual Information",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: Confusion matrix: predicting plural allomorph from singular phonology and gender",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: Partial Pointwise Mutual Information (PMI) shared by word form and class for each allomorph class",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Distribution of Maltese nominal plural allomorphs by lexeme etymology and concatenative type",
"figure_data": "SemiticTotalLexemeLexeme(%)Non-Semitic Affix1,2742142%Semitic Affix41668435%Semitic Template24053723%Total (%)62%38%100%",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Normalized Mutual Information measures for plural class C defined with respect to TYPE vs.",
"figure_data": "TYPE ALLO.",
"figure_id": "tab_1",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Model accuracy for all models predicting Etymology E or Plural Class C (Type vs. Allomorph) using the Phonology W of singular nouns in Maltese identify a lexeme's concatenative type before predicting, possibly incorrectly, an allomorph of that specific type.",
"figure_data": "TargetModelAccuracyETYM. (E)MI(E; W |G) Baseline0.90 0.62MI(C; W |G)0.80TYPE (C)MI(C; E; W |G)0.81Baseline0.77MI(C; W |G)0.65ALLOMORPHMI(C; E; W |G)0.68(C)Baseline0.40",
"figure_id": "tab_2",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Sound plural allomorphs in Maltese, fromNieder et al. (2021b) ",
"figure_data": "Sound PluralSingular PluralGlossAllomorphkartakarti'paper'-iommommijiet 'mother'-ijietrixarixiet'feather'-ietgiddieb giddieba 'liar'-ameèlusmeèlusin 'freed'-inkuxinkuxins'cushions' -striqtriqat'street'-atsidsidien'owner'-ienbaèribaèrin'sailor'-nèatièatjin'guilty'-jinspallaspallejn'shoulder' -ejnsieqsaqajn'foot'-ajnqiegèqiegèan'bottom'-anBroken PluralSingular PluralGlossAllomorphfardalfradal'apron'CCVVCVCbirrabirer'beer'(C)CVCVCkbirkbar'big'CCVVCftiraftajjar'type of bread' CCVjjVCbitèabtieèi'yard'CCVVCVsiderisdra'chest'VCCCVmaridmorda'sick person'CVCCVgèodda gèodod'tool'(gè)VCVCelfeluf'thousand'VCVCgèarefgèorrief 'wise man'CVCCVVC(V)gèamagèomja'blind person' (gè)VCCV",
"figure_id": "tab_3",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Broken plural allomorphs in Maltese, fromNieder et al. (2021b) ",
"figure_data": "",
"figure_id": "tab_4",
"figure_label": "5",
"figure_type": "table"
}
] | Sara Court; Andrea D Sims; Micha Elsner | [
{
"authors": "James P Farrell Ackerman; Robert Blevins; Malouf",
"journal": "Oxford University Press",
"ref_id": "b0",
"title": "Parts and wholes: Patterns of relatedness in complex morphological systems and why they matter",
"year": "2009"
},
{
"authors": "Maria Alegre; Peter Gordon",
"journal": "Brain and Language",
"ref_id": "b1",
"title": "Rule-based versus associative processes in derivational morphology",
"year": "1999"
},
{
"authors": "Raimo Anttila",
"journal": "Mouton",
"ref_id": "b2",
"title": "Analogy",
"year": "1977"
},
{
"authors": "Joseph Aquilina",
"journal": "Midsea Books",
"ref_id": "b3",
"title": "Concise Maltese-English, English-Maltese Dictionary",
"year": "2006"
},
{
"authors": "Karim Bensoukas",
"journal": "International Journal of Arabic Linguistics",
"ref_id": "b4",
"title": "Concurrent cognate and contact-induced plural traits in Afro-Asiatic: Amazigh id-and Arabic -at plurals",
"year": "2018"
},
{
"authors": "Albert J Borg; Marie Azzopardi-Alexander",
"journal": "Routledge",
"ref_id": "b5",
"title": "Maltese",
"year": "1997"
},
{
"authors": "Claudia Borg; Albert Gatt",
"journal": "Association for Computational Linguistics",
"ref_id": "b6",
"title": "Morphological analysis for the Maltese language: The challenges of a hybrid system",
"year": "2017"
},
{
"authors": "Joseph M Brincat",
"journal": "Midsea Books",
"ref_id": "b7",
"title": "Maltese and other languages: A linguistic history of Malta",
"year": "2011"
},
{
"authors": "Joseph M Brincat",
"journal": "Lexicographica",
"ref_id": "b8",
"title": "Maltese: Blending Semitic, Romance and Germanic lexemes",
"year": "2017"
},
{
"authors": "F Peter; Stephen A Brown; Vincent J Della Pietra; Jennifer C Della Pietra; Robert L Lai; Mercer",
"journal": "Computational Linguistics",
"ref_id": "b9",
"title": "An estimate of an upper bound for the entropy of English",
"year": "1992"
},
{
"authors": "G Greville; Norman M Corbett; Fraser",
"journal": "Cambridge University Press",
"ref_id": "b10",
"title": "Gender assignment: A typology and a model",
"year": "2000"
},
{
"authors": "Lisa Garnand; Dawdy-Hesterberg ; Janet Breckenridge; Pierrehumbert ",
"journal": "Language, Cognition and Neuroscience",
"ref_id": "b11",
"title": "Learnability and generalisation of Arabic broken plural nouns",
"year": "2014"
},
{
"authors": "Ray Fabri",
"journal": "Revue belge de Philologie et d'Histoire",
"ref_id": "b12",
"title": "Maltese",
"year": "2010"
},
{
"authors": "L David; Fertig",
"journal": "Edinburgh University Press",
"ref_id": "b13",
"title": "Analogy and morphological change",
"year": "2013"
},
{
"authors": "A Victor; Friedman",
"journal": "Romani Studies",
"ref_id": "b14",
"title": "Compartmentalized grammar: The variable (non)-integration of Turkish verbal conjugation in Romani dialects",
"year": "2013"
},
{
"authors": "Francesco Gardani",
"journal": "Morphology",
"ref_id": "b15",
"title": "Borrowing matter and pattern in morphology: An overview",
"year": "2020"
},
{
"authors": "Francesco Gardani",
"journal": "Word Structure",
"ref_id": "b16",
"title": "On how morphology spreads",
"year": "2021"
},
{
"authors": "Albert Gatt; Slavomír Čéplö",
"journal": "",
"ref_id": "b17",
"title": "Digital corpora and other electronic resources for Maltese",
"year": "2013"
},
{
"authors": "Guzmán Matías; Naranjo",
"journal": "Morphology",
"ref_id": "b18",
"title": "Analogy, complexity and predictability in the Russian nominal inflection system",
"year": "2020"
},
{
"authors": "Einar Haugen",
"journal": "Language",
"ref_id": "b19",
"title": "The analysis of linguistic borrowing",
"year": "1950"
},
{
"authors": "Robert D Hoberman",
"journal": "Morphologies of Asia and Africa",
"ref_id": "b20",
"title": "Maltese morphology",
"year": "2007"
},
{
"authors": "Sepp Hochreiter; Jürgen Schmidhuber",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b21",
"title": "LSTM can solve hard long time lag problems",
"year": "1996"
},
{
"authors": "Hans Henrich; Hock ",
"journal": "De Gruyter Mouton",
"ref_id": "b22",
"title": "Principles of historical linguistics",
"year": "1991"
},
{
"authors": "Brian D Joseph",
"journal": "Working papers in linguistics",
"ref_id": "b23",
"title": "On phonically based analogy",
"year": "2013"
},
{
"authors": "P Diederik; Jimmy Kingma; Ba",
"journal": "",
"ref_id": "b24",
"title": "Adam: A method for stochastic optimization",
"year": "2015"
},
{
"authors": "Anne Krause-Lerche",
"journal": "Corpus Linguistics and Linguistic Theory",
"ref_id": "b25",
"title": "Conservation in ongoing analogical change: The measurement and effect(s) of token frequency",
"year": "2022"
},
{
"authors": "Yaron Matras",
"journal": "Cambridge University Press",
"ref_id": "b26",
"title": "Language contact",
"year": "2009"
},
{
"authors": "Yaron Matras",
"journal": "De Gruyter Mouton",
"ref_id": "b27",
"title": "Why is the borrowing of inflectional morphology dispreferred?",
"year": "2015"
},
{
"authors": "Yaron Matras; Evangelia Adamou",
"journal": "Routledge",
"ref_id": "b28",
"title": "Borrowing",
"year": "2020"
},
{
"authors": "Carol Myers-Scotton",
"journal": "Clarendon",
"ref_id": "b29",
"title": "Duelling languages: Grammatical structure in codeswitching",
"year": "1997"
},
{
"authors": "Jessica Nieder; Fabian Tomaschek; Enum Cohrs; Ruben Van De Vijver",
"journal": "Language, Cognition and Neuroscience",
"ref_id": "b30",
"title": "Modelling Maltese noun plural classes without morphemes",
"year": "2021"
},
{
"authors": "Jessica Nieder; Fabian Tomaschek; Enum Cohrs; Ruben Van De Vijver",
"journal": "",
"ref_id": "b31",
"title": "Modeling Maltese noun plural classes without morphemes: Supplemental materials",
"year": "2021"
},
{
"authors": "Jessica Nieder; Ruben Van De; Yu-Ying Vijver; R H Chuang; Baayen",
"journal": "",
"ref_id": "b32",
"title": "A discriminative lexicon approach to word comprehension, production and processing: Maltese plurals: Supplemental materials",
"year": "2021"
},
{
"authors": "Jessica Nieder; Ruben Van De; Holger Vijver; Mitterer",
"journal": "Morphology",
"ref_id": "b33",
"title": "d. Knowledge of Maltese singular-plural mappings",
"year": "2021"
},
{
"authors": "Robert M Nosofsky",
"journal": "Journal of Mathematical Psychology",
"ref_id": "b34",
"title": "Relations between exemplar-similarity and likelihood models of classification",
"year": "1990"
},
{
"authors": "Jeff Parker; Robert Reynolds; Andrea D Sims",
"journal": "Cambridge University Press",
"ref_id": "b35",
"title": "Network structure and inflection class predictability: Modeling the emergence of Marginal Detraction",
"year": "2022"
},
{
"authors": "Shana Poplack; David Sankoff; Christophe Miller",
"journal": "Linguistics",
"ref_id": "b36",
"title": "The social correlates and linguistic processes of lexical borrowing and assimilation",
"year": "1988"
},
{
"authors": "Tamara Schembri",
"journal": "Brockmeyer",
"ref_id": "b37",
"title": "The broken plural in Maltese: A description",
"year": "2012"
},
{
"authors": "Andrea D Sims; Jeff Parker",
"journal": "Word Structure",
"ref_id": "b38",
"title": "How inflection class systems work: On the informativity of implicative structure",
"year": "2016"
},
{
"authors": "Michael Spagnol",
"journal": "",
"ref_id": "b39",
"title": "A tale of two morphologies: Verb structure and argument alternations in Maltese",
"year": "2011"
},
{
"authors": "Uriel Weinreich",
"journal": "De Gruyter Mouton",
"ref_id": "b40",
"title": "Languages in contact: Findings and problems",
"year": "1953"
},
{
"authors": "Adina Williams; Tiago Pimentel; Hagen Blix; D Arya; Eleanor Mccarthy; Ryan Chodroff; Cotterell",
"journal": "Association for Computational Linguistics",
"ref_id": "b41",
"title": "Predicting declension class from form and meaning",
"year": "2020"
},
{
"authors": "Zoubida Ziani",
"journal": "International Journal of Arabic Linguistics",
"ref_id": "b42",
"title": "The morphology of borrowings and its relevance to lexical organization in Moroccan Arabic",
"year": "2020"
}
] | [
{
"formula_coordinates": [
5,
344.83,
235.88,
179.58,
9.81
],
"formula_id": "formula_0",
"formula_text": "MI(C; W ) = H(C) -H(C|W )(1)"
},
{
"formula_coordinates": [
5,
311.6,
304.77,
212.81,
9.81
],
"formula_id": "formula_1",
"formula_text": "MI(C; E; W ) = MI(C; W ) -MI(C; W |E) (2)"
},
{
"formula_coordinates": [
5,
321.06,
404.33,
203.35,
9.81
],
"formula_id": "formula_2",
"formula_text": "MI(C; W |G) = H(C|G) -H(C|W, G) (3)"
},
{
"formula_coordinates": [
6,
105.67,
440.98,
183.46,
33.71
],
"formula_id": "formula_3",
"formula_text": "H(C|W ) ≤ - 1 M M i=1 log q(c i |w i ) (4)"
},
{
"formula_coordinates": [
6,
119.77,
626.97,
169.36,
24.43
],
"formula_id": "formula_4",
"formula_text": "NMI(C; W ) = MI(C; W ) H(C) (5)"
}
] |
10.18653/v1/2020.acl-main.632 | 2023-11-08 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b21",
"b3",
"b26",
"b32",
"b2",
"b8",
"b20",
"b27",
"b12",
"b25",
"b24",
"b23",
"b14",
"b11",
"b19",
"b19",
"b11",
"b11",
"b9",
"b13",
"b30",
"b17",
"b20",
"b6",
"b7",
"b22",
"b15",
"b27",
"b12",
"b12"
],
"table_ref": [],
"text": "Reliable analysis of arguments in natural language holds the promise to support applications such as automated grading (Ludwig et al., 2021), and tackling misinformation and targeted speech (Alhindi, Tariq, 2023). Computational argument analysis has been relatively popular through tasks like argument extraction (Chakrabarty et al., 2019), evidence mining (Rinott et al., 2015), relation assignment (Trautmann et al., 2020), writing support (Stab and Gurevych, 2014b) and claim generation (Bilu and Slonim, 2016). A particularly challenging task is argument quality assessment (Fromm et al., 2022), which addresses the cogency, effectiveness, and reasonableness of an argument (Wachsmuth et al., 2017b) pertaining to a topic. Assessing the quality of the argument involves analyzing the objective evidence, relevant assumptions, and structural soundness, making the overall task difficult.\nResearch on argument quality assessment has focused on extracting textual patterns using various learning frameworks and content features (Lauscher et al., 2022). It has been widely recognized that contextualizing arguments with implicit knowledge, such as extracting claim revisions (Skitalinskaya et al., 2021) and generating explicit conclusions (Gurcke et al., 2021) can be informative for reasoning models. However, we note that: 1) these methods fail to generalize to novel arguments where this information is not available, and 2) no prior work has considered jointly a comprehensive set of such contextualization strategies.\nTo bridge these gaps, we propose a novel framework called SPARK (Scoring the Pragmatics of Arguments via Relevant Knowledge), which incorporates augmentation strategies based on a large language model (LLM), GPT 3.5 (OpenAI, 2022), and elements from argumentation literature (Nickerson, 2020;Mulyati et al., 2023;Harvey, 2009), specifically, feedback, assumptions, arguments with similar quality, and counter-arguments. SPARK processes the original argument and topic and its augmentations separately using a dualencoder Transformer architecture with a multi-head cross-attention layer. We demonstrate the effectiveness of SPARK 's augmentations and architecture using both in-domain and out-of-domain datasets. Our entire code is made available at (anonymized).\nTask formulation. Inspired by prior work (Gretz et al., 2020;Lauscher et al., 2020), we formalize argumentation quality assessment as a regression task of predicting the quality of a natural language argument. Given a topic and an argument, we consider three quality indicators (Lauscher et al., 2020): 1) cogency, which evaluates the relevance and sufficiency of the argument's premise in relation to the conclusion, 2) effectiveness, which measures the argument's persuasive power based on factors like arrangement, clarity, and appropriateness, and 3) reasonableness, which determines the argument's ability to resolve the debate's issue (Wachsmuth et al., 2017a). The overall quality of an argument can be estimated by averaging these three metrics (Gretz et al., 2020). Connection to prior studies. The introduction of benchmarks for argument quality (Stab and Gurevych, 2014a;Gretz et al., 2020) has inspired various methods based on logistic regression (Ghosh et al., 2016), fully connected and recurrent neural networks (Habernal and Gurevych, 2016), and fine-tuned Transformers (Toledo et al., 2019). Hulpus et al. (2019) explained that contextualizing an argument with implicit knowledge is essential to understanding its quality. Lauscher et al. (2022) categorized knowledge used by current argument assessment research so far under linguistic, task-specific, and argument-specific sections. To mimic human reasoning over arguments, prior work has incorporated users' prior beliefs as predictors of argument persuasiveness (Durmus and Cardie, 2018), trained classifiers for different audience groups (El Baff et al., 2020), utilized user history to predict persuasion (Al Khatib et al., 2020), augmented arguments with supporting or refuting documents (Marro et al., 2022), and augmented arguments with visual features (Hasan et al., 2021). Most similar to SPARK, Skitalinskaya et al. (2021) leverage comparison between revisions of the same claims, while Gurcke et al. (2021) generate conclusions to assess argument sufficiency. However, revisions are rarely provided for novel arguments, whereas generated conclusions are argument-specific and may not generalize well (Gurcke et al., 2021). Addressing prior work limitations, SPARK implements four wellmotivated augmentation strategies to enhance novel arguments and utilizes an attention-based dual encoder model for effective reasoning."
},
{
"figure_ref": [
"fig_0",
"fig_0",
"fig_0",
"fig_0",
"fig_1"
],
"heading": "SPARK",
"publication_ref": [
"b23",
"b24",
"b27",
"b16",
"b14",
"b10",
"b33"
],
"table_ref": [],
"text": "Augmentation strategies. We devise four augmentation techniques to contextualize arguments. We generate the augmentations by prompting GPT-3.5 (OpenAI, 2022) (see appendix for details). Feedback. Constructive feedback in the form of comments and suggestions helps comprehension and domain knowledge acquisition (Mulyati et al., 2023). We hypothesize that assessing argument strengths and weaknesses helps argument ranking, as in Figure 1, where feedback identifies emotional appeal, insufficient evidence, and generalization. We prompt the LLM to generate writing feedback for a topic-argument pair in a zero-shot setting. Assumptions. Unstated assumptions frequently introduce bias in arguments (Nickerson, 2020). Making assumptions explicit can reveal these hidden biases, which may aid in assessing argument persuasiveness and relevance. One such assumption in Figure 1 is that people do not take driving seriously. We employ an LLM to extract the argument's underlying assumptions in a zero-shot setting. Similar-quality instance. Inspired by prior work on claim revisions (Skitalinskaya et al., 2021), we hypothesize that retrieving arguments with similar quality at training time leads to generalizable model learning. For this purpose, we derive a synthetic argument with similar reasonableness, cogency, and effectiveness to the original one (Figure 1). We generate this synthetic argument in a few-shot setting, where the LLM has access to example arguments alongside their quality scores covering the full 1-5 range. Since this augmentation uses ground-truth information that is not available during inference, we randomly replace synthetic arguments with None at training time with a probability of P = 0.5, thus familiarizing the model with the absence of similar arguments during testing where it only sees None. This technique is similar to distillation (Hinton et al., 2015), where the encoder's (student) goal is to utilize the LLM's (teacher) ranking-based argument generations (soft labels) to learn the argument quality ranking task. Counter-arguments. Counter-arguments provide objections, alternatives, and doubts of a skeptical reader (Harvey, 2009). We expect that contrasting the strengths and weaknesses of two opposite arguments will aid quality assessment. The example counter-argument in Figure 1 makes a firm claim that alternative activities (e.g., eating) lead to distractions, and solely blaming phones is unfair. We ask the LLM to provide a counter-argument for a topic and an argument in a zero-shot setting. Dual-encoder architecture. To consider the argument together with the augmentations, we employ a dual BERT encoder (Figure 2) as an improvement to the architecture by Gillick et al. (2018). The first encoder embeds the topic and argument, whereas the second embeds the augmentations. The second encoder can store individual augmentations or their concatenation, arbitrarily fixed to Similar quality argument [SEP] Feedback [SEP] Assumptions [SEP] Counter-argument. Notably, the dual encoder can effectively store all of the augmentation data without truncating information in practically all cases (see subsection A.4 for the augmentation lengths). We use a multi-head crossattention layer (Vaswani et al., 2017) to enable the model to weigh each augmentation according to the argument-topic pair. We pass the attention outputs to a mean pooler, whose output is fed into three separate regressor heads, one per quality metric."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Baselines",
"publication_ref": [
"b5",
"b37",
"b18",
"b4",
"b31"
],
"table_ref": [],
"text": "Scoring models. We compare our dual BERT with a standard BERT model to study the effect of disjoint embeddings. We use bert-base-uncased (Devlin et al., 2019) as an encoder. We compare dual BERT to XLNet (Yang et al., 2019), as they can both handle more than 512 tokens. Finally, we utilize GPT-3.5 in a zero-shot setting to gauge the model's ability to accomplish the task directly, without a dual encoder. We provide GPT-3.5 with the definitions of each metric and prompt it to individually rate each argument by a float between 1 and 5 with respect to the topic. Alternative augmentation strategies. We evaluate the impact of using all augmentations together or one at a time, against ablated baselines without augmentations. We include two alternative augmentation methods: Wikipedia paragraphs extracted using dense passage retrieval (DPR) (Karpukhin et al., 2020), and augmentations generated using smaller models, namely, Flan-T5-XL (Chung et al., 2022) and Llama-2 (7B) (Touvron et al., 2023)."
},
{
"figure_ref": [],
"heading": "Datasets and Evaluation",
"publication_ref": [
"b19",
"b11",
"b19"
],
"table_ref": [],
"text": "We use GAQCorpus (Lauscher et al., 2020) as our training dataset for its diversity of domains (reviews, QA, and debates) and quality metrics (cogency, effectiveness, and reasonableness). We also use IBM-30K (Gretz et al., 2020) to test SPARK 's generalization on out-of-domain data. For IBM-30K, we perform weighted averaging (WA) of the three metric scores, which is supported by the high correlation between IBM-30K's WA and the GAQ-Corpus metrics (Lauscher et al., 2020). We report the Pearson (ρ) and Spearman (σ) correlation coefficients between the predictions and ground truth."
},
{
"figure_ref": [],
"heading": "Results",
"publication_ref": [],
"table_ref": [
"tab_0"
],
"text": "We investigate whether augmentations help language models assess the quality of arguments more effectively (Q1); how augmentation strategies compare to each other (Q2); and whether human quality judgments align with their model utility (Q3).\nEffect of augmentation on in-domain performance (Q1). Table 1 shows that the overall best-performing combination uses Dual BERT with all four augmentations combined, which improves the Spearman correlation over the baseline BERT by 0.08-0.17 across the three metrics. The improvement is the largest for effectiveness, where the Spearman correlation increases by 61%. While both single BERT and XLNet benefit from SPARK 's augmentations as well, their performance is consistently lower than using the dual encoder. GPT-3 alone often performs better than the other baselines but lags significantly behind SPARK, showing the importance of the dual encoder. Among the augmentations, the benefit of our four augmentations declines when using a smaller generative model (Flan T5 and Llama-2), augmenting via DPR, or using the dual BERT with a masked second encoder. Thus, merely adding text to the second encoder does not by itself bring higher performance. The gap between SPARK and the baselines increases in the zero-shot setting on the IBM-Rank-30K dataset. Here, augmentation with DPR, Flan-T5, and Llama-2 is consistently inferior, as is the SPARK augmentation of the singleencoder methods. In summary, SPARK effectively combines dual encoding and data augmentation for strong task accuracy and generalization."
},
{
"figure_ref": [],
"heading": "Comparison of augmentation variants (Q2).",
"publication_ref": [],
"table_ref": [
"tab_1"
],
"text": "On the in-domain task, the dual encoder performs the best when it has access to the information from the four SPARK augmentations simultaneously. Among them, feedback is most effective for predicting argument cogency and reasonableness as it exposes flaws that directly relate to these metrics. Meanwhile, contextualizing through similar-quality arguments is optimal for predicting effectiveness, which we attribute to illustrating the connection between quality scores and the argument structure, format, and wording. On out-ofdomain data, the best performance is obtained by the feedback-augmented dual BERT, which even outperforms using all augmentations. This is illustrated in Table 3, where the first two arguments receive positive feedback, with space for improvement by further elaboration or addressing of alternatives, directing SPARK to increase its score.\nThe third argument receives more critical feedback, causing SPARK to decrease its score. While GPT-3.5 is generally able to highlight the salient points of an argument and provide valid criticism, we also note an occasional bias of the model towards maintaining a neutral or positive argument stance (as in the case of the libertarianism argument).\nHuman judgment of augmentations (Q3). To validate the alignment of the augmentations with human utility, we performed a human study where we asked participants to score the validity, infor-mativeness, and relevance of each augmentation strategy. The participants were asked to score 50 randomly sampled in-domain data points on these three metrics using a Likert scale of 1 (lowest) to 5 (highest). The results in Table 2 show that the augmentations are perceived by people as highly valid, informative, and relevant. Assumptions and counter-arguments were found to be consistently more valid, informative, and relevant than the other augmentations. Curiously, the participants judged the feedback informativeness to be lower, explaining that it often summarizes the argument instead of giving writing suggestions. This finding provides a cue that highly effective augmentations for models may not be perceived as informative by people. SPARK alleviates this issue by effectively combining the complementary augmentations and delegating the weighting of their utility to the model."
},
{
"figure_ref": [],
"heading": "Conclusions",
"publication_ref": [],
"table_ref": [],
"text": "This paper enhances argument quality estimation models by providing contextualized feedback, inferred assumptions, and similar quality arguments or counter-arguments. We employ a dual-encoder Transformer to compare the argument and additional evidence effectively. Experimental results demonstrate that the best performance is achieved with the combination of all augmentations, indicating their complementary insights. SPARK outperforms single BERT, and XLNet in handling longer sequences, surpassing baselines and alternative augmentations in both in-and out-of-domain scenarios. Feedback augmentation is the single most effective augmentation strategy for models, despite being scored the least informative by human participants.\nThe limitations in this paper primarily stem from the generative capabilities and hallucination tendencies of the LLM used for augmentation. For example, despite constraining the output format for the assumption augmentation task, the LLM still generates \"No assumptions\" as output after listing a set of valid assumptions. As future work, further augmentation studies must be performed to analyze and improve the prompts to minimize the misunderstanding and biases of the LLM. Moreover, evaluating on a larger set of domains and other argumentative tasks such as logical fallacy detection is an important next step in investigating the generalizability of our method. While in theory, SPARK can be directly applied to such tasks, it remains to be seen to which extent the current architecture and augmentation strategies will generalize to other tasks.\nFinally, our analysis in this paper focuses on the comparison of methods and strategies, yet, we do not dive deep into the specific differences in performance across the three quality metrics which is an important future direction.\nTable 3 shows augmentation examples for one topic-argument pair. The topic questions if cell phones distract people and the argument agrees with it in context to distracted driving due to cell phones.\nFeedback: The feedback discusses how the argument takes a clear position but appears overly emotional while answering, which may influence the persuasiveness of the argument. Additionally, the feedback regarding lack of evidence other than personal experience and the lack of discussion on phones causing a distraction beyond driving aim to help improve understanding while ensuring that the Feedback -The argument takes a clear position that cell phones cause distraction while driving.\n-The writer expresses strong emotions regarding the topic, which may influence the persuasiveness of the argument.\n-The argument does not provide any evidence or supporting details beyond the writer's personal experience and emotions.\n-The argument does not address whether cell phones cause distractions in other contexts beyond driving. Assumptions -Cell phones cause a distraction when driving.\n-Bluetooth or speaker use of cell phones can also be a distraction while driving.\n-People don't take driving as seriously as they should. Similar quality argument Yes, cell phones can be a major distraction to people. People are constantly checking their phones for notifications, messages, and other tasks that can take away from what they should be focusing on. This can lead to people not paying attention in class, not being productive at work, and not being present in conversations. Cell phones are an amazing invention, but they can also be a major distraction."
},
{
"figure_ref": [],
"heading": "Counter-Argument",
"publication_ref": [],
"table_ref": [],
"text": "While it is true that distracted driving can be a major problem, it is unfair to blame cell phones as the sole cause of distraction. Many other factors can also contribute to distracted driving, such as eating, applying makeup, or even talking with passengers. Furthermore, cell phones can also be used in a responsible manner while driving, such as with the use of hands-free devices or by pulling over to take a call. Therefore, instead of demonizing cell phones, we should focus on educating drivers on safe driving practices and the dangers of distracted driving in general.\nTable 3: Augmentation examples for the topic Do cell phones cause a distraction to people? with argument: If you mean do they cause a distraction when driving the answer is a resounding YES!!! That includes bluetooth or speaker. Nothing makes me more angry when people don't take driving as seriously as they should. It is a major killer and if you don't follow the rules and focus you are putting other people in danger.\nEffectiveness Score Generated similar quality argument to: Do you equate legality with morality? 1.0 Yes, I think that if the law says something is wrong, then it is wrong. People should follow the law and not do anything illegal."
},
{
"figure_ref": [],
"heading": "2.0",
"publication_ref": [],
"table_ref": [],
"text": "No, I do not think that legality necessarily equates to morality. Laws can often be outdated, biased, or simply not reflective of what is right and wrong. For example, slavery used to be legal in many countries, yet it was obviously immoral. So legality and morality do not always align."
},
{
"figure_ref": [],
"heading": "3.0",
"publication_ref": [],
"table_ref": [],
"text": "No, I do not equate legality with morality. Just because something is legal does not make it moral, and just because something is illegal does not make it immoral. Laws are made by people and can be changed, while morality is often determined by an individual's personal values and beliefs. Therefore, legality should not be used as a measure of morality. 4.0\nNo, I do not equate legality with morality. I believe that there are laws that should be followed in order to maintain the peace and order of a society, but there are also moral standards that must be adhered to for the same purpose. For example, it may be legal in some places to drive over the speed limit, but it is still morally wrong to do so. 5.0 I do not necessarily equate legality with morality. Legality is based on laws, which can be changed to reflect the morality of a society. Morality, in contrast, is based on principles and values that don't necessarily have to be enforced by law. For example, while it may be legal to drive over the speed limit, it is not necessarily moral to do so. model pays careful attention to the topic in context."
},
{
"figure_ref": [],
"heading": "Assumptions:",
"publication_ref": [],
"table_ref": [
"tab_2",
"tab_2"
],
"text": "The LLM lists several important assumptions in the proposed argument with respect to the topic. The first is the base assumption of the author's perspective, which is that cell phones cause distraction during driving. The LLM extracts the sentence \" people don't take driving as seriously as they should\" and labels it as an assumption because this is a faulty generalization to apply a general rule to all people, which is, in this case, people not taking driving seriously.\nSimilar quality argument: The generated similar quality argument tries to replicate the structural pattern of the given argument. The similar quality instance contains open-ended, long-winded sentences such as \"Cell phones are an amazing invention, but they can also be a major distraction\" which reduce the score of the argument, similar to the original argument. We also see that the LLM understands the ranking progression in a few shot setting and similar to the original argument, the similar quality argument also focuses solely on driving based cell phone distractions.\nCounter argument: We notice that the LLM recognizes that the original argument only discusses distracted driving and so it only produces a counter argument of the stance that distracted driving is not the only cause for distraction. The response discusses the safe use of cell phones such as hands free, etc and advocates educating drivers on the effects of cell phone based distracted driving.\nA.3 Impact of effectiveness score on GPT 3.5 outputs for similar quality arguments Table 4 discusses the impact of effectiveness score on the generated similar quality argument. As can be seen in Table 4, the generated argument with an effectiveness score of 1.0 oversimplifies the relationship between legality and morality and treats related laws as fixed. Comparatively, the argument with an effectiveness score of 2.0 provides an example of slavery which enhances the effectiveness of the argument. Despite the addition of this example, the second argument lacks elaboration on why the example is immoral and fails to provide relevant evidence. The argument generated, given an effectiveness score of 3.0, recognizes that the law is not the sole arbiter of morality and that laws are subject to change. It does not only highlight the potential flaws in legal systems but also addresses the distinction between personal values and the law. However, this argument oversimplifies morality by implying that personal values and beliefs solely determine morality and lacks supporting evidence for the statement: just because something is legal does not make it moral, and just because something is illegal does not make it immoral.. The argument generated with an effectiveness score of 4.0 considers the coexistence of legal and moral standards. The addition of a specific example in this argument adds concreteness and strengthens its persua-siveness. However, the argument can be further strengthened by acknowledging a broader range of situations where legality and morality may diverge. Finally, the argument ranked with the highest effectiveness score emphasizes the independence of morality from legal enforcement, which makes it even more persuasive. The contrasting comparison adds clarity to the flow of the argument and hence makes it better than all previously generated arguments."
},
{
"figure_ref": [
"fig_2"
],
"heading": "A.4 Distribution analysis of augmentation lengths across splits",
"publication_ref": [],
"table_ref": [],
"text": "In this subsection, we conduct a distribution analysis on the augmentation input sizes to justify the use of the dual BERT architecture. Based on the findings presented in The validation split has a minimum tokenized length of 144 tokens, a maximum length of 703 tokens, and an average length of 318.55 tokens. The testing split, on the other hand, has a minimum length of 136 tokens, a maximum length of 613 tokens, and an average length of 324.60 tokens. The percentage of data points in the validation and testing splits as seen in Figure 3 that exceed BERT's Hence, we can conclude that the second BERT encoder tasked with embedding the augmentations is able to capture all the information in the augmentations without truncating the augmentations."
},
{
"figure_ref": [],
"heading": "A.5 Impact of feedback",
"publication_ref": [],
"table_ref": [
"tab_6"
],
"text": "Table 6 shows the impact of providing feedback on the quality scores given by SPARK for three arguments.\nA.6 Questions for the human study Our human study poses the following three targeted questions to the participants:\n1. How valid is the information provided by the augmentation with respect to the background of the argument?\n2. How informative is the augmentation for the task of argument quality analysis?\n3. How relevant is the augmentation to help with the task of assessing the quality of the argument?"
},
{
"figure_ref": [],
"heading": "A Appendix",
"publication_ref": [],
"table_ref": [],
"text": "A.1 GPT 3.5 prompt templates"
},
{
"figure_ref": [],
"heading": "Feedback",
"publication_ref": [],
"table_ref": [],
"text": "The feedback on writing is sampled by considering both the topic and the argument related to the topic. To ensure brevity, we output the feedback in bullet point format. We follow the format below for sampling the feedback from the LLM:\nGive concise writing feedback for the following argument in context with the topic, preferably in bullet points: Topic: topic Argument: argument."
},
{
"figure_ref": [],
"heading": "Assumptions",
"publication_ref": [],
"table_ref": [],
"text": "Similar to feedback, assumptions are sampled in bullet point format to ensure brevity. Additionally, to constrain the hallucinations of the LLM, we restrict it to output \"No assumptions\" for the cases where it does not find assumptions or biases. We use the below prompt to sample this assumptions list:\nSummarize the assumptions, if any, in the following argument in a bullet format otherwise return \"No assumptions\" Topic: topic Argument: argument."
},
{
"figure_ref": [],
"heading": "Similar quality argument",
"publication_ref": [],
"table_ref": [],
"text": "To sample a similar quality argument, we use the following template:\nCogency Score: cogency score Effectiveness Score: effectiveness score Reasonableness Score: reasonableness score Topic: topic\nWe use ten samples in the few shot setting with two each from every integer ranking from 1-5 on the ranking scale for each metric. Finally, we prompt the LLM to generate the argument with respect to the cogency, effectiveness and robustness scores."
},
{
"figure_ref": [],
"heading": "Counter-argument",
"publication_ref": [],
"table_ref": [],
"text": "The counter-argument is generated using the given argument and topic, and the following template:\nGive a counter-argument for the following argument with respect to the Topic: topic Argument: argument"
},
{
"figure_ref": [],
"heading": "A.2 Augmentation examples",
"publication_ref": [],
"table_ref": [],
"text": ""
}
] | Automatic assessment of the quality of arguments has been recognized as a challenging task with significant implications for misinformation and targeted speech. While real-world arguments are tightly anchored in context, existing computational methods analyze their quality in isolation, which affects their accuracy and generalizability. We propose SPARK : a novel method for scoring argument quality based on contextualization via relevant knowledge. We devise four augmentations that leverage large language models to provide feedback, infer hidden assumptions, supply a similar-quality argument, or give a counter-argument. SPARK uses a dual-encoder Transformer architecture to enable the original argument and its augmentation to be considered jointly. Our experiments in both in-domain and zero-shot setups show that SPARK consistently outperforms existing techniques across multiple metrics. | Contextualizing Argument Quality Assessment with Relevant Knowledge | [
{
"figure_caption": "Figure 1 :1Figure 1: Overview of SPARK.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: Dual BERT encoder architecture.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: Distributions for augmentation lengths for the training, validation and testing splits respectively.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Performance of Dual-BERT model with augmentations applied compared to the baseline models. The performance of the model achieving the best scores per metric is boldfaced and the second best score is underlined.",
"figure_data": "GAQ Corpus (in-domain)IBM-30K (ZS)",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Validity, informativeness, and relevance scores of the augmentations for argument scoring.",
"figure_data": "",
"figure_id": "tab_1",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Similar quality examples.",
"figure_data": "",
"figure_id": "tab_2",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Distribution properties across splits for the concatenated augmentations.",
"figure_data": "",
"figure_id": "tab_4",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": ", the training split ex-",
"figure_id": "tab_5",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "Feedback examples for the out-of-domain dataset.",
"figure_data": "TopicArgumentScoreScoreGroundFeedback augmentationbe-aftertruthforeaug-scoreaug-men-men-tationtationThe use of publicA centralized system of0.6280.6750.828 -Clear and concise argument presenteddefenders shouldcriminal defense would-Supports the idea of a centralized system ofbe mandatorymean that all people wouldcriminal defensehave access to the same-Highlights the importance of equal accessstandard of legal counsel,to legal counselmeaning that wealth and-Addresses the issue of wealth and powerpower can't be used to avoidinfluencing justicejustice.-Could benefit from further elaboration orevidence to strengthen the argumentWe should ban al-algorithmic trading has been0.5850.6310.948 -Clear and concise argument presentedgorithmic tradingresponsible for several mini-Provides specific examples to support argu-market collapses, since com-mentputer systems lack the hu--Could benefit from further elaboration onman sensitivity to look out-the potential consequences of mini marketside the stream of meaning-collapsesless numbers to a wider con--Could also benefit from addressing potentialtext.counterarguments or alternative solutions tothe issueWe should adoptlibertarianism is a justifica-0.5400.5190.666The argument is not specific enough aboutlibertarianismtion for greed and exploita-what adopting libertarianism entailstion-It assumes that libertarianism automaticallyleads to working together for a greater goodand favoring the less well off, which is notnecessarily true-The argument could benefit from provid-ing concrete examples of how libertarianismwould benefit society as a whole-It is unclear how giving freedom of choicewould lead to greater societal benefits, andthis point could be expanded upon.token limit are only 1.1% (13 examples) and 1.31%(15 examples), respectively.",
"figure_id": "tab_6",
"figure_label": "6",
"figure_type": "table"
}
] | Darshan Deshpande; Zhivar Sourati; Filip Ilievski; Fred Morstatter | [
{
"authors": "Al Khalid; Michael Khatib; Shahbaz Völske; Nikolay Syed; Benno Kolyada; Stein",
"journal": "Association for Computational Linguistics",
"ref_id": "b0",
"title": "Exploiting personal characteristics of debaters for predicting persuasiveness",
"year": "2020"
},
{
"authors": " Alhindi",
"journal": "",
"ref_id": "b1",
"title": "Computational models of argument structure and argument quality for understanding misinformation",
"year": "2023"
},
{
"authors": "Yonatan Bilu; Noam Slonim",
"journal": "Association for Computational Linguistics",
"ref_id": "b2",
"title": "Claim synthesis via predicate recycling",
"year": "2016"
},
{
"authors": "Tuhin Chakrabarty; Christopher Hidey; Smaranda Muresan; Kathy Mckeown; Alyssa Hwang",
"journal": "",
"ref_id": "b3",
"title": "AMPERSAND: Argument mining for PERSuAsive oNline discussions",
"year": "2019"
},
{
"authors": "Chung Hyung Won; Le Hou; Shayne Longpre; Barret Zoph; Yi Tay; William Fedus; Eric Li; Xuezhi Wang; Mostafa Dehghani; Siddhartha Brahma",
"journal": "",
"ref_id": "b4",
"title": "Scaling instruction-finetuned language models",
"year": "2022"
},
{
"authors": "Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova",
"journal": "Association for Computational Linguistics",
"ref_id": "b5",
"title": "BERT: Pre-training of deep bidirectional transformers for language understanding",
"year": "2019"
},
{
"authors": "Esin Durmus; Claire Cardie",
"journal": "Association for Computational Linguistics",
"ref_id": "b6",
"title": "Exploring the role of prior beliefs for argument persuasion",
"year": "2018"
},
{
"authors": "Roxanne El Baff; Henning Wachsmuth; Khalid ; Al Khatib; Benno Stein",
"journal": "Association for Computational Linguistics",
"ref_id": "b7",
"title": "Analyzing the Persuasive Effect of Style in News Editorial Argumentation",
"year": "2020"
},
{
"authors": "Michael Fromm; Max Berrendorf; Johanna Reiml; Isabelle Mayerhofer; Siddharth Bhargava; Evgeniy Faerman; Thomas Seidl",
"journal": "",
"ref_id": "b8",
"title": "Towards a holistic view on argument quality prediction",
"year": "2022"
},
{
"authors": "Debanjan Ghosh; Aquila Khanam; Yubo Han; Smaranda Muresan",
"journal": "Association for Computational Linguistics",
"ref_id": "b9",
"title": "Coarse-grained argumentation features for scoring persuasive essays",
"year": "2016"
},
{
"authors": "Daniel Gillick; Alessandro Presta; Gaurav Singh Tomar",
"journal": "",
"ref_id": "b10",
"title": "End-to-end retrieval in continuous space",
"year": "2018"
},
{
"authors": "Shai Gretz; Roni Friedman; Edo Cohen-Karlik; Assaf Toledo; Dan Lahav; Ranit Aharonov; Noam Slonim",
"journal": "",
"ref_id": "b11",
"title": "A large-scale dataset for argument quality ranking: Construction and analysis",
"year": "2020"
},
{
"authors": "Timon Gurcke; Milad Alshomary; Henning Wachsmuth",
"journal": "",
"ref_id": "b12",
"title": "Assessing the sufficiency of arguments through conclusion generation",
"year": "2021"
},
{
"authors": "Ivan Habernal; Iryna Gurevych",
"journal": "Association for Computational Linguistics",
"ref_id": "b13",
"title": "What makes a convincing argument? empirical analysis and detecting attributes of convincingness in web argumentation",
"year": "2016"
},
{
"authors": "Gordon Harvey",
"journal": "",
"ref_id": "b14",
"title": "A brief guide to the elements of the academic essay",
"year": "2009"
},
{
"authors": "Md Kamrul Hasan; James Spann; Masum Hasan; Md Saiful Islam; Kurtis Haut; Rada Mihalcea; Ehsan Hoque",
"journal": "Association for Computational Linguistics",
"ref_id": "b15",
"title": "Hitting your MARQ: Multimodal ARgument quality assessment in long debate video",
"year": "2021"
},
{
"authors": "Geoffrey Hinton; Oriol Vinyals; Jeff Dean",
"journal": "",
"ref_id": "b16",
"title": "Distilling the knowledge in a neural network",
"year": "2015"
},
{
"authors": "Ioana Hulpus; Jonathan Kobbe; Christian Meilicke; Heiner Stuckenschmidt; Maria Becker; Juri Opitz; Vivi Nastase; Anette Frank",
"journal": "",
"ref_id": "b17",
"title": "Towards explaining natural language arguments with background knowledge",
"year": "2019"
},
{
"authors": "Vladimir Karpukhin; Barlas Oguz; Sewon Min; Patrick Lewis; Ledell Wu; Sergey Edunov; Danqi Chen; Wen-Tau Yih",
"journal": "Association for Computational Linguistics",
"ref_id": "b18",
"title": "Dense passage retrieval for opendomain question answering",
"year": "2020"
},
{
"authors": "Anne Lauscher; Lily Ng; Courtney Napoles; Joel Tetreault",
"journal": "International Committee on Computational Linguistics",
"ref_id": "b19",
"title": "Rhetoric, logic, and dialectic: Advancing theory-based argument quality assessment in natural language processing",
"year": "2020"
},
{
"authors": "Anne Lauscher; Henning Wachsmuth; Iryna Gurevych; Goran Glavaš",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b20",
"title": "Scientia Potentia Est-On the Role of Knowledge in Computational Argumentation",
"year": "2022"
},
{
"authors": "Sabrina Ludwig; Christian Mayer; Christopher Hansen; Kerstin Eilers; Steffen Brandt",
"journal": "Psych",
"ref_id": "b21",
"title": "Automated essay scoring using transformer models",
"year": "2021"
},
{
"authors": "Santiago Marro; Elena Cabrio; Serena Villata",
"journal": "Association for Computational Linguistics",
"ref_id": "b22",
"title": "Graph embeddings for argumentation quality assessment",
"year": "2022"
},
{
"authors": "Yeti Mulyati; Daris Hadianto",
"journal": "International Journal of Instruction",
"ref_id": "b23",
"title": "Enhancing argumentative writing via online peer feedbackbased essay: A quasi-experiment study",
"year": "2023"
},
{
"authors": "Raymond S Nickerson",
"journal": "Cambridge University Press",
"ref_id": "b24",
"title": "Biases, Misconceptions, and the Like",
"year": "2020"
},
{
"authors": " Openai",
"journal": "",
"ref_id": "b25",
"title": "Chatgpt",
"year": "2022-04-30"
},
{
"authors": "Ruty Rinott; Lena Dankin; Carlos Alzate Perez; Mitesh M Khapra; Ehud Aharoni; Noam Slonim",
"journal": "Association for Computational Linguistics",
"ref_id": "b26",
"title": "Show me your evidence -an automatic method for context dependent evidence detection",
"year": "2015"
},
{
"authors": "Gabriella Skitalinskaya; Jonas Klaff; Henning Wachsmuth",
"journal": "",
"ref_id": "b27",
"title": "Learning from revisions: Quality assessment of claims in argumentation at scale",
"year": "2021"
},
{
"authors": "Christian Stab; Iryna Gurevych",
"journal": "Dublin City University and Association for Computational Linguistics",
"ref_id": "b28",
"title": "a. Annotating argument components and relations in persuasive essays",
"year": "2014"
},
{
"authors": "Christian Stab; Iryna Gurevych",
"journal": "Association for Computational Linguistics",
"ref_id": "b29",
"title": "Identifying argumentative discourse structures in persuasive essays",
"year": "2014"
},
{
"authors": "Assaf Toledo; Shai Gretz; Edo Cohen-Karlik; Roni Friedman; Elad Venezian; Dan Lahav; Michal Jacovi; Ranit Aharonov; Noam Slonim",
"journal": "Association for Computational Linguistics",
"ref_id": "b30",
"title": "Automatic argument quality assessment -new datasets and methods",
"year": "2019"
},
{
"authors": "Hugo Touvron; Louis Martin; Kevin Stone; Peter Albert; Amjad Almahairi; Yasmine Babaei; Nikolay Bashlykov; Soumya Batra; Prajjwal Bhargava; Shruti Bhosale; Dan Bikel; Lukas Blecher; Cristian Canton Ferrer; Moya Chen; Guillem Cucurull; David Esiobu; Jude Fernandes; Jeremy Fu; Wenyin Fu; Brian Fuller; Cynthia Gao; Vedanuj Goswami; Naman Goyal; Anthony Hartshorn; Saghar Hosseini; Rui Hou; Hakan Inan; Marcin Kardas; Viktor Kerkez; Madian Khabsa; Isabel Kloumann; Artem Korenev; Punit Singh Koura; Marie-Anne Lachaux; Thibaut Lavril; Jenya Lee; Diana Liskovich; Yinghai Lu; Yuning Mao; Xavier Martinet; Todor Mihaylov; Pushkar Mishra; Igor Molybog; Yixin Nie; Andrew Poulton; Jeremy Reizenstein; Rashi Rungta; Kalyan Saladi; Alan Schelten; Ruan Silva; Eric Michael Smith; Ranjan Subramanian; Ellen Xiaoqing; Binh Tan; Ross Tang; Adina Taylor; Jian Williams; Puxin Xiang Kuan; Zheng Xu; Iliyan Yan; Yuchen Zarov; Angela Zhang; Melanie Fan; Sharan Kambadur; Aurelien Narang; Robert Rodriguez; Sergey Stojnic; Thomas Edunov; Scialom",
"journal": "",
"ref_id": "b31",
"title": "Llama 2: Open foundation and finetuned chat models",
"year": "2023"
},
{
"authors": "Dietrich Trautmann; Michael Fromm; Thomas Volker Tresp; Hinrich Seidl; Schütze",
"journal": "Datenbank-Spektrum",
"ref_id": "b32",
"title": "Relational and fine-grained argument mining",
"year": "2020"
},
{
"authors": "Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Ł Ukasz Kaiser; Illia Polosukhin",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b33",
"title": "Attention is all you need",
"year": "2017"
},
{
"authors": "Curran Associates; Inc ",
"journal": "",
"ref_id": "b34",
"title": "",
"year": ""
},
{
"authors": "Henning Wachsmuth; Nona Naderi; Ivan Habernal; Yufang Hou; Graeme Hirst; Iryna Gurevych; Benno Stein",
"journal": "Association for Computational Linguistics",
"ref_id": "b35",
"title": "Argumentation quality assessment: Theory vs. practice",
"year": "2017"
},
{
"authors": "Henning Wachsmuth; Nona Naderi; Yufang Hou; Yonatan Bilu; Tim Alberdingk Vinodkumar Prabhakaran; Graeme Thijm; Benno Hirst; Stein",
"journal": "Association for Computational Linguistics",
"ref_id": "b36",
"title": "Computational argumentation quality assessment in natural language",
"year": "2017"
},
{
"authors": "Zhilin Yang; Zihang Dai; Yiming Yang; Jaime Carbonell; Russ R Salakhutdinov; Quoc V Le",
"journal": "",
"ref_id": "b37",
"title": "Xlnet: Generalized autoregressive pretraining for language understanding",
"year": "2019"
}
] | [] |
10.18653/v1/2021.emnlp-main.532 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b18",
"b15",
"b38",
"b47",
"b52"
],
"table_ref": [],
"text": "When recurrent neural networks such as LSTM (Hochreiter and Schmidhuber, 1997) are the mainstream language model (LM) architecture, pointer networks, or so-called copy mechanisms (Gu et al., 2016), have been shown to improve the state-of-the-art LMs for next word prediction (Merity et al., 2017) and summarizations (See et al., 2017) by a large margin. However, after transformer (Vaswani et al., 2017) becomes the dominating LM architectures, the pointer networks are rarely used in the state-of-the-art pretrained LMs.One major reason is that the attention mechanism in every transformer layer can learn to copy the words from the context, so it "
},
{
"figure_ref": [
"fig_7",
"fig_7",
"fig_0"
],
"heading": "Input context Ct",
"publication_ref": [
"b5",
"b44",
"b45",
"b60",
"b5",
"b5",
"b15",
"b60",
"b5"
],
"table_ref": [],
"text": "Figure 1: Illustration of the softmax bottleneck and pointer network using an example from Chang and McCallum (2022). GPT-2 cannot output both king or woman as the possible next word due to the parallelogram structure in the output word embedding space, while the pointer network could solve this by directly copying words from the context. The standard softmax estimate the probabilities of outputting king and woman by the dot products between the hidden state h ct,V and their global word embeddings. By contrast, The pointer networks compute the dot products between the projected current hidden state h ct,S and projected hidden states h e,. for king and woman to estimate their probabilities.\nseems to be redundant to add a copying mechanism on top of the transformer.\nIn this paper, we demonstrate that the architectures like pointer networks can still substantially improve the state-of-the-art transformer LM architectures such as GPT-2 (Radford et al., 2019) and T5 (Raffel et al., 2020) mainly due to breaking the bottleneck of their final softmax layer (Yang et al., 2018;Chang and McCallum, 2022).\nIn Figure 1, we illustrate a simple example from Chang and McCallum (2022) word, most LMs would try to output a hidden state h ct,V that is close to all the next word possibilities. For example, when the next word should be either king or woman with similar probabilities, the ideal hidden state is supposed to be the average of the global output word embeddings of king and woman. However, there might be other interfering words (queen and man in this case) between the ideal next word candidates, which force the LM to output the wrong distribution.\nTo solve this problem, we can let the LMs predict the probability of copying the words in the context separately by paying attention to the previous hidden states (Gu et al., 2016) and we call this kind of architecture pointer networks in this paper. That is, we can compute the dot products with the hidden states of king h e,k and the hidden states of woman h e,w rather than with their global output word embeddings in order to estimate the probabilities of copying these two words in the context. Our experiments show that the pointer networks consistently improve the performance of GPT-2 in next word prediction and the quality of summarization from T5 and BART.\nContrary to the mainstream explanation in previous pointer network literature, we discover that most of the improvements in our experiments do not come from the attention mechanism. To study these improvements, we propose a very simple pointer network variant that does not use any previous hidden states and we show that the proposed method can achieve similar improvements.\nAs shown in Figure 2, we simply project the last hidden state into two embeddings. One embedding h ct,S is to compute the dot product with the context words, and h ct,V is for the dot product of the other words. Then, the GPT-2 can output the hidden state for context words h ct,S as the average embedding of the king and woman without interfered by the words of man and queen that are handled by h ct,V . We call this method context partition. In addition to words in the context, we can also use another embedding for the top-k likely next words. This can be viewed as a very simple and efficient alternative to a reranker, so we call it reranker partition.\nIn our experiments, we show that the context partition performs similarly to pointer networks while combining a pointer network, context partition, and reranker partition would significantly outperform each individual method. Compared to the state-ofthe-art solutions for alleviating the softmax bottleneck such as mixture of softmax (Yang et al., 2018;Chang and McCallum, 2022), our proposed method is more efficient while achieving lower perplexity on GPT-2. Furthermore, we find that adding a very expensive word-by-word reranker only improves our method slightly, which suggested the difficulty of further improving the final softmax layer over the proposed alternatives.\nIn the text completion task using GPT-2, we find that the proposed softmax alternatives reduce hallucination by copying more proper nouns from the context even though we did not provide any partof-speech information during training. In summarization, our methods and pointer networks output a more specific summary, increase the factuality, and consistently improve 9 metrics, especially in the smaller language models. Finally, we show that the softmax bottleneck problem is not completely solved in GPT-3.5 in the limitation section."
},
{
"figure_ref": [],
"heading": "Main Contributions",
"publication_ref": [],
"table_ref": [],
"text": "• We propose a series of efficient softmax alternatives that unify the ideas of pointer network, reranker, multiple embeddings, and vocabulary partitioning. 1• We evaluate the proposed softmax alternatives in text completion tasks and summarization tasks using various metrics to identify where our methods improve the most.\n• Our experiments indicate pointer networks and our proposed alternatives can still improve the modern transformer-based LMs. By breaking the softmax bottleneck, our methods learn to sometimes copy the context words to reduce generation hallucination and sometimes exclude the context words to reduce the repetition. Besides, we find that the softmax bottleneck problem won't be completely solved by the huge size of GPT-3.5."
},
{
"figure_ref": [],
"heading": "Background",
"publication_ref": [],
"table_ref": [],
"text": "Before introducing our method, we would first briefly review the problem we are solving and its state-of-the-art solutions."
},
{
"figure_ref": [],
"heading": "Softmax Bottleneck Problem",
"publication_ref": [
"b5"
],
"table_ref": [],
"text": "Most LMs use a softmax layer to compute the final probability of predicting the word x:\nP M (x|c t ) = exp(Logit(x, c t ))\nx exp(Logit(x , c t ))\n,\nwhere c t is the context words. Typically, the logit Logit(x, c t ) = (h M ct ) T w x , h M ct is the M th-layer hidden state given the input context c t and w x is the output word embeddings for x.\nOne problem is that the output word embeddings w x are global and independent to the context. After pretraining, the similar words would have similar output word embeddings. However, the similarity structure in the word embedding space might prevent LMs from outputting the desired distribution. The parallelogram structure among the embeddings of king, queen, woman, and man is a simple example. Chang and McCallum (2022) generalize this observation and show that some words in a small subspace would create some multi-mode distributions that a LM cannot output using a single hidden state h ct in the softmax layer."
},
{
"figure_ref": [],
"heading": "Mixture of Softmax Method",
"publication_ref": [
"b60"
],
"table_ref": [],
"text": "To overcome the bottleneck, one natural solution is to have multiple hidden states and each hidden state corresponds to a group of possible words (Yang et al., 2018). For example, we can have one hidden state for king and another hidden state for woman.\nOne major concern of this mixture of softmax (MoS) approach is the computational overhead. MoS needs to compute the final softmax multiple times and merge their resulting distributions. That is, we need to compute the dot products between every hidden state and all the words in the vocabulary, which is expensive especially when the vocabulary size is large. "
},
{
"figure_ref": [],
"heading": "Multiple Input State Enhancement",
"publication_ref": [
"b5",
"b5"
],
"table_ref": [],
"text": "In MoS, the multiple hidden states come from the linear projections of the last hidden state. Chang and McCallum (2022) point out that the total degree of freedom among the multiple hidden states is limited by the dimensionality of the hidden state.\nTo allow LMs to move multiple hidden states more freely, Chang and McCallum (2022) propose to concatenate the projection of a block of hidden state with the last hidden state h M ct so as to increase its dimensionality:\nq ct = h M ct ⊕ GELU L h (⊕ i,m h M -m c t-i ) ,(2)\nwhere GELU is the non-linear transformation used in GPT-2 and L h is a linear transformation that allows us to consider more hidden states without significantly increasing the model size.\n⊕ i,m h M -m c t-i\nis the concatenation of a block of hidden states. We set the block size to be 3x3 in our GPT-2 experiments and 1x3 in our summarization experiments (i.e., considering the last 3 hidden states in the last layer as shown in Figure 3)."
},
{
"figure_ref": [],
"heading": "Methods",
"publication_ref": [],
"table_ref": [],
"text": "To break the softmax bottleneck more efficiently compared to MoS, our overall strategy is simple.\nIf we can identify a small partition of words that are very likely to become the next word, we can just compute the dot products between a hidden state and the embeddings of these likely words instead of all the words as in MoS. For example, if we can identify king and woman are much more likely to appear than queen and man, we can only compute the dot product between a hidden state and the embeddings of king and woman without being interfered by other words. Specifically, when we compute the next word probability in Equation 1, the logit of the word x given the context c \nt Logit(x, c t ) = f T ct,S e x if x ∈ S f T ct,V w x O/W ,(3)\nL f LD L f C\nFigure 3: Architectures of our method for T5/BART that computes Logit CEP R in Equation 6. In GPT-2, we use same architecture except that we take the 3x3 input hidden state block rather than the 1x3 block and there are no encoder-related components, which are marked by dotted lines.\nwhere f ct,S = L f S (q ct ) and f ct,V = L f V (q ct ) are the linear projections of the hidden state concatenation q ct in Equation 2. As shown in Table 1, different softmax alternatives have different ways of constructing this set S and use different word embeddings e x .\nTo simplify our explanation, we will focus on the decoder-only LM (i.e., GPT-2) first and extend our method to encoder-decoder LM (i.e., T5 and BART)."
},
{
"figure_ref": [],
"heading": "GPT-2",
"publication_ref": [],
"table_ref": [],
"text": "We will explain each softmax alternative individually and their connections to previous work such as pointer networks or rerankers."
},
{
"figure_ref": [],
"heading": "Pointer Network (P) as Local Word Embedding",
"publication_ref": [
"b38",
"b38"
],
"table_ref": [],
"text": "Similar to Pointer Sentinel (PS) (Merity et al., 2017), we treat the words in the context differently (S = {x|x ∈ c t }) and let their word embeddings e x come from the previous hidden states:\ne x = f x,ct,LD = t i=1 1 c i t =x L f LD (q c i t ) t i=1 1 c i t =x ,(4)\nwhere c i t is the ith input words in the context c t , L f LD is a linear layer, and\n1 c i t =x = 1 if c i t = x.\nAs a result, we can use the GPT-2 model to not only predict the hidden state f ct,S = f ct,P D = L f P D (q ct ) and f ct,V but also predict the word embedding of context words e x . Unlike the global word embedding w x , the local word embedding e x is context-dependent, so the LM can break the softmax bottleneck by adjusting the similarity of words based on the context. For example, GPT-2 could increase the similarity between e king and e woman to output the high probability for both words easily.\nWe call this version of pointer network local decoder (LD) embedding, which has some minor differences compared to PS (Merity et al., 2017) and other variants. For example, we merge their logits while PS merges their probabilities. PS does not do normalization when computing e x . In our experiments, we would show that these pointer network variants all have very similar improvements in modern LMs."
},
{
"figure_ref": [],
"heading": "Context Partition (C)",
"publication_ref": [],
"table_ref": [],
"text": "To understand the source of the improvements from pointer networks, we simplify their architectures by setting the word embedding e x = w x and the partition S is still the set of context words. Although much simpler, the LM with this context partition method can still break the softmax bottleneck by properly coordinating the hidden state specifically for the context words f ct,S = f ct,C = L f C (q ct ) and the hidden state for other words f ct,V . Compared to the pointer network, one advantage of context partition is that the LM can still leverage the learned global word similarity when estimating the probabilities of context words."
},
{
"figure_ref": [],
"heading": "Reranker Partition (R)",
"publication_ref": [],
"table_ref": [],
"text": "In some cases, the possible next words might not be mentioned in the context. For example, in the context My favorite actor is Ryan [MASK], the next word could be Reynolds, Gosling, or the last names of other Ryan. Hence, using only the context partition does not completely solve the multimodal distribution problem.\nInspired by the idea of the reranker, we set S to be the top k words with the highest logits f T ct,V w x . In practice, finding an ideal k could be difficult. When k is small, the reranker partition might not include the very likely next word. When k is large, the reranker partition might not be able to separate the output candidates and the interfering words. To alleviate the problem, we can have multiple reranker partitions and use different hidden state embeddings (e.g., f ct,R1 and f ct,R2 ) for different partitions."
},
{
"figure_ref": [],
"heading": "Hybrid Approach (CPR)",
"publication_ref": [],
"table_ref": [],
"text": "Local embeddings in the pointer networks and global embeddings in the context partition are complementary. Using local embeddings is representational powerful while using global embedding can leverage the global similarity of words. Hence, we can combine the two methods by summing their dot products.\nFor the methods that use different S, we can simply determine an order of computing the dot products and let the later dot products overwrite the existing values. In our experiments, we always use the order illustrated in Figure 3. That is, we compute the logits (Logit CP R (x, c t )) by\n f T ct,C w x + f T ct,P D f x,ct,LD if x ∈ c t f T ct,R1 w x if x ∈ W (k 1 ) -c t f T ct,R2 w x if x ∈ W (k 2 ) -W (k 1 ) -c t f T ct,V w x O/W ,(5)\nwhere W (k 2 ) is the top k 2 words with the highest f T ct,V w x and W (k 1 ) is the top k 1 words with the highest max(f T ct,V w x , f T ct,R2 w x )."
},
{
"figure_ref": [],
"heading": "T5 and BART",
"publication_ref": [],
"table_ref": [],
"text": "In the encoder-decoder architectures, our local decoder embedding, context partition, and reranker partitions are still applicable. Besides, we can leverage the words in the encoder input to further improve the performance."
},
{
"figure_ref": [],
"heading": "Encoder Partition (E) and Local Encoder Embedding (P)",
"publication_ref": [],
"table_ref": [],
"text": "Similar to the context partition, the encoder partition handles the words in the encoder input I differently by setting S = {x|x ∈ I} and using the global word embedding e x = w x .\nAs in Equation 4, we can also let the hidden states in the last layer pass through another linear layer L f LE () to predict the embeddings of the words in the encoder input. The method is called local encoder (LE) embedding."
},
{
"figure_ref": [],
"heading": "Hybrid Approach (CEPR)",
"publication_ref": [],
"table_ref": [],
"text": "Similar to GPT-2, we combine local encoder embedding and encoder partition for computing the probabilities of the words that are in the encoder context but not in the decoder context. As shown in Figure 3, we compute Logit CEP R (x, c t ) by\n f T ct,C w x + f T ct,P D f x,ct,LD if x ∈ c t f T ct,E w x + f T ct,P E f x,I,LE if x ∈ I -c t f T ct,R1 w x if x ∈ W (k 1 ) -c t -I f T ct,V w x O/W ,(6)\nwhich is the same as Equation 5 except that we add the encoder partition and local encoder embedding, and we remove the second reranker partition."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [
"b47"
],
"table_ref": [],
"text": "The pointer network was a popular technique in language modeling (Merity et al., 2017) and summarization (See et al., 2017). Thus, we also focus on these two fundamental applications."
},
{
"figure_ref": [],
"heading": "GPT-2",
"publication_ref": [
"b5",
"b44"
],
"table_ref": [],
"text": "We follow the setup in Chang and McCallum (2022) to continue training GPT-2 on Wikipedia 2021 and OpenWebText (Radford et al., 2019)."
},
{
"figure_ref": [],
"heading": "Perplexity Comparison",
"publication_ref": [],
"table_ref": [
"tab_3",
"tab_5",
"tab_6",
"tab_13"
],
"text": "In Table 2, we first compare their predictions on the next word distribution using the testing data perplexity, which is a standard metric in the LM architecture studies. In the To know how well our method breaks the softmax bottleneck, we implement a word-by-word reranker model on GPT-2, which appends the most likely 100 words to the context when predicting each next word (see Appendix C.3 for more details). In Table 3, we show that our efficient softmax alternative Softmax + CPR:20,100 + Mi achieves significantly lower perplexity. Furthermore, the word-by-word reranker is at least 10 times slower during training. Combining word-by-word reranker with our method only improves the perplexity very 2 Notice that the pointer networks from the previous work were originally designed for RNN. To add them on top of the transformer based LMs and make it more comparable to our methods, we simplify their architectures a little. Please see Appendix C.2 for more details. Table 4: ROUGE-1 F1 (%) of different methods on GPT-2. We compare the scores between the generated text and the reference (i.e., continuation), and between the generation and context. More methods and metrics are reported in Table 8.\nslightly, which suggests the challenges of further improving LM by breaking softmax bottleneck."
},
{
"figure_ref": [],
"heading": "Generated Text Comparison",
"publication_ref": [],
"table_ref": [
"tab_13"
],
"text": "Next, we would like to understand how the distribution improvement affects the text generation. We sample some contexts in the test set of Wikipedia 2021 and compare the generated text quality of the different models given the contexts. The quality is measured by the ROUGE-1 F1 scores between the generated text and the actual continuation. To know how much the different models copy from the context, we also report the ROUGE-1 scores between the generation and the contexts. 8, we compare methods using more metrics to further support the conclusion."
},
{
"figure_ref": [],
"heading": "The results in",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Qualitative Analysis",
"publication_ref": [],
"table_ref": [
"tab_8"
],
"text": "In Table 5, we visualize some distributions to explain our improvements. The softmax layer of GPT-2 is unable to properly learn to copy or exclude the word from the input context. For example, Softmax + Mi and MoS + Mi might output \"There are plates, keys, scissors, toys, and balloons in front of me, and I pick up the phone\", which causes a hallucination problem, while Softmax + CPR:20,100 + Mi and Pointer Sentinel (PS) + Mi can output the mentioned options with similar probability by copying the words in the context. In addition, GPT-2, MoS, and PS + Mi are very likely to output \"I like tennis, baseball, golf, basketball, and tennis\". This repetition problem happens because the next word should be some words similar to the listed sports names except for the sports that have been mentioned and the softmax layer has difficulties in outputting a donut-shape next word distribution in embedding space. In contrast, Softmax + CPR:20,100 + Mi can learn to exclude the listed sports by putting very negative logits on the context words, which yield the desired donut-shape distribution."
},
{
"figure_ref": [],
"heading": "T5 and BART in Summarization",
"publication_ref": [
"b34",
"b26",
"b43",
"b24",
"b4"
],
"table_ref": [
"tab_15",
"tab_16",
"tab_9"
],
"text": "We In the main paper, we evaluate the quality of summaries using four metrics. ROUGE-1 F1 (Lin, 2004) measures the unigram overlapping between the generated summary and the ground truth summary; CIDEr (Vedantam et al., 2015) adds a tfidf weighting on the n-gram overlapping score to emphasize correct prediction of rare phrases; factCC (Kryscinski et al., 2020) evaluates the factuality of the summary; MAUVE (Pillutla et al., 2021) compares the word distribution of summary and ground truth in a quantized embedding space. To further support our conclusions, we also compare the quality measured by several other metrics and their model sizes in Table 9 andTable 10.\nThe results are reported in Table 6. Similar to the GPT-2 experiments, the results are generally better as we combine more partitions and local embedding approaches. This demonstrates that we can directly fine-tune the LMs with our softmax alternatives without expensive pretraining.\nUnlike the GPT-2 experiments, multiple input hidden state enhancement (Mi) is not very effective, so we mainly compare the methods without Mi (i.e., q ct = h M ct , unlike Equation 2). We hypothesize one possible reason is that we haven't pretrained the T5 and BART with our softmax alternatives.\nOur improvements are larger in smaller models. This is probably because in a smaller word embedding space, there are more likely to be interfering words between the desired next word possibilities. Compared to our methods, the pointer networks perform well in BART-base but usually perform worse in other LMs. We need further investigations in the future to explore the reasons.\nCompared to ROUGE-1 score, the improvement percentage of CIDEr is overall higher. One major problem of the summarization LMs is that the generated summary contains too many commonly used phrases (King et al., 2022) and our considerably higher CIDEr scores indicate the alleviation of the problem. Our improvement on the factCC is also significant (Cao and Wang, 2021). Finally, our MAUVE improvement percentage on Book-Sum Paragraph dataset could reach around 30% in T5-Small. We hypothesize this is because we often mention the global entity names in the news (e.g., Obama) while the meaning of names in stories (e.g., John) is often defined by the context."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b5",
"b5"
],
"table_ref": [],
"text": "Repetition and hallucination are two common problems in language generation tasks. One common solution for repetition is to avoid outputting the words in the context, which is often called unlike- Our analyses demonstrate that parts of the hallucination and repetition problem come from the softmax bottleneck. The findings provide an explanation for the effectiveness of prior studies such as the above reranker approaches and pointer networks (Li et (Chang and McCallum, 2022) to explain the improvement of a pointer network. Their empirical results also support our conclusion that softmax bottleneck is a major reason that causes the factuality problem of LMs.\nOur work is motivated and inspired by Chang and McCallum (2022). In their work, they also propose to use different hidden states for different vocabulary partitions, but their partitioning is global and needs to be combined with the mixture of softmax (MoS) approach, which adds a significant overhead compared to the standard softmax layer. Our dynamic partitioning methods not only perform better but greatly reduce the overhead by removing the reliance on MoS."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [
"b5"
],
"table_ref": [],
"text": "Since the transformer becomes the mainstream encoder and decoder for LMs, the output softmax layer seems to be the only reasonable option for computing the word probability distribution. Although being simple and efficient, the softmax layer is inherently limited while the existing solutions are relatively slow (Chang and McCallum, 2022). This work proposes a series of softmax alternatives that can improve the text generation models without increasing the computational costs significantly. Our experiments suggest that the main improvement of the pointer network on top of a transformer comes from breaking the softmax bottleneck. Our results also indicate that the alternatives could alleviate some problems of hallucination, repetition, and too generic generation. Furthermore, all of the proposed alternatives can be applied to the LMs that have already been pretrained using softmax without requiring retraining from scratch. For the practitioner, we recommend using all the partitioning methods together to get the best performance, or using only the simple context partition to keep the architecture simple while getting the majority of the gain."
},
{
"figure_ref": [],
"heading": "Acknowledgement",
"publication_ref": [],
"table_ref": [],
"text": "We thank Nadar Akoury and the anonymous reviewers for their constructive feedback. This work was supported in part by the Center for Data Science and the Center for Intelligent Information Retrieval, in part by the Chan Zuckerberg Initiative under the project Scientific Knowledge Base Construction, in part by the IBM Research AI through the AI Horizons Network, in part using high performance computing equipment obtained under a grant from the Collaborative R&D Fund managed by the Massachusetts Technology Collaborative, and in part by the National Science Foundation (NSF) grant numbers IIS-1922090 and IIS-1763618. Any opinions, findings, conclusions, or recommendations expressed in this material are those of the authors and do not necessarily reflect those of the sponsor."
},
{
"figure_ref": [
"fig_2",
"fig_2",
"fig_2",
"fig_2"
],
"heading": "Limitations",
"publication_ref": [
"b33",
"b23"
],
"table_ref": [
"tab_8",
"tab_8"
],
"text": "In our experiments, we find that the improvement of our methods tend to be larger in relatively smaller language models. Due to our limited access of computational resources, we are not able to try our methods on larger LMs. To know if a larger LM still suffers from the softmax bottleneck problem, we input the examples we used in Table 5 to GPT-3.5 and report their results in Figure 4.\nWe find that although GPT-3.5 greatly reduces the chance of hallucination compared to GPT-2, the next word distribution is still not ideal. For example, in Figure 4a, although the incorrect answer queen receives only a small probability, GPT-3.5 puts around 67% probability on woman. Similarly, even though GPT-3.5 is unlikely to hallucinate the sentence: There are plates, keys, scissors, toys, and balloons in front of me, and I pick up the phone as GPT-2, Figure 4b and Figure 4d show that the output distribution is still heavily biased toward one of the options and the most likely next word could change if the order of the options in the context changes. These results suggest that increasing model size indeed alleviates the softmax bottleneck problem but the problem is not completely solved even if a huge hidden state size (12k) and model size (175B) are used (Brown et al., 2020). We expect that adding our methods to the large LMs could rectify the biased distributions as shown in our experiments on smaller LMs (Table 5). Therefore, although improving smaller LMs has already had wide applications in practice, trying our methods on a larger LM is a promising next step, which we haven't been able to do.\nThe current implementation of our methods also has some room for improvements. Our codes currently contain some unnecessary computation to circumvent the restrictions of PyTorch library, so we should be able to further accelerate it by writing CUDA code. Furthermore, our codes haven't supported the pretraining of BART or T5. We expect that completing the future work could make our method faster and better.\nSince the focus of this paper is improving the architecture of general transformer decoder, our evaluation of each application is not as comprehensive as the studies for a particular application. For example, although we test our methods using many metrics and the metrics show a consistent trend, there are many other factuality metrics we haven't tried (Li et al., 2022). We also haven't conducted human evaluation to further verify our conclusion because conducting human evaluation properly is challenging (Karpinska et al., 2021) and time-consuming. In addition, if we include more words in a context partition, the performance might be better at the cost of extra computational overhead. We leave the analyses of the tradeoff as future work."
},
{
"figure_ref": [],
"heading": "Ethics Statement",
"publication_ref": [
"b29"
],
"table_ref": [],
"text": "In our experiments, we find that our methods usually copy more words from the context or encoder input. The tendency might have some potential issues. For example, our improvements might be reduced on the languages with more morphology. Furthermore, in some summarization applications, increasing the factuality by increasing the extractiveness might not be ideal (Ladhak et al., 2022;Goyal et al., 2022a).\nAs described in Section 2.1, one major limita-tion of the popular softmax layer is its global word embeddings. The problem would become more serious when there are more tokens whose meanings are locally defined (e.g., names in the BookSum dataset). Our methods would be more useful in those circumstances and might alleviate some biases described in Finally, our results show that when there are some uncertainties in the next word (e.g., could be king or woman), existing LMs could have some difficulties of copying the words from the context and our methods alleviate the problem. Thus, our methods should also be able to improve the lexically controllable language generation models that put the desired keywords into the context such as Goldfarb-Tarrant et al. ( 2019 "
},
{
"figure_ref": [],
"heading": "A Appendix Overview",
"publication_ref": [],
"table_ref": [],
"text": "In the appendix, we first analyze our methods using more metrics in Appendix B and describe what we learn from the results. Next, we provide some details of our methods and baselines in Appendix C. Finally, we specify some experiment setups and hyperparameters in Appendix D."
},
{
"figure_ref": [
"fig_4"
],
"heading": "B More Results and Analysis",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we will report more results and provide more detailed analyses accordingly to investigate the advantages of different methods. 2020) demonstrate that the loss decreases linearly as the log of the model size increases. Therefore, a new architecture needs to perform better than the old architecture with a similar model size to verify that the improvement does not come from memorizing more information through the extra parameters. From the loss versus log(model size) curve in Figure 5, we can see that our proposed methods are significantly better than MoS and slightly better than a pointer network baseline as the model becomes larger."
},
{
"figure_ref": [],
"heading": "B.1 GPT-2 Experiments",
"publication_ref": [
"b48",
"b8",
"b5",
"b39",
"b5",
"b5"
],
"table_ref": [
"tab_13",
"tab_12"
],
"text": "We use the following metrics to measure the text generated by GPT-2.\n• ROUGE-1 (R1): The prediction F1 for unigram in the actual continuation.\n• ROUGE-1 Context (R1C): The prediction F1 for unigram in the context. • ROUGE-1 Proper (R1P): The same as ROUGE-1 except that only the proper nouns are considered. We measure this metric because the correctness of the entity name prediction is critical to the factuality of the generation.\n• ROUGE-1 Proper Context (R1PC): The same as ROUGE-1 Context (R1C) except that only the proper nouns are considered.\n• ROUGE-2 (R2): The prediction F1 for bigram in the actual continuation.\n• Proper Noun Ratio (P Ratio): The average number of proper nouns in the generation divided by the average number of proper nouns in the actual continuation. The LMs usually generate fewer proper nouns compared to the actual continuation (See et al., 2019), so the values are usually lower than 1. The P Ratio closer to 1 is better.\n• CIDEr (Vedantam et al., 2015): A metric for measuring the quality and specificity of the generation.\n• NIST (Doddington, 2002): Similar to CIDEr. CIDEr uses tf-idf to weigh the n-gram while NIST measures the information gain.\nThe results are reported in Table 8. In terms of R1, R2, CIDEr, and NIST, our proposed methods such as Softmax + C + Mi and Softmax + CPR:20,100 + Mi are significantly better than the pointer network baselines PS + Mi and PG + Mi. Comparing with Softmax + CPR:20,100 + Mi, PS + Mi has a significantly higher P Ratio and R1PC but similar R1P. This indicates that PS + Mi copies more proper nouns from the context while there is a similar number of proper nouns that are in actual continuation, so Softmax + CPR:20,100 + Mi actually has a higher accuracy on the proper noun prediction.\nIn text corpus such as Wikipedia, we do not know the ground truth next word distribution and which context leads to multiple probable next words, so we cannot quantitatively analyze the improvement on the ambiguous contexts. To alleviate the concern, we test our methods on the synthetic dataset constructed by Chang and McCallum (2022). The dataset is built using templates and Google analogy dataset (Mikolov et al., 2013), so we know the ground truth next word distribution. The dataset consists of the ambiguous contexts such as I went to Paris and Germany before, and I love one of the places more, which is, where the next word is either the diagonal words of the parallelogram such as Paris and Germany or the edge words such as Paris and France. For the details of the experimental setup, please refer to Chang and McCallum (2022).\nIn Table 7, we can see that Softmax + CPR:20,100 + Mi achieves the lowest perplexity in all subsets and outperforms the Softmax + Mi baseline by a large margin, especially in the diagonal subset where the ground truth word embedding distribution has multiple modes. Notice that the performance of MoS + Mi is worse than what reported in Chang and McCallum (2022) probably because we shared the input and output word embeddings."
},
{
"figure_ref": [
"fig_4",
"fig_6"
],
"heading": "B.2 Summarization",
"publication_ref": [
"b47"
],
"table_ref": [
"tab_15",
"tab_16",
"tab_9"
],
"text": "Compared to Figure 5, Figure 6 shows that our methods improve the loss of T5 in CNN/DM more than GPT-2 in Wikipedia.\nIn Table 9 and Table 10, we compare the different summarization models by their model size, evaluation losses, inference time, and other metrics which we use in subsection B.1. The pointer network baselines and our methods significantly improve most metrics over the softmax baseline, which is used ubiquitously in nearly all LMs. Although our method generally improves less on the T5-Base model, the percentages of additional parameters and inference time overhead are much smaller. Although our methods tend to improve less in larger language model, we still improve BART Large very significantly in NIST, CIDEr, and MAUVE, and Mi seems to become more effective in BART Large.\nThe testing set of SAMSUM dataset only has 819 samples, so some metrics such as R1 and R2 are not as stable as other three datasets. PG (See et al., 2017) for T5-Small and T5-Base perform much worse in SAMSUM dataset. We hypothesize that it is because the dialog input in SAMSUM dataset is very different from the pretraining data of T5, which makes training PG unstable.\nIn most datasets and models, the R Ratio from our method is significantly closer to 1 than the softmax baseline, which means the average number of proper nouns in our summaries is much closer to the average number of proper nouns in the humanwritten summary. For example, in BookSum Paragraph, we improve its R Ratio by 26%, which partially explains our large MAUVE improvement in Table 6. Notice that our methods do not always output more proper nouns. For example, for BART Base in CNN/DM dataset, our methods reduce the R Ratio of the softmax baseline, which is larger than 1. This shows that our methods could learn when we should copy the proper nouns according to the training data."
},
{
"figure_ref": [],
"heading": "C Method Details",
"publication_ref": [],
"table_ref": [],
"text": "We describe some details of our methods and baselines in this section. "
},
{
"figure_ref": [],
"heading": "C.1 Proposed Methods",
"publication_ref": [],
"table_ref": [],
"text": "To allow us to start from existing LMs that are pretrained using softmax, we keep the modified softmax layer initially working almost the same as the original softmax layer. We initialize the linear transformation weights of L f P D (), L f LD (), L f P E (), and L f LE () as 10 -10 • I. The other linear weights L f . () are initialized as the identity matrix I.\nIn the local decoder embedding method Softmax + P + Mi, the initialization would give the 0 logit to all context words. To solve the issue, we revise Equation 3 a little and compute Logit P (x, c t ) by\nf T ct,V w x + f T ct,P D f x,ct,LD if x ∈ c t f T ct,V w x O/W .(7)\nThat is, we initially rely on the original softmax layer to compute all the logits and let the term f T ct,P D f x,ct,LD gradually influences the logits of the context words.\nIn MoS + CPR:20,100 + Mi, our proposed method only revises the logit in one of the softmax."
},
{
"figure_ref": [],
"heading": "C.2 Pointer Network Baselines",
"publication_ref": [
"b15",
"b47",
"b38"
],
"table_ref": [
"tab_8"
],
"text": "The pointer networks are originally designed for RNN, so we are unable to use exactly the same formula proposed in the papers. Nevertheless, we try our best to adapt the pointer networks for the transformer encoder while keeping the gist of the formulas. In all methods, to let the results more comparable to our methods, we use f ct,P E and L f LE to determine the probability of copying the words from the context, and use f T ct,V w x to determine the probability of generating all the words in the vocabulary.\nIn CopyNet (Gu et al., 2016), we compute the probability of outputting the word x as Notice that CopyNet needs to sum up the exponential of dot products, which often causes overflow problems in GPT-2. We can set b to be a large negative value initially to solve the problem, but its perplexity is much worse than the other two pointer network variants. Thus, we choose to skip the CopyNet in the GPT-2 experiments.\nIn Pointer Generator (See et al., 2017), we compute the probability of x using\nP rob(x|I, c t ) = p gen exp f T ct,V w x Z V +(1 -p gen ) |I| j=1 1 I j =x P E (j|I, c t ),(9)\nwhere P E (j|I, c t ) = exp v T tanh(f c t ,P E +L f LE (h M I j )+b) Z E , p gen = σ(q T h M ct + b ptr ), the normalization term Z V =\nx∈V exp f T ct,V w x and Z E = |I| j=1 exp v T tanh(f ct,P E + L f LE (h M I j ) + b) . We skip the coverage mechanism in the pointer generator paper to make it more comparable to other methods. In T5 experiments, its training loss is sometimes very large, so we set b ptr as 3 initially to keep the p gen close to 1 (i.e., turn pointer part off initially). In other experiments, we set b ptr = 0.\nIn Pointer Sentinel (Merity et al., 2017), the probability of x is computed by\nP rob(x|I, ct) = g exp f T c t ,V wx ZV + |I| j=1 1 I j =x exp f T c t ,P E tanh(L f LE (h M I j )) + b Zp ,(10)\ng = exp(q T h M c t ) Zp , and Z p exp f T ct,P E tanh(L f LE (h M I j )) + b . In our experiments, we find that the pointer network variants usually have similar performance (except that PG sometimes performs much worse in summarization due to some training stability issues). This suggests that the differences in the pointer network variants often do not influence the performance significantly, which justifies our simplification of the formulas in the original paper and supports our conclusion that the improvement comes from breaking the softmax bottleneck.\nNotice that in the above pointer network variants, the pointer part can only increase the probability of the context words from the generator part. As a result, it cannot alleviate the repetition problem in the last example of Table 5."
},
{
"figure_ref": [],
"heading": "C.3 Word-by-word Reranker Baseline",
"publication_ref": [],
"table_ref": [],
"text": "We illustrate our word-by-word reranker (wbwR) in Figure 7. The method has two stages. In the first stage, we compute the logits using the projected hidden state f ct,V and retrieve the top k words. At the second stage, we append the top k words to the input context along with the hidden state f ct,R for reranking the context words. 3 We use the same positional embeddings for all candidates to encourage the model to change the ranking of the words. Next, we use the hidden states corresponding to the candidates to compute their local word embeddings as f x,ct,LD . Finally, we re-estimate the probabilities of top k words by\nf T ct,V w x + f T ct,R f x,ct,LD if x ∈ W (k) f T ct,V w x O/W .(11)\nTo improve the quality of our top k candidates, the final loss is the addition of the wbwR loss at the second stage and the loss of the original softmax layer that only uses the logits from f T ct,V w x at the first stage. When we combine the wbwR with Softmax + CPR:20,100 + Mi, we simply use Softmax + CPR:20,100 + Mi at the first stage and use the wbwR to overwrite the logits of Softmax + CPR:20,100 + Mi at the second stage.\nUsing this method, we can update the embeddings of the words that are not in the context and allow the candidates to interact with the input context to determine their probabilities as the classic two-stage reranker while keeping the model size roughly unchanged. Nevertheless, the method can only change the probability of the top k words and its computational overhead and memory requirement prevents us from using a very large k."
},
{
"figure_ref": [],
"heading": "Global Word Embeddings",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "GPT-2 encoder ……",
"publication_ref": [],
"table_ref": [],
"text": "After debating whether to bow to the king or the woman first, the jester decided on the Unlike the standard GPT-2, we cannot get the probability of all positions in one forward pass because the input contexts are different when computing the probability at each position and the input of the second stage reranker depends on the results of the previous forward pass at the first stage. To speed up, we reuse the computed hidden states and batchify the forward passes.\nIn our implementation, we first get the top k candidates corresponding to all tokens in the stage1 (just original GPT2) as the input of stage2 reranker. To avoid recalculating the hidden states of the context at stage2, we store the hidden states using the past-key-value in Hugging Face and only compute the hidden states corresponding to the top k candidate tokens at stage2.\nWe divide the computation of the whole input sequence into several blocks as shown in Figure 8. In each block, we input a batch containing the last few tokens and top k candidates into the GPT-2 while reusing the hidden states of their common contexts from stage1. In this way, we can increase parallelism by increasing the block size if the GPU memory allows it.\nEven though we spent substantial effort on optimizing the wbwR, the method is still too slow to be practically useful. Even if we use four RTX 8000 (a faster GPU with a larger memory), our wbwR implementation is still around 10 times slower than our proposed Softmax + CPR:20,100 + Mi that uses only one RTX 2080."
},
{
"figure_ref": [],
"heading": "D Experiment Details",
"publication_ref": [],
"table_ref": [],
"text": "For the reproducibility, we provide some experimental configuration in this section.\nPlease see our codes (https: //github.com/iesl/Softmax-CPR) for more details."
},
{
"figure_ref": [],
"heading": "D.1 GPT-2 Experimental Details",
"publication_ref": [
"b5"
],
"table_ref": [],
"text": "We mostly follow the experimental setup Chang and McCallum (2022) except that we share the input and output word embeddings as in the standard GPT-2 models. As in Chang and McCallum (2022), we use the last 2% of the corpus as the test set and the 2% before that as the validation set. 4 In 4 We do not shuffle the corpus before splitting the datasets. We found that our improvement could be even larger if we In the text completion experiment, we generate 360k continuations with a length of 50 given the prompts in Wikipedia. We first sample 40k sequences in the test data of Wikipedia 2021. Next, we use the first 20, 70, and 120 words in the sequence as our context and let the different models generate the next 50 words as continuations. The references are the actual next 50 words. All the methods use Top-K sampling and K=5."
},
{
"figure_ref": [],
"heading": "D.2 Summarization Experimental Details",
"publication_ref": [
"b27"
],
"table_ref": [],
"text": "BookSum dataset (Kryściński et al., 2021) includes three summarization tasks: Summarizing a book, a chapter, and a paragraph. We test our methods using the paragraph summarization task due to the input length restriction of BART and T5. The dataset is constructed by automatically aligning the paragraphs in a chapter with the sentences in a chapter summary, which introduces noise to the dataset. Similarly, XSUM uses the first sentence in news instead of manually-written summary as the ground truth reference. The relatively noisier datasets such as XSUM and BookSum Paragraph, and smaller dataset like SAMSUM could test the stability of the methods. The sizes of the summarization datasets could be found in Table 11.\nshuffle the corpus to let the training data distribution closer to the testing data distribution.\nWe conduct the summarization experiments based on a summarization example code from Hugging Face 5 . Most of our hyperparameters use the default value in the code. In our preliminary study, our improvement is not sensitive to the hyperparameter choice (e.g., the improvement gap is similar across different numbers of epochs). Thus, we do not tune the hyperparameters for each method or for each dataset unless we cannot reach a low training loss at the end.\nIn CNN/DM, XSUM, and SAMSUM datasets, We train models for 3 epochs. In BookSum datasets, We trained models for 5 epochs. 6 The learning rate is set to be 5e -05 except for BART Large model in BookSum, where we use 1e -05 to stablize the training of all methods.\nAll the experiments use batch size 8 and AdamW with betas=(0.9,0.999), epsilon=1e -06, weight-decay=1.2e -6. During the generation, we used Top-K sampling (K=10) as our decoding method. The maximum summary length is set as 128 and maximum input length is 1024. We use warmup for the first 1000 steps in all the experiments, which allows us to change the architecture of T5 and BART more significantly (e.g., using Mi) without having a training stability issue.\nThe k in the reranker partition and the block size of multiple input hidden states (Mi) is coarsely tuned based on validation performance of CNN/DM. Unlike considering the top 100 words in the open-end text completion using GPT-2, we find that reranking the top 20 words is sufficient for our summarization models, probably because next words are easier to predict in the summarization task.\nFor our evaluation metrics, we use the default setting for ROUGE 7 and set use_stemmer=True. When reporting the ROUGE scores, we follow the conventions to show their percentages. We use the default setting for MAUVE 8 , CIDER 9 , NIST 10 . For MAUVE, we insert a new line symbol after every sentence as in the original Hugging Face "
},
{
"figure_ref": [],
"heading": "D.3 Computational Environment and Software",
"publication_ref": [
"b59",
"b19"
],
"table_ref": [],
"text": "We implement our methods by revising the Hugging Face library (Wolf et al., 2020). From Hugging Face, we load the pretrained LMs including GPT-2 Small 12 , GPT-2 Medium 13 , T5-Small 14 , T5-Base 15 , BART Base 16 , and BART Large 17 . We use SpaCy (Honnibal et al., 2020) to detect the proper nouns.\nFor GPT-2 Medium, T5-Base, and BART Large, we use NVIDIA GeForce RTX 8000 to train the model and for other smaller models, we use NVIDIA GeForce RTX 2080. Most of experiments could be done within one week. In all the inference time experiments, we use NVIDIA GeForce GTX TITAN X, batch size 4 for GPT-2, and batch size 8 for BART and T5."
}
] | Is the output softmax layer, which is adopted by most language models (LMs), always the best way to compute the next word probability? Given so many attention layers in a modern transformer-based LM, are the pointer networks redundant nowadays? In this study, we discover that the answers to both questions are no. This is because the softmax bottleneck sometimes prevents the LMs from predicting the desired distribution and the pointer networks can be used to break the bottleneck efficiently. Based on the finding, we propose several softmax alternatives by simplifying the pointer networks and accelerating the word-by-word rerankers. In GPT-2, our proposals are significantly better and more efficient than mixture of softmax, a state-of-theart softmax alternative. In summarization experiments, without significantly decreasing its training/testing speed, our best method based on T5-Small improves factCC score by 2 points in CNN/DM and XSUM dataset, and improves MAUVE scores by 30% in Book-Sum paragraph-level dataset. | Revisiting the Architectures like Pointer Networks to Efficiently Improve the Next Word Distribution, Summarization Factuality, and Beyond | [
{
"figure_caption": "Figure 2 :2Figure 2: We simplify the pointer network / reranker by using another embedding h ct,S for the words in the context / the top-k likely words.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "(a) The example where the next word should be either woman or king (or their synonym such as former and latter). (b) The example where the next word plates, keys, scissors, toys, and balloons should receive similar probabilities. (c) The example where the next word John, Alex, Mary, Kathryn, and Jack should receive similar probabilities. (d) Same as above except that the order of the objects in the context is different.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: The next word probabilities outputted by GPT-3.5 (text-davinci-003).",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Shwartz et al. (2020) andLadhak et al. (2023). Moreover, the meaning of tokens are also locally defined in many other applications such as variables in code or math problems, the new terminologies in a scientific paper, or the products in a sequential recommendation problem. We believe that our methods could become an efficient alternative of reranker(Cobbe et al., 2021; Welleck et al., 2022) and create impacts in those areas.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 :5Figure 5: The model size versus the model loss in Wikipedia test data after training for 0.4 epochs. The left side points are the results from GPT-2 Small and the right side points come from GPT-2 Medium. The lower curves are better.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Kaplan et al. (2020); Henighan et al. (",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 6 :6Figure 6: The model size versus the model loss in CNN/DM test set. The left side points are the results from T5-Small and the right side points come from T5-Base. The lower curves are better.",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "P 1 I1rob(x|I, c t ) ∝ exp f T ct,V w x + |I| j=1 j =x exp f T ct,P E L f LE (h M I j ) + b . (8)",
"figure_data": "",
"figure_id": "fig_7",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Comparison of different methods on top of GPT-2. Wiki and OWT refer to the testing perplexity of Wikipedia 2021 and OpenWebText, respectively. Lower perplexity is better. Time is the inference time of a batch; Mi is the multiple input hidden state enhancement; C is the context partition; R:20,100 is the reranker partition with k Pointer Generator (PG + Mi) (See et al., 2017), and Pointer Sentinel (PS + Mi) (Merity et al., 2017). 2 Their similar performances indicate that the improvement of pointer networks come from breaking the softmax bottleneck. The significantly better performance of PS + Mi compared to PS further supports the finding.",
"figure_data": "GPT-2 SmallGPT-2 MediumModel NameSizeTime (ms) OWT (↓) Wiki (↓)SizeTime (ms) OWT (↓) Wiki (↓)Softmax (GPT-2)125.0M82.918.9624.28355.9M207.815.8120.12Softmax + Mi130.9M85.618.7424.08366.4M213.815.7120.07Mixture of Softmax (MoS) (Yang et al., 2018) 126.2M130.218.9724.10358.0M262.915.7119.95MoS + Mi (Chang and McCallum, 2022)133.3M133.218.6823.82370.6M268.215.6119.86Pointer Generator (PG) (See et al., 2017)126.2M106.018.6723.70358.0M237.815.7219.95Pointer Sentinel (PS) (Merity et al., 2017)126.2M94.118.7023.79358.0M218.315.7219.95Softmax + R:20 + Mi132.1M90.418.6724.03368.5M203.615.6419.94Softmax + R:20,100 + Mi133.3M101.118.6923.93370.6M228.515.6119.89Softmax + C + Mi132.1M94.818.4823.56368.5M222.715.6019.83Softmax + P + Mi133.3M99.118.5823.66370.6M214.715.6319.90PG + Mi133.3M111.218.4323.43370.6M242.515.6019.89PS + Mi133.3M98.018.4823.53370.6M224.615.6019.87Softmax + CR:20,100 + Mi134.5M113.318.4623.48372.7M234.515.5419.75Softmax + CPR:20,100 + Mi136.8M119.918.4323.42376.9M249.915.5319.71MoS + CPR:20,100 + Mi139.2M165.118.3923.29381.1M300.615.4419.57",
"figure_id": "tab_3",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": "AllProper NounModel NameRef Context Ref ContextSoftmax + Mi22.9024.047.4914.84MoS + Mi22.8823.987.7015.49PS + Mi22.8525.018.1618.21Softmax + CPR:20,100 + Mi 23.0525.368.1617.92",
"figure_id": "tab_5",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": "show that different meth-",
"figure_id": "tab_6",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Prediction visualization of three input contexts. We show the top five words with the highest prediction probabilities of each model. The reasonable next word predictions are boldfaced.",
"figure_data": "There are plates, keys, scissors, toys,Choosing between John, Alex,I like tennis, baseball, golf, basketball,Input Contextand balloons in front of me, and IMary, Kathryn, and Jack, I decidedandpick up theto first talk toSoftmax + Mikeys 0.108, pieces 0.045, key 0.036, phone 0.020, balloons 0.019John 0.108, the 0.102, them 0.095, him 0.045, my 0.032tennis 0.089, baseball 0.075, football 0.041, basketball 0.036, I 0.032Mixture of Softmax (MoS) + Mikeys 0.085, phone 0.035, key 0.031, pieces 0.029, balloons 0.016John 0.099, the 0.097, them 0.083, Alex 0.055, Mary 0.040baseball 0.076, basketball 0.062, tennis 0.059, golf 0.037, bad 0.035Pointer Sentinel (PS) + Mikeys 0.091, plates 0.079, scissors 0.050, balloons 0.034, toys 0.033John 0.130, the 0.105, Alex 0.076, them 0.076, Mary 0.037tennis 0.095, golf 0.050, baseball 0.043, I 0.038, other 0.038Softmax + CPR:20,100 + Mikeys 0.077, balloons 0.052, plates 0.036, toys 0.030, pieces 0.030the 0.106, John 0.099, my 0.060, Alex 0.057, them 0.044football 0.075, volleyball 0.058, soccer 0.056, I 0.047, bad 0.038CNN/DMXSUMBookSum ParagraphSAMSUMModel NameR1CIDEr factCC MAUVER1CIDEr factCC MAUVER1CIDEr factCC MAUVER1CIDEr factCC MAUVET5-SmallSoftmax (S)38.255 0.4420.4620.86128.713 0.4460.2540.93916.313 0.0830.4240.32839.472 0.8170.5770.898CopyNet (Gu et al., 2016) 37.990 0.4380.4820.86528.573 0.4420.2740.94016.666 0.0920.4390.40239.525 0.8530.5790.924PG (See et al., 2017)37.913 0.4420.4670.87428.777 0.4500.2570.93116.432 0.0880.4290.37632.451 0.5850.5520.153PS (Merity et al., 2017)38.058 0.4440.4660.85428.442 0.4350.2670.93216.408 0.0900.4360.39538.731 0.8170.5780.865S + R:2037.881 0.4330.4740.87228.557 0.4400.2560.93116.336 0.0860.4310.37039.073 0.7520.5790.847S + E38.137 0.4410.4770.86628.723 0.4440.2720.94216.542 0.0900.4350.39039.056 0.7840.5790.904S + CE38.461 0.4600.4750.87429.155 0.4640.2700.94816.628 0.0930.4360.40340.055 0.8350.5830.943S + CER:2038.346 0.4500.4820.89029.067 0.4590.2760.94216.638 0.0930.4360.40040.505 0.8460.5800.915S + CEPR:2038.807 0.4560.4810.87729.395 0.4740.2730.94216.894 0.0980.4400.41840.127 0.8910.5820.946S + CEPR:20 + Mi38.675 0.4510.4750.87829.348 0.4700.2750.94616.738 0.0960.4380.42640.328 0.8740.5820.932T5-BaseSoftmax (S)40.198 0.5040.4780.90733.571 0.6670.2490.97916.761 0.0960.4240.46744.348 1.0460.5740.986CopyNet (Gu et al., 2016) 39.940 0.5070.4840.90333.557 0.6660.2530.97916.918 0.1010.4300.53144.141 1.0520.5700.973PG (See et al., 2017)39.982 0.4890.4850.91133.605 0.6630.2550.98216.611 0.0950.4230.46337.597 0.7840.5480.140PS (Merity et al., 2017)40.018 0.4950.4830.91433.638 0.6720.2490.98316.905 0.1000.4280.50443.098 1.0080.5750.946S + CEPR:2040.354 0.5110.4870.91933.700 0.6750.2600.98016.997 0.1000.4320.54944.860 1.0640.5730.963S + CEPR:20 + Mi40.510 0.5060.4810.91833.853 0.6830.2630.98316.975 0.1010.4310.54644.488 1.0550.5760.980BART BaseSoftmax (S)39.390 0.4280.4790.90035.675 0.8140.2410.98516.393 0.0940.4140.40445.132 1.1290.5670.966CopyNet (Gu et al., 2016) 39.385 0.4380.4840.90635.515 0.8140.2510.98816.642 0.1000.4220.49544.316 1.1030.5770.970PG (See et al., 2017)39.264 0.4440.4890.90935.653 0.8100.2420.98716.402 0.0940.4140.40245.278 1.1530.5780.977PS (Merity et al., 2017)39.471 0.4590.4900.90635.411 0.8090.2470.98616.718 0.0990.4220.49244.575 1.0840.5730.974S + R:2039.181 0.4340.4750.90535.586 0.8080.2470.98816.419 0.0960.4180.43945.024 1.1540.5720.970S + E39.267 0.4390.4830.90735.698 0.8190.2410.98816.442 0.0970.4150.42944.825 1.1060.5720.981S + CE39.416 0.4420.4810.90835.727 0.8120.2410.98816.555 0.0960.4170.43544.295 1.1160.5720.985S + CER:2039.421 0.4390.4820.90035.576 0.8120.2360.98716.553 0.0960.4180.45445.054 1.1500.5760.988S + CEPR:2039.723 0.4410.4830.90835.732 0.8220.2420.98616.664 0.0980.4200.46744.732 1.1150.5750.974S + CEPR:20 + Mi39.626 0.4420.4820.90735.846 0.8280.2450.98616.597 0.0970.4190.46644.728 1.1320.5740.988BART LargeSoftmax (S)40.749 0.4240.4950.89938.828 0.9210.2630.98817.271 0.1030.4200.46147.384 1.1870.5740.975CopyNet (Gu et al., 2016) 40.622 0.4070.4870.89038.576 0.9200.2580.98917.342 0.1060.4250.51247.911 1.2320.5730.980PG (See et al., 2017)40.766 0.4070.4890.90238.869 0.9440.2560.99017.289 0.1030.4240.47047.737 1.1990.5730.964PS (Merity et al., 2017)40.643 0.4240.5020.90738.886 0.9520.2550.98817.382 0.1050.4260.52748.253 1.2460.5740.986S + CEPR:2040.876 0.4580.5000.92538.991 0.9550.2480.99017.337 0.1060.4230.46747.253 1.2980.5720.976S + CEPR:20 + Mi40.441 0.4630.5000.92738.705 0.9650.2420.99116.995 0.1050.4210.48247.488 1.2710.5710.986",
"figure_id": "tab_8",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "",
"figure_data": "lihood training (Welleck et al., 2020; Jiang et al.,Jiang et al., 2022a). Although being effective, the2022b; Su et al., 2022). However, when LM shouldrerankers usually slow down significantly the train-mention some names in the context, this might ex-ing and/or inference speed (as our word-by-wordacerbate the hallucination problem. In contrast, ourreranker baseline) and might occupy extra memorymethod can learn to copy and exclude the words inresources.context as in Table 5.To alleviate the hallucination problem or sat-isfy some constraints, many recent generation mod-els rerank the generated text (Deng et al., 2020;Gabriel et al., 2021; Cobbe et al., 2021; Ravautet al., 2022; Krishna et al., 2022; Glass et al.,2022; An et al., 2022; Arora et al., 2022; Adolphset al., 2022; Meng et al., 2022; Mireshghallah et al.,2022; Kumar et al., 2022; Wan and Bansal, 2022;",
"figure_id": "tab_9",
"figure_label": "6",
"figure_type": "table"
},
{
"figure_caption": "al., 2021; Zhong et al., 2022; Ma et al., 2023). Another example is encouraging the word embeddings to be isotropy (Wang et al., 2020; Su et al., 2022). Their improvement might also come from reducing linear dependency of the candidate word embeddings. Nevertheless, their side effect of breaking the similarity structure in the word embedding space might hurt the generation quality in some cases. Concurrently to our work, Wan et al. (2023) also use the softmax bottleneck theory",
"figure_data": "",
"figure_id": "tab_10",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": ") and Lu et al. (2021).",
"figure_data": "3.20 3.25Softmax Softmax + Mi MoS + Mi Softmax + C + Mi PS + Mi Softmax + CPR:100,20 + Mi3.15loss3.103.053.0018.618.819.0 ln(number of parameters) 19.2 19.419.619.8",
"figure_id": "tab_11",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Diagonal (e.g., king or woman) Comparing the perplexity of different GPT-2 Small models using the synthetic dataset fromChang and McCallum (2022).",
"figure_data": "Edge (e.g., king or queen)",
"figure_id": "tab_12",
"figure_label": "7",
"figure_type": "table"
},
{
"figure_caption": "Comparison of the continuation generated by GPT-2 Small in Wikipedia test data. Table 4 is a short summary of this table. The meaning of the metrics is described in Appendix B.1. Higher R1C and R1PC mean copying more words from the context. A higher P Ratio means generating more proper nouns. All ROUGE scores are percentages.",
"figure_data": "",
"figure_id": "tab_13",
"figure_label": "8",
"figure_type": "table"
},
{
"figure_caption": "Comparison of the summaries generated by different models in the test sets of CNN/DM and XSUM datasets. We also report the number of parameters of each model. From top to bottom, the four sections are the results of T5-Small, T5-Base, BART Base, and BART Large. The meaning of the metrics are described in Appendix B.1. R2 (ROUGE 2-F1) scores are percentages. Within each section, we highlight the smallest loss, the P Ratio that is closest to 1, and highest numbers in the other metrics.",
"figure_data": "",
"figure_id": "tab_15",
"figure_label": "9",
"figure_type": "table"
},
{
"figure_caption": "Comparison of the summaries generated by different models in the test sets of BookSum and SAM-SUM datasets. We also report the inference time of one samples. The meaning of the metrics are described in Appendix B.1. R2 (ROUGE 2-F1) scores are percentages. Within each section, we highlight the smallest loss, the P Ratio that is closest to 1, and highest numbers in the other metrics.",
"figure_data": "BookSum ParagraphSAMSUMModelsTime (ms) Loss (↓)R2R1P P Ratio NIST Loss (↓)R2R1P P Ratio NISTT5-SmallSoftmax (S)30.10.6541.673 0.149 0.589 1.3830.38313.806 0.605 0.873 3.945CopyNet (Gu et al., 2016)37.00.6461.722 0.183 0.747 1.4400.38114.210 0.594 0.809 3.965PG (See et al., 2017)43.40.6481.669 0.160 0.631 1.4130.39210.673 0.542 0.711 1.665PS (Merity et al., 2017)37.60.6461.627 0.177 0.700 1.4170.38313.817 0.583 0.794 3.960S + R:2032.90.6521.663 0.159 0.677 1.4030.38013.728 0.598 0.870 3.995S + E33.80.6451.710 0.171 0.673 1.4210.37013.557 0.602 0.892 3.906S + CE34.00.6441.734 0.173 0.680 1.4360.36814.136 0.619 0.892 3.971S + CER:2035.80.6421.710 0.174 0.693 1.4340.36714.281 0.627 0.911 3.968S + CEPR:2038.40.6411.768 0.184 0.725 1.4610.36514.451 0.639 0.909 4.034S + CEPR:20 + Mi41.70.6411.733 0.185 0.721 1.4580.36514.193 0.630 0.922 4.011T5-BaseSoftmax (S)102.40.5871.876 0.160 0.650 1.4430.30817.662 0.672 0.915 4.559CopyNet (Gu et al., 2016)0.5820.744 1.4810.30717.556 0.678 0.901 4.544PG (See et al., 2017)117.70.5851.832 0.159 0.647 1.4340.31714.649 0.611 0.740 1.870PS (Merity et al., 2017)112.00.5821.899 0.176 0.718 1.4650.30817.502 0.660 0.897 4.453S + CEPR:20115.30.5801.842 0.191 0.771 1.4820.30018.082 0.677 0.950 4.553S + CEPR:20 + Mi116.30.5841.860 0.187 0.770 1.4770.30117.617 0.677 0.938 4.521BART BaseSoftmax (S)46.60.6241.807 0.141 0.656 1.4250.32719.379 0.672 0.995 4.546CopyNet (Gu et al., 2016)57.80.6131.866 0.166 0.728 1.4540.32618.227 0.662 0.944 4.535PG (See et al., 2017)64.80.6241.864 0.140 0.668 1.4280.32818.791 0.673 0.963 4.537PS (Merity et al., 2017)57.90.6131.867 0.163 0.723 1.4610.32418.367 0.674 0.951 4.573S + R:2050.50.6271.807 0.154 0.720 1.4300.32619.022 0.671 0.971 4.608S + E54.20.6201.825 0.150 0.688 1.4290.32418.902 0.680 0.970 4.501S + CE56.50.6191.847 0.153 0.685 1.4410.32318.739 0.672 0.949 4.537S + CER:2057.20.6181.834 0.156 0.727 1.4440.32119.267 0.678 0.981 4.561S + CEPR:2058.80.6181.865 0.157 0.742 1.4570.32118.631 0.670 0.992 4.516S + CEPR:20 + Mi63.20.6201.827 0.158 0.733 1.4420.32218.681 0.670 0.987 4.439BART LargeSoftmax (S)143.50.5542.094 0.171 0.722 1.4720.30320.848 0.711 1.006 4.621CopyNet (Gu et al., 2016)168.90.5482.087 0.184 0.762 1.4900.29821.703 0.708 1.026 4.727PG (See et al., 2017)178.30.7312.090 0.174 0.725 1.4790.30121.428 0.706 1.051 4.604PS (Merity et al., 2017)168.50.7262.083 0.184 0.760 1.4930.30022.144 0.710 1.036 4.779S + CEPR:20169.90.5522.069 0.178 0.763 1.5050.30221.326 0.691 1.017 4.595S + CEPR:20 + Mi177.40.5442.024 0.175 0.737 1.5000.29421.244 0.713 0.959 4.746",
"figure_id": "tab_16",
"figure_label": "10",
"figure_type": "table"
}
] | Haw-Shiuan Chang; Zonghai Yao; Alolika Gon; Hong Yu; Andrew Mccallum; ‡ Cics | [
{
"authors": "Leonard Adolphs; Tianyu Gao; Jing Xu; Kurt Shuster; Sainbayar Sukhbaatar; Jason Weston",
"journal": "",
"ref_id": "b0",
"title": "The cringe loss: Learning what language not to model",
"year": "2022"
},
{
"authors": "Chenxin An; Jiangtao Feng; Kai Lv; Lingpeng Kong; Xipeng Qiu; Xuanjing Huang",
"journal": "",
"ref_id": "b1",
"title": "Cont: Contrastive neural text generation",
"year": "2022"
},
{
"authors": "Kushal Arora; Kurt Shuster; Sainbayar Sukhbaatar; Jason Weston",
"journal": "",
"ref_id": "b2",
"title": "Director: Generator-classifiers for supervised language modeling",
"year": "2022"
},
{
"authors": "B Tom; Benjamin Brown; Nick Mann; Melanie Ryder; Jared Subbiah; Prafulla Kaplan; Arvind Dhariwal; Pranav Neelakantan; Girish Shyam; Amanda Sastry; Sandhini Askell; Ariel Agarwal; Gretchen Herbert-Voss; Tom Krueger; Rewon Henighan; Aditya Child; Daniel M Ramesh; Jeffrey Ziegler; Clemens Wu; Christopher Winter; Mark Hesse; Eric Chen; Mateusz Sigler; Scott Litwin; Benjamin Gray; Jack Chess; Christopher Clark; Sam Berner; Alec Mc-Candlish; Ilya Radford; Dario Sutskever; Amodei",
"journal": "",
"ref_id": "b3",
"title": "Language models are few-shot learners",
"year": "2020-12-06"
},
{
"authors": "Shuyang Cao; Lu Wang",
"journal": "Association for Computational Linguistics",
"ref_id": "b4",
"title": "CLIFF: Contrastive learning for improving faithfulness and factuality in abstractive summarization",
"year": "2021"
},
{
"authors": "Haw-Shiuan Chang; Andrew Mccallum",
"journal": "Association for Computational Linguistics",
"ref_id": "b5",
"title": "Softmax bottleneck makes language models unable to represent multi-mode word distributions",
"year": "2022"
},
{
"authors": "Karl Cobbe; Vineet Kosaraju; Mohammad Bavarian; Mark Chen; Heewoo Jun; Lukasz Kaiser; Matthias Plappert; Jerry Tworek; Jacob Hilton; Reiichiro Nakano",
"journal": "",
"ref_id": "b6",
"title": "Training verifiers to solve math word problems",
"year": "2021"
},
{
"authors": "Yuntian Deng; Anton Bakhtin; Myle Ott; Arthur Szlam; Marc'aurelio Ranzato",
"journal": "",
"ref_id": "b7",
"title": "Residual energybased models for text generation",
"year": "2020-04-26"
},
{
"authors": "George Doddington",
"journal": "",
"ref_id": "b8",
"title": "Automatic evaluation of machine translation quality using n-gram cooccurrence statistics",
"year": "2002"
},
{
"authors": "Saadia Gabriel; Antoine Bosselut; Jeff Da; Ari Holtzman; Jan Buys; Kyle Lo; Asli Celikyilmaz; Yejin Choi",
"journal": "",
"ref_id": "b9",
"title": "Discourse understanding and factual consistency in abstractive summarization",
"year": "2021"
},
{
"authors": "Michael Glass; Gaetano Rossiello; Md Faisal; Mahbub Chowdhury; Ankita Naik; Pengshan Cai; Alfio Gliozzo",
"journal": "Association for Computational Linguistics",
"ref_id": "b10",
"title": "Re2G: Retrieve, rerank, generate",
"year": "2022"
},
{
"authors": "Bogdan Gliwa; Iwona Mochol; Maciej Biesek; Aleksander Wawer",
"journal": "Association for Computational Linguistics",
"ref_id": "b11",
"title": "SAMSum corpus: A human-annotated dialogue dataset for abstractive summarization",
"year": "2019"
},
{
"authors": "Seraphina Goldfarb-Tarrant; Haining Feng; Nanyun Peng",
"journal": "Association for Computational Linguistics",
"ref_id": "b12",
"title": "Plan, write, and revise: an interactive system for open-domain story generation",
"year": "2019"
},
{
"authors": "Tanya Goyal; Junyi ; Jessy Li; Greg Durrett",
"journal": "",
"ref_id": "b13",
"title": "News summarization and evaluation in the era of gpt-3",
"year": "2022"
},
{
"authors": "Tanya Goyal; Junyi ; Jessy Li; Greg Durrett",
"journal": "",
"ref_id": "b14",
"title": "Snac: Coherence error detection for narrative summarization",
"year": "2022"
},
{
"authors": "Jiatao Gu; Zhengdong Lu; Hang Li; O K Victor; Li",
"journal": "Association for Computational Linguistics",
"ref_id": "b15",
"title": "Incorporating copying mechanism in sequence-to-sequence learning",
"year": "2016"
},
{
"authors": "Jian Guan; Zhenyu Yang; Rongsheng Zhang; Zhipeng Hu; Minlie Huang",
"journal": "",
"ref_id": "b16",
"title": "Generating coherent narratives by learning dynamic and discrete entity states with a contrastive framework",
"year": "2022"
},
{
"authors": "Tom Henighan; Jared Kaplan; Mor Katz; Mark Chen; Christopher Hesse; Jacob Jackson; Heewoo Jun; Prafulla Tom B Brown; Scott Dhariwal; Gray",
"journal": "",
"ref_id": "b17",
"title": "Scaling laws for autoregressive generative modeling",
"year": "2020"
},
{
"authors": "Sepp Hochreiter; Jürgen Schmidhuber",
"journal": "Neural computation",
"ref_id": "b18",
"title": "Long short-term memory",
"year": "1997"
},
{
"authors": "Matthew Honnibal; Ines Montani; Sofie Van Landeghem; Adriane Boyd",
"journal": "",
"ref_id": "b19",
"title": "spaCy: Industrial-strength Natural Language Processing in Python",
"year": "2020"
},
{
"authors": "Dongfu Jiang; Bill Yuchen Lin; Xiang Ren",
"journal": "",
"ref_id": "b20",
"title": "Pairreranker: Pairwise reranking for natural language generation",
"year": "2022"
},
{
"authors": "Shaojie Jiang; Ruqing Zhang; Svitlana Vakulenko; Maarten De Rijke",
"journal": "",
"ref_id": "b21",
"title": "A simple contrastive learning objective for alleviating neural text degeneration",
"year": "2022"
},
{
"authors": "Jared Kaplan; Sam Mccandlish; Tom Henighan; Tom B Brown; Benjamin Chess; Rewon Child; Scott Gray; Alec Radford; Jeffrey Wu; Dario Amodei",
"journal": "",
"ref_id": "b22",
"title": "Scaling laws for neural language models",
"year": "2020"
},
{
"authors": "Marzena Karpinska; Nader Akoury; Mohit Iyyer",
"journal": "Association for Computational Linguistics",
"ref_id": "b23",
"title": "The perils of using Mechanical Turk to evaluate open-ended text generation",
"year": "2021"
},
{
"authors": "Daniel King; Zejiang Shen; Nishant Subramani; Iz Daniel S Weld; Doug Beltagy; Downey",
"journal": "",
"ref_id": "b24",
"title": "Don't say what you don't know: Improving the consistency of abstractive summarization by constraining beam search",
"year": "2022"
},
{
"authors": "Kalpesh Krishna; Yapei Chang; John Wieting; Mohit Iyyer",
"journal": "",
"ref_id": "b25",
"title": "Rankgen: Improving text generation with large ranking models",
"year": "2022"
},
{
"authors": "Wojciech Kryscinski; Bryan Mccann; Caiming Xiong; Richard Socher",
"journal": "Association for Computational Linguistics",
"ref_id": "b26",
"title": "Evaluating the factual consistency of abstractive text summarization",
"year": "2020"
},
{
"authors": "Wojciech Kryściński; Nazneen Rajani; Divyansh Agarwal; Caiming Xiong; Dragomir Radev",
"journal": "",
"ref_id": "b27",
"title": "BookSum: A collection of datasets for long-form narrative summarization",
"year": "2021"
},
{
"authors": "Sachin Kumar; Biswajit Paria; Yulia Tsvetkov",
"journal": "",
"ref_id": "b28",
"title": "Gradient-based constrained sampling from language models",
"year": "2022"
},
{
"authors": "Faisal Ladhak; Esin Durmus; He He; Claire Cardie; Kathleen Mckeown",
"journal": "Association for Computational Linguistics",
"ref_id": "b29",
"title": "Faithful or extractive? on mitigating the faithfulness-abstractiveness tradeoff in abstractive summarization",
"year": "2022"
},
{
"authors": "Faisal Ladhak; Esin Durmus; Mirac Suzgun; Tianyi Zhang; Dan Jurafsky; Kathleen Mckeown; Tatsunori B Hashimoto",
"journal": "",
"ref_id": "b30",
"title": "When do pre-training biases propagate to downstream tasks? a case study in text summarization",
"year": "2023"
},
{
"authors": "Mike Lewis; Yinhan Liu; Naman Goyal; Marjan Ghazvininejad; Abdelrahman Mohamed; Omer Levy; Veselin Stoyanov; Luke Zettlemoyer",
"journal": "Association for Computational Linguistics",
"ref_id": "b31",
"title": "BART: Denoising sequence-to-sequence pretraining for natural language generation, translation, and comprehension",
"year": "2020"
},
{
"authors": "Haoran Li; Song Xu; Peng Yuan; Yujia Wang; Youzheng Wu; Xiaodong He; Bowen Zhou",
"journal": "",
"ref_id": "b32",
"title": "Learn to copy from the copying history: Correlational copy network for abstractive summarization",
"year": "2021"
},
{
"authors": "Wei Li; Wenhao Wu; Moye Chen; Jiachen Liu; Xinyan Xiao; Hua Wu",
"journal": "",
"ref_id": "b33",
"title": "Faithfulness in natural language generation: A systematic survey of analysis, evaluation and optimization methods",
"year": "2022"
},
{
"authors": "Chin-Yew Lin",
"journal": "Association for Computational Linguistics",
"ref_id": "b34",
"title": "ROUGE: A package for automatic evaluation of summaries",
"year": "2004"
},
{
"authors": "Ximing Lu; Peter West; Rowan Zellers; Le Ronan; Chandra Bras; Yejin Bhagavatula; Choi",
"journal": "Association for Computational Linguistics",
"ref_id": "b35",
"title": "Neuro-Logic decoding: (un)supervised neural text generation with predicate logic constraints",
"year": "2021"
},
{
"authors": "Yeyun Xinbei ; Ma; Pengcheng Gong; Hai He; Nan Zhao; Duan",
"journal": "",
"ref_id": "b36",
"title": "Prom: A phrase-level copying mechanism with pre-training for abstractive summarization",
"year": "2023"
},
{
"authors": "Tao Meng; Sidi Lu; Nanyun Peng; Kai-Wei Chang",
"journal": "",
"ref_id": "b37",
"title": "Controllable text generation with neurally-decomposed oracle",
"year": "2022"
},
{
"authors": "Stephen Merity; Caiming Xiong; James Bradbury; Richard Socher",
"journal": "",
"ref_id": "b38",
"title": "Pointer sentinel mixture models",
"year": "2017-04-24"
},
{
"authors": "Tomás Mikolov; Ilya Sutskever; Kai Chen; Gregory S Corrado; Jeffrey Dean",
"journal": "",
"ref_id": "b39",
"title": "Distributed representations of words and phrases and their compositionality",
"year": "2013-12-05"
},
{
"authors": "Fatemehsadat Mireshghallah; Kartik Goyal; Taylor Berg-Kirkpatrick",
"journal": "Association for Computational Linguistics",
"ref_id": "b40",
"title": "Mix and match: Learningfree controllable text generationusing energy language models",
"year": "2022"
},
{
"authors": "Shashi Narayan; Shay B Cohen; Mirella Lapata",
"journal": "Association for Computational Linguistics",
"ref_id": "b41",
"title": "Don't give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization",
"year": "2018"
},
{
"authors": "Pinelopi Papalampidi; Kris Cao; Tomas Kocisky",
"journal": "",
"ref_id": "b42",
"title": "Towards coherent and consistent use of entities in narrative generation",
"year": "2022"
},
{
"authors": "Krishna Pillutla; Swabha Swayamdipta; Rowan Zellers; John Thickstun; Sean Welleck; Yejin Choi; Zaid Harchaoui",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b43",
"title": "MAUVE: Measuring the gap between neural text and human text using divergence frontiers",
"year": "2021"
},
{
"authors": "Alec Radford; Jeffrey Wu; Rewon Child; David Luan; Dario Amodei; Ilya Sutskever",
"journal": "",
"ref_id": "b44",
"title": "Language models are unsupervised multitask learners",
"year": "2019"
},
{
"authors": "Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu",
"journal": "Journal of Machine Learning Research",
"ref_id": "b45",
"title": "Exploring the limits of transfer learning with a unified text-totext transformer",
"year": "2020"
},
{
"authors": "Shafiq Mathieu Ravaut; Nancy Joty; Chen",
"journal": "Association for Computational Linguistics",
"ref_id": "b46",
"title": "SummaReranker: A multi-task mixture-of-experts re-ranking framework for abstractive summarization",
"year": "2022"
},
{
"authors": "Abigail See; Peter J Liu; Christopher D Manning",
"journal": "Association for Computational Linguistics",
"ref_id": "b47",
"title": "Get to the point: Summarization with pointergenerator networks",
"year": "2017"
},
{
"authors": "Abigail See; Aneesh Pappu; Rohun Saxena; Akhila Yerukola; Christopher D Manning",
"journal": "Association for Computational Linguistics",
"ref_id": "b48",
"title": "Do massively pretrained language models make better storytellers?",
"year": "2019"
},
{
"authors": "Kurt Shuster; Jack Urbanek; Arthur Szlam; Jason Weston",
"journal": "Association for Computational Linguistics",
"ref_id": "b49",
"title": "Am I me or you? state-of-the-art dialogue models cannot maintain an identity",
"year": "2022"
},
{
"authors": "Vered Shwartz; Rachel Rudinger; Oyvind Tafjord",
"journal": "Association for Computational Linguistics",
"ref_id": "b50",
"title": "you are grounded!\": Latent name artifacts in pre-trained language models",
"year": "2020"
},
{
"authors": "Yixuan Su; Tian Lan; Yan Wang; Dani Yogatama; Lingpeng Kong; Nigel Collier",
"journal": "",
"ref_id": "b51",
"title": "A contrastive framework for neural text generation",
"year": "2022"
},
{
"authors": "Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Lukasz Kaiser; Illia Polosukhin",
"journal": "",
"ref_id": "b52",
"title": "Attention is all you need",
"year": "2017-09"
},
{
"authors": "C Lawrence Ramakrishna Vedantam; Devi Zitnick; Parikh",
"journal": "IEEE Computer Society",
"ref_id": "b53",
"title": "Cider: Consensus-based image description evaluation",
"year": "2015-06-07"
},
{
"authors": "David Wan; Mohit Bansal",
"journal": "",
"ref_id": "b54",
"title": "Factpegasus: Factuality-aware pre-training and fine-tuning for abstractive summarization",
"year": "2022"
},
{
"authors": "David Wan; Shiyue Zhang; Mohit Bansal",
"journal": "",
"ref_id": "b55",
"title": "Histalign: Improving context dependency in language generation by aligning with history",
"year": "2023"
},
{
"authors": "Lingxiao Wang; Jing Huang; Kevin Huang; Ziniu Hu; Guangtao Wang; Quanquan Gu",
"journal": "",
"ref_id": "b56",
"title": "Improving neural language generation with spectrum control",
"year": "2020-04-26"
},
{
"authors": "Sean Welleck; Ilia Kulikov; Stephen Roller; Emily Dinan; Kyunghyun Cho; Jason Weston",
"journal": "",
"ref_id": "b57",
"title": "Neural text generation with unlikelihood training",
"year": "2020-04-26"
},
{
"authors": "Sean Welleck; Jiacheng Liu; Ximing Lu; Hannaneh Hajishirzi; Yejin Choi",
"journal": "",
"ref_id": "b58",
"title": "Naturalprover: Grounded mathematical proof generation with language models",
"year": "2022"
},
{
"authors": "Thomas Wolf; Lysandre Debut; Victor Sanh; Julien Chaumond; Clement Delangue; Anthony Moi; Pierric Cistac; Tim Rault; Remi Louf; Morgan Funtowicz; Joe Davison; Sam Shleifer; Clara Patrick Von Platen; Yacine Ma; Julien Jernite; Canwen Plu; Teven Xu; Sylvain Le Scao; Mariama Gugger; Quentin Drame; Alexander Lhoest; Rush",
"journal": "",
"ref_id": "b59",
"title": "Transformers: State-of-the-art natural language processing",
"year": "2020"
},
{
"authors": "Zhilin Yang; Zihang Dai; Ruslan Salakhutdinov; William W Cohen",
"journal": "",
"ref_id": "b60",
"title": "Breaking the softmax bottleneck: A high-rank RNN language model",
"year": "2018-04-30"
},
{
"authors": "Haopeng Zhang; Semih Yavuz; Wojciech Kryscinski; Kazuma Hashimoto; Yingbo Zhou",
"journal": "Association for Computational Linguistics",
"ref_id": "b61",
"title": "Improving the faithfulness of abstractive summarization via entity coverage control",
"year": "2022"
},
{
"authors": "Zexuan Zhong; Tao Lei; Danqi Chen",
"journal": "Association for Computational Linguistics",
"ref_id": "b62",
"title": "Training language models with memory augmentation",
"year": "2022-12-07"
}
] | [
{
"formula_coordinates": [
3,
98.94,
305.41,
145.85,
18.06
],
"formula_id": "formula_0",
"formula_text": "P M (x|c t ) = exp(Logit(x, c t ))"
},
{
"formula_coordinates": [
3,
316.45,
361.48,
207.96,
15.02
],
"formula_id": "formula_2",
"formula_text": "q ct = h M ct ⊕ GELU L h (⊕ i,m h M -m c t-i ) ,(2)"
},
{
"formula_coordinates": [
3,
471.86,
426.15,
52.05,
14.47
],
"formula_id": "formula_3",
"formula_text": "⊕ i,m h M -m c t-i"
},
{
"formula_coordinates": [
3,
335.83,
732.87,
188.58,
43.63
],
"formula_id": "formula_4",
"formula_text": "t Logit(x, c t ) = f T ct,S e x if x ∈ S f T ct,V w x O/W ,(3)"
},
{
"formula_coordinates": [
4,
252.34,
194.84,
41.13,
25.83
],
"formula_id": "formula_5",
"formula_text": "L f LD L f C"
},
{
"formula_coordinates": [
4,
84.79,
672,
204.34,
35.52
],
"formula_id": "formula_6",
"formula_text": "e x = f x,ct,LD = t i=1 1 c i t =x L f LD (q c i t ) t i=1 1 c i t =x ,(4)"
},
{
"formula_coordinates": [
4,
184.9,
733.35,
92.98,
16.34
],
"formula_id": "formula_7",
"formula_text": "1 c i t =x = 1 if c i t = x."
},
{
"formula_coordinates": [
5,
79.39,
568.46,
209.74,
56.98
],
"formula_id": "formula_8",
"formula_text": " f T ct,C w x + f T ct,P D f x,ct,LD if x ∈ c t f T ct,R1 w x if x ∈ W (k 1 ) -c t f T ct,R2 w x if x ∈ W (k 2 ) -W (k 1 ) -c t f T ct,V w x O/W ,(5)"
},
{
"formula_coordinates": [
5,
311.98,
327.04,
212.43,
56.98
],
"formula_id": "formula_9",
"formula_text": " f T ct,C w x + f T ct,P D f x,ct,LD if x ∈ c t f T ct,E w x + f T ct,P E f x,I,LE if x ∈ I -c t f T ct,R1 w x if x ∈ W (k 1 ) -c t -I f T ct,V w x O/W ,(6)"
},
{
"formula_coordinates": [
18,
94.68,
713.67,
194.46,
28.45
],
"formula_id": "formula_10",
"formula_text": "f T ct,V w x + f T ct,P D f x,ct,LD if x ∈ c t f T ct,V w x O/W .(7)"
},
{
"formula_coordinates": [
19,
97.9,
326.43,
191.23,
70.27
],
"formula_id": "formula_11",
"formula_text": "P rob(x|I, c t ) = p gen exp f T ct,V w x Z V +(1 -p gen ) |I| j=1 1 I j =x P E (j|I, c t ),(9)"
},
{
"formula_coordinates": [
19,
82.6,
624.52,
206.53,
56.91
],
"formula_id": "formula_12",
"formula_text": "P rob(x|I, ct) = g exp f T c t ,V wx ZV + |I| j=1 1 I j =x exp f T c t ,P E tanh(L f LE (h M I j )) + b Zp ,(10)"
},
{
"formula_coordinates": [
19,
321.85,
450.99,
202.56,
28.45
],
"formula_id": "formula_13",
"formula_text": "f T ct,V w x + f T ct,R f x,ct,LD if x ∈ W (k) f T ct,V w x O/W .(11)"
}
] |
|
10.3115/v1/D14-1082 | 2023-05-21 | [
{
"figure_ref": [
"fig_1"
],
"heading": "Introduction",
"publication_ref": [
"b10",
"b17",
"b8",
"b17",
"b10",
"b2",
"b3",
"b9",
"b5"
],
"table_ref": [],
"text": "Fine-grained entity typing (FET) is an important task in text analysis. Assigning fine-grained semantic types to parsed entity mention spans based on the local context enables effective and structured analysis of unstructured text data, such as entity linking (Ling and Weld, 2012;Onoe and Durrett, 2020), relation extraction (Koch et al., 2014), and coreference resolution (Onoe and Durrett, 2020). Example 1. Given a sentence S 1 : \"Sammy Sosa got a standing ovation at Wrigley Field.\" and a parsed entity mention span \"Sammy Sosa\" in the sentence, an FET method aims to assign it not only the coarse-grained type \"Person\" but also the finegrained types \"Athlete\" or \"Player\".\nFET on large text corpora is challenging due to (1) the high cost to obtain a large amount of humanannotated training data, especially in dynamic and emerging domains, and (2) inaccurate annotations due to (i) different annotators marking concepts at different granularity (e.g., person vs. politician vs. president), and (ii) contextual subtlety on finegrained types (e.g., Boston vs. Detroit could be two sport teams, instead of two cities). Most existing methods utilize weak or distant supervision to automatically generate training data for the FET tasks. There are three major approaches to obtaining weakly-labeled training data to tackle these challenges. The first is to automatically match the mentions in text with the concepts in some existing knowledge bases (e.g., Wikipedia) (Ling and Weld, 2012). The typical workflow is to first detect entity mentions from a corpus, map these mentions to knowledge base (KB) entities of target types, and then leverage the confidently mapped types as labeled data to infer the final type. The second is to directly obtain the head words of nominal mentions as its fine-grained type (Choi et al., 2018). It leverages the head words of the entity mention to consolidate context-aware types derived from the KB matching. However, both approaches suffer from the label sparsity and context-agnostic problems, resulting in the inability to generate high-quality training data for FET.\nThe third approach is to probe the pre-trained language models through the use of masked patterns and entailment templates. Leveraging masked language model (MLM) prompting to generate rich and context-aware weak supervisions for finegrained entity typing is a recent trend, aiming to reduce expensive human labor (Dai et al., 2021;Li et al., 2022).For example, given a sentence that con-tains a mention, a short piece of text that contains a \"[MASK]\" token is added to generate candidate entity types. This method conducts labeling in a context-aware manner and greatly enriches the fine-grained types labeled for each mention. However, such a process can still generate tokens unsuitable for typing (e.g., {Teams, Thing} for \"Wrigley Field\" in S 1 ) or a mixture of rough and fine-grained types (e.g., {Location, Building, Stadium}). Due to the lack of further hierarchical knowledge of the generated tokens/types, such problems cannot be resolved automatically.\nIn this study, we vision that an ontology structure, which provides a semantics-rich, hierarchical structure, may help select the best results generated by multiple PLM models. We propose an zero-shot, ontology-guided, fine-grained entity typing (FET) method, ONTOTYPE, that leverages an input ontology and the power of MLM prompting and Natural Language Inference (NLI). We first ensemble multiple Hearst patterns to perform MLM prompting, reducing the noise in the initial candidate type generation. Since the generated candidate labels for a given mention are likely a mixture of fine and coarse-grained labels, or tokens unsuitable for typing, we propose to automatically match the generated candidate labels to a coarse-grained type in our type ontology and then rank and select a coarsegrained type with a pre-trained entailment model under the local context. Such a type resolution process will progress deeper to finer levels, based on the same principle of entailment score-based type selection, following the type ontology, until the finest possible label can be consolidated.\nExample 2. For sentence S 1 in Ex. 1, candidate type generation (Step 1) ensembles prompting results of multiple Hearst patterns and generates a set of candidate labels {\"Stadium\", \"Venues\", \"Locations\", \"Games\"} for \"Wrigley Field\" (Fig. 2). Using a given ontology structure (Fig. 1), this set of types is resolved to the course-grained type \"Location\" by leveraging a pre-trained entailment model and the local context (Fig. 3). Note that without the ontology type structure, \"Stadium\", \"Venues\", and \"Locations\" are rivals; but with the structure, \"Stadium\" and \"Venues\" are fine-grained \"Locations\" and they support each other at multilevel resolution. By the same principle, the type resolution proceeds deeper to finer-grained levels, along the type ontology, from \"Location\" to \"Building\" and further down to \"Stadium\" for \"Wrigley Field\", leading to the most accurate fine-grained type (Fig. 4). Our contributions are summarized as follows.\n1. A fully zero-shot, ontology-guided, fine-grained typing method, ONTOTYPE, is proposed 2. ONTOTYPE improves fine-grained entity typing (FET) by leveraging candidate labels generated and refined with three information sources: (i) pre-trained language models, (ii) a fine-grained type ontology, and (iii) head words 3. Experiments on the Ontonotes, FIGER, and NYT datasets (Gillick et al., 2014) using their associated ontological structures show that ON-TOTYPE clearly outperforms existing zero-shot named entity typing methods and even rivals supervised methods. Our error analysis shows that refinement of ontology structures will further improve fine-grained entity typing."
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b23",
"b10",
"b6",
"b2",
"b10",
"b24",
"b2",
"b3",
"b22",
"b25",
"b28",
"b19",
"b13",
"b4",
"b3",
"b11",
"b9",
"b3",
"b9",
"b21"
],
"table_ref": [],
"text": "Supervised FET. Fine-grained entity typing benefits various downstream tasks and has received extensive attention in natural language research. Recent studies focus on different contexts from the phrase level (Yao et al., 2013) to considering specific entity mentions in the sentence or document level (Ling and Weld, 2012;Gupta et al., 2017;Choi et al., 2018). Entity typing has been generally studied under supervised learning settings (Ling and Weld, 2012;Yosef et al., 2012;Choi et al., 2018;Dai et al., 2021). These studies require annotated data to achieve high performance and lack the flexibility to identify newly defined entities.\nWeakly supervised or zero-shot FET. To handle such difficulties, the zero-shot learning setting has been introduced for named entity typing (Xia et al., 2018). Several studies (Yuan and Downey, 2018;Zhou et al., 2018) address the problems by grounding the mentions with Wikipedia entries from the assembled related pages. These methods achieve good performance but also require a lot of resources. Other studies explore the pre-trained semantic word-level embeddings (Ren et al., 2016;Ma et al., 2016) or extract raw embeddings without auxiliary information and utilize end-to-end neural networks (Zhang et al., 2020b). However, they still suffer from low accuracy and inefficiency in zeroshot settings. As a result, ONTOTYPE turns to the weak supervision from pre-trained language models, such as BERT (Devlin et al., 2019), that have substantial knowledge in language understanding.\nLeveraging pre-trained language models. Some recent studies (e.g., (Zhang et al., 2020a;Dai et al., 2021;Liu et al., 2021;Li et al., 2022)) leverage pre-trained language models and prompting templates to obtain knowledge for entity mentions in given sentences. (Dai et al., 2021) improves ultrafine entity typing with BERT Masked Language Model (MLM). However, they use a single prompt for each mention to generate labels for weak supervision, which may generate erroneous types. (Li et al., 2022) improves ultra-fine entity typing by treating the task of predicting an entity type as an NLI task. However, they ignore rich semantic information contained in a type ontology during the typing process. Our method ensembles multiple MLM prompting and NLI results to reduce noises in candidate type generation. Further, we utilize the fine-grained type ontology structure as guidance to progressively resolve candidate labels from coarse to fine under the local context. ChemNER (Wang et al., 2021) leverages a type ontology structure to guide fine-grained Chemistry entity recognition. It relies on an existing knowledge base whereas this study leverages the pre-trained language models as rich and context-aware weak supervision."
},
{
"figure_ref": [
"fig_1"
],
"heading": "Methodology",
"publication_ref": [
"b12"
],
"table_ref": [],
"text": "We propose ONTOTYPE, a zero-shot, ontologyguided, fine-grained entity typing method using pre-trained language models and a fine-grained ontology structure. Given an input sentence and a set of pre-identified mentions in the sentence, ONTO-TYPE consists of the following steps: (1) generating a set of candidate labels for each input mention with both head word parsing and an ensemble of MLM prompting (Fig. 2); (2) resolving the coarsegrained types by matching and ranking the generated labels to the type ontology using an entailment model (Fig. 3); and (3) progressively refining the fine-grained types along the type ontology following the principle of entailment score-based type selection (Fig. 4). We utilize the inherent structure of the fine-grained type ontology and a pre-trained entailment model (RoBERTa model pre-trained on the MNLI dataset (Liu et al., 2019)) to guide our fine-grained entity typing."
},
{
"figure_ref": [],
"heading": "Problem Definition",
"publication_ref": [],
"table_ref": [],
"text": "The input to our proposed ONTOTYPE framework is a text corpus D and input fine-grained type ontology O. In this study, we assume our input text corpus D includes a set of pre-identified entity mentions. An entity mention, e, is a token span in the text document that refers to a real-world entity. Given a sentence S and a parsed entity mention e ∈ S, the fine-grained entity typing (FET) task is to identify one or more types t from the label space T (provided in a structured ontology G) for the entity mention e.\nAs an example of our FET task, in the sentence S 1 of Example 1, the entity mention to be typed is e 1 : \"Wrigley Field\". It should be labeled progressively deeper as \"Location→Building→Stadium\" as opposed to other labels like \"Organization\", \"Person\", or \"School\".\nFine-Grained Type Ontology Structure. The structure of the type ontology is fundamental to the ONTOTYPE algorithm. In this work, we leverage the fine-grained type ontologies provided in the OntoNotes and FIGER datasets. These ontologies are structured as forests or a disjoint union of trees. Each tree consists of a type hierarchy stemming from a handful of \"root\" coarse-grained types that include but are not limited to \"Organization\", \"Person\" and \"Location\". Each tree imposes the following structural conditions: (1) Each type has a singular parent type; (2) each type can have an infinite (potentially) number of children; and (3) each type is connected by a directional edge indicating a hypernym-hyponym type relationship between the parent and child nodes. In our experiments, we find that general types such as \"things\" lead to inaccurate labels as pre-trained entailment models often rank vague types higher than specific types. Note that in the given OntoNotes type ontology (Fig. 1), Building and Country are sibling types since they share the same parent type Location. Such a hypernym-hyponym or \"is-a\" relationship is critical for guiding fine-grained type selection."
},
{
"figure_ref": [
"fig_1",
"fig_1"
],
"heading": "Candidate Type Generation",
"publication_ref": [
"b2",
"b3",
"b2",
"b1",
"b4",
"b7",
"b20"
],
"table_ref": [
"tab_0"
],
"text": "To generate a set of candidate labels for each input mention, we leverage two main information sources: (1) head word parsing, and (2) an ensemble of MLM prompting. Head words and hypernyms can serve as powerful context-aware type indicators that can be leveraged as weak supervision sources for entity typing (Choi et al., 2018;Dai et al., 2021). Our method adopts both information sources to first select a candidate label to guide our fine-grained type refinement. First, we generate and parse the head words for the given entity mention. Second, we generate candidate entity types with the use of ensembled MLM prompting1 . The generated candidate labels are used as input to the following steps of ontology-guided type resolution.\nHead Word Parsing. As discussed by (Choi et al., 2018), the input text often contains cues that explicitly match a mention to its type, in the form of the mention's head word. Thus, given the pre-identified entity mentions in the input sentence, we first identify the head word of the input mention. We utilize the Stanford Dependency Parser (Chen and Manning, 2014) to extract the head word of the entity that we are interested in typing. Consider the sentence \"Governor Arnold Schwarzenegger gives a speech at Mission Serve's service project on Veterans Day 2010\" with \"Governor Arnold Schwarzenegger\" as an entity mention. We can describe \"Arnold Schwarzenegger\" with numerous types including actor, father, and governor. However with the use of the head word \"Governor\", we consolidate the fine-grained type of governor given this input sentence. Thus, ONTOTYPE leverages the head words of the input entity, if any, to select an initial context-sensitive coarse-grained type and guide the selection of the final fine-grained type.\nEnsembled MLM Prompting. While head words can provide strong type indicators, they do not always provide sufficient information to consolidate a high-level type. In some cases, head words cannot directly provide accurate type information as they are not directly present in the input type ontology. Furthermore, head words are not always available in the input sentences. Thus, with the parsed entity mention span as input, we propose to leverage context-aware hypernyms as initial type labels for the target mention. With an ensembled cloze prompting method, ONTOTYPE generates candidate labels of the mention and performs an initial high-level typing on the input mention. Specifically, we leverage a BERT masked language model (Devlin et al., 2019) and artificial Hearst patterns (Hearst, 1992) to generate context-aware hypernyms that serve as labels for the mentions. We first modify the input sentence by inserting a Hearst pattern and [MASK] token into the sentence. Then we use the BERT MLM to generate candidate labels for the target mention under the local context. For example, in Fig. 2, we first insert Hearst patterns such as \"and the other [MASK]\" in the input sentence and then use the BERT MLM model to generate candidate labels such as \"Venues\", \"Teams\", and \"Stadiums\".\nWe evaluated the quality of hypernyms generated with the 44 patterns proposed in (Seitner et al., 2016) on the Ontonotes Development Dataset. When generating hypernyms, we expect high-quality candidates to be semantically equivalent to concepts contained in our input type ontology. Overall, we found that the four patterns in Table 1 provided the highest quality hypernyms under a simple direct matching to types in the OntoNotes ontology. However, based on the syntax and grammar of the sentence, hearst patterns can generate tokens that are unsuitable to serve as fine-grained entity types. For example, with a single prompt, the MLM can generate \"Famous\", \"Actors\", \"Celebrities\" and \"People\" as the most probable words. Note that \"Famous\" is unsuitable to serve as a fine-grained entity type. To reduce the noises caused by the use of a single Hearst pattern, we ensemble n Hearst Patterns to consolidate the most commonly generated candidate labels. For each pattern in the pattern list P , we collect the top k most probable tokens from the probability distribution predicted by the BERT MLM. Then, we aggregate the tokens and identify the set of labels that have the largest overlap. We perform the voting ensemble as follows:\ny = count (m) {H 1 (x), H 2 (x), ..., H n (x)} (1)\nwhere H n (x) is the candidate type generated by the n th hearst pattern and count (m) is the function that selects all candidate labels generated at least m times. In our experiments, we use m = n 2 + 1. In addition, we find that the number of ensembled patterns is not sensitive as ensembling ensures we retain the most confident candidate labels. Example 3: As shown in Fig. 2, by ensembling the results from prompting with several Hearst patterns, the quality types for e 1 \"Stadiums, Venues, Locations, Games\" retain but the noisy types \"Things\" and \"Teams\" are removed."
},
{
"figure_ref": [],
"heading": "High-Level Type Resolution",
"publication_ref": [
"b14"
],
"table_ref": [],
"text": "Given the generated candidate labels and head words for each mention in the sentence, we seek to resolve the concrete type for this candidate type set at the high levels of the type ontology. Specifically, we first align the generated candidate labels to several high-level types in the type ontology, and then select the most accurate high-level types with a pre-trained entailment language model.\nCandidate Type Alignment. Following the previous step of candidate type generation, we combine the generated candidate labels from both the parsed head words and the ensembled cloze prompting to form a candidate type set. These labels are generally noisy and may not exist directly in our type ontology. However, we observe that the generated candidates will usually cluster closely around a high-level concept in the ontology. Thus, we perform our high-level type alignment with a cosinesimilarity-based word matching. We use Word2Vec 2 (Mikolov et al., 2013) embeddings to construct our type embeddings for the cosine-similarity-based type alignment. We construct a verbalizer by selecting at least l seman-2 https://code.google.com/archive/p/word2vec/ tically related words for each coarse type. For a high-level type of \"Organization\", we might include seed types such as \"Corporation\", \"University\", \"Firm\", \"Business\", and \"Government\" in its verbalizer. This verbalizer is then utilized to systematically map the MLM hypernym prediction to the most relevant type. In our experiments, we provide at least five seed types S for each type node c in the first level of the type ontology. Increasing the number of seed types increases the coverage and confidence of the verbalizer. Since the most commonly generated hypernyms for each concept exists in the input ontologies, ONTOTYPE is not so sensitive to the number of seed types collected. With the seed types, we construct a node embedding N by taking the average of word embeddings from both the first-level type and its corresponding seed types provided by human,\nN = s i ∈S emb(s i ) + emb(c) |S| + 1 . (2\n)\nwhere emb(•) indicates the Word2Vec embeddings. Then we rank each generated candidate type to a first-level type on our type ontology by calculating the cosine similarity between the embeddings of the generated candidate labels b and that of the first-level types T , score(b, T ) rank = cosine(emb(b), emb(T )).\n(3)\nFinally, the first-level type that has the highest similarity is selected as the aligned high-level type."
},
{
"figure_ref": [
"fig_1"
],
"heading": "⋮ Figure 3: Candidate Type Generation",
"publication_ref": [
"b12"
],
"table_ref": [],
"text": "Example 4: In Figures 2 and3, the consolidated candidate labels \"Locations\", \"Stadiums\", \"Venues\", and \"Games\" are closely related to the high-level type \"location\" in our ontology. By performing the cosine-similarity-based matching, \"Location\" is identified as e 1 's high-level type over \"Organization\" or \"Person\". Thus, we finally select the most similar high-level type given by the head word, generated candidate labels, and the entailment model.\nHigh-Level Type Selection After the previous step of candidate type alignment, we obtain several high-level types for each entity mention in the sentence. Given these types, we seek to select the most accurate high-level type for each entity mention under the local context. We observe that the task of selecting the most appropriate entity type can be viewed as a Natural Language Inference (NLI) task. Thus, we treat the input sentence as the premise in NLI and generate the hypothesis using a pre-defined declarative template. To resolve the type of the input mention we utilize the following template: \"In this sentence, [MENTION] is a [TYPE]\". ONTOTYPE then ranks each type in the first level of the ontology with the entailment score from a RoBERTa NLI model (Liu et al., 2019).\nExample 5: In Fig. 3, we have aligned a majority of the generated candidate type set to the Location seed types. The NLI model further ranks Location over Organization or Person. By utilizing the information in conjunction, e 1 is solidly aligned to the high-level type \"Location\"."
},
{
"figure_ref": [],
"heading": "Fine-Grained Type Resolution",
"publication_ref": [],
"table_ref": [],
"text": "Given the high-level types of the entity mentions, ONTOTYPE further leverages the ontology structure to progressively resolve the fine-grained label.\nFollowing the same principle of entailment-based type selection for the high-level types, we utilize the entailment model to automatically select the most accurate fine-grained entity types along our type ontology. Specifically, we first examine the child types of the previously determined higherlevel types and then select the child type with the highest ranked score as the fine-grained type.\nIn addition to the entailment model, we also utilize the candidate type set to refine our fine-grained type selection. If the parsed head word is present in our type ontology, the entailment scores of that parent and its child types are weighted higher. Similarly, if the generated candidate labels are in our type ontology, the entailment scores of those parents and their child types are also weighted higher. Finally, we select the fine-grained type for each mention with the highest weighted entailment score.\nWe calculate the ranking score for the entities at the current level of the ontology as follows:\nrank(type) = σ entail + σ cand(4)\nWe first leverage our NLI pre-trained model to find the entailment score σ entail for each entity type. Then if a type in the candidate type set is a descendent to the entity type, we add a weight σ cand . We repeat this entailment-based selection process recursively along the type ontology to select the best fine-grained type.\nDefinition (Entity Type Granularity Parameter θ):\nWe assume there is a scalar of θ indicating how granular or specific the final selected entity type label should be. The smaller θ is, the more granular entity type labels are consolidated as the final label. Thus, if the child types do not have a sufficient gain of at least θ in ranking score over the parent type at a certain level, we stop the recursion and select the parent type as the final fine-grained type. We conduct a parameter study to explore the sensitivity of ONTOTYPE to the parameter θ in Appendix A.3\n⋮ Figure 4: Fine-Grained Type Refinement\nExample 6: In Fig. 4, we consider all descendent types of \"location\" as potential fine-grained entity types for e 1 . To begin, ONTOTYPE generates the hypotheses and ranks all child types of \"location\". With the resultant entailment score rankings of these sibling types, we consolidate and select the \"building\" type as the highest-ranked label. Then, a similar process is done at a deeper level of the ontology to select the final type of \"stadium\"."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Datasets",
"publication_ref": [
"b5",
"b10"
],
"table_ref": [],
"text": "We compare the performance of ONTOTYPE and several baseline methods on three benchmark FET datasets: NYT, Ontonotes (Gillick et al., 2014) and FIGER (Ling and Weld, 2012). The basic statistics of the datasets are shown in Table 3: Results (%) on Three Test Sets (Some slots in supervised methods marked \"-\" due to no fully annotated training data).\nOntoNotes and FIGER datasets, we leverage the included type ontologies while the NYT dataset leverages the input FIGER ontology. All NER test sets are annotated using the ontologies as a set of type labels. Thus, each entity mention is labeled with a fine-grained label represented as a path within the ontology. The ontologies have a maximum depth of three and contain four to six high-level types (e.g., Location, Person, and Organization)."
},
{
"figure_ref": [],
"heading": "Baselines",
"publication_ref": [],
"table_ref": [],
"text": "ONTOTYPE is a zero-shot FET method that does not require human annotation as supervision. We compare ONTOTYPE with two groups of baseline methods for FET: supervised methods and zeroshot methods. Details of the parameter settings for each model is listed in Appendix A.2. We also conduct parameter and ablation studies listed in Appendices A.3 and A.4 respectively."
},
{
"figure_ref": [],
"heading": "Supervised FET Methods",
"publication_ref": [
"b19"
],
"table_ref": [],
"text": "We compare the performance of ONTOTYPE with three supervised FET methods as discussed below.\nAFET (Ren et al., 2016) MZET (Zhang et al., 2020b): This zero-shot baseline leverages the semantic meaning and the hierarchical structure into the type representation. The method leverages the knowledge from seen types to label the zero-shot types through the use of a memory component."
},
{
"figure_ref": [],
"heading": "Main Results",
"publication_ref": [
"b28"
],
"table_ref": [],
"text": "Table 3 strates the benefit of leveraging the knowledge embedded in pre-trained language models as a form of minimal supervision to identify entity mention types. ZOE leverages a type ontology and maps a given mention to type-compatible Wikipedia Entries. As a result, ZOE relies on surface-level information from the mention string. OntoType ensembles contextual information from various PLMs to consolidate the final entity type. Given a sentence: \"The biggest cause for concern for McGuff is the bruised hamstring Regina Rogers suffered against (Utah) last Saturday\", ZOE incorrectly utilizes the surface string to label \"Utah\" as a location. With PLMs, OntoType recognizes teams or opponents as candidates and finally consolidates \"sports_team\". On the FIGER dataset, ONTOTYPE achieves 0.3 absolute Macro-F1 improvement over state-of-the-art zero-shot fine-grained entity typing method ZOE (Zhou et al., 2018). However, our method trails ZOE in strict accuracy and Micro-F1. In the FIGER dataset, predictions are made based on both the surface-level information and the contextual information in the sentence. The major advantage of ONTOTYPE is to accurately capture the contextual information to provide a fine-grained entity type. However, ONTOTYPE does not involve a mechanism to capture the surface-level informa-tion of an entity mention. We discuss this issue further in our error analysis (Appendix A.6)."
},
{
"figure_ref": [],
"heading": "Comparative Case Study",
"publication_ref": [],
"table_ref": [
"tab_6"
],
"text": "Table 4 presents a sentence from a recent news article with tagged mentions and predicted entity types. We find that methods like MZET sometimes predict incompatible types due to incorrect or misleading surface information. For example, when typing \"The White House\", MZET and ZOE leverage the surface-level information from large KBs to identify the mention as a location. However, given the local context, the White House clearly refers to the U.S. government. When considering the mention \"US President Joe Biden\", our method utilizes the type information from the candidate type set (Officials, Leaders, Politicians, Individuals) to explore the \"Person\" branch within the ontology. We search the \"Politician\" branch for the best fine-grained context-aware type to label our given entity mention. Thus, with the assistance of the pre-trained language models, we can incorporate contextual information to derive more context-aware types."
},
{
"figure_ref": [],
"heading": "Error Analysis",
"publication_ref": [],
"table_ref": [],
"text": "The structure of the fine-grained type ontology is important to the performance of ONTOTYPE. Input ontologies should utilize hypernym-type relationships where each parent type is a generalization of the child type. The provided FIGER ontology contains logical inconsistencies in how various types of relations are organized. Specifically, the FIGER ontology considers the parent of the \"Road\" type to be \"Transportation\" rather than \"Location\". This logical inconsistency leads to erroneous typing from our method. We further examine our method's errors and evaluate ONTOTYPE using a modified type ontology in Appendix A.6."
},
{
"figure_ref": [],
"heading": "Large Language Models",
"publication_ref": [
"b18"
],
"table_ref": [],
"text": "Traditional language models (LM) aim at assigning a Probability P (w) to an input sequence of n words s = [w 1 , w 2 , ..., w n ]. Recent works include large language models (LLMs) such as ChatGPT and GPT-4 (OpenAI, 2023), which are pre-trained on with a large text corpora enables them to generate contextualized word representation to achieve state of the art performance on various NLP tasks including language translation, text summarization, and question-answering. These recent works provide a richer knowledge base for ONTOTYPE's candidate type generation. As a result, we can similarly leverage models like T-5 (Raffel et al., 2020) or GPT-3 (Brown et al., 2020) to obtain high-quality candidate types through hypernym prompting methods.\nExample 6: Given a sentence S 1 : \"Sammy Sosa got a standing ovation at Wrigley Field.\" and a parsed entity mention span \"Sammy Sosa\", Chat-GPT generates \"Athletes\", \"Players\", \"Teammates\", \"Stars\" and \"Figures\" while T-5 generates \"Athletes\", \"Players\", \"Teammates\" and \"Individuals\". Overall, we find that even LLMs can leverage the surrounding context and learned \"common-sense\" knowledge to generate high quality candidate types. ONTOTYPE provides a structured approach to provide guidance for LLMs like ChatGPT and GPT-4 which are extremely susceptible to hallucinations. By grounding our types to mention hypernyms and a fine-grained type ontology, we can confidently extract the most fine-grained type for the given entity mention span."
},
{
"figure_ref": [],
"heading": "Conclusions",
"publication_ref": [],
"table_ref": [],
"text": "We propose ONTOTYPE, which leverages weak supervision setting of pre-trained language model prompting. We use head words and MLM cloze prompting for fine-grained candidate label generation. Then we automatically match the generated fine-grained types to our type ontology with an inference method to select the most appropriate finegrained types under the local context. Extensive experiments on real-world datasets show that ON-TOTYPE is highly effective and substantially outperforms the state-of-the-art zero-shot FET methods. In the future, we plan to further refine and enrich a type ontology which will enable us to incorporate more type information for even better performance."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [],
"table_ref": [],
"text": "ONTOTYPE utilizes pre-trained language models to perform context-aware fine-grained entity typing leveraging the rich structural information from type ontologies, head words and cloze prompting. However, this method still has a few limitations. Since the pre-trained language models utilized are not fine-tuned, some candidate labels and fine-grained types generated from the MLM and NLI models are noisy and unsuitable to serve as fine-grained types. One future improvement of ONTOTYPE is to finetune the MLM and NLI models based on the local corpus and type ontology. Furthermore, ONTO-TYPE ignores the surface-level information of the entity mentions. Prior distantly-supervised methods have been able to leverage the surface-level information through entity linking to the knowledge bases. Another future improvement of ONTOTYPE is to incorporate these rich information sources to perform typing on the entity mentions with the additional information from the knowledge bases."
},
{
"figure_ref": [],
"heading": "Ethical Considerations",
"publication_ref": [
"b15"
],
"table_ref": [],
"text": "Fine-grained entity typing is an important task for many downstream applications. We are not aware of any ongoing research on potential biases in finegrained entity typing. However, bias has been reported and explored in the use of Masked Language Models (Nangia et al., 2020), which is leveraged to generate ONTOTYPE's candidate types. In addition, we present a newly annotated NYT dataset which we will release for use in future studies. Finally, we also leverage the existing Ontonotes and FIGER test data sets for baseline comparisons."
},
{
"figure_ref": [],
"heading": "A Appendix",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "A.1 Efficiency & Scalability Study",
"publication_ref": [],
"table_ref": [
"tab_8"
],
"text": "We conduct an efficiency and scalability study to gauge the impact of larger scale data sets on our method and include the results in Table 5. "
},
{
"figure_ref": [],
"heading": "Number of Mentions",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "A.2 Parameter Settings",
"publication_ref": [
"b19",
"b2",
"b3",
"b28",
"b25",
"b16",
"b4",
"b12",
"b14"
],
"table_ref": [],
"text": "For AFET (Ren et al., 2016), UFET (Choi et al., 2018), and MLMET (Dai et al., 2021), we utilize the default parameters as detailed in their studies.\nFor the zero-shot methods, ZOE (Zhou et al., 2018), OTyper (Yuan and Downey, 2018), DZET (Obeidat et al., 2019) and MZET (Zhang et al., 2020b), we leverage the default parameters as detailed in their respective studies. ONTOTYPE leverages a pre-trained BERT (Devlin et al., 2019) (BERT-base, uncased) and pre-trained RoBERTa fine-tuned on the MNLI dataset (Liu et al., 2019) available in the HuggingFace Library. In addition, ONTOTYPE utilizes Word2Vec 3 (Mikolov et al., 2013) embeddings to construct our type embeddings."
},
{
"figure_ref": [],
"heading": "A.3 Study of Sensitivity to Parameters",
"publication_ref": [],
"table_ref": [],
"text": "A potential concern with the experimental setup can be overtly high sensitivity of ONTOTYPE to the Entity Type Specificity parameter θ. For all experiments in Table 3, we leverage the same θ value of 0.3. Additionally, from the plot in Figure 5, we clearly see that F1 scores are not drastically sensitive to theta with standard deviations of 0.3631 and 0.6262 on FIGER and OntoNotes respectively."
},
{
"figure_ref": [],
"heading": "A.4 Ablation Study",
"publication_ref": [
"b19",
"b2",
"b3",
"b9",
"b28",
"b25",
"b16"
],
"table_ref": [],
"text": "We also include the prediction results of our ablation models to demonstrate how the NLI module contributes to the final type assignment. We utilize a simple type mapping to evaluate our ensembled BERT Cloze Prompting module. Figure 6 shows the results of ablation studies on the test set of the Ontonotes dataset. We compared the ONTOTYPE full model with various ablations and variations. We find that direct prompting\\NLI performs the (Ren et al., 2016) 55.3 66.4 69.3 UFET (Choi et al., 2018) ---BERT-MLMET (Dai et al., 2021) ---LITE (Li et al., 2022) 66.2 74.7 80.1\nZero-Shot ZOE (Zhou et al., 2018) 58.8 71.3 74.8 OTyper (Yuan and Downey, 2018) 47.2 67.2 69.1 DZET (Obeidat et al., 2019) 28.5 56.0 55.1 MZET (Zhang et al., 2020b) 31.9 57.9 55.5 ONTOTYPE (Ours) 49.1 67.4 75.1 ONTOTYPE + Modified FIGER Type Ontology (Ours) 51.1 68.9 77.2\nTable 7: Results (%) on the FIGER test set with Provided and Modified Type Ontology ontologies where the parent type is a generalization of the child type. The included FIGER type ontology considers the parent type of the \"Road\" type to be \"Other\\Transportation\" and the parent type of \"Building\" to be \"Other\". While \"Other\" can be considered a valid generalization, it is extremely broad and the coarse-grained type of \"Location\" serves as a stronger parent for both finegrained types. Thus, we modify the included ontology by reorganizing the fine-grained types under parent types that have stronger generalizations (e.g. Building & Road under Location rather than Other). ONTOTYPE achieves 2.1 absolute Ma-F1 improvement given this modified type ontology. Thus, we find that ONTOTYPE can be further improved with correct type ontologies to leverage the inherent hierarchical relationships between coarse and fine-grained types.\nA.6 Analysis of Typing Errors ONTOTYPE, though having high performance, still generates nontrivial errors. We analyze the reasons behind the errors on the Ontonotes and FIGER test sets and categorize them into three types. Note the Gold Type for each mention is highlighted in blue.\nError Type 1: Lack of Knowledge Base Sentence E1: He was a caseworker in Minnesota [\\Location\\Region] but left the job because he found himself perpetually sick from the environments in which he worked. In E1, ONTOTYPE incorrectly types Minnesota as a Country rather than the gold type of Region. In this context, we might consider State to be a more reasonable type to predict. With a simple KB-matching method, we would be able to capture the surface-level information to type it as a state/province. ONTOTYPE relies on neither human annotations nor knowledge base. Obviously, introducing a knowledge base and KB-matching mechanism will further improve its performance.\nError Type 2: Incomplete Fine-Grained Type Ontology Sentence E2: Valley Federal Savings & Loan Association said Imperial Corp. of America withdrew from regulators its application to buy five Valley Federal branches [\\Location\\Structure], leaving the transaction in limbo. In E2, ONTOTYPE generates \\Other. Even when ONTOTYPE is able to generate high-quality candidate labels, it can sometimes fail to align to the correct entity type due to an incomplete type set. In this example, our Candidate Type set generated by prompting consists of Asset, Property, Facility, Bank, and Branches. Clearly, the best fine-grained type should be Asset or Property (rather than the Gold Type: Location or Structure). Since the provided OntoNotes ontology does not include such fine-grained types, ONTOTYPE is unable to generate the correct answer. Clearly, a refined ontology will further improve its performance.\nError Type 3: Incapability to Type Nested Entities Sentence E3: The 33-year-old Billings [\\Loca-tion\\City] native enlisted as a military veterinarian.\nIn E3, ONTOTYPE identifies Billings as \\Person. This mistake can be caused by confusing the type of the whole entity \"Billings native\" with the type of the nested entity \"Billings\". PLM is good at generating candidate types for the whole entity based on its contextual structure, whereas the knowledge about a nested entity like \"Billings\" (in front of head word \"native\") can be more easily derived from a knowledge-base or from some nested entity type patterns. We will leave the issue of typing nested entities to future work. The Kremlin's \\Location\\Building effort to shield the Russian \\Person population from the impact of its invasion of Ukraine \\Location\\Country are wearing thin as the war \\Other heads into its second year.\nTrailing two games to one in the NBA Finals \\Other\\Event and facing the daunting task of trying to beat the Boston Celtics \\Organiza-tion\\Company in the hostile environment of TD Garden \\Location\\Building on Friday night, the Warriors knew they needed to summon one of the best efforts of their dynastic run in order to even the best-of-seven series."
},
{
"figure_ref": [],
"heading": "ZOE",
"publication_ref": [],
"table_ref": [],
"text": "The Kremlin's \\Location\\Building effort to shield the Russian \\Person\\Ethnicity population from the impact of its invasion of Ukraine \\Location\\Country are wearing thin as the war \\Other\\Event heads into its second year.\nTrailing two games to one in the NBA Finals \\Other\\Event and facing the daunting task of trying to beat the Boston Celtics \\Organiza-tion\\Sports_Team in the hostile environment of TD Garden \\Location\\Building\\Sports_Facility on Friday night, the Warriors knew they needed to summon one of the best efforts of their dynastic run in order to even the best-of-seven series. ONTOTYPE The Kremlin's \\Organization\\Government effort to shield the Russian \\Person\\Ethnicity population from the impact of its invasion of Ukraine \\Location\\Country are wearing thin as the war \\Other\\Event\\Conflict heads into its second year.\nTrailing two games to one in the NBA Finals \\Other\\Event\\Finals and facing the daunting task of trying to beat the Boston Celtics \\Organization\\Sports_Team\\Basketball_Team in the hostile environment of TD Garden \\Loca-tion\\Building\\Sports_Facility on Friday night, the Warriors knew they needed to summon one of the best efforts of their dynastic run in order to even the best-of-seven series. "
},
{
"figure_ref": [],
"heading": "A.7 Additional Comparative Case Study",
"publication_ref": [],
"table_ref": [
"tab_10"
],
"text": "Table 8 presents further sentences from two recent news articles with tagged mentions and predicted entity types. In the first sentence, ONTO-TYPE learns the best label for the entity mention span \"the war\" through the use of our NLI typing method. ZOE is able to leverage similar entries in a knowledge base to align the event type correctly. However, the candidate types generated by MLM prompting allow ONTOTYPE to traverse the ontology structure and find the most fine-grained type that fits the given entity mention. As before, MZET and ZOE heavily leverage the surface level information from large KB to identify \"The Kremlin\" as a fortress or physical building in Moscow. However, the local context makes it clear that \"The Kremlin\" refers to the Russian government. With the aid of pre-trained language models, ONTOTYPE is able to leverage the local context to consolidate the most appropriate fine-grained types.\nIn order to mitigate an over-reliance on surfacelevel information, ZOE attempts to match the concepts that best align given the local context as a constraint. This enables ZOE to recognize that TD Garden refers to a stadium or sports facility. When ONTOTYPE types TD Garden, we learn the most fine-grained label through the use of local context and pre-trained language models. The type indicators (Venues, Places, Sites, Locations) allow us to consolidate the high-level type of Location. However, by leveraging the surrounding contextual clues and knowledge learned during the training, NLI models enable ONTOTYPE to establish that TD Garden refers to a building or specifically a sports facility.\nFinally, ONTOTYPE is able to derive the most fine-grained and accurate types by leveraging the inherent structure of the input type ontology. Consider the mention span of \"The NBA Finals\". The predicted candidate labels and head words (Finals, Playoffs, Semifinals, Rounds) indicate that the type is related to an Event. Thus, with the assistance of the pre-trained language models and head words, we select the most context-aware type of \\Other\\Event. Thus with these information sources and the ontology structure as guidance, ONTOTYPE can refine and select the final type of \\Other\\Event\\Finals."
},
{
"figure_ref": [],
"heading": "A.8 Evaluation Metrics",
"publication_ref": [
"b3",
"b10",
"b13"
],
"table_ref": [],
"text": "Following the prior FET studies ( (Dai et al., 2021), (Ling and Weld, 2012), (Ma et al., 2016)), we evaluate our methods and the baselines using three evaluation metrics: Strict Accuracy (Acc), Micro-F1 (Mi-F1), and Macro-F1 . Accuracy. Given a set of entity mentions M , we denote the set of ground truths and predicted types as t M and t M respectively. Given σ as an indicator function, strict accuracy is defined as\nAcc = Σ m∈M σ(t m == t m ) M\nMacro-F1. Macro-F1 is calculated using Macro-Precision (P ma ) and Macro-Recall (R ma ) where\nP ma = 1 |M | Σ m∈M |t m ∩ t m | t m R ma = 1 |M | Σ m∈M |t m ∩ t m | t m\nMicro-F1. Micro-F1 is calculated using Micro-Precision (P mi ) and Micro-Recall (R mi ) where\nP mi = Σ m∈M |t m ∩ t m | Σ m∈M t m R mi = Σ m∈M |t m ∩ t m | Σ m∈M t m\nMacro-F1 and Micro-F1 are calculated using the F1 score formula with their respective granular precision and recall scores. "
}
] | Fine-grained entity typing (FET), which assigns entities in text with context-sensitive, fine-grained semantic types, will play an important role in natural language understanding. A supervised FET method, which typically relies on human-annotated corpora for training, is costly and difficult to scale. Recent studies leverage pre-trained language models (PLMs) to generate rich and context-aware weak supervision for FET. However, a PLM may still generate a mixture of rough and fine-grained types, or tokens unsuitable for typing. In this study, we vision that an ontology provides a semantics-rich, hierarchical structure, which will help select the best results generated by multiple PLM models and head words. Specifically, we propose a novel zero-shot, ontologyguided FET method, ONTOTYPE, which follows a type ontological structure, from coarse to fine, ensembles multiple PLM prompting results to generate a set of type candidates, and refines its type resolution, under the local context with a natural language inference model. Our experiments on the Ontonotes, FIGER, and NYT datasets using their associated ontological structures demonstrate that our method outperforms the state-of-the-art zero-shot fine-grained entity typing methods. Our error analysis shows that refinement of the existing ontology structures will further improve fine-grained entity typing. | ONTOTYPE: Ontology-Guided Zero-Shot Fine-Grained Entity Typing with Weak Supervision from Pre-Trained Language Models | [
{
"figure_caption": "Figure 1: OntoNotes Type Ontology",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: Candidate Type Generation",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Performance of Hearst Patterns with simple type mapping on Ontonotes Dataset.",
"figure_data": "Hearst PatternPrec Rec F1[MASK] such as53.372.4 61.4such [MASK] as47.968.7 56.5and some other [MASK] 48.866.6 56.4and the other [MASK]47.668.3 56.1",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Dataset Statistics",
"figure_data": "",
"figure_id": "tab_2",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "For the",
"figure_data": "SettingsModelNYT Acc Mi-F1 Ma-F1 Acc Mi-F1 Ma-F1 Acc FIGEROntonotes Mi-F1 Ma-F1AFET (Ren et al., 2016)---55.3 66.469.355.164.771.1SupervisedUFET (Choi et al., 2018)------59.571.876.8BERT-MLMET (Dai et al., 2021)------67.44 80.3585.44LITE (Li et al., 2022)---66.2 74.780.168.281.486.6ZOE (Zhou et al., 2018)62.1 73.776.958.8 71.374.850.760.866.9Zero-ShotOTyper (Yuan and Downey, 2018) 46.4 65.767.347.2 67.269.131.836.039.1DZET (Obeidat et al., 2019)27.3 53.151.628.5 56.055.123.128.127.6MZET (Zhang et al., 2020b)30.7 58.256.731.9 57.955.533.743.742.3ONTOTYPE + Original Ontology69.6 78.482.849.1 67.475.165.773.481.5",
"figure_id": "tab_3",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Examples showing how ONTOTYPE and other Zero-Shot FET methods perform on recent news articles with a modified FIGER type ontology. Entity mentions are bolded and predicted labels are in color.",
"figure_data": "",
"figure_id": "tab_6",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "Results of the efficiency & scalability study.",
"figure_data": "",
"figure_id": "tab_8",
"figure_label": "5",
"figure_type": "table"
},
{
"figure_caption": "Examples showing how ONTOTYPE performs on recent news articles with a modified FIGER type ontology. Entity mentions are bolded in the given sentences. The predicted labels are in color.",
"figure_data": "",
"figure_id": "tab_10",
"figure_label": "8",
"figure_type": "table"
}
] | Tanay Komarlu; Minhao Jiang; Xuan Wang; Jiawei Han | [
{
"authors": "B Tom; Benjamin Brown; Nick Mann; Melanie Ryder; Jared Subbiah; Prafulla Kaplan; Arvind Dhariwal; Pranav Neelakantan; Girish Shyam; Amanda Sastry; Sandhini Askell; Ariel Agarwal; Gretchen Herbert-Voss; Tom Krueger; Rewon Henighan; Aditya Child; Daniel M Ramesh; Jeffrey Ziegler; Clemens Wu; Christopher Winter; Mark Hesse; Eric Chen; Mateusz Sigler; Scott Litwin; Benjamin Gray; Jack Chess; Christopher Clark; Sam Berner; Alec Mc-Candlish; Ilya Radford; Dario Sutskever; Amodei",
"journal": "",
"ref_id": "b0",
"title": "Language Models are Few-Shot Learners",
"year": "2020"
},
{
"authors": "Danqi Chen; Christopher Manning",
"journal": "",
"ref_id": "b1",
"title": "A Fast and Accurate Dependency Parser using Neural Networks",
"year": "2014"
},
{
"authors": "Eunsol Choi; Omer Levy; Yejin Choi; Luke Zettlemoyer",
"journal": "Association for Computational Linguistics",
"ref_id": "b2",
"title": "Ultra-Fine Entity Typing",
"year": "2018"
},
{
"authors": "Hongliang Dai; Yangqiu Song; Haixun Wang",
"journal": "",
"ref_id": "b3",
"title": "Ultra-Fine Entity Typing with Weak Supervision from a Masked Language Model",
"year": "2021"
},
{
"authors": "Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova",
"journal": "Association for Computational Linguistics",
"ref_id": "b4",
"title": "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding",
"year": "2019"
},
{
"authors": "Daniel Gillick; Nevena Lazic; Kuzman Ganchev; Jesse Kirchner; David Huynh",
"journal": "",
"ref_id": "b5",
"title": "Context-Dependent Fine-Grained Entity Type Tagging",
"year": "2014"
},
{
"authors": "Nitish Gupta; Sameer Singh; Dan Roth",
"journal": "",
"ref_id": "b6",
"title": "Entity Linking via Joint Encoding of Types, Descriptions, and Context",
"year": "2017"
},
{
"authors": "Marti A Hearst",
"journal": "",
"ref_id": "b7",
"title": "Automatic Acquisition of Hyponyms from Large Text Corpora",
"year": "1992"
},
{
"authors": "Mitchell Koch; John Gilmer; Stephen Soderland; Daniel S Weld",
"journal": "",
"ref_id": "b8",
"title": "Type-aware distantly supervised relation extraction with linked arguments",
"year": "2014"
},
{
"authors": "Bangzheng Li; Wenpeng Yin; Muhao Chen",
"journal": "Transactions of the Association for Computational Linguistics",
"ref_id": "b9",
"title": "Ultra-fine Entity Typing with Indirect Supervision from Natural Language Inference",
"year": "2022"
},
{
"authors": "Xiao Ling; Daniel S Weld",
"journal": "AAAI Press",
"ref_id": "b10",
"title": "Fine-Grained Entity Recognition",
"year": "2012"
},
{
"authors": "Pengfei Liu; Weizhe Yuan; Jinlan Fu; Zhengbao Jiang; Hiroaki Hayashi; Graham Neubig",
"journal": "",
"ref_id": "b11",
"title": "Pretrain, Prompt, and Predict: A Systematic Survey of Prompting Methods in Natural Language Processing",
"year": "2021"
},
{
"authors": "Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov",
"journal": "",
"ref_id": "b12",
"title": "RoBERTa: A Robustly Optimized BERT Pretraining Approach",
"year": "2019"
},
{
"authors": "Yukun Ma; Erik Cambria; Sa Gao",
"journal": "",
"ref_id": "b13",
"title": "Label Embedding for Zero-shot Fine-grained Named Entity Typing",
"year": "2016"
},
{
"authors": "Tomas Mikolov; Ilya Sutskever; Kai Chen; Greg S Corrado; Jeff Dean",
"journal": "Curran Associates, Inc",
"ref_id": "b14",
"title": "Distributed Representations of Words and Phrases and their Compositionality",
"year": "2013"
},
{
"authors": "Nikita Nangia; Clara Vania; Rasika Bhalerao; Samuel R Bowman",
"journal": "",
"ref_id": "b15",
"title": "CrowS-Pairs: A Challenge Dataset for Measuring Social Biases in Masked Language Models",
"year": "1953"
},
{
"authors": "Rasha Obeidat; Xiaoli Fern; Hamed Shahbazi; Prasad Tadepalli",
"journal": "Association for Computational Linguistics",
"ref_id": "b16",
"title": "Description-Based Zeroshot Fine-Grained Entity Typing",
"year": "2019"
},
{
"authors": "Yasumasa Onoe; Greg Durrett",
"journal": "OpenAI",
"ref_id": "b17",
"title": "Interpretable Entity Representations through Large-Scale Typing",
"year": "2020"
},
{
"authors": "Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu",
"journal": "",
"ref_id": "b18",
"title": "Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer",
"year": "2020"
},
{
"authors": "Xiang Ren; Wenqi He; Meng Qu; Lifu Huang; Ji Heng; Jiawei Han",
"journal": "Association for Computational Linguistics",
"ref_id": "b19",
"title": "AFET: Automatic Fine-Grained Entity Typing by Hierarchical Partial-Label Embedding",
"year": "2016"
},
{
"authors": "Julian Seitner; Christian Bizer; Kai Eckert; Stefano Faralli; Robert Meusel; Heiko Paulheim; Simone Paolo; Ponzetto ",
"journal": "",
"ref_id": "b20",
"title": "A Large DataBase of Hypernymy Relations Extracted from the Web",
"year": "2016"
},
{
"authors": "Xuan Wang; Vivian Hu; Xiangchen Song; Shweta Garg; Jinfeng Xiao; Jiawei Han",
"journal": "",
"ref_id": "b21",
"title": "CHEMNER: Fine-Grained Chemistry Named Entity Recognition with Ontology-Guided Distant Supervision",
"year": "2021"
},
{
"authors": "Congying Xia; Chenwei Zhang; Xiaohui Yan; Yi Chang; Philip Yu",
"journal": "",
"ref_id": "b22",
"title": "Zero-shot User Intent Detection via Capsule Neural Networks",
"year": "2018"
},
{
"authors": "Limin Yao; Sebastian Riedel; Andrew Mccallum",
"journal": "Association for Computing Machinery",
"ref_id": "b23",
"title": "Universal Schema for Entity Type Prediction (AKBC '13)",
"year": "2013"
},
{
"authors": "Mohamed Amir Yosef; Sandro Bauer; Johannes Hoffart; Marc Spaniol; Gerhard Weikum",
"journal": "",
"ref_id": "b24",
"title": "HYENA: Hierarchical Type Classification for Entity Names",
"year": "2012"
},
{
"authors": "Zheng Yuan; Doug Downey",
"journal": "AAAI Press",
"ref_id": "b25",
"title": "OTyper: A Neural Architecture for Open Named Entity Typing",
"year": "2018"
},
{
"authors": "Tao Zhang; Congying Xia; Chun-Ta Lu; Philip Yu",
"journal": "",
"ref_id": "b26",
"title": "MZET: Memory Augmented Zero-Shot Fine-grained Named Entity Typing",
"year": "2020"
},
{
"authors": "Yunyi Zhang; Jiaming Shen; Jingbo Shang; Jiawei Han",
"journal": "",
"ref_id": "b27",
"title": "Empower Entity Set Expansion via Language Model Probing",
"year": "2020"
},
{
"authors": "Ben Zhou; Daniel Khashabi; Chen-Tse Tsai; Dan Roth",
"journal": "",
"ref_id": "b28",
"title": "Zero-Shot Open Entity Typing as Type-Compatible Grounding",
"year": "2018"
}
] | [
{
"formula_coordinates": [
4,
317.61,
656.33,
206.8,
11.22
],
"formula_id": "formula_0",
"formula_text": "y = count (m) {H 1 (x), H 2 (x), ..., H n (x)} (1)"
},
{
"formula_coordinates": [
5,
343.64,
327.85,
176.53,
25.14
],
"formula_id": "formula_1",
"formula_text": "N = s i ∈S emb(s i ) + emb(c) |S| + 1 . (2"
},
{
"formula_coordinates": [
5,
520.17,
336.28,
4.24,
9.46
],
"formula_id": "formula_2",
"formula_text": ")"
},
{
"formula_coordinates": [
6,
113.29,
730.83,
175.85,
10.77
],
"formula_id": "formula_3",
"formula_text": "rank(type) = σ entail + σ cand(4)"
},
{
"formula_coordinates": [
6,
333.75,
455.3,
163.06,
67.23
],
"formula_id": "formula_4",
"formula_text": "⋮ Figure 4: Fine-Grained Type Refinement"
},
{
"formula_coordinates": [
15,
116.47,
96.27,
125.86,
26.44
],
"formula_id": "formula_5",
"formula_text": "Acc = Σ m∈M σ(t m == t m ) M"
},
{
"formula_coordinates": [
15,
114.77,
167.9,
129.27,
66.64
],
"formula_id": "formula_6",
"formula_text": "P ma = 1 |M | Σ m∈M |t m ∩ t m | t m R ma = 1 |M | Σ m∈M |t m ∩ t m | t m"
},
{
"formula_coordinates": [
15,
125.69,
276.92,
107.43,
66.93
],
"formula_id": "formula_7",
"formula_text": "P mi = Σ m∈M |t m ∩ t m | Σ m∈M t m R mi = Σ m∈M |t m ∩ t m | Σ m∈M t m"
}
] |
10.1145/nnnnnnn.nnnnnnn | 2023-05-21 | [
{
"figure_ref": [],
"heading": "INTRODUCTION",
"publication_ref": [
"b22",
"b16",
"b28",
"b12",
"b16",
"b22",
"b28"
],
"table_ref": [],
"text": "Nowadays, the ever-expanding new breeds of content, e.g., pictures, live streams, and short videos, drive the recommender system to drift from the traditional homogeneous form into an integrated form. Integrated recommender systems (IRSs) aim to simultaneously recommend heterogeneous items from multiple sources/channels in a row. This integrated form greatly expands users' choices on different types of content thereby satisfying users' diversified preferences on both item-level and channel-level. Therefore, IRS has nowadays been widely deployed in various online platforms such as the homepage feeds in Kuaishou [23], XiaohongShu [17], Taobao [29], and AliExpress [13]. In these products, users continuously slide down to browse and interact with heterogeneous items in a sequential manner, as shown in Figure 1.\nThough attractive, integrated feed recommendation faces more challenges than conventional recommendation with homogeneous items. First, real-world applications usually impose upper or lower exposure guarantees on different channels, such as lower constraints for sponsored/new content (e.g., ads and cold-start items) or upper constraints for individual channels to ensure diversity. These constraints lead to greater difficulty in the joint ranking of heterogeneous items. Second, heterogeneous items from multiple channels usually have different features and ranking strategies. Hence, it is difficult to directly compare items from different channels for joint ranking. Third, users' interests on different channels have a great impact on their behaviors, such that traditional user-item prediction models need to evolve into user-channel-item prediction models by considering both intra-channel and inter-channel information and their correlation with user interests. Finally, in feed products, users tend to review a large number of items in a row such that the previously viewed items have a great impact on the users' behavior towards the next item [17,23,29]. Therefore, it is of vital importance to consider the influence from page context when determining the item order in the return list.\nAlthough integrated feed recommendation has been widely deployed in practice, there are still few works focusing on the above challenges systematically. In this paper, we propose a general framework named Multi-channel Integrated Recommendation with Exposure Constraints (MIREC) to deal with the multi-channel integrated recommendation task under resource constraints in feed products. MIREC consists of two layers: an allocation-layer which optimizes the exposure of different channels from a global view over all user requests, and a ranking-layer which determines the optimal item layout of a given user request from a local view. These two layers operate in an iterative manner to make online decisions along with the arrival of user requests. The main contributions are as follows.\n• This work formulates the integrated recommendation task with exposure constraints as a binary online linear programming problem and proposes a two-layer framework named MIREC to obtain the optimal solution. We also describe a practical system architecture for its implementation in industrial platforms. • This work proposes an efficient multi-channel allocation algorithm to obtain the optimal exposure assignment of different channels over a fixed time horizon. The proposed algorithm is able to reach an optimal solution with linear complexity w.r.t. the number of constraints. We also prove that this algorithm admits a regret bound of O ( √ 𝑇 ) towards the global optimal point under certain assumptions.\n• This work proposes a series of collaborative models to determine the optimal layout of heterogeneous items on a page, with joint modeling of user interests, cross-channel correlation, and page context. This aligns more with the browsing nature of feed products than existing models. • This work conducts extensive experiments on both offline datasets and online A/B tests to verify the superiority of our proposed method.\nMIREC has been implemented on the homepage of Taobao to serve the main traffic. It brings 3% lift on user clicks, 1.56% lift on purchase, and 1.42% lift on stay time. It now serves hundreds of millions of users towards billions of items every day."
},
{
"figure_ref": [],
"heading": "RELATED WORK",
"publication_ref": [
"b9",
"b38",
"b27",
"b40",
"b3",
"b0",
"b26",
"b15",
"b19",
"b21",
"b31",
"b10",
"b14",
"b2",
"b8",
"b24",
"b33",
"b34",
"b1",
"b17",
"b39",
"b13",
"b5",
"b23",
"b12",
"b32",
"b36",
"b20"
],
"table_ref": [],
"text": "Re-ranking Methods. The main objective of re-ranking methods is to consider the mutual influence among a list of items in order to refine the prediction results produced by point-wise ranking models. Three prevalent models are commonly adopted in the existing literature: RNN-based methods, attention-based methods, and evaluator-generator-based methods. The first two methods feed an initial ranking list produced by point-wise models (e.g., Wide&Deep [10], DIN [39] and SIM [28]) into RNN-based (e.g., MiDNN [41], Seq2Slate [4] and DLCM [1]) or attention-based structure (e.g., PRM [27], PFRN [16], PEAR [20], and Raiss [22]) sequentially and output the encoded vector at each subsequent layer to model the mutual influences among items. The evaluator-generatorbased methods (e.g., SEG [32] and GRN [11]), use a generator to generate feasible permutations and use an evaluator to evaluate their list-wise utility to determine the optimal permutation. However, most re-ranking methods mainly focus on capturing the mutual influence among homogeneous items provided by one channel, instead of heterogeneous items provided by multiple channels. Moreover, they only optimize the item order at a single time slot, instead of considering a cumulative utility over a broad time horizon under resource constraints.\nOnline Allocation Methods. The online allocation problem with resource constraints has been mostly studied in online convex optimization [15]. The primal-dual methods [3,9,25,34,35] avoid taking expensive projection iterations by penalizing the violation of constraints through duality. The BwK methods [2,18] determines an optimal action from a finite set of possible actions and then optimize the policy of decision-making according to the observed rewards and costs over a fixed period of time. Several recent works studied the practical performance of online allocation in advertising recommendations. For example, PDOA [40] adopts the primal-dual framework by optimizing the dual prices with online gradient descent to eliminate the online max-min problem's regret. However, it assumes that the utility and cost values can be optimally estimated and only verify the performance through offline simulations. MSBCB [14] and HCA2E [6] proposed a two-level optimization framework based on BwK methods, where the high-level determines whether to present ads on the given request while the low-level searches the optimal position to insert ads. Most related works on online convex optimization focus on theoretical analysis (e.g., regret bound) instead of real-world applications. Other related works on advertising mainly consider binary content, i.e., ads or non-ads, instead of heterogeneous content. Directly extending them to deal with multi-channel recommendations in IRSs may lead to sub-optimal results.\nIntegrated Recommendation. The integrated recommendation is a newly emerged but rapidly developing domain driven by practical problems [24]. Integrated recommendation methods need to consider both intra-channel and inter-channel features within the heterogeneous content and provide recommendation results continuously along with user arrivals. Recently, DHANR [13] proposed a hierarchical self-attention structure to consider the cross-channel interactions. HRL-Rec [33] decomposed the integrated re-ranking problem into two subtasks: source selection and item ranking, and use hierarchical reinforcement learning to solve the problem. DEAR [37] proposed to interpolate ads and organic items by deep Q-networks. Cross-DQN [21] also adopt a reinforcement learning solution with a cross-channel attention unit. However, many integrated methods only focus on ranking at a single time slot instead of over a continuous time horizon. The joint consideration of both integrated ranking and online allocation of limited resources still remains to be explored."
},
{
"figure_ref": [],
"heading": "PROBLEM FORMULATION",
"publication_ref": [
"b11",
"b5",
"b25"
],
"table_ref": [],
"text": "This section formulates the integrated recommendation task with exposure constraints as a binary online linear programming problem. Specifically, we consider a generic IRS setting where user requests arrive sequentially during a finite time horizon. For each request, the IRS needs to rank a list of heterogeneous items provided by multiple channels. The aim is to maximize the overall utilities (e.g., the sum of clicks and pays) of all channels over the entire time horizon, subject to multiple resource constraints.\nFormally, the request of user 𝑢 triggered at time 𝑡 is described as 𝑒 𝑡 = (𝑢, 𝑓 , 𝑔, X 𝑡 ), where 𝑓 ∈ R + is a non-negative utility function, 𝑔 ∈ R + is a non-negative resource consumption function, and X 𝑡 ⊂ R 𝑑 + is a compact set denoting all possible item layouts for decisionmaking. For each request 𝑒 𝑡 , the IRS needs to choose a number of 𝑁 heterogeneous items from a candidate set 𝐼 𝑡 and place them into 𝑁 slots to form a complete page and return it to the user. This action 𝒙 𝑡 ∈ X 𝑡 can be represented as a decision matrix 𝒙 𝑡 ∈ [0, 1] 𝑁 ×|𝐼 𝑡 | , where each entry 𝑥 𝑡,𝑛,𝑖 is indexed by a slot 𝑛 and a card index 𝑖. Once the user finished viewing the current page, a new user request will be triggered to ask the platform to return to the next page. Therefore, this decision-making process will be performed repeatedly. Moreover, in real-world applications, the item layouts need to satisfy the following constraints:\nX 𝑡 = 𝑖 ∈𝐼 𝑡 𝑥 𝑡,𝑛,𝑖 = 1, ∀𝑡 ∈ T , ∀𝑛 ∈ N 𝑛 𝑥 𝑡,𝑛,𝑖 ≤ 1, ∀𝑡 ∈ T , ∀𝑖 ∈ 𝐼,(1)\nwhere the upper constraint restricts that each slot must be assigned to one item and the lower constraint restricts that each item can be assigned to at most one slot.\nAfter executing an action 𝒙 𝑡 at request 𝑒 𝑡 , the IRS consumes a resource cost 𝑔(𝒙 𝑡 ) and obtains an utility 𝑓 (𝒙 𝑡 ). In IRS, the utility function 𝑓 (𝒙 𝑡 ) is defined according to the concerned metrics. For example, it can be defined as a combination of stay time, adds to cart, and favorites to encourage user engagement, or defined as a combination of clicks and purchases to encourage user conversion. On the other hand, the consumption function 𝑔(𝒙 𝑡 ) is defined based on the concerned resource constraints. For example, the platform may need to allocate a certain amount of exposure to new channels in order to support the growth of new content [12]. Meanwhile, a too large proportion of exposures on one specific channel will damage the recommendation diversity thereby harming user experience [6,26]. In this case, the IRS needs to guarantee both a lower exposure limit and an upper exposure limit for the heterogeneous items from different channels.\nIn this paper, we focus on the exposure constraints in practical systems which lead to the following optimization problem\nP 0 : OPT(S) = max 𝒙 𝑡 ∈X 𝑡 ∑︁ 𝑇 𝑡 =1 𝑓 (𝒙 𝑡 )(2)\ns.t. 𝐶 1 : ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≤ 𝐺 max 𝑚,𝑡ℎ 𝑁 (S), ∀𝑚 ∈ M,(3)\n𝐶 2 : ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≥ 𝐺 min 𝑚,𝑡ℎ 𝑁 (S), ∀𝑚 ∈ M,(4)\nwhere 𝑁 (S) denotes the total available exposures to allocate over the entire time horizon, 𝐺 max 𝑚,𝑡ℎ and 𝐺 min 𝑚,𝑡ℎ denote the proportion of upper exposure limits and lower exposure limits for each channel 𝑚 ∈ M, respectively, and 𝑔 𝑚 (𝒙 𝑡 ) denotes the consumed exposures of cards from channel 𝑚 after executing 𝒙 𝑡 at request 𝑒 𝑡 . Although this paper mainly focuses on the exposure guarantee, the above formulation is generally applicable to other problems with different resource constraints, e.g., the number of coupons to allocate."
},
{
"figure_ref": [],
"heading": "METHODOLOGY 4.1 Framework Overview",
"publication_ref": [],
"table_ref": [],
"text": "Directly solving problem P 0 is challenging. On one hand, the estimation accuracy of utility 𝑓 (𝒙 𝑡 ) and consumption 𝑔(𝒙 𝑡 ) suffer influence from multi-factors, including the user's personal interest, the page context, and the cross-correlation between different channels. On the other hand, the determination of each 𝒙 𝑡 needs to consider the cumulative exposures over the entire time horizon due to the exposure guarantees. Therefore, the optimization of exposure allocation must be performed from a global view over the entire timeline instead of a single time slot.\nTo this end, we propose the MIREC framework which solves P 0 through online primal-dual iterations. Specifically, MIREC consists of two layers, i.e., the allocation-layer and the ranking-layer, which correspond to the dual and primal problem of P 0 , respectively. The allocation-layer optimizes dual variables to control the cumulative exposure of different channels on all user requests from a global view to guarantee the exposure limits. Meanwhile, the rankinglayer optimizes the item layout under fixed dual variables given by the allocation-layer, with the aim to maximize the instant utility on a single user request from a local view. These two layers operate in an iterative manner along with the arrival of online user requests to determine the optimal item layout at user requests continuously. The general workflow of MIREC is shown in Figure . 2.\nFor the allocation-layer, we propose a simple but efficient Mirrordescent based Multi-channel Exposure Allocation (ME2A) algorithm to adaptively balance the utility gain and the exposure cost of presenting heterogeneous items from different channels. The proposed M2EA algorithm has a closed form solution that can be computed in linear time and admits a regret bound of O ( √ 𝑇 ) towards the global optimal point under certain assumptions.\nFor the ranking-layer, we propose a personalized cross-channel ranking (PCR) model and a context-aware reranking (CAR) model to jointly determine the optimal item layout on a given user request, with fixed dual parameters given by the allocation-layer. In particular, PCR gives point-wise estimation of the quality of candidate items by joint modeling the influence from user interests, intra-channel information, and inter-channel correlations. Afterward, CAR refines the point-wise estimation generated by PCR into context-aware estimation by making use of both the context information and the high-level knowledge extracted from PCR."
},
{
"figure_ref": [],
"heading": "Global: Online Exposure Allocation",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we introduce the primal-dual formulation of P 0 and propose the ME2A algorithm to obtain the optimal solution for online systems. The complete algorithm is presented in Algorithm 1. \n+ ∑︁ 𝑚 𝜆 𝑚 ∑︁ 𝑡 𝑔 𝑚 (𝒙 𝑡 ) -𝐺 min 𝑚 .(5)\nHere 𝝁 ≥ 0 and 𝝀 ≥ 0 are the introduced dual parameters, 𝐺 max 𝑚 and 𝐺 min 𝑚 are short for 𝐺 max 𝑚,𝑡ℎ 𝑁 (S) and 𝐺 min 𝑚,𝑡ℎ 𝑁 (S), respectively. Note that the parameters 𝝁 and 𝝀 are related to the violation of exposure consumption over the upper bound limit and the lower bound limit, respectively, which are mutually exclusive. In particular, if one of them is positive, the other one must be zero; otherwise, both of them are zero. Hence, it is viable to only introduce one real number dual variable 𝝁 ∈ R 1×𝑀 to replace 𝝁 and 𝝀 in the dual problem, which simplifies (5) into\nmin 𝝁 𝐷 (𝝁) = ∑︁ 𝑡 𝑓 (𝒙 𝑡 ) - ∑︁ 𝑚 [𝜇 𝑚 ] + ∑︁ 𝑡 𝑔 𝑚 (𝒙 𝑡 ) -𝐺 max 𝑚 (6a) + ∑︁ 𝑚 [-𝜇 𝑚 ] + ∑︁ 𝑡 𝑔 𝑚 (𝒙 𝑡 ) -𝐺 min 𝑚 (6b) = ∑︁ 𝑡 𝑓 (𝒙 𝑡 ) - ∑︁ 𝑚 𝜇 𝑚 𝑔 𝑚 (𝒙 𝑡 ) + ∑︁ 𝑚 [𝜇 𝑚 ] + 𝐺 max 𝑚 -[-𝜇 𝑚 ] + 𝐺 min 𝑚 ,(6c)\nwhere\n[𝜇 𝑚 ] + = max{𝜇 𝑚 , 0}."
},
{
"figure_ref": [],
"heading": "Dual",
"publication_ref": [
"b4",
"b9"
],
"table_ref": [],
"text": "Optimization for Exposure Constraints. The dual problem in (6) can be solved optimally via primal-dual updates. Specifically, given a user request 𝑒 𝑡 = (𝑢, 𝑓 , 𝑏, X 𝑡 ), we assume that the utility 𝑓 (𝒙 𝑡 ) and cost 𝝁 𝑇 𝑡 𝑔(𝒙 𝑡 ) under different item layout 𝒙 𝑡 can be properly estimated by the models at the ranking layer. As such, the optimal item layout 𝒙 𝑡 under a fixed dual variable 𝝁 𝑡 can be obtained by solving the following primal problem:\nP 1 : x𝑡 = arg max 𝒙 𝑡 ∈X 𝑓 (𝒙 𝑡 ) -𝝁 𝑇 𝑡 𝑔(𝒙 𝑡 ) .(7)\nThis is the focus of the ranking-layer to be discussed layer. After the optimal item layout 𝒙 𝑡 at user request 𝑒 𝑡 is properly determined, the next step is to update the dual variable 𝝁 𝑡 to adjust the exposure of different channels in future user requests. Specifically, the remained exposure resource of different channels after the presentation of 𝒙 𝑡 at 𝑒 𝑡 is updated by\n𝐺 max 𝑚,𝑡 +1 = 𝐺 max 𝑚,𝑡 -𝑔 𝑚 (𝒙 𝑡 ), ∀𝑚 ∈ M.(8)\nThe sub-gradient of the dual function in ( 6) can be obtained via Danskin's theorem [5] by 9) where 1(𝑥 ∈ 𝐴) is an indicator function which equals to one if 𝑥 ∈ 𝐴 otherwise zero. As such, the dual variable 𝝁 𝑡 can be updated based on the mirror-descent method as\n∇𝜇 𝑚,𝑡 = -𝑔 𝑚 (𝒙 𝑡 ) +𝐺 max 𝑚,𝑡 +1 • 1(𝜇 𝑚,𝑡 ≥ 0) +𝐺 min 𝑚 • 1(𝜇 𝑚,𝑡 ≤ 0), (\n𝜇 𝑚,𝑡 +1 = arg min 𝜇 𝑚 ∈R 𝜇 𝑚 ∇𝜇 𝑚,𝑡 + 1 𝜂 𝑉 ℎ (𝜇 𝑚 , 𝜇 𝑚,𝑡 ),(10)\nwhere 𝑉 ℎ (𝑥, 𝑦) = ℎ(𝑥)ℎ(𝑦)∇ℎ(𝑦) 𝑇 (𝑥-𝑦) is the Bregman divergence based on reference function ℎ(•) and 𝜂 ∈ R is a fixed step-size. Note that this mirror descent step can be computed in linear time since (10) admits a closed-form solution. For example, if we use ℎ(𝝁) = 1 2 ∥𝝁 ∥ 2 as the reference function, the dual update in (10) becomes\n𝝁 𝑡 +1 = [𝝁 𝑡 -𝜂∇𝝁 𝑡 ] + ,(11)\nwhich recovers the online projected gradient descent method. Moreover, in order to guarantee the upper exposure constraints, one needs to examine the violation of upper limits of each channel before the determination of 𝒙 𝑡 at each user request. If the sum of exposures of a specific channel exceeds its upper bound, one needs to remove all candidate items from this channel to forbid allocate more exposures when determining 𝒙 𝑡 . We present the optimality of this proposed ME2A algorithm and its feasibility to guarantee exposure constraints of different channels as follows. Detailed proofs are deferred to the appendix."
},
{
"figure_ref": [],
"heading": "Algorithm 1",
"publication_ref": [],
"table_ref": [],
"text": "The proposed ME2A algorithm of MIREC Receive request 𝑒 𝑡 = (𝑢, 𝑓 , 𝑏, X 𝑡 )."
},
{
"figure_ref": [],
"heading": "6:",
"publication_ref": [],
"table_ref": [],
"text": "Update the candidate set 𝐼 𝑡 provided by multi-channels."
},
{
"figure_ref": [],
"heading": "7:",
"publication_ref": [
"b6"
],
"table_ref": [],
"text": "Determine the optimal item list 𝒙 𝑡 by solving the primal problem in (7) at the ranking-layer."
},
{
"figure_ref": [],
"heading": "8:",
"publication_ref": [],
"table_ref": [],
"text": "Update the remained exposure resource via (8)."
},
{
"figure_ref": [],
"heading": "9:",
"publication_ref": [
"b8",
"b10",
"b2",
"b24",
"b39",
"b9"
],
"table_ref": [],
"text": "Obtain the sub-gradient of the dual variable via (9). Update the dual variable based on mirror descent via (11). 11: end for 4.2.3 Optimality. It is viable to prove that Algorithm 1 is asymptotically optimal and admits a regret bound scales as O ( √ 𝑇 ) when the user requests arrive from an i.i.d unknown distribution. This assumption is reasonable when the number of requests is numerous [3,25,40]. Specifically, we denote Algorithm 1 as 𝜋 and the overall utility over all user requests in set S under the running of 𝜋 as 𝑅(𝜋 |S) = 𝑇 𝑡 =1 𝑓 (𝒙 𝑡 ). The regret of model 𝜋 is defined as the worst-case difference over S between the expected performance of the global optimal solution and the model 𝜋:\nRegret(𝜋 |S) = sup E S [OPT(S) -𝑅(𝜋 |S)] ,(12)\nwhere OPT(S) denotes the optimal utilities one can obtain under the request set S. The regret bound can be given as follows.\nTheorem 1. Suppose that the requests come from an i.i.d model with unknown distribution. Then, Regret(𝜋 |S) ≤ 𝐶 1 +𝐶 2 𝜂𝑇 + 𝐶 3 𝜂 with 𝜂 > 0 holds for any 𝑇 ≥ 1. Here 𝐶 1 , 𝐶 2 and 𝐶 3 are constant values depending on the numerical bounds of the utility 𝑓 , the consumption 𝑔, and terms from the dual iterates in Eq. (10).\nFrom Theorem 1, we obtain Regret(𝜋 |S) ≤ 𝑂 ( √ 𝑇 ) when using a step-size 𝜂 ∝ 𝑐/"
},
{
"figure_ref": [],
"heading": "√",
"publication_ref": [],
"table_ref": [],
"text": "𝑇 with any constant 𝑐 > 0. We defer the proof and detailed definitions of 𝐶 1 , 𝐶 2 , and 𝐶 3 into the appendix."
},
{
"figure_ref": [],
"heading": "Exposure Feasibility.",
"publication_ref": [
"b9"
],
"table_ref": [],
"text": "In Algorithm 1, if the upper exposure limit of a specific channel is violated, we will forbid the exposure of any item from this channel when determining the item list 𝒙 𝑡 . Therefore, the exposure can never be overspent. On the other hand, the lower exposure limits are soft-restricted by adaptively adjusting the dual variable 𝝁. This may cause exposure underspent. However, it is viable to prove that the violation of the lower exposure limit of any channel also admits a convergence rate of 𝑂 ( √ 𝑇 ). In other words, even if the violations on lower exposure limits may occur, their growth is considerably smaller than 𝑇 . Proposition 1. Suppose the requests come from an i.i.d model with unknown distribution. Then, it holds for any 𝑇 ≥ 1 and any\nchannel 𝑚 ∈ M that 𝐺 min 𝑚 -E 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≤ 𝐶 4 + 𝐶 5\n𝜂 , where 𝐶 4 and 𝐶 5 are constant values depends on the numerical bounds of utility 𝑓 , consumption 𝑔, and terms from the dual iterates (10).\nProposition 1 states that when using 𝜂 ∝ 𝑐/"
},
{
"figure_ref": [],
"heading": "√",
"publication_ref": [],
"table_ref": [],
"text": "𝑇 with 𝑐 > 0, the exposure underspend of any channel is bounded by 𝑂 ( √ 𝑇 ). We defer the proof and definitions of 𝐶 4 and 𝐶 5 into the appendix."
},
{
"figure_ref": [],
"heading": "Local: Context-Aware Integrated Ranking",
"publication_ref": [
"b5",
"b6"
],
"table_ref": [],
"text": "Different from the allocation-layer which optimizes an objective with accumulative utilities over the entire time horizon as defined in (6), the ranking-layer focus on maximizing the utilities on a single time slot. This corresponds to the primal problem given in (7):\nP 1 : x𝑡 = arg max 𝒙 𝑡 ∈X 𝑓 (𝒙 𝑡 ) -𝝁 𝑇 𝑡 𝑔(𝒙 𝑡 ) .\nIn other words, the allocation layer adjusts the exposure of items from different channels by optimizing the dual parameter 𝝁 𝑡 from a global view of all user requests. While the ranking-layer determines the optimal item list 𝒙 𝑡 under a fixed dual variable 𝝁 𝑡 from a local view of a given user request 𝑒 𝑡 .\nThere are two common characteristics that are strongly related to the estimation of 𝑓 (𝒙 𝑡 ) and 𝑔(𝒙 𝑡 ) in the integrated recommendation. First, users' preference on different channels has a great impact on the utility (e.g., prefer to click or not) and exposure (e.g., prefer to view or not) estimations. Therefore, it is of vital importance to consider both intra-channel and inter-channel correlations with reference to user interests during the estimation. Second, in feed products, users tend to review a large number of items in a row such that the previously viewed items have a great impact on users' behavior towards the next item. Therefore, it is necessary to consider page context when determining the item order.\nTherefore, we propose two models to deal with the above two challenges, respectively. First, we propose PCR model to deal with the joint modeling of user interests and inter/intra-channel correlation. It gives a point-wise estimation of the utility/exposure value of presenting each candidate item. Second, we propose CAR model to refine the point-wise estimation from PCR into contextaware estimation by considering both context information and the high-level knowledge obtained from PCR. It is also responsible for selecting optimal items from a set of candidate items to generate the final return list. In real-world systems, for each user request, we only need to run PCR once to get the point-wise scores, and then run CAR multiple times to generate the return list. Next, we mainly focus on the estimation of 𝑓 (𝑥), the estimation of 𝑔(𝑥) can be performed in a similar way by changing the learning goals."
},
{
"figure_ref": [
"fig_0"
],
"heading": "Personalized",
"publication_ref": [
"b30",
"b26",
"b12",
"b6",
"b27",
"b37",
"b38",
"b19",
"b18"
],
"table_ref": [],
"text": "Cross-Channel Ranking Model. PCR takes four types of features as input, i.e., the user profile feature 𝑋 𝑢 , the user behavior sequences 𝑋 𝑏 , the candidate items provided by each channel 𝑋 𝑙 , and the item-level features of target item 𝑋 𝑖 . As shown in Figure 2, We use an embedding layer to transform these features into dense embedding vectors, denoted as 𝐸 𝑢 , 𝐸 𝑏 , 𝐸 𝑙 and 𝐸 𝑖 , respectively. These embedding vectors are then fed into three components, i.e., the intra-channel encoding layer, interest-aware evolution layer, and inter-channel encoding layer in order, which are described below. Intra-Channel Encoding layer. This layer aims at extracting the mutual influence of item pairs and other extra information within the channel. We adopt the well-known multi-head attention [31] as the basic learning unit for intra-channel encoding. This is due to that the self-attention mechanism is able to directly capture the mutual influences between any two items, and is robust to far distance within the sequence. Formally, the formulation of this attention-based encoding can be written as\n𝑉 𝑚 𝑙 = [ℎ𝑒𝑎𝑑 1 , ℎ𝑒𝑎𝑑 2 , ..., ℎ𝑒𝑎𝑑 ℎ ]𝑊 𝑂 ,(13a)\nℎ𝑒𝑎𝑑 𝑖 = Softmax (𝐸 𝑏 𝑊 𝑄 )(𝐸 𝑏 𝑊 𝐾 ) 𝑇 √︁ 𝑑 ℎ /ℎ (𝐸 𝑏 𝑊 𝑉 ) ,(13b)\nwhere 𝑊 𝑂 ∈ R 𝑑 ℎ ×𝑑 ℎ denotes the learnable parameters for each head with 𝑑 ℎ being the length of projected embedding vector after attention, 𝑊 𝑄 ,𝑊 𝐾 ,𝑊 𝑉 ∈ R 𝑑×𝑑 ℎ /ℎ are the vectors of query, key and value with 𝑑 being the length of original embedding vector and ℎ being the number of heads, 𝑉 𝑚 𝑙 represents the encoded candidate items of each channel 𝑚 ∈ M. Interest-Aware Evolution Layer. Existing works such as PRM [27] and DHANR [13] directly apply the self-attention mechanism to model the inter-dependencies among items and channels without considering user's recent behavior. However, the interests hidden in user's behavior items usually have a great impact on the prediction accuracy in recommendation tasks [7,28,38,39]. The recently proposed PEAR [20] firstly models the dependency between the candidate item list and the user's historical behaviors based on a transformer-like structure, which, however, suffers from two limitations. First, directly mixing the raw item-level features from user behavior items may introduce redundant or noisy information to degrade the learning performance. Second, each user may exhibit multiple interest points, such that it is beneficial to reinforce the interest related to the target item before feature-crossing to avoid drifting. Therefore, we first reinforce the interest vector according to the correlation between behavior items and the target item as\n𝑉 𝑈 = 𝑓 (𝐸 𝑏 ; 𝐸 𝑖 ) = ∑︁ 𝐵 𝑖=1 𝐴(𝑏 𝑖 , 𝐸 𝑖 )𝑏 𝑖 = ∑︁ 𝐵 𝑖=1 𝑤 𝑖 𝑏 𝑖 , (14\n)\nwhere 𝐵 is the length of behavior sequence, 𝑏 𝑖 is the behavior item, 𝑉 𝑈 denotes the user representation feature with respect to 𝐸 𝑖 , and 𝐴(•) is a feed-forward network whose output is the activation weight 𝑤 𝑖 . Then, we make use of this reinforced interest vector to extract useful information from the candidate items of different channels. Formally, for each channel 𝑚, given 𝑉 𝑚 𝑙 and 𝑉 𝑈 as inputs, we use scaled dot-product attention formulated as follows:\n𝐻 𝑖 𝑠 = Softmax (𝑉 𝑚 𝑙 𝑊 𝑞 ) [𝑉 𝑈 𝑊 𝑘1 , 𝑉 𝑚 𝑙 𝑊 𝑘2 ] 𝑇 √︁ 𝑑 ℎ [𝑉 𝑈 𝑊 𝑣1 , 𝑉 𝑚 𝑙 𝑊 𝑣2 ],(15)\nwhere 𝑊 𝑘1 ,𝑊 𝑣1 ∈ R 𝑑×𝑑 ℎ and 𝑊 𝑘2 ,𝑊 𝑞 ,𝑊 𝑣2 ∈ R 𝑑 ℎ ×𝑑 ℎ are all learnable parameters, [•] denotes the concatenation operation. After the above operations, we successfully merged the information from the candidate item lists and the user's historical behaviors into a series of evolved embedding vectors for further processing. Inter-Channel Encoding Layer. Previous layers mainly extract intra-channel correlations. We now focus on the modeling of interchannel correlation. First, we feed the embedding vector of each channel and the target item embedding into the MLP layer with softmax function to obtain the importance weights on each channel that is related to the target item:\n𝑊 𝐶𝐻 = Softmax(MLP[𝐻 𝑚 𝑠 , 𝐸 𝑖 ]),(16)\nwhere 𝑊 𝐶𝐻 ∈ R 1×𝑚 is the importance weights, and 𝐻 𝑚 𝑠 denotes the concatenation of all channels' output from the Interest-Aware Evolution Layer. Second, we perform multi-head self-attention on the evolved embedding vector 𝐻 𝑚 𝑠 of each channel 𝑚 ∈ M to obtain the mixed embedding H𝑚 𝑠 which contains rich inter-channel information. Then, we perform the weighted sum on the mixed embedding of all channels based on𝑊 𝐶𝐻 to get the final representation of multi-channel modeling:\n𝑉 𝐿 = 𝑊 𝐶𝐻 • [ H𝑚 𝑠 ] 𝑇 , 𝑚 ∈ M,(17)\nwhere [ H𝑚 𝑠 ] represents the concatenation of the mixed embeddings of all channels.\nFinally, we concatenate all vectors as input and feed it into the MLP layers with a sigmoid function to predict the utility of presenting a given target item to a given target user as We maintain two types of context information in CAR, i.e., the context of previous items and the context of remaining candidate items. Specifically, when selecting the 𝑘-th item in a page, we represent the context of previously presented 𝑘 -1 items ℎ 𝑝𝑟𝑒 by mean-pooling over their embeddings. Meanwhile, we represent the context of all candidate items ℎ 𝑐𝑎𝑛 by mean-pooling over the embeddings of all remained candidates. These two context vectors are updated and repeated along with the item selection process. Furthermore, we perform a series of embedding crossing operations between the target item embedding 𝑒 𝑖 and the context embeddings to model the influence from page context. In specific, the operations in the context of previous items can be formulated as follows:\n𝑌 𝑃𝐶𝑅 = Sigmoid(Concat(𝐸 𝑢 , 𝐸 𝑖 , 𝑉 𝐿 ))(18\n𝐻 𝑝𝑡 = Concat(ℎ 𝑝𝑟𝑒 ⊕ 𝑒 𝑖 , ℎ 𝑝𝑟𝑒 ⊗ 𝑒 𝑖 , ℎ 𝑝𝑟𝑒 ⊖ 𝑒 𝑖 ),(19)\nwhere ⊕, ⊗, and ⊖ denote the addition, subtraction, and dot product between embedding vectors, respectively. The same goes for 𝐻 𝑐𝑡 by replacing ℎ 𝑝𝑟𝑒 in (19) with the context of candidate items ℎ 𝑐𝑎𝑛 . Additionally, we also perform embedding-crossing between the context embeddings and the high-level knowledge 𝑉 𝐿 from PCR to obtain another two vectors, i.e., 𝐻 𝑣 𝑐𝑡 and 𝐻 𝑣 𝑝𝑟𝑒 . Finally, for each candidate item 𝑖 ∈ 𝐼 𝑐𝑎𝑛𝑑 , we predict the contextaware utility score by feeding these embedding vectors along with the point-wise utility score 𝑌 𝑃𝐶𝑅 from PCR and user profile features 𝐸 𝑢 into one MLP layer as\n𝐻 all = Concat(𝐻 𝑝𝑡 , 𝐻 𝑐𝑡 , 𝐻 𝑣 𝑐𝑡 , 𝐻 𝑣 𝑝𝑟𝑒 , ℎ 𝑝𝑟𝑒 , ℎ 𝑐𝑎𝑛 , 𝐸 𝑢 , 𝑌 𝑃𝐶𝑅 ), (20a) 𝑌 𝐶𝐴𝑅 =𝜎 (MLP(𝐻 all )),(20b)\nwhere 𝜎 represents the sigmoid activation function. After scoring all candidate items, we choose the item with the highest score as the optimal item for slot 𝑘, and update the context vectors and the remained candidate items accordingly. This process will be repeated 𝐾 times to generate a return item list of length 𝐾. Note that the above operations in CAR only involve linear computations, such that this item selection process is cost-efficient in online systems. Both PCR and CAR can be trained with the commonly used crossentropy loss as in other ranking models, the learning objective can be given as\n𝐽 = ∑︁ 𝑒 𝑡 ∈ D 𝑦 𝑒 𝑡 𝑢,𝑖 log ŷ𝑒 𝑡 𝑢,𝑖 + (1 -𝑦 𝑒 𝑡 𝑢,𝑖 ) log(1 -ŷ𝑒 𝑡 𝑢,𝑖 ) ,(21)\nwhere D denotes the training dataset, 𝑦 𝑒 𝑡 𝑢,𝑖 is the real user-item recommendation label (equals 1 or 0) between user 𝑢 and item 𝑖 at request 𝑒 𝑡 , and ŷ𝑒 𝑡 𝑢,𝑖 is the predicted label given by 𝑌 𝑃𝐶𝑅 or 𝑌 𝐶𝐴𝑅 . In our experiments, when predicting the utility function 𝑓 with user clicks, 𝑦 𝑒 𝑡 𝑢,𝑖 refers to the click label; when predicting the cost function 𝑔 with exposure constraints, 𝑦 𝑒 𝑡 𝑢,𝑖 refers to the exposure label between user 𝑢 and item 𝑖 at request 𝑒 𝑡 , i.e., whether user 𝑢 has seen item 𝑖 at request 𝑒 𝑡 . One can readily change the learning objective according to actual demands."
},
{
"figure_ref": [],
"heading": "Online Implementation",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we introduce the online implementation of our proposed MIREC model in the homepage feed of Taobao. The presented system architecture is able to handle 120, 000 QPS at traffic peak and respond within 20 milliseconds in general. It now serves the main traffic of Taobao to provide services to hundreds of millions of users towards billions of items in Taobao every day.\nFigure . 3 gives a general architecture to implement our proposed MIREC model in real-world IRS. Each time a user request is triggered from the device, the upstream RS of each channel will run its own recommendation models to determine the top items to return. The Integrated Recommendation Controller uses the top items from all channels as the candidates. It retrieves user/item features from a feature center in real-time and ranks candidate items by solving based on our proposed LCPR and PCR model. Meanwhile, the dual variable 𝜇 is estimated by an Exposure Controller. This module monitors the completeness of exposure guarantees based on the real-time exposures collected from user logs and updates the dual variable to adjust the exposures on different channels periodically."
},
{
"figure_ref": [],
"heading": "EXPERIMENTAL RESULTS",
"publication_ref": [
"b7"
],
"table_ref": [],
"text": "This section conducts extensive experiments on both offline datasets and real-world applications with the goal of answering the following research questions: Q1: Does our proposed PCR and CAR outperform other baseline models in integrated ranking tasks? Q2: Does our proposed MIREC framework outperform other methods in integrated recommendation tasks with exposure constraints? Q3: How does MIREC perform in real-world applications?\n5.1 Experimental Setup 5.1.1 Datasets. We use one public dataset named MicroVideo-1.7M and one industrial dataset named Taobao for experiments. The public available MicroVideo-1.7M dataset1 released by [8] contains 12, 737, 619 interactions that 10, 986 users have made on 1, 704, 880 micro-videos. This dataset provides rich user behavior data and timestamps to evaluate the performance on both interest modeling and context-aware reranking. The Taobao dataset is an industrial private dataset that contains users' behaviors and feedback logs from multiple channels in the homepage feed of Taobao Mobile App. It is one of the largest feed scenarios for online merchandise in China. The feed provides items in form of the streams, videos, pictures, etc, from various channels. Users can slide to view more items in a row. This dataset contains about ten billion interactions that one hundred million of users have made on sixty million items. We also conduct online A/B tests on the platform Taobao to examine the performance of MIREC in real-world applications."
},
{
"figure_ref": [],
"heading": "Comparing",
"publication_ref": [
"b38",
"b38",
"b0",
"b26",
"b19",
"b29",
"b12",
"b35",
"b5",
"b5"
],
"table_ref": [],
"text": "Methods. We compare MIREC with two mainstreams of baselines. The first steam of baselines are the methods for ranking tasks with different goals on user interest modeling (i.e., DIN and DIEN), re-ranking (i.e., DLCM, PRM, and PEAR), or multichannel recommendation (i.e., STAR and DHANR). Specifically, DIN [39] is a widely used benchmark for sequential user data modeling in point-wise CTR predictions, which models short behavior sequences with target attention. DIEN [39] combines GRUs and attention to capture temporal interests from users' historical behaviors with respect to the target item. DLCM [1] uses gated recurrent units (GRU) to sequentially encode the top-ranked items with their feature vectors. PRM [27]: directly optimizes the whole recommendation list by employing a Transformer structure to efficiently encode the information of all items in the list. PEAR [20] not only captures feature-level and item-level interactions but also models item contexts from both the candidate list and the historical clicked item list. STAR [30] trains a unified model to serve all channels simultaneously, which consists of shared centered parameters and channel-specific parameters. DHANR [13] proposes a hierarchical self-attention structure to consider cross-channel interactions.\nThe second stream of baselines is the online allocation methods which have been successfully applied in industrial applications for online resource allocation. Fixed is the fixed-positions strategy, where the positions of recommended items and ads are manually pre-determined for every request. 𝛽-WPO: is based on the Whole-Page Optimization (WPO) [36]. WPO ranks recommended and ad candidates jointly according to the predefined ranking scores. Similar to [6], we introduce an adjustable variable 𝛽 to control the proportion of different channels on each request to satisfy the resource constraint. In general, 𝛽-WPO can be regarded as a heuristic list merging algorithm. Each list from one channel is assigned a priority weight. The algorithm merges the top items of each list based on both their ranking scores and the priority weights into a final return list. HCA2E [6]: proposed a two-level optimization framework based on BwK methods. The high-level determines whether to present ads on the page while the low-level searches the optimal position to insert ads heuristically."
},
{
"figure_ref": [],
"heading": "Metrics.",
"publication_ref": [
"b18"
],
"table_ref": [],
"text": "For offline experiments, we use user clicks to measure the utility function 𝑓 . For online experiments, we consider a joint measurement of user click, purchase, and stay-time for utility function 𝑓 . For all experiments, we use the exposure of items to measure cost function 𝑔. In this case, we compare the performance of integrated ranking in offline evaluation using the widely used Area Under ROC (AUC) and normalized discounted cumulative gain (nDCG) [19]. Remark that nDCG@K refers to the performance of top-k recommended items in the return list. The online performance is evaluated by CLICK, Click-Through-Rate (CTR), Gross Merchandise Volume (GMV), and Stay Time. Here, CLICK refers to the total number of clicked items. CTR is defined as CLICK/PV with PV denoting the total number of impressed items. CTR measures users' willingness to click and is therefore a widely used metric in practical applications. GMV is a term used in online retailing to indicate a total sales monetary-value for merchandise sold over a certain period of time. Stay Time denotes the time period of users' average stay time in the product, averaged on all users.\n5.1.4 Reproducibility. Our source codes have been made public to ensure reproducibility 2 . In all experiments, we use the validation set to tune the hyper-parameters to generate the best performance for different methods. The learning rate is searched from 10 -4 to 10 -2 . The L2 regularization term is searched from 10 -4 to 1. All models use Adam as the optimizer."
},
{
"figure_ref": [],
"heading": "Offline Evaluation",
"publication_ref": [],
"table_ref": [
"tab_3"
],
"text": "5.2.1 Q1: Performance on Integrated Ranking. We first compare the performance with the first stream of baselines on item ranking. The results are shown in Table 1, which leads to the following findings. First, the re-ranking methods perform generally better than the point-wise user interest methods, indicating that modeling mutual influence among the input ranking list is of vital importance for the ranking. Therefore, it is essential to consider the influence from page context in feed recommendations. Second, the multi-channel methods perform better than the reranking methods, which verifies that exploiting the distinction and mutual influence among different channels has a great impact on integrated recommendations. Besides, we also notice that DHANR performs better than STAR, which may be due to that DHANR considers both the correlation among different channels and the influence from the candidate list. Finally, our proposed MIREC model achieves superior performance than other competitors on all datasets, verifying the effectiveness of joint modeling the cross-channel information, user interest, context information, and candidate list."
},
{
"figure_ref": [],
"heading": "Ablation Study.",
"publication_ref": [
"b19"
],
"table_ref": [],
"text": "The results in Table . 2 investigates the impact of each component of MIREC on item quality estimation. Specifically, PCR * replaces the attention mechanism for user behaviors in the Merged-Sequence Evolution layer with a self-attention mechanism which is in accordance with PEAR [20]. PCR outperforms PCR * indicates that the tailored attention mechanism in PCR can filter out noisy or redundant information from historical behaviors to benefit the subsequent modeling of bi-sequence interaction. PCR † removes the scaled dot-product attention mechanism (i.e. there is no explicit interaction between initial lists and user behaviors) and achieved a worse performance. This demonstrates the necessity of this direct modeling between sequences, directly guiding the reordering of the initial lists. PCR w/o IntraCE removes the Intra-Channel Encoding module, i.e., directly feeding the embeddings of the initial item lists into subsequent layers for learning. The result shows that PCR achieves superior performance than PCR w/o IntraCE, verifying that it is of vital importance to model the mutual information inside each channel for final prediction. PCR w/o InterCE removes the Inter-Channel Encoding, which also leads to worse performance. It verifies that without considering the relationship and distinction between different channels will degrade the model performance considerably. Moreover, the joint learning of PCR and CAR performs better than only using PCR. This verifies that the modeling of page-wise context information can improve prediction accuracy effectively."
},
{
"figure_ref": [],
"heading": "Q2:",
"publication_ref": [],
"table_ref": [
"tab_5",
"tab_5"
],
"text": "Performance with Exposure Constraints. To the best of our knowledge, there does not exist publicly available datasets which has rich user logs and multi-channel features to examine the joint performance of integrated recommendation and exposure allocation. Therefore, we only perform experiments on Taobao dataset, +1.42% using the complete platform logs. In this experiment, we assume that the IRS needs to allocate exposures to satisfy the exposure guarantees of four distinct channels. The aim is to maximize the overall user-click utility of all channels. The compared fixed, WPO, and HCA2E methods all use point-wise scores to be consistent with their original proposals. For HCA2E, we use their proposed heuristic search method to determine the final order of the item list. The results are shown in Table 3, which are averaged on multiple runs to give a fair comparison. The simulated time horizon is one complete day with more than one billion user requests from real productive environment. We evaluate the performance using two different sets of lower bounds: 1) Channel 1=55%, Channel 2 = 20%, Channel 3 = 15%, Channel 4 = 10%; 2) Channel 1=70%, Channel 2 = 15%, Channel 3 = 10%, Channel 4 = 5%. The parameters of all comparing methods are carefully tuned to satisfy the exposure constraints. The completeness in Table 3 shows that all methods can control the violation of constraints to a low-level. HCA2E and our proposed MIREC perform slightly better than the fixed method and the WPO method. Noticeably, our proposed method outperforms other comparing methods considerably in terms of CTR enhancement, which verifies that the joint use of the allocation and estimation algorithm can bring a remarkable improvement in practical environments."
},
{
"figure_ref": [],
"heading": "Online Evaluation",
"publication_ref": [],
"table_ref": [
"tab_6"
],
"text": "MIREC has been fully deployed in the homepage feed of Taobao named Guess-you-like to serve the main traffic. Guess-you-like is one of the largest merchandise feed recommendation platform in China, which serves more than hundreds of millions of users toward billions of items every day. We deploy MIREC at the integrated recommendation stage in Guess-you-like platform, which takes hundreds of candidate items provided by multiple channels as input and outputs the final item list to return to the user. The online performance is compared against our previous baseline which is similar as a combination of 𝛽-WPO and HCA2E. In particular, the baseline uses a point-wise model for item quality estimation and uses a PID-based feedback control to automatically adjust parameter 𝛽 to guarantee the exposures for different channels. For each user request, the baseline also runs an MDP-based search method to determine the optimal card layout based on the estimated scores, which is similar as the heuristic search method in HCA2E. The overall performance in Table 4 is averaged over two consecutive weeks. The results show that compared with the baseline method, MIREC brings an improvement of 3.00% for CLICK, 1.75% for CTR, 1.56% for GMV, and 1.42% for stay time. Compared with the fixed method, MIREC brings an improvement of 3.00% for CLICK, 1.75% for CTR, 1.56% for GMV, and 1.42% for stay time. These improvements indicate that our framework is able to increase user's willingness to stay and interact with the recommended items in practical applications. It is noteworthy that 1% improvement on CLICK in Guess-you-like brings millions of clicks every day. Figure . 4 shows a detailed comparison of the exposure allocation results of a specific channel, where the items from this channel have a generally lower CTR than others. Each line in Figure . 4(a) represents the robustness of long-term exposure guarantee of this channel within two consecutive weeks. It is clear that compared with the baseline, our proposed MIREC is more robust to alleviate daily exposure fluctuations. The distribution of exposures on different positions in the feed is given in Figure . 4(b). The result shows that our proposed framework tends to put lower-quality items backward to increase the overall utilities of all channels. Consequently, as shown in Figure . 4(c), the averaged CTR of all channels on each position can be improved remarkably. This verifies that MIREC is superior in optimizing the item layout from a global perspective."
},
{
"figure_ref": [],
"heading": "A PROOF OF REGRET BOUND",
"publication_ref": [
"b2",
"b24",
"b2",
"b24",
"b2",
"b24",
"b24",
"b24",
"b29",
"b33",
"b24",
"b24"
],
"table_ref": [],
"text": "Our proof shares the same spirit as that of Theorem 1 in [3,25]. The difference is that [3] does not consider a lower resource limit while [25] develops proof with an additional learnable parameter within 𝑓 (𝒙 𝑡 ) and 𝑔(𝒙 𝑡 ). Therefore, we here develop a separate proof that is consistent with our formulation. We directly refer to a few propositions in [3,25] as preliminaries for simplicity. It is noteworthy that developing new proof of the online revenue maximization problem is not the main focus of this paper.\nRecall that the integrated recommendation problem is\nP 0 : OPT(S) = max 𝒙 𝑡 ∈X ∑︁ 𝑇 𝑡 =1 𝑓 (𝒙 𝑡 )(22)\ns.t. 𝐶 1 : ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≤ 𝐺 max 𝑚,𝑡ℎ 𝑁 (S), ∀𝑚 ∈ M,(23)\n𝐶 2 : ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≥ 𝐺 min 𝑚,𝑡ℎ 𝑁 (S), ∀𝑚 ∈ M.(24)\nSince 𝑁 (S) denotes the sum exposures over the entire time horizon from 𝑡 = 1 to 𝑇 , we can replace the the upper exposure limit 𝐺 max 𝑚,𝑡ℎ 𝑁 (S) and lower exposure limit 𝐺 min 𝑚,𝑡ℎ 𝑁 (S) with 𝑇𝐺 𝑚 and 𝛼𝑇𝐺 𝑚 for simplicity, respectively, where 𝐺 𝑚 , 𝛼 ∈ [0, 1] are constants. As such, problem P 0 can be reformulated as\nP 1 : OPT(S) = max 𝒙 𝑡 ∈X ∑︁ 𝑇 𝑡 =1 𝑓 (𝒙 𝑡 )(25) s\n.t. 𝛼𝑇𝐺 𝑚 ≤ ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≤ 𝑇𝐺 𝑚 , ∀𝑚 ∈ M.(26)\nBefore our analysis, we define constants f , ḡ > 0, 𝐺 > 0 and Ḡ > 0 such that sup 𝒙 ∈X 𝑓 (𝒙) ≤ f , sup 𝒙 ∈X 𝑔(𝒙) ≤ ḡ, 𝐺 := min 𝑚 ∈M 𝐺 𝑚 and Ḡ := max 𝑚 ∈M 𝐺 𝑚 . Also, 𝜃 refers to the strongly-convexity parameter of the reference function ℎ(•).\nFirst, we bound the dual iterates as follows.\nAssumption A.1. There exists a constant 𝐶 ℎ > 0 such that the dual iterates\n𝜇 𝑡 satisfy E[||∇ℎ(𝜇 𝑡 )|| ∞ ] ≤ 𝐶 ℎ , ∀𝑡 ∈ [𝑇 ].\nRemark 1. Note that, when choosing the reference function ℎ(𝜆) := 1 2 ∥𝜆∥ 2 , Assumption A.1 can be omited according to Proposition 3 in [25].\nDenote the online Algorithm 1 as 𝜋 which makes a real-time decision 𝒙 𝑡 at time 𝑡. Define the stopping time 𝜏 𝜋 ≤ 𝑇 as the minimum between 𝑇 and the smallest time 𝑡 such that there exists 𝑚 ∈ M with 𝜏 𝜋 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) + ḡ > 𝑇𝐺 𝑘 . In other words, 𝜏 𝜋 refers to the first time the violation of one resource constraint happens. We bound the averaged gap between 𝑇 and 𝜏 𝜋 as follows.\nProposition 2. Suppose that Assumption A.1 holds, using a constant step-size 𝜂 > 0 in Algorithm 1 yields\nE [𝑇 -𝜏 𝜋 ] ≤ ḡ 𝐺 + 𝐶 ℎ + ∥∇ℎ(𝜆 1 )∥ ∞ 𝜂𝐺 .(27)\nProof. According to Step. 9 in Algorithm 1 we have\n∇𝜇 𝑘,𝑡 = -𝑔 𝑘 (𝒙 𝑡 ) + 𝐺 𝑘 (1(𝝁 𝑘 ≥ 0) + 𝛼 𝑘 1(𝝁 𝑘 < 0)) , ≤ -𝑔 𝑘 (𝒙 𝑡 ) + 𝐺 𝑘 , ∀𝑘 ∈ [𝑚].(28)\nAssume that the stopping time 𝜏 𝜋 is activated due to the violation of constraint on 𝑘-th channel, we have\n𝜏 𝜋 ∑︁ 𝑡 =1 ∇𝜇 𝑘,𝑡 ≤𝐺 𝑘 𝜏 𝜋 - 𝜏 𝜋 ∑︁ 𝑡 =1 𝑔 𝑘 (𝒙 𝑡 ) ≤ 𝐺 𝑘 𝜏 𝜋 -𝑇𝐺 𝑘 + ḡ,(29)\nwhich leads to\n𝑇 -𝜏 𝜋 ≤ 1 𝐺 𝑘 ḡ - 𝜏 𝜋 ∑︁ 𝑡 =1 ∇𝜇 𝑘,𝑡 .(30)\nAccording to Proposition 6 in [25], the gradients of mirror descent satisfy ∇ℎ 𝑘 (𝜇 𝑡 +1 𝑘 ) ≥ ∇ℎ 𝑘 (𝜇 𝑡 𝑘 ) -𝜂∇𝜇 𝑡 𝑘,𝑡 , ∀𝑡 ≤ 𝜏 𝜋 , such that\n-𝜏 𝜋 𝑡 =1 ∇𝜇 𝑘,𝑡 ≤ 1 𝜂 ∇ℎ 𝑘 (𝜇 𝜏 𝜋 +1 𝑘 ) -∇ℎ 𝑘 (𝜇 1 𝑘 ) .\nCombing with the inequality in (30), we obtain\nE [𝑇 -𝜏 𝜋 ] ≤ ḡ 𝐺 𝑘 + E ∇ℎ 𝑘 (𝜇 𝜏 𝜋 +1 𝑘 ) -∇ℎ 𝑘 (𝜇 1 𝑘 ) 𝜂𝐺 𝑘 (31) ≤ ḡ 𝐺 + 𝐶 ℎ + ∥∇ℎ(𝜆 1 )∥ ∞ 𝜂𝐺 ,(32)\nas required. ■ Let us denote the random variable 𝛾 𝑡 to be the type of the request at period 𝑡, which can determine the sample of the request. 𝐺 𝑘 ([𝝁 𝑘 ] + -𝛼 𝑘 [-𝝁 𝑘 ] + ) , (34) where 𝜑 (𝝁) = 𝑓 * (𝝁) -𝝁 𝑇 𝑔(𝒙 𝑡 ). Considering that 𝒙 𝑡 is an optimal solution of 𝜑 (𝝁 𝑡 ) not of 𝜑 (𝝁), we have 𝑓 (𝒙 𝑡 ) -𝝁 𝑇 𝑔(𝒙 𝑡 ) ≤ 𝜑 (𝝁). Then, by taking 𝝁 = [0, 0, . . . , 0], and summing from one to 𝜏 𝜋 , we obtain \nCombining ( 37) and ( 38), we complete the proof of Proposition 3. ■ Before providing more details on the proof of regret bound, we introduce a new benchmark of problem P 0 as in [25] due to that problem P 0 may be infeasible due the presence of both lower and upper exposure constraints. Specifically, we define OPT(S, 𝜆).\nThis benchmark is an interpolate value between the expected optimal value of problem P 0 and a deterministic problem which replaces the varying utility values 𝑓 (𝒙 𝑡 ) and cost values 𝑔(𝒙 𝑡 ) with their expected values. For this benchmark, we have \nwhere the inequality uses the fact that OPT(S) ≤ 𝐷 (𝝁 |S) according to Proposition 1 in [25]and that OPT(S) ≤ 𝑇 f . Combining all findings together, we have \nwhere the second inequality comes from Proposition 2."
}
] | Integrated recommendation, which aims at jointly recommending heterogeneous items from different channels in a main feed, has been widely applied to various online platforms. Though attractive, integrated recommendation requires the ranking methods to migrate from conventional user-item models to the new user-channelitem paradigm in order to better capture users' preferences on both item and channel levels. Moreover, practical feed recommendation systems usually impose exposure constraints on different channels to ensure user experience. This leads to greater difficulty in the joint ranking of heterogeneous items. In this paper, we investigate the integrated recommendation task with exposure constraints in practical recommender systems. Our contribution is forth-fold. First, we formulate this task as a binary online linear programming problem and propose a two-layer framework named Multi-channel Integrated Recommendation with Exposure Constraints (MIREC) to obtain the optimal solution. Second, we propose an efficient online allocation algorithm to determine the optimal exposure assignment of different channels from a global view of all user requests over the entire time horizon. We prove that this algorithm reaches the optimal point under a regret bound of O ( √ 𝑇 ) with linear complexity. Third, we propose a series of collaborative models to determine the optimal layout of heterogeneous items at each user request. The joint modeling of user interests, cross-channel correlation, and page context in our models aligns more with the browsing nature of feed products than existing models. Finally, we conduct extensive experiments on both offline datasets and online A/B tests to verify the effectiveness of MIREC. The proposed framework has now been implemented on the homepage of Taobao to serve the main traffic, providing service to hundreds of millions of users towards billions of items every day. | Multi-channel Integrated Recommendation with Exposure Constraints | [
{
"figure_caption": "Figure 2 :2Figure 2: An overview of the MIREC framework.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "10:",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: Online system architecture.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "CTR on different positions.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: Online Performance Analysis.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Proposition 3 .3Using a constant step-size rule 𝜂 > 0 for 𝑡 > 1 in Algorithm 1, it holdsE 𝜏 𝜋 𝐷 ( μ𝜏 𝜋 ) -𝜏 𝜋 ∑︁ 𝑡 =1 𝑓 (𝒙 𝑡 ) ≤ 2( ḡ2 + Ḡ2 ) 𝜃 𝜂E[𝜏 𝜋 ] + 1 𝜂 𝑉 ℎ (𝝁, 𝝁 1 ), (33)whereμ𝜏 𝜋 = 𝜏𝜋 𝑡 =1 𝝁 𝑡 𝜏 𝜋 .Proof. According to the definition of ∇𝝁 𝑡 and the subgradient inequality, we have(∇𝝁 𝑡 ) 𝑇 (𝝁 𝑡 -𝝁) ≥ 𝐷 (𝝁 𝑡 ) -𝐷 (𝝁) ≥ 𝐷 (𝝁 𝑡 ) -E 𝛾 𝑡 [𝜑 (𝝁)] + ∑︁ 𝑘 ∈ [𝐾 ]",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "F(𝒙 𝑡 , 𝜆) = (1 -𝜆)𝑓 (𝒙 𝑡 ) + 𝜆E S [𝑓 (𝒙 𝑡 )] G(𝒙 𝑡 , 𝜆) = (1 -𝜆)𝑔(𝒙 𝑡 ) + 𝜆E S [𝑔(𝒙 𝑡 )],where 𝜆 ∈ [0, 1] is the interpolation parameter. We defineOPT(S, 𝜆) = E S 𝑇 max 𝒙 𝑡 ,𝑡 ∈ [𝑇 ] 𝑇 𝑡 =1 𝐹 (𝒙 𝑡 , 𝜆) s.t. 𝑇 𝛼 ⊙ 𝐺 ≤ 𝑇 𝑡 =1 𝐺 (𝒙 𝑡 , 𝜆) ≤ 𝑇𝐺whereS 𝑇 := S × • • • × S isa product distribution of length 𝑇 . Now we give the definition of the new benchmark as OPT(S) := max 𝜆 ∈ [0,1]",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "ES [OPT(S)]= 𝜏 𝜋 𝑇 E S [OPT(S)] + 𝑇 -𝜏 𝜋 𝑇 E S [OPT(S)] ≤ 𝜏 𝜋 D (𝝁 𝜏 𝜋 |S) + (𝑇 -𝜏 𝜋 ) f ,",
"figure_data": "",
"figure_id": "fig_7",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "(Regret (𝜋 |S) = E S [OPT(S) -𝑅(𝜋 |S)](41a)≤ E S 𝜏 𝜋 D (𝝁 𝜏 𝜋 |S) + (𝑇 -𝜏 𝜋 ) f -𝜏 𝜋 ∑︁ 𝑡 =1 𝑓 (𝒙 𝑡 ) (41b) = E S 𝜏 𝜋 D (𝝁 𝜏 𝜋 |S) -𝜏 𝜋 ∑︁ 𝑡 =1 𝑓 (𝒙 𝑡 ) + E S [𝑇 -𝜏 𝜋 ] f (41c) ≤ 2( ḡ2 + Ḡ2 ) 𝜃 𝜂E[𝜏 𝜋 ] + 1 𝜂 𝑉 ℎ (𝝁, 𝝁 1 ) + ḡ 𝐺 + 𝐶 ℎ + ∥∇ℎ(𝜆 1 )∥ ∞ 𝜂𝐺 ,(41d)where the first inequality is from(40) and the second inequality is from Proposition 2 and Proposition 3. Therefore, the constants inTheorem 1 are 𝐶 1 = ḡ 𝐺 , 𝐶 2 = 2( C2 + b2 ) 𝜃 𝜂, and 𝐶 3 = 𝑉 ℎ (𝝁, 𝝁 1 ) + ḡ 𝐺 + 𝐶 ℎ + ∥ ∇ℎ (𝜆 1 ) ∥ ∞ 𝐺, respectively. Moreover, recall Remark 1, we choose ℎ(𝜆) := 1 2 ∥𝜆∥ 2 and dual iterates are bounded. Hence, we complete the proof of Theorem 1. ■B PROOF OF COST FEASIBILITYProposition 1 shows that a solution obtained using Algorithm 1 can not overspend, but may underspend. Based on the definition of subgradient ∇𝜇 𝑡 𝑘 , we have ∇ℎ 𝑘 (𝜇 1 ) -∇ℎ 𝑘 (𝜇 𝜏 𝜋 +1 ) 𝐺 𝑘 (1(𝜇 𝑘 ≥ 0) +𝛼 𝑘 1(𝜇 𝑘 < 0)) -𝑔 𝑘 (𝒙 𝑡 )) (42) Now, given that 1(𝜇 ≥ 0) + 𝛼 𝑘 1(𝜇 < 0) ≥ 𝛼 𝑘 for any 𝜇 ∈ R and that 𝜏 𝜋 ≤ 𝑇 by definition, we have𝜏 𝐴 ∑︁ 𝑡 =1 𝐺 𝑘 (1(𝜇 𝑘 ≥ 0) +𝛼 𝑘 1(𝜇 𝑘 < 0)) + (𝑇 -𝜏 𝐴 )𝛼 𝑘 𝑏 𝑘 ≥𝑇 𝛼 𝑘 𝑏 𝑘 . (43) Combining (42) and (43) and taking expectation, we get 𝑇 𝛼 𝑘 𝐺 𝑘 -E[",
"figure_data": "",
"figure_id": "fig_8",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "",
"figure_data": "Inter-channel Encoding Layer Interest-Aware Evolution LayerLinear Linear V KLinear Linear V K MHSALinear Linear V KMLP & Softmax Weighted Sum QQPCR EmbsConcate & Flatten MLP & Sigmoid (} Concate & MLP Target Item Prev Context Cand Context PCR Embs PCR Scores (c) CAR Model (b) PCR CAR Point-wise Scores Context-aware Scores | ) ( ) ( max{ arg t t t t t I x g x f x Optimal Item Selection Ranking-layerEncoding Layer Intra-channelMHSAMHSAMHSALinear Linear V KConcate&LinearConstraints Update Allocation-layerDuals Update Real-time Duals...Real-timeMulti-ChannelFinalExposuresCandidate ItemsItem ListCandidate Seq Channel-1Candidate Seq Channel-2Candidate Seq ... Channel-N Embedding Layer Behavior Seq UserItem TargetInfo ChannelProfile UserStreaming Data CenterCH-1 Upstream RSs ... CH-NReportRequestMHSAMulti-head Self-attentionScaled Dot-productEmbedding VectorWeightsEmbedding CrossingDot ProductConcateUser Device",
"figure_id": "tab_0",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": ") 4.3.2 Context-Aware Refinement Model. In this section, we propose the CAR model to refine the point-wise utility scores estimated by PCR into context-aware utility scores. Given a candidate item set 𝐼 𝑐𝑎𝑛𝑑 with size 𝑁 , the aim of CAR is to optimally choose 𝐾 items from 𝐼 𝑐𝑎𝑛𝑑 and allocate them to the 𝐾 slots in a page based on the learning results from PCR.",
"figure_data": "",
"figure_id": "tab_2",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Comparison of ranking performance (bold: best; underline: runner-up).",
"figure_data": "DatasetMethodAUCLogloss NDCG@20 NDCG@30DIN0.68310.59220.54030.6535DIEN0.68420.59090.54080.6537DLCM0.68720.58980.55820.6698MicroVideo 3PRM0.69790.58720.55910.6708PEAR0.70210.58210.56320.6745Ours0.7084 0.57870.56670.6826DIN0.76810.49820.52030.6481DIEN0.76920.49710.52020.6479DLCM0.76990.49650.52090.6482TaobaoPRM0.77220.49410.52110.6489PEAR0.77480.49190.52320.6511STAR0.77380.49310.52190.6492DHANR 0.77530.49230.52430.6513Ours0.7791 0.48990.52750.6545",
"figure_id": "tab_3",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Ablation study of the ranking components.",
"figure_data": "AUCLogloss NDCG@20 NDCG@30PCR *0.77580.49330.52220.6511PCR †0.77610.49320.52460.6513PCR w/o IntraCE0.77630.49280.52470.6516PCR w/o InterCE0.77560.49350.52390.6509PCR0.77780.49130.52620.6531PCR+CAR (propsed) 0.77910.48990.52750.6545",
"figure_id": "tab_4",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Joint performance of allocation and ranking (bold: best; dagger: baseline).",
"figure_data": "Exposure completenessExp. Settings MethodCTRCTR LiftCH1CH2CH3CH4Fixed0.44% 1.35% 0.20% 0.50% 5.54% †-WPO0.15% 0.15% 0.13% 0.30% 6.09%+9.93%Setting-1HCA2E 0.24% 0.95% 0.33% 0.10% 6.34%+14.44%Ours0.02% 0.65% 0.67% 0.20% 6.56% * +18.41% *Fixed0.10% 0.53% 0.10% 0.20% 5.91% †-WPO0.16% 0.27% 1.40% 0.20% 6.28%+6.26%Setting-2HCA2E 0.19% 0.53% 0.30% 0.40% 6.53%+10.49%Ours0.17% 0.53% 0.40% 0.20% 6.76% * +14.38% *",
"figure_id": "tab_5",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "Results of online A/B tests.",
"figure_data": "CLICKCTRGMV Stay TimeOurs vs Fixed+4.02% +2.15% +1.98%+2.01%Ours vs Baseline +3.00% +1.75% +1.56%",
"figure_id": "tab_6",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "𝛾 𝑡 [𝑓 (𝒙 𝑡 )] ≥𝜏 𝜋 𝐷 ( μ𝜏 𝜋 ) -𝜏𝜋 𝑡 =1 𝝁 𝑡 𝜏 𝜋and the inequality is based on the fact that the dual function is convex. In this paper, we adopt 𝜃 -strongly convex function as the relation function ℎ(•) in mirror descents. According to Step. 2 of Proposition 8 in[25], we have 𝛾 𝑡 [𝑓 (𝒙 𝑡 )] = E",
"figure_data": "According to Step. 3 of Proposition 8 in [25], we haveE𝜏 𝜋 ∑︁E 𝜏 𝜋 ∑︁𝑡 =1𝑡 =1𝜏 𝜋𝜏 𝜋𝜏 𝜋∑︁ 𝑡 =1∑︁ 𝑡 =1 E 𝜏 𝜋 𝐷 (𝝁 𝑡 ) -∑︁ 𝑡 =1∑︁E 𝛾 𝑡 [𝑓 (𝒙 𝑡 ))],(35)𝑡 =1where μ𝜏 𝜋 =E𝑡 =1 𝜏 𝜋 ∑︁(∇𝝁 𝑡 ) 𝑇 (𝝁 𝑡 -𝝁) ≤𝜃 2( ḡ2 + Ḡ2 )𝜂E[𝜏 𝐴 ] +𝜂 𝑉 ℎ (𝜆, 𝜆 1 ). (36)Combining (36) with (35), we get(37)",
"figure_id": "tab_7",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "≤ ∇ℎ 𝑘 (𝜇 1 ) -E[∇ℎ 𝑘 (𝜇 𝜏 𝜋 +1 )] 𝜂 + E[𝑇 -𝜏 𝐴 ]𝛼 𝑘 𝑏 𝑘",
"figure_data": "𝜏 𝜋∑︁𝑔 𝑘 (𝒙 𝑡 )](44a)𝑡 =1(44b)≤∥∇ℎ(𝜆 1 )∥ ∞ + 𝐶 ℎ𝐺 + 𝛼 𝑘 𝑏 𝑘+𝛼 𝑘 𝑏 𝑘ḡ,𝜂𝐺𝐺",
"figure_id": "tab_8",
"figure_label": "",
"figure_type": "table"
}
] | Yue Xu; Qijie Shen; Dimin Wang; Hao Chen; Lixiang Lai; Tao Zhuang | [
{
"authors": "Qingyao Ai; Keping Bi; Jiafeng Guo; Bruce Croft",
"journal": "",
"ref_id": "b0",
"title": "Learning a deep listwise context model for ranking refinement",
"year": "2018"
},
{
"authors": "Ashwinkumar Badanidiyuru; Robert Kleinberg; Aleksandrs Slivkins",
"journal": "Journal of the ACM (JACM)",
"ref_id": "b1",
"title": "Bandits with knapsacks",
"year": "2018"
},
{
"authors": "Haihao Santiago R Balseiro; Vahab Lu; Mirrokni",
"journal": "Operations Research",
"ref_id": "b2",
"title": "The best of many worlds: Dual mirror descent for online allocation problems",
"year": "2022"
},
{
"authors": "Irwan Bello; Sayali Kulkarni; Sagar Jain; Craig Boutilier; Ed Chi; Elad Eban; Xiyang Luo; Alan Mackey; Ofer Meshi",
"journal": "",
"ref_id": "b3",
"title": "Seq2slate: Re-ranking and slate optimization with rnns",
"year": "2018"
},
{
"authors": "Dimitri P Bertsekas",
"journal": "Journal of the Operational Research Society",
"ref_id": "b4",
"title": "Nonlinear programming",
"year": "1997"
},
{
"authors": "Dagui Chen; Qi Yan; Chunjie Chen; Zhenzhe Zheng; Yangsu Liu; Zhenjia Ma; Chuan Yu; Jian Xu; Bo Zheng",
"journal": "",
"ref_id": "b5",
"title": "Hierarchically constrained adaptive ad exposure in feeds",
"year": "2022"
},
{
"authors": "Qiwei Chen; Yue Xu; Changhua Pei; Shanshan Lv; Tao Zhuang; Junfeng Ge",
"journal": "",
"ref_id": "b6",
"title": "Efficient long sequential user data modeling for click-through rate prediction",
"year": "2022"
},
{
"authors": "Xusong Chen; Dong Liu; Zheng-Jun Zha; Wengang Zhou; Zhiwei Xiong; Yan Li",
"journal": "MM",
"ref_id": "b7",
"title": "Temporal hierarchical attention at category-and item-level for micro-video click-through prediction",
"year": "2018"
},
{
"authors": "Yuwei Chen; Zengde Deng; Yinzhi Zhou; Zaiyi Chen; Yujie Chen; Haoyuan Hu",
"journal": "IEEE",
"ref_id": "b8",
"title": "An online algorithm for chance constrained resource allocation",
"year": "2023"
},
{
"authors": "Heng-Tze Cheng; Levent Koc; Jeremiah Harmsen; Tal Shaked; Tushar Chandra; Hrishi Aradhye; Glen Anderson; Greg Corrado; Wei Chai; Mustafa Ispir",
"journal": "",
"ref_id": "b9",
"title": "Wide & deep learning for recommender systems",
"year": "2016"
},
{
"authors": "Yufei Feng; Binbin Hu; Yu Gong; Fei Sun; Qingwen Liu; Wenwu Ou",
"journal": "",
"ref_id": "b10",
"title": "GRN: Generative rerank network for context-wise recommendation",
"year": "2021"
},
{
"authors": "Jyotirmoy Gope; Sanjay Kumar; Jain ",
"journal": "IEEE",
"ref_id": "b11",
"title": "A survey on solving cold start problem in recommender systems",
"year": "2017"
},
{
"authors": "Qi Hao; Tianze Luo; Guangda Huzhang",
"journal": "",
"ref_id": "b12",
"title": "Re-ranking with constraints on diversified exposures for homepage recommender system",
"year": "2021"
},
{
"authors": "Xiaotian Hao; Zhaoqing Peng; Yi Ma; Guan Wang; Junqi Jin; Jianye Hao; Shan Chen; Rongquan Bai; Mingzhou Xie; Miao Xu",
"journal": "",
"ref_id": "b13",
"title": "Dynamic knapsack optimization towards efficient multi-channel sequential advertising",
"year": "2020"
},
{
"authors": "Elad Hazan",
"journal": "Foundations and Trends® in Optimization",
"ref_id": "b14",
"title": "Introduction to online convex optimization",
"year": "2016"
},
{
"authors": "Jinhong Huang; Yang Li; Shan Sun; Bufeng Zhang; Jin Huang",
"journal": "",
"ref_id": "b15",
"title": "Personalized flight itinerary ranking at fliggy",
"year": "2020"
},
{
"authors": "Yanhua Huang; Weikun Wang; Lei Zhang; Ruiwen Xu",
"journal": "",
"ref_id": "b16",
"title": "Sliding spectrum decomposition for diversified recommendation",
"year": "2021"
},
{
"authors": "Nicole Immorlica; Karthik Abinav Sankararaman; Robert Schapire; Aleksandrs Slivkins",
"journal": "",
"ref_id": "b17",
"title": "Adversarial bandits with knapsacks",
"year": "2019"
},
{
"authors": "Kalervo Järvelin; Jaana Kekäläinen",
"journal": "",
"ref_id": "b18",
"title": "IR evaluation methods for retrieving highly relevant documents",
"year": "2017"
},
{
"authors": "Yi Li; Jieming Zhu; Weiwen Liu; Liangcai Su; Guohao Cai; Qi Zhang; Ruiming Tang; Xi Xiao; Xiuqiang He",
"journal": "",
"ref_id": "b19",
"title": "Pear: Personalized re-ranking with contextualized transformer for recommendation",
"year": "2022"
},
{
"authors": "Guogang Liao; Ze Wang; Xiaoxu Wu; Xiaowen Shi; Chuheng Zhang; Yongkang Wang; Xingxing Wang; Dong Wang",
"journal": "",
"ref_id": "b20",
"title": "Cross DQN: Cross deep Q network for ads allocation in feed",
"year": "2022"
},
{
"authors": "Zhuoyi Lin; Sheng Zang; Rundong Wang; Zhu Sun; Chi Xu; Chee-Keong Kwoh",
"journal": "",
"ref_id": "b21",
"title": "Attention over self-attention: Intention-aware re-ranking with dynamic transformer encoders for recommendation",
"year": "2022"
},
{
"authors": "Zihan Lin; Hui Wang; Jingshu Mao; Wayne Xin Zhao; Cheng Wang; Peng Jiang; Ji-Rong Wen",
"journal": "",
"ref_id": "b22",
"title": "Feature-aware diversified re-ranking with disentangled representations for relevant recommendation",
"year": "2022"
},
{
"authors": "Weiwen Liu; Yunjia Xi; Jiarui Qin; Fei Sun; Bo Chen; Weinan Zhang; Rui Zhang; Ruiming Tang",
"journal": "",
"ref_id": "b23",
"title": "Neural re-ranking in multi-stage recommender systems: A review",
"year": "2020"
},
{
"authors": "Alfonso Lobos; Paul Grigas; Zheng Wen",
"journal": "",
"ref_id": "b24",
"title": "Joint online learning and decisionmaking via dual mirror descent",
"year": "2021"
},
{
"authors": "Xingyu Lu; Qintong Wu; Wenliang Zhong",
"journal": "PMLR",
"ref_id": "b25",
"title": "Multi-slots online matching with high entropy",
"year": "2022"
},
{
"authors": "Changhua Pei; Yi Zhang; Yongfeng Zhang; Fei Sun; Xiao Lin; Hanxiao Sun; Jian Wu; Peng Jiang; Junfeng Ge; Wenwu Ou",
"journal": "",
"ref_id": "b26",
"title": "Personalized re-ranking for recommendation",
"year": "2019"
},
{
"authors": "Qi Pi; Weijie Bian; Guorui Zhou; Xiaoqiang Zhu; Kun Gai",
"journal": "",
"ref_id": "b27",
"title": "Practice on long sequential user behavior modeling for click-through rate prediction",
"year": "2019"
},
{
"authors": "Xufeng Qian; Yue Xu; Fuyu Lv; Shengyu Zhang; Ziwen Jiang; Qingwen Liu; Xiaoyi Zeng; Tat-Seng Chua; Fei Wu",
"journal": "",
"ref_id": "b28",
"title": "Intelligent request strategy design in recommender system",
"year": "2022"
},
{
"authors": "Xiang-Rong Sheng; Liqin Zhao; Guorui Zhou; Xinyao Ding; Binding Dai; Qiang Luo; Siran Yang; Jingshan Lv; Chi Zhang; Hongbo Deng",
"journal": "",
"ref_id": "b29",
"title": "One model to serve all: Star topology adaptive recommender for multi-domain ctr prediction",
"year": "2021"
},
{
"authors": "Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Łukasz Kaiser; Illia Polosukhin",
"journal": "Advances in neural information processing systems",
"ref_id": "b30",
"title": "Attention is all you need",
"year": "2017"
},
{
"authors": "Fan Wang; Xiaomin Fang; Lihang Liu; Yaxue Chen; Jiucheng Tao; Zhiming Peng; Cihang Jin; Hao Tian",
"journal": "",
"ref_id": "b31",
"title": "Sequential evaluation and generation framework for combinatorial recommender system",
"year": "2019"
},
{
"authors": "Ruobing Xie; Shaoliang Zhang; Rui Wang; Feng Xia; Leyu Lin",
"journal": "",
"ref_id": "b32",
"title": "Hierarchical reinforcement learning for integrated recommendation",
"year": "2021"
},
{
"authors": "Biao Yuan; Zengde Deng; Na Geng; Yujie Chen; Haoyuan Hu",
"journal": "INFORMS Journal on Applied Analytics",
"ref_id": "b33",
"title": "Practice summary: Cainiao optimizes the fulfillment routes of parcels",
"year": "2023"
},
{
"authors": "Jianjun Yuan; Andrew Lamperski",
"journal": "",
"ref_id": "b34",
"title": "Online convex optimization for cumulative constraints",
"year": "2018"
},
{
"authors": "Weiru Zhang; Chao Wei; Xiaonan Meng; Yi Hu; Hao Wang",
"journal": "",
"ref_id": "b35",
"title": "The whole-page optimization via dynamic ad allocation",
"year": "2018"
},
{
"authors": "Xiangyu Zhao; Changsheng Gu; Haoshenglun Zhang; Xiwang Yang; Xiaobing Liu; Jiliang Tang; Hui Liu",
"journal": "",
"ref_id": "b36",
"title": "Dear: Deep reinforcement learning for online advertising impression in recommender systems",
"year": "2021"
},
{
"authors": "Guorui Zhou; Na Mou; Ying Fan; Qi Pi; Weijie Bian; Chang Zhou; Xiaoqiang Zhu; Kun Gai",
"journal": "",
"ref_id": "b37",
"title": "Deep interest evolution network for click-through rate prediction",
"year": "2019"
},
{
"authors": "Guorui Zhou; Xiaoqiang Zhu; Chenru Song; Ying Fan; Han Zhu; Xiao Ma; Yanghui Yan; Junqi Jin; Han Li; Kun Gai",
"journal": "",
"ref_id": "b38",
"title": "Deep interest network for click-through rate prediction",
"year": "2018"
},
{
"authors": "Yu-Hang Zhou; Peng Hu; Chen Liang; Huan Xu; Guangda Huzhang; Yinfu Feng; Qing Da; Xinshang Wang; An-Xiang Zeng",
"journal": "",
"ref_id": "b39",
"title": "A primal-dual online algorithm for online matching problem in dynamic environments",
"year": "2021"
},
{
"authors": "Tao Zhuang; Wenwu Ou; Zhirong Wang",
"journal": "",
"ref_id": "b40",
"title": "Globally optimized mutual influence aware ranking in e-commerce search",
"year": "2018"
}
] | [
{
"formula_coordinates": [
3,
94.41,
394.83,
199.64,
19.41
],
"formula_id": "formula_0",
"formula_text": "X 𝑡 = 𝑖 ∈𝐼 𝑡 𝑥 𝑡,𝑛,𝑖 = 1, ∀𝑡 ∈ T , ∀𝑛 ∈ N 𝑛 𝑥 𝑡,𝑛,𝑖 ≤ 1, ∀𝑡 ∈ T , ∀𝑖 ∈ 𝐼,(1)"
},
{
"formula_coordinates": [
3,
83.4,
649.45,
210.64,
17.89
],
"formula_id": "formula_1",
"formula_text": "P 0 : OPT(S) = max 𝒙 𝑡 ∈X 𝑡 ∑︁ 𝑇 𝑡 =1 𝑓 (𝒙 𝑡 )(2)"
},
{
"formula_coordinates": [
3,
87.11,
672.85,
206.94,
15.89
],
"formula_id": "formula_2",
"formula_text": "s.t. 𝐶 1 : ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≤ 𝐺 max 𝑚,𝑡ℎ 𝑁 (S), ∀𝑚 ∈ M,(3)"
},
{
"formula_coordinates": [
3,
106.4,
692.7,
187.64,
15.89
],
"formula_id": "formula_3",
"formula_text": "𝐶 2 : ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≥ 𝐺 min 𝑚,𝑡ℎ 𝑁 (S), ∀𝑚 ∈ M,(4)"
},
{
"formula_coordinates": [
4,
119.42,
361.16,
174.62,
39.44
],
"formula_id": "formula_4",
"formula_text": "+ ∑︁ 𝑚 𝜆 𝑚 ∑︁ 𝑡 𝑔 𝑚 (𝒙 𝑡 ) -𝐺 min 𝑚 .(5)"
},
{
"formula_coordinates": [
4,
62.85,
509.68,
231.19,
78.43
],
"formula_id": "formula_5",
"formula_text": "min 𝝁 𝐷 (𝝁) = ∑︁ 𝑡 𝑓 (𝒙 𝑡 ) - ∑︁ 𝑚 [𝜇 𝑚 ] + ∑︁ 𝑡 𝑔 𝑚 (𝒙 𝑡 ) -𝐺 max 𝑚 (6a) + ∑︁ 𝑚 [-𝜇 𝑚 ] + ∑︁ 𝑡 𝑔 𝑚 (𝒙 𝑡 ) -𝐺 min 𝑚 (6b) = ∑︁ 𝑡 𝑓 (𝒙 𝑡 ) - ∑︁ 𝑚 𝜇 𝑚 𝑔 𝑚 (𝒙 𝑡 ) + ∑︁ 𝑚 [𝜇 𝑚 ] + 𝐺 max 𝑚 -[-𝜇 𝑚 ] + 𝐺 min 𝑚 ,(6c)"
},
{
"formula_coordinates": [
4,
79.05,
596.72,
77.01,
8.79
],
"formula_id": "formula_6",
"formula_text": "[𝜇 𝑚 ] + = max{𝜇 𝑚 , 0}."
},
{
"formula_coordinates": [
4,
103.58,
695.94,
190.46,
15.08
],
"formula_id": "formula_7",
"formula_text": "P 1 : x𝑡 = arg max 𝒙 𝑡 ∈X 𝑓 (𝒙 𝑡 ) -𝝁 𝑇 𝑡 𝑔(𝒙 𝑡 ) .(7)"
},
{
"formula_coordinates": [
4,
373.66,
400.35,
184.54,
10.98
],
"formula_id": "formula_8",
"formula_text": "𝐺 max 𝑚,𝑡 +1 = 𝐺 max 𝑚,𝑡 -𝑔 𝑚 (𝒙 𝑡 ), ∀𝑚 ∈ M.(8)"
},
{
"formula_coordinates": [
4,
322.53,
442.11,
229.33,
12.25
],
"formula_id": "formula_9",
"formula_text": "∇𝜇 𝑚,𝑡 = -𝑔 𝑚 (𝒙 𝑡 ) +𝐺 max 𝑚,𝑡 +1 • 1(𝜇 𝑚,𝑡 ≥ 0) +𝐺 min 𝑚 • 1(𝜇 𝑚,𝑡 ≤ 0), ("
},
{
"formula_coordinates": [
4,
356.03,
495.43,
202.17,
17.62
],
"formula_id": "formula_10",
"formula_text": "𝜇 𝑚,𝑡 +1 = arg min 𝜇 𝑚 ∈R 𝜇 𝑚 ∇𝜇 𝑚,𝑡 + 1 𝜂 𝑉 ℎ (𝜇 𝑚 , 𝜇 𝑚,𝑡 ),(10)"
},
{
"formula_coordinates": [
4,
398.84,
588.64,
159.36,
8.79
],
"formula_id": "formula_11",
"formula_text": "𝝁 𝑡 +1 = [𝝁 𝑡 -𝜂∇𝝁 𝑡 ] + ,(11)"
},
{
"formula_coordinates": [
5,
90.65,
378.9,
203.4,
9.43
],
"formula_id": "formula_12",
"formula_text": "Regret(𝜋 |S) = sup E S [OPT(S) -𝑅(𝜋 |S)] ,(12)"
},
{
"formula_coordinates": [
5,
53.8,
675.97,
212.23,
11.72
],
"formula_id": "formula_13",
"formula_text": "channel 𝑚 ∈ M that 𝐺 min 𝑚 -E 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≤ 𝐶 4 + 𝐶 5"
},
{
"formula_coordinates": [
5,
367.74,
197.82,
140.75,
15.08
],
"formula_id": "formula_14",
"formula_text": "P 1 : x𝑡 = arg max 𝒙 𝑡 ∈X 𝑓 (𝒙 𝑡 ) -𝝁 𝑇 𝑡 𝑔(𝒙 𝑡 ) ."
},
{
"formula_coordinates": [
6,
86.07,
135.89,
207.98,
10.42
],
"formula_id": "formula_15",
"formula_text": "𝑉 𝑚 𝑙 = [ℎ𝑒𝑎𝑑 1 , ℎ𝑒𝑎𝑑 2 , ..., ℎ𝑒𝑎𝑑 ℎ ]𝑊 𝑂 ,(13a)"
},
{
"formula_coordinates": [
6,
78.94,
152.44,
215.11,
23.78
],
"formula_id": "formula_16",
"formula_text": "ℎ𝑒𝑎𝑑 𝑖 = Softmax (𝐸 𝑏 𝑊 𝑄 )(𝐸 𝑏 𝑊 𝐾 ) 𝑇 √︁ 𝑑 ℎ /ℎ (𝐸 𝑏 𝑊 𝑉 ) ,(13b)"
},
{
"formula_coordinates": [
6,
83.55,
433.83,
207.08,
14.69
],
"formula_id": "formula_17",
"formula_text": "𝑉 𝑈 = 𝑓 (𝐸 𝑏 ; 𝐸 𝑖 ) = ∑︁ 𝐵 𝑖=1 𝐴(𝑏 𝑖 , 𝐸 𝑖 )𝑏 𝑖 = ∑︁ 𝐵 𝑖=1 𝑤 𝑖 𝑏 𝑖 , (14"
},
{
"formula_coordinates": [
6,
290.62,
439.44,
3.42,
4.09
],
"formula_id": "formula_18",
"formula_text": ")"
},
{
"formula_coordinates": [
6,
59.33,
535.71,
234.72,
32.96
],
"formula_id": "formula_19",
"formula_text": "𝐻 𝑖 𝑠 = Softmax (𝑉 𝑚 𝑙 𝑊 𝑞 ) [𝑉 𝑈 𝑊 𝑘1 , 𝑉 𝑚 𝑙 𝑊 𝑘2 ] 𝑇 √︁ 𝑑 ℎ [𝑉 𝑈 𝑊 𝑣1 , 𝑉 𝑚 𝑙 𝑊 𝑣2 ],(15)"
},
{
"formula_coordinates": [
6,
114.79,
700.48,
179.25,
9.16
],
"formula_id": "formula_20",
"formula_text": "𝑊 𝐶𝐻 = Softmax(MLP[𝐻 𝑚 𝑠 , 𝐸 𝑖 ]),(16)"
},
{
"formula_coordinates": [
6,
386.35,
179.89,
171.85,
9.16
],
"formula_id": "formula_21",
"formula_text": "𝑉 𝐿 = 𝑊 𝐶𝐻 • [ H𝑚 𝑠 ] 𝑇 , 𝑚 ∈ M,(17)"
},
{
"formula_coordinates": [
6,
371.35,
255.18,
183.43,
8.03
],
"formula_id": "formula_22",
"formula_text": "𝑌 𝑃𝐶𝑅 = Sigmoid(Concat(𝐸 𝑢 , 𝐸 𝑖 , 𝑉 𝐿 ))(18"
},
{
"formula_coordinates": [
6,
355.65,
476.43,
202.55,
7.93
],
"formula_id": "formula_23",
"formula_text": "𝐻 𝑝𝑡 = Concat(ℎ 𝑝𝑟𝑒 ⊕ 𝑒 𝑖 , ℎ 𝑝𝑟𝑒 ⊗ 𝑒 𝑖 , ℎ 𝑝𝑟𝑒 ⊖ 𝑒 𝑖 ),(19)"
},
{
"formula_coordinates": [
6,
327.06,
605.41,
231.15,
22.71
],
"formula_id": "formula_24",
"formula_text": "𝐻 all = Concat(𝐻 𝑝𝑡 , 𝐻 𝑐𝑡 , 𝐻 𝑣 𝑐𝑡 , 𝐻 𝑣 𝑝𝑟𝑒 , ℎ 𝑝𝑟𝑒 , ℎ 𝑐𝑎𝑛 , 𝐸 𝑢 , 𝑌 𝑃𝐶𝑅 ), (20a) 𝑌 𝐶𝐴𝑅 =𝜎 (MLP(𝐻 all )),(20b)"
},
{
"formula_coordinates": [
7,
73.99,
280.04,
220.06,
15.76
],
"formula_id": "formula_25",
"formula_text": "𝐽 = ∑︁ 𝑒 𝑡 ∈ D 𝑦 𝑒 𝑡 𝑢,𝑖 log ŷ𝑒 𝑡 𝑢,𝑖 + (1 -𝑦 𝑒 𝑡 𝑢,𝑖 ) log(1 -ŷ𝑒 𝑡 𝑢,𝑖 ) ,(21)"
},
{
"formula_coordinates": [
11,
83.16,
203.43,
210.88,
17.89
],
"formula_id": "formula_26",
"formula_text": "P 0 : OPT(S) = max 𝒙 𝑡 ∈X ∑︁ 𝑇 𝑡 =1 𝑓 (𝒙 𝑡 )(22)"
},
{
"formula_coordinates": [
11,
86.87,
226.84,
207.17,
15.89
],
"formula_id": "formula_27",
"formula_text": "s.t. 𝐶 1 : ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≤ 𝐺 max 𝑚,𝑡ℎ 𝑁 (S), ∀𝑚 ∈ M,(23)"
},
{
"formula_coordinates": [
11,
106.17,
246.68,
187.88,
15.89
],
"formula_id": "formula_28",
"formula_text": "𝐶 2 : ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≥ 𝐺 min 𝑚,𝑡ℎ 𝑁 (S), ∀𝑚 ∈ M.(24)"
},
{
"formula_coordinates": [
11,
86.25,
326.24,
207.8,
33.1
],
"formula_id": "formula_29",
"formula_text": "P 1 : OPT(S) = max 𝒙 𝑡 ∈X ∑︁ 𝑇 𝑡 =1 𝑓 (𝒙 𝑡 )(25) s"
},
{
"formula_coordinates": [
11,
92.63,
349.65,
201.41,
15.89
],
"formula_id": "formula_30",
"formula_text": ".t. 𝛼𝑇𝐺 𝑚 ≤ ∑︁ 𝑇 𝑡 =1 𝑔 𝑚 (𝒙 𝑡 ) ≤ 𝑇𝐺 𝑚 , ∀𝑚 ∈ M.(26)"
},
{
"formula_coordinates": [
11,
100.11,
439.94,
148.82,
9.18
],
"formula_id": "formula_31",
"formula_text": "𝜇 𝑡 satisfy E[||∇ℎ(𝜇 𝑡 )|| ∞ ] ≤ 𝐶 ℎ , ∀𝑡 ∈ [𝑇 ]."
},
{
"formula_coordinates": [
11,
107.58,
587.15,
186.47,
18.88
],
"formula_id": "formula_32",
"formula_text": "E [𝑇 -𝜏 𝜋 ] ≤ ḡ 𝐺 + 𝐶 ℎ + ∥∇ℎ(𝜆 1 )∥ ∞ 𝜂𝐺 .(27)"
},
{
"formula_coordinates": [
11,
74.11,
626.56,
219.94,
24.19
],
"formula_id": "formula_33",
"formula_text": "∇𝜇 𝑘,𝑡 = -𝑔 𝑘 (𝒙 𝑡 ) + 𝐺 𝑘 (1(𝝁 𝑘 ≥ 0) + 𝛼 𝑘 1(𝝁 𝑘 < 0)) , ≤ -𝑔 𝑘 (𝒙 𝑡 ) + 𝐺 𝑘 , ∀𝑘 ∈ [𝑚].(28)"
},
{
"formula_coordinates": [
11,
84.19,
681.37,
209.85,
26.33
],
"formula_id": "formula_34",
"formula_text": "𝜏 𝜋 ∑︁ 𝑡 =1 ∇𝜇 𝑘,𝑡 ≤𝐺 𝑘 𝜏 𝜋 - 𝜏 𝜋 ∑︁ 𝑡 =1 𝑔 𝑘 (𝒙 𝑡 ) ≤ 𝐺 𝑘 𝜏 𝜋 -𝑇𝐺 𝑘 + ḡ,(29)"
},
{
"formula_coordinates": [
11,
384.19,
101.83,
174.01,
26.33
],
"formula_id": "formula_35",
"formula_text": "𝑇 -𝜏 𝜋 ≤ 1 𝐺 𝑘 ḡ - 𝜏 𝜋 ∑︁ 𝑡 =1 ∇𝜇 𝑘,𝑡 .(30)"
},
{
"formula_coordinates": [
11,
317.96,
158.71,
160.14,
14.08
],
"formula_id": "formula_36",
"formula_text": "-𝜏 𝜋 𝑡 =1 ∇𝜇 𝑘,𝑡 ≤ 1 𝜂 ∇ℎ 𝑘 (𝜇 𝜏 𝜋 +1 𝑘 ) -∇ℎ 𝑘 (𝜇 1 𝑘 ) ."
},
{
"formula_coordinates": [
11,
351.57,
189.55,
206.63,
52.52
],
"formula_id": "formula_37",
"formula_text": "E [𝑇 -𝜏 𝜋 ] ≤ ḡ 𝐺 𝑘 + E ∇ℎ 𝑘 (𝜇 𝜏 𝜋 +1 𝑘 ) -∇ℎ 𝑘 (𝜇 1 𝑘 ) 𝜂𝐺 𝑘 (31) ≤ ḡ 𝐺 + 𝐶 ℎ + ∥∇ℎ(𝜆 1 )∥ ∞ 𝜂𝐺 ,(32)"
}
] |
10.1161/CIR.0000000000001039 | [
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b0",
"b1",
"b2",
"b3",
"b4",
"b5",
"b6",
"b7",
"b8",
"b9",
"b10",
"b11",
"b11",
"b12"
],
"table_ref": [],
"text": "Coronary artery disease (CAD) has long been the leading cause of death in the United States [1] and is rapidly becoming the top killer in other counties in the world [2,3]. CAD is characterized by the accumulation of cholesterol-rich plaque within the inner layers of the coronary artery wall. This plaque buildup leads to varying degrees of stenosis, or narrowing, in the arteries. As a consequence, blood flow to the myocardium, the heart muscle, becomes restricted, resulting in myocardial ischemia -a condition marked by insufficient oxygen and nutrient supply to the heart. As the blood supply-demand mismatch worsens, symptoms of CAD such as chest pain (angina), shortness of breath, and others may appear. Moreover, instability in the areas of stenosis can lead to acute occlusion of the coronary artery, resulting in a heart attack. Currently, the treatment for significant coronary artery stenosis involves percutaneous coronary intervention (PCI) or coronary artery bypass grafting (CABG) along with aggressive medical therapies [4].\nInvasive coronary angiography (ICA) indeed remains the gold standard for diagnosing CAD [5]. It is an established imaging technique that plays a crucial role in both the diagnosis and treatment of heart conditions, particularly CAD. During an ICA procedure, a thin, flexible tube called a catheter is inserted into an artery, usually in the groin or arm, and guiding it through the blood vessels to the heart. Once the catheter is in place, a special dye is injected into the coronary arteries, which makes them visible. These images support cardiologists identifying any blockages or narrowing in the coronary arteries. Depending on the findings, the cardiologists may use the same catheter to perform treatments such as angioplasty or stent placement to open blocked arteries and improve blood flow to the heart. While ICA is highly effective in providing detailed images of the coronary arteries, it is important to acknowledge the limitations associated with subjective visual assessment. Human interpretation of the angiograms can introduce variability and subjectivity into the diagnosis, potentially leading to unreliable assessments [6].\nThe coronary vascular tree is indeed complex and contains two major systems: the left coronary artery (LCA) and the right coronary artery (RCA) systems. The LCA is more clinically relevant, given it provides most of the blood supply to the left ventricle. LCA system and codominance are associated with modestly increased PCI in-hospital mortality in patients with stable CAD [7]. According to the 15-segments model defined by the American Heart Association [8], the LCA system can be further subdivided into three main coronary arteries: the left anterior descending (LAD) artery, the left circumflex (LCX) artery, and the left main artery (LMA), which act as the main blood suppliers of the myocardium. In normal coronary anatomy, the LMA bifurcates into two main branches: LAD and LCX. The LAD artery supplies blood to the front and part of the left side of the myocardium muscle while the LCX courses around the left side of the myocardium and supplies blood to the left atrium and part of the left ventricle [9]. The length of LAD varies between 10 to 13 cm and gives rise to the diagonal branches (D), which further contribute to the blood supply of the myocardium. Similarly, the length of the LCX artery varies from 5 to 8 cm and gives rise to obtuse marginal (OM) branches, extending its reach to specific areas of the left ventricle [10].\nHowever, it is challenging to identify individual coronary artery using ICAs because of the morphological similarity among different segments [11] and loss of partial high-frequency detail information due to the projection of the 3D vascular tree into a 2D plane during ICA imaging acquisition [12]. This loss of information limits the ability to precisely discern the intricate structures and boundaries of the coronary arteries, making their identification more challenging. In addition, the coronary arteries not only span over a long distance but also show similar semantic features with each other [12], making it challenging to associate them with the exact branches. This causes the confusion in judgement of whether a coronary segment belongs to main branch or side branch. ICA image understanding is a complex task due to several factors that degrade the visual quality of the images. Furthermore, ICA images suffer from various factors that degrade their visual quality. The contrast degradation occurs as the contrast dye dissipates, leading to reduced visibility of the coronary arteries. Additionally, spatial blurring and overlapping caused by nonvessel tissues and structures further obscure the vessel boundaries and hinder accurate interpretation.\nThe limitations of existing coronary artery semantic labeling methods that rely solely on position and imaging features are evident when processing complex coronary vasculature from different view angles of ICA images [13]. The morphological similarity among different types of arteries poses a challenge for pixel-intensity-based models to accurately discern each arterial segment and generate meaningful semantic segmentation. To address these challenges, we propose an innovative approach that leverages the concept of graphs to represent the coronary arteries and their connections. By converting the arteries into graph structures, we can incorporate non-traditional features, such as node degrees and graph-theory distances, in addition to the pixel-derived features, to enhance the accuracy of semantic segmentation. A key aspect of our method is the use of graph matching techniques, which aim to find the semantic correspondence between arteries in the labeled graphs. By comparing the graph structures and identifying similar patterns, we can establish meaningful associations between arterial segments and assign appropriate semantic labels.\nIn this paper, we propose a novel algorithm to perform coronary artery semantic labeling using ICA images. We propose an edge attention graph matching network (EAGMN) to build the semantic correspondence between coronary arterial segments from ICAs. The problem of semantic segmentation is converted into a problem that classifies the type of an unlabeled arterial segment by searching for a labelled arterial segment with the maximal similarity in a database of arterial graphs generated from a large number of ICAs. In detail, the individual graph of the coronary artery is generated according to coronary artery binary segmentation result. The node in the individual graph represents a segment of the coronary artery and the edge indicates the connectivity between arterial segments according to the physical connection of the vascular tree. Then an association graph is constructed from two individual graphs, where each vertex is built from two nodes in the individual graph, representing the node-to-node correspondence of two arterial segments from two individual graphs. Thus, the coronary artery semantic labeling task is converted into vertex classification task using the generated association graph. EAGMN incorporates an encoder module responsible for embedding vertex and edge features using a graph attention convolution network. This module extracts meaningful representations from the association graph by considering the interactions and relationships between the vertices and edges. Additionally, the EAGMN employs a decoder module that facilitates the readout of the feature representations generated by the encoder. By examining the positive vertices in the association graph, which indicate matched nodes from the individual graphs, the EAGMN accomplishes the semantic labeling task by assigning appropriate labels to the coronary arterial segments. The workflow of the proposed EAGMN for coronary artery semantic labeling is shown in Figure 1. 1) The utilization of graph matching for coronary artery semantic labeling: This paper introduces a novel approach that employs graph matching techniques to establish the node-to-node correspondence between labeled and unlabeled coronary arterial segments. This enables accurate and reliable semantic labeling of the coronary artery.\n2) The EAGMN utilizes an edge attention mechanism to dynamically aggregate features of adjacency edges, which in turn update vertex features for vertex classification using the association graph.\n3) We demonstrate the robustness of the proposed model by conducting experiments using corrupted ICAs, which simulates challenging real-world scenarios."
},
{
"figure_ref": [],
"heading": "4)",
"publication_ref": [],
"table_ref": [],
"text": "We employ ZORRO to provide interpretability and explainability of the graph matching model, shedding light on the decision-making process. This paper is organized as follows. Section 1 introduces the background, challenges of coronary artery semantic labeling and the highlights of the paper. Section 2 reviews existing algorithms on coronary artery semantic labeling and state-of-the-art graph matching algorithms. In Section 3, the proposed EAGMN is described in detail. The enrolled subjects, implementation details, experimental results, and discussion are presented in Section 4. Section 5 illustrates the conclusion."
},
{
"figure_ref": [],
"heading": "Related work",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Coronary artery semantic labeling",
"publication_ref": [
"b14",
"b5",
"b11",
"b15",
"b16",
"b17",
"b12"
],
"table_ref": [],
"text": "Coronary artery semantic labeling can be categorized into pixel-to-pixel/voxel-to-voxel based semantic segmentation methods and segment identification based semantic labeling methods. Pixel-to-pixel based methods focus on achieving dense pixel-level or voxel-level classification by mapping each individual pixel or voxel to its corresponding semantic category. On the other hand, segment identification based methods take a different approach by classifying the entire arterial segment as a whole into a specific semantic category.\nPixel-to-pixel based methods are straightforward because the algorithm assigns unique labels to each pixel in the arterial images. Jun et al. proposed a T-Net for main artery segmentation using ICAs and T-Net achieved a Dice coefficient similarity of 0.8377 using 4700 ICAs for LAD, LCX and RCA segmentation [15]. Xian et al. proposed a U-Net based residual attention network, which integrates attention mechanism and stacks multiple attention modules for main arteries segmentation. The model achieved an F1 score of 0.921 using an dataset with 3,200 ICAs [6]. Zhang et al. proposed the progressive perception learning framework to capture the long-distance semantic relationship, enhance the foreground and suppress the background pixels, and highlight the boundary details for main artery semantic segmentation [12]. The model was validated using a dataset containing 1085 subjects with ICAs and a Dice coefficient similarity of 0.9585 was achieved. Although the aforementioned studies have demonstrated impressive performance in segmenting the main arteries, they rely on separate deep learning models for extracting each type of artery and they cannot label side branches, such as D and OM arteries. Focusing solely on the main branches, such as LMA, LAD, and LCX, may not provide adequate support for downstream CAD analysis.\nSegment identification based semantic labeling methods usually contain a vascular tree binary segmentation step and a segment classification step. Cao et al. proposed an automatic coronary artery semantic labeling algorithm based on blood flow and logical rules using coronary computed tomography angiography (CCTA) images [16]. Wu et al. extracted the arterial features according to the spatial locations and directions and organized the arterial segments as tree-structured sequential data. Then, a bi-directional tree-structured long short-term memory network (LSTM) was employed to classify each segment to perform coronary artery semantic labeling [17]. Yang et al. integrated the imaging features extracted by convolutional LSTM and position features as the feature embedding and employed a partial-residual graph convolution network (GCN) for coronary artery semantic labeling [18]. Existing coronary artery semantic labeling methods have demonstrated remarkable performance when applied to 3D CCTA data. However, their direct applicability to 2D ICA images is limited due to the loss of partial high-frequency detail information resulting from the projection of the 3D vascular tree into a 2D plane. Consequently, the labeling and segmentation of coronary arteries in 2D ICA images may suffer from inaccuracies. Our previous work performed coronary artery semantic labeling by learning the semantic correspondence of arterial branches from different individual graphs using graph matching [13]. The information from neighbors is uniformly averaged during the graph matching. However, different types of arteries contribute unequally which induced a lower accuracy. It is crucial to develop methods specifically designed for 2D ICA images to enhance the accuracy and reliability of coronary artery disease diagnosis and treatment."
},
{
"figure_ref": [],
"heading": "Graph matching",
"publication_ref": [
"b18",
"b19",
"b20",
"b21",
"b22",
"b23",
"b24"
],
"table_ref": [],
"text": "Graph-matching is a combinatorial optimization problem based on graph structure, which is NP-hard or practically intractable for large-scale settings. The graph matching aims to establish a meaningful node-tonode and edge-to-edge correspondence between different graphs, which is often formulated as a graph edit problem [19], subgraph isomorphism problem and quadradic assignment problem [20]. Traditional graph matching algorithms primarily emphasize combinatorial matching techniques, which involve comparing and matching the structural components of graphs. On the other hand, learning-based methods take a different approach by incorporating feature extraction and affinity learning. These methods are particularly advantageous when dealing with large-scale and high-dimensional data, as they enable the identification of meaningful patterns and relationships within graphs [21]. With the advert development of graph neural network (GNN), it has been embedded into the learning-based graph matching problem because GNN models the structured information and helps transform the graph matching problem into a linear assignment task [22]. Nowak et al. employed GNN to extract structured features of nodes and employed Sinkhorn network as a differentiable layer to solve the linear assignment problem [23]. Wang et al. employed the GNN for both intra-graph node embedding and cross-graph node embedding iteratively for graph matching [24]. Later, Wang et al. directly employed the association graph induced affinity matrix for embedding learning and the Koopmans-Beckmann's QAP was adopted for affinity learning [25]. While existing graph matching algorithms have demonstrated promising performance in various domains, such as nature images, their application to medical images with topological features is still relatively limited and understudied. Given the significance of topological features in medical image analysis, there is a growing need for dedicated research efforts to develop graph matching algorithms specifically tailored to medical imaging. These algorithms should consider the complex interplay between structural elements and leverage the intrinsic knowledge of anatomical connectivity to achieve more accurate and meaningful matching results."
},
{
"figure_ref": [],
"heading": "Methodology",
"publication_ref": [],
"table_ref": [],
"text": "The approach outlined in this study centers on segment identification-based coronary artery semantic labeling. Our proposed EAGMN aims to establish the node-to-node correspondence between two individual graphs, taking into account the edge information. In this framework, each node represents a coronary arterial segment. Consequently, the task of coronary artery semantic labeling is transformed into the problem of finding one-to-one or one-to-zero mappings for arterial segments from two distinct vascular trees. Our method consists of two main steps: vascular tree binary segmentation and segment classification."
},
{
"figure_ref": [
"fig_1"
],
"heading": "Vascular tree extraction and graph generation",
"publication_ref": [
"b13",
"b25",
"b26"
],
"table_ref": [],
"text": "Our previous work, Feature Pyramid U-Net++ (FP-U-Net++) [14], is used to extract the vascular tree. To analyze the arterial anatomy and perform the semantic labeling, it is essential to extract the arterial centerline from the extracted vascular tree. The arterial graph generation includes the centerline extraction and arterial segment separation. Centerline extraction is the process of removing the redundant foreground pixels using binary image while preserving the connectivity and the topology of the vascular tree. We adopt the hit-and-mass transformation algorithm [26], which is based on morphological thinning, to extract the arterial centerline. The edge linking algorithm is employed to identify the end points and bifurcation points within the arterial segments. The end points denote the termination of an arterial segment, while the bifurcation points represent the junctions where the arterial segment branches out into sub-branches [27]. Based on the detected end points and bifurcation points, the arterial centerline is divided into distinct centerline segments. Each centerline segment is then associated with an arterial segment, and the semantic labeling involves assigning a label to each arterial segment. An example of the individual graph generation pipeline is shown in Figure 2.\nThe individual coronary arterial graph is constructed based on the extracted arterial segments, centerline segments, end points, and bifurcation points. The graph's connectivity, represented by its edges, is determined by the connections between arterial segments. Each node in the arterial graph is formed by the endpoints and the bifurcation points, and the semantic labels should be assigned to each edge. However, in the context of the graph matching neural network, the focus is on establishing the correspondence between nodes. Therefore, in practical terms, we interchange the concepts of nodes and edges in the individual graph. In this adjusted perspective, each node represents an arterial segment, and each edge represents an endpoint or a bifurcation point in the arterial centerline, which preserves the connectivity of the arterial tree. "
},
{
"figure_ref": [
"fig_2"
],
"heading": "Graph matching for coronary artery semantic labeling",
"publication_ref": [
"b24",
"b27",
"b10",
"b28",
"b29",
"b10",
"b30",
"b31",
"b32",
"b33",
"b34"
],
"table_ref": [
"tab_7"
],
"text": "Graph matching aims to establish a meaningful node-to-node correspondence between individual graphs. Given two undirected attributed individual graphs 𝒢 1 = (𝕍 1 , 𝔼 1 , 𝒱 1 , ℰ 1 ) and 𝒢 2 = (𝕍 2 , 𝔼 2 , 𝒱 2 , ℰ 2 ), where |𝕍 1 | = 𝑛 1 and |𝕍 2 | = 𝑛 2 , and |𝔼 1 | = 𝑛 𝑒1 and |𝔼 2 | = 𝑛 𝑒2 . We aim to find the node correspondence between them by considering the node-to-node affinities and edge-to-edge affinities. Without loss of the generality, we assume 𝑛 1 ≤ 𝑛 2 . The node-to-node correspondence can be represented by an assignment matrix 𝑀 ∈ {0,1} 𝑛 1 ×𝑛 2 that is defined in Eq. 1.\n𝑀 𝑖𝑗 = { 1 𝑖𝑓 𝑉 𝑖 𝑚𝑎𝑡𝑐ℎ𝑒𝑠 𝑉 𝑗 0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒(1)\nInspired by the success of applying association graphs in combinatorial optimization [25,28], we transform the graph matching problem between two individual graphs 𝒢 1 and 𝒢 2 into a vertex binary classification problem using an attributed undirected association graph 𝒢 𝐴 = (𝕍 𝐴 , 𝔼 𝐴 , 𝒱 𝐴 , ℰ 𝐴 ). 𝒢 𝐴 contains all candidate correspondence between nodes from two individual graphs. Each vertex of 𝒢 𝐴 is built by the nodes of the individual graphs that 𝕍 𝐴 = {(𝑉 1 1 , 𝑉 2 have the same semantic labels, then the vertex 𝑉 𝑖𝑎 ∈ 𝒢 𝐴 is a positive vertex, indicating that these two nodes are matched and these two arterial segments are matched; and verse visa. Our association-graph based edge attention graph matching network for coronary artery semantic labeling contains 6 modules, as shown in Figure 3. 1) Feature extraction in individual graphs. According to the extracted coronary artery segment, segment centerline, endpoints, bifurcation points and topology of the vascular graph, we extract the imaging features, arterial position features, and topology features for each node in individual graph. For the imaging features, we extract the texture features [11,29] according to masked ICA images for the artery using PyRadiomics [30], including first-ordered features, shape-based features, and gray-level based features. To measure the absolute position of the arterial segment relative to the entire vascular tree, we design 20 hand-crafted position features. These features capture various spatial attributes, such as the distance from the segment to the root of the tree, and the relative position within the tree hierarchy [11]. For the topology feature, we use the degree of the two endpoints of the artery segment as the features. The detailed feature information is shown in Table S1. In total, 121 features are extracted. Without loss of the generality, the dimension of node feature is denoted as 𝑑 and 𝑣 𝑖 ∈ ℝ 𝑑 .\n2) Feature extraction in association graph. The vertex is generated according to the nodes from the individual graphs. Then, the vertex feature 𝑣 𝑖𝑎 is generated by concatenating features from features of node 𝑉 𝑖 ∈ 𝒢 1 and features of node 𝑉 𝑎 ∈ 𝒢 2 , as defined in Eq. 2.\n𝑣 𝑖𝑎 = [𝑣 𝑖 , 𝑣 𝑎 ] ∈ ℝ 2𝑑 s.t. 𝑖 ∈ [1, ⋯ , 𝑛 1 ] and 𝑗 ∈ [1, ⋯ , 𝑛 2 ](2)\nwhere [⋅] denotes feature concatenation. In the same manner, the feature of edge 𝐸 𝑖𝑎,𝑗𝑏 in 𝒢 𝐴 is constructed by concatenating the features of edge 𝐸 𝑖𝑗 in 𝒢 1 and the features of edge 𝐸 𝑎𝑏 in 𝒢 2 , as defined in Eq. 3.\n𝑒 𝑖𝑎,𝑗𝑏 = [𝑒 𝑖𝑗 1 , 𝑒 𝑎𝑏 2 ] ∈ ℝ 4𝑑 s.t. 𝑖, 𝑗 ∈ [1, ⋯ , 𝑛 1 ] and 𝑎, 𝑏 ∈ [1, ⋯ , 𝑛 2 ](3)\nwhere\n𝑒 𝑖𝑗 𝑔 = [𝑣 𝑖 , 𝑣 𝑗 ] s.t. 𝑔 ∈ [1,2].\n3) Feature embedding. In the feature embedding stage, we aim to transform the input features into a lowerdimensional representation which captures the essential information for the task of coronary artery semantic labeling. This process involves mapping the input features into a feature space where relationships and patterns relevant to the task can be more easily learned and utilized by the model. Using the constructed attribute association graph 𝒢 𝐴 , two multi-layer perceptron (MLP) based encoders are employed to perform vertex and edge feature embedding, as defined in Eq. 4.\n𝑣 𝑖𝑎 𝑒𝑚𝑏 = 𝑓 𝑣 𝑒𝑚𝑏 (𝑣 𝑖𝑎 ) 𝑒 𝑖𝑎,𝑗𝑏 𝑒𝑚𝑏 = 𝑓 𝑒 𝑒𝑚𝑏 (𝑒 𝑖𝑎,𝑗𝑏 )(4)\nwhere 𝑓 𝑣 𝑒𝑚𝑏 and 𝑓 𝑒 𝑒𝑚𝑏 are MLPs with layer-wise instance normalization [31].\n4) Attention score calculation for dynamic feature aggregation. GCNs are a type of neural networks specifically designed to operate on graph-structured data. The key idea of GCN is to perform convolution on graphs by aggregating information from neighborhood nodes and edges [32]. One of the popular GCNs is graph sample and aggregated (GraphSAGE) GCN [33], which aggregates information from a node's local neighborhood using various aggregation functions, such as mean, max or concatenation. The information from neighbors is uniformly averaged because GCN treats the neighbors equally. However, in coronary artery semantic labeling task, different types of arteries contribute unequally to the central artery segment. Thus, we employed graph attention network (GAT) which incorporates attention mechanisms to improve information aggregation using graph-structured data [34,35]. In detail, an edge attention network is employed to calculate the edge attention score dynamically according to the embedded features in Eq. 4.\nIn detail, the two MLP based encoders are employed to further embed the vertex and edge features, as defined in Eq. 5.\n𝑣̅ 𝑖𝑎 𝑒𝑚𝑏 = 𝑔 𝑣 𝑒𝑚𝑏 (𝑣 𝑖𝑎 𝑒𝑚𝑏 ) 𝑒̅ 𝑖𝑎,𝑗𝑏 𝑒𝑚𝑏 = 𝑔 𝑒 𝑒𝑚𝑏 (𝑒 𝑖𝑎,𝑗𝑏 𝑒𝑚𝑏 )(5)\nwhere 𝑔 𝑣 𝑒𝑚𝑏 and 𝑔 𝑒 𝑒𝑚𝑏 are MLPs with layer-wise instance normalization. Then, a GCN layer is employed to perform information aggregation. The GCN layer contains an edge convolution layer and a vertex convolution layer. The edge convolution layer aggregates the information of the connected vertices and updates edge features according to the aggregated features and the previous edge features. Formally, the edge convolution layer contains an aggregation function and an update function, as defined in Eq. 6. "
},
{
"figure_ref": [],
"heading": "𝑒̅ 𝑖𝑎",
"publication_ref": [],
"table_ref": [],
"text": "where 𝑔 𝑒 𝑔𝑎𝑡 is an MLP encoder which aggregates features based on the embedded vertex features in Eq. 5\nof the connected two vertices of edge 𝐸 𝑖𝑎,𝑗𝑏 ; and 𝑔 𝑒 𝑢 is another MLP encoder which updates edge features according to updated features and the edge features defined in Eq. 5. The purpose of this aggregation process is to enhance the representation of the current edge by considering the characteristics and interactions of the connected vertices. This helps to capture more comprehensive and discriminative information about the graph structure and facilitates the subsequent vertex classification task.\nThe vertex convolution layer aggregates the information of the connected edges and updates its attributes according to the aggregated features and the embedded features in Eq. 5. Formally, node convolution layer contains an information aggregation layer and an update function, as defined in Eq. 7.\n𝑣̅ \nwhere 𝑓 𝑐𝑙𝑓 is an MLP decoder with ReLU activation functions and 𝑦 ̂𝑖𝑎 ∈ ℝ represents the predicted probability of the vertex 𝑉 𝑖𝑎 ."
},
{
"figure_ref": [],
"heading": "Training and testing",
"publication_ref": [
"b7",
"b35",
"b36"
],
"table_ref": [],
"text": "To train the proposed EAGMN, we need to prepare pairs of individual graphs, and then generate the association graph based on these selected pairs. In our study, we initially manually annotated the coronary arterial segments using ICA images. Subsequently, we employed the method described in Section 3.1 of our paper to generate the labeled individual graphs. According to the coronary artery anatomy defined by the American Heart Association [8], two cardiologists assigned semantic labels to each arterial segment. Then, the node correspondences between arterial segments are automatically identified and the ground truth of assignment matrix 𝑀 is generated and 𝑦 𝑖𝑎 = 1 if two arterial segments are matched.\nHowever, it should be noted that the main arterial branches, such as LAD and LCX, are often divided into multiple segments due to the presence of side branches like D and OM. As a result, the individual graph representation of these main branches consists of multiple nodes, all of which share the same semantic labels. This introduces a unique challenge in the graph matching task, as it transforms into a one-to-many or many-to-many mapping problem. Unlike the simpler one-to-one or one-to-zero mappings, the presence of multiple possible mappings between nodes significantly increases the complexity of the search space [36,37]. To simplify the complexity of the graph matching task, our approach focuses on one-to-one graph matching. We handle the issue of multiple nodes representing the separated main branches by assigning semantic labels with increasing indices along the blood flow. By assigning incremental semantic labels, we ensure that each node in the individual graph representing a main branch segment has a unique label. This allows us to establish a one-to-one correspondence between the nodes in the individual graphs, simplifying the graph matching process. For example, if LCX and OM1 exist, then LCX is separated into two arterial segments with the semantic labels of LCX1 and LCX2. Then, each arterial segment in 𝒢 1 is only matched with one exact segment in 𝒢 2 .\nFor model training, a batch of individual graph pairs is selected, and the association graphs are used as the input of the EAGMN. In our study, we specifically focus on ICAs obtained from two different view angles: left anterior oblique (LAO) and right anterior oblique (RAO). We acknowledge that the anatomy and visual characteristics of the coronary arteries can differ between these two view angles. To ensure consistency and accuracy in our analysis, we select individual graphs exclusively from ICAs captured using the same view angle. This allows us to maintain a more homogenous dataset for our analysis and ensure reliable results in coronary artery semantic labeling. The mean squared error (MSE) between the predicted vertex classification probability and the ground truth is used to optimize the model, as defined in Eq. 12.\n𝐿 = 𝑀𝑆𝐸(𝑀, 𝑀 ̂) = ∑ ∑(𝑀 𝑖𝑎 -𝑀 ̂𝑖𝑎 ) 2 𝑛 2 𝑎=1 𝑛 1 𝑖=1(12)\nwhere 𝑀 𝑖𝑎 = 𝑦 𝑖𝑎 is the ground truth and 𝑀 ̂𝑖𝑎 = 𝑦 ̂𝑖𝑎 is the predicted vertex class. An Adam optimizer with the learning rate of 0.0001 was used to optimize the model. The training algorithm is shown in Algorithm 1."
},
{
"figure_ref": [],
"heading": "Algorithm 1.",
"publication_ref": [],
"table_ref": [],
"text": "The training process of our proposed EAGMN.\nDuring the model testing phase, we employ a voting strategy to assign semantic labels to artery branches in the tested ICAs. Our proposed EAGMN establishes the node-to-node semantic correspondence between the individual graphs. During the testing, 𝒢 1 is generated using ICAs in the testing set and 𝒢 2 is generated using ICAs in the template set, where template set contains a set of representative ICAs selected by experienced cardiologists. We simulate the learning procedure by graph matching that the cardiologist learns the coronary artery anatomy by comparing the anatomy of the testing case with the reference cases in the template set. By applying the EAGMN to the association graph generated by the tested ICA and template ICA, the artery segments are matched, and the semantic labeling is achieved. Given the complex and diverse structure of coronary artery anatomy, clinical decisions are often made based on multiple ICA frames. Therefore, we perform graph matching between the tested ICA and every ICA in the template set. Each arterial segment of the tested ICA is matched to multiple arterial segments in the template set, and a majority voting strategy is employed to assign the label according to the labels of the matched arterial segments in the template set. This approach helps ensure robustness and accuracy in the final semantic labeling process. The testing algorithm is shown in Algorithm 2. 4. Embed vertex and edge features 𝑣̅ 𝑖𝑎 𝑒𝑚𝑏 and 𝑒̅ 𝑖𝑎,𝑗𝑏 𝑒𝑚𝑏 with Eq. 5."
},
{
"figure_ref": [],
"heading": "Input",
"publication_ref": [],
"table_ref": [],
"text": "For 𝑢 = 1 ⋯ 𝑁 𝑎𝑡𝑡 do 5. Update edge features 𝑣̅ 𝑖𝑎 𝑔𝑎𝑡 using Eq. 6 and vertex features 𝑣̅ 𝑖𝑎 𝑔𝑎𝑡 with Eq. 7.\n6. Calculate edge attention score 𝜃 𝑖𝑎,𝑗𝑏 with Eq. 8.\nFor 𝑣 = 1 ⋯ 𝑁 𝑚𝑝 do 7. Update edge features using 𝑒 𝑖𝑎,𝑗𝑏 𝑔𝑐𝑛 with Eq. 9 and vertex features 𝑣 𝑖𝑎 𝑔𝑐𝑛 with Eq. 10 8. Perform vertex classification with Eq. 11. 9. Optimize EAGMN using Eq .12.\nAlgorithm 2. The testing process of our proposed EAGMN."
},
{
"figure_ref": [],
"heading": "Evaluation",
"publication_ref": [
"b15"
],
"table_ref": [],
"text": "The × 𝑛 𝑐 (16) where 𝑇𝑃 𝑐 , 𝑇𝑁 𝑐 , 𝐹𝑃 𝑐 and 𝐹𝑁 𝑐 represent the true positive, true negative, false positive, and false negative arterial segments, respectively. 𝐶 is the total number of classes. 𝑛 𝐶 is the number of arterial segments in class 𝐶, and 𝑛 is the total number of arterial segments."
},
{
"figure_ref": [],
"heading": "Model interpretation",
"publication_ref": [
"b37",
"b38"
],
"table_ref": [],
"text": "Interpreting the decisions made by EAGMN is crucial for building trust and confidence in its predictions, especially in clinical practice. Existing approaches for explaining both node and edge features in GNN focus on gradient-based approaches and perturbation-based methods [38]. Gradient-based methods suffer from saturation problems in that the model output changes minimally with respect to any input change; thus, we adopt a perturbation-based method, ZORRO [39], which compares the output variations with respect to different input perturbations to calculate the node and feature importance for our graph matching network. Given a generated association graph, ZORRO iteratively and recursively adds the important features and nodes according to the fidelity score, where the fidelity score measures the difference between the raw"
},
{
"figure_ref": [],
"heading": "Input:",
"publication_ref": [
"b38"
],
"table_ref": [],
"text": "𝐷 𝑡𝑒 = {𝒢 𝑖 = (𝕍 𝑖 , 𝔼 model predictions and the predictions after masking out the important features and nodes [39]. Formally, the fidelity score is defined as\n𝐹(𝕍 𝑠 , 𝔽 𝑠 ) = 1 𝑛 1 𝑛 2 ∑ ∑ 𝕀(𝑦 𝑖𝑎 𝑠 = 𝑦 ̂𝑖𝑎 ) 𝑛 2 𝑎=1 𝑛 1 𝑖=1 (17\n)\nwhere 𝕀 is the indicator function. 𝑦 ̂𝑖𝑎 is the original prediction of the model and 𝑦 𝑖𝑎 𝑠 is the prediction using the masked nodes by 𝕍 𝑠 and the masked features by 𝔽 𝑠 . 𝕍 𝑠 and 𝔽 𝑠 are binary vectors in which 1\nindicates the node/feature is selected and vice versa. An explanation for GNN-based model indicates that using the masked features and the masked nodes in graph, the fidelity score measured by the new prediction and the original prediction reached 𝜏, i.e. 𝜏 ≥ 𝐹(𝕍 𝑠 , 𝔽 𝑠 ).\nExplaining feature importance. In section 3.1, we manually designed 121 hand-crafted features. The vertex features are generated by the concatenation of the nodes in the individual graph and identifying the concatenated features for each vertex cannot explain the feature importance for the node in the individual graph. To explain the feature importance, a unified feature mask is applied to the feature of every node in the individual graphs, then the masked features are concatenated as the vertex features using Eq. 2. ZORRO masks out all features at the beginning of the explanation and gradually adding features with the highest fidelity scores until the fidelity score is reached at 𝜏. By applying ZORRO to all individual pairs generated by the testing set and the template set, the frequency of the selected important features indicates the feature importance used for graph matching.\nExplaining node importance. For the original ZORRO, neighbor nodes are removed first and then gradually added to test the importance of the node by the improvement of the fidelity scores. The aim of explaining node importance is to interpret the importance of different arterial segments for graph matching. During the testing phase, the individual graph 𝒢 1 is the tested case and 𝒢 2 is the template ICA graph. However, our graph matching algorithm is applied to the association graph, examining the importance of vertex in the association graph cannot reflect the node importance. Thus, we modify the ZORRO to explain the node importance that at the beginning of the explanation, all nodes in 𝒢 1 is retained and all nodes in 𝒢 2 are removed. ZORRO iteratively adds the node from 𝒢 2 with the highest improvement of the fidelity score.\nAccording to the improvement of the fidelity score, the importance of the node is obtained. The node importance in our approach reflects the significance of each arterial segment in the template ICAs for both graph matching and semantic labeling. By assigning importance scores to the nodes, we can capture the relevance and influence of each arterial segment in the template set. This allows us to prioritize and weight the contributions of different segments during the graph matching process."
},
{
"figure_ref": [],
"heading": "Experiments and discussions",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Dataset",
"publication_ref": [
"b12"
],
"table_ref": [],
"text": "In this study, we manually annotated 204 and 59 ICAs from site 1 at The First Affiliated Hospital of Nanjing Medical University and site 2 at Chang Bing Show Chwan Memorial Hospital, respectively. In total, this retrospective study enrolled 263 ICA images. The detailed description of the image acquisition was illustrated in our previous work [13]. For each patient, a frame that was used for anatomical structure analysis in clinical practice was selected from the view video for semantic labeling. In this study, we only focus on semantic labeling for the main branches of LMA, LAD, and LCX, and the side branches of D and OM."
},
{
"figure_ref": [],
"heading": "Implementation details",
"publication_ref": [
"b39"
],
"table_ref": [],
"text": "We implemented our EAGMN using TensorFlow and GraphNets [40]. The vertex feature embedding module 𝑓 𝑣 𝑒𝑚𝑏 , the edge feature embedding module 𝑓 𝑒 𝑒𝑚𝑏 , the graph convolution module 𝑓 𝑒 𝑔𝑐𝑛 , 𝑓 𝑒 𝑢 , 𝑓 𝑣 𝑔𝑐𝑛 and 𝑓 𝑣 𝑢 were implemented using two MLP layers with layer-wise instance normalization and the number of hidden units were set as 64. For the attention score calculation, the vertex feature embedding module 𝑔 𝑣 𝑒𝑚𝑏 , the edge feature embedding module 𝑔 𝑒 𝑒𝑚𝑏 , the graph attention module 𝑔 𝑒 𝑔𝑎𝑡 , 𝑔 𝑒 𝑢 , 𝑔 𝑣 𝑔𝑎𝑡 and 𝑔 𝑣 𝑢 were also implemented using two MLP layers with the 64 hidden unites. The edge attention score readout module 𝑔 𝑒 𝑜𝑢𝑡 and the vertex classification layer 𝑓 𝑐𝑙𝑓 were implemented using two MLP layers with ReLU activation functions. The number of message-passing steps for graph attention score calculation, 𝑁 𝑎𝑡𝑡 , was set as 3 and the number of message-passing steps for graph convolution, 𝑁 𝑚𝑝 , was set as 2."
},
{
"figure_ref": [],
"heading": "Coronary artery semantic labeling performance",
"publication_ref": [
"b10",
"b16",
"b17",
"b23",
"b24",
"b12",
"b23",
"b24",
"b12"
],
"table_ref": [
"tab_6",
"tab_8",
"tab_8"
],
"text": "Our ICA dataset contains 263 annotated ICA graphs, and 79 of them were selected as the template set. The remining 184 ICAs were used for five-fold cross validation, which indicates that each fold 147 ICAs were used as the training set and 37 ICAs were used as the testing set. We trained our EAGMN for 𝑁 = 100,000 steps. The averaged performance on the five-folds were reported. The achieved performance of our EAGMN for coronary artery semantic labeling is shown in Table 1. We also compared the EAGMN with three existing segment identification based semantic labeling methods.\nIn addition, we implemented 3 graph matching neural networks and applied them into coronary artery semantic labeling task using the same training and testing algorithms developed in this study.\n• SVM [11]: In our previous work, a support vector machine was employed to classify coronary artery segments using the 20 position features and 2 topology features as described in section 3.1.\n• BiTreeLSTM [17]: The bidirectional tree LSTM (BiTreeLSTM) was initially applied for coronary artery semantic labeling using CCTA according to the spatial locations and directions of arteries in 3D. We adopted the same network architecture but extracted the spatial locations and directions of coronary arteries in 2D. • CPR-GCN [18]: CPR-GCN was proposed to perform coronary artery semantic labeling using CCTA with 3D convolution LSTM for arterial imaging feature extraction and GCN for artery semantic labeling. We adopted the same network architecture but extracted the arterial imaging features using 2D convolution LSTM. • IPCA [24]: Iterative Permutation loss and Cross-graph Affinity (IPCA) for graph matching employed the GNN for both intra-graph node embedding and cross-graph node embedding iteratively for graph matching using individual graphs. • NGM [25]: Neural Graph Matching (NGM) network employed the association graph induced affinity matrix for embedding learning. In our EAGMN, the connectivity of the association was generated by the connectivity of the individual graphs; however, NGM adopted the Koopmans-Beckmann's QAP for affinity learning to calculate the assignment matrix. • AGMN [13]: Our previous work adopted the association graph for graph matching without the graph attention module. For each baseline, we performed five-fold cross validation, and the performance comparisons of the coronary artery semantic labeling are illustrated in Table 2. According to Table 2, the proposed EAGMN achieved the highest accuracy of coronary artery semantic labeling for 0.8653. However, the BiTreeLSTM baseline achieved stable performance in labeling LMA branches. For BiTreeLSTM and CPR-GCN models, they were initially designed for coronary artery semantic labeling using CCTA datasets. Though they showed impressive performance with an ACC greater than 0.9, they may not be suitable for coronary artery semantic labeling using ICAs. Understanding the relationships between arteries and their spatial orientation requires a three-dimensional perspective; however, coronary artery labeling using 2D images is challenging and these two methods cannot achieve satisfactory performance. In detail, the CPR-GCN achieved a weighted ACC of 0.4581 while the BiTreeLSTM achieved an ACC of 0.7492, which were significantly lower than the graph matching based methods. The results indicate that the complex branching patterns and variations in anatomy make it more challenging to understand coronary artery anatomy, and learning the ICA anatomy using individual graphs is difficult.\nFor the graph matching based methods, we use IPCA, NGM and AGMN as baselines. IPCA [24] first performs feature embedding using individual graphs, and then performs cross-graph feature embeddings using GCNs. Compared to association graph-based graph matching algorithms, individual graph-based matching methods are computationally efficient and suitable for simpler graphs but may struggle with complex graph structures. IPCA achieved an ACC of 0.8039, which was lower than the association graphbased methods. NGM [25] first embeds features of nodes of the individual graphs to calculate the affinity matrix, and uses the calculated affinity matrix as the adjacency matrix for the association graph. Then, NGM performs the feature embedding according to the node features for vertex classification.\nOn the contrary, AGMN [13] and our proposed EAGMN concatenate the node features of the individual graphs and perform the node embedding using the concatenated features. The adjacency matrix of the association graph is generated by the connectivity of the individual graphs rather than the affinity matrix used by NGM. AGMN and EAGMN achieved the ACC of 0.8264 and 0.8653, which were higher than that of NGM. Experimental results indicate that the physical connectivity of the coronary artery plays an important role in graph matching, where the physical connectivity refers to the actual anatomical connections between different arterial segments within the coronary artery network. The physical connectivity of the coronary artery provides valuable cues and constraints that guide the matching process. It helps ensure that the matched arterial segments are not only similar in appearance or position but also consistent with the underlying anatomical structure. Compared to AGMN, our proposed EAGMN employs the edge attention mechanisms to aggregate information from the adjacent edges dynamically. By employing edge attention mechanisms, EAGMN can adaptively prioritize and weight the importance of different edges based on their relevance to the graph matching task. This dynamic attention mechanism allows the model to focus on the most informative and discriminative edges while suppressing the influence of less relevant edges. Experimental results indicate that EAGMN improved the ACC by 0.0389 for all types of arteries compared to that of AGMN, which outperformed other baseline models significantly."
},
{
"figure_ref": [
"fig_3",
"fig_3"
],
"heading": "Robustness test",
"publication_ref": [],
"table_ref": [],
"text": "In clinical practice, due to the imaging quality, artery overlapping and human subjective annotation, the algorithm proposed in section 3.1 may not generate a perfect coronary arterial graph. To reflect the robustness of the proposed EAGMN and test the clinical flexibility, we performed the robustness test using the corrupted individual graphs with random arterial segment dropping. Note that the graph-based approaches use the topology information and the connected graphs as the input, thus, we only randomly removed the arterial branches with at least one endpoint during the robustness test. In addition, the LMA was retained in all experiments due to the importance of this in-let artery segmentation for the coronary artery anatomy. We set the probability of 5%, 7.5%, 10%, 12.5%, 15%, 17.5% and 20% to randomly remove the arteries, and the averaged performance for the robustness test among baselines and EAGMN of ICAs in testing set is shown in Figure 4. According to Figure 4, the proposed EAGMN was robust since the achieved ACCs were greater than 0.81 during these robust testing experiments, even with 20% missing arterial segments. The BiTreeLSTM and CPR-GCN showed a minor performance dropping, which indicated that these two methods did not fully utilize the graph structure of coronary artery. In comparison to the graph matching-based methods, such as IPCA, NGM, and AGMN, our proposed EAGMN consistently demonstrated superior performance in terms of ACC, PREC, REC, and F1 across various missing probabilities. These results highlight the robustness of the proposed EAGMN in handling missing data scenarios."
},
{
"figure_ref": [
"fig_4",
"fig_5",
"fig_5"
],
"heading": "Result explanation",
"publication_ref": [],
"table_ref": [],
"text": "We adopted 37 ICAs as the testing set and 70 ICAs as the template set. Each ICA in testing set was used as 𝒢 1 to perform graph matching with every ICA as 𝒢 2 in template set. Note that number of nodes in 𝒢 1 is smaller than 𝒢 2 , and 1148 individual graph matching pairs were generated. We applied ZORRO to explain each graph matching pair to iteratively and recursively calculate the feature and node importance for graph matching. The threshold of the fidelity score used to measure the new prediction and the original prediction, 𝜏, was set as 0.8. We ranked the frequency of the selected important features to measure the feature importance, as shown in Figure 5. For explaining the node importance, we visualized the graph matching results and the fidelity score improvement when adding the node in the 𝒢 2 into the node mask. Example results are shown in Figure 6.\nIn Figure 6, the top four pairs of ICA graphs demonstrate a successful and accurate matching with only minor errors. However, the bottom two pairs do not exhibit a satisfactory level of matching. Upon analyzing the fidelity scores depicted in the pseudo color visualization, we observed that the LMA branch plays a crucial role in the identification of coronary arteries. The successful identification of the LMA branch, along with the first segment of LAD (LAD1) and the first segment of LCX (LCX1), is of paramount importance for the semantic labeling of the entire vascular graph.\nIn contrast, for the bottom two pairs of graphs, the significance of the LMA branch was diminished compared to the top four pairs, leading to a degradation in performance. The analysis of the node importance further confirms that our EAGMN closely mimics the approach of cardiologists in identifying coronary arteries. It starts by focusing on the main LMA branch and subsequently considers the two most vital subbranches, namely LAD1 and LCX1, in order to achieve accurate semantic labeling. Overall, these findings highlight the importance of the LMA branch and the initial segments of LAD and LCX in the graph matching process. They demonstrate that our EAGMN model effectively emulates the decisionmaking process followed by cardiologists when identifying and labeling coronary arteries. "
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "In this paper, we present an edge attention graph matching network for coronary artery semantic labeling. By performing the graph matching between individual ICA generated graphs, the relationship between the arterial segments is obtained and the unlabeled coronary arterial segments are labeled by template ICAs. Experimental results showed that our approach is powerful and robust. By employing the ZORRO, we explained the graph matching results and improved the interpretability of coronary artery semantic labeling using our proposed approach."
},
{
"figure_ref": [],
"heading": "Acknowledgement",
"publication_ref": [],
"table_ref": [],
"text": "This research was supported in part by a research seed fund from Michigan Technological University Health Research Institute and an NIH grant (U19AG055373)."
}
] | Coronary artery disease (CAD) is one of the primary causes leading deaths worldwide. The presence of atherosclerotic lesions in coronary arteries is the underlying pathophysiological basis of CAD, and accurate extraction of individual arterial branches using invasive coronary angiography (ICA) is crucial for stenosis detection and CAD diagnosis. However, deep-learning-based models face challenges in generating semantic segmentation for coronary arteries due to the morphological similarity among different types of arteries. To address this challenge, we propose an innovative approach called the Edge Attention Graph Matching Network (EAGMN) for coronary artery semantic labeling. Inspired by the learning process of interventional cardiologists in interpreting ICA images, our model compares arterial branches between two individual graphs generated from different ICAs. We begin with extracting individual graphs based on the vascular tree obtained from the ICA. Each node in the individual graph represents an arterial segment, and the EAGMN aims to learn the similarity between nodes from the two individual graphs. By converting the coronary artery semantic segmentation task into a graph node similarity comparison task, identifying the node-to-node correspondence would assign semantic labels for each arterial branch. More specifically, The EAGMN utilizes the association graph constructed from the two individual graphs as input. A graph attention module is employed for feature embedding and aggregation, while a decoder generates the linear assignment for node-to-node semantic mapping. Based on the learned node-to-node relationships, unlabeled coronary arterial segments are classified using the labeled coronary arterial segments, thereby achieving semantic labeling. A dataset with 263 labeled ICAs is used to train and validate the EAGMN. Experimental results indicate the EAGMN achieved a weighted accuracy of 0.8653, a weighted precision of 0.8656, a weighted recall of 0.8653 and a weighted F1-score of 0.8643. Furthermore, we employ ZORRO to provide interpretability and explainability of the graph matching for artery semantic labeling. These findings highlight the potential of the EAGMN for accurate and efficient coronary artery semantic labeling using ICAs. By leveraging the inherent characteristics of ICAs and incorporating graph matching techniques, our proposed model provides a promising solution for improving CAD diagnosis and treatment. | Coronary Artery Semantic Labeling using Edge Attention Graph Matching Network | [
{
"figure_caption": "Figure 1 .1Figure 1. Workflow of the proposed EAGMN for coronary artery semantic labeling using EAGMN. (a) Coronary artery binary segmentation by our previous work, feature pyramid U-Net++ [14], and individual graph generation. (b) Coronary artery semantic labeling using EAGMN by comparing the artery-to-artery correspondence between the unlabeled artery in 𝒢 1 and labeled arteries in 𝒢 2 . The positive vertex in yellow and '1' in assignment matrix represent two arterial branches are matched and the type of the unlabeled artery is assigned. The highlights of this paper are shown below:",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 .2Figure 2. Pipeline of the individual graph generation. (a) Original selected ICA frame; (b) Vascular tree generated by FP-U-Net++; (c) Centerline extraction and key points detection; (d) Artery segments with semantic labels; (e) Connectivity of the individual graph; (f) Switched individual graph that each node represents an arterial segment with semantic label and each edge represents the connectivity of the arteries. Formally, the individual graph represented by an attributed undirected graph as 𝒢 = (𝕍, 𝔼, 𝒱, ℰ), where • 𝕍 = {𝑉 𝑖 } 𝑠. 𝑡. 𝑖 ∈ [1, ⋯ , 𝑛] represents node set, and |𝕍| = 𝑛. • 𝔼 = {𝐸 𝑖𝑗 } 𝑠. 𝑡. 𝑖, 𝑗 ∈ [1, ⋯ , 𝑛] indicates the set of edges and |𝔼| = 𝑛 𝑒 . • 𝒱 = {𝑣 𝑖 } 𝑠. 𝑡. 𝑖 ∈ [1, ⋯ , 𝑛] indicates the attribute vectors associated with each node. • ℰ = {𝑒 𝑖𝑗 } s.t. 𝑖, 𝑗 ∈ [1, ⋯ , 𝑛] indicates the attribute vectors associated with each edge.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 .3Figure 3. Architecture of EAGMN for coronary artery semantic labeling. (a) Individual graph generation and feature extraction; (b) Association graph generation and feature extraction; (c) Vertex and edge feature embedding; (d) Edge attention score calculation using GCN; (e) Feature representation learning according to edge attention score derived in (d); (f) Vertex classification according to the learned feature representations.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 .4Figure 4. Robustness test of the coronary artery semantic labeling among baselines and the proposed EAGMN. The horizontal axis indicates the probability of dropping an artery segment and the vertical axis represents its corresponding performance.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 .5Figure 5. Feature importance explained by modified ZORRO. The vertical axis represents the name of the hand-crafted features, and the horizontal axis represents the frequency of the selected features when the explaination was reached. According to Figure 5. The top 15 features with the most importance are topology-based features, i.e. p2_degree and p1_degree, and position-based features. The p1_degree and p2_degree represent the degree of the endpoint of the left and right end of the arterial segment. The results indicate that the topology is an extremely important factor in arterial identification. The position features capture the absolute and relative position of the arterial segment in relation to the overall vascular tree, which also provide concrete information for artery semantic labeling. This information helps in distinguishing and identifying the individual arterial segments based on their unique spatial locations.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Figure 6 .6Figure 6. Visualization of graph matching results and the improvement of the fidelity score when adding the arterial segment in template graph for graph matching. The green line indicates a correct match, and the red line represents an error.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "𝑛 2 ]) from 𝒢 2 exist, then edge 𝐸 𝑖𝑎,𝑗𝑏 is constructed in 𝒢 𝐴 . As such, |𝔼 𝐴 | = 2 × 𝑛 𝑒1 × 𝑛 𝑒2 . If the node 𝑉 𝑖 ∈ 𝒢 1 and node 𝑉 𝑎 ∈ 𝒢",
"figure_data": "1 2 ), (𝑉 1 1 , 𝑉 2 2 ), ⋯ , (𝑉 𝑛 1 1 , 𝑉 𝑛 2",
"figure_id": "tab_0",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "where 𝑔 𝑣 𝑔𝑎𝑡 is an MLP encoder which aggregates the updated edge embeddings in Eq. 6 and 𝐴(𝑉 𝑖𝑎 )represents the set of the connected vertices of vertex 𝑉 𝑖𝑎 . The summation indicates a non-parametric function for feature aggregation. 𝑔 𝑣 𝑢 is another MLP encoder which updates vertex features according to aggregated edge features and node embedding 𝑣̅ 𝑖𝑎 𝑒𝑚𝑏 defined in Eq. 5.By iteratively applying Eqs. 6 and 7 for 𝑁 𝑎𝑡𝑡 times, the edge features are updated according to both the local and global edge and vertex features. Then, a readout module is employed to calculate the edge attention score, as defined in Eq. 8. 𝑔 𝑒 𝑜𝑢𝑡 is an MLP decoder to calculate the attention score of edge 𝐸 𝑖𝑗,𝑎𝑏 ∈ 𝒢 𝐴 , which is denoted as 𝜃 𝑖𝑎,𝑗𝑏 ∈ ℝ. The readout module takes the aggregated features from the previous step and applies further transformations to capture the importance or relevance of each edge. The edge attention score plays a crucial role in guiding the subsequent steps of the model. It helps prioritize the edges that contribute the most to the overall graph structure and assists in making accurate and meaningful vertex classification predictions.5) Feature representation learning using GCN. The feature representation learning module employs a GCN and the calculated graph attention score in Eq. 8. Similarly, this GCN also contains an edge convolution layer and a vertex convolution layer. Formally, edge convolution layer and vertex convolution layer for feature representation learning are defined in Eqs. 9 and 10.",
"figure_data": "𝑣̅ 𝑖𝑎 𝑔𝑎𝑡 ← 𝑔 𝑣 𝑢 ([𝑣̅ 𝑖𝑎 𝑔𝑎𝑡 , 𝑣̅ 𝑖𝑎 𝑒𝑚𝑏 ])𝜃 𝑖𝑎,𝑗𝑏 = 𝑔 𝑒 𝑜𝑢𝑡 (𝑒̅ 𝑖𝑎,𝑗𝑏 𝑔𝑎𝑡 )(8)𝑔𝑐𝑛 = 𝑓 𝑒 𝑔𝑐𝑛 ([𝑣 𝑖𝑎 𝑒𝑚𝑏 , 𝑣 𝑗𝑏 𝑒𝑚𝑏 ]) where the 𝑒 𝑖𝑎,𝑗𝑏 𝑒 𝑖𝑎,𝑗𝑏 𝑔𝑐𝑛 ← 𝑓 𝑒 𝑢 ([exp (-𝜃 𝑖𝑎,𝑗𝑏 ) ⋅ 𝑒 𝑖𝑎,𝑗𝑏 𝑔𝑐𝑛 , exp (-𝜃 𝑖𝑎,𝑗𝑏 ) ⋅ 𝑒 𝑖𝑎,𝑗𝑏 𝑒𝑚𝑏 ])(9)𝑣 𝑖𝑎 𝑔𝑐𝑛 =∑𝑓 𝑣 𝑔𝑐𝑛 (exp (-𝜃 𝑖𝑎,𝑗𝑏 ) ⋅ 𝑒 𝑖𝑎,𝑗𝑏 𝑔𝑐𝑛 )∀𝑉 𝑗𝑏 ∈𝐴(𝑉 𝑖𝑎 )(10)𝑣 𝑖𝑎 𝑔𝑐𝑛 ← 𝑓 𝑣 𝑢 ([𝑣 𝑖𝑎 𝑔𝑐𝑛 , 𝑣 𝑖𝑎 𝑒𝑚𝑏 ])where 𝑓 𝑒 𝑔𝑐𝑛 , 𝑓 𝑒 𝑢 , 𝑓 𝑣 𝑔𝑐𝑛 and 𝑓 𝑣𝑦 ̂𝑖𝑎 = 𝑓 𝑐𝑙𝑓 (𝑣 𝑖𝑎 𝑔𝑐𝑛 )𝑖𝑎 𝑔𝑎𝑡 =∑𝑔 𝑣 𝑔𝑎𝑡 (𝑒̅ 𝑖𝑎,𝑗𝑏 𝑔𝑎𝑡 )(7)∀𝑉 𝑗𝑏 ∈𝐴(𝑉 𝑖𝑎 )",
"figure_id": "tab_2",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "𝐷 𝑡𝑟 = {𝒢 𝑖 = (𝕍 𝑖 , 𝔼 𝑖 , 𝒱 𝑖 , ℰ 𝑖 )} 𝑖=1 𝑛 𝑡𝑟 : training set contains 𝑛 𝑡𝑟 individual graphs with labeled arteries 𝑁 𝑎𝑡𝑡 : number of message-passing steps for graph attention score calculation 𝑁 𝑚𝑝 : number of message-passing steps for graph convolution 𝑁: number of the training iterations Random select two individual graphs 𝒢 1 and 𝒢 2 from the same view angles. 2. Construct association graph 𝒢 𝐴 using 𝒢 1 and 𝒢 2 with Eqs. 2 and 3. 3. Perform feature embedding for 𝑣 𝑖𝑎 𝑒𝑚𝑏 and 𝑒 𝑖𝑎,𝑗𝑏 𝑒𝑚𝑏 with Eq. 4 .",
"figure_data": "Output:Φ: Trained EAGMNFor 𝑖𝑡𝑒𝑟 = 1 ⋯ 𝑁 do1.",
"figure_id": "tab_3",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "𝑖 , 𝒱 𝑖 , ℰ 𝑖 )} 𝑖=1 𝑛 𝑡𝑒 : testing set contains 𝑛 𝑡𝑒 individual graphs without labeled arteries. 𝐷 𝑡𝑝 = {𝒢 𝑗 = (𝕍 𝑗 , 𝔼 𝑗 , 𝒱 𝑗 , ℰ 𝑗 )} 𝑗=1 𝑛 𝑡𝑝 : template set contains 𝑛 𝑡𝑝 individual graphs with labeled arteries. Labels for each segment of 𝑛 𝑡𝑒 ICA graphs in 𝐷 𝑡𝑒 For 𝑖 = 1 ⋯ 𝑛 𝑡𝑒 do For 𝑖 = 𝑗 ⋯ 𝑛 𝑡𝑝 do If |𝕍 𝑖 | ≤ |𝕍 𝑗 | and 𝒢 𝑖 and 𝒢 𝑗 are from the same view angles, then 1. Construct association graph 𝒢 𝐴 using 𝒢 𝑖 and 𝒢 𝑗 with Eqs. 2 and 3 2. Calculate assignment matrix 𝑀 ̂𝑖𝑗 = Φ(𝒢 𝐴 ) 3. Assign labels to 𝒢 𝑖 according to the majority voting among set of the 𝑀 ̂𝑖𝑗 , s.t. 𝑗 ∈ [1, ⋯ 𝑛 𝑡𝑝 ]",
"figure_data": "Φ: trained AGAMNOutput:",
"figure_id": "tab_5",
"figure_label": "",
"figure_type": "table"
},
{
"figure_caption": "Achieved performance for coronary artery semantic labeling using the proposed EAGMN.",
"figure_data": "Artery typeACCPRECRECF1-scoreLMA1.0000±0.00000.9841±0.01301.0000±0.00000.9919±0.0066LAD0.8905±0.06610.8692±0.04910.8905±0.06610.8795±0.0563LCX0.8831±0.06520.8444±0.04050.8831±0.06520.8629±0.0497D0.8142±0.07140.8335±0.08210.8142±0.07140.8233±0.0741OM0.7597±0.04080.8472±0.07690.7597±0.04080.7999±0.0505weighted average0.8653±0.04880.8656±0.04990.8653±0.04880.8643±0.0485As illustrated in",
"figure_id": "tab_6",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "our EAGMN achieved an average accuracy of 0.8653, which indicating 86.53% of the arterial segments in the testing set were correctly classified according to the trained EAGMN and the selected ICAs in the template set.",
"figure_data": "",
"figure_id": "tab_7",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Comparison of coronary artery semantic labeling between the proposed EAGMN and existing methods using our ICA dataset. The achieved highest performance is shown in bold.",
"figure_data": "MethodMetricLMALADLCXDOMmeanML0.9925±0.01510.6331±0.05400.6388±0.03900.6147±0.04100.5907±0.04190.6651±0.0080BiTreeLSTM1.0000±0.00000.8845±0.01500.9871±0.01200.0000±0.00000.5981±0.01650.7492±0.0085CPR-GCN0.5361±0.29960.5319±0.12390.5072±0.14470.0624±0.09530.5341±0.30450.4581±0.0536IPCAACC0.9820±0.02200.8222±0.06990.7837±0.06340.7924±0.08570.7140±0.06760.8039±0.0537NGM0.9953±0.00930.8316±0.05860.8154±0.07390.7629±0.08700.7105±0.10360.8130±0.0636AGMN0.9956±0.00890.8432±0.03060.8046±0.04520.7956±0.04120.7565±0.08250.8264±0.0302EAGMN1.0000±0.00000.8905±0.06610.8831±0.06520.8142±0.07140.7597±0.04080.8653±0.0488ML0.9778±0.00710.6586±0.01740.6375±0.03780.5554±0.01010.6278±0.02610.6679±0.0081BiTreeLSTM1.0000±0.00000.8562±0.01900.5853±0.00990.0000±0.00000.9808±0.01220.6927±0.0074CPR-GCN0.6208±0.32400.5675±0.05400.3964±0.05700.2802±0.37270.3821±0.01390.4463±0.1075IPCAPREC0.9779±0.01350.7865±0.06340.8008±0.05520.7233±0.05720.7932±0.06730.8046±0.0545NGM0.9910±0.01100.8269±0.07060.7985±0.06690.7546±0.06880.7492±0.08660.8122±0.0644AGMN0.9911±0.01090.8476±0.04810.8256±0.03070.7536±0.04930.7613±0.03190.8276±0.0298EAGMN0.9841±0.01300.8692±0.04910.8444±0.04050.8335±0.08210.8472±0.07690.8656±0.0499ML0.9925±0.01510.6331±0.05400.6388±0.03900.6147±0.04100.5907±0.04190.6651±0.0080BiTreeLSTM1.0000±0.00000.8845±0.01500.9871±0.01200.0000±0.00000.5981±0.01650.7492±0.0085CPR-GCN0.5361±0.29960.5319±0.12390.5072±0.14470.0624±0.09530.5341±0.30450.4581±0.0536IPCAREC0.9820±0.02200.8222±0.06990.7837±0.06340.7924±0.08570.7140±0.06760.8039±0.0537NGM0.9953±0.00930.8316±0.05860.8154±0.07390.7629±0.08700.7105±0.10360.8130±0.0636AGMN0.9956±0.00890.8432±0.03060.8046±0.04520.7956±0.04120.7565±0.08250.8264±0.0302EAGMN1.0000±0.00000.8905±0.06610.8831±0.06520.8142±0.07140.7597±0.04080.8653±0.0488ML0.9850±0.00760.6437±0.02130.6360±0.00870.5832±0.02340.6071±0.01830.6646±0.0077BiTreeLSTM1.0000±0.00000.8699±0.01010.7348±0.00930.0000±0.00000.7429±0.01410.6967±0.0085CPR-GCN0.5698±0.30260.5455±0.08990.4353±0.06320.0742±0.09570.3924±0.16600.4192±0.0661IPCAF10.9798±0.01330.8038±0.06540.7921±0.05910.7559±0.06950.7510±0.06450.8033±0.0538NGM0.9932±0.00920.8290±0.06370.8067±0.06990.7583±0.07600.7289±0.09460.8123±0.0640AGMN0.9933±0.00890.8452±0.03860.8143±0.03100.7736±0.04240.7569±0.05080.8262±0.0301EAGMN0.9919±0.00660.8795±0.05630.8629±0.04970.8233±0.07410.7999±0.05050.8643±0.0485",
"figure_id": "tab_8",
"figure_label": "2",
"figure_type": "table"
}
] | Chen Zhao; Zhihui Xu; Guang-Uei Hung; Weihua Zhou | [
{
"authors": "J S Lawton; J E Tamis-Holland; S Bangalore; E R Bates; T M Beckie; J M Bischoff; J A Bittl; M G Cohen; J M Dimaio; C W Don; S E Fremes; M F Gaudino; Z D Goldberger; M C Grant; J B Jaswal; P A Kurlansky; R Mehran; T S Metkus; L C Nnacheta; S V Rao; F W Sellke; G Sharma; C M Yong; B A Zwischenberger",
"journal": "Circulation",
"ref_id": "b0",
"title": "ACC/AHA/SCAI Guideline for Coronary Artery Revascularization: Executive Summary: A Report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines",
"year": "2021"
},
{
"authors": "W E Boden; R A O'rourke; K K Teo; P M Hartigan; D J Maron; W J Kostuk; M Knudtson; M Dada; P Casperson; C L Harris; B R Chaitman; L Shaw; G Gosselin; S Nawaz; L M Title; G Gau; A S Blaustein; D C Booth; E R Bates; J A Spertus; D S Berman; G B J Mancini; W S Weintraub",
"journal": "N Engl J Med",
"ref_id": "b1",
"title": "Optimal Medical Therapy with or without PCI for Stable Coronary Disease",
"year": "2007"
},
{
"authors": "The Discharge Trial Group; P Maurovich-Horvat; M Bosserdt; K F Kofoed; N Rieckmann; T Benedek; P Donnelly; J Rodriguez-Palomares; A Erglis; C Štěchovský; G Šakalyte; N Čemerlić Adić; M Gutberlet; J D Dodd; I Diez; G Davis; E Zimmermann; C Kępka; R Vidakovic; M Francone; M Ilnicka-Suckiel; F Plank; J Knuuti; R Faria; S Schröder; C Berry; L Saba; B Ruzsics; C Kubiak; I Gutierrez-Ibarluzea; K Schultz Hansen; J Müller-Nordhorn; B Merkely; A D Knudsen; I Benedek; C Orr; F Xavier; L Valente; V Zvaigzne; L Suchánek; F Zajančkauskiene; M Adić; M Woinke; I Hensey; E Lecumberri; M Thwaite; M Laule; A N Kruk; M Neskovic; D Mancone; G Kuśmierz; M Feuchtner; V Pietilä; T Gama Ribeiro; C Drosch; G Delles; M Matta; B Fisher; L Szilveszter; M Larsen; S Ratiu; B Kelly; A Garcia Del Blanco; Z D Rubio; B Drobni; I Jurlander; S Rodean; H Regan; M Cuéllar Calabria; T Boussoussou; R Engstrøm; A E Hodas; R Napp; S Haase; L M Feger; K Serna-Higuita; H Neumann; M Dreger; V Rief; M Wieske; P Estrella; M Martus; Dewey",
"journal": "N Engl J Med",
"ref_id": "b2",
"title": "CT or Invasive Coronary Angiography in Stable Chest Pain",
"year": "2022"
},
{
"authors": "C Spadaccio; U Benedetto",
"journal": "Ann Cardiothorac Surg",
"ref_id": "b3",
"title": "Coronary artery bypass grafting (CABG) vs. percutaneous coronary intervention (PCI) in the treatment of multivessel coronary disease: quo vadis? -a review of the evidences on coronary artery disease",
"year": "2018"
},
{
"authors": "Z Li; Y Zhang; G Liu; H Shao; W Li; X Tang",
"journal": "Biomedical Signal Processing and Control",
"ref_id": "b4",
"title": "A robust coronary artery identification and centerline extraction method in angiographies",
"year": "2015"
},
{
"authors": "Z Xian; X Wang; S Yan; D Yang; J Chen; C Peng",
"journal": "Mathematical Problems in Engineering",
"ref_id": "b5",
"title": "Main Coronary Vessel Segmentation Using Deep Learning in Smart Medical",
"year": "2020"
},
{
"authors": "N I Parikh; E F Honeycutt; M T Roe; M Neely; E J Rosenthal; M A Mittleman; J P Carrozza; K K L Ho",
"journal": "Circ: Cardiovascular Quality and Outcomes",
"ref_id": "b6",
"title": "Left and Codominant Coronary Artery Circulations Are Associated With Higher In-Hospital Mortality Among Patients Undergoing Percutaneous Coronary Intervention for Acute Coronary Syndromes: Report From the National Cardiovascular Database Cath Percutaneous Coronary Intervention (CathPCI) Registry",
"year": "2012"
},
{
"authors": "W Austen; J Edwards; R Frye; G Gensini; V Gott; L Griffith; D Mcgoon; M Murphy; B Roe",
"journal": "American Heart Association, Circulation",
"ref_id": "b7",
"title": "A reporting system on patients evaluated for coronary artery disease. Report of the Ad Hoc Committee for Grading of Coronary Artery Disease, Council on Cardiovascular Surgery",
"year": "1975"
},
{
"authors": "S Kastellanos; K Aznaouridis; C Vlachopoulos; E Tsiamis; E Oikonomou; D Tousoulis",
"journal": "WJC",
"ref_id": "b8",
"title": "Overview of coronary artery variants, aberrations and anomalies",
"year": "2018"
},
{
"authors": "G Yang; A Broersen; R Petr; P Kitslaar; M A De Graaf; J J Bax; J H C Reiber; J Dijkstra",
"journal": "Computing in Cardiology",
"ref_id": "b9",
"title": "Automatic coronary artery tree labeling in coronary computed tomographic angiography datasets",
"year": "2011"
},
{
"authors": "C Zhao; R Bober; H Tang; J Tang; M Dong; C Zhang; Z He; M Esposito; Z Xu; W Zhou",
"journal": "Journal of Advances in Applied & Computational Mathematics",
"ref_id": "b10",
"title": "Semantic Segmentation to Extract Coronary Arteries in Invasive Coronary Angiograms",
"year": "2022"
},
{
"authors": "H Zhang; Z Gao; D Zhang; W K Hau; H Zhang",
"journal": "IEEE Transactions on Medical Imaging",
"ref_id": "b11",
"title": "Progressive Perception Learning for Main Coronary Segmentation in X-ray Angiography",
"year": "2022"
},
{
"authors": "C Zhao; Z Xu; J Jiang; M Esposito; D Pienta; G.-U Hung; W Zhou",
"journal": "",
"ref_id": "b12",
"title": "AGMN: Association Graph-based Graph Matching Network for Coronary Artery Semantic Labeling on Invasive Coronary Angiograms",
"year": "2023-02-10"
},
{
"authors": "C Zhao; A Vij; S Malhotra; J Tang; H Tang; D Pienta; Z Xu; W Zhou",
"journal": "Computers in Biology and Medicine",
"ref_id": "b13",
"title": "Automatic extraction and stenosis evaluation of coronary arteries in invasive coronary angiograms",
"year": "2021"
},
{
"authors": "T J Jun; J Kweon; Y.-H Kim; D Kim; T-Net ",
"journal": "Neural Networks",
"ref_id": "b14",
"title": "Nested encoder-decoder architecture for the main vessel segmentation in coronary angiography",
"year": "2020"
},
{
"authors": "Q Cao; A Broersen; M A De Graaf; P H Kitslaar; G Yang; A J Scholte; B P F Lelieveldt; J H C Reiber; J Dijkstra",
"journal": "Int J Cardiovasc Imaging",
"ref_id": "b15",
"title": "Automatic identification of coronary tree anatomy in coronary computed tomography angiography",
"year": "2017"
},
{
"authors": "D Wu; X Wang; J Bai; X Xu; B Ouyang; Y Li; H Zhang; Q Song; K Cao; Y Yin",
"journal": "Int J CARS",
"ref_id": "b16",
"title": "Automated anatomical labeling of coronary arteries via bidirectional tree LSTMs",
"year": "2019"
},
{
"authors": "H Yang; X Zhen; Y Chi; L Zhang; X.-S Hua",
"journal": "IEEE",
"ref_id": "b17",
"title": "CPR-GCN: Conditional Partial-Residual Graph Convolutional Network in Automated Anatomical Labeling of Coronary Arteries",
"year": "2020"
},
{
"authors": "H Bunke",
"journal": "Pattern Recognition Letters",
"ref_id": "b18",
"title": "On a relation between graph edit distance and maximum common subgraph",
"year": "1997"
},
{
"authors": "E L Lawler",
"journal": "Management Science",
"ref_id": "b19",
"title": "The quadratic assignment problem",
"year": "1963"
},
{
"authors": "J Yan; S Yang; E Hancock",
"journal": "",
"ref_id": "b20",
"title": "Learning for Graph Matching and Related Combinatorial Optimization Problems",
"year": "2020"
},
{
"authors": "H Dai; E B Khalil; Y Zhang; B Dilkina; L Song",
"journal": "",
"ref_id": "b21",
"title": "Learning Combinatorial Optimization Algorithms over Graphs",
"year": "2018-05-09"
},
{
"authors": "A Nowak; S Villar; A S Bandeira; J Bruna",
"journal": "IEEE",
"ref_id": "b22",
"title": "Revised Note on Learning Quadratic Assignment with Graph Neural Networks",
"year": "2018"
},
{
"authors": "R Wang; J Yan; X Yang",
"journal": "IEEE Trans. Pattern Anal. Mach. Intell",
"ref_id": "b23",
"title": "Combinatorial Learning of Robust Deep Graph Matching: an Embedding based Approach",
"year": "2020"
},
{
"authors": "R Wang; J Yan; X Yang",
"journal": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"ref_id": "b24",
"title": "Neural Graph Matching Network: Learning Lawler's Quadratic Assignment Problem With Extension to Hypergraph and Multiple-Graph Matching",
"year": "2022"
},
{
"authors": "E R Dougherty",
"journal": "SPIE, Optical Engineering Press",
"ref_id": "b25",
"title": "An introduction to morphological image processing",
"year": "1992"
},
{
"authors": "J Xie; Y Zhao; Y Liu; P Su; Y Zhao; J Cheng; Y Zheng; J Liu",
"journal": "",
"ref_id": "b26",
"title": "Topology reconstruction of treelike structure in images via structural similarity measure and dominant set clustering",
"year": "2019"
},
{
"authors": "T Wang; H Liu; Y Li; Y Jin; X Hou; H Ling",
"journal": "IEEE",
"ref_id": "b27",
"title": "Learning Combinatorial Solver for Graph Matching",
"year": "2020"
},
{
"authors": "C Zhao; Y Xu; Z He; J Tang; Y Zhang; J Han; Y Shi; W Zhou",
"journal": "Pattern Recognition",
"ref_id": "b28",
"title": "Lung segmentation and automatic detection of COVID-19 using radiomic features from chest CT images",
"year": "2021"
},
{
"authors": "J J M Van Griethuysen; A Fedorov; C Parmar; A Hosny; N Aucoin; V Narayan; R G H Beets-Tan; J.-C Fillion-Robin; S Pieper; H J W L Aerts",
"journal": "Cancer Res",
"ref_id": "b29",
"title": "Computational Radiomics System to Decode the Radiographic Phenotype",
"year": "2017"
},
{
"authors": "D Ulyanov; A Vedaldi; V Lempitsky",
"journal": "",
"ref_id": "b30",
"title": "Instance normalization: The missing ingredient for fast stylization",
"year": "2016"
},
{
"authors": "J Gilmer; S S Schoenholz; P F Riley; O Vinyals; G E Dahl",
"journal": "",
"ref_id": "b31",
"title": "Neural Message Passing for Quantum Chemistry",
"year": "2017"
},
{
"authors": "W Hamilton; Z Ying; J Leskovec",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b32",
"title": "Inductive representation learning on large graphs",
"year": "2017"
},
{
"authors": "P Veličković; G Cucurull; A Casanova; A Romero; P Lio; Y Bengio",
"journal": "",
"ref_id": "b33",
"title": "Graph attention networks",
"year": "2017"
},
{
"authors": "J Qu; H Ling; C Zhang; X Lyu; Z Tang",
"journal": "",
"ref_id": "b34",
"title": "Adaptive Edge Attention for Graph Matching with Outliers",
"year": "2021"
},
{
"authors": "H Zhang; R Yanagi; R Togo; T Ogawa; M Haseyama",
"journal": "IEEE Access",
"ref_id": "b35",
"title": "Cross-Modal Image Retrieval Considering Semantic Relationships With Many-to-Many Correspondence Loss",
"year": "2023"
},
{
"authors": "M Zaslavskiy; F Bach; J.-P Vert",
"journal": "",
"ref_id": "b36",
"title": "Many-to-Many Graph Matching: a Continuous Relaxation Approach",
"year": "2010-06-21"
},
{
"authors": "H Yuan; H Yu; S Gui; S Ji",
"journal": "IEEE Trans. Pattern Anal. Mach. Intell",
"ref_id": "b37",
"title": "Explainability in Graph Neural Networks: A Taxonomic Survey",
"year": "2022"
},
{
"authors": "T Funke; M Khosla; M Rathee; A Anand; Zorro ",
"journal": "IEEE Trans. Knowl. Data Eng",
"ref_id": "b38",
"title": "Valid, Sparse, and Stable Explanations in Graph Neural Networks",
"year": "2023"
},
{
"authors": "P W Battaglia; J B Hamrick; V Bapst; A Sanchez-Gonzalez; V Zambaldi; M Malinowski; A Tacchetti; D Raposo; A Santoro; R Faulkner; C Gulcehre; F Song; A Ballard; J Gilmer; G Dahl; A Vaswani; K Allen; C Nash; V Langston; C Dyer; N Heess; D Wierstra; P Kohli; M Botvinick; O Vinyals; Y Li; R Pascanu",
"journal": "",
"ref_id": "b39",
"title": "Relational inductive biases, deep learning, and graph networks",
"year": "2018"
}
] | [
{
"formula_coordinates": [
8,
228.29,
596.31,
304.46,
26.04
],
"formula_id": "formula_0",
"formula_text": "𝑀 𝑖𝑗 = { 1 𝑖𝑓 𝑉 𝑖 𝑚𝑎𝑡𝑐ℎ𝑒𝑠 𝑉 𝑗 0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒(1)"
},
{
"formula_coordinates": [
9,
166.37,
589.16,
367.82,
14.4
],
"formula_id": "formula_1",
"formula_text": "𝑣 𝑖𝑎 = [𝑣 𝑖 , 𝑣 𝑎 ] ∈ ℝ 2𝑑 s.t. 𝑖 ∈ [1, ⋯ , 𝑛 1 ] and 𝑗 ∈ [1, ⋯ , 𝑛 2 ](2)"
},
{
"formula_coordinates": [
9,
147.62,
644.72,
387.17,
15.12
],
"formula_id": "formula_2",
"formula_text": "𝑒 𝑖𝑎,𝑗𝑏 = [𝑒 𝑖𝑗 1 , 𝑒 𝑎𝑏 2 ] ∈ ℝ 4𝑑 s.t. 𝑖, 𝑗 ∈ [1, ⋯ , 𝑛 1 ] and 𝑎, 𝑏 ∈ [1, ⋯ , 𝑛 2 ](3)"
},
{
"formula_coordinates": [
9,
101.66,
665,
121.35,
17.27
],
"formula_id": "formula_3",
"formula_text": "𝑒 𝑖𝑗 𝑔 = [𝑣 𝑖 , 𝑣 𝑗 ] s.t. 𝑔 ∈ [1,2]."
},
{
"formula_coordinates": [
10,
246.53,
129.84,
288.26,
29.52
],
"formula_id": "formula_4",
"formula_text": "𝑣 𝑖𝑎 𝑒𝑚𝑏 = 𝑓 𝑣 𝑒𝑚𝑏 (𝑣 𝑖𝑎 ) 𝑒 𝑖𝑎,𝑗𝑏 𝑒𝑚𝑏 = 𝑓 𝑒 𝑒𝑚𝑏 (𝑒 𝑖𝑎,𝑗𝑏 )(4)"
},
{
"formula_coordinates": [
10,
246.17,
348.99,
288.62,
36.24
],
"formula_id": "formula_5",
"formula_text": "𝑣̅ 𝑖𝑎 𝑒𝑚𝑏 = 𝑔 𝑣 𝑒𝑚𝑏 (𝑣 𝑖𝑎 𝑒𝑚𝑏 ) 𝑒̅ 𝑖𝑎,𝑗𝑏 𝑒𝑚𝑏 = 𝑔 𝑒 𝑒𝑚𝑏 (𝑒 𝑖𝑎,𝑗𝑏 𝑒𝑚𝑏 )(5)"
},
{
"formula_coordinates": [
12,
197.69,
425.57,
335.42,
40.2
],
"formula_id": "formula_8",
"formula_text": "𝐿 = 𝑀𝑆𝐸(𝑀, 𝑀 ̂) = ∑ ∑(𝑀 𝑖𝑎 -𝑀 ̂𝑖𝑎 ) 2 𝑛 2 𝑎=1 𝑛 1 𝑖=1(12)"
},
{
"formula_coordinates": [
15,
204.89,
108.72,
324.81,
40.2
],
"formula_id": "formula_9",
"formula_text": "𝐹(𝕍 𝑠 , 𝔽 𝑠 ) = 1 𝑛 1 𝑛 2 ∑ ∑ 𝕀(𝑦 𝑖𝑎 𝑠 = 𝑦 ̂𝑖𝑎 ) 𝑛 2 𝑎=1 𝑛 1 𝑖=1 (17"
},
{
"formula_coordinates": [
15,
529.7,
124.6,
4.61,
9.94
],
"formula_id": "formula_10",
"formula_text": ")"
}
] |
|
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b0",
"b1",
"b2",
"b3",
"b4",
"b5",
"b6"
],
"table_ref": [],
"text": "Nowadays, router devices are so important with Internet Service Provider (ISP) in core network. Juniper router that is used in SCTV core network enables a wide range of business and residential applications and services (e.g., high-speed transport, Virtual Private Network services, high-speed Internet) [1]. To detect and prevent abnormal operations of Juniper router devices, there are some solutions such as monitoring systems, checking syslog server [2]. Abnormal operations of devices may be cause of routing errors in the network, chassis errors, Distributed Denial-of-service (DDos) attack or some fail processes in router devices. These abnormal operations of router devices are quite reported to syslog server if they are configured. It is possible to detect abnormal operation by inspecting manually these logs. Anomaly detection has been a practical research topic and has great importance in many application domains in university and in industry (e.g., [3], [4], [5]). For conventional small systems, engineers manually define rules or check system logs to detect anomalies based on their domain knowledge. Additionally, they can use regular expression match or keyword searches (e.g., \"error\", \"fail\"). However, the extremely large sizes of log data generated (up to million logs) by hundreds of devices make manual analysis impossible.\nAs a result, automated log analysis methods for anomaly detection of Juniper router devices based on machine learning systems are highly in demand. However, we found no resources about this problem. This paper is the first work about anomaly detection using One-Class SVM for logs of Juniper router devices. We present a new way to get features from log data of Juniper router devices based on importance characteristics of log messages. The standard support vector machine (SVM) [6] is the machine learning method to classify two class or multiple class of data. But the log data of anomaly detection are very special, the abnormal logs are much less than the normal logs. Therefore, the standard SVM does not work well on our task, we use One-Class SVM described in [7] to handle the logs classification."
},
{
"figure_ref": [],
"heading": "Collecting Log Data, Preprocessing and Feature Extraction",
"publication_ref": [],
"table_ref": [],
"text": "Collecting log data for training and testing is necessary because we use the real logs from real router Juniper devices. Logs must be made clean before they can be used for feature extraction."
},
{
"figure_ref": [
"fig_0"
],
"heading": "Collecting Log Data and Text Preprocessing",
"publication_ref": [],
"table_ref": [],
"text": "Juniper router devices routinely generate logs to record device states and runtime information, each including a content indicating what has happened and a timestamp. These devices are configured to send logs to syslog server that always receives data on corresponding port. The log messages have some common characteristics although the format of them is not fixed length. This valuable information could be utilized for anomaly detection and other purposes; thereby logs are collected first and saved as a file in syslog server for further usage. For example, Fig. 1 depicts some log messages of Juniper router devices. As the example above, a timestamp is the beginning of each message, a name of Juniper router device with the same format and other characteristics: the log messages at syslog server are written in English and comprise digits, lower, upper case letters and many special characters. The raw logs are pre-processed by Python program. Our text normalization procedures are given below: Remove timestamps: Remove whole timestamp before each message (both date and time). Remove router device's name: Remove device's name. Remove digits, special characters: Remove any special character, including punctuations, all digits. Replace continuous spaces with a single space: The length of log message depends on how many spaces in log message, so that replacing continuous spaces with a single space is necessary. Lower cases: Replace all upper case letters by lower case.\nAlthough timestamp information, which is the time the event happened, is very important according to the SCTV engineers, we still remove it from the log message. After anomaly detection, the abnormal original log messages (including timestamp information) will be sent to the engineers."
},
{
"figure_ref": [],
"heading": "Feature Extraction",
"publication_ref": [
"b7",
"b8",
"b9"
],
"table_ref": [],
"text": "Some methods used to perform feature extraction for text classification [8][9][10] are not effective with logs of Juniper router devices. In our works, we use three characteristics of each log message to be three elements of the feature vector: the length of log message, the number of different words and the sum of TF-IDF in log message.\nThe Length of Log.\nThe length of log message is a significant characteristic. In the process of manually defining abnormal logs, the length of logs shows that the logs with irregular length have a higher probability of being abnormal logs. We denote i S as the length of log i . We use the spaces of each log message to calculate i S ."
},
{
"figure_ref": [],
"heading": "The Number of Different Words in Log Message.",
"publication_ref": [
"b10",
"b11",
"b12"
],
"table_ref": [],
"text": "The number of different words in log message: the number of words which are different from dictionary of each log message. Bag-of-Words (BoW) model and the length of log are used to calculate this characteristic. Some of new document representation methods [11][12][13] are developed based on BoW.\nThe dictionary is built based on normal logs that we classify from original logs. To get the number of words that are in dictionary, BoW model is the accordant and simple classical model for our purpose. Based on BoW, each vector represents a log message; each element denotes the normalized number of occurrence of a word in log. The words that do not appear in dictionary are not counted by BoW model.\nBased on the dictionary and BoW model, each log is converted from text to a vector space. The log data become the matrix that is given by: where m is the length of dictionary, n is the size of log data. Each row of the matrix A represents a log message in log data. The number of words which are different from dictionary of each log message is given by the formula:\n1 m i i ij j L S a (2)\nwhere i S is length of the i-th log."
},
{
"figure_ref": [],
"heading": "The Sum of TF-IDF.",
"publication_ref": [
"b13",
"b14",
"b15"
],
"table_ref": [],
"text": "Term frequency-inverse document frequency (TF-IDF) is one of the most famous algorithms used in document mining research; it is used for calculating the weight of each word. The word frequency means the number of time a term is repeated in a log message, and Inverse Document Frequency is an algorithm used to calculate the inverse probability of finding a word in log data [14]. Some improvements of feature extraction based on TF-IDF are mentioned in [15,16]. TF-IDF Formula:\nlog ij ij j N g tf df (3)\nwhere ij g is the weight of the word j in the log i , N is the total number of log mes- sages, ij tf is the frequency of the word j in the log i , j df is the number of logs con- taining the word j . The sum of TF-IDF in log is the third element of feature vector. Equation 4is the summary equation used to calculate the sum of TF-IDF:\n1 h i ij j Gg (4)\nwhere i G is the sum of TF-IDF of log i , h is the number of words in log i . The feature vector of log i :"
},
{
"figure_ref": [],
"heading": " ",
"publication_ref": [],
"table_ref": [],
"text": ",,\ni i i i x S L G (5)\nThe feature vectors are the input data of One-Class SVM model, which will be discussed below."
},
{
"figure_ref": [],
"heading": "One-Class SVM",
"publication_ref": [
"b5",
"b6"
],
"table_ref": [],
"text": "One famous machine learning method is the support vector machine (SVM), which was invented by Vladimir Vapnik and Alexey Ya. Chervonenkis in 1963, and is widely applied for pattern classification and data analyzing. However, Vladimir Vapnik and Corinna Cortes proposed the current standard incarnation in 1993 [6].\nThe labels associated with training data are based to group anomaly detection techniques for log data into two broad categories: one-class and multi-class anomaly detection technique. The labels in log messages of Juniper router devices are grouped into two types: abnormal log and normal log, so that one-class anomaly detection technique was chosen for our purpose.\nFor the case of one-class classification, Scholkopf et al. proposed a maximum margin based classifier that is an adaptation of the Support Vector Machine algorithm [7]. The data are separated from the origin by a separating hyperplane ,0 wz with maximum margin (where w and are respectively the normal vector of the hyper- plane and the distance from the hyperplane to the origin).\nThe maximum margin from the origin is found by solving the below quadratic optimization problem:\n 2 ,, 1 1 min 2 subject to ( ) , 0. i i w i i i w vl w x (6)\nwhere i are so-called slack variables that are used to model the separation errors. The 0,1 v is a parameter that adjusts the balance between maximizing the distance from the origin and the region created by the hyperplane containing most of the data. () is a non-linear projection is evaluated through a kernel function that is used as a mapping from the original feature space to a possibly higher dimensional feature space:\n ( , ) ( ) ( k x y x y . In our works, we consider the kernel Radial Basis Function (RBF), linear kernel, polynomial kernel and sigmoid kernel. They are expressed respectively by these following equations ( 7)-( 10): x y e\n (7) , T linear k x y x y C (8) ( , ) d T polynomial k x y x y C (9)\n \n( , ) tanh\nT sigmoid k x y x y C (10)"
},
{
"figure_ref": [],
"heading": "Experimental Results",
"publication_ref": [
"b16",
"b17",
"b18"
],
"table_ref": [],
"text": "In this section, we descript the log datasets, the evaluate method, perform an experimental evaluation and comparison of our anomaly detection method for log data of Juniper router devices based on One-Class SVM model. We have chosen Python programming language for our convenience purpose. Python is a powerful interpreted and popular language and supports some power libraries for data science, machine learning (numpy, matplotlib, scikit-learn, etc) [17][18][19]."
},
{
"figure_ref": [
"fig_3",
"fig_3"
],
"heading": "Log Datasets",
"publication_ref": [],
"table_ref": [
"tab_0"
],
"text": "Publicly available production logs are scarce data, especially log data of router devices because companies and ISPs rarely publish them to community due to confidential issues. So that we collected log data from real Juniper router devices in core network of SCTV (Saigontourist Cable Television Company). Logs from devices were sent through the internet network to syslog server and saved as txt files. These files can be read directly and easily by most popular programs. The log data contains 12907 log messages and 266 abnormal log messages, which are manually labeled by us and the SCTV experts. Using this log dataset, we evaluated the performance and compared the results of the models. More statistical information of the log dataset is provided in Table 1. After the raw logs are collected using Python script, they are called from our main Python program. Feature extraction is applied to these log messages, each log is represented by a feature vector include: the length of log message, the number of words which are different from dictionary and the sum of TF-IDF value in log message. We choose 60% of log messages as the training data and the remainders as the testing data.\nFigure 2 shows the training data and testing data on 3-dimensional feature space. Blue points are the normal data, red points are abnormal data. Due to the fact that the log data come from over a hundred router Juniper devices, we receive a lot of similar log messages. There are some differences between them such as the timestamp, device's name, etc. So a lot of log messages become the same after they are passed the preprocess step that we mentioned above. Because of that, the data points in the figure may look less than reality. Although the visualizations in figure 2 shows that the data are naturally well-separated after they are converted to feature vectors, we still propose the One-Class SVM method to separate the normal log data and abnormal log data automatically, and make sure the model still works efficiently in case there are more abnormal log data appeared when the hardware system is extended."
},
{
"figure_ref": [],
"heading": "Method Evaluation",
"publication_ref": [
"b19"
],
"table_ref": [],
"text": "The Precision, Recall and F-measure, which are the most commonly used metrics, are used to evaluate the accuracy of One-Class SVM anomaly detection method using different kernels (Radial Basis Function (RBF), linear kernel, polynomial kernel and sigmoid kernel) as we have already the ground truth for the log data. As shown below, Precision shows the percentage of true anomalies among all anomalies detected, Recall measures the percentage of how many real anomalies are detected, and F-measure denotes the harmonic mean of precision and recall [20]. They are expressed respectively by the following equations ( 11) -( 13):"
},
{
"figure_ref": [],
"heading": "Anomalies detected Precision",
"publication_ref": [],
"table_ref": [],
"text": "Anomalies reported (11)"
},
{
"figure_ref": [],
"heading": "Anomalies detected Recall",
"publication_ref": [],
"table_ref": [],
"text": "All anomalies\n (12) 2 Precision Recall F measure Precision Recall (13)"
},
{
"figure_ref": [
"fig_4"
],
"heading": "Result Evaluation",
"publication_ref": [],
"table_ref": [],
"text": "Based on evaluate methods that we mentioned, we show the results which evaluate of each model applied on training data and testing data. Figure 3 shows the accuracy of anomaly detection on training log data. Three models (with linear, polynomial, sigmoid kernels) have no good performance on training data with the F-measure close to 0.75. We can observe that recall measures of these models are low (close to 0.65). We give priority to minimize the number of abnormal log messages which the model predicted wrongly, One-Class SVM method with RBG kernel is the best model in our case. Recall measure of this model is equal to 1, it shows that the model predicted rightly all abnormal logs. One-Class SVM RBF kernel model's performance on training data is better than others models. However, their accuracy on testing data varies with different kernels. When applying these models for testing data, three models (with linear, polynomial, sigmoid kernels) become unacceptable. These kernels are not suitable for our purpose. We can observe that One-Class SVM methods with RBF and polynomial kernel achieve high Precision (over 0.95), which implies that normal instances and abnormal instances are well separated by using our feature representation. As we observe on "
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "Logs are widely utilized to detection anomalies in Juniper router device systems. However, traditional anomaly detection that depends heavily on manual log inspection becomes impossible due to the limit of human ability and the sharp increase of log size.\nTo reduce manual effort, automated log analysis and anomaly detection methods have been widely studied in recent years.\nIn this paper, we successfully created the new feature extraction for log data of Juniper router devices and used the One-Class SVM model with different kernels for anomaly detection. We also compared their accuracy and efficiency on training and testing real log datasets. We find that the One-Class SVM model with RBF kernel has the best accuracy in terms of precision, recall and F-measure."
},
{
"figure_ref": [],
"heading": "Acknowledgments",
"publication_ref": [],
"table_ref": [],
"text": "This work was partially supported by the Ministry of Science and Technology, Taiwan, under grant MOST 107-2221-E-006-222. In addition, this work got the encouragement from Posts and Telecommunications Institute of Technology, Vietnam."
}
] | The article deals with anomaly detection of Juniper router logs. Abnormal Juniper router logs include logs that are usually different from the normal operation, and they often reflect the abnormal operation of router devices. To prevent router devices from being damaged and help administrator to grasp the situation of error quickly, detecting abnormal operation soon is very important. In this work, we present a new way to get important features from log data of Juniper router devices and use machine learning method (basing on One-Class SVM model) for anomaly detection. One-Class SVM model requires some knowledge and comprehension about logs of Juniper router devices so that it can analyze, interpret, and test the knowledge acquired. We collect log data from a lot of real Juniper router devices and classify them based on our knowledge. Before these logs are used for training and testing the One-Class SVM model, the feature extraction phase for these data was carried out. Finally, with the proposed method, the system errors of the routers were dectected quickly and accurately. This may help our company to reduce the operation cost for the router systems. | Anomaly Detection Using One-Class SVM for Logs of Juniper Router Devices | [
{
"figure_caption": "Fig. 1 .1Fig. 1. Some log messages of Juniper router devices.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 2 .2Fig. 2. Training data and testing data on 3-dimensional feature space.",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 3 .3Fig. 3. Training accuracy of One-Class SVM method.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 4 .4Fig. 4. Testing accuracy of One-Class SVM method.",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 ,Fig. 5 .45Fig. 5. The decision boundary of One-Class SVM model with RBF kernel.",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "45",
"figure_type": "figure"
},
{
"figure_caption": "",
"figure_data": "",
"figure_id": "",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Summary of log datasets.",
"figure_data": "DevicesData sizeNumber of messages AnomaliesRouter Juniper1,7 Mb12 907266",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
}
] | Tat-Bao-Thien Nguyen; Teh-Lu Liao; Tuan-Anh Vu | [
{
"authors": "",
"journal": "",
"ref_id": "b0",
"title": "MX960 3D Universal Edge Router Hardware Guide",
"year": "2018-01-22"
},
{
"authors": "",
"journal": "",
"ref_id": "b1",
"title": "Network Performance Monitor Getting Started Guide",
"year": "2018-01-25"
},
{
"authors": "Q Lin; J G Lou; H Zhang",
"journal": "",
"ref_id": "b2",
"title": "Log Clustering Based Problem Identification for Online Service Systems",
"year": "2016"
},
{
"authors": "M Macit; E Delibaş; B Karanlık",
"journal": "",
"ref_id": "b3",
"title": "Real time distributed analysis of MPLS network logs for anomaly detection",
"year": "2016"
},
{
"authors": "Y Gao; Y Ma; D Li",
"journal": "",
"ref_id": "b4",
"title": "Anomaly detection of malicious users' behaviors for web applications based on web logs",
"year": "2017"
},
{
"authors": "C Cortes; V Vapnik",
"journal": "Journal Machine Learning",
"ref_id": "b5",
"title": "Support-Vector Networks",
"year": "1995"
},
{
"authors": "B Scholkopf; J Platt; J Shawe-Taylor; A J Smola; R Williamson",
"journal": "Neural Computation",
"ref_id": "b6",
"title": "Estimating the support of a high-dimentional distribution",
"year": "2001"
},
{
"authors": "C Y Chang; S J Lee; C C Lai",
"journal": "",
"ref_id": "b7",
"title": "Weighted word2vec based on the distance of words",
"year": "2017"
},
{
"authors": "W Tian; J Li; H Li",
"journal": "",
"ref_id": "b8",
"title": "A Method of Feature Selection Based on Word2Vec in Text Categorization",
"year": "2018"
},
{
"authors": "T P Van; T M Thanh",
"journal": "",
"ref_id": "b9",
"title": "Vietnamese news classification based on BoW with keywords extraction and neural network",
"year": "2017"
},
{
"authors": "R Zhao; K Mao",
"journal": "IEEE Transactions on Fuzzy Systems",
"ref_id": "b10",
"title": "Fuzzy Bag-of-Words Model for Document Representation",
"year": "2018"
},
{
"authors": "L Wu; S C H Hoi; N Yu",
"journal": "IEEE Transactions on Image Processing",
"ref_id": "b11",
"title": "Semantics-Preserving Bag-of-Words Models and Applications",
"year": "2010"
},
{
"authors": "A Alahmadi; A Joorabchi; A E Mahdi",
"journal": "",
"ref_id": "b12",
"title": "A new text representation scheme combining Bag-of-Words and Bag-of-Concepts approaches for automatic text classification",
"year": "2013"
},
{
"authors": "S M H Dadgar; M S Araghi; M M Farahani",
"journal": "",
"ref_id": "b13",
"title": "A novel text mining approach based on TF-IDF and Support Vector Machine for news classification",
"year": "2016"
},
{
"authors": "X Huang; Q Wu",
"journal": "",
"ref_id": "b14",
"title": "Micro-blog commercial word extraction based on improved TF-IDF algorithm",
"year": "2013"
},
{
"authors": "A Guo; T Yang",
"journal": "",
"ref_id": "b15",
"title": "Research and improvement of feature words weight based on TFIDF algorithm",
"year": "2016"
},
{
"authors": "F Dubosson; S Bromuri; M Schumacher",
"journal": "",
"ref_id": "b16",
"title": "A Python Framework for Exhaustive Machine Learning Algorithms and Features Evaluations",
"year": "2016"
},
{
"authors": "E Patterson; R Mcburney; H Schmidt",
"journal": "IBM Journal of Research and Development",
"ref_id": "b17",
"title": "Dataflow representation of data analyses: Toward a platform for collaborative data science",
"year": "2017"
},
{
"authors": "C P Hwang; M S Chen; C M Shih",
"journal": "",
"ref_id": "b18",
"title": "Apply Scikit-Learn in Python to Analyze Driver Behavior Based on OBD Data",
"year": "2018"
},
{
"authors": "M Sokolova; G Lapalme",
"journal": "Information Processing and Management",
"ref_id": "b19",
"title": "A systematic analysis of performance measures for classification tasks",
"year": "2009"
}
] | [
{
"formula_coordinates": [
4,
261.91,
468.63,
208.79,
19.33
],
"formula_id": "formula_0",
"formula_text": "1 m i i ij j L S a (2)"
},
{
"formula_coordinates": [
4,
251.02,
640.75,
219.67,
29.21
],
"formula_id": "formula_1",
"formula_text": "log ij ij j N g tf df (3)"
},
{
"formula_coordinates": [
5,
267.75,
232.74,
202.94,
20.83
],
"formula_id": "formula_2",
"formula_text": "1 h i ij j Gg (4)"
},
{
"formula_coordinates": [
5,
260.01,
303.34,
210.68,
10.78
],
"formula_id": "formula_3",
"formula_text": "i i i i x S L G (5)"
},
{
"formula_coordinates": [
5,
225.5,
603.34,
245.2,
32.26
],
"formula_id": "formula_4",
"formula_text": " 2 ,, 1 1 min 2 subject to ( ) , 0. i i w i i i w vl w x (6)"
},
{
"formula_coordinates": [
6,
232.06,
224.16,
238.64,
83.98
],
"formula_id": "formula_5",
"formula_text": " (7) , T linear k x y x y C (8) ( , ) d T polynomial k x y x y C (9)"
},
{
"formula_coordinates": [
6,
229.04,
318.06,
241.66,
17.02
],
"formula_id": "formula_6",
"formula_text": "T sigmoid k x y x y C (10)"
},
{
"formula_coordinates": [
8,
215.21,
194.46,
255.49,
52.68
],
"formula_id": "formula_7",
"formula_text": " (12) 2 Precision Recall F measure Precision Recall (13)"
}
] |
||
10.18653/v1/S17-2001 | 2023-05-21 | [
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b3",
"b14",
"b17",
"b20",
"b30",
"b20",
"b7",
"b31",
"b11",
"b6",
"b24",
"b34",
"b35",
"b18",
"b22",
"b10",
"b25",
"b28",
"b29",
"b21",
"b16",
"b27"
],
"table_ref": [],
"text": "Pretrained language models (LMs) powered by finetuning have achieved remarkable performance on a wide range of downstream tasks (Devlin et al., 2019;Liu et al., 2019;Radford et al., 2019). Driven by the pursued task-agnostic property, distillation of LMs has witnessed a paradigm shift from taskspecific to task-agnostic distillation (Sanh et al., 2019). Under a teacher-student regime, taskagnostic distillation distils pretrained LMs into ones of small compute on pretraining data so that Figure 1: The failures of prior distillation methods. The setting is to distil a base-scale teacher to a 6-layer student. Either distilling last layer self-attention distributions (Wang et al., 2021) or logits (Sanh et al., 2019) for encoder-decoder LMs yields severe degradation or only marginal gain compared to pretraining from scratch, in contrast to significant improvements for either encoderonly or decoder-only LMs. Note that the lower the perplexity, the better.\nthese small LMs can be applied to tasks by finetuning (Jiao et al., 2020;Wang et al., 2020;Liang et al., 2023). In contrast, task-specific distillation distils finetuned LMs on finetuning data and consumed resource can be even huge when the number of tasks explode (Hinton et al., 2015;Sun et al., 2019;Xia et al., 2022;Yang et al., 2022). Additionally, it is acknowledged that task-agnostic distillation typically brings performance gain over task-specific distillation does (Zhang et al., 2022a).\nDespite so many merits, prior studies mostly lie in distillation of either encoder-only LMs (e.g., BERT, Devlin et al., 2019) or decoder-only LMs (e.g., GPT, Radford et al., 2019) and largely ignore the signifance of task-agnostic distillation of encoder-decoder LMs (e.g., T5, Raffel et al., 2020) given recent advances in task-specific distillation of encoder-decoder LMs though (Shleifer and Rush, 2020;Zhang et al., 2022b;Li et al., 2022;Tao et al., 2022). Frustratingly, we find that existing distillation methods may fail to handle taskagnostic distillation of encoder-decoder LMs since encoder-decoder LMs can behave very differently in comparison with encoder-only and decoder-only LMs (e.g., the use of cross-attention, Vaswani et al., 2017). The failures of prior methods are showcased in Figure 1.\nTo the end, we investigate to, in a task-agnostic style, save the distillation of encoder-decoder LMs from the awkward position. Specifically, we reveal that the key to unlocking the expressiveness of distillation is the interplay between the encoder and the decoder. Therefore, we offer a path named as MINIEND that successfully tackles the distillation of encoder-decoder LMs by alternatively distilling the cross-attention to explicitly fall to both the encoder and the decoder.\nWe check MINIEND on language understanding and abstractive summarization in sense that encoder-decoder LMs are more capable of sequence-to-sequence tasks. For evaluation on language understanding, we take GLUE (Wang et al., 2019) to benchmark the performance. For evaluation on abstractive summarization, we adopt CNN/DailyMail (See et al., 2017) and XSum (Narayan et al., 2018) as two testbeds. The results of both distilling T5 and BART indicate that MINIEND is effective and competitive to other compression options such as quantization. We further scale our method up to the distillation of 3B T5 xlarge with the aid of progressive distillation. The results suggest that distilling large language models (e.g., LLaMA, Touvron et al., 2023) should be promising but can be challenging."
},
{
"figure_ref": [],
"heading": "Encoder-Decoder Interplay",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Architecture Perspective",
"publication_ref": [
"b28"
],
"table_ref": [],
"text": "Typically, an encoder-decoder LM is composed of an encoder and a decoder, each of which is essentially a stack of transformer layers (Vaswani et al., 2017). Concretely, a transformer layer in the encoder contains a multihead self-attention (MSA) module and a feedforward network (FFN) module. Similarly, a transformer layer in the decoder comprises an MSA module, an FFN module, and additionally a multihead cross-attention (MCA) module that is inserted between the MSA and the FFN modules and accounts for absorption of encoded information from the encoder. Around each of these modules is attached necessarily a layer normaliza-tion and a residual connection.\nMSA and FFN Mathematically, the procedure that a transformer encoder layer consumes an intermediate encoder input X ∈ R n×d containing a n-length sequence of d-dimension vectors from last layer and gives an output to next layer can be depicted as a composition of MSA and FFN:\nMSA(X; W Q , W K , W V ) = A i SelfAttn(X; W Q i , W K i )XW V i W O i , SelfAttn(X; W Q i , W K i ) = Softmax(XW Q i W K i X /d A ), FFN(X; W I , W O ) = I j g(XW I j )W O j ,\nwhere potential details (e.g., linear bias and layer normalization) are omitted. i is used to indicate i-th head parameterized by MCA Likewise, the procedure that a transformer decoder layer processes an intermediate decoder input Z ∈ R m×d based on the final encoder output E ∈ R n×d can be incrementally described as an insertion of MCA:\nW Q i , W K i , W V i ∈ R d×d A , W O i ∈ R d A ×d among A heads,\nMCA(Z, E; W Q , W K , W V ) = A i CrossAttn(Z, E; W Q i , W K i )EW V i W O i , CrossAttn(Z, E; W Q i , W K i ) = Softmax(ZW Q i W K i E /d A ),\nHere, each cross-attention head is parameterized by another set of parameters\nW Q i , W K i , W V i ∈ R d×d A , W O i ∈ R d A ×d .\nInterplay through Architecture In this architectural sense, the decoder is tightly connected to the encoder through MCA modules. In spite that state-of-the-art methods mainly manipulate the decoder during distillation (e.g., logits, Zhang et al., 2022b), the encoder could be learned anyway through the connections offered by MCA modules. However, it is still not clear to what extent the encoder-decoder interplay is significant in the distillation and whether the implicit connections mentioned above are enough for alignment of the interplay."
},
{
"figure_ref": [],
"heading": "Gradient Perspective",
"publication_ref": [
"b9"
],
"table_ref": [],
"text": "More thoroughly, we take a closer look at the connections between the encoder and the decoder through the lens of gradients.\nWe examine the gradient norms of last layer hidden states of both the encoder and the decoder under two distinguished distillation objectives when distilling from BART (Lewis et al., 2020). The intuition lies in that, in contrast to implicit consideration of the encoder-decoder interplay, a distillation objective explicitly involving the encoder-decoder interplay alignment could behave much differently in terms of gradients if the interplay is central to the distillation of encoder-decoder LMs. And naturally, if suboptimal cases are identified in the implicit objective, we can further highlight that the implicit objective suffers from the limited interplay alignment and the explicit objective can provide a more effective one."
},
{
"figure_ref": [
"fig_0"
],
"heading": "Implicit versus Explicit Objective",
"publication_ref": [
"b30",
"b36"
],
"table_ref": [],
"text": "We instantiate the implicit objective as aligning logits and last decoder layer self-attention distributions, and the explicit objective as aligning logits, last decoder layer self-attention distributions, and last decoder layer cross-attention distributions. The core idea of last layer attention distribution alignment is borrowed from MiniLM (Wang et al., 2021). Any alignment can be abstracted as L(S; T , D * ), where D * denotes, with slight abuse of notation, the distribution of the input. As a crucial part, the alignment of self-attention is like the following:\nL SelfAttn (S; T , D Z ) = E Z∼D Z R k=1 KL(Reln(Z; T W Q k ), Reln(Z; S W Q k )) + KL(Reln(Z; T W K k ), Reln(Z; S W K k )) + KL(Reln(Z; T W V k ), Reln(Z; S W V k )), Reln(Z; T W Q k ) = Softmax(Z T W Q k T W Q k Z /d R ),\nwhere S and T are the teacher and the student, and KL stands for kullback-leibler divergence. Particularly, attention heads are first merged from the original A attention heads and then split to R heads for alignment of the number of attention heads.\nT /S W Q k is the redistributed query parameter of the k-th head within totally R heads from the last decoder layer, likewise T /S W K k and T /S W V k are the key and value parameters.\nThe alignment of cross-attention is similar but sort of different in that the keys and the values are aligned in fact from the encoder side, as the following:\nL CrossAttn (S; T , D Z , D E ) = E Z∼D Z ,E∼D E R k=1 KL(Reln(Z; T W Q k ), Reln(Z; S W Q k )) + KL(Reln(E; T W K k ), Reln(E; S W K k )) + KL(Reln(E; T W V k ), Reln(E; S W V k )),\nHere, the notations should be self-contained by referring to previously mentioned ones.\nInterplay through Gradient To recap, the implicit objective is: Contrarily, the explicit objective is derived by adding a cross-attention term as:\nL Logit (S; T , D Z ) + L SelfAttn (S; T , D Z ),\nL Logit (S; T , D Z ) + L SelfAttn (S; T , D Z ) + L CrossAttn (S; T , D Z , D E ),\nPreliminary results are shown in Figure 2, from which we can see that 1) the implicit objective and the explicit objective lead to distinct gradient variations, and 2) the implicit objective exhibits gradient spikes, compared with smooth gradient transitions from the explicit objective, that may result in instability for a nice convergence (Zeng et al., 2022). Thereby, from the gradient perspective, we safely conclude that the encoder-decoder interplay is of importance in the distillation of encoder-decoder LMs and an explicit correspondence to the interplay is superior to an implicit one."
},
{
"figure_ref": [
"fig_1"
],
"heading": "MINIEND",
"publication_ref": [],
"table_ref": [],
"text": "With aforementioned justifications in mind, we propose a path dubbed as MINIEND that tackles the distillation of encoder-decoder LMs under the guidance of the encoder-decoder interplay alignment. The path can be built in two directions. An overview of these two directions is given in Figure 3."
},
{
"figure_ref": [],
"heading": "Decoder Cross-Attention",
"publication_ref": [],
"table_ref": [],
"text": "The first is the one used in our pilot study. That said, we should always plus a fraction towards the alignment of output logits and the overall distillation objective is therefore depicted as:\nL(S; T , D Z , D E ) = L Logit (S; T , D Z )+ L SelfAttn (S; T , D Z ) + L CrossAttn (S; T , D Z , D E ),\nThe alignment of logits can be further detailed as:\nL Logit (S; T , D Z ) = E Z∼D Z CE(Z S W E , Z T W E ),\nwhere CE stands for soft cross entropy and T /S W E denotes output embedding.\nEncoder Self-Attention The second is an alternative to the first one where the interplay part is accounted by the last encoder self-attention distributions instead as:\nL(S; T , D Z , D X ) = L Logit (S; T , D Z )+ L SelfAttn (S; T , D Z ) + L EncSelfAttn (S; T , D X ),\nThe rationale of introducing the encoder selfattention alignment abides in that this term together with the decoder self-attention alignment can sufficiently replace the cross-attention term and align the encoder-decoder interplay by aligning both the encoder and the decoder."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Data and Metrics",
"publication_ref": [
"b18",
"b5",
"b29",
"b23",
"b32",
"b4",
"b1",
"b33",
"b19",
"b0",
"b8"
],
"table_ref": [
"tab_2"
],
"text": "Following the pretraining of T5 and BART, we use C4 (Raffel et al., 2020) as the corpus for task-agnostic distillation of T5 and OpenWeb-Text (Gokaslan et al., 2019) for that of BART. They are separately processed to follow the pretraining styles of T5 and BART. That is, C4 is converted to the masked language modeling style and Open-WebText is converted to the denoising style.\nFor evaluation of MINIEND, we mainly take GLUE (Wang et al., 2019) for language understanding. The GLUE benchmark consists of two sequence classification tasks, SST-2 (Socher et al., 2013), i.e., CoLA (Warstadt et al., 2019), and seven sequence-pair classification tasks, i.e., MRPC (Dolan and Brockett, 2005), STS-B (Cer et al., 2017), QQP, MNLI (Williams et al., 2018), QNLI (Rajpurkar et al., 2016), RTE (Bentivogli et al., 2011), WNLI (Levesque et al., 2012). We exclude WNLI and CoLA due to the evaluation 1, where the corpora used for distillation is also attached."
},
{
"figure_ref": [],
"heading": "Implementation",
"publication_ref": [
"b13",
"b39",
"b31"
],
"table_ref": [
"tab_3"
],
"text": "The distillation is carried out on 16 Nvidia A100s. The number of relation heads is set to 32. After the distillation, the finetuning is carried out on one Nvidia A100. For language understanding tasks, T5 is finetuned with simplicity and performance guarantee following EncT5 (Liu et al., 2021) which uses the very first token (i.e., [BOS]) representation from the decoder, while BART is finetued following its original paper which uses the very last token (i.e., [EOS]) representation from the decoder. As for abstractive summarization tasks, both T5 and BART are finetuned in a sequence-to-sequence manner. For fast development, we use greedy search for T5 and beam search for BART only. The beam search setting strictly follows the original paper. In order to achieve higher training efficiency, we utilize fully-sharded data parallel (Zhao et al., 2023) to shard both the teacher and the student across GPUs during the distillation. For all cases, students are always randomly initialized before the distillation following MiniLM (Wang et al., 2020).\nThe details of hyperparameters for distillation and finetuning are shown in Table 2. We will be releasing our code and scripts in the final version for exact reproducibility."
},
{
"figure_ref": [],
"heading": "Baselines",
"publication_ref": [
"b6",
"b30",
"b10"
],
"table_ref": [],
"text": "We name two variants of MINIEND as MINIEND-D and MINIEND-E respectively, where MINIEND-D uses decoder cross-attention for interplay alignment and MINIEND-E uses encoder self-attention instead. As there are no existing work in taskagnostic distillation of encoder-decoder LMs, we mainly compare MINIEND to task-agnostic baselines that are heavily adapted to encoder-decoder LMs and task-specific baselines that may be not super fair for comparison.\nWe compare MINIEND-D and MINIEND-E distilled from T5 to task-agnostic baselines on GLUE, CNN/DailyMail, and XSum: MlmKD (Hinton et al., 2015) that directly distils masked language modeling logits; MiniLM (Wang et al., 2021) that distils last decoder layer attention distributions; MlmKD+MiniLM that is essentially a combination of preceding two. We also compare MINIEND-D and MINIEND-E distilled from T5 to a task- specific baseline that is as far as we know the most comparable one on GLUE: MiniDisc (Zhang et al., 2022a) that exploits a teacher assistant for large compression.\nOn the other hand, we compare MINIEND-D distilled from BART to two recent task-specific baselines on CNN/DailyMail and XSum: LogitKD and DQ-BART (Li et al., 2022) that jointly quantizes and distils from the teacher.\nFor MINIEND above baselines, student structures are denoted either with *L;*H for number of layers and dimension of hidden states in random initialization, or with *% for preserved portion of parameters in pruning initialization."
},
{
"figure_ref": [],
"heading": "Main Results",
"publication_ref": [],
"table_ref": [
"tab_4",
"tab_5"
],
"text": "Baselines fail, yet MINIEND triumphs. From results in Table 3 andTable 4, we can tell that baselines fail to handle the distillation of encoder-decoder LMs since they either underperform the baseline pretrained from scratch or out-perform it by only a small margin. For example, MlmKD+MiniLM achieves 84.5 versus 84.6 from T5 in GLUE Score, and 35.8 versus 35.7 from T5 in CNN/DailyMail Rg-1.\nContrarily, MiniEnD can safely escape from performance degradation and bring further performance increment. For example, MINIEND-D reaches 0.1 absolute improvement in GLUE Score, and 0.9 absolute improvement in XSum Rg-1. The improvement in GLUE Score seems to be not very significant, but can be boosted according to the ablation. That is, MINIEND-E w/o L Logit goes up to 85.0, which is notably better than 84.6 from T5 in the average sense. All count, and interplay forms the key. On another note, removing L Logit will consistently produce performance deterioration on CNN/DailyMail and XSum. We conjecture there is a tradeoff of using between using L Logit or not. Namely, the use of L Logit will offer better generative ability but worse discriminative ability, and the removal of it will work reversely. Anyway, either L CrossAttn in MINIEND-D or L SelfEncAttn in MINIEND-E shall be a crucial ingredient as the interplay alignment term is the only difference between MINIEND and MlmKD+MiniLM but results in a considerable performance gap.\nAnd it may be suspected that whether L SelfAttn is still important given that MiniLM is not an ideal choice for the distillation of encoder-decoder LMs. We suggest the use of it in two aspects: 1) MlmKD+MiniLM is better than MlmKD alone; 2) the interplay alignment will witeness a subtle performance drop after the removal of L SelfAttn , say MINIEND-D will decrease from 84.7 to 83.0 in GLUE Score. Quantization has two sides. MINIEND surpasses most of them except DQ-BART. However, we should emphasize that quantized LMs usually perform better but run much slower than distilled LMs do when compression is the same. In our case, DQ-BART uses 8 bit precision and gives rise to a 4× model size reduction which is the same as that of MINIEND. In addition to that, MINIEND is orthogonal to quantization and thus can be enhanced with other quantization schemes."
},
{
"figure_ref": [
"fig_2"
],
"heading": "Analyses",
"publication_ref": [
"b15",
"b12"
],
"table_ref": [
"tab_6"
],
"text": "Data Scaling Some would wonder whether the huge amounts of GPU hours due to the large pre-training corpus is necessary. So we inspect the performance variation of MINIEND-D by varying data scale, which is shown in Figure 4.\nThe results generally hint that using a portion of data could hardly approximate the full data performance, though half data can achieve acceptable performance. Therefore, we suggest the use of full data in the distillation.\nModel Scaling Inspired by pioneering work finding a curse that larger teachers induces worse students, we double check the existence of the curse and offer a trial solution to the curse so that we can scale the teacher up to 3B T5 xlarge .\nFrom the results in Table 5, we observe that the curse of capacity gap still exists in our case. With the increase of teacher scale, the student performance decreases. We attempt to apply common solutions the circumvent the curse. The first is to make the student learn from a teacher assistant distilled from the teacher (Mirzadeh et al., 2020). The second is to make the student to learn from a smaller teacher and then from the teacher (Lin et al., 2023). Both two solutions inherit the idea of inserting an additional distillation step thus progressive distillation. We reveal that teacher assistantbased distillation is somewhat useful but not as excepted since T5 xlarge ⇒T5 12L;384H ⇒T5 6L;384H still does not imrpove over T5 xlarge ⇒T5 6L;384H . We claim that distilling large language models like LLaMA can therefore be appealing but challenging."
},
{
"figure_ref": [],
"heading": "Conclusions",
"publication_ref": [],
"table_ref": [],
"text": "In this paper, we aim to provide a path that successfully tackles the distillation of encoder-decoder LMs, which fails most previous methods in the area. We find through a pilot study that the encoderdecoder interplay is a key component that should be aligned in the distillation so that the distilled encoder-decoder LMs are promising. Based on the idea, we propose two directions that the encoderdecoder interplay alignment can be incorporated and verify their effectiveness on a language understanding benchmark and two abstractive summarization datasets. We further scale the distillation of encoder-decoder LMs to a 3B teacher that requires additional distillation steps. In this sense, we recommend future research to devote more efforts to exploring how large language models can be distilled."
},
{
"figure_ref": [],
"heading": "Limitations",
"publication_ref": [
"b2",
"b26"
],
"table_ref": [],
"text": "This paper lacks a validation study on more recently advanced encoder-decoder LMs such as Flan (Chung et al., 2022) and UL2 (Tay et al., 2022) as well as their instruction-tuned version."
}
] | Finetuning pretrained language models (LMs) have enabled appealing performance on a diverse array of tasks. The intriguing taskagnostic property has driven a shifted focus from task-specific to task-agnostic distillation of LMs. While task-agnostic, computeefficient, performance-preserved LMs can be yielded by task-agnostic distillation, previous studies mainly sit in distillation of either encoder-only LMs (e.g., BERT) or decoderonly ones (e.g., GPT) yet largely neglect that distillation of encoder-decoder LMs (e.g., T5) can posit very distinguished behaviors. Frustratingly, we discover that existing taskagnostic distillation methods can fail to handle the distillation of encoder-decoder LMs. To the demand, we explore a few paths and uncover a path named as MINIEND that successfully tackles the distillation of encoderdecoder LMs in a task-agnostic fashion. We examine MINIEND on language understanding and abstractive summarization. The results showcase that MINIEND is generally effective and is competitive compared to other alternatives. We further scale MINIEND up to distillation of 3B encoder-decoder language models with interpolated distillation. The results imply the opportunities and challenges in distilling large language models (e.g., LLaMA). | Task-agnostic Distillation of Encoder-Decoder Language Models | [
{
"figure_caption": "Figure 2 :2Figure2: The preliminary results of gradient norms when using the implicit or explicit objective. The implicit objective imposes distinct gradient variations and unexpected gradient spikes during the distillation.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: The overview of MINIEND. Two directions are proposed to consider the encoder-decoder interplay alignment.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: The results of data scaling using MINIEND-D.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "The data statistics, maximum sequence lengths, and metrics. The maximum decoder sequence lengths of T5 and BART are indicated differently for language understanding tasks since they use different finetuning strategies.",
"figure_data": "Dataset#Train exam #Dev exam Max enc len Max dec lenMetricC4364.9M-512114-OpenWebText37.8M-512512-SST-267.3K0.9K641 / 64AccuracyMRPC3.7K0.4K1281 / 128F1STS-B7.0K1.5K1281 / 128 Spearman CorrelationQQP364.0K40.0K1281 / 128F1MNLI-m/mm393.0K20.0K1281 / 128AccuracyQNLI105.0K5.5K1281 / 128AccuracyRTE2.5K0.3K1281 / 128AccuracyCNN/DailyMail287.1K13.4K512128F1XSum204.0K11.3K512128F1inconsistency (in other words, MiniLMs get dra-matically worse results while LMs get much betterones as found out in Xia et al., 2022) and use theleft tasks. Following BERT (Devlin et al., 2019),we report Accuracy (Acc) on SST-2, MNLI, QNLI,RTE, Spearman Correlation scores (SpCorr) onSTS-B, and F1 on MRPC, QQP, CoNLL. Aver-age score over tasks from GLUE (GLUE Score) isadditionally computed. Regarding that one of themost promising properties of encoder-decoder LMsis sequence-to-sequence modeling, we addition-ally adopt CNN/DailyMail (See et al., 2017) andXSum (Narayan et al., 2018) for abstractive sum-marization. We report Rouge-{1,2,L} (Rg-{1,2,L})on both of them. Results are reported on develop-ment sets. GFLOPs are also attached as theoreticalspeedup references.The detailed data statistics, maximum sequencelengths, and metrics for datasets we use are shownin Table",
"figure_id": "tab_2",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "The hyperparameters for both distillation and finetuning. In order to realize the global batch size, necessary gradient accumulations should be used. The beam search setting applies to BART only.",
"figure_data": "HyperparameterC4Distillation OpenWebTextGLUEFinetuning CNN/DailyMailXSumBatch size10241024{16,32}{16,32}{16,32}OptimizerAdamWAdamWAdamWAdamWAdamWLearning rate3e-43e-4{1e-5,2e-5,3e-5} {1e-4,2e-4,3e-4} {1e-4,2e-4,3e-4}Training epochs15101010Earlystop epochs--555Warmup proportion0.010.010.10.10.1Weight decay0.010.010.010.010.01Number of beams---46Length penalty---2.01.0",
"figure_id": "tab_3",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "The results on GLUE. The best results are boldfaced.MiniDisc is distilled from T5 xlarge , and owns larger GFLOPs.",
"figure_data": "MethodGFLOPsSST-2 MRPC STS-B QQP MNLI-m/mm QNLI RTE GLUE Acc F1 SpCorr F1 Acc Acc Acc ScoreT5 base25.41×94.693.090.088.986.7/86.892.974.788.5T5 6L;384H3.1892.290.286.087.381.2/81.788.270.084.6MiniDisc 5% MlmKD 6L;384H MiniLM 6L;384H7.80 3.18 3.183∼8×93.8 92.3 92.189.8 88.7 89.685.3 86.2 85.286.7 87.5 87.082.9/82.7 81.6/82.1 81.2/81.589.2 88.2 88.064.6 67.9 68.684.4 84.3 84.1MlmKD+MiniLM 6L;384H 3.1892.489.286.087.381.7/82.189.167.984.5MINIEND-D 6L;384H3.1892.190.685.887.781.8/82.389.068.684.7w/o L Logit MINIEND-E 6L;384H3.18 3.188×92.2 92.790.1 90.086.6 86.187.6 87.482.2/82.8 81.8/82.189.1 88.868.6 69.384.9 84.8w/o L Logit3.1892.389.986.687.782.5/83.189.269.085.0",
"figure_id": "tab_4",
"figure_label": "3",
"figure_type": "table"
},
{
"figure_caption": "The results on CNN/DailyMail and XSum. The best results are boldfaced.LogitKD is distilled with an asymmetric layer setting, i.e., more encoder layers than decoder layers, for saved performance decline. DQ-BART only quantizes parameter precision to lower one, i.e., 8 bit, but does not reduce parameter amount. Quantization would not give any speedup in GFLOPs though nice reduction in model size.",
"figure_data": "MethodGFLOPsCNN/DailyMail Rg-1 Rg-2 Rg-L Rg-1 Rg-2 Rg-L XSumT5 base25.41×40.1 19.4 31.534.7 12.4 29.7T5 6L;384H3.1835.7 16.8 28.428.68.924.8MlmKD 6L;384H MiniLM 6L;384H3.18 3.188×36.0 17.0 28.7 35.0 16.5 28.028.9 25.99.2 7.525.0 22.5MlmKD+MiniLM 6L;384H 3.1835.8 17.0 28.729.09.125.1MINIEND-D 6L;384H3.1836.2 17.2 28.929.59.225.4w/o L Logit MINIEND-E 6L;384H3.18 3.188×35.7 17.0 28.6 36.1 17.3 28.927.3 28.98.2 9.123.7 24.9w/o L Logit3.1835.8 17.1 28.727.28.023.6BART base12.71×39.4 18.5 30.636.9 14.7 31.9LogitKD 3/1L;768H DQ-BART 8bit4.23 12.71∼3×38.0 16.0 25.2 42.4 19.3 28.832.9 12.4 26.9 38.2 15.7 30.7MINIEND-D 6L;384H3.184×38.5 18.5 29.733.6 12.9 29.2",
"figure_id": "tab_5",
"figure_label": "4",
"figure_type": "table"
},
{
"figure_caption": "The results of model scaling using MINIEND-D. ⇒ denotes a distillation step, which should be operated sequentially otherwise {} is prioritized.• • • ⇒ • • • ⇒ • • • indicates teacher assistant-based distillation and • • • ⇒ {• • • ⇒ • • • } indicates progressive distillation.",
"figure_data": "MethodGLUE Score Rg-1 Rg-2 Rg-L Rg-1 Rg-2 Rg-L CNN/DailyMail XSumT5 6L;384H84.635.7 16.8 28.428.68.924.8T5 12L;384H85.037.2 17.9 29.631.2 10.5 27.0T5 base88.540.1 19.4 31.534.7 12.4 29.7T5 large90.740.6 19.4 31.738.2 15.1 32.9T5 xlarge92.040.8 19.7 32.141.1 17.6 35.5T5 base ⇒T5 6L;384H84.736.2 17.2 28.929.59.225.4T5 large ⇒T5 6L;384H84.536.4 17.4 29.029.49.325.3T5 xlarge ⇒T5 6L;384H84.236.1 17.2 28.829.19.125.1T5 xlarge ⇒T5 12L;384H ⇒T5 6L;384H84.636.6 17.5 29.229.29.125.1T5 large ⇒T5 12L;384H85.538.3 18.4 30.432.4 11.2 27.9T5 xlarge ⇒T5 12L;384H85.238.0 18.4 30.332.2 11.1 27.7T5 xlarge ⇒{T5 large ⇒T5 12L;384H }85.838.4 18.5 30.632.9 11.5 28.3in some cases. Nonetheless, we unearth thatprogressive distillation is more promising interms of consistent performance gains whencomparing T5 xlarge ⇒{T5 large ⇒T5 12L;384H } toT5 xlarge ⇒T5 12L;384H",
"figure_id": "tab_6",
"figure_label": "5",
"figure_type": "table"
}
] | Chen Zhang; Yang Yang; Jingang Wang; Dawei Song | [
{
"authors": "Luisa Bentivogli; Peter Clark; Ido Dagan; Danilo Giampiccolo",
"journal": "",
"ref_id": "b0",
"title": "The seventh PASCAL recognizing textual entailment challenge",
"year": "2011-11-14"
},
{
"authors": "M Daniel; Mona T Cer; Eneko Diab; Iñigo Agirre; Lucia Lopez-Gazpio; Specia",
"journal": "",
"ref_id": "b1",
"title": "Semeval-2017 task 1: Semantic textual similarity multilingual and crosslingual focused evaluation",
"year": "2017-08-03"
},
{
"authors": "Chung Hyung Won; Le Hou; Shayne Longpre; Barret Zoph; Yi Tay; William Fedus; Eric Li; Xuezhi Wang; Mostafa Dehghani; Siddhartha Brahma; Albert Webson; Shane Shixiang; Zhuyun Gu; Mirac Dai; Xinyun Suzgun; Aakanksha Chen; Sharan Chowdhery; Gaurav Narang; Adams Mishra; Vincent Y Yu; Yanping Zhao; Andrew M Huang; Hongkun Dai; Slav Yu; Ed H Petrov; Jeff Chi; Jacob Dean; Adam Devlin; Denny Roberts; Quoc V Zhou; Jason Le; Wei",
"journal": "",
"ref_id": "b2",
"title": "Scaling instruction-finetuned language models",
"year": "2022"
},
{
"authors": "Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova",
"journal": "Association for Computational Linguistics",
"ref_id": "b3",
"title": "BERT: pre-training of deep bidirectional transformers for language understanding",
"year": "2019-06-02"
},
{
"authors": "William B Dolan; Chris Brockett",
"journal": "",
"ref_id": "b4",
"title": "Automatically constructing a corpus of sentential paraphrases",
"year": "2005-10"
},
{
"authors": "Aaron Gokaslan; Vanya Cohen; Ellie Pavlick; Stefanie Tellex",
"journal": "",
"ref_id": "b5",
"title": "Openwebtext corpus",
"year": "2019"
},
{
"authors": "Geoffrey E Hinton; Oriol Vinyals; Jeffrey Dean",
"journal": "",
"ref_id": "b6",
"title": "Distilling the knowledge in a neural network",
"year": "2015"
},
{
"authors": "Xiaoqi Jiao; Yichun Yin; Lifeng Shang; Xin Jiang; Xiao Chen; Linlin Li; Fang Wang; Qun Liu",
"journal": "Association for Computational Linguistics",
"ref_id": "b7",
"title": "Tinybert: Distilling BERT for natural language understanding",
"year": "2020-11"
},
{
"authors": "Hector J Levesque; Ernest Davis; Leora Morgenstern",
"journal": "",
"ref_id": "b8",
"title": "The winograd schema challenge",
"year": "2012-06-10"
},
{
"authors": "Mike Lewis; Yinhan Liu; Naman Goyal; Marjan Ghazvininejad; Abdelrahman Mohamed; Omer Levy; Veselin Stoyanov; Luke Zettlemoyer",
"journal": "Association for Computational Linguistics",
"ref_id": "b9",
"title": "BART: denoising sequence-to-sequence pretraining for natural language generation, translation, and comprehension",
"year": "2020-07-05"
},
{
"authors": "Zheng Li; Zijian Wang; Ming Tan; Ramesh Nallapati; Parminder Bhatia; Andrew O Arnold; Bing Xiang; Dan Roth",
"journal": "Association for Computational Linguistics",
"ref_id": "b10",
"title": "DQ-BART: efficient sequenceto-sequence model via joint distillation and quantization",
"year": "2022-05-22"
},
{
"authors": "Chen Liang; Haoming Jiang; Zheng Li; Xianfeng Tang; Bin Yin; Tuo Zhao",
"journal": "",
"ref_id": "b11",
"title": "Homodistil: Homotopic task-agnostic distillation of pre-trained transformers",
"year": "2023"
},
{
"authors": "Zhenghao Lin; Yeyun Gong; Xiao Liu; Hang Zhang; Chen Lin; Anlei Dong; Jian Jiao; Jingwen Lu; Daxin Jiang; Rangan Majumder; Nan Duan",
"journal": "ACM",
"ref_id": "b12",
"title": "PROD: progressive distillation for dense retrieval",
"year": "2023-04-30"
},
{
"authors": "Frederick Liu; Siamak Shakeri; Hongkun Yu; Jing Li",
"journal": "",
"ref_id": "b13",
"title": "Enct5: Fine-tuning T5 encoder for nonautoregressive tasks",
"year": "2021"
},
{
"authors": "Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov",
"journal": "",
"ref_id": "b14",
"title": "Roberta: A robustly optimized BERT pretraining approach",
"year": "2019"
},
{
"authors": "Seyed-Iman Mirzadeh; Mehrdad Farajtabar; Ang Li; Nir Levine; Akihiro Matsukawa; Hassan Ghasemzadeh",
"journal": "AAAI Press",
"ref_id": "b15",
"title": "Improved knowledge distillation via teacher assistant",
"year": "2020-02-07"
},
{
"authors": "Shashi Narayan; Shay B Cohen; Mirella Lapata",
"journal": "Association for Computational Linguistics",
"ref_id": "b16",
"title": "Don't give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization",
"year": "2018-10-31"
},
{
"authors": "Alec Radford; Jeff Wu; Rewon Child; David Luan; Dario Amodei; Ilya Sutskever",
"journal": "",
"ref_id": "b17",
"title": "Language models are unsupervised multitask learners",
"year": "2019"
},
{
"authors": "Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu",
"journal": "J. Mach. Learn. Res",
"ref_id": "b18",
"title": "Exploring the limits of transfer learning with a unified text-to-text transformer",
"year": "2020"
},
{
"authors": "Pranav Rajpurkar; Jian Zhang; Konstantin Lopyrev; Percy Liang",
"journal": "",
"ref_id": "b19",
"title": "Squad: 100, 000+ questions for machine comprehension of text",
"year": "2016-11-01"
},
{
"authors": "Victor Sanh; Lysandre Debut; Julien Chaumond; Thomas Wolf",
"journal": "",
"ref_id": "b20",
"title": "Distilbert, a distilled version of BERT: smaller, faster, cheaper and lighter",
"year": "2019"
},
{
"authors": "Abigail See; Peter J Liu; Christopher D Manning",
"journal": "Association for Computational Linguistics",
"ref_id": "b21",
"title": "Get to the point: Summarization with pointergenerator networks",
"year": "2017-07-30"
},
{
"authors": "Sam Shleifer; Alexander M Rush",
"journal": "",
"ref_id": "b22",
"title": "Pre-trained summarization distillation",
"year": "2020"
},
{
"authors": "Richard Socher; Alex Perelygin; Jean Wu; Jason Chuang; Christopher D Manning; Andrew Y Ng; Christopher Potts",
"journal": "",
"ref_id": "b23",
"title": "Recursive deep models for semantic compositionality over a sentiment treebank",
"year": "2013-10"
},
{
"authors": "Siqi Sun; Yu Cheng; Zhe Gan; Jingjing Liu",
"journal": "Association for Computational Linguistics",
"ref_id": "b24",
"title": "Patient knowledge distillation for BERT model compression",
"year": "2019-11-03"
},
{
"authors": "Chaofan Tao; Lu Hou; Wei Zhang; Lifeng Shang; Xin Jiang; Qun Liu; Ping Luo; Ngai Wong",
"journal": "Association for Computational Linguistics",
"ref_id": "b25",
"title": "Compression of generative pre-trained language models via quantization",
"year": "2022-05-22"
},
{
"authors": "Yi Tay; Mostafa Dehghani; Q Vinh; Xavier Tran; Dara Garcia; Tal Bahri; Huaixiu Schuster; Neil Steven Zheng; Donald Houlsby; Metzler",
"journal": "",
"ref_id": "b26",
"title": "Unifying language learning paradigms",
"year": "2022"
},
{
"authors": "Hugo Touvron; Thibaut Lavril; Gautier Izacard; Xavier Martinet; Marie-Anne Lachaux; Timothée Lacroix; Baptiste Rozière; Naman Goyal; Eric Hambro; Faisal Azhar; Aurélien Rodriguez; Armand Joulin; Edouard Grave; Guillaume Lample",
"journal": "",
"ref_id": "b27",
"title": "Llama: Open and efficient foundation language models",
"year": "2023"
},
{
"authors": "Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Lukasz Kaiser; Illia Polosukhin",
"journal": "",
"ref_id": "b28",
"title": "Attention is all you need",
"year": "2017-09"
},
{
"authors": "Alex Wang; Amanpreet Singh; Julian Michael; Felix Hill; Omer Levy; Samuel R Bowman",
"journal": "",
"ref_id": "b29",
"title": "GLUE: A multi-task benchmark and analysis platform for natural language understanding",
"year": "2019-05-06"
},
{
"authors": "Wenhui Wang; Hangbo Bao; Shaohan Huang; Li Dong; Furu Wei",
"journal": "Association for Computational Linguistics",
"ref_id": "b30",
"title": "Minilmv2: Multi-head selfattention relation distillation for compressing pretrained transformers",
"year": "2021-08-01"
},
{
"authors": "Wenhui Wang; Furu Wei; Li Dong; Hangbo Bao; Nan Yang; Ming Zhou",
"journal": "",
"ref_id": "b31",
"title": "Minilm: Deep selfattention distillation for task-agnostic compression of pre-trained transformers",
"year": "2020-12-06"
},
{
"authors": "Alex Warstadt; Amanpreet Singh; Samuel R Bowman",
"journal": "Transactions on Association for Computational Linguistics",
"ref_id": "b32",
"title": "Neural network acceptability judgments",
"year": "2019"
},
{
"authors": "Adina Williams; Nikita Nangia; Samuel R Bowman",
"journal": "",
"ref_id": "b33",
"title": "A broad-coverage challenge corpus for sentence understanding through inference",
"year": "2018-06-01"
},
{
"authors": "Mengzhou Xia; Zexuan Zhong; Danqi Chen",
"journal": "Association for Computational Linguistics",
"ref_id": "b34",
"title": "Structured pruning learns compact and accurate models",
"year": "2022-05-22"
},
{
"authors": "Yi Yang; Chen Zhang; Dawei Song",
"journal": "Association for Computational Linguistics",
"ref_id": "b35",
"title": "Sparse teachers can be dense with knowledge",
"year": "2022-12-07"
},
{
"authors": "Aohan Zeng; Xiao Liu; Zhengxiao Du; Zihan Wang; Hanyu Lai; Ming Ding; Zhuoyi Yang; Yifan Xu; Wendi Zheng; Xiao Xia; Weng Lam Tam; Zixuan Ma; Yufei Xue; Jidong Zhai; Wenguang Chen; Peng Zhang; Yuxiao Dong; Jie Tang",
"journal": "",
"ref_id": "b36",
"title": "GLM-130B: an open bilingual pre-trained model",
"year": "2022"
},
{
"authors": "Chen Zhang; Yang Yang; Qifan Wang; Jiahao Liu; Jingang Wang; Yunsen Xian; Wei Wu; Dawei Song",
"journal": "",
"ref_id": "b37",
"title": "Minidisc: Minimal distillation schedule for language model compression",
"year": "2022"
},
{
"authors": "Shengqiang Zhang; Xingxing Zhang; Hangbo Bao; Furu Wei",
"journal": "Association for Computational Linguistics",
"ref_id": "b38",
"title": "Attention temperature matters in abstractive summarization distillation",
"year": "2022-05-22"
},
{
"authors": "Yanli Zhao; Andrew Gu; Rohan Varma; Liang Luo; Chien-Chin Huang; Min Xu; Less Wright; Hamid Shojanazeri; Myle Ott; Sam Shleifer; Alban Desmaison; Can Balioglu; Bernard Nguyen; Geeta Chauhan; Yuchen Hao; Shen Li",
"journal": "",
"ref_id": "b39",
"title": "Pytorch FSDP: experiences on scaling fully sharded data parallel",
"year": "2023"
}
] | [
{
"formula_coordinates": [
2,
321.28,
182.32,
187.99,
147.04
],
"formula_id": "formula_0",
"formula_text": "MSA(X; W Q , W K , W V ) = A i SelfAttn(X; W Q i , W K i )XW V i W O i , SelfAttn(X; W Q i , W K i ) = Softmax(XW Q i W K i X /d A ), FFN(X; W I , W O ) = I j g(XW I j )W O j ,"
},
{
"formula_coordinates": [
2,
305.49,
363.85,
220.28,
30.7
],
"formula_id": "formula_1",
"formula_text": "W Q i , W K i , W V i ∈ R d×d A , W O i ∈ R d A ×d among A heads,"
},
{
"formula_coordinates": [
2,
307.26,
515.89,
216.04,
97.94
],
"formula_id": "formula_2",
"formula_text": "MCA(Z, E; W Q , W K , W V ) = A i CrossAttn(Z, E; W Q i , W K i )EW V i W O i , CrossAttn(Z, E; W Q i , W K i ) = Softmax(ZW Q i W K i E /d A ),"
},
{
"formula_coordinates": [
2,
306.14,
630.13,
218.27,
30.09
],
"formula_id": "formula_3",
"formula_text": "W Q i , W K i , W V i ∈ R d×d A , W O i ∈ R d A ×d ."
},
{
"formula_coordinates": [
3,
80.52,
579.07,
198.96,
146.21
],
"formula_id": "formula_4",
"formula_text": "L SelfAttn (S; T , D Z ) = E Z∼D Z R k=1 KL(Reln(Z; T W Q k ), Reln(Z; S W Q k )) + KL(Reln(Z; T W K k ), Reln(Z; S W K k )) + KL(Reln(Z; T W V k ), Reln(Z; S W V k )), Reln(Z; T W Q k ) = Softmax(Z T W Q k T W Q k Z /d R ),"
},
{
"formula_coordinates": [
3,
312.97,
595.11,
204.45,
90.12
],
"formula_id": "formula_5",
"formula_text": "L CrossAttn (S; T , D Z , D E ) = E Z∼D Z ,E∼D E R k=1 KL(Reln(Z; T W Q k ), Reln(Z; S W Q k )) + KL(Reln(E; T W K k ), Reln(E; S W K k )) + KL(Reln(E; T W V k ), Reln(E; S W V k )),"
},
{
"formula_coordinates": [
3,
326.31,
760.99,
177.94,
13.29
],
"formula_id": "formula_6",
"formula_text": "L Logit (S; T , D Z ) + L SelfAttn (S; T , D Z ),"
},
{
"formula_coordinates": [
4,
92.55,
282.11,
174.91,
31.38
],
"formula_id": "formula_7",
"formula_text": "L Logit (S; T , D Z ) + L SelfAttn (S; T , D Z ) + L CrossAttn (S; T , D Z , D E ),"
},
{
"formula_coordinates": [
4,
72.56,
683.83,
214.87,
31.38
],
"formula_id": "formula_8",
"formula_text": "L(S; T , D Z , D E ) = L Logit (S; T , D Z )+ L SelfAttn (S; T , D Z ) + L CrossAttn (S; T , D Z , D E ),"
},
{
"formula_coordinates": [
4,
118.83,
744.31,
121.34,
30.48
],
"formula_id": "formula_9",
"formula_text": "L Logit (S; T , D Z ) = E Z∼D Z CE(Z S W E , Z T W E ),"
},
{
"formula_coordinates": [
4,
313.6,
346.46,
203.36,
31.38
],
"formula_id": "formula_10",
"formula_text": "L(S; T , D Z , D X ) = L Logit (S; T , D Z )+ L SelfAttn (S; T , D Z ) + L EncSelfAttn (S; T , D X ),"
}
] |
10.3390/s22239384 | [
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b0",
"b1",
"b2",
"b3",
"b4",
"b5",
"b6",
"b7",
"b8",
"b9",
"b10",
"b11",
"b13",
"b14",
"b15",
"b13",
"b16",
"b17",
"b18",
"b19",
"b20",
"b21",
"b22",
"b23",
"b15",
"b24",
"b25",
"b26",
"b27",
"b28",
"b31",
"b32",
"b28",
"b23",
"b25",
"b26",
"b29",
"b30",
"b26",
"b29",
"b32",
"b27",
"b33",
"b14",
"b14",
"b15",
"b13",
"b34",
"b30",
"b34",
"b12"
],
"table_ref": [],
"text": "In recent years, object detection with Unmanned Aerial Vehicles (UAVs) has attracted much attention in computer vision research and has provided many benefits in various domains. Such as fire smoke detection [1], military [2], urban surveillance [3], and agriculture [4], [5]. However, it is not easy to accurately detect objects with UAVs that capture object images using the camera from a very high followed by a widely geographic one. Most of the current traditional object detection methods are based only on the sliding-window paradigm and handcrafted features. Like Viola-Jones [6], Histogram of Oriented Gradients (HOG) [7], Scale-Invariant Feature Transform (SIFT) [8], [9], Haar [10], [11], which has made significant progress in the research field of object detection. However, this method takes time and effort to achieve the robustness of feature representation and is still vulnerable to failure when handling variations in data obtained from the UAVs. What is urgently needed by object detection systems with UAVs today is an accurate method capable of processing image data end-to-end. Currently, deep learning [12] is one of the solutions to answer these needs. The CNN can process visual data accurately without the need to go through a separately feature extraction process and has proven can outperform traditional methods in ImageNet Large Scale Visual Recognition Challenge (ILSVRC) [14]. The progress is inseparable from the availability of large-scale data, such as Microsoft Common Objects in Context (COCO) [15], Pascal Visual Object Classes (PASCAL VOC) [16], ImageNet [14], as well as the availability of computing resources, and driven by ongoing research with the proposed various network architectures. Such as VGG [17], GoogLeNet [18], Residual Networks (ResNets) [19], [20], ResNeXt [21], Cross Stage Partial Network (CSPNet) [22], and EfficientNet [23] in the classification tasks which is widely used as a backbone layer for feature extraction in the object detection tasks. Object detection based on deep learning methods generally divided into two: the one-stage detector and the two-stage detector. The two-stage detector method predicts the bounding box through the process of region proposal and then classifies it to detect the class from the object. Such as the Region-based Convolutional Neural Network (R-CNN) proposed by Ross Girshick et al. [24] is the first deep learning based object detection method. R-CNN In the PASCAL VOC 2010 challenge [16] was able to outperform traditional detector methods, such as Deformable Parts Model (DPM) [25], which at that time occupied the first position. This progress is also driven by the development of other popular methods, such as Fast R-CNN [26] and Faster R-CNN [27], which is average have a high prediction accuracy. However, that method is still relatively slow in the detection process. That is deficiency can overcome by one-stage detector methods, such as RetinaNet [28], You Only Look Once (YOLO) [29]- [32] and Single Shot MultiBox Detector (SSD) [33], which are very fast when predicting objects. Such as the YOLO method proposed by Joseph Redmon et al. [29] can predict multiple bounding boxes and class probabilities simultaneously, which makes it very fast during the detection process. However, YOLO in the first version still has several localization errors compared with the region proposal method [24], [26], [27]. So development was also carried out to reduce the shortcomings of previous versions, such as YOLOv2 [30] and YOLOv3 [31]. YOLOv2 uses darknet-19 as backbone layers that consist of 19 convolutional layers and 5 max-pooling. While YOLOv3 is a further development of YOLOv2, which can predict the bounding boxes with multi-scale prediction and uses darknet-53 in the backbone layer. YOLOv3 can produce a balance between accuracy and detection speed. The result of YOLOv3 can get the average precision better than Faster R-CNN [27], YOLOv2 [30], SSD [33], and faster than RetinaNet [28] and Region-based Fully Convolutional Network (R-FCN) [34] on the testing of the COCO dataset [15]. However, the data obtained from the UAVs is not like that of data from COCO [15], PASCAL VOC [16], and ImageNet [14], that dominated by global image objects with large individual objects. UAVs capture object images from a very high camera and produce data with varying perspectives viewing. As illustrated in Figure 1, the image data captured by UAVs is dominated by small object sizes, which makes the image contain less clear features and the density of objects with different illumination levels is also a challenge for detecting objects with the images obtained by UAVs.\nMotivated by the challenge above, we aim to improve the performance of the YOLOv3 method for detecting objects images obtained from UAVs.\nWe added Spatial pyramid pooling (SPP) [35] at the end of the darknet-53 backbone architecture to achieve a more efficient feature extraction process. The details objective and the contribution of this study are explained as follows:\n1) We improved the performance of YOLOv3 [31] by adding SPP [35] on the end layer of the darknet-53 backbone to obtain more efficient feature extraction process in object detection tasks with UAVs. 2) We also show an evaluation study of different versions YOLOv3 method on object detection tasks with UAVs, including YOLOv3 with SPP, YOLOv3, and YOLOv3-tiny which we analyzed with the VisDrone2019-Det dataset [13]."
},
{
"figure_ref": [],
"heading": "Research Method",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "YOLOv3",
"publication_ref": [
"b28"
],
"table_ref": [],
"text": "You Only Look Once (YOLO) [29] consists of a backbone layer for feature extraction and a head layer for detection. YOLO predicts objects by mapping the image input pixels to 𝑆𝑥𝑆 grid. Each grid cell predicts B bounding box and confidence score, which is described in the following equation,\n𝑐𝑜𝑛𝑓𝑖𝑑𝑒𝑛𝑐𝑒 = P r (Object)*IoU ( truth predict )(1)\nP r (Object) shows the probability of an object inside the bounding box, and 𝐼𝑜𝑈 𝑝𝑟𝑒𝑑𝑖𝑐𝑡 𝑡𝑟𝑢𝑡ℎ shows the Intersection over Union (IoU) of ground truth and box prediction. The confidence will have a value of 0 if there are no objects in the grid cell and a value of 1 if there are objects. The bounding box consists of 5 parameters (𝑥, 𝑦, 𝑤, ℎ, 𝑐𝑜𝑛𝑓𝑖𝑑𝑒𝑛𝑐𝑒), the width and height are represented by 𝑤, ℎ, and 𝑥, 𝑦 represents the center coordinates of the bounding box. In the end, the results of predicted confidence will represent the Intersections over Union (𝐼𝑜𝑈) between the predicted box and the ground truth boxes. At the same time, each grid cell also predicts C conditional class probabilities that described in the following equation,\n𝐶𝑙𝑎𝑠𝑠 𝑝𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦 = 𝑃𝑟(𝐶𝑙𝑎𝑠𝑠 𝑖 |𝑂𝑏𝑗𝑒𝑐𝑡) (2)\nThe predicting process of conditional class probabilities C in each grid cell is conditioned if there are objects in the grid cell. And the testing process will multiply of conditional class probability with the predicted value of the box confidence to get the confidence score class specific in each box. As represented by equation (3), which encodes the probability of the class appearing in the box and also represents how well the predicted box matches the object."
},
{
"figure_ref": [
"fig_2"
],
"heading": "𝑃𝑟(𝐶𝑙𝑎𝑠𝑠",
"publication_ref": [
"b28",
"b29",
"b29",
"b18",
"b35"
],
"table_ref": [],
"text": "𝑖 |𝑂𝑏𝑗𝑒𝑐𝑡) * P r (Object) * IoU ( truth predict ) = 𝑃𝑟(𝐶𝑙𝑎𝑠𝑠 𝑖 ) * IoU ( truth predict ) (3)\nYOLOv3 is an improvement over its predecessors [29], [30], which involves different architecture and is more accurate in the detection process. YOLOv3 uses darknet-53 for the feature extraction process, as represented by Figure 3. Darknet-53 uses 3x3 and 1x1 convolutional layers of darknet-19 in YOLOv2 [30], which is organized by residual networks [19]. YOLOv3 predicts bounding boxes with three different scales using ideas from Feature Pyramid Network (FPN) [36], where the final feature map results from the convolutional layers will predict 3D tensors which are coded as bounding boxes, objectness, and class predictions. Each scales predict 3 squares which are represented as 𝑆𝑥𝑆𝑥(3 * (4 + 1 + 80)), where 𝑆𝑥𝑆 represent the size of the feature map, 4 bounding boxes, 1 objectness prediction, and 80 which is illustrated as the total class prediction. "
},
{
"figure_ref": [
"fig_3"
],
"heading": "Spatial Pyramid Pooling",
"publication_ref": [
"b36",
"b37",
"b34",
"b34"
],
"table_ref": [],
"text": "Spatial Pyramid Pooling (SPP) [37], [38] in CNN was first introduced by [35]. The process of SPP is represented in Figure 4, which receives the input feature map from the convolutional layers. Afterward, in each spatial bin, the pooling layer responds to each filter to produce output (kMvector). M represents the number of bins, k is the number of filters in the last convolutional layer, and the fixed dimensional vector is the input to the fully connected layers. SPP has some extraordinary properties for deep CNN compared with general networks that use pooling sliding windows. Based on research that conducted by He et al. [35] SPP-net is capable of producing the output of fixed-length regardless of the input size and uses multi-level spatial bins. Meanwhile, pooling sliding windows only use single window sizes. In this study, we aim to add SPP to the final layer of darknet-53 in YOLOv3 to improve performance in object detection tasks with UAVs. "
},
{
"figure_ref": [
"fig_1"
],
"heading": "Architecture Model",
"publication_ref": [],
"table_ref": [],
"text": "In this study, we aim to add an SPP layer to the final darknet-53 layer to improve the performance of YOLOv3 in object detection tasks with data obtained from UAVs. The details of the architecture in this study is represented in Figure 2. The first process is feature extraction from the input image with darknet-53. Then the SPP layer is added to the final darknet-53 layer to improve the feature extraction process. In the end, the results of SPP is a feature map that uses as input into head detection of YOLOv3 for predicting the bounding boxes and class probabilities."
},
{
"figure_ref": [],
"heading": "Loss Function",
"publication_ref": [],
"table_ref": [],
"text": "In Los Function is used to determine the state of the training model in each iteration to calculate the difference between the value of predicted and the value of ground truth. As represented in equation ( 4), this study split three loss functions: (𝑙 𝑐𝑜𝑜𝑟𝑑 , 𝑙 𝐼𝑜𝑈 , 𝑙 𝑐𝑙𝑎𝑠𝑠 ). The notation of 𝑙 𝑐𝑜𝑜𝑟𝑑 represents the coordinate prediction errors, 𝑙 𝐼𝑜𝑈 is IoU errors, and 𝑙 𝑐𝑙𝑎𝑠𝑠 is the classification errors."
},
{
"figure_ref": [],
"heading": "𝐿𝑜𝑠𝑠 = 𝑙 𝑐𝑜𝑜𝑟𝑑 + 𝑙 𝐼𝑜𝑈 + 𝑙 𝑐𝑙𝑎𝑠𝑠 (4)",
"publication_ref": [],
"table_ref": [],
"text": "The coordinate prediction error is represented in the following equation,\n𝑙 𝑐𝑜𝑜𝑟𝑑 = 𝜆 𝑐𝑜𝑜𝑟𝑑 ∑ ∑ 𝛪 𝑖𝑗 𝑜𝑏𝑗 [(𝑥 𝑖 -𝑥 ̂𝑖) 2 + (𝑦 𝑖 - 𝐵 𝑗=0 𝑆 2 𝑖=0 𝑦 ̂𝑖) 2 ] + 𝜆 𝑐𝑜𝑜𝑟𝑑 ∑ ∑ 𝛪 𝑖𝑗 𝑜𝑏𝑗 [(√𝑤 𝑖 -√𝑤 ̂𝑖) 2 + 𝐵 𝑗=0 𝑆 2 𝑖=0 (√ℎ 𝑖 -√ℎ ̂𝑖) 2 ] (5)\nWhere 𝜆 𝑐𝑜𝑜𝑟𝑑 is the weight coordinate error, 𝑆 2 is the number of grid cells for each detection layer, and 𝐵 is the number of bounding boxes in each grid cell. (𝑥 𝑖 , 𝑦 𝑖 , 𝑥 ̂𝑖, 𝑦 ̂𝑖) represents the center coordinate of the ground truth and the target object. Whereas (ℎ 𝑖 , 𝑤 𝑖 , ℎ ̂𝑖, 𝑤 ̂𝑖) represents the width and height of the ground truth and the target prediction box. For IoU errors and Classification errors are denoted by equations ( 6) and ( 7) as follows,\n𝑙 𝐼𝑜𝑈 = 𝜆 𝐼𝑜𝑈 ∑ ∑ 𝛪 𝑖𝑗 𝑜𝑏𝑗 [(𝐶 𝑖 -𝐶 ̂𝑖) 2 ] 𝐵 𝑗=0 𝑆 2 𝑖=0 () + 𝜆 𝑛𝑜𝑜𝑏𝑗 ∑ ∑ 𝛪 𝑖𝑗 𝑛𝑜𝑜𝑏𝑗 (𝐶 𝑖 -𝐶 ̂𝑖) 2 𝐵 𝑗=0 𝑆 2 𝑖=0 (6) 𝑙 𝑐𝑙𝑎𝑠𝑠 = 𝜆 𝑐𝑙𝑎𝑠𝑠 ∑ 𝛪 𝑖 𝑜𝑏𝑗 ∑ (𝑝 𝑖 (𝑐) - 𝑐 ∈ 𝑐𝑙𝑎𝑠𝑠𝑒𝑠 𝑆 2 𝑖=0 𝑝̂𝑖(𝑐)) 2 (7)\nThe IoU error indicates the degree of overlap between the ground truth and the prediction box. If the anchor box indicates that there is a target located in grid cells (𝑖, 𝑗), then the value of Ι 𝑖𝑗 𝑜𝑏𝑗 is 1, and otherwise, the value is 0. The notation of 𝜆 𝑛𝑜𝑜𝑏𝑗 represents a belief penalty if the prediction box contains no objects, and also misclassification, which represents classification accuracy. Where 𝑝 𝑖 (𝑐) is the value of true probability and 𝑝̂𝑖(𝑐) is the predicted value of the target."
},
{
"figure_ref": [],
"heading": "Experiment",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [
"fig_4"
],
"heading": "Dataset",
"publication_ref": [
"b12"
],
"table_ref": [],
"text": "In this study, we used the VisDrone2019-Det dataset [13], which consisted of 10,209 images, 6,471 for training, 548 for validation, and 3,190 for testing. The VisDrone2019-Det dataset consists of ten object categories: pedestrian, person, bicycle, car, van, truck, tricycle, awningtricycle, bus, and motorcycle. As shown in Figure 5. The VisDrone dataset has several objects with different levels of occlusion in each category, which becomes a challenge to detect objects with UAVs. In this study, we use a training set for the training process and evaluate it with set validation. "
},
{
"figure_ref": [],
"heading": "Metric Evaluation",
"publication_ref": [
"b10"
],
"table_ref": [],
"text": "To evaluate each method, we used the parameters Precision (𝑃), Recall (𝑅), Average Precision (𝐴𝑃), and mean Average Precision (𝑚𝐴𝑃) with 0.5 Intersections over Union (𝐼𝑜𝑈). The details of 𝑃 and 𝑅 parameters are described by the following equation,\nPrecision (P) = TP TP+FP (8) Recall (R) = TP TP+FN (9)\nWhere 𝑇𝑃 is true positive, that is the correct detection of the ground truth bounding box, 𝐹𝑃 is false positive, that is object was detected but misplaced. 𝐹𝑁 is false negative, which means that the basic ground truth of the bounding box was not detected. 𝐴𝑃 and 𝑚𝐴𝑃 parameters are described by the following equation, (11) Where 𝐴𝑃 is the average value of 𝑃 and 𝑅. 𝑚𝐴𝑃 is the average of the 𝐴𝑃 used to measure all class categories in the dataset and is the metric used to measure the accuracy of object detection with UAV."
},
{
"figure_ref": [],
"heading": "Experimental Details",
"publication_ref": [
"b14"
],
"table_ref": [],
"text": "For the experimental procedures in this study, we use a pre-trained model from the COCO dataset [15]. In the training phase, we use stochastic gradient descent as optimization with momentum 0.9, batch size 16, learning rate 0.01, and training iterations of 50 epochs with an input scale 640x640. The framework used in this study is PyTorch with a Tesla T4 GPU for the training and validation process. "
},
{
"figure_ref": [
"fig_6"
],
"heading": "Results and Discussion",
"publication_ref": [],
"table_ref": [
"tab_0",
"tab_0",
"tab_1",
"tab_2"
],
"text": "Table 1, shows the results of training with an input scale of 640x640, which can be concluded in several findings. First, the YOLOv3 with SPP obtained a higher mAP of 0.6% than YOLOv3. These results prove that the addition of the SPP architecture to the YOLOv3 can improve the performance of the object detection model. Second, the YOLOv3-tiny obtained the mAP value of 26.6% much lower than YOLOv3 with SPP and 26% from YOLOv3. These results, one of which is influenced by the depth of the network. YOLOv3-tiny is a lightweight model with fewer parameters and depth. So that able to obtain faster detection processing. However, inversely proportional to the obtained accuracy. The details of the detection results in Table 1, are shown in Table 2. When observed from the total of 10 detection classes, YOLOv3 with SPP excels in six classes: pedestrian, bicycle, van, tricycle, awn, and bus compared to YOLOv3, which only excels in four classes: people, cars, trucks, and motorcycles. Whereas the results of YOLOv3-tiny is lower than YOLOv3 with SPP and YOLOv3 from all detections. For one of the results of visualization from the detection is represented in Figure 6..\nTo obtain a more in-depth analysis, we also validate each model with different input scales. Our goal is to find out if the image scale also affects each object detection model. As reported in Table 3. The YOLOv3 with SPP that we propose is still superior to YOLOv3, with an mAP difference of 0.8% on a 960x960 scale, and 0.9% on a 1280x1280 scale. Whereas YOLOv3-tiny is still lower on both scales compared to the results of YOLOv3 with SPP and YOLOv3."
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "This study aims to improve the performance of YOLOv3 in object detection tasks with UAVs by adding an SPP layer at the end of the darknet-53. We trained three different models: YOLOv3 with SPP, YOLOv3, and YOLOv3-tiny with Visdrone2019-Det training set and evaluated them with a validation set at an input scale of 640x640. The results of YOLOv3 with SPP can improve the performance object detection model, with the results of mAP accuracy of 0.6% more height than YOLOv3 and 26.6% than YOLOv3-tiny. The YOLOv3 with SPP also can maintain accuracy at different input scales, which can outperform the results of YOLOv3 with a difference of 0.8% mAP accuracy on a 960x960 input scale and 0.9% on a 1280x1280 scale. Meanwhile, YOLOv3-tiny is still lower on both scales compared to the results of YOLOv3 with SPP and YOLOv3. The results of YOLOv3 with SPP prove that the addition of the SPP layers to YOLOv3 can improve the performance of object detection models with data obtained from UAVs even with different input scales of image."
}
] | Object detection with Unmanned Aerial Vehicles (UAVs) has attracted much attention in the research field of computer vision. However, not easy to accurately detect objects with data obtained from UAVs, which capture images from very high altitudes, making the image dominated by small object sizes, that difficult to detect. Motivated by that challenge, we aim to improve the performance of the one-stage detector YOLOv3 by adding a Spatial Pyramid Pooling (SPP) layer on the end of the backbone darknet-53 to obtain more efficient feature extraction process in object detection tasks with UAVs. We also conducted an evaluation study on different versions of YOLOv3 methods. Includes YOLOv3 with SPP, YOLOv3, and YOLOv3-tiny, which we analyzed with the VisDrone2019-Det dataset. Here we show that YOLOv3 with SPP can get results mAP 0.6% higher than YOLOv3 and 26.6% than YOLOv3-Tiny at 640x640 input scale and is even able to maintain accuracy at different input image scales than other versions of the YOLOv3 method. Those results prove that the addition of SPP layers to YOLOv3 can be an efficient solution for improving the performance of the object detection method with data obtained from UAVs. | YOLOv3 with Spatial Pyramid Pooling for Object Detection with Unmanned Aerial Vehicles | [
{
"figure_caption": "Fig. 1 .1Fig. 1. Object detection challenges with UAVs: (a) small objects, (b) object density, and (c) differentilluminations[13] ",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 2 .2Fig. 2. Architecture method in this study Currently, deep learning methods have become a focus of research in the field of object detection. In particular, deep learning methods based on Convolutional Neural Networks (CNN).The CNN can process visual data accurately without the need to go through a separately feature extraction process and has proven can outperform traditional methods in ImageNet Large Scale Visual Recognition Challenge (ILSVRC)[14]. The progress is inseparable from the availability of large-scale data, such as Microsoft Common Objects in Context (COCO)[15], Pascal Visual Object Classes (PASCAL VOC)[16], ImageNet[14], as well as the availability of computing resources, and driven by ongoing research with the proposed various network architectures. Such as VGG[17], GoogLeNet[18], Residual Networks (ResNets)[19],[20], ResNeXt[21], Cross Stage Partial Network (CSPNet)[22], and EfficientNet[23] in the classification tasks which is widely used as a backbone layer for feature extraction in the object detection tasks. Object detection based on deep learning methods generally divided into two: the one-stage detector and the two-stage detector. The two-stage detector method predicts the bounding box through the process of region proposal and then classifies it to detect the class from the object. Such as the Region-based Convolutional Neural Network (R-CNN) proposed by Ross Girshick et al.[24] is the first deep learning based object detection method. R-CNN In the PASCAL VOC 2010 challenge[16] was able to outperform traditional detector methods, such as Deformable Parts Model (DPM)[25], which at that time occupied the first position. This progress is also driven by the development of other popular methods, such as Fast R-CNN[26] and Faster R-CNN[27], which is average have a high prediction accuracy. However, that method is still relatively slow in the detection process. That is deficiency can overcome by one-stage detector",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 3 .3Fig. 3. Darknet-53 Architecture",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 4 .4Fig. 4. Spatial Pyramid Pooling architecture",
"figure_data": "",
"figure_id": "fig_3",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 5 .5Fig. 5. Visdrone dataset with different levels of occlusion",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "AP = ∑ (𝑅 𝑛+1 -𝑅 𝑛 ) 𝑛 𝑃(𝑅 ̃) 𝑅 ̃:𝑅 ̃≥𝑅 𝑛+1 𝑚𝑎𝑚𝑎𝑥𝑠𝑎𝑥",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 6 .6Fig. 6. Detection visualization.",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "Training results.",
"figure_data": "ModelPrecisionRecallmAP_50YOLOv350.140.239.7YOLOv3-tiny22.917.913.7YOLOv3-SPP49.341.440.3",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Detection results.",
"figure_data": "ModelPedestrianPeopleBicycleCarVanTruckTricycleAwnBusMotorYOLOv349.339.816.378.341.838.325.312.249.845.8YOLOv3-tiny1615.62.446.7110.16.52.810.616.6YOLOv3-SPP49.439.417.878.142.837.926.714.350.645.6",
"figure_id": "tab_1",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "Validation results with different input scales.",
"figure_data": "ModelInput scalesPrecisionRecallmAP_50YOLOv3960x96048.740.2381280x128049.439.838.2YOLOv3-tiny960x96025.821.115.61280x128025.621.916.1YOLOv3-SPP960x96048.341.338.81280x128047.242.839.1",
"figure_id": "tab_2",
"figure_label": "3",
"figure_type": "table"
}
] | Wahyu Pebrianto; Panca Mudjirahardjo; Sholeh Hadi Pramono; Raden Arief Rahmadwati; Setyawan | [
{
"authors": "M Mukhiddinov; A B Abdusalomov; J Cho",
"journal": "Sensors",
"ref_id": "b0",
"title": "A Wildfire Smoke Detection System Using Unmanned Aerial Vehicle Images Based on the Optimized YOLOv5",
"year": "2022-12"
},
{
"authors": "P Gupta; B Pareek; G Singal; D V Rao",
"journal": "Multimed. Tools Appl",
"ref_id": "b1",
"title": "Edge device based Military Vehicle Detection and Classification from UAV",
"year": "2022"
},
{
"authors": "W L Leong; N Martinel; S Huang; C Micheloni; G L Foresti; R S H Teo",
"journal": "J. Intell. Robot. Syst",
"ref_id": "b2",
"title": "An Intelligent Auto-Organizing Aerial Robotic Sensor Network System for Urban Surveillance",
"year": "2021-06"
},
{
"authors": "U R Mogili; B B V L Deepak",
"journal": "Procedia Comput. Sci",
"ref_id": "b3",
"title": "Review on Application of Drone Systems in Precision Agriculture",
"year": "2018"
},
{
"authors": "L El Hoummaidi; A Larabi; K Alam",
"journal": "Heliyon",
"ref_id": "b4",
"title": "Using unmanned aerial systems and deep learning for agriculture mapping in Dubai",
"year": "2021"
},
{
"authors": "P Viola; M Jones",
"journal": "Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit",
"ref_id": "b5",
"title": "Rapid object detection using a boosted cascade of simple features",
"year": "2001"
},
{
"authors": "N Dalal; B Triggs",
"journal": "",
"ref_id": "b6",
"title": "Histograms of oriented gradients for human detection",
"year": "2005"
},
{
"authors": "D G Lowe",
"journal": "",
"ref_id": "b7",
"title": "Object recognition from scaleinvariant features",
"year": "1999"
},
{
"authors": "T Nguyen; E.-A Park; J Han; D.-C Park; S.-Y Min",
"journal": "",
"ref_id": "b8",
"title": "Object Detection Using Scale Invariant Feature Transform",
"year": "2014"
},
{
"authors": "R Lienhart; J Maydt",
"journal": "",
"ref_id": "b9",
"title": "An extended set of Haar-like features for rapid object detection",
"year": "2002"
},
{
"authors": "L Arreola; G Gudino; G Flores",
"journal": "",
"ref_id": "b10",
"title": "Object recognition and tracking using Haar-like Features Cascade Classifiers: Application to a quad-rotor UAV",
"year": "2022-05"
},
{
"authors": "Y Lecun; Y Bengio; G Hinton",
"journal": "Nature",
"ref_id": "b11",
"title": "Deep learning",
"year": "2015-05"
},
{
"authors": "D Du",
"journal": "",
"ref_id": "b12",
"title": "VisDrone-DET2019: The Vision Meets Drone Object Detection in Image Challenge Results",
"year": "2019-10"
},
{
"authors": "O Russakovsky",
"journal": "Int. J. Comput. Vis",
"ref_id": "b13",
"title": "ImageNet Large Scale Visual Recognition Challenge",
"year": "2015-12"
},
{
"authors": "T.-Y Lin",
"journal": "Eccv",
"ref_id": "b14",
"title": "Microsoft COCO: Common Objects in Context",
"year": "2014-06"
},
{
"authors": "M Everingham; L Van Gool; C K I Williams; J Winn; A Zisserman",
"journal": "Int. J. Comput. Vis",
"ref_id": "b15",
"title": "The Pascal Visual Object Classes (VOC) Challenge",
"year": "2010-06"
},
{
"authors": "K Simonyan; A Zisserman",
"journal": "",
"ref_id": "b16",
"title": "Very Deep Convolutional Networks for Large-Scale Image Recognition",
"year": "2014-09"
},
{
"authors": "C Szegedy",
"journal": "",
"ref_id": "b17",
"title": "Going deeper with convolutions",
"year": "2015-06"
},
{
"authors": "K He; X Zhang; S Ren; J Sun",
"journal": "",
"ref_id": "b18",
"title": "Deep Residual Learning for Image Recognition",
"year": "2016-06"
},
{
"authors": "K He; X Zhang; S Ren; J Sun",
"journal": "LNCS",
"ref_id": "b19",
"title": "Identity Mappings in Deep Residual Networks",
"year": "2016"
},
{
"authors": "S Xie; R Girshick; P Dollar; Z Tu; K He",
"journal": "Janua",
"ref_id": "b20",
"title": "Aggregated Residual Transformations for Deep Neural Networks",
"year": "2017-07"
},
{
"authors": "C.-Y Wang; H.-Y. Mark Liao; Y.-H Wu; P.-Y Chen; J.-W Hsieh; I.-H Yeh",
"journal": "",
"ref_id": "b21",
"title": "CSPNet: A New Backbone that can Enhance Learning Capability of CNN",
"year": "2020-06"
},
{
"authors": "M Tan; Q V Le",
"journal": "",
"ref_id": "b22",
"title": "EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks",
"year": "2019-05"
},
{
"authors": "R Girshick; J Donahue; T Darrell; J Malik",
"journal": "",
"ref_id": "b23",
"title": "Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation",
"year": "2014-06"
},
{
"authors": "P F Felzenszwalb; R B Girshick; D Mcallester; D Ramanan",
"journal": "IEEE Trans. Pattern Anal. Mach. Intell",
"ref_id": "b24",
"title": "Object Detection with Discriminatively Trained Part-Based Models",
"year": "2010-09"
},
{
"authors": "R Girshick",
"journal": "",
"ref_id": "b25",
"title": "Fast R-CNN",
"year": "2015-12"
},
{
"authors": "S Ren; K He; R Girshick; J Sun",
"journal": "IEEE Trans. Pattern Anal. Mach. Intell",
"ref_id": "b26",
"title": "Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks",
"year": "2017-06"
},
{
"authors": "T.-Y Lin; P Goyal; R Girshick; K He; P Dollar",
"journal": "IEEE Trans. Pattern Anal. Mach. Intell",
"ref_id": "b27",
"title": "Focal Loss for Dense Object Detection",
"year": "2020-02"
},
{
"authors": "J Redmon; S Divvala; R Girshick; A Farhadi",
"journal": "",
"ref_id": "b28",
"title": "You Only Look Once: Unified, Real-Time Object Detection",
"year": "2016-06"
},
{
"authors": "J Redmon; A Farhadi",
"journal": "",
"ref_id": "b29",
"title": "YOLO9000: Better, Faster, Stronger",
"year": "2017-07"
},
{
"authors": "J Redmon; A Farhadi",
"journal": "",
"ref_id": "b30",
"title": "YOLOv3: An Incremental Improvement",
"year": "2018-04"
},
{
"authors": "W Pebrianto; P Mudjirahardjo; S H Pramono",
"journal": "",
"ref_id": "b31",
"title": "YOLO Method Analysis and Comparison for Real-Time Human Face Detection",
"year": "2022-08"
},
{
"authors": "W Liu",
"journal": "",
"ref_id": "b32",
"title": "SSD: Single Shot MultiBox Detector",
"year": "2016"
},
{
"authors": "J Dai; Y Li; K He; J Sun",
"journal": "Adv. Neural Inf. Process. Syst",
"ref_id": "b33",
"title": "R-FCN: Object detection via region-based fully convolutional networks",
"year": "2016"
},
{
"authors": "K He; X Zhang; S Ren; J Sun",
"journal": "IEEE Trans. Pattern Anal. Mach. Intell",
"ref_id": "b34",
"title": "Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition",
"year": "2015"
},
{
"authors": "T.-Y Lin; P Dollár; R Girshick; K He; B Hariharan; S Belongie",
"journal": "",
"ref_id": "b35",
"title": "Feature Pyramid Networks for Object Detection",
"year": "2016-12"
},
{
"authors": "K Grauman; T Darrell",
"journal": "",
"ref_id": "b36",
"title": "The pyramid match kernel: Discriminative classification with sets of image features",
"year": "2005"
},
{
"authors": "S Lazebnik; C Schmid; J Ponce",
"journal": "Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit",
"ref_id": "b37",
"title": "Beyond bags of features: Spatial pyramid matching for recognizing natural scene categories",
"year": "2006"
}
] | [
{
"formula_coordinates": [
3,
96.08,
422.05,
191.7,
14.79
],
"formula_id": "formula_0",
"formula_text": "𝑐𝑜𝑛𝑓𝑖𝑑𝑒𝑛𝑐𝑒 = P r (Object)*IoU ( truth predict )(1)"
},
{
"formula_coordinates": [
3,
96.08,
637.78,
191.7,
11.39
],
"formula_id": "formula_1",
"formula_text": "𝐶𝑙𝑎𝑠𝑠 𝑝𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦 = 𝑃𝑟(𝐶𝑙𝑎𝑠𝑠 𝑖 |𝑂𝑏𝑗𝑒𝑐𝑡) (2)"
},
{
"formula_coordinates": [
3,
350.68,
122.57,
159.54,
32.81
],
"formula_id": "formula_2",
"formula_text": "𝑖 |𝑂𝑏𝑗𝑒𝑐𝑡) * P r (Object) * IoU ( truth predict ) = 𝑃𝑟(𝐶𝑙𝑎𝑠𝑠 𝑖 ) * IoU ( truth predict ) (3)"
},
{
"formula_coordinates": [
4,
308.56,
108.49,
201.66,
63.49
],
"formula_id": "formula_3",
"formula_text": "𝑙 𝑐𝑜𝑜𝑟𝑑 = 𝜆 𝑐𝑜𝑜𝑟𝑑 ∑ ∑ 𝛪 𝑖𝑗 𝑜𝑏𝑗 [(𝑥 𝑖 -𝑥 ̂𝑖) 2 + (𝑦 𝑖 - 𝐵 𝑗=0 𝑆 2 𝑖=0 𝑦 ̂𝑖) 2 ] + 𝜆 𝑐𝑜𝑜𝑟𝑑 ∑ ∑ 𝛪 𝑖𝑗 𝑜𝑏𝑗 [(√𝑤 𝑖 -√𝑤 ̂𝑖) 2 + 𝐵 𝑗=0 𝑆 2 𝑖=0 (√ℎ 𝑖 -√ℎ ̂𝑖) 2 ] (5)"
},
{
"formula_coordinates": [
4,
330.16,
306.93,
180.06,
81.11
],
"formula_id": "formula_4",
"formula_text": "𝑙 𝐼𝑜𝑈 = 𝜆 𝐼𝑜𝑈 ∑ ∑ 𝛪 𝑖𝑗 𝑜𝑏𝑗 [(𝐶 𝑖 -𝐶 ̂𝑖) 2 ] 𝐵 𝑗=0 𝑆 2 𝑖=0 () + 𝜆 𝑛𝑜𝑜𝑏𝑗 ∑ ∑ 𝛪 𝑖𝑗 𝑛𝑜𝑜𝑏𝑗 (𝐶 𝑖 -𝐶 ̂𝑖) 2 𝐵 𝑗=0 𝑆 2 𝑖=0 (6) 𝑙 𝑐𝑙𝑎𝑠𝑠 = 𝜆 𝑐𝑙𝑎𝑠𝑠 ∑ 𝛪 𝑖 𝑜𝑏𝑗 ∑ (𝑝 𝑖 (𝑐) - 𝑐 ∈ 𝑐𝑙𝑎𝑠𝑠𝑒𝑠 𝑆 2 𝑖=0 𝑝̂𝑖(𝑐)) 2 (7)"
},
{
"formula_coordinates": [
5,
127.7,
582.04,
160.11,
47.06
],
"formula_id": "formula_5",
"formula_text": "Precision (P) = TP TP+FP (8) Recall (R) = TP TP+FN (9)"
}
] |
|
2023-05-21 | [
{
"figure_ref": [
"fig_0",
"fig_1",
"fig_1",
"fig_2"
],
"heading": "Introduction",
"publication_ref": [
"b6",
"b15",
"b33",
"b24",
"b28",
"b4",
"b27",
"b24",
"b3",
"b26",
"b22",
"b28",
"b28",
"b26"
],
"table_ref": [],
"text": "Transformers, which have gained far-flung fame in natural language processing (NLP) area [8,28], are also attracting increasing attention in lots of computer vision (CV) tasks, such as object detection [4], image classification [9] and many others [13,31], impelling the widespread research on vision transformers (ViTs). There has a natural fit for ViTs to achieve better performance simply by training a larger model on a larger data set. For example, historical records show better performance of a ViT-H model [9] accompanying with astonishing 632M parameters and 162G FLOPs. Such a high model complexity poses a great challenge to deploy models on platforms with short resource supplies. Therefore, both academia and industry call for an ultimate compression of these large models, and the DeiT-Small on ImageNet [17], respectively. Here 8-bit DeiT is quantized with PTQ method [22] and 2/3/4 bit DeiT is trained with QAT method [18]. The binarized ViT is conducted with the baseline method Bi-Real Net [26].\npast years have witnessed some promising techniques such as network pruning [38,5], low-rank decomposition [7], knowledge distillation [12], and quantization [18,19].\nNetwork quantization, which represents weights and activations in a low-bit format, has got great earnestness of many researchers for its reduced memory access costs and increased compute efficiency as well as performance benefit. Using the lower-bit quantized data, in particular to the extreme 1-bit case, requires less data movement, both on-chip and off-chip, and therefore reduces memory bandwidth and saves significant energy. Existing documentary records observe 32× less network size and 58× speedups beneficial from xnor and bit-count logics for 1-bit networks [30]. Earlier attempts [25,22] apply post-training quantization (PTQ) [1,40] directly to ViTs without datadriven fine-tuning, causing sub-optimal performance, in particular to impotent 1-bit ViTs. Therefore, by quantizing while training, quantization-aware training (QAT) methods are more congenial to 1-bit ViTs. Extensive empirical studies [24,20,36,28] have well demonstrated the efficacy of QAT methods in 1-bit convolutional neural networks (CNNs) or BERTs, however, the application to 1-bit ViTs remains not to be fully explored so far.\nIn this paper, we first build a fully-binarized ViT baseline, a straightforward solution constructed upon popular binarized QAT method of Bi-Real Net [26]. Through an empirical study of this baseline, we observe significant performance drops on the ImageNet dataset [17], as shown in Fig. 1. For instance, extending Bi-Real Net to binarize DeiT-Tiny [32] incurs a tremendous performance gap of 52.6% in the Top-1 accuracy compared to the 2-bit quantized counterpart. Similar performance drops occur in DeiT-Small as well. Delving into a deeper analysis, we find that the incompatibility of existing QAT methods mainly stems from the binarized self-attention module in ViTs, where a simple application of existing binarization methods [26] leads to severe attention distortion, as plotted in Fig. 2 (a) and Fig. 2 (b), especially in the diagonal scores of the map which are supposed to be the most attentive.\nIn this paper we dig deeper into this attention distortion problem. Through empirical analysis, we find that this phenomenon is mainly caused by gradient vanishing due to the straight-through-estimator (STE) [2] and non-scaled binarization in self-attention. Meanwhile, a simple distillation utilizing distillation token in DeiT [32] and KL-divergence in ReActNet [24] is ineffective in dismissing the ranking disorder, since it neglects the relative order of the attention map between the binarized ViTs and their real-valued counterpart. To address the aforementioned issues, a fullybinarized ViT (Bi-ViT) is developed by reactivating the vanished gradients through a learnable scaling factor in selfattention and a ranking-aware distillation to further effectively rectify the disordered ranking of attention (see the overview in Fig. 3). In addition, we also provide both empirical and theoretical analysis about how our method can rectify the distorted attention and thus promote the optimization of Bi-ViT. The contributions of our work are summarized as:\n• We identify the bottleneck of a fully-binarized ViT through empirical analyses and formulate the problem in a theoretical perspective. Based on these, we introduce learnable head-wise scaling factor into binarized self-attention to reactivate the vanished gradients.\n• We develop a ranking-aware distillation scheme to eliminate attention distortion. Our distillation method fully utilizes the ranking-aware knowledge from the real-valued teacher to promote the optimization of Bi-ViT.\n• Our Bi-ViT is the first promising way to push the limit of ViT quantization to the fully-binarized version. Extensive experiments on the ImageNet benchmark demonstrate that Bi-ViT surpasses both the baseline and prior binarized methods by a significant margin, achieving a remarkable acceleration rate of up to 61.5×. "
},
{
"figure_ref": [],
"heading": "Related Work",
"publication_ref": [
"b37",
"b31",
"b39",
"b8",
"b28",
"b26"
],
"table_ref": [],
"text": "Vision Transformer. Unlike traditional CNN-based models, ViTs are capable of capturing long-range visual relationships through the self-attention mechanism, and offer a more generalizable paradigm without inductive bias specific to images. The starting ViT [9] views an image as a sequence of 16 × 16 patches and uses a unique class token to predict the classification, yielding promising results. Subsequently, many works, such as DeiT [32] and PVT [35], have improved upon ViT, making it more efficient and applicable to downstream tasks. However, these high-performing ViTs have also accompanied with a significant number of parameters and high computational overhead, limiting their widespread applications. Thus, designing smaller and faster ViTs has become a new trend. LeViT [11] makes progress in faster inference through down-sampling, patch descriptors, and a redesign of the Attention-MLP block.\nDynamicViT [29] proposes a dynamic token sparsification framework to progressively and dynamically prune redundant tokens, achieving a competitive complexity and accuracy trade-off. Evo-ViT [37] proposes a slow-fast updating mechanism that ensures information flow and spatial structure, reducing both the training and inference complexity.\nWhile the aforementioned works focus on efficient model design, this paper aims to boost compression and acceleration through binarization. Network Binarization. BinaryNet is a technique originally proposed to train convolutional neural networks (CNNs) with binary weights. BinaryConnect [6] is the precursor to BinaryNet, where the parameters are binary while the activations remain in full-precision states. Local binary convolution layers (LBC) [16] were introduced to binarize the non-linear activations, and XNOR-Net [30] was introduced to improve convolution efficiency by binarizing the weights and inputs of convolution kernels. Bi-Real Net [26] explores a new variant of residual structure to preserve the information of real activations before the sign function, with a tight approximation to the derivative of the nondifferentiable sign function. Real-to-binary [27] re-scales the feature maps on the channels according to the input before binarized operations and adds an SE-Net [15] like gating module. ReActNet [24] replaces the conventional PReLU and the sign function of the BNNs with RPReLU and RSign with a learnable threshold, thus improving the performance of BNNs. RBONN [36] introduces a recurrent bilinear optimization to address the asynchronous convergence problem for BNNs, which further improves the performance of BNNs. These techniques improve the efficiency and accuracy of binary neural networks (BNNs) and allow them to be applied in practical applications. Majorities of these techniques consider non-scaled binarization in activations, which is beneficial to conventional CNNs while causing gradient mismatch issue for the pecularity of selfattention mechanism in ViTs."
},
{
"figure_ref": [],
"heading": "Background",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Multi-Head Self-Attention and Binarization",
"publication_ref": [
"b28"
],
"table_ref": [],
"text": "For a multi-head self-attention (MHSA) module, we denote its query, key, and value set as {a {q,k,v} ∈ R h×N ×d }, where h denotes head number, N and d represent the patch and channel numbers of each head. Specifically, N = (W in //W P in ) × (H in //H P in ) where W in and H in are the width and height of the feature, W P in , H P in are the width and height of patch maps respectively. Then, the attention score A and MHSA module output a out are computed as follows [34]:\nA = softmax[(a q • a k )/ √ d], a out = A • a v ,(1)\nwhere softmax(•) represents the softmax operation. Intending to represent query, key, value and attention score, i.e., a q , a k , a v and A, in a 1-bit format, Eq. (1) changes into:\nA = softmax[(b aq • b a k )/ √ d], a out = b A • b av .(2)\nWe follow the common network binarization methods [30] that use the sign function b • = sign(•) in the binary forward pass, and STE\n[2] ∂b• ∂• = 1 |•|≤1\nto compute the gradient for sign function in its backward pass. We omit the non-linear function here for simplicity. For all the projection and linear layers in binarized ViTs, we conduct binarization following [28,26] as\na out = b ain • (α w • b w ) = α w •(b ain •b w ) where α w = {α 1 w , α 2 w , ..., α Cout w } ∈ R Cout +\nis known as the channel-wise scaling factor vector [30] and • represents channel-wise channel-wise multiplication. The matrix multiplication process, i.e., b ain • b w , can be executed by the efficient XNOR and Bit-count instructions on edge devices."
},
{
"figure_ref": [],
"heading": "Bottleneck of Fully-Binarized ViTs",
"publication_ref": [],
"table_ref": [],
"text": "The high-performing ViTs are built on premise of transformer's supreme ability to model the long-range relationships thanks to the attention mechanism within the MHSA module. Unfortunately, a binarized version of weights and inputs significantly weakens the representation ability. In Module Degradation. By gradually replacing the multilayer perceptron (MLP) and MHSA modules with realvalued weights or activations, we have discovered that maintaining the MLP as \"w1a1\" (all weights and activations in the MLP are binarized) still results in satisfactory performance. For instance, keeping MLP as \"w1a1\" while keeping MHSA as \"w1a32\" obtains 26.3% Top-1 accuracy, which might be acceptable comparing to the 55.2% of realvalued DeiT-Tiny when taking into consideration 47.3× acceleration rates. On the contrast, when maintaining MHSA module as \"w1a1\", we observe a significant drop in performance. To be more specific, even when the MLP was maintained as \"w32a32\", we still observe a significant 50.8% decrease in Top-1 accuracy (from 55.2% to 4.4%). This result indicates that using binarized weights and activations in the MHSA module can have a substantial negative impact on the model's performance, even when other parts retain in real-valued states.\nOperation Degradation. To better understand the impact of fully-binarized ViT's performance, we conduct further analyses by examining the operations within the MHSA module. Specifically, when we maintain the self-attention activations in Eq. ( 1) as real-valued (\"a32\"), we observe only a relatively small decrease in performance from 48.8% to 37.6%. However, when the self-attention activations in Eq. (2) are binarized, significant drops in accuracy occur from 48.8% to 7.6%. This finding highlights the importance of the self-attention process within the MHSA module and suggests more efforts to mitigate the negative impact of binarization on the MHSA module."
},
{
"figure_ref": [
"fig_5"
],
"heading": "Gradient Mismatch in Self-Attention",
"publication_ref": [],
"table_ref": [],
"text": "With conclusion from the experimental results in Sec. 3.2 that self-attention process, i.e., Eq. (2), is the most critical part causing the performance drops. We attempt to analyze the underlying reasons for this phenomenon from an optimization perspective. For simplicity, we derive the gradient mismatch in a q as an example, and the analysis can be applicable to explain a k as well. We first represent the features before softmax(•) in Eq. (2) as:\np = (b aq • b a k .)/ √ d.(3)\nThe gradient of a hi,n,c q w.r.t. A is formulated as:\n∂A ∂a h i ,n,c q = ∂A ∂p h i ,n,n • ∂p h i ,n,n ∂b h i ,n,c aq • ∂b h i ,n,c aq ∂a h i ,n,c q , (4\n)\nwhere\nh i ∈ R h , n & n ∈ R N , c ∈ R d\nand the gradient of a k is likewise. The explicit form of the first item ∂A ∂p h i ,n,n in Eq. ( 4) is:\n∂A ∂p hi,n,n = ∂ softmax(p hi,n,n ) ∂p hi,n,n = A hi,n,n ⊗ (1 -A hi,n,n ),(5)\nwhere ⊗ denotes Hadamard product. And the second item is formulated as:\n∂p h i ,n,n ∂b h i ,n,c aq = ∂b h i ,n,c aq • b h i ,c,n a k ∂b h i ,n,c aq = b h i ,c,n a k ,(6)\nresult of which is therefore correlated with b a k . The third item is solved through STE [2] as:\n∂b hi,n,c aq ∂a hi,n,c q = 1 |a h i ,n,c q |≤1 .(7)\nCombing Eq. (5)-Eq. ( 7), we have the final gradient form in fully-binarized ViTs as:\n∂A ∂a h i ,n,c q = ∂A ∂p h i ,n,n • ∂p h i ,n,n ∂b h i ,n,c aq • ∂b h i ,n,c aq ∂a h i ,n,c q = A h i ,n,n (1 -A h i ,n,n ) • b h i ,c,n a k • 1 |a h i ,n,c q |≤1 .(8)\nConsidering b hi,n,:\naq = [1, • • • , 1] and •b hi,n ,: a k = [1, • • • , 1] as the extreme condition, b hi,n,: aq • b hi,:,n a k = d. Therefore, a specific element in b aq •b a k is ∈ {-d, • • • , d}.\nWe plot the curve of a specific element in the first item between [-64, 64] in Fig. 5 = 0, likewise for a k . Therefore we formulate the gradient mismatch phenomenon in the aforementioned theoretical analysis. And such gradient mismatch leads to distorted gradient in the optimization of a q & a k and therefore degrades performance of fully-binarized ViTs."
},
{
"figure_ref": [],
"heading": "Our Bi-ViT",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we propose to dismiss the affect of gradient mismatch mentioned in Sec. 3.3 from perspectives of gradient approximation in Sec. 4.1 and intermediate distillation in Sec. 4.2."
},
{
"figure_ref": [
"fig_5",
"fig_5"
],
"heading": "Learnable Head-wise Scaling Factor",
"publication_ref": [],
"table_ref": [],
"text": "As one of the solution to the above mentioned problem, we propose a head-wise scaling factor binarization scheme for the self-attention process, where the scaling factors are learned during training to first modify the gradient clip range in Fig. 5(b). Eq. ( 2) is changed into:\n= softmax(p), p = (α q ⊗ α k ) • (b aq • b a k )/ √ d = α q;k • (b aq • b a k )/ √ d,(9)\nand\nãout = (αA • bA) • (αv • ba v ) = (αA ⊗ αv) • (bA • b av ) = αA;v • (bA • b av ),(10)\nwhere b a• = sign( a• α• ), α q , α k , α v and α A are the headwise learnable scaling factors in binarized MHSA, where\nα {q,k,v,A} = {α 1 {q,k,v,A} , α 2 {q,k,v,A} , • • • , α h {q,k,v,A} } ∈ R h + .\nThe second rows in Eq. ( 9) & Eq. ( 10) are established since the scaling factors are aligned with the head dimension, which is independent with the matrix multiplication operation. Thus,\nα q;k = {α 1 q;k , α 2 q;k , • • • , α h q;k } ∈ R h + and α A;v = {α 1 A;v , α 2 A;v , • • • , α h A;v } ∈ R h + .\nConsequently, the gradient ∂ Ã ∂a :,n,c q in Eq. ( 8) is further formulated in our Bi-ViT as:\n∂ Ã ∂a h i ,n,c q = Ãh i ,n,n (1 -Ãh i ,n,n ) ∂ Ã ∂p h i ,n,n • α h i q;k • b h i ,c,n a k ∂p h i ,n,n ∂b h i ,n,c aq • 1 |a h i ,n,c q |≤αq ∂b h i ,n,c aq ∂a h i ,n,c q . (11\n)\nSince softmax(.) and • are aligned with different dimensions, the value of Eq. ( 5) remains unchanged (softmax(p) = softmax(α q;k • p)). As can be seen, the threshold of gradient clip in Eq. ( 7) changes from 1 into α q , which means that we can surpass the occurance of gradient mismatch by modifying the value of α q . Note that the scaling factor (α q ) is to imitate the magnitude of the latent activations. When p has a large magnitude, i.e., in the circled part of Fig. 5 (a), α q also tends to be larger and a hi,n,c q locates in the field that ∂b h i ,n,c aq ∂a h i ,n,c q > 0. Thus the vanishing gradients are reactivated through the introduced learnable scaling factor."
},
{
"figure_ref": [
"fig_1",
"fig_2"
],
"heading": "Ranking-aware Distillation for Bi-ViT",
"publication_ref": [],
"table_ref": [],
"text": "Fig. 2 illustrates a significant difference in the attention map's relative order between Bi-RealNet (a) and its realvalued counterpart (c). This difference could result in a notable decrease in performance. To address this issue during binarized training, a ranking-aware distillation in a teacherstudent framework is introduced:\nL ranking = L l=1 ψ(A T ) -ψ(A S ) 2 ,(12)\nwhere A T and A S represents the attention scores from the real-valued teacher and binarized student. ψ(•) denotes the function for obtaining the ranking, i.e., relative order of an attention score, which is formulated as:\nψ(A :,n,: ) = A :,n,: -A :,n-1,: , if 0 < n ≤ N -1 A :,0,: -A :,N -1,: , otherwise .(13)\nDetailed relative order computation can be seen in the right part of Fig. 3. We implement our Bi-ViT under the teacherstudent framework [32], thus the final objective of our method is formulated as:\nL = L dist + λL ranking ,(14)\nwhere λ is a hyper-parameter to balance these two loss functions."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [
"b16",
"b5",
"b28"
],
"table_ref": [],
"text": "In this section, we evaluate the performance of the proposed Bi-ViT model for image classification task using pop-ular DeiT [32] & Swin [23] backbones and object detection task using Mask R-CNN [14] & Cascade [3] Mask R-CNN with Swin-Tiny [23] backbone. To the best of our knowledge, there is no publicly available source codebase on fully-binarized ViTs at this point, so we implement the baseline i.e., Bi-Real Net [26] methods by ourselves."
},
{
"figure_ref": [],
"heading": "Datasets and Implementation Details",
"publication_ref": [
"b23",
"b23",
"b35",
"b12"
],
"table_ref": [],
"text": "Datasets. The experiments are carried out on the ImageNet ILSVRC12 dataset [17] for image classification task and COCO dataset [21] for object detection task. The ImageNet dataset is challenging due to its large scale and greater diversity. There are 1000 classes and 1.2 million training images, and 50k validation images in it. In our experiments, we use the classic data augmentation method described in [32].\nThe COCO [21] dataset includes images from 80 different categories. All of our COCO dataset experiments are performed on the object detection track of the COCO trainval35k training dataset, which consists of 80k images from the COCO train2014 dataset and 35k images sampled from the COCO val2014 dataset. We report the average precision (AP) for IoUs∈ [0.5:0.05:0.95], designated as mAP@[.5,.95], using COCO's standard evaluation metric. For further analyzing our method, we also report AP 50 , AP 75 , AP s , AP m , and AP l . Experimental settings. In our experiments, we initialize the weights of binarized model with the corresponding pretrained real-valued model. The binarized model is trained for 300 epochs with batch-size 512 and the base learning rate 5e -4. We do not use warm-up scheme. For all the experiments, we apply LAMB [39] optimizer with weight decay set as 0, following DeiT III [33]. Note that we keep the patch embedding (first) layer and the classification (last) layer as real-valued, following [10]. Backbone. We evaluate our binarized method on two popular vision transformer networks: DeiT [32] and Swin Transformer [23]. The DeiT-Tiny, DeiT-Small, DeiT-Base, Swin-Tiny and Swin-Small are adopted as the backbone models, whose Top-1 accuracy on ImageNet dataset are 72.2%, 79.9%, 81.8%, 81.2%, and 83.2% respectively. For a fair comparison, we utilize the official implementation of DeiT and Swin Transformer."
},
{
"figure_ref": [
"fig_6"
],
"heading": "Ablation Study",
"publication_ref": [
"b28"
],
"table_ref": [],
"text": "Hyper-parameter Selection. We λ of Eq. ( 14) in this part, with experiments conducted on ImageNet [17] dataset. We show the model performance (Top-1 accuracy) with different setups of hyper-parameter λ in Fig. 6, in which the performances increase first and then decrease with the uplift of λ from left to right. Since λ controls the importance of L ranking , we show that the vanilla baseline (λ = 0) performs worse than any versions with Ranking-aware Distillation loss (λ > 0), showing the proposed distillation Effectiveness of components. We conduct the ablative experiments regarding the proposed components on DeiT-Tiny network. Firstly, we compose the baseline network using the binarization method following Bi-Real Net [26]. As shown in the third row of Tab. 1, the baseline networks only obtains 6.6% Top-1 accuracy, which is far from satisfactory. With the introduction of our first novelty, i.e., learnable scaling factor (LSF), the baseline network is boosted by 17.8%, achieving 24.4% Top-1 accuracy. We also observe the other contribution Ranking-aware Disitllation (RD) singly promotes the baseline network by 5.9%, which is also significant on ImageNet dataset. By combining the two main contributions together, we get Bi-ViT, outperforming the vanilla baseline by 22.1%. + Ranking-aware Distillation (RD) 1-1 12.5 +5.9\n+ LSF + RD (Bi-ViT) 1-1 28.7 +22.1"
},
{
"figure_ref": [],
"heading": "Results on Image Classification",
"publication_ref": [
"b28",
"b28",
"b28",
"b16",
"b5",
"b23"
],
"table_ref": [],
"text": "The experimental results are shown in Tab. 2. We compare our method with 1-bit methods including BiB-ERT [28], RBONN [36], and Bi-Real Net [26] based on the same frameworks for the task of image classification with the ImageNet dataset. We also report the classification per- formance of the low-bit training-aware quantization method Q-ViT [18] for further reference. We use model size and OPs following [26] in comparison to other bit-widthe models for further reference. We firstly evaluate the proposed method on DeiT models.\nFor DeiT-Tiny backbone, compared with other binary methods, our Bi-ViT achieves significant performance improvements. For example, our Bi-ViT surpasses the base-line Bi-Real Net [26] by 22.1% Top-1 accuracy, which is significant and meaningful for real-world applications. And it is worth noting that the proposed 1-bit model significantly compresses the DeiT-Tiny by 61.5× on OPs. The proposed method also boosts the performance of baseline by 21.7% with the same architecture and bit-width using DeiT-Small bacobone, a significant improvement on the Im-ageNet dataset. For larger DeiT-B, as shown in Tab. 2, the Table 3. Experiments with Mask R-CNN [14] and Cascade R-CNN [3] using Swin [23] backbones on COCO [21]. \"#Bits\" denotes the bit-width of weights and activations. We report the AP (%) with different IoU threshold and AP for objects in various sizes. The bold denotes the best result with binarized weights and activations."
},
{
"figure_ref": [],
"heading": "Framework",
"publication_ref": [],
"table_ref": [],
"text": "Backbone Method # Bits Size "
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "In this paper, we present Bi-ViT, an improved version of fully-binarized ViTs that offers a high compression ratio and acceptable performance. Initially, we establish a empirical framework for fully-binarized ViT and analyze the bottlenecks of the baseline. Our empirical analysis shows that attention distortion in MHSA is the primary cause of the significant drop in ViT binarization, which results from gradient vanishing and ranking disorder. To address these issues, we introduce a learnable scaling factor that reactivates vanished gradients, which we illustrate through both theoretical and experimental analysis. Additionally, we propose ranking-aware distillation for Bi-ViT, which rectifies disordered ranking in a teacher-student framework. Our work provides a comprehensive analysis and effective solutions for the crucial issues in ViT full binarization, paving the way for the extreme compression of ViT."
}
] | Vision transformers (ViTs) quantization offers a promising prospect to facilitate deploying large pre-trained networks on resource-limited devices. Fully-binarized ViTs (Bi-ViT) that pushes the quantization of ViTs to its limit remain largely unexplored and a very challenging task yet, due to their unacceptable performance. Through extensive empirical analyses, we identify the severe drop in ViT binarization is caused by attention distortion in self-attention, which technically stems from the gradient vanishing and ranking disorder. To address these issues, we first introduce a learnable scaling factor to reactivate the vanished gradients and illustrate its effectiveness through theoretical and experimental analyses. We then propose a ranking-aware distillation method to rectify the disordered ranking in a teacher-student framework. Bi-ViT achieves significant improvements over popular DeiT and Swin backbones in terms of Top-1 accuracy and FLOPs. For example, with DeiT-Tiny and Swin-Tiny, our method significantly outperforms baselines by 22.1% and 21.4% respectively, while 61.5× and 56.1× theoretical acceleration in terms of FLOPs compared with real-valued counterparts on ImageNet. | Bi-ViT: Pushing the Limit of Vision Transformer Quantization | [
{
"figure_caption": "Figure 1 .1Figure 1. Performance of real-valued and quantized DeiT [32] with varying bit-widths. We report results with (a) DeiT-Tiny and (b)DeiT-Small on ImageNet [17], respectively. Here 8-bit DeiT is quantized with PTQ method[22] and 2/3/4 bit DeiT is trained with QAT method[18]. The binarized ViT is conducted with the baseline method Bi-Real Net[26].",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 .2Figure 2. Visualization of the attention map before softmax in the first block of DeiT-Tiny [32] on ImageNet [17]. From the left to right, is the baseline method [26], previous binarization method [36], our Bi-ViT and real-valued counterpart.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 .3Figure 3. Overview of the proposed Bi-ViT framework. We introduce the learnable scaling factor in an architecture perspective and a ranking-aware distillation scheme incorporated in the optimization process. From left to right, we respectively show the detailed architecture of single block in Bi-ViT and the distillation framework of Bi-ViT.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 .4Figure 4. Performance of fully-binarized DeiT-Tiny on Ima-geNet [17] with different binarized/real-valued settings.",
"figure_data": "",
"figure_id": "fig_4",
"figure_label": "4",
"figure_type": "figure"
},
{
"figure_caption": "Figure 5 .5Figure 5. Illustration of gradient mismatch between Eq. (5) and Eq. (7).",
"figure_data": "",
"figure_id": "fig_5",
"figure_label": "5",
"figure_type": "figure"
},
{
"figure_caption": "Figure 6 .6Figure 6. Effect of hyper-parameter λ on ImageNet [17].",
"figure_data": "",
"figure_id": "fig_6",
"figure_label": "6",
"figure_type": "figure"
},
{
"figure_caption": "Evaluating the components of Bi-ViT based on DeiT-Tiny [32] backbone. \"#Bits\" denotes the bit-width of weights and activations. We report the Top-1 (%) accuracy performances.",
"figure_data": "Method#Bits Top-1 (%)Real-valued32-32 72.1Baseline (Bi-Real Net [26])1-16.6+ Learnable Scaling Factor (LSF)1-124.4 +17.8",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Experiments withDeiT [32] and Swin[23] on ImageNet[17]. \"#Bits\" denotes the bit-width of weights and activations. We report the Top-1 (%) and Top-5 (%) accuracy performances. The bold denotes the best result with binarized weights and activations.",
"figure_data": "NetworkMethod#BitsSize (MB)OPs (10 8 )Top-1 (%)Top-5 (%)Real-valued32-3222.812.372.291.14-43.01.674.391.7Q-ViT [18]3-32.30.871.591.2DeiT-TinyBiBERT [28]2-21.70.459.0 5.981.8 16.0RBONN [36] Bi-Real Net [26]1-11.00.26.3 6.616.9 17.1Bi-ViT28.7 +22.151.7 +34.6Real-valued32-3288.245.579.995.04-411.45.880.994.9Q-ViT [18]3-38.73.079.094.2DeiT-SmallBiBERT [28]2-26.01.572.1 17.490.3 29.7RBONN [36] Bi-Real Net [26]1-13.40.818.5 19.230.0 30.3Bi-ViT40.9 +21.765.0 +34.7Real-valued32-32346.2174.781.895.64-444.122.083.096.1Q-ViT [18]3-333.411.181.095.1DeiT-BaseBiBERT [28]2-222.75.774.2 24.592.2 36.3RBONN [36] Bi-Real Net [26]1-112.12.926.1 26.538.6 38.8Bi-ViT47.3 +20.872.8 +34.0Real-valued32-32114.244.981.295.54-414.65.882.597.3Q-ViT [18]3-311.23.080.996.1Swin-TinyBiBERT [28]2-210.01.674.7 34.092.5 46.9RBONN [36] Bi-Real Net [26]1-14.20.833.8 34.146.7 46.9Bi-ViT55.5 +21.479.4 +32.5Real-valued32-32199.887.583.296.24-425.311.184.498.3Q-ViT [18]3-319.25.682.797.5Swin-SmallBiBERT [28]2-213.02.976.9 39.494.9 53.0RBONN [36] Bi-Real Net [26]1-16.91.539.0 39.252.7 52.8Bi-ViT60.7 +21.583.9 +31.1",
"figure_id": "tab_1",
"figure_label": "2",
"figure_type": "table"
},
{
"figure_caption": "(MB) AP AP 50 AP 75 AP s AP m AP l",
"figure_data": "Real-valued32-32 191.343.7 66.647.728.5 47.0 57.34-494.943.3 66.347.128.2 46.5 57.5Q-ViT [18]3-391.440.1 63.543.925.4 42.4 54.9Mask R-CNN [14] Swin-TinyBiBERT [28]2-288.030.2 53.7 8.9 25.033.4 8.615.2 32.0 45.2 1.7 9.0 15.9RBONN [36] Bi-Real Net [26]1-184.59.5 9.925.2 25.28.9 9.21.9 2.19.1 9.116.0 16.4Bi-ViT20.7 38.719.912.0 20.9 27.6Cascade Mask R-CNN [3]Swin-Tiny",
"figure_id": "tab_2",
"figure_label": "",
"figure_type": "table"
}
] | Yanjing Li; Sheng Xu; Mingbao Lin; Xianbin Cao; Chuanjian Liu; Xiao Sun; Baochang Zhang | [
{
"authors": "",
"journal": "Real-valued",
"ref_id": "b0",
"title": "",
"year": ""
},
{
"authors": "",
"journal": "",
"ref_id": "b1",
"title": "Bi-Real Net",
"year": ""
},
{
"authors": "",
"journal": "",
"ref_id": "b2",
"title": "To provide a basis for comparison, we also included the performance of 2/3/4-bit Q-ViT in Table 3. Our method outperform BiBERT, RBONN",
"year": null
},
{
"authors": "Ron Banner; Yury Nahshan; Daniel Soudry",
"journal": "",
"ref_id": "b3",
"title": "Post training 4-bit quantization of convolutional networks for rapiddeployment",
"year": "2019"
},
{
"authors": "Yoshua Bengio; Nicholas Léonard; Aaron Courville",
"journal": "",
"ref_id": "b4",
"title": "Estimating or propagating gradients through stochastic neurons for conditional computation",
"year": "2013"
},
{
"authors": "Zhaowei Cai; Nuno Vasconcelos",
"journal": "",
"ref_id": "b5",
"title": "Cascade r-cnn: Delving into high quality object detection",
"year": "2018"
},
{
"authors": "Nicolas Carion; Francisco Massa; Gabriel Synnaeve; Nicolas Usunier; Alexander Kirillov; Sergey Zagoruyko",
"journal": "",
"ref_id": "b6",
"title": "End-toend object detection with transformers",
"year": "2020"
},
{
"authors": "Mengzhao Chen; Mingbao Lin; Ke Li; Yunhang Shen; Yongjian Wu; Fei Chao; Rongrong Ji",
"journal": "",
"ref_id": "b7",
"title": "Cf-vit: A general coarse-to-fine method for vision transformer",
"year": "2023"
},
{
"authors": "Matthieu Courbariaux; Yoshua Bengio; Jean-Pierre David",
"journal": "",
"ref_id": "b8",
"title": "Binaryconnect: Training deep neural networks with binary weights during propagations",
"year": "2015"
},
{
"authors": "Misha Denil; Babak Shakibi; Laurent Dinh; Marc'aurelio Ranzato; Nando De Freitas",
"journal": "",
"ref_id": "b9",
"title": "Predicting parameters in deep learning",
"year": "2013"
},
{
"authors": "Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova",
"journal": "",
"ref_id": "b10",
"title": "Bert: Pre-training of deep bidirectional transformers for language understanding",
"year": "2018"
},
{
"authors": "Alexey Dosovitskiy; Lucas Beyer; Alexander Kolesnikov; Dirk Weissenborn; Xiaohua Zhai; Thomas Unterthiner; Mostafa Dehghani; Matthias Minderer; Georg Heigold; Sylvain Gelly",
"journal": "",
"ref_id": "b11",
"title": "An image is worth 16x16 words: Transformers for image recognition at scale",
"year": "2020"
},
{
"authors": "Jeffrey L Steven K Esser; Deepika Mckinstry; Rathinakumar Bablani; Dharmendra S Appuswamy; Modha",
"journal": "",
"ref_id": "b12",
"title": "Learned step size quantization",
"year": "2019"
},
{
"authors": "Benjamin Graham; Alaaeldin El-Nouby; Hugo Touvron; Pierre Stock; Armand Joulin; Hervé Jégou; Matthijs Douze",
"journal": "",
"ref_id": "b13",
"title": "Levit: a vision transformer in convnet's clothing for faster inference",
"year": "2021"
},
{
"authors": "Zhiwei Hao; Jianyuan Guo; Ding Jia; Kai Han; Yehui Tang; Chao Zhang; Han Hu; Yunhe Wang",
"journal": "",
"ref_id": "b14",
"title": "Learning efficient vision transformers via fine-grained manifold distillation",
"year": "2021"
},
{
"authors": "Kaiming He; Xinlei Chen; Saining Xie; Yanghao Li; Piotr Dollár; Ross Girshick",
"journal": "",
"ref_id": "b15",
"title": "Masked autoencoders are scalable vision learners",
"year": "2022"
},
{
"authors": "Kaiming He; Georgia Gkioxari",
"journal": "",
"ref_id": "b16",
"title": "Piotr Dollár, and Ross Girshick",
"year": "2017"
},
{
"authors": "Jie Hu; Li Shen; Gang Sun",
"journal": "",
"ref_id": "b17",
"title": "Squeeze-and-excitation networks",
"year": "2018"
},
{
"authors": "Felix Juefei-Xu; Vishnu Naresh Boddeti; Marios Savvides",
"journal": "",
"ref_id": "b18",
"title": "Local binary convolutional neural networks",
"year": "2017"
},
{
"authors": "Alex Krizhevsky; Ilya Sutskever; Geoffrey E Hinton",
"journal": "",
"ref_id": "b19",
"title": "Imagenet classification with deep convolutional neural networks",
"year": "2012"
},
{
"authors": "Yanjing Li; Sheng Xu; Baochang Zhang; Xianbin Cao; Peng Gao; Guodong Guo",
"journal": "",
"ref_id": "b20",
"title": "Q-vit: Accurate and fully quantized low-bit vision transformer",
"year": "2022"
},
{
"authors": "Zhexin Li; Tong Yang; Peisong Wang; Jian Cheng",
"journal": "",
"ref_id": "b21",
"title": "Qvit: Fully differentiable quantization for vision transformer",
"year": "2022"
},
{
"authors": "Mingbao Lin; Rongrong Ji; Zihan Xu; Baochang Zhang; Yan Wang; Yongjian Wu; Feiyue Huang; Chia-Wen Lin",
"journal": "",
"ref_id": "b22",
"title": "Rotated binary neural network",
"year": "2020"
},
{
"authors": "Tsung-Yi Lin; Michael Maire; Serge Belongie; James Hays; Pietro Perona; Deva Ramanan; Piotr Dollár; C Lawrence; Zitnick ",
"journal": "",
"ref_id": "b23",
"title": "Microsoft coco: Common objects in context",
"year": "2014"
},
{
"authors": "Yang Lin; Tianyu Zhang; Peiqin Sun; Zheng Li; Shuchang Zhou",
"journal": "",
"ref_id": "b24",
"title": "Fq-vit: Fully quantized vision transformer without retraining",
"year": "2022"
},
{
"authors": "Ze Liu; Yutong Lin; Yue Cao; Han Hu; Yixuan Wei; Zheng Zhang; Stephen Lin; Baining Guo",
"journal": "",
"ref_id": "b25",
"title": "Swin transformer: Hierarchical vision transformer using shifted windows",
"year": "2021"
},
{
"authors": "Zechun Liu; Zhiqiang Shen; Marios Savvides; Kwang-Ting Cheng",
"journal": "",
"ref_id": "b26",
"title": "Reactnet: Towards precise binary neural network with generalized activation functions",
"year": "2020"
},
{
"authors": "Zhenhua Liu; Yunhe Wang; Kai Han; Wei Zhang; Siwei Ma; Wen Gao",
"journal": "",
"ref_id": "b27",
"title": "Post-training quantization for vision transformer",
"year": "2021"
},
{
"authors": "Zechun Liu; Baoyuan Wu; Wenhan Luo; Xin Yang; Wei Liu; Kwang-Ting Cheng",
"journal": "",
"ref_id": "b28",
"title": "Bi-real net: Enhancing the performance of 1-bit cnns with improved representational capability and advanced training algorithm",
"year": "2007"
},
{
"authors": "Brais Martinez; Jing Yang; Adrian Bulat; Georgios Tzimiropoulos",
"journal": "",
"ref_id": "b29",
"title": "Training binary neural networks with real-tobinary convolutions",
"year": "2020"
},
{
"authors": "Yifu Haotong Qin; Mingyuan Ding; Qinghua Zhang; Aishan Yan; Qingqing Liu; Ziwei Dang; Xianglong Liu; Liu",
"journal": "",
"ref_id": "b30",
"title": "Bibert: Accurate fully binarized bert",
"year": "2022"
},
{
"authors": "Yongming Rao; Wenliang Zhao; Benlin Liu; Jiwen Lu; Jie Zhou; Cho-Jui Hsieh",
"journal": "",
"ref_id": "b31",
"title": "Dynamicvit: Efficient vision transformers with dynamic token sparsification",
"year": "2021"
},
{
"authors": "Mohammad Rastegari; Vicente Ordonez; Joseph Redmon; Ali Farhadi",
"journal": "",
"ref_id": "b32",
"title": "Xnor-net: Imagenet classification using binary convolutional neural networks",
"year": "2016"
},
{
"authors": "Yunjie Tian; Lingxi Xie; Zhaozhi Wang; Longhui Wei; Xiaopeng Zhang; Jianbin Jiao; Yaowei Wang; Qi Tian; Qixiang Ye",
"journal": "",
"ref_id": "b33",
"title": "Integrally pre-trained transformer pyramid networks",
"year": "2022"
},
{
"authors": "Hugo Touvron; Matthieu Cord; Matthijs Douze; Francisco Massa; Alexandre Sablayrolles; Hervé Jégou",
"journal": "",
"ref_id": "b34",
"title": "Training data-efficient image transformers & distillation through attention",
"year": "2007"
},
{
"authors": "Hugo Touvron; Matthieu Cord; Hervé Jégou",
"journal": "",
"ref_id": "b35",
"title": "Deit iii: Revenge of the vit",
"year": "2022"
},
{
"authors": "Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Łukasz Kaiser; Illia Polosukhin",
"journal": "",
"ref_id": "b36",
"title": "Attention is all you need",
"year": "2017"
},
{
"authors": "Wenhai Wang; Enze Xie; Xiang Li; Deng-Ping Fan; Kaitao Song; Ding Liang; Tong Lu; Ping Luo; Ling Shao",
"journal": "",
"ref_id": "b37",
"title": "Pyramid vision transformer: A versatile backbone for dense prediction without convolutions",
"year": "2021"
},
{
"authors": "Sheng Xu; Yanjing Li; Tiancheng Wang; Teli Ma; Baochang Zhang; Peng Gao; Yu Qiao; Jinhu Lü; Guodong Guo",
"journal": "",
"ref_id": "b38",
"title": "Recurrent bilinear optimization for binary neural networks",
"year": "2007"
},
{
"authors": "Yifan Xu; Zhijie Zhang; Mengdan Zhang; Kekai Sheng; Ke Li; Weiming Dong; Liqing Zhang; Changsheng Xu; Xing Sun",
"journal": "",
"ref_id": "b39",
"title": "Evo-vit: Slow-fast token evolution for dynamic vision transformer",
"year": "2022"
},
{
"authors": "Huanrui Yang; Hongxu Yin; Pavlo Molchanov; Hai Li; Jan Kautz",
"journal": "",
"ref_id": "b40",
"title": "Nvit: Vision transformer compression and parameter redistribution",
"year": "2021"
},
{
"authors": "Yang You; Jing Li; Sashank Reddi; Jonathan Hseu; Sanjiv Kumar; Srinadh Bhojanapalli; Xiaodan Song; James Demmel; Kurt Keutzer; Cho-Jui Hsieh",
"journal": "",
"ref_id": "b41",
"title": "Large batch optimization for deep learning: Training bert in 76 minutes",
"year": "2020"
},
{
"authors": "Yunshan Zhong; Mingbao Lin; Mengzhao Chen; Ke Li; Yunhang Shen; Fei Chao; Yongjian Wu; Rongrong Ji",
"journal": "",
"ref_id": "b42",
"title": "Finegrained data distribution alignment for post-training quantization",
"year": "2022"
}
] | [
{
"formula_coordinates": [
3,
366.84,
347.17,
178.27,
36.13
],
"formula_id": "formula_0",
"formula_text": "A = softmax[(a q • a k )/ √ d], a out = A • a v ,(1)"
},
{
"formula_coordinates": [
3,
362.63,
430.27,
182.48,
36.13
],
"formula_id": "formula_1",
"formula_text": "A = softmax[(b aq • b a k )/ √ d], a out = b A • b av .(2)"
},
{
"formula_coordinates": [
3,
401.07,
497.51,
64.29,
13.64
],
"formula_id": "formula_2",
"formula_text": "[2] ∂b• ∂• = 1 |•|≤1"
},
{
"formula_coordinates": [
3,
308.86,
547.34,
236.25,
22.62
],
"formula_id": "formula_3",
"formula_text": "a out = b ain • (α w • b w ) = α w •(b ain •b w ) where α w = {α 1 w , α 2 w , ..., α Cout w } ∈ R Cout +"
},
{
"formula_coordinates": [
4,
377.37,
191.2,
167.75,
20.27
],
"formula_id": "formula_4",
"formula_text": "p = (b aq • b a k .)/ √ d.(3)"
},
{
"formula_coordinates": [
4,
342.23,
245.1,
199.4,
24.74
],
"formula_id": "formula_5",
"formula_text": "∂A ∂a h i ,n,c q = ∂A ∂p h i ,n,n • ∂p h i ,n,n ∂b h i ,n,c aq • ∂b h i ,n,c aq ∂a h i ,n,c q , (4"
},
{
"formula_coordinates": [
4,
541.63,
254.24,
3.48,
7.77
],
"formula_id": "formula_6",
"formula_text": ")"
},
{
"formula_coordinates": [
4,
335.97,
281.07,
130.08,
11.23
],
"formula_id": "formula_7",
"formula_text": "h i ∈ R h , n & n ∈ R N , c ∈ R d"
},
{
"formula_coordinates": [
4,
347.26,
329.62,
197.85,
38.43
],
"formula_id": "formula_8",
"formula_text": "∂A ∂p hi,n,n = ∂ softmax(p hi,n,n ) ∂p hi,n,n = A hi,n,n ⊗ (1 -A hi,n,n ),(5)"
},
{
"formula_coordinates": [
4,
365.18,
410.74,
179.93,
43.06
],
"formula_id": "formula_9",
"formula_text": "∂p h i ,n,n ∂b h i ,n,c aq = ∂b h i ,n,c aq • b h i ,c,n a k ∂b h i ,n,c aq = b h i ,c,n a k ,(6)"
},
{
"formula_coordinates": [
4,
379.33,
496.14,
165.78,
27.74
],
"formula_id": "formula_10",
"formula_text": "∂b hi,n,c aq ∂a hi,n,c q = 1 |a h i ,n,c q |≤1 .(7)"
},
{
"formula_coordinates": [
4,
317.04,
567.05,
228.07,
54.41
],
"formula_id": "formula_11",
"formula_text": "∂A ∂a h i ,n,c q = ∂A ∂p h i ,n,n • ∂p h i ,n,n ∂b h i ,n,c aq • ∂b h i ,n,c aq ∂a h i ,n,c q = A h i ,n,n (1 -A h i ,n,n ) • b h i ,c,n a k • 1 |a h i ,n,c q |≤1 .(8)"
},
{
"formula_coordinates": [
4,
308.86,
623.4,
236.25,
41.56
],
"formula_id": "formula_12",
"formula_text": "aq = [1, • • • , 1] and •b hi,n ,: a k = [1, • • • , 1] as the extreme condition, b hi,n,: aq • b hi,:,n a k = d. Therefore, a specific element in b aq •b a k is ∈ {-d, • • • , d}."
},
{
"formula_coordinates": [
5,
101.31,
505.79,
185.05,
45.98
],
"formula_id": "formula_13",
"formula_text": "= softmax(p), p = (α q ⊗ α k ) • (b aq • b a k )/ √ d = α q;k • (b aq • b a k )/ √ d,(9)"
},
{
"formula_coordinates": [
5,
106.18,
570.36,
180.18,
40.45
],
"formula_id": "formula_14",
"formula_text": "ãout = (αA • bA) • (αv • ba v ) = (αA ⊗ αv) • (bA • b av ) = αA;v • (bA • b av ),(10)"
},
{
"formula_coordinates": [
5,
50.11,
638.72,
236.25,
26.6
],
"formula_id": "formula_15",
"formula_text": "α {q,k,v,A} = {α 1 {q,k,v,A} , α 2 {q,k,v,A} , • • • , α h {q,k,v,A} } ∈ R h + ."
},
{
"formula_coordinates": [
5,
50.11,
688.99,
236.25,
26.11
],
"formula_id": "formula_16",
"formula_text": "α q;k = {α 1 q;k , α 2 q;k , • • • , α h q;k } ∈ R h + and α A;v = {α 1 A;v , α 2 A;v , • • • , α h A;v } ∈ R h + ."
},
{
"formula_coordinates": [
5,
310.06,
104.28,
252.5,
59.27
],
"formula_id": "formula_17",
"formula_text": "∂ Ã ∂a h i ,n,c q = Ãh i ,n,n (1 -Ãh i ,n,n ) ∂ Ã ∂p h i ,n,n • α h i q;k • b h i ,c,n a k ∂p h i ,n,n ∂b h i ,n,c aq • 1 |a h i ,n,c q |≤αq ∂b h i ,n,c aq ∂a h i ,n,c q . (11"
},
{
"formula_coordinates": [
5,
541.38,
155.78,
3.73,
7.77
],
"formula_id": "formula_18",
"formula_text": ")"
},
{
"formula_coordinates": [
5,
349.59,
422.92,
195.52,
30.55
],
"formula_id": "formula_19",
"formula_text": "L ranking = L l=1 ψ(A T ) -ψ(A S ) 2 ,(12)"
},
{
"formula_coordinates": [
5,
315.14,
518.42,
229.97,
39.48
],
"formula_id": "formula_20",
"formula_text": "ψ(A :,n,: ) = A :,n,: -A :,n-1,: , if 0 < n ≤ N -1 A :,0,: -A :,N -1,: , otherwise .(13)"
},
{
"formula_coordinates": [
5,
377.57,
617.26,
167.54,
9.65
],
"formula_id": "formula_21",
"formula_text": "L = L dist + λL ranking ,(14)"
}
] |
|
2023-05-21 | [
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [],
"table_ref": [],
"text": "In the past decade, the development of information technology has led to a shift in the main carrier of information from text to images and then to videos. Moreover, with the rise of User-generated Content (UGC), the producer of information has shifted from Occupationally-generated Content (OGC) to UGC. As a result, a large number of videos have emerged on social media platforms and have been widely shared, leading to the increasingly important and challenging problems of video copyright protection. The core challenges of video copy detection are twofold: effective video descriptors and computational costs. Copied videos often involve edited portions, making it difficult for general visual models to differentiate between copied and original content. A powerful model must be capable of discriminating between videos, even when significant editing has taken place. Additionally, the cost of time and resources required to process each query video and identify the most similar reference video is a significant concern, necessitating the development of efficient and cost-effective methods.\nIn this paper, we summarize our proposed method for Meta AI Video Similarity Challenge, which tackle the core challenges of video copy detection through a dual-level detection method. In Fig. 1, there are three typical situations of query videos, including unedited video, copied video with general editing operation and copied video with multiple scenes in each frame. Our proposed dual-level detection method first identify if the video has been edited in video-level. For unedited videos, we use random vectors with small norm as their descriptors. What's more, we replace the bias term of these descriptors with a negative value during score normalization. For edited videos, we notice that it is necessary to deal with the situation that multiple scenes are concatenated along edge. We adopt traditional image processing method to detect and split the scenes in one frame. With our dual-level detection method, we can reduce the storage cost for unedited videos and improve the efficiency and accuracy of retrieval.\nThe main contributions are summarized as follows:\n• We propose a dual-level detection method for Descriptor Track, which detects edited videos at video-level and multiple scenes at frame-level. With the dual-level detection, we can reduce the computational cost and improve the performance.\n• The proposed method achieve outstanding performance on Meta AI Video Similarity Challenge and we got second prize on Descriptor Track. Our ablation study shows the effective of each module. For video with general editing, we extract features of each frame by our pre-trained basic model as descriptors. For video with multiple scenes, we detect and split the scenes in each frame, then use basic model to generate descriptors."
},
{
"figure_ref": [],
"heading": "Video Editing Detection",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Method",
"publication_ref": [],
"table_ref": [],
"text": "In this section, we first introduce the design of our basic model. Then we explain the details of our proposed dualdetection method including video editing detection, frame scenes detection and score filter normalization."
},
{
"figure_ref": [],
"heading": "Basic Model",
"publication_ref": [
"b7",
"b0",
"b7"
],
"table_ref": [],
"text": "We train a basic model to extract descriptors for video copy detection. There are two potential types of descriptors that can be employed in video copy detection: videolevel features or frame-level features. Given that our objective is not only to discriminate copied videos but also to identify copied portions between query and reference videos, we selected frame-level features as the video descriptor. Therefore, we adopt image transformer [4, 7] as backbone. We follow SSCD [8] to train our basic model in a self-supervised manner. As SSCD uses, we combine SimCLR [1] method with entropy loss [10]. The InfoNCE Loss. We use the InfoNCE loss in SimCLR, which is softmax cross-entropy loss with temprature. The loss function is formulated as follow:\nL InfoNCE = - 1 |P | i,j∈P log exp(cos(z i , z j )/τ ) Σ k =i exp(cos(z i , z k )/τ ) (1)\nWhere P is the set of positive pairs, z i represents the descriptor, τ is the temperature. The Entropy Loss. We follow SSCD using the entropy loss proposed in [10]. The loss function is formulated as:\nL KoLeo = - 1 N N i=1 log(min j =i ||z i -z j ||)(2)\nWhere N is the size of training set.\nThe Final Loss. The final loss function is:\nL = L InfoNCE + λL KoLeo (3\n)\nWhere λ is the weights of Entropy Loss term. Cause our training process of basic model is based on SSCD [8], more details can be found in this paper."
},
{
"figure_ref": [],
"heading": "Video Editing Detection",
"publication_ref": [
"b8"
],
"table_ref": [],
"text": "To address the issue of high computational costs in video copy detection and provide an efficient solution, we propose a straightforward method to identify edited videos before generating frame-level descriptors. We observe that videos with copies are often edited, incorporating techniques such as blending, blurring, rotations, and other manipulations. This is due to the fact that copied videos must frequently merge multiple clips, necessitating additional editing operations. By filtering out videos that have not been edited among the query videos, we can reduce computational costs. To achieve this goal, we have developed a model capable of discriminating between edited and unedited videos. Since editing operations can be viewed as strong augmentations, we aim to identify videos with such augmentations using a binary classification approach. We utilize CLIP [9] to extract frame features without any post-processing and feed these features into RoBERTa [6]. And we employ the class token to calculate cross entropy. We find that the edited video detection can achieve high accuracy, and using a small value α as threshold can filter most of unedited videos. For the unedited video, we use a random vector with very small value as descriptor. This processing can reduce the storage cost for query videos and speed up searching."
},
{
"figure_ref": [
"fig_2"
],
"heading": "Frame Scenes Detection",
"publication_ref": [],
"table_ref": [],
"text": "We notice that stacking multiple scenes in one frames is an obvious augmentation of copied videos, and simple traditional image processing method can deal with this situation well. Due to the continuity of the video, the combination of multiple scenes in one frames are limited. As shown in Fig. 3, scenes are usually concatenated along one side and one frame usually has an even number of scenes, most of them have two or four scenes. "
},
{
"figure_ref": [],
"heading": "Score Filter Normalization",
"publication_ref": [],
"table_ref": [],
"text": "We follow [5, 8] using similarity normalization in our evaluation. It introduce a background image dataset and only queries whose similarity score with reference is much higher than images in background dataset will have high scores. Based on it, we modify the integrated bias to suppress the score of unedited videos. In Sec. 2.2, we use a random vector with very small value as descriptor for unedited video, but scores of these videos are clustered around 0. Because scores of hard positive pairs are clustered around 0 too, we should further suppress the score of unedited videos. Inspired by the integrated bias term of similarity normalization, we can replace it by a negative value and the similarity score with any reference videos will reduce to a negative value."
},
{
"figure_ref": [],
"heading": "Experiments",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Implementation Details",
"publication_ref": [],
"table_ref": [],
"text": "In video editing detection, we adopt ViT-L-16 of CLIP to extract frame features. The initial weights for RoBERTa is chinese-roberta-wwm-ext [2, 3] in huggingface."
},
{
"figure_ref": [],
"heading": "Results",
"publication_ref": [],
"table_ref": [],
"text": "Our proposed method achieve outstanding performance on Meta AI Video Similarity Challenge. The results of Phase 1 and Phase 2 on Descriptor Tracks are presented in Tab. 1. On Descriptor Track, our method got second place Phases 1, just 0.021 away from the first place. Although we got the first place on Phase 2, we notice that the performance drop a lot when transfer the model to Phase 2. The reason is that we only ensemble 4 models and our ensemble results are not much better than single model. Without a strong ensemble method, the transfer ability is limited. "
},
{
"figure_ref": [],
"heading": "User or teams",
"publication_ref": [],
"table_ref": [],
"text": ""
},
{
"figure_ref": [],
"heading": "Ablation Study",
"publication_ref": [],
"table_ref": [],
"text": "To validate the effective of our proposed method, we split validation set by ourselves and analyze each module on validation set. At the beginning of the competition, we randomly divide queries in training set into 8:2 as offline training set and validation set. And the trend of performance on validation set can reflect the trend on test set. We use single basic model in ablation study because our ensemble method do not improve much. The results are shown in Tab. 2, our basic model can achieve 0.8580 on µAP , it shows that our basic model is a very strong baseline for video copy detection. Then combining frame scenes detection with basic model, the performance increased by 5%. And with the video editing detection and frame scenes detection, the performance achieve 0.9492. "
},
{
"figure_ref": [],
"heading": "Conclusion",
"publication_ref": [],
"table_ref": [],
"text": "We introduce a dual-detection method for Video Copy Detection in this paper. The video editing detection in video-level can identify unedited videos and use random vectors with small norm and negative bias term as descriptors. The frame scenes detection in frame-level can detect scenes and split them into multiple videos, where video only have one scenes in each frame. Thought the dual-detection method, we got second place on the Descriptor Track of Meta AI Video Similarity Challenge 2022. And our descriptor give strong support to our first-place solution on Matching Track."
}
] | Copy Detection has been a crucial problem for social media platforms. Meta AI hold Video Similarity Challenge on CVPR 2023 to push the technology forward. In this paper, we share our winner solutions on both tracks to help progress in this area. For Descriptor Track, we propose a dual-level detection method with Video Editing Detection (VED) and Frame Scenes Detection (FSD) to tackle the core challenges on Video Copy Detection. Experimental results demonstrate the effectiveness and efficiency of our proposed method. Code is available at https://github. | A Dual-level Detection Method for Video Copy Detection | [
{
"figure_caption": "Figure 1 .1Figure 1. Three typical situations of query videos. (a) is an unedited video, which is the most of query videos. (b) is a copied video with general editing operation. (c) is a copied video with multiple scenes in each frame.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 .2Figure2. Overview of our proposed pipeline. For unedited video, we directly use a random vector with samll norm and negative bias term. For video with general editing, we extract features of each frame by our pre-trained basic model as descriptors. For video with multiple scenes, we detect and split the scenes in each frame, then use basic model to generate descriptors.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 .3Figure 3. Multiple scenes in one frame.",
"figure_data": "",
"figure_id": "fig_2",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Leaderboard results on Descriptor Track. Bold indicates the best result and underline indicates the second best result.",
"figure_data": "Phase 1 µAP Phase 2 µAPdo something(Ours)0.91760.8717FriendshipFirst0.91970.8514cvl-descriptor0.85340.8362Zihao0.78410.7729",
"figure_id": "tab_0",
"figure_label": "1",
"figure_type": "table"
},
{
"figure_caption": "Ablation study.",
"figure_data": "MethodµAPBasic model 0.8580+ FSD0.9075+ VED0.9492",
"figure_id": "tab_1",
"figure_label": "2",
"figure_type": "table"
}
] | Tianyi Wang; Feipeng Ma; Zhenhua Liu; Fengyun Rao | [
{
"authors": "Ting Chen; Simon Kornblith; Mohammad Norouzi; Geoffrey Hinton",
"journal": "PMLR",
"ref_id": "b0",
"title": "A simple framework for contrastive learning of visual representations",
"year": "2020"
},
{
"authors": "Yiming Cui; Wanxiang Che; Ting Liu; Bing Qin; Shijin Wang; Guoping Hu",
"journal": "",
"ref_id": "b1",
"title": "Revisiting pre-trained models for Chinese natural language processing",
"year": "2020-11"
},
{
"authors": "Yiming Cui; Wanxiang Che; Ting Liu; Bing Qin; Ziqing Yang; Shijin Wang; Guoping Hu",
"journal": "",
"ref_id": "b2",
"title": "Pre-training with whole word masking for chinese bert",
"year": "2019"
},
{
"authors": "Alexey Dosovitskiy; Lucas Beyer; Alexander Kolesnikov; Dirk Weissenborn; Xiaohua Zhai; Thomas Unterthiner; Mostafa Dehghani; Matthias Minderer; Georg Heigold; Sylvain Gelly",
"journal": "",
"ref_id": "b3",
"title": "An image is worth 16x16 words: Transformers for image recognition at scale",
"year": "2020"
},
{
"authors": "Matthijs Douze; Giorgos Tolias; Ed Pizzi; Zoë Papakipos; Lowik Chanussot; Filip Radenovic; Tomas Jenicek; Maxim Maximov; Laura Leal-Taixé; Ismail Elezi",
"journal": "",
"ref_id": "b4",
"title": "The 2021 image similarity dataset and challenge",
"year": "2021"
},
{
"authors": "Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov",
"journal": "",
"ref_id": "b5",
"title": "Roberta: A robustly optimized bert pretraining approach",
"year": "2019"
},
{
"authors": "Ze Liu; Han Hu; Yutong Lin; Zhuliang Yao; Zhenda Xie; Yixuan Wei; Jia Ning; Yue Cao; Zheng Zhang; Li Dong; Furu Wei; Baining Guo",
"journal": "",
"ref_id": "b6",
"title": "Swin transformer v2: Scaling up capacity and resolution",
"year": "2022"
},
{
"authors": "Ed Pizzi; Dutta Sreya; Sugosh Roy; Priya Nagavara Ravindra; Matthijs Goyal; Douze",
"journal": "",
"ref_id": "b7",
"title": "A self-supervised descriptor for image copy detection",
"year": "2022"
},
{
"authors": "Alec Radford; Jong Wook Kim; Chris Hallacy; Aditya Ramesh; Gabriel Goh; Sandhini Agarwal; Girish Sastry; Amanda Askell; Pamela Mishkin; Jack Clark",
"journal": "PMLR",
"ref_id": "b8",
"title": "Learning transferable visual models from natural language supervision",
"year": "2021"
},
{
"authors": "Alexandre Sablayrolles; Matthijs Douze; Cordelia Schmid; Hervé Jégou",
"journal": "",
"ref_id": "b9",
"title": "Spreading vectors for similarity search",
"year": "2018"
}
] | [
{
"formula_coordinates": [
2,
60.79,
561.08,
225.57,
37.74
],
"formula_id": "formula_0",
"formula_text": "L InfoNCE = - 1 |P | i,j∈P log exp(cos(z i , z j )/τ ) Σ k =i exp(cos(z i , z k )/τ ) (1)"
},
{
"formula_coordinates": [
2,
88.78,
661.4,
197.58,
30.32
],
"formula_id": "formula_1",
"formula_text": "L KoLeo = - 1 N N i=1 log(min j =i ||z i -z j ||)(2)"
},
{
"formula_coordinates": [
2,
374.11,
304.2,
167.13,
9.65
],
"formula_id": "formula_2",
"formula_text": "L = L InfoNCE + λL KoLeo (3"
},
{
"formula_coordinates": [
2,
541.24,
304.52,
3.87,
8.64
],
"formula_id": "formula_3",
"formula_text": ")"
}
] |